Merge Collector transport with image-list values and native cache reuse

Use the existing images/audio Collector for typed serverless pools. Preserve saved legacy_png defaults, map them to serverless png, and retain every transfer format with exact opt-in. Remove the separate Media Collector.

Send direct exact frames through bounded raw tensor multipart transport without 8-bit fallback. Reject incompatible legacy pools, invalid payloads and incomplete exact results. Cover real CPU HTTP round trips, native ComfyUI execution, mixed participant ordering and failure paths.

README.md

@@ -27,6 +27,11 @@
 - Intelligent distribution
 - Handles single images and videos
 
+#### Dynamic Image-List Processing
+- Process a whole image list in parallel: the master and every worker take the next image when they free up, and results come back in input order
+- Use **Distributed Image List Start / End** around a per-image processing region
+- [Wiring, supported inputs, and first-version limits](/docs/image-list-queue.md)
+
 #### Ease of Use
 - Auto-setup local workers; easily add remote/cloud ones
 - Convert any workflow to distributed with 2 nodes
@@ -38,11 +43,12 @@
 
 <img width="200" align="right" alt="ComfyUI_temp_khvcc_00034_@0 25x" src="https://github.com/user-attachments/assets/651e4912-7c23-4e32-bd88-250f5175e129" />
 
-ComfyUI Distributed supports three types of workers:
+ComfyUI Distributed supports four types of workers:
 
 - **Local Workers** - Additional GPUs on the same machine (auto-configured on first launch)
 - **Remote Workers** - GPUs on other computers in your network
 - **Cloud Workers** - GPUs hosted on a cloud service like Runpod, accessible via secure tunnels
+- **Serverless (Runpod)** - Queue-based endpoints for image Collector jobs; no Distributed install or inbound tunnel on the endpoint. [Setup, endpoint contract, and preview limitations](/docs/serverless-workers.md).
 
 > For detailed setup instructions, see the [setup guide](/docs/worker-setup-guides.md)
 
@@ -178,6 +184,8 @@ Use **Distributed Value** when you want per-worker overrides (for example, diffe
 | **Audio Segment Divider** | Splits an audio waveform into up to ten sequential time segments |
 | **Distributed Model Name** | Passes model paths to workers, enabling workflows to use models not present on the master in orchestrator-only mode |
 | **Distributed Empty Image** | Produces an empty IMAGE batch used when the master delegates all work |
+| **Distributed Image List Start** | Turns an image list or batch into one job per image, scheduled to the next available participant; mask and value inputs are shared across images or paired one per image |
+| **Distributed Image List End** | Closes the paired region and returns images, masks, and texts as lists in the original input order |
 
 ---
 

api/__init__.py

@@ -1,5 +1,7 @@
 from . import config_routes  # noqa: F401
+from . import serverless_routes  # noqa: F401
 from . import tunnel_routes  # noqa: F401
 from . import worker_routes  # noqa: F401
 from . import job_routes  # noqa: F401
 from . import usdu_routes  # noqa: F401
+from . import image_list_routes  # noqa: F401

api/config_routes.py

@@ -1,5 +1,6 @@
 import json
 from contextlib import asynccontextmanager
+from typing import Any
 
 from aiohttp import web
 import server
@@ -24,6 +25,12 @@ except ImportError:
             save_config(config)
 from ..utils.logging import debug_log, log
 from ..utils.network import handle_api_error, normalize_host
+from ..utils.secret_fields import is_secret_field_name
+from ..utils.serverless_config import (
+    ServerlessSecretChanges,
+    public_config,
+    validate_serverless_worker,
+)
 
 
 def _positive_int(value):
@@ -86,22 +93,169 @@ def _apply_field_patch(target: dict, data: dict, field_rules: list) -> None:
             target[key] = normalizer(value) if (normalizer and value is not None) else value
 
 
+_SECRET_FIELDS = {"api_key", "clear_api_key"}
+_SERVERLESS_LOCAL_FIELDS = {"host", "port", "listen_port", "cuda_device", "extra_args"}
+_SERVERLESS_ONLY_FIELDS = {
+    "provider", "endpoint_id", "jobs_per_run", "concurrency", "timeout_seconds", "api_key_configured",
+    "media_profile",
+}
+_SERVERLESS_CONFIG_FIELDS = {
+    "worker_id", "type", "provider", "endpoint_id", "jobs_per_run", "concurrency",
+    "timeout_seconds", "enabled", "name", "api_key", "clear_api_key", "media_profile",
+}
+
+
+class _ActivePoolCheckError(RuntimeError):
+    pass
+
+
+def _error_response(message, status):
+    return web.json_response({"status": "error", "message": message}, status=status)
+
+
+def _contains_secret_fields(value, *, allow_root_serverless_secrets=False, depth=0):
+    if isinstance(value, dict):
+        for key, item in value.items():
+            field_name = key.casefold() if isinstance(key, str) else ""
+            if is_secret_field_name(field_name):
+                if not (allow_root_serverless_secrets and depth == 0 and key in _SECRET_FIELDS):
+                    return True
+            if _contains_secret_fields(
+                item, allow_root_serverless_secrets=allow_root_serverless_secrets, depth=depth + 1
+            ):
+                return True
+    elif isinstance(value, (list, tuple)):
+        return any(
+            _contains_secret_fields(
+                item, allow_root_serverless_secrets=allow_root_serverless_secrets, depth=depth + 1
+            )
+            for item in value
+        )
+    return False
+
+
+def _serverless_secret_store():
+    from .orchestration.serverless_runtime import get_secret_store
+    return get_secret_store()
+
+
+def _serverless_pool_is_active(worker_id):
+    from .orchestration.serverless_runtime import get_manager
+    return get_manager().pool_is_active(worker_id)
+
+
+def _ensure_pool_idle(worker_id):
+    try:
+        active = _serverless_pool_is_active(worker_id)
+    except Exception:
+        raise _ActivePoolCheckError from None
+    if active:
+        raise _ActivePoolCheckError
+
+
+def _worker_by_id(workers, worker_id):
+    for worker in workers:
+        if isinstance(worker, dict) and worker.get("id") == worker_id:
+            return worker
+    return None
+
+
+def _worker_type(worker):
+    return worker.get("type", "local") if isinstance(worker, dict) else None
+
+
+def _strip_sensitive_fields(value: Any) -> Any:
+    if isinstance(value, dict):
+        return {
+            key: _strip_sensitive_fields(item)
+            for key, item in value.items()
+            if not is_secret_field_name(key)
+        }
+    if isinstance(value, list):
+        return [_strip_sensitive_fields(item) for item in value]
+    return value
+
+
+def _clean_serverless_worker(worker):
+    clean = _strip_sensitive_fields(dict(worker))
+    for field in _SERVERLESS_LOCAL_FIELDS | {"api_key_configured"}:
+        clean.pop(field, None)
+    return clean
+
+
+def _validate_bulk_workers(workers):
+    validated = []
+    for worker in workers:
+        if isinstance(worker, dict) and worker.get("type") == "serverless":
+            clean = _clean_serverless_worker(worker)
+            validated.append(validate_serverless_worker(clean))
+        else:
+            validated.append(worker)
+    return validated
+
+
+def _clean_bulk_type_conversions(old_workers, new_workers):
+    old_by_id = {str(w.get("id")): w for w in old_workers if isinstance(w, dict) and w.get("id") is not None}
+    for index, worker in enumerate(new_workers):
+        if not isinstance(worker, dict) or worker.get("id") is None:
+            continue
+        old = old_by_id.get(str(worker["id"]))
+        if old is None or old.get("type") != "serverless":
+            continue
+        if worker.get("type") == "serverless":
+            # Older clients do not know this field. Only an explicit selection
+            # may change the endpoint contract, including on bulk saves.
+            if "media_profile" not in worker and "media_profile" in old:
+                worker["media_profile"] = old["media_profile"]
+            continue
+        converted = dict(worker)
+        for field in _SERVERLESS_ONLY_FIELDS:
+            converted.pop(field, None)
+        new_workers[index] = converted
+    return new_workers
+
+
+def _bulk_serverless_changes(old_workers, new_workers):
+    old_by_id = {str(w.get("id")): w for w in old_workers if isinstance(w, dict) and w.get("id") is not None}
+    new_by_id = {str(w.get("id")): w for w in new_workers if isinstance(w, dict) and w.get("id") is not None}
+    secret_ids = set()
+    for worker_id, old in old_by_id.items():
+        new = new_by_id.get(worker_id)
+        if old.get("type") == "serverless":
+            if new is None or new.get("type") != "serverless":
+                _ensure_pool_idle(worker_id)
+                secret_ids.add(worker_id)
+            elif (
+                old.get("endpoint_id") != new.get("endpoint_id")
+                or old.get("media_profile", "legacy-image-v1") != new.get("media_profile", "legacy-image-v1")
+            ):
+                _ensure_pool_idle(worker_id)
+        elif new is not None and new.get("type") == "serverless":
+            secret_ids.add(worker_id)
+    return secret_ids
+
+
 @server.PromptServer.instance.routes.get("/distributed/config")
 async def get_config_endpoint(request):
-    config = load_config()
-    return web.json_response(config)
+    try:
+        config = load_config()
+        return web.json_response(public_config(config, _serverless_secret_store()))
+    except Exception:
+        return _error_response("Unable to read config", 500)
 
 
 @server.PromptServer.instance.routes.post("/distributed/config")
 async def update_config_endpoint(request):
-    """Bulk config update with schema validation."""
+    """Bulk config update with schema validation and credential redaction."""
     try:
         data = await request.json()
-    except Exception as e:
-        return await handle_api_error(request, f"Invalid JSON payload: {e}", 400)
+    except Exception:
+        return _error_response("Invalid JSON payload", 400)
 
     if not isinstance(data, dict):
-        return await handle_api_error(request, "Config payload must be an object", 400)
+        return _error_response("Config payload must be an object", 400)
+    if _contains_secret_fields(data):
+        return _error_response("API keys cannot be set through bulk config", 400)
 
     validated_settings = {}
     validated_root = {}
@@ -118,9 +272,15 @@ async def update_config_endpoint(request):
             continue
 
         if validator and not validator(value):
-            errors.append(f"{key}: value {value!r} failed validation")
+            errors.append(f"{key}: failed validation")
             continue
 
+        if key == "workers":
+            try:
+                value = _validate_bulk_workers(value)
+            except (TypeError, ValueError):
+                return _error_response("Invalid serverless worker configuration", 400)
+
         if key in _SETTINGS_FIELDS:
             validated_settings[key] = value
         else:
@@ -133,15 +293,41 @@ async def update_config_endpoint(request):
             "message": "; ".join(errors),
         }, status=400)
 
+    store = None
+    response_config = None
     try:
-        async with config_transaction() as config:
-            settings = config.setdefault("settings", {})
-            settings.update(validated_settings)
-            for key, value in validated_root.items():
-                config[key] = value
-            return web.json_response({"status": "success", "config": config})
-    except Exception as e:
-        return await handle_api_error(request, e)
+        with ServerlessSecretChanges() as secret_changes:
+            async with config_transaction() as config:
+                if "workers" in validated_root:
+                    old_workers = config.get("workers", [])
+                    validated_root["workers"] = _clean_bulk_type_conversions(
+                        old_workers, validated_root["workers"]
+                    )
+                    changed_ids = _bulk_serverless_changes(old_workers, validated_root["workers"])
+                    for worker in validated_root["workers"]:
+                        if isinstance(worker, dict) and worker.get("type") == "serverless" and worker.get("enabled", False):
+                            store = store or _serverless_secret_store()
+                            if worker["id"] in changed_ids or not store.has(worker["id"]):
+                                return _error_response("Save an API key before enabling this worker", 400)
+                    if changed_ids:
+                        store = store or _serverless_secret_store()
+                        for worker_id in sorted(changed_ids):
+                            prior_secret = secret_changes.capture(store, worker_id)
+                            if prior_secret is not None:
+                                secret_changes.delete(store, worker_id)
+
+                settings = config.setdefault("settings", {})
+                settings.update(validated_settings)
+                for key, value in validated_root.items():
+                    config[key] = value
+                response_config = public_config(config, store or _serverless_secret_store())
+            response = web.json_response({"status": "success", "config": response_config})
+            secret_changes.commit()
+            return response
+    except _ActivePoolCheckError:
+        return _error_response("Finish or reconcile unresolved serverless jobs before changing this worker", 409)
+    except Exception:
+        return _error_response("Unable to update config", 500)
 
 
 @server.PromptServer.instance.routes.get("/distributed/queue_status/{job_id}")
@@ -164,79 +350,206 @@ async def queue_status_endpoint(request):
 
 @server.PromptServer.instance.routes.post("/distributed/config/update_worker")
 async def update_worker_endpoint(request):
+    return await _run_with_serverless_secret_changes(_update_worker_endpoint, request)
+
+
+async def _run_with_serverless_secret_changes(endpoint, request):
+    with ServerlessSecretChanges() as secret_changes:
+        response = await endpoint(request, secret_changes)
+        if response.status < 400:
+            secret_changes.commit()
+        return response
+
+
+async def _update_worker_endpoint(request, secret_changes):
+    store = None
+    worker_id = None
     try:
-        data = await request.json()
+        try:
+            data = await request.json()
+        except Exception:
+            return _error_response("Invalid JSON payload", 400)
+        if not isinstance(data, dict):
+            return _error_response("Worker payload must be an object", 400)
+
         worker_id = data.get("worker_id")
-        
         if worker_id is None:
-            return await handle_api_error(request, "Missing worker_id", 400)
+            return _error_response("Missing worker_id", 400)
 
+        if _contains_secret_fields(data, allow_root_serverless_secrets=True):
+            return _error_response("Invalid worker credentials", 400)
+
+        result = None
         async with config_transaction() as config:
-            worker_found = False
             workers = config.setdefault("workers", [])
+            existing = _worker_by_id(workers, worker_id)
+            current_type = _worker_type(existing)
+            target_type = data.get("type", current_type or "local")
 
-            for worker in workers:
-                if worker["id"] == worker_id:
-                    _apply_field_patch(worker, data, _WORKER_FIELDS)
-                    worker_found = True
-                    break
+            if target_type == "serverless":
+                unknown = set(data) - _SERVERLESS_CONFIG_FIELDS - _SERVERLESS_LOCAL_FIELDS
+                if unknown:
+                    return _error_response("Unknown serverless worker field", 400)
+                if not isinstance(worker_id, str) or not worker_id.strip():
+                    return _error_response("worker_id must be a non-empty string", 400)
+                if "api_key" in data and not isinstance(data["api_key"], str):
+                    return _error_response("Invalid worker credentials", 400)
+                if "clear_api_key" in data and type(data["clear_api_key"]) is not bool:
+                    return _error_response("Invalid worker credentials", 400)
+                supplied_key = data.get("api_key", "")
+                clear_key = data.get("clear_api_key", False)
+                if supplied_key and clear_key:
+                    return _error_response("Choose either api_key or clear_api_key", 400)
+                if "enabled" in data and type(data["enabled"]) is not bool:
+                    return _error_response("enabled must be a boolean", 400)
+                if "name" in data and not isinstance(data["name"], str):
+                    return _error_response("name must be a string", 400)
 
-            if not worker_found:
-                # If worker not found and all required fields are provided, create new worker
-                if all(key in data for key in ["name", "port", "cuda_device"]):
-                    new_worker = {
-                        "id": worker_id,
-                        "name": data["name"],
-                        "host": normalize_host(data.get("host", "localhost")),
-                        "port": data["port"],
-                        "cuda_device": data["cuda_device"],
-                        "enabled": data.get("enabled", False),
-                        "extra_args": data.get("extra_args", ""),
-                        "type": data.get("type", "local")
-                    }
-                    workers.append(new_worker)
+                candidate = dict(existing or {})
+                candidate["id"] = worker_id
+                candidate["type"] = "serverless"
+                candidate.setdefault("provider", "runpod")
+                candidate.setdefault("enabled", False)
+                # Legacy "concurrency" is accepted (see _SERVERLESS_CONFIG_FIELDS) but ignored.
+                for field in ("provider", "endpoint_id", "jobs_per_run", "timeout_seconds", "enabled", "name", "media_profile"):
+                    if field in data:
+                        candidate[field] = data[field]
+                candidate = _clean_serverless_worker(candidate)
+                validated = validate_serverless_worker(candidate)
+                if "enabled" in validated and type(validated["enabled"]) is not bool:
+                    return _error_response("enabled must be a boolean", 400)
+                if "name" in validated and not isinstance(validated["name"], str):
+                    return _error_response("name must be a string", 400)
+
+                store = _serverless_secret_store()
+                prior_secret = secret_changes.capture(store, worker_id)
+                converting_to_serverless = current_type != "serverless"
+                next_secret = None if converting_to_serverless else prior_secret
+                if supplied_key:
+                    next_secret = supplied_key
+                elif clear_key:
+                    next_secret = None
+
+                endpoint_changed = (
+                    current_type == "serverless"
+                    and existing.get("endpoint_id") != validated.get("endpoint_id")
+                )
+                media_profile_changed = (
+                    current_type == "serverless"
+                    and existing.get("media_profile", "legacy-image-v1") != validated.get("media_profile", "legacy-image-v1")
+                )
+                if current_type == "serverless" and (
+                    endpoint_changed or media_profile_changed or next_secret != prior_secret
+                ):
+                    _ensure_pool_idle(worker_id)
+
+                if validated.get("enabled", False) and next_secret is None:
+                    return _error_response("Save an API key before enabling this worker; turn the worker off before clearing its key", 400)
+
+                if next_secret != prior_secret:
+                    if next_secret is None:
+                        secret_changes.delete(store, worker_id)
+                    else:
+                        secret_changes.set(store, worker_id, next_secret)
+
+                if existing is None:
+                    workers.append(validated)
                 else:
-                    return await handle_api_error(
-                        request,
-                        f"Worker {worker_id} not found and missing required fields for creation",
-                        404,
-                    )
+                    existing.clear()
+                    existing.update(validated)
+                result = {"status": "success", "api_key_configured": next_secret is not None}
+
+            elif existing is not None:
+                cleaned_existing = _strip_sensitive_fields(existing)
+                existing.clear()
+                existing.update(cleaned_existing)
+                if current_type == "serverless":
+                    if _contains_secret_fields(data):
+                        return _error_response("Invalid worker credentials", 400)
+                    _ensure_pool_idle(worker_id)
+                    store = _serverless_secret_store()
+                    prior_secret = secret_changes.capture(store, worker_id)
+                    if prior_secret is not None:
+                        secret_changes.delete(store, worker_id)
+                    for field in _SERVERLESS_ONLY_FIELDS:
+                        existing.pop(field, None)
+                elif _contains_secret_fields(data):
+                    return _error_response("API keys are only accepted for serverless workers", 400)
+                _apply_field_patch(existing, data, _WORKER_FIELDS)
+                result = {"status": "success"}
+            elif all(key in data for key in ["name", "port", "cuda_device"]):
+                if _contains_secret_fields(data):
+                    return _error_response("API keys are only accepted for serverless workers", 400)
+                new_worker = {
+                    "id": worker_id,
+                    "name": data["name"],
+                    "host": normalize_host(data.get("host", "localhost")),
+                    "port": data["port"],
+                    "cuda_device": data["cuda_device"],
+                    "enabled": data.get("enabled", False),
+                    "extra_args": data.get("extra_args", ""),
+                    "type": data.get("type", "local")
+                }
+                workers.append(new_worker)
+                result = {"status": "success"}
+            else:
+                return _error_response(
+                    "Worker not found and missing required fields for creation", 404
+                )
+
+        return web.json_response(result)
+    except _ActivePoolCheckError:
+        return _error_response("Finish or reconcile unresolved serverless jobs before changing this worker", 409)
+    except ValueError:
+        return _error_response("Invalid worker configuration", 400)
+    except Exception:
+        return _error_response("Unable to update worker configuration", 500)
 
-            return web.json_response({"status": "success"})
-    except Exception as e:
-        return await handle_api_error(request, e, 400)
 
 @server.PromptServer.instance.routes.post("/distributed/config/delete_worker")
 async def delete_worker_endpoint(request):
+    return await _run_with_serverless_secret_changes(_delete_worker_endpoint, request)
+
+
+async def _delete_worker_endpoint(request, secret_changes):
+    store = None
+    worker_id = None
     try:
-        data = await request.json()
+        try:
+            data = await request.json()
+        except Exception:
+            return _error_response("Invalid JSON payload", 400)
+        if not isinstance(data, dict):
+            return _error_response("Worker payload must be an object", 400)
         worker_id = data.get("worker_id")
-        
         if worker_id is None:
-            return await handle_api_error(request, "Missing worker_id", 400)
-            
+            return _error_response("Missing worker_id", 400)
+
+        result = None
         async with config_transaction() as config:
             workers = config.get("workers", [])
-
-            # Find and remove the worker
-            worker_index = -1
-            for i, worker in enumerate(workers):
-                if worker["id"] == worker_id:
-                    worker_index = i
-                    break
-
-            if worker_index == -1:
-                return await handle_api_error(request, f"Worker {worker_id} not found", 404)
-
-            # Remove the worker
-            removed_worker = workers.pop(worker_index)
-
-            return web.json_response({
+            removed = _worker_by_id(workers, worker_id)
+            if removed is None:
+                return _error_response("Worker not found", 404)
+            if _worker_type(removed) == "serverless":
+                _ensure_pool_idle(worker_id)
+                store = _serverless_secret_store()
+                secret_worker_id = str(worker_id)
+                prior_secret = secret_changes.capture(store, secret_worker_id)
+                if prior_secret is not None:
+                    secret_changes.delete(store, secret_worker_id)
+            workers.remove(removed)
+            result = {
                 "status": "success",
-                "message": f"Worker {removed_worker.get('name', worker_id)} deleted"
-            })
-    except Exception as e:
-        return await handle_api_error(request, e, 400)
+                "message": "Worker deleted" if _worker_type(removed) == "serverless"
+                else f"Worker {removed.get('name', worker_id)} deleted",
+            }
+
+        return web.json_response(result)
+    except _ActivePoolCheckError:
+        return _error_response("Finish or reconcile unresolved serverless jobs before changing this worker", 409)
+    except Exception:
+        return _error_response("Unable to delete worker", 500)
 
 @server.PromptServer.instance.routes.post("/distributed/config/update_setting")
 async def update_setting_endpoint(request):

api/image_list_routes.py

@@ -0,0 +1,92 @@
+"""Bounded, typed HTTP transport for dynamic image-list worker tasks."""
+import json
+
+from aiohttp import web
+import server
+
+from ..image_list.runtime import (
+    MAX_REQUEST_BYTES,
+    PROTOCOL_VERSION,
+    decode_result,
+    delete_worker_item,
+    set_master_result,
+    store_worker_item,
+)
+
+prompt_server = server.PromptServer.instance
+
+
+async def _json_body(request):
+    length = request.content_length
+    if length is not None and length > MAX_REQUEST_BYTES:
+        raise ValueError(f"Request exceeds the {MAX_REQUEST_BYTES}-byte image-list limit")
+    raw = bytearray()
+    async for chunk in request.content.iter_chunked(64 * 1024):
+        raw.extend(chunk)
+        if len(raw) > MAX_REQUEST_BYTES:
+            raise ValueError(f"Request exceeds the {MAX_REQUEST_BYTES}-byte image-list limit")
+    try:
+        return json.loads(raw.decode("utf-8"))
+    except (UnicodeDecodeError, json.JSONDecodeError) as exc:
+        raise ValueError(f"Invalid image-list JSON body: {exc}") from exc
+
+
+@prompt_server.routes.get("/distributed/image-list/capabilities")
+async def image_list_capabilities_endpoint(_request):
+    return web.json_response({
+        "supported": True,
+        "execution_targets": True,
+        "value_results": True,
+        "protocol_version": PROTOCOL_VERSION,
+        "node_ids": ["DistributedImageListStart", "DistributedImageListEnd", "DistributedImageListGetValue"],
+        "transport": "typed-float-tensor-json-v1",
+    })
+
+
+@prompt_server.routes.post("/distributed/image-list/item")
+async def image_list_item_endpoint(request):
+    try:
+        data = await _json_body(request)
+        if not isinstance(data, dict):
+            raise ValueError("Expected an image-list item object")
+        context_id = data.get("context_id")
+        if not isinstance(context_id, str) or not context_id:
+            raise ValueError("context_id must be a non-empty string")
+        item = data.get("item")
+        identity = store_worker_item(data.get("token"), item)
+        return web.json_response({"status": "stored", "item_index": item["index"],
+                                  "item_identity": identity})
+    except (ValueError, TypeError) as exc:
+        return web.json_response({"error": str(exc)}, status=400)
+    except Exception as exc:
+        return web.json_response({"error": str(exc)}, status=500)
+
+
+@prompt_server.routes.delete("/distributed/image-list/item/{token}")
+async def image_list_delete_item_endpoint(request):
+    token = request.match_info.get("token", "")
+    return web.json_response({"status": "deleted", "removed": delete_worker_item(token)})
+
+
+@prompt_server.routes.post("/distributed/image-list/result")
+async def image_list_result_endpoint(request):
+    try:
+        data = await _json_body(request)
+        if not isinstance(data, dict):
+            raise ValueError("Expected an image-list result object")
+        context_id = data.get("context_id")
+        token = data.get("token")
+        item_index = data.get("item_index")
+        if not isinstance(context_id, str) or not context_id:
+            raise ValueError("context_id must be a non-empty string")
+        if not isinstance(token, str) or not token:
+            raise ValueError("token must be a non-empty string")
+        if isinstance(item_index, bool) or not isinstance(item_index, int) or item_index < 0:
+            raise ValueError("item_index must be a non-negative integer")
+        result = decode_result(data.get("result"))
+        set_master_result(context_id, item_index, token, result)
+        return web.json_response({"status": "accepted", "item_index": item_index})
+    except (ValueError, TypeError) as exc:
+        return web.json_response({"error": str(exc)}, status=400)
+    except Exception as exc:
+        return web.json_response({"error": str(exc)}, status=500)

api/job_routes.py

@@ -37,6 +37,8 @@ prompt_server = server.PromptServer.instance
 #   envelope: JSON { "job_id", "worker_id", "format", "start_index", "count", "width", "height",
 #                    "is_last", "audio"?: {...} }
 #   media:    one part per frame for still formats, or one video file for video formats
+#   format "exact" adds envelope "descriptors" (one tensor-le-v1 descriptor per
+#   frame) and sends one raw tensor part per frame; values are not quantized.
 
 
 def _decode_image_sync(image_path):
@@ -299,6 +301,8 @@ async def _enqueue_worker_items(multi_job_id, items):
 
 def _decode_compressed_chunk_sync(envelope, parts):
     """Decode a compressed chunk into per-frame IMAGE tensors (threadpool)."""
+    if envelope.get("format") == transfer_codec.EXACT_FORMAT:
+        return _decode_exact_chunk_sync(envelope, parts)
     frames = transfer_codec.decode_chunk(envelope, parts)
     return [
         ensure_contiguous(torch.from_numpy(transfer_codec.uint8_frame_to_tensor_array(frame)))
@@ -306,6 +310,35 @@ def _decode_compressed_chunk_sync(envelope, parts):
     ]
 
 
+def _decode_exact_chunk_sync(envelope, parts):
+    """Decode raw exact IMAGE frames, checking each against the envelope geometry."""
+    from ..serverless_bridge import codec as bridge_codec
+    count, width, height = envelope.get("count"), envelope.get("width"), envelope.get("height")
+    for name, value in (("count", count), ("width", width), ("height", height)):
+        if not isinstance(value, int) or isinstance(value, bool) or value <= 0:
+            raise transfer_codec.TransferCodecError(f"envelope.{name}: expected positive integer")
+    descriptors = envelope.get("descriptors")
+    if not isinstance(descriptors, list) or len(descriptors) != count or len(parts) != count:
+        raise transfer_codec.TransferCodecError(
+            f"exact chunk expected {count} descriptors and frames, got "
+            f"{len(descriptors) if isinstance(descriptors, list) else 'no'} descriptors and {len(parts)} frames"
+        )
+    tensors = []
+    for descriptor, payload in zip(descriptors, parts):
+        if not isinstance(descriptor, dict) or descriptor.get("encoding") != "tensor-le-v1":
+            raise transfer_codec.TransferCodecError("exact frames require tensor-le-v1 encoding")
+        shape = descriptor.get("shape") if isinstance(descriptor, dict) else None
+        if not isinstance(shape, list) or len(shape) != 4 or shape[:3] != [1, height, width]:
+            raise transfer_codec.TransferCodecError(
+                f"exact frame shape {shape} does not match the envelope's 1x{height}x{width} frames"
+            )
+        try:
+            tensors.append(bridge_codec.decode_payload("IMAGE", descriptor, payload))
+        except ValueError as exc:
+            raise transfer_codec.TransferCodecError(f"exact frame: {exc}") from None
+    return tensors
+
+
 def _validate_compressed_envelope(envelope):
     errors = []
     if not isinstance(envelope, dict):
@@ -384,6 +417,7 @@ async def _job_complete_compressed(request):
             "image_index": start_index + pos,
             "is_last": bool(is_last and pos == last_pos),
             "audio": decoded_audio if pos == last_pos else None,
+            "format": envelope["format"],
         }
         for pos, tensor in enumerate(tensors)
     ]
@@ -462,6 +496,7 @@ async def job_complete_endpoint(request):
                     "image_index": int(batch_idx),
                     "is_last": is_last,
                     "audio": decoded_audio,
+                    "format": transfer_codec.LEGACY_FORMAT,
                 }
             ],
         )

api/orchestration/dispatch.py

@@ -6,6 +6,9 @@ import aiohttp
 
 from ...utils.logging import debug_log, log
 from ...utils.network import build_worker_url, get_client_session, probe_worker
+# Must match image_list.runtime.PROTOCOL_VERSION (kept local so this module
+# stays importable without the image-list runtime; a test pins them equal).
+IMAGE_LIST_PROTOCOL_VERSION = 2
 try:
     from ...utils.trace_logger import trace_debug, trace_info
 except ImportError:
@@ -55,6 +58,11 @@ async def worker_ws_is_active(worker):
 
 async def _probe_worker_active(worker, use_websocket, semaphore):
     async with semaphore:
+        if worker.get("type") == "serverless":
+            from .serverless_runtime import get_manager
+            # Authentication/configuration failures must surface, not silently skip paid work.
+            status = await get_manager().health(worker)
+            return worker, status["ok"]
         is_active = await (worker_ws_is_active(worker) if use_websocket else worker_is_active(worker))
         return worker, is_active
 
@@ -69,30 +77,49 @@ async def _dispatch_via_websocket(worker_url, payload, client_id, timeout=60.0):
         "workflow": payload.get("workflow"),
         "client_id": client_id,
     }
+    if "partial_execution_targets" in payload:
+        ws_payload["partial_execution_targets"] = payload["partial_execution_targets"]
     ws_url = worker_url.replace("http://", "ws://").replace("https://", "wss://")
     ws_url = f"{ws_url}/distributed/worker_ws"
     session = await get_client_session()
+    loop = asyncio.get_running_loop()
+    deadline = loop.time() + timeout
+    ws = None
 
-    async with session.ws_connect(ws_url, heartbeat=20, timeout=timeout) as ws:
-        await ws.send_json(ws_payload)
-        async for msg in ws:
-            if msg.type == aiohttp.WSMsgType.TEXT:
-                data = json.loads(msg.data or "{}")
-                if data.get("type") == "dispatch_ack" and data.get("request_id") == request_id:
-                    if data.get("ok"):
-                        return
-                    error_text = data.get("error") or "Worker rejected websocket dispatch."
-                    validation_error = data.get("validation_error")
-                    node_errors = data.get("node_errors")
-                    if validation_error:
-                        error_text = f"{error_text} | validation_error={validation_error}"
-                    if node_errors:
-                        error_text = f"{error_text} | node_errors={node_errors}"
-                    raise RuntimeError(error_text)
-            elif msg.type in (aiohttp.WSMsgType.ERROR, aiohttp.WSMsgType.CLOSED):
-                raise RuntimeError(f"Worker websocket closed unexpectedly: {msg.type}")
-
-    raise RuntimeError("Worker websocket closed before dispatch_ack was received.")
+    try:
+        async with asyncio.timeout_at(deadline):
+            ws = await session.ws_connect(ws_url, heartbeat=20)
+            await ws.send_json(ws_payload)
+            while True:
+                msg = await ws.receive()
+                if msg.type == aiohttp.WSMsgType.TEXT:
+                    data = json.loads(msg.data or "{}")
+                    if data.get("type") == "dispatch_ack" and data.get("request_id") == request_id:
+                        if data.get("ok"):
+                            return data.get("prompt_id")
+                        error_text = data.get("error") or "Worker rejected websocket dispatch."
+                        validation_error = data.get("validation_error")
+                        node_errors = data.get("node_errors")
+                        if validation_error:
+                            error_text = f"{error_text} | validation_error={validation_error}"
+                        if node_errors:
+                            error_text = f"{error_text} | node_errors={node_errors}"
+                        raise RuntimeError(error_text)
+                elif msg.type in (aiohttp.WSMsgType.ERROR, aiohttp.WSMsgType.CLOSE, aiohttp.WSMsgType.CLOSED):
+                    raise RuntimeError(f"Worker websocket closed before acknowledgement: {msg.type}; submission status is ambiguous")
+    except asyncio.TimeoutError as exc:
+        raise RuntimeError(
+            "Worker websocket dispatch timed out before dispatch_ack; submission status is ambiguous."
+        ) from exc
+    finally:
+        if ws is not None:
+            # Closing must not discard an already acknowledged prompt ID or
+            # replace a definite rejection with an ambiguous timeout.
+            try:
+                async with asyncio.timeout_at(min(deadline, loop.time() + 1.0)):
+                    await ws.close()
+            except (asyncio.TimeoutError, aiohttp.ClientError):
+                pass
 
 
 async def dispatch_worker_prompt(
@@ -102,11 +129,18 @@ async def dispatch_worker_prompt(
     client_id=None,
     use_websocket=False,
     trace_execution_id=None,
+    return_prompt_id=False,
+    *,
+    partial_execution_targets=None,
 ):
     """Send the prepared prompt to a worker ComfyUI instance."""
     worker_url = build_worker_url(worker)
     url = build_worker_url(worker, "/prompt")
     payload = {"prompt": prompt_obj}
+    target_payload = {}
+    if partial_execution_targets is not None:
+        target_payload["partial_execution_targets"] = list(partial_execution_targets)
+        payload.update(target_payload)
     extra_data = {}
     if workflow_meta:
         extra_data.setdefault("extra_pnginfo", {})["workflow"] = workflow_meta
@@ -115,15 +149,18 @@ async def dispatch_worker_prompt(
 
     if use_websocket:
         try:
-            await _dispatch_via_websocket(
+            prompt_id = await _dispatch_via_websocket(
                 worker_url,
                 {
                     "prompt": prompt_obj,
                     "workflow": workflow_meta,
+                    **target_payload,
                 },
                 client_id,
             )
-            return
+            if return_prompt_id and (not isinstance(prompt_id, str) or not prompt_id):
+                raise RuntimeError("Worker websocket acknowledgement omitted prompt_id")
+            return prompt_id if return_prompt_id else None
         except Exception as exc:
             worker_id = worker.get("id")
             if trace_execution_id:
@@ -139,6 +176,108 @@ async def dispatch_worker_prompt(
         timeout=aiohttp.ClientTimeout(total=60),
     ) as resp:
         resp.raise_for_status()
+        if not return_prompt_id:
+            return None
+        try:
+            response_data = await resp.json()
+        except Exception as exc:
+            raise RuntimeError("Worker /prompt response did not contain JSON acknowledgement") from exc
+        prompt_id = response_data.get("prompt_id") if isinstance(response_data, dict) else None
+        if not isinstance(prompt_id, str) or not prompt_id:
+            raise RuntimeError("Worker /prompt response omitted prompt_id")
+        return prompt_id if return_prompt_id else None
+
+
+async def preflight_image_list_workers(workers, require_idle=True, timeout=3.0, *, require_value_results=False):
+    """Require protocol/node support and report which direct workers are idle.
+
+    This performs read-only GETs only; it never submits a prompt or contacts a
+    serverless provider. Unsupported worker releases fail before master inputs run.
+    """
+    if not workers:
+        raise RuntimeError("Distributed Image List requires at least one selected direct worker")
+    session = await get_client_session()
+    required_nodes = {"DistributedImageListStart", "DistributedImageListEnd"}
+    semaphore = asyncio.Semaphore(8)
+
+    async def inspect(worker):
+        worker_id = str(worker.get("id") or "unknown")
+        check = "capability"
+        try:
+            async with semaphore:
+                cap_url = build_worker_url(worker, "/distributed/image-list/capabilities")
+                async with session.get(cap_url, timeout=aiohttp.ClientTimeout(total=timeout)) as response:
+                    if response.status != 200:
+                        raise RuntimeError(
+                            f"Image-list worker {worker_id} does not expose the capability endpoint "
+                            f"(HTTP {response.status}); upgrade that worker before running this feature"
+                        )
+                    capability = await response.json()
+                if not isinstance(capability, dict):
+                    raise RuntimeError(f"Image-list worker {worker_id} returned malformed capability data")
+                supported = capability.get("supported")
+                protocol_version = capability.get("protocol_version")
+                nodes = capability.get("node_ids")
+                if (
+                    not isinstance(supported, bool)
+                    or type(protocol_version) is not int
+                    or not isinstance(nodes, list)
+                    or any(not isinstance(node_id, str) for node_id in nodes)
+                ):
+                    raise RuntimeError(f"Image-list worker {worker_id} returned malformed capability data")
+                if (
+                    supported is not True
+                    or protocol_version != IMAGE_LIST_PROTOCOL_VERSION
+                    or not required_nodes.issubset(nodes)
+                ):
+                    raise RuntimeError(
+                        f"Image-list worker {worker_id} is incompatible: requires protocol "
+                        f"v{IMAGE_LIST_PROTOCOL_VERSION} and nodes "
+                        f"{sorted(required_nodes)}; no work was submitted"
+                    )
+
+                if capability.get("execution_targets") is not True:
+                    raise RuntimeError(
+                        f"Image-list worker {worker_id} lacks explicit execution_targets support; "
+                        "upgrade that worker before running this feature; no work was submitted"
+                    )
+
+                if require_value_results and capability.get("value_results") is not True:
+                    raise RuntimeError(
+                        f"Image-list worker {worker_id} lacks explicit value_results support; "
+                        "upgrade that worker before using End overrides; no work was submitted"
+                    )
+
+                check = "queue"
+                queue_url = build_worker_url(worker, "/queue")
+                async with session.get(queue_url, timeout=aiohttp.ClientTimeout(total=timeout)) as response:
+                    if response.status != 200:
+                        raise RuntimeError(f"Image-list worker {worker_id} queue preflight returned HTTP {response.status}")
+                    queue = await response.json()
+                if (
+                    not isinstance(queue, dict)
+                    or not isinstance(queue.get("queue_running"), list)
+                    or not isinstance(queue.get("queue_pending"), list)
+                ):
+                    raise RuntimeError(f"Image-list worker {worker_id} returned malformed /queue state")
+                idle = not queue["queue_running"] and not queue["queue_pending"]
+                return worker_id, idle
+        except (asyncio.TimeoutError, aiohttp.ClientConnectionError) as exc:
+            log(
+                f"[Distributed] Image-list worker {worker_id} unavailable during {check} check; "
+                f"skipping: {exc}"
+            )
+            return worker_id, None
+
+    statuses = await asyncio.gather(*(inspect(worker) for worker in workers))
+    idle_workers = {
+        worker_id: True
+        for worker_id, idle in statuses
+        if idle is True
+    }
+    if require_idle and not idle_workers:
+        raise RuntimeError("No selected feature-compatible direct worker is idle; unrelated worker prompts were left untouched")
+    return idle_workers
 
 
 async def select_active_workers(
@@ -203,6 +342,10 @@ def _extract_queue_remaining(payload):
 
 async def _probe_worker_queue(worker, semaphore, probe_timeout):
     async with semaphore:
+        if worker.get("type") == "serverless":
+            from .serverless_runtime import get_manager
+            status = await get_manager().health(worker)
+            return {"worker": worker, "queue_remaining": status["queueRemaining"]}
         worker_url = build_worker_url(worker)
         payload = await probe_worker(worker_url, timeout=probe_timeout)
         if payload is None:

api/orchestration/media_sync.py

@@ -29,6 +29,11 @@ def _normalize_media_reference(value):
         return None
     cleaned = re.sub(r"\s*\[\w+\]$", "", value).strip().replace("\\", "/")
     if MEDIA_FILE_RE.search(cleaned):
+        # Output/temp annotations select a different master-side media root.
+        # Preserve them through discovery/loading and as distinct rewrite keys.
+        annotation = re.search(r"\s*\[(output|temp)\]$", value)
+        if annotation:
+            return f"{cleaned} [{annotation.group(1)}]"
         return cleaned
     return None
 
@@ -171,6 +176,12 @@ async def _upload_media_to_worker(worker, filename, file_bytes, file_hash, mime_
     """Upload one media file to worker iff missing or hash-mismatched."""
     session = await get_client_session()
     normalized = filename.replace("\\", "/")
+    annotation = re.search(r"\s*\[(output|temp)\]$", normalized)
+    if annotation:
+        # Worker uploads live under input/. Do not collide across master roots,
+        # or overwrite a different reference with the same basename.
+        basename = normalized[:annotation.start()].rstrip().rsplit("/", 1)[-1]
+        normalized = f"distributed-media/{annotation.group(1)}/{file_hash}/{basename}"
 
     check_url = build_worker_url(worker, "/distributed/check_file")
     try:

api/orchestration/prompt_transform.py

@@ -458,11 +458,55 @@ def prepare_delegate_master_prompt(prompt_obj, collector_ids):
     return pruned_prompt
 
 
+def _resolve_direct_workers_for_image_list(config, requested_ids):
+    """Resolve enabled workers by default, or exact explicitly selected direct workers."""
+    requested = None if requested_ids is None else {str(worker_id) for worker_id in requested_ids}
+    configured_workers = (config or {}).get("workers", [])
+    configured_ids = {
+        str(worker.get("id") or "").strip()
+        for worker in configured_workers
+        if str(worker.get("id") or "").strip()
+    }
+    if requested is not None:
+        unknown_ids = requested - configured_ids
+        if unknown_ids:
+            raise RuntimeError(
+                "Unknown explicitly selected image-list worker ID(s): "
+                + ", ".join(sorted(unknown_ids))
+            )
+    direct = []
+    for worker in configured_workers:
+        worker_id = str(worker.get("id") or "").strip()
+        if not worker_id:
+            continue
+        if requested is None:
+            if not worker.get("enabled", False):
+                continue
+        elif worker_id not in requested:
+            continue
+        if str(worker.get("type") or "local").strip().lower() == "serverless":
+            continue
+        try:
+            port = int(worker.get("port", worker.get("listen_port", 8188)) or 8188)
+        except (TypeError, ValueError) as exc:
+            raise RuntimeError(f"Invalid port for image-list worker {worker_id}") from exc
+        direct.append({
+            "id": worker_id,
+            "name": str(worker.get("name") or worker_id),
+            "host": worker.get("host"),
+            "port": port,
+            "type": worker.get("type", "local"),
+        })
+    return direct
+
+
 def generate_job_id_map(prompt_index, prefix):
     """Create stable per-node job IDs for distributed nodes."""
     job_map = {}
-    distributed_nodes = prompt_index.nodes_for_class("DistributedCollector") + prompt_index.nodes_for_class(
-        "UltimateSDUpscaleDistributed"
+    distributed_nodes = (
+        prompt_index.nodes_for_class("DistributedCollector")
+        + prompt_index.nodes_for_class("UltimateSDUpscaleDistributed")
+        + prompt_index.nodes_for_class("DistributedImageListStart")
     )
     for node_id in distributed_nodes:
         job_map[node_id] = f"{prefix}_{node_id}"
@@ -557,6 +601,22 @@ def _override_value_nodes(prompt_copy, prompt_index, is_master, participant_id,
             inputs["worker_id"] = f"worker_{worker_index_map.get(participant_id, 0)}"
 
 
+def _override_image_list_nodes(prompt_copy, prompt_index, participant_id, worker_ids,
+                               job_id_map, master_url, delegate_master, workflow_meta=None):
+    """Attach protocol context to Start without changing its visible V3 sockets."""
+    for node_id in prompt_index.nodes_for_class("DistributedImageListStart"):
+        node = prompt_copy.get(node_id)
+        if not isinstance(node, dict):
+            continue
+        inputs = node.setdefault("inputs", {})
+        inputs["image_list_enabled_worker_ids"] = json.dumps(list(worker_ids))
+        inputs["image_list_master_participates"] = participant_id == "master" and not delegate_master
+        inputs["image_list_master_callback"] = str(master_url or "")
+        inputs["image_list_execution_id"] = job_id_map.get(node_id, node_id)
+        if workflow_meta is not None:
+            inputs["image_list_workflow_meta"] = workflow_meta
+
+
 def apply_participant_overrides(
     prompt_copy,
     participant_id,
@@ -565,6 +625,7 @@ def apply_participant_overrides(
     master_url,
     delegate_master,
     prompt_index,
+    workflow_meta=None,
 ):
     """Return a prompt copy with hidden inputs configured for master/worker."""
     is_master = participant_id == "master"
@@ -592,5 +653,15 @@ def apply_participant_overrides(
         master_url,
         enabled_json,
     )
+    _override_image_list_nodes(
+        prompt_copy,
+        prompt_index,
+        participant_id,
+        enabled_worker_ids,
+        job_id_map,
+        master_url,
+        delegate_master,
+        workflow_meta=workflow_meta,
+    )
 
     return prompt_copy

api/orchestration/runpod_client.py

@@ -0,0 +1,127 @@
+"""Small Runpod queue API client. Never retries a potentially billable POST."""
+import asyncio
+import json
+import re
+
+import aiohttp
+
+API_BASE = "https://api.runpod.ai/v2"
+MAX_REQUEST_BYTES = 10_000_000
+_ID = re.compile(r"[A-Za-z0-9_-]{1,200}\Z")
+
+
+class RunpodError(RuntimeError):
+    """Sanitized provider failure safe to display in the panel."""
+
+
+class AmbiguousSubmission(RunpodError):
+    """Submission may have been accepted: never retry it automatically."""
+
+
+def serialize_request(workflow, images, timeout_seconds, *, media=None):
+    """Serialize and size-check the exact JSON body used by Runpod ``/run``."""
+    payload = {
+        "input": {"workflow": workflow, "images": images},
+        "policy": {
+            "executionTimeout": timeout_seconds * 1000,
+            "ttl": timeout_seconds * 1000,
+        },
+    }
+    if media is not None:
+        from ...serverless_bridge.handler import _validate_request
+        payload["input"]["media"] = media
+        _validate_request(payload["input"])
+    body = json.dumps(
+        payload,
+        allow_nan=False,
+        ensure_ascii=False,
+        separators=(",", ":"),
+    ).encode("utf-8")
+    size = len(body)
+    if size > MAX_REQUEST_BYTES:
+        raise ValueError(
+            f"Serialized serverless request body is {size:,} bytes; Runpod allows at most "
+            f"{MAX_REQUEST_BYTES:,} bytes. Base64-encoded images add about 33% over raw "
+            "image bytes; reduce image or workflow size."
+        )
+    return body
+
+
+def checked_id(value):
+    if not isinstance(value, str) or not _ID.fullmatch(value):
+        raise ValueError("Invalid Runpod endpoint or job ID")
+    return value
+
+
+class RunpodClient:
+    def __init__(self, session, endpoint_id, api_key):
+        self.session = session
+        self.endpoint_id = checked_id(endpoint_id)
+        if not isinstance(api_key, str) or not api_key or any(c in api_key for c in "\r\n"):
+            raise ValueError("A valid Runpod API key is required")
+        self._api_key = api_key
+
+    async def _request(self, method, path, payload=None, request_body=None):
+        # Base origin is code-owned, never a user-supplied URL. Do not follow redirects
+        # with credentials, or include response bodies/transport errors in exceptions.
+        url = f"{API_BASE}/{self.endpoint_id}/{path}"
+        attempts = 3 if method == "GET" else 1
+        for attempt in range(attempts):
+            try:
+                request_kwargs = {
+                    "headers": {"Authorization": f"Bearer {self._api_key}", "Content-Type": "application/json"},
+                    "timeout": aiohttp.ClientTimeout(total=30),
+                    "allow_redirects": False,
+                }
+                if request_body is None:
+                    request_kwargs["json"] = payload
+                else:
+                    request_kwargs["data"] = request_body
+                async with self.session.request(
+                    method, url,
+                    **request_kwargs,
+                ) as response:
+                    status = response.status
+                    if status != 200:
+                        if method == "GET" and (status == 429 or status >= 500) and attempt + 1 < attempts:
+                            await asyncio.sleep(0.5 * (2 ** attempt))
+                            continue
+                        if method == "POST" and path == "run" and status >= 500:
+                            raise AmbiguousSubmission("Runpod submission returned a server error; acceptance is unknown. Check the provider before retrying.")
+                        raise RunpodError(f"Runpod {path.split('/')[0]} request failed (HTTP {status})")
+                    try:
+                        result = await response.json()
+                    except (ValueError, aiohttp.ContentTypeError):
+                        if method == "POST" and path == "run":
+                            raise AmbiguousSubmission("Runpod submission returned an unreadable response; check the provider before retrying.") from None
+                        raise RunpodError("Runpod returned an invalid JSON response") from None
+                    if not isinstance(result, dict):
+                        if method == "POST" and path == "run":
+                            raise AmbiguousSubmission("Runpod submission did not return a job record; check the provider before retrying.")
+                        raise RunpodError("Runpod returned an invalid response")
+                    return result
+            except (asyncio.TimeoutError, aiohttp.ClientError, OSError):
+                if method == "POST" and path == "run":
+                    raise AmbiguousSubmission("Runpod submission connection was interrupted; acceptance is unknown. Check the provider before retrying.") from None
+                if method == "GET" and attempt + 1 < attempts:
+                    await asyncio.sleep(0.5 * (2 ** attempt))
+                    continue
+                raise RunpodError(f"Runpod {path.split('/')[0]} connection failed") from None
+        raise RunpodError("Runpod request failed")
+
+    async def health(self):
+        return await self._request("GET", "health")
+
+    async def submit(self, workflow, images, timeout_seconds, *, media=None):
+        body = serialize_request(workflow, images, timeout_seconds, media=media)
+        result = await self._request("POST", "run", request_body=body)
+        try:
+            return checked_id(result.get("id"))
+        except ValueError:
+            raise AmbiguousSubmission("Runpod submission returned no valid job ID; check the provider before retrying.") from None
+
+    async def status(self, job_id):
+        return await self._request("GET", f"status/{checked_id(job_id)}")
+
+    async def cancel(self, job_id):
+        return await self._request("POST", f"cancel/{checked_id(job_id)}")

api/orchestration/serverless_ledger.py

@@ -0,0 +1,57 @@
+"""Private, atomic job-ID journal. Contains identifiers/status only, never payloads."""
+import json
+import os
+from pathlib import Path
+import tempfile
+import threading
+
+from .serverless_state import TERMINAL
+FIELDS = {"record_id", "execution_id", "worker_id", "participant_id", "endpoint_id", "job_id", "status", "created_at", "updated_at", "error", "prompt_id"}
+
+
+class JobLedger:
+    def __init__(self, path):
+        self.path = Path(path)
+        self._lock = threading.RLock()
+
+    def _read(self):
+        if self.path.is_symlink():
+            raise RuntimeError("Unsafe serverless job journal path")
+        if not self.path.exists():
+            return {}
+        try:
+            data = json.loads(self.path.read_text(encoding="utf-8"))
+            if not isinstance(data, dict) or any(not isinstance(v, dict) or set(v) - FIELDS for v in data.values()):
+                raise ValueError()
+            return data
+        except (ValueError, OSError):
+            raise RuntimeError("Serverless job journal is unreadable; preserve it for recovery before submitting more work") from None
+
+    def all(self):
+        with self._lock:
+            return list(self._read().values())
+
+    def update(self, record_id, **fields):
+        if set(fields) - FIELDS:
+            raise ValueError("Unexpected serverless journal field")
+        with self._lock:
+            data = self._read()
+            data.setdefault(record_id, {"record_id": record_id}).update(fields)
+            # Keep every unresolved job, plus a bounded terminal history.
+            completed = sorted((k for k, v in data.items() if v.get("status") in TERMINAL), key=lambda k: data[k].get("updated_at", 0))
+            for key in completed[:-200]:
+                data.pop(key)
+            if self.path.parent.is_symlink():
+                raise RuntimeError("Unsafe serverless journal directory")
+            self.path.parent.mkdir(mode=0o700, parents=True, exist_ok=True)
+            fd, temporary = tempfile.mkstemp(prefix=".jobs-", dir=self.path.parent)
+            try:
+                with os.fdopen(fd, "w", encoding="utf-8") as stream:
+                    json.dump(data, stream, allow_nan=False)
+                    stream.flush()
+                    os.fsync(stream.fileno())
+                os.replace(temporary, self.path)
+            finally:
+                if os.path.exists(temporary):
+                    os.unlink(temporary)
+            return dict(data[record_id])

api/orchestration/serverless_media.py

@@ -0,0 +1,601 @@
+"""Bounded image transport for generic Runpod-backed ComfyUI jobs."""
+
+import asyncio
+import base64
+import binascii
+import hashlib
+import io
+import ipaddress
+import os
+import re
+import socket
+import stat
+import warnings
+from pathlib import Path, PurePosixPath
+from urllib.parse import urljoin, urlsplit
+
+import aiohttp
+import numpy as np
+from PIL import Image, UnidentifiedImageError
+
+
+MAX_INPUT_IMAGE_BYTES = 25 * 1024 * 1024
+MAX_INPUT_TOTAL_BYTES = 10_000_000
+MAX_OUTPUT_FILES = 1024
+MAX_OUTPUT_IMAGE_BYTES = 128 * 1024 * 1024
+MAX_OUTPUT_TOTAL_BYTES = 256 * 1024 * 1024
+MAX_IMAGE_PIXELS = 40_000_000
+MAX_OUTPUT_ARRAY_BYTES = 256 * 1024 * 1024
+MAX_REDIRECTS = 3
+_DOWNLOAD_TIMEOUT_SECONDS = 60
+_CHUNK_SIZE = 64 * 1024
+
+_IMAGE_EXTENSIONS = {".png", ".jpg", ".jpeg", ".webp", ".bmp"}
+_UNSUPPORTED_IMAGE_EXTENSIONS = {".gif", ".tif", ".tiff", ".avif", ".heic", ".ico", ".svg"}
+_AUDIO_EXTENSIONS = {".wav", ".mp3", ".flac", ".m4a", ".aac", ".ogg", ".opus", ".aiff", ".aif", ".wma"}
+_VIDEO_EXTENSIONS = {".mp4", ".avi", ".mov", ".mkv", ".webm"}
+_MEDIA_EXTENSIONS = _IMAGE_EXTENSIONS | _UNSUPPORTED_IMAGE_EXTENSIONS | _AUDIO_EXTENSIONS | _VIDEO_EXTENSIONS
+_MEDIA_ANNOTATION = re.compile(r"\s+\[([A-Za-z0-9_-]{1,64})\]$")
+_PREFIX_RE = re.compile(r"[A-Za-z0-9][A-Za-z0-9_.-]{0,199}\Z", re.ASCII)
+_SAFE_FILENAME_RE = re.compile(r"[A-Za-z0-9_.-]{1,255}\Z", re.ASCII)
+
+
+class ServerlessMediaError(ValueError):
+    """Sanitized input, artifact, or download validation failure."""
+
+
+def _folder_roots(folder_paths):
+    roots = {}
+    for kind, getter_name in (
+        ("input", "get_input_directory"),
+        ("output", "get_output_directory"),
+        ("temp", "get_temp_directory"),
+    ):
+        getter = getattr(folder_paths, getter_name, None)
+        if not callable(getter):
+            raise ServerlessMediaError("ComfyUI media directories are unavailable")
+        try:
+            roots[kind] = Path(getter()).resolve(strict=True)
+        except Exception:
+            raise ServerlessMediaError("ComfyUI media directories are unavailable") from None
+    return roots
+
+
+def _inside_roots(path, roots):
+    for root in roots.values():
+        try:
+            path.relative_to(root)
+            return True
+        except ValueError:
+            continue
+    return False
+
+
+def _normalize_media_reference(value):
+    if not isinstance(value, str) or not value or value != value.strip() or "\x00" in value:
+        raise ServerlessMediaError("Image input reference is invalid")
+    value = value.replace("\\", "/")
+    explicit_root = None
+    annotation = _MEDIA_ANNOTATION.search(value)
+    if annotation:
+        marker = annotation.group(1).lower()
+        if marker not in {"input", "output", "temp"}:
+            raise ServerlessMediaError("Image input annotation is unsupported")
+        explicit_root = marker
+        value = value[: annotation.start()]
+    if "://" in value:
+        raise ServerlessMediaError("Remote media references are not supported as image inputs")
+    if re.match(r"^[A-Za-z]:", value) or value.startswith("//"):
+        raise ServerlessMediaError("Image input path is outside the allowed ComfyUI media directories")
+
+    path = PurePosixPath(value)
+    if any(part == ".." for part in path.parts):
+        raise ServerlessMediaError("Image input path traversal is not allowed")
+    if not path.parts or value in {".", ""}:
+        raise ServerlessMediaError("Image input reference is invalid")
+
+    if explicit_root:
+        value = f"{value} [{explicit_root}]"
+    suffix = Path(path.name).suffix.lower()
+    return value, suffix
+
+
+def _resolve_input_path(value, suffix, folder_paths, roots):
+    if suffix not in _IMAGE_EXTENSIONS:
+        if suffix in _MEDIA_EXTENSIONS:
+            raise ServerlessMediaError("Audio and video inputs are not supported by serverless image transport")
+        raise ServerlessMediaError("Unsupported serverless image file type")
+
+    try:
+        source = Path(value)
+        if source.is_absolute():
+            annotation = _MEDIA_ANNOTATION.search(value)
+            unannotated = value[:annotation.start()] if annotation else value
+            resolved = Path(unannotated).resolve(strict=True)
+        else:
+            resolver = getattr(folder_paths, "get_annotated_filepath", None)
+            if not callable(resolver):
+                raise ServerlessMediaError("ComfyUI media path resolver is unavailable")
+            resolved = Path(resolver(value)).resolve(strict=True)
+    except ServerlessMediaError:
+        raise
+    except Exception:
+        raise ServerlessMediaError("Image input file is missing or invalid") from None
+
+    if not _inside_roots(resolved, roots):
+        raise ServerlessMediaError("Image input path is outside the allowed ComfyUI media directories")
+    try:
+        mode = resolved.stat().st_mode
+    except OSError:
+        raise ServerlessMediaError("Image input file is missing or invalid") from None
+    if not stat.S_ISREG(mode):
+        raise ServerlessMediaError("Image input is not a regular file")
+    return resolved
+
+
+def _iter_image_references(workflow):
+    if not isinstance(workflow, dict):
+        raise ServerlessMediaError("Serverless workflow must be an API graph")
+    references = []
+    for node_id, node in workflow.items():
+        if not isinstance(node, dict):
+            continue
+        class_type = node.get("class_type")
+        inputs = node.get("inputs", {})
+        if class_type == "LoadImage" and not isinstance(inputs, dict):
+            raise ServerlessMediaError("LoadImage requires a literal image filename; linked image inputs are unsupported")
+        if not isinstance(inputs, dict):
+            continue
+        if class_type == "LoadImage":
+            value = inputs.get("image")
+            if not isinstance(value, str):
+                raise ServerlessMediaError("LoadImage requires a literal image filename; linked image inputs are unsupported")
+            references.append((node_id, inputs, "image", value, True))
+
+        for key in ("image", "file"):
+            if key == "image" and class_type == "LoadImage":
+                continue
+            value = inputs.get(key)
+            if not isinstance(value, str):
+                continue
+            normalized, suffix = _normalize_media_reference(value)
+            if suffix in _IMAGE_EXTENSIONS:
+                references.append((node_id, inputs, key, normalized, False))
+            elif suffix in _AUDIO_EXTENSIONS | _VIDEO_EXTENSIONS:
+                raise ServerlessMediaError("Audio and video inputs are not supported by serverless image transport")
+            elif suffix in _UNSUPPORTED_IMAGE_EXTENSIONS or (key == "image" and suffix):
+                raise ServerlessMediaError("Unsupported serverless image file type")
+
+        for key in ("audio", "video"):
+            value = inputs.get(key)
+            if not isinstance(value, str):
+                continue
+            _normalized, suffix = _normalize_media_reference(value)
+            if suffix in _MEDIA_EXTENSIONS:
+                raise ServerlessMediaError("Audio and video inputs are not supported by serverless image transport")
+    return references
+
+
+def _read_limited(path, expected_size):
+    try:
+        descriptor = os.open(path, os.O_RDONLY | getattr(os, "O_NOFOLLOW", 0))
+        with os.fdopen(descriptor, "rb") as handle:
+            info = os.fstat(handle.fileno())
+            if not stat.S_ISREG(info.st_mode) or info.st_size > MAX_INPUT_IMAGE_BYTES:
+                raise ServerlessMediaError("Image input exceeds the permitted size")
+            if info.st_size > MAX_INPUT_TOTAL_BYTES:
+                raise ServerlessMediaError(
+                    f"Total raw serverless image input size exceeds {MAX_INPUT_TOTAL_BYTES:,} bytes; "
+                    "base64 encoding adds about 33%, and the complete request body is also limited to 10,000,000 bytes"
+                )
+            data = handle.read(MAX_INPUT_TOTAL_BYTES + 1)
+    except ServerlessMediaError:
+        raise
+    except OSError:
+        raise ServerlessMediaError("Image input file could not be read safely") from None
+    if len(data) > MAX_INPUT_TOTAL_BYTES:
+        raise ServerlessMediaError(
+            f"Total raw serverless image input size exceeds {MAX_INPUT_TOTAL_BYTES:,} bytes; "
+            "base64 encoding adds about 33%, and the complete request body is also limited to 10,000,000 bytes"
+        )
+    if len(data) != expected_size:
+        raise ServerlessMediaError("Image input changed while being read")
+    return data
+
+
+async def prepare_images(workflow: dict) -> list[dict]:
+    """Pack referenced local images and rewrite only their graph input references.
+
+    ``workflow`` is expected to be a private API-graph copy owned by the caller.
+    Returned ``name`` values are flat generated names suitable for ``input.images``.
+    """
+    references = _iter_image_references(workflow)
+    if not references:
+        return []
+
+    # Import only in ComfyUI, keeping this module importable in the serverless client.
+    try:
+        import folder_paths
+    except Exception:
+        raise ServerlessMediaError("ComfyUI media directories are unavailable") from None
+
+    roots = _folder_roots(folder_paths)
+    resolved_refs = []
+    unique_paths = {}
+    total_source_bytes = 0
+    for node_id, inputs, key, value, _is_load_image in references:
+        normalized, suffix = _normalize_media_reference(value)
+        if key == "image" and suffix not in _IMAGE_EXTENSIONS:
+            if suffix in _AUDIO_EXTENSIONS | _VIDEO_EXTENSIONS:
+                raise ServerlessMediaError("Audio and video inputs are not supported by serverless image transport")
+            raise ServerlessMediaError("Unsupported serverless image file type")
+        path = _resolve_input_path(normalized, suffix, folder_paths, roots)
+        if path not in unique_paths:
+            try:
+                size = path.stat().st_size
+            except OSError:
+                raise ServerlessMediaError("Image input file is missing or invalid") from None
+            if size < 0 or size > MAX_INPUT_IMAGE_BYTES:
+                raise ServerlessMediaError("Image input exceeds the permitted size")
+            total_source_bytes += size
+            if total_source_bytes > MAX_INPUT_TOTAL_BYTES:
+                raise ServerlessMediaError(
+                    f"Total raw serverless image input size is {total_source_bytes:,} bytes; "
+                    f"the input read guard is {MAX_INPUT_TOTAL_BYTES:,} bytes. Base64 encoding "
+                    "adds about 33%, and the complete request body is also limited to 10,000,000 bytes"
+                )
+            unique_paths[path] = {"size": size, "suffix": suffix}
+        resolved_refs.append((inputs, key, path))
+
+    packed_by_name = {}
+    path_to_name = {}
+    for path, metadata in unique_paths.items():
+        data = _read_limited(path, metadata["size"])
+        digest = hashlib.sha256(data).hexdigest()
+        name = f"serverless_{digest}{metadata['suffix']}"
+        path_to_name[path] = name
+        packed_by_name.setdefault(name, {"name": name, "image": base64.b64encode(data).decode("ascii")})
+
+    for inputs, key, path in resolved_refs:
+        inputs[key] = path_to_name[path]
+    return [packed_by_name[name] for name in sorted(packed_by_name)]
+
+
+def _validate_artifact_url(url):
+    if not isinstance(url, str) or not url or any(ord(char) < 0x20 or char.isspace() for char in url):
+        raise ServerlessMediaError("Serverless image artifact URL is invalid")
+    try:
+        parsed = urlsplit(url)
+        port = parsed.port
+    except ValueError:
+        raise ServerlessMediaError("Serverless image artifact URL is invalid") from None
+    if (
+        parsed.scheme.lower() != "https"
+        or not parsed.hostname
+        or parsed.username is not None
+        or parsed.password is not None
+        or port not in (None, 443)
+        or parsed.fragment
+    ):
+        raise ServerlessMediaError("Serverless image artifact URL must be a public HTTPS URL")
+    host = parsed.hostname.rstrip(".").lower()
+    if host in {"localhost", "metadata", "metadata.google.internal"} or host.endswith((".localhost", ".local", ".internal")):
+        raise ServerlessMediaError("Serverless image artifact URL must be a public HTTPS URL")
+    try:
+        address = ipaddress.ip_address(host)
+    except ValueError:
+        return url
+    if not address.is_global:
+        raise ServerlessMediaError("Serverless image artifact URL must resolve to a public address")
+    return url
+
+
+class _PublicResolver(aiohttp.abc.AbstractResolver):
+    """Resolve once, reject non-public answers, and pin the checked address."""
+
+    async def _getaddrinfo(self, host, port, family):
+        return await asyncio.get_running_loop().getaddrinfo(
+            host, port, family=family, type=socket.SOCK_STREAM
+        )
+
+    async def resolve(self, host, port=0, family=socket.AF_INET):
+        try:
+            results = await self._getaddrinfo(host, port, family)
+        except (OSError, asyncio.TimeoutError):
+            raise OSError("Public artifact host could not be resolved") from None
+        if not results:
+            raise OSError("Public artifact host could not be resolved")
+
+        resolved = []
+        seen = set()
+        for result_family, _socktype, proto, _canonname, sockaddr in results:
+            try:
+                address = ipaddress.ip_address(sockaddr[0])
+            except (ValueError, TypeError, IndexError):
+                raise OSError("Public artifact host resolved to an invalid address") from None
+            if not address.is_global:
+                raise OSError("Public artifact host resolved to a non-public address")
+            key = (str(address), result_family)
+            if key in seen:
+                continue
+            seen.add(key)
+            resolved.append({
+                "hostname": host,
+                "host": str(address),
+                "port": port,
+                "family": result_family,
+                "proto": proto,
+                "flags": 0,
+            })
+        if not resolved:
+            raise OSError("Public artifact host could not be resolved")
+        return resolved
+
+    async def close(self):
+        return None
+
+
+async def download_public_artifact(url: str, max_bytes: int = MAX_OUTPUT_IMAGE_BYTES) -> bytes:
+    """Download a signed public HTTPS artifact without provider credentials."""
+    if isinstance(max_bytes, bool) or not isinstance(max_bytes, int) or not 0 < max_bytes <= MAX_OUTPUT_IMAGE_BYTES:
+        raise ServerlessMediaError("Serverless image artifact exceeds the permitted size")
+    current_url = _validate_artifact_url(url)
+    resolver = _PublicResolver()
+    connector = aiohttp.TCPConnector(resolver=resolver, use_dns_cache=False, limit=4)
+    timeout = aiohttp.ClientTimeout(
+        total=_DOWNLOAD_TIMEOUT_SECONDS,
+        connect=10,
+        sock_read=15,
+    )
+    try:
+        async with aiohttp.ClientSession(
+            connector=connector,
+            timeout=timeout,
+            trust_env=False,
+            cookie_jar=aiohttp.DummyCookieJar(),
+        ) as session:
+            for redirect_count in range(MAX_REDIRECTS + 1):
+                _validate_artifact_url(current_url)
+                async with session.get(current_url, allow_redirects=False) as response:
+                    if response.status in {301, 302, 303, 307, 308}:
+                        location = response.headers.get("Location")
+                        if not location or redirect_count >= MAX_REDIRECTS:
+                            raise ServerlessMediaError("Serverless image artifact redirect is invalid")
+                        current_url = _validate_artifact_url(urljoin(current_url, location))
+                        continue
+                    if response.status != 200:
+                        raise ServerlessMediaError("Serverless image artifact download failed")
+                    payload = bytearray()
+                    async for chunk in response.content.iter_chunked(_CHUNK_SIZE):
+                        if not isinstance(chunk, bytes):
+                            raise ServerlessMediaError("Serverless image artifact response is invalid")
+                        if len(payload) + len(chunk) > max_bytes:
+                            raise ServerlessMediaError("Serverless image artifact exceeds the permitted size")
+                        payload.extend(chunk)
+                    if not payload:
+                        raise ServerlessMediaError("Serverless image artifact is empty")
+                    return bytes(payload)
+    except ServerlessMediaError:
+        raise
+    except Exception:
+        raise ServerlessMediaError("Serverless image artifact download failed") from None
+    raise ServerlessMediaError("Serverless image artifact download failed")
+
+
+def _validated_manifest(manifest):
+    if not isinstance(manifest, list) or len(manifest) > MAX_OUTPUT_FILES:
+        raise ServerlessMediaError("Serverless image manifest is invalid")
+    normalized = []
+    seen_prefixes = set()
+    prefixes = []
+    for item in manifest:
+        if not isinstance(item, dict):
+            raise ServerlessMediaError("Serverless image manifest is invalid")
+        if item.get("kind") != "image":
+            raise ServerlessMediaError("Only image collectors are supported by serverless media transport")
+        for key in ("collector_id", "job_id", "node_id"):
+            value = item.get(key)
+            if not isinstance(value, (str, int)) or isinstance(value, bool) or not str(value):
+                raise ServerlessMediaError("Serverless image manifest is invalid")
+        prefix = item.get("prefix")
+        if not isinstance(prefix, str) or not _PREFIX_RE.fullmatch(prefix) or prefix in {".", ".."}:
+            raise ServerlessMediaError("Serverless image manifest contains an unsafe filename prefix")
+        folded = prefix.casefold()
+        if folded in seen_prefixes:
+            raise ServerlessMediaError("Serverless image manifest prefixes must be unique")
+        seen_prefixes.add(folded)
+        prefixes.append(prefix)
+        normalized.append({
+            "collector_id": str(item["collector_id"]),
+            "job_id": str(item["job_id"]),
+            "node_id": str(item["node_id"]),
+            "prefix": prefix,
+        })
+    for first in prefixes:
+        for second in prefixes:
+            if first != second and second.casefold().startswith(first.casefold() + "_"):
+                raise ServerlessMediaError("Serverless image manifest prefixes are ambiguous")
+    return normalized
+
+
+def _artifact_filename_matches(filename, prefixes):
+    if not isinstance(filename, str):
+        return None
+    basename = filename.replace("\\", "/").rsplit("/", 1)[-1]
+    matches = [item for item in prefixes if basename.startswith(item["prefix"] + "_")]
+    if len(matches) > 1:
+        raise ServerlessMediaError("Serverless image artifact matches multiple collectors")
+    return matches[0] if matches else None
+
+
+def _canonical_artifact_identity(entry, filename):
+    artifact_path = entry.get("artifact_path")
+    artifact_type = entry.get("artifact_type", "output")
+    if artifact_path is not None:
+        if not isinstance(artifact_path, str) or not artifact_path or "\x00" in artifact_path:
+            raise ServerlessMediaError("Serverless image artifact identity is invalid")
+        normalized_path = os.path.normcase(os.path.normpath(artifact_path.replace("\\", "/")))
+    else:
+        subfolder = entry.get("subfolder", "")
+        if not isinstance(subfolder, str) or "\x00" in subfolder:
+            raise ServerlessMediaError("Serverless image artifact identity is invalid")
+        normalized_path = os.path.normcase(os.path.normpath(f"{subfolder}/{filename}"))
+    return str(artifact_type), normalized_path
+
+
+def _decode_base64_data(value, max_bytes=MAX_OUTPUT_IMAGE_BYTES):
+    if not isinstance(value, str) or not value:
+        raise ServerlessMediaError("Serverless image artifact data is invalid")
+    max_bytes = min(max_bytes, MAX_OUTPUT_IMAGE_BYTES)
+    max_encoded = ((max_bytes + 2) // 3) * 4
+    if len(value) > max_encoded:
+        raise ServerlessMediaError("Serverless image artifact exceeds the permitted size")
+    padding = len(value) - len(value.rstrip("="))
+    estimated_size = (len(value) // 4) * 3 - padding
+    if estimated_size > max_bytes:
+        raise ServerlessMediaError("Serverless image artifact exceeds the permitted size")
+    try:
+        payload = base64.b64decode(value, validate=True)
+    except (binascii.Error, ValueError):
+        raise ServerlessMediaError("Serverless image artifact data is invalid") from None
+    if not payload:
+        raise ServerlessMediaError("Serverless image artifact is empty")
+    if len(payload) > max_bytes:
+        raise ServerlessMediaError("Serverless image artifact exceeds the permitted size")
+    return payload
+
+
+def _decode_image_array(payload, total_array_bytes):
+    try:
+        with warnings.catch_warnings():
+            warnings.simplefilter("error", Image.DecompressionBombWarning)
+            image = Image.open(io.BytesIO(payload))
+            width, height = image.size
+            if width <= 0 or height <= 0 or width * height > MAX_IMAGE_PIXELS:
+                raise ServerlessMediaError("Serverless image dimensions exceed the permitted limit")
+            has_alpha = image.mode in {"RGBA", "LA"} or "transparency" in image.info
+            mode = "RGBA" if has_alpha else "RGB"
+            output_bytes = width * height * (4 if has_alpha else 3) * 4
+            if total_array_bytes + output_bytes > MAX_OUTPUT_ARRAY_BYTES:
+                raise ServerlessMediaError("Decoded serverless images exceed the permitted memory limit")
+            converted = image.convert(mode)
+            converted.load()
+            array = np.array(converted, dtype=np.float32, copy=True)
+            np.multiply(array, 1.0 / 255.0, out=array)
+            return array, output_bytes
+    except ServerlessMediaError:
+        raise
+    except (UnidentifiedImageError, OSError, ValueError, Image.DecompressionBombError, Image.DecompressionBombWarning):
+        raise ServerlessMediaError("Serverless image artifact could not be decoded") from None
+    except Exception:
+        raise ServerlessMediaError("Serverless image artifact could not be decoded") from None
+
+
+def _image_output(output):
+    """Unwrap handler output and reject explicit execution errors."""
+    if not isinstance(output, dict):
+        raise ServerlessMediaError("Serverless job output is invalid")
+    nested = output.get("output")
+    data = nested if isinstance(nested, dict) else output
+    for candidate in (output, data):
+        if candidate.get("error") or candidate.get("errors"):
+            raise ServerlessMediaError("Serverless job reported an execution error")
+
+    return data
+
+
+async def _select_image_payloads(data, expected, download):
+    """Match artifact identities and fetch bounded bytes, without image decoding."""
+    artifacts = data.get("images", [])
+    if not isinstance(artifacts, list) or len(artifacts) > MAX_OUTPUT_FILES:
+        raise ServerlessMediaError("Serverless job returned too many image files")
+
+    selected = {item["prefix"]: [] for item in expected}
+    identities = set()
+    total_payload_bytes = 0
+    for entry in artifacts:
+        if not isinstance(entry, dict):
+            continue
+        filename = entry.get("filename")
+        manifest_entry = _artifact_filename_matches(filename, expected)
+        if manifest_entry is None:
+            continue
+        normalized_filename = filename.replace("\\", "/")
+        if "/" in normalized_filename or not _SAFE_FILENAME_RE.fullmatch(filename):
+            raise ServerlessMediaError("Serverless image artifact filename is invalid")
+        if Path(filename).suffix.lower() not in _IMAGE_EXTENSIONS:
+            raise ServerlessMediaError("Serverless image artifact has an unsupported file type")
+        if "artifact_type" in entry and entry["artifact_type"] != "output":
+            raise ServerlessMediaError("Expected serverless image artifact was not an output image")
+        identity = _canonical_artifact_identity(entry, filename)
+        if identity in identities:
+            raise ServerlessMediaError("Duplicate serverless image artifact identity")
+        identities.add(identity)
+        match = re.match(rf"^{re.escape(manifest_entry['prefix'])}_(\d+)(?:_|\.)", filename)
+        if match is None:
+            raise ServerlessMediaError("Expected serverless image filename counter is invalid")
+        artifact_type = entry.get("type")
+        remaining_bytes = MAX_OUTPUT_TOTAL_BYTES - total_payload_bytes
+        if remaining_bytes <= 0:
+            raise ServerlessMediaError("Serverless image artifacts exceed the permitted total size")
+        if artifact_type == "base64":
+            payload = _decode_base64_data(entry.get("data"), remaining_bytes)
+        elif artifact_type in {"gcs_url", "s3_url", "url"}:
+            url = entry.get("data") or entry.get("url")
+            url = _validate_artifact_url(url)
+            try:
+                if download is None:
+                    payload = await download_public_artifact(
+                        url, max_bytes=min(MAX_OUTPUT_IMAGE_BYTES, remaining_bytes)
+                    )
+                else:
+                    payload = await download(url)
+            except ServerlessMediaError:
+                raise
+            except Exception:
+                raise ServerlessMediaError("Serverless image artifact download failed") from None
+            if not isinstance(payload, bytes) or not payload:
+                raise ServerlessMediaError("Serverless image artifact response is invalid")
+            if len(payload) > MAX_OUTPUT_IMAGE_BYTES:
+                raise ServerlessMediaError("Serverless image artifact exceeds the permitted size")
+        else:
+            raise ServerlessMediaError("Serverless image artifact transport type is unsupported")
+        total_payload_bytes += len(payload)
+        if total_payload_bytes > MAX_OUTPUT_TOTAL_BYTES:
+            raise ServerlessMediaError("Serverless image artifacts exceed the permitted total size")
+        selected[manifest_entry["prefix"]].append((int(match.group(1)), filename, identity, payload))
+
+    if any(not selected[item["prefix"]] for item in expected):
+        raise ServerlessMediaError("Serverless job is missing an expected collector image")
+    return selected
+
+
+def _decode_image_groups(selected, expected):
+    """Decode in stable counter order, bounding array memory across collectors."""
+    result = {}
+    total_array_bytes = 0
+    for manifest_entry in expected:
+        prefix = manifest_entry["prefix"]
+        rows = sorted(selected[prefix], key=lambda row: (row[0], row[1], row[2]))
+        arrays = []
+        shape = None
+        for _counter, _filename, _identity, payload in rows:
+            array, added_bytes = _decode_image_array(payload, total_array_bytes)
+            total_array_bytes += added_bytes
+            if shape is None:
+                shape = array.shape
+            elif array.shape != shape:
+                raise ServerlessMediaError("Images from one serverless collector have inconsistent dimensions")
+            arrays.append(array)
+        result.setdefault(manifest_entry["job_id"], []).extend(arrays)
+    return result
+
+
+async def collect_images(output: dict, manifest: list, download=None) -> dict[str, list[np.ndarray]]:
+    """Validate and decode expected image artifacts, grouped by serverless job ID."""
+    data = _image_output(output)
+    expected = _validated_manifest(manifest)
+    if not expected:
+        return {}
+    selected = await _select_image_payloads(data, expected, download)
+    return _decode_image_groups(selected, expected)

api/orchestration/serverless_media_prompt.py

@@ -0,0 +1,217 @@
+"""Pure compilation of prepared Distributed Collectors for comfy-media-v1 pools.
+
+The legacy image compiler remains separate. This compiler emits one MediaExportV1
+execution target per connected Collector socket (``images`` and/or ``audio``),
+retaining only its upstream computation. It never imports ComfyUI or
+inspects/executes runtime node classes.
+"""
+
+from __future__ import annotations
+
+import copy
+import hashlib
+import json
+from graphlib import CycleError, TopologicalSorter
+
+from ...serverless_bridge.contract import SOCKET_KINDS, identifier, validate_bindings
+from ...serverless_bridge.image_transfer import DEFAULT_FORMAT, validate_options
+from .serverless_prompt import (
+    _PRIMITIVE_NODES,
+    _REMOVED_SCALAR_NODES,
+    _is_link,
+    _is_unsupported_distributed,
+    _make_scalar_resolver,
+    _prune_graph,
+)
+
+
+_COLLECTOR = "DistributedCollector"
+# Collector inputs, in output-slot order, that a typed pool returns.
+COLLECTOR_SOCKETS = ("images", "audio")
+# Collector widget defaults (the saved-workflow contract is legacy_png / 90).
+COLLECTOR_DEFAULT_FORMAT = "legacy_png"
+COLLECTOR_DEFAULT_QUALITY = 90
+
+
+def collector_image_policy(inputs):
+    """Return the endpoint IMAGE (format, quality) for one Collector.
+
+    legacy_png is the Collector's saved default and names a JSON/base64 framing
+    that only applies to direct workers; serverless returns the same 8-bit
+    pixels as compressed png. exact stays opt-in. No other format is altered.
+    """
+    fmt, quality = validate_options(inputs.get("transfer_format", COLLECTOR_DEFAULT_FORMAT),
+                                    inputs.get("transfer_quality", COLLECTOR_DEFAULT_QUALITY))
+    return ("png" if fmt == "legacy_png" else fmt), quality
+
+
+def _validate_identifier(value, context):
+    # Validate the original token before any lookup. In particular, neither
+    # integer node IDs nor whitespace in prepared identities is normalized.
+    try:
+        identifier(value)
+    except ValueError:
+        raise ValueError(f"{context} must be a valid nonempty media identifier") from None
+
+
+def _validate_prompt(prepared_prompt):
+    """Validate the entire source graph, including branches later discarded."""
+    if not isinstance(prepared_prompt, dict) or not prepared_prompt:
+        raise ValueError("prepared prompt must be a nonempty API graph mapping")
+
+    collectors = []
+    for node_id, node in prepared_prompt.items():
+        _validate_identifier(node_id, "prompt node ID")
+        if not isinstance(node, dict):
+            raise ValueError(f"prompt node {node_id!r} must be an object")
+        class_type = node.get("class_type")
+        if not isinstance(class_type, str) or not class_type.strip():
+            raise ValueError(f"prompt node {node_id!r} must have a nonempty class_type")
+        if class_type == _COLLECTOR:
+            collectors.append(node_id)
+        elif _is_unsupported_distributed(class_type):
+            raise ValueError(f"unsupported distributed or batch node: {class_type}")
+        inputs = node.get("inputs", {})
+        if not isinstance(inputs, dict):
+            raise ValueError(f"prompt node {node_id!r} inputs must be an object")
+        if any(not isinstance(name, str) or not name for name in inputs):
+            raise ValueError(f"prompt node {node_id!r} input names must be nonempty strings")
+
+    if not collectors:
+        raise ValueError("prepared prompt contains no DistributedCollector nodes")
+
+    dependencies = {}
+    for node_id, node in prepared_prompt.items():
+        parents = set()
+        for input_name, value in node.get("inputs", {}).items():
+            # Ordinary inputs may contain literal lists. As in the legacy
+            # compiler, exactly two elements denotes an API link, not media.
+            if not _is_link(value):
+                continue
+            source_id, slot = value
+            context = f"link at node {node_id!r} input {input_name!r}"
+            _validate_identifier(source_id, context + " source ID")
+            if source_id not in prepared_prompt:
+                raise ValueError(f"invalid {context}: source {source_id!r} does not exist")
+            if type(slot) is not int or slot < 0:
+                raise ValueError(f"invalid {context}: output slot must be a nonnegative integer")
+            source_class = prepared_prompt[source_id]["class_type"]
+            if source_class in _REMOVED_SCALAR_NODES or source_class in _PRIMITIVE_NODES:
+                if slot != 0:
+                    raise ValueError(f"invalid output slot {slot} for scalar node {source_id!r}")
+            elif source_class == _COLLECTOR and slot >= len(COLLECTOR_SOCKETS):
+                raise ValueError(f"invalid output slot {slot} for collector {source_id!r}")
+            parents.add(source_id)
+        dependencies[node_id] = parents
+
+    for collector_id in collectors:
+        inputs = prepared_prompt[collector_id].get("inputs", {})
+        collector_image_policy(inputs)
+        for field in ("multi_job_id", "worker_id"):
+            _validate_identifier(inputs.get(field), f"collector {collector_id!r} {field}")
+        connected = [socket for socket in COLLECTOR_SOCKETS if socket in inputs]
+        if not connected:
+            raise ValueError(f"collector {collector_id!r} requires at least one connected media socket")
+        for socket in connected:
+            value = inputs[socket]
+            if not _is_link(value):
+                raise ValueError(f"collector {collector_id!r} {socket} must be linked; use PrimitiveString for literal text")
+            if prepared_prompt[value[0]]["class_type"] in _REMOVED_SCALAR_NODES:
+                # Export sockets must remain actual links after scalar folding.
+                # A PrimitiveString wrapper supports distributed text values.
+                raise ValueError(f"collector {collector_id!r} {socket} cannot link directly to a distributed scalar; use a typed node or PrimitiveString")
+
+    # Iterative topological sorting checks ordinary and unused cycles too, not
+    # only the primitive chains that the shared scalar resolver traverses.
+    try:
+        ordered = list(TopologicalSorter(dependencies).static_order())
+    except CycleError:
+        raise ValueError("cycle in prepared media prompt") from None
+    collector_set = set(collectors)
+    has_collector_ancestor = {}
+    for node_id in ordered:
+        has_collector_ancestor[node_id] = any(
+            parent in collector_set or has_collector_ancestor[parent]
+            for parent in dependencies[node_id]
+        )
+        if node_id in collector_set and has_collector_ancestor[node_id]:
+            raise ValueError("collector-to-collector dependence is unsupported")
+
+    return sorted(collectors)
+
+
+def _build_exports(source_nodes, collector_ids, request_id):
+    bindings = []
+    exports = {}
+    reserved = set(source_nodes)
+    for collector_id in collector_ids:
+        inputs = source_nodes[collector_id].get("inputs", {})
+        for socket in COLLECTOR_SOCKETS:
+            kind = SOCKET_KINDS[socket]
+            if socket not in inputs:
+                continue
+            # Hashing avoids separator ambiguity and keeps long valid source
+            # identities within the bridge's identifier length bound. Identity
+            # is stable across request IDs and dictionary insertion orders.
+            identity = json.dumps([collector_id, socket], separators=(",", ":"))
+            digest = hashlib.sha256(identity.encode("utf-8")).hexdigest()
+            base_id = f"serverless:media:{digest}"
+            node_id = base_id
+            suffix = 0
+            while node_id in reserved:
+                suffix += 1
+                node_id = f"{base_id}:{suffix}"
+            reserved.add(node_id)
+            binding = {
+                "binding_id": f"b:{digest}",
+                "collector_id": collector_id,
+                "job_id": inputs["multi_job_id"],
+                "participant_id": inputs["worker_id"],
+                "node_id": node_id,
+                "socket": socket,
+                "kind": kind,
+            }
+            bindings.append(binding)
+            exports[node_id] = {
+                "class_type": "MediaExportV1",
+                "inputs": {
+                    "request_id": request_id,
+                    "binding_json": json.dumps(binding),
+                    socket: copy.deepcopy(inputs[socket]),
+                },
+            }
+            if kind == "IMAGE":
+                fmt, quality = collector_image_policy(inputs)
+                if fmt != DEFAULT_FORMAT:
+                    exports[node_id]["inputs"].update(transfer_format=fmt, transfer_quality=quality)
+    validate_bindings(bindings)
+    return bindings, exports
+
+
+def compile_media_prompt(prepared_prompt: dict, request_id: str) -> tuple[dict, list[dict]]:
+    """Return a worker graph and canonical typed-media export bindings.
+
+    Each DistributedCollector must have prepared literal ``multi_job_id`` and
+    ``worker_id`` identities and at least one linked ``images``/``audio`` input.
+    Present media inputs must be API links. Collector transport controls become
+    the IMAGE export policy (``legacy_png`` is sent as ``png``); other controls
+    are not exported.
+
+    Only MediaExportV1 targets and their ancestors survive. Required ordinary
+    output/utility nodes are preserved regardless of their runtime OUTPUT_NODE
+    flag; callers must dispatch exactly the returned binding ``node_id`` values
+    as partial_execution_targets. Distributed scalar dependencies use the same
+    resolver as the legacy compiler. Direct scalar-to-media links need a typed
+    wrapper so that each export input remains a link.
+
+    Validation and compilation never modify the caller's source graph. This
+    structural compiler does not infer socket types or unknown nodes' output
+    counts; ComfyUI performs runtime type validation at the endpoint.
+    """
+    _validate_identifier(request_id, "request_id")
+    collector_ids = _validate_prompt(prepared_prompt)
+    bindings, exports = _build_exports(prepared_prompt, collector_ids, request_id)
+    compiled = _prune_graph(
+        prepared_prompt, bindings, exports, _make_scalar_resolver(prepared_prompt)
+    )
+    return {node_id: compiled[node_id] for node_id in sorted(compiled)}, bindings

api/orchestration/serverless_prompt.py

@@ -0,0 +1,420 @@
+"""Compile prepared distributed prompts into worker-side ComfyUI image jobs.
+
+This module is deliberately pure: it only copies and rewrites an API prompt
+mapping and never imports or executes ComfyUI nodes.
+"""
+
+from __future__ import annotations
+
+import copy
+import hashlib
+import json
+import re
+
+
+_SUPPORTED_DISTRIBUTED = {
+    "DistributedCollector",
+    "DistributedSeed",
+    "DistributedValue",
+    "DistributedModelName",
+}
+_REMOVED_SCALAR_NODES = {
+    "DistributedSeed",
+    "DistributedValue",
+    "DistributedModelName",
+}
+_PRIMITIVE_NODES = {
+    "PrimitiveBoolean": False,
+    "PrimitiveFloat": 0.0,
+    "PrimitiveInt": 0,
+    "PrimitiveString": "",
+}
+_REMOVED_NODE_OUTPUTS = {
+    "DistributedCollector": {0, 1},
+    "DistributedSeed": {0},
+    "DistributedValue": {0},
+    "DistributedModelName": {0},
+}
+_SEED_DEFAULT = 1125899906842
+
+
+def _is_link(value):
+    return isinstance(value, list) and len(value) == 2
+
+
+def _is_unsupported_distributed(class_type):
+    if not isinstance(class_type, str) or class_type in _SUPPORTED_DISTRIBUTED:
+        return False
+    lowered = class_type.lower()
+    return "distributed" in lowered or "batchdivider" in lowered
+
+
+def _safe_component(value):
+    cleaned = re.sub(r"[^A-Za-z0-9_-]+", "_", str(value)).strip("_-")
+    return cleaned[:24] or "empty"
+
+
+def _filename_prefix(job_id, worker_id, collector_id, kind):
+    identity = json.dumps(
+        [job_id, worker_id, collector_id, kind],
+        ensure_ascii=False,
+        separators=(",", ":"),
+    )
+    digest = hashlib.sha256(identity.encode("utf-8")).hexdigest()
+    parts = (
+        "serverless",
+        _safe_component(job_id),
+        _safe_component(worker_id),
+        _safe_component(collector_id),
+        kind,
+        digest,
+    )
+    return "_".join(parts)
+
+
+def _worker_index(worker_id):
+    """Mirror nodes.utilities worker-index parsing; None means fallback."""
+    try:
+        if worker_id.startswith("worker_"):
+            return int(worker_id.split("_")[1])
+        return int(worker_id)
+    except (ValueError, IndexError):
+        return None
+
+
+def _coerce(value, value_type):
+    if value_type == "INT":
+        return int(float(value))
+    if value_type == "FLOAT":
+        return float(value)
+    return value
+
+
+def _coerce_safe(value, value_type):
+    try:
+        return _coerce(value, value_type)
+    except (TypeError, ValueError):
+        return value
+
+
+def _stringify_model_name(value):
+    if isinstance(value, str):
+        return value
+    if isinstance(value, (int, float, bool)):
+        return str(value)
+    try:
+        return json.dumps(value, indent=4)
+    except (TypeError, ValueError):
+        return str(value)
+
+
+def _model_name_result(text):
+    if isinstance(text, list):
+        values = [_stringify_model_name(value) for value in text]
+        return values[0] if len(values) == 1 else values
+    return _stringify_model_name(text)
+
+
+def _validate_prepared_prompt(prepared_prompt):
+    """Validate graph shape, links, collectors, and supported topology."""
+    if not isinstance(prepared_prompt, dict):
+        raise ValueError("prepared prompt must be an API graph mapping")
+
+    source_nodes = {}
+    for raw_id, node in prepared_prompt.items():
+        node_id = str(raw_id)
+        if node_id in source_nodes:
+            raise ValueError("prompt has duplicate node IDs after string normalization")
+        if not isinstance(node, dict):
+            raise ValueError(f"prompt node {node_id!r} must be an object")
+        class_type = node.get("class_type")
+        if not isinstance(class_type, str):
+            raise ValueError(f"prompt node {node_id!r} is missing a class_type")
+        if _is_unsupported_distributed(class_type):
+            raise ValueError(f"unsupported distributed or batch node: {class_type}")
+        inputs = node.get("inputs", {})
+        if not isinstance(inputs, dict):
+            raise ValueError(f"prompt node {node_id!r} inputs must be an object")
+        source_nodes[node_id] = node
+
+    if not source_nodes:
+        raise ValueError("prepared prompt contains no nodes")
+
+    # Validate every API link before any graph rewriting. Slots on the nodes
+    # removed by this compiler have a known schema and are checked exactly.
+    for node_id, node in source_nodes.items():
+        for input_name, value in node.get("inputs", {}).items():
+            if not _is_link(value):
+                continue
+            upstream_id, output_slot = value
+            upstream_id = str(upstream_id)
+            if upstream_id not in source_nodes:
+                raise ValueError(
+                    f"invalid link at node {node_id!r} input {input_name!r}: "
+                    f"source {upstream_id!r} does not exist"
+                )
+            if isinstance(output_slot, bool) or not isinstance(output_slot, int) or output_slot < 0:
+                raise ValueError(
+                    f"invalid link at node {node_id!r} input {input_name!r}: "
+                    "output slot must be a nonnegative integer"
+                )
+            upstream_class = source_nodes[upstream_id]["class_type"]
+            valid_slots = _REMOVED_NODE_OUTPUTS.get(upstream_class)
+            if valid_slots is not None and output_slot not in valid_slots:
+                raise ValueError(
+                    f"invalid output slot {output_slot} for removed node "
+                    f"{upstream_class} {upstream_id!r}"
+                )
+
+    collector_ids = sorted(
+        (node_id for node_id, node in source_nodes.items() if node["class_type"] == "DistributedCollector"),
+        key=str,
+    )
+    if not collector_ids:
+        raise ValueError("prepared prompt contains no DistributedCollector nodes")
+
+    # Only image collectors are supported by the current generic handler.
+    for collector_id in collector_ids:
+        inputs = source_nodes[collector_id].get("inputs", {})
+        if "audio" in inputs:
+            raise ValueError("collector audio input is unsupported by this image-only compiler")
+        if not _is_link(inputs.get("images")):
+            raise ValueError(f"collector {collector_id!r} is missing a linked image input")
+        source_id = str(inputs["images"][0])
+        source_class = source_nodes[source_id]["class_type"]
+        if source_class in _REMOVED_SCALAR_NODES:
+            raise ValueError(f"collector {collector_id!r} image input must link to an image node")
+
+    # A collector cannot be used, directly or through processing, as another
+    # collector's input: this first compiler handles disconnected jobs only.
+    for collector_id in collector_ids:
+        pending = [str(source_nodes[collector_id]["inputs"]["images"][0])]
+        seen = set()
+        while pending:
+            node_id = pending.pop()
+            if node_id in seen:
+                continue
+            seen.add(node_id)
+            if source_nodes[node_id]["class_type"] == "DistributedCollector":
+                raise ValueError("collector-to-collector dependence is unsupported")
+            for value in source_nodes[node_id].get("inputs", {}).values():
+                if _is_link(value):
+                    pending.append(str(value[0]))
+
+    return source_nodes, collector_ids
+
+
+def _make_scalar_resolver(source_nodes):
+    """Build pure resolvers for distributed scalar nodes and stock primitives."""
+    # Resolve only stock primitive nodes. A visited set makes malformed
+    # primitive-to-primitive cycles fail clearly instead of recursing forever.
+    def resolve_primitive(node_id, visiting=None):
+        node_id = str(node_id)
+        node = source_nodes[node_id]
+        class_type = node["class_type"]
+        if class_type not in _PRIMITIVE_NODES:
+            raise ValueError("linked value must be literal or supported PrimitiveInt/Float/String/Boolean")
+        if visiting is None:
+            visiting = set()
+        if node_id in visiting:
+            raise ValueError(f"cycle while resolving primitive value at node {node_id!r}")
+        visiting.add(node_id)
+        value = node.get("inputs", {}).get("value", _PRIMITIVE_NODES[class_type])
+        if _is_link(value):
+            parent_id, output_slot = str(value[0]), value[1]
+            if output_slot != 0:
+                raise ValueError(f"invalid output slot {output_slot} for primitive node {parent_id!r}")
+            if source_nodes[parent_id]["class_type"] not in _PRIMITIVE_NODES:
+                raise ValueError("linked primitive value must come from a stock Primitive node")
+            result = resolve_primitive(parent_id, visiting)
+        else:
+            result = copy.deepcopy(value)
+        visiting.remove(node_id)
+        return result
+
+    def resolve_literal_or_primitive(value, context):
+        if _is_link(value):
+            source_id, output_slot = str(value[0]), value[1]
+            if output_slot != 0 or source_nodes[source_id]["class_type"] not in _PRIMITIVE_NODES:
+                raise ValueError(f"{context} must be a literal or supported Primitive link")
+            try:
+                return resolve_primitive(source_id)
+            except ValueError as exc:
+                if "cycle" in str(exc):
+                    raise
+                raise ValueError(f"{context} must be a literal or supported Primitive link") from None
+        return copy.deepcopy(value)
+
+    def resolve_hidden(node_id, name, default):
+        value = source_nodes[node_id].get("inputs", {}).get(name, default)
+        if _is_link(value):
+            raise ValueError(f"{name} on distributed node {node_id!r} must be a literal")
+        return copy.deepcopy(value)
+
+    resolved_distributed = {}
+
+    def resolve_distributed(node_id, output_slot):
+        node_id = str(node_id)
+        class_type = source_nodes[node_id]["class_type"]
+        if output_slot != 0:
+            raise ValueError(f"invalid output slot {output_slot} for {class_type} {node_id!r}")
+        cache_key = (node_id, output_slot)
+        if cache_key in resolved_distributed:
+            return copy.deepcopy(resolved_distributed[cache_key])
+        inputs = source_nodes[node_id].get("inputs", {})
+        if class_type == "DistributedSeed":
+            seed = resolve_literal_or_primitive(inputs.get("seed", _SEED_DEFAULT), f"seed on {node_id!r}")
+            is_worker = resolve_hidden(node_id, "is_worker", False)
+            worker_id = resolve_hidden(node_id, "worker_id", "")
+            if is_worker:
+                worker_index = _worker_index(worker_id)
+                if worker_index is not None:
+                    seed = seed + worker_index + 1
+            result = seed
+        elif class_type == "DistributedValue":
+            default_value = resolve_literal_or_primitive(
+                inputs.get("default_value", ""), f"default_value on {node_id!r}"
+            )
+            worker_values = resolve_literal_or_primitive(
+                inputs.get("worker_values", "{}"), f"worker_values on {node_id!r}"
+            )
+            is_worker = resolve_hidden(node_id, "is_worker", False)
+            worker_id = resolve_hidden(node_id, "worker_id", "")
+            try:
+                values = json.loads(worker_values) if isinstance(worker_values, str) else worker_values
+            except json.JSONDecodeError:
+                values = {}
+            if not isinstance(values, dict):
+                values = {}
+            value_type = values.get("_type", "STRING")
+            fallback = _coerce_safe(default_value, value_type)
+            result = fallback
+            if is_worker:
+                worker_index = _worker_index(worker_id)
+                if worker_index is not None:
+                    raw_value = values.get(str(worker_index + 1), "")
+                    if raw_value:
+                        try:
+                            result = _coerce(raw_value, value_type)
+                        except (ValueError, IndexError):
+                            result = fallback
+        elif class_type == "DistributedModelName":
+            text = resolve_literal_or_primitive(
+                inputs.get("text", ""), f"DistributedModelName text on {node_id!r}"
+            )
+            result = _model_name_result(text)
+        else:
+            raise ValueError(f"unsupported distributed node: {class_type}")
+        resolved_distributed[cache_key] = copy.deepcopy(result)
+        return result
+
+    return resolve_distributed
+
+
+def _build_savers(source_nodes, collector_ids):
+    """Construct stable SaveImage nodes and their collector metadata."""
+    # The prepared collector values define the artifact identity. The fallback
+    # for absent hidden values is stable and retains the collector's job scope.
+    collector_specs = []
+    max_numeric_id = 0
+    for node_id in source_nodes:
+        try:
+            max_numeric_id = max(max_numeric_id, int(node_id))
+        except ValueError:
+            pass
+    next_numeric_id = max_numeric_id + 1
+    for collector_id in collector_ids:
+        inputs = source_nodes[collector_id].get("inputs", {})
+        raw_job_id = inputs.get("multi_job_id", collector_id)
+        if _is_link(raw_job_id):
+            raise ValueError(f"collector {collector_id!r} multi_job_id must be a literal")
+        job_id = str(collector_id if raw_job_id in (None, "") else raw_job_id)
+        raw_worker_id = inputs.get("worker_id", "")
+        if _is_link(raw_worker_id):
+            raise ValueError(f"collector {collector_id!r} worker_id must be a literal")
+        worker_id = "" if raw_worker_id is None else str(raw_worker_id)
+        saver_id = str(next_numeric_id)
+        next_numeric_id += 1
+        prefix = _filename_prefix(job_id, worker_id, collector_id, "image")
+        collector_specs.append(
+            {
+                "collector_id": collector_id,
+                "job_id": job_id,
+                "node_id": saver_id,
+                "kind": "image",
+                "prefix": prefix,
+                "source_link": copy.deepcopy(inputs["images"]),
+            }
+        )
+
+    saver_nodes = {}
+    for spec in collector_specs:
+        saver_nodes[spec["node_id"]] = {
+            "class_type": "SaveImage",
+            "inputs": {
+                "images": spec["source_link"],
+                "filename_prefix": spec["prefix"],
+            },
+            "_meta": {"title": f"Serverless output: collector {spec['collector_id']} (image)"},
+        }
+
+    return collector_specs, saver_nodes
+
+
+def _prune_graph(source_nodes, collector_specs, saver_nodes, resolve_distributed):
+    """Keep saver ancestors, replacing distributed scalar links with literals."""
+    all_nodes = {node_id: copy.deepcopy(node) for node_id, node in source_nodes.items()}
+    all_nodes.update(copy.deepcopy(saver_nodes))
+    needed = set()
+    pending = [spec["node_id"] for spec in collector_specs]
+    while pending:
+        node_id = pending.pop()
+        if node_id in needed:
+            continue
+        needed.add(node_id)
+        node = all_nodes[node_id]
+        if node["class_type"] == "DistributedCollector":
+            raise ValueError("collector-to-collector dependence is unsupported")
+        for input_name, value in list(node.get("inputs", {}).items()):
+            if not _is_link(value):
+                continue
+            upstream_id, output_slot = str(value[0]), value[1]
+            upstream_class = source_nodes[upstream_id]["class_type"]
+            if upstream_class in _REMOVED_SCALAR_NODES:
+                node["inputs"][input_name] = resolve_distributed(upstream_id, output_slot)
+            else:
+                pending.append(upstream_id)
+
+    compiled = {
+        node_id: copy.deepcopy(node)
+        for node_id, node in all_nodes.items()
+        if node_id in needed
+    }
+    if any("distributed" in node["class_type"].lower() for node in compiled.values()):
+        raise ValueError("compiled prompt still contains a distributed dependency")
+
+    return compiled
+
+
+def compile_serverless_prompt(prepared_prompt: dict) -> tuple[dict, list[dict]]:
+    """Compile a prepared worker prompt to an image-only serverless graph.
+
+    Every DistributedCollector's original ``images`` input is connected to a
+    new SaveImage node. Only those savers and their upstream dependencies are
+    retained. DistributedSeed/Value/ModelName links are resolved as scalar
+    literals; unsupported node execution is never attempted.
+    """
+    source_nodes, collector_ids = _validate_prepared_prompt(prepared_prompt)
+    resolve_distributed = _make_scalar_resolver(source_nodes)
+    collector_specs, saver_nodes = _build_savers(source_nodes, collector_ids)
+    compiled = _prune_graph(
+        source_nodes,
+        collector_specs,
+        saver_nodes,
+        resolve_distributed,
+    )
+    manifest = [
+        {key: spec[key] for key in ("collector_id", "job_id", "node_id", "kind", "prefix")}
+        for spec in collector_specs
+    ]
+    return compiled, manifest

api/orchestration/serverless_runtime.py

@@ -0,0 +1,411 @@
+"""Master-owned asynchronous Runpod jobs, independent of worker callbacks."""
+import asyncio
+import time
+import uuid
+from pathlib import Path
+
+from .runpod_client import RunpodClient, RunpodError, AmbiguousSubmission
+from .serverless_ledger import JobLedger
+from .serverless_state import (
+    PROVIDER_STATES, PROVIDER_TERMINAL, belongs_to_pool, is_unresolved,
+    occupies_capacity, terminal_provider_status,
+)
+
+
+# Fixed per-worker cap on this master's outstanding Runpod jobs; equals the
+# Images per run maximum (MAX_JOBS_PER_RUN in utils/serverless_config.py), so
+# every image of a run is submitted at once. Deliberately not configurable or
+# derived from jobs_per_run: unresolved jobs from earlier runs (ambiguous
+# submissions, unconfirmed cancels) still count against it, and a constant
+# never has to change while such jobs exist.
+PER_WORKER_JOB_LIMIT = 16
+
+
+async def decode_results(output, manifest):
+    from .serverless_media import collect_images
+    import torch
+    images = await collect_images(output, manifest)
+    return {job_id: [torch.from_numpy(image).unsqueeze(0) for image in batch] for job_id, batch in images.items()}
+
+
+def state_directory():
+    import folder_paths
+    return Path(folder_paths.get_user_directory()) / "distributed" / "serverless"
+
+
+def get_secret_store():
+    from ...utils.serverless_config import SecretStore
+    return SecretStore(state_directory() / "credentials.json")
+
+
+def get_manager():
+    import server
+    from ...utils.network import get_client_session
+    ps = server.PromptServer.instance
+    manager = getattr(ps, "distributed_serverless", None)
+    if manager is None:
+        manager = ServerlessManager(ps, get_secret_store(), state_directory() / "jobs.json", get_client_session)
+        ps.distributed_serverless = manager
+    return manager
+
+
+class ServerlessManager:
+    def __init__(self, prompt_server, secrets, ledger_path, session_factory):
+        self.server = prompt_server
+        self.secrets = secrets
+        self.ledger = JobLedger(ledger_path)
+        self.session_factory = session_factory
+        self.tasks = {}
+        self.runs = {}
+        self.pool_limits = {}
+        self.health_cache = {}
+        self.health_locks = {}
+        self.poll_seconds = 2.0
+        self.capacity_changed = asyncio.Condition()
+        self.recovery_lock = asyncio.Lock()
+        self._recover_local_records()
+
+    def _recover_local_records(self):
+        # WAITING is committed before admission; it cannot have reached Runpod.
+        # SUBMITTING is an intent record, so a lost ID must remain reserved.
+        for record in self.ledger.all():
+            if record.get("status") == "WAITING" and not record.get("job_id"):
+                self.ledger.update(record["record_id"], status="CANCELLED", updated_at=time.time())
+            elif record.get("status") == "SUBMITTING" and not record.get("job_id"):
+                self.ledger.update(record["record_id"], status="UNKNOWN", updated_at=time.time())
+
+    def _task_is_active(self, record_id):
+        task = self.tasks.get(record_id)
+        return task is not None and not task.done()
+
+    def pool_is_active(self, worker_id):
+        """Protect both live local work and unresolved pre-restart provider work."""
+        for rid, run in self.runs.items():
+            pool_id = run["worker"].get("pool_id", run["worker"]["id"])
+            if str(pool_id) == str(worker_id) and self._task_is_active(rid):
+                return True
+        return any(belongs_to_pool(record, worker_id) and is_unresolved(record)
+                   for record in self.ledger.all())
+
+    async def _reserve_provider_capacity(self, run):
+        pool_id = run["worker"].get("pool_id", run["worker"]["id"])
+        async with self.capacity_changed:
+            while True:
+                self._check_master(run)
+                outstanding = sum(belongs_to_pool(record, pool_id) and occupies_capacity(record)
+                                  for record in self.ledger.all())
+                if outstanding < PER_WORKER_JOB_LIMIT:
+                    # Claim atomically before any await that could submit a billed job.
+                    self.ledger.update(run["record_id"], status="SUBMITTING", updated_at=time.time())
+                    return
+                await self.capacity_changed.wait()
+
+    async def _notify_capacity(self):
+        async with self.capacity_changed:
+            self.capacity_changed.notify_all()
+
+    async def _client(self, worker):
+        pool_id = worker.get("pool_id", worker["id"])
+        key = self.secrets.get(pool_id)
+        return RunpodClient(await self.session_factory(), worker["endpoint_id"], key)
+
+    async def health(self, worker, *, fresh=False):
+        pool_id = worker.get("pool_id", worker["id"])
+        cache_key = (pool_id, worker["endpoint_id"])
+        lock = self.health_locks.setdefault(cache_key, asyncio.Lock())
+        async with lock:
+            cached = self.health_cache.get(cache_key)
+            if cached and not fresh and time.monotonic() - cached[0] < 10:
+                return cached[1]
+            client = await self._client(worker)
+            data = await client.health()
+            jobs = data.get("jobs", {})
+            if not isinstance(jobs, dict) or not isinstance(data.get("workers"), dict):
+                raise RunpodError("Runpod returned an invalid endpoint health record")
+            try:
+                pending = max(0, int(jobs.get("inQueue", 0))) + max(0, int(jobs.get("inProgress", 0)))
+            except (ValueError, TypeError):
+                raise RunpodError("Runpod returned invalid queue counts") from None
+            result = {"ok": True, "queueRemaining": pending, "message": "Queued / processing" if pending else "Ready (scales from zero)"}
+            self.health_cache[cache_key] = (time.monotonic(), result)
+            return result
+
+    def _semaphore(self, worker):
+        # The limit is a constant, so it never has to be swapped while this
+        # worker still has unresolved jobs.
+        pool_id = worker.get("pool_id", worker["id"])
+        semaphore = self.pool_limits.get(pool_id)
+        if semaphore is None:
+            semaphore = asyncio.Semaphore(PER_WORKER_JOB_LIMIT)
+            self.pool_limits[pool_id] = semaphore
+        return semaphore
+
+    def ensure_capacity(self, count):
+        if sum(not task.done() for task in self.tasks.values()) + count > 64:
+            raise ValueError("Distributed already has too many pending serverless jobs (limit 64)")
+
+    async def start(self, execution_id, prompt_id, prepared_jobs):
+        self.ensure_capacity(len(prepared_jobs))
+        from .serverless_typed_results import validate_prepared_job
+        # Validate the entire batch before creating even the first local slot.
+        for prepared in prepared_jobs:
+            validate_prepared_job(prepared)
+        new_runs = []
+        # Journal every local slot before starting any billed submission.
+        for prepared in prepared_jobs:
+            worker = prepared["worker"]
+            semaphore = self._semaphore(worker)
+            record_id = uuid.uuid4().hex
+            now = time.time()
+            self.ledger.update(record_id, execution_id=execution_id, prompt_id=prompt_id,
+                               worker_id=worker.get("pool_id", worker["id"]), participant_id=worker["id"],
+                               endpoint_id=worker["endpoint_id"], job_id=None, status="WAITING", created_at=now, updated_at=now)
+            run = dict(prepared, execution_id=execution_id, prompt_id=prompt_id, record_id=record_id,
+                       provider_job_id=None, client=None, deadline=time.monotonic() + worker["timeout_seconds"])
+            self.runs[record_id] = run
+            new_runs.append((record_id, run, semaphore))
+        for record_id, run, semaphore in new_runs:
+            self.tasks[record_id] = asyncio.create_task(self._run(run, semaphore), name=f"distributed-serverless-{record_id}")
+        # Bound in-memory completed tracking as well as the on-disk journal.
+        done = [rid for rid, task in self.tasks.items() if task.done()]
+        for rid in done[:-200]:
+            self.tasks.pop(rid, None)
+            self.runs.pop(rid, None)
+
+    def is_pending(self, collector_job_id, participant_id):
+        return any(not task.done() and run["worker"]["id"] == participant_id and
+                   any(m["job_id"] == collector_job_id for m in run["manifest"])
+                   for rid, task in self.tasks.items() for run in [self.runs[rid]])
+
+    def owns(self, collector_job_id, participant_id):
+        return any(run["worker"]["id"] == participant_id and
+                   any(m["job_id"] == collector_job_id for m in run["manifest"]) for run in self.runs.values())
+
+    async def _deliver(self, run, images=None, error=None, *, media=None):
+        if media is not None and not error:
+            # Validate all destinations before exposing any partial success.
+            for binding in run["manifest"]:
+                if binding["participant_id"] != run["worker"]["id"]:
+                    raise ValueError("Serverless media participant identity mismatch")
+                values = media[binding["job_id"]][binding["socket"]]
+                if not isinstance(values, list) or not values:
+                    raise ValueError("Missing serverless media values")
+        async with self.server.distributed_jobs_lock:
+            delivered = set()
+            for binding in run["manifest"]:
+                job_id = binding["job_id"]
+                if error or media is not None:
+                    if job_id in delivered:
+                        continue
+                    delivered.add(job_id)
+                queue = self.server.distributed_pending_jobs.get(binding["job_id"])
+                if queue is None:
+                    continue  # A cancelled/finished collector must never be resurrected.
+                if error:
+                    queue.put_nowait({"worker_id": run["worker"]["id"], "serverless_error": error, "is_last": True})
+                elif media is not None:
+                    queue.put_nowait({"worker_id": run["worker"]["id"], "media": media[job_id], "is_last": True})
+                else:
+                    batch = images[binding["job_id"]]
+                    for index, tensor in enumerate(batch):
+                        queue.put_nowait({"worker_id": run["worker"]["id"], "image_index": index,
+                                          "tensor": tensor, "is_last": index == len(batch) - 1})
+
+    async def _cancel_provider(self, run):
+        if run["provider_job_id"] and run["client"]:
+            try:
+                result = await run["client"].cancel(run["provider_job_id"])
+                return terminal_provider_status(result)
+            except Exception:
+                return None
+        return None
+
+    def _check_master(self, run):
+        queue = getattr(self.server, "prompt_queue", None)
+        if queue is None:
+            return
+        history = queue.get_history(run["prompt_id"])
+        record = history.get(run["prompt_id"], {})
+        if record.get("status", {}).get("status_str") == "error":
+            raise RunpodError("Master workflow failed; serverless work stopped")
+        if not record and hasattr(queue, "get_current_queue"):
+            running, pending = queue.get_current_queue()
+            if not any(str(item[1]) == run["prompt_id"] for item in [*running, *pending]):
+                raise RunpodError("Master workflow was removed from the queue; serverless work stopped")
+
+    async def _run(self, run, semaphore):
+        rid = run["record_id"]
+        submitted = False
+        provider_terminal = False
+        try:
+            from .serverless_typed_results import validate_prepared_job, collect_typed_results
+            typed = validate_prepared_job(run)
+            async with asyncio.timeout(max(0, run["deadline"] - time.monotonic())):
+                async with semaphore:
+                    await self._reserve_provider_capacity(run)
+                    client = await self._client(run["worker"])
+                    run["client"] = client
+                    submitted = True
+                    if typed:
+                        job_id = await client.submit(run["workflow"], run["images"], run["worker"]["timeout_seconds"], media=run["media"])
+                    else:
+                        job_id = await client.submit(run["workflow"], run["images"], run["worker"]["timeout_seconds"])
+                    run["provider_job_id"] = job_id
+                    # Persist immediately: all subsequent actions use this exact ID.
+                    self.ledger.update(rid, job_id=job_id, status="IN_QUEUE", updated_at=time.time())
+                    while True:
+                        self._check_master(run)
+                        response = await client.status(job_id)
+                        status = response.get("status")
+                        if not isinstance(status, str) or status not in PROVIDER_STATES:
+                            raise RunpodError("Runpod returned an unknown job state")
+                        provider_terminal = status in PROVIDER_TERMINAL
+                        if status == "COMPLETED":
+                            output = response.get("output")
+                            if not isinstance(output, dict) or output.get("error") or output.get("errors"):
+                                raise RunpodError("Serverless workflow reported an output error; inspect the provider job")
+                            self.ledger.update(rid, status="DOWNLOADING", updated_at=time.time())
+                            if typed:
+                                media = await collect_typed_results(output, run["media"], run["manifest"])
+                                await self._deliver(run, media=media)
+                            else:
+                                images = await decode_results(output, run["manifest"])
+                                await self._deliver(run, images=images)
+                            self.ledger.update(rid, status="COMPLETED", updated_at=time.time())
+                            return
+                        self.ledger.update(rid, status=status, updated_at=time.time())
+                        if provider_terminal:
+                            raise RunpodError(f"Serverless job ended with {status}")
+                        await asyncio.sleep(self.poll_seconds)
+        except AmbiguousSubmission as exc:
+            self.ledger.update(rid, status="UNKNOWN", error=str(exc), updated_at=time.time())
+            await self._deliver(run, error=str(exc))
+        except (asyncio.TimeoutError, asyncio.CancelledError) as exc:
+            terminal = "FAILED" if provider_terminal else await self._cancel_provider(run)
+            if submitted and not run["provider_job_id"]:
+                status = "UNKNOWN"
+                error = "Submission interrupted before its job ID was received; check Runpod before retrying"
+            elif isinstance(exc, asyncio.TimeoutError):
+                status = terminal or ("CANCEL_REQUESTED" if run["provider_job_id"] else "TIMED_OUT")
+                error = "Serverless job exceeded its queue/execution deadline"
+            else:
+                status = terminal or ("CANCEL_REQUESTED" if run["provider_job_id"] else "CANCELLED")
+                error = "Serverless work stopped" if terminal or not submitted else "Serverless cancellation requested; provider confirmation is pending"
+            self.ledger.update(rid, status=status, error=error, updated_at=time.time())
+            await self._deliver(run, error=error)
+        except Exception as exc:
+            # If polling/decoding failed, stop only our own provider job, never the pool.
+            terminal = "FAILED" if provider_terminal else await self._cancel_provider(run)
+            error = str(exc) if isinstance(exc, RunpodError) else "Serverless processing failed; check the workflow and endpoint contract"
+            status = terminal or ("CANCEL_REQUESTED" if run["provider_job_id"] else "FAILED")
+            self.ledger.update(rid, status=status, error=error, updated_at=time.time())
+            await self._deliver(run, error=error)
+        finally:
+            # Completed tracking retains IDs/bindings only, not potentially large
+            # base64 inputs, workflow metadata, or a client holding credentials.
+            run.pop("workflow", None)
+            run.pop("images", None)
+            run.pop("media", None)
+            run["client"] = None
+            # A released local semaphore alone cannot release a durable reservation.
+            await self._notify_capacity()
+
+    async def wait_for_execution(self, execution_id):
+        tasks = [task for rid, task in self.tasks.items() if self.runs[rid]["execution_id"] == execution_id]
+        await asyncio.gather(*tasks, return_exceptions=False)
+
+    def _detached_records(self, worker_id=None, execution_id=None):
+        return [record for record in self.ledger.all()
+                if is_unresolved(record) and not self._task_is_active(record["record_id"])
+                and (worker_id is None or belongs_to_pool(record, worker_id))
+                and (execution_id is None or record.get("execution_id") == execution_id)]
+
+    async def _recover_provider_record(self, record, *, cancel=False):
+        """Reconcile an exact saved ID, never submit or resume a workflow."""
+        if not record.get("job_id"):
+            return False
+        status = None
+        try:
+            client = await self._client({"id": record["worker_id"], "endpoint_id": record["endpoint_id"]})
+            response = await (client.cancel(record["job_id"]) if cancel else client.status(record["job_id"]))
+            status = terminal_provider_status(response)
+        except Exception:
+            pass  # Unreachable/expired is not proof of provider termination.
+        if status is not None or cancel:
+            self.ledger.update(record["record_id"], status=status or "CANCEL_REQUESTED", updated_at=time.time())
+        return status is not None
+
+    async def reconcile(self, worker_id=None):
+        """Read terminal states for detached jobs and release only proven capacity."""
+        resolved = 0
+        async with self.recovery_lock:
+            for record in self._detached_records(worker_id):
+                resolved += await self._recover_provider_record(record)
+        await self._notify_capacity()
+        return resolved
+
+    async def acknowledge_recovery(self, worker_id, record_id, *, provider_checked=False):
+        """Explicit operator escape hatch for lost IDs/expired provider records.
+
+        This records a human's verification, not a cancellation or provider
+        status. It is never called automatically or allowed for live tasks.
+        """
+        if provider_checked is not True:
+            raise ValueError("Check Runpod and confirm that this job is no longer running before resolving it")
+        async with self.recovery_lock:
+            record = next((r for r in self._detached_records(worker_id) if r["record_id"] == record_id), None)
+            if record is None:
+                raise ValueError("Unresolved inactive job not found for this worker")
+            self.ledger.update(record_id, status="RESOLVED", updated_at=time.time(),
+                               error="Recovery explicitly acknowledged after operator checked Runpod; not provider-confirmed")
+        await self._notify_capacity()
+
+    async def cancel(self, worker_id=None, execution_id=None, collector_job_id=None):
+        targets = {}
+        for rid, task in self.tasks.items():
+            run = self.runs[rid]
+            if worker_id and run["worker"].get("pool_id", run["worker"]["id"]) != worker_id:
+                continue
+            if execution_id and run["execution_id"] != execution_id:
+                continue
+            if collector_job_id and not any(m["job_id"] == collector_job_id for m in run["manifest"]):
+                continue
+            if not task.done():
+                # Repeated requests must not interrupt the first cancellation's cleanup.
+                if not task.cancelling():
+                    task.cancel()
+                targets[rid] = task
+        if targets:
+            await asyncio.gather(*targets.values(), return_exceptions=True)
+        recovered = 0
+        async with self.recovery_lock:
+            for record in self._detached_records(worker_id, execution_id):
+                rid = record["record_id"]
+                # A task cancelled before its first step never enters _run's handlers.
+                # A failed start() may also leave a journalled but unstarted local slot.
+                if record.get("status") == "WAITING" and not record.get("job_id"):
+                    run = self.runs.get(rid)
+                    if run is not None and (collector_job_id is None or any(m["job_id"] == collector_job_id for m in run["manifest"])):
+                        self.ledger.update(rid, status="CANCELLED", updated_at=time.time())
+                        run.pop("workflow", None)
+                        run.pop("images", None)
+                        run.pop("media", None)
+                        await self._deliver(run, error="Serverless job cancelled before submission")
+                    continue
+                # Do not send two cancellation requests for the same live job in
+                # one action. Its unresolved reservation survives for later recovery.
+                if rid in targets or collector_job_id or not record.get("job_id"):
+                    continue
+                await self._recover_provider_record(record, cancel=True)
+                recovered += 1
+        await self._notify_capacity()
+        return len(targets) + recovered
+
+    def public_jobs(self, worker_id):
+        records = [record for record in reversed(self.ledger.all()) if belongs_to_pool(record, worker_id)]
+        # Never hide unresolved IDs behind the bounded completed history.
+        visible = [record for record in records if is_unresolved(record)]
+        visible.extend(record for record in records if not is_unresolved(record))
+        unresolved_count = sum(is_unresolved(record) for record in records)
+        return [{key: record.get(key) for key in ("record_id", "job_id", "status", "worker_id", "execution_id", "error")}
+                for record in visible[:max(20, unresolved_count)]]

api/orchestration/serverless_state.py

@@ -0,0 +1,29 @@
+"""Shared rules for provider capacity and durable recovery protection.
+
+A local task ending is not evidence that provider work ended. RESOLVED is an
+explicit operator acknowledgment, never a provider-reported state.
+"""
+
+PROVIDER_TERMINAL = frozenset({"COMPLETED", "FAILED", "CANCELLED", "TIMED_OUT"})
+TERMINAL = PROVIDER_TERMINAL | {"RESOLVED"}
+PROVIDER_STATES = PROVIDER_TERMINAL | {"IN_QUEUE", "IN_PROGRESS"}
+
+
+def terminal_provider_status(response):
+    status = response.get("status") if isinstance(response, dict) else None
+    return status if isinstance(status, str) and status in PROVIDER_TERMINAL else None
+
+
+def is_unresolved(record):
+    return record.get("status") not in TERMINAL
+
+
+def occupies_capacity(record):
+    # WAITING is journalled before admission and cannot have been submitted.
+    # Every other unresolved state is conservatively reserved, including a
+    # submission whose provider ID was lost and an unconfirmed cancellation.
+    return is_unresolved(record) and (record.get("status") != "WAITING" or bool(record.get("job_id")))
+
+
+def belongs_to_pool(record, worker_id):
+    return str(record.get("worker_id")) == str(worker_id)

api/orchestration/serverless_typed_inputs.py

@@ -0,0 +1,225 @@
+"""Bounded local native inputs for comfy-media-v1 requests.
+
+Only stock LoadAudio/LoadVideo schemas are adapted. Original nodes and links
+remain in the private request graph for endpoint validation; only the existing
+image preparer rewrites filenames. No upload, transcode, resample or provider
+operation belongs here. The caller still enforces the complete request-body cap.
+"""
+import hashlib
+import io
+import json
+from pathlib import Path
+
+from ...serverless_bridge import codec, contract, video_codec
+from . import serverless_media
+
+
+_LOADERS = {"LoadAudio": ("audio", "AUDIO"), "LoadVideo": ("file", "VIDEO")}
+Error = serverless_media.ServerlessMediaError
+
+
+def _loader_references(workflow):
+    if not isinstance(workflow, dict):
+        raise Error("Serverless workflow must be an API graph")
+    loaders = {}
+    for node_id, node in workflow.items():
+        if not isinstance(node, dict):
+            continue
+        class_type = node.get("class_type")
+        if not isinstance(class_type, str) or class_type not in _LOADERS:
+            continue
+        contract.identifier(node_id)
+        key, kind = _LOADERS[class_type]
+        inputs = node.get("inputs")
+        if not isinstance(inputs, dict) or set(inputs) != {key}:
+            raise Error(f"{class_type} inputs must contain only its native filename; extra controls are unsupported")
+        if not isinstance(inputs[key], str):
+            raise Error(f"{class_type} requires a literal filename; linked inputs are unsupported")
+        normalized, _suffix = serverless_media._normalize_media_reference(inputs[key])
+        loaders[node_id] = (kind, normalized)
+
+    references = []
+    used = set()
+    for node_id, node in workflow.items():
+        if not isinstance(node, dict) or not isinstance(node.get("inputs"), dict):
+            continue
+        for key, value in node["inputs"].items():
+            if not isinstance(value, (list, tuple)) or not value or not isinstance(value[0], str):
+                continue
+            if value[0] not in loaders:
+                continue
+            if type(value) is not list or len(value) != 2 or type(value[1]) is not int or value[1] != 0:
+                raise Error("Native media loader links require output slot 0")
+            contract.identifier(node_id)
+            contract.identifier(key)
+            references.append((node_id, key, value[0]))
+            used.add(value[0])
+    if used != set(loaders):
+        raise Error("Native media loader has no consumer")
+    return loaders, references
+
+
+def _source_path(value, folder_paths):
+    """Resolve native annotations; authorization/opening is owned by _read_path."""
+    annotation = serverless_media._MEDIA_ANNOTATION.search(value)
+    unannotated = value[:annotation.start()] if annotation else value
+    if Path(unannotated).is_absolute():
+        return unannotated
+    try:
+        return folder_paths.get_annotated_filepath(value)
+    except Exception:
+        raise Error("Native media input path is missing or invalid") from None
+
+
+def _decode_audio(data, limit):
+    """Match core nodes_audio.load, bounding each frame *before* accumulation.
+
+    Float32 output bytes are charged even for narrow integer PCM. Native float64
+    remains float64, as in f32_pcm (there is no silent dtype conversion). Raw
+    retained arrays and each live AV plane allocation have a separate <= limit
+    guard, so wider/padded raw frames cannot bypass the output budget. PyAV owns
+    the one frame being decoded; it is checked before to_ndarray or torch.cat.
+    Concatenation/conversion use bounded temporary copies, not an RSS guarantee.
+    """
+    if limit <= 0:
+        raise Error("AUDIO decoded input exceeds byte limit")
+    try:
+        import av
+        import torch
+        from comfy_extras.nodes_audio import f32_pcm
+    except ImportError:
+        raise Error("Native AUDIO decoding is unavailable") from None
+
+    try:
+        with av.open(io.BytesIO(data)) as container:
+            if not container.streams.audio:
+                raise Error("Native AUDIO input has no audio stream")
+            stream = container.streams.audio[0]
+            rate = stream.codec_context.sample_rate
+            channels = stream.channels
+            if type(rate) is not int or rate <= 0 or type(channels) is not int or channels <= 0:
+                raise Error("Native AUDIO stream shape or sample rate is invalid")
+            frames = []
+            output_bytes = raw_bytes = 0
+            dtype = None
+            for frame in container.decode(streams=stream.index):
+                samples = frame.samples
+                if (type(samples) is not int or samples <= 0 or frame.sample_rate != rate
+                        or len(frame.layout.channels) != channels):
+                    raise Error("Native AUDIO frame shape or sample rate changed")
+                if frame.format.name not in {"s16", "s16p", "s32", "s32p", "flt", "fltp", "dbl", "dblp"}:
+                    raise Error("Unsupported native AUDIO PCM format")
+                itemsize = frame.format.bytes
+                count = samples * channels
+                # Native f32_pcm promotes integer PCM to float32, not float64.
+                decoded_bytes = count * max(4, itemsize)
+                frame_bytes = count * itemsize
+                plane_bytes = sum(plane.buffer_size for plane in frame.planes)
+                if (decoded_bytes > limit - output_bytes or frame_bytes > limit - raw_bytes
+                        or plane_bytes > limit - raw_bytes):
+                    raise Error("AUDIO decoded input or raw frame exceeds byte limit")
+                output_bytes += decoded_bytes
+                raw_bytes += frame_bytes
+                array = frame.to_ndarray()
+                if array.ndim != 2 or array.size != count or array.nbytes != frame_bytes:
+                    raise Error("Native AUDIO frame shape is invalid")
+                buf = torch.from_numpy(array)
+                if buf.dtype not in (torch.int16, torch.int32, torch.float32, torch.float64):
+                    raise Error("Unsupported native AUDIO PCM dtype")
+                if dtype is not None and buf.dtype != dtype:
+                    raise Error("Native AUDIO frame dtype changed")
+                dtype = buf.dtype
+                if buf.shape[0] != channels:
+                    buf = buf.view(-1, channels).t()
+                if tuple(buf.shape) != (channels, samples):
+                    raise Error("Native AUDIO frame shape is invalid")
+                frames.append(buf)
+            if not frames:
+                raise Error("Native AUDIO input has no decoded frames")
+            waveform = f32_pcm(torch.cat(frames, dim=1)).unsqueeze(0)
+        return {"waveform": waveform, "sample_rate": rate}
+    except Error:
+        raise
+    except Exception:
+        raise Error("Native AUDIO input could not be decoded safely") from None
+
+
+def _artifact_id(kind, descriptor, payload):
+    digest = hashlib.sha256()
+    digest.update(kind.encode("ascii") + b"\0")
+    # Include interpretation (not just waveform bytes), notably sample_rate.
+    digest.update(json.dumps(descriptor, sort_keys=True, separators=(",", ":"), allow_nan=False).encode())
+    digest.update(b"\0")
+    digest.update(payload)
+    return "input-" + kind.lower() + "-" + digest.hexdigest()
+
+
+async def prepare_media_inputs(workflow: dict) -> tuple[list, list, list]:
+    """Return legacy images, inline typed artifacts, and every consumer reference.
+
+    Paths are read once into bounded bytes through the video adapter's no-follow
+    directory-fd reader. Decoders receive BytesIO, never a second path open.
+    Source files and unique decoded artifacts have independent aggregate caps.
+    Repeated content of the same kind is decoded once and charged once.
+    """
+    loaders, references = _loader_references(workflow)
+    image_view = {node_id: node for node_id, node in workflow.items() if node_id not in loaders}
+    # Retain the strict scan for every other loader before expensive decoding.
+    serverless_media._iter_image_references(image_view)
+    if not loaders:
+        return await serverless_media.prepare_images(image_view), [], []
+    try:
+        import folder_paths
+    except ImportError:
+        raise Error("ComfyUI media directories are unavailable") from None
+    roots = list(serverless_media._folder_roots(folder_paths).values())
+    artifacts = {}
+    source_artifacts = {}
+    source_digests = set()
+    filenames = {}
+    loader_artifacts = {}
+    file_total = decoded_total = 0
+    for node_id, (kind, filename) in loaders.items():
+        # Avoid retaining file buffers merely to deduplicate repeat references.
+        if (kind, filename) in filenames:
+            loader_artifacts[node_id] = filenames[kind, filename]
+            continue
+        path = _source_path(filename, folder_paths)
+        data = video_codec._read_path(path, contract.MAX_ARTIFACT_BYTES, roots)
+        source_digest = hashlib.sha256(data).hexdigest()
+        if source_digest not in source_digests:
+            if len(data) > contract.MAX_TOTAL_BYTES - file_total:
+                raise Error("Native media source files exceed total byte limit")
+            file_total += len(data)
+            source_digests.add(source_digest)
+        source_key = (kind, source_digest)
+        if source_key not in source_artifacts:
+            if len(artifacts) >= contract.MAX_ARTIFACTS:
+                raise Error("Native media input artifact count exceeds limit")
+            limit = min(contract.MAX_ARTIFACT_BYTES, contract.MAX_TOTAL_BYTES - decoded_total)
+            if kind == "AUDIO":
+                value = _decode_audio(data, limit)
+                descriptor, payload = codec.encode_payload(kind, value, max_bytes=limit)
+                del value
+            else:
+                from comfy_api.latest import InputImpl
+                value = InputImpl.VideoFromFile(io.BytesIO(data))
+                descriptor, payload = video_codec.encode_video(value, max_bytes=limit, allowed_roots=roots)
+                del value
+            aid = _artifact_id(kind, descriptor, payload)
+            if aid not in artifacts:
+                if len(payload) > limit:
+                    raise Error("Native media decoded inputs exceed total byte limit")
+                artifacts[aid] = contract.artifact_from_payload(
+                    aid, kind, descriptor, payload, max_bytes=contract.MAX_ARTIFACT_BYTES)
+                decoded_total += len(payload)
+            source_artifacts[source_key] = aid
+            del payload
+        aid = source_artifacts[source_key]
+        loader_artifacts[node_id] = filenames[kind, filename] = aid
+        del data
+    refs = [{"node_id": node_id, "input_name": key, "artifact_id": loader_artifacts[source]}
+            for node_id, key, source in references]
+    # Shared node dictionaries deliberately propagate only legacy image rewrites.
+    images = await serverless_media.prepare_images(image_view)
+    return images, list(artifacts.values()), refs

api/orchestration/serverless_typed_results.py

@@ -0,0 +1,79 @@
+"""Explicit media-v1 preflight and atomic, lossless result materialization."""
+from urllib.parse import urljoin
+
+import aiohttp
+
+from ...serverless_bridge import contract
+from ...serverless_bridge.contract import MAX_ARTIFACT_BYTES, PROFILE, validate_bindings
+from .serverless_media import (
+    MAX_REDIRECTS, _CHUNK_SIZE, _DOWNLOAD_TIMEOUT_SECONDS,
+    _PublicResolver, _validate_artifact_url,
+)
+
+
+def validate_prepared_job(prepared):
+    """Fail before journalling/submission; legacy jobs keep their original path."""
+    profile = prepared["worker"].get("media_profile", "legacy-image-v1")
+    media = prepared.get("media")
+    if profile == "legacy-image-v1" and media is None:
+        return False
+    if profile != PROFILE or media is None:
+        raise ValueError("Serverless media profile/request mismatch")
+    from .runpod_client import serialize_request
+    # The shared endpoint validator also checks exporter controls and input refs.
+    serialize_request(prepared["workflow"], prepared["images"],
+                      prepared["worker"]["timeout_seconds"], media=media)
+    expected = validate_bindings(prepared["manifest"])
+    if expected != validate_bindings(media["outputs"]):
+        raise ValueError("Serverless media manifest identity mismatch")
+    if any(binding["participant_id"] != prepared["worker"]["id"] for binding in expected.values()):
+        raise ValueError("Serverless media participant identity mismatch")
+    return True
+
+
+async def download_public_artifact(url, *, max_bytes=MAX_ARTIFACT_BYTES):
+    """Contract-sized public HTTPS download, never using a provider session/key."""
+    if type(max_bytes) is not int or not 0 <= max_bytes <= MAX_ARTIFACT_BYTES:
+        raise ValueError("Media artifact exceeds the permitted size")
+    current_url = _validate_artifact_url(url)
+    connector = aiohttp.TCPConnector(resolver=_PublicResolver(), use_dns_cache=False, limit=4)
+    timeout = aiohttp.ClientTimeout(total=_DOWNLOAD_TIMEOUT_SECONDS, connect=10, sock_read=15)
+    try:
+        async with aiohttp.ClientSession(connector=connector, timeout=timeout, trust_env=False,
+                                         cookie_jar=aiohttp.DummyCookieJar()) as session:
+            for redirect_count in range(MAX_REDIRECTS + 1):
+                async with session.get(current_url, allow_redirects=False) as response:
+                    if response.status in {301, 302, 303, 307, 308}:
+                        location = response.headers.get("Location")
+                        if not location or redirect_count >= MAX_REDIRECTS:
+                            raise ValueError("Media artifact redirect is invalid")
+                        current_url = _validate_artifact_url(urljoin(current_url, location))
+                        continue
+                    if response.status != 200:
+                        raise ValueError("Media artifact download failed")
+                    payload = bytearray()
+                    async for chunk in response.content.iter_chunked(_CHUNK_SIZE):
+                        if not isinstance(chunk, bytes) or len(payload) + len(chunk) > max_bytes:
+                            raise ValueError("Media artifact exceeds the permitted size")
+                        payload.extend(chunk)
+                    return bytes(payload)
+    except Exception:
+        # Signed URLs and transport exception details must not reach the ledger.
+        raise ValueError("Media artifact download failed or exceeded its byte limit") from None
+    raise ValueError("Media artifact download failed")
+
+
+async def collect_typed_results(output, media, manifest, *, download=None):
+    """Decode every binding before returning any collector's native value lists."""
+    if not isinstance(output, dict) or output.get("error") or output.get("errors"):
+        raise ValueError("Invalid serverless media output")
+    decoded = await contract.decode_response(output.get("media"), request_id=media["request_id"],
+                                    bindings=manifest,
+                                    download=download if download is not None else download_public_artifact)
+    grouped = {}
+    for binding in manifest:
+        values = decoded[binding["binding_id"]]
+        if not isinstance(values, list) or not values:
+            raise ValueError("Missing serverless media values")
+        grouped.setdefault(binding["job_id"], {}).setdefault(binding["socket"], []).extend(values)
+    return grouped

api/queue_orchestration.py

@@ -1,4 +1,5 @@
 import asyncio
+import copy
 import time
 import uuid
 
@@ -35,7 +36,10 @@ from .orchestration.prompt_transform import (
     prepare_delegate_master_prompt,
     prune_prompt_for_worker,
     WORKER_SINK_CLASS,
+    _resolve_direct_workers_for_image_list,
 )
+from ..image_list.runtime import validate_image_list_regions
+from ..image_list.execution_plan import prepare_image_list_execution
 
 
 prompt_server = server.PromptServer.instance
@@ -80,6 +84,18 @@ def _resolve_enabled_workers(config, requested_ids=None):
         elif not worker.get("enabled", False):
             continue
 
+        if worker.get("type") == "serverless":
+            from ..utils.serverless_config import validate_serverless_worker
+            pool = validate_serverless_worker(worker)
+            if pool.get("media_profile") == "comfy-media-v1":
+                from ..serverless_bridge.contract import identifier
+                # Do not repair malformed media identities by stripping them.
+                identifier(pool["id"])
+            for slot in range(pool["jobs_per_run"]):
+                workers.append(dict(pool, pool_id=worker_id, id=f"serverless:{worker_id}:{slot + 1}",
+                                    name=f"{pool.get('name', worker_id)} / job {slot + 1}"))
+            continue
+
         raw_port = worker.get("port", worker.get("listen_port", 8188))
         try:
             port = int(raw_port or 8188)
@@ -144,6 +160,32 @@ def _prompt_requests_load_balance(prompt_index):
     return False
 
 
+def _validate_media_workers(workers, prompt_index=None):
+    """Reject unknown serverless media profiles before probing or dispatching.
+
+    A Distributed Collector accepts direct workers, legacy-image-v1 pools and
+    comfy-media-v1 pools in the same run; each pool's profile selects its
+    endpoint contract. Unknown profiles fail rather than guessing a contract.
+    legacy-image-v1 endpoints return 8-bit files, so a Collector set to exact
+    cannot use them.
+    """
+    exact = prompt_index is not None and any(
+        str(prompt_index.inputs_by_node.get(node_id, {}).get("transfer_format", "")).strip().lower() == "exact"
+        for node_id in prompt_index.nodes_for_class("DistributedCollector")
+    )
+    for worker in workers:
+        if worker.get("type") != "serverless":
+            continue
+        profile = worker.get("media_profile", "legacy-image-v1")
+        if profile not in ("legacy-image-v1", "comfy-media-v1"):
+            raise ValueError("Unsupported serverless media_profile")
+        if exact and profile == "legacy-image-v1":
+            raise ValueError(
+                "transfer_format exact requires comfy-media-v1 serverless pools; "
+                "legacy-image-v1 endpoints return 8-bit images"
+            )
+
+
 async def _prepare_worker_payload(
     worker,
     prompt_index,
@@ -159,7 +201,44 @@ async def _prepare_worker_payload(
 ):
     """Prepare one worker prompt payload with bounded concurrency and media-sync timeout."""
     async with worker_prep_semaphore:
+        _validate_media_workers([worker], prompt_index)
         worker_prompt = prompt_index.copy_prompt()
+        if worker.get("type") == "serverless":
+            from .orchestration.runpod_client import serialize_request
+            if worker.get("media_profile", "legacy-image-v1") == "comfy-media-v1":
+                from ..serverless_bridge.contract import PROFILE, VERSION
+                from .orchestration.serverless_media_prompt import compile_media_prompt
+                from .orchestration.serverless_typed_inputs import prepare_media_inputs
+                # The typed compiler validates the complete private graph and
+                # prunes it itself, retaining every connected Collector branch.
+                worker_prompt = apply_participant_overrides(
+                    worker_prompt, worker["id"], enabled_ids, job_id_map,
+                    "", delegate_master, prompt_index,
+                )
+                request_id = str(uuid.uuid4())
+                workflow, manifest = compile_media_prompt(worker_prompt, request_id)
+                images, artifacts, refs = await prepare_media_inputs(workflow)
+                media = {
+                    "version": VERSION, "profile": PROFILE, "request_id": request_id,
+                    "artifacts": artifacts, "inputs": refs, "outputs": copy.deepcopy(manifest),
+                }
+                serialize_request(workflow, images, worker["timeout_seconds"], media=media)
+                return worker, {"worker": worker, "workflow": workflow, "manifest": manifest,
+                                "images": images, "media": media}
+
+            from .orchestration.serverless_prompt import compile_serverless_prompt
+            from .orchestration.serverless_media import prepare_images
+            worker_prompt = apply_participant_overrides(
+                prune_prompt_for_worker(worker_prompt), worker["id"], enabled_ids,
+                job_id_map, "", delegate_master, prompt_index,
+            )
+            workflow, manifest = compile_serverless_prompt(worker_prompt)
+            images = await prepare_images(workflow)
+            # Fail during preparation, before regular workers or the master are
+            # dispatched. submit() uses this same serializer for the sent bytes.
+            serialize_request(workflow, images, worker["timeout_seconds"])
+            return worker, {"worker": worker, "workflow": workflow, "manifest": manifest, "images": images}
+
         worker_master_url = build_master_callback_url(
             worker,
             config=config,
@@ -229,6 +308,8 @@ async def orchestrate_distributed_execution(
     execution_trace_id = trace_execution_id or _generate_execution_trace_id()
 
     config = load_config()
+    if delegate_master is None:
+        delegate_master = bool(config.get("settings", {}).get("master_delegate_only", False))
     use_websocket = bool(config.get("settings", {}).get("websocket_orchestration", False))
     master_url = build_master_url(config=config, prompt_server_instance=prompt_server)
     (
@@ -238,8 +319,53 @@ async def orchestrate_distributed_execution(
         media_sync_timeout_seconds,
     ) = _resolve_orchestration_limits(config)
     requested_ids = enabled_worker_ids if enabled_worker_ids is not None else None
-    workers = _resolve_enabled_workers(config, requested_ids)
+    image_list_mode = any(isinstance(node, dict) and node.get("class_type") == "DistributedImageListStart"
+                          for node in prompt_obj.values())
+    target_kwargs = {}
+    if image_list_mode:
+        # Check the source graph before execution-root pruning can remove an
+        # incompatible mode (or mask it with an unrelated boundary error).
+        if any(isinstance(node, dict) and node.get("class_type") in (
+                "DistributedCollector", "UltimateSDUpscaleDistributed") for node in prompt_obj.values()):
+            raise ValueError(
+                "Distributed Image List v1 cannot be mixed with Distributed Collector or "
+                "Ultimate SD Upscale nodes in one prompt"
+            )
+        prompt_obj, outer_targets = prepare_image_list_execution(prompt_obj)
+        target_kwargs["partial_execution_targets"] = outer_targets
     prompt_index = PromptIndex(prompt_obj)
+    if image_list_mode:
+        # Validate before probing or queueing anything so malformed boundaries and
+        # old worker releases fail before any upstream image is materialized.
+
+        validate_image_list_regions(prompt_obj, explicit_targets=True)
+        configured_serverless = [
+            str(worker.get("id")) for worker in config.get("workers", [])
+            if worker.get("enabled", False)
+            and str(worker.get("type") or "local").lower() == "serverless"
+            and (requested_ids is None or str(worker.get("id")) in {str(x) for x in requested_ids})
+        ]
+        if configured_serverless:
+            trace_debug(
+                execution_trace_id,
+                "Image List v1 explicitly excludes unsupported serverless pools: "
+                + ", ".join(configured_serverless),
+            )
+        workers = _resolve_direct_workers_for_image_list(config, requested_ids)
+        if not workers and delegate_master:
+            raise RuntimeError(
+                "Distributed Image List in orchestrator-only mode requires a selected direct worker; "
+                "serverless pools are unsupported"
+            )
+    else:
+        workers = _resolve_enabled_workers(config, requested_ids)
+        _validate_media_workers(workers, prompt_index)
+        if any(worker.get("media_profile") == "comfy-media-v1" for worker in workers):
+            from ..serverless_bridge.contract import identifier
+            # Validate the source, not PromptIndex's JSON copy (which coerces
+            # integer keys to strings), before probing or dispatching workers.
+            for node_id in prompt_obj:
+                identifier(node_id)
     load_balance_requested = _prompt_requests_load_balance(prompt_index)
     trace_debug(
         execution_trace_id,
@@ -255,26 +381,42 @@ async def orchestrate_distributed_execution(
         ),
     )
 
-    # Respect master delegate-only configuration
-    if delegate_master is None:
-        delegate_master = bool(config.get("settings", {}).get("master_delegate_only", False))
-
-    if not workers and delegate_master:
+    # Keep the legacy Collector fallback, but never override image-list
+    # orchestrator-only mode merely because its direct workers are unavailable.
+    if not image_list_mode and not workers and delegate_master:
         trace_debug(
             execution_trace_id,
             "Delegate-only requested but no workers are enabled. Falling back to master execution.",
         )
         delegate_master = False
 
-    active_workers, delegate_master = await select_active_workers(
+    active_workers, selected_delegate_master = await select_active_workers(
         workers,
         use_websocket,
-        delegate_master,
+        False if image_list_mode else delegate_master,
         trace_execution_id=execution_trace_id,
         probe_concurrency=worker_probe_concurrency,
     )
+    if image_list_mode:
+        from .orchestration.dispatch import preflight_image_list_workers
+        if not active_workers and delegate_master:
+            raise RuntimeError(
+                "No selected direct worker is reachable for Distributed Image List in orchestrator-only mode"
+            )
+        from ..image_list.runtime import END_CLASS, end_override_names
+        requires_values = any(end_override_names(node.get("inputs", {}))
+                              for node in prompt_obj.values() if node.get("class_type") == END_CLASS)
+        value_preflight = {"require_value_results": True} if requires_values else {}
+        # Persist candidates, not this submission-time idle snapshot: a preceding
+        # master prompt may still be using them. The image-list runtime rechecks
+        # capability and availability before assigning the first item.
+        if active_workers:
+            await preflight_image_list_workers(active_workers, require_idle=False,
+                                               **value_preflight)
+    else:
+        delegate_master = selected_delegate_master
 
-    if load_balance_requested:
+    if load_balance_requested and not image_list_mode:
         candidate_workers = list(active_workers)
         if not delegate_master:
             # Include master in load balancing only when master participation is enabled.
@@ -337,6 +479,7 @@ async def orchestrate_distributed_execution(
             workflow_meta,
             client_id,
             include_queue_metadata=True,
+            **target_kwargs,
         )
         return (
             queue_result["prompt_id"],
@@ -345,9 +488,6 @@ async def orchestrate_distributed_execution(
             queue_result.get("node_errors", {}),
         )
 
-    for job_id in job_id_map.values():
-        await _ensure_distributed_queue(job_id)
-
     master_prompt = prompt_index.copy_prompt()
     master_prompt = apply_participant_overrides(
         master_prompt,
@@ -357,9 +497,10 @@ async def orchestrate_distributed_execution(
         master_url,
         delegate_master,
         prompt_index,
+        workflow_meta=workflow_meta,
     )
 
-    if delegate_master:
+    if delegate_master and not image_list_mode:
         collector_ids = find_nodes_by_class(master_prompt, "DistributedCollector")
         upscale_nodes = find_nodes_by_class(master_prompt, "UltimateSDUpscaleDistributed")
         if upscale_nodes:
@@ -380,7 +521,7 @@ async def orchestrate_distributed_execution(
             + ", ".join(f"{worker['name']} ({worker['id']})" for worker in active_workers),
         )
     worker_payloads = []
-    if active_workers:
+    if active_workers and not image_list_mode:
         worker_prep_semaphore = asyncio.Semaphore(worker_prep_concurrency)
         media_sync_semaphore = asyncio.Semaphore(media_sync_concurrency)
         worker_payloads = await asyncio.gather(
@@ -402,7 +543,29 @@ async def orchestrate_distributed_execution(
             ]
         )
 
-    if worker_payloads:
+    serverless_jobs = [payload for worker, payload in worker_payloads if worker.get("type") == "serverless"]
+    regular_payloads = [(worker, payload) for worker, payload in worker_payloads if worker.get("type") != "serverless"]
+    manager = None
+    if serverless_jobs:
+        from .orchestration.serverless_runtime import get_manager
+        # Repeat the manager's whole-batch identity/profile preflight *before*
+        # any direct dispatch, master queueing or collector queue allocation.
+        # Pure legacy jobs are a no-op in that validator; keep their path free
+        # of typed imports (and compatible with existing legacy media adapters).
+        for prepared in serverless_jobs:
+            if (prepared["worker"].get("media_profile", "legacy-image-v1") != "legacy-image-v1"
+                    or prepared.get("media") is not None):
+                from .orchestration.serverless_typed_results import validate_prepared_job
+                validate_prepared_job(prepared)
+        manager = get_manager()
+        manager.ensure_capacity(len(serverless_jobs))
+
+    # Do not leave collector queues behind when offline compilation/media validation fails.
+    if not image_list_mode:
+        for job_id in job_id_map.values():
+            await _ensure_distributed_queue(job_id)
+
+    if regular_payloads:
         await asyncio.gather(
             *[
                 dispatch_worker_prompt(
@@ -413,21 +576,45 @@ async def orchestrate_distributed_execution(
                     use_websocket=use_websocket,
                     trace_execution_id=execution_trace_id,
                 )
-                for worker, wprompt in worker_payloads
+                for worker, wprompt in regular_payloads
             ]
         )
 
-    queue_result = await queue_prompt_payload(
-        master_prompt,
-        workflow_meta,
-        client_id,
-        include_queue_metadata=True,
-    )
+    try:
+        queue_result = await queue_prompt_payload(
+            master_prompt,
+            workflow_meta,
+            client_id,
+            include_queue_metadata=True,
+            **target_kwargs,
+        )
+    except Exception:
+        if serverless_jobs:
+            async with prompt_server.distributed_jobs_lock:
+                for job_id in job_id_map.values():
+                    prompt_server.distributed_pending_jobs.pop(job_id, None)
+        raise
     prompt_id = queue_result["prompt_id"]
+    if manager is not None:
+        try:
+            await manager.start(execution_trace_id, prompt_id, serverless_jobs)
+        except Exception:
+            # Wake queued collectors with an explicit error if journal/capacity setup failed.
+            await manager.cancel(execution_id=execution_trace_id)
+            async with prompt_server.distributed_jobs_lock:
+                for prepared in serverless_jobs:
+                    # Typed manifests have several sockets per collector, but
+                    # each participant gets only one terminal collector event.
+                    for job_id in {binding["job_id"] for binding in prepared["manifest"]}:
+                        q = prompt_server.distributed_pending_jobs.get(job_id)
+                        if q is not None:
+                            q.put_nowait({"worker_id": prepared["worker"]["id"], "is_last": True,
+                                          "serverless_error": "Serverless job setup failed; no automatic retry was attempted"})
+            raise
     prompt_number = queue_result["number"]
     node_errors = queue_result.get("node_errors", {})
     trace_debug(
         execution_trace_id,
         f"Orchestration complete: prompt_id={prompt_id}, dispatched_workers={len(worker_payloads)}, delegate_master={delegate_master}",
     )
-    return prompt_id, prompt_number, len(worker_payloads), node_errors
+    return prompt_id, prompt_number, (len(active_workers) if image_list_mode else len(worker_payloads)), node_errors

api/serverless_routes.py

@@ -0,0 +1,98 @@
+"""Master-only Runpod pool monitoring. Health checks never submit inference."""
+from aiohttp import web
+import server
+
+from ..utils.config import load_config
+from .orchestration.serverless_runtime import get_manager
+from .orchestration.runpod_client import RunpodError
+
+
+def _pool(worker_id):
+    return next((w for w in load_config().get("workers", [])
+                 if w.get("id") == worker_id and w.get("type") == "serverless"), None)
+
+
+async def _status(worker_id, fresh=False):
+    worker = _pool(worker_id)
+    if worker is None:
+        return web.json_response({"ok": False, "message": "Serverless worker not found"}, status=404)
+    jobs = []
+    try:
+        manager = get_manager()
+        jobs = manager.public_jobs(worker_id)
+        result = await manager.health(worker, fresh=fresh)
+        return web.json_response(dict(result, jobs=jobs))
+    except Exception as exc:
+        message = str(exc) if isinstance(exc, RunpodError) else "Check the saved endpoint ID and API key"
+        return web.json_response({"ok": False, "queueRemaining": 0, "message": message, "jobs": jobs})
+
+
+@server.PromptServer.instance.routes.post("/distributed/serverless/test")
+async def test_serverless_endpoint(request):
+    try:
+        data = await request.json()
+        worker_id = data.get("worker_id")
+    except Exception:
+        return web.json_response({"ok": False, "message": "Invalid request"}, status=400)
+    return await _status(worker_id, fresh=True)
+
+
+@server.PromptServer.instance.routes.get("/distributed/serverless/status/{worker_id}")
+async def serverless_status_endpoint(request):
+    return await _status(request.match_info["worker_id"])
+
+
+@server.PromptServer.instance.routes.post("/distributed/serverless/cancel")
+async def cancel_serverless_endpoint(request):
+    try:
+        data = await request.json()
+        worker_id = data.get("worker_id")
+    except Exception:
+        return web.json_response({"ok": False, "message": "Invalid request"}, status=400)
+    if _pool(worker_id) is None:
+        return web.json_response({"ok": False, "message": "Serverless worker not found"}, status=404)
+    try:
+        manager = get_manager()
+        count = await manager.cancel(worker_id=worker_id)
+        return web.json_response({"ok": True, "cancelled": count, "jobs": manager.public_jobs(worker_id),
+                                  "message": "Cancellation requested for this master's jobs only"})
+    except Exception:
+        return web.json_response({"ok": False, "message": "Cancellation could not be confirmed; check provider jobs"}, status=502)
+
+
+async def _recover(request, *, acknowledge=False):
+    try:
+        data = await request.json()
+        worker_id = data.get("worker_id")
+        if not isinstance(worker_id, str) or not worker_id.strip():
+            raise ValueError()
+        record_id = data.get("record_id")
+        if acknowledge:
+            if not isinstance(record_id, str) or not record_id.strip() or data.get("provider_checked") is not True:
+                raise ValueError()
+    except Exception:
+        return web.json_response({"ok": False, "message": "Provide a worker ID; manual resolution also requires a record ID and provider_checked: true"}, status=400)
+    if _pool(worker_id) is None:
+        return web.json_response({"ok": False, "message": "Serverless worker not found"}, status=404)
+    try:
+        manager = get_manager()
+        if acknowledge:
+            await manager.acknowledge_recovery(worker_id, record_id, provider_checked=True)
+            result = {"message": "Recovery acknowledged by operator; not provider-confirmed cancellation"}
+        else:
+            result = {"resolved": await manager.reconcile(worker_id=worker_id)}
+        return web.json_response(dict(result, ok=True, jobs=manager.public_jobs(worker_id)))
+    except ValueError:
+        return web.json_response({"ok": False, "message": "Unresolved inactive job not found for this worker"}, status=409)
+    except Exception:
+        return web.json_response({"ok": False, "message": "Recovery could not be confirmed; retained jobs remain protected"}, status=502)
+
+
+@server.PromptServer.instance.routes.post("/distributed/serverless/reconcile")
+async def reconcile_serverless_endpoint(request):
+    return await _recover(request)
+
+
+@server.PromptServer.instance.routes.post("/distributed/serverless/resolve")
+async def resolve_serverless_endpoint(request):
+    return await _recover(request, acknowledge=True)

api/worker_routes.py

@@ -61,6 +61,13 @@ async def worker_ws_endpoint(request):
                 })
                 continue
 
+            if not isinstance(data, dict):
+                await ws.send_json({
+                    "type": "dispatch_ack", "request_id": None, "ok": False,
+                    "error": "Expected a websocket message object.",
+                })
+                continue
+
             if data.get("type") != "dispatch_prompt":
                 await ws.send_json({
                     "type": "dispatch_ack",
@@ -81,10 +88,16 @@ async def worker_ws_endpoint(request):
                 continue
 
             try:
+                target_kwargs = {}
+                if "partial_execution_targets" in data:
+                    from ..utils.async_helpers import validate_execution_targets
+                    target_kwargs["partial_execution_targets"] = validate_execution_targets(
+                        prompt, data["partial_execution_targets"])
                 prompt_id = await queue_prompt_payload(
                     prompt,
                     workflow_meta=data.get("workflow"),
                     client_id=data.get("client_id"),
+                    **target_kwargs,
                 )
                 await ws.send_json({
                     "type": "dispatch_ack",
@@ -460,6 +473,9 @@ async def launch_worker_endpoint(request):
         if not worker:
             return await handle_api_error(request, f"Worker {worker_id} not found", 404)
         
+        if worker.get("type") == "serverless":
+            return await handle_api_error(request, "Serverless workers are managed by Runpod, not local processes", 400)
+
         # Ensure consistent string ID
         worker_id_str = worker_id
         
@@ -554,6 +570,9 @@ async def get_local_worker_status_endpoint(request):
         worker_statuses = {}
         
         for worker in config.get("workers", []):
+            # Serverless pools have no host, port, or local process.
+            if worker.get("type") == "serverless":
+                continue
             # Only check local workers
             host = normalize_host(worker.get("host")) or ""
             if not host or host in ["localhost", "127.0.0.1"]:

docs/comfyui-distributed-api.md

@@ -141,13 +141,16 @@ Also, for collector-based flows:
 
 ## Endpoint: `POST /distributed/job_complete`
 
-Submit one completed worker image back to the master collector queue.
+Submit direct-worker image/audio results to the master Collector queue. The route
+accepts canonical JSON completions or multipart image chunks, including opt-in
+exact floating-point IMAGE transport. Serverless typed results use their separate
+bridge/provider contract.
 
 ### URL
 
 - `http://<master-host>:<master-port>/distributed/job_complete`
 
-### Request Body
+### JSON request body
 
 ```json
 {
@@ -164,13 +167,35 @@ Submit one completed worker image back to the master collector queue.
 - `job_id` (string, required)
 - `worker_id` (string, required)
 - `batch_idx` (integer >= 0, required)
-- `image` (string, required)
+- `image` (string, optional if `audio` is present)
   - PNG payload as either:
     - data URL: `data:image/png;base64,...`
     - raw base64 PNG bytes
 - `is_last` (boolean, required)
 
-Legacy multipart/tensor payload formats are no longer accepted on this endpoint.
+At least one of `image` or `audio` must be present. Optional `audio` is an object
+with `sample_rate` (positive integer), `shape` (`[batch, channels, samples]`),
+`dtype: "float32"`, and `data` (base64 waveform bytes). Audio-only completions
+still include `batch_idx`; the Collector uses `0` and `is_last: true`.
+
+### Multipart image chunks
+
+Send `multipart/form-data` with a JSON `envelope` field and repeated binary
+`media` file fields. The envelope contains `job_id`, `worker_id`, `format`,
+`start_index` (nonnegative integer), `count`, `width`, `height`, `is_last`, and
+optional canonical `audio`.
+
+- `png`, `jpeg`, `webp`: one `media` part per frame.
+- `ffv1`, `h264`, `h265` (experimental), `av1`: one video part per chunk.
+- `exact`: one raw tensor part per frame and an envelope `descriptors` list with
+  one `tensor-le-v1` IMAGE descriptor per part. Each descriptor shape must be
+  `[1, height, width, channels]`; decoding preserves supported float dtype and
+  values instead of quantizing to 8-bit.
+
+`legacy_png` uses JSON, not multipart. Exact transport requires support on both
+sides and never retries as an 8-bit format when an older master rejects it.
+See [result transfer](compressed-result-transfer.md) for format controls,
+non-exact compatibility fallbacks and request limits.
 
 ### CORS note
 

docs/compressed-result-transfer.md

@@ -1,38 +1,49 @@
-# Compressed Result Transfer
+# Compressed and Exact Result Transfer
 
-**Distributed Collector** has two optional widgets that control how workers send their images back to the master. This matters most when workers run in the cloud and you pay for data leaving it (egress).
+**Distributed Collector** (`DistributedCollector`) collects images and/or audio. Its `transfer_format` and `transfer_quality` widgets control **remote image returns only**; they do not re-encode master-local images or change audio transport. The saved defaults remain `legacy_png` and quality `90`.
 
-| `transfer_format` | Type | Notes |
+| `transfer_format` | Fidelity | Direct-worker transport |
 |---|---|---|
-| `legacy_png` (default) | lossless | Original behaviour: one JSON request per frame, uncompressed PNG in base64. Works with any master version. |
-| `png` | lossless | Compressed PNG, binary, several frames per request. Same tensors as `legacy_png`. |
-| `ffv1` | lossless | Whole batch as one FFV1 video. Same tensors as `legacy_png`; best lossless choice for frame sequences. |
-| `jpeg`, `webp` | lossy | Per-frame stills. Good for image batches. |
-| `h264`, `h265`, `av1` | lossy | Whole batch as one video. By far the smallest for video frame batches. |
+| `legacy_png` (default) | 8-bit pixels | Original behaviour: one JSON request per frame, uncompressed PNG in base64. Compatible with legacy masters. |
+| `png` | Lossless 8-bit pixels | Compressed PNG, binary, several frames per request. |
+| `ffv1` | Lossless 8-bit pixels | FFV1 video chunks; useful for frame sequences. |
+| `jpeg`, `webp` | Lossy 8-bit images | Per-frame stills. |
+| `h264`, `h265` (**experimental**), `av1` | Lossy 8-bit images | Video chunks; useful for similar consecutive frames. |
+| `exact` (opt-in) | Original floating-point tensors | Binary multipart chunks with one raw tensor payload and `tensor-le-v1` descriptor per frame. No 8-bit quantization. |
 
-`transfer_quality` (1–100, default 90) applies to the lossy formats. JPEG/WebP use it directly; video codecs map it to CRF (90 → CRF 13 for h264/h265, CRF 16 for av1; 100 → CRF 0).
+“Lossless” PNG/FFV1 means lossless **after 8-bit quantization**, not preservation of the source float tensor. Direct non-exact returns decode to RGB float32. `exact` preserves supported floating-point dtype, shape, channels and values: float16, bfloat16, float32 or float64, finite dense BHWC tensors with 1, 3 or 4 channels. It generally costs substantially more bandwidth.
 
-Measured on a 121-frame 1280×720 batch (one worker's result):
+`transfer_quality` (1–100, default 90) applies to lossy formats, not `exact`, PNG or FFV1. JPEG/WebP use it directly; video codecs map it to CRF (90 → CRF 13 for h264/h265, CRF 16 for av1; 100 → CRF 0). A quality of 100 is not a substitute for exact float transport.
+
+## Serverless profiles
+
+- **`comfy-media-v1`:** the same Collector supplies the endpoint's image-return policy. The saved `legacy_png` default is mapped to compressed **`png`** with the same 8-bit pixel precision; normal serverless returns are not exact by default. All other formats retain their selection, including opt-in `exact` and experimental `h265`. Missing codecs or rejected formats fail instead of selecting another codec. Typed image returns preserve channel count and carry alpha separately where needed. See [serverless images and audio](serverless-media.md).
+- **`legacy-image-v1`:** retains its image-file/`SaveImage` contract, not the typed return-format controls. Selecting a legacy pool when any Collector requests `exact` is rejected before dispatch; there is no 8-bit fallback. Audio requires direct workers or `comfy-media-v1`.
+
+## Direct-worker compatibility and limits
+
+- Choose the format on the master; the value is sent to workers with the workflow.
+- **`exact` requires support on both master and worker.** If an older master rejects an exact chunk, the worker raises an error and never resends it as `legacy_png` or another 8-bit format. An exact Collector also rejects direct-worker image returns labelled with a non-exact format. Update both sides or explicitly select a non-exact format.
+- For **non-exact compressed** formats only, the specific first-request older-master rejection (`400` with `Invalid JSON payload`) causes a retry using `legacy_png`. Other rejections, including HTTP 413, remain errors.
+- For direct video returns only, an unavailable encoder falls back to `h264` for lossy video if available, otherwise `png`; unavailable `ffv1` falls back to `png`. The worker logs this. This compatibility fallback does **not** apply to `exact` or typed serverless returns.
+- Multipart requests use a 32 MiB chunk target (`COMFYUI_DISTRIBUTED_TRANSFER_CHUNK_BYTES`) and a 96 MiB request budget (`COMFYUI_DISTRIBUTED_TRANSFER_MAX_REQUEST_BYTES`), reserving space for the envelope and accounting separately for audio. A single large frame may exceed the chunk target but must fit the request budget. Keep the request budget below the master's `--max-upload-size` (default 100 MB).
+- Non-exact compressed chunks are all encoded before sending. Exact transfer instead plans frame sizes from shape/dtype before sending and encodes one chunk at a time. An oversized individual frame is rejected before any chunks are sent; choose another format or raise the worker and master limits together.
+- Audio is attached to the final image chunk when it fits, otherwise sent separately. Audio-only completions use the canonical JSON envelope. Direct audio uses float32 waveform bytes regardless of the image format; `exact` is an IMAGE option, not a new audio codec.
+- Each worker logs the format and bytes sent. These request limits are for direct multipart transfers; serverless request/artifact limits are documented separately.
+
+## Previously documented size example
+
+The following measurements are retained from a 121-frame 1280×720 slow-pan batch (one worker's result); they are not new validation of the current implementation or exact transport.
 
 | Format | Sent | vs legacy | PSNR |
 |---|---|---|---|
-| `legacy_png` | 446 MB | 1× | lossless |
-| `png` | 163 MB | 2.7× | lossless |
-| `ffv1` | 128 MB | 3.5× | lossless |
+| `legacy_png` | 446 MB | 1× | lossless at 8-bit precision |
+| `png` | 163 MB | 2.7× | lossless at 8-bit precision |
+| `ffv1` | 128 MB | 3.5× | lossless at 8-bit precision |
 | `jpeg` q90 | 27 MB | 17× | 42.7 dB |
 | `webp` q90 | 13 MB | 33× | 40.6 dB |
 | `h264` q90 | 2.4 MB | 186× | 40.2 dB |
-| `h265` q90 | 1.0 MB | 434× | 40.1 dB |
+| `h265` q90 (experimental) | 1.0 MB | 434× | 40.1 dB |
 | `av1` q90 | 0.4 MB | 1031× | 40.5 dB |
 
-The test clip is a slow pan, which video codecs compress especially well; expect smaller (still large) gains on content with more motion.
-
-Notes:
-- Choose the format on the master; the value is sent to workers with the workflow.
-- The master must run this version to accept compressed results. A worker sending to an older master gets a 400 on its first request and automatically resends with `legacy_png`.
-- If a worker's PyAV build lacks the requested video encoder, it falls back to `h264` (lossy formats) or `png` (`ffv1`), and logs it.
-- Lossy formats only affect frames produced by workers; the master's own frames are never re-encoded.
-- Request size is bounded on the final encoded bytes. Frames are batched into requests of up to 32 MiB (`COMFYUI_DISTRIBUTED_TRANSFER_CHUNK_BYTES`); a single large frame may go alone in a request of up to 96 MiB (`COMFYUI_DISTRIBUTED_TRANSFER_MAX_REQUEST_BYTES`). Keep the latter below the master's `--max-upload-size` (default 100 MB) if you change either.
-- All chunks are encoded before anything is sent. If a single frame cannot fit within the request limit, the worker fails with an error and sends nothing; use a lossy format or raise both limits.
-- A rejection other than the older-master case above (for example the master's size limit, HTTP 413) is reported as an error rather than retried as `legacy_png`.
-- Each worker logs what it sent, e.g. `Sent 121 image(s) as h264 in 1 request(s): 2.40 MB (raw 334.49 MB, 139.4x smaller)`.
+Video codecs compress this slow pan especially well; gains depend on content and motion.

docs/image-list-queue.md

@@ -0,0 +1,137 @@
+# Dynamic image-list queue
+
+Process a list of images across the master and direct local or remote ComfyUI workers without assigning fixed chunks to each participant.
+
+The master evaluates the input list once. By default, the master and each selected idle worker receive one image job at a time; when a participant finishes, it gets the next unassigned image. Faster participants can therefore process more images. The master restores input order before running downstream nodes.
+
+The existing **Orchestrator Only** setting (`settings.master_delegate_only`) excludes the master from per-image processing. The master still creates the input list and runs downstream nodes. API callers can override the setting for a run with `delegate_master`. There is no second image-list-specific toggle.
+
+## Wiring
+
+```text
+Image list or IMAGE batch
+          |
+Distributed Image List Start -------- context ---------+
+          | image                                     |
+          v                                           |
+    Per-image processing                              |
+          | IMAGE                                     |
+          v                                           v
+                 Distributed Image List End
+                            | image_list
+                            v
+                     Preview / Save / next stage
+```
+
+1. Install the same feature-capable version of Distributed on the master and every selected worker. Workers on an older image-list protocol fail preflight before any work is sent. Restart those instances and refresh the frontend after installation.
+2. Enable Distributed and select the **direct** workers to use. The master also participates unless **Orchestrator Only** is enabled. With the master participating, an empty worker selection runs locally through the same image-list scheduler; orchestrator-only mode requires at least one reachable, idle direct worker. The master must be reachable from remote workers at its configured callback address.
+3. Feed your image list or tensor batch into **Distributed Image List Start → images**.
+4. Connect its `image` output through the per-image processing nodes to **Distributed Image List End → images**.
+5. Connect Start's **context** directly to End's **context**. This is a dedicated `DISTRIBUTED_IMAGE_LIST_CONTEXT` socket, not editable text.
+6. Put preview/save nodes **after End**, then queue normally. API callers should submit through `/distributed/queue`, not directly through `/prompt`.
+
+ComfyUI represents these as ordinary `IMAGE` sockets with list semantics, not a separate `IMAGE_LIST` socket type. Outside the region there is a list; for each assigned item, Start provides one frame and its aligned values. End returns ordered lists, not a concatenated tensor batch, so differently sized images can be preserved.
+
+## Inputs and outputs
+
+Start accepts:
+
+- `images`: one image, an IMAGE batch, or a ComfyUI image list. A batch is split into individual frames; each frame becomes one job.
+- Optional `mask`.
+- Optional `value_1`, `value_2`, … inputs. Only `value_1` is shown at first; connecting it adds `value_2`, and so on, up to `value_10`. Each accepts an IMAGE, MASK, STRING, INT, FLOAT or BOOLEAN.
+
+`mask` and every value follow one rule:
+
+- **1 value** is shared by every image.
+- **One value per image** pairs them in order.
+- **Any other count** is an error. Values are never padded or repeated silently.
+
+Only `images` is split into frames. A batch connected to a value input stays a single value, so a 2-image batch can travel with each item (for example, two reference images per job). To pass a different batch to each image, connect a list with one batch per image.
+
+Start outputs, in order: `image`, `mask`, `item_index` (zero-based), `context`, then one `value_N` output per connected `value_N` input, with the same type. Unused value outputs are hidden. Seeds and strengths are ordinary values: connect a single number to share it, a list for one per image, or derive it from `item_index`.
+
+Values that cannot be transported exactly (models, latents, conditioning and other objects) are rejected. Load models inside the region from a loader connected directly to the processing nodes, as described below.
+
+End accepts `images`, optional `masks` / `texts`, and `context`. Its first three outputs remain `image_list`, `masks`, and `texts` in original item order. Each job must produce exactly one image, at most one mask, and at most one text value. Unconnected optional mask/text outputs contain `None` entries.
+
+### Keep values through End without extra wires
+
+Start's existing **context** connection now also lets End recover each item's original `value_1` … `value_10`. You do not need to wire those values individually into End. Context remains a lightweight execution reference, not a general-purpose data pipe.
+
+End appends a **values** output (`DISTRIBUTED_IMAGE_LIST_VALUES`): one value bundle per image, in the same source order. Connect it to **Distributed Image List Get Value**, then choose `value_name` to extract the field you need. The getter preserves its type and ComfyUI maps it across the bundles, producing a list aligned with End's images. Missing fields fail clearly instead of producing empty placeholders.
+
+For example, to preserve filenames:
+
+```text
+filenames → Start.value_1
+                 Start.context ─────────────→ End.context
+                 Start.image → processing → End.images
+                                               ├─ image_list → Save.images
+                                               └─ values → Get Value [value_1] → Save.filename_prefix
+```
+
+**New or changed per-item values:** connect the calculated result to End's optional `value_N` input. It replaces the original field with that number, or adds it if Start did not supply it. Every connected override must produce exactly one supported value for each item. A tensor batch remains one value, just as it does at Start. Image, mask, string, integer, float and boolean values are supported; arbitrary objects/models/latents and non-finite numbers are not.
+
+End shows override inputs progressively. Use **Add value override** to expose a later slot without connecting the earlier ones. Disconnecting an override never renumbers the other End fields: `value_3` always means `value_3`. Unlike Start's native Autogrow inputs, End overrides must keep these identities stable. Existing image/mask/text/context sockets and output indices are unchanged. If the frontend extension is unavailable, Python still declares all optional override inputs and the values output.
+
+Original values are retained on the master; workers only return explicitly connected overrides, avoiding a redundant return trip for unchanged reference images or other values. Overrides use the existing exact value transport and size limits. Selected workers must advertise `value_results: true` when overrides are requested, checked before execution and again at dispatch. Original-value passthrough alone does not require that additional capability.
+
+Values leaving the region must come from **End**, not directly from Start. To chain regions, extract an End value and connect it to the next Start's corresponding input. Values remain available downstream after End cleans up the execution context; they do not depend on a later lookup of expired job state.
+
+Models are loaded and cached on each processing participant, not serialized between machines. Connect loader outputs directly to the processing nodes. The master needs the region's model files when participating, as does each selected worker. The master still validates the prompt: it needs the relevant node definitions and any model-name/file availability required by their validation (including combo lists). This is not a model-free-master guarantee.
+
+Per-image data must pass through Start; do not bypass it with upstream images, latents or generated batches connected directly into the processing region. Shared dependencies across Start are restricted to registered model outputs or scalar constants, not arbitrary image-producing leaf nodes.
+
+End's optional `texts` is a connection-only socket. For workflows saved before this socket became connection-only, recreate End and reconnect its wires if an old text widget or literal `texts: ""` remains in the compiled prompt. Arbitrary text literals are not accepted at End.
+
+File-based dependencies keep ComfyUI's `[output]` / `[temp]` root annotations and are read on the master. Their worker copies use separate content-addressed paths under `input/distributed-media/`; each worker's rewritten prompt is reused across items without reusing item tokens.
+
+## Native cache reuse
+
+Repeated items can reuse **ComfyUI's native processing cache** on the same participant. Start uses a deterministic identity of the exact image, optional mask, all supplied value slots and source index; tensor dtype, shape and bytes are preserved. Fresh job IDs, assignment tokens and callback addresses live at End, so they no longer invalidate otherwise unchanged upstream processing. End still validates the current assignment and delivers results in source order on every run.
+
+- Processing settings, seeds, links and node fingerprints still participate in ComfyUI's normal signatures. Changed content invalidates reuse; nodes that declare themselves noncacheable still run.
+- The master's private executor retains one native LRU output cache between successful slots, capped at **128 entries when retained**. An active graph may temporarily exceed that target; oversized generations are not retained. Failed or cancelled local slots do not seed reuse, and `--cache-none` disables the local output cache. Direct workers use their own configured native caches.
+- Cache hits are best-effort, not guaranteed: eviction, restarts, changed ordering/index, changed settings and assignment to a different worker can cause misses. There is no worker affinity, disk cache or separate completed-result store.
+- Reuse retains tensor/model references and can increase RAM/VRAM use; the entry cap is **not a byte limit**. Exact identity calculation also adds CPU hashing and, for GPU-resident inputs, a CPU copy.
+- Remote items are still uploaded. Each upload returns a worker-computed identity; if a compatible older worker does not return a valid identity, that assignment uses the legacy graph without stable-input reuse. A worker downgrade between items also falls back safely. Assignment expiry and current-item validation apply even when processing hits the cache.
+- Keep side-effect nodes outside the region when they must run on every submission. A cached processing node follows ordinary ComfyUI semantics and does not repeat its save/preview side effects; End's delivery and downstream saving still run.
+
+## Scope and failure behavior
+
+- This version supports the master and direct workers, not Runpod serverless pools. Unsupported pools are excluded without contacting the provider for this job.
+- Queue submission checks selected direct workers for reachability and feature compatibility, but does not freeze the worker list to that moment's idle snapshot. A selected worker busy with an earlier prompt remains a candidate for the queued run. Immediately before a region starts assigning images, its saved candidates are checked again: workers still busy or transiently unreachable are excluded; healthy idle candidates can continue. Malformed capability/queue responses and incompatible versions fail explicitly. Workers excluded by the initial reachability probe are not retained, and newly selected workers are not added to an already-queued run. Workers becoming free after the region starts are not added mid-region.
+- **Orchestrator Only** requires a selected, reachable, feature-compatible direct worker at submission, but can queue while all selected workers are busy. At region start it requires at least one idle compatible direct worker and fails if none are available; it never falls back to master processing. The queue response's `worker_count` counts saved remote candidates, not guaranteed execution-time participants.
+- Omitted worker IDs select enabled direct workers. Explicit IDs select the named configured direct workers, including disabled ones; an empty list stays empty, unknown IDs fail, and serverless workers are excluded. Local workers use the master's loopback callback; remote workers use its configured public address.
+- Only the enclosed processing is distributed. List creation and downstream saving run on the master in both modes. With participation enabled, the master is a normal scheduler participant and can run alone when no direct workers are available. **Orchestrator-only mode never silently falls back to master processing.** Failed items are not reassigned automatically.
+- Local items execute while the enclosing master prompt waits; they are not queued behind that prompt on the master's `/prompt` queue. The enclosed branch is not also evaluated as a full list by the outer prompt.
+- Start/End regions may be chained sequentially and share model-loader dependencies. Nested/overlapping per-image regions and data escaping a region without End are rejected. Collector and distributed-upscaler nodes cannot be mixed into the same prompt in this first version.
+- Queue preparation works on an execution copy, leaving the saved canvas unchanged. Unused per-item preview/save branches are pruned by dependency reachability, not node names. Required computational utilities marked `OUTPUT_NODE` stay in the graph; explicit `partial_execution_targets` prevent them from also becoming independent outer-master roots. End and legitimate post-End or unrelated outputs are retained. Bypass/rejoin branches still fail boundary validation rather than being silently removed.
+- Every selected worker must advertise `execution_targets: true` as well as the existing image-list protocol. Older workers fail preflight before dispatch. Both HTTP and WebSocket item submissions target End only; the participating master uses its existing End-only local executor. Submit through `/distributed/queue`: an ordinary all-output `/prompt` request does not receive this relaxation.
+- **Target selection is not a sandbox.** A required utility still runs its own save, preview or other side effects when it executes for an item. Keep intentional final saving after End.
+- The entire input list is materialized before scheduling; this is not streaming input. The limit is 4,096 frames. Remote transport preserves floating-point tensor dtype and values rather than encoding to lossy images. Each transported tensor is limited to 50 MiB, and the complete JSON/base64 HTTP request to 101 MiB; image/mask/value combinations can hit the aggregate request limit first. In-process master items use tensors directly and are not subject to those HTTP limits.
+- Results are accepted once per item using assignment tokens. Missing, duplicate, malformed and wrong-index results fail rather than silently returning a partial list. No failed or ambiguous submission is retried automatically.
+- Worker errors fail the region. The worker timeout measures loss of evidence for the owned prompt ID, not maximum render duration. Pending/running evidence refreshes liveness; unrelated busy jobs and failed polls do not. Healthy cold loads and long renders can continue.
+- User interruption is checked during submission and result waiting. Cleanup targets only this region's acknowledged prompt IDs and staged input tokens. Submission may wait briefly for an acknowledgement before cancellation; remote transport and cleanup have finite time limits. Local cleanup stops at node/progress boundaries and waits for its execution thread to finish before the parent prompt can resume. A synchronous node that does not cooperate with cancellation can delay this beyond the configured deadline. Lost acknowledgements are ambiguous, never retried automatically. Unreachable cleanup is logged as unconfirmed and may require checking the worker manually; cancellation does not guarantee remote termination through a broken network.
+- State is in memory. Restarting the master does not resume an interrupted region. Abandoned master contexts and expired worker staging are actively swept on the server loop; active/queued master work is protected from age-only expiry. End also removes its context on success, failure or interruption. This releases feature-owned references, not ComfyUI's separate execution cache.
+
+### Time and size limits
+
+`settings.image_list_job_timeout_seconds` defaults to `0` (no arbitrary total-run cap). A positive finite number opts into an overall deadline with owned-work cleanup; negative, non-finite or malformed settings fail explicitly. There is no fixed one-hour region deadline. The setting is separate from the existing worker liveness timeout.
+
+Uploads, media preparation and submission acknowledgements are independently bounded. WebSocket acknowledgement waiting includes connect/send/receive and is not extended by heartbeats or unrelated messages. A user interrupt allows up to one second for a late acknowledgement; cleanup requests use five-second bounds each.
+
+**4K float32 RGB outputs cannot be returned by direct workers under these transport limits.** A 3840×2160 RGB float32 tensor exceeds 50 MiB before base64 encoding. Mixed master/worker 1080p-to-4K upscaling therefore cannot use this configuration; master-only execution does not use the HTTP transport. Dtypes/values are never silently reduced to fit. Complete encoded input requests are checked before upload, including image/mask/value combinations; output limits are checked when the worker has computed the result. Reverse proxies may impose lower HTTP limits.
+
+Use trusted workers and trusted workflows, as with the existing ComfyUI API. Capability tokens protect assignment/result matching; they are not a replacement for network access control.
+
+## Verification
+
+The generic acceptance helper starts two isolated loopback workers using real ComfyUI HTTP handlers, `PromptQueue` records and CPU `PromptExecutor` instances. Production queue preparation, list nodes, graph extraction, HTTP item/result routes, dynamic scheduling, float transport, ordering and cancellation are exercised. The harness checks orchestrator-only, mixed master/worker and master-only execution, including exact-once required output utilities, discarded previews, unchanged source graphs, End-only worker roots, post-End capture, source-once materialization and cached local model-loader dependencies. WebSocket target propagation is covered by focused route/dispatch tests. No managed ComfyUI service or model is started.
+
+```bash
+COMFYUI_SOURCE_ROOT=/path/to/ComfyUI /path/to/ComfyUI/.venv/bin/python -m pytest tests -q
+bash scripts/test-web.sh
+```
+
+Without `COMFYUI_SOURCE_ROOT`, real-framework tests are explicitly skipped. This coverage is not a substitute for checking your model workflow and connectivity on the actual GPU worker machines.

docs/serverless-media.md

@@ -0,0 +1,187 @@
+# Serverless images and audio
+
+**Distributed Collector** (`DistributedCollector`) collects `IMAGE` and/or `AUDIO`
+from direct workers and serverless pools. There is no separate
+`DistributedMediaCollector`; replace that removed node with `DistributedCollector`
+for image/audio workflows. The Collector has no MASK, STRING, or native VIDEO
+output sockets.
+
+`comfy-media-v1` is an **explicit endpoint profile**, not an automatically detected
+capability. It returns typed image/audio items to the Collector. The standalone
+bridge's generic contract also supports `MASK`, `STRING`, and native ComfyUI
+`VIDEO`, but that does not add those outputs to the Collector. Omitting
+`media_profile` still selects the image-only `legacy-image-v1` contract.
+See [legacy serverless setup](serverless-workers.md) for credentials and recovery.
+
+## Use the Collector with a typed endpoint
+
+1. Prepare an endpoint with the [standalone media bridge](../serverless_bridge/README.md)
+   and the workflow's models/custom nodes. The endpoint needs only the bridge,
+   **not the full ComfyUI-Distributed extension**.
+2. In the serverless worker settings, set **Endpoint media profile** to
+   **Typed media v1 (requires endpoint bridge)** (`media_profile: "comfy-media-v1"`).
+   Start with **Images per run = 1**. This existing setting counts remote jobs,
+   not the number of typed items each job returns.
+3. Use **Distributed Collector** (`DistributedCollector`), connect `images`,
+   `audio`, or both, and connect the corresponding outputs downstream. Image-only
+   and audio-only workflows use the same node.
+4. Queue through Distributed. **Delegate Only** excludes local upstream
+   generation; normal collection includes the master's local values.
+
+**Test Saved Endpoint** checks provider authentication/health, not bridge
+installation, model availability, or typed execution. Selecting a profile does
+not install or upgrade anything on the endpoint.
+
+The Collector accepts optional `images` and `audio` inputs, including ComfyUI list
+inputs, and returns two ordinary outputs—not list outputs:
+
+| Socket / output index | Native type | Aggregation |
+| --- | --- | --- |
+| `images` / 0 | `IMAGE` | Concatenate BHWC tensors along the batch dimension |
+| `audio` / 1 | `AUDIO` | Concatenate waveforms along the sample dimension |
+
+For each connected socket, order is **master first, then prepared participant
+order, then that participant's item order**—never completion order. Delegate-only
+mode omits the master. Image dimensions/channels and audio batch/channel dimensions
+must be compatible for concatenation. Use a common audio sample rate; the Collector
+does not resample or align streams. Audio-only collection returns `None` for images;
+when no audio is supplied it returns a silent placeholder, not an empty list.
+
+The typed wire contract carries 1–64 items per connected socket. The entire remote
+response is validated and decoded before any of its results are delivered;
+missing, duplicate, mismatched, or corrupt typed results fail collection. Validated
+wire lists are then combined into the Collector's ordinary outputs.
+
+Direct workers, `legacy-image-v1` pools and `comfy-media-v1` pools can share an
+image-only Collector run. Audio requires direct workers or `comfy-media-v1` pools.
+If any Collector requests `exact`, a selected `legacy-image-v1` pool is rejected
+before dispatch instead of silently returning 8-bit images.
+
+## Image return format and quality
+
+The Collector's saved `transfer_format` default remains **`legacy_png`**, with
+`transfer_quality` **90**. For `comfy-media-v1`, the compiler maps `legacy_png` to
+**`png`**, so normal serverless returns use compressed 8-bit PNG. `exact` is opt-in.
+These controls affect **remote IMAGE returns only**, not master-local images,
+uploaded inputs or AUDIO. The bridge's generic MASK, STRING and native VIDEO
+codecs are also unaffected.
+
+- `exact`: preserve supported floating-point dtype and values, with no image
+  quantization. Available for direct workers and `comfy-media-v1` pools; it is
+  never downgraded to an 8-bit format. An older master that rejects exact transfer
+  causes an error, not a legacy retry. Update both sides before selecting it.
+- `legacy_png`, `png`, `ffv1`: lossless **8-bit image pixels**, not exact float
+  tensors. For typed serverless returns, image values are clipped to [0,1] and
+  quantized; `legacy_png` is sent as `png`. Direct workers retain the original
+  JSON/base64 framing for `legacy_png`.
+- `jpeg`, `webp`: lossy still-image compression.
+- `h264`, `h265` (**experimental**), `av1`: lossy video encoding of each IMAGE
+  tensor's frame batch. This is an image-return encoding, not a native VIDEO output.
+- Quality 1–100 affects lossy formats only, using the existing Collector's
+  quality-to-CRF mapping for video. Dimensions, frame order and channel count
+  are preserved for typed serverless returns; alpha is carried separately where
+  needed, not silently dropped.
+
+Both master and endpoint must have this bridge revision for compressed image
+returns. Pillow handles still formats and CPU/software PyAV encoders handle video;
+unavailable codecs fail clearly rather than silently changing the selection.
+Compression can reduce result egress and inline response size, but does not relax
+encoded/decoded byte budgets or the input request limit. Compressed IMAGE batches
+are capped at 240 frames and 8192 pixels per dimension, with a conservative
+256 MiB aggregate decode budget (including codec scratch and alpha). Direct
+workers have different limits and non-exact compatibility fallbacks; see
+[result transfer](compressed-result-transfer.md). The bridge's internal
+`MediaExportV1` default is still `exact` for generic callers that omit controls;
+the Collector compiler explicitly supplies `png` for the normal saved default.
+
+## Inputs and fidelity
+
+Automatic local-file transfer supports stock `LoadImage`, plus explicit stock
+`LoadAudio` (`audio` filename) and `LoadVideo` (`file` filename) adapters. Audio and
+video loaders require a literal filename, no extra controls, and output slot 0.
+Files must be under ComfyUI's input/output/temp roots. Custom file loaders are not
+supported; detected unsupported file references fail rather than silently
+becoming endpoint-local paths. Models and custom nodes are never transferred.
+
+- `LoadImage` uses the existing bounded file/base64 image staging path.
+- `LoadAudio` is decoded on the master using native PCM semantics with bounded
+  accumulation, then transferred as a waveform and sample rate. It does not
+  resample; integer PCM conversion follows ComfyUI's `f32_pcm` behavior.
+- `LoadVideo` transfers opaque file bytes as native `VideoFromFile`, without
+  decoding frames or transcoding the container.
+- Typed tensor transport is `tensor-le-v1`: contiguous little-endian float16,
+  bfloat16, float32, or float64 bytes with exact dtype/shape/layout metadata.
+  Only dense, finite tensors with positive dimensions are supported. IMAGE has
+  1, 3, or 4 channels. Bridge export explicitly detaches/stages tensors to CPU;
+  the pure codec itself rejects non-CPU storage. Remote decoding returns
+  independently owned CPU values. STRING transport is exact UTF-8 (`utf8-v1`).
+
+### Bridge-level native VIDEO support
+
+The generic bridge can import/export native VIDEO, and the typed input adapter
+can stage `LoadVideo` for upstream processing. This is not a Collector VIDEO
+output or a generic video converter.
+
+The adapter supports the pinned native classes **exactly**, not arbitrary
+subclasses or objects exposing `save_to()` / `get_components()`:
+
+- `VideoFromComponents`: images, optional matching BHW or BHW1 alpha, optional
+  audio, positive rational frame rate, bounded JSON metadata, positive integer
+  bit depth, and `sRGB`, `HDR`, or `HDR PQ` color space are preserved.
+- `VideoFromFile`: regular local files within allowed media roots or an exact
+  `BytesIO`; opaque bytes, finite signed start/duration, and optional four-integer
+  crop are preserved without normalization.
+- `VideoFromList`: 1–64 flat exact `VideoFromFile` segments, plus optional
+  `complete_audio`, in order.
+
+Unknown classes, incompatible native state, nested/component segment lists,
+unsupported tensors, invalid metadata, and oversized payloads are rejected.
+There is no fallback transcode, frame extraction, color conversion, or container
+repair/validation. A downstream native consumer still needs to read the original
+container. Decoded file-backed VIDEO values own independent `BytesIO` storage;
+component tensors also own their storage, so staging-file cleanup cannot
+invalidate a returned VIDEO.
+
+## Size and operational limits
+
+- The typed contract caps each artifact at **256 MiB**, aggregate artifact bytes
+  at **512 MiB**, and each artifact list at **128** entries. These are validation
+  ceilings, not a promise that inline requests can carry that much data.
+- The built-in master packs inputs inline and enforces **10,000,000 bytes for the
+  complete serialized Runpod `/run` request**, including workflow, policy and
+  base64. Legacy image reads also retain their 10,000,000-byte aggregate guard.
+  There is **no built-in master-side large-input uploader** in PR2.
+- The endpoint defaults to inline results and caps its result JSON at
+  **10,000,000 bytes**. Larger results require an operator-supplied publisher
+  returning expiring public HTTPS artifact URLs. See the bridge guide for
+  bounded download/upload responsibilities and retention.
+- The generic five-kind typed envelope is a serverless bridge contract. Direct
+  workers support Collector images/audio, including opt-in exact IMAGE transport,
+  not generic MASK/STRING/VIDEO Collector outputs. The typed compiler rejects
+  unsupported distributed/batch nodes and collector-to-collector dependencies.
+  Downstream processing stays on the master. [Image-list scheduling](image-list-queue.md)
+  is a separate execution mode.
+
+## Timeout is not proof that Comfy stopped
+
+Provider cancellation, a handler timeout, or cancellation of its Python task
+**does not establish that the local Comfy prompt stopped**. The bridge never
+posts `/interrupt`: the pinned core route checks a queue snapshot and then
+interrupts globally, so even a prompt ID does not make it race-free.
+
+If submission/termination or callback teardown cannot be confirmed, the bridge
+raises `MediaExecutionUncertain` and retains owned input/exporter files—even
+after terminal Comfy history if a downloader/publisher ignores cancellation.
+Callback teardown has a separate 10-second grace; late callback exit does not
+clean retained files. Callbacks must be cooperative async code; synchronous
+event-loop blocking cannot be hard-timed-out.
+
+The endpoint operator must **quarantine the owning worker/environment, reconcile
+the local prompt and readers or terminate that environment, and establish that
+they have stopped before cleanup or worker reuse**. Do not automatically retry
+the request or turn every handler exception into a safely reusable terminal
+provider job. Master-side provider status/reconciliation alone is not proof
+that a surviving local Comfy process or callback has stopped.
+
+These docs describe the implementation, not a test result or live Runpod/provider
+or GPU end-to-end validation.

docs/serverless-workers.md

@@ -0,0 +1,184 @@
+# Runpod serverless workers (preview)
+
+A **Serverless (Runpod)** worker is a queue-based endpoint, not a permanent
+ComfyUI machine. Runpod chooses the actual GPU workers. This page describes the
+default **`legacy-image-v1`** profile: Distributed submits normal ComfyUI API
+workflows and collects returned image files. For image/audio collection and
+opt-in exact IMAGE returns, use [the `comfy-media-v1` profile](serverless-media.md)
+with the same **Distributed Collector** (`DistributedCollector`). The endpoint
+workers do not need the full ComfyUI-Distributed extension installed.
+
+## Setup
+
+1. Add a worker in the Distributed panel and select **Serverless (Runpod)**.
+2. Enter its name, Runpod **endpoint ID** (not a URL; a pasted `https://api.runpod.ai/v2/<id>/…`
+   URL is trimmed to the ID), and API key.
+3. Start with **Images per run = 1** and **Timeout = 1800**.
+4. Save, then choose **Test Saved Endpoint**. This checks authentication and
+   endpoint health only; it does not start GPU inference or prove model/node compatibility.
+5. Enable the worker when you want queued Distributed workflows to use it.
+6. Connect the **Distributed Collector**'s `images` input and queue through
+   Distributed as usual. Leave `audio` unconnected for this legacy profile.
+   Enable **Delegate Only** if the master should collect without generating locally.
+
+Keys are write-only. Leave the key field blank to retain a saved key, or select
+**Clear saved API key** to remove it. The checkbox is unavailable while the
+worker is enabled, so turn the worker off first. A saved key is required before
+enabling a worker. Use HTTPS or a trusted local connection when
+entering keys. Like ComfyUI itself, the panel must not be exposed to untrusted
+users: anyone able to queue arbitrary workflows can run code and spend endpoint credits.
+
+## Endpoint contract
+
+The endpoint must accept this generic input shape; named application modes are
+not used:
+
+```json
+{
+  "input": {
+    "workflow": {
+      "1": {"class_type": "EmptyImage", "inputs": {"width": 32, "height": 32, "batch_size": 1, "color": 0}},
+      "2": {"class_type": "SaveImage", "inputs": {"images": ["1", 0], "filename_prefix": "example"}}
+    },
+    "images": []
+  }
+}
+```
+
+For image inputs, `images` contains `{ "name": "input.png", "image": "<base64>" }`
+entries and the handler must rewrite `LoadImage` references to the staged files.
+The adapter transfers local input images; it does not copy models or install nodes.
+All referenced models and custom nodes must already exist in the endpoint image.
+
+On completion, Runpod's `output.images` must contain artifact records such as:
+
+```json
+{
+  "filename": "example_00001_.png",
+  "subfolder": "",
+  "artifact_type": "output",
+  "artifact_path": "example_00001_.png",
+  "type": "base64",
+  "data": "<base64>"
+}
+```
+
+Image data can alternatively use `type: "s3_url"`, `"gcs_url"`, or `"url"`
+with an HTTPS URL in `data`/`url`. Signed artifact URLs are downloaded without
+Runpod credentials; private-network destinations and unsafe redirects are refused.
+The handler must preserve injected `SaveImage.filename_prefix` values so results
+can be matched to the right collector and slot. Duplicate basenames in different
+subfolders remain different artifact identities.
+
+A Runpod job marked `COMPLETED` is still rejected if `output.error` or
+`output.errors` is nonempty, required collector outputs are missing, or image
+artifacts are invalid. Failures are surfaced rather than silently returning an
+incomplete serverless batch.
+
+## Scheduling and cancellation
+
+- **Images per run** (1–16) is how many separate Runpod jobs this worker submits
+  for each queued run. Each returns one image and receives its own Distributed
+  Seed/Value overrides. All of them are submitted at once.
+- How many of those jobs actually run in parallel is decided by the Runpod
+  endpoint's **max workers**, not by Distributed. Jobs beyond that wait in
+  Runpod's queue. To run every image at the same time, keep Images per run at or
+  below the endpoint's max workers.
+- Distributed also keeps a fixed limit of 16 outstanding jobs per serverless
+  worker (and 64 across all of them). Retained jobs after a restart, ambiguous
+  submissions, and unconfirmed cancellations keep counting against it until
+  reconciled. Other applications using the same endpoint are unaffected.
+- Older configs may contain a `concurrency` field from the previous separate
+  setting. It is still accepted when loading or saving, but ignored and dropped.
+- **Load balance** selects one participant, so it sends one job instead of
+  Images per run.
+- **Timeout** includes Runpod queue/cold-start time, execution, and output
+  retrieval. Provider policy limits are submitted as well.
+- A zero-worker endpoint can be healthy: Runpod scales from zero on submission.
+- Interrupting a collector or using the panel cancellation action requests
+  cancellation only for jobs submitted by this master. It never purges the endpoint.
+- A network failure during submission is not retried automatically: the job
+  may already be billable. Check the Runpod job list before queuing again.
+
+## Persistence and recovery
+
+Credentials live outside `gpu_config.json` and Git in the master ComfyUI user
+folder under `distributed/serverless/credentials.json`, with restrictive POSIX
+permissions. They are omitted from public configuration, workflow exports, and
+browser configuration state. The first version requires a POSIX (Linux/macOS)
+master for serverless secret storage; existing non-serverless workers remain
+available on Windows.
+
+`distributed/serverless/jobs.json` records provider job IDs and sanitized states,
+not image data, workflow payloads, or API keys. A restarted master cannot resume
+an interrupted collector. It does not automatically resubmit jobs. Retained
+IDs are available through `/distributed/serverless/status/<worker_id>`; the worker's
+cancel action can request cancellation of known pre-restart jobs. If no job ID
+was received, check Runpod manually. Cancellation is best-effort and may race
+with completion; `CANCEL_REQUESTED` is not a provider-confirmed cancellation.
+
+Unresolved records protect the worker's endpoint, API key, worker type,
+and deletion even when no local task remains. Disabling the worker is still allowed;
+it does not cancel existing jobs. On restart, never-submitted `WAITING` slots are
+closed locally, while interrupted `SUBMITTING` records without IDs become
+`UNKNOWN` and remain protected.
+
+### Reconcile or acknowledge recovery
+
+Recovery uses the same trusted-master API as the panel; it never resubmits a job.
+
+1. Read `GET /distributed/serverless/status/<worker_id>`. The `jobs` list includes
+   every unresolved record with its local `record_id`, provider `job_id` (if
+   received), and state. These records remain visible if endpoint health fails.
+2. Request cancellation from the panel, or send
+   `POST /distributed/serverless/reconcile` with the body below to read the
+   retained provider IDs. Only a confirmed terminal state releases protection.
+   Network errors, expired IDs, or ambiguous responses do not.
+
+   ```json
+   {"worker_id": "YOUR_WORKER_ID"}
+   ```
+
+3. If no ID was returned, or Runpod no longer retains its status, inspect Runpod
+   manually first. **Only after verifying that the job is no longer running or
+   queued**, send `POST /distributed/serverless/resolve` for that one record:
+
+   ```json
+   {"worker_id": "YOUR_WORKER_ID", "record_id": "LOCAL_RECORD_ID", "provider_checked": true}
+   ```
+
+   This records `RESOLVED`: an explicit operator acknowledgment, not a
+   provider-confirmed cancellation. It cannot resolve a live local task and does
+   not stop remote compute. It releases the record's reservation and may let
+   already-waiting jobs proceed. Never automate this acknowledgment as a retry
+   mechanism; an unverified job can still be billable.
+
+Worker setting changes become available once all its local tasks and unresolved records
+are finished. No credentials or ledger files need to be edited by hand.
+
+## Legacy profile limits
+
+- Only the Collector's image branch is supported by `legacy-image-v1`; connected
+  audio is rejected. `comfy-media-v1` adds audio through the same Collector and
+  supports the configured image-return formats.
+- `exact` is opt-in and requires direct workers or `comfy-media-v1` pools. If any
+  Collector requests it, a selected legacy pool is rejected before dispatch,
+  not used as an 8-bit fallback. Other Collector transfer-format controls do not
+  change this legacy profile's `SaveImage`/image-file contract. With
+  `comfy-media-v1`, the saved `legacy_png` default is mapped to compressed `png`.
+- Coordinated `UltimateSDUpscaleDistributed` workloads and unsupported Distributed
+  node types are rejected before submission.
+- Downstream post-collector processing remains local; the remote graph contains
+  only the upstream compute required for image saves.
+- Inputs must be supported images under ComfyUI's input/output/temp roots.
+- Raw image input reads are guarded at 10,000,000 bytes total, and the complete
+  serialized `/run` JSON request body (workflow, base64 images, and policy) must
+  also fit within 10,000,000 bytes. Base64 adds about 33%, so the usable raw-image
+  budget is lower and depends on workflow size. Large-media/object-storage upload
+  support is not included.
+- No remote GPU discovery, endpoint provisioning, model synchronization, automatic
+  scaling changes, or inbound master callback/tunnel is needed.
+
+These docs describe the implementation, not a test result. A successful health
+check is not a paid end-to-end generation test; validate your deployed handler
+and model inventory with one image per run before increasing it.

image_list/__init__.py

@@ -0,0 +1 @@
+"""Dynamic per-image image-list execution internals."""

image_list/execution_plan.py

@@ -0,0 +1,71 @@
+"""Prepare image-list execution without changing the saved/API prompt or registry."""
+from __future__ import annotations
+
+import copy
+import sys
+
+from .runtime import (
+    END_CLASS,
+    END_RESULT_INPUTS,
+    _ancestors_for_inputs,
+    _downstream_until,
+    _image_list_pairs,
+    _is_output_node,
+    _prompt_links,
+    validate_image_list_regions,
+)
+
+
+def _reachable(roots: set[str], edges: dict[str, set[str]]) -> set[str]:
+    visited = set()
+    pending = list(roots)
+    while pending:
+        node_id = pending.pop()
+        if node_id in visited:
+            continue
+        visited.add(node_id)
+        pending.extend(edges[node_id] - visited)
+    return visited
+
+
+def prepare_image_list_execution(prompt: dict) -> tuple[dict, list[str]]:
+    """Return an execution copy and its outer-master output targets.
+
+    Disposable per-item side branches are removed before boundary validation.
+    Required output-node utilities remain normal dependencies, but must not be
+    independent outer roots. Callers must use these targets (and End-only targets
+    for item execution); this does not suppress side effects inside required nodes.
+    """
+    pairs = _image_list_pairs(prompt)
+    prepared = copy.deepcopy(prompt)
+    upstream = {node_id: set() for node_id in prepared}
+    downstream = {node_id: set() for node_id in prepared}
+    for node_id, node in prepared.items():
+        for _, source, _ in _prompt_links(node, prepared):
+            upstream[node_id].add(source)
+            downstream[source].add(node_id)
+
+    # Never erase another region or its dependencies to make a bad graph pass.
+    boundaries = {node_id for pair in pairs for node_id in pair}
+    required = _reachable(boundaries, upstream)
+    discarded = set()
+    for start_id, end_id in pairs:
+        region_nodes = _ancestors_for_inputs(prepared, end_id, start_id, END_RESULT_INPUTS)
+        outside = _downstream_until(prepared, start_id, end_id) - region_nodes
+        # Preserve bypass joins and their whole chain into post-End computation.
+        # They must fail boundary validation, not disappear with debug previews.
+        post_end = _reachable({end_id}, downstream)
+        protected = required | _reachable(post_end, upstream)
+        discarded.update(outside - protected)
+    for node_id in discarded:
+        del prepared[node_id]
+
+    regions = validate_image_list_regions(prepared, explicit_targets=True)
+    per_item = set().union(*(region["branch_ids"] for region in regions))
+    registry = getattr(sys.modules.get("nodes"), "NODE_CLASS_MAPPINGS", {})
+    targets = [
+        node_id for node_id, node in prepared.items()
+        if node_id not in per_item and isinstance(node, dict)
+        and (node.get("class_type") == END_CLASS or _is_output_node(node, registry))
+    ]
+    return prepared, targets

image_list/lifecycle.py

@@ -0,0 +1,174 @@
+"""Server-owned background cleanup for image-list runtime state."""
+from __future__ import annotations
+
+import asyncio
+import logging
+import math
+import time
+from collections.abc import Iterable
+from typing import Any, cast
+
+LOGGER = logging.getLogger(__name__)
+DEFAULT_SWEEP_INTERVAL_SECONDS = 30.0
+DEFAULT_ORPHAN_GRACE_SECONDS = 300.0
+_LIFECYCLE_MARKER = object()
+
+
+def has_safe_execution_targets(prompt_server: Any, prompt_id: str | None, prompt: dict) -> bool:
+    """Prove outer roots from the running queue, not a forgeable prompt marker.
+
+    Plain /prompt remains strict when ComfyUI selected all outputs, including a
+    flagged per-item dependency. No saved workflow field grants this exemption.
+    """
+    if not prompt_id:
+        return False
+    from .execution_plan import prepare_image_list_execution
+    prepared, targets = prepare_image_list_execution(prompt)
+    if prepared != prompt or not targets:
+        return False
+    queue = getattr(prompt_server, "prompt_queue", None)
+    for name in ("get_current_queue_volatile", "get_current_queue"):
+        reader = getattr(queue, name, None)
+        if not callable(reader):
+            continue
+        try:
+            running, _pending = reader()
+            for entry in running:
+                if (isinstance(entry, (tuple, list)) and len(entry) > 4
+                        and entry[1] == prompt_id and entry[2] == prompt):
+                    roots = entry[4]
+                    return (isinstance(roots, (list, tuple))
+                            and all(isinstance(root, str) for root in roots)
+                            and len(roots) == len(targets) and set(roots) == set(targets))
+            return False
+        except Exception:
+            continue
+    return False
+
+
+def _queue_prompt_ids(prompt_server: Any) -> set[str]:
+    """Read running/pending prompt IDs, preferring ComfyUI's volatile view."""
+    queue = getattr(prompt_server, "prompt_queue", None)
+    readers = (
+        getattr(queue, "get_current_queue_volatile", None),
+        getattr(queue, "get_current_queue", None),
+    )
+    errors: list[Exception] = []
+    for reader in readers:
+        if not callable(reader):
+            continue
+        try:
+            snapshot = reader()
+            if not isinstance(snapshot, (tuple, list)) or len(snapshot) < 2:
+                raise TypeError("prompt queue snapshot must contain running and pending entries")
+            running, pending = snapshot[0], snapshot[1]
+            prompt_ids: set[str] = set()
+            for entries in (running, pending):
+                if not isinstance(entries, Iterable):
+                    raise TypeError("prompt queue entries must be iterable")
+                for entry in entries:
+                    if isinstance(entry, (tuple, list)) and len(entry) > 1:
+                        prompt_id = entry[1]
+                        if isinstance(prompt_id, str):
+                            prompt_ids.add(prompt_id)
+            return prompt_ids
+        except Exception as exc:
+            errors.append(exc)
+
+    if errors:
+        raise RuntimeError("unable to read the prompt queue") from errors[-1]
+    raise RuntimeError("prompt queue has no supported queue reader")
+
+
+def _sweep_orphan_contexts(prompt_server: Any, runtime: Any, orphan_grace_seconds: float) -> None:
+    queue_read_ok = True
+    try:
+        queued_prompt_ids = _queue_prompt_ids(prompt_server)
+    except Exception:
+        # A failed queue read must not make a submitted context look orphaned.
+        LOGGER.warning("Could not read prompt queue during image-list lifecycle sweep", exc_info=True)
+        queued_prompt_ids = set()
+        queue_read_ok = False
+
+    now = time.monotonic()
+    with runtime._LOCK:
+        for context_id, state in list(runtime._MASTER_JOBS.items()):
+            if getattr(state, "active", False):
+                continue
+            prompt_id = getattr(state, "prompt_id", None)
+            if prompt_id is not None:
+                if not queue_read_ok or prompt_id in queued_prompt_ids:
+                    continue
+            created_at = getattr(state, "created_at", None)
+            expired = (
+                isinstance(created_at, (int, float))
+                and not isinstance(created_at, bool)
+                and math.isfinite(created_at)
+                and now - created_at >= orphan_grace_seconds
+            )
+            if expired and runtime._MASTER_JOBS.get(context_id) is state:
+                runtime.remove_master_job(context_id)
+
+
+def _sweep_once(prompt_server: Any, runtime: Any, orphan_grace_seconds: float) -> None:
+    try:
+        runtime._sweep_worker_items()
+    except Exception:
+        LOGGER.exception("Could not sweep expired image-list worker staging")
+    _sweep_orphan_contexts(prompt_server, runtime, orphan_grace_seconds)
+
+
+async def _sweep_loop(
+    prompt_server: Any,
+    runtime: Any,
+    interval_seconds: float,
+    orphan_grace_seconds: float,
+) -> None:
+    while True:
+        try:
+            _sweep_once(prompt_server, runtime, orphan_grace_seconds)
+        except Exception:
+            # Keep the server-owned task alive if one pass encounters unexpected state.
+            LOGGER.exception("Image-list lifecycle sweep failed")
+        await asyncio.sleep(interval_seconds)
+
+
+def install_lifecycle(
+    prompt_server: Any,
+    runtime: Any,
+    *,
+    interval_seconds: float = DEFAULT_SWEEP_INTERVAL_SECONDS,
+    orphan_grace_seconds: float = DEFAULT_ORPHAN_GRACE_SECONDS,
+) -> None:
+    """Register one cancellable periodic sweep in the prompt server's app context."""
+    interval_seconds = float(interval_seconds)
+    orphan_grace_seconds = float(orphan_grace_seconds)
+    if not math.isfinite(interval_seconds) or interval_seconds <= 0:
+        raise ValueError("interval_seconds must be a positive finite value")
+    if not math.isfinite(orphan_grace_seconds) or orphan_grace_seconds < 0:
+        raise ValueError("orphan_grace_seconds must be a non-negative finite value")
+
+    app = getattr(prompt_server, "app", None)
+    cleanup_ctx = getattr(app, "cleanup_ctx", None)
+    if cleanup_ctx is None:
+        raise TypeError("prompt_server must expose an app with cleanup_ctx")
+    if any(getattr(factory, "_image_list_lifecycle_marker", None) is _LIFECYCLE_MARKER
+           for factory in cleanup_ctx):
+        return
+
+    async def lifecycle_context(_app):
+        task = asyncio.create_task(
+            _sweep_loop(prompt_server, runtime, interval_seconds, orphan_grace_seconds),
+            name="image-list-lifecycle-sweeper",
+        )
+        try:
+            yield
+        finally:
+            task.cancel()
+            try:
+                await task
+            except asyncio.CancelledError:
+                pass
+
+    cast(Any, lifecycle_context)._image_list_lifecycle_marker = _LIFECYCLE_MARKER
+    cleanup_ctx.append(lifecycle_context)

image_list/local_executor.py

@@ -0,0 +1,223 @@
+"""A queue-free Comfy graph slot, only while the synchronous outer End is blocked.
+
+Do NOT use PromptExecutor.execute[_async] here. Its prompt lifecycle clears the
+process interrupt, resets progress handlers, changes RAM-cache callbacks and
+starts/ends a model tracker. Those belong to the suspended parent executor.
+Instead use core's ExecutionList/execute (including lazy/async/expanded nodes)
+with private provider-free caches and a private server facade. Model outputs are
+registered with the existing parent tracker, whose lifetime includes this work.
+
+Core progress and GraphBuilder defaults are process-global. Borrow them only
+inside the owned execution thread, restore them in finally, and join that thread
+before End returns. Cancellation never clears or raises the global interrupt:
+it stops locally at node/progress boundaries. A non-cooperative synchronous node
+must finish before cleanup can return; abandoning its thread is not safe.
+"""
+from __future__ import annotations
+
+import asyncio
+from concurrent.futures import ThreadPoolExecutor
+import threading
+from types import SimpleNamespace
+
+from .runtime import ImageListCancelledError
+
+
+_EXECUTION_LOCK = threading.Lock()
+_CACHE_LOCK = threading.Lock()
+# One process-local native output cache, not a second completed-result cache.
+# Core's LRU can temporarily exceed its target for the active graph; oversized
+# generations are discarded rather than retained after an item/run finishes.
+NATIVE_CACHE_MAX_ENTRIES = 128
+_RETAINED_OUTPUTS = None
+
+
+def _take_outputs(execution):
+    global _RETAINED_OUTPUTS
+    with _CACHE_LOCK:
+        outputs, _RETAINED_OUTPUTS = _RETAINED_OUTPUTS, None
+    if getattr(execution.args, "cache_none", False):
+        return execution.NullCache()
+    return outputs if outputs is not None else execution.LRUCache(
+        execution.CacheKeySetInputSignature, max_size=NATIVE_CACHE_MAX_ENTRIES)
+
+
+def _retain_outputs(outputs):
+    global _RETAINED_OUTPUTS
+    # Only the bounded native LRU is eligible. NullCache and test doubles are not.
+    if getattr(outputs, "max_size", None) != NATIVE_CACHE_MAX_ENTRIES:
+        return
+    outputs.clean_unused()
+    if len(outputs.cache) > NATIVE_CACHE_MAX_ENTRIES:
+        return
+    with _CACHE_LOCK:
+        _RETAINED_OUTPUTS = outputs
+
+
+class LocalExecutionSlot:
+    def __init__(self, *, model_callback=None):
+        self._pool = ThreadPoolExecutor(max_workers=1, thread_name_prefix="image-list-master")
+        self._stop = threading.Event()
+        self._future = None
+        self._caches = None
+        self._failed = False
+        self._model_callback = model_callback
+        self._closed = False
+        self._loop = None
+        self._task = None
+        self._state_lock = threading.Lock()
+
+    def submit(self, prompt, prompt_id, end_id, workflow_meta=None):
+        if self._closed or self._stop.is_set():
+            raise ImageListCancelledError("Local image-list slot has stopped")
+        if self._future is not None and not self._future.done():
+            raise RuntimeError("Local image-list slot already has an in-flight item")
+        self._future = self._pool.submit(self._run, prompt, prompt_id, end_id, workflow_meta)
+        return self._future
+
+    def request_stop(self):
+        self._stop.set()
+        # Wakes a graph waiting on an async node as well as stopping synchronous
+        # nodes at their next progress/between-node checkpoint.
+        with self._state_lock:
+            if self._loop is not None and self._task is not None:
+                self._loop.call_soon_threadsafe(self._task.cancel)
+
+    async def close(self):
+        if self._closed:
+            return
+        # A stop requested before close indicates cancellation, not normal drain.
+        cancelled = self._stop.is_set()
+        self.request_stop()
+        try:
+            if self._future is not None:
+                # Do not cancel this Future: cancelling an asyncio wrapper cannot
+                # kill its running thread, and falsely makes cleanup look done.
+                waiter = asyncio.wrap_future(self._future)
+                while not waiter.done():
+                    try:
+                        await asyncio.shield(waiter)
+                    except asyncio.CancelledError:
+                        continue
+                    except Exception:
+                        break  # The scheduler propagates execution errors.
+                if waiter.done() and not waiter.cancelled():
+                    waiter.exception()
+        finally:
+            self._pool.shutdown(wait=True, cancel_futures=True)
+            if not cancelled and not self._failed and self._caches is not None:
+                _retain_outputs(self._caches.outputs)
+            self._caches = None
+            self._closed = True
+
+    def _run(self, prompt, prompt_id, end_id, workflow_meta):
+        # The normal Comfy prompt worker is single-threaded and blocked in End.
+        # Fail closed if another local slot is nevertheless executing, rather
+        # than interleave writes to core's process-global execution state.
+        if not _EXECUTION_LOCK.acquire(blocking=False):
+            raise RuntimeError("Another local image-list executor is already active")
+        try:
+            asyncio.run(self._execute(prompt, prompt_id, end_id, workflow_meta))
+        except asyncio.CancelledError as exc:
+            self._failed = True
+            raise ImageListCancelledError("Local image-list execution cancelled") from exc
+        except BaseException:
+            self._failed = True
+            raise
+        finally:
+            with self._state_lock:
+                self._loop = self._task = None
+            _EXECUTION_LOCK.release()
+
+    def _check_stopped(self):
+        import comfy.model_management
+        if self._stop.is_set() or comfy.model_management.processing_interrupted():
+            raise ImageListCancelledError("Local image-list execution interrupted")
+
+    async def _execute(self, prompt, prompt_id, end_id, workflow_meta):
+        import execution
+        import torch
+        import comfy.model_management
+        import comfy.utils
+        from comfy_execution import progress
+        from comfy_execution.graph_utils import GraphBuilder
+
+        self._check_stopped()
+        with self._state_lock:
+            self._loop, self._task = asyncio.get_running_loop(), asyncio.current_task()
+        dynamic_prompt = execution.DynamicPrompt(prompt)
+        private_progress = progress.ProgressRegistry(prompt_id, dynamic_prompt)
+        parent_progress = progress.global_progress_registry
+        parent_hook = comfy.utils.PROGRESS_BAR_HOOK
+        prefix_keys = ("_default_prefix_root", "_default_prefix_call_index", "_default_prefix_graph_index")
+        parent_prefix = {key: getattr(GraphBuilder, key) for key in prefix_keys}
+        progress.global_progress_registry = private_progress
+
+        def progress_hook(value, total, preview=None, prompt_id=None, node_id=None):
+            self._check_stopped()
+            # Never route a child progress/preview event through the parent's
+            # client or reset its handlers. Keep the private progress accurate.
+            if node_id is not None:
+                private_progress.update_progress(node_id, value, total, preview)
+
+        comfy.utils.PROGRESS_BAR_HOOK = progress_hook
+        try:
+            with torch.inference_mode():
+                if self._caches is None:
+                    outputs = _take_outputs(execution)
+                    objects = execution.HierarchicalCache(execution.CacheKeySetID)
+                    self._caches = SimpleNamespace(outputs=outputs, objects=objects, all=[outputs, objects])
+                caches = self._caches
+                changed = execution.IsChangedCache(prompt_id, dynamic_prompt, caches.outputs)
+                for cache in caches.all:
+                    await cache.set_prompt(dynamic_prompt, prompt.keys(), changed)
+                    cache.clean_unused()
+                execution_list = execution.ExecutionList(dynamic_prompt, caches.outputs, self._model_callback)
+                execution_list.add_node(end_id)
+                server = SimpleNamespace(client_id=None, last_node_id=None, send_sync=lambda *a, **kw: None)
+                assets = execution.NoAssets(execution.args)
+                extra = {"extra_pnginfo": {"workflow": workflow_meta}} if workflow_meta is not None else {}
+                executed, pending_subgraphs, pending_async, ui = set(), {}, {}, {}
+                while not execution_list.is_empty():
+                    self._check_stopped()
+                    node_id, error, exc = await execution_list.stage_node_execution()
+                    if error is not None:
+                        raise RuntimeError(f"Local image-list graph failed: {error}") from exc
+                    self._check_stopped()
+                    result, error, exc = await execution.execute(
+                        server, dynamic_prompt, caches, node_id, extra, executed, prompt_id,
+                        execution_list, pending_subgraphs, pending_async, ui, assets)
+                    if result == execution.ExecutionResult.FAILURE:
+                        if isinstance(exc, (comfy.model_management.InterruptProcessingException,
+                                            ImageListCancelledError)):
+                            raise ImageListCancelledError("Local image-list execution interrupted") from exc
+                        raise RuntimeError(f"Local image-list node {node_id} failed: {error}") from exc
+                    if result == execution.ExecutionResult.PENDING:
+                        execution_list.unstage_node_execution()
+                    else:
+                        execution_list.complete_node_execution()
+                self._check_stopped()
+        finally:
+            # Core execute creates async-node and unblock tasks. This is a private
+            # event loop, so every other task is owned here. Drain/cancel them
+            # before restoring globals, including when the graph failed early.
+            with self._state_lock:
+                self._task = None  # Stop requests must not interrupt this drain.
+            try:
+                while True:
+                    pending = [task for task in asyncio.all_tasks() if task is not asyncio.current_task()]
+                    if not pending:
+                        break
+                    for task in pending:
+                        task.cancel()
+                    drain = asyncio.gather(*pending, return_exceptions=True)
+                    while not drain.done():
+                        try:
+                            await asyncio.shield(drain)
+                        except asyncio.CancelledError:
+                            continue
+            finally:
+                progress.global_progress_registry = parent_progress
+                comfy.utils.PROGRESS_BAR_HOOK = parent_hook
+                for key, value in parent_prefix.items():
+                    setattr(GraphBuilder, key, value)

nodes/collector.py

@@ -27,6 +27,39 @@ def _encode_all_chunks(frames, fmt, quality, max_chunk_bytes, max_frame_bytes):
     return list(transfer_codec.encode_chunks(frames, fmt, quality, max_chunk_bytes, max_frame_bytes))
 
 
+def _exact_frame_bytes(image_batch):
+    """Wire size of one exact frame, computed from shape/dtype before encoding."""
+    from ..serverless_bridge import codec as bridge_codec
+    if not isinstance(image_batch, torch.Tensor) or image_batch.layout != torch.strided or image_batch.is_nested:
+        raise transfer_codec.TransferCodecError("exact transfer requires a dense IMAGE tensor")
+    if image_batch.dtype not in bridge_codec._DTYPE_NAMES:
+        raise transfer_codec.TransferCodecError(f"exact transfer does not support dtype {image_batch.dtype}")
+    if image_batch.ndim != 4:
+        raise transfer_codec.TransferCodecError(
+            f"Expected IMAGE tensor [B,H,W,C], got shape {tuple(image_batch.shape)}")
+    batch, height, width, channels = (int(value) for value in image_batch.shape)
+    if min(batch, height, width) <= 0 or channels not in (1, 3, 4):
+        raise transfer_codec.TransferCodecError("exact IMAGE dimensions must be positive with 1, 3 or 4 channels")
+    size = height * width * channels * image_batch.element_size()
+    if size > bridge_codec.MAX_PAYLOAD_BYTES:
+        raise transfer_codec.FrameTooLargeError("exact frame exceeds the tensor codec byte limit")
+    return size
+
+
+def _encode_exact_frames(frames):
+    """Encode [1,H,W,C] CPU slices losslessly (runs in an executor)."""
+    from ..serverless_bridge import codec as bridge_codec
+    descriptors, parts = [], []
+    for frame in frames:
+        try:
+            descriptor, payload = bridge_codec.encode_payload("IMAGE", frame)
+        except ValueError as exc:
+            raise transfer_codec.TransferCodecError(f"exact transfer cannot encode this IMAGE: {exc}") from None
+        descriptors.append(descriptor)
+        parts.append(payload)
+    return descriptors, parts
+
+
 class DistributedCollectorNode:
     INPUT_IS_LIST = True
     EMPTY_AUDIO = {"waveform": torch.zeros(1, 2, 1), "sample_rate": 44100}
@@ -178,6 +211,15 @@ class DistributedCollectorNode:
         )
         return result
 
+    @staticmethod
+    def _get_serverless_manager():
+        """Return the serverless manager only for collector jobs with serverless slots."""
+        manager = getattr(prompt_server, "distributed_serverless", None)
+        if manager is not None:
+            return manager
+        from ..api.orchestration.serverless_runtime import get_manager
+        return get_manager()
+
     async def send_batch_to_master(
         self,
         image_batch,
@@ -191,6 +233,10 @@ class DistributedCollectorNode:
         """Send an image batch, optionally with audio, or an audio-only completion."""
         batch_size = 0 if image_batch is None else image_batch.shape[0]
         fmt = self._resolve_transfer_format(transfer_format) if batch_size else transfer_codec.LEGACY_FORMAT
+        if fmt == transfer_codec.EXACT_FORMAT:
+            # Never downgrade exact to 8-bit: an older master fails the job clearly.
+            await self._send_exact_batch(image_batch, audio, multi_job_id, master_url, worker_id)
+            return
         if fmt != transfer_codec.LEGACY_FORMAT:
             sent = await self._send_compressed_batch(
                 image_batch, audio, multi_job_id, master_url, worker_id, fmt, transfer_quality
@@ -320,6 +366,100 @@ class DistributedCollectorNode:
         )
         return True
 
+    async def _send_exact_batch(self, image_batch, audio, multi_job_id, master_url, worker_id):
+        """Send original floating-point frames as binary multipart chunks.
+
+        Frame sizes are known from shape and dtype, so every request is planned
+        (and an oversized frame rejected) before anything is sent. Frames are
+        encoded one chunk at a time to avoid holding a second copy of the batch.
+        Any master rejection raises: there is no lossy or 8-bit fallback.
+        """
+        encoded_audio = encode_audio_payload(audio)
+        session = await get_client_session()
+        url = f"{master_url}/distributed/job_complete"
+        loop = asyncio.get_running_loop()
+        media_limit = max(1, TRANSFER_MAX_REQUEST_BYTES - self.TRANSFER_ENVELOPE_RESERVE_BYTES)
+        chunk_limit = min(TRANSFER_MAX_CHUNK_BYTES, media_limit)
+
+        frame_bytes = _exact_frame_bytes(image_batch)
+        total = int(image_batch.shape[0])
+        if frame_bytes > media_limit:
+            raise transfer_codec.FrameTooLargeError(
+                f"Frame 0 is {frame_bytes} bytes as exact, above the {media_limit}-byte request limit. "
+                "Use png or a lossy format, or raise the master's --max-upload-size together with "
+                "COMFYUI_DISTRIBUTED_TRANSFER_MAX_REQUEST_BYTES."
+            )
+        # Bound descriptor/part-header overhead even for tiny frames.
+        per_chunk = min(128, max(1, chunk_limit // frame_bytes))
+        ranges = [(start, min(per_chunk, total - start)) for start in range(0, total, per_chunk)]
+        audio_bytes = len(json.dumps(encoded_audio)) if encoded_audio is not None else 0
+        last_media = ranges[-1][1] * frame_bytes
+        audio_inline = encoded_audio is not None and last_media + audio_bytes <= media_limit
+        audio_separate = encoded_audio is not None and not audio_inline
+        height, width = int(image_batch.shape[1]), int(image_batch.shape[2])
+        if audio_separate and audio_bytes > media_limit:
+            raise transfer_codec.FrameTooLargeError("exact completion audio exceeds the request limit")
+        cpu_batch = ensure_contiguous(image_batch.detach().cpu())
+
+        started = time.time()
+        sent_bytes = 0
+        for index, (start, count) in enumerate(ranges):
+            frames = [cpu_batch[i:i + 1] for i in range(start, start + count)]
+            descriptors, parts = await loop.run_in_executor(None, _encode_exact_frames, frames)
+            is_final_chunk = index == len(ranges) - 1
+            envelope = {
+                "job_id": str(multi_job_id),
+                "worker_id": str(worker_id),
+                "format": transfer_codec.EXACT_FORMAT,
+                "start_index": start,
+                "count": count,
+                "width": width,
+                "height": height,
+                "descriptors": descriptors,
+                "is_last": bool(is_final_chunk and not audio_separate),
+            }
+            if is_final_chunk and audio_inline:
+                envelope["audio"] = encoded_audio
+            chunk_bytes = sum(len(p) for p in parts)
+            body = self._build_multipart(envelope, parts)()
+            if body.size is None or body.size > TRANSFER_MAX_REQUEST_BYTES:
+                raise transfer_codec.FrameTooLargeError("exact multipart body exceeds the request limit")
+            timeout_seconds = min(600, 60 + chunk_bytes // (256 * 1024))
+            try:
+                async with session.post(
+                    url,
+                    data=body,
+                    timeout=aiohttp.ClientTimeout(total=timeout_seconds),
+                ) as response:
+                    if response.status >= 400:
+                        body = await response.text()
+                        raise RuntimeError(
+                            f"Master rejected exact chunk {index + 1}/{len(ranges)} ({chunk_bytes} bytes) "
+                            f"with HTTP {response.status}: {body[:300]}. exact needs a master with exact "
+                            "transfer support and requests within its --max-upload-size; it is never "
+                            "resent as an 8-bit format. Update the master or choose png."
+                        )
+            except Exception as e:
+                log(f"Worker - Failed to send exact chunk to master: {e}")
+                debug_log(f"Worker - Full error details: URL={url}")
+                raise
+            sent_bytes += chunk_bytes
+
+        requests = len(ranges)
+        if audio_separate:
+            await self._post_json_completion(session, url, {
+                "job_id": str(multi_job_id),
+                "worker_id": str(worker_id),
+                "batch_idx": 0,
+                "audio": encoded_audio,
+                "is_last": True,
+            })
+            requests += 1
+        log(
+            f"Worker - Sent {total} image(s) as exact in {requests} request(s): "
+            f"{sent_bytes / 1e6:.2f} MB in {time.time() - started:.1f}s"
+        )
+
     @staticmethod
     async def _post_json_completion(session, url, payload):
         timeout_seconds = 60 if "image" in payload else 600
@@ -448,6 +588,31 @@ class DistributedCollectorNode:
         worker_images[worker_id][image_index] = tensor
         return 1
 
+    # Collector sockets a comfy-media-v1 serverless pool may return.
+    TYPED_MEDIA_SOCKETS = ("images", "audio")
+
+    def _store_typed_media(self, worker_images: dict, worker_audio: dict, worker_id: str, media) -> int:
+        """Store one atomic typed serverless delivery: ordered IMAGE/AUDIO item lists.
+
+        Each IMAGE item is a BHWC batch kept at its list position; AUDIO items are
+        joined in order, exactly as list inputs are on the master.
+        """
+        if (not isinstance(media, dict) or not media
+                or any(socket not in self.TYPED_MEDIA_SOCKETS for socket in media)):
+            raise RuntimeError(f"Invalid serverless media delivery from {worker_id}")
+        for socket, items in media.items():
+            if not isinstance(items, list) or not items or any(item is None for item in items):
+                raise RuntimeError(f"Serverless participant {worker_id} returned no {socket} items")
+        stored = 0
+        for index, tensor in enumerate(media.get("images", [])):
+            if not isinstance(tensor, torch.Tensor) or tensor.ndim != 4:
+                raise RuntimeError(f"Serverless participant {worker_id} returned an invalid IMAGE item")
+            worker_images.setdefault(worker_id, {})[index] = tensor
+            stored += 1
+        if "audio" in media:
+            worker_audio[worker_id] = self._normalize_audio_input(media["audio"])
+        return stored
+
     def _reorder_and_combine_tensors(
         self,
         worker_images: dict,
@@ -514,6 +679,21 @@ class DistributedCollectorNode:
                 seen_enabled.add(worker_id_str)
                 enabled_workers.append(worker_id_str)
             expected_workers = set(enabled_workers)
+            serverless_workers = {
+                worker_id for worker_id in expected_workers if worker_id.startswith("serverless:")
+            }
+            serverless_manager = None
+
+            def get_serverless_manager():
+                nonlocal serverless_manager
+                if serverless_workers and serverless_manager is None:
+                    serverless_manager = self._get_serverless_manager()
+                return serverless_manager
+
+            try:
+                exact_requested = transfer_codec.normalize_format(transfer_format) == transfer_codec.EXACT_FORMAT
+            except transfer_codec.TransferCodecError:
+                exact_requested = False
             num_workers = len(expected_workers)
             if num_workers == 0:
                 return (images, audio if audio is not None else self.EMPTY_AUDIO)
@@ -580,152 +760,202 @@ class DistributedCollectorNode:
                 workers_done.add(done_worker_id)
                 p.update(1)  # +1 per completed expected worker
 
+            def consume_result(result):
+                nonlocal collected_count
+                if "serverless_error" in result:
+                    error = result.get("serverless_error")
+                    message = str(error).strip() if error is not None else "unspecified serverless failure"
+                    participant = result.get("worker_id", "unknown")
+                    raise RuntimeError(f"Serverless participant {participant} failed: {message}")
+
+                result_worker_id = str(result["worker_id"])
+                if "media" in result:
+                    if result_worker_id not in serverless_workers:
+                        raise RuntimeError(f"Unexpected serverless media participant {result_worker_id}")
+                    if result_worker_id in workers_done:
+                        raise RuntimeError(f"Duplicate serverless media delivery from {result_worker_id}")
+                    if result.get("is_last") is not True:
+                        raise RuntimeError("Serverless media delivery must be atomic and final")
+                    collected_count += self._store_typed_media(
+                        worker_images, worker_audio, result_worker_id, result["media"]
+                    )
+                    debug_log(f"Master - Got typed media from serverless participant {result_worker_id}")
+                    mark_worker_done(result_worker_id)
+                    return
+
+                if (exact_requested and result.get("tensor") is not None
+                        and result.get("format") != transfer_codec.EXACT_FORMAT):
+                    raise RuntimeError(
+                        f"Worker {result_worker_id} sent {result.get('format')} images, but this Collector "
+                        "requested exact. Update ComfyUI-Distributed on that worker; exact is never "
+                        "replaced by an 8-bit format."
+                    )
+                is_last = result.get("is_last", False)
+                collected_count += self._store_worker_result(worker_images, result)
+                debug_log(
+                    f"Master - Got canonical result from worker {result_worker_id}, "
+                    f"image {result.get('image_index', 0)}, is_last={is_last}"
+                )
+                result_audio = result.get("audio")
+                if result_audio is not None:
+                    worker_audio[result_worker_id] = result_audio
+                    debug_log(f"Master - Got audio from worker {result_worker_id}")
+                if is_last:
+                    mark_worker_done(result_worker_id)
+
             try:
                 while len(workers_done) < num_workers:
-                    # Check for user interruption to abort collection promptly
+                    # Check for user interruption to abort collection promptly.
                     comfy.model_management.throw_exception_if_processing_interrupted()
                     try:
-                        # Get the queue again each time to ensure we have the right reference
                         async with prompt_server.distributed_jobs_lock:
                             q = prompt_server.distributed_pending_jobs[multi_job_id]
-                            current_size = q.qsize()
-                        
                         result = await asyncio.wait_for(q.get(), timeout=slice_timeout)
-                        worker_id = result['worker_id']
-                        is_last = result.get('is_last', False)
-                        count = self._store_worker_result(worker_images, result)
-                        collected_count += count
-                        debug_log(
-                            f"Master - Got canonical result from worker {worker_id}, "
-                            f"image {result.get('image_index', 0)}, is_last={is_last}"
-                        )
+                        consume_result(result)
 
-                        # Collect audio data if present
-                        result_audio = result.get('audio')
-                        if result_audio is not None:
-                            worker_audio[worker_id] = result_audio
-                            debug_log(f"Master - Got audio from worker {worker_id}")
-
-                        # Record activity and refresh timeout baseline
+                        # Real results refresh the regular-worker timeout baseline.
                         last_activity = time.time()
                         base_timeout = float(get_worker_timeout_seconds())
 
-                        if is_last:
-                            mark_worker_done(worker_id)
-                        
                     except asyncio.TimeoutError:
-                        # If we still have time, continue polling; otherwise handle timeout
                         if (time.time() - last_activity) < base_timeout:
                             comfy.model_management.throw_exception_if_processing_interrupted()
                             continue
-                        # Re-check for user interruption after timeout expiry
                         comfy.model_management.throw_exception_if_processing_interrupted()
-                        missing_workers = set(str(w) for w in enabled_workers) - workers_done
+                        missing_workers = expected_workers - workers_done
+
+                        # A queue delivery can race the timed wait. Process it before
+                        # declaring an owned provider slot stopped without a result.
+                        async with prompt_server.distributed_jobs_lock:
+                            current_q = prompt_server.distributed_pending_jobs.get(multi_job_id)
+                            queue_has_results = current_q is not None and not current_q.empty()
+                        if queue_has_results:
+                            continue
+
+                        missing_serverless = missing_workers & serverless_workers
+                        if missing_serverless:
+                            manager = get_serverless_manager()
+                            if manager is None:
+                                raise RuntimeError("Serverless manager is unavailable for pending collector slots")
+                            pending_serverless = []
+                            for participant_id in sorted(missing_serverless):
+                                if manager.is_pending(multi_job_id, participant_id):
+                                    pending_serverless.append(participant_id)
+                                elif manager.owns(multi_job_id, participant_id):
+                                    raise RuntimeError(
+                                        f"Serverless participant {participant_id} stopped without delivering a result"
+                                    )
+                            # Provider deadlines are managed by the serverless runtime;
+                            # do not apply the regular worker heartbeat timeout to it.
+                            if pending_serverless:
+                                continue
+
+                        missing_regular = missing_workers - serverless_workers
                         elapsed = time.time() - last_activity
-                        for missing_worker_id in sorted(missing_workers):
+                        for missing_worker_id in sorted(missing_regular):
                             log(
                                 "Master - Heartbeat timeout: "
                                 f"worker={missing_worker_id}, elapsed={elapsed:.1f}s"
                             )
-                        log(
-                            f"Master - Heartbeat timeout. Still waiting for workers: {list(missing_workers)} "
-                            f"(elapsed={elapsed:.1f}s)"
-                        )
+                        if missing_regular:
+                            log(
+                                f"Master - Heartbeat timeout. Still waiting for workers: {list(sorted(missing_regular))} "
+                                f"(elapsed={elapsed:.1f}s)"
+                            )
 
-                        # Probe missing workers' /prompt endpoints to check if they are actively processing
+                        # Probe only regular worker /prompt endpoints.
                         any_busy = False
-                        try:
-                            cfg = load_config()
-                            cfg_workers = cfg.get('workers', [])
-                            for wid in list(missing_workers):
-                                wrec = next((w for w in cfg_workers if str(w.get('id')) == str(wid)), None)
-                                if not wrec:
-                                    debug_log(f"Collector probe: worker {wid} not found in config")
-                                    continue
-                                worker_url = build_worker_url(wrec)
-                                try:
-                                    payload = await probe_worker(worker_url, timeout=2.0)
-                                    queue_remaining = None
-                                    if payload is not None:
-                                        queue_remaining = int(payload.get('exec_info', {}).get('queue_remaining', 0))
-                                    debug_log(
-                                        "Collector probe: worker "
-                                        f"{wid} online={payload is not None} queue_remaining={queue_remaining}"
-                                    )
-                                    if payload is not None and queue_remaining and queue_remaining > 0:
-                                        any_busy = True
-                                        log(
-                                            f"Master - Probe grace: worker {wid} appears busy "
-                                            f"(queue_remaining={queue_remaining}). Continuing to wait."
+                        if missing_regular:
+                            try:
+                                cfg = load_config()
+                                cfg_workers = cfg.get('workers', [])
+                                for wid in sorted(missing_regular):
+                                    wrec = next((w for w in cfg_workers if str(w.get('id')) == str(wid)), None)
+                                    if not wrec:
+                                        debug_log(f"Collector probe: worker {wid} not found in config")
+                                        continue
+                                    worker_url = build_worker_url(wrec)
+                                    try:
+                                        payload = await probe_worker(worker_url, timeout=2.0)
+                                        queue_remaining = None
+                                        if payload is not None:
+                                            queue_remaining = int(payload.get('exec_info', {}).get('queue_remaining', 0))
+                                        debug_log(
+                                            "Collector probe: worker "
+                                            f"{wid} online={payload is not None} queue_remaining={queue_remaining}"
                                         )
-                                        break
-                                except Exception as e:
-                                    debug_log(f"Collector probe failed for worker {wid}: {e}")
-                        except Exception as e:
-                            debug_log(f"Collector probe setup error: {e}")
+                                        if payload is not None and queue_remaining and queue_remaining > 0:
+                                            any_busy = True
+                                            log(
+                                                f"Master - Probe grace: worker {wid} appears busy "
+                                                f"(queue_remaining={queue_remaining}). Continuing to wait."
+                                            )
+                                            break
+                                    except Exception as e:
+                                        debug_log(f"Collector probe failed for worker {wid}: {e}")
+                            except Exception as e:
+                                debug_log(f"Collector probe setup error: {e}")
 
                         if any_busy:
-                            # Refresh last_activity and continue waiting
                             last_activity = time.time()
-                            # Refresh base timeout in case the user changed it in UI
                             base_timeout = float(get_worker_timeout_seconds())
                             continue
-                        
-                        # Check queue size again with lock
+
                         async with prompt_server.distributed_jobs_lock:
-                            if multi_job_id in prompt_server.distributed_pending_jobs:
-                                final_q = prompt_server.distributed_pending_jobs[multi_job_id]
-                                final_size = final_q.qsize()
-                                
-                                # Try to drain any remaining items
-                                remaining_items = []
+                            final_q = prompt_server.distributed_pending_jobs.get(multi_job_id)
+                            remaining_items = []
+                            if final_q is not None:
                                 while not final_q.empty():
                                     try:
-                                        item = final_q.get_nowait()
-                                        remaining_items.append(item)
+                                        remaining_items.append(final_q.get_nowait())
                                     except asyncio.QueueEmpty:
                                         break
-                                
-                                if remaining_items:
-                                    # Process them
-                                    for item in remaining_items:
-                                        worker_id = item['worker_id']
-                                        is_last = item.get('is_last', False)
-
-                                        collected_count += self._store_worker_result(worker_images, item)
-                                        
-                                        if is_last:
-                                            mark_worker_done(worker_id)
-                            else:
-                                log(f"Master - Queue {multi_job_id} no longer exists!")
+                        for item in remaining_items:
+                            consume_result(item)
                         break
-            except comfy.model_management.InterruptProcessingException:
-                # Cleanup queue on interruption and re-raise to abort prompt cleanly
+
+                if exact_requested and workers_done != expected_workers:
+                    raise RuntimeError(
+                        f"Exact transfer incomplete: missing completion from {sorted(expected_workers - workers_done)}"
+                    )
+                total_collected = sum(len(imgs) for imgs in worker_images.values())
+
+                # Remove the queue before combining so a late provider callback cannot
+                # resurrect results for a collector that has completed.
                 async with prompt_server.distributed_jobs_lock:
-                    if multi_job_id in prompt_server.distributed_pending_jobs:
-                        del prompt_server.distributed_pending_jobs[multi_job_id]
+                    prompt_server.distributed_pending_jobs.pop(multi_job_id, None)
+
+                combined_audio = self._combine_audio(master_audio, worker_audio, self.EMPTY_AUDIO, enabled_workers)
+                serverless_results = set(worker_images) & serverless_workers
+                try:
+                    combined = self._reorder_and_combine_tensors(
+                        worker_images, enabled_workers, master_batch_size, images_on_cpu, delegate_mode, images
+                    )
+                    if combined is None:
+                        debug_log(f"Master - Job {multi_job_id} complete with audio only")
+                    else:
+                        debug_log(f"Master - Job {multi_job_id} complete. Combined {combined.shape[0]} images total "
+                                  f"(master: {master_batch_size}, workers: {combined.shape[0] - master_batch_size})")
+                    return (combined, combined_audio)
+                except Exception as e:
+                    if serverless_results or exact_requested:
+                        raise RuntimeError(
+                            "Serverless or exact result batch shape mismatch while combining participants "
+                            f"{sorted(worker_images)}: {e}"
+                        ) from e
+                    log(f"Master - Error combining images: {e}")
+                    # Preserve collected audio for regular workers when image assembly fails.
+                    return (images, combined_audio)
+            except BaseException:
+                # Remove the queue first, preventing cancelled work from delivering into
+                # a collector that is already unwinding.
+                async with prompt_server.distributed_jobs_lock:
+                    prompt_server.distributed_pending_jobs.pop(multi_job_id, None)
+                if serverless_workers:
+                    try:
+                        manager = get_serverless_manager()
+                        if manager is not None:
+                            await manager.cancel(collector_job_id=multi_job_id)
+                    except Exception as cancel_error:
+                        debug_log(f"Collector - Serverless cancellation failed: {cancel_error}")
                 raise
-            
-            total_collected = sum(len(imgs) for imgs in worker_images.values())
-            
-            # Clean up job queue
-            async with prompt_server.distributed_jobs_lock:
-                if multi_job_id in prompt_server.distributed_pending_jobs:
-                    del prompt_server.distributed_pending_jobs[multi_job_id]
-
-            combined_audio = self._combine_audio(master_audio, worker_audio, self.EMPTY_AUDIO, enabled_workers)
-            try:
-                combined = self._reorder_and_combine_tensors(
-                    worker_images, enabled_workers, master_batch_size, images_on_cpu, delegate_mode, images
-                )
-                if combined is None:
-                    debug_log(f"Master - Job {multi_job_id} complete with audio only")
-                else:
-                    debug_log(f"Master - Job {multi_job_id} complete. Combined {combined.shape[0]} images total "
-                              f"(master: {master_batch_size}, workers: {combined.shape[0] - master_batch_size})")
-
-                return (combined, combined_audio)
-            except Exception as e:
-                log(f"Master - Error combining images: {e}")
-                # Preserve collected audio even when image assembly fails.
-                return (images, combined_audio)

nodes/image_list.py

@@ -0,0 +1,376 @@
+"""Public V3 Start/End nodes for dynamic per-image image-list execution."""
+from __future__ import annotations
+
+import json
+
+from comfy_api.v0_0_2 import io
+
+from ..utils.async_helpers import run_async_in_server_loop
+from ..image_list.runtime import (
+    MAX_VALUE_SLOTS,
+    VALUE_NAMES,
+    VALUES_INPUT,
+    end_override_names,
+    create_master_job,
+    normalize_image_list_inputs,
+    normalize_worker_result,
+    run_image_list_job,
+    encode_result,
+    encode_request_payload,
+    unwrap_single,
+    is_valid_item_identity,
+)
+
+
+def _stable_worker_binding(prompt, start_id, identity):
+    """Resolve transient state only from the exact Start -> End context wire."""
+    if not is_valid_item_identity(identity):
+        raise ValueError("Invalid worker item identity")
+    start = prompt.get(start_id) if isinstance(prompt, dict) else None
+    if (not isinstance(start, dict) or start.get("class_type") != "DistributedImageListStart"
+            or start.get("inputs", {}).get("worker_item_identity") != identity):
+        raise RuntimeError("Worker item identity is missing its paired Start")
+    candidates = []
+    for end_id, node in prompt.items():
+        if not isinstance(node, dict) or node.get("class_type") != "DistributedImageListEnd":
+            continue
+        inputs = node.get("inputs", {})
+        link = inputs.get("context")
+        if isinstance(link, (list, tuple)) and list(link) == [start_id, 3]:
+            candidates.append((end_id, inputs))
+    if len(candidates) != 1:
+        raise RuntimeError("Worker Start must have exactly one paired End context wire")
+    end_id, inputs = candidates[0]
+    if inputs.get("worker_item_identity") != identity:
+        raise ValueError("Paired End worker item identity does not match Start")
+    for name in ("worker_result_token", "worker_context_id"):
+        if not isinstance(inputs.get(name), str) or not inputs[name]:
+            raise RuntimeError(f"Worker paired End is missing {name}")
+    index = inputs.get("worker_item_index")
+    if isinstance(index, bool) or not isinstance(index, int) or index < 0:
+        raise ValueError("Worker paired End item index is invalid")
+    return end_id, inputs
+
+
+# Types the transport can carry exactly. Anything else (models, latents, ...)
+# must be loaded inside the region or shared as a model dependency instead.
+_VALUE_TYPES = [io.Image, io.Mask, io.String, io.Int, io.Float, io.Boolean]
+assert len(VALUE_NAMES) == MAX_VALUE_SLOTS
+
+
+class DistributedImageListStart(io.ComfyNode):
+    """Materialize a batch once, then expose ordered per-item typed values."""
+
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id="DistributedImageListStart",
+            display_name="Distributed Image List Start",
+            category="image/distributed",
+            description=("Starts one per-image processing region. Each image becomes one job. "
+                         "mask and value_N inputs take 1 value (shared by every image) or exactly "
+                         "one value per image, paired in order."),
+            inputs=[
+                io.Image.Input("images", optional=True, tooltip="Image list or image batch; each frame becomes one job."),
+                io.Mask.Input("mask", optional=True,
+                              tooltip="1 mask shared by every image, or one mask per image."),
+                io.Autogrow.Input(VALUES_INPUT, optional=True, template=io.Autogrow.TemplateNames(
+                    io.MultiType.Input("value", types=_VALUE_TYPES,
+                                       tooltip=("Any image, mask, text, number or boolean. 1 value is shared by "
+                                                "every image; one value per image pairs them in order. "
+                                                "A batch stays one value.")),
+                    names=list(VALUE_NAMES), min=0)),
+            ],
+            outputs=[
+                io.Image.Output(display_name="image", is_output_list=True),
+                io.Mask.Output(display_name="mask", is_output_list=True),
+                io.Int.Output(display_name="item_index", is_output_list=True),
+                io.Custom("DISTRIBUTED_IMAGE_LIST_CONTEXT").Output(display_name="context", is_output_list=True),
+                # value_N outputs mirror value_N inputs. The web extension hides
+                # unused ones and shows the connected input's type.
+                *[io.AnyType.Output(id=name, display_name=name, is_output_list=True) for name in VALUE_NAMES],
+            ],
+            is_input_list=True,
+            accept_all_inputs=True,
+            not_idempotent=True,
+            hidden=[io.Hidden.prompt, io.Hidden.unique_id],
+        )
+
+    @classmethod
+    def execute(cls, images=None, mask=None, values=None,
+                worker_item_token="", worker_context_id="", image_list_enabled_worker_ids="[]",
+                image_list_master_callback="", image_list_execution_id="",
+                image_list_workflow_meta=None, image_list_master_participates=None,
+                image_list_local_execution=False, worker_item_identity=None):
+        worker_item_token = unwrap_single(worker_item_token)
+        worker_context_id = unwrap_single(worker_context_id)
+        worker_item_identity = unwrap_single(worker_item_identity)
+        image_list_workflow_meta = unwrap_single(image_list_workflow_meta)
+        if worker_item_identity is not None:
+            prompt = getattr(cls.hidden, "prompt", None)
+            start_id = str(getattr(cls.hidden, "unique_id", ""))
+            _, delivery = _stable_worker_binding(prompt, start_id, worker_item_identity)
+            # Only cache misses execute Start. Never put this token/context in
+            # Start's literal inputs or returned context: both would poison reuse.
+            from ..image_list.runtime import consume_worker_item
+            item = consume_worker_item(delivery["worker_result_token"],
+                                       item_identity=worker_item_identity,
+                                       item_index=delivery["worker_item_index"])
+            item_values = item.get("values") or {}
+            return io.NodeOutput(
+                [item["image"]], [item.get("mask")], [item["index"]], [worker_item_identity],
+                *[[item_values.get(name)] for name in VALUE_NAMES],
+            )
+        if worker_item_token:
+            if unwrap_single(image_list_local_execution):
+                from ..image_list.runtime import consume_worker_item
+                item = consume_worker_item(worker_item_token)
+            else:
+                item = run_async_in_server_loop(_consume_worker_item(worker_item_token), timeout=10)
+            context_id = worker_context_id
+            if not isinstance(context_id, str) or not context_id:
+                raise RuntimeError("Worker image-list prompt is missing its paired context")
+            item_values = item.get("values") or {}
+            return io.NodeOutput(
+                [item["image"]], [item.get("mask")], [item["index"]], [context_id],
+                *[[item_values.get(name)] for name in VALUE_NAMES],
+            )
+
+        items = normalize_image_list_inputs(images, mask, values)
+        prompt = getattr(cls.hidden, "prompt", None)
+        start_id = str(getattr(cls.hidden, "unique_id", ""))
+        if not prompt or not start_id:
+            raise RuntimeError("Distributed Image List Start needs a submitted ComfyUI prompt context")
+        try:
+            worker_ids = json.loads(str(unwrap_single(image_list_enabled_worker_ids or "[]")))
+        except (TypeError, ValueError) as exc:
+            raise RuntimeError("Invalid preflighted image-list worker list") from exc
+        if not isinstance(worker_ids, list) or not all(isinstance(item, str) for item in worker_ids):
+            raise RuntimeError("Invalid preflighted image-list worker list")
+        from comfy_execution.utils import get_executing_context
+        from ..image_list.lifecycle import has_safe_execution_targets
+        import server
+        executing = get_executing_context()
+        explicit_targets = has_safe_execution_targets(
+            getattr(server.PromptServer, "instance", None),
+            executing.prompt_id if executing is not None else None, prompt)
+        from ..image_list.runtime import validate_image_list_regions
+        regions = validate_image_list_regions(prompt, explicit_targets=explicit_targets)
+        region = next((row for row in regions if row["start_id"] == start_id), None)
+        if region is None:
+            raise RuntimeError(f"No safe paired Image List End was found for Start node {start_id}")
+        master_participates = unwrap_single(image_list_master_participates)
+        if master_participates is None:
+            from ..utils.config import load_config
+            master_participates = not bool((load_config().get("settings", {}) or {}).get("master_delegate_only", False))
+        if not isinstance(master_participates, bool):
+            raise ValueError("image_list_master_participates must be a boolean")
+        context_id = create_master_job(
+            start_id, region["end_id"], prompt, items, worker_ids,
+            str(unwrap_single(image_list_master_callback or "")),
+            image_list_workflow_meta if isinstance(image_list_workflow_meta, dict) else None,
+            prompt_id=executing.prompt_id if executing is not None else None,
+            master_participates=master_participates,
+            explicit_targets=explicit_targets,
+        )
+        return io.NodeOutput(
+            [item["image"] for item in items],
+            [item.get("mask") for item in items],
+            [item["index"] for item in items],
+            [context_id for _ in items],
+            *[[item["values"].get(name) for item in items] for name in VALUE_NAMES],
+        )
+
+
+async def _consume_worker_item(token):
+    # Item state is local to each worker process; source nodes consume it exactly once.
+    from ..image_list.runtime import consume_worker_item
+    return consume_worker_item(token)
+
+
+class DistributedImageListEnd(io.ComfyNode):
+    """Collect branch results in source order without evaluating the branch on master."""
+
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id="DistributedImageListEnd",
+            display_name="Distributed Image List End",
+            category="image/distributed",
+            description=("Closes a paired image-list region and returns ordered image, mask, text and values lists. "
+                         "Start values are retained; connected value_N inputs replace or add that key per item."),
+            inputs=[
+                io.Image.Input("images", lazy=True),
+                io.Mask.Input("masks", optional=True, lazy=True),
+                io.String.Input("texts", optional=True, lazy=True, force_input=True),
+                io.Custom("DISTRIBUTED_IMAGE_LIST_CONTEXT").Input("context"),
+                *[io.MultiType.Input(name, types=_VALUE_TYPES, optional=True, lazy=True,
+                                     tooltip="Replace or add this value for each item; exactly one value per item.")
+                  for name in VALUE_NAMES],
+            ],
+            outputs=[
+                io.Image.Output(display_name="image_list", is_output_list=True),
+                io.Mask.Output(display_name="masks", is_output_list=True),
+                io.String.Output(display_name="texts", is_output_list=True),
+                io.Custom("DISTRIBUTED_IMAGE_LIST_VALUES").Output(
+                    id="values", display_name="values", is_output_list=True),
+            ],
+            is_input_list=True,
+            is_output_node=True,
+            accept_all_inputs=True,
+            not_idempotent=True,
+            hidden=[io.Hidden.prompt, io.Hidden.unique_id, io.Hidden.execution_list],
+        )
+
+    @classmethod
+    def check_lazy_status(cls, images=None, masks=None, texts=None, context=None,
+                          worker_result_token="", **kwargs):
+        worker_result_token = unwrap_single(worker_result_token)
+        if worker_result_token:
+            # Decide which optional sockets are connected from the actual worker prompt;
+            # absent optional inputs must not be demanded from Comfy's lazy executor.
+            prompt = getattr(cls.hidden, "prompt", {}) or {}
+            end_inputs = next((node.get("inputs", {}) for node in prompt.values()
+                               if isinstance(node, dict)
+                               and node.get("class_type") == "DistributedImageListEnd"
+                               and node.get("inputs", {}).get("worker_result_token") == worker_result_token), {})
+            required = []
+            for name, value in (("images", images), ("masks", masks), ("texts", texts),
+                                *((name, kwargs.get(name)) for name in end_override_names(end_inputs))):
+                connected = name == "images" or name in end_inputs
+                missing = value is None or (isinstance(value, (list, tuple)) and all(v is None for v in value))
+                if connected and missing:
+                    required.append(name)
+            return required
+        # Master only needs Start's context. Never request branch outputs here.
+        if context is None or (isinstance(context, (tuple, list)) and all(item is None for item in context)):
+            return ["context"]
+        return []
+
+    @classmethod
+    def execute(cls, images=None, masks=None, texts=None, context=None,
+                worker_result_token="", worker_item_index=-1, master_callback_url="",
+                image_list_local_execution=False, worker_context_id="",
+                worker_item_identity=None, **values):
+        worker_result_token = unwrap_single(worker_result_token)
+        master_callback_url = unwrap_single(master_callback_url)
+        worker_context_id = unwrap_single(worker_context_id)
+        worker_item_identity = unwrap_single(worker_item_identity)
+        if worker_item_identity is not None and not worker_result_token:
+            raise RuntimeError("Worker item identity is missing its delivery token")
+        if worker_result_token:
+            context_values = [item for item in (context if isinstance(context, (list, tuple)) else [context])
+                              if isinstance(item, str) and item]
+            if not context_values or len(set(context_values)) != 1:
+                raise RuntimeError("Worker Image List End is missing its paired context")
+            prompt = getattr(cls.hidden, "prompt", {}) or {}
+            context_id = context_values[0]
+            if worker_item_identity is not None:
+                end_id = str(getattr(cls.hidden, "unique_id", ""))
+                node = prompt.get(end_id, {})
+                link = node.get("inputs", {}).get("context")
+                if (node.get("class_type") != "DistributedImageListEnd"
+                        or not isinstance(link, (list, tuple)) or len(link) != 2):
+                    raise RuntimeError("Worker End is missing its paired Start context wire")
+                paired_id, end_inputs = _stable_worker_binding(prompt, link[0], worker_item_identity)
+                if (paired_id != end_id or context_values != [worker_item_identity]
+                        or end_inputs["worker_context_id"] != worker_context_id
+                        or end_inputs["worker_result_token"] != worker_result_token
+                        or end_inputs["worker_item_index"] != unwrap_single(worker_item_index)
+                        or end_inputs.get("master_callback_url", "") != master_callback_url):
+                    raise RuntimeError("Worker End identity/context does not match its current delivery")
+                from ..image_list.runtime import finish_worker_item
+                # On a miss this consumes Start's receipt; on a hit it validates
+                # and drops the current staged tensors. Both paths enforce TTL.
+                finish_worker_item(worker_result_token, worker_item_identity,
+                                   unwrap_single(worker_item_index))
+                context_id = worker_context_id
+            else:
+                end_inputs = next((node.get("inputs", {}) for node in prompt.values()
+                                   if isinstance(node, dict) and node.get("class_type") == "DistributedImageListEnd"
+                                   and node.get("inputs", {}).get("worker_result_token") == worker_result_token), {})
+            expected = end_override_names(end_inputs)
+            # Missing connected lazy values must fail instead of becoming an
+            # implicit fallback to the original Start value.
+            supplied = {name: values.get(name) for name in expected}
+            supplied.update(values)
+            result = normalize_worker_result(images, masks, texts, supplied)
+            if unwrap_single(image_list_local_execution):
+                from ..image_list.runtime import set_master_result
+                set_master_result(context_id, int(unwrap_single(worker_item_index)),
+                                  str(worker_result_token), result)
+            else:
+                payload = encode_result(result)
+                run_async_in_server_loop(
+                    submit_worker_result(str(master_callback_url), str(worker_result_token),
+                                         context_id, int(unwrap_single(worker_item_index)), payload), timeout=60,
+                )
+            return io.NodeOutput([result["image"]], [result.get("mask")], [result.get("text")],
+                                 [result.get("values", {})])
+
+        context_values = [item for item in (context if isinstance(context, (list, tuple)) else [context])
+                          if isinstance(item, str) and item]
+        if not context_values or len(set(context_values)) != 1:
+            raise RuntimeError("Distributed Image List End requires the paired Start context wire")
+        context_id = context_values[0]
+        from ..image_list.runtime import ImageListCancelledError
+        try:
+            # The runtime owns the configurable deadline and bounded remote cleanup.
+            execution_list = getattr(cls.hidden, "execution_list", None)
+            # Local model outputs belong to this still-running outer prompt.
+            # Add them to its existing tracker; never start/end another tracker
+            # that could unpin a model still needed by the parent.
+            model_callback = getattr(execution_list, "output_link_callback", None)
+            result_lists = run_async_in_server_loop(run_image_list_job(
+                context_id, model_callback=model_callback, include_values=True))
+        except ImageListCancelledError:
+            import comfy.model_management
+            raise comfy.model_management.InterruptProcessingException()
+        images_out, masks_out, texts_out, values_out = result_lists
+        if len({len(output) for output in result_lists}) != 1:
+            raise RuntimeError("Image-list collector received mismatched image, mask, text or values result counts")
+        return io.NodeOutput(images_out, masks_out, texts_out, values_out)
+
+
+class DistributedImageListGetValue(io.ComfyNode):
+    """Ordinary scalar execution lets core map over End's ordered values list."""
+
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id="DistributedImageListGetValue",
+            display_name="Distributed Image List Get Value",
+            category="image/distributed",
+            inputs=[io.Custom("DISTRIBUTED_IMAGE_LIST_VALUES").Input("values"),
+                    io.Combo.Input("value_name", options=list(VALUE_NAMES), default="value_1")],
+            outputs=[io.AnyType.Output(id="value", display_name="value")],
+        )
+
+    @classmethod
+    def execute(cls, values, value_name="value_1"):
+        if value_name not in VALUE_NAMES:
+            raise ValueError(f"Unknown image-list value slot {value_name!r}")
+        if not isinstance(values, dict):
+            raise TypeError("Image List Get Value expects one values mapping from Image List End")
+        if value_name not in values:
+            raise ValueError(f"Image List Get Value: {value_name} is absent from this item's values")
+        return io.NodeOutput(values[value_name])
+
+
+async def submit_worker_result(callback_url, item_token, context_id, item_index, result):
+    if not callback_url.startswith(("http://", "https://")):
+        raise RuntimeError("Worker image-list callback URL is missing or invalid")
+    import aiohttp
+    from ..utils.network import get_client_session
+    session = await get_client_session()
+    payload = {"context_id": context_id, "token": item_token,
+               "item_index": item_index, "result": result}
+    async with session.post(
+        f"{callback_url.rstrip('/')}/distributed/image-list/result",
+        data=encode_request_payload(payload),
+        headers={"Content-Type": "application/json"},
+        timeout=aiohttp.ClientTimeout(total=60),
+    ) as response:
+        if response.status >= 400:
+            body = (await response.text())[:500]
+            raise RuntimeError(f"Master rejected image-list result (HTTP {response.status}): {body}")

nodes/v3.py

@@ -10,6 +10,7 @@ from . import utilities as _utilities
 from .collector import DistributedCollectorNode as _CollectorRuntime
 from ..utils import transfer_codec as _transfer_codec
 from .distributed_upscale import UltimateSDUpscaleDistributed as _UpscaleRuntime
+from .image_list import DistributedImageListEnd, DistributedImageListStart, DistributedImageListGetValue
 
 
 class DistributedSeed(io.ComfyNode):
@@ -211,4 +212,4 @@ class DistributedWorkerSink(io.ComfyNode):
 
 NODES = [DistributedCollector, DistributedSeed, DistributedModelName, DistributedValue,
          ImageBatchDivider, AudioBatchDivider, DistributedEmptyImage, UltimateSDUpscaleDistributed,
-         DistributedWorkerSink]
+         DistributedWorkerSink, DistributedImageListStart, DistributedImageListEnd, DistributedImageListGetValue]

runtime/bootstrap.py

@@ -24,6 +24,9 @@ def initialize():
 
     ensure_distributed_state(server.PromptServer.instance)
     ensure_tile_jobs_initialized()
+    from ..image_list import runtime as image_list_runtime
+    from ..image_list.lifecycle import install_lifecycle
+    install_lifecycle(server.PromptServer.instance, image_list_runtime)
 
     if not os.environ.get('COMFYUI_IS_WORKER'):
         atexit.register(sync_cleanup)

serverless_bridge/README.md

@@ -0,0 +1,266 @@
+# Standalone `comfy-media-v1` endpoint bridge
+
+This directory is independently installable. It registers two internal Comfy V3
+nodes, `MediaImportV1` and `MediaExportV1`, and provides the asynchronous HTTP
+executor `execute_media_request`. It does **not** start ComfyUI, provision an
+endpoint, install models, or provide a provider worker/supervisor. The endpoint
+needs this bridge, not the full ComfyUI-Distributed extension.
+
+The bridge's generic contract supports `IMAGE`, `AUDIO`, `MASK`, `STRING`, and
+native ComfyUI `VIDEO`. **Distributed Collector** (`DistributedCollector`) uses
+only the image/audio outputs and combines their returned lists into ordinary
+IMAGE/AUDIO outputs. It has no MASK/STRING/VIDEO outputs, and the separate
+`DistributedMediaCollector` node has been removed. For Collector setup, ordering,
+and limits, see [serverless images and audio](../docs/serverless-media.md).
+
+## Install in the endpoint image
+
+1. Copy this complete directory to `<ComfyUI>/custom_nodes/serverless_bridge`.
+   Use a compatible ComfyUI with `comfy_api.v0_0_2` and the native VIDEO classes
+   used by [video_codec.py](video_codec.py). These exact-class/private-state
+   adapters are tied to the supported core implementation, not arbitrary VIDEO
+   plugins or future core layouts.
+2. Restart ComfyUI so `comfy_entrypoint()` registers the two nodes. Keep the
+   workflow's own models and custom nodes installed separately.
+3. Make `<ComfyUI>/custom_nodes` importable by the handler process (for example,
+   via its `PYTHONPATH`). The handler uses `aiohttp`; the nodes/codecs use
+   ComfyUI's PyTorch/native media dependencies. The handler module itself does
+   not import ComfyUI, PyTorch, Runpod, or Distributed bootstrap code.
+4. Give the handler filesystem access to **the same input directory** used by
+   that ComfyUI instance, plus its trusted HTTP URL. These are operator settings,
+   never request-controlled paths/URLs. Staging/cleanup requires POSIX directory
+   descriptor and no-follow file operations.
+
+Installing the nodes alone is insufficient: the endpoint must route media-v1
+inputs through this executor and implement the uncertainty policy below. A
+legacy image handler cannot interpret the typed envelope.
+
+## Callable integration
+
+```python
+from serverless_bridge.handler import execute_media_request
+
+
+async def execute_endpoint_input(
+    input_data, *, comfy_url, input_dir, temp_dir=None,
+    download=None, publish=None,
+):
+    # Invoke only under a supervisor implementing the uncertainty policy below.
+    # Exceptions intentionally propagate; this is NOT a provider worker loop.
+    return await execute_media_request(
+        input_data,
+        comfy_url=comfy_url,
+        input_dir=input_dir,
+        timeout_seconds=1800,
+        session=None,
+        download=download,
+        publish=publish,
+        temp_dir=temp_dir,
+    )
+```
+
+Pass the provider job's **`input` object**, not the whole job. Return the resulting
+`{"media": ...}` object as provider output; Runpod exposes it as `output.media`.
+Do not add another `output` wrapper inside it. `timeout_seconds` must be finite
+and positive. If supplied, `session` is a caller-owned `aiohttp.ClientSession`
+for local Comfy HTTP and remains open; otherwise the executor creates/closes its
+own session. Input dictionaries are copied before graph rewriting.
+
+**This callable is not a turnkey Runpod handler.** An outer SDK that catches every
+exception and immediately reuses the worker is unsafe. The supervisor must
+recognize `serverless_bridge.handler.MediaExecutionUncertain`, stop intake, and
+quarantine/reconcile or terminate the owning environment before reuse. Merely
+returning `{"error": ...}` or marking a provider job `FAILED` is not that policy.
+
+## Request and response
+
+A minimal generic bridge text request builder (no files or model required by this
+graph). This illustrates the bridge contract, not a Collector text output:
+
+```python
+import json
+import uuid
+
+
+def make_text_request(text="hello"):
+    request_id = uuid.uuid4().hex
+    binding = {
+        "binding_id": "text-binding",
+        "collector_id": "collector-1",
+        "job_id": "collector-job-1",
+        "participant_id": "serverless:example:0",
+        "node_id": "export-1",
+        "socket": "texts",
+        "kind": "STRING",
+    }
+    return {
+        "workflow": {
+            "text-1": {"class_type": "PrimitiveString", "inputs": {"value": text}},
+            "export-1": {
+                "class_type": "MediaExportV1",
+                "inputs": {
+                    "request_id": request_id,
+                    "binding_json": json.dumps(binding),
+                    "texts": ["text-1", 0],
+                },
+            },
+        },
+        "images": [],
+        "media": {
+            "version": 1,
+            "profile": "comfy-media-v1",
+            "request_id": request_id,
+            "artifacts": [],
+            "inputs": [],
+            "outputs": [binding],
+        },
+    }
+```
+
+`workflow` is an API graph, not an editor workflow. Top-level input fields are
+exactly `workflow`, `media`, and optional `images`. `media` has exactly the six
+fields shown above. `media.outputs` declares one exporter per connected socket;
+its binding must match that node's `binding_json` and request ID exactly. Socket
+names/types are `images/IMAGE`, `audio/AUDIO`, `masks/MASK`, `texts/STRING`, and
+`video/VIDEO`. Use unique request IDs. Identifiers are 1–200 characters from
+`[A-Za-z0-9_.:/-]`; request scopes additionally reject empty, `.` or `..` slash
+components. The executor validates the graph and submits **once**, with exactly
+the declared exporter IDs as `partial_execution_targets`. Unrelated graph roots,
+cycles, undeclared exporters, or caller-supplied `MediaImportV1` nodes fail.
+
+For typed inputs, `media.artifacts` contains records with exactly:
+
+| Field | Value |
+| --- | --- |
+| `id`, `kind` | Unique artifact ID and one of the five native types |
+| `descriptor` | Kind-specific codec metadata; use the shared codec, not pickle |
+| `byte_length`, `sha256` | Exact payload byte count and lowercase SHA-256 hex |
+| `transport` | `{"type": "inline", "data_base64": "..."}` or the URL transport below |
+
+Each `media.inputs` entry is
+`{"node_id": "consumer", "input_name": "socket", "artifact_id": "artifact-id"}`.
+It replaces an existing consumer input with a generated import link; every
+artifact must be used, and destinations cannot repeat. The executor verifies
+bytes/checksums, stages owned files, inserts imports, then prunes disconnected
+loader ancestors. Optional legacy `images` entries are `{ "name": "input.png",
+"image": "<base64>" }`, must match stock `LoadImage` references, and are rewritten
+to owned staging paths. They are not an arbitrary filename-upload interface.
+
+The returned `media` object has exactly `version`, `profile`, `request_id`,
+`bindings`, `artifacts`, and `items`:
+
+- `bindings`: every declared binding, unchanged, plus `count` (1–64).
+- `artifacts`: the same artifact schema above with inline or URL transport.
+- `items`: `{ "binding_id": "...", "item_index": 0, "artifact_id": "..." }`
+  records covering every index exactly once, ordered by `item_index` when decoded.
+
+All identity, completeness, type, size, and checksum checks must pass; do not
+consume a partial result. The internal history transport `comfy-file` is local
+only and must not be returned to the master. The executor reads validated
+`media_exports` history records and bounded `/view` payloads before constructing
+the external envelope. [contract.py](contract.py) is the schema authority.
+
+## Optional image return encoding
+
+IMAGE `MediaExportV1` nodes may also contain literal `transfer_format` and
+`transfer_quality` controls. The **bridge-level defaults** are `exact` and `90`;
+generic callers that omit the controls retain that representation. Other media
+exports must not carry image controls. Callers should not alter binding identity
+metadata.
+
+The **Collector default is different**: its saved widget value remains
+`legacy_png` with quality `90`. For a `comfy-media-v1` pool, the master compiler
+maps that value to an explicit `png` export, so normal serverless image returns
+use compressed 8-bit PNG. `exact` is opt-in at the Collector. Other selected
+formats are unchanged. Direct workers also support opt-in exact IMAGE transport;
+an older master rejection is an error, not an 8-bit retry. A Collector requesting
+`exact` cannot dispatch to a `legacy-image-v1` pool.
+
+Formats: `exact`, `legacy_png`, `png`, `ffv1`, `jpeg`, `webp`, `h264`,
+`h265` (**experimental**), `av1`.
+`exact` preserves float tensor bits. The image/video codecs quantize RGB/grayscale
+to 8-bit; PNG/FFV1 are lossless at that precision, while the other codecs are
+lossy. Quality is an integer 1–100 and affects only lossy formats. IMAGE batch
+order, shape and channel count are retained, and alpha is not discarded.
+Native VIDEO/MASK/AUDIO/STRING bridge values still use their original exact codecs;
+the image-format controls do not add those sockets to the Collector.
+
+Copy/update the complete bridge directory on both sides, including
+`image_transfer.py` and `image_codec.py`. Pillow and the selected CPU/software
+PyAV codec must be available; missing encoders fail without automatic fallback.
+Compressed IMAGE artifacts use `image-transfer-v1`; decode through the shared
+`codec.decode_payload`, which enforces payload and decoded-shape limits.
+See [image transfer choices](../docs/serverless-media.md#image-return-format-and-quality).
+
+## Transport budgets and ownership
+
+- Contract maxima: 256 MiB per artifact, 512 MiB aggregate artifact bytes,
+  128 artifacts, 64 items per binding. Request input totals include decoded
+  legacy-image bytes. Decoders validate tensor/VIDEO sizes and layouts before
+  allocating native storage; these are payload limits, **not peak-RSS guarantees**.
+- The built-in master currently emits inline inputs only. Its complete Runpod
+  request JSON is capped at **10,000,000 bytes**, including base64, graph and
+  policy. The larger contract limits do not bypass this cap; there is no
+  built-in large-input/object-storage uploader.
+- The executor's result JSON is capped at **10,000,000 bytes**. Inline results
+  are the default. When needed, `publish` is called for selected larger
+  artifacts to fit that budget; supplying it does not force all results to URLs.
+  Without a publisher, an oversized response fails instead of being truncated.
+- Optional **async** `download(url, *, max_bytes)` returns `bytes`. The executor
+  supplies the declared artifact length and checks exact length and SHA-256.
+  The callback must enforce public HTTPS, DNS/IP restrictions, redirect checks,
+  and bounded streaming **before allocation**; syntactic contract validation
+  does not provide SSRF protection. No downloader is installed by default.
+- Optional **async** `publish(metadata, payload)` uploads the exact `bytes` and
+  returns exactly `{"type": "url", "url": "https://...", "expires_at": "..."}`.
+  `metadata` is the artifact record without `transport`. Expiry must be a future,
+  timezone-aware ISO timestamp; URLs use HTTPS port 443, no userinfo/fragment,
+  and at most 8192 characters. Keep objects available long enough for collection.
+  The callback/operator owns upload credentials, object expiry and cleanup,
+  including orphan uploads after failure. There is no bridge storage backend.
+- The master uses a separate bounded public-HTTPS downloader without provider
+  credentials. Never send provider credentials to artifact hosts or log signed
+  URLs. Endpoint callbacks must enforce equivalent protections themselves.
+
+Callbacks must cooperate with async cancellation, finish/abort their own I/O,
+and not detach readers/uploads that outlive the callback. Synchronous code that
+blocks the event loop cannot be hard-timed-out by this executor.
+
+## Cancellation, quarantine, and file retention
+
+`timeout_seconds` covers staging, submission, polling, downloads and publication.
+On failure/cancellation, there is a separate **10-second callback-drain grace**
+and, when needed, a **10-second prompt-reconciliation grace**. These are recovery
+allowances, not evidence that execution stopped. Ambiguous `/prompt` submission
+is never retried. An identified pending prompt may be deleted from `/queue`,
+with exact read-back; running prompts are never sent `/interrupt`. The pinned
+core route has a snapshot/global-interrupt race even when given a prompt ID.
+
+`MediaExecutionUncertain` means prompt acceptance/termination **or callback
+teardown** is unconfirmed. Even successful terminal Comfy history is insufficient
+if an async downloader/publisher ignores cancellation. The bridge retains input
+and exporter files in that case; late callback exit does not automatically
+remove retained files. Before accepting another job on that environment:
+
+1. Quarantine it and stop new intake; do not automatically resubmit the request.
+2. Reconcile the local prompt and all readers, or terminate the owning execution
+   environment when their termination cannot be established.
+3. Only then clean retained files and allow reuse/replacement. A provider terminal
+   job status alone cannot certify a surviving local Comfy process or callback.
+
+Staged inputs use random exclusive files under
+`input_dir/distributed-media-v1/<sha256(request_id)>/`. Cleanup removes only files
+created by that invocation, after confirmed terminal state (or before submission)
+and reader teardown; it never recursively deletes a shared namespace.
+
+`temp_dir=None` leaves exporter retention to the Comfy temp-directory/environment
+lifecycle. Optionally pass the **trusted local Comfy temp root** to enable
+best-effort cleanup of validated, invocation-owned exporter UUID `.bin` files
+under the corresponding hashed namespace, only after terminal history and
+collection/publisher teardown. Symlinked, hard-linked, replaced, missing, or
+unidentifiable files are not claimed; directories and all unidentified leftovers
+remain lifecycle-owned. This is not a temp-directory sweep. Independently decoded
+VIDEO objects do not depend on either staging or exporter files.
+
+This guide describes the source contract, not a test result. No live provider/GPU
+validation is claimed, and this directory is not a turnkey provider deployment.

serverless_bridge/__init__.py

@@ -0,0 +1,14 @@
+"""Standalone typed-media bridge; copy this directory into endpoint custom_nodes.
+
+Importing codecs does not load ComfyUI, providers or Distributed bootstrap.
+"""
+
+async def comfy_entrypoint():
+    from comfy_api.v0_0_2 import ComfyExtension
+    from .nodes import MediaImportV1, MediaExportV1
+
+    class MediaBridgeExtension(ComfyExtension):
+        async def get_node_list(self):
+            return [MediaImportV1, MediaExportV1]
+
+    return MediaBridgeExtension()

serverless_bridge/codec.py

@@ -0,0 +1,189 @@
+"""Lossless, bounded CPU payloads shared with the standalone endpoint bridge.
+
+Descriptors contain only kind-specific metadata. Tensor bytes are contiguous,
+little-endian floating-point values, never a pickle or an image/audio container.
+The optional byte budget can lower, but never raise, MAX_PAYLOAD_BYTES.
+"""
+
+import ctypes
+import sys
+
+import torch
+
+
+MAX_PAYLOAD_BYTES = 256 * 1024 * 1024
+
+_DTYPES = {
+    "float16": (torch.float16, 2),
+    "bfloat16": (torch.bfloat16, 2),
+    "float32": (torch.float32, 4),
+    "float64": (torch.float64, 8),
+}
+_DTYPE_NAMES = {dtype: name for name, (dtype, _) in _DTYPES.items()}
+_LAYOUTS = {"IMAGE": "BHWC", "MASK": "BHW", "AUDIO": "BCT"}
+
+
+def _validate_kind(kind):
+    if type(kind) is not str or kind not in ("IMAGE", "MASK", "STRING", "AUDIO"):
+        raise ValueError("Unsupported payload kind")
+
+
+def _byte_limit(max_bytes):
+    if type(max_bytes) is not int or max_bytes < 0:
+        raise ValueError("max_bytes must be a nonnegative integer")
+    return min(max_bytes, MAX_PAYLOAD_BYTES)
+
+
+def _sample_rate(value):
+    if type(value) is not int or value <= 0:
+        raise ValueError("AUDIO sample_rate must be a positive integer")
+    return value
+
+
+def _tensor_size(kind, shape, itemsize, limit):
+    if type(shape) is not list or len(shape) != len(_LAYOUTS[kind]):
+        raise ValueError("Tensor shape has an invalid rank")
+    if any(type(dimension) is not int or dimension <= 0 for dimension in shape):
+        raise ValueError("Tensor shape dimensions must be positive integers")
+    if kind == "IMAGE" and shape[-1] not in (1, 3, 4):
+        raise ValueError("IMAGE shape must have 1, 3 or 4 channels")
+    size = itemsize
+    for dimension in shape:
+        # Divide before multiplying, including for arbitrary-size Python ints.
+        if dimension > limit // size:
+            raise ValueError("Tensor payload exceeds byte limit")
+        size *= dimension
+    return size
+
+
+def _native_little_endian(payload, itemsize):
+    """Convert native to wire byte order, or vice versa (an involution)."""
+    if sys.byteorder == "little":
+        return payload
+    swapped = bytearray(len(payload))
+    for offset in range(itemsize):
+        swapped[offset::itemsize] = payload[itemsize - 1 - offset::itemsize]
+    return bytes(swapped)
+
+
+def _encode_string(value, limit):
+    if not isinstance(value, str):
+        raise ValueError("STRING value must be text")
+    # Check the exact UTF-8 size before allocating the encoded payload.
+    if len(value) > limit:
+        raise ValueError("STRING payload exceeds byte limit")
+    size = 0
+    for character in value:
+        codepoint = ord(character)
+        if 0xD800 <= codepoint <= 0xDFFF:
+            raise ValueError("STRING value must be valid UTF-8 text")
+        size += 1 + (codepoint > 0x7F) + (codepoint > 0x7FF) + (codepoint > 0xFFFF)
+        if size > limit:
+            raise ValueError("STRING payload exceeds byte limit")
+    return {"encoding": "utf8-v1"}, value.encode("utf-8")
+
+
+def encode_payload(kind: str, value, *, max_bytes: int = MAX_PAYLOAD_BYTES) -> tuple[dict, bytes]:
+    """Encode a typed value without dtype conversion, normalization or GPU use.
+
+    IMAGE/MASK accept dense CPU tensors; AUDIO accepts a waveform CPU tensor
+    plus a positive integer sample_rate. STRING accepts exact Unicode text.
+    Invalid values and exceeded budgets raise nonsensitive ValueError messages.
+    """
+    _validate_kind(kind)
+    limit = _byte_limit(max_bytes)
+    if kind == "STRING":
+        return _encode_string(value, limit)
+
+    sample_rate = None
+    if kind == "AUDIO":
+        if not isinstance(value, dict) or set(value) != {"waveform", "sample_rate"}:
+            raise ValueError("AUDIO value must contain waveform and sample_rate")
+        sample_rate = _sample_rate(value["sample_rate"])
+        value = value["waveform"]
+    if not isinstance(value, torch.Tensor):
+        raise ValueError("Tensor value must be a torch.Tensor")
+    if value.device.type != "cpu":
+        raise ValueError("Tensor value must be on CPU")
+    if value.layout != torch.strided or value.is_nested:
+        raise ValueError("Tensor value must have a dense strided layout")
+    if value.dtype not in _DTYPE_NAMES:
+        raise ValueError("Unsupported tensor dtype")
+
+    dtype_name = _DTYPE_NAMES[value.dtype]
+    _, itemsize = _DTYPES[dtype_name]
+    shape = list(value.shape)
+    size = _tensor_size(kind, shape, itemsize, limit)
+    tensor = value.detach().resolve_neg().contiguous()
+    if not torch.isfinite(tensor).all().item():
+        raise ValueError("Tensor values must be finite")
+    # A uint16 view preserves BF16 bits without conversion through float32.
+    if tensor.dtype == torch.bfloat16:
+        tensor = tensor.view(torch.uint16)
+    # The live, contiguous CPU tensor owns this pointer for the entire copy.
+    payload = _native_little_endian(ctypes.string_at(tensor.data_ptr(), size), itemsize)
+    descriptor = {
+        "encoding": "tensor-le-v1", "dtype": dtype_name,
+        "shape": shape, "layout": _LAYOUTS[kind],
+    }
+    if kind == "AUDIO":
+        descriptor["sample_rate"] = sample_rate
+    return descriptor, payload
+
+
+def decode_payload(kind: str, descriptor: dict, payload: bytes, *, max_bytes: int = MAX_PAYLOAD_BYTES):
+    """Decode validated wire bytes into independent contiguous CPU tensors/text.
+
+    All metadata, shape products and byte lengths are checked before tensor
+    allocation. No GPU device discovery, transfers or object deserialization
+    occur. AUDIO returns {'waveform': tensor, 'sample_rate': int}.
+    """
+    _validate_kind(kind)
+    limit = _byte_limit(max_bytes)
+    if not isinstance(descriptor, dict):
+        raise ValueError("Payload descriptor must be a dictionary")
+    if not isinstance(payload, bytes):
+        raise ValueError("Payload must be bytes")
+    if len(payload) > limit:
+        raise ValueError("Payload exceeds byte limit")
+    if kind == "IMAGE" and descriptor.get("encoding") == "image-transfer-v1":
+        # Lazy and IMAGE-only: exact tensors and other kinds stay unchanged.
+        from .image_codec import decode_image
+        return decode_image(descriptor, payload, max_bytes=limit)
+    if kind == "STRING":
+        if set(descriptor) != {"encoding"} or descriptor["encoding"] != "utf8-v1":
+            raise ValueError("Invalid STRING descriptor encoding or fields")
+        try:
+            return payload.decode("utf-8")
+        except UnicodeDecodeError:
+            raise ValueError("STRING payload must be valid UTF-8 bytes") from None
+
+    fields = {"encoding", "dtype", "shape", "layout"}
+    if kind == "AUDIO":
+        fields.add("sample_rate")
+    if set(descriptor) != fields:
+        raise ValueError("Invalid tensor descriptor fields")
+    if descriptor["encoding"] != "tensor-le-v1":
+        raise ValueError("Unsupported tensor descriptor encoding")
+    if descriptor["layout"] != _LAYOUTS[kind]:
+        raise ValueError("Invalid tensor descriptor layout")
+    dtype_name = descriptor["dtype"]
+    if type(dtype_name) is not str or dtype_name not in _DTYPES:
+        raise ValueError("Unsupported tensor descriptor dtype")
+    dtype, itemsize = _DTYPES[dtype_name]
+    size = _tensor_size(kind, descriptor["shape"], itemsize, limit)
+    if len(payload) != size:
+        raise ValueError("Tensor payload length does not match shape and dtype")
+    sample_rate = _sample_rate(descriptor["sample_rate"]) if kind == "AUDIO" else None
+
+    # frombuffer is always CPU. Copy immutable wire bytes into a writable private
+    # buffer, then clone so the returned tensor owns its storage independently.
+    buffer = bytearray(_native_little_endian(payload, itemsize))
+    storage_dtype = torch.uint16 if dtype == torch.bfloat16 else dtype
+    tensor = torch.frombuffer(buffer, dtype=storage_dtype).view(dtype)
+    if not torch.isfinite(tensor).all().item():
+        raise ValueError("Tensor payload values must be finite")
+    tensor = tensor.clone().reshape(descriptor["shape"])
+    if kind == "AUDIO":
+        return {"waveform": tensor, "sample_rate": sample_rate}
+    return tensor

serverless_bridge/contract.py

@@ -0,0 +1,249 @@
+"""Versioned, bounded media envelopes shared by master and endpoint bridge.
+
+No provider or ComfyUI bootstrap imports. Network I/O is supplied by the caller,
+which must enforce public-HTTPS resolution, redirect and streaming byte limits.
+"""
+import base64
+import binascii
+import hashlib
+import json
+import re
+from datetime import datetime, timezone
+from urllib.parse import urlsplit
+
+PROFILE = "comfy-media-v1"
+VERSION = 1
+MAX_ARTIFACT_BYTES = 256 * 1024 * 1024
+MAX_TOTAL_BYTES = 512 * 1024 * 1024
+MAX_ARTIFACTS = 128
+MAX_ITEMS = 64
+MAX_DESCRIPTOR_BYTES = 128 * 1024
+SOCKET_KINDS = {"images": "IMAGE", "audio": "AUDIO", "masks": "MASK", "texts": "STRING", "video": "VIDEO"}
+BINDING_FIELDS = {"binding_id", "collector_id", "job_id", "participant_id", "node_id", "socket", "kind"}
+_ID = re.compile(r"[A-Za-z0-9_.:/-]{1,200}\Z")
+_HASH = re.compile(r"[a-f0-9]{64}\Z")
+
+
+def identifier(value):
+    if not isinstance(value, str) or not _ID.fullmatch(value):
+        raise ValueError("Invalid media identifier")
+    return value
+
+
+def bounded_int(value, maximum, *, minimum=0):
+    if type(value) is not int or not minimum <= value <= maximum:
+        raise ValueError("Invalid media count or size")
+    return value
+
+
+def validate_json(value, *, max_bytes=MAX_DESCRIPTOR_BYTES):
+    """Bound metadata depth and size without accepting non-finite JSON numbers."""
+    def walk(item, depth=0):
+        if depth > 12:
+            raise ValueError("Media metadata nesting exceeds limit")
+        if isinstance(item, dict):
+            if len(item) > 4096 or any(not isinstance(k, str) for k in item):
+                raise ValueError("Invalid media metadata object")
+            for child in item.values():
+                walk(child, depth + 1)
+        elif isinstance(item, list):
+            if len(item) > 4096:
+                raise ValueError("Media metadata list exceeds limit")
+            for child in item:
+                walk(child, depth + 1)
+        elif item is not None and type(item) not in (str, int, float, bool):
+            raise ValueError("Invalid media metadata value")
+    walk(value)
+    try:
+        raw = json.dumps(value, allow_nan=False, separators=(",", ":")).encode("utf-8")
+    except (ValueError, TypeError, UnicodeError, RecursionError):
+        raise ValueError("Invalid media JSON") from None
+    if len(raw) > max_bytes:
+        raise ValueError("Media JSON exceeds byte limit")
+    return raw
+
+
+def validate_binding(binding):
+    if not isinstance(binding, dict) or set(binding) != BINDING_FIELDS:
+        raise ValueError("Invalid media binding")
+    for field in BINDING_FIELDS - {"socket", "kind"}:
+        identifier(binding[field])
+    if SOCKET_KINDS.get(binding["socket"]) != binding["kind"]:
+        raise ValueError("Media socket/type mismatch")
+    return binding
+
+
+def validate_bindings(bindings):
+    if not isinstance(bindings, list) or not 1 <= len(bindings) <= MAX_ARTIFACTS:
+        raise ValueError("Invalid media bindings count")
+    result = {}
+    destinations = set()
+    for binding in bindings:
+        validate_binding(binding)
+        bid = binding["binding_id"]
+        destination = (binding["job_id"], binding["participant_id"], binding["collector_id"], binding["socket"])
+        if bid in result or destination in destinations:
+            raise ValueError("Duplicate media binding")
+        result[bid] = binding
+        destinations.add(destination)
+    return result
+
+
+def _validate_url_transport(transport):
+    if set(transport) != {"type", "url", "expires_at"}:
+        raise ValueError("Invalid media URL transport")
+    url = transport["url"]
+    if not isinstance(url, str) or len(url) > 8192:
+        raise ValueError("Invalid media URL")
+    try:
+        parsed = urlsplit(url)
+        if parsed.scheme != "https" or not parsed.hostname or parsed.username or parsed.password or parsed.fragment:
+            raise ValueError()
+        if parsed.port not in (None, 443):
+            raise ValueError()
+        expiry = datetime.fromisoformat(transport["expires_at"].replace("Z", "+00:00"))
+        if expiry.tzinfo is None or expiry <= datetime.now(timezone.utc):
+            raise ValueError()
+    except (ValueError, TypeError, AttributeError):
+        raise ValueError("Invalid or expired media URL") from None
+
+
+def validate_artifact(artifact, *, max_bytes=MAX_ARTIFACT_BYTES):
+    fields = {"id", "kind", "descriptor", "byte_length", "sha256", "transport"}
+    if not isinstance(artifact, dict) or set(artifact) != fields:
+        raise ValueError("Invalid media artifact")
+    identifier(artifact["id"])
+    if artifact["kind"] not in SOCKET_KINDS.values():
+        raise ValueError("Unsupported media kind")
+    if not isinstance(artifact["descriptor"], dict):
+        raise ValueError("Invalid media descriptor")
+    validate_json(artifact["descriptor"])
+    size = bounded_int(artifact["byte_length"], max_bytes)
+    digest = artifact["sha256"]
+    if not isinstance(digest, str) or not _HASH.fullmatch(digest):
+        raise ValueError("Invalid media checksum")
+    transport = artifact["transport"]
+    if not isinstance(transport, dict):
+        raise ValueError("Invalid media transport")
+    if transport.get("type") == "inline":
+        if set(transport) != {"type", "data_base64"} or not isinstance(transport["data_base64"], str):
+            raise ValueError("Invalid inline media transport")
+        if len(transport["data_base64"]) != 4 * ((size + 2) // 3):
+            raise ValueError("Inline media length mismatch")
+    elif transport.get("type") == "url":
+        _validate_url_transport(transport)
+    else:
+        raise ValueError("Unsupported media transport")
+    return artifact
+
+
+def artifact_from_payload(artifact_id, kind, descriptor, payload, *, transport=None, max_bytes=MAX_ARTIFACT_BYTES):
+    if not isinstance(payload, bytes):
+        raise ValueError("Media payload must be bytes")
+    bounded_int(len(payload), max_bytes)
+    result = {"id": artifact_id, "kind": kind, "descriptor": descriptor,
+              "byte_length": len(payload), "sha256": hashlib.sha256(payload).hexdigest(),
+              "transport": transport if transport is not None else {
+                  "type": "inline", "data_base64": base64.b64encode(payload).decode("ascii")}}
+    return validate_artifact(result, max_bytes=max_bytes)
+
+
+async def resolve_artifact(artifact, *, download=None, max_bytes=MAX_ARTIFACT_BYTES):
+    """download(url, *, max_bytes) must return bounded bytes; never log its URL."""
+    validate_artifact(artifact, max_bytes=max_bytes)
+    transport = artifact["transport"]
+    if transport["type"] == "inline":
+        try:
+            payload = base64.b64decode(transport["data_base64"], validate=True)
+        except (ValueError, binascii.Error):
+            raise ValueError("Invalid media base64") from None
+    else:
+        if download is None:
+            raise ValueError("Media URL download is not configured")
+        payload = await download(transport["url"], max_bytes=artifact["byte_length"])
+    if not isinstance(payload, bytes) or len(payload) != artifact["byte_length"]:
+        raise ValueError("Media byte length mismatch")
+    if hashlib.sha256(payload).hexdigest() != artifact["sha256"]:
+        raise ValueError("Media checksum mismatch")
+    return payload
+
+
+def validate_response(response, *, request_id, bindings, max_total_bytes=MAX_TOTAL_BYTES):
+    """Validate ALL identity/completeness before downloading or delivering anything."""
+    expected = validate_bindings(bindings)
+    identifier(request_id)
+    if not isinstance(response, dict) or set(response) != {"version", "profile", "request_id", "bindings", "artifacts", "items"}:
+        raise ValueError("Invalid media response envelope")
+    if type(response["version"]) is not int or response["version"] != VERSION or response["profile"] != PROFILE:
+        raise ValueError("Unsupported media response version/profile")
+    if response["request_id"] != request_id:
+        raise ValueError("Media request identity mismatch")
+    returned = response["bindings"]
+    if not isinstance(returned, list) or len(returned) != len(expected):
+        raise ValueError("Missing or extra media bindings")
+    counts = {}
+    for row in returned:
+        if not isinstance(row, dict) or set(row) != BINDING_FIELDS | {"count"}:
+            raise ValueError("Invalid result binding")
+        binding = {key: row[key] for key in BINDING_FIELDS}
+        validate_binding(binding)
+        bid = binding["binding_id"]
+        if bid in counts or expected.get(bid) != binding:
+            raise ValueError("Media binding identity mismatch")
+        # Connected outputs must produce at least one value. Empty STRING is a value.
+        counts[bid] = bounded_int(row["count"], MAX_ITEMS, minimum=1)
+    artifacts = response["artifacts"]
+    if not isinstance(artifacts, list) or not 1 <= len(artifacts) <= MAX_ARTIFACTS:
+        raise ValueError("Invalid media artifacts count")
+    indexed = {}
+    total = 0
+    for artifact in artifacts:
+        validate_artifact(artifact)
+        aid = artifact["id"]
+        if aid in indexed:
+            raise ValueError("Duplicate media artifact")
+        total += artifact["byte_length"]
+        if total > max_total_bytes:
+            raise ValueError("Media response exceeds total byte limit")
+        indexed[aid] = artifact
+    items = response["items"]
+    if not isinstance(items, list) or len(items) != sum(counts.values()):
+        raise ValueError("Missing or extra media items")
+    ordered = {bid: [None] * count for bid, count in counts.items()}
+    used = set()
+    for item in items:
+        if not isinstance(item, dict) or set(item) != {"binding_id", "item_index", "artifact_id"}:
+            raise ValueError("Invalid media item")
+        bid, aid = identifier(item["binding_id"]), identifier(item["artifact_id"])
+        if bid not in counts or aid not in indexed:
+            raise ValueError("Unknown media item identity")
+        index = bounded_int(item["item_index"], counts[bid] - 1)
+        if ordered[bid][index] is not None or indexed[aid]["kind"] != expected[bid]["kind"]:
+            raise ValueError("Duplicate or mistyped media item")
+        ordered[bid][index] = indexed[aid]
+        used.add(aid)
+    if used != set(indexed) or any(value is None for values in ordered.values() for value in values):
+        raise ValueError("Incomplete media result manifest")
+    return ordered
+
+
+async def decode_response(response, *, request_id, bindings, download=None, max_total_bytes=MAX_TOTAL_BYTES):
+    ordered = validate_response(response, request_id=request_id, bindings=bindings, max_total_bytes=max_total_bytes)
+    # Resolve and decode before exposing a single result to the collector queue.
+    from .codec import decode_payload
+    cache = {}
+    decoded = {}
+    for bid, artifacts in ordered.items():
+        values = []
+        for artifact in artifacts:
+            aid = artifact["id"]
+            if aid not in cache:
+                payload = await resolve_artifact(artifact, download=download)
+                if artifact["kind"] == "VIDEO":
+                    from .video_codec import decode_video
+                    cache[aid] = decode_video(artifact["descriptor"], payload)
+                else:
+                    cache[aid] = decode_payload(artifact["kind"], artifact["descriptor"], payload)
+            values.append(cache[aid])
+        decoded[bid] = values
+    return decoded

serverless_bridge/handler.py

@@ -0,0 +1,794 @@
+"""Pure HTTP executor for media-v1 requests against a local ComfyUI instance.
+
+No ComfyUI, tensor, or provider imports. ``download(url, *, max_bytes)`` must
+provide public-HTTPS/redirect/DNS protection and bounded streaming. Optional
+``publish(metadata, payload)`` returns a contract-valid, expiring HTTPS transport.
+
+Only identified invocation-owned files are deleted. Optional trusted local
+``temp_dir`` enables exporter cleanup after terminal history and collection.
+Without it, exporter retention belongs to the ComfyUI temp-directory lifecycle.
+The pinned native /interrupt route is not atomic: running-prompt cancellation
+fails closed and retains files instead of risking an unrelated successor prompt.
+"""
+import asyncio
+import base64
+import binascii
+import copy
+import hashlib
+import json
+import math
+import os
+import re
+import stat
+import uuid
+from pathlib import Path
+from urllib.parse import quote, urlsplit
+
+import aiohttp
+
+from .contract import (
+    BINDING_FIELDS, MAX_ARTIFACT_BYTES, MAX_ARTIFACTS, MAX_TOTAL_BYTES,
+    PROFILE, VERSION, artifact_from_payload, identifier, resolve_artifact,
+    validate_artifact, validate_bindings, validate_json, validate_response,
+)
+
+MAX_RESPONSE_JSON_BYTES = 10_000_000
+MAX_HTTP_JSON_BYTES = 32 * 1024 * 1024
+POLL_INTERVAL_SECONDS = 0.25
+CANCEL_GRACE_SECONDS = 10.0
+DRAIN_GRACE_SECONDS = 10.0
+HTTP_TIMEOUT_SECONDS = 15.0
+_UNCERTAIN_WORK = set()
+GET_ATTEMPTS = 3
+_IMPORT_SLOTS = {"IMAGE": 0, "AUDIO": 1, "MASK": 2, "STRING": 3, "VIDEO": 4}
+# Used ONLY to run the shared manifest validator before /view. Never fetched.
+_VALIDATION_TRANSPORT = {"type": "url", "url": "https://media-validation.invalid/artifact",
+                         "expires_at": "9999-01-01T00:00:00Z"}
+
+
+class MediaExecutionUncertain(RuntimeError):
+    """The local prompt may still execute; staged files were deliberately retained."""
+
+
+class _HTTPError(RuntimeError):
+    def __init__(self, status):
+        super().__init__(f"Local Comfy HTTP request failed (status {status})")
+        self.status = status
+
+
+def _scope_id(value):
+    identifier(value)
+    if any(part in ("", ".", "..") for part in value.split("/")):
+        raise ValueError("Unsafe media scope identifier")
+    return value
+
+
+def _json_bytes(value):
+    return json.dumps(value, allow_nan=False).encode("utf-8")
+
+
+def _link(value, workflow):
+    if (not isinstance(value, list) or len(value) != 2 or
+            not isinstance(value[0], str) or value[0] not in workflow or
+            type(value[1]) is not int or value[1] < 0):
+        raise ValueError("Invalid workflow socket link")
+    return value[0]
+
+
+def _validate_request(input_data):
+    if (not isinstance(input_data, dict) or not {"workflow", "media"} <= set(input_data)
+            or set(input_data) - {"workflow", "media", "images"}):
+        raise ValueError("Invalid media request envelope")
+    media = input_data["media"]
+    if (not isinstance(media, dict) or set(media) !=
+            {"version", "profile", "request_id", "artifacts", "inputs", "outputs"}):
+        raise ValueError("Invalid media request controls")
+    if type(media["version"]) is not int or media["version"] != VERSION or media["profile"] != PROFILE:
+        raise ValueError("Unsupported media request version/profile")
+    _scope_id(media["request_id"])
+    bindings = validate_bindings(media["outputs"])
+    targets = []
+    for binding in bindings.values():
+        for field in BINDING_FIELDS - {"socket", "kind"}:
+            _scope_id(binding[field])
+        targets.append(binding["node_id"])
+    if len(set(targets)) != len(targets):
+        raise ValueError("Duplicate media exporter node")
+
+    workflow = input_data["workflow"]
+    if not isinstance(workflow, dict) or not workflow:
+        raise ValueError("Invalid API workflow")
+    validate_json(workflow, max_bytes=16 * 1024 * 1024)
+    dependencies = {}
+    for node_id, node in workflow.items():
+        _scope_id(node_id)
+        if (not isinstance(node, dict) or not isinstance(node.get("class_type"), str)
+                or not node["class_type"] or not isinstance(node.get("inputs"), dict)):
+            raise ValueError("Invalid API workflow node")
+        if node["class_type"] == "MediaImportV1":
+            raise ValueError("Request may not supply its own media import filenames")
+        if node["class_type"] == "MediaExportV1" and node_id not in targets:
+            raise ValueError("Undeclared media exporter")
+        dependencies[node_id] = set()
+        for name, value in node["inputs"].items():
+            _scope_id(name)
+            if isinstance(value, list):
+                dependencies[node_id].add(_link(value, workflow))
+    for binding in bindings.values():
+        node = workflow.get(binding["node_id"])
+        if node is None or node["class_type"] != "MediaExportV1":
+            raise ValueError("Missing declared MediaExportV1 node")
+        controls = node["inputs"]
+        from .image_transfer import DEFAULT_FORMAT, DEFAULT_QUALITY, validate_options
+        required = {"request_id", "binding_json", binding["socket"]}
+        optional = {"transfer_format", "transfer_quality"} if binding["kind"] == "IMAGE" else set()
+        if not required <= set(controls) or set(controls) - required - optional:
+            raise ValueError("Media exporter must have exactly its declared socket and controls")
+        validate_options(controls.get("transfer_format", DEFAULT_FORMAT),
+                         controls.get("transfer_quality", DEFAULT_QUALITY))
+        if controls["request_id"] != media["request_id"]:
+            raise ValueError("Media exporter request identity mismatch")
+        text = controls["binding_json"]
+        if not isinstance(text, str) or len(text) > 128 * 1024:
+            raise ValueError("Invalid media exporter binding JSON")
+        try:
+            declared = json.loads(text)
+        except (ValueError, RecursionError):
+            raise ValueError("Invalid media exporter binding JSON") from None
+        if declared != binding:
+            raise ValueError("Media exporter binding identity mismatch")
+        _link(controls[binding["socket"]], workflow)
+    # The execution roots are exactly the declared exporters, not every class
+    # with OUTPUT_NODE=True. Required upstream output utilities may remain.
+    visited = set()
+    pending = list(targets)
+    while pending:
+        node_id = pending.pop()
+        if node_id not in visited:
+            visited.add(node_id)
+            pending.extend(dependencies[node_id])
+    if visited != set(workflow):
+        raise ValueError("Workflow contains nodes outside declared output roots")
+    if any(set(targets) & sources for sources in dependencies.values()):
+        raise ValueError("A media exporter cannot be an upstream dependency")
+    # Kahn traversal bounds cycle checking without relying on recursion depth.
+    counts = {node: len(sources) for node, sources in dependencies.items()}
+    consumers = {node: [] for node in workflow}
+    for node, sources in dependencies.items():
+        for source in sources:
+            consumers[source].append(node)
+    pending = [node for node, count in counts.items() if not count]
+    processed = 0
+    while pending:
+        processed += 1
+        for consumer in consumers[pending.pop()]:
+            counts[consumer] -= 1
+            if not counts[consumer]:
+                pending.append(consumer)
+    if processed != len(workflow):
+        raise ValueError("Workflow contains a dependency cycle")
+
+    artifacts = media["artifacts"]
+    if not isinstance(artifacts, list) or len(artifacts) > MAX_ARTIFACTS:
+        raise ValueError("Invalid input artifacts count")
+    indexed, total = {}, 0
+    for artifact in artifacts:
+        validate_artifact(artifact)
+        aid = _scope_id(artifact["id"])
+        if aid in indexed:
+            raise ValueError("Duplicate input artifact")
+        indexed[aid] = artifact
+        total += artifact["byte_length"]
+    inputs = media["inputs"]
+    if not isinstance(inputs, list) or len(inputs) > 4096:
+        raise ValueError("Invalid input references")
+    destinations, used = set(), set()
+    exporters = {binding["node_id"]: binding for binding in bindings.values()}
+    for ref in inputs:
+        if not isinstance(ref, dict) or set(ref) != {"node_id", "input_name", "artifact_id"}:
+            raise ValueError("Invalid media input reference")
+        for value in ref.values():
+            _scope_id(value)
+        node_id, name, aid = ref["node_id"], ref["input_name"], ref["artifact_id"]
+        if (node_id not in workflow or name not in workflow[node_id]["inputs"]
+                or aid not in indexed or (node_id, name) in destinations):
+            raise ValueError("Unknown or duplicate media input reference")
+        binding = exporters.get(node_id)
+        if binding is not None and (name != binding["socket"] or indexed[aid]["kind"] != binding["kind"]):
+            raise ValueError("Exporter input must match its declared media socket and kind")
+        destinations.add((node_id, name))
+        used.add(aid)
+    if used != set(indexed):
+        raise ValueError("Every input artifact must be used")
+
+    images = input_data.get("images", [])
+    if not isinstance(images, list) or len(images) > MAX_ARTIFACTS:
+        raise ValueError("Invalid legacy images count")
+    names = set()
+    legacy = []
+    for image in images:
+        if not isinstance(image, dict) or set(image) != {"name", "image"}:
+            raise ValueError("Invalid legacy image")
+        name, encoded = image["name"], image["image"]
+        if (not isinstance(name, str) or not re.fullmatch(r"[A-Za-z0-9][A-Za-z0-9_. -]{0,254}", name)
+                or name in names or not isinstance(encoded, str)):
+            raise ValueError("Invalid or duplicate legacy image basename")
+        names.add(name)
+        if len(encoded) > 4 * ((MAX_ARTIFACT_BYTES + 2) // 3):
+            raise ValueError("Legacy image exceeds byte limit")
+        # Exact decoded length, including padding, before allocating bytes.
+        size = (len(encoded) // 4) * 3 - (len(encoded) - len(encoded.rstrip("=")))
+        if len(encoded) % 4 or size < 0 or size > MAX_ARTIFACT_BYTES:
+            raise ValueError("Invalid legacy image base64 length")
+        total += size
+        matches = [node_id for node_id, node in workflow.items()
+                   if node["class_type"] == "LoadImage" and node["inputs"].get("image") == name]
+        if not matches or any((node_id, "image") in destinations for node_id in matches):
+            raise ValueError("Unused or conflicting legacy image binding")
+        legacy.append((encoded, size, matches))
+    if total > MAX_TOTAL_BYTES:
+        raise ValueError("Media inputs exceed total byte limit")
+    return copy.deepcopy(workflow), copy.deepcopy(media), legacy, targets
+
+
+class _Staging:
+    """Exclusive random files under a hashed namespace; never delete siblings."""
+    def __init__(self, input_dir, request_id):
+        self.root = Path(input_dir).resolve()
+        self.key = hashlib.sha256(request_id.encode()).hexdigest()
+        self.relative = f"distributed-media-v1/{self.key}"
+        self.names = []
+        self.fd = None
+
+    def _open(self):
+        self.root.mkdir(parents=True, exist_ok=True)
+        directory = self.root
+        for component in ("distributed-media-v1", self.key):
+            directory = directory / component
+            if directory.is_symlink():
+                raise ValueError("Unsafe media staging directory symlink")
+            directory.mkdir(exist_ok=True, mode=0o700)
+            if not directory.resolve().is_relative_to(self.root):
+                raise ValueError("Unsafe media staging directory")
+        self.fd = os.open(directory, os.O_RDONLY | os.O_DIRECTORY | os.O_NOFOLLOW)
+
+    def write(self, payload):
+        if self.fd is None:
+            self._open()
+        name = uuid.uuid4().hex + ".bin"
+        fd = os.open(name, os.O_WRONLY | os.O_CREAT | os.O_EXCL | os.O_NOFOLLOW, 0o600, dir_fd=self.fd)
+        self.names.append(name)
+        with os.fdopen(fd, "wb") as stream:
+            stream.write(payload)
+        return f"{self.relative}/{name}"
+
+    def close(self, *, cleanup):
+        if self.fd is not None:
+            try:
+                if cleanup:
+                    for name in self.names:
+                        try:
+                            os.unlink(name, dir_fd=self.fd)
+                        except FileNotFoundError:
+                            pass
+            finally:
+                os.close(self.fd)
+                self.fd = None
+            if cleanup:
+                # Never recursively remove a shared namespace or another run.
+                for directory in (self.root / self.relative, self.root / "distributed-media-v1"):
+                    try:
+                        directory.rmdir()
+                    except OSError:
+                        pass
+
+
+class _OwnedExports:
+    """Best-effort retirement of validated terminal-history files, never a sweep.
+
+    Walk every directory component with openat/O_NOFOLLOW, then pin the directory
+    fd and each regular file's identity. Missing, linked, replaced, or unidentifiable
+    files stay with the endpoint lifecycle. Never use history-supplied path strings.
+    """
+    def __init__(self, temp_dir, request_id):
+        self.root = Path(temp_dir).absolute() if temp_dir is not None else None
+        self.key = hashlib.sha256(request_id.encode()).hexdigest()
+        self.fd = None
+        self.files = {}
+
+    def _open(self):
+        if self.fd is not None:
+            return True
+        if self.root is None or ".." in self.root.parts:
+            return False
+        flags = os.O_RDONLY | os.O_DIRECTORY | os.O_NOFOLLOW
+        fd = None
+        try:
+            fd = os.open(self.root.anchor, flags)
+            for component in (*self.root.parts[1:], "distributed-media-v1", self.key):
+                child = os.open(component, flags, dir_fd=fd)
+                os.close(fd)
+                fd = child
+            self.fd, fd = fd, None
+            return True
+        except OSError:
+            return False
+        finally:
+            if fd is not None:
+                os.close(fd)
+
+    def _identity(self, name):
+        info = os.stat(name, dir_fd=self.fd, follow_symlinks=False)
+        if not stat.S_ISREG(info.st_mode) or info.st_nlink != 1:
+            return None
+        return (info.st_dev, info.st_ino, info.st_size, info.st_mtime_ns, info.st_ctime_ns)
+
+    def identify(self, entry, media):
+        if self.root is None or not isinstance(entry.get("outputs"), dict):
+            return
+        # A failed prompt may contain valid completed exporters alongside missing
+        # or malformed ones. Claim only independently validated declared records.
+        for binding in media["outputs"]:
+            node_id = binding["node_id"]
+            try:
+                _, files = _manifest({"outputs": {node_id: entry["outputs"].get(node_id)}},
+                                     dict(media, outputs=[binding]))
+            except ValueError:
+                continue
+            if not files or not self._open():
+                continue
+            for artifact in files:
+                name = artifact["id"] + ".bin"  # _manifest enforces UUID hex.
+                try:
+                    identity = self._identity(name)
+                except OSError:
+                    continue
+                if identity is not None:
+                    self.files.setdefault(name, identity)
+
+    def close(self, *, cleanup):
+        if self.fd is None:
+            return
+        try:
+            if cleanup:
+                for name, identity in self.files.items():
+                    try:
+                        if self._identity(name) == identity:
+                            os.unlink(name, dir_fd=self.fd)
+                    except OSError:
+                        pass  # Endpoint lifecycle owns inaccessible leftovers.
+        finally:
+            os.close(self.fd)
+            self.fd = None
+        # Deliberately retain directories: they may belong to concurrent runs.
+
+
+async def _body(response, maximum):
+    length = response.content_length
+    if length is not None and length > maximum:
+        raise ValueError("Local Comfy response exceeds declared byte length")
+    body = bytearray()
+    async for chunk in response.content.iter_chunked(64 * 1024):
+        if len(body) + len(chunk) > maximum:
+            raise ValueError("Local Comfy response exceeds declared byte length")
+        body.extend(chunk)
+    return bytes(body)
+
+
+class _Execution:
+    def __init__(self, session, url, stage, media, targets, deadline, exports):
+        self.session, self.url, self.stage = session, url, stage
+        self.media, self.targets = media, targets
+        self.deadline = deadline
+        self.run_id = uuid.uuid4().hex
+        self.prompt_id = None
+        self.attempted = False
+        self.terminal = False
+        self.exports = exports
+        self.interruption_blocked = False
+
+    def remaining(self):
+        remaining = self.deadline - asyncio.get_running_loop().time()
+        if remaining <= 0:
+            raise asyncio.TimeoutError("Media execution deadline expired")
+        return remaining
+
+    async def json_request(self, method, path, *, body=None, permit_rejection=False):
+        async with self.session.request(method, self.url + path, json=body,
+                timeout=aiohttp.ClientTimeout(total=min(HTTP_TIMEOUT_SECONDS, self.remaining())),
+                allow_redirects=False) as response:
+            raw = await _body(response, MAX_HTTP_JSON_BYTES)
+            # Native queue-delete and interrupt routes return an empty 200.
+            if response.status == 200 and not raw and method == "POST" and path in {"/queue", "/interrupt"}:
+                return {}
+            try:
+                value = json.loads(raw)
+            except (ValueError, UnicodeError, RecursionError):
+                if response.status != 200:
+                    raise _HTTPError(response.status) from None
+                raise ValueError("Invalid local Comfy JSON response") from None
+            if response.status != 200:
+                # Native /prompt rejects the entire graph with a 400 error.
+                # Other failures may hide acceptance and are never retried.
+                if (permit_rejection and response.status == 400 and isinstance(value, dict)
+                        and value.get("error") and "prompt_id" not in value):
+                    self.terminal = True
+                raise _HTTPError(response.status)
+            if not isinstance(value, dict):
+                raise ValueError("Invalid local Comfy JSON object")
+            return value
+
+    async def get_json(self, path):
+        for attempt in range(GET_ATTEMPTS):
+            try:
+                return await self.json_request("GET", path)
+            except (_HTTPError, aiohttp.ClientError, asyncio.TimeoutError) as exc:
+                if isinstance(exc, _HTTPError) and exc.status != 429 and exc.status < 500:
+                    raise
+                if attempt == GET_ATTEMPTS - 1:
+                    raise
+                await asyncio.sleep(POLL_INTERVAL_SECONDS * (2 ** attempt))
+
+    async def history(self):
+        response = await self.get_json("/history/" + quote(self.prompt_id, safe=""))
+        entry = response.get(self.prompt_id)
+        if entry is None:
+            return None
+        if not isinstance(entry, dict) or not isinstance(entry.get("status"), dict):
+            raise ValueError("Malformed Comfy prompt history")
+        status = entry["status"]
+        messages = status.get("messages", [])
+        if not isinstance(messages, list):
+            raise ValueError("Malformed Comfy history messages")
+        failed = status.get("status_str") in {"error", "interrupted"} or any(
+            isinstance(message, list) and message and message[0] in {"execution_error", "execution_interrupted"}
+            for message in messages)
+        if failed or status.get("completed") is True:
+            self.terminal = True
+            self.exports.identify(entry, self.media)
+            if failed or status.get("status_str") != "success":
+                raise RuntimeError("Local Comfy prompt execution failed")
+            return entry
+        return None
+
+    async def submit_and_wait(self, workflow):
+        body = {"prompt": workflow, "client_id": self.run_id,
+                "extra_data": {"distributed_media_run_id": self.run_id},
+                "partial_execution_targets": list(self.targets)}
+        self.remaining()  # CPU staging must not spend past the absolute deadline.
+        self.attempted = True
+        response = await self.json_request("POST", "/prompt", body=body, permit_rejection=True)
+        self.prompt_id = _scope_id(response.get("prompt_id"))
+        if response.get("node_errors"):
+            raise RuntimeError("Local Comfy accepted an incomplete set of execution targets")
+        while True:
+            entry = await self.history()
+            if entry is not None:
+                return entry
+            await asyncio.sleep(POLL_INTERVAL_SECONDS)
+
+    async def settle(self):
+        """Bounded recovery; caller shields this from task cancellation."""
+        deleted = False
+        while True:
+            if self.prompt_id:
+                try:
+                    await self.history()
+                except RuntimeError:
+                    if not self.terminal:
+                        raise
+                if self.terminal:
+                    return
+            queue = await self.get_json("/queue")
+            groups = []
+            for key in ("queue_running", "queue_pending"):
+                rows = queue.get(key)
+                if not isinstance(rows, list):
+                    raise ValueError("Malformed Comfy queue")
+                group = {}
+                for row in rows:
+                    if not isinstance(row, list) or len(row) < 4 or not isinstance(row[3], dict):
+                        raise ValueError("Malformed Comfy queue entry")
+                    pid = _scope_id(row[1])
+                    if pid in group:
+                        raise ValueError("Duplicate Comfy queue identity")
+                    group[pid] = row[3]
+                groups.append(group)
+            running, pending = groups
+            if self.prompt_id is None:
+                own = [pid for group in groups for pid, extra in group.items()
+                       if extra.get("distributed_media_run_id") == self.run_id
+                       and extra.get("client_id") == self.run_id]
+                if len(own) > 1:
+                    raise ValueError("Ambiguous Comfy queue ownership")
+                if own:
+                    self.prompt_id = own[0]
+            if self.prompt_id in pending and not deleted:
+                await self.json_request("POST", "/queue", body={"delete": [self.prompt_id]})
+                deleted = True
+            elif self.prompt_id in running:
+                # Pinned ComfyUI server.py checks a snapshot then globally
+                # interrupts, outside the queue mutex. Supplying prompt_id does
+                # NOT close that race. No safe capability negotiation exists on
+                # this route; do not invoke it, even after our own queue read.
+                self.interruption_blocked = True
+                raise RuntimeError("Native prompt-scoped interruption unavailable")
+            elif deleted and self.prompt_id not in running and self.prompt_id not in pending:
+                self.terminal = True  # queue deletion confirmed by exact read-back
+                return
+            await asyncio.sleep(POLL_INTERVAL_SECONDS)
+
+    async def fetch(self, artifact):
+        transport = artifact["transport"]
+        params = {"filename": transport["filename"], "subfolder": transport["subfolder"], "type": "temp"}
+        for attempt in range(GET_ATTEMPTS):
+            try:
+                async with self.session.get(self.url + "/view", params=params, allow_redirects=False,
+                        timeout=aiohttp.ClientTimeout(total=min(HTTP_TIMEOUT_SECONDS, self.remaining()))) as response:
+                    if response.status != 200:
+                        raise _HTTPError(response.status)
+                    payload = await _body(response, artifact["byte_length"])
+                break
+            except (_HTTPError, aiohttp.ClientError, asyncio.TimeoutError) as exc:
+                if isinstance(exc, _HTTPError) and exc.status != 429 and exc.status < 500:
+                    raise
+                if attempt == GET_ATTEMPTS - 1:
+                    raise
+                await asyncio.sleep(POLL_INTERVAL_SECONDS * (2 ** attempt))
+        if len(payload) != artifact["byte_length"]:
+            raise ValueError("Media output byte length mismatch")
+        if hashlib.sha256(payload).hexdigest() != artifact["sha256"]:
+            raise ValueError("Media output checksum mismatch")
+        return payload
+
+
+def _manifest(entry, media):
+    outputs = entry.get("outputs")
+    if not isinstance(outputs, dict):
+        raise ValueError("Missing media history outputs")
+    expected_ids = {binding["node_id"] for binding in media["outputs"]}
+    if any(node_id not in expected_ids and isinstance(value, dict) and "media_exports" in value
+           for node_id, value in outputs.items()):
+        raise ValueError("Unexpected media export history")
+    response = {"version": VERSION, "profile": PROFILE, "request_id": media["request_id"],
+                "bindings": [], "artifacts": [], "items": []}
+    files = []
+    subfolder = "distributed-media-v1/" + hashlib.sha256(media["request_id"].encode()).hexdigest()
+    for binding in media["outputs"]:
+        node = outputs.get(binding["node_id"], {})
+        records = node.get("media_exports") if isinstance(node, dict) else None
+        if not isinstance(records, list) or len(records) != 1:
+            raise ValueError("Missing or duplicate media export result")
+        record = records[0]
+        if (not isinstance(record, dict) or set(record) != {"request_id", "binding", "artifacts", "items"}
+                or record["request_id"] != media["request_id"]):
+            raise ValueError("Media export request identity mismatch")
+        if not isinstance(record["artifacts"], list) or len(record["artifacts"]) > MAX_ARTIFACTS:
+            raise ValueError("Invalid exported artifacts count")
+        converted = []
+        for artifact in record["artifacts"]:
+            if not isinstance(artifact, dict) or not isinstance(artifact.get("transport"), dict):
+                raise ValueError("Invalid exported artifact")
+            transport = artifact["transport"]
+            aid = artifact.get("id")
+            if (not isinstance(aid, str) or not re.fullmatch(r"[a-f0-9]{32}", aid)
+                    or transport != {"type": "comfy-file", "filename": aid + ".bin",
+                                     "subfolder": subfolder, "folder_type": "temp"}):
+                raise ValueError("Unsafe or unowned exported media file")
+            converted.append(dict(artifact, transport=dict(_VALIDATION_TRANSPORT)))
+        partial = dict(response, bindings=[record["binding"]], artifacts=converted, items=record["items"])
+        validate_response(partial, request_id=media["request_id"], bindings=[binding])
+        response["bindings"].append(copy.deepcopy(record["binding"]))
+        response["artifacts"].extend(converted)
+        response["items"].extend(copy.deepcopy(record["items"]))
+        files.extend(copy.deepcopy(record["artifacts"]))
+    # Validate the whole manifest BEFORE the first /view or publisher callback.
+    validate_response(response, request_id=media["request_id"], bindings=media["outputs"])
+    return response, files
+
+
+def _publication_plan(response, files, publish):
+    skeleton = dict(response, artifacts=[dict(a, transport={"type": "inline", "data_base64": ""}) for a in files])
+    base_size = len(_json_bytes({"media": skeleton}))
+    sizes = {a["id"]: 4 * ((a["byte_length"] + 2) // 3) for a in files}
+    predicted = base_size + sum(sizes.values())
+    published = set()
+    if predicted > MAX_RESPONSE_JSON_BYTES and publish is None:
+        raise ValueError("Media response exceeds inline JSON limit; a publisher is required")
+    # Reserve enough space for a maximum-size escaped URL, including metadata.
+    # Small remaining inline values preserve their exact bytes and ordering.
+    for aid in sorted(sizes, key=sizes.get, reverse=True):
+        if predicted <= MAX_RESPONSE_JSON_BYTES:
+            break
+        if sizes[aid] > 65_536:
+            published.add(aid)
+            predicted -= sizes[aid] - 65_536
+    if predicted > MAX_RESPONSE_JSON_BYTES:
+        raise ValueError("Media response metadata exceeds JSON limit")
+    return published
+
+
+async def _run(execution, workflow, legacy, download, publish):
+    media = execution.media
+    imports = {}
+    for artifact in media["artifacts"]:
+        execution.remaining()
+        payload = await resolve_artifact(artifact, download=download)
+        execution.remaining()
+        filename = execution.stage.write(payload)
+        node_id = "distributed-media-import:" + uuid.uuid4().hex
+        while node_id in workflow:
+            node_id = "distributed-media-import:" + uuid.uuid4().hex
+        metadata = {key: value for key, value in artifact.items() if key != "transport"}
+        workflow[node_id] = {"class_type": "MediaImportV1", "inputs": {
+            "artifact_json": json.dumps(metadata, allow_nan=False), "filename": filename}}
+        imports[artifact["id"]] = [node_id, _IMPORT_SLOTS[artifact["kind"]]]
+    for ref in media["inputs"]:
+        workflow[ref["node_id"]]["inputs"][ref["input_name"]] = list(imports[ref["artifact_id"]])
+    for encoded, size, matches in legacy:
+        execution.remaining()
+        try:
+            payload = base64.b64decode(encoded, validate=True)
+        except (ValueError, binascii.Error):
+            raise ValueError("Invalid legacy image base64") from None
+        if len(payload) != size:
+            raise ValueError("Legacy image byte length mismatch")
+        execution.remaining()
+        filename = execution.stage.write(payload)
+        for node_id in matches:
+            workflow[node_id]["inputs"]["image"] = filename
+    # Replacements can disconnect file loaders whose original endpoint-local
+    # paths are unavailable. Keep exactly the post-rewrite exporter closure;
+    # shared ancestors still needed by another branch must survive.
+    reachable, pending = set(), list(execution.targets)
+    while pending:
+        node_id = pending.pop()
+        if node_id in reachable:
+            continue
+        reachable.add(node_id)
+        pending.extend(_link(value, workflow) for value in workflow[node_id]["inputs"].values()
+                       if isinstance(value, list))
+    workflow = {node_id: node for node_id, node in workflow.items() if node_id in reachable}
+    entry = await execution.submit_and_wait(workflow)
+    response, files = _manifest(entry, media)
+    published = _publication_plan(response, files, publish)
+    artifacts = []
+    for artifact in files:
+        payload = await execution.fetch(artifact)
+        transport = None
+        if artifact["id"] in published:
+            metadata = {key: copy.deepcopy(value) for key, value in artifact.items() if key != "transport"}
+            transport = await publish(metadata, payload)
+            if not isinstance(transport, dict) or transport.get("type") != "url":
+                raise ValueError("Publisher must return an expiring HTTPS URL transport")
+        artifacts.append(artifact_from_payload(artifact["id"], artifact["kind"], artifact["descriptor"],
+                                               payload, transport=transport))
+    response["artifacts"] = artifacts
+    validate_response(response, request_id=media["request_id"], bindings=media["outputs"])
+    result = {"media": response}
+    if len(_json_bytes(result)) > MAX_RESPONSE_JSON_BYTES:
+        raise ValueError("Media response exceeds JSON byte limit")
+    execution.remaining()
+    return result
+
+
+def _release_uncertain_work(task):
+    _UNCERTAIN_WORK.discard(task)
+    try:
+        task.result()
+    except BaseException:
+        pass
+
+
+async def _drain_work(work):
+    """Cancel once; a misbehaving callback cannot make teardown unbounded."""
+    if not work.done():
+        work.cancel()
+    deadline = asyncio.get_running_loop().time() + DRAIN_GRACE_SECONDS
+    while not work.done():
+        remaining = deadline - asyncio.get_running_loop().time()
+        if remaining <= 0:
+            # Keep a strong reference until it exits, but never delete its files
+            # asynchronously: uncertain ownership belongs to endpoint recovery.
+            _UNCERTAIN_WORK.add(work)
+            work.add_done_callback(_release_uncertain_work)
+            return False
+        try:
+            # Unlike wait_for, wait's timeout does not wait for cancellation.
+            await asyncio.wait({work}, timeout=remaining)
+        except asyncio.CancelledError:
+            continue
+    _release_uncertain_work(work)
+    return True
+
+
+async def execute_media_request(input_data, *, comfy_url, input_dir, temp_dir=None,
+                                timeout_seconds=1800, session=None, download=None, publish=None) -> dict:
+    """Validate, stage, submit once, collect exact media, and retire owned inputs.
+
+    ``timeout_seconds`` covers staging, HTTP, polling, and publishing. A separate
+    bounded cancellation grace is allowed ONLY to establish the local prompt is
+    terminal. On uncertain acceptance/termination, raise MediaExecutionUncertain
+    and retain input files rather than racing an executing prompt. Never retry an
+    ambiguous submission, even when the exception is a timeout.
+    Async callback teardown has its own bounded grace; non-cooperative callbacks
+    also raise MediaExecutionUncertain and retain files, even after terminal history.
+
+    ``temp_dir`` is a trusted endpoint setting for this ComfyUI's local temp root,
+    never a request-supplied path. Only validated exporter UUID.bin files beneath
+    distributed-media-v1/sha256(request_id) are retired, after all collection and
+    publisher work has unwound. Unidentified files remain lifecycle-owned. With
+    ``temp_dir=None`` all exporter files remain lifecycle-owned as before.
+
+    The pinned native /interrupt route has a check/global-interrupt race. Running
+    prompts therefore fail closed with MediaExecutionUncertain, without posting
+    /interrupt. A prompt-scoped request body alone is not a safety guarantee.
+    """
+    if (type(timeout_seconds) not in (int, float) or not math.isfinite(timeout_seconds)
+            or timeout_seconds <= 0):
+        raise ValueError("Invalid media execution timeout")
+    if not isinstance(comfy_url, str):
+        raise ValueError("Invalid local Comfy URL")
+    try:
+        parsed = urlsplit(comfy_url)
+        if (parsed.scheme not in {"http", "https"} or not parsed.hostname or parsed.username
+                or parsed.password or parsed.query or parsed.fragment):
+            raise ValueError()
+        parsed.port
+    except ValueError:
+        raise ValueError("Invalid local Comfy URL") from None
+    deadline = asyncio.get_running_loop().time() + timeout_seconds
+    workflow, media, legacy, targets = _validate_request(input_data)
+    stage = _Staging(input_dir, media["request_id"])
+    exports = _OwnedExports(temp_dir, media["request_id"])
+    owned_session = session is None
+    if owned_session:
+        session = aiohttp.ClientSession()
+    execution = _Execution(session, comfy_url.rstrip("/"), stage, media, targets, deadline, exports)
+    work = None
+    try:
+        remaining = execution.remaining()
+        work = asyncio.create_task(_run(execution, workflow, legacy, download, publish))
+        return await asyncio.wait_for(asyncio.shield(work), remaining)
+    except BaseException as exc:
+        # Retire files only after readers unwind. Cooperative async callbacks
+        # normally drain immediately; an uncooperative one requires quarantine.
+        if work is not None and not await _drain_work(work):
+            raise MediaExecutionUncertain(
+                "Media callback teardown unconfirmed; owned files retained; "
+                "quarantine this endpoint before reuse; do not resubmit automatically"
+            ) from exc
+        if execution.attempted and not execution.terminal:
+            execution.deadline = asyncio.get_running_loop().time() + CANCEL_GRACE_SECONDS
+            recovery = asyncio.create_task(asyncio.wait_for(execution.settle(), CANCEL_GRACE_SECONDS))
+            # Shield the bounded recovery even from repeated task.cancel() calls.
+            while not recovery.done():
+                try:
+                    await asyncio.shield(recovery)
+                except asyncio.CancelledError:
+                    continue
+                except Exception:
+                    break
+            try:
+                recovery.result()
+            except BaseException:
+                pass
+            if not execution.terminal:
+                raise MediaExecutionUncertain(
+                    "Local Comfy prompt terminal state unconfirmed; owned staged files retained "
+                    f"under {stage.relative}; prompt_id={execution.prompt_id or 'unknown'}; "
+                    "do not resubmit this request automatically"
+                    + ("; native prompt-scoped interruption unavailable (non-atomic /interrupt route)"
+                       if execution.interruption_blocked else "")
+                ) from exc
+        raise
+    finally:
+        try:
+            try:
+                exports.close(cleanup=(work is None or work.done()) and execution.terminal)
+            finally:
+                stage.close(cleanup=(work is None or work.done()) and
+                            (not execution.attempted or execution.terminal))
+        finally:
+            if owned_session:
+                await session.close()

serverless_bridge/image_codec.py

@@ -0,0 +1,634 @@
+"""Bounded optional IMAGE returns, independent of ComfyUI and Distributed.
+
+Exact mode is the existing tensor wire format, unchanged. All other modes clip
+and truncate RGB/luma to uint8 (as the Collector does), normalize in float32,
+then cast to the source dtype. RGBA always carries an exact tensor alpha
+sidecar, including PNG/WebP. ``legacy_png`` means uncompressed binary PNG here,
+not the old Collector's JSON/base64 transport.
+
+The compressed wire format concatenates length-delimited stills or one Matroska
+video, followed by optional alpha bytes. No archive, path, URL or pickle is
+accepted. Limits cover wire bytes, descriptor structure, output tensor bytes,
+and aggregate padded pixel buffers, not process RSS/native codec overhead.
+Video preflight reads EBML track dimensions and block counts before creating a
+decoder. We extract packets ourselves, avoiding native container stream probing
+(which can decode before caller checks). FFmpeg max_pixels also bounds
+bitstream-driven resolution changes, including native buffer alignment.
+Software codecs only, one thread, no fallback to another encoder/format.
+"""
+
+import io
+import struct
+import zlib
+from fractions import Fraction
+
+import torch
+
+from . import codec
+from .codec import MAX_PAYLOAD_BYTES
+from .image_transfer import DEFAULT_FORMAT, DEFAULT_QUALITY, validate_options
+
+MAX_FRAMES = 240
+MAX_DIMENSION = 8192
+MAX_DESCRIPTOR_BYTES = 64 * 1024
+MAX_CONTAINER_ELEMENTS = 16384
+_VIDEO_CODECS = {
+    "ffv1": ("ffv1", None, "bgr0"),
+    "h264": ("libx264", 51, "yuv420p"),
+    "h265": ("libx265", 51, "yuv420p"),
+    "av1": ("libsvtav1", 63, "yuv420p"),
+}
+_CODEC_IDS = {"ffv1": b"V_FFV1", "h264": b"V_MPEG4/ISO/AVC",
+              "h265": b"V_MPEGH/ISO/HEVC", "av1": b"V_AV1"}
+
+
+class _BoundedIO(io.BytesIO):
+    """Enforce the final file size on every write, including muxer flush/seek."""
+    def __init__(self, limit):
+        super().__init__()
+        self.limit = limit
+
+    def write(self, data):
+        if self.tell() + len(data) > self.limit:
+            raise ValueError("Encoded IMAGE exceeds byte limit")
+        return super().write(data)
+
+    def seek(self, offset, whence=0):
+        if whence == 0:
+            position = offset
+        elif whence == 1:
+            position = self.tell() + offset
+        elif whence == 2:
+            position = self.getbuffer().nbytes + offset
+        else:
+            raise ValueError("Invalid image buffer seek")
+        if not 0 <= position <= self.limit:
+            raise ValueError("Encoded IMAGE exceeds byte limit")
+        return super().seek(offset, whence)
+
+
+def _quality_to_crf(quality, maximum):
+    return int(round(maximum * (1.0 - quality / 100.0) ** 0.6))
+
+
+def _padded_size(fmt, height, width):
+    if fmt in ("h264", "h265", "av1"):
+        height, width = height + height % 2, width + width % 2
+    # libx265 / SVT-AV1 cannot safely encode arbitrarily small pictures.
+    if fmt in ("h265", "av1"):
+        height, width = max(64, height), max(64, width)
+    return height, width
+
+
+def _shape_budget(shape, fmt, limit, itemsize=4, alpha_size=0, coded_frames=None):
+    size = codec._tensor_size("IMAGE", shape, itemsize, limit)
+    count, height, width, _ = shape
+    if count > MAX_FRAMES or max(height, width) > MAX_DIMENSION:
+        raise ValueError("IMAGE frame count or dimensions exceed limit")
+    ph, pw = _padded_size(fmt, height, width)
+    # Reserve output and padded RGBA-sized decoded buffers BEFORE allocation.
+    bh, bw = ((ph + 63) // 64 * 64, (pw + 63) // 64 * 64) if fmt in _VIDEO_CODECS else (ph, pw)
+    # Slice/OBU counts conservatively bound native frames, including AV1 hidden
+    # frames and multiple outputs from one packet. Reserve 16 bytes per pixel
+    # for hostile high-bit-depth streams, not only our usual 8-bit RGB frames.
+    native_count = max(count, coded_frames or count)
+    if native_count > MAX_FRAMES * 4:
+        raise ValueError("IMAGE coded frame count exceeds limit")
+    pixels = codec._tensor_size("IMAGE", [native_count, bh, bw, 4], 4, limit)
+    if size + pixels + alpha_size > limit:
+        raise ValueError("IMAGE decoded aggregate exceeds byte limit")
+    return ph, pw
+
+
+def _descriptor_budget(value):
+    """Bound the JSON-shaped tree without serializing attacker-sized values."""
+    remaining = MAX_DESCRIPTOR_BYTES
+    def walk(obj, depth):
+        nonlocal remaining
+        if depth > 5:
+            raise ValueError("IMAGE descriptor nesting exceeds limit")
+        remaining -= 8
+        if remaining < 0:
+            raise ValueError("IMAGE descriptor exceeds byte limit")
+        if obj is None:
+            return
+        if type(obj) is str:
+            remaining -= len(obj) * 6
+        elif type(obj) is int:
+            if obj.bit_length() > 63:
+                raise ValueError("IMAGE descriptor integer exceeds limit")
+            remaining -= 20
+        elif type(obj) is list:
+            if len(obj) > MAX_FRAMES:
+                raise ValueError("IMAGE descriptor list exceeds limit")
+            for item in obj:
+                walk(item, depth + 1)
+        elif type(obj) is dict:
+            if len(obj) > 10:
+                raise ValueError("IMAGE descriptor fields exceed limit")
+            for key, item in obj.items():
+                if type(key) is not str:
+                    raise ValueError("Invalid IMAGE descriptor key")
+                walk(key, depth + 1)
+                walk(item, depth + 1)
+        else:
+            raise ValueError("Invalid IMAGE descriptor value")
+        if remaining < 0:
+            raise ValueError("IMAGE descriptor exceeds byte limit")
+    walk(value, 0)
+
+
+def _alpha_metadata(alpha, shape, limit):
+    if shape[-1] != 4:
+        if alpha is not None:
+            raise ValueError("Unexpected IMAGE alpha sidecar")
+        return 0, torch.float32
+    if type(alpha) is not dict or set(alpha) != {"descriptor", "size"}:
+        raise ValueError("Missing or invalid IMAGE alpha sidecar")
+    d = alpha["descriptor"]
+    if type(d) is not dict or set(d) != {"encoding", "dtype", "shape", "layout"}:
+        raise ValueError("Invalid IMAGE alpha descriptor")
+    if d["encoding"] != "tensor-le-v1" or d["layout"] != "BHWC" or d["shape"] != shape[:3] + [1]:
+        raise ValueError("Invalid IMAGE alpha shape or encoding")
+    if type(d["dtype"]) is not str or d["dtype"] not in codec._DTYPES:
+        raise ValueError("Invalid IMAGE alpha dtype")
+    dtype, itemsize = codec._DTYPES[d["dtype"]]
+    size = codec._tensor_size("IMAGE", d["shape"], itemsize, limit)
+    if type(alpha["size"]) is not int or alpha["size"] != size:
+        raise ValueError("Invalid IMAGE alpha size")
+    return size, torch.float64 if dtype == torch.float64 else torch.float32
+
+
+def _import_av():
+    try:
+        import av
+    except ImportError:
+        raise ValueError("PyAV is required for IMAGE video transfer") from None
+    return av
+
+
+def _encode_still(frame, fmt, quality, limit):
+    from PIL import Image
+    sink = _BoundedIO(limit)
+    if frame.shape[-1] == 1:
+        frame = frame[..., 0]
+    image = Image.fromarray(frame)
+    if fmt in ("legacy_png", "png"):
+        image.save(sink, format="PNG", compress_level=0 if fmt == "legacy_png" else 4)
+    elif fmt == "jpeg":
+        image.save(sink, format="JPEG", quality=quality, subsampling=0 if quality >= 90 else 2)
+    else:
+        image.save(sink, format="WEBP", quality=quality, method=4)
+    return sink.getvalue()
+
+
+def _encode_video(frames, fmt, quality, height, width, limit):
+    import numpy as np
+    av = _import_av()
+    encoder, maximum, pix_fmt = _VIDEO_CODECS[fmt]
+    try:
+        av.codec.Codec(encoder, "w")
+    except Exception:
+        raise ValueError("IMAGE software encoder unavailable: " + encoder) from None
+    sink = _BoundedIO(limit)
+    options = {}
+    if maximum is not None:
+        options["crf"] = str(_quality_to_crf(quality, maximum))
+    if fmt == "h264":
+        options["preset"] = "medium"
+    elif fmt == "h265":
+        options.update(preset="fast", **{"x265-params": "log-level=error:pools=none:frame-threads=1"})
+    elif fmt == "av1":
+        options.update(preset="8", **{"svtav1-params": "lp=1"})
+    else:
+        options.update(level="3", slicecrc="1")
+    try:
+        with av.open(sink, mode="w", format="matroska") as container:
+            stream = container.add_stream(encoder, rate=24)
+            stream.width, stream.height = width, height
+            stream.pix_fmt = pix_fmt
+            stream.codec_context.thread_count = 1
+            stream.options = options
+            for index, frame in enumerate(frames):
+                if frame.shape[-1] == 1:
+                    frame = np.repeat(frame, 3, axis=-1)
+                h, w = frame.shape[:2]
+                frame = np.pad(frame, ((0, height - h), (0, width - w), (0, 0)), mode="edge")
+                vf = av.VideoFrame.from_ndarray(frame, format="rgb24")
+                vf.pts = index
+                vf.time_base = Fraction(1, 24)
+                for packet in stream.encode(vf):
+                    container.mux(packet)
+            for packet in stream.encode():
+                container.mux(packet)
+        data = sink.getvalue()
+        if len(data) > limit:
+            raise ValueError("Encoded IMAGE exceeds byte limit")
+        return data
+    except av.error.FFmpegError:
+        raise ValueError("IMAGE software video encode failed (codec, dimensions or byte limit)") from None
+    except ValueError:
+        raise
+    except Exception:
+        raise ValueError("IMAGE software video encode failed (codec, dimensions or byte limit)") from None
+
+
+def encode_image(value, *, transfer_format=DEFAULT_FORMAT, transfer_quality=DEFAULT_QUALITY,
+                 max_bytes=MAX_PAYLOAD_BYTES):
+    """Return (descriptor, bytes); reject unsupported inputs, never fall back."""
+    fmt, quality = validate_options(transfer_format, transfer_quality)
+    if fmt == "exact":
+        return codec.encode_payload("IMAGE", value, max_bytes=max_bytes)
+    limit = codec._byte_limit(max_bytes)
+    if not isinstance(value, torch.Tensor) or value.device.type != "cpu":
+        raise ValueError("IMAGE must be a CPU tensor")
+    if value.layout != torch.strided or value.is_nested:
+        raise ValueError("IMAGE must have dense strided layout")
+    if value.dtype not in codec._DTYPE_NAMES:
+        raise ValueError("Unsupported IMAGE dtype")
+    shape = list(value.shape)
+    codec._tensor_size("IMAGE", shape, value.element_size(), limit)
+    alpha_size = shape[0] * shape[1] * shape[2] * value.element_size() if shape[-1] == 4 else 0
+    itemsize = value.element_size()
+    ph, pw = _shape_budget(shape, fmt, limit, itemsize, alpha_size)
+    if not torch.isfinite(value).all().item():
+        raise ValueError("IMAGE values must be finite")
+    value = value.detach().resolve_neg()
+    alpha, alpha_payload = None, b""
+    if shape[-1] == 4:
+        ad, alpha_payload = codec.encode_payload("IMAGE", value[..., 3:4], max_bytes=limit)
+        alpha = {"descriptor": ad, "size": len(alpha_payload)}
+    import numpy as np
+    def frames():
+        for frame in value:
+            arr = frame[..., :3].float().numpy()
+            with np.errstate(over="ignore", invalid="ignore"):
+                yield np.ascontiguousarray(np.clip(arr * 255.0, 0, 255).astype(np.uint8))
+    budget = limit - len(alpha_payload)
+    if fmt in _VIDEO_CODECS:
+        parts = [_encode_video(frames(), fmt, quality, ph, pw, budget)]
+        header = _video_header(parts[0], fmt, ph, pw, shape[0])
+        coded_frames = _coded_frame_bound(parts[0], fmt, header)
+        _shape_budget(shape, fmt, limit, itemsize, alpha_size, coded_frames)
+    else:
+        parts = []
+        for frame in frames():
+            part = _encode_still(frame, fmt, quality, budget)
+            budget -= len(part)
+            parts.append(part)
+    descriptor = {"encoding": "image-transfer-v1", "format": fmt, "quality": quality,
+                  "shape": shape, "layout": "BHWC", "dtype": codec._DTYPE_NAMES[value.dtype],
+                  "parts": [len(p) for p in parts], "alpha": alpha}
+    _descriptor_budget(descriptor)
+    payload = b"".join(parts) + alpha_payload
+    if len(payload) > limit:
+        raise ValueError("Encoded IMAGE exceeds byte limit")
+    return descriptor, payload
+
+
+def _still_header(data, fmt):
+    """Read dimensions before Pillow (WebP may allocate pixels during open)."""
+    if fmt in ("png", "legacy_png"):
+        if data[:8] != b"\x89PNG\r\n\x1a\n":
+            raise ValueError("Invalid IMAGE PNG header")
+        offset, dimensions, saw_data, chunks = 8, None, False, 0
+        while offset < len(data):
+            chunks += 1
+            if chunks > MAX_CONTAINER_ELEMENTS:
+                raise ValueError("IMAGE PNG chunk count exceeds limit")
+            if offset + 12 > len(data):
+                raise ValueError("Truncated IMAGE PNG chunk")
+            length = int.from_bytes(data[offset:offset + 4], "big")
+            end = offset + 12 + length
+            if end > len(data):
+                raise ValueError("Truncated IMAGE PNG chunk")
+            tag = data[offset + 4:offset + 8]
+            body = data[offset + 8:end - 4]
+            if zlib.crc32(data[offset + 4:end - 4]) != int.from_bytes(data[end - 4:end], "big"):
+                raise ValueError("Invalid IMAGE PNG checksum")
+            if tag == b"IHDR" and offset == 8 and length == 13:
+                width, height, depth, color, compression, filtering, interlace = struct.unpack(">IIBBBBB", body)
+                if depth != 8 or color not in (0, 2) or (compression, filtering, interlace) != (0, 0, 0):
+                    raise ValueError("Unsupported IMAGE PNG layout")
+                dimensions = (height, width)
+            elif tag == b"IDAT" and dimensions:
+                saw_data = True
+            elif tag == b"IEND" and not length and end == len(data) and saw_data:
+                return dimensions
+            else:
+                raise ValueError("Unsupported IMAGE PNG chunk")
+            offset = end
+        raise ValueError("Truncated IMAGE PNG")
+    if fmt == "webp":
+        # Our encoder writes a single lossy VP8 still. Reject animation/alpha/
+        # extended WebP before WebPAnimDecoder can allocate its canvas.
+        if len(data) < 30 or data[:4] != b"RIFF" or data[8:16] != b"WEBPVP8 ":
+            raise ValueError("Unsupported IMAGE WebP header")
+        size = int.from_bytes(data[16:20], "little")
+        if int.from_bytes(data[4:8], "little") + 8 != len(data) or 20 + size + size % 2 != len(data):
+            raise ValueError("Invalid IMAGE WebP length")
+        if data[23:26] != b"\x9d\x01\x2a":
+            raise ValueError("Invalid IMAGE WebP frame")
+        width, height = struct.unpack("<HH", data[26:30])
+        return height & 0x3fff, width & 0x3fff
+    if data[:2] != b"\xff\xd8" or data[-2:] != b"\xff\xd9":
+        raise ValueError("Invalid IMAGE JPEG header or trailer")
+    offset, dimensions = 2, None
+    while offset + 4 <= len(data):
+        if data[offset] != 0xff:
+            raise ValueError("Invalid IMAGE JPEG marker")
+        marker = data[offset + 1]
+        size = int.from_bytes(data[offset + 2:offset + 4], "big")
+        end = offset + 2 + size
+        if size < 2 or end > len(data):
+            raise ValueError("Truncated IMAGE JPEG marker")
+        if marker in (0xc0, 0xc2):
+            if dimensions or size < 8 or data[offset + 4] != 8 or data[offset + 9] not in (1, 3):
+                raise ValueError("Unsupported IMAGE JPEG layout")
+            dimensions = struct.unpack(">HH", data[offset + 5:offset + 9])
+        elif marker == 0xda:
+            if dimensions is None:
+                raise ValueError("Missing IMAGE JPEG dimensions")
+            return dimensions
+        offset = end
+    raise ValueError("Missing IMAGE JPEG pixels")
+
+
+def _video_header(data, fmt, height, width, count):
+    """Strict bounded EBML preflight: one track, no lacing/attachments/encodings.
+
+    Native stream probing is not the first dimension check: it may instantiate a
+    decoder. Validate the Matroska header and count blocks without decompressing.
+    """
+    nodes = 0
+    def vint(pos, end, identifier=False):
+        if pos >= end or data[pos] == 0:
+            raise ValueError("Invalid IMAGE EBML integer")
+        length = 9 - data[pos].bit_length()
+        if length > (4 if identifier else 8) or pos + length > end:
+            raise ValueError("Truncated IMAGE EBML integer")
+        value = int.from_bytes(data[pos:pos + length], "big")
+        if not identifier:
+            value &= (1 << (7 * length)) - 1
+            if value == (1 << (7 * length)) - 1:
+                raise ValueError("Unknown-sized IMAGE EBML elements are forbidden")
+        return value, pos + length
+    def elements(start, end):
+        nonlocal nodes
+        while start < end:
+            nodes += 1
+            if nodes > MAX_CONTAINER_ELEMENTS:
+                raise ValueError("IMAGE container element count exceeds limit")
+            tag, pos = vint(start, end, True)
+            length, body = vint(pos, end)
+            finish = body + length
+            if finish > end:
+                raise ValueError("Truncated IMAGE EBML element")
+            yield tag, body, finish
+            start = finish
+    def integer(start, end):
+        if not 1 <= end - start <= 8:
+            raise ValueError("Invalid IMAGE EBML numeric field")
+        return int.from_bytes(data[start:end], "big")
+    roots = list(elements(0, len(data)))
+    if len(roots) != 2 or [r[0] for r in roots] != [0x1a45dfa3, 0x18538067]:
+        raise ValueError("Invalid IMAGE Matroska structure")
+    tracks, blocks, packets = [], 0, []
+    for tag, start, end in elements(roots[1][1], roots[1][2]):
+        if tag == 0x1654ae6b:
+            for entry, es, ee in elements(start, end):
+                if entry in (0xbf, 0xec):
+                    continue
+                if entry != 0xae:
+                    raise ValueError("Unexpected IMAGE track element")
+                fields, dims = {}, {}
+                for field, fs, fe in elements(es, ee):
+                    if field in fields:
+                        raise ValueError("Duplicate IMAGE track field")
+                    fields[field] = (fs, fe)
+                    if field == 0x6d80:
+                        raise ValueError("IMAGE track encodings are forbidden")
+                    if field == 0xe0:
+                        for dim, ds, de in elements(fs, fe):
+                            if dim in dims:
+                                raise ValueError("Duplicate IMAGE video field")
+                            dims[dim] = (ds, de)
+                if any(key not in fields for key in (0xd7, 0x83, 0x86)) or any(key not in dims for key in (0xb0, 0xba)):
+                    raise ValueError("Incomplete IMAGE video track")
+                if integer(*fields[0xd7]) != 1 or integer(*fields[0x83]) != 1:
+                    raise ValueError("Unsupported IMAGE video track")
+                if fields[0x86][1] - fields[0x86][0] > 32:
+                    raise ValueError("IMAGE codec identifier exceeds limit")
+                codec_id = bytes(data[slice(*fields[0x86])])
+                if codec_id != _CODEC_IDS[fmt] and not (fmt == "ffv1" and codec_id == b"V_MS/VFW/FOURCC"):
+                    raise ValueError("IMAGE video codec mismatch")
+                if integer(*dims[0xb0]) != width or integer(*dims[0xba]) != height:
+                    raise ValueError("IMAGE video header dimensions mismatch")
+                if codec_id == b"V_MS/VFW/FOURCC":
+                    cs, ce = fields.get(0x63a2, (0, 0))
+                    if ce - cs < 40 or data[cs + 16:cs + 20] != b"FFV1":
+                        raise ValueError("Invalid IMAGE FFV1 codec header")
+                    if struct.unpack("<ii", data[cs + 4:cs + 12]) != (width, height):
+                        raise ValueError("IMAGE FFV1 header dimensions mismatch")
+                tracks.append(fields)
+        elif tag == 0x1f43b675:
+            def block(bs, be):
+                nonlocal blocks
+                track, pos = vint(bs, be)
+                if track != 1 or pos + 3 >= be or data[pos + 2] & 0x06:
+                    raise ValueError("Invalid or laced IMAGE video block")
+                blocks += 1
+                if blocks > count:
+                    raise ValueError("IMAGE video frame count exceeds descriptor")
+                packets.append((pos + 3, be))
+            for field, fs, fe in elements(start, end):
+                if field == 0xa3:
+                    block(fs, fe)
+                elif field == 0xa0:
+                    group = [e for e in elements(fs, fe) if e[0] == 0xa1]
+                    if len(group) != 1:
+                        raise ValueError("Invalid IMAGE video block group")
+                    block(group[0][1], group[0][2])
+                elif field not in (0xe7, 0xec, 0xbf):
+                    raise ValueError("Unsupported IMAGE cluster field")
+        elif tag not in (0x114d9b74, 0x1549a966, 0x1c53bb6b, 0x1254c367, 0xec, 0xbf):
+            raise ValueError("Unsupported IMAGE Matroska element")
+    if len(tracks) != 1 or blocks != count:
+        raise ValueError("IMAGE video track or frame count mismatch")
+    cs, ce = tracks[0].get(0x63a2, (0, 0))
+    if not 0 < ce - cs <= MAX_DESCRIPTOR_BYTES:
+        raise ValueError("IMAGE codec configuration exceeds limit or is missing")
+    if bytes(data[slice(*tracks[0][0x86])]) == b"V_MS/VFW/FOURCC":
+        cs += 40  # Skip BITMAPINFOHEADER; FFV1 receives only its codec config.
+    return bytes(data[cs:ce]), packets
+
+
+def _coded_frame_bound(data, fmt, header):
+    """Conservative per-packet allocation bound before native decode.
+
+    FFV1 has one picture per packet. Every AVC/HEVC picture needs at least one
+    VCL NAL; every AV1 output (including show-existing) needs a frame/header OBU.
+    Counting slices overestimates pictures, which is safe. Do not trust a
+    Matroska block count alone: a malicious packet may contain several pictures.
+    """
+    config, packets = header
+    if fmt == "ffv1":
+        return len(packets)
+    total, units = 0, 0
+    if fmt in ("h264", "h265"):
+        index = 4 if fmt == "h264" else 21
+        if len(config) <= index or config[0] != 1:
+            raise ValueError("Invalid IMAGE video configuration")
+        length_size = (config[index] & 3) + 1
+    for start, end in packets:
+        while start < end:
+            units += 1
+            if units > MAX_CONTAINER_ELEMENTS:
+                raise ValueError("IMAGE coded unit count exceeds limit")
+            if fmt in ("h264", "h265"):
+                if start + length_size > end:
+                    raise ValueError("Truncated IMAGE NAL length")
+                length = int.from_bytes(data[start:start + length_size], "big")
+                start += length_size
+                if length < (1 if fmt == "h264" else 2) or length > end - start:
+                    raise ValueError("Truncated IMAGE NAL unit")
+                kind = (data[start] & 31) if fmt == "h264" else (data[start] >> 1) & 63
+                total += (1 <= kind <= 5) if fmt == "h264" else (kind <= 31)
+            else:
+                tag = data[start]
+                start += 1
+                if tag & 0x81 or not tag & 2:
+                    raise ValueError("Unsupported IMAGE AV1 OBU header")
+                kind = (tag >> 3) & 15
+                if tag & 4:
+                    start += 1
+                length = 0
+                for shift in range(0, 56, 7):
+                    if start >= end:
+                        raise ValueError("Truncated IMAGE AV1 OBU length")
+                    value = data[start]
+                    start += 1
+                    length |= (value & 127) << shift
+                    if not value & 128:
+                        break
+                else:
+                    raise ValueError("IMAGE AV1 OBU length exceeds limit")
+                if length > end - start:
+                    raise ValueError("Truncated IMAGE AV1 OBU")
+                total += kind in (3, 6)
+            start += length
+            if total > MAX_FRAMES * 4:
+                raise ValueError("IMAGE coded frame count exceeds limit")
+    if total < len(packets):
+        raise ValueError("IMAGE video has missing coded frames")
+    return total
+
+
+def _decode_video(data, fmt, height, width, count, header):
+    av = _import_av()
+    maximum = ((height + 63) // 64 * 64) * ((width + 63) // 64 * 64)
+    ctx = av.CodecContext.create("libdav1d" if fmt == "av1" else "hevc" if fmt == "h265" else fmt, "r")
+    ctx.width, ctx.height = width, height
+    ctx.extradata, packets = header
+    ctx.thread_count = 1
+    ctx.options = {"max_pixels": str(maximum), "threads": "1", "err_detect": "explode"}
+    ctx.open()
+    if "max_pixels" in ctx.options:
+        raise ValueError("IMAGE decoder does not support max_pixels safety limit")
+    if (ctx.width, ctx.height) != (width, height):
+        raise ValueError("IMAGE bitstream dimensions mismatch")
+    decoded = 0
+    for bounds in packets + [None]:
+        packet = av.Packet(data[slice(*bounds)]) if bounds is not None else None
+        for frame in ctx.decode(packet):
+            if decoded >= count or (frame.width, frame.height) != (width, height):
+                raise ValueError("IMAGE decoded dimensions or frame count mismatch")
+            decoded += 1
+            yield frame.to_ndarray(format="rgb24")
+    if decoded != count:
+        raise ValueError("IMAGE decoded frame count mismatch")
+
+
+def decode_image(descriptor, payload, *, max_bytes=MAX_PAYLOAD_BYTES):
+    """Decode to an independent contiguous CPU tensor after bounded preflight."""
+    limit = codec._byte_limit(max_bytes)
+    if not isinstance(descriptor, dict):
+        raise ValueError("IMAGE descriptor must be a dictionary")
+    if descriptor.get("encoding") == "tensor-le-v1":
+        return codec.decode_payload("IMAGE", descriptor, payload, max_bytes=limit)
+    _descriptor_budget(descriptor)
+    if set(descriptor) != {"encoding", "format", "quality", "shape", "layout", "dtype", "parts", "alpha"}:
+        raise ValueError("Invalid IMAGE descriptor fields")
+    if descriptor["encoding"] != "image-transfer-v1" or descriptor["layout"] != "BHWC":
+        raise ValueError("Invalid IMAGE descriptor encoding or layout")
+    fmt, _ = validate_options(descriptor["format"], descriptor["quality"])
+    if fmt == "exact":
+        raise ValueError("Exact IMAGE requires the tensor encoding")
+    if not isinstance(payload, bytes):
+        raise ValueError("IMAGE payload must be bytes")
+    if len(payload) > limit:
+        raise ValueError("IMAGE payload exceeds byte limit")
+    shape = descriptor["shape"]
+    if type(descriptor["dtype"]) is not str or descriptor["dtype"] not in codec._DTYPES:
+        raise ValueError("Unsupported IMAGE descriptor dtype")
+    dtype, itemsize = codec._DTYPES[descriptor["dtype"]]
+    codec._tensor_size("IMAGE", shape, itemsize, limit)
+    alpha_size, _ = _alpha_metadata(descriptor["alpha"], shape, limit)
+    if alpha_size and descriptor["alpha"]["descriptor"]["dtype"] != descriptor["dtype"]:
+        raise ValueError("IMAGE alpha dtype mismatch")
+    ph, pw = _shape_budget(shape, fmt, limit, itemsize, alpha_size)
+    count, height, width, channels = shape
+    lengths = descriptor["parts"]
+    if type(lengths) is not list or len(lengths) != (1 if fmt in _VIDEO_CODECS else count):
+        raise ValueError("IMAGE part count mismatch")
+    total = alpha_size
+    for length in lengths:
+        if type(length) is not int or not 0 < length <= limit - total:
+            raise ValueError("IMAGE part length exceeds byte limit")
+        total += length
+    if total != len(payload):
+        raise ValueError("IMAGE payload length does not match parts")
+    parts, offset = [], 0
+    for length in lengths:
+        parts.append(memoryview(payload)[offset:offset + length])
+        offset += length
+    try:
+        if fmt in _VIDEO_CODECS:
+            video_header = _video_header(parts[0], fmt, ph, pw, count)
+            coded_frames = _coded_frame_bound(parts[0], fmt, video_header)
+            _shape_budget(shape, fmt, limit, itemsize, alpha_size, coded_frames)
+        else:
+            for part in parts:
+                if _still_header(part, fmt) != (height, width):
+                    raise ValueError("IMAGE header dimensions mismatch")
+        # All headers and aggregate budgets are checked before output allocation.
+        alpha = None
+        if alpha_size:
+            alpha = codec.decode_payload("IMAGE", descriptor["alpha"]["descriptor"], payload[offset:], max_bytes=limit)
+        result = torch.empty(shape, dtype=dtype, device="cpu")
+        if fmt in _VIDEO_CODECS:
+            frames = _decode_video(parts[0], fmt, ph, pw, count, video_header)
+        else:
+            def stills():
+                import numpy as np
+                from PIL import Image
+                for part in parts:
+                    with Image.open(io.BytesIO(part)) as image:
+                        if image.size != (width, height) or getattr(image, "n_frames", 1) != 1:
+                            raise ValueError("IMAGE still dimensions or frame count mismatch")
+                        if image.mode not in ("L", "RGB"):
+                            raise ValueError("Unsupported IMAGE still pixel mode")
+                        yield np.array(image.convert("RGB"), dtype=np.uint8)
+            frames = stills()
+        color = 3 if channels == 4 else channels
+        for index, frame in enumerate(frames):
+            # Float32 division followed by the source dtype cast reproduces
+            # Collector normalized pixels without narrowing exact alpha.
+            result[index, ..., :color] = torch.from_numpy(frame[:height, :width, :color].copy()).float().div_(255)
+        if alpha is not None:
+            result[..., 3:4] = alpha
+        return result
+    except ValueError:
+        raise
+    except Exception:
+        raise ValueError("Malformed or unsupported IMAGE compressed payload") from None

serverless_bridge/image_transfer.py

@@ -0,0 +1,30 @@
+"""Image-return policy shared by the master and standalone endpoint bridge.
+
+This module is deliberately dependency-free so graph compilation does not import
+Torch, Pillow, PyAV or ComfyUI. Existing Collector settings stay unchanged.
+"""
+
+FORMATS = ("exact", "legacy_png", "png", "ffv1", "jpeg", "webp", "h264", "h265", "av1")
+DEFAULT_FORMAT = "exact"
+DEFAULT_QUALITY = 90
+FORMAT_TOOLTIP = (
+    "Image returns only; master-local images and other media stay unchanged. "
+    "exact: original floating-point tensor, no quantization. "
+    "legacy_png / png / ffv1: lossless 8-bit pixels, not exact float tensors. "
+    "jpeg / webp: lossy stills; h264 / h265 / av1: lossy frame-batch video. "
+    "Alpha is preserved separately when the codec cannot carry it. "
+    "Requires matching transfer support on the worker/endpoint."
+)
+QUALITY_TOOLTIP = (
+    "Quality 1–100 for lossy image returns; ignored by exact and lossless modes. "
+    "JPEG/WebP use it directly; video formats use the existing Collector CRF mapping."
+)
+
+
+def validate_options(transfer_format=DEFAULT_FORMAT, transfer_quality=DEFAULT_QUALITY):
+    """Reject malformed controls rather than repairing or silently substituting."""
+    if type(transfer_format) is not str or transfer_format not in FORMATS:
+        raise ValueError("Unsupported image transfer_format")
+    if type(transfer_quality) is not int or not 1 <= transfer_quality <= 100:
+        raise ValueError("Image transfer_quality must be an integer from 1 to 100")
+    return transfer_format, transfer_quality

serverless_bridge/nodes.py

@@ -0,0 +1,175 @@
+"""Internal ComfyUI boundary nodes for the standalone media-v1 handler."""
+import hashlib
+import json
+import uuid
+from pathlib import Path
+
+import folder_paths
+from comfy_api.v0_0_2 import io
+
+from .codec import decode_payload, encode_payload
+from .image_transfer import (FORMATS, DEFAULT_FORMAT, DEFAULT_QUALITY,
+                             FORMAT_TOOLTIP, QUALITY_TOOLTIP, validate_options)
+from .contract import (
+    MAX_ARTIFACT_BYTES, MAX_ITEMS, MAX_TOTAL_BYTES, SOCKET_KINDS,
+    bounded_int, identifier, validate_binding, validate_json,
+)
+
+
+def _scalar(value):
+    if isinstance(value, list):
+        if len(value) != 1:
+            raise ValueError("Media bridge control must have exactly one value")
+        return value[0]
+    return value
+
+
+def _json(text):
+    if not isinstance(text, str) or len(text) > 256 * 1024:
+        raise ValueError("Invalid media bridge JSON")
+    try:
+        value = json.loads(text)
+    except (ValueError, TypeError, RecursionError):
+        raise ValueError("Invalid media bridge JSON") from None
+    validate_json(value)
+    return value
+
+
+def _cpu(kind, value):
+    # Runtime transfer only, never sampling or device discovery. Pure codecs
+    # deliberately reject non-CPU storage rather than making hidden transfers.
+    if kind in ("IMAGE", "MASK"):
+        return value.detach().cpu()
+    if kind == "AUDIO":
+        return {"waveform": value["waveform"].detach().cpu(), "sample_rate": value["sample_rate"]}
+    return value
+
+
+def encode_value(kind, value, *, transfer_format=DEFAULT_FORMAT, transfer_quality=DEFAULT_QUALITY):
+    fmt, quality = validate_options(transfer_format, transfer_quality)
+    if kind == "IMAGE" and fmt != DEFAULT_FORMAT:
+        from .image_codec import encode_image
+        return encode_image(_cpu(kind, value), transfer_format=fmt, transfer_quality=quality)
+    if kind == "VIDEO":
+        from .video_codec import encode_video_cpu
+        return encode_video_cpu(value)
+    return encode_payload(kind, _cpu(kind, value))
+
+
+def decode_value(kind, descriptor, payload):
+    if kind == "VIDEO":
+        from .video_codec import decode_video
+        return decode_video(descriptor, payload)
+    return decode_payload(kind, descriptor, payload)
+
+
+def export_values(request_id, binding, values, *, transfer_format=DEFAULT_FORMAT, transfer_quality=DEFAULT_QUALITY):
+    """Write bounded owned payloads; history contains metadata, never base64."""
+    identifier(request_id)
+    validate_binding(binding)
+    fmt, quality = validate_options(transfer_format, transfer_quality)
+    if binding["kind"] != "IMAGE" and fmt != DEFAULT_FORMAT:
+        raise ValueError("Image transfer controls apply only to IMAGE exports")
+    if not isinstance(values, list) or not 1 <= len(values) <= MAX_ITEMS:
+        raise ValueError("Connected media output has no values or exceeds item limit")
+    request_key = hashlib.sha256(request_id.encode()).hexdigest()
+    subfolder = f"distributed-media-v1/{request_key}"
+    root = Path(folder_paths.get_temp_directory()).resolve()
+    directory = (root / subfolder).resolve()
+    if not directory.is_relative_to(root):
+        raise ValueError("Unsafe media output directory")
+    directory.mkdir(parents=True, exist_ok=True)
+    written, artifacts, items = [], [], []
+    total = 0
+    try:
+        for index, value in enumerate(values):
+            descriptor, payload = encode_value(binding["kind"], value,
+                                               transfer_format=fmt, transfer_quality=quality)
+            validate_json(descriptor)
+            total += len(payload)
+            if total > MAX_TOTAL_BYTES:
+                raise ValueError("Media output exceeds total byte limit")
+            aid = uuid.uuid4().hex
+            filename = aid + ".bin"
+            path = directory / filename
+            with path.open("xb") as stream:
+                written.append(path)
+                stream.write(payload)
+            artifacts.append({"id": aid, "kind": binding["kind"], "descriptor": descriptor,
+                              "byte_length": len(payload), "sha256": hashlib.sha256(payload).hexdigest(),
+                              "transport": {"type": "comfy-file", "filename": filename,
+                                            "subfolder": subfolder, "folder_type": "temp"}})
+            items.append({"binding_id": binding["binding_id"], "item_index": index, "artifact_id": aid})
+    except BaseException:
+        for path in written:
+            path.unlink(missing_ok=True)
+        raise
+    return {"request_id": request_id, "binding": dict(binding, count=len(values)),
+            "artifacts": artifacts, "items": items}
+
+
+def import_value(artifact, filename):
+    """Read a handler-staged, checksummed file within native media roots only."""
+    fields = {"id", "kind", "descriptor", "byte_length", "sha256"}
+    if not isinstance(artifact, dict) or set(artifact) != fields:
+        raise ValueError("Invalid staged media metadata")
+    identifier(artifact["id"])
+    if artifact["kind"] not in SOCKET_KINDS.values():
+        raise ValueError("Invalid staged media kind")
+    size = bounded_int(artifact["byte_length"], MAX_ARTIFACT_BYTES)
+    if not isinstance(filename, str):
+        raise ValueError("Invalid staged media filename")
+    roots = [Path(getter()).resolve() for getter in (
+        folder_paths.get_input_directory, folder_paths.get_output_directory, folder_paths.get_temp_directory)]
+    path = Path(folder_paths.get_annotated_filepath(filename)).resolve()
+    if not any(path.is_relative_to(root) for root in roots) or not path.is_file():
+        raise ValueError("Staged media file is outside allowed roots or unavailable")
+    if path.stat().st_size != size:
+        raise ValueError("Staged media file length mismatch")
+    with path.open("rb") as stream:
+        payload = stream.read(size + 1)
+    if len(payload) != size or hashlib.sha256(payload).hexdigest() != artifact["sha256"]:
+        raise ValueError("Staged media file failed integrity validation")
+    return decode_value(artifact["kind"], artifact["descriptor"], payload)
+
+
+class MediaImportV1(io.ComfyNode):
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(node_id="MediaImportV1", category="distributed/internal", is_deprecated=True,
+            inputs=[io.String.Input("artifact_json"), io.String.Input("filename")],
+            outputs=[io.Image.Output(), io.Audio.Output(), io.Mask.Output(), io.String.Output(), io.Video.Output()])
+
+    @classmethod
+    def execute(cls, artifact_json, filename):
+        artifact = _json(artifact_json)
+        value = import_value(artifact, filename)
+        return io.NodeOutput(*(value if kind == artifact["kind"] else None for kind in SOCKET_KINDS.values()))
+
+
+class MediaExportV1(io.ComfyNode):
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(node_id="MediaExportV1", category="distributed/internal", is_deprecated=True,
+            inputs=[io.String.Input("request_id"), io.String.Input("binding_json"),
+                    io.Image.Input("images", optional=True), io.Audio.Input("audio", optional=True),
+                    io.Mask.Input("masks", optional=True), io.String.Input("texts", optional=True, force_input=True),
+                    io.Video.Input("video", optional=True),
+                    io.Combo.Input("transfer_format", options=list(FORMATS), default=DEFAULT_FORMAT,
+                                   optional=True, tooltip=FORMAT_TOOLTIP),
+                    io.Int.Input("transfer_quality", default=DEFAULT_QUALITY, min=1, max=100,
+                                 optional=True, tooltip=QUALITY_TOOLTIP)],
+            outputs=[], is_output_node=True, is_input_list=True)
+
+    @classmethod
+    def execute(cls, request_id, binding_json, images=None, audio=None, masks=None, texts=None, video=None,
+                transfer_format=DEFAULT_FORMAT, transfer_quality=DEFAULT_QUALITY):
+        binding = _json(_scalar(binding_json))
+        validate_binding(binding)
+        provided = dict(images=images, audio=audio, masks=masks, texts=texts, video=video)
+        if {name for name, value in provided.items() if value is not None} != {binding["socket"]}:
+            raise ValueError("Media export must receive exactly its declared typed socket")
+        exported = export_values(_scalar(request_id), binding, provided[binding["socket"]],
+                                 transfer_format=_scalar(transfer_format),
+                                 transfer_quality=_scalar(transfer_quality))
+        return io.NodeOutput(ui={"media_exports": [exported]})

serverless_bridge/video_codec.py

@@ -0,0 +1,482 @@
+"""Lossless transport of pinned native ComfyUI VIDEO representations.
+
+No save_to/get_components fallback is allowed: an unknown class must not silently
+transcode. Tensor components use the shared CPU codec. File bytes are opaque and
+retain signed trim/crop parameters; decoded files own independent BytesIOs.
+VideoFromList's native constructor flattens its inputs, so wire segments are a
+bounded, flat list of exact VideoFromFile objects, never recursive playlists.
+
+Offsets are relative to the containing VIDEO payload. Canonical part order is
+images/alpha/audio, or segments/complete_audio. Every byte is covered exactly
+once. All descriptor structure and tensor sizes are preflighted before decoding
+any media storage. File containers are not parsed until a native consumer reads
+them; this adapter does not attempt to validate or repair container codecs.
+"""
+from fractions import Fraction
+import io
+import json
+import math
+import os
+from pathlib import Path
+import stat
+
+from .codec import MAX_PAYLOAD_BYTES, decode_payload, encode_payload
+
+
+MAX_SEGMENTS = 64
+MAX_METADATA_BYTES = 64 * 1024
+MAX_METADATA_DEPTH = 16
+_ENCODING = "native-video-v1"
+_BASE_FIELDS = {"encoding", "representation", "byte_length"}
+_COMPONENT_FIELDS = {"images", "alpha", "audio", "frame_rate", "metadata", "bit_depth", "color_space"}
+_FILE_FIELDS = {"data", "start_time", "duration", "crop"}
+_SEGMENT_FIELDS = {"segments", "complete_audio"}
+_TENSOR_BYTES = {"float16": 2, "bfloat16": 2, "float32": 4, "float64": 8}
+
+
+def _limit(max_bytes):
+    if type(max_bytes) is not int or max_bytes < 0:
+        raise ValueError("VIDEO max_bytes must be a nonnegative integer")
+    return min(max_bytes, MAX_PAYLOAD_BYTES)
+
+
+def _native_types():
+    # Lazy imports keep non-VIDEO users independent of ComfyUI and PyAV.
+    from comfy_api.latest._input_impl.video_types import (
+        VideoFromComponents, VideoFromFile, VideoFromList,
+    )
+    from comfy_api.latest._util import VideoComponents
+    return VideoFromComponents, VideoFromFile, VideoFromList, VideoComponents
+
+
+def _fields(value, fields, label):
+    if type(value) is not dict or set(value) != fields:
+        raise ValueError(f"Invalid VIDEO {label} fields")
+
+
+def _integer(value, *, positive=False):
+    if type(value) is not int or value < (1 if positive else 0):
+        raise ValueError("VIDEO length, offset and dimensions must be integers in range")
+    return value
+
+
+def _number(value):
+    if type(value) not in (int, float) or (type(value) is float and not math.isfinite(value)):
+        raise ValueError("VIDEO trim values must be finite numbers")
+    return value
+
+
+def _crop(value, *, wire=False):
+    if value is None:
+        return None
+    if type(value) is not (list if wire else tuple) or len(value) != 4:
+        raise ValueError("VIDEO crop must contain four integers")
+    if any(type(part) is not int for part in value):
+        raise ValueError("VIDEO crop must contain four integers")
+    # Native crop normalization is a consumer concern, not a transport operation.
+    return list(value) if not wire else tuple(value)
+
+
+def _frame_rate(value):
+    _fields(value, {"numerator", "denominator"}, "frame_rate")
+    _integer(value["numerator"], positive=True)
+    _integer(value["denominator"], positive=True)
+
+
+def _encoding_options(bit_depth, color_space):
+    _integer(bit_depth, positive=True)
+    if type(color_space) is not str or color_space not in ("sRGB", "HDR", "HDR PQ"):
+        raise ValueError("Unsupported native VIDEO color_space")
+
+
+def _metadata(value):
+    """Validate bounded JSON before allocating its serialized form; return a copy."""
+    if value is None:
+        return None
+    if type(value) is not dict:
+        raise ValueError("VIDEO metadata must be a JSON object or null")
+    remaining = MAX_METADATA_BYTES
+
+    def charge(size):
+        nonlocal remaining
+        remaining -= size
+        if remaining < 0:
+            raise ValueError("VIDEO metadata exceeds byte limit")
+
+    def visit(item, depth):
+        if depth > MAX_METADATA_DEPTH:
+            raise ValueError("VIDEO metadata exceeds depth limit")
+        kind = type(item)
+        if kind in (dict, list):
+            # Each member costs at least one JSON byte. Also bounds iteration
+            # before walking potentially huge containers or cyclic structures.
+            if len(item) > remaining:
+                raise ValueError("VIDEO metadata exceeds byte limit")
+            charge(2 + max(0, len(item) - 1))
+            if kind is dict:
+                for key, child in item.items():
+                    if type(key) is not str:
+                        raise ValueError("VIDEO metadata keys must be strings")
+                    visit(key, depth + 1)
+                    charge(1)
+                    visit(child, depth + 1)
+            else:
+                for child in item:
+                    visit(child, depth + 1)
+        elif kind in (str, int, float, bool) or item is None:
+            if kind is str and len(item) > remaining:
+                raise ValueError("VIDEO metadata exceeds byte limit")
+            # Avoid converting an unbounded Python integer to decimal first.
+            if kind is int and item.bit_length() > MAX_METADATA_BYTES * 4:
+                raise ValueError("VIDEO metadata exceeds byte limit")
+            if kind is float and not math.isfinite(item):
+                raise ValueError("VIDEO metadata numbers must be finite")
+            try:
+                charge(len(json.dumps(item, ensure_ascii=True, allow_nan=False)))
+            except (ValueError, OverflowError):
+                raise ValueError("Invalid or oversized VIDEO metadata scalar") from None
+        else:
+            raise ValueError("VIDEO metadata must contain only JSON values")
+
+    visit(value, 0)
+    return json.loads(json.dumps(value, ensure_ascii=True, allow_nan=False, separators=(",", ":")))
+
+
+def _attrs(value, names):
+    attributes = vars(value)
+    if not all(name in attributes for name in names):
+        raise ValueError("Native VIDEO attributes do not match the pinned adapter")
+    return [attributes[name] for name in names]
+
+
+def _roots(allowed_roots):
+    if allowed_roots is None:
+        import folder_paths
+        allowed_roots = [folder_paths.get_input_directory(), folder_paths.get_output_directory(),
+                         folder_paths.get_temp_directory()]
+    if type(allowed_roots) not in (list, tuple):
+        raise ValueError("VIDEO allowed_roots must be a list or tuple")
+    try:
+        roots = [Path(root).resolve() for root in allowed_roots]
+    except (OSError, TypeError, ValueError, RuntimeError):
+        raise ValueError("Invalid VIDEO allowed roots") from None
+    return roots
+
+
+def _read_path(value, limit, allowed_roots):
+    """Resolve authorization, reject special files, and bound all reads.
+
+    Open each resolved path component with NOFOLLOW using directory fds, so an
+    intermediate symlink swapped in after resolve() cannot bypass authorization.
+    NONBLOCK prevents an attacker swapping in a FIFO from hanging the worker.
+    """
+    roots = _roots(allowed_roots)
+    directory_fd = file_fd = None
+    try:
+        path = Path(value).resolve(strict=True)
+        if not any(path.is_relative_to(root) for root in roots):
+            raise ValueError("VIDEO file is outside allowed roots")
+        directory_flags = os.O_RDONLY | os.O_DIRECTORY | os.O_NOFOLLOW
+        directory_fd = os.open(path.anchor, directory_flags)
+        for part in path.parts[1:-1]:
+            next_fd = os.open(part, directory_flags, dir_fd=directory_fd)
+            os.close(directory_fd)
+            directory_fd = next_fd
+        file_fd = os.open(path.name, os.O_RDONLY | os.O_NOFOLLOW | os.O_NONBLOCK, dir_fd=directory_fd)
+        info = os.fstat(file_fd)
+        if not stat.S_ISREG(info.st_mode):
+            raise ValueError("VIDEO source must be a regular file")
+        if info.st_size > limit:
+            raise ValueError("VIDEO file exceeds byte limit")
+        # Read at most the stat size plus a sentinel; reject growth or truncation.
+        chunks = []
+        count = 0
+        while count <= info.st_size:
+            chunk = os.read(file_fd, min(1024 * 1024, info.st_size + 1 - count))
+            if not chunk:
+                break
+            chunks.append(chunk)
+            count += len(chunk)
+        if count != info.st_size:
+            raise ValueError("VIDEO source changed size during bounded read")
+        return b"".join(chunks)
+    except (OSError, RuntimeError):
+        raise ValueError("VIDEO file cannot be read safely") from None
+    finally:
+        if file_fd is not None:
+            os.close(file_fd)
+        if directory_fd is not None:
+            os.close(directory_fd)
+
+
+def _file_bytes(value, limit, allowed_roots):
+    if type(value) is io.BytesIO:
+        try:
+            with value.getbuffer() as view:
+                if view.nbytes > limit:
+                    raise ValueError("VIDEO file exceeds byte limit")
+                return bytes(view)
+        except ValueError:
+            raise ValueError("VIDEO BytesIO is closed or exceeds byte limit") from None
+    if isinstance(value, (str, os.PathLike)):
+        return _read_path(value, limit, allowed_roots)
+    raise ValueError("VIDEO file must be a local path or exact BytesIO")
+
+
+def encode_video(value, *, max_bytes=MAX_PAYLOAD_BYTES, allowed_roots=None):
+    """Encode an exact native VIDEO without opening/transcoding its container.
+
+    Paths are authorized only under allowed_roots (when supplied), otherwise
+    under ComfyUI input/output/temp roots. Budgets may lower, never raise, the
+    shared maximum. Unsupported native state raises ValueError, not a fallback.
+    Non-CPU tensors are rejected; runtime staging belongs to the bridge boundary.
+    """
+    return _encode_video(value, max_bytes=max_bytes, allowed_roots=allowed_roots, stage_cpu=False)
+
+
+def encode_video_cpu(value, *, max_bytes=MAX_PAYLOAD_BYTES, allowed_roots=None):
+    """Bridge-only boundary: detach/stage native tensors, then losslessly encode.
+
+    Share the exact-type adapter with encode_video rather than reconstructing
+    native videos, calling get_components, or duplicating private attributes.
+    Images, alpha and both audio representations move without dtype conversion;
+    metadata/options and opaque file bytes/trim/crop take the unchanged path.
+    """
+    return _encode_video(value, max_bytes=max_bytes, allowed_roots=allowed_roots, stage_cpu=True)
+
+
+def _stage_tensor_cpu(value):
+    import torch
+    if not isinstance(value, torch.Tensor):
+        raise ValueError("Tensor value must be a torch.Tensor")
+    return value.detach().cpu()
+
+
+def _encode_video(value, *, max_bytes, allowed_roots, stage_cpu):
+    limit = _limit(max_bytes)
+    components_class, file_class, list_class, components_type = _native_types()
+    descriptor = {"encoding": _ENCODING}
+    chunks = []
+    size = 0
+
+    def add(payload, nested=None):
+        nonlocal size
+        if len(payload) > limit - size:
+            raise ValueError("VIDEO payload exceeds byte limit")
+        slot = {"offset": size, "length": len(payload)}
+        if nested is not None:
+            slot["descriptor"] = nested
+        chunks.append(payload)
+        size += len(payload)
+        return slot
+
+    def tensor(kind, tensor_value):
+        if stage_cpu:
+            if kind == "AUDIO":
+                if not isinstance(tensor_value, dict) or set(tensor_value) != {"waveform", "sample_rate"}:
+                    raise ValueError("AUDIO value must contain waveform and sample_rate")
+                tensor_value = {**tensor_value, "waveform": _stage_tensor_cpu(tensor_value["waveform"])}
+            else:
+                tensor_value = _stage_tensor_cpu(tensor_value)
+        nested, payload = encode_payload(kind, tensor_value, max_bytes=limit - size)
+        return add(payload, nested)
+
+    if type(value) is components_class:
+        component, bit_depth, color_space = _attrs(value, [
+            "_VideoFromComponents__components", "_VideoFromComponents__bit_depth",
+            "_VideoFromComponents__color_space",
+        ])
+        if type(component) is not components_type:
+            raise ValueError("Unsupported native VIDEO components class")
+        images, fps, audio, metadata, alpha = _attrs(component, [
+            "images", "frame_rate", "audio", "metadata", "alpha",
+        ])
+        if type(fps) is not Fraction or fps <= 0:
+            raise ValueError("VIDEO frame_rate must be a positive Fraction")
+        _encoding_options(bit_depth, color_space)
+        descriptor.update(representation="components", bit_depth=bit_depth, color_space=color_space,
+                          frame_rate={"numerator": fps.numerator, "denominator": fps.denominator},
+                          metadata=_metadata(metadata))
+        descriptor["images"] = tensor("IMAGE", images)
+        # Native get_components() returns BHW1 alpha; MASK nodes use BHW.
+        # Select an existing lossless tensor layout without squeezing/converting.
+        alpha_kind = "IMAGE" if getattr(alpha, "ndim", None) == 4 else "MASK"
+        descriptor["alpha"] = tensor(alpha_kind, alpha) if alpha is not None else None
+        if alpha is not None:
+            _alpha_header(descriptor["alpha"], descriptor["images"])
+        descriptor["audio"] = tensor("AUDIO", audio) if audio is not None else None
+    elif type(value) is file_class:
+        source, start, duration, crop = _attrs(value, [
+            "_VideoFromFile__file", "_VideoFromFile__start_time", "_VideoFromFile__duration",
+            "_VideoFromFile__crop",
+        ])
+        descriptor.update(representation="file", start_time=_number(start), duration=_number(duration),
+                          crop=_crop(crop))
+        descriptor["data"] = add(_file_bytes(source, limit, allowed_roots))
+    elif type(value) is list_class:
+        videos, complete_audio = _attrs(value, ["videos", "complete_audio"])
+        if type(videos) is not list or not 1 <= len(videos) <= MAX_SEGMENTS:
+            raise ValueError("VIDEO must have between 1 and 64 segments")
+        if any(type(segment) is not file_class for segment in videos):
+            raise ValueError("VIDEO segments must be exact native VideoFromFile objects")
+        descriptor.update(representation="segments", segments=[])
+        for segment in videos:
+            nested, payload = encode_video(segment, max_bytes=limit - size, allowed_roots=allowed_roots)
+            descriptor["segments"].append(add(payload, nested))
+        descriptor["complete_audio"] = tensor("AUDIO", complete_audio) if complete_audio is not None else None
+    else:
+        raise ValueError("Unsupported native VIDEO class")
+    descriptor["byte_length"] = size
+    return descriptor, b"".join(chunks)
+
+
+def _tensor_header(kind, descriptor, length):
+    """Allocation-free preflight of the shared tensor-le-v1 descriptor.
+
+    The shared decoder still validates the bytes/finite values. Doing its small
+    structural check here first prevents allocation for an early part when a
+    later part's descriptor is already known to be invalid.
+    """
+    fields = {"encoding", "dtype", "shape", "layout"}
+    if kind == "AUDIO":
+        fields.add("sample_rate")
+    _fields(descriptor, fields, "tensor")
+    layout = {"IMAGE": "BHWC", "MASK": "BHW", "AUDIO": "BCT"}[kind]
+    if type(descriptor["encoding"]) is not str or descriptor["encoding"] != "tensor-le-v1":
+        raise ValueError("Invalid VIDEO tensor encoding")
+    if type(descriptor["layout"]) is not str or descriptor["layout"] != layout:
+        raise ValueError("Invalid VIDEO tensor layout")
+    dtype = descriptor["dtype"]
+    if type(dtype) is not str or dtype not in _TENSOR_BYTES:
+        raise ValueError("Invalid VIDEO tensor dtype")
+    shape = descriptor["shape"]
+    if type(shape) is not list or len(shape) != len(layout):
+        raise ValueError("Invalid VIDEO tensor rank")
+    size = _TENSOR_BYTES[dtype]
+    for dimension in shape:
+        _integer(dimension, positive=True)
+        if dimension > length // size:
+            raise ValueError("VIDEO tensor shape exceeds byte limit")
+        size *= dimension
+    if kind == "IMAGE" and shape[-1] not in (1, 3, 4):
+        raise ValueError("Invalid VIDEO image channels")
+    if size != length:
+        raise ValueError("VIDEO tensor length does not match shape")
+    if kind == "AUDIO":
+        _integer(descriptor["sample_rate"], positive=True)
+
+
+def _alpha_header(alpha, images):
+    _fields(alpha, {"offset", "length", "descriptor"}, "alpha part")
+    descriptor = alpha["descriptor"]
+    if type(descriptor) is not dict:
+        raise ValueError("Invalid VIDEO alpha descriptor")
+    layout = descriptor.get("layout")
+    if type(layout) is not str or layout not in ("BHW", "BHWC"):
+        raise ValueError("Invalid VIDEO alpha layout")
+    kind = "MASK" if layout == "BHW" else "IMAGE"
+    _tensor_header(kind, descriptor, _integer(alpha["length"]))
+    expected = images["descriptor"]["shape"][:3] + ([1] if kind == "IMAGE" else [])
+    if descriptor["shape"] != expected:
+        raise ValueError("VIDEO alpha shape must match images")
+    return kind
+
+
+def _validate_video(descriptor, size, *, file_only=False):
+    if type(descriptor) is not dict:
+        raise ValueError("VIDEO descriptor must be a dictionary")
+    representation = descriptor.get("representation")
+    if type(representation) is not str or representation not in ("components", "file", "segments"):
+        raise ValueError("Unsupported VIDEO representation")
+    if file_only and representation != "file":
+        # Depth is at most one: native lists already flatten their segments.
+        raise ValueError("VIDEO segments cannot contain nested lists or components")
+    fields = {"components": _COMPONENT_FIELDS, "file": _FILE_FIELDS, "segments": _SEGMENT_FIELDS}[representation]
+    _fields(descriptor, _BASE_FIELDS | fields, "descriptor")
+    if type(descriptor["encoding"]) is not str or descriptor["encoding"] != _ENCODING:
+        raise ValueError("Unsupported VIDEO encoding")
+    if _integer(descriptor["byte_length"]) != size:
+        raise ValueError("VIDEO payload length does not match descriptor")
+    result = dict(descriptor)
+    cursor = 0
+
+    def part(slot, kind=None, video=False):
+        nonlocal cursor
+        fields = {"offset", "length", "descriptor"} if kind or video else {"offset", "length"}
+        _fields(slot, fields, "part")
+        offset = _integer(slot["offset"])
+        length = _integer(slot["length"])
+        if offset != cursor or length > size - cursor:
+            raise ValueError("VIDEO parts must not overlap, leave gaps or exceed bounds")
+        cursor += length
+        if kind:
+            _tensor_header(kind, slot["descriptor"], length)
+        if video:
+            return {**slot, "descriptor": _validate_video(slot["descriptor"], length, file_only=True)}
+        return slot
+
+    if representation == "components":
+        _frame_rate(descriptor["frame_rate"])
+        _encoding_options(descriptor["bit_depth"], descriptor["color_space"])
+        result["metadata"] = _metadata(descriptor["metadata"])
+        part(descriptor["images"], "IMAGE")
+        if descriptor["alpha"] is not None:
+            part(descriptor["alpha"], _alpha_header(descriptor["alpha"], descriptor["images"]))
+        if descriptor["audio"] is not None:
+            part(descriptor["audio"], "AUDIO")
+    elif representation == "file":
+        _number(descriptor["start_time"])
+        _number(descriptor["duration"])
+        _crop(descriptor["crop"], wire=True)
+        part(descriptor["data"])
+    else:
+        segments = descriptor["segments"]
+        if type(segments) is not list or not 1 <= len(segments) <= MAX_SEGMENTS:
+            raise ValueError("VIDEO must have between 1 and 64 segments")
+        result["segments"] = [part(segment, video=True) for segment in segments]
+        if descriptor["complete_audio"] is not None:
+            part(descriptor["complete_audio"], "AUDIO")
+    if cursor != size:
+        raise ValueError("VIDEO payload contains trailing bytes")
+    return result
+
+
+def _decode_validated(descriptor, payload, types):
+    components_class, file_class, list_class, components_type = types
+
+    def data(slot):
+        return payload[slot["offset"]:slot["offset"] + slot["length"]]
+
+    def tensor(kind, slot):
+        if slot is None:
+            return None
+        return decode_payload(kind, slot["descriptor"], data(slot), max_bytes=slot["length"])
+
+    if descriptor["representation"] == "file":
+        # VideoFromFile retains the BytesIO, unlike a temporary filesystem path.
+        return file_class(io.BytesIO(data(descriptor["data"])), start_time=descriptor["start_time"],
+                          duration=descriptor["duration"], crop=_crop(descriptor["crop"], wire=True))
+    if descriptor["representation"] == "components":
+        fps = descriptor["frame_rate"]
+        return components_class(components_type(
+            images=tensor("IMAGE", descriptor["images"]),
+            frame_rate=Fraction(fps["numerator"], fps["denominator"]),
+            audio=tensor("AUDIO", descriptor["audio"]),
+            metadata=descriptor["metadata"],
+            alpha=tensor("IMAGE" if descriptor["alpha"] is not None and
+                         descriptor["alpha"]["descriptor"]["layout"] == "BHWC" else "MASK",
+                         descriptor["alpha"]),
+        ), bit_depth=descriptor["bit_depth"], color_space=descriptor["color_space"])
+    videos = [_decode_validated(slot["descriptor"], data(slot), types) for slot in descriptor["segments"]]
+    return list_class(videos, complete_audio=tensor("AUDIO", descriptor["complete_audio"]))
+
+
+def decode_video(descriptor, payload, *, max_bytes=MAX_PAYLOAD_BYTES):
+    """Decode a bounded VIDEO into independently owned native CPU media objects."""
+    limit = _limit(max_bytes)
+    if type(payload) is not bytes:
+        raise ValueError("VIDEO payload must be bytes")
+    if len(payload) > limit:
+        raise ValueError("VIDEO payload exceeds byte limit")
+    validated = _validate_video(descriptor, len(payload))
+    return _decode_validated(validated, payload, _native_types())

tests/api/test_config_routes.py

@@ -38,17 +38,10 @@ def _load_config_routes_module():
     sys.modules[f"{package_name}.api"] = api_pkg
 
     utils_pkg = types.ModuleType(f"{package_name}.utils")
-    utils_pkg.__path__ = []
+    utils_pkg.__path__ = [str(module_path.parents[1] / "utils")]
     sys.modules[f"{package_name}.utils"] = utils_pkg
 
-    created_aiohttp_stub = False
-    if "aiohttp" not in sys.modules:
-        created_aiohttp_stub = True
-        aiohttp_module = types.ModuleType("aiohttp")
-        aiohttp_module.web = types.SimpleNamespace(
-            json_response=lambda payload, status=200: _FakeResponse(payload, status=status)
-        )
-        sys.modules["aiohttp"] = aiohttp_module
+    import aiohttp  # Load the real dependency; isolate the fake response at the tested module.
 
     class _Routes:
         def get(self, _path):
@@ -96,8 +89,14 @@ def _load_config_routes_module():
     assert spec is not None and spec.loader is not None
     spec.loader.exec_module(module)
 
-    if created_aiohttp_stub:
-        sys.modules.pop("aiohttp", None)
+    class _EmptySecretStore:
+        def has(self, _worker_id):
+            return False
+
+    module.web = types.SimpleNamespace(
+        json_response=lambda payload, status=200: _FakeResponse(payload, status=status)
+    )
+    module._serverless_secret_store = lambda: _EmptySecretStore()
 
     return module
 

tests/api/test_distributed_queue.py

@@ -219,6 +219,10 @@ def _load_job_routes_module():
     module = importlib.util.module_from_spec(spec)
     assert spec is not None and spec.loader is not None
     spec.loader.exec_module(module)
+    # Keep response assertions independent of real aiohttp import ordering.
+    module.web = types.SimpleNamespace(
+        json_response=lambda payload, status=200: _FakeResponse(payload, status=status)
+    )
 
     if created_aiohttp_stub:
         sys.modules.pop("aiohttp", None)

tests/api/test_media_sync.py

@@ -243,6 +243,55 @@ class RewritePromptMediaInputsTests(unittest.TestCase):
         self.assertEqual(prompt["1"]["inputs"]["audio"], "song.wav")
 
 
+class AnnotatedMediaUploadTests(unittest.IsolatedAsyncioTestCase):
+    async def test_output_temp_and_input_uploads_do_not_collide(self):
+        from unittest.mock import AsyncMock, patch
+
+        class Response:
+            status = 200
+            def __init__(self, data):
+                self.data = data
+            async def __aenter__(self):
+                return self
+            async def __aexit__(self, *_args):
+                return None
+            async def json(self):
+                return self.data
+            def raise_for_status(self):
+                pass
+
+        class Form:
+            def __init__(self):
+                self.fields = {}
+            def add_field(self, key, value, **kwargs):
+                self.fields[key] = value
+                if key == 'image':
+                    self.fields['name'] = kwargs['filename']
+
+        checked = []
+        def post(url, *, json=None, data=None, **kwargs):
+            if url.endswith('/check_file'):
+                checked.append(json['filename'])
+                return Response({'exists': False})
+            self.assertEqual(data.fields['type'], 'input')
+            self.assertEqual(data.fields['image'], b'real file bytes')
+            return Response({'name': data.fields['name'], 'subfolder': data.fields['subfolder']})
+
+        refs = ['shared/ref.png', 'shared/ref.png [output]', 'shared/ref.png [temp]']
+        prompt = {str(i): {'inputs': {'image': ref}} for i, ref in enumerate(refs)}
+        self.assertEqual(ms._find_media_references(prompt), refs)
+        with patch.object(ms, 'get_client_session', AsyncMock(return_value=types.SimpleNamespace(post=post))), \
+             patch.object(ms, 'aiohttp', types.SimpleNamespace(FormData=Form, ClientTimeout=lambda **kw: kw)):
+            paths = [await ms._upload_media_to_worker({}, ref, b'real file bytes', 'fixture-hash', 'image/png')
+                     for ref in refs]
+        expected = ['shared/ref.png', 'distributed-media/output/fixture-hash/ref.png',
+                    'distributed-media/temp/fixture-hash/ref.png']
+        self.assertEqual(checked, expected)
+        self.assertEqual(paths, [(True, path) for path in expected])
+        ms._rewrite_prompt_media_inputs(prompt, dict(zip(refs, expected)))
+        self.assertEqual([node['inputs']['image'] for node in prompt.values()], expected)
+
+
 if __name__ == "__main__":
     unittest.main()
 

tests/api/test_transfer_integration.py

@@ -35,6 +35,9 @@ def _load_modules():
         return web.json_response({"error": str(error)}, status=status)
 
     job_routes.handle_api_error = handle_api_error
+    # The queue-test harness also fakes aiohttp.web for route modules; the
+    # endpoint must build real responses for the real server used here.
+    job_routes.web = web
     # The collector harness stubs aiohttp; this test needs the real client.
     sys.modules["aiohttp"] = real_aiohttp
     collector.aiohttp = real_aiohttp

tests/api/test_worker_routes.py

@@ -257,6 +257,8 @@ def _load_worker_routes_module():
     module = importlib.util.module_from_spec(spec)
     assert spec is not None and spec.loader is not None
     spec.loader.exec_module(module)
+    # Keep this module's response double independent of aiohttp import order.
+    module.web = types.SimpleNamespace(json_response=lambda payload, status=200: _FakeResponse(payload, status))
 
     if created_aiohttp_stub:
         sys.modules.pop("aiohttp", None)
@@ -269,6 +271,74 @@ def _load_worker_routes_module():
 worker_routes = _load_worker_routes_module()
 
 
+class WebsocketExecutionTargetTests(unittest.IsolatedAsyncioTestCase):
+    def setUp(self):
+        from test_async_helpers import _load_async_helpers_module
+        self.modules = patch.dict(sys.modules)
+        self.modules.start()
+        self.addCleanup(self.modules.stop)
+        self.helper, self.server = _load_async_helpers_module()
+        sys.modules['dist_api_worker_testpkg.utils.async_helpers'] = self.helper
+        self.prompt = {'9': {'class_type': 'Output'}, '10': {'class_type': 'Output'}}
+
+    async def receive(self, payload):
+        import json
+        sent = []
+        class Socket:
+            async def prepare(self, request):
+                pass
+            async def send_json(self, value):
+                sent.append(value)
+            def __aiter__(self):
+                async def messages():
+                    yield types.SimpleNamespace(type=worker_routes.aiohttp.WSMsgType.TEXT,
+                                                data=json.dumps(payload))
+                return messages()
+        with patch.object(worker_routes, 'web', types.SimpleNamespace(WebSocketResponse=lambda **kw: Socket())), \
+             patch.object(worker_routes, 'queue_prompt_payload', self.helper.queue_prompt_payload), \
+             patch.object(worker_routes, 'PromptValidationError', self.helper.PromptValidationError):
+            await worker_routes.worker_ws_endpoint(_FakeRequest())
+        return sent[0]
+
+    def payload(self, targets):
+        return {'type': 'dispatch_prompt', 'request_id': 'req', 'prompt': self.prompt,
+                'partial_execution_targets': targets}
+
+    async def test_explicit_targets_reach_queue_and_acknowledgement(self):
+        ack = await self.receive(self.payload(['9']))
+        self.assertTrue(ack['ok'], ack)
+        self.assertEqual(ack['request_id'], 'req')
+        self.assertEqual(self.server.prompt_queue.items[0][4], ['9'])
+
+    async def test_malformed_targets_and_messages_are_rejected(self):
+        for targets in (None, [], '9', [True], [9], ['absent'], ['9', '9']):
+            with self.subTest(targets=targets):
+                ack = await self.receive(self.payload(targets))
+                self.assertFalse(ack['ok'])
+                self.assertIn('partial_execution_targets', ack['error'])
+        self.assertFalse((await self.receive([]))['ok'])
+        self.assertEqual(self.server.prompt_queue.items, [])
+
+    async def test_validation_cannot_silently_queue_a_partial_subset(self):
+        ack = await self.receive(self.payload(['9', '10']))
+        self.assertFalse(ack['ok'])
+        self.assertEqual(ack['validation_error']['type'], 'partial_execution_targets_failed_validation')
+        self.assertEqual(self.server.prompt_queue.items, [])
+
+    async def test_missing_target_field_keeps_legacy_queue_call(self):
+        from unittest.mock import AsyncMock
+        queued = AsyncMock(return_value='legacy')
+        original = self.helper.queue_prompt_payload
+        self.helper.queue_prompt_payload = queued
+        try:
+            payload = self.payload(['9'])
+            payload.pop('partial_execution_targets')
+            self.assertTrue((await self.receive(payload))['ok'])
+            self.assertNotIn('partial_execution_targets', queued.call_args.kwargs)
+        finally:
+            self.helper.queue_prompt_payload = original
+
+
 class WorkerRoutesTests(unittest.IsolatedAsyncioTestCase):
     async def test_network_info_reports_actual_master_and_available_ports(self):
         with patch.object(worker_routes, "_get_cuda_info", return_value=(0, 1, 4)), \

tests/helpers/image_list_native_cache_identity.py

@@ -0,0 +1,172 @@
+"""Offline CPU proof against ComfyUI's real signatures and PromptExecutor.
+
+Exercises one native worker executor and separate private master-slot lifetimes,
+including their shared stable Start / fresh End contract.
+"""
+import asyncio
+import copy
+import importlib
+from pathlib import Path
+import sys
+import types
+
+ROOT = Path(__file__).resolve().parents[2]
+
+
+def main(core):
+    sys.path.insert(0, str(Path(core).resolve()))
+    import comfy.cli_args
+    comfy.cli_args.args.cpu = True  # Must precede all model-management imports.
+    comfy.cli_args.args.disable_assets = True
+    import execution
+    import nodes
+    import torch
+    from comfy_execution.caching import CacheKeySetInputSignature
+
+    name = '_native_cache_identity_probe'
+    for suffix in ('', '.image_list', '.nodes', '.utils'):
+        package = types.ModuleType(name + suffix)
+        package.__path__ = [str(ROOT / suffix.lstrip('.').replace('.', '/'))]
+        sys.modules[name + suffix] = package
+
+    def forbidden(*args, **kwargs):
+        raise AssertionError('offline native proof attempted a server-loop bridge')
+
+    sys.modules[name + '.utils.async_helpers'] = types.SimpleNamespace(run_async_in_server_loop=forbidden)
+    rt = importlib.import_module(name + '.image_list.runtime')
+    boundary = importlib.import_module(name + '.nodes.image_list')
+    calls = []
+
+    class Operation:
+        @classmethod
+        def INPUT_TYPES(cls):
+            return {'required': {'image': ('IMAGE',), 'mask': ('MASK',),
+                                 'text': ('STRING', {'forceInput': True})}}
+        RETURN_TYPES = ('IMAGE', 'MASK', 'STRING')
+        FUNCTION = 'run'
+
+        def run(self, image, mask, text):
+            calls.append(text)
+            return image + 1, mask + 1, text + ':done'
+
+    nodes.NODE_CLASS_MAPPINGS.update(DistributedImageListStart=boundary.DistributedImageListStart,
+                                    DistributedImageListEnd=boundary.DistributedImageListEnd,
+                                    Operation=Operation)
+    prompt = {
+        'start': {'class_type': rt.START_CLASS, 'inputs': {'image_list_execution_id': 'outer-A'}},
+        'op': {'class_type': 'Operation', 'inputs': {
+            'image': ['start', 0], 'mask': ['start', 1], 'text': ['start', 4]}},
+        'end': {'class_type': rt.END_CLASS, 'inputs': {
+            'images': ['op', 0], 'masks': ['op', 1], 'texts': ['op', 2],
+            'context': ['start', 3], 'value_3': ['start', 2]}},
+    }
+    data = rt.normalize_image_list_inputs(torch.zeros(1, 2, 3, 3, dtype=torch.float64),
+                                          torch.ones(1, 2, 3, dtype=torch.bfloat16),
+                                          {'value_1': 'original', 'value_2': 3})[0]
+    region = rt.validate_image_list_regions(prompt)[0]
+
+    def prepare(data):
+        token = rt.create_worker_task_token()
+        context = rt.create_master_job('start', 'end', prompt, [data], ['worker'], '')
+        rt.register_assignment(context, data['index'], 'worker', token)
+        identity = rt.store_worker_item(token, rt.encode_item(data))
+        result = rt.build_worker_prompt(prompt, region, token, context, data['index'], '', item_identity=identity)
+        result['start']['inputs']['image_list_local_execution'] = True
+        result['end']['inputs']['image_list_local_execution'] = True
+        return token, context, result
+
+    async def signature(graph, executor):
+        dynamic = execution.DynamicPrompt(graph)
+        changed = execution.IsChangedCache('signature-only', dynamic, executor.caches.outputs)
+        keys = CacheKeySetInputSignature(dynamic, graph.keys(), changed)
+        await keys.add_keys(graph.keys())
+        return {node: keys.get_data_key(node) for node in graph}
+
+    async def probe():
+        server = types.SimpleNamespace(client_id=None, last_node_id=None, send_sync=lambda *a, **kw: None)
+        executor = execution.PromptExecutor(server, cache_type=execution.CacheType.LRU,
+                                            cache_args={'lru': 64, 'ram': 0, 'ram_inactive': 0})
+        ta, ca, a = prepare(data)
+        prompt['start']['inputs']['image_list_execution_id'] = 'outer-B'
+        tb, cb, b = prepare(data)
+        assert ta != tb and ca != cb
+        ka, kb = await signature(a, executor), await signature(b, executor)
+        assert ka['start'] == kb['start'] and ka['op'] == kb['op']
+        assert ka['end'] != kb['end']
+        assert a['start']['inputs'] == b['start']['inputs']
+        # Exercise actual V3 list mapping, hidden.prompt/unique_id, lazy inputs,
+        # native node signatures and cache hits, not a reimplementation of them.
+        for number, (token, context, graph) in enumerate(((ta, ca, a), (tb, cb, b))):
+            valid = await execution.validate_prompt(f'native-{number}', graph, ['end'])
+            assert valid[0], valid
+            await executor.execute_async(graph, f'native-{number}', {}, ['end'])
+            assert executor.success, executor.status_messages
+            assert token not in rt._WORKER_ITEMS
+            result = rt.get_master_job(context).results[0]
+            assert torch.equal(result['image'], data['image'] + 1)
+            assert torch.equal(result['mask'], data['mask'] + 1)
+            assert result['text'] == 'original:done'
+            assert result['values'] == {'value_3': 0}
+            cached = next(entry['nodes'] for event, entry in executor.status_messages if event == 'execution_cached')
+            if number:
+                assert 'start' in cached and 'op' in cached and 'end' not in cached, cached
+            assert calls == ['original'], calls
+        for failure in ('expired', 'mismatched'):
+            token, context, graph = prepare(data)
+            expires, staged = rt._WORKER_ITEMS[token]
+            if failure == 'expired':
+                rt._WORKER_ITEMS[token] = (0, staged)
+            else:
+                staged['image'] = staged['image'] + 9
+            await executor.execute_async(graph, f'rejected-{failure}', {}, ['end'])
+            assert not executor.success, failure
+            assert not rt.get_master_job(context).received_indexes
+            assert token not in rt._WORKER_ITEMS
+            cached = next(entry['nodes'] for event, entry in executor.status_messages if event == 'execution_cached')
+            assert 'start' in cached and 'op' in cached and 'end' not in cached, cached
+            assert calls == ['original'], calls
+        for part in ('image', 'mask', 'value', 'index'):
+            changed = copy.deepcopy(data)
+            if part in ('image', 'mask'):
+                changed[part].reshape(-1)[0] += 1
+            elif part == 'value':
+                changed['values']['value_2'] = 4  # Even an unconnected slot is content.
+            else:
+                changed['index'] = 1
+            token, context, graph = prepare(changed)
+            keys = await signature(graph, executor)
+            assert keys['start'] != ka['start'] and keys['op'] != ka['op'], part
+            await executor.execute_async(graph, f'changed-{part}', {}, ['end'])
+            assert executor.success, executor.status_messages
+            assert token not in rt._WORKER_ITEMS
+            assert rt.get_master_job(context).received_indexes == {changed['index']}
+        assert len(calls) == 5, calls
+        # Combine the real Start/End identity path with cache transfer between
+        # private master slots, not just independent signature/retention probes.
+        local = importlib.import_module(name + '.image_list.local_executor')
+        before = len(calls)
+        for number in range(2):
+            token, context, graph = prepare(data)
+            slot = local.LocalExecutionSlot()
+            try:
+                await asyncio.wrap_future(slot.submit(graph, f'local-{number}', 'end'))
+                result = rt.get_master_job(context).results[0]
+                assert torch.equal(result['image'], data['image'] + 1)
+                assert result['values'] == {'value_3': 0}
+                assert token not in rt._WORKER_ITEMS
+            finally:
+                await slot.close()
+                rt.remove_master_job(context)
+            assert local._RETAINED_OUTPUTS is not None
+        assert len(calls) == before + 1, calls
+        assert not rt._WORKER_ITEMS
+        assert not torch.cuda.is_initialized()
+        print('NATIVE_CACHE_IDENTITY_OK stable Start/op keys; fresh End key; '
+              '1 operation for 2 same-data jobs; 4 changed-content misses; '
+              'expired/mismatched cached inputs rejected; current delivery and cleanup')
+
+    asyncio.run(probe())
+
+
+if __name__ == '__main__':
+    main(sys.argv[1])

tests/helpers/image_list_network_check.py

@@ -0,0 +1,688 @@
+"""Model-free multi-process acceptance using real ComfyUI executors and loopback HTTP.
+
+Isolated listeners use real core /prompt, /queue, /history and /interrupt handlers.
+Start/End, graph extraction, scheduling, payload codecs and callbacks are production code.
+No managed ComfyUI service, user workflow, model, or remote worker is touched.
+"""
+import asyncio
+from collections import Counter
+import copy
+import importlib
+import json
+import os
+from pathlib import Path
+import sys
+import tempfile
+import time
+import uuid
+from unittest.mock import patch
+
+ROOT = Path(__file__).resolve().parents[2]
+COMFY_ROOT = Path(sys.argv[1]).resolve()
+WORKER = '--worker' in sys.argv
+DELAY = float(sys.argv[3]) if WORKER else 0
+sys.path.insert(0, str(COMFY_ROOT))
+os.environ['COMFYUI_IS_WORKER'] = '1'
+import comfy.cli_args
+comfy.cli_args.args.cpu = True
+comfy.cli_args.args.disable_assets = True
+from app.assets.manager import default_asset_manager
+import comfy.model_management
+import execution
+import nodes
+import server
+import torch
+from aiohttp import web
+import aiohttp
+
+torch.set_num_threads(1)
+SOURCE_CALLS, BRANCH_CALLS, CAPTURED, MODEL_CALLS = [], [], [], []
+UTILITY_CALLS, ITEM_EXECUTIONS, VALUE_CAPTURES = [], [], []
+ALLOW_MASTER = False
+COUNT = 7
+
+
+def source_rows():
+    return [{'image': torch.full((1, 5 + i, 7, 3 if i % 2 == 0 else 4), i / 16, dtype=torch.float32),
+             'reference_image': torch.full((2, 2, 3, 3), (i + 1) / 16),
+             'mask': torch.full((1, 5 + i, 7), (i + 1) / 16),
+             'text': f'item-{i}', 'strength': (i + 1) / 16, 'seed': 100 + i, 'index': i}
+            for i in range(COUNT)]
+
+
+class TestImageSource:
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {'required': {}}
+    RETURN_TYPES = ('IMAGE', 'IMAGE', 'MASK', 'STRING', 'FLOAT', 'INT')
+    OUTPUT_IS_LIST = (True,) * 6
+    FUNCTION = 'run'
+    def run(self):
+        assert not WORKER, 'upstream image source leaked into worker prompt'
+        SOURCE_CALLS.append(True)
+        rows = source_rows()
+        return tuple([row[key] for row in rows] for key in
+                     ('image', 'reference_image', 'mask', 'text', 'strength', 'seed'))
+
+
+class TestSharedFlag:
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {'required': {}}
+    RETURN_TYPES = ('BOOLEAN',)
+    FUNCTION = 'run'
+    def run(self):
+        assert not WORKER, 'shared value source leaked into worker prompt'
+        return (True,)
+
+
+class TestCachedModel:
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {'required': {}}
+    RETURN_TYPES = ('MODEL',)
+    FUNCTION = 'run'
+    def run(self):
+        MODEL_CALLS.append(True)
+        return (object(),)
+
+
+def assert_item_execution(prompt, index):
+    """Prove the operation belongs to an item, not merely a participating host."""
+    from comfy_execution.utils import get_executing_context
+    context = get_executing_context()
+    assert context is not None
+    assert prompt['4']['inputs'].get('worker_result_token'), 'per-item node ran in outer graph'
+    assert (prompt['2']['inputs'].get('worker_item_identity')
+            or prompt['2']['inputs'].get('worker_item_token')), 'per-item source binding missing'
+    running, _ = server.PromptServer.instance.prompt_queue.get_current_queue_volatile()
+    assert len(running) == 1, running
+    entry = running[0]
+    if WORKER:
+        assert entry[1] == context.prompt_id and entry[2] == prompt
+        assert entry[4] == ['4'], entry[4]
+    else:
+        assert ALLOW_MASTER, 'per-item node ran on delegate-only master'
+        assert entry[1] != context.prompt_id, 'per-item node ran in outer executor'
+        assert set(entry[4]) == {'4', '5'}, entry[4]
+    return {'index': index, 'prompt_id': context.prompt_id,
+            'role': 'worker' if WORKER else 'master-item', 'queue_prompt_id': entry[1]}
+
+
+class TestRequiredOutputUtility:
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {'required': {'image': ('IMAGE',), 'index': ('INT', {'forceInput': True})},
+                'hidden': {'prompt': 'PROMPT'}}
+    OUTPUT_NODE = True
+    RETURN_TYPES = ('IMAGE',)
+    FUNCTION = 'run'
+    def run(self, image, index, prompt):
+        assert_item_execution(prompt, index)
+        UTILITY_CALLS.append(index)
+        return (image,)
+
+
+class TestDiscardedPreview:
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {'required': {'image': ('IMAGE',)}}
+    OUTPUT_NODE = True
+    RETURN_TYPES = ()
+    FUNCTION = 'run'
+    def run(self, image):
+        raise AssertionError('discarded per-item output preview executed')
+
+
+class TestImageOperation:
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {'required': {'image': ('IMAGE',), 'reference_image': ('IMAGE',), 'mask': ('MASK',),
+                             'text': ('STRING', {'forceInput': True}),
+                             'strength': ('FLOAT', {'forceInput': True}),
+                             'seed': ('INT', {'forceInput': True}), 'index': ('INT', {'forceInput': True}),
+                             'flag': ('BOOLEAN', {'forceInput': True}), 'model': ('MODEL',)},
+                'hidden': {'prompt': 'PROMPT'}}
+    RETURN_TYPES = ('IMAGE', 'MASK', 'STRING')
+    FUNCTION = 'run'
+    def run(self, image, reference_image, mask, text, strength, seed, index, flag, model, prompt):
+        assert WORKER or ALLOW_MASTER, 'per-image branch executed in orchestrator-only mode'
+        ITEM_EXECUTIONS.append(assert_item_execution(prompt, index))
+        BRANCH_CALLS.append(index)
+        # Sleep in short chunks so the real Comfy interrupt signal can be tested.
+        until = time.monotonic() + DELAY
+        while time.monotonic() < until:
+            comfy.model_management.throw_exception_if_processing_interrupted()
+            time.sleep(0.01)
+        assert seed == 100 + index
+        assert flag is True
+        assert text == f'item-{index}'
+        assert strength == (index + 1) / 16
+        assert torch.all(reference_image == strength)
+        return image + strength, mask * 0.5, f'{text}:seed-{seed}'
+
+
+class TestValueOverride:
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {'required': {'text': ('STRING', {'forceInput': True}),
+                             'index': ('INT', {'forceInput': True})}}
+    RETURN_TYPES = ('STRING', 'INT')
+    FUNCTION = 'run'
+    def run(self, text, index):
+        assert WORKER or ALLOW_MASTER, 'override branch executed on delegate-only master'
+        return (text + ':override', 200 + index)
+
+
+class TestCapture:
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {'required': {'images': ('IMAGE',), 'masks': ('MASK',), 'texts': ('STRING', {'forceInput': True})},
+                'optional': {'values': ('DISTRIBUTED_IMAGE_LIST_VALUES',),
+                             'picked': ('STRING', {'forceInput': True})}}
+    INPUT_IS_LIST = True
+    OUTPUT_NODE = True
+    RETURN_TYPES = ()
+    FUNCTION = 'run'
+    def run(self, images, masks, texts, values=None, picked=None):
+        assert not WORKER, 'post-End capture leaked into a worker prompt'
+        CAPTURED.append((images, masks, texts))
+        VALUE_CAPTURES.append((values, picked))
+        return ()
+
+
+async def setup():
+    assets = default_asset_manager()
+    ps = server.PromptServer(asyncio.get_running_loop(), assets)
+    assert await nodes.load_custom_node(str(ROOT))
+    module = sys.modules[str(ROOT).replace('.', '_x_')]
+    boundary = importlib.import_module(module.__name__ + '.nodes.image_list')
+    encode_result = boundary.encode_result
+    def checked_encode(result):
+        encoded = encode_result(result)
+        if 'values' in encoded:
+            assert set(encoded['values']) == {f'value_{n}' for n in (2, 3, 4, 5, 7, 8, 9, 10)}
+        return encoded
+    boundary.encode_result = checked_encode
+    nodes.NODE_CLASS_MAPPINGS.update(TestImageSource=TestImageSource,
+                                   TestImageOperation=TestImageOperation, TestCapture=TestCapture,
+                                   TestCachedModel=TestCachedModel, TestSharedFlag=TestSharedFlag,
+                                   TestRequiredOutputUtility=TestRequiredOutputUtility,
+                                   TestDiscardedPreview=TestDiscardedPreview, TestValueOverride=TestValueOverride)
+    # Match the CORS-enabled direct-worker launch contract for canvas probes.
+    app = web.Application(middlewares=[server.create_cors_middleware('*')])
+    # Register the actual feature routes, not test replacements.
+    app.add_routes([route for route in ps.routes if route.path.startswith('/distributed/image-list/')])
+    return ps, assets, module, app
+
+
+async def listen(app):
+    runner = web.AppRunner(app)
+    await runner.setup()
+    site = web.TCPSite(runner, '127.0.0.1', 0)
+    await site.start()
+    return runner, site._server.sockets[0].getsockname()[1]
+
+
+def finish_queue_item(ps, item_id, executor):
+    ps.prompt_queue.task_done(item_id, executor.history_result,
+        status=execution.PromptQueue.ExecutionStatus(
+            status_str='success' if executor.success else 'error',
+            completed=executor.success, messages=executor.status_messages),
+        process_item=lambda item: item[:5] + item[6:])
+
+
+async def worker_main():
+    ps, assets, module, app = await setup()
+    wake = asyncio.Event()
+    interrupted_ids, submissions = [], []
+    core_handlers = {(route.method, route.path): route.handler for route in ps.routes}
+    async def prompt(request):
+        data = await request.json()
+        # Use core's actual validation/enqueue/response schema, including
+        # partial_execution_targets; only observe it and wake our CPU consumer.
+        response = await core_handlers['POST', '/prompt'](request)
+        if response.status == 200:
+            accepted = json.loads(response.body)
+            _, pending = ps.prompt_queue.get_current_queue_volatile()
+            entry = next(item for item in pending if item[1] == accepted['prompt_id'])
+            submissions.append({'prompt_id': entry[1], 'requested': data.get('partial_execution_targets'),
+                                'queued': entry[4], 'node_ids': sorted(entry[2]),
+                                'end_ids': [key for key, node in entry[2].items()
+                                            if node['class_type'] == 'DistributedImageListEnd']})
+            wake.set()
+        return response
+    async def interrupt(request):
+        interrupted_ids.append((await request.json())['prompt_id'])
+        return await core_handlers['POST', '/interrupt'](request)
+    async def stats(request):
+        running, pending = ps.prompt_queue.get_current_queue_volatile()
+        return web.json_response({'calls': BRANCH_CALLS, 'utility_calls': UTILITY_CALLS,
+                                  'item_executions': ITEM_EXECUTIONS, 'submissions': submissions,
+                                  'running': [e[1] for e in running], 'pending': [e[1] for e in pending],
+                                  'interrupted': interrupted_ids, 'history': ps.prompt_queue.get_history()})
+    async def consume():
+        executor = execution.PromptExecutor(ps, cache_args={'ram': 0, 'ram_inactive': 0}, asset_manager=assets)
+        while True:
+            await wake.wait()
+            while (queued := ps.prompt_queue.get(timeout=0)) is not None:
+                entry, item_id = queued
+                await asyncio.to_thread(executor.execute, entry[2], entry[1], entry[3], entry[4])
+                finish_queue_item(ps, item_id, executor)
+            wake.clear()
+    app.add_routes([route for route in ps.routes if
+                    (route.path in ('/queue', '/history/{prompt_id}') or
+                     (route.path == '/prompt' and route.method == 'GET'))])
+    app.router.add_post('/prompt', prompt)
+    app.router.add_post('/interrupt', interrupt)
+    app.router.add_get('/test-stats', stats)
+    runner, port = await listen(app)
+    consumer = asyncio.create_task(consume())
+    print(f'READY:{port}', flush=True)
+    try:
+        await asyncio.Event().wait()
+    finally:
+        consumer.cancel()
+        await asyncio.gather(consumer, return_exceptions=True)
+        await runner.cleanup()
+
+
+def graph(callback, worker_ids):
+    return {
+        '1': {'class_type': 'TestImageSource', 'inputs': {}},
+        # Generic Autogrow value slots carry per-item reference images, text,
+        # strength and seed (one value per image), plus a shared boolean.
+        '2': {'class_type': 'DistributedImageListStart', 'inputs': {
+            'images': ['1', 0], 'mask': ['1', 2],
+            'values.value_1': ['1', 1], 'values.value_2': ['1', 3],
+            'values.value_3': ['1', 4], 'values.value_4': ['1', 5], 'values.value_5': ['7', 0],
+            'values.value_6': ['1', 2],
+            'image_list_enabled_worker_ids': json.dumps(worker_ids), 'image_list_master_callback': callback}},
+        '3': {'class_type': 'TestImageOperation', 'inputs': {
+            key: ['2', index] for key, index in [('image', 0), ('mask', 1), ('index', 2),
+                ('reference_image', 4), ('text', 5), ('strength', 6), ('seed', 7), ('flag', 8)]}
+              | {'image': ['8', 0], 'model': ['6', 0]}},
+        '4': {'class_type': 'DistributedImageListEnd', 'inputs': {
+            'images': ['3', 0], 'masks': ['3', 1], 'texts': ['3', 2], 'context': ['2', 3],
+            'value_2': ['10', 0], 'value_3': ['2', 6], 'value_4': ['2', 7], 'value_5': ['2', 8],
+            'value_7': ['3', 1], 'value_8': ['2', 4], 'value_9': ['2', 9], 'value_10': ['10', 1]}},
+        '5': {'class_type': 'TestCapture', 'inputs': {'images': ['4', 0], 'masks': ['4', 1], 'texts': ['4', 2],
+                                                     'values': ['4', 3], 'picked': ['get', 0]}},
+        '6': {'class_type': 'TestCachedModel', 'inputs': {}},
+        '7': {'class_type': 'TestSharedFlag', 'inputs': {}},
+        '8': {'class_type': 'TestRequiredOutputUtility', 'inputs': {'image': ['2', 0], 'index': ['2', 2]}},
+        '9': {'class_type': 'TestDiscardedPreview', 'inputs': {'image': ['8', 0]}},
+        '10': {'class_type': 'TestValueOverride', 'inputs': {'text': ['2', 5], 'index': ['2', 2]}},
+        'get': {'class_type': 'DistributedImageListGetValue', 'inputs': {'values': ['4', 3], 'value_name': 'value_2'}},
+    }
+
+
+def check_carried_values():
+    values, picked = VALUE_CAPTURES[-1]
+    assert len(values) == len(picked) == COUNT
+    for index, (actual, source) in enumerate(zip(values, source_rows())):
+        assert set(actual) == {f'value_{n}' for n in range(1, 11)}, actual.keys()
+        assert torch.equal(actual['value_1'], source['reference_image'])
+        assert actual['value_2'] == picked[index] == f'item-{index}:override'
+        assert actual['value_3'] == source['strength']
+        assert actual['value_4'] == source['seed']
+        assert actual['value_5'] is True
+        assert torch.equal(actual['value_6'], source['mask'])
+        assert torch.equal(actual['value_7'], source['mask'] * 0.5)
+        assert torch.equal(actual['value_8'], source['reference_image'])
+        assert actual['value_8'].shape[0] == 2, 'value tensor batch was split'
+        assert torch.equal(actual['value_9'], source['mask'])
+        assert actual['value_10'] == 200 + index
+
+
+async def coordinator():
+    global ALLOW_MASTER
+    ps, assets, module, app = await setup()
+    runtime = importlib.import_module(module.__name__ + '.image_list.runtime')
+    execution_plan = importlib.import_module(module.__name__ + '.image_list.execution_plan')
+    lifecycle = importlib.import_module(module.__name__ + '.image_list.lifecycle')
+    config_mod = importlib.import_module(module.__name__ + '.utils.config')
+    orchestration = importlib.import_module(module.__name__ + '.api.queue_orchestration')
+    runner, port = await listen(app)
+    ps.port = port  # Match the listener used by local-worker callback resolution.
+    processes, stderr_files = [], []
+    # Old worker capability refusal is read-only and precedes input evaluation.
+    legacy_app = web.Application()
+    legacy_requests = []
+    async def legacy(request):
+        legacy_requests.append(request.method)
+        return web.json_response({'error': 'not found'}, status=404)
+    legacy_app.router.add_route('*', '/{path:.*}', legacy)
+    legacy_runner, legacy_port = await listen(legacy_app)
+    dispatch_module = importlib.import_module(module.__name__ + '.api.orchestration.dispatch')
+    try:
+        try:
+            await dispatch_module.preflight_image_list_workers([
+                {'id': 'legacy', 'type': 'local', 'host': '127.0.0.1', 'port': legacy_port}])
+            raise AssertionError('legacy worker accepted')
+        except RuntimeError as exc:
+            assert 'upgrade' in str(exc), str(exc)
+        assert legacy_requests == ['GET'] and not SOURCE_CALLS
+    finally:
+        await legacy_runner.cleanup()
+    print('IMAGE_LIST_PREFLIGHT_OK legacy worker rejected before materialization or writes')
+    scratch = Path(os.environ.get('TMPDIR', Path.home() / '.hermes/cache/scratch'))
+    scratch.mkdir(parents=True, exist_ok=True)
+    try:
+        with tempfile.TemporaryDirectory(prefix='image-list-network-', dir=scratch) as tmp:
+            workers = []
+            for name, delay in [('fast', 0.02), ('slow', 0.5)]:
+                err = open(Path(tmp) / f'{name}.log', 'w+')
+                stderr_files.append(err)
+                proc = await asyncio.create_subprocess_exec(sys.executable, str(Path(__file__).resolve()),
+                    str(COMFY_ROOT), '--worker', str(delay), stdout=asyncio.subprocess.PIPE, stderr=err,
+                    env={**os.environ, 'TMPDIR': tmp})
+                processes.append(proc)
+                async def ready():
+                    while True:
+                        line = await proc.stdout.readline()
+                        if not line:
+                            err.seek(0)
+                            raise AssertionError(f'Worker exited: {err.read()}')
+                        if line.startswith(b'READY:'):
+                            return int(line.split(b':')[1])
+                worker_port = await asyncio.wait_for(ready(), 60)
+                workers.append({'id': name, 'type': 'local', 'host': '127.0.0.1',
+                                'port': worker_port, 'enabled': True})
+            config = {'workers': workers, 'settings': {'websocket_orchestration': False, 'master_delegate_only': True},
+                      'master': {'host': f'http://127.0.0.1:{port}'}}
+            prompt = graph(f'http://127.0.0.1:{port}', [w['id'] for w in workers])
+            valid = await execution.validate_prompt('network-acceptance', prompt, None)
+            assert valid[0], valid
+            assert {'4', '5', '8', '9'} == set(valid[2]), valid
+            executor = execution.PromptExecutor(ps, cache_args={'ram': 0, 'ram_inactive': 0}, asset_manager=assets)
+            # Exercise production queue preparation; only master enqueue is redirected
+            # to this isolated executor, not the managed server's live queue.
+            prompt['2']['inputs'].pop('image_list_enabled_worker_ids')
+            prompt['2']['inputs'].pop('image_list_master_callback')
+            saved_prompt = copy.deepcopy(prompt)
+            outer_executions = []
+            async def enqueue(prepared, workflow, client_id, **kwargs):
+                targets = kwargs['partial_execution_targets']
+                planned, planned_targets = execution_plan.prepare_image_list_execution(prepared)
+                assert planned == prepared and targets == planned_targets
+                if '9' in saved_prompt and 'TestImageOperation' in {
+                        node['class_type'] for node in prepared.values()}:
+                    assert set(prepared) == set(saved_prompt) - {'9'}, prepared
+                    assert prepared['8'] == saved_prompt['8'], 'required utility changed'
+                    assert prepared['5'] == saved_prompt['5'], 'post-End output changed'
+                    assert set(targets) == {'4', '5'}, targets
+                    assert nodes.NODE_CLASS_MAPPINGS['TestRequiredOutputUtility'].OUTPUT_NODE is True
+                    assert nodes.NODE_CLASS_MAPPINGS['TestDiscardedPreview'].OUTPUT_NODE is True
+                pid = 'network-acceptance' if not outer_executions else str(uuid.uuid4())
+                checked = await execution.validate_prompt(pid, prepared, targets)
+                assert checked[0], checked
+                assert not checked[3] and Counter(checked[2]) == Counter(targets), checked
+                assert not lifecycle.has_safe_execution_targets(ps, pid, prepared)
+                # Exactly the core queue tuple/put/get/task_done lifecycle. Start
+                # must prove its roots from this real running record, not a mock.
+                number = len(outer_executions)
+                ps.prompt_queue.put((number, pid, prepared, {'client_id': client_id}, checked[2], {}))
+                entry, item_id = ps.prompt_queue.get(timeout=0)
+                assert lifecycle.has_safe_execution_targets(ps, pid, prepared)
+                before_captures = len(CAPTURED)
+                await asyncio.to_thread(executor.execute, entry[2], entry[1], entry[3], entry[4])
+                finish_queue_item(ps, item_id, executor)
+                assert ps.prompt_queue.get_current_queue_volatile() == ([], [])
+                history = ps.prompt_queue.get_history(pid)[pid]
+                assert Counter(history['prompt'][4]) == Counter(targets)
+                assert not lifecycle.has_safe_execution_targets(ps, pid, prepared)
+                outer_executions.append({'prompt_id': pid, 'targets': targets,
+                                         'captures': len(CAPTURED) - before_captures})
+                if executor.success:
+                    assert len(CAPTURED) == before_captures + 1, 'post-End Capture must run exactly once'
+                return {'prompt_id': pid, 'number': number, 'node_errors': checked[3]}
+            with patch.object(config_mod, 'load_config', return_value=config), \
+                 patch.object(orchestration, 'load_config', return_value=config), \
+                 patch.object(orchestration, 'queue_prompt_payload', enqueue), \
+                 patch.object(config_mod, 'get_worker_timeout_seconds', return_value=15):
+                queued = await asyncio.wait_for(orchestration.orchestrate_distributed_execution(
+                    prompt, None, 'test', [w['id'] for w in workers]), 45)
+                assert queued[0] == 'network-acceptance' and queued[2] == 2, queued
+                assert not ps.distributed_pending_jobs, 'image-list created unused Collector queues'
+            assert prompt == saved_prompt, 'queue preparation changed the saved/source graph'
+            if not executor.success:
+                for err in stderr_files:
+                    err.flush(); err.seek(0); print(err.read(), file=sys.stderr)
+            assert executor.success, executor.status_messages
+            assert len(SOURCE_CALLS) == 1 and not BRANCH_CALLS and not UTILITY_CALLS
+            assert len(CAPTURED) == 1
+            check_carried_values()
+            images, masks, texts = CAPTURED[0]
+            assert len(images) == len(masks) == len(texts) == COUNT
+            for i, row in enumerate(source_rows()):
+                assert torch.equal(images[i], row['image'] + row['strength']), (i, images[i])
+                assert torch.equal(masks[i], row['mask'] * 0.5), i
+                assert texts[i] == f'item-{i}:seed-{100 + i}', texts
+            async with aiohttp.ClientSession() as session:
+                stats = []
+                for worker in workers:
+                    async with session.get(f"http://127.0.0.1:{worker['port']}/test-stats") as response:
+                        stats.append(await response.json())
+                counts = [len(row['calls']) for row in stats]
+                assert counts[0] > counts[1] > 0, counts
+                assert Counter(i for row in stats for i in row['calls']) == Counter(range(COUNT))
+                for row in stats:
+                    assert Counter(row['utility_calls']) == Counter(row['calls']), row
+                    assert len(row['submissions']) == len(row['calls'])
+                    assert len({event['prompt_id'] for event in row['item_executions']}) == len(row['calls'])
+                    for submission in row['submissions']:
+                        assert submission['requested'] == submission['queued'] == ['4'], submission
+                        assert '8' in submission['node_ids'] and '9' not in submission['node_ids']
+                        assert '5' not in submission['node_ids']
+            assert not runtime._MASTER_JOBS
+            # Three sequential regions on one worker, including optional outputs
+            # left unconnected. Region N+1 must not race the prior worker End.
+            chained = {'1': {'class_type': 'EmptyImage', 'inputs': {
+                'height': 8, 'width': 8, 'batch_size': 2, 'color': 0}}}
+            chained['constant'] = {'class_type': 'TestSharedFlag', 'inputs': {}}
+            previous = '1'
+            value_source = 'constant'
+            for start, end in [('2', '3'), ('4', '5'), ('6', '7')]:
+                chained[start] = {'class_type': 'DistributedImageListStart', 'inputs': {
+                    'images': [previous, 0], 'values.value_1': [value_source, 0]}}
+                chained[end] = {'class_type': 'DistributedImageListEnd', 'inputs': {
+                    'images': [start, 0], 'context': [start, 3]}}
+                value_source = f'get-{end}'
+                chained[value_source] = {'class_type': 'DistributedImageListGetValue', 'inputs': {
+                    'values': [end, 3], 'value_name': 'value_1'}}
+                previous = end
+            chained['8'] = {'class_type': 'TestCapture', 'inputs': {
+                'images': ['7', 0], 'masks': ['7', 1], 'texts': ['7', 2], 'values': ['7', 3]}}
+            saved_chained = copy.deepcopy(chained)
+            with patch.object(config_mod, 'load_config', return_value=config), \
+                 patch.object(orchestration, 'load_config', return_value=config), \
+                 patch.object(orchestration, 'queue_prompt_payload', enqueue), \
+                 patch.object(config_mod, 'get_worker_timeout_seconds', return_value=15):
+                await asyncio.wait_for(orchestration.orchestrate_distributed_execution(
+                    chained, None, 'test', ['slow']), 45)
+            assert chained == saved_chained
+            assert executor.success, executor.status_messages
+            assert len(CAPTURED) == 2 and len(CAPTURED[1][0]) == 2
+            assert all(torch.equal(image, torch.zeros((1, 8, 8, 3))) for image in CAPTURED[1][0])
+            assert CAPTURED[1][1:] == ([None, None], [None, None])
+            assert VALUE_CAPTURES[1][0] == [{'value_1': True}, {'value_1': True}]
+            assert not runtime._MASTER_JOBS
+            print('IMAGE_LIST_SEQUENTIAL_OK three regions; one worker; optional outputs omitted')
+            # Same public queue path, now with master participation enabled. The
+            # parent queue stays occupied while its local slot processes items.
+            ALLOW_MASTER = True
+            config['settings']['master_delegate_only'] = False
+            participation_counts = {}
+            for mode, selected in [('mixed', [w['id'] for w in workers]), ('master-only', [])]:
+                # Keep cold-run exact-once assertions independent of other modes.
+                importlib.import_module(module.__name__ + '.image_list.local_executor')._RETAINED_OUTPUTS = None
+                executor = execution.PromptExecutor(ps, cache_args={'ram': 0, 'ram_inactive': 0}, asset_manager=assets)
+                before_sources, before_local, before_models = len(SOURCE_CALLS), len(BRANCH_CALLS), len(MODEL_CALLS)
+                before_utilities, before_events = len(UTILITY_CALLS), len(ITEM_EXECUTIONS)
+                async with aiohttp.ClientSession() as session:
+                    before_worker_stats = []
+                    for worker in workers:
+                        async with session.get(f"http://127.0.0.1:{worker['port']}/test-stats") as response:
+                            before_worker_stats.append(await response.json())
+                with patch.object(config_mod, 'load_config', return_value=config), \
+                     patch.object(orchestration, 'load_config', return_value=config), \
+                     patch.object(orchestration, 'queue_prompt_payload', enqueue), \
+                     patch.object(config_mod, 'get_worker_timeout_seconds', return_value=15):
+                    queued = await asyncio.wait_for(orchestration.orchestrate_distributed_execution(
+                        prompt, None, 'test', selected), 45)
+                assert prompt == saved_prompt, (mode, 'saved/source graph changed')
+                assert executor.success, (mode, executor.status_messages)
+                assert queued[2] == len(selected)
+                assert len(SOURCE_CALLS) == before_sources + 1, 'source was reevaluated per local item'
+                local_calls = BRANCH_CALLS[before_local:]
+                assert Counter(UTILITY_CALLS[before_utilities:]) == Counter(local_calls), mode
+                assert len({event['prompt_id'] for event in ITEM_EXECUTIONS[before_events:]}) == len(local_calls)
+                assert local_calls, f'{mode}: master did not process any items'
+                assert len(MODEL_CALLS) == before_models + 1, 'local model dependency did not stay cached'
+                remote_calls = []
+                async with aiohttp.ClientSession() as session:
+                    for worker, before_stats in zip(workers, before_worker_stats):
+                        async with session.get(f"http://127.0.0.1:{worker['port']}/test-stats") as response:
+                            current = await response.json()
+                        calls = current['calls'][len(before_stats['calls']):]
+                        utilities = current['utility_calls'][len(before_stats['utility_calls']):]
+                        submissions = current['submissions'][len(before_stats['submissions']):]
+                        events = current['item_executions'][len(before_stats['item_executions']):]
+                        assert Counter(utilities) == Counter(calls), (mode, current)
+                        assert len(submissions) == len(calls) == len({event['prompt_id'] for event in events})
+                        for submission in submissions:
+                            assert submission['requested'] == submission['queued'] == ['4'], submission
+                            assert '8' in submission['node_ids'] and '9' not in submission['node_ids']
+                            assert '5' not in submission['node_ids']
+                        remote_calls.extend(calls)
+                        assert not current['running'] and not current['pending'], current
+                if mode == 'mixed':
+                    assert remote_calls, 'mixed execution failed to use direct workers'
+                else:
+                    assert not remote_calls and len(local_calls) == COUNT
+                assert Counter(local_calls + remote_calls) == Counter(range(COUNT)), (local_calls, remote_calls)
+                check_carried_values()
+                images, masks, texts = CAPTURED[-1]
+                assert len(images) == len(masks) == len(texts) == COUNT
+                for i, row in enumerate(source_rows()):
+                    assert torch.equal(images[i], row['image'] + row['strength']), (mode, i)
+                    assert torch.equal(masks[i], row['mask'] * 0.5), (mode, i)
+                    assert texts[i] == f'item-{i}:seed-{100 + i}', (mode, texts)
+                assert not runtime._MASTER_JOBS and not runtime._WORKER_ITEMS
+                participation_counts[mode] = {'master': len(local_calls), 'workers': len(remote_calls)}
+            # Repeat the completed master-only queue with a fresh outer executor.
+            # Every item must be delivered again without recomputing its branch.
+            before_sources, before_local = len(SOURCE_CALLS), len(BRANCH_CALLS)
+            before_models, before_utilities = len(MODEL_CALLS), len(UTILITY_CALLS)
+            before_captures = len(CAPTURED)
+            previous_outputs = CAPTURED[-1]
+            executor = execution.PromptExecutor(ps, cache_args={'ram': 0, 'ram_inactive': 0}, asset_manager=assets)
+            with patch.object(config_mod, 'load_config', return_value=config), \
+                 patch.object(orchestration, 'load_config', return_value=config), \
+                 patch.object(orchestration, 'queue_prompt_payload', enqueue), \
+                 patch.object(config_mod, 'get_worker_timeout_seconds', return_value=15):
+                await asyncio.wait_for(orchestration.orchestrate_distributed_execution(
+                    prompt, None, 'test', []), 45)
+            assert executor.success, executor.status_messages
+            assert len(SOURCE_CALLS) == before_sources + 1
+            assert len(BRANCH_CALLS) == before_local
+            assert len(MODEL_CALLS) == before_models
+            assert len(UTILITY_CALLS) == before_utilities
+            assert len(CAPTURED) == before_captures + 1
+            for before, after in zip(previous_outputs[:2], CAPTURED[-1][:2]):
+                assert len(before) == len(after) == COUNT
+                assert all(torch.equal(a, b) for a, b in zip(before, after))
+            assert previous_outputs[2] == CAPTURED[-1][2]
+            check_carried_values()
+            assert not runtime._MASTER_JOBS and not runtime._WORKER_ITEMS
+            assert prompt == saved_prompt
+            print('IMAGE_LIST_CACHE_REQUEUE_OK unchanged items reused; fresh ordered delivery and cleanup')
+            print('IMAGE_LIST_MASTER_OK ' + json.dumps(participation_counts))
+            print('IMAGE_LIST_VALUES_OK passthrough all six types; sparse override/add; Get; sequential Get-to-Start; direct/mixed/master-only')
+            assert len(outer_executions) == 5 and all(row['captures'] == 1 for row in outer_executions)
+            print('IMAGE_LIST_TARGETS_OK ' + json.dumps({
+                'modes': ['delegate-only', 'mixed', 'master-only'],
+                'outer_targets': outer_executions[0]['targets'], 'worker_targets': ['4'],
+                'required_utility_once_per_item': True, 'preview_discarded': True,
+                'saved_prompt_unchanged': prompt == saved_prompt, 'capture_once_per_outer_prompt': True}))
+            ALLOW_MASTER = False
+            config['settings']['master_delegate_only'] = True
+            # Preserve the original direct-runtime cancellation/failure fixture:
+            # it has no outer queue and must not claim target-control evidence.
+            direct_prompt = copy.deepcopy(prompt)
+            del direct_prompt['8'], direct_prompt['9']
+            direct_prompt['3']['inputs']['image'] = ['2', 0]
+            # Real Comfy user-interrupt flag while a real worker executor is active.
+            rows = source_rows()
+            items = runtime.normalize_image_list_inputs(
+                [row['image'] for row in rows], [row['mask'] for row in rows],
+                {f'value_{n}': [row[key] for row in rows]
+                 for n, key in ((1, 'reference_image'), (2, 'text'), (3, 'strength'), (4, 'seed'), (6, 'mask'))}
+                | {'value_5': True})
+            slow = workers[1]
+            async with aiohttp.ClientSession() as session:
+                async def slow_stats():
+                    async with session.get(f"http://127.0.0.1:{slow['port']}/test-stats") as response:
+                        return await response.json()
+                async def until(predicate):
+                    for _ in range(150):
+                        current = await slow_stats()
+                        if predicate(current):
+                            return current
+                        await asyncio.sleep(0.02)
+                    raise AssertionError('worker state transition timed out')
+                before = await until(lambda state: not state['running'] and not state['pending'])
+                context = runtime.create_master_job('2', '4', direct_prompt, items, ['slow'],
+                                                  f'http://127.0.0.1:{port}')
+                with patch.object(config_mod, 'load_config', return_value=config), \
+                     patch.object(config_mod, 'get_worker_timeout_seconds', return_value=15):
+                    task = asyncio.create_task(runtime.run_image_list_job(context, include_values=True))
+                    active = await until(lambda state: bool(state['running']))
+                    owned_id = active['running'][0]
+                    comfy.model_management.interrupt_current_processing(True)
+                    try:
+                        await asyncio.wait_for(task, 5)
+                        raise AssertionError('interrupt was ignored')
+                    except runtime.ImageListCancelledError:
+                        pass
+                    finally:
+                        comfy.model_management.interrupt_current_processing(False)
+                after = await until(lambda state: not state['running'] and not state['pending'])
+                assert after['interrupted'] == [owned_id], after
+                assert len(after['calls']) <= len(before['calls']) + 1
+                assert set(before['history']) <= set(after['history'])
+                assert not runtime._MASTER_JOBS
+            print('IMAGE_LIST_CANCEL_OK owned prompt stopped; no further dispatch; context removed')
+            # A worker exception before End must fail, not hang or return a partial list.
+            bad_rows = copy.deepcopy(items)
+            bad_rows[0]['values']['value_4'] = -1
+            context = runtime.create_master_job('2', '4', direct_prompt, bad_rows, ['fast'],
+                                              f'http://127.0.0.1:{port}')
+            with patch.object(config_mod, 'load_config', return_value=config), \
+                 patch.object(config_mod, 'get_worker_timeout_seconds', return_value=15):
+                try:
+                    await asyncio.wait_for(runtime.run_image_list_job(context, include_values=True), 5)
+                    raise AssertionError('worker exception was ignored')
+                except RuntimeError as exc:
+                    assert 'failed image-list item' in str(exc), str(exc)
+            assert not runtime._MASTER_JOBS
+            print('IMAGE_LIST_FAILURE_OK worker exception surfaced; context removed')
+            print('IMAGE_LIST_NETWORK_OK ' + json.dumps({'workers': 2, 'items': COUNT,
+                'assignments': dict(zip(['fast', 'slow'], counts)), 'ordered_exact_outputs': True,
+                'master_branch_calls': len(BRANCH_CALLS), 'source_calls': len(SOURCE_CALLS)}))
+    finally:
+        for proc in processes:
+            if proc.returncode is None:
+                proc.terminate()
+        await asyncio.gather(*(proc.wait() for proc in processes))
+        for err in stderr_files:
+            err.close()
+        await runner.cleanup()
+        network = importlib.import_module(module.__name__ + '.utils.network')
+        session = await network.get_client_session()
+        await session.close()
+
+
+if __name__ == '__main__':
+    asyncio.run(worker_main() if WORKER else coordinator())

tests/helpers/local_cache_retention_check.py

@@ -0,0 +1,114 @@
+"""CPU-only native Comfy execution across private-slot lifetimes."""
+import asyncio
+import importlib
+import os
+from pathlib import Path
+import sys
+import types
+
+ROOT = Path(__file__).resolve().parents[2]
+sys.path.insert(0, sys.argv[1])
+import comfy.cli_args
+comfy.cli_args.args.cpu = True
+comfy.cli_args.args.disable_assets = True
+import execution
+import nodes
+
+pkg = '_local_retention_probe'
+for name, path in [(pkg, ROOT), (pkg + '.image_list', ROOT / 'image_list')]:
+    module = types.ModuleType(name)
+    module.__path__ = [str(path)]
+    sys.modules[name] = module
+local = importlib.import_module(pkg + '.image_list.local_executor')
+
+CALLS = []
+OUTPUTS = []
+
+class Counted:
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {'required': {'value': ('INT',)}}
+    RETURN_TYPES = ('INT',)
+    FUNCTION = 'run'
+    def run(self, value):
+        if value < 0:
+            raise ValueError('intentional retention failure')
+        CALLS.append(value)
+        return (value,)
+
+class Uncached(Counted):
+    @classmethod
+    def IS_CHANGED(cls, **kwargs):
+        return float('nan')
+
+class Capture:
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {'required': {'value': ('INT',), 'delivery': ('STRING',)}}
+    RETURN_TYPES = ()
+    OUTPUT_NODE = True
+    FUNCTION = 'run'
+    def run(self, value, delivery):
+        OUTPUTS.append((value, delivery))
+        return ()
+
+nodes.NODE_CLASS_MAPPINGS.update(RetentionCounted=Counted, RetentionUncached=Uncached, RetentionCapture=Capture)
+
+async def run(slot, value, delivery, uncached=False):
+    prompt = {
+        'source': {'class_type': 'RetentionUncached' if uncached else 'RetentionCounted', 'inputs': {'value': value}},
+        'end': {'class_type': 'RetentionCapture', 'inputs': {'value': ['source', 0], 'delivery': delivery}},
+    }
+    await asyncio.wrap_future(slot.submit(prompt, delivery, 'end'))
+
+async def main():
+    first = local.LocalExecutionSlot()
+    await run(first, 1, 'first-1')
+    await run(first, 2, 'first-2')
+    await first.close()
+    assert local._RETAINED_OUTPUTS is not None
+    second = local.LocalExecutionSlot()
+    await run(second, 1, 'second-1')
+    await run(second, 2, 'second-2')
+    await run(second, 3, 'changed')
+    await second.close()
+    assert CALLS == [1, 2, 3], CALLS
+    assert OUTPUTS == [(1, 'first-1'), (2, 'first-2'), (1, 'second-1'), (2, 'second-2'), (3, 'changed')], OUTPUTS
+    third = local.LocalExecutionSlot()
+    await run(third, 4, 'uncached-1', True)
+    await run(third, 4, 'uncached-2', True)
+    await third.close()
+    assert CALLS[-2:] == [4, 4]
+    comfy.cli_args.args.cache_none = True
+    disabled = local.LocalExecutionSlot()
+    await run(disabled, 1, 'disabled-1')
+    await run(disabled, 1, 'disabled-2')
+    await disabled.close()
+    assert CALLS[-2:] == [1, 1]
+    assert local._RETAINED_OUTPUTS is None
+    comfy.cli_args.args.cache_none = False
+    failed = local.LocalExecutionSlot()
+    await run(failed, 7, 'before-failure')
+    try:
+        await run(failed, -1, 'failure')
+    except RuntimeError:
+        pass
+    else:
+        raise AssertionError('failure fixture unexpectedly succeeded')
+    await failed.close()
+    assert local._RETAINED_OUTPUTS is None
+    cancelled = local.LocalExecutionSlot()
+    await run(cancelled, 8, 'cancelled')
+    cancelled.request_stop()
+    await cancelled.close()
+    assert local._RETAINED_OUTPUTS is None
+    oversized = execution.LRUCache(execution.CacheKeySetInputSignature, max_size=local.NATIVE_CACHE_MAX_ENTRIES)
+    empty = execution.DynamicPrompt({})
+    await oversized.set_prompt(empty, [], execution.IsChangedCache('bound', empty, oversized))
+    oversized.min_generation = oversized.generation
+    oversized.cache = {n: n for n in range(local.NATIVE_CACHE_MAX_ENTRIES + 1)}
+    local._retain_outputs(oversized)
+    assert local._RETAINED_OUTPUTS is None
+    print('PASS: cross-run native cache reuse, fresh delivery, input change, noncacheable nodes, cache_none, cancellation, retention bound')
+
+asyncio.run(main())

tests/helpers/media_end_to_end_check.py

@@ -0,0 +1,512 @@
+"""Real CPU Comfy V3 + HTTP proof; invoked only by the opt-in parent test.
+
+The only harness pieces are deterministic source/assertion nodes and a bounded
+queue worker. /prompt, /history and /view are Comfy's real handlers; compilation,
+packing, staging, codecs, bridge execution, response decoding and Collector are
+unmodified production code. No normal instance or Runpod client is contacted.
+"""
+import asyncio
+import copy
+import importlib
+import io as bytes_io
+import json
+import os
+from pathlib import Path
+import socket
+import sys
+import types
+import wave
+from contextlib import asynccontextmanager
+from fractions import Fraction
+
+ROOT = Path(__file__).resolve().parents[2]
+
+
+def bootstrap():
+    if len(sys.argv) != 3:
+        raise SystemExit("Use test_media_end_to_end.py with explicit COMFYUI_SOURCE_ROOT")
+    core, run_root = (Path(value).resolve() for value in sys.argv[1:])
+    assert run_root.is_dir() and run_root != core and not run_root.is_relative_to(core)
+    for name in ("CUDA_VISIBLE_DEVICES", "HIP_VISIBLE_DEVICES", "ROCR_VISIBLE_DEVICES"):
+        assert os.environ.get(name) == "", f"CPU isolation missing: {name}"
+    assert not any(name in sys.modules for name in ("torch", "execution", "folder_paths"))
+    sys.path.insert(0, str(core))
+    import comfy.cli_args
+    args = comfy.cli_args.args
+    # Set CPU and all writable roots BEFORE any framework/tensor import.
+    args.cpu = True
+    args.disable_assets = True
+    args.disable_all_custom_nodes = True
+    args.enable_manager = False
+    args.offline = True
+    args.base_directory = str(run_root)
+    args.front_end_root = str(run_root)
+    args.database_url = "sqlite:///:memory:"
+    for name in ("input", "output", "temp", "user"):
+        (run_root / name).mkdir()
+        setattr(args, name + "_directory", str(run_root / name))
+    return run_root
+
+
+async def check(base):
+    import av
+    import numpy as np
+    import torch
+    import execution
+    import folder_paths
+    import nodes
+    import server
+    from aiohttp import web
+    from app.assets.manager import default_asset_manager
+    from comfy_api.v0_0_2 import io
+    from comfy_api.latest._input_impl.video_types import (
+        VideoFromComponents, VideoFromFile, VideoFromList,
+    )
+    from comfy_api.latest._util import VideoComponents
+    from comfy_extras import nodes_audio
+
+    assert not torch.cuda.is_initialized()
+    torch.set_num_threads(1)
+    assert execution.comfy.model_management.get_torch_device().type == "cpu"
+    for name in ("input", "output", "temp", "user"):
+        assert Path(getattr(folder_paths, "get_" + name + "_directory")()).is_relative_to(base)
+
+    # Namespace scaffolding avoids the unrelated pack bootstrap (service/config
+    # side effects). Every imported module below is the real source on disk.
+    package_name = "_media_e2e_acceptance"
+    for suffix in ("", ".api", ".api.orchestration", ".nodes"):
+        package = types.ModuleType(package_name + suffix)
+        package.__path__ = [str(ROOT.joinpath(*suffix.lstrip(".").split(".")))] if suffix else [str(ROOT)]
+        sys.modules[package.__name__] = package
+    def load(name):
+        return importlib.import_module(package_name + "." + name)
+
+    asset_manager = default_asset_manager()
+    ps = server.PromptServer(asyncio.get_running_loop(), asset_manager)
+    ps.distributed_pending_jobs = {}
+    ps.distributed_jobs_lock = asyncio.Lock()
+    compiler = load("api.orchestration.serverless_media_prompt")
+    packer = load("api.orchestration.serverless_typed_inputs")
+    decoder = load("api.orchestration.serverless_typed_results")
+    serializer = load("api.orchestration.runpod_client").serialize_request
+    handler = load("serverless_bridge.handler")
+    contract = load("serverless_bridge.contract")
+    video_codec = load("serverless_bridge.video_codec")
+    bridge = load("serverless_bridge")
+    collector = load("nodes.v3").DistributedCollector
+    extension = await bridge.comfy_entrypoint()
+    bridge_classes = await extension.get_node_list()
+    assert {cls.GET_SCHEMA().node_id for cls in bridge_classes} == {"MediaImportV1", "MediaExportV1"}
+    sockets = list(contract.SOCKET_KINDS)
+    collector_sockets = ["images", "audio"]
+    kinds = list(contract.SOCKET_KINDS.values())
+    assert kinds == ["IMAGE", "AUDIO", "MASK", "STRING", "VIDEO"]
+    io_types = [io.Image, io.Audio, io.Mask, io.String, io.Video]
+    fixtures, expected_sinks, seen_sinks, source_calls = {}, {}, {}, []
+
+    class AcceptanceMediaSource(io.ComfyNode):
+        @classmethod
+        def define_schema(cls):
+            return io.Schema(node_id="AcceptanceMediaSource", inputs=[io.String.Input("key")],
+                             outputs=[kind.Output(is_output_list=True) for kind in io_types])
+
+        @classmethod
+        def execute(cls, key):
+            source_calls.append(key)
+            return io.NodeOutput(*(fixtures[key][socket] for socket in sockets))
+
+    def tensor_equal(actual, expected):
+        assert actual.device.type == "cpu" and actual.dtype == expected.dtype
+        assert actual.shape == expected.shape
+        # Byte comparison preserves signed zero and catches dtype narrowing.
+        assert torch.equal(actual.contiguous().view(torch.uint8), expected.contiguous().view(torch.uint8))
+
+    def exact(kind, actual, expected):
+        if kind in ("IMAGE", "MASK"):
+            tensor_equal(actual, expected)
+        elif kind == "AUDIO":
+            assert actual["sample_rate"] == expected["sample_rate"]
+            tensor_equal(actual["waveform"], expected["waveform"])
+        elif kind == "STRING":
+            assert type(actual) is str and actual.encode("utf-8") == expected.encode("utf-8")
+        else:
+            assert video_codec.encode_video(actual) == expected
+            if type(actual) is VideoFromFile:
+                buffer = vars(actual)["_VideoFromFile__file"]
+                assert type(buffer) is bytes_io.BytesIO and not buffer.closed
+                position = buffer.tell()
+                buffer.seek(0)
+                with av.open(buffer) as container:
+                    assert len(list(container.decode(video=0))) == 3
+                buffer.seek(position)
+            elif type(actual) is VideoFromList:
+                for item in actual.videos:
+                    exact("VIDEO", item, video_codec.encode_video(item))
+            else:
+                assert type(actual) is VideoFromComponents
+                components = actual.get_components()
+                assert type(components.frame_rate) is Fraction
+                assert components.images.dtype == torch.float64
+                assert components.alpha.dtype == torch.bfloat16
+                assert components.audio["waveform"].dtype == torch.float64
+                assert actual.get_bit_depth() == 10 and actual.get_color_space() == "HDR PQ"
+
+    class AcceptanceMediaAssert(io.ComfyNode):
+        @classmethod
+        def define_schema(cls):
+            return io.Schema(node_id="AcceptanceMediaAssert", inputs=[io.String.Input("tag"),
+                *[kind.Input(socket, optional=True, **({"force_input": True} if socket == "texts" else {}))
+                  for socket, kind in zip(sockets, io_types)]], outputs=[],
+                is_input_list=True, is_output_node=True)
+
+        @classmethod
+        def execute(cls, tag, **media):
+            assert len(tag) == 1 and tag[0] not in seen_sinks
+            expected = expected_sinks[tag[0]]
+            assert set(media) == set(expected)
+            for socket in expected:
+                assert len(media[socket]) == len(expected[socket]), (tag, socket)
+                # Existing Collector outputs are a batched IMAGE and one AUDIO,
+                # not lists. Native V3 wraps each output once for this list sink.
+                for actual, wanted in zip(media[socket], expected[socket]):
+                    exact(contract.SOCKET_KINDS[socket], actual, wanted)
+            counts = {socket: len(media[socket]) for socket in expected}
+            seen_sinks[tag[0]] = counts
+            return io.NodeOutput(ui={"accepted": [counts]})
+
+    for cls in [*bridge_classes, collector, AcceptanceMediaSource, AcceptanceMediaAssert]:
+        assert issubclass(cls, io.ComfyNode)
+        cls.GET_NODE_INFO_V1()
+        nodes.NODE_CLASS_MAPPINGS[cls.GET_SCHEMA().node_id] = cls
+
+    def node(class_type, **inputs):
+        return {"class_type": class_type, "inputs": inputs}
+
+    def add_bridge_exports(workflow, manifest, request_id, participant, job_id, links):
+        """Exercise bridge-only kinds without inventing Collector sockets."""
+        for socket, link in links.items():
+            export_id = "bridge-export:" + socket
+            binding = {"binding_id": "bridge-binding:" + socket, "collector_id": "bridge-only",
+                       "job_id": job_id, "participant_id": participant, "node_id": export_id,
+                       "socket": socket, "kind": contract.SOCKET_KINDS[socket]}
+            assert export_id not in workflow
+            manifest.append(binding)
+            workflow[export_id] = node("MediaExportV1", request_id=request_id,
+                                      binding_json=json.dumps(binding), **{socket: link})
+        contract.validate_bindings(manifest)
+
+    def tiny_video(variant):
+        buffer = bytes_io.BytesIO()
+        with av.open(buffer, "w", format="matroska") as container:
+            stream = container.add_stream("ffv1", rate=Fraction(24, 1))
+            stream.width = stream.height = 16
+            stream.pix_fmt = "yuv420p"
+            for level in (16 + variant, 64 + variant, 128 + variant):
+                frame = av.VideoFrame.from_ndarray(np.full((16, 16, 3), level, dtype=np.uint8), format="rgb24")
+                for packet in stream.encode(frame):
+                    container.mux(packet)
+            for packet in stream.encode():
+                container.mux(packet)
+        return buffer.getvalue()
+
+    def native_fixtures(variant):
+        image = torch.arange(48, dtype=torch.float64).reshape(2, 2, 3, 4) / 997 + variant / 100
+        image[0, 0, 0, 0] = -0.0
+        mask = torch.tensor([[[-0.0, 0.125, 0.875]]], dtype=torch.bfloat16)
+        audio = {"waveform": torch.arange(38, dtype=torch.float64).reshape(1, 2, 19) / 997,
+                 "sample_rate": 44100}
+        audio["waveform"][0, 0, 0] = -0.0
+        component = VideoFromComponents(VideoComponents(
+            images=image[..., :3], frame_rate=Fraction(30000, 1001), alpha=mask.reshape(1, 1, 3, 1).repeat(2, 2, 1, 1),
+            audio=audio, metadata={"title": "雪", "nested": [True, None, -4, 0.125]}),
+            bit_depth=10, color_space="HDR PQ")
+        payload = tiny_video(variant)
+        file = VideoFromFile(bytes_io.BytesIO(payload), start_time=-0.0, duration=0, crop=(0, 0, 8, 8))
+        sequence = VideoFromList([VideoFromFile(bytes_io.BytesIO(payload)), file], complete_audio=audio)
+        return {"images": [image, image[:1].clone()], "audio": [audio], "masks": [mask],
+                "texts": [f"  café 🎥 {variant}\r\n雪  ", ""], "video": [component, sequence]}
+
+    for key, variant in (("local", 0), ("worker-a", 1), ("worker-b", 2)):
+        fixtures[key] = native_fixtures(variant)
+    native_root = base / "master-input"
+    native_root.mkdir()
+    wav = native_root / "stereo.wav"
+    pcm = np.array([-32768, -12001, -1, 0, 1, 8193, 32767, 17] * 2, dtype="<i2")
+    with wave.open(str(wav), "wb") as out:
+        out.setnchannels(2)
+        out.setsampwidth(2)
+        out.setframerate(44100)
+        out.writeframes(pcm.tobytes())
+    waveform, sample_rate = nodes_audio.load(str(wav))
+    clip = native_root / "native.mkv"
+    clip.write_bytes(tiny_video(3))
+    native_expected = copy.copy(fixtures["worker-a"])
+    native_expected["audio"] = [{"waveform": waveform.unsqueeze(0), "sample_rate": sample_rate}]
+    native_expected["video"] = [VideoFromFile(str(clip))]
+    expectations = {}
+    for key, values in (("worker-a", native_expected), ("worker-b", fixtures["worker-b"])):
+        expectations[key] = {socket: [video_codec.encode_video(v, allowed_roots=[native_root]) if socket == "video" else v
+                                      for v in values[socket]] for socket in sockets}
+
+    # Native routes only, bound to a new ephemeral loopback listener. No fake
+    # history, permissive validator, bridge/codec mock, or existing service.
+    allowed_paths = {"/prompt", "/history/{prompt_id}", "/view", "/queue"}
+    ps.app.add_routes([route for route in ps.routes if route.path in allowed_paths])
+    completed = []
+    stop = asyncio.Event()
+
+    async def queue_worker():
+        while not stop.is_set():
+            item = await asyncio.to_thread(ps.prompt_queue.get, 0.1)
+            if item is None:
+                continue
+            queued, item_id = item
+            assert len(queued) in (5, 6), "Unsupported core PromptQueue tuple; adapt acceptance worker explicitly"
+            executor = execution.PromptExecutor(ps, cache_args={"ram": 0, "ram_inactive": 0}, asset_manager=asset_manager)
+            await asyncio.to_thread(executor.execute, queued[2], queued[1], queued[3], queued[4])
+            status = execution.PromptQueue.ExecutionStatus(
+                status_str="success" if executor.success else "error", completed=executor.success,
+                messages=executor.status_messages)
+            ps.prompt_queue.task_done(item_id, executor.history_result, status=status)
+            completed.append((queued, executor.success, executor.status_messages))
+
+    @asynccontextmanager
+    async def serve():
+        runner = web.AppRunner(ps.app)
+        await runner.setup()
+        site = web.TCPSite(runner, "127.0.0.1", 0)
+        await site.start()
+        port = site._server.sockets[0].getsockname()[1]
+        ALLOWED_CONNECTIONS.add(("127.0.0.1", port))
+        worker = asyncio.create_task(queue_worker())
+        try:
+            yield f"http://127.0.0.1:{port}"
+        finally:
+            stop.set()
+            await asyncio.wait_for(worker, 10)
+            await runner.cleanup()
+            ALLOWED_CONNECTIONS.clear()
+
+    grouped_results = {}
+    async with serve() as url:
+        for key in ("worker-a", "worker-b"):
+            participant = "serverless:" + key
+            request_id = "acceptance-" + key
+            graph = {"source": node("AcceptanceMediaSource", key=key),
+                     "collector": node("DistributedCollector", load_balance=False, transfer_format="exact",
+                         multi_job_id="job-main", worker_id=participant,
+                         **{socket: ["source", index] for index, socket in enumerate(collector_sockets)}),
+                     "audio-collector": node("DistributedCollector", load_balance=False,
+                         multi_job_id="job-audio", worker_id=participant, audio=["source", 1])}
+            if key == "worker-a":
+                graph.update({"wav": node("LoadAudio", audio=wav.name), "clip": node("LoadVideo", file=clip.name)})
+                graph["collector"]["inputs"]["audio"] = ["wav", 0]
+            before = copy.deepcopy(graph)
+            workflow, manifest = compiler.compile_media_prompt(graph, request_id)
+            assert graph == before
+            bridge_links = {socket: ["source", sockets.index(socket)] for socket in sockets[2:]}
+            if key == "worker-a":
+                workflow["clip"] = copy.deepcopy(graph["clip"])
+                bridge_links["video"] = ["clip", 0]
+            add_bridge_exports(workflow, manifest, request_id, participant, "job-bridge", bridge_links)
+            assert len(manifest) == 6 and {b["kind"] for b in manifest} == set(kinds)
+            folder_paths.set_input_directory(str(native_root))
+            workflow_before = copy.deepcopy(workflow)
+            images, artifacts, refs = await packer.prepare_media_inputs(workflow)
+            assert images == [] and workflow == workflow_before
+            if key == "worker-a":
+                assert {a["kind"] for a in artifacts} == {"AUDIO", "VIDEO"} and len(artifacts) == 2
+                assert len(refs) == 2
+                assert {ref["input_name"] for ref in refs} == {"audio", "video"}
+                assert all(workflow[ref["node_id"]]["class_type"] == "MediaExportV1" for ref in refs)
+                # Endpoint cannot fall back to the master's local files/loaders.
+                wav.unlink()
+                clip.unlink()
+                print("NATIVE_INPUTS_AND_DIRECT_EXPORT_REFS_OK WAV + native file VIDEO")
+            else:
+                assert not artifacts and not refs
+            folder_paths.set_input_directory(str(base / "input"))
+            media = {"version": 1, "profile": contract.PROFILE, "request_id": request_id,
+                     "artifacts": artifacts, "inputs": refs, "outputs": manifest}
+            body = json.loads(serializer(workflow, images, 30, media=media))
+            assert decoder.validate_prepared_job({"worker": {"id": participant, "media_profile": contract.PROFILE,
+                       "timeout_seconds": 30}, "workflow": workflow, "images": images,
+                       "media": media, "manifest": manifest}) is True
+            before = copy.deepcopy(body["input"])
+            result = await handler.execute_media_request(body["input"], comfy_url=url,
+                        input_dir=base / "input", temp_dir=base / "temp", timeout_seconds=30)
+            assert body["input"] == before
+            assert completed and completed[-1][1], completed
+            queued = completed[-1][0]
+            assert set(queued[4]) == {b["node_id"] for b in manifest}
+            submitted = queued[2]
+            if key == "worker-a":
+                assert "wav" not in submitted and "clip" not in submitted
+                imports = [n for n in submitted.values() if n["class_type"] == "MediaImportV1"]
+                assert len(imports) == 2
+                assert {json.loads(n["inputs"]["artifact_json"])["kind"] for n in imports} == {"AUDIO", "VIDEO"}
+            assert not list((base / "input").rglob("*.bin"))
+            assert not list((base / "temp").rglob("*.bin"))
+            if key == "worker-b":
+                # Retire the originals as well as handler-owned staging/export
+                # files before constructing the result's lazy native videos.
+                for original in fixtures[key]["video"][1].videos:
+                    vars(original)["_VideoFromFile__file"].close()
+            # Exercise the wire format, not an in-memory handler-only object.
+            result = json.loads(json.dumps(result, allow_nan=False))
+            grouped = await decoder.collect_typed_results(result, media, manifest)
+            assert set(grouped) == {"job-main", "job-audio", "job-bridge"}
+            assert set(grouped["job-main"]) == set(collector_sockets)
+            assert set(grouped["job-bridge"]) == set(sockets[2:])
+            exact("AUDIO", grouped["job-audio"]["audio"][0], fixtures[key]["audio"][0])
+            all_media = {**grouped["job-main"], **grouped["job-bridge"]}
+            for socket, kind in zip(sockets, kinds):
+                values, expected = all_media[socket], expectations[key][socket]
+                assert len(values) == len(expected), (key, socket)
+                for actual, wanted in zip(values, expected):
+                    exact(kind, actual, wanted)
+            if key == "worker-b":
+                original = fixtures[key]["video"][0].get_components()
+                decoded = grouped["job-bridge"]["video"][0].get_components()
+                for field in ("images", "alpha"):
+                    assert getattr(decoded, field).data_ptr() != getattr(original, field).data_ptr()
+                assert decoded.audio["waveform"].data_ptr() != original.audio["waveform"].data_ptr()
+                assert decoded.metadata == original.metadata and decoded.metadata is not original.metadata
+            grouped_results[participant] = grouped
+            # Late invalid artifact must reject the entire grouped response.
+            corrupt = copy.deepcopy(result)
+            corrupt["media"]["artifacts"][-1]["sha256"] = "0" * 64
+            try:
+                await decoder.collect_typed_results(corrupt, media, manifest)
+            except ValueError:
+                pass
+            else:
+                raise AssertionError("Result decoder accepted a corrupt final artifact")
+            assert ps.distributed_pending_jobs == {}
+        # Exercise user-selected image return encoding through the actual
+        # compiler, HTTP handler, V3 exporter and atomic response decoder.
+        for fmt in ("legacy_png", "png", "jpeg", "h264"):
+            key = "encoded-" + fmt
+            fixtures[key] = native_fixtures(3)
+            # Keep full alpha precision and use nontrivial image dimensions.
+            torch.manual_seed(41)
+            rgba = torch.rand((2, 64, 64, 4), dtype=torch.float64)
+            fixtures[key]["images"] = [rgba]
+            request_id = "request-" + key
+            graph = {"source": node("AcceptanceMediaSource", key=key),
+                     "collector": node("DistributedCollector", transfer_format=fmt,
+                         transfer_quality=73, multi_job_id="job-" + key,
+                         worker_id="serverless:" + key,
+                         **{socket: ["source", i] for i, socket in enumerate(collector_sockets)})}
+            if fmt == "legacy_png":
+                # A saved Collector with no newly selected controls still uses
+                # legacy_png / 90, mapped to PNG by the serverless compiler.
+                graph["collector"]["inputs"].pop("transfer_format")
+                graph["collector"]["inputs"].pop("transfer_quality")
+            original = copy.deepcopy(graph)
+            workflow, manifest = compiler.compile_media_prompt(graph, request_id)
+            if fmt == "legacy_png":
+                binding = next(b for b in manifest if b["kind"] == "IMAGE")
+                controls = workflow[binding["node_id"]]["inputs"]
+                assert (controls["transfer_format"], controls["transfer_quality"]) == ("png", 90)
+            add_bridge_exports(workflow, manifest, request_id, "serverless:" + key, "job-bridge-" + key,
+                               {socket: ["source", sockets.index(socket)] for socket in sockets[2:]})
+            media = {"version": 1, "profile": contract.PROFILE, "request_id": request_id,
+                     "artifacts": [], "inputs": [], "outputs": manifest}
+            body = json.loads(serializer(workflow, [], 30, media=media))
+            result = await handler.execute_media_request(body["input"], comfy_url=url,
+                        input_dir=base / "input", temp_dir=base / "temp", timeout_seconds=30)
+            assert graph == original
+            grouped = await decoder.collect_typed_results(result, media, manifest)
+            actual = {**grouped["job-" + key], **grouped["job-bridge-" + key]}
+            image = actual["images"][0]
+            assert image.shape == rgba.shape and image.dtype == rgba.dtype
+            assert torch.equal(image[..., 3], rgba[..., 3]), "alpha must not be discarded/narrowed"
+            assert not torch.equal(image[..., :3], rgba[..., :3]), "selection must actually encode"
+            if fmt in ("legacy_png", "png"):
+                pixels = (rgba[..., :3].float().clamp(0, 1) * 255).to(torch.uint8)
+                assert torch.equal(image[..., :3], (pixels.float() / 255).to(rgba.dtype))
+            for socket, kind in zip(sockets[1:], kinds[1:]):
+                for got, wanted in zip(actual[socket], fixtures[key][socket]):
+                    exact(kind, got, video_codec.encode_video(wanted) if kind == "VIDEO" else wanted)
+            assert not list((base / "temp").rglob("*.bin"))
+        print("IMAGE_TRANSFER_HTTP_OK default legacy_png -> png/90; png jpeg h264; alpha and other media exact")
+        print("REAL_HTTP_BRIDGE_RESULTS_OK Collector IMAGE/AUDIO + independent MASK/STRING/VIDEO bridge exports; native VIDEO survives cleanup")
+        print("ATOMIC_DECODE_REJECTION_OK corrupt final artifact rejected before delivery")
+
+    participants = list(grouped_results)
+
+    def expected_batches(ordered, include_images=True):
+        audio_items = [item for media in ordered for item in media["audio"]]
+        expected = {"audio": [{"waveform": torch.cat([item["waveform"] for item in audio_items], dim=-1),
+                                "sample_rate": audio_items[0]["sample_rate"]}]}
+        if include_images:
+            expected["images"] = [torch.cat([image for media in ordered for image in media["images"]])]
+        return expected
+
+    for delegate in (False, True):
+        tag = "delegate" if delegate else "master"
+        job = "collect-" + tag
+        queue = asyncio.Queue()
+        for participant in reversed(participants):
+            await queue.put({"worker_id": participant, "media": grouped_results[participant]["job-main"], "is_last": True})
+        ps.distributed_pending_jobs[job] = queue
+        ordered = ([] if delegate else [fixtures["local"]]) + [grouped_results[p]["job-main"] for p in participants]
+        expected_sinks[tag] = expected_batches(ordered)
+        graph = {"collector": node("DistributedCollector", load_balance=False, multi_job_id=job,
+                    enabled_worker_ids=json.dumps(participants), delegate_only=delegate, transfer_format="exact"),
+                 "assert": node("AcceptanceMediaAssert", tag=tag,
+                    **{socket: ["collector", index] for index, socket in enumerate(collector_sockets)})}
+        if not delegate:
+            graph["source"] = node("AcceptanceMediaSource", key="local")
+            graph["collector"]["inputs"].update({socket: ["source", index] for index, socket in enumerate(collector_sockets)})
+        # A second audio-only Collector consumes its own job group from the same
+        # atomic response. It must return None for images and batch audio in order.
+        audio_job, audio_tag = job + "-audio", tag + "-audio"
+        audio_queue = asyncio.Queue()
+        for participant in reversed(participants):
+            await audio_queue.put({"worker_id": participant,
+                "media": grouped_results[participant]["job-audio"], "is_last": True})
+        ps.distributed_pending_jobs[audio_job] = audio_queue
+        audio_ordered = ([] if delegate else [fixtures["local"]]) + [grouped_results[p]["job-audio"] for p in participants]
+        expected_sinks[audio_tag] = expected_batches(audio_ordered, include_images=False)
+        graph["audio-collector"] = node("DistributedCollector", load_balance=False, multi_job_id=audio_job,
+            enabled_worker_ids=json.dumps(participants), delegate_only=delegate)
+        graph["audio-assert"] = node("AcceptanceMediaAssert", tag=audio_tag, audio=["audio-collector", 1])
+        if not delegate:
+            graph["audio-collector"]["inputs"]["audio"] = ["source", 1]
+        valid = await execution.validate_prompt(job, graph, None)
+        assert valid[0] and not valid[3], valid
+        executor = execution.PromptExecutor(ps, cache_args={"ram": 0, "ram_inactive": 0}, asset_manager=asset_manager)
+        await asyncio.to_thread(executor.execute, graph, job, {}, valid[2])
+        assert executor.success, executor.status_messages
+        assert executor.history_result["outputs"]["assert"]["accepted"] == [seen_sinks[tag]]
+        assert executor.history_result["outputs"]["audio-assert"]["accepted"] == [seen_sinks[audio_tag]]
+        assert job not in ps.distributed_pending_jobs and audio_job not in ps.distributed_pending_jobs
+    assert source_calls == ["worker-a", "worker-b", "encoded-legacy_png", "encoded-png", "encoded-jpeg", "encoded-h264", "local"], source_calls
+    assert set(seen_sinks) == {"master", "delegate", "master-audio", "delegate-audio"}
+
+    assert not list((base / "output").rglob("*"))
+    assert not torch.cuda.is_initialized()
+    print("MASTER_AND_DELEGATE_ORDER_OK", json.dumps(seen_sinks, sort_keys=True))
+    print("MEDIA_END_TO_END_CPU_OK real PromptExecutor/HTTP; no queue-orchestration or provider acceptance claimed")
+
+
+ALLOWED_CONNECTIONS = set()
+
+
+def network_guard(event, args):
+    """Fail closed on provider/normal-instance traffic, including during imports."""
+    if event == "socket.connect":
+        sock, address = args
+        if sock.family in (socket.AF_INET, socket.AF_INET6):
+            assert address in ALLOWED_CONNECTIONS, "Acceptance forbids non-owned network connections"
+    elif event == "socket.bind":
+        sock, address = args
+        if sock.family in (socket.AF_INET, socket.AF_INET6):
+            assert address == ("127.0.0.1", 0), "Acceptance binds only ephemeral loopback"
+
+
+if __name__ == "__main__":
+    sys.addaudithook(network_guard)
+    asyncio.run(asyncio.wait_for(check(bootstrap()), timeout=120))

tests/helpers/v3_runtime_check.py

@@ -36,7 +36,7 @@ BASELINE = json.loads((ROOT / 'tests/fixtures/v1_node_contracts.json').read_text
 ADDED_OPTIONAL_INPUTS = {
     'DistributedCollector': {
         'transfer_format': [
-            ['legacy_png', 'png', 'ffv1', 'jpeg', 'webp', 'h264', 'h265', 'av1'],
+            ['legacy_png', 'png', 'ffv1', 'jpeg', 'webp', 'h264', 'h265', 'av1', 'exact'],
             {'default': 'legacy_png'},
         ],
         'transfer_quality': ['INT', {'default': 90, 'min': 1, 'max': 100, 'step': 1}],
@@ -45,7 +45,8 @@ ADDED_OPTIONAL_INPUTS = {
 
 # Nodes added after the V1 baseline; not part of saved-workflow contracts.
 # DistributedWorkerSink is internal (dev-only) and only inserted by the master.
-ADDED_NODES = {'DistributedWorkerSink'}
+ADDED_NODES = {'DistributedWorkerSink', 'DistributedImageListStart', 'DistributedImageListEnd',
+               'DistributedImageListGetValue'}
 
 
 def normalized_input(value):
@@ -254,6 +255,109 @@ async def check_execution(mapping, module, prompt_server, asset_manager):
     print('V3_EXECUTION_OK nine nodes; upscale GPU boundary mocked; preview artifact verified; sink writes nothing')
 
 
+async def check_image_list_lazy_executor(mapping, module, prompt_server, asset_manager):
+    """Execute real native V3 list/lazy flow; master must skip the processing branch."""
+    branch_calls = []
+    class ImageListProbe:
+        @classmethod
+        def INPUT_TYPES(cls):
+            return {'required': {'image': ('IMAGE',)}}
+        RETURN_TYPES = ('IMAGE',)
+        FUNCTION = 'run'
+        CATEGORY = 'tests'
+        def run(self, image):
+            branch_calls.append(image)
+            return (image,)
+
+    nodes.NODE_CLASS_MAPPINGS['ImageListTestProbe'] = ImageListProbe
+    runtime = sys.modules[module.__name__ + '.image_list.runtime']
+    original_contexts = set(runtime._MASTER_JOBS)
+    image_list_nodes = sys.modules[module.__name__ + '.nodes.image_list']
+    workflow_meta = {'nodes': [{'id': 2, 'type': 'DistributedImageListStart'}], 'extra': {'acceptance': True}}
+    async def fake_worker_results(_context_id, *, model_callback=None, include_values=False):
+        state = runtime.get_master_job(_context_id)
+        assert state.prompt_id == 'image-list-lazy', state.prompt_id
+        assert state.workflow_meta == workflow_meta, state.workflow_meta
+        assert len(prompt_server.app.cleanup_ctx) >= 1, 'lifecycle cleanup was not registered'
+        return (
+            [torch.full((1, 8, 8, 3), value, dtype=torch.float32) for value in (0.125, 0.5, 0.875)],
+            [None, None, None], [None, None, None],
+            [{}, {}, {}],
+        )
+
+    graph = {
+        '1': {'class_type': 'EmptyImage', 'inputs': {'height': 8, 'width': 8, 'batch_size': 3, 'color': 0}},
+        '2': {'class_type': 'DistributedImageListStart', 'inputs': {
+            'images': ['1', 0], 'image_list_enabled_worker_ids': '["test-worker"]',
+            'image_list_master_callback': 'http://127.0.0.1:1',
+            'image_list_workflow_meta': workflow_meta,
+        }},
+        '3': {'class_type': 'ImageListTestProbe', 'inputs': {'image': ['2', 0]}},
+        '4': {'class_type': 'DistributedImageListEnd', 'inputs': {
+            'images': ['3', 0], 'context': ['2', 3],
+        }},
+        '5': {'class_type': 'PreviewImage', 'inputs': {'images': ['4', 0]}},
+    }
+    valid = await execution.validate_prompt('image-list-lazy', graph, None)
+    assert valid[0], valid
+    validator = runtime.validate_image_list_regions
+    assert 'texts' not in graph['4']['inputs']
+    assert len(validator(graph)) == 1
+    connected_text_graph = json.loads(json.dumps(graph))
+    connected_text_graph['6'] = {
+        'class_type': 'ImageListTextProbe', 'inputs': {'text': ['2', 4]},
+    }
+    connected_text_graph['4']['inputs']['texts'] = ['6', 0]
+    assert len(validator(connected_text_graph)) == 1
+    for literal_text in ('', 'per-item text'):
+        literal_text_graph = json.loads(json.dumps(graph))
+        literal_text_graph['4']['inputs']['texts'] = literal_text
+        try:
+            validator(literal_text_graph)
+        except ValueError:
+            pass
+        else:
+            raise AssertionError(f'End accepted unconnected literal text: {literal_text!r}')
+    print('IMAGE_LIST_END_TEXT_INPUT_OK native forceInput; unconnected omitted; linked accepted; literals rejected')
+    import folder_paths
+    scratch = Path(os.environ.get('TMPDIR', Path.home() / '.hermes/cache/scratch'))
+    try:
+        # Resolve genuine Comfy output/temp annotations against the pinned core,
+        # not only the controlled folder_paths seam in unit tests.
+        media = sys.modules[module.__name__ + '.api.orchestration.media_sync']
+        with tempfile.TemporaryDirectory(prefix='media-roots-', dir=scratch) as media_root:
+            for kind in ('input', 'output', 'temp'):
+                source = Path(media_root) / kind / 'shared/ref.png'
+                source.parent.mkdir(parents=True)
+                source.write_bytes(kind.encode())
+            with patch.object(folder_paths, 'input_directory', str(Path(media_root) / 'input')), \
+                 patch.object(folder_paths, 'output_directory', str(Path(media_root) / 'output')), \
+                 patch.object(folder_paths, 'temp_directory', str(Path(media_root) / 'temp')):
+                for kind in ('output', 'temp'):
+                    refs = media._find_media_references({'1': {'inputs': {'image': f'shared/ref.png [{kind}]'}}})
+                    assert len(refs) == 1
+                    assert media._load_media_file_sync(refs[0])[0] == kind.encode()
+        with tempfile.TemporaryDirectory(prefix='image-list-lazy-', dir=scratch) as temp:
+            with patch.object(folder_paths, 'temp_directory', temp), \
+                 patch.object(image_list_nodes, 'run_image_list_job', fake_worker_results):
+                executor = execution.PromptExecutor(
+                    prompt_server, cache_args={'ram': 0, 'ram_inactive': 0}, asset_manager=asset_manager)
+                await asyncio.to_thread(executor.execute, graph, 'image-list-lazy', {}, valid[2])
+                assert executor.success, executor.status_messages
+                assert branch_calls == [], 'master evaluated the lazy per-image branch'
+                records = executor.history_result['outputs']['5']['images']
+                assert len(records) == 3, records
+                for record in records:
+                    preview = Path(temp) / record.get('subfolder', '') / record['filename']
+                    with Image.open(preview) as image:
+                        assert image.size == (8, 8) and image.mode == 'RGB'
+    finally:
+        nodes.NODE_CLASS_MAPPINGS.pop('ImageListTestProbe', None)
+        for context_id in set(runtime._MASTER_JOBS) - original_contexts:
+            runtime.remove_master_job(context_id)
+    print('IMAGE_LIST_LAZY_EXECUTION_OK real PromptExecutor; branch skipped on master; 3 previews verified')
+
+
 async def main():
     asset_manager = default_asset_manager()
     prompt_server = server.PromptServer(asyncio.get_running_loop(), asset_manager)
@@ -264,10 +368,16 @@ async def main():
     extension = await module.comfy_entrypoint()
     classes = await extension.get_node_list()
     mapping = {cls.GET_SCHEMA().node_id: cls for cls in classes}
-    assert len(classes) == len(mapping) == len(BASELINE) + len(ADDED_NODES) == 9
+    assert len(classes) == len(mapping) == len(BASELINE) + len(ADDED_NODES) == 12
     assert set(mapping) == set(BASELINE) | ADDED_NODES
     for node_id in ADDED_NODES:
         assert nodes.NODE_CLASS_MAPPINGS[node_id] is mapping[node_id]
+    assert 'DistributedMediaCollector' not in mapping
+    collector_info = mapping['DistributedCollector'].GET_NODE_INFO_V1()
+    assert collector_info['output'] == ['IMAGE', 'AUDIO'], collector_info
+    assert not any(collector_info['output_is_list']), collector_info
+    assert list(collector_info['input']['optional']) == [
+        'images', 'audio', 'transfer_format', 'transfer_quality']
     sink_info = mapping['DistributedWorkerSink'].GET_NODE_INFO_V1()
     assert sink_info['output_node'] and sink_info['dev_only'], sink_info
     for node_id, cls in mapping.items():
@@ -315,9 +425,93 @@ async def main():
         if old['output_node']:
             expected_hidden.extend(name for name in ['prompt', 'extra_pnginfo'] if name not in expected_hidden)
         assert list(new['input'].get('hidden', {})) == expected_hidden, (node_id, new['input'].get('hidden'))
+    start_schema = mapping['DistributedImageListStart'].GET_SCHEMA()
+    end_schema = mapping['DistributedImageListEnd'].GET_SCHEMA()
+    assert start_schema.node_id == 'DistributedImageListStart'
+    assert start_schema.display_name == 'Distributed Image List Start'
+    assert end_schema.node_id == 'DistributedImageListEnd'
+    assert end_schema.display_name == 'Distributed Image List End'
+    runtime = sys.modules[module.__name__ + '.image_list.runtime']
+    assert [input.id for input in end_schema.inputs] == ['images', 'masks', 'texts', 'context'] + list(runtime.VALUE_NAMES)
+    assert [output.io_type for output in end_schema.outputs] == [
+        'IMAGE', 'MASK', 'STRING', 'DISTRIBUTED_IMAGE_LIST_VALUES']
+    assert end_schema.outputs[3].id == end_schema.outputs[3].display_name == 'values'
+    assert all(output.is_output_list for output in end_schema.outputs)
+    for override in end_schema.inputs[4:]:
+        assert override.optional and override.lazy
+        assert override.get_io_type() == 'IMAGE,MASK,STRING,INT,FLOAT,BOOLEAN'
+    getter = mapping['DistributedImageListGetValue']
+    getter_schema = getter.GET_SCHEMA()
+    assert getter_schema.display_name == 'Distributed Image List Get Value'
+    assert not getter_schema.is_input_list
+    assert getter_schema.outputs[0].id == getter_schema.outputs[0].display_name == 'value'
+    assert getter_schema.outputs[0].io_type == '*'
+    assert getter_schema.inputs[1].options == list(runtime.VALUE_NAMES)
+    assert getter_schema.inputs[1].default == 'value_1'
+    # Real core maps scalar mappings, never arbitrary tensor batches.
+    tensor = torch.ones(3, 2, 2, 3)
+    data, missing, hidden = execution.get_input_data({'value_name': 'value_1'}, getter, 'get-value')
+    data['values'] = [{'value_1': tensor}, {'value_1': 17}, {'value_1': False}]
+    mapped = await execution._async_map_node_over_list('get-value', 'get-value', getter, data,
+                                                       getter.FUNCTION, v3_data=hidden)
+    mapped = await execution.resolve_map_node_over_list_results(mapped)
+    assert len(mapped) == 3 and mapped[0].result[0] is tensor
+    assert mapped[1].result == (17,) and mapped[2].result == (False,)
+    try:
+        getter.execute({}, 'value_1')
+        raise AssertionError('missing Get Value key accepted')
+    except ValueError as exc:
+        assert 'absent' in str(exc)
+    texts_input = next(input for input in end_schema.inputs if input.id == 'texts')
+    texts_schema = texts_input.as_dict()
+    assert texts_schema.get('forceInput') is True, texts_schema
+    assert 'default' not in texts_schema, texts_schema
+    assert [out.io_type for out in start_schema.outputs] == [
+        'IMAGE', 'MASK', 'INT', 'DISTRIBUTED_IMAGE_LIST_CONTEXT'] + ['*'] * 10
+    assert [out.display_name for out in start_schema.outputs][4:] == [f'value_{i}' for i in range(1, 11)]
+    assert all(out.is_output_list for out in start_schema.outputs)
+    assert start_schema.is_input_list and start_schema.accept_all_inputs
+    start_info = json.loads(json.dumps(mapping['DistributedImageListStart'].GET_NODE_INFO_V1()))
+    assert list(start_info['input']['optional']) == ['images', 'mask', 'values'], start_info['input']
+    values_spec = start_info['input']['optional']['values']
+    assert values_spec[0] == 'COMFY_AUTOGROW_V3', values_spec
+    template = values_spec[1]['template']
+    assert template['names'] == [f'value_{i}' for i in range(1, 11)] and template['min'] == 0, template
+    assert template['input']['required']['value'][0] == 'IMAGE,MASK,STRING,INT,FLOAT,BOOLEAN', template
+    # Autogrow + INPUT_IS_LIST through the real parser/list mapper, values
+    # still split 1-or-N, unconnected slots ignored.
+    # get_input_data must accept Autogrow names; constants are wrapped once.
+    parsed, missing, hidden = execution.get_input_data(
+        {'values.value_1': 'shared', 'values.value_3': 4},
+        mapping['DistributedImageListStart'], unique_id='autogrow-probe', extra_data={})
+    assert not missing and parsed['values.value_1'] == ['shared'], parsed
+    # Linked inputs arrive as the upstream output list (one element per value).
+    prepared = {'images': [torch.zeros((3, 2, 2, 3))], 'values.value_1': ['shared'],
+                'values.value_3': [4, 5, 6]}
+    seen = {}
+    runtime_mod = sys.modules[module.__name__ + '.image_list.runtime']
+    def capture(_images, _mask, values):
+        seen['values'] = values
+        raise RuntimeError('captured')
+    nodes_mod = sys.modules[module.__name__ + '.nodes.image_list']
+    with patch.object(nodes_mod, 'normalize_image_list_inputs', capture):
+        try:
+            await asyncio.to_thread(lambda: asyncio.run(execution._async_map_node_over_list(
+                'autogrow-probe', 'autogrow-probe', mapping['DistributedImageListStart'], prepared,
+                'execute', v3_data=hidden)))
+        except RuntimeError as exc:
+            assert 'captured' in str(exc), exc
+    assert seen['values'] == {'value_1': ['shared'], 'value_3': [4, 5, 6]}, seen
+    rows = runtime_mod.normalize_image_list_inputs(prepared['images'], None, seen['values'])
+    assert [row['values'] for row in rows] == [{'value_1': 'shared', 'value_3': n} for n in (4, 5, 6)]
+    print('IMAGE_LIST_AUTOGROW_OK real parser; value_N slots shared/per-item')
+    assert all(out.is_output_list for out in end_schema.outputs)
+    assert end_schema.is_input_list and end_schema.is_output_node
+    print('IMAGE_LIST_SCHEMA_OK exact IDs, names, typed list outputs')
     print('SCHEMA_PARITY_OK', len(mapping))
     check_saved_workflows(mapping)
     await check_execution(mapping, module, prompt_server, asset_manager)
+    await check_image_list_lazy_executor(mapping, module, prompt_server, asset_manager)
     print('V3_ACCEPTANCE_OK')
 
 

tests/test_async_helpers.py

@@ -3,6 +3,7 @@ import sys
 import types
 import unittest
 from pathlib import Path
+from unittest.mock import AsyncMock, patch
 
 
 class _PromptQueue:
@@ -87,5 +88,59 @@ class QueuePromptPayloadTests(unittest.IsolatedAsyncioTestCase):
         self.assertEqual(extra_data["extra_pnginfo"]["workflow"], {"id": "workflow-1"})
 
 
+class ExplicitTargetTests(unittest.IsolatedAsyncioTestCase):
+    def setUp(self):
+        self.module, self.server = _load_async_helpers_module()
+        self.prompt = {"9": {"class_type": "Output"}, "10": {"class_type": "Output"}}
+        self.validate = AsyncMock(return_value=(True, None, ["9"], {}))
+        self.module.execution.validate_prompt = self.validate
+
+    async def test_targets_survive_callback_omission_and_are_queued_exactly(self):
+        seen = []
+        def callback(payload):
+            seen.append(payload["partial_execution_targets"][:])
+            payload["partial_execution_targets"].append("10")
+            return {"prompt": payload["prompt"]}
+        self.server.trigger_on_prompt = callback
+        targets = ["9"]
+        await self.module.queue_prompt_payload(self.prompt, partial_execution_targets=targets)
+        self.assertEqual(seen, [["9"]])
+        self.assertEqual(targets, ["9"])
+        self.assertEqual(self.validate.call_args.args[2], ["9"])
+        self.assertEqual(self.server.prompt_queue.items[0][4], ["9"])
+
+    async def test_invalid_requests_never_reach_validation_or_queue(self):
+        for targets in ([], "9", [9], [True], ["9", "9"], ["missing"], {}):
+            with self.subTest(targets=targets):
+                with self.assertRaises(self.module.PromptValidationError):
+                    await self.module.queue_prompt_payload(self.prompt, partial_execution_targets=targets)
+        self.validate.assert_not_awaited()
+        self.assertEqual(self.server.prompt_queue.items, [])
+
+    async def test_removed_target_after_callback_is_rejected(self):
+        self.server.trigger_on_prompt = lambda payload: {"prompt": {"10": self.prompt["10"]}}
+        with self.assertRaises(self.module.PromptValidationError):
+            await self.module.queue_prompt_payload(self.prompt, partial_execution_targets=["9"])
+        self.validate.assert_not_awaited()
+        self.assertEqual(self.server.prompt_queue.items, [])
+
+    async def test_partial_subset_extra_roots_and_target_errors_fail_closed(self):
+        for roots, errors in ((["9"], {}), (["9", "10", "extra"], {}),
+                              (["9", "10"], {"9": {"errors": [{"message": "bad"}]}}),
+                              (["9", "10"], {"dep": {"dependent_outputs": ["10"], "errors": []}})):
+            with self.subTest(roots=roots, errors=errors):
+                self.validate.return_value = (True, None, roots, errors)
+                with self.assertRaises(self.module.PromptValidationError):
+                    await self.module.queue_prompt_payload(self.prompt, partial_execution_targets=["9", "10"])
+        self.assertEqual(self.server.prompt_queue.items, [])
+
+    async def test_none_preserves_legacy_payload_and_validation(self):
+        seen = []
+        self.server.trigger_on_prompt = lambda payload: seen.append(payload.copy()) or payload
+        await self.module.queue_prompt_payload(self.prompt)
+        self.assertNotIn("partial_execution_targets", seen[0])
+        self.assertIsNone(self.validate.call_args.args[2])
+
+
 if __name__ == "__main__":
     unittest.main()

tests/test_collector_exact_transfer.py

@@ -0,0 +1,524 @@
+"""Exact Collector transport through real serializers, HTTP, route and queue.
+
+Only ComfyUI's process-wide server/model-management and unrelated orchestration
+are replaced. Tensor/audio codecs, multipart bodies and aiohttp are real; no
+running ComfyUI service, GPU, provider or external network access is required.
+"""
+import asyncio
+from contextlib import asynccontextmanager
+import importlib
+import importlib.util
+import json
+from pathlib import Path
+import sys
+import types
+from unittest.mock import AsyncMock
+
+import pytest
+import torch
+
+
+ROOT = Path(__file__).resolve().parents[1]
+PKG = "dist_collector_exact_testpkg"
+JOB_ID = "exact-job"
+WORKER_ID = "worker-1"
+
+
+@pytest.fixture
+def transport(monkeypatch):
+    # Other legacy test loaders may leave aiohttp stubs behind. Replace only
+    # those entries, never unload/reload real extension-backed dependencies.
+    for name, loaded_module in list(sys.modules.items()):
+        if (name == "aiohttp" or name.startswith("aiohttp.")) and not getattr(loaded_module, "__file__", None):
+            monkeypatch.delitem(sys.modules, name)
+    aiohttp = importlib.import_module("aiohttp")
+    web = importlib.import_module("aiohttp.web")
+
+    def module(name, **attrs):
+        result = types.ModuleType(name)
+        result.__dict__.update(attrs)
+        monkeypatch.setitem(sys.modules, name, result)
+        return result
+
+    def load(relative):
+        name = f"{PKG}.{relative}"
+        spec = importlib.util.spec_from_file_location(name, ROOT.joinpath(*relative.split(".")).with_suffix(".py"))
+        assert spec is not None and spec.loader is not None
+        result = importlib.util.module_from_spec(spec)
+        monkeypatch.setitem(sys.modules, name, result)
+        spec.loader.exec_module(result)
+        parent, _, child = name.rpartition(".")
+        setattr(sys.modules[parent], child, result)
+        return result
+
+    module(PKG, __path__=[str(ROOT)])
+    for package in ("nodes", "api", "utils", "serverless_bridge"):
+        module(f"{PKG}.{package}", __path__=[str(ROOT / package)])
+    server = types.SimpleNamespace(routes=web.RouteTableDef(), distributed_pending_jobs={})
+    module("server", PromptServer=types.SimpleNamespace(instance=server))
+    mm = module("comfy.model_management", throw_exception_if_processing_interrupted=lambda: None)
+    progress = module("comfy.utils", ProgressBar=lambda _total: types.SimpleNamespace(update=lambda _n: None))
+    module("comfy", __path__=[], model_management=mm, utils=progress)
+    module(f"{PKG}.utils.logging", debug_log=lambda *_a, **_kw: None, log=lambda *_a, **_kw: None)
+    module(f"{PKG}.utils.config", get_worker_timeout_seconds=lambda: 0.1,
+           load_config=lambda: {"workers": []}, is_master_delegate_only=lambda: False)
+
+    async def handle_api_error(_request, error, status=500):
+        return web.json_response({"error": str(error)}, status=status)
+
+    module(f"{PKG}.utils.network", handle_api_error=handle_api_error,
+           get_client_session=AsyncMock(), build_worker_url=lambda _worker: "http://unused.invalid",
+           probe_worker=AsyncMock(return_value=None))
+    module(f"{PKG}.utils.async_helpers", run_async_in_server_loop=asyncio.run)
+    module(f"{PKG}.api.queue_orchestration", ensure_distributed_state=lambda: None,
+           orchestrate_distributed_execution=AsyncMock())
+    module(f"{PKG}.api.queue_request", parse_queue_request_payload=lambda value: value)
+    for name in ("utils.constants", "utils.image", "utils.audio_payload", "utils.transfer_codec"):
+        load(name)
+    bridge_codec = load("serverless_bridge.codec")
+    load("serverless_bridge.image_transfer")
+    image_codec = load("serverless_bridge.image_codec")
+    collector = load("nodes.collector")
+    routes = load("api.job_routes")
+    return types.SimpleNamespace(aiohttp=aiohttp, web=web, server=server,
+                                 collector=collector, routes=routes, bridge_codec=bridge_codec,
+                                 image_codec=image_codec)
+
+
+@asynccontextmanager
+async def endpoint(transport, monkeypatch, *, request_limit=2 * 1024 * 1024,
+                   chunk_target=1024 * 1024, client_max_size=None, handler=None):
+    """A disposable loopback endpoint; record actual HTTP lengths, not estimates."""
+    queue = asyncio.Queue()
+    transport.server.distributed_jobs_lock = asyncio.Lock()
+    transport.server.distributed_pending_jobs = {JOB_ID: queue}
+    requests = []
+    web = transport.web
+
+    @web.middleware
+    async def record_request(request, handler):
+        record = {"bytes": request.content_length, "content_type": request.content_type}
+        requests.append(record)
+        response = await handler(request)
+        record["status"] = response.status
+        if response.status < 400 and request.content_type.startswith("multipart/"):
+            form = await request.post()
+            record["envelope"] = json.loads(form["envelope"])
+        return response
+
+    app = web.Application(client_max_size=client_max_size or request_limit, middlewares=[record_request])
+    app.router.add_post("/distributed/job_complete", handler or transport.routes.job_complete_endpoint)
+    runner = web.AppRunner(app)
+    await runner.setup()
+    try:
+        site = web.TCPSite(runner, "127.0.0.1", 0)
+        await site.start()
+        port = site._server.sockets[0].getsockname()[1]
+        async with transport.aiohttp.ClientSession() as session:
+            monkeypatch.setattr(transport.collector, "get_client_session", AsyncMock(return_value=session))
+            monkeypatch.setattr(transport.collector, "TRANSFER_MAX_REQUEST_BYTES", request_limit)
+            monkeypatch.setattr(transport.collector, "TRANSFER_MAX_CHUNK_BYTES", chunk_target)
+            yield types.SimpleNamespace(
+                session=session, url=f"http://127.0.0.1:{port}", requests=requests, queue=queue,
+                node=transport.collector.DistributedCollectorNode(),
+            )
+    finally:
+        await runner.cleanup()
+
+
+def frames(dtype=torch.float32, channels=3, count=5, height=7, width=9):
+    # Non-contiguous, out-of-range and non-8-bit-representable values exercise
+    # detach/contiguous without permitting clipping, alpha loss or quantization.
+    values = torch.arange(count * height * width * channels, dtype=torch.float64)
+    values = (values.remainder(41) - 11) / 17
+    images = values.reshape(count, width, height, channels).transpose(1, 2).to(dtype)
+    images[0, 0, 0, 0] = -0.0
+    return images.requires_grad_(True)
+
+
+def assert_exact(actual, expected):
+    assert actual.dtype == expected.dtype
+    assert actual.shape == expected.shape
+    assert actual.device.type == "cpu"
+    assert actual.is_contiguous()
+    # Compare raw bits as well as values, including BF16 and signed zero.
+    assert torch.equal(actual, expected)
+    assert torch.equal(actual.view(torch.uint8), expected.detach().contiguous().view(torch.uint8))
+
+
+def queued_items(queue):
+    return [queue.get_nowait() for _ in range(queue.qsize())]
+
+
+async def send(ctx, images, *, audio=None, fmt="exact"):
+    await ctx.node.send_batch_to_master(images, audio, JOB_ID, ctx.url, WORKER_ID, fmt, 17)
+
+
+@pytest.mark.parametrize("delegate_only", [False, True])
+@pytest.mark.parametrize("dtype", [torch.float16, torch.bfloat16, torch.float32, torch.float64])
+@pytest.mark.parametrize("channels", [1, 3, 4])
+def test_exact_sender_route_and_collector_preserve_dtype_values_and_channels(
+        transport, monkeypatch, dtype, channels, delegate_only):
+    images = frames(dtype, channels)
+    frame_bytes = images[0].numel() * images.element_size()
+    reserve = transport.collector.DistributedCollectorNode.TRANSFER_ENVELOPE_RESERVE_BYTES
+
+    async def scenario():
+        async with endpoint(transport, monkeypatch, request_limit=reserve + 2 * frame_bytes,
+                            chunk_target=2 * frame_bytes) as ctx:
+            await send(ctx, images)
+            items = queued_items(ctx.queue)
+            assert [item["image_index"] for item in items] == list(range(len(images)))
+            assert [item["is_last"] for item in items] == [False] * 4 + [True]
+            assert all(item["worker_id"] == WORKER_ID and item["format"] == "exact" for item in items)
+            for item, expected in zip(items, images.split(1)):
+                assert_exact(item["tensor"], expected)
+                await ctx.queue.put(item)
+            envelopes = [record["envelope"] for record in ctx.requests]
+            assert [(e["start_index"], e["count"]) for e in envelopes] == [(0, 2), (2, 2), (4, 1)]
+            assert all(record["bytes"] <= reserve + 2 * frame_bytes for record in ctx.requests)
+            master_images = None if delegate_only else images[:1].detach().clone().add(0.03125)
+            actual, _audio = await ctx.node.execute(
+                master_images, None, multi_job_id=JOB_ID, enabled_worker_ids=json.dumps([WORKER_ID]),
+                delegate_only=delegate_only, transfer_format="exact",
+            )
+            expected = images if delegate_only else torch.cat([master_images, images])
+            assert_exact(actual, expected)
+            assert JOB_ID not in transport.server.distributed_pending_jobs
+
+    asyncio.run(scenario())
+
+
+def test_exact_single_frame_can_exceed_chunk_target_but_not_request_limit(transport, monkeypatch):
+    images = frames(count=3)
+    frame_bytes = images[0].numel() * images.element_size()
+    reserve = transport.collector.DistributedCollectorNode.TRANSFER_ENVELOPE_RESERVE_BYTES
+
+    async def scenario():
+        async with endpoint(transport, monkeypatch, request_limit=reserve + frame_bytes,
+                            chunk_target=frame_bytes // 2) as ctx:
+            await send(ctx, images)
+            assert [record["envelope"]["count"] for record in ctx.requests] == [1, 1, 1]
+            assert all(record["bytes"] <= reserve + frame_bytes for record in ctx.requests)
+            assert_exact(torch.cat([item["tensor"] for item in queued_items(ctx.queue)]), images)
+
+    asyncio.run(scenario())
+
+
+def test_exact_oversized_frame_is_rejected_before_encoding_or_post(transport, monkeypatch):
+    images = frames(count=3)
+    frame_bytes = images[0].numel() * images.element_size()
+    reserve = transport.collector.DistributedCollectorNode.TRANSFER_ENVELOPE_RESERVE_BYTES
+    encode = AsyncMock(side_effect=AssertionError("oversized frame must not be encoded"))
+    monkeypatch.setattr(transport.collector, "_encode_exact_frames", encode)
+
+    async def scenario():
+        async with endpoint(transport, monkeypatch, request_limit=reserve + frame_bytes - 1) as ctx:
+            with pytest.raises(transport.collector.transfer_codec.FrameTooLargeError, match="request limit"):
+                await send(ctx, images)
+            assert ctx.requests == []
+            assert ctx.queue.empty()
+            encode.assert_not_called()
+
+    asyncio.run(scenario())
+
+
+@pytest.mark.parametrize("shape", [(0, 2, 3, 3), (1, 0, 3, 3), (1, 2, 0, 3), (1, 2, 3, 0),
+                                    (1, 2, 3, 2), (1, 2, 3, 5), (1, 2, 3), (2, 3)])
+def test_exact_sender_rejects_invalid_geometry_before_post(transport, monkeypatch, shape):
+    image = torch.empty(shape)
+    with pytest.raises(transport.collector.transfer_codec.TransferCodecError):
+        transport.collector._exact_frame_bytes(image)
+
+    async def scenario():
+        async with endpoint(transport, monkeypatch) as ctx:
+            # An empty batch is also a no-media completion error (ValueError).
+            with pytest.raises(ValueError):
+                await send(ctx, image)
+            assert ctx.requests == []
+            assert ctx.queue.empty()
+
+    asyncio.run(scenario())
+
+
+@pytest.mark.parametrize("dtype", [torch.uint8, torch.int32, torch.bool, torch.complex64])
+def test_exact_sender_rejects_unsupported_dtype_before_post(transport, monkeypatch, dtype):
+    async def scenario():
+        async with endpoint(transport, monkeypatch) as ctx:
+            with pytest.raises(transport.collector.transfer_codec.TransferCodecError, match="dtype"):
+                await send(ctx, torch.zeros(1, 2, 3, 3, dtype=dtype))
+            assert ctx.requests == []
+
+    asyncio.run(scenario())
+
+
+def test_exact_frame_planning_enforces_bridge_byte_limit(transport, monkeypatch):
+    monkeypatch.setattr(transport.bridge_codec, "MAX_PAYLOAD_BYTES", 32)
+    with pytest.raises(transport.collector.transfer_codec.TransferCodecError, match="[Bb]yte|limit"):
+        transport.collector._exact_frame_bytes(torch.zeros(1, 2, 3, 3))
+
+
+def test_exact_frame_planning_rejects_non_dense_layout(transport):
+    image = torch.zeros(1, 2, 3, 3).to_sparse()
+    with pytest.raises(transport.collector.transfer_codec.TransferCodecError, match="dense|layout|strided"):
+        transport.collector._exact_frame_bytes(image)
+
+
+@pytest.mark.parametrize("value", [float("nan"), float("inf"), -float("inf")])
+def test_exact_sender_rejects_nonfinite_payload(transport, monkeypatch, value):
+    images = torch.full((1, 2, 3, 3), value)
+
+    async def scenario():
+        async with endpoint(transport, monkeypatch) as ctx:
+            with pytest.raises(transport.collector.transfer_codec.TransferCodecError, match="finite"):
+                await send(ctx, images)
+            assert ctx.requests == []
+            assert ctx.queue.empty()
+
+    asyncio.run(scenario())
+
+
+def exact_chunk(transport, count=2):
+    images = frames(count=count).detach()
+    encoded = [transport.bridge_codec.encode_payload("IMAGE", frame) for frame in images.split(1)]
+    return {
+        "job_id": JOB_ID, "worker_id": WORKER_ID, "format": "exact",
+        "start_index": 0, "count": count, "width": images.shape[2], "height": images.shape[1],
+        "is_last": True, "descriptors": [descriptor for descriptor, _ in encoded],
+    }, [payload for _, payload in encoded]
+
+
+@pytest.mark.parametrize("case", [
+    "zero_count", "bool_width", "negative_height", "negative_index", "bool_index",
+    "missing_descriptor", "extra_descriptor", "missing_part", "extra_part", "not_a_descriptor",
+    "wrong_height", "wrong_batch", "bad_channels", "bad_dtype", "bad_layout", "bad_encoding",
+    "extra_descriptor_field", "truncated", "trailing_bytes", "nonfinite", "compressed_spoof",
+])
+def test_exact_route_rejects_bad_geometry_and_payload_atomically(transport, monkeypatch, case):
+    envelope, parts = exact_chunk(transport)
+    descriptor = envelope["descriptors"][-1]
+    if case == "zero_count":
+        envelope["count"] = 0
+    elif case == "bool_width":
+        envelope["width"] = True
+    elif case == "negative_height":
+        envelope["height"] = -1
+    elif case == "negative_index":
+        envelope["start_index"] = -1
+    elif case == "bool_index":
+        envelope["start_index"] = True
+    elif case == "missing_descriptor":
+        envelope["descriptors"].pop()
+    elif case == "extra_descriptor":
+        envelope["descriptors"].append(dict(descriptor))
+    elif case == "missing_part":
+        parts.pop()
+    elif case == "extra_part":
+        parts.append(parts[-1])
+    elif case == "not_a_descriptor":
+        envelope["descriptors"][-1] = None
+    elif case == "wrong_height":
+        descriptor["shape"][1] += 1
+    elif case == "wrong_batch":
+        descriptor["shape"][0] = 2
+    elif case == "bad_channels":
+        descriptor["shape"][-1] = 2
+    elif case == "bad_dtype":
+        descriptor["dtype"] = "uint8"
+    elif case == "bad_layout":
+        descriptor["layout"] = "BCHW"
+    elif case == "bad_encoding":
+        descriptor["encoding"] = "pickle"
+    elif case == "extra_descriptor_field":
+        descriptor["unknown"] = True
+    elif case == "truncated":
+        parts[-1] = parts[-1][:-1]
+    elif case == "trailing_bytes":
+        parts[-1] += b"\0"
+    elif case == "nonfinite":
+        bad = torch.full(tuple(descriptor["shape"]), float("nan"))
+        parts[-1] = bad.numpy().tobytes()
+    elif case == "compressed_spoof":
+        # A valid serverless image-transfer descriptor must NOT be accepted as
+        # exact by the direct route, even though the shared bridge can decode it.
+        envelope["descriptors"][-1], parts[-1] = transport.image_codec.encode_image(
+            frames(count=1).detach(), transfer_format="png", transfer_quality=90,
+        )
+        # Prove the spoof is a valid compressed bridge payload, not simply junk.
+        decoded = transport.bridge_codec.decode_payload("IMAGE", envelope["descriptors"][-1], parts[-1])
+        assert list(decoded.shape) == descriptor["shape"]
+
+    async def scenario():
+        async with endpoint(transport, monkeypatch) as ctx:
+            async with ctx.session.post(
+                ctx.url + "/distributed/job_complete",
+                data=ctx.node._build_multipart(envelope, parts),
+            ) as response:
+                body = await response.text()
+                assert response.status == 400, body
+            # A corrupt second frame must not leave the first in the queue.
+            assert ctx.queue.empty()
+
+    asyncio.run(scenario())
+
+
+@pytest.mark.parametrize("status,body", [(400, "Invalid JSON payload: old master"),
+                                        (413, "upload limit"), (500, "unexpected failure")])
+def test_exact_master_rejection_never_falls_back(transport, monkeypatch, status, body):
+    async def reject(_request):
+        return transport.web.Response(status=status, text=body)
+
+    async def scenario():
+        async with endpoint(transport, monkeypatch, handler=reject, chunk_target=1) as ctx:
+            legacy = AsyncMock(side_effect=AssertionError("exact must never use legacy transport"))
+            compressed = AsyncMock(side_effect=AssertionError("exact must never use compressed transport"))
+            monkeypatch.setattr(ctx.node, "_send_legacy_batch", legacy)
+            monkeypatch.setattr(ctx.node, "_send_compressed_batch", compressed)
+            with pytest.raises(RuntimeError, match=f"HTTP {status}") as raised:
+                await send(ctx, frames(count=3))
+            assert "never" in str(raised.value) and "8-bit" in str(raised.value)
+            assert len(ctx.requests) == 1
+            assert ctx.requests[0]["content_type"].startswith("multipart/")
+            assert ctx.queue.empty()
+            legacy.assert_not_called()
+            compressed.assert_not_called()
+
+    asyncio.run(scenario())
+
+
+@pytest.mark.parametrize("fmt", ["legacy_png", "png"])
+def test_exact_collector_rejects_old_worker_8bit_result(transport, monkeypatch, fmt):
+    async def scenario():
+        async with endpoint(transport, monkeypatch) as ctx:
+            await send(ctx, frames(count=1).detach(), fmt=fmt)
+            assert ctx.queue.qsize() == 1
+            with pytest.raises(RuntimeError, match="requested exact"):
+                await ctx.node.execute(
+                    None, None, multi_job_id=JOB_ID, enabled_worker_ids=json.dumps([WORKER_ID]),
+                    delegate_only=True, transfer_format="exact",
+                )
+            assert JOB_ID not in transport.server.distributed_pending_jobs
+
+    asyncio.run(scenario())
+
+
+def test_exact_collector_shape_mismatch_raises_instead_of_returning_local_images(transport, monkeypatch):
+    async def scenario():
+        async with endpoint(transport, monkeypatch) as ctx:
+            await send(ctx, frames(count=1, channels=4))
+            with pytest.raises(RuntimeError, match="shape|combin|mismatch"):
+                await ctx.node.execute(
+                    frames(count=1, channels=3), None, multi_job_id=JOB_ID,
+                    enabled_worker_ids=json.dumps([WORKER_ID]), transfer_format="exact",
+                )
+            assert JOB_ID not in transport.server.distributed_pending_jobs
+
+    asyncio.run(scenario())
+
+
+def test_exact_incomplete_worker_raises_and_cleans_queue(transport, monkeypatch):
+    """A timed-out worker's first frame is not a successful partial result."""
+    envelope, parts = exact_chunk(transport, count=1)
+    envelope["is_last"] = False
+
+    async def scenario():
+        async with endpoint(transport, monkeypatch) as ctx:
+            async with ctx.session.post(
+                ctx.url + "/distributed/job_complete", data=ctx.node._build_multipart(envelope, parts),
+            ) as response:
+                assert response.status == 200, await response.text()
+            assert ctx.queue.qsize() == 1
+            with pytest.raises(RuntimeError, match="Exact transfer incomplete"):
+                await asyncio.wait_for(ctx.node.execute(
+                    frames(count=1).detach(), None, multi_job_id=JOB_ID,
+                    enabled_worker_ids=json.dumps([WORKER_ID]), transfer_format="exact",
+                ), timeout=2)
+            assert JOB_ID not in transport.server.distributed_pending_jobs
+
+    asyncio.run(scenario())
+
+
+def test_exact_oversized_multipart_metadata_is_rejected_before_post(transport, monkeypatch):
+    images = frames(count=1)
+    reserve = transport.collector.DistributedCollectorNode.TRANSFER_ENVELOPE_RESERVE_BYTES
+    limit = reserve + images.numel() * images.element_size()
+
+    async def scenario():
+        async with endpoint(transport, monkeypatch, request_limit=limit) as ctx:
+            with pytest.raises(transport.collector.transfer_codec.FrameTooLargeError, match="request limit"):
+                await ctx.node.send_batch_to_master(
+                    images, None, JOB_ID, ctx.url, "w" * limit, "exact", 90,
+                )
+            assert ctx.requests == []
+            assert ctx.queue.empty()
+
+    asyncio.run(scenario())
+
+
+def test_exact_actual_server_upload_rejection_is_413_without_retry(transport, monkeypatch):
+    async def scenario():
+        async with endpoint(transport, monkeypatch, client_max_size=1024) as ctx:
+            with pytest.raises(RuntimeError, match="HTTP 413"):
+                await send(ctx, frames(count=3, height=64, width=64))
+            assert len(ctx.requests) == 1
+            assert ctx.requests[0]["status"] == 413
+            assert ctx.queue.empty()
+
+    asyncio.run(scenario())
+
+
+def test_exact_oversized_audio_is_rejected_before_any_image_post(transport, monkeypatch):
+    images = frames(count=3)
+    audio = {"waveform": torch.zeros(1, 1, 100000), "sample_rate": 48000}
+    reserve = transport.collector.DistributedCollectorNode.TRANSFER_ENVELOPE_RESERVE_BYTES
+
+    async def scenario():
+        async with endpoint(transport, monkeypatch, request_limit=reserve + 1024) as ctx:
+            with pytest.raises(transport.collector.transfer_codec.FrameTooLargeError, match="audio.*request limit"):
+                await send(ctx, images, audio=audio)
+            assert ctx.requests == []
+            assert ctx.queue.empty()
+
+    asyncio.run(scenario())
+
+
+def test_exact_tiny_frames_cannot_overflow_multipart_metadata_budget(transport, monkeypatch):
+    images = torch.full((2048, 1, 1, 1), 0.125, dtype=torch.float16)
+    reserve = transport.collector.DistributedCollectorNode.TRANSFER_ENVELOPE_RESERVE_BYTES
+    limit = reserve + images.numel() * images.element_size()
+
+    async def scenario():
+        async with endpoint(transport, monkeypatch, request_limit=limit, chunk_target=limit) as ctx:
+            await send(ctx, images)
+            assert len(ctx.requests) > 1
+            assert all(record["bytes"] <= limit for record in ctx.requests)
+            assert_exact(torch.cat([item["tensor"] for item in queued_items(ctx.queue)]), images)
+
+    asyncio.run(scenario())
+
+
+@pytest.mark.parametrize("separate", [False, True])
+def test_exact_audio_is_preserved_and_only_final_completion_is_terminal(transport, monkeypatch, separate):
+    images = frames(count=3)
+    audio = {"waveform": torch.linspace(-0.9, 0.9, 60).reshape(1, 1, -1), "sample_rate": 48000}
+    frame_bytes = images[0].numel() * images.element_size()
+    reserve = transport.collector.DistributedCollectorNode.TRANSFER_ENVELOPE_RESERVE_BYTES
+    media_budget = frame_bytes if separate else frame_bytes + 8192
+    limit = reserve + media_budget
+
+    async def scenario():
+        async with endpoint(transport, monkeypatch, request_limit=limit, chunk_target=frame_bytes) as ctx:
+            await send(ctx, images, audio=audio)
+            items = queued_items(ctx.queue)
+            assert len(ctx.requests) == 3 + int(separate)
+            assert [item["is_last"] for item in items] == [False] * (len(items) - 1) + [True]
+            assert all(item["audio"] is None for item in items[:-1])
+            assert items[-1]["audio"]["sample_rate"] == audio["sample_rate"]
+            assert_exact(items[-1]["audio"]["waveform"], audio["waveform"])
+            assert_exact(torch.cat([item["tensor"] for item in items if item["tensor"] is not None]), images)
+            assert all(record["bytes"] <= limit for record in ctx.requests)
+            if separate:
+                assert items[-1]["tensor"] is None
+                assert ctx.requests[-1]["content_type"] == "application/json"
+
+    asyncio.run(scenario())

tests/test_collector_list_inputs.py

@@ -4,6 +4,7 @@ import types
 import asyncio
 from pathlib import Path
 
+import pytest
 import torch
 
 
@@ -19,6 +20,13 @@ def _register_real_transfer_codec(package_name, utils_pkg):
 
 
 def _load_collector_module():
+    # Restore shared dependencies even when a different test imports this loader.
+    # Keep real modules imported during loading (notably extension modules) cached.
+    with pytest.MonkeyPatch.context() as monkeypatch:
+        return _load_collector_module_with_stubs(monkeypatch)
+
+
+def _load_collector_module_with_stubs(monkeypatch):
     module_path = Path(__file__).resolve().parents[1] / "nodes" / "collector.py"
     package_name = "dist_collector_list_testpkg"
 
@@ -53,7 +61,7 @@ def _load_collector_module():
     )
     server_module = types.ModuleType("server")
     server_module.PromptServer = types.SimpleNamespace(instance=prompt_server)
-    sys.modules["server"] = server_module
+    monkeypatch.setitem(sys.modules, "server", server_module)
 
     comfy_module = types.ModuleType("comfy")
     model_management = types.ModuleType("comfy.model_management")
@@ -77,13 +85,13 @@ def _load_collector_module():
 
     comfy_utils.ProgressBar = ProgressBar
     comfy_module.utils = comfy_utils
-    sys.modules["comfy"] = comfy_module
-    sys.modules["comfy.model_management"] = model_management
-    sys.modules["comfy.utils"] = comfy_utils
+    monkeypatch.setitem(sys.modules, "comfy", comfy_module)
+    monkeypatch.setitem(sys.modules, "comfy.model_management", model_management)
+    monkeypatch.setitem(sys.modules, "comfy.utils", comfy_utils)
 
     aiohttp_module = types.ModuleType("aiohttp")
     aiohttp_module.ClientTimeout = lambda total: types.SimpleNamespace(total=total)
-    sys.modules["aiohttp"] = aiohttp_module
+    monkeypatch.setitem(sys.modules, "aiohttp", aiohttp_module)
 
     logging_module = types.ModuleType(f"{package_name}.utils.logging")
     logging_module.debug_log = lambda *_args, **_kwargs: None

tests/test_collector_typed_media.py

@@ -0,0 +1,112 @@
+"""Real Collector aggregation of typed bridge deliveries and direct results.
+
+Only Comfy's host services are stubbed by the shared Collector loader. Tensor
+normalization, participant ordering, queue consumption and cleanup are real.
+"""
+import asyncio
+import json
+import types
+from unittest.mock import AsyncMock
+
+import pytest
+import torch
+
+from test_collector_list_inputs import _load_collector_module
+
+
+@pytest.fixture
+def collector_module():
+    module = _load_collector_module()
+    module.prompt_server.distributed_jobs_lock = asyncio.Lock()
+    module.prompt_server.distributed_serverless = types.SimpleNamespace(cancel=AsyncMock())
+    return module
+
+
+def native_media(value=0):
+    image = torch.arange(24, dtype=torch.float64).reshape(1, 2, 3, 4) / 997 + value
+    image[0, 0, 0, 0] = -0.0
+    audio = {"waveform": torch.arange(6, dtype=torch.float64).reshape(1, 2, 3) / 997 + value,
+             "sample_rate": 24000}
+    return {"images": image, "audio": audio}
+
+
+def assert_exact(actual, expected):
+    assert actual.dtype == expected.dtype and actual.shape == expected.shape
+    assert torch.equal(actual.contiguous().view(torch.uint8), expected.contiguous().view(torch.uint8))
+
+
+def collect_queue(module, messages, participants, *, local=None, delegate=True):
+    queue = asyncio.Queue()
+    for message in messages:
+        queue.put_nowait(message)
+    module.prompt_server.distributed_pending_jobs["typed-job"] = queue
+    values = local or {"images": None, "audio": None}
+    return asyncio.run(asyncio.wait_for(module.DistributedCollectorNode().execute(
+        **values, multi_job_id="typed-job", enabled_worker_ids=json.dumps(participants),
+        delegate_only=delegate), timeout=2))
+
+
+@pytest.mark.parametrize("delegate", [False, True])
+@pytest.mark.parametrize("socket_set", [("images",), ("audio",), ("images", "audio")])
+def test_typed_and_direct_results_preserve_batch_and_audio_order(collector_module, delegate, socket_set):
+    module = collector_module
+    local, typed_a, typed_b, direct = [native_media(n) for n in range(4)]
+    typed = {socket: [typed_a[socket], typed_b[socket]] for socket in socket_set}
+    direct_message = {"worker_id": "direct", "is_last": True}
+    if "images" in socket_set:
+        direct_message.update(tensor=direct["images"], image_index=0)
+    if "audio" in socket_set:
+        direct_message["audio"] = direct["audio"]
+    messages = [direct_message, {"worker_id": "serverless:pool:1", "media": typed, "is_last": True}]
+    # Delivery order differs from participant order. Duplicate configured IDs
+    # must not duplicate a participant's data or leave collection waiting.
+    participants = ["serverless:pool:1", "direct", "serverless:pool:1"]
+    local_inputs = {socket: local[socket] if socket in socket_set else None for socket in ("images", "audio")}
+    images, audio = collect_queue(module, messages, participants, local=local_inputs, delegate=delegate)
+    expected = ([] if delegate else [local]) + [typed_a, typed_b, direct]
+    if "images" in socket_set:
+        assert_exact(images, torch.cat([row["images"] for row in expected]))
+    else:
+        assert images is None
+    if "audio" in socket_set:
+        assert audio["sample_rate"] == 24000
+        assert_exact(audio["waveform"], torch.cat([row["audio"]["waveform"] for row in expected], dim=-1))
+    else:
+        assert torch.equal(audio["waveform"], module.DistributedCollectorNode.EMPTY_AUDIO["waveform"])
+    assert module.prompt_server.distributed_pending_jobs == {}
+    module.prompt_server.distributed_serverless.cancel.assert_not_awaited()
+
+
+@pytest.mark.parametrize("media", [{}, {"texts": ["not a Collector socket"]}, {"images": []},
+                                      {"audio": [None]}, {"images": "not a list"},
+                                      {"images": [torch.zeros(2, 2, 3)]}])
+def test_invalid_typed_delivery_cancels_and_removes_queue(collector_module, media):
+    module = collector_module
+    with pytest.raises(RuntimeError, match="media|items|IMAGE"):
+        collect_queue(module, [{"worker_id": "serverless:pool:1", "media": media, "is_last": True}],
+                      ["serverless:pool:1"])
+    assert module.prompt_server.distributed_pending_jobs == {}
+    module.prompt_server.distributed_serverless.cancel.assert_awaited_once_with(collector_job_id="typed-job")
+
+
+@pytest.mark.parametrize("failure", ["nonfinal", "unexpected", "duplicate", "provider-error", "shape"])
+def test_typed_delivery_protocol_failures_are_not_partial_success(collector_module, failure):
+    media = {"images": [native_media()["images"]]}
+    first = {"worker_id": "serverless:pool:1", "media": media, "is_last": True}
+    messages = [first]
+    participants = ["serverless:pool:1", "serverless:pool:2"]
+    if failure == "nonfinal":
+        first["is_last"] = False
+    elif failure == "unexpected":
+        first["worker_id"] = "other"
+    elif failure == "duplicate":
+        messages.append(dict(first))
+    elif failure == "provider-error":
+        messages.append({"worker_id": "serverless:pool:2", "serverless_error": "offline failure", "is_last": True})
+    else:
+        messages.append({"worker_id": "serverless:pool:2", "media": {"images": [torch.zeros(1, 4, 5, 3)]},
+                         "is_last": True})
+    with pytest.raises(RuntimeError, match="atomic|Unexpected|Duplicate|failed|shape mismatch"):
+        collect_queue(collector_module, messages, participants)
+    assert collector_module.prompt_server.distributed_pending_jobs == {}
+    collector_module.prompt_server.distributed_serverless.cancel.assert_awaited_once_with(collector_job_id="typed-job")

tests/test_config.py

@@ -18,7 +18,7 @@ def _load_config_module():
             del sys.modules[mod_name]
 
     root_pkg = types.ModuleType(package_name)
-    root_pkg.__path__ = []
+    root_pkg.__path__ = [str(module_path.parent)]
     sys.modules[package_name] = root_pkg
 
     logging_module = types.ModuleType(f"{package_name}.logging")

tests/test_dispatch_selection.py

@@ -4,7 +4,7 @@ import sys
 import types
 import unittest
 from pathlib import Path
-from unittest.mock import patch
+from unittest.mock import AsyncMock, patch
 
 
 def _load_dispatch_module():
@@ -55,7 +55,13 @@ def _load_dispatch_module():
             def __init__(self, total=None):
                 self.total = total
 
-        class _ClientConnectorError(Exception):
+        class _ClientError(Exception):
+            pass
+
+        class _ClientConnectorError(_ClientError):
+            pass
+
+        class _ClientConnectionError(_ClientError):
             pass
 
         class _WSMsgType:
@@ -76,7 +82,9 @@ def _load_dispatch_module():
                 self.closed = True
 
         aiohttp_module.ClientTimeout = _ClientTimeout
+        setattr(aiohttp_module, "ClientError", _ClientError)
         aiohttp_module.ClientConnectorError = _ClientConnectorError
+        setattr(aiohttp_module, "ClientConnectionError", _ClientConnectionError)
         aiohttp_module.WSMsgType = _WSMsgType
         aiohttp_module.TCPConnector = _TCPConnector
         aiohttp_module.ClientSession = _ClientSession
@@ -100,6 +108,113 @@ def _load_dispatch_module():
 dispatch = _load_dispatch_module()
 
 
+def _load_prompt_transform_module():
+    module_path = Path(__file__).resolve().parents[1] / "api" / "orchestration" / "prompt_transform.py"
+    package_name = "prompt_transform_worker_selection_testpkg"
+    for suffix in ("", ".api", ".api.orchestration", ".utils"):
+        package = types.ModuleType(package_name + suffix)
+        package.__path__ = []
+        sys.modules[package_name + suffix] = package
+    logging_module = types.ModuleType(f"{package_name}.utils.logging")
+    setattr(logging_module, "debug_log", lambda *_args, **_kwargs: None)
+    sys.modules[f"{package_name}.utils.logging"] = logging_module
+
+    spec = importlib.util.spec_from_file_location(
+        f"{package_name}.api.orchestration.prompt_transform",
+        module_path,
+    )
+    assert spec is not None and spec.loader is not None
+    module = importlib.util.module_from_spec(spec)
+    spec.loader.exec_module(module)
+    return module
+
+
+prompt_transform = _load_prompt_transform_module()
+
+
+class _FakeResponse:
+    def __init__(self, status=200, payload=None, error=None):
+        self.status = status
+        self.payload = payload
+        self.error = error
+
+    async def __aenter__(self):
+        if self.error is not None:
+            raise self.error
+        return self
+
+    async def __aexit__(self, *_args):
+        return None
+
+    async def json(self):
+        return self.payload
+
+
+class _FakeSession:
+    def __init__(self, responses):
+        self.responses = responses
+
+    def get(self, url, timeout=None):
+        del timeout
+        worker_id = url.removeprefix("http://").split("/", 1)[0]
+        endpoint = "capabilities" if url.endswith("/capabilities") else "queue"
+        response = self.responses[(worker_id, endpoint)]
+        if isinstance(response, BaseException):
+            return _FakeResponse(error=response)
+        status, payload = response
+        return _FakeResponse(status=status, payload=payload)
+
+
+def _valid_capability():
+    return {
+        "supported": True,
+        "execution_targets": True,
+        "protocol_version": 2,
+        "node_ids": ["DistributedImageListStart", "DistributedImageListEnd"],
+    }
+
+
+def _valid_queue(running=None, pending=None):
+    return {"queue_running": running or [], "queue_pending": pending or []}
+
+
+class ExplicitTargetTransportTests(unittest.IsolatedAsyncioTestCase):
+    async def test_http_preserves_targets_and_omits_default_field(self):
+        response = types.SimpleNamespace(raise_for_status=lambda: None)
+        class Response:
+            async def __aenter__(self): return response
+            async def __aexit__(self, *args): pass
+        calls = []
+        def post(url, **kwargs):
+            calls.append(kwargs['json'])
+            return Response()
+        session = types.SimpleNamespace(post=post)
+        with patch.object(dispatch, 'get_client_session', AsyncMock(return_value=session)):
+            await dispatch.dispatch_worker_prompt({'id': 'worker'}, {'9': {}}, None,
+                                                  partial_execution_targets=['9'])
+            await dispatch.dispatch_worker_prompt({'id': 'worker'}, {'9': {}}, None)
+        self.assertEqual(calls[0]['partial_execution_targets'], ['9'])
+        self.assertNotIn('partial_execution_targets', calls[1])
+
+    async def test_websocket_sends_targets_all_the_way_to_wire(self):
+        import json
+        sent = []
+        async def receive():
+            return types.SimpleNamespace(type=dispatch.aiohttp.WSMsgType.TEXT, data=json.dumps({
+                'type': 'dispatch_ack', 'request_id': sent[-1]['request_id'],
+                'ok': True, 'prompt_id': 'worker-prompt'}))
+        ws = types.SimpleNamespace(send_json=AsyncMock(side_effect=sent.append),
+                                   receive=receive, close=AsyncMock())
+        session = types.SimpleNamespace(ws_connect=AsyncMock(return_value=ws))
+        with patch.object(dispatch, 'get_client_session', AsyncMock(return_value=session)):
+            result = await dispatch.dispatch_worker_prompt({'id': 'worker'}, {'9': {}}, None,
+                use_websocket=True, return_prompt_id=True, partial_execution_targets=['9'])
+            await dispatch.dispatch_worker_prompt({'id': 'worker'}, {'9': {}}, None, use_websocket=True)
+        self.assertEqual(result, 'worker-prompt')
+        self.assertEqual(sent[0]['partial_execution_targets'], ['9'])
+        self.assertNotIn('partial_execution_targets', sent[1])
+
+
 class DispatchSelectionTests(unittest.IsolatedAsyncioTestCase):
     async def test_select_active_workers_filters_offline(self):
         workers = [
@@ -249,6 +364,231 @@ class DispatchSelectionTests(unittest.IsolatedAsyncioTestCase):
 
         self.assertIsNone(selected)
 
+    async def test_image_list_preflight_skips_unreachable_worker_and_keeps_healthy_worker(self):
+        workers = [{"id": "healthy"}, {"id": "offline"}]
+        responses = {
+            (worker_id, "capabilities"): (200, _valid_capability())
+            for worker_id in ("healthy", "offline")
+        }
+        responses.update({
+            (worker_id, "queue"): (200, _valid_queue())
+            for worker_id in ("healthy", "offline")
+        })
+        responses[("offline", "capabilities")] = dispatch.aiohttp.ClientConnectionError("connection refused")
+        session = _FakeSession(responses)
+
+        async def fake_session():
+            return session
+
+        with patch.object(dispatch, "get_client_session", side_effect=fake_session), patch.object(
+            dispatch, "build_worker_url", side_effect=lambda worker, path: f"http://{worker['id']}{path}"
+        ), patch.object(dispatch, "log") as log_mock:
+            result = await dispatch.preflight_image_list_workers(workers)
+
+        self.assertEqual(result, {"healthy": True})
+        self.assertTrue(any("offline" in str(call) for call in log_mock.call_args_list))
+
+    async def test_image_list_preflight_skips_timeout_and_keeps_healthy_worker(self):
+        workers = [{"id": "healthy"}, {"id": "slow"}]
+        responses: dict = {
+            (worker_id, "capabilities"): (200, _valid_capability())
+            for worker_id in ("healthy", "slow")
+        }
+        responses.update({
+            (worker_id, "queue"): (200, _valid_queue())
+            for worker_id in ("healthy", "slow")
+        })
+        responses[("slow", "queue")] = asyncio.TimeoutError()
+        session = _FakeSession(responses)
+
+        async def fake_session():
+            return session
+
+        with patch.object(dispatch, "get_client_session", side_effect=fake_session), patch.object(
+            dispatch, "build_worker_url", side_effect=lambda worker, path: f"http://{worker['id']}{path}"
+        ), patch.object(dispatch, "log") as log_mock:
+            result = await dispatch.preflight_image_list_workers(workers)
+
+        self.assertEqual(result, {"healthy": True})
+        self.assertTrue(any("slow" in str(call) for call in log_mock.call_args_list))
+
+    async def test_image_list_preflight_excludes_busy_workers(self):
+        workers = [{"id": "idle"}, {"id": "busy"}]
+        responses = {
+            (worker_id, "capabilities"): (200, _valid_capability())
+            for worker_id in ("idle", "busy")
+        }
+        responses.update({
+            ("idle", "queue"): (200, _valid_queue()),
+            ("busy", "queue"): (200, _valid_queue(running=[{"prompt": "existing"}])),
+        })
+        session = _FakeSession(responses)
+
+        async def fake_session():
+            return session
+
+        with patch.object(dispatch, "get_client_session", side_effect=fake_session), patch.object(
+            dispatch, "build_worker_url", side_effect=lambda worker, path: f"http://{worker['id']}{path}"
+        ):
+            result = await dispatch.preflight_image_list_workers(workers)
+
+        self.assertEqual(result, {"idle": True})
+
+    async def test_image_list_preflight_reports_no_idle_workers_when_all_are_unavailable(self):
+        workers = [{"id": "offline-a"}, {"id": "offline-b"}]
+        responses = {
+            (worker_id, "capabilities"): dispatch.aiohttp.ClientConnectionError("unreachable")
+            for worker_id in ("offline-a", "offline-b")
+        }
+        session = _FakeSession(responses)
+
+        async def fake_session():
+            return session
+
+        with patch.object(dispatch, "get_client_session", side_effect=fake_session), patch.object(
+            dispatch, "build_worker_url", side_effect=lambda worker, path: f"http://{worker['id']}{path}"
+        ):
+            self.assertEqual(
+                await dispatch.preflight_image_list_workers(workers, require_idle=False),
+                {},
+            )
+            with self.assertRaisesRegex(RuntimeError, "No selected feature-compatible direct worker is idle"):
+                await dispatch.preflight_image_list_workers(workers, require_idle=True)
+
+    async def test_image_list_preflight_requires_exact_execution_targets_support(self):
+        for value in (None, False, 1, "true", []):
+            with self.subTest(value=value):
+                capability = _valid_capability()
+                if value is None:
+                    capability.pop("execution_targets")
+                else:
+                    capability["execution_targets"] = value
+                session = _FakeSession({("worker", "capabilities"): (200, capability)})
+                async def fake_session():
+                    return session
+                with patch.object(dispatch, "get_client_session", side_effect=fake_session), patch.object(
+                    dispatch, "build_worker_url", side_effect=lambda worker, path: f"http://{worker['id']}{path}"
+                ):
+                    with self.assertRaisesRegex(RuntimeError, "execution_targets.*upgrade"):
+                        await dispatch.preflight_image_list_workers([{"id": "worker"}])
+
+    async def test_image_list_value_results_requires_exact_true_only_for_overrides(self):
+        for value in (None, False, 1, "true", [], True):
+            capability = _valid_capability()
+            if value is not None:
+                capability['value_results'] = value
+            session = _FakeSession({('worker', 'capabilities'): (200, capability),
+                                    ('worker', 'queue'): (200, _valid_queue())})
+            async def fake_session():
+                return session
+            with patch.object(dispatch, 'get_client_session', side_effect=fake_session), patch.object(
+                dispatch, 'build_worker_url', side_effect=lambda worker, path: f"http://{worker['id']}{path}"
+            ):
+                self.assertEqual(await dispatch.preflight_image_list_workers([{'id': 'worker'}]), {'worker': True})
+                if value is True:
+                    self.assertEqual(await dispatch.preflight_image_list_workers(
+                        [{'id': 'worker'}], require_value_results=True), {'worker': True})
+                else:
+                    with self.assertRaisesRegex(RuntimeError, 'value_results.*upgrade'):
+                        await dispatch.preflight_image_list_workers([{'id': 'worker'}], require_value_results=True)
+
+    async def test_image_list_preflight_fails_closed_for_malformed_capabilities(self):
+        malformed = {"supported": True, "protocol_version": 2, "node_ids": None}
+        session = _FakeSession({("worker", "capabilities"): (200, malformed)})
+
+        async def fake_session():
+            return session
+
+        with patch.object(dispatch, "get_client_session", side_effect=fake_session), patch.object(
+            dispatch, "build_worker_url", side_effect=lambda worker, path: f"http://{worker['id']}{path}"
+        ):
+            with self.assertRaisesRegex(RuntimeError, "malformed.*capabilit"):
+                await dispatch.preflight_image_list_workers([{"id": "worker"}])
+
+    async def test_image_list_preflight_fails_closed_for_incompatible_protocol(self):
+        incompatible = _valid_capability()
+        incompatible["protocol_version"] = 1
+        session = _FakeSession({("worker", "capabilities"): (200, incompatible)})
+
+        async def fake_session():
+            return session
+
+        with patch.object(dispatch, "get_client_session", side_effect=fake_session), patch.object(
+            dispatch, "build_worker_url", side_effect=lambda worker, path: f"http://{worker['id']}{path}"
+        ):
+            with self.assertRaisesRegex(RuntimeError, "incompatible"):
+                await dispatch.preflight_image_list_workers([{"id": "worker"}])
+
+    async def test_image_list_preflight_fails_closed_for_missing_required_node(self):
+        incompatible = _valid_capability()
+        incompatible["node_ids"] = ["DistributedImageListStart"]
+        session = _FakeSession({("worker", "capabilities"): (200, incompatible)})
+
+        async def fake_session():
+            return session
+
+        with patch.object(dispatch, "get_client_session", side_effect=fake_session), patch.object(
+            dispatch, "build_worker_url", side_effect=lambda worker, path: f"http://{worker['id']}{path}"
+        ):
+            with self.assertRaisesRegex(RuntimeError, "incompatible"):
+                await dispatch.preflight_image_list_workers([{"id": "worker"}])
+
+    async def test_image_list_preflight_fails_closed_for_malformed_queue(self):
+        responses = {
+            ("worker", "capabilities"): (200, _valid_capability()),
+            ("worker", "queue"): (200, {"queue_running": [], "queue_pending": None}),
+        }
+        session = _FakeSession(responses)
+
+        async def fake_session():
+            return session
+
+        with patch.object(dispatch, "get_client_session", side_effect=fake_session), patch.object(
+            dispatch, "build_worker_url", side_effect=lambda worker, path: f"http://{worker['id']}{path}"
+        ):
+            with self.assertRaisesRegex(RuntimeError, "malformed /queue"):
+                await dispatch.preflight_image_list_workers([{"id": "worker"}])
+
+    async def test_image_list_preflight_does_not_hide_unexpected_errors(self):
+        session = _FakeSession({("worker", "capabilities"): ValueError("programming error")})
+
+        async def fake_session():
+            return session
+
+        with patch.object(dispatch, "get_client_session", side_effect=fake_session), patch.object(
+            dispatch, "build_worker_url", side_effect=lambda worker, path: f"http://{worker['id']}{path}"
+        ):
+            with self.assertRaisesRegex(ValueError, "programming error"):
+                await dispatch.preflight_image_list_workers([{"id": "worker"}])
+
+
+class ImageListDirectWorkerResolutionTests(unittest.TestCase):
+    def setUp(self):
+        self.config = {"workers": [
+            {"id": "enabled", "enabled": True, "host": "enabled-host", "port": 8188},
+            {"id": "disabled", "enabled": False, "host": "disabled-host", "port": 8189},
+            {"id": "serverless", "enabled": True, "type": "serverless"},
+            {"id": "other", "enabled": True, "host": "other-host", "port": 8190},
+        ]}
+
+    def test_implicit_selection_uses_only_enabled_direct_workers(self):
+        workers = prompt_transform._resolve_direct_workers_for_image_list(self.config, None)
+        self.assertEqual([worker["id"] for worker in workers], ["enabled", "other"])
+
+    def test_explicit_selection_includes_exact_disabled_direct_worker(self):
+        workers = prompt_transform._resolve_direct_workers_for_image_list(self.config, ["disabled"])
+        self.assertEqual([worker["id"] for worker in workers], ["disabled"])
+
+    def test_explicit_empty_selection_returns_no_workers(self):
+        self.assertEqual(prompt_transform._resolve_direct_workers_for_image_list(self.config, []), [])
+
+    def test_unknown_explicit_id_raises_instead_of_falling_back(self):
+        with self.assertRaisesRegex(RuntimeError, "missing"):
+            prompt_transform._resolve_direct_workers_for_image_list(self.config, ["enabled", "missing"])
+
+    def test_explicit_serverless_selection_is_excluded(self):
+        self.assertEqual(prompt_transform._resolve_direct_workers_for_image_list(self.config, ["serverless"]), [])
+
 
 if __name__ == "__main__":
     unittest.main()

tests/test_image_list.py

@@ -0,0 +1,341 @@
+"""Regression tests for typed dynamic image-list normalization and graph scheduling."""
+import asyncio
+import unittest
+
+import torch
+
+from image_list.runtime import (
+    END_CLASS,
+    IMAGE_OUT,
+    CONTEXT_OUT,
+    START_CLASS,
+    VALUE_OUT_BASE,
+    ImageListCancelledError,
+    build_worker_prompt,
+    create_master_job,
+    decode_item,
+    decode_result,
+    dynamic_schedule,
+    encode_item,
+    encode_result,
+    normalize_image_list_inputs,
+    normalize_worker_result,
+    remove_master_job,
+    set_master_result,
+    validate_image_list_regions,
+)
+
+
+def start_node(image_link):
+    return {"class_type": START_CLASS, "inputs": {"images": image_link}}
+
+
+def graph_one_region():
+    return {
+        "1": {"class_type": "CheckpointLoaderSimple", "inputs": {}},
+        "2": {"class_type": "EmptyImage", "inputs": {"height": 8, "width": 8, "batch_size": 2}},
+        "3": start_node(["2", 0]),
+        "4": {"class_type": "GenericFrameOperation", "inputs": {
+            "image": ["3", IMAGE_OUT], "model": ["1", 0], "seed": ["3", VALUE_OUT_BASE],
+        }},
+        "5": {"class_type": END_CLASS, "inputs": {
+            "images": ["4", 0], "context": ["3", CONTEXT_OUT],
+        }},
+        "6": {"class_type": "SaveImage", "inputs": {"images": ["5", 0]}},
+    }
+
+
+def graph_two_sequential_regions():
+    return {
+        "1": {"class_type": "EmptyImage", "inputs": {"height": 4, "width": 4, "batch_size": 2}},
+        "2": start_node(["1", 0]),
+        "3": {"class_type": "FirstFrameOperation", "inputs": {"image": ["2", IMAGE_OUT]}},
+        "4": {"class_type": END_CLASS, "inputs": {"images": ["3", 0], "context": ["2", CONTEXT_OUT]}},
+        "5": start_node(["4", 0]),
+        "6": {"class_type": "SecondFrameOperation", "inputs": {"image": ["5", IMAGE_OUT]}},
+        "7": {"class_type": END_CLASS, "inputs": {"images": ["6", 0], "context": ["5", CONTEXT_OUT]}},
+        "8": {"class_type": "SaveImage", "inputs": {"images": ["7", 0]}},
+    }
+
+
+class ImageListNormalizationTests(unittest.TestCase):
+    def test_batch_frame_normalization_preserves_sizes_channels_and_float_values(self):
+        images = [
+            torch.tensor([[[[0.125, 0.25, 0.5]]]], dtype=torch.float64),
+            torch.arange(2 * 3 * 4, dtype=torch.float32).reshape(1, 2, 3, 4) / 17,
+        ]
+        refs = torch.ones((2, 5, 7, 3), dtype=torch.float16) * 0.333
+        masks = [torch.full((1, 2, 3), 0.125, dtype=torch.float64),
+                 torch.full((1, 8, 9), 0.875, dtype=torch.float32)]
+        rows = normalize_image_list_inputs(images, masks, {
+            "value_1": [refs[:1], refs[1:]], "value_2": ["first", "second"],
+            "value_3": [0.2], "value_4": [37],
+        })
+        self.assertEqual(len(rows), 2)
+        self.assertEqual([tuple(row["image"].shape) for row in rows], [(1, 1, 1, 3), (1, 2, 3, 4)])
+        self.assertEqual([row["image"].dtype for row in rows], [torch.float64, torch.float32])
+        self.assertTrue(torch.equal(rows[0]["values"]["value_1"], refs[:1]))
+        self.assertEqual(rows[1]["mask"].dtype, torch.float32)
+        self.assertEqual([row["index"] for row in rows], [0, 1])
+        self.assertEqual([row["values"]["value_2"] for row in rows], ["first", "second"])
+        self.assertEqual([row["values"]["value_3"] for row in rows], [0.2, 0.2])
+        self.assertEqual([row["values"]["value_4"] for row in rows], [37, 37])
+
+    def test_one_value_is_shared_and_n_values_pair_in_order(self):
+        image = torch.zeros((3, 2, 4, 3), dtype=torch.float32)
+        rows = normalize_image_list_inputs(image, values={
+            "value_1": ["same"], "value_2": [0.4], "value_3": [9], "value_4": [True]})
+        self.assertEqual([row["values"]["value_1"] for row in rows], ["same"] * 3)
+        self.assertEqual([row["values"]["value_2"] for row in rows], [0.4] * 3)
+        self.assertEqual([row["values"]["value_3"] for row in rows], [9] * 3)
+        self.assertEqual([row["values"]["value_4"] for row in rows], [True] * 3)
+        rows = normalize_image_list_inputs(image, values={"value_2": [1, 2, 3]})
+        self.assertEqual([row["values"] for row in rows], [{"value_2": 1}, {"value_2": 2}, {"value_2": 3}])
+        with self.assertRaisesRegex(ValueError, "value_1 has 2 values; connect exactly 1.*or 3"):
+            normalize_image_list_inputs(image, values={"value_1": ["one", "two"]})
+        with self.assertRaisesRegex(ValueError, "mask has 2 values"):
+            normalize_image_list_inputs(image, mask=torch.zeros((2, 2, 4)))
+        shared_mask = normalize_image_list_inputs(image, mask=torch.ones((1, 2, 4)))
+        self.assertTrue(all(torch.equal(row["mask"], torch.ones((1, 2, 4))) for row in shared_mask))
+
+    def test_value_batches_stay_whole_and_only_images_split(self):
+        images = torch.zeros((2, 2, 2, 3), dtype=torch.float32)
+        pair = torch.rand((2, 4, 4, 3))  # e.g. two references for one item
+        rows = normalize_image_list_inputs(images, values={"value_1": [pair]})
+        self.assertTrue(all(torch.equal(row["values"]["value_1"], pair) for row in rows))
+        per_item = [torch.rand((2, 4, 4, 3)), torch.rand((2, 4, 4, 3))]
+        rows = normalize_image_list_inputs(images, values={"value_1": per_item})
+        self.assertTrue(torch.equal(rows[1]["values"]["value_1"], per_item[1]))
+
+    def test_rejects_unsupported_values_and_counts(self):
+        images = torch.zeros((2, 2, 2, 3), dtype=torch.float32)
+        with self.assertRaisesRegex(ValueError, "value_1 has 3 values"):
+            normalize_image_list_inputs(images, values={"value_1": [1, 2, 3]})
+        with self.assertRaisesRegex(TypeError, "unsupported type dict"):
+            normalize_image_list_inputs(images, values={"value_1": [{"samples": 1}]})
+        with self.assertRaisesRegex(TypeError, "floating-point"):
+            normalize_image_list_inputs(images, values={"value_1": [torch.zeros(2, dtype=torch.int64)]})
+        with self.assertRaisesRegex(TypeError, "finite"):
+            normalize_image_list_inputs(images, values={"value_1": [float("nan")]})
+        with self.assertRaisesRegex(ValueError, "Unknown Start value slot"):
+            normalize_image_list_inputs(images, values={"value_11": [1]})
+        with self.assertRaisesRegex(ValueError, "at least one image"):
+            normalize_image_list_inputs(None)
+        # Unconnected Autogrow slots arrive as None and are ignored.
+        rows = normalize_image_list_inputs(images, values={"value_1": [None], "value_2": None})
+        self.assertEqual([row["values"] for row in rows], [{}, {}])
+
+    def test_typed_transport_round_trips_float_images_masks_and_scalar_values(self):
+        item = {"image": torch.arange(24, dtype=torch.float64).reshape(1, 2, 4, 3) / 7,
+                "mask": torch.full((1, 5, 6), 0.375, dtype=torch.float32),
+                "values": {"value_1": torch.ones((2, 1, 3, 4), dtype=torch.float16),
+                           "value_2": "safe text", "value_3": 0.625, "value_4": 42,
+                           "value_5": False},
+                "index": 3}
+        restored = decode_item(encode_item(item))
+        self.assertTrue(torch.equal(restored["image"], item["image"]))
+        self.assertEqual(restored["image"].dtype, torch.float64)
+        self.assertTrue(torch.equal(restored["values"]["value_1"], item["values"]["value_1"]))
+        self.assertEqual(restored["values"]["value_1"].dtype, torch.float16)
+        self.assertTrue(torch.equal(restored["mask"], item["mask"]))
+        self.assertEqual(restored["values"]["value_2"], "safe text")
+        self.assertEqual(restored["values"]["value_3"], 0.625)
+        self.assertIs(restored["values"]["value_4"], 42)
+        self.assertIs(restored["values"]["value_5"], False)
+        with self.assertRaisesRegex(ValueError, "unknown slot"):
+            decode_item({**encode_item(item), "values": {"strength": {"type": "FLOAT", "value": 1.0}}})
+        with self.assertRaisesRegex(ValueError, "unsupported type"):
+            decode_item({**encode_item(item), "values": {"value_1": {"type": "PATH", "value": "/x"}}})
+        with self.assertRaisesRegex(TypeError, "floating-point tensor"):
+            encode_item({**item, "image": torch.ones((1, 2, 4, 3), dtype=torch.int32)})
+        with self.assertRaisesRegex(ValueError, "Unsupported image-list transport type"):
+            from image_list.runtime import encode_value
+            encode_value("/untrusted/path", "PATH")
+
+    def test_result_transport_preserves_ragged_result_shapes(self):
+        result = normalize_worker_result(
+            [torch.ones((1, 3, 5, 4), dtype=torch.float32)],
+            [torch.ones((1, 3, 5), dtype=torch.float16)],
+            ["done"],
+        )
+        encoded = encode_result(result)
+        decoded = decode_result(encoded)
+        self.assertEqual(tuple(decoded["image"].shape), (1, 3, 5, 4))
+        self.assertEqual(decoded["image"].dtype, torch.float32)
+        self.assertEqual(decoded["mask"].dtype, torch.float16)
+        self.assertEqual(decoded["text"], "done")
+        with self.assertRaisesRegex(ValueError, "exactly one IMAGE"):
+            normalize_worker_result(torch.zeros((2, 3, 4, 3)))
+
+
+class ImageListGraphBoundaryTests(unittest.TestCase):
+    def test_prunes_upstream_item_source_but_keeps_external_model_loaders(self):
+        prompt = graph_one_region()
+        region = validate_image_list_regions(prompt)[0]
+        worker = build_worker_prompt(prompt, region, "token_12345678901234567890123456789012",
+                                     "ctx", 1, "http://master.invalid")
+        self.assertEqual(set(worker), {"1", "3", "4", "5"})
+        self.assertEqual(worker["3"]["inputs"]["worker_item_token"], "token_12345678901234567890123456789012")
+        self.assertNotIn("images", worker["3"]["inputs"])
+
+    def test_worker_prompt_drops_per_item_value_links(self):
+        prompt = graph_one_region()
+        prompt["8"] = {"class_type": "StringList", "inputs": {}}
+        prompt["3"]["inputs"].update({"values.value_1": ["8", 0], "values.value_2": ["8", 0]})
+        import sys, types
+        from unittest.mock import patch
+        registry = {"CheckpointLoaderSimple": type("Loader", (), {"RETURN_TYPES": ("MODEL",)})}
+        with patch.dict(sys.modules, {"nodes": types.SimpleNamespace(NODE_CLASS_MAPPINGS=registry)}):
+            region = validate_image_list_regions(prompt)[0]
+        worker = build_worker_prompt(prompt, region, "token_12345678901234567890123456789012", "ctx", 0, "http://m")
+        self.assertFalse([name for name in worker["3"]["inputs"] if name.startswith("values.")])
+        self.assertNotIn("8", worker)
+        self.assertEqual(worker["4"]["inputs"]["model"], ["1", 0])
+        self.assertNotIn("6", worker)
+
+    def test_rejects_upstream_processing_and_leaf_image_bypasses(self):
+        # Both a leaf source and a generated batch must cross Start exactly once.
+        for generated in (False, True):
+            with self.subTest(generated=generated):
+                prompt = graph_one_region()
+                if generated:
+                    prompt["2"] = {"class_type": "VAEDecode", "inputs": {"samples": ["7", 0]}}
+                    prompt["7"] = {"class_type": "KSampler", "inputs": {"model": ["1", 0]}}
+                prompt["4"]["inputs"]["reference"] = ["2", 0]
+                for explicit_targets in (False, True):
+                    with self.subTest(explicit_targets=explicit_targets):
+                        with self.assertRaisesRegex(ValueError, "bypasses Start.*2"):
+                            validate_image_list_regions(prompt, explicit_targets=explicit_targets)
+
+    def test_shared_typed_loader_and_leaf_constant_remain_valid(self):
+        import sys
+        import types
+        from unittest.mock import patch
+        prompt = graph_one_region()
+        prompt["7"] = {"class_type": "SeedConstant", "inputs": {"value": 9}}
+        prompt["2"]["inputs"].update({"model": ["1", 0], "seed": ["7", 0]})
+        prompt["4"]["inputs"]["shared_seed"] = ["7", 0]
+        registry = {
+            "CheckpointLoaderSimple": type("Loader", (), {"RETURN_TYPES": ("MODEL", "CLIP", "VAE")}),
+            "SeedConstant": type("Constant", (), {"RETURN_TYPES": ("INT",)}),
+        }
+        with patch.dict(sys.modules, {"nodes": types.SimpleNamespace(NODE_CLASS_MAPPINGS=registry)}):
+            region = validate_image_list_regions(prompt)[0]
+        worker = build_worker_prompt(prompt, region, "token", "ctx", 0, "http://master")
+        self.assertNotIn("2", worker)
+        self.assertTrue({"1", "7"}.issubset(worker))
+
+    def test_unknown_or_list_producing_shared_dependency_is_not_exempt(self):
+        import sys
+        import types
+        from unittest.mock import patch
+        for node_class in (None, type("ListLoader", (), {
+                "RETURN_TYPES": ("MODEL",), "OUTPUT_IS_LIST": (True,)})):
+            prompt = graph_one_region()
+            prompt["2"]["inputs"]["model"] = ["1", 0]
+            registry = {} if node_class is None else {"CheckpointLoaderSimple": node_class}
+            with patch.dict(sys.modules, {"nodes": types.SimpleNamespace(NODE_CLASS_MAPPINGS=registry)}):
+                with self.assertRaisesRegex(ValueError, "bypasses Start.*1"):
+                    validate_image_list_regions(prompt)
+
+    def test_two_sequential_regions_are_valid_and_isolated(self):
+        regions = validate_image_list_regions(graph_two_sequential_regions())
+        self.assertEqual([(r["start_id"], r["end_id"]) for r in regions], [("2", "4"), ("5", "7")])
+        p = graph_two_sequential_regions()
+        second = build_worker_prompt(p, regions[1], "item_token_123456789012345678901234", "ctx", 0,
+                                     "http://master.invalid")
+        self.assertEqual(set(second), {"5", "6", "7"})
+
+    def test_rejects_missing_context_pair_nested_region_and_cross_boundary_branch(self):
+        prompt = graph_one_region()
+        prompt["5"]["inputs"]["context"] = ["2", 0]
+        for explicit_targets in (False, True):
+            with self.subTest(explicit_targets=explicit_targets):
+                with self.assertRaisesRegex(ValueError, "typed context output"):
+                    validate_image_list_regions(prompt, explicit_targets=explicit_targets)
+        nested = graph_one_region()
+        nested["7"] = start_node(["4", 0])
+        nested["8"] = {"class_type": "NestedOperation", "inputs": {"image": ["7", IMAGE_OUT]}}
+        nested["5"]["inputs"]["images"] = ["8", 0]
+        nested["9"] = {"class_type": END_CLASS, "inputs": {"images": ["8", 0], "context": ["7", CONTEXT_OUT]}}
+        for explicit_targets in (False, True):
+            with self.subTest(explicit_targets=explicit_targets):
+                with self.assertRaisesRegex(ValueError, "Nested/ambiguous"):
+                    validate_image_list_regions(nested, explicit_targets=explicit_targets)
+        outside = graph_one_region()
+        outside["7"] = {"class_type": "PreviewImage", "inputs": {"images": ["3", IMAGE_OUT]}}
+        for explicit_targets in (False, True):
+            with self.subTest(explicit_targets=explicit_targets):
+                with self.assertRaisesRegex(ValueError, "cross-boundary"):
+                    validate_image_list_regions(outside, explicit_targets=explicit_targets)
+
+    def test_rejects_duplicate_or_missing_results(self):
+        prompt = graph_one_region()
+        prompt["5"]["inputs"].pop("images")
+        with self.assertRaisesRegex(ValueError, "images must be connected"):
+            validate_image_list_regions(prompt)
+
+
+class DynamicImageSchedulingTests(unittest.IsolatedAsyncioTestCase):
+    async def test_fast_workers_get_more_items_and_order_is_stable(self):
+        assignments = {"fast": [], "slow": []}
+        async def dispatch(worker, index):
+            assignments[worker].append(index)
+            delay = 0.002 if worker == "fast" else 0.025
+            task = asyncio.create_task(asyncio.sleep(delay, result=(index, f"result-{index}")))
+            return {"index": index, "task": task}
+        async def wait_next(inflight):
+            done, _ = await asyncio.wait([h["task"] for h in inflight.values()], return_when=asyncio.FIRST_COMPLETED)
+            task = next(iter(done))
+            worker = next(w for w, handle in inflight.items() if handle["task"] is task)
+            index, result = task.result()
+            return worker, index, result
+        async def cleanup(_worker, handle):
+            handle["task"].cancel()
+        ordered = await dynamic_schedule(12, ["fast", "slow"], dispatch, wait_next, cleanup)
+        self.assertEqual(ordered, [f"result-{i}" for i in range(12)])
+        self.assertGreater(len(assignments["fast"]), len(assignments["slow"]))
+        self.assertEqual(sorted(assignments["fast"] + assignments["slow"]), list(range(12)))
+
+    async def test_failure_timeout_and_cancel_clean_only_current_assignments(self):
+        async def dispatch(worker, index):
+            return {"index": index, "prompt_id": f"{worker}-{index}"}
+        cancelled = []
+        async def cleanup(worker, handle):
+            cancelled.append((worker, handle["prompt_id"]))
+        async def timeout(_inflight):
+            raise TimeoutError("worker timed out")
+        with self.assertRaisesRegex(TimeoutError, "timed out"):
+            await dynamic_schedule(4, ["one", "two"], dispatch, timeout, cleanup)
+        self.assertCountEqual(cancelled, [("one", "one-0"), ("two", "two-1")])
+
+        cancelled.clear()
+        async def wrong_index(inflight):
+            worker, handle = next(iter(inflight.items()))
+            return worker, handle["index"] + 99, "bad"
+        with self.assertRaisesRegex(RuntimeError, "unexpected image index"):
+            await dynamic_schedule(2, ["one"], dispatch, wrong_index, cleanup)
+        self.assertEqual(len(cancelled), 1)
+
+        cancelled.clear()
+        with self.assertRaises(ImageListCancelledError):
+            await dynamic_schedule(2, ["one", "two"], dispatch, timeout, cleanup,
+                                   interrupted=lambda: True)
+        self.assertEqual(len(cancelled), 0)
+
+    async def test_no_workers_and_duplicate_results_fail_closed(self):
+        async def unused(*_args):
+            self.fail("must not dispatch without workers")
+        with self.assertRaisesRegex(RuntimeError, "No free"):
+            await dynamic_schedule(2, [], unused, unused, unused)
+        async def dispatch(worker, index):
+            return {"index": index}
+        async def duplicate(inflight):
+            worker, _handle = next(iter(inflight.items()))
+            return worker, 0, "same"
+        with self.assertRaisesRegex(RuntimeError, "unexpected image index"):
+            await dynamic_schedule(2, ["one"], dispatch, duplicate, lambda *_: asyncio.sleep(0))
+
+
+if __name__ == "__main__":
+    unittest.main()

tests/test_image_list_execution_plan.py

@@ -0,0 +1,245 @@
+"""CPU-only execution-copy planning; no ComfyUI imports or prompt execution."""
+import copy
+import sys
+import types
+
+import pytest
+
+from image_list.execution_plan import prepare_image_list_execution
+from image_list.runtime import (
+    CONTEXT_OUT,
+    END_CLASS,
+    IMAGE_OUT,
+    START_CLASS,
+    build_worker_prompt,
+    validate_image_list_regions,
+)
+
+
+class OutputNode:
+    OUTPUT_NODE = True
+
+
+class SchemaOutputNode:
+    OUTPUT_NODE = False  # V3's schema, not this fallback, is authoritative.
+
+    @classmethod
+    def GET_NODE_INFO_V1(cls):
+        return {"output_node": True, "output": ("IMAGE",)}
+
+
+class ModelLoader:
+    RETURN_TYPES = ("MODEL", "CLIP", "VAE")
+
+
+@pytest.fixture
+def registry(monkeypatch):
+    classes = {
+        END_CLASS: SchemaOutputNode,
+        "SaveImage": OutputNode,
+        "PreviewImage": OutputNode,
+        "OutputUtility": OutputNode,
+        "SchemaUtility": SchemaOutputNode,
+        "ModelLoader": ModelLoader,
+    }
+    monkeypatch.setitem(sys.modules, "nodes", types.SimpleNamespace(NODE_CLASS_MAPPINGS=classes))
+    return classes
+
+
+def node(class_type, **inputs):
+    return {"class_type": class_type, "inputs": inputs}
+
+
+def graph():
+    return {
+        "loader": node("ModelLoader"),
+        "source": node("ImageSource"),
+        "start": node(START_CLASS, images=["source", 0]),
+        "process": node("FrameOperation", image=["start", IMAGE_OUT], model=["loader", 0]),
+        "end": node(END_CLASS, images=["process", 0], context=["start", CONTEXT_OUT]),
+        "save": node("SaveImage", images=["end", 0]),
+    }
+
+
+def test_discards_multistep_previews_but_preserves_other_output_roots(registry):
+    prompt = graph()
+    prompt.update({
+        "debug-transform": node("DebugTransform", images=["start", IMAGE_OUT]),
+        "debug-preview": node("PreviewImage", images=["debug-transform", 0]),
+        "process-preview": node("PreviewImage", images=["process", 0]),
+        "unrelated-source": node("ImageSource"),
+        "unrelated-preview": node("PreviewImage", images=["unrelated-source", 0]),
+        "upstream-preview": node("PreviewImage", images=["source", 0]),
+        "post-transform": node("FrameOperation", images=["end", 0]),
+        "post-preview": node("PreviewImage", images=["post-transform", 0]),
+    })
+    before = copy.deepcopy(prompt)
+    with pytest.raises(ValueError, match="cross-boundary"):
+        validate_image_list_regions(prompt)
+    prepared, targets = prepare_image_list_execution(prompt)
+    assert set(prepared) == set(prompt) - {"debug-transform", "debug-preview", "process-preview"}
+    assert targets == ["end", "save", "unrelated-preview", "upstream-preview", "post-preview"]
+    validate_image_list_regions(prepared, explicit_targets=True)
+    assert prompt == before
+
+
+@pytest.mark.parametrize("class_type", ["OutputUtility", "SchemaUtility"])
+def test_required_output_utility_keeps_complete_dependency_chain(registry, class_type):
+    prompt = graph()
+    prompt.update({
+        "utility-config": node("Constant", value=7),
+        "utility-first": node(class_type, images=["process", 0], config=["utility-config", 0]),
+        "utility-second": node(class_type, images=["utility-first", 0]),
+        "result": node("FrameOperation", image=["utility-second", 0]),
+        "discarded-preview": node("PreviewImage", images=["utility-first", 0]),
+    })
+    prompt["end"]["inputs"]["images"] = ["result", 0]
+    before = copy.deepcopy(prompt)
+    registered = dict(registry)
+    prepared, targets = prepare_image_list_execution(prompt)
+    assert set(prepared) == set(prompt) - {"discarded-preview"}
+    assert targets == ["end", "save"]
+    with pytest.raises(ValueError, match="Output node"):
+        validate_image_list_regions(prepared)
+    region = validate_image_list_regions(prepared, explicit_targets=True)[0]
+    worker = build_worker_prompt(prepared, region, "token", "context", 0, "http://master.invalid")
+    assert set(worker) == set(prepared) - {"source", "save"}
+    assert worker["utility-second"]["inputs"] == {"images": ["utility-first", 0]}
+    assert prompt == before
+    assert registry == registered
+    assert OutputNode.OUTPUT_NODE is True
+    assert SchemaOutputNode.OUTPUT_NODE is False
+    assert SchemaOutputNode.GET_NODE_INFO_V1()["output_node"] is True
+
+
+@pytest.mark.parametrize("source", ["start", "process"])
+@pytest.mark.parametrize("direct", [False, True])
+def test_internal_bypass_into_post_end_computation_is_rejected(registry, source, direct):
+    prompt = graph()
+    prompt["post-process"] = node("Composite", images=["end", 0])
+    if not direct:
+        prompt["bypass"] = node("FrameOperation", image=[source, 0])
+        source = "bypass"
+    prompt["post-process"]["inputs"]["reference"] = [source, 0]
+    prompt["save"]["inputs"]["images"] = ["post-process", 0]
+    before = copy.deepcopy(prompt)
+    with pytest.raises(ValueError, match="cross-boundary"):
+        prepare_image_list_execution(prompt)
+    assert prompt == before
+
+
+def test_sequential_namespaced_regions_and_intermediate_output_are_preserved(registry):
+    prompt = graph()
+    prompt.update({
+        "next-start": node(START_CLASS, images=["end", 0]),
+        "next-process": node("SchemaUtility", images=["next-start", IMAGE_OUT], model=["loader", 0]),
+        "next-end": node(END_CLASS, images=["next-process", 0], context=["next-start", CONTEXT_OUT]),
+        "next-save": node("SaveImage", images=["next-end", 0]),
+        "first-debug": node("PreviewImage", images=["process", 0]),
+        "second-debug": node("PreviewImage", images=["next-process", 0]),
+    })
+    renamed = {key: f"group:7/{key}" for key in prompt}
+    for spec in prompt.values():
+        for name, value in spec["inputs"].items():
+            if isinstance(value, list):
+                spec["inputs"][name] = [renamed[value[0]], value[1]]
+    prompt = {renamed[key]: spec for key, spec in prompt.items()}
+    prepared, targets = prepare_image_list_execution(prompt)
+    assert set(prepared) == set(prompt) - {renamed["first-debug"], renamed["second-debug"]}
+    assert targets == [renamed[name] for name in ("end", "save", "next-end", "next-save")]
+    regions = validate_image_list_regions(prepared, explicit_targets=True)
+    assert {(r["start_id"], r["end_id"]) for r in regions} == {
+        (renamed["start"], renamed["end"]), (renamed["next-start"], renamed["next-end"]),
+    }
+    second = next(r for r in regions if r["start_id"] == renamed["next-start"])
+    worker = build_worker_prompt(prepared, second, "token", "context", 0, "http://master.invalid")
+    assert set(worker) == {renamed[name] for name in ("loader", "next-start", "next-process", "next-end")}
+
+
+def test_shared_loader_and_its_independent_output_root_are_preserved(registry):
+    registry["ModelLoader"] = type("OutputLoader", (ModelLoader, OutputNode), {})
+    prompt = graph()
+    prompt["source"]["inputs"]["model"] = ["loader", 0]
+    prompt["independent-save"] = node("SaveImage", model=["loader", 0])
+    prepared, targets = prepare_image_list_execution(prompt)
+    assert prepared == prompt
+    assert targets == ["loader", "end", "save", "independent-save"]
+    region = validate_image_list_regions(prepared, explicit_targets=True)[0]
+    assert "loader" in region["node_ids"]
+    assert "loader" not in region["branch_ids"]
+
+
+@pytest.mark.parametrize("unsafe_class", [None, type("ListLoader", (), {
+    "RETURN_TYPES": ("MODEL",), "OUTPUT_IS_LIST": (True,),
+})])
+def test_unknown_or_list_producing_shared_loader_still_fails_closed(registry, unsafe_class):
+    if unsafe_class is None:
+        registry.pop("ModelLoader")
+    else:
+        registry["ModelLoader"] = unsafe_class
+    prompt = graph()
+    prompt["source"]["inputs"]["model"] = ["loader", 0]
+    prompt["debug"] = node("PreviewImage", images=["process", 0])
+    with pytest.raises(ValueError, match="bypasses Start.*loader"):
+        prepare_image_list_execution(prompt)
+
+
+def test_upstream_per_item_bypass_is_not_hidden_by_preview_pruning(registry):
+    prompt = graph()
+    prompt["process"]["inputs"]["reference"] = ["source", 0]
+    prompt["debug"] = node("PreviewImage", images=["process", 0])
+    with pytest.raises(ValueError, match="bypasses Start.*source"):
+        prepare_image_list_execution(prompt)
+
+
+@pytest.mark.parametrize("nested", [False, True])
+def test_overlapping_or_nested_regions_cannot_be_discarded_as_side_branches(registry, nested):
+    prompt = graph()
+    prompt["other-start"] = node(START_CLASS, images=["process" if nested else "source", 0])
+    prompt["other-process"] = node("FrameOperation", image=["other-start", IMAGE_OUT], reference=["process", 0])
+    prompt["other-end"] = node(END_CLASS, images=["other-process", 0], context=["other-start", CONTEXT_OUT])
+    with pytest.raises(ValueError, match="Nested/ambiguous|cross-boundary|ambiguous processing"):
+        prepare_image_list_execution(prompt)
+
+
+@pytest.mark.parametrize("invalid", ["missing-end", "wrong-context", "missing-images", "unrelated-images"])
+def test_pairing_and_result_validation_cannot_be_pruned_away(registry, invalid):
+    prompt = graph()
+    if invalid == "missing-end":
+        del prompt["end"]
+    elif invalid == "wrong-context":
+        prompt["end"]["inputs"]["context"] = ["source", 0]
+    elif invalid == "missing-images":
+        del prompt["end"]["inputs"]["images"]
+    else:
+        prompt["end"]["inputs"]["images"] = ["source", 0]
+    before = copy.deepcopy(prompt)
+    with pytest.raises(ValueError):
+        prepare_image_list_execution(prompt)
+    assert prompt == before
+
+
+def test_execution_copy_is_deep_and_repeated_preparation_is_stable(registry):
+    prompt = graph()
+    prompt["process"]["_meta"] = {"title": "Original", "tags": ["saved"]}
+    before = copy.deepcopy(prompt)
+    prepared, targets = prepare_image_list_execution(prompt)
+    prepared_again, targets_again = prepare_image_list_execution(prepared)
+    assert prepared_again == prepared
+    assert targets_again == targets
+    prepared["start"]["inputs"]["images"][0] = "changed"
+    prepared["process"]["_meta"]["tags"].append("execution-only")
+    assert prompt == before
+    assert prepared_again == before
+
+
+def test_v3_false_output_flag_overrides_legacy_true(registry):
+    registry["NotOutput"] = type("NotOutput", (), {
+        "OUTPUT_NODE": True,
+        "GET_NODE_INFO_V1": classmethod(lambda cls: {"output_node": False}),
+    })
+    prompt = graph()
+    prompt["not-output"] = node("NotOutput", images=["end", 0])
+    prepared, targets = prepare_image_list_execution(prompt)
+    assert prepared == prompt
+    assert targets == ["end", "save"]

tests/test_image_list_integration_gaps.py

@@ -0,0 +1,189 @@
+"""Regression coverage at the scheduler and real HTTP transport boundaries."""
+import asyncio
+import importlib.util
+import json
+from pathlib import Path
+import sys
+import types
+import unittest
+from unittest.mock import patch
+
+import torch
+from aiohttp import web
+from aiohttp.test_utils import TestClient, TestServer
+
+from image_list import runtime
+from test_image_list import graph_two_sequential_regions, start_node
+
+
+def load_routes():
+    root = Path(__file__).resolve().parents[1]
+    name = '_image_list_transport_tests'
+    routes = web.RouteTableDef()
+    modules = {
+        name: types.ModuleType(name),
+        name + '.api': types.ModuleType(name + '.api'),
+        name + '.image_list': types.ModuleType(name + '.image_list'),
+        name + '.image_list.runtime': runtime,
+        'server': types.SimpleNamespace(PromptServer=types.SimpleNamespace(
+            instance=types.SimpleNamespace(routes=routes))),
+    }
+    spec = importlib.util.spec_from_file_location(name + '.api.image_list_routes', root / 'api/image_list_routes.py')
+    module = importlib.util.module_from_spec(spec)
+    with patch.dict(sys.modules, modules):
+        spec.loader.exec_module(module)
+    return module, routes
+
+
+class CapabilityTests(unittest.IsolatedAsyncioTestCase):
+    async def test_capability_advertises_target_control_without_tensor_protocol_bump(self):
+        module, _routes = load_routes()
+        response = await module.image_list_capabilities_endpoint(None)
+        payload = json.loads(response.text)
+        self.assertIs(payload['execution_targets'], True)
+        self.assertEqual(payload['protocol_version'], 2)
+
+
+class TransportTests(unittest.IsolatedAsyncioTestCase):
+    async def asyncSetUp(self):
+        self.module, routes = load_routes()
+        app = web.Application()
+        app.add_routes(routes)
+        self.client = TestClient(TestServer(app))
+        await self.client.start_server()
+
+    async def asyncTearDown(self):
+        await self.client.close()
+        runtime._WORKER_ITEMS.clear()
+
+    async def test_fragmented_chunked_item_upload_is_consumable(self):
+        token = runtime.create_worker_task_token()
+        item = {'image': torch.full((1, 2, 3, 3), 0.25), 'index': 2}
+        data = json.dumps({'token': token, 'context_id': 'context', 'item': runtime.encode_item(item)}).encode()
+        async def chunks():
+            yield data[:8]
+            await asyncio.sleep(0.02)
+            yield data[8:]
+        response = await self.client.post('/distributed/image-list/item', data=chunks())
+        self.assertEqual(response.status, 200, await response.text())
+        result = runtime.consume_worker_item(token)
+        self.assertTrue(torch.equal(result['image'], item['image']))
+        self.assertEqual(result['index'], 2)
+
+    async def test_complete_item_upload_and_replay_rejection(self):
+        token = runtime.create_worker_task_token()
+        data = {'token': token, 'context_id': 'context', 'item': runtime.encode_item(
+            {'image': torch.zeros((1, 2, 3, 3)), 'index': 0})}
+        response = await self.client.post('/distributed/image-list/item', json=data)
+        self.assertEqual(response.status, 200, await response.text())
+        response = await self.client.post('/distributed/image-list/item', json=data)
+        self.assertEqual(response.status, 400)
+        runtime.consume_worker_item(token)
+        with self.assertRaisesRegex(RuntimeError, 'already consumed'):
+            runtime.consume_worker_item(token)
+
+    async def test_chunked_body_limit_is_aggregate(self):
+        async def chunks():
+            yield b'{"context_id":"'
+            await asyncio.sleep(0.01)
+            yield b'x' * 60
+            yield b'"}'
+        with patch.object(self.module, 'MAX_REQUEST_BYTES', 64):
+            response = await self.client.post('/distributed/image-list/item', data=chunks())
+            self.assertEqual(response.status, 400)
+            self.assertIn('exceeds', await response.text())
+
+    async def test_invalid_result_shapes_are_rejected(self):
+        for kind, shape in [('image', (1, 2, 3)), ('mask', (1, 2, 3, 1))]:
+            result = runtime.encode_result({'image': torch.zeros((1, 2, 3, 3)),
+                                             'mask': torch.zeros((1, 2, 3))})
+            result[kind]['shape'] = list(shape)
+            with self.assertRaises(ValueError):
+                runtime.decode_result(result)
+
+
+class SchedulingBoundaryTests(unittest.IsolatedAsyncioTestCase):
+    async def test_user_interrupt_during_result_wait_cleans_up_promptly(self):
+        entered = asyncio.Event()
+        cancelled = asyncio.Event()
+        wait_cancelled = asyncio.Event()
+        dispatches, cleanups = [], []
+        async def dispatch(worker, index):
+            dispatches.append(index)
+            return {'index': index}
+        async def wait_next(_inflight):
+            entered.set()
+            try:
+                await asyncio.Event().wait()
+            finally:
+                wait_cancelled.set()
+        async def cleanup(worker, handle):
+            cleanups.append((worker, handle['index']))
+        task = asyncio.create_task(runtime.dynamic_schedule(
+            3, ['one'], dispatch, wait_next, cleanup, cancelled.is_set))
+        await entered.wait()
+        cancelled.set()
+        with self.assertRaises(runtime.ImageListCancelledError):
+            await asyncio.wait_for(task, timeout=1)
+        self.assertTrue(wait_cancelled.is_set())
+        self.assertEqual(dispatches, [0])
+        self.assertEqual(cleanups, [('one', 0)])
+
+    async def test_already_interrupted_does_not_submit_any_work(self):
+        async def unused(*args):
+            self.fail('pre-cancelled job must not dispatch or wait')
+        with self.assertRaises(runtime.ImageListCancelledError):
+            await runtime.dynamic_schedule(2, ['one'], unused, unused, unused, lambda: True)
+
+    async def test_task_cancel_during_submit_waits_for_owned_handle(self):
+        entered, release = asyncio.Event(), asyncio.Event()
+        cleanups = []
+        async def dispatch(worker, index):
+            entered.set()
+            await release.wait()
+            return {'index': index, 'prompt_id': 'owned'}
+        async def unused(*args):
+            self.fail('cancelled dispatch must not enter result wait')
+        async def cleanup(worker, handle):
+            cleanups.append(handle['prompt_id'])
+        task = asyncio.create_task(runtime.dynamic_schedule(2, ['one'], dispatch, unused, cleanup))
+        await entered.wait()
+        task.cancel()
+        release.set()
+        with self.assertRaises(asyncio.CancelledError):
+            await task
+        self.assertEqual(cleanups, ['owned'])
+
+    def test_output_nodes_inside_region_are_rejected(self):
+        prompt = graph_two_sequential_regions()
+        output_class = type('OutputOperation', (), {'OUTPUT_NODE': True})
+        registry = types.SimpleNamespace(NODE_CLASS_MAPPINGS={'FirstFrameOperation': output_class})
+        with patch.dict(sys.modules, {'nodes': registry}):
+            with self.assertRaisesRegex(ValueError, 'must be placed after'):
+                runtime.validate_image_list_regions(prompt)
+
+    def test_shared_scalars_are_validated_before_dispatch(self):
+        for value in (float('nan'), object(), torch.zeros(2, dtype=torch.int32)):
+            with self.assertRaises(TypeError):
+                runtime.normalize_image_list_inputs(torch.zeros((1, 2, 2, 3)), values={'value_1': [value]})
+
+    def test_three_sequential_regions_can_share_worker_loader(self):
+        prompt = graph_two_sequential_regions()
+        prompt.pop('8')
+        prompt.update({
+            '8': start_node(['7', 0]),
+            '9': {'class_type': 'ThirdOperation', 'inputs': {'image': ['8', 0]}},
+            '10': {'class_type': runtime.END_CLASS,
+                   'inputs': {'images': ['9', 0], 'context': ['8', runtime.CONTEXT_OUT]}},
+            '11': {'class_type': 'GenericModelLoader', 'inputs': {}},
+        })
+        for node_id in ['3', '6', '9']:
+            prompt[node_id]['inputs']['model'] = ['11', 0]
+        # Shared dependencies are verified against the registered output schema.
+        loader = type('GenericModelLoader', (), {'RETURN_TYPES': ('MODEL',)})
+        registry = types.SimpleNamespace(NODE_CLASS_MAPPINGS={'GenericModelLoader': loader})
+        with patch.dict(sys.modules, {'nodes': registry}):
+            regions = runtime.validate_image_list_regions(prompt)
+        self.assertEqual(len(regions), 3)
+        for region in regions:
+            self.assertIn('11', region['node_ids'])

tests/test_image_list_lifecycle.py

@@ -0,0 +1,254 @@
+"""Tests for periodic image-list server lifecycle cleanup."""
+import asyncio
+import time
+import unittest
+from types import SimpleNamespace
+from typing import Any, cast
+from unittest.mock import patch
+
+from aiohttp import web
+
+from image_list import lifecycle, runtime
+
+
+class ExecutionTargetProofTests(unittest.TestCase):
+    def setUp(self):
+        from test_image_list_master_runtime import graph
+        self.prompt = graph()
+        self.registry = SimpleNamespace(NODE_CLASS_MAPPINGS={
+            'Operation': type('Flagged', (), {'OUTPUT_NODE': True})})
+        self.modules = patch.dict('sys.modules', {'nodes': self.registry})
+        self.modules.start()
+        self.addCleanup(self.modules.stop)
+
+    def server(self, roots, *, prompt=None, running=True):
+        item = (0, 'owned', self.prompt if prompt is None else prompt, {}, roots, {})
+        return SimpleNamespace(prompt_queue=SimpleNamespace(
+            get_current_queue_volatile=lambda: ([item], []) if running else ([], [item])))
+
+    def test_outer_roots_in_running_queue_are_authoritative(self):
+        self.assertTrue(lifecycle.has_safe_execution_targets(self.server(['end']), 'owned', self.prompt))
+        for roots in (['op', 'end'], [], ['end', 'end'], None, [True]):
+            with self.subTest(roots=roots):
+                self.assertFalse(lifecycle.has_safe_execution_targets(self.server(roots), 'owned', self.prompt))
+        self.assertFalse(lifecycle.has_safe_execution_targets(self.server(['end']), 'other', self.prompt))
+        self.assertFalse(lifecycle.has_safe_execution_targets(self.server(['end'], running=False), 'owned', self.prompt))
+        self.assertFalse(lifecycle.has_safe_execution_targets(self.server(['end'], prompt={}), 'owned', self.prompt))
+
+    def test_prompt_marker_is_not_an_execution_guarantee(self):
+        self.prompt['start']['inputs']['image_list_explicit_targets'] = True
+        self.assertFalse(lifecycle.has_safe_execution_targets(self.server(['op', 'end']), 'owned', self.prompt))
+        with self.assertRaisesRegex(ValueError, 'must be placed after'):
+            runtime.create_master_job('start', 'end', self.prompt, [], [], '', master_participates=True)
+        self.assertEqual(runtime._MASTER_JOBS, {})
+
+    def test_start_node_uses_running_queue_proof(self):
+        import importlib.util
+        import sys
+        from pathlib import Path
+        from types import ModuleType
+        import torch
+
+        root = Path(__file__).resolve().parents[1]
+        package = '_image_list_start_targets'
+        modules = {}
+        for suffix in ('', '.nodes', '.image_list', '.utils'):
+            module = ModuleType(package + suffix)
+            module.__path__ = [str(root / suffix.lstrip('.'))]
+            modules[package + suffix] = module
+        modules[package + '.image_list.runtime'] = runtime
+        modules[package + '.image_list.lifecycle'] = lifecycle
+        modules[package + '.utils.async_helpers'] = SimpleNamespace(run_async_in_server_loop=None)
+        io = SimpleNamespace(ComfyNode=type('ComfyNode', (), {}), NodeOutput=lambda *args: args,
+                             **{name: object() for name in ('Image', 'Mask', 'String', 'Int', 'Float', 'Boolean')})
+        modules['comfy_api.v0_0_2'] = SimpleNamespace(io=io)
+        modules['comfy_execution.utils'] = SimpleNamespace(
+            get_executing_context=lambda: SimpleNamespace(prompt_id='owned'))
+        server = SimpleNamespace(PromptServer=SimpleNamespace(instance=self.server(['end'])))
+        modules['server'] = server
+        spec = importlib.util.spec_from_file_location(package + '.nodes.image_list', root / 'nodes/image_list.py')
+        node_module = importlib.util.module_from_spec(spec)
+        with patch.dict(sys.modules, modules):
+            spec.loader.exec_module(node_module)
+            start = node_module.DistributedImageListStart
+            start.hidden = SimpleNamespace(prompt=self.prompt, unique_id='start')
+            result = start.execute(images=torch.zeros((1, 2, 2, 3)), image_list_master_participates=True)
+            context = result[3][0]
+            self.assertIsNotNone(runtime.get_master_job(context))
+            runtime.remove_master_job(context)
+            server.PromptServer.instance = self.server(['op', 'end'])
+            with self.assertRaisesRegex(ValueError, 'must be placed after'):
+                start.execute(images=torch.zeros((1, 2, 2, 3)), image_list_master_participates=True)
+        self.assertEqual(runtime._MASTER_JOBS, {})
+
+    def test_owned_target_controlled_start_can_create_job(self):
+        context = runtime.create_master_job('start', 'end', self.prompt, [], [], '',
+                                            master_participates=True, explicit_targets=True)
+        self.addCleanup(runtime.remove_master_job, context)
+        self.assertIsNotNone(runtime.get_master_job(context))
+
+
+class ImageListLifecycleTests(unittest.IsolatedAsyncioTestCase):
+    def setUp(self):
+        with runtime._LOCK:
+            runtime._MASTER_JOBS.clear()
+            runtime._WORKER_ITEMS.clear()
+
+    def tearDown(self):
+        with runtime._LOCK:
+            runtime._MASTER_JOBS.clear()
+            runtime._WORKER_ITEMS.clear()
+
+    @staticmethod
+    def _state(prompt_id=None, *, active=False, age=10_000.0):
+        return SimpleNamespace(
+            prompt_id=prompt_id,
+            active=active,
+            created_at=time.monotonic() - age,
+        )
+
+    @staticmethod
+    def _server(queue):
+        return SimpleNamespace(app=web.Application(), prompt_queue=queue)
+
+    async def _start(self, server, *, interval=0.01, grace=0.01):
+        lifecycle.install_lifecycle(
+            server,
+            runtime,
+            interval_seconds=interval,
+            orphan_grace_seconds=grace,
+        )
+        runner = web.AppRunner(server.app)
+        await runner.setup()
+        return runner
+
+    async def test_unattended_sweep_removes_orphan_and_expired_staging(self):
+        context_id = "orphan"
+        with runtime._LOCK:
+            cast(Any, runtime._MASTER_JOBS)[context_id] = self._state()
+            runtime._WORKER_ITEMS["expired-token"] = (time.monotonic() - 1, {"image": "stale"})
+
+        queue = SimpleNamespace(
+            get_current_queue_volatile=lambda: ([], []),
+        )
+        server = self._server(queue)
+        swept = asyncio.Event()
+        sweep_worker_items = runtime._sweep_worker_items
+
+        def track_worker_sweep():
+            swept.set()
+            sweep_worker_items()
+
+        with patch.object(runtime, "_sweep_worker_items", side_effect=track_worker_sweep):
+            runner = await self._start(server)
+            try:
+                await asyncio.wait_for(swept.wait(), timeout=0.5)
+                self.assertNotIn(context_id, runtime._MASTER_JOBS)
+                self.assertNotIn("expired-token", runtime._WORKER_ITEMS)
+                with self.assertRaisesRegex(ValueError, "Unknown or expired"):
+                    runtime.set_master_result(context_id, 0, "late-token", {})
+            finally:
+                await runner.cleanup()
+
+    async def test_old_active_and_queued_contexts_survive_but_orphans_expire(self):
+        states = {
+            "active": self._state("active-prompt", active=True),
+            "queued": self._state("queued-prompt"),
+            "orphan": self._state("missing-prompt"),
+            "unbound": self._state(None),
+        }
+        with runtime._LOCK:
+            cast(Any, runtime._MASTER_JOBS).update(states)
+        queue = SimpleNamespace(
+            get_current_queue_volatile=lambda: (
+                [[1, "active-prompt", {}]],
+                [(2, "queued-prompt", {}, {}, None, None)],
+            ),
+        )
+        runner = await self._start(self._server(queue))
+        try:
+            await asyncio.sleep(0.04)
+            self.assertIs(runtime._MASTER_JOBS.get("active"), states["active"])
+            self.assertIs(runtime._MASTER_JOBS.get("queued"), states["queued"])
+            self.assertNotIn("orphan", runtime._MASTER_JOBS)
+            self.assertNotIn("unbound", runtime._MASTER_JOBS)
+        finally:
+            await runner.cleanup()
+
+    async def test_falls_back_to_nonvolatile_queue_reader(self):
+        state = self._state("queued-prompt")
+        with runtime._LOCK:
+            cast(Any, runtime._MASTER_JOBS)["queued"] = state
+        fallback_reads = 0
+
+        def volatile_failure():
+            raise RuntimeError("volatile view unavailable")
+
+        def current_queue():
+            nonlocal fallback_reads
+            fallback_reads += 1
+            return ([], [[1, "queued-prompt", {}]])
+
+        queue = SimpleNamespace(
+            get_current_queue_volatile=volatile_failure,
+            get_current_queue=current_queue,
+        )
+        runner = await self._start(self._server(queue))
+        try:
+            await asyncio.sleep(0.04)
+            self.assertGreaterEqual(fallback_reads, 1)
+            self.assertIs(runtime._MASTER_JOBS.get("queued"), state)
+        finally:
+            await runner.cleanup()
+
+    async def test_queue_read_error_protects_bound_contexts(self):
+        with runtime._LOCK:
+            cast(Any, runtime._MASTER_JOBS)["bound"] = self._state("bound-prompt")
+            cast(Any, runtime._MASTER_JOBS)["unbound"] = self._state(None)
+
+        def fail_read():
+            raise RuntimeError("queue unavailable")
+
+        queue = SimpleNamespace(
+            get_current_queue_volatile=fail_read,
+            get_current_queue=fail_read,
+        )
+        runner = await self._start(self._server(queue))
+        try:
+            await asyncio.sleep(0.04)
+            self.assertIn("bound", runtime._MASTER_JOBS)
+            self.assertNotIn("unbound", runtime._MASTER_JOBS)
+        finally:
+            await runner.cleanup()
+
+    async def test_install_and_removal_are_idempotent_and_task_stops_on_cleanup(self):
+        queue_reads = 0
+
+        def read_queue():
+            nonlocal queue_reads
+            queue_reads += 1
+            return ([], [])
+
+        server = self._server(SimpleNamespace(get_current_queue_volatile=read_queue))
+        lifecycle.install_lifecycle(server, runtime, interval_seconds=0.01, orphan_grace_seconds=0.01)
+        lifecycle.install_lifecycle(server, runtime, interval_seconds=0.01, orphan_grace_seconds=0.01)
+        self.assertEqual(len(server.app.cleanup_ctx), 1)
+
+        with runtime._LOCK:
+            cast(Any, runtime._MASTER_JOBS)["removed"] = self._state()
+        runtime.remove_master_job("removed")
+        runtime.remove_master_job("removed")
+        self.assertNotIn("removed", runtime._MASTER_JOBS)
+
+        runner = web.AppRunner(server.app)
+        await runner.setup()
+        await asyncio.sleep(0.035)
+        self.assertGreaterEqual(queue_reads, 1)
+        await runner.cleanup()
+        reads_after_cleanup = queue_reads
+        await asyncio.sleep(0.035)
+        self.assertEqual(queue_reads, reads_after_cleanup)
+
+
+if __name__ == "__main__":
+    unittest.main()

tests/test_image_list_local_cache_retention.py

@@ -0,0 +1,20 @@
+"""Run the real framework's CPU-only private-cache retention regression."""
+import os
+from pathlib import Path
+import subprocess
+import sys
+
+import pytest
+
+
+def test_native_local_cache_retention():
+    core = os.environ.get('COMFYUI_SOURCE_ROOT')
+    if not core:
+        pytest.skip('Set COMFYUI_SOURCE_ROOT for CPU-only real-framework validation')
+    helper = Path(__file__).parent / 'helpers' / 'local_cache_retention_check.py'
+    env = dict(os.environ, CUDA_VISIBLE_DEVICES='', HIP_VISIBLE_DEVICES='')
+    env.pop('PYTHONPATH', None)
+    result = subprocess.run([sys.executable, str(helper), core], env=env,
+                            text=True, capture_output=True, timeout=90)
+    assert result.returncode == 0, result.stdout + result.stderr
+    assert 'PASS: cross-run native cache reuse' in result.stdout

tests/test_image_list_master_runtime.py

@@ -0,0 +1,506 @@
+"""Master slot regressions, plus an opt-in real CPU ComfyUI subprocess probe."""
+import asyncio
+import importlib
+import json
+import os
+from pathlib import Path
+import subprocess
+import sys
+import types
+import unittest
+from unittest.mock import AsyncMock, patch
+
+import torch
+
+ROOT = Path(__file__).resolve().parents[1]
+
+
+def graph():
+    return {
+        'source': {'class_type': 'Source', 'inputs': {}},
+        'start': {'class_type': 'DistributedImageListStart', 'inputs': {'images': ['source', 0]}},
+        'op': {'class_type': 'Operation', 'inputs': {'image': ['start', 0]}},
+        'end': {'class_type': 'DistributedImageListEnd', 'inputs': {
+            'images': ['op', 0], 'context': ['start', 3]}},
+    }
+
+
+class Response:
+    status = 200
+
+    def __init__(self, data=None):
+        self.data = data
+
+    async def __aenter__(self):
+        return self
+
+    async def __aexit__(self, *args):
+        return None
+
+    async def json(self):
+        return self.data
+
+    def raise_for_status(self):
+        pass
+
+
+class MasterRuntimeTests(unittest.IsolatedAsyncioTestCase):
+    def setUp(self):
+        name = '_master_slot_tests'
+        self.modules = patch.dict(sys.modules)
+        self.modules.start()
+        self.addCleanup(self.modules.stop)
+        for suffix in ('', '.image_list', '.api', '.api.orchestration', '.utils'):
+            module = types.ModuleType(name + suffix)
+            module.__path__ = [str(ROOT / suffix.lstrip('.').replace('.', '/'))]
+            sys.modules[name + suffix] = module
+        self.rt = importlib.import_module(name + '.image_list.runtime')
+        self.workers = []
+        self.preflight = AsyncMock(side_effect=lambda workers, **kw: {w['id']: True for w in workers})
+        self.dispatched = []
+        self.local_calls = []
+        self.tokens = []
+        self.closed = 0
+        self.local_error = None
+        self.local_delay = 0.005
+        self.interrupted = False
+        self.local_started = asyncio.Event()
+        self.local_finished = asyncio.Event()
+        self.deadline = 0
+        self.slot_instances = []
+        self.config = {'workers': self.workers, 'settings': {}}
+        fixture = self
+
+        class FakeSlot:
+            def __init__(self, model_callback=None):
+                fixture.slot_instances.append(self)
+                self.tasks = []
+
+            def submit(self, prompt, prompt_id, end_id, workflow_meta=None):
+                async def execute():
+                    inputs = prompt[end_id]['inputs']
+                    index = inputs['worker_item_index']
+                    token = inputs['worker_result_token']
+                    fixture.local_calls.append(index)
+                    fixture.tokens.append(token)
+                    fixture.local_started.set()
+                    try:
+                        await asyncio.sleep(fixture.local_delay)
+                        if fixture.local_error:
+                            raise fixture.local_error
+                        item = fixture.rt.consume_worker_item(token)
+                        fixture.rt.set_master_result(fixture.context, index, token, {
+                            'image': item['image'], 'mask': item.get('mask'),
+                            'text': item['values'].get('value_1')})
+                    finally:
+                        fixture.local_finished.set()
+                task = asyncio.create_task(execute())
+                self.tasks.append(task)
+                return task
+
+            def request_stop(self):
+                pass
+
+            async def close(self):
+                # Model a synchronous node that must finish before its thread can join.
+                await asyncio.gather(*self.tasks, return_exceptions=True)
+                fixture.closed += 1
+
+        dependencies = {
+            '.image_list.local_executor': types.SimpleNamespace(LocalExecutionSlot=FakeSlot),
+            '.api.orchestration.dispatch': types.SimpleNamespace(
+                preflight_image_list_workers=self.preflight,
+                dispatch_worker_prompt=AsyncMock(side_effect=self.dispatch)),
+            '.api.orchestration.prompt_transform': types.SimpleNamespace(
+                _resolve_direct_workers_for_image_list=lambda config, ids: self.workers),
+            '.api.orchestration.media_sync': types.SimpleNamespace(
+                sync_worker_media=AsyncMock(), fetch_worker_path_separator=AsyncMock(return_value=None)),
+            '.utils.config': types.SimpleNamespace(load_config=lambda: self.config,
+                get_worker_timeout_seconds=lambda: 60,
+                get_image_list_job_timeout_seconds=lambda: self.deadline),
+            '.utils.network': types.SimpleNamespace(
+                build_master_callback_url=lambda *a, **kw: 'http://master',
+                build_worker_url=lambda worker, path: 'http://remote' + path,
+                get_client_session=AsyncMock(return_value=self)),
+            '.utils.logging': types.SimpleNamespace(log=lambda *args: None),
+        }
+        for suffix, module in dependencies.items():
+            sys.modules[name + suffix] = module
+        self.network = dependencies['.utils.network']
+        sys.modules['aiohttp'] = types.SimpleNamespace(ClientTimeout=lambda **kw: kw, ClientError=ConnectionError)
+        comfy = types.ModuleType('comfy')
+        comfy.model_management = types.SimpleNamespace(processing_interrupted=lambda: self.interrupted)
+        sys.modules['comfy'] = comfy
+        sys.modules['comfy.model_management'] = comfy.model_management
+
+    def create(self, participates=True, count=5):
+        self.items = [{'image': torch.full((1, i + 1, 2, 3), float(i)),
+                       'mask': torch.full((1, i + 1, 2), float(i)), 'values': {'value_1': str(i)},
+                       'index': i}
+                      for i in range(count)]
+        self.context = self.rt.create_master_job('start', 'end', graph(), self.items,
+            [w['id'] for w in self.workers], '', master_participates=participates)
+        self.state = self.rt.get_master_job(self.context)
+        return self.context
+
+    def post(self, url, *, data=None, **kwargs):
+        if url.endswith('/distributed/image-list/item'):
+            payload = json.loads(data)
+            self.rt.store_worker_item(payload['token'], payload['item'])
+        return Response({})
+
+    def delete(self, url, **kwargs):
+        self.rt.delete_worker_item(url.rsplit('/', 1)[-1])
+        return Response({})
+
+    def get(self, url, **kwargs):
+        pid = url.rsplit('/', 1)[-1]
+        return Response({pid: {'status': {'completed': True, 'status_str': 'success'}}})
+
+    async def dispatch(self, worker, prompt, meta, **kwargs):
+        inputs = prompt['end']['inputs']
+        index, token = inputs['worker_item_index'], inputs['worker_result_token']
+        self.dispatched.append(index)
+        # A remote callback is ready immediately; the local slot is slower.
+        item = self.rt.consume_worker_item(token)
+        self.rt.set_master_result(self.context, index, token, {
+            'image': item['image'], 'mask': item['mask'], 'text': item['values']['value_1']})
+        return f'remote-{index}'
+
+    async def test_local_only_preserves_order_and_uses_no_http(self):
+        result = await self.rt.run_image_list_job(self.create())
+        self.assertEqual(result[2], [str(i) for i in range(5)])
+        self.assertEqual(self.local_calls, list(range(5)))
+        self.assertEqual(len(self.slot_instances), 1)
+        self.assertEqual(len(set(self.tokens)), 5)
+        self.network.get_client_session.assert_not_awaited()
+        self.assertEqual(self.closed, 1)
+        self.assertFalse(self.rt._WORKER_ITEMS)
+        self.assertFalse(self.rt._MASTER_JOBS)
+        for i, (image, mask) in enumerate(zip(*result[:2])):
+            self.assertTrue(torch.equal(image, self.items[i]['image']))
+            self.assertTrue(torch.equal(mask, self.items[i]['mask']))
+
+    async def test_mixed_pool_claims_from_one_queue_and_relaxes_remote_idle(self):
+        self.workers.append({'id': 'master', 'type': 'local'})  # Not the in-process master slot.
+        self.local_delay = 0.04
+        result = await self.rt.run_image_list_job(self.create(count=9))
+        self.assertTrue(self.local_calls)
+        self.assertTrue(self.dispatched)
+        self.assertEqual(sorted(self.local_calls + self.dispatched), list(range(9)))
+        self.assertEqual(result[2], [str(i) for i in range(9)])
+        self.preflight.assert_awaited_once_with(self.workers, require_idle=False)
+
+    async def test_orchestrator_mode_never_instantiates_local_slot(self):
+        self.workers.append({'id': 'worker', 'type': 'remote'})
+        await self.rt.run_image_list_job(self.create(participates=False))
+        self.assertFalse(self.slot_instances)
+        self.preflight.assert_awaited_once_with(self.workers, require_idle=True)
+        with self.assertRaisesRegex(RuntimeError, 'worker|fallback'):
+            self.rt.create_master_job('start', 'end', graph(), self.items, [], '')
+
+    async def test_values_reconstructed_in_source_order_without_transporting_originals(self):
+        self.workers.append({'id': 'remote', 'type': 'remote'})
+        context = self.create(count=6)
+        tensor = torch.ones(4, 2, 2, 3, dtype=torch.float64)
+        for index, item in enumerate(self.items):
+            item['values'].update(value_2=tensor, value_3=True, value_4=index / 8)
+        self.local_delay = 0.1  # Remote results arrive ahead of master item zero.
+        with patch.object(self.rt, 'encode_result', side_effect=AssertionError('originals must not return')):
+            result = await self.rt.run_image_list_job(context, include_values=True)
+        self.assertFalse(self.rt._MASTER_JOBS)
+        self.assertEqual(len(result), 4)
+        for index, row in enumerate(result[3]):
+            self.assertEqual(row['value_1'], str(index))
+            self.assertIs(row['value_2'], tensor)
+            self.assertIs(row['value_3'], True)
+            self.assertEqual(row['value_4'], index / 8)
+        self.assertTrue(self.dispatched and self.local_calls)
+
+    async def test_overrides_merge_add_and_recheck_capability_before_every_item(self):
+        self.workers.append({'id': 'remote', 'type': 'remote'})
+        context = self.create(participates=False, count=3)
+        self.state.original_prompt['end']['inputs'].update(value_1=['start', 4], value_10=['start', 2])
+        original_set = self.rt.set_master_result
+        def override(context, index, token, result):
+            result['values'] = {'value_1': f'new-{index}', 'value_10': index}
+            return original_set(context, index, token, result)
+        with patch.object(self.rt, 'set_master_result', side_effect=override):
+            result = await self.rt.run_image_list_job(context, include_values=True)
+        self.assertEqual(result[3], [{'value_1': f'new-{index}', 'value_10': index} for index in range(3)])
+        self.assertEqual(self.preflight.await_count, 4)
+        self.assertTrue(all(call.kwargs['require_value_results'] is True
+                            for call in self.preflight.await_args_list))
+        self.assertEqual(self.items[0]['values'], {'value_1': '0'})
+
+    async def test_override_capability_downgrade_stops_before_upload(self):
+        self.workers.append({'id': 'remote', 'type': 'remote'})
+        context = self.create(participates=False)
+        self.state.original_prompt['end']['inputs']['value_1'] = ['start', 4]
+        self.preflight.side_effect = [{'remote': True}, RuntimeError('value_results unsupported')]
+        with patch.object(self, 'post', wraps=self.post) as upload:
+            with self.assertRaisesRegex(RuntimeError, 'value_results'):
+                await self.rt.run_image_list_job(context, include_values=True)
+            upload.assert_not_called()
+        self.assertFalse(self.dispatched or self.rt._MASTER_JOBS)
+
+    async def test_missing_expected_override_fails_without_partial_output(self):
+        self.workers.append({'id': 'remote', 'type': 'remote'})
+        context = self.create(participates=False)
+        self.state.original_prompt['end']['inputs']['value_1'] = ['start', 4]
+        with self.assertRaisesRegex(ValueError, 'missing'):
+            await self.rt.run_image_list_job(context, include_values=True)
+        self.assertFalse(self.rt._MASTER_JOBS)
+
+    async def test_values_failure_and_cancellation_return_no_partial_lists(self):
+        self.local_error = RuntimeError('branch failed')
+        with self.assertRaisesRegex(RuntimeError, 'branch failed'):
+            await self.rt.run_image_list_job(self.create(), include_values=True)
+        self.local_error = None
+        self.local_started.clear()
+        self.local_delay = 0.1
+        task = asyncio.create_task(self.rt.run_image_list_job(self.create(), include_values=True))
+        await self.local_started.wait()
+        task.cancel()
+        with self.assertRaises(self.rt.ImageListCancelledError):
+            await task
+        self.assertFalse(self.rt._MASTER_JOBS)
+
+    async def test_local_failure_surfaces_and_reclaims_staged_token(self):
+        self.local_error = RuntimeError('local branch failed')
+        with self.assertRaisesRegex(RuntimeError, 'local branch failed'):
+            await self.rt.run_image_list_job(self.create())
+        self.assertEqual(self.closed, 1)
+        self.assertFalse(self.rt._WORKER_ITEMS)
+        self.assertFalse(self.rt._MASTER_JOBS)
+
+    async def test_deadline_joins_local_execution_before_removing_context(self):
+        self.deadline, self.local_delay = 0.03, 0.1
+        with self.assertRaisesRegex(TimeoutError, 'total deadline'):
+            await self.rt.run_image_list_job(self.create())
+        self.assertTrue(self.local_finished.is_set())
+        self.assertEqual(self.closed, 1)
+        self.assertFalse(self.rt._WORKER_ITEMS)
+        self.assertFalse(self.rt._MASTER_JOBS)
+
+
+    async def test_cancelled_scheduler_joins_owned_execution_even_if_cancelled_again(self):
+        self.local_delay = 0.12
+        task = asyncio.create_task(self.rt.run_image_list_job(self.create()))
+        await self.local_started.wait()
+        task.cancel()
+        await asyncio.sleep(0.02)
+        task.cancel()
+        with self.assertRaises(self.rt.ImageListCancelledError):
+            await task
+        self.assertTrue(self.local_finished.is_set())
+        self.assertFalse(self.rt._WORKER_ITEMS)
+        self.assertFalse(self.rt._MASTER_JOBS)
+
+
+class RealFrameworkTests(unittest.TestCase):
+    @unittest.skipUnless(os.environ.get('COMFYUI_SOURCE_ROOT'), 'set COMFYUI_SOURCE_ROOT for the real CPU framework probe')
+    def test_real_framework_local_slot(self):
+        result = subprocess.run([sys.executable, str(Path(__file__).resolve()), '--real-probe',
+                                 os.environ['COMFYUI_SOURCE_ROOT']], text=True, capture_output=True, timeout=60)
+        self.assertEqual(result.returncode, 0, result.stdout + result.stderr)
+        self.assertIn('MASTER_SLOT_REAL_OK', result.stdout)
+
+
+def real_framework_probe(comfy_root):
+    import threading
+    import time
+    sys.path.insert(0, str(Path(comfy_root).resolve()))
+    import comfy.cli_args
+    comfy.cli_args.args.cpu = True
+    comfy.cli_args.args.disable_assets = True
+    import execution
+    import nodes
+    import comfy.memory_management as memory
+    import comfy.model_management as models
+    import comfy.utils as utils
+    from comfy_execution import progress
+    from comfy_execution.graph_utils import GraphBuilder
+    from comfy_api.v0_0_2 import io
+
+    name = '_real_master_slot'
+    for suffix in ('', '.image_list', '.utils', '.nodes'):
+        module = types.ModuleType(name + suffix)
+        module.__path__ = [str(ROOT / suffix.lstrip('.').replace('.', '/'))]
+        sys.modules[name + suffix] = module
+    # No HTTP/service/extension bootstrap is needed by the local slot. Fail if
+    # either boundary tries to use a server-loop bridge instead of direct memory.
+    def forbidden(*args, **kwargs):
+        raise AssertionError('local boundary used the server loop or /prompt')
+    sys.modules[name + '.utils.async_helpers'] = types.SimpleNamespace(run_async_in_server_loop=forbidden)
+    config = {'settings': {}}
+    sys.modules[name + '.utils.config'] = types.SimpleNamespace(load_config=lambda: config)
+    rt = importlib.import_module(name + '.image_list.runtime')
+    local = importlib.import_module(name + '.image_list.local_executor')
+    boundary = importlib.import_module(name + '.nodes.image_list')
+    nodes.NODE_CLASS_MAPPINGS.update(DistributedImageListStart=boundary.DistributedImageListStart,
+                                   DistributedImageListEnd=boundary.DistributedImageListEnd)
+    loads, calls = [], []
+    entered, stopped = threading.Event(), threading.Event()
+    mode = ['normal']
+    model = object()
+
+    class Loader:
+        @classmethod
+        def INPUT_TYPES(cls):
+            return {'required': {}}
+        RETURN_TYPES = ('MODEL',)
+        FUNCTION = 'run'
+        def run(self):
+            loads.append(True)
+            return (model,)
+
+    class Operation:
+        @classmethod
+        def INPUT_TYPES(cls):
+            return {'required': {'image': ('IMAGE',), 'model': ('MODEL',),
+                                 'mask': ('MASK',), 'text': ('STRING', {'forceInput': True})}}
+        RETURN_TYPES = ('IMAGE', 'MASK', 'STRING')
+        FUNCTION = 'run'
+        async def run(self, image, model, mask, text):
+            calls.append(text)
+            entered.set()
+            if mode[0] == 'error':
+                raise ValueError('probe branch failure')
+            if mode[0] in ('cancel', 'interrupt'):
+                try:
+                    while True:
+                        utils.ProgressBar(1).update(1)
+                        await asyncio.sleep(0.01)
+                finally:
+                    await asyncio.sleep(0.02)
+                    stopped.set()
+            if mode[0] == 'sync':
+                # No progress/interrupt checks: cleanup still must join, not
+                # release the parent just because an asyncio Future is cancelled.
+                time.sleep(0.08)
+                stopped.set()
+            await asyncio.sleep(0)
+            return image + 1, mask + 1, text + ':done'
+
+    class Expanded:
+        @classmethod
+        def INPUT_TYPES(cls):
+            return Operation.INPUT_TYPES()
+        RETURN_TYPES = Operation.RETURN_TYPES
+        FUNCTION = 'run'
+        def run(self, image, model, mask, text):
+            builder = GraphBuilder()
+            node = builder.node('Operation', image=image, model=model, mask=mask, text=text)
+            return {'result': (node.out(0), node.out(1), node.out(2)), 'expand': builder.finalize()}
+
+    nodes.NODE_CLASS_MAPPINGS.update(Loader=Loader, Operation=Operation, Expanded=Expanded)
+    prompt = graph()
+    prompt['loader'] = {'class_type': 'Loader', 'inputs': {}}
+    prompt['op']['inputs'].update(model=['loader', 0], mask=['start', rt.MASK_OUT],
+                                  text=['start', rt.VALUE_OUT_BASE])
+    prompt['end']['inputs'].update(masks=['op', 1], texts=['op', 2])
+    items = [{'image': torch.full((1, i + 1, 2, 3), float(i)),
+              'mask': torch.full((1, i + 1, 2), float(i)), 'values': {'value_1': str(i)}, 'index': i}
+             for i in range(3)]
+    region = rt.validate_image_list_regions(prompt)[0]
+    parent_progress = progress.ProgressRegistry('outer', execution.DynamicPrompt({}))
+    parent_progress.nodes['outer'] = {'state': progress.NodeState.Running, 'value': 3, 'max': 7}
+    progress.global_progress_registry = parent_progress
+    parent_hook = lambda *a, **kw: None
+    utils.PROGRESS_BAR_HOOK = parent_hook
+    ram_callback = lambda *a, **kw: None
+    memory.set_ram_cache_release_state(ram_callback, 123)
+    GraphBuilder.set_default_prefix('outer-prefix', 7, 9)
+
+    def restored():
+        assert progress.global_progress_registry is parent_progress
+        assert parent_progress.nodes == {'outer': {'state': progress.NodeState.Running, 'value': 3, 'max': 7}}
+        assert utils.PROGRESS_BAR_HOOK is parent_hook
+        assert memory.extra_ram_release_callback is ram_callback and memory.RAM_CACHE_HEADROOM == 123
+        assert (GraphBuilder._default_prefix_root, GraphBuilder._default_prefix_call_index,
+                GraphBuilder._default_prefix_graph_index) == ('outer-prefix', 7, 9)
+        assert not any(t.name.startswith('image-list-master') for t in threading.enumerate())
+
+    async def probe():
+        # Verify actual V3 input parsing and list mapping, not only direct execute.
+        for supplied, delegated, expected in [(None, False, True), (None, True, False),
+                                               (False, False, False), (True, True, True)]:
+            config['settings']['master_delegate_only'] = delegated
+            inputs = {'image_list_enabled_worker_ids': '["remote"]'}
+            if supplied is not None:
+                inputs['image_list_master_participates'] = supplied
+            cls = boundary.DistributedImageListStart
+            data, _, v3 = execution.get_input_data(inputs, cls, 'start', dynprompt=execution.DynamicPrompt(prompt))
+            data['images'] = [torch.zeros((3, 2, 2, 3))]
+            result = await execution._async_map_node_over_list('outer', 'start', cls, data, cls.FUNCTION, v3_data=v3)
+            result = await execution.resolve_map_node_over_list_results(result)
+            context = result[0].result[rt.CONTEXT_OUT][0]
+            assert rt.get_master_job(context).master_participates is expected
+            rt.remove_master_job(context)
+
+        for behavior in ('normal', 'expand', 'error', 'cancel', 'interrupt', 'sync', 'already-interrupted'):
+            # Independent lifecycle cases start cold; retention is tested separately.
+            local._RETAINED_OUTPUTS = None
+            mode[0] = behavior
+            prompt['op']['class_type'] = 'Expanded' if behavior == 'expand' else 'Operation'
+            loads.clear()
+            calls.clear()
+            entered.clear()
+            stopped.clear()
+            tracked = []
+            slot = local.LocalExecutionSlot(model_callback=tracked.append)
+            context = rt.create_master_job('start', 'end', prompt, items, [], '', master_participates=True)
+            try:
+                for index in range(3 if behavior in ('normal', 'expand') else 1):
+                    token = rt.create_worker_task_token()
+                    rt.register_assignment(context, index, 'local', token)
+                    rt._WORKER_ITEMS[token] = (time.monotonic() + 60, items[index])
+                    worker_prompt = rt.build_worker_prompt(prompt, region, token, context, index, '')
+                    for key in ('start', 'end'):
+                        worker_prompt[key]['inputs']['image_list_local_execution'] = True
+                    if behavior == 'already-interrupted':
+                        models.interrupt_current_processing(True)
+                    future = slot.submit(worker_prompt, f'local-{index}', 'end')
+                    if behavior in ('cancel', 'interrupt', 'sync'):
+                        assert await asyncio.to_thread(entered.wait, 5)
+                        if behavior == 'interrupt':
+                            models.interrupt_current_processing(True)
+                        else:
+                            slot.request_stop()
+                    try:
+                        await asyncio.wait_for(asyncio.shield(asyncio.wrap_future(future)), 5)
+                    except rt.ImageListCancelledError:
+                        assert behavior in ('cancel', 'interrupt', 'sync', 'already-interrupted'), behavior
+                    except RuntimeError as exc:
+                        assert behavior == 'error' and 'probe branch failure' in str(exc), str(exc)
+                    else:
+                        assert behavior in ('normal', 'expand'), behavior
+                        result = rt.get_master_job(context).results[index]
+                        assert torch.equal(result['image'], items[index]['image'] + 1)
+                        assert torch.equal(result['mask'], items[index]['mask'] + 1)
+                        assert result['text'] == f'{index}:done'
+                if behavior in ('normal', 'expand'):
+                    assert len(loads) == 1 and calls == ['0', '1', '2'], (loads, calls)
+                    assert any(model in outputs for outputs in tracked), tracked
+                if behavior in ('cancel', 'interrupt', 'sync'):
+                    assert stopped.is_set(), behavior
+                if behavior == 'already-interrupted':
+                    assert models.processing_interrupted() and not calls
+            finally:
+                await slot.close()
+                for token in rt.get_master_job(context).expected_tokens.values():
+                    rt.delete_worker_item(token)
+                rt.remove_master_job(context)
+                models.interrupt_current_processing(False)  # Only test setup/teardown.
+            restored()
+        assert not rt._MASTER_JOBS and not rt._WORKER_ITEMS
+    asyncio.run(probe())
+    print('MASTER_SLOT_REAL_OK V3 metadata; cache reuse; async/expanded graph; errors; cancellation; state restoration')
+
+
+if __name__ == '__main__':
+    if len(sys.argv) > 1 and sys.argv[1] == '--real-probe':
+        real_framework_probe(sys.argv[2])
+    else:
+        unittest.main()

tests/test_image_list_master_selection.py

@@ -0,0 +1,286 @@
+"""Image-list participation uses the existing master orchestrator setting."""
+import json
+import sys
+import types
+import unittest
+from unittest.mock import AsyncMock, Mock, patch
+
+from test_serverless_integration import PKG, load_modules
+
+
+class MasterSelectionTests(unittest.IsolatedAsyncioTestCase):
+    async def asyncSetUp(self):
+        self.q, modules, self.ps = load_modules()
+        self.modules_patch = patch.dict(sys.modules, modules)
+        self.modules_patch.start()
+        self.addCleanup(self.modules_patch.stop)
+        self.worker = {"id": "worker", "name": "Worker", "type": "local", "enabled": True}
+        self.config = {"workers": [self.worker], "settings": {}}
+        self.q.load_config = lambda: self.config
+        # Model the legacy helper's fallback: image-list must not inherit it.
+        self.q.select_active_workers = AsyncMock(side_effect=lambda workers, ws, delegate, **kw:
+                                                (workers, delegate if workers else False))
+        self.preflight = AsyncMock(return_value={"worker": True})
+        sys.modules[PKG + ".api.orchestration.dispatch"].preflight_image_list_workers = self.preflight
+        self.q.queue_prompt_payload = AsyncMock(return_value={"prompt_id": "master-prompt", "number": 1})
+        self.prompt = {
+            "1": {"class_type": "EmptyImage", "inputs": {}},
+            "2": {"class_type": "DistributedImageListStart", "inputs": {"images": ["1", 0]}},
+            "3": {"class_type": "DistributedImageListEnd", "inputs": {"images": ["2", 0], "context": ["2", 3]}},
+        }
+
+    async def queue(self, **kwargs):
+        return await self.q.orchestrate_distributed_execution(self.prompt, None, "client", **kwargs)
+
+    def prepared_inputs(self):
+        return self.q.queue_prompt_payload.call_args.args[0]["2"]["inputs"]
+
+    async def test_prepared_master_prunes_side_outputs_without_mutating_input(self):
+        import copy
+        self.prompt['op'] = {'class_type': 'Flagged', 'inputs': {'image': ['2', 0]}}
+        self.prompt['3']['inputs']['images'] = ['op', 0]
+        self.prompt['preview'] = {'class_type': 'Flagged', 'inputs': {'image': ['op', 0]}}
+        self.prompt['save'] = {'class_type': 'Flagged', 'inputs': {'image': ['3', 0]}}
+        original = copy.deepcopy(self.prompt)
+        registry = type('Registry', (), {'NODE_CLASS_MAPPINGS': {
+            'Flagged': type('Flagged', (), {'OUTPUT_NODE': True})}})
+        with patch.dict(sys.modules, {'nodes': registry}):
+            await self.queue()
+        queued = self.q.queue_prompt_payload.call_args
+        self.assertEqual(queued.kwargs['partial_execution_targets'], ['3', 'save'])
+        self.assertIn('op', queued.args[0])
+        self.assertNotIn('preview', queued.args[0])
+        self.assertEqual(self.prompt, original)
+
+    async def test_default_includes_master_and_direct_workers(self):
+        result = await self.queue()
+        self.assertIs(self.prepared_inputs()["image_list_master_participates"], True)
+        self.assertEqual(json.loads(self.prepared_inputs()["image_list_enabled_worker_ids"]), ["worker"])
+        self.assertEqual(result[2], 1)  # Public worker_count still counts remote participants only.
+        self.preflight.assert_awaited_once_with(
+            [{"id": "worker", "name": "Worker", "type": "local", "host": None, "port": 8188}],
+            require_idle=False)
+        self.q.dispatch_worker_prompt.assert_not_awaited()
+        self.assertEqual(self.ps.distributed_pending_jobs, {})
+
+    async def test_end_overrides_require_value_capability_before_master_queue(self):
+        self.prompt['3']['inputs']['value_10'] = ['2', 2]
+        await self.queue()
+        self.assertIs(self.preflight.call_args.kwargs['require_value_results'], True)
+        self.q.queue_prompt_payload.reset_mock()
+        self.preflight.side_effect = RuntimeError('value_results unsupported')
+        with self.assertRaisesRegex(RuntimeError, 'value_results'):
+            await self.queue()
+        self.q.queue_prompt_payload.assert_not_awaited()
+
+    async def test_orchestrator_setting_excludes_master_processing(self):
+        self.config["settings"]["master_delegate_only"] = True
+        await self.queue()
+        self.assertIs(self.prepared_inputs()["image_list_master_participates"], False)
+        self.assertIs(self.preflight.call_args.kwargs["require_idle"], False)
+        # Upstream source stays on master even in orchestrator-only mode.
+        self.assertIn("1", self.q.queue_prompt_payload.call_args.args[0])
+
+    async def test_explicit_override_wins_over_config_in_both_directions(self):
+        for configured, requested in [(True, False), (False, True)]:
+            with self.subTest(configured=configured, requested=requested):
+                self.config["settings"]["master_delegate_only"] = configured
+                await self.queue(delegate_master=requested)
+                self.assertIs(self.prepared_inputs()["image_list_master_participates"], not requested)
+
+    async def test_master_only_with_explicit_empty_worker_selection(self):
+        result = await self.queue(enabled_worker_ids=[])
+        self.assertIs(self.prepared_inputs()["image_list_master_participates"], True)
+        self.assertEqual(json.loads(self.prepared_inputs()["image_list_enabled_worker_ids"]), [])
+        self.assertEqual(result[2], 0)
+        self.preflight.assert_not_awaited()
+
+    async def test_master_only_when_no_workers_are_configured(self):
+        self.config["workers"] = []
+        await self.queue()
+        self.assertIs(self.prepared_inputs()["image_list_master_participates"], True)
+
+    async def test_unavailable_workers_do_not_prevent_enabled_master(self):
+        self.q.select_active_workers.return_value = ([], False)
+        self.q.select_active_workers.side_effect = None
+        self.preflight.return_value = {}
+        await self.queue()
+        self.assertEqual(json.loads(self.prepared_inputs()["image_list_enabled_worker_ids"]), [])
+        self.assertIs(self.prepared_inputs()["image_list_master_participates"], True)
+
+    def use_real_preflight(self, queues):
+        """Exercise capability/queue parsing with read-only fake HTTP responses."""
+        from test_dispatch_selection import dispatch, _FakeSession, _valid_capability
+
+        responses = {(worker_id, "capabilities"): (200, _valid_capability())
+                     for worker_id in queues}
+        responses.update({(worker_id, "queue"): (200, queue)
+                          for worker_id, queue in queues.items()})
+        session = _FakeSession(responses)
+        for target, value in (
+            ("get_client_session", AsyncMock(return_value=session)),
+            ("build_worker_url", lambda worker, path: f"http://{worker['id']}{path}"),
+        ):
+            patched = patch.object(dispatch, target, value)
+            patched.start()
+            self.addCleanup(patched.stop)
+        self.preflight = AsyncMock(wraps=dispatch.preflight_image_list_workers)
+        sys.modules[PKG + ".api.orchestration.dispatch"].preflight_image_list_workers = self.preflight
+        return responses
+
+    async def run_prepared_selection(self):
+        """Run real runtime selection up to the scheduler, without any item I/O."""
+        rt = sys.modules[PKG + ".image_list.runtime"]
+        inputs = self.prepared_inputs()
+        image = object()
+        context = rt.create_master_job(
+            "2", "3", self.q.queue_prompt_payload.call_args.args[0],
+            [{"image": image, "index": 0}],
+            json.loads(inputs["image_list_enabled_worker_ids"]),
+            inputs["image_list_master_callback"],
+            master_participates=inputs["image_list_master_participates"],
+            explicit_targets=True,
+        )
+        self.addCleanup(rt.remove_master_job, context)
+        self.scheduler = AsyncMock(return_value=[{"image": image}])
+        self.local_slot = Mock(return_value=types.SimpleNamespace(close=AsyncMock()))
+        self.runtime_session = AsyncMock(return_value=object())
+        config = sys.modules[PKG + ".utils.config"]
+        network = sys.modules[PKG + ".utils.network"]
+        with patch.object(config, "load_config", lambda: self.config), \
+             patch.object(config, "get_worker_timeout_seconds", lambda: 60, create=True), \
+             patch.object(config, "get_image_list_job_timeout_seconds", lambda: 0, create=True), \
+             patch.object(network, "get_client_session", self.runtime_session, create=True), \
+             patch.object(network, "build_worker_url", Mock(), create=True), \
+             patch.object(rt, "dynamic_schedule", self.scheduler), \
+             patch.dict(sys.modules, {PKG + ".image_list.local_executor": types.SimpleNamespace(
+                 LocalExecutionSlot=self.local_slot)}):
+            await rt.run_image_list_job(context)
+        self.assertIsNone(rt.get_master_job(context))
+        return self.scheduler.call_args.args[1]
+
+    async def test_busy_worker_at_enqueue_is_available_to_scheduler_when_idle_at_start(self):
+        responses = self.use_real_preflight({
+            "worker": {"queue_running": [[1, "prompt-a"]], "queue_pending": []},
+        })
+        result = await self.queue()
+        self.assertEqual(json.loads(self.prepared_inputs()["image_list_enabled_worker_ids"]), ["worker"])
+        self.assertEqual(result[2], 1)  # Candidates, not a promise of execution-time availability.
+        self.q.dispatch_worker_prompt.assert_not_awaited()
+
+        responses[("worker", "queue")] = (200, {"queue_running": [], "queue_pending": []})
+        # A newly enabled worker is not implicitly added to this queued selection.
+        self.config["workers"].append({"id": "later", "enabled": True})
+        slots = await self.run_prepared_selection()
+        self.assertEqual([slot for slot in slots if isinstance(slot, str)], ["worker"])
+        self.assertEqual(len(slots), 2)
+        self.assertEqual(self.preflight.await_count, 2)
+        self.assertEqual(self.preflight.await_args_list[0].args, self.preflight.await_args_list[1].args)
+
+    async def test_worker_still_busy_at_start_is_excluded_from_scheduler(self):
+        self.use_real_preflight({
+            "worker": {"queue_running": [], "queue_pending": [[1, "prompt-a"]]},
+        })
+        await self.queue()
+        self.assertEqual(json.loads(self.prepared_inputs()["image_list_enabled_worker_ids"]), ["worker"])
+        slots = await self.run_prepared_selection()
+        self.assertEqual(len(slots), 1)
+        self.assertNotIn("worker", slots)
+        self.assertEqual(self.preflight.await_count, 2)
+        self.runtime_session.assert_not_awaited()
+
+    async def test_explicit_empty_selection_stays_master_only_at_start(self):
+        self.use_real_preflight({"worker": {"queue_running": [], "queue_pending": []}})
+        await self.queue(enabled_worker_ids=[])
+        slots = await self.run_prepared_selection()
+        self.assertEqual(len(slots), 1)
+        self.assertNotIn("worker", slots)
+        self.preflight.assert_not_awaited()
+        self.runtime_session.assert_not_awaited()
+
+    async def test_busy_incompatible_worker_fails_before_master_queue(self):
+        responses = self.use_real_preflight({
+            "worker": {"queue_running": [[1, "prompt-a"]], "queue_pending": []},
+        })
+        capability = responses[("worker", "capabilities")][1]
+        for field, invalid, message in (
+            ("protocol_version", 1, "incompatible"),
+            ("execution_targets", False, "execution_targets"),
+            ("value_results", False, "value_results"),
+        ):
+            with self.subTest(field=field), patch.dict(capability, {field: invalid}):
+                if field == "value_results":
+                    self.prompt["3"]["inputs"]["value_10"] = ["2", 2]
+                with self.assertRaisesRegex(RuntimeError, message):
+                    await self.queue()
+                self.q.queue_prompt_payload.assert_not_awaited()
+                self.q.dispatch_worker_prompt.assert_not_awaited()
+
+    async def test_orchestrator_with_all_workers_busy_queues_for_idle_start(self):
+        responses = self.use_real_preflight({
+            "worker": {"queue_running": [[1, "prompt-a"]], "queue_pending": []},
+        })
+        result = await self.queue(delegate_master=True)
+        self.q.queue_prompt_payload.assert_awaited_once()
+        self.assertEqual(json.loads(self.prepared_inputs()["image_list_enabled_worker_ids"]), ["worker"])
+        self.assertIs(self.prepared_inputs()["image_list_master_participates"], False)
+        self.assertEqual(result[2], 1)
+        self.q.dispatch_worker_prompt.assert_not_awaited()
+
+        responses[("worker", "queue")] = (200, {"queue_running": [], "queue_pending": []})
+        self.assertEqual(await self.run_prepared_selection(), ["worker"])
+        self.local_slot.assert_not_called()
+        self.assertEqual(self.preflight.await_count, 2)
+        self.assertIs(self.preflight.await_args_list[0].kwargs["require_idle"], False)
+        self.assertIs(self.preflight.await_args_list[1].kwargs["require_idle"], True)
+
+    async def test_orchestrator_rechecks_availability_without_master_fallback(self):
+        responses = self.use_real_preflight({"worker": {"queue_running": [], "queue_pending": []}})
+        await self.queue(delegate_master=True)
+        responses[("worker", "queue")] = (200, {"queue_running": [[1, "other"]], "queue_pending": []})
+        with self.assertRaisesRegex(RuntimeError, "No selected feature-compatible direct worker is idle"):
+            await self.run_prepared_selection()
+        self.assertEqual(self.preflight.await_count, 2)
+        self.scheduler.assert_not_awaited()
+        self.local_slot.assert_not_called()
+
+    async def test_orchestrator_retains_busy_candidates_when_another_worker_is_idle(self):
+        self.config["workers"].append({"id": "idle", "enabled": True})
+        responses = self.use_real_preflight({
+            "worker": {"queue_running": [[1, "prompt-a"]], "queue_pending": []},
+            "idle": {"queue_running": [], "queue_pending": []},
+        })
+        await self.queue(delegate_master=True)
+        self.assertEqual(json.loads(self.prepared_inputs()["image_list_enabled_worker_ids"]), ["worker", "idle"])
+        responses[("worker", "queue")] = (200, {"queue_running": [], "queue_pending": []})
+        responses[("idle", "queue")] = (200, {"queue_running": [[1, "other"]], "queue_pending": []})
+        self.assertEqual(await self.run_prepared_selection(), ["worker"])
+        self.local_slot.assert_not_called()
+
+    async def test_orchestrator_without_workers_fails_before_queue(self):
+        self.config["settings"]["master_delegate_only"] = True
+        with self.assertRaisesRegex(RuntimeError, "orchestrator|Orchestrator"):
+            await self.queue(enabled_worker_ids=[])
+        self.q.queue_prompt_payload.assert_not_awaited()
+
+    async def test_orchestrator_never_inherits_legacy_offline_master_fallback(self):
+        self.config["settings"]["master_delegate_only"] = True
+        self.q.select_active_workers.side_effect = None
+        self.q.select_active_workers.return_value = ([], False)
+        self.preflight.return_value = {}
+        with self.assertRaisesRegex(RuntimeError, "orchestrator|Orchestrator"):
+            await self.queue()
+        self.q.queue_prompt_payload.assert_not_awaited()
+
+    async def test_serverless_only_uses_master_without_provider_dispatch(self):
+        self.config["workers"] = [{"id": "pool", "type": "serverless", "enabled": True}]
+        result = await self.queue()
+        self.assertIs(self.prepared_inputs()["image_list_master_participates"], True)
+        self.assertEqual(result[2], 0)
+        self.q.dispatch_worker_prompt.assert_not_awaited()
+
+    async def test_incompatible_worker_is_not_silently_hidden_by_master(self):
+        self.preflight.side_effect = RuntimeError("upgrade worker")
+        with self.assertRaisesRegex(RuntimeError, "upgrade worker"):
+            await self.queue()
+        self.q.queue_prompt_payload.assert_not_awaited()

tests/test_image_list_native_cache_identity.py

@@ -0,0 +1,527 @@
+"""Stable content signatures with fresh, expiring delivery capabilities."""
+import asyncio
+import copy
+import importlib.util
+import json
+import os
+from pathlib import Path
+import subprocess
+import sys
+import types
+from unittest.mock import AsyncMock
+
+import pytest
+import torch
+
+from image_list import runtime as rt
+
+ROOT = Path(__file__).resolve().parents[1]
+
+
+def item():
+    return rt.normalize_image_list_inputs(
+        torch.arange(18, dtype=torch.float64).reshape(1, 2, 3, 3),
+        torch.zeros(1, 2, 3, dtype=torch.bfloat16),
+        {'value_1': 'original', 'value_2': True, 'value_3': 7, 'value_4': 0.25,
+         'value_10': torch.ones(2, 3, dtype=torch.float16)})[0]
+
+
+def graph():
+    return {
+        'source': {'class_type': 'Source', 'inputs': {}},
+        'start': {'class_type': rt.START_CLASS, 'inputs': {
+            'images': ['source', 0], 'image_list_execution_id': 'transient',
+            'image_list_workflow_meta': {'job': 'metadata'},
+            'image_list_enabled_worker_ids': '["worker"]',
+            'image_list_master_callback': 'http://old-master',
+            'values.value_1': 'inherited', 'worker_context_id': 'old-context'}},
+        'op': {'class_type': 'Operation', 'inputs': {'image': ['start', 0]}},
+        'end': {'class_type': rt.END_CLASS, 'inputs': {
+            'images': ['op', 0], 'context': ['start', 3], 'value_3': ['start', 2]}},
+    }
+
+
+def worker_prompt(token, context, data, *, local=True):
+    prompt = graph()
+    result = rt.build_worker_prompt(prompt, rt.validate_image_list_regions(prompt)[0],
+                                    token, context, data['index'], 'http://master',
+                                    item_identity=rt.item_content_identity(data))
+    if local:
+        result['start']['inputs']['image_list_local_execution'] = True
+        result['end']['inputs']['image_list_local_execution'] = True
+    return result
+
+
+@pytest.fixture(autouse=True)
+def clean_state():
+    yield
+    rt._WORKER_ITEMS.clear()
+    rt._MASTER_JOBS.clear()
+
+
+@pytest.fixture
+def boundaries(monkeypatch):
+    """Load only the boundary module; restore every replaced module on exit."""
+    name = '_native_cache_identity_tests'
+    for suffix in ('', '.nodes', '.utils', '.image_list', '.api'):
+        module = types.ModuleType(name + suffix)
+        module.__path__ = [str(ROOT / suffix.lstrip('.').replace('.', '/'))]
+        monkeypatch.setitem(sys.modules, name + suffix, module)
+    monkeypatch.setitem(sys.modules, name + '.image_list.runtime', rt)
+    io = types.SimpleNamespace(ComfyNode=type('ComfyNode', (), {}),
+                              NodeOutput=lambda *args: args)
+    for cls in ('Image', 'Mask', 'String', 'Int', 'Float', 'Boolean'):
+        setattr(io, cls, object())
+    monkeypatch.setitem(sys.modules, 'comfy_api', types.ModuleType('comfy_api'))
+    monkeypatch.setitem(sys.modules, 'comfy_api.v0_0_2', types.SimpleNamespace(io=io))
+    monkeypatch.setitem(sys.modules, name + '.utils.async_helpers',
+                        types.SimpleNamespace(run_async_in_server_loop=lambda coro, **kw: asyncio.run(coro)))
+    spec = importlib.util.spec_from_file_location(name + '.nodes.image_list', ROOT / 'nodes/image_list.py')
+    module = importlib.util.module_from_spec(spec)
+    monkeypatch.setitem(sys.modules, spec.name, module)
+    spec.loader.exec_module(module)
+    return module
+
+
+def start(boundaries, prompt):
+    boundaries.DistributedImageListStart.hidden = types.SimpleNamespace(prompt=prompt, unique_id='start')
+    return boundaries.DistributedImageListStart.execute(**prompt['start']['inputs'])
+
+
+def end(boundaries, prompt, cached):
+    boundaries.DistributedImageListEnd.hidden = types.SimpleNamespace(prompt=prompt, unique_id='end')
+    inputs = {k: v for k, v in prompt['end']['inputs'].items()
+              if k not in ('images', 'context', 'value_3')}
+    return boundaries.DistributedImageListEnd.execute(
+        images=cached[0], context=cached[3], value_3=cached[2], **inputs)
+
+
+def stage(data):
+    token = rt.create_worker_task_token()
+    rt.store_worker_item(token, rt.encode_item(data))
+    return token
+
+
+def master(data, token):
+    context = rt.create_master_job('start', 'end', graph(), [data], ['worker'], '')
+    rt.register_assignment(context, data['index'], 'worker', token)
+    return context
+
+
+def test_identity_is_exact_typed_order_independent_and_not_transport_bounded(monkeypatch):
+    original = item()
+    identity = rt.item_content_identity(original)
+    assert rt.is_valid_item_identity(identity)
+    reordered = copy.deepcopy(original)
+    reordered['values'] = dict(reversed(list(reordered['values'].items())))
+    assert rt.item_content_identity(reordered) == identity
+    strided = copy.deepcopy(original)
+    strided['image'] = original['image'].transpose(1, 2).contiguous().transpose(1, 2)
+    assert not strided['image'].is_contiguous()
+    assert rt.item_content_identity(strided) == identity
+    monkeypatch.setattr(rt, 'MAX_PAYLOAD_BYTES', 1)
+    with pytest.raises(ValueError, match='transport limit'):
+        rt.encode_item(original)
+    assert rt.item_content_identity(original) == identity
+    scalar_identities = set()
+    for value in (True, 1, 1.0, '1', 0.0, -0.0):
+        changed = copy.deepcopy(original)
+        changed['values']['value_2'] = value
+        scalar_identities.add(rt.item_content_identity(changed))
+    assert len(scalar_identities) == 6
+
+
+@pytest.mark.parametrize('change', ['image', 'mask', 'value', 'index', 'dtype', 'shape', 'slot', 'missing-mask'])
+def test_every_item_component_invalidates_identity(change):
+    original, changed = item(), item()
+    if change in ('image', 'mask'):
+        changed[change].reshape(-1)[0] = 1
+    elif change == 'value':
+        changed['values']['value_10'][0, 0] = 2
+    elif change == 'index':
+        changed['index'] = 1
+    elif change == 'dtype':
+        changed['image'] = changed['image'].float()
+    elif change == 'shape':
+        changed['values']['value_10'] = changed['values']['value_10'].reshape(3, 2)
+    elif change == 'slot':
+        changed['values']['value_9'] = changed['values'].pop('value_10')
+    else:
+        changed['mask'] = None
+    assert rt.item_content_identity(original) != rt.item_content_identity(changed)
+
+
+def test_float64_low_bits_and_bfloat16_bytes_are_not_quantized():
+    data, changed = item(), item()
+    changed['image'].reshape(-1)[0] = torch.nextafter(torch.tensor(0., dtype=torch.float64),
+                                                  torch.tensor(1., dtype=torch.float64))
+    assert rt.item_content_identity(data) != rt.item_content_identity(changed)
+    changed = item()
+    changed['mask'].view(torch.int16).reshape(-1)[0] = 1
+    assert rt.item_content_identity(data) != rt.item_content_identity(changed)
+
+
+def test_stable_start_replaced_wholesale_and_legacy_graph_preserved():
+    prompt = graph()
+    frozen = copy.deepcopy(prompt)
+    region = rt.validate_image_list_regions(prompt)[0]
+    legacy = rt.build_worker_prompt(prompt, region, 'a', 'ctx-a', 0, 'http://a')
+    assert legacy['start']['inputs']['worker_item_token'] == 'a'
+    assert legacy['start']['inputs']['image_list_execution_id'] == 'transient'
+    identity = rt.item_content_identity(item())
+    a = rt.build_worker_prompt(prompt, region, 'a', 'ctx-a', 0, 'http://a', item_identity=identity)
+    b = rt.build_worker_prompt(prompt, region, 'b', 'ctx-b', 0, 'http://b', item_identity=identity)
+    assert a['start']['inputs'] == b['start']['inputs'] == {'worker_item_identity': identity}
+    assert a['end']['inputs']['worker_context_id'] == 'ctx-a'
+    assert b['end']['inputs']['worker_result_token'] == 'b'
+    assert b['end']['inputs']['worker_item_identity'] == identity
+    assert b['end']['inputs']['value_3'] == ['start', 2]
+    assert prompt == frozen
+
+
+@pytest.mark.parametrize('bad', ['', 'sha256:' + 'z' * 64, 'a' * 64, False, ['bad']])
+def test_malformed_identity_rejected(bad):
+    assert not rt.is_valid_item_identity(bad)
+    prompt = graph()
+    with pytest.raises((ValueError, TypeError), match='identity'):
+        rt.build_worker_prompt(prompt, rt.validate_image_list_regions(prompt)[0],
+                               'token', 'context', 0, '', item_identity=bad)
+
+
+def test_cache_miss_then_hit_delivers_current_context_token_and_values(boundaries):
+    data = item()
+    a = stage(data)
+    context_a = master(data, a)
+    prompt_a = worker_prompt(a, context_a, data)
+    cached = start(boundaries, prompt_a)
+    assert cached[3] == [rt.item_content_identity(data)]
+    assert 'image' not in rt._WORKER_ITEMS[a][1]  # only a small, expiring receipt remains
+    end(boundaries, prompt_a, cached)
+    assert a not in rt._WORKER_ITEMS
+    assert rt.get_master_job(context_a).results[0]['values'] == {'value_3': 0}
+    b = stage(data)
+    context_b = master(data, b)
+    prompt_b = worker_prompt(b, context_b, data)
+    assert prompt_a['start'] == prompt_b['start']
+    # A native hit does NOT execute Start; End must consume this current upload.
+    end(boundaries, prompt_b, cached)
+    assert b not in rt._WORKER_ITEMS
+    state = rt.get_master_job(context_b)
+    assert state.expected_tokens[0] == b
+    assert torch.equal(state.results[0]['image'], data['image'])
+    assert rt._collect_output_lists([state.results[0]], state=state, include_values=True)[3][0]['value_3'] == 0
+    with pytest.raises(RuntimeError, match='missing|expired|consumed'):
+        end(boundaries, prompt_b, cached)
+
+
+@pytest.mark.parametrize('miss', [False, True])
+def test_expired_item_cannot_deliver_even_with_cached_start(boundaries, miss):
+    data = item()
+    token = stage(data)
+    context = master(data, token)
+    prompt = worker_prompt(token, context, data)
+    cached = start(boundaries, prompt) if miss else ([data['image']], [data['mask']], [0], [rt.item_content_identity(data)])
+    _, staged = rt._WORKER_ITEMS[token]
+    rt._WORKER_ITEMS[token] = (0, staged)
+    with pytest.raises(RuntimeError, match='missing|expired|consumed'):
+        end(boundaries, prompt, cached)
+    assert not rt.get_master_job(context).received_indexes
+    assert token not in rt._WORKER_ITEMS
+
+
+def test_identity_mismatch_on_start_and_on_cache_hit_fails_closed(boundaries):
+    data, different = item(), item()
+    different['image'][0, 0, 0, 0] = 3
+    for hit in (False, True):
+        token = stage(different)
+        context = master(different, token)
+        prompt = worker_prompt(token, context, data)
+        cached = ([data['image']], [data['mask']], [0], [rt.item_content_identity(data)])
+        with pytest.raises((ValueError, RuntimeError), match='identity'):
+            end(boundaries, prompt, cached) if hit else start(boundaries, prompt)
+        assert not rt.get_master_job(context).received_indexes
+        rt.delete_worker_item(token)
+
+
+@pytest.mark.parametrize('change', ['ambiguous', 'wrong-link', 'wrong-start-id', 'end-identity', 'end-index'])
+def test_start_requires_exact_paired_end_binding(boundaries, change):
+    data = item()
+    token = stage(data)
+    prompt = worker_prompt(token, 'context', data)
+    if change == 'ambiguous':
+        prompt['second-end'] = copy.deepcopy(prompt['end'])
+    elif change == 'wrong-link':
+        prompt['end']['inputs']['context'] = ['start', 2]
+    elif change == 'wrong-start-id':
+        prompt['renamed'] = prompt.pop('start')
+    elif change == 'end-identity':
+        prompt['end']['inputs']['worker_item_identity'] = 'sha256:' + '0' * 64
+    else:
+        prompt['end']['inputs']['worker_item_index'] = 1
+    boundaries.DistributedImageListStart.hidden = types.SimpleNamespace(prompt=prompt, unique_id='start')
+    with pytest.raises((RuntimeError, ValueError), match='paired|identity|index'):
+        boundaries.DistributedImageListStart.execute(worker_item_identity=rt.item_content_identity(data),
+                                                     image_list_local_execution=True)
+
+
+def test_end_checks_stable_context_before_delivering(boundaries):
+    data = item()
+    token = stage(data)
+    context = master(data, token)
+    prompt = worker_prompt(token, context, data)
+    cached = list(start(boundaries, prompt))
+    cached[3] = ['sha256:' + '0' * 64]
+    with pytest.raises((RuntimeError, ValueError), match='identity|context'):
+        end(boundaries, prompt, cached)
+    assert not rt.get_master_job(context).received_indexes
+
+
+def test_legacy_worker_path_still_consumes_and_delivers_job_context(boundaries):
+    data = item()
+    token = stage(data)
+    context = master(data, token)
+    prompt = graph()
+    prompt = rt.build_worker_prompt(prompt, rt.validate_image_list_regions(prompt)[0], token, context, 0, '')
+    prompt['start']['inputs']['image_list_local_execution'] = True
+    prompt['end']['inputs']['image_list_local_execution'] = True
+    cached = start(boundaries, prompt)
+    assert cached[3] == [context]
+    assert token not in rt._WORKER_ITEMS
+    end(boundaries, prompt, cached)
+    assert rt.get_master_job(context).received_indexes == {0}
+
+
+def test_remote_cache_hit_submits_fresh_callback_metadata(boundaries, monkeypatch):
+    data = item()
+    token = stage(data)
+    prompt = worker_prompt(token, 'fresh-context', data, local=False)
+    submit = AsyncMock()
+    monkeypatch.setattr(boundaries, 'submit_worker_result', submit)
+    cached = ([data['image']], [data['mask']], [0], [rt.item_content_identity(data)])
+    end(boundaries, prompt, cached)
+    assert submit.await_args.args[:4] == ('http://master', token, 'fresh-context', 0)
+    assert submit.await_args.args[4]['values']['value_3'] == {'type': 'INT', 'value': 0}
+    assert token not in rt._WORKER_ITEMS
+
+
+def test_item_route_returns_server_computed_identity(boundaries, monkeypatch):
+    from aiohttp import web
+    name = '_native_cache_identity_tests'
+    server = types.SimpleNamespace(PromptServer=types.SimpleNamespace(instance=types.SimpleNamespace(routes=web.RouteTableDef())))
+    monkeypatch.setitem(sys.modules, 'server', server)
+    spec = importlib.util.spec_from_file_location(name + '.api.image_list_routes', ROOT / 'api/image_list_routes.py')
+    module = importlib.util.module_from_spec(spec)
+    monkeypatch.setitem(sys.modules, spec.name, module)
+    spec.loader.exec_module(module)
+    data = item()
+    token = rt.create_worker_task_token()
+    monkeypatch.setattr(module, '_json_body', AsyncMock(return_value={
+        'token': token, 'context_id': 'context', 'item': rt.encode_item(data),
+        'item_identity': 'sha256:' + '0' * 64}))
+    response = asyncio.run(module.image_list_item_endpoint(object()))
+    assert response.status == 200
+    assert json.loads(response.body)['item_identity'] == rt.item_content_identity(data)
+    assert len(rt._WORKER_ITEMS[token]) == 2
+
+
+@pytest.mark.parametrize('responses', [
+    ('valid', 'valid'), ('valid', 'absent'), ('invalid', 'valid'),
+    ('bad-json', 'valid'), ('non-object', 'valid'),
+])
+def test_runtime_opts_in_per_upload_and_can_downgrade(monkeypatch, responses):
+    name = '_native_cache_identity_runtime'
+    for suffix in ('', '.image_list', '.api', '.api.orchestration', '.utils'):
+        module = types.ModuleType(name + suffix)
+        module.__path__ = [str(ROOT / suffix.lstrip('.').replace('.', '/'))]
+        monkeypatch.setitem(sys.modules, name + suffix, module)
+    spec = importlib.util.spec_from_file_location(name + '.image_list.runtime', ROOT / 'image_list/runtime.py')
+    runtime = importlib.util.module_from_spec(spec)
+    monkeypatch.setitem(sys.modules, spec.name, runtime)
+    spec.loader.exec_module(runtime)
+    workers = [{'id': 'worker', 'type': 'local'}]
+    dispatched, uploads = [], []
+    preflight = AsyncMock(return_value={'worker': True})
+    sync_media = AsyncMock()
+
+    class Response:
+        status = 200
+
+        def __init__(self, data):
+            self.data = data
+
+        async def __aenter__(self):
+            return self
+
+        async def __aexit__(self, *args):
+            pass
+
+        async def json(self):
+            if self.data == 'bad-json':
+                raise ValueError('old response is not JSON')
+            return self.data
+
+        def raise_for_status(self):
+            pass
+
+    class Session:
+        def post(self, url, *, data, **kw):
+            assert url.endswith('/distributed/image-list/item')
+            body = json.loads(data)
+            identity = runtime.store_worker_item(body['token'], body['item'])
+            mode = responses[len(uploads)]
+            uploads.append(body)
+            response = {'valid': {'item_identity': identity}, 'absent': {'status': 'stored'},
+                        'invalid': {'item_identity': 'bad'}, 'bad-json': 'bad-json', 'non-object': []}[mode]
+            return Response(response)
+
+        def get(self, url, **kw):
+            pid = url.rsplit('/', 1)[-1]
+            return Response({pid: {'status': {'completed': True, 'status_str': 'success'}}})
+
+        def delete(self, url, **kw):
+            runtime.delete_worker_item(url.rsplit('/', 1)[-1])
+            return Response({})
+
+    async def dispatch(worker, prompt, meta, **kwargs):
+        dispatched.append(copy.deepcopy(prompt))
+        delivery = prompt['end']['inputs']
+        token, index = delivery['worker_result_token'], delivery['worker_item_index']
+        identity = prompt['start']['inputs'].get('worker_item_identity')
+        options = {'item_identity': identity, 'item_index': index} if identity else {}
+        data = runtime.consume_worker_item(token, **options)
+        if identity:
+            runtime.finish_worker_item(token, identity, index)
+        runtime.set_master_result(context, index, token, {
+            'image': data['image'], 'text': str(index), 'values': {'value_3': index}})
+        assert kwargs['partial_execution_targets'] == ['end']
+        return f'worker-{index}'
+
+    dependencies = {
+        '.api.orchestration.dispatch': types.SimpleNamespace(
+            preflight_image_list_workers=preflight, dispatch_worker_prompt=dispatch),
+        '.api.orchestration.prompt_transform': types.SimpleNamespace(
+            _resolve_direct_workers_for_image_list=lambda *args: workers),
+        '.api.orchestration.media_sync': types.SimpleNamespace(sync_worker_media=sync_media),
+        '.utils.config': types.SimpleNamespace(load_config=lambda: {'settings': {}},
+                                              get_worker_timeout_seconds=lambda: 60),
+        '.utils.network': types.SimpleNamespace(build_master_callback_url=lambda *a, **kw: 'http://master',
+            build_worker_url=lambda worker, path: 'http://worker' + path,
+            get_client_session=AsyncMock(return_value=Session())),
+        '.utils.logging': types.SimpleNamespace(log=lambda *args: None),
+    }
+    for suffix, module in dependencies.items():
+        monkeypatch.setitem(sys.modules, name + suffix, module)
+    monkeypatch.setitem(sys.modules, 'aiohttp', types.SimpleNamespace(
+        ClientTimeout=lambda **kwargs: kwargs, ClientError=ConnectionError))
+    comfy = types.ModuleType('comfy')
+    comfy.model_management = types.SimpleNamespace(processing_interrupted=lambda: False)
+    monkeypatch.setitem(sys.modules, 'comfy', comfy)
+    monkeypatch.setitem(sys.modules, 'comfy.model_management', comfy.model_management)
+    items = [item(), item()]
+    items[1]['index'] = 1
+    context = runtime.create_master_job('start', 'end', graph(), items, ['worker'], '')
+    state = runtime.get_master_job(context)
+    try:
+        result = asyncio.run(runtime._run_image_list_job(state, include_values=True))
+        assert result[2] == ['0', '1']
+        assert [row['value_3'] for row in result[3]] == [0, 1]
+        assert len(uploads) == len(dispatched) == 2
+        assert sync_media.await_count == 1
+        for index, prompt in enumerate(dispatched):
+            start_inputs = prompt['start']['inputs']
+            if responses[index] == 'valid':
+                assert start_inputs == {'worker_item_identity': runtime.item_content_identity(items[index])}
+                assert prompt['end']['inputs']['worker_context_id'] == context
+            else:
+                assert 'worker_item_identity' not in start_inputs
+                assert start_inputs['worker_item_token'] == uploads[index]['token']
+                assert start_inputs['worker_context_id'] == context
+                assert 'worker_item_identity' not in prompt['end']['inputs']
+            assert prompt['end']['inputs']['worker_result_token'] == uploads[index]['token']
+        assert not runtime._WORKER_ITEMS
+    finally:
+        runtime.remove_master_job(context)
+        runtime._WORKER_ITEMS.clear()
+
+
+def test_runtime_local_path_hashes_without_transport(monkeypatch):
+    name = '_native_cache_identity_local_runtime'
+    for suffix in ('', '.image_list', '.api', '.api.orchestration', '.utils'):
+        module = types.ModuleType(name + suffix)
+        module.__path__ = [str(ROOT / suffix.lstrip('.').replace('.', '/'))]
+        monkeypatch.setitem(sys.modules, name + suffix, module)
+    spec = importlib.util.spec_from_file_location(name + '.image_list.runtime', ROOT / 'image_list/runtime.py')
+    runtime = importlib.util.module_from_spec(spec)
+    monkeypatch.setitem(sys.modules, spec.name, runtime)
+    spec.loader.exec_module(runtime)
+    dispatched = []
+
+    class Slot:
+        def __init__(self, **kw):
+            pass
+
+        def submit(self, prompt, prompt_id, end_id, workflow_meta):
+            dispatched.append(prompt)
+            delivery = prompt[end_id]['inputs']
+            identity = prompt['start']['inputs']['worker_item_identity']
+            data = runtime.consume_worker_item(delivery['worker_result_token'], item_identity=identity,
+                                                item_index=delivery['worker_item_index'])
+            runtime.finish_worker_item(delivery['worker_result_token'], identity, data['index'])
+            runtime.set_master_result(delivery['worker_context_id'], data['index'], delivery['worker_result_token'],
+                                      {'image': data['image'], 'values': {'value_3': data['index']}})
+            future = asyncio.get_running_loop().create_future()
+            future.set_result(None)
+            return future
+
+        def request_stop(self):
+            pass
+
+        async def close(self):
+            pass
+
+    def forbidden(*args, **kwargs):
+        raise AssertionError('local path attempted HTTP/transport')
+
+    dependencies = {
+        '.image_list.local_executor': types.SimpleNamespace(LocalExecutionSlot=Slot),
+        '.api.orchestration.dispatch': types.SimpleNamespace(
+            preflight_image_list_workers=forbidden, dispatch_worker_prompt=forbidden),
+        '.api.orchestration.prompt_transform': types.SimpleNamespace(
+            _resolve_direct_workers_for_image_list=lambda *args: []),
+        '.utils.config': types.SimpleNamespace(load_config=lambda: {'settings': {}},
+                                              get_worker_timeout_seconds=lambda: 60),
+        '.utils.network': types.SimpleNamespace(build_master_callback_url=forbidden,
+                                                build_worker_url=forbidden, get_client_session=forbidden),
+        '.utils.logging': types.SimpleNamespace(log=lambda *args: None),
+    }
+    for suffix, module in dependencies.items():
+        monkeypatch.setitem(sys.modules, name + suffix, module)
+    monkeypatch.setitem(sys.modules, 'aiohttp', types.SimpleNamespace(
+        ClientTimeout=lambda **kwargs: kwargs, ClientError=ConnectionError))
+    comfy = types.ModuleType('comfy')
+    comfy.model_management = types.SimpleNamespace(processing_interrupted=lambda: False)
+    monkeypatch.setitem(sys.modules, 'comfy', comfy)
+    monkeypatch.setitem(sys.modules, 'comfy.model_management', comfy.model_management)
+    monkeypatch.setattr(runtime, 'MAX_PAYLOAD_BYTES', 1)
+    monkeypatch.setattr(runtime, 'encode_item', forbidden)
+    data = item()
+    context = runtime.create_master_job('start', 'end', graph(), [data], [], '', master_participates=True)
+    try:
+        result = asyncio.run(runtime._run_image_list_job(runtime.get_master_job(context), include_values=True))
+        assert dispatched[0]['start']['inputs'] == {'worker_item_identity': runtime.item_content_identity(data),
+                                                   'image_list_local_execution': True}
+        assert result[3][0]['value_3'] == 0
+        assert not runtime._WORKER_ITEMS
+    finally:
+        runtime.remove_master_job(context)
+        runtime._WORKER_ITEMS.clear()
+
+
+def test_real_native_signatures_and_cached_delivery():
+    core = Path(os.environ.get('COMFYUI_SOURCE_ROOT', str(Path(sys.executable).parents[2])))
+    if not (core / 'comfy_execution/caching.py').exists():
+        pytest.skip('real ComfyUI source not available beside this interpreter')
+    result = subprocess.run([sys.executable, str(ROOT / 'tests/helpers/image_list_native_cache_identity.py'), str(core)],
+                            text=True, capture_output=True, timeout=90)
+    assert result.returncode == 0, result.stdout + result.stderr
+    assert 'NATIVE_CACHE_IDENTITY_OK' in result.stdout

tests/test_image_list_runtime_boundaries.py

@@ -0,0 +1,358 @@
+"""Exercise production per-item orchestration through controlled provider seams."""
+import asyncio
+import importlib.util
+import json
+from pathlib import Path
+import sys
+import types
+import unittest
+from unittest.mock import AsyncMock, Mock, patch
+
+import torch
+
+from test_image_list import graph_one_region
+
+ROOT = Path(__file__).resolve().parents[1]
+
+
+class Response:
+    def __init__(self, data=None, status=200):
+        self.data, self.status = data, status
+
+    async def __aenter__(self):
+        return self
+
+    async def __aexit__(self, *_args):
+        return None
+
+    async def json(self):
+        return self.data
+
+    def raise_for_status(self):
+        if self.status >= 400:
+            raise RuntimeError(f'HTTP {self.status}')
+
+
+class RuntimeBoundaryTests(unittest.IsolatedAsyncioTestCase):
+    async def asyncSetUp(self):
+        name = '_image_list_boundaries'
+        self.modules = {}
+        for suffix in ('', '.image_list', '.api', '.api.orchestration', '.utils'):
+            module = types.ModuleType(name + suffix)
+            module.__path__ = []
+            self.modules[name + suffix] = module
+        self.patch_modules = patch.dict(sys.modules, self.modules)
+        self.patch_modules.start()
+        self.addCleanup(self.patch_modules.stop)
+        spec = importlib.util.spec_from_file_location(name + '.image_list.runtime', ROOT / 'image_list/runtime.py')
+        self.rt = importlib.util.module_from_spec(spec)
+        sys.modules[spec.name] = self.rt
+        spec.loader.exec_module(self.rt)
+        self.worker = {'id': 'worker', 'host': 'worker', 'type': 'remote', 'port': 8188}
+        self.config = {'workers': [self.worker], 'settings': {}}
+        self.prompts, self.uploads, self.deletions, self.cancel_requests = [], [], [], []
+        self.reject = False
+        self.interrupted = False
+        self.completed = True
+        self.queue = {'queue_running': [], 'queue_pending': []}
+        self.callback = Mock(side_effect=lambda worker, **_kwargs:
+                             'http://127.0.0.1:8188' if worker['type'] == 'local' else 'http://public-master')
+        self.sync = AsyncMock(side_effect=self.sync_media)
+        self.dispatch = AsyncMock(side_effect=self.dispatch_prompt)
+        self.logs = []
+        dependencies = {
+            '.api.orchestration.dispatch': types.SimpleNamespace(
+                preflight_image_list_workers=AsyncMock(return_value={'worker': True}),
+                dispatch_worker_prompt=self.dispatch),
+            '.api.orchestration.prompt_transform': types.SimpleNamespace(
+                _resolve_direct_workers_for_image_list=lambda config, ids: config['workers']),
+            '.api.orchestration.media_sync': types.SimpleNamespace(
+                sync_worker_media=self.sync, fetch_worker_path_separator=AsyncMock(return_value='/'),
+                convert_paths_for_platform=lambda prompt, separator: prompt),
+            '.utils.config': types.SimpleNamespace(load_config=lambda: self.config,
+                get_worker_timeout_seconds=lambda: 60,
+                get_image_list_job_timeout_seconds=lambda config=None: 0),
+            '.utils.network': types.SimpleNamespace(build_master_url=lambda **kwargs: 'http://public-master',
+                build_master_callback_url=self.callback,
+                build_worker_url=lambda worker, path='': f"http://{worker['id']}{path}",
+                get_client_session=AsyncMock(return_value=self)),
+            '.utils.logging': types.SimpleNamespace(log=self.logs.append),
+        }
+        for suffix, module in dependencies.items():
+            sys.modules[name + suffix] = module
+        self.config_module = dependencies['.utils.config']
+        # Other legacy test modules replace aiohttp during collection. This
+        # controlled provider fixture needs only its explicit timeout/error seam.
+        sys.modules['aiohttp'] = types.SimpleNamespace(
+            ClientTimeout=lambda **kwargs: kwargs, ClientError=ConnectionError, ClientConnectionError=ConnectionError)
+        comfy = types.ModuleType('comfy')
+        comfy.model_management = types.SimpleNamespace(processing_interrupted=lambda: self.interrupted)
+        sys.modules['comfy'] = comfy
+        sys.modules['comfy.model_management'] = comfy.model_management
+
+    async def sync_media(self, worker, prompt):
+        prompt['7']['inputs']['image'] = 'shared/ref.png'
+
+    async def dispatch_prompt(self, worker, prompt, meta, **kwargs):
+        self.assertEqual(kwargs.get('partial_execution_targets'), ['5'])
+        if self.reject:
+            raise RuntimeError('prompt rejected')
+        self.prompts.append((json.loads(json.dumps(prompt)), meta))
+        inputs = prompt['5']['inputs']
+        token = inputs['worker_result_token']
+        item = self.rt.consume_worker_item(token)
+        self.rt.set_master_result(self.context, inputs['worker_item_index'], token, {'image': item['image']})
+        return f'prompt-{len(self.prompts)}'
+
+    def post(self, url, *, json=None, data=None, **kwargs):
+        if data is not None:
+            json = __import__('json').loads(data)
+        if url.endswith('/distributed/image-list/item'):
+            self.uploads.append(json['token'])
+            self.rt.store_worker_item(json['token'], json['item'])
+        else:
+            self.cancel_requests.append((url, json))
+        return Response({})
+
+    def delete(self, url, **kwargs):
+        token = url.rsplit('/', 1)[-1]
+        self.deletions.append(token)
+        self.rt.delete_worker_item(token)
+        return Response({})
+
+    def get(self, url, **kwargs):
+        if url.endswith('/queue'):
+            return Response(self.queue)
+        pid = url.rsplit('/', 1)[-1]
+        return Response({pid: {'status': {'completed': True, 'status_str': 'success'}}}
+                        if self.completed else {})
+
+    def create(self, count=2):
+        prompt = graph_one_region()
+        prompt['7'] = {'class_type': 'LoadImage', 'inputs': {'image': 'shared/ref.png [output]'}}
+        prompt['4']['inputs']['reference'] = ['7', 0]
+        self.metadata = {'nodes': [{'id': 42}], 'extra': {'label': 'generic fixture'}}
+        self.context = self.rt.create_master_job('3', '5', prompt,
+            [{'image': torch.zeros((1, 2, 2, 3)), 'index': i} for i in range(count)],
+            ['worker'], 'http://public-master', self.metadata)
+        return self.context
+
+    async def test_runtime_preflight_rejects_old_target_capability_before_upload(self):
+        from test_dispatch_selection import _load_dispatch_module, _FakeSession, _valid_capability
+        real_dispatch = _load_dispatch_module()
+        capability = _valid_capability()
+        capability.pop('execution_targets')
+        session = _FakeSession({('worker', 'capabilities'): (200, capability)})
+        module = sys.modules['_image_list_boundaries.api.orchestration.dispatch']
+        module.preflight_image_list_workers = real_dispatch.preflight_image_list_workers
+        with patch.object(real_dispatch, 'get_client_session', AsyncMock(return_value=session)), patch.object(
+            real_dispatch, 'build_worker_url', side_effect=lambda worker, path: f"http://{worker['id']}{path}"
+        ):
+            with self.assertRaisesRegex(RuntimeError, 'execution_targets.*upgrade'):
+                await self.rt.run_image_list_job(self.create(1))
+        self.assertFalse(self.uploads)
+        self.dispatch.assert_not_awaited()
+        self.assertFalse(self.rt._MASTER_JOBS)
+
+    async def test_each_assignment_reuses_media_paths_with_fresh_tokens_and_metadata(self):
+        context = self.create()
+        await self.rt.run_image_list_job(context)
+        self.assertEqual(self.sync.await_count, 1)
+        self.assertEqual([p['7']['inputs']['image'] for p, _ in self.prompts],
+                         ['shared/ref.png', 'shared/ref.png'])
+        tokens = [p['5']['inputs']['worker_result_token'] for p, _ in self.prompts]
+        self.assertEqual(len(set(tokens)), 2)
+        self.assertEqual([p['5']['inputs']['worker_item_index'] for p, _ in self.prompts], [0, 1])
+        self.assertEqual([meta for _, meta in self.prompts], [self.metadata, self.metadata])
+
+    async def test_annotated_media_roots_survive_windows_prep_and_repeated_items(self):
+        import tempfile
+
+        context = self.create(2)
+        state = self.rt.get_master_job(context)
+        state.original_prompt['8'] = {'class_type': 'LoadImage', 'inputs': {'image': 'shared\\ref.png [temp]'}}
+        state.original_prompt['4']['inputs']['second_reference'] = ['8', 0]
+        name = '_image_list_boundaries.api.orchestration.media_sync'
+        spec = importlib.util.spec_from_file_location(name, ROOT / 'api/orchestration/media_sync.py')
+        assert spec is not None and spec.loader is not None
+        media = importlib.util.module_from_spec(spec)
+        sys.modules['_image_list_boundaries.utils.logging'].debug_log = self.logs.append
+        sys.modules['_image_list_boundaries.utils.trace_logger'] = types.SimpleNamespace(
+            trace_debug=lambda *args: None, trace_info=lambda *args: None)
+        spec.loader.exec_module(media)
+        media.fetch_worker_path_separator = AsyncMock(return_value='\\')
+        converted = Mock(wraps=media.convert_paths_for_platform)
+        media.convert_paths_for_platform = converted
+        uploads = []
+
+        async def upload(worker, filename, content, digest, mime, **kwargs):
+            uploads.append((filename, content))
+            return True, f'worker-input/{len(uploads)}.png'
+
+        media._upload_media_to_worker = AsyncMock(side_effect=upload)
+        media.sync_worker_media = AsyncMock(wraps=media.sync_worker_media)
+        with tempfile.TemporaryDirectory(prefix='image-list-media-') as directory:
+            root = Path(directory)
+            for kind in ('input', 'output', 'temp'):
+                path = root / kind / 'shared/ref.png'
+                path.parent.mkdir(parents=True)
+                path.write_bytes(kind.encode())
+
+            def resolve_annotated(filename):
+                for kind in ('output', 'temp', 'input'):
+                    suffix = f' [{kind}]'
+                    if filename.endswith(suffix):
+                        return str(root / kind / filename[:-len(suffix)])
+                return str(root / 'input' / filename)
+
+            with patch.dict(sys.modules, {name: media, 'folder_paths': types.SimpleNamespace(
+                    get_annotated_filepath=resolve_annotated)}):
+                await self.rt.run_image_list_job(context)
+        self.assertEqual(uploads, [('shared/ref.png [output]', b'output'),
+                                   ('shared/ref.png [temp]', b'temp')])
+        media.sync_worker_media.assert_awaited_once()
+        media.fetch_worker_path_separator.assert_awaited_once()
+        converted.assert_called_once()
+        self.assertEqual(converted.call_args.args[1], '\\')
+        for prompt, _ in self.prompts:
+            self.assertEqual(prompt['7']['inputs']['image'], 'worker-input/1.png')
+            self.assertEqual(prompt['8']['inputs']['image'], 'worker-input/2.png')
+        self.assertEqual(state.original_prompt['7']['inputs']['image'], 'shared/ref.png [output]')
+
+    async def test_local_worker_callback_does_not_use_public_master_address(self):
+        self.worker['type'] = 'local'
+        await self.rt.run_image_list_job(self.create(1))
+        self.assertEqual(self.prompts[0][0]['5']['inputs']['master_callback_url'], 'http://127.0.0.1:8188')
+        self.callback.assert_called()
+
+    async def test_remote_worker_uses_topology_callback_helper(self):
+        await self.rt.run_image_list_job(self.create(1))
+        self.assertEqual(self.prompts[0][0]['5']['inputs']['master_callback_url'], 'http://public-master')
+        self.callback.assert_called()
+
+    def test_metadata_is_forwarded_into_master_start_prompt(self):
+        from test_prompt_transform import pt
+        prompt = graph_one_region()
+        metadata = {'nodes': [{'id': 1}]}
+        index = pt.PromptIndex(prompt)
+        transformed = pt.apply_participant_overrides(
+            prompt, 'master', ['worker'], {'3': 'run'}, 'http://public-master', False,
+            index, workflow_meta=metadata)
+        self.assertEqual(transformed['3']['inputs']['image_list_workflow_meta'], metadata)
+
+    async def test_aggregate_input_limit_fails_before_upload(self):
+        self.rt.MAX_REQUEST_BYTES = 64
+        with self.assertRaisesRegex(ValueError, 'complete.*request'):
+            await self.rt.run_image_list_job(self.create(1))
+        self.assertFalse(self.uploads)
+        self.assertFalse(self.rt._WORKER_ITEMS)
+
+    def test_tensor_and_combined_request_limits_preserve_dtype(self):
+        item = {'image': torch.zeros((1, 4, 4, 3)), 'index': 0}
+        encoded = self.rt.encode_item(item)
+        payload = {'context_id': 'example', 'token': 't' * 48, 'item': encoded}
+        single = self.rt.encode_request_payload(payload)
+        self.assertEqual(json.loads(single)['item']['image']['dtype'], 'float32')
+        self.rt.MAX_REQUEST_BYTES = len(single)
+        self.assertEqual(self.rt.encode_request_payload(payload), single)
+        item['values'] = {'value_1': item['image']}
+        item['mask'] = torch.zeros((1, 4, 4))
+        payload['item'] = self.rt.encode_item(item)
+        with self.assertRaisesRegex(ValueError, 'complete.*request'):
+            self.rt.encode_request_payload(payload)
+        # Size is checked before allocating/copying a 4K tensor onto the CPU.
+        with self.assertRaisesRegex(ValueError, 'transport limit'):
+            self.rt.encode_value(torch.empty((1, 2160, 3840, 3), device='meta'), 'IMAGE')
+
+    async def test_rejected_submission_reclaims_uploaded_token_without_prompt_id(self):
+        self.reject = True
+        with self.assertRaisesRegex(RuntimeError, 'prompt rejected'):
+            await self.rt.run_image_list_job(self.create(1))
+        self.assertEqual(self.deletions, self.uploads)
+        self.assertFalse(self.rt._WORKER_ITEMS)
+        self.assertFalse(self.rt._MASTER_JOBS)
+        self.assertFalse(self.cancel_requests)
+
+    async def test_user_interrupt_during_pending_submission_is_bounded(self):
+        entered = asyncio.Event()
+        async def never_ack(*args, **kwargs):
+            entered.set()
+            await asyncio.Event().wait()
+        self.dispatch.side_effect = never_ack
+        self.rt.SUBMISSION_CANCEL_GRACE_SECONDS = 0.02
+        task = asyncio.create_task(self.rt.run_image_list_job(self.create(2)))
+        await asyncio.wait_for(entered.wait(), 1)
+        self.interrupted = True
+        with self.assertRaises(self.rt.ImageListCancelledError):
+            await asyncio.wait_for(task, 0.8)
+        self.assertEqual(len(self.uploads), 1)
+        self.assertEqual(self.deletions, self.uploads)
+        self.assertFalse(self.rt._MASTER_JOBS)
+        self.assertTrue(any('ambiguous' in line for line in self.logs), self.logs)
+
+    async def test_owned_running_prompt_survives_cold_load_and_old_hour_cap(self):
+        self.completed = False
+        context = self.create(2)
+        clock = [0.0]
+        calls = [0]
+        original_get = self.get
+        def get(url, **kwargs):
+            if '/history/' in url:
+                calls[0] += 1
+                clock[0] += 3700
+                self.completed = calls[0] >= 3
+            if url.endswith('/queue'):
+                return Response({'queue_running': [[0, 'prompt-1']] if calls[0] == 2 else [],
+                                 'queue_pending': [[0, 'prompt-1']] if calls[0] == 1 else []})
+            return original_get(url, **kwargs)
+        self.rt.time = types.SimpleNamespace(monotonic=lambda: clock[0])
+        with patch.object(self, 'get', side_effect=get):
+            result = await self.rt.run_image_list_job(context)
+        self.assertEqual(len(result[0]), 2)
+        self.assertGreaterEqual(clock[0], 7200)
+        self.assertFalse(self.cancel_requests)
+
+    async def test_unrelated_busy_prompt_does_not_refresh_owned_liveness(self):
+        self.completed = False
+        clock = [0.0]
+        original_get = self.get
+        def get(url, **kwargs):
+            if '/history/' in url:
+                clock[0] += 70
+            if url.endswith('/queue'):
+                return Response({'queue_running': [[0, 'not-our-prompt']], 'queue_pending': []})
+            return original_get(url, **kwargs)
+        self.rt.time = types.SimpleNamespace(monotonic=lambda: clock[0])
+        with patch.object(self, 'get', side_effect=get):
+            with self.assertRaisesRegex(TimeoutError, 'liveness'):
+                await self.rt.run_image_list_job(self.create(1))
+        self.assertEqual(self.cancel_requests, [
+            ('http://worker/queue', {'delete': ['prompt-1']}),
+            ('http://worker/interrupt', {'prompt_id': 'prompt-1'})])
+
+    async def test_failed_queue_cleanup_still_interrupts_and_reclaims_staging(self):
+        self.completed = False
+        self.config_module.get_image_list_job_timeout_seconds = lambda config=None: 0.03
+        original_post = self.post
+        def post(url, **kwargs):
+            result = original_post(url, **kwargs)
+            return Response(status=500) if url.endswith('/queue') else result
+        with patch.object(self, 'post', side_effect=post):
+            with self.assertRaisesRegex(TimeoutError, 'total deadline'):
+                await self.rt.run_image_list_job(self.create(2))
+        self.assertEqual(self.cancel_requests, [
+            ('http://worker/queue', {'delete': ['prompt-1']}),
+            ('http://worker/interrupt', {'prompt_id': 'prompt-1'})])
+        self.assertEqual(self.deletions, self.uploads)
+        self.assertTrue(any('cleanup unconfirmed' in line for line in self.logs), self.logs)
+        self.assertEqual(self.dispatch.await_count, 1)
+
+    async def test_configured_total_deadline_cleans_owned_work(self):
+        self.completed = False
+        self.config_module.get_image_list_job_timeout_seconds = lambda config=None: 0.03
+        self.queue = {'queue_running': [[0, 'prompt-1']], 'queue_pending': []}
+        with self.assertRaisesRegex(TimeoutError, 'total deadline'):
+            await asyncio.wait_for(self.rt.run_image_list_job(self.create(2)), 1)
+        self.assertEqual(len(self.uploads), 1)
+        self.assertEqual(self.deletions, self.uploads)
+        self.assertFalse(self.rt._MASTER_JOBS)

tests/test_image_list_timeout_config.py

@@ -0,0 +1,106 @@
+import importlib.util
+import sys
+import types
+from pathlib import Path
+
+import pytest
+
+
+def _load_config_module():
+    module_path = Path(__file__).resolve().parents[1] / "utils" / "config.py"
+    package_name = "image_list_timeout_config_testpkg"
+
+    for mod_name in list(sys.modules):
+        if mod_name == package_name or mod_name.startswith(f"{package_name}."):
+            del sys.modules[mod_name]
+
+    root_pkg = types.ModuleType(package_name)
+    root_pkg.__path__ = [str(module_path.parent)]
+    sys.modules[package_name] = root_pkg
+
+    logging_module = types.ModuleType(f"{package_name}.logging")
+    logging_module.log = lambda *_args, **_kwargs: None
+    logging_module.debug_log = lambda *_args, **_kwargs: None
+    sys.modules[f"{package_name}.logging"] = logging_module
+
+    constants_module = types.ModuleType(f"{package_name}.constants")
+    constants_module.HEARTBEAT_TIMEOUT = 30
+    sys.modules[f"{package_name}.constants"] = constants_module
+
+    spec = importlib.util.spec_from_file_location(f"{package_name}.config", module_path)
+    module = importlib.util.module_from_spec(spec)
+    assert spec is not None and spec.loader is not None
+    spec.loader.exec_module(module)
+    return module
+
+
+config_module = _load_config_module()
+
+
+SETTING_KEY = "image_list_job_timeout_seconds"
+
+
+def _config_with_timeout(value):
+    return {"settings": {SETTING_KEY: value}}
+
+
+def test_default_config_sets_image_list_job_timeout_to_zero():
+    assert config_module.get_default_config()["settings"][SETTING_KEY] == 0.0
+
+
+def test_missing_value_defaults_to_disabled_timeout():
+    assert config_module.get_image_list_job_timeout_seconds({}) == 0.0
+    assert config_module.get_image_list_job_timeout_seconds({"settings": {}}) == 0.0
+
+
+def test_none_config_loads_current_config(monkeypatch):
+    cfg = _config_with_timeout("12.5")
+    monkeypatch.setattr(config_module, "load_config", lambda: cfg)
+
+    assert config_module.get_image_list_job_timeout_seconds() == 12.5
+
+
+@pytest.mark.parametrize(
+    ("value", "expected"),
+    [
+        (0, 0.0),
+        (0.0, 0.0),
+        ("0", 0.0),
+        (" 0.0 ", 0.0),
+        (1, 1.0),
+        (2.5, 2.5),
+        ("3", 3.0),
+        (" 4.25 ", 4.25),
+    ],
+)
+def test_accepts_zero_and_positive_numeric_values(value, expected):
+    result = config_module.get_image_list_job_timeout_seconds(_config_with_timeout(value))
+
+    assert result == expected
+    assert isinstance(result, float)
+
+
+@pytest.mark.parametrize(
+    "value",
+    [
+        True,
+        False,
+        -1,
+        -0.01,
+        float("nan"),
+        float("inf"),
+        float("-inf"),
+        "not-a-number",
+        "",
+        "   ",
+        "-1",
+        "nan",
+        "inf",
+        None,
+        [],
+        {},
+    ],
+)
+def test_rejects_invalid_values_with_setting_name(value):
+    with pytest.raises(ValueError, match=SETTING_KEY):
+        config_module.get_image_list_job_timeout_seconds(_config_with_timeout(value))

tests/test_image_list_values.py

@@ -0,0 +1,165 @@
+"""Context-carried values and sparse, per-item End overrides."""
+import copy
+from unittest.mock import patch
+
+import pytest
+import torch
+
+from image_list import runtime as rt
+from image_list.execution_plan import prepare_image_list_execution
+
+
+def value_graph():
+    # Use a minimal, closed region with an override-only ancestor.
+    return {
+        'source': {'class_type': 'Source', 'inputs': {}},
+        'start': {'class_type': rt.START_CLASS, 'inputs': {'images': ['source', 0]}},
+        'override': {'class_type': 'Transform', 'inputs': {'value': ['start', 4]}},
+        'end': {'class_type': rt.END_CLASS, 'inputs': {
+            'images': ['start', 0], 'context': ['start', 3], 'value_1': ['override', 0],
+            'value_10': ['start', 5]}},
+    }
+
+
+def values():
+    return {'value_1': torch.arange(72, dtype=torch.float64).reshape(2, 3, 4, 3),
+            'value_2': torch.ones(3, 2, 2, dtype=torch.bfloat16), 'value_3': 'name',
+            'value_4': 42, 'value_5': 0.125, 'value_6': True}
+
+
+def assert_values(actual, expected):
+    assert actual.keys() == expected.keys()
+    for key, value in expected.items():
+        assert type(actual[key]) is type(value)
+        if isinstance(value, torch.Tensor):
+            assert actual[key].dtype == value.dtype
+            assert torch.equal(actual[key], value)
+        else:
+            assert actual[key] == value
+
+
+def test_normalize_and_roundtrip_all_override_kinds_keeps_batches_whole():
+    expected = values()
+    result = rt.normalize_worker_result(torch.zeros(1, 2, 2, 3),
+                                       values={key: [value] for key, value in expected.items()})
+    assert_values(result['values'], expected)
+    wire = rt.encode_result(result)
+    assert_values(rt.decode_result(wire)['values'], expected)
+    assert wire['values']['value_1']['shape'][0] == 2
+    assert wire['values']['value_2']['shape'][0] == 3
+
+
+@pytest.mark.parametrize('bad', [None, [], [None], [1, 2], [[1]], object(),
+                                     float('nan'), float('inf'), torch.tensor([float('nan')]),
+                                     torch.tensor([float('inf')]), torch.ones(1, dtype=torch.int32)])
+def test_invalid_override_rejected(bad):
+    with pytest.raises((ValueError, TypeError)):
+        rt.normalize_worker_result(torch.zeros(1, 2, 2, 3), values={'value_1': bad})
+
+
+def test_unknown_override_names_and_wire_types_rejected():
+    for name in ('value_0', 'value_11', 'values.value_1'):
+        with pytest.raises(ValueError, match='slot'):
+            rt.normalize_worker_result(torch.zeros(1, 2, 2, 3), values={name: 1})
+    base = rt.encode_result({'image': torch.zeros(1, 2, 2, 3)})
+    for payload in (None, [], {'value_0': {'type': 'INT', 'value': 1}},
+                    {'value_1': {'type': 'LATENT', 'value': {}}},
+                    {'value_1': {'type': 'FLOAT', 'value': float('nan')}},
+                    {'value_1': {'type': 'STRING', 'value': None}},
+                    {'value_1': rt.encode_value(torch.tensor([float('inf')]), 'TENSOR')}):
+        with pytest.raises((ValueError, TypeError)):
+            rt.decode_result({**base, 'values': payload})
+
+
+def test_no_overrides_keeps_legacy_wire_and_does_not_encode_original_tensors():
+    image = torch.zeros(1, 2, 2, 3)
+    result = rt.normalize_worker_result(image)
+    assert set(result) == {'image', 'mask', 'text'}
+    with patch.object(rt, 'encode_value', wraps=rt.encode_value) as encode:
+        wire = rt.encode_result(result)
+    assert set(wire) == {'image', 'mask', 'text'}
+    assert [call.args[1] for call in encode.call_args_list] == ['IMAGE', 'MASK', 'STRING']
+    assert set(rt.decode_result(wire)) == {'image', 'mask', 'text'}
+
+
+def test_override_ancestors_survive_plan_and_worker_prompt():
+    prompt = value_graph()
+    frozen = copy.deepcopy(prompt)
+    prepared, targets = prepare_image_list_execution(prompt)
+    assert prepared == frozen and prompt == frozen
+    assert targets == ['end']
+    region = rt.validate_image_list_regions(prepared)[0]
+    assert 'override' in region['branch_ids']
+    worker = rt.build_worker_prompt(prepared, region, 'token', 'context', 0, 'http://master')
+    assert worker['end']['inputs']['value_1'] == ['override', 0]
+    assert worker['end']['inputs']['value_10'] == ['start', 5]
+    assert 'source' not in worker
+    for bad in ('value_0', 'value_11', 'values.value_1'):
+        invalid = copy.deepcopy(prompt)
+        invalid['end']['inputs'][bad] = ['start', 4]
+        with pytest.raises(ValueError, match='slot'):
+            rt.validate_image_list_regions(invalid)
+
+
+def test_missing_or_unexpected_result_override_fails_before_recording():
+    prompt = value_graph()
+    items = rt.normalize_image_list_inputs(torch.zeros(1, 2, 2, 3), values={'value_1': 'original'})
+    context = rt.create_master_job('start', 'end', prompt, items, ['worker'], '')
+    try:
+        rt.register_assignment(context, 0, 'worker', 'token')
+        for result_values in ({}, {'value_1': 'updated'},
+                              {'value_1': 'updated', 'value_10': 10, 'value_9': False}):
+            with pytest.raises(ValueError, match='override'):
+                rt.set_master_result(context, 0, 'token', {'image': items[0]['image'], 'values': result_values})
+            assert rt.get_master_job(context).received_indexes == set()
+        rt.set_master_result(context, 0, 'token', {'image': items[0]['image'],
+                                                  'values': {'value_1': 'updated', 'value_10': 10}})
+        assert rt.get_master_job(context).received_indexes == {0}
+    finally:
+        rt.remove_master_job(context)
+
+
+def test_override_return_request_budget_is_aggregate():
+    result = rt.normalize_worker_result(torch.zeros(1, 2, 2, 3), values={'value_1': ['x' * 500]})
+    with patch.object(rt, 'MAX_REQUEST_BYTES', 400):
+        with pytest.raises(ValueError, match='request'):
+            rt.encode_request_payload({'result': rt.encode_result(result)})
+
+
+class TestResultHTTP:
+    def test_missing_and_invalid_overrides_fail_through_real_http(self):
+        import asyncio
+        from aiohttp import web
+        from aiohttp.test_utils import TestClient, TestServer
+        from test_image_list_integration_gaps import load_routes
+        async def check():
+            _module, routes = load_routes()
+            app = web.Application()
+            app.add_routes(routes)
+            prompt = value_graph()
+            items = rt.normalize_image_list_inputs(torch.zeros(1, 2, 2, 3), values={'value_1': 'original'})
+            context = rt.create_master_job('start', 'end', prompt, items, ['worker'], '')
+            token = rt.create_worker_task_token()
+            rt.register_assignment(context, 0, 'worker', token)
+            try:
+                async with TestClient(TestServer(app)) as client:
+                    caps = await (await client.get('/distributed/image-list/capabilities')).json()
+                    assert caps['value_results'] is True and caps['protocol_version'] == 2
+                    result = rt.encode_result({'image': items[0]['image']})
+                    envelope = {'context_id': context, 'token': token, 'item_index': 0, 'result': result}
+                    response = await client.post('/distributed/image-list/result', json=envelope)
+                    assert response.status == 400
+                    assert 'missing' in (await response.json())['error']
+                    assert not rt.get_master_job(context).received_indexes
+                    result['values'] = {'value_1': {'type': 'INT', 'value': [1, 2]},
+                                        'value_10': {'type': 'INT', 'value': 10}}
+                    response = await client.post('/distributed/image-list/result', json=envelope)
+                    assert response.status == 400
+                    assert not rt.get_master_job(context).received_indexes
+                    result['values']['value_1'] = {'type': 'STRING', 'value': 'new'}
+                    response = await client.post('/distributed/image-list/result', json=envelope)
+                    assert response.status == 200
+                    assert rt.get_master_job(context).results[0]['values'] == {'value_1': 'new', 'value_10': 10}
+            finally:
+                rt.remove_master_job(context)
+        asyncio.run(check())

tests/test_image_list_websocket_deadline.py

@@ -0,0 +1,212 @@
+import asyncio
+import importlib.util
+import sys
+import types
+import unittest
+import uuid
+from pathlib import Path
+from typing import Any, cast
+from unittest.mock import patch
+
+import aiohttp
+from aiohttp import web
+
+
+def _load_dispatch_module():
+    module_path = Path(__file__).resolve().parents[1] / "api" / "orchestration" / "dispatch.py"
+    package_name = f"image_list_ws_deadline_testpkg_{uuid.uuid4().hex}"
+    package_modules = {}
+
+    for suffix in ("", ".api", ".api.orchestration", ".utils"):
+        package = types.ModuleType(package_name + suffix)
+        package.__path__ = []
+        package_modules[package_name + suffix] = package
+
+    logging_module = types.ModuleType(f"{package_name}.utils.logging")
+    setattr(logging_module, "debug_log", lambda *_args, **_kwargs: None)
+    setattr(logging_module, "log", lambda *_args, **_kwargs: None)
+    package_modules[logging_module.__name__] = logging_module
+
+    network_module = types.ModuleType(f"{package_name}.utils.network")
+    setattr(network_module, "build_worker_url", lambda *_args, **_kwargs: "http://example.invalid")
+
+    async def _unused_probe_worker(*_args, **_kwargs):
+        return None
+
+    async def _unused_get_client_session(*_args, **_kwargs):
+        raise RuntimeError("get_client_session must be patched by this test")
+
+    setattr(network_module, "probe_worker", _unused_probe_worker)
+    setattr(network_module, "get_client_session", _unused_get_client_session)
+    package_modules[network_module.__name__] = network_module
+
+    module_name = f"{package_name}.api.orchestration.dispatch"
+    spec = importlib.util.spec_from_file_location(module_name, module_path)
+    assert spec is not None and spec.loader is not None
+    module = importlib.util.module_from_spec(spec)
+    package_modules[module_name] = module
+
+    previous_modules = {name: sys.modules.get(name) for name in package_modules}
+    try:
+        sys.modules.update(package_modules)
+        spec.loader.exec_module(module)
+    finally:
+        for name, previous in previous_modules.items():
+            if previous is None:
+                sys.modules.pop(name, None)
+            else:
+                sys.modules[name] = previous
+    return module
+
+
+dispatch = _load_dispatch_module()
+
+
+class WebsocketDispatchDeadlineTests(unittest.IsolatedAsyncioTestCase):
+    async def asyncSetUp(self):
+        self.mode = None
+        self.received_payload = None
+        self.dispatch_received = asyncio.Event()
+        self.peer_finished = asyncio.Event()
+        self.irrelevant_messages_sent = 0
+
+        app = web.Application()
+        app.router.add_get("/distributed/worker_ws", self._handle_websocket)
+        self.runner = web.AppRunner(app)
+        await self.runner.setup()
+        self.site = web.TCPSite(self.runner, "127.0.0.1", 0)
+        await self.site.start()
+        server = cast(Any, self.site._server)
+        port = server.sockets[0].getsockname()[1]
+        self.worker_url = f"http://127.0.0.1:{port}"
+        self.session = aiohttp.ClientSession()
+
+    async def asyncTearDown(self):
+        await self.session.close()
+        await self.runner.cleanup()
+
+    async def _get_client_session(self):
+        return self.session
+
+    async def _handle_websocket(self, request):
+        ws = web.WebSocketResponse()
+        await ws.prepare(request)
+        try:
+            payload = await ws.receive_json()
+            self.received_payload = payload
+            self.dispatch_received.set()
+
+            if self.mode in ("success", "slow-close"):
+                await ws.send_json({
+                    "type": "dispatch_ack",
+                    "request_id": "unrelated-request",
+                    "ok": True,
+                    "prompt_id": "wrong-prompt",
+                })
+                await ws.send_json({"type": "heartbeat"})
+                await ws.send_json({
+                    "type": "dispatch_ack",
+                    "request_id": payload["request_id"],
+                    "ok": True,
+                    "prompt_id": "prompt-123",
+                })
+                if self.mode == "slow-close":
+                    await asyncio.sleep(0.15)
+            elif self.mode == "rejection":
+                await ws.send_json({
+                    "type": "dispatch_ack",
+                    "request_id": payload["request_id"],
+                    "ok": False,
+                    "error": "Invalid prompt",
+                    "validation_error": {"node": "bad"},
+                    "node_errors": {"bad": "missing input"},
+                })
+            elif self.mode in ("no-ack", "wait-for-cancel"):
+                async def send_irrelevant_messages():
+                    while True:
+                        try:
+                            await ws.send_json({"type": "heartbeat"})
+                            self.irrelevant_messages_sent += 1
+                            await ws.send_json({
+                                "type": "dispatch_ack",
+                                "request_id": "unrelated-request",
+                                "ok": True,
+                                "prompt_id": "not-this-request",
+                            })
+                            self.irrelevant_messages_sent += 1
+                        except (ConnectionError, RuntimeError):
+                            return
+                        await asyncio.sleep(0.01)
+
+                sender = asyncio.create_task(send_irrelevant_messages())
+                try:
+                    async for _message in ws:
+                        pass
+                finally:
+                    sender.cancel()
+                    try:
+                        await sender
+                    except asyncio.CancelledError:
+                        pass
+        finally:
+            self.peer_finished.set()
+        return ws
+
+    async def _dispatch(self, timeout):
+        with patch.object(dispatch, "get_client_session", self._get_client_session):
+            return await dispatch._dispatch_via_websocket(
+                self.worker_url,
+                {"prompt": {"node": {"class_type": "Example"}}, "workflow": {"name": "test"}},
+                "client-456",
+                timeout=timeout,
+            )
+
+    async def test_returns_matching_acknowledgement_prompt_id(self):
+        self.mode = "success"
+
+        prompt_id = await self._dispatch(timeout=0.5)
+
+        self.assertEqual(prompt_id, "prompt-123")
+        received_payload = self.received_payload
+        assert received_payload is not None
+        self.assertEqual(received_payload["type"], "dispatch_prompt")
+        self.assertTrue(received_payload["request_id"])
+        self.assertEqual(received_payload["client_id"], "client-456")
+        self.assertEqual(received_payload["prompt"], {"node": {"class_type": "Example"}})
+        self.assertEqual(received_payload["workflow"], {"name": "test"})
+
+    async def test_acknowledged_prompt_id_survives_slow_close(self):
+        self.mode = "slow-close"
+        self.assertEqual(await self._dispatch(timeout=0.03), "prompt-123")
+
+    async def test_matching_rejection_preserves_worker_error_details(self):
+        self.mode = "rejection"
+
+        with self.assertRaisesRegex(RuntimeError, "Invalid prompt.*validation_error=.*node_errors="):
+            await self._dispatch(timeout=0.5)
+
+    async def test_repeated_irrelevant_messages_do_not_reset_ack_deadline(self):
+        self.mode = "no-ack"
+        started = asyncio.get_running_loop().time()
+
+        with self.assertRaisesRegex(RuntimeError, "submission status is ambiguous"):
+            await asyncio.wait_for(self._dispatch(timeout=0.12), timeout=0.5)
+
+        elapsed = asyncio.get_running_loop().time() - started
+        self.assertLess(elapsed, 0.5)
+        self.assertGreaterEqual(self.irrelevant_messages_sent, 4)
+        await asyncio.wait_for(self.peer_finished.wait(), timeout=0.5)
+
+    async def test_outer_cancellation_propagates_and_closes_peer(self):
+        self.mode = "wait-for-cancel"
+        task = asyncio.create_task(self._dispatch(timeout=2.0))
+        await asyncio.wait_for(self.dispatch_received.wait(), timeout=0.5)
+
+        task.cancel()
+        with self.assertRaises(asyncio.CancelledError):
+            await asyncio.wait_for(task, timeout=0.5)
+        await asyncio.wait_for(self.peer_finished.wait(), timeout=0.5)
+
+
+if __name__ == "__main__":
+    unittest.main()

tests/test_media_bridge_nodes.py

@@ -0,0 +1,198 @@
+"""Real ComfyUI V3 bridge mapping, file ownership and exact CPU round trips."""
+import os
+from pathlib import Path
+import subprocess
+import sys
+
+import pytest
+
+ROOT = Path(__file__).resolve().parents[1]
+
+
+def test_real_media_bridge_nodes():
+    core = os.environ.get("COMFYUI_SOURCE_ROOT")
+    if core is None:
+        pytest.skip("Set COMFYUI_SOURCE_ROOT for CPU media bridge acceptance")
+    assert core and (Path(core) / "comfy_api/latest/_input_impl/video_types.py").is_file(), (
+        "Configured COMFYUI_SOURCE_ROOT must provide native video classes"
+    )
+    probe = r'''
+import asyncio, importlib.util, io, json, os, sys, tempfile
+from fractions import Fraction
+import av
+import numpy as np
+from pathlib import Path
+from unittest.mock import patch
+import torch
+sys.path.insert(0, sys.argv[1])
+import comfy.cli_args
+comfy.cli_args.args.cpu = True
+comfy.cli_args.args.disable_assets = True
+def forbidden(*args, **kwargs):
+    raise AssertionError('Bridge tests must not query or initialize a GPU')
+torch.cuda.is_available = lambda: False
+torch.cuda.device_count = lambda: 0
+for name in ('init', '_lazy_init', 'current_device'):
+    setattr(torch.cuda, name, forbidden)
+import execution
+import folder_paths
+from comfy_api.latest._input_impl.video_types import VideoFromComponents, VideoFromFile, VideoFromList
+from comfy_api.latest._util import VideoComponents
+for name in ('is_available', 'device_count'):
+    setattr(torch.cuda, name, forbidden)
+spec = importlib.util.spec_from_file_location('typed_bridge_probe', Path(sys.argv[2])/'serverless_bridge/__init__.py', submodule_search_locations=[str(Path(sys.argv[2])/'serverless_bridge')])
+bridge = importlib.util.module_from_spec(spec)
+sys.modules[spec.name] = bridge
+spec.loader.exec_module(bridge)
+from typed_bridge_probe.nodes import MediaImportV1, MediaExportV1, export_values, import_value
+from typed_bridge_probe.contract import SOCKET_KINDS
+from typed_bridge_probe.video_codec import encode_video
+
+def tiny_video():
+    buffer = io.BytesIO()
+    with av.open(buffer, 'w', format='matroska') as container:
+        stream = container.add_stream('ffv1', rate=Fraction(24, 1))
+        stream.width = stream.height = 32
+        stream.pix_fmt = 'yuv420p'
+        for level in (16, 64, 128):
+            frame = av.VideoFrame.from_ndarray(np.full((32, 32, 3), level, dtype=np.uint8), format='rgb24')
+            for packet in stream.encode(frame): container.mux(packet)
+        for packet in stream.encode(): container.mux(packet)
+    return buffer.getvalue()
+
+def exact_tensor(actual, expected):
+    assert actual.device.type == 'cpu' and actual.dtype == expected.dtype
+    assert actual.shape == expected.shape
+    assert torch.equal(actual.contiguous().view(torch.uint8), expected.contiguous().view(torch.uint8))
+
+def exact_video(actual, expected, wire):
+    assert type(actual) is type(expected)
+    assert encode_video(actual) == wire
+    if type(actual) is VideoFromComponents:
+        a = vars(actual)['_VideoFromComponents__components']
+        e = vars(expected)['_VideoFromComponents__components']
+        assert type(a) is VideoComponents and type(a.frame_rate) is Fraction
+        assert a.frame_rate == e.frame_rate and a.metadata == e.metadata
+        assert a.metadata is not e.metadata
+        for name in ('images', 'alpha'): exact_tensor(getattr(a, name), getattr(e, name))
+        exact_tensor(a.audio['waveform'], e.audio['waveform'])
+        assert a.audio['sample_rate'] == e.audio['sample_rate']
+        assert actual.get_bit_depth() == 10 and actual.get_color_space() == 'HDR PQ'
+    else:
+        values = actual.videos if type(actual) is VideoFromList else [actual]
+        for value in values:
+            buffer = vars(value)['_VideoFromFile__file']
+            assert type(buffer) is io.BytesIO and not buffer.closed
+            buffer.seek(0)
+            with av.open(buffer) as container:
+                assert len(list(container.decode(video=0))) == 3
+        if type(actual) is VideoFromList:
+            exact_tensor(actual.complete_audio['waveform'], expected.complete_audio['waveform'])
+
+async def check():
+    extension = await bridge.comfy_entrypoint()
+    assert await extension.get_node_list() == [MediaImportV1, MediaExportV1]
+    imp = MediaImportV1.GET_NODE_INFO_V1()
+    exp = MediaExportV1.GET_NODE_INFO_V1()
+    assert imp['output'] == ['IMAGE','AUDIO','MASK','STRING','VIDEO'], imp
+    assert exp['output_node'] and MediaExportV1.define_schema().is_input_list, exp
+    with tempfile.TemporaryDirectory(prefix='typed-media-', dir=os.environ.get('TMPDIR')) as scratch:
+        base = Path(scratch)
+        roots = {key: base/key for key in ('input','output','temp')}
+        for path in roots.values(): path.mkdir()
+        def annotated(name):
+            for key, root in roots.items():
+                suffix = f' [{key}]'
+                if name.endswith(suffix): return str(root/name[:-len(suffix)])
+            return str(roots['input']/name)
+        with patch.object(folder_paths,'get_input_directory',lambda: str(roots['input'])), patch.object(folder_paths,'get_output_directory',lambda: str(roots['output'])), patch.object(folder_paths,'get_temp_directory',lambda: str(roots['temp'])), patch.object(folder_paths,'get_annotated_filepath',annotated):
+            examples = {'images': torch.rand(2,2,3,4,dtype=torch.float64),
+                        'masks': torch.tensor([[[0.123456789,0.987654321]]],dtype=torch.float64),
+                        'audio': {'waveform':torch.randn(1,6,31,dtype=torch.float64),'sample_rate':44100},
+                        'texts': '  café\n🎥\r\n'}
+            media = tiny_video()
+            sources = [roots['input']/f'source-{n}.mkv' for n in range(3)]
+            for path in sources: path.write_bytes(media)
+            images = torch.arange(36, dtype=torch.float64, device='cpu').reshape(2,2,3,3) / 997
+            images[0,0,0,0] = -0.0
+            component_video = VideoFromComponents(VideoComponents(
+                images=images, frame_rate=Fraction(30000,1001),
+                alpha=torch.full((2,2,3,1),0.123456789,dtype=torch.float64,device='cpu'),
+                audio=examples['audio'], metadata={'title':'雪', 'nested':[True,None,-4,0.125]}),
+                bit_depth=10,color_space='HDR PQ')
+            videos = [component_video,
+                      VideoFromFile(str(sources[0]),start_time=-0.0,duration=0,crop=(0,0,2,2)),
+                      VideoFromList([VideoFromFile(str(sources[1]),start_time=-1.25,duration=-2.5,crop=(-3,5,7,-9)),
+                                     VideoFromFile(str(sources[2]),start_time=0,duration=0)],
+                                    complete_audio=examples['audio'])]
+            for socket,value in [*examples.items(), *[('video',v) for v in videos]]:
+                kind = SOCKET_KINDS[socket]
+                expected_wire = encode_video(value) if kind == 'VIDEO' else None
+                binding = dict(binding_id='b-'+socket,collector_id='collector',job_id='job',participant_id='worker',node_id='export-'+socket,socket=socket,kind=kind)
+                controls = {'request_id':'request','binding_json':json.dumps(binding)}
+                data, missing, v3 = execution.get_input_data(controls,MediaExportV1,'export-'+socket)
+                assert not missing, missing
+                data[socket] = [value]
+                results = await execution._async_map_node_over_list('probe','export-'+socket,MediaExportV1,data,MediaExportV1.FUNCTION,v3_data=v3)
+                results = await execution.resolve_map_node_over_list_results(results)
+                raw = results[0].ui
+                exported = raw['media_exports'][0]
+                assert exported['binding']['count'] == 1
+                artifact = exported['artifacts'][0]
+                transport = artifact.pop('transport')
+                filename = transport['subfolder']+'/'+transport['filename']+' [temp]'
+                decoded = import_value(artifact,filename)
+                if kind in ('IMAGE','MASK'):
+                    assert decoded.dtype == value.dtype and torch.equal(decoded,value)
+                elif kind == 'AUDIO':
+                    assert decoded['sample_rate'] == value['sample_rate']
+                    assert torch.equal(decoded['waveform'],value['waveform'])
+                elif kind == 'VIDEO': exact_video(decoded,value,expected_wire)
+                else: assert decoded == value
+                mapped, missing, hidden = execution.get_input_data({'artifact_json':json.dumps(artifact),'filename':filename},MediaImportV1,'import')
+                got = await execution._async_map_node_over_list('probe','import',MediaImportV1,mapped,MediaImportV1.FUNCTION,v3_data=hidden)
+                got = await execution.resolve_map_node_over_list_results(got)
+                assert len(got[0].result) == 5
+                assert all(x is None for n,x in enumerate(got[0].result) if n != list(SOCKET_KINDS).index(socket))
+                bad = dict(artifact,sha256='0'*64)
+                try: import_value(bad,filename)
+                except ValueError: pass
+                else: raise AssertionError('checksum accepted')
+                try: import_value(artifact,str(base/'outside.bin'))
+                except ValueError: pass
+                else: raise AssertionError('outside path accepted')
+                if kind == 'VIDEO':
+                    mapped_video = got[0].result[list(SOCKET_KINDS).index(socket)]
+                    # Remove both the exported artifact and the original input
+                    # containers before checking independently owned BytesIOs.
+                    Path(annotated(filename)).unlink()
+                    originals = value.videos if type(value) is VideoFromList else [value]
+                    for original in originals:
+                        if type(original) is VideoFromFile:
+                            Path(vars(original)['_VideoFromFile__file']).unlink()
+                    exact_video(decoded,value,expected_wire)
+                    exact_video(mapped_video,value,expected_wire)
+                    if type(value) is VideoFromFile:
+                        assert tuple(decoded.get_components().images.shape) == (3,2,2,3)
+            text_binding = dict(binding_id='empty',collector_id='collector',job_id='job',participant_id='worker',node_id='export-empty',socket='texts',kind='STRING')
+            out = export_values('request',text_binding,[''])
+            assert out['artifacts'][0]['byte_length'] == 0
+            before = set(roots['temp'].rglob('*.bin'))
+            try: export_values('request',text_binding,['first',None])
+            except ValueError: pass
+            else: raise AssertionError('invalid text accepted')
+            assert set(roots['temp'].rglob('*.bin')) == before
+    print('MEDIA_BRIDGE_CPU_OK')
+asyncio.run(check())
+'''
+    env = dict(os.environ)
+    env.pop("PYTHONPATH", None)
+    env.pop("COMFYUI_SOURCE_ROOT", None)
+    scratch = Path(os.environ.get("TMPDIR", str(Path.home() / ".hermes/cache/scratch")))
+    scratch.mkdir(parents=True, exist_ok=True)
+    env.update(CUDA_VISIBLE_DEVICES="", HIP_VISIBLE_DEVICES="", PYTHONDONTWRITEBYTECODE="1",
+               TMPDIR=str(scratch))
+    result = subprocess.run([sys.executable, "-c", probe, str(Path(core).resolve()), str(ROOT)],
+                            capture_output=True, text=True, timeout=90, env=env)
+    assert result.returncode == 0, result.stdout + result.stderr
+    assert "MEDIA_BRIDGE_CPU_OK" in result.stdout

tests/test_media_codec.py

@@ -0,0 +1,328 @@
+import importlib.util
+import math
+import struct
+import types
+from pathlib import Path
+
+import pytest
+import torch
+
+
+MODULE_PATH = Path(__file__).resolve().parents[1] / "serverless_bridge" / "codec.py"
+spec = importlib.util.spec_from_file_location("dist_media_codec_test", MODULE_PATH)
+codec = importlib.util.module_from_spec(spec)
+spec.loader.exec_module(codec)
+
+DTYPES = [torch.float16, torch.bfloat16, torch.float32, torch.float64]
+SHAPES = [
+    ("IMAGE", (2, 3, 5, 1), "BHWC"),
+    ("IMAGE", (2, 3, 5, 3), "BHWC"),
+    ("IMAGE", (2, 3, 5, 4), "BHWC"),
+    ("MASK", (2, 3, 5), "BHW"),
+    ("AUDIO", (2, 1, 17), "BCT"),
+    ("AUDIO", (2, 6, 17), "BCT"),
+]
+
+
+def audio(tensor, sample_rate=48000):
+    return {"waveform": tensor, "sample_rate": sample_rate}
+
+
+def descriptor(_kind="MASK", **changes):
+    result = {
+        "encoding": "tensor-le-v1", "dtype": "float32",
+        "shape": [1, 1, 1], "layout": "BHW",
+    }
+    if _kind == "IMAGE":
+        result.update(shape=[1, 1, 1, 3], layout="BHWC")
+    elif _kind == "AUDIO":
+        result.update(layout="BCT", sample_rate=48000)
+    result.update(changes)
+    return result
+
+
+@pytest.fixture(autouse=True)
+def forbid_gpu_queries(monkeypatch):
+    def forbidden(*args, **kwargs):
+        pytest.fail("codec must not query or initialize a GPU")
+    for name in ("is_available", "init", "_lazy_init", "device_count"):
+        monkeypatch.setattr(torch.cuda, name, forbidden)
+
+
+@pytest.mark.parametrize("dtype", DTYPES)
+@pytest.mark.parametrize("kind,shape,layout", SHAPES)
+def test_exact_roundtrip_preserves_dtype_shape_channels_and_bits(dtype, kind, shape, layout):
+    count = math.prod(shape)
+    tensor = (torch.arange(count, device="cpu", dtype=torch.float64) / 997 - 0.125).to(dtype).reshape(shape)
+    tensor.flatten()[0] = -0.0
+    tensor.requires_grad_(True)
+    value = audio(tensor) if kind == "AUDIO" else tensor
+
+    metadata, payload = codec.encode_payload(kind, value)
+    expected = {
+        "encoding": "tensor-le-v1", "dtype": str(dtype).removeprefix("torch."),
+        "shape": list(shape), "layout": layout,
+    }
+    if kind == "AUDIO":
+        expected["sample_rate"] = 48000
+    assert metadata == expected
+    assert type(payload) is bytes
+    assert len(payload) == count * tensor.element_size()
+    result = codec.decode_payload(kind, metadata, payload)
+    if kind == "AUDIO":
+        assert set(result) == {"waveform", "sample_rate"}
+        assert result["sample_rate"] == 48000
+        result = result["waveform"]
+    assert result.device.type == "cpu"
+    assert result.is_contiguous()
+    assert not result.requires_grad
+    assert result.dtype == dtype
+    assert result.shape == shape
+    assert torch.equal(result.view(torch.uint8), tensor.detach().view(torch.uint8))
+
+
+@pytest.mark.parametrize("dtype,fmt,expected_bits", [
+    (torch.float16, "e", None), (torch.bfloat16, "H", [0x3F80, 0x8000, 0xC020]),
+    (torch.float32, "f", None), (torch.float64, "d", None),
+])
+def test_wire_bytes_are_little_endian(dtype, fmt, expected_bits):
+    values = [1.0, -0.0, -2.5]
+    tensor = torch.tensor(values, dtype=dtype, device="cpu").reshape(1, 1, 3)
+    metadata, payload = codec.encode_payload("MASK", tensor)
+    expected = struct.pack("<3" + fmt, *(expected_bits if expected_bits else values))
+    assert payload == expected
+    assert torch.equal(codec.decode_payload("MASK", metadata, expected).view(torch.uint8), tensor.view(torch.uint8))
+
+
+def test_bfloat16_extremes_and_subnormal_bits_are_lossless():
+    bits = torch.tensor([0, 0x8000, 1, 0x8001, 0x3F81, 0x7F7F, 0xFF7F], dtype=torch.uint16, device="cpu")
+    value = bits.view(torch.bfloat16).reshape(1, 1, -1)
+    metadata, payload = codec.encode_payload("MASK", value)
+    assert payload == struct.pack("<7H", *bits.tolist())
+    assert torch.equal(codec.decode_payload("MASK", metadata, payload).view(torch.uint16).flatten(), bits)
+
+
+def test_fractional_mask_is_not_quantized_clipped_or_normalized():
+    value = torch.tensor([0.1234567890123, 1.0000000000001, -2.5], dtype=torch.float64, device="cpu").reshape(1, 1, 3)
+    metadata, payload = codec.encode_payload("MASK", value)
+    assert torch.equal(codec.decode_payload("MASK", metadata, payload), value)
+
+
+@pytest.mark.parametrize("text", ["", " ", "\t\n  ", "café 雪 🎵", "e\u0301", "\x00\ufeff\r\n"])
+def test_strings_preserve_exact_utf8(text):
+    metadata, payload = codec.encode_payload("STRING", text)
+    assert metadata == {"encoding": "utf8-v1"}
+    assert payload == text.encode("utf-8")
+    assert codec.decode_payload("STRING", metadata, payload) == text
+
+
+def test_noncontiguous_and_lazy_negative_tensors_roundtrip():
+    value = torch.arange(24, dtype=torch.float32, device="cpu").reshape(2, 3, 4).transpose(1, 2)
+    for tensor in (value, torch._neg_view(value)):
+        metadata, payload = codec.encode_payload("MASK", tensor)
+        result = codec.decode_payload("MASK", metadata, payload)
+        assert result.is_contiguous()
+        assert torch.equal(result, tensor)
+
+
+def test_decode_is_independent_and_ignores_default_device(monkeypatch):
+    source = torch.tensor([0.25], device="cpu").reshape(1, 1, 1)
+    metadata, payload = codec.encode_payload("MASK", source)
+    buffers = []
+    original = torch.frombuffer
+    def capture(buffer, *args, **kwargs):
+        buffers.append(buffer)
+        return original(buffer, *args, **kwargs)
+    monkeypatch.setattr(torch, "frombuffer", capture)
+    with torch.device("meta"):
+        first = codec.decode_payload("MASK", metadata, payload)
+        second = codec.decode_payload("MASK", metadata, payload)
+    assert first.device.type == second.device.type == "cpu"
+    first.fill_(9)
+    source.fill_(7)
+    for buffer in buffers:
+        if isinstance(buffer, bytearray):
+            buffer[:] = b"\x00" * len(buffer)
+    assert second.item() == 0.25
+    assert codec.decode_payload("MASK", metadata, payload).item() == 0.25
+
+
+@pytest.mark.parametrize("kind", ["image", " MASK", "LATENT", "", None, 1, [], {}])
+def test_unknown_kinds_raise_value_error(kind):
+    with pytest.raises(ValueError, match="kind"):
+        codec.encode_payload(kind, "")
+    with pytest.raises(ValueError, match="kind"):
+        codec.decode_payload(kind, {}, b"")
+
+
+@pytest.mark.parametrize("kind,value", [
+    ("STRING", None), ("STRING", b"text"), ("STRING", 123),
+    ("IMAGE", [[[[0.5]]]]), ("MASK", None),
+    ("AUDIO", torch.ones(1, 1, 1, device="cpu")), ("AUDIO", {}),
+    ("AUDIO", {"waveform": "invalid", "sample_rate": 48000}),
+    ("AUDIO", {"waveform": torch.ones(1, 1, 1, device="cpu")}),
+])
+def test_unsupported_values_raise_value_error(kind, value):
+    with pytest.raises(ValueError):
+        codec.encode_payload(kind, value)
+
+
+@pytest.mark.parametrize("dtype", [torch.uint8, torch.int64, torch.bool, torch.complex64])
+def test_unsupported_tensor_dtype(dtype):
+    with pytest.raises(ValueError, match="dtype"):
+        codec.encode_payload("MASK", torch.ones(1, 1, 1, dtype=dtype, device="cpu"))
+
+
+@pytest.mark.parametrize("kind,shape", [
+    ("IMAGE", (1, 2, 3)), ("IMAGE", (1, 1, 1, 2)), ("IMAGE", (1, 1, 1, 5)),
+    ("MASK", (1, 2)), ("MASK", (1, 2, 3, 1)), ("AUDIO", (2, 4)),
+    ("IMAGE", (1, 0, 2, 3)), ("MASK", (0, 1, 1)), ("AUDIO", (1, 0, 4)),
+    ("AUDIO", (1, 2, 0)),
+])
+def test_invalid_tensor_shape(kind, shape):
+    tensor = torch.empty(shape, dtype=torch.float32, device="cpu")
+    with pytest.raises(ValueError, match="shape|rank|channel|dimension"):
+        codec.encode_payload(kind, audio(tensor) if kind == "AUDIO" else tensor)
+
+
+def test_rejects_non_cpu_and_sparse_without_transferring(monkeypatch):
+    meta = torch.empty((1, 1, 1), device="meta")
+    sparse = torch.ones((1, 1, 1), device="cpu").to_sparse()
+    def forbidden(*args, **kwargs):
+        pytest.fail("codec must not move tensors between devices")
+    monkeypatch.setattr(torch.Tensor, "cpu", forbidden)
+    monkeypatch.setattr(torch.Tensor, "to", forbidden)
+    for tensor in (meta, sparse):
+        with pytest.raises(ValueError, match="CPU|layout|strided"):
+            codec.encode_payload("MASK", tensor)
+
+
+@pytest.mark.parametrize("sample_rate", [True, False, 0, -1, 44100.0, "44100", None])
+def test_invalid_sample_rate(sample_rate):
+    with pytest.raises(ValueError, match="sample_rate"):
+        codec.encode_payload("AUDIO", audio(torch.ones(1, 1, 1, device="cpu"), sample_rate))
+    with pytest.raises(ValueError, match="sample_rate"):
+        codec.decode_payload("AUDIO", descriptor("AUDIO", sample_rate=sample_rate), b"\x00" * 4)
+
+
+@pytest.mark.parametrize("bad", [None, [], "descriptor", {}, {"encoding": "utf8-v1"}])
+def test_malformed_tensor_descriptor(bad):
+    with pytest.raises(ValueError, match="descriptor"):
+        codec.decode_payload("MASK", bad, b"\x00" * 4)
+
+
+@pytest.mark.parametrize("changes", [
+    {"encoding": "pickle"}, {"encoding": None}, {"encoding": []},
+    {"dtype": "uint8"}, {"dtype": torch.float32}, {"dtype": []},
+    {"layout": "BHWC"}, {"layout": None}, {"layout": []},
+    {"shape": (1, 1, 1)}, {"shape": None}, {"shape": "1,1,1"},
+    {"shape": [1, 1]}, {"shape": [1, 1, 1, 1]}, {"shape": [1, 0, 1]},
+    {"shape": [1, -1, 1]}, {"shape": [1, True, 1]}, {"shape": [1, 1.0, 1]},
+    {"shape": [1, "1", 1]}, {"shape": [1, [], 1]}, {"kind": "MASK"},
+])
+def test_invalid_tensor_descriptor_fields(changes):
+    with pytest.raises(ValueError):
+        codec.decode_payload("MASK", descriptor(**changes), b"\x00" * 4)
+
+
+@pytest.mark.parametrize("metadata", [{}, {"encoding": "tensor-le-v1"}, {"encoding": "utf8-v1", "kind": "STRING"}, {"encoding": []}])
+def test_invalid_string_descriptor(metadata):
+    with pytest.raises(ValueError):
+        codec.decode_payload("STRING", metadata, b"")
+
+
+@pytest.mark.parametrize("payload", [b"", b"\x00" * 3, b"\x00" * 5, b"\x00" * 8])
+def test_shape_length_mismatch(payload):
+    with pytest.raises(ValueError, match="length"):
+        codec.decode_payload("MASK", descriptor(), payload)
+
+
+@pytest.mark.parametrize("payload", [None, "text", [0, 0, 0, 0], 4])
+def test_nonbyte_payload_rejected(payload):
+    with pytest.raises(ValueError, match="bytes"):
+        codec.decode_payload("MASK", descriptor(), payload)
+
+
+@pytest.mark.parametrize("dtype", DTYPES)
+@pytest.mark.parametrize("value", [float("nan"), float("inf"), -float("inf")])
+def test_nonfinite_values_fail_on_encode_and_decode(dtype, value):
+    tensor = torch.tensor([value], dtype=dtype, device="cpu").reshape(1, 1, 1)
+    with pytest.raises(ValueError, match="finite"):
+        codec.encode_payload("MASK", tensor)
+    fmt = {torch.float16: "e", torch.float32: "f", torch.float64: "d"}
+    if dtype == torch.bfloat16:
+        payload = struct.pack("<H", 0x7FC0 if math.isnan(value) else (0x7F80 if value > 0 else 0xFF80))
+    else:
+        payload = struct.pack("<" + fmt[dtype], value)
+    metadata = descriptor(dtype=str(dtype).removeprefix("torch."))
+    with pytest.raises(ValueError, match="finite"):
+        codec.decode_payload("MASK", metadata, payload)
+
+
+def test_invalid_utf8_errors_do_not_echo_content():
+    with pytest.raises(ValueError) as error:
+        codec.encode_payload("STRING", "private-value\ud800")
+    assert "private-value" not in str(error.value)
+    with pytest.raises(ValueError) as error:
+        codec.decode_payload("STRING", {"encoding": "utf8-v1"}, b"private-value\xff")
+    assert "private-value" not in str(error.value)
+
+
+@pytest.mark.parametrize("max_bytes", [True, False, -1, 4.0, "4", None])
+def test_invalid_byte_budget(max_bytes):
+    with pytest.raises(ValueError, match="max_bytes"):
+        codec.encode_payload("STRING", "", max_bytes=max_bytes)
+    with pytest.raises(ValueError, match="max_bytes"):
+        codec.decode_payload("STRING", {"encoding": "utf8-v1"}, b"", max_bytes=max_bytes)
+
+
+def test_exact_byte_budget_and_zero_budget_for_empty_string():
+    tensor = torch.ones(1, 1, 1, device="cpu")
+    metadata, payload = codec.encode_payload("MASK", tensor, max_bytes=4)
+    assert codec.decode_payload("MASK", metadata, payload, max_bytes=4).item() == 1
+    with pytest.raises(ValueError, match="limit|budget"):
+        codec.encode_payload("MASK", tensor, max_bytes=3)
+    with pytest.raises(ValueError, match="limit|budget"):
+        codec.decode_payload("MASK", metadata, payload, max_bytes=3)
+    metadata, payload = codec.encode_payload("STRING", "", max_bytes=0)
+    assert codec.decode_payload("STRING", metadata, payload, max_bytes=0) == ""
+    for text in ("a", "é", "🎵"):
+        byte_count = len(text.encode("utf-8"))
+        metadata, payload = codec.encode_payload("STRING", text, max_bytes=byte_count)
+        assert codec.decode_payload("STRING", metadata, payload, max_bytes=byte_count) == text
+        with pytest.raises(ValueError, match="limit|budget"):
+            codec.encode_payload("STRING", text, max_bytes=byte_count - 1)
+        with pytest.raises(ValueError, match="limit|budget"):
+            codec.decode_payload("STRING", metadata, payload, max_bytes=byte_count - 1)
+
+
+def test_oversized_shape_rejected_before_tensor_allocation(monkeypatch):
+    assert codec.MAX_PAYLOAD_BYTES == 256 * 1024 * 1024
+    def forbidden(*args, **kwargs):
+        pytest.fail("oversized shape must fail before tensor allocation")
+    for name in ("frombuffer", "empty", "tensor", "zeros", "isfinite"):
+        monkeypatch.setattr(torch, name, forbidden)
+    for shape in ([1, 1, codec.MAX_PAYLOAD_BYTES], [2**100, 2**100, 1]):
+        with pytest.raises(ValueError, match="limit|budget"):
+            codec.decode_payload("MASK", descriptor(shape=shape), b"\x00" * 4)
+
+
+def test_oversized_encode_rejected_before_materialization(monkeypatch):
+    huge_view = torch.ones(1, device="cpu").expand(1, 1, codec.MAX_PAYLOAD_BYTES)
+    def forbidden(*args, **kwargs):
+        pytest.fail("oversized tensor must fail before materialization or scanning")
+    monkeypatch.setattr(torch.Tensor, "contiguous", forbidden)
+    monkeypatch.setattr(torch, "isfinite", forbidden)
+    with pytest.raises(ValueError, match="limit|budget"):
+        codec.encode_payload("MASK", huge_view)
+
+
+@pytest.mark.parametrize("itemsize", [2, 4, 8])
+def test_native_little_endian_conversion_handles_big_endian(monkeypatch, itemsize):
+    raw = bytes(range(itemsize * 3))
+    monkeypatch.setattr(codec, "sys", types.SimpleNamespace(byteorder="little"))
+    assert codec._native_little_endian(raw, itemsize) == raw
+    monkeypatch.setattr(codec, "sys", types.SimpleNamespace(byteorder="big"))
+    expected = b"".join(raw[i:i + itemsize][::-1] for i in range(0, len(raw), itemsize))
+    assert codec._native_little_endian(raw, itemsize) == expected
+    assert codec._native_little_endian(expected, itemsize) == raw

tests/test_media_contract.py

@@ -0,0 +1,152 @@
+"""Offline contract tests: malformed results never reach transport/delivery."""
+import asyncio
+import copy
+import importlib.util
+from pathlib import Path
+
+import pytest
+
+ROOT = Path(__file__).resolve().parents[1]
+spec = importlib.util.spec_from_file_location("media_contract_test", ROOT / "serverless_bridge" / "contract.py")
+media = importlib.util.module_from_spec(spec)
+spec.loader.exec_module(media)
+
+
+def binding(**changes):
+    return dict(binding_id="binding-1", collector_id="100", job_id="job-1",
+                participant_id="serverless:worker:0", node_id="export-1", socket="texts", kind="STRING", **changes)
+
+
+def response():
+    artifact = media.artifact_from_payload("a", "STRING", {"encoding": "utf8-v1"}, b" \n")
+    return {"version": 1, "profile": media.PROFILE, "request_id": "request-1",
+            "bindings": [dict(binding(), count=1)], "artifacts": [artifact],
+            "items": [{"binding_id": "binding-1", "item_index": 0, "artifact_id": "a"}]}
+
+
+def validate(value):
+    return media.validate_response(value, request_id="request-1", bindings=[binding()])
+
+
+def test_valid_manifest_and_exact_inline_payload():
+    value = response()
+    ordered = validate(value)
+    assert asyncio.run(media.resolve_artifact(ordered["binding-1"][0])) == b" \n"
+
+
+def test_order_uses_indices_not_arrival():
+    value = response()
+    value["bindings"][0]["count"] = 2
+    value["artifacts"].append(media.artifact_from_payload("b", "STRING", {"encoding": "utf8-v1"}, b"second"))
+    value["items"] = [{"binding_id": "binding-1", "item_index": 1, "artifact_id": "b"}, value["items"][0]]
+    assert [item["id"] for item in validate(value)["binding-1"]] == ["a", "b"]
+
+
+@pytest.mark.parametrize("field,value", [("version", True), ("version", 2), ("profile", "legacy"), ("request_id", "another")])
+def test_envelope_identity(field, value):
+    obj = response()
+    obj[field] = value
+    with pytest.raises(ValueError):
+        validate(obj)
+
+
+@pytest.mark.parametrize("field", ["collector_id", "participant_id", "job_id", "node_id", "binding_id"])
+def test_binding_identity(field):
+    value = response()
+    value["bindings"][0][field] = "another"
+    with pytest.raises(ValueError):
+        validate(value)
+
+
+@pytest.mark.parametrize("count", [0, -1, True, 1.0, "1", 65])
+def test_invalid_count(count):
+    value = response()
+    value["bindings"][0]["count"] = count
+    with pytest.raises(ValueError):
+        validate(value)
+
+
+@pytest.mark.parametrize("mutation", ["missing_binding", "missing_item", "extra_item", "duplicate_binding", "duplicate_artifact", "unused_artifact", "wrong_type", "wrong_index"])
+def test_incomplete_or_ambiguous_manifest(mutation):
+    value = response()
+    if mutation == "missing_binding":
+        value["bindings"] = []
+    elif mutation == "missing_item":
+        value["items"] = []
+    elif mutation == "extra_item":
+        value["items"] *= 2
+    elif mutation == "duplicate_binding":
+        value["bindings"] *= 2
+    elif mutation == "duplicate_artifact":
+        value["artifacts"] *= 2
+    elif mutation == "unused_artifact":
+        artifact = copy.deepcopy(value["artifacts"][0])
+        artifact["id"] = "unused"
+        value["artifacts"].append(artifact)
+    elif mutation == "wrong_type":
+        value["artifacts"][0]["kind"] = "MASK"
+    elif mutation == "wrong_index":
+        value["items"][0]["item_index"] = True
+    with pytest.raises(ValueError):
+        validate(value)
+
+
+def test_duplicate_item_indices():
+    value = response()
+    value["bindings"][0]["count"] = 2
+    value["items"] *= 2
+    with pytest.raises(ValueError, match="Duplicate"):
+        validate(value)
+
+
+def test_total_budget():
+    with pytest.raises(ValueError, match="total byte"):
+        media.validate_response(response(), request_id="request-1", bindings=[binding()], max_total_bytes=1)
+
+
+def test_checksum_and_base64_validation():
+    artifact = response()["artifacts"][0]
+    artifact["sha256"] = "0" * 64
+    with pytest.raises(ValueError, match="checksum"):
+        asyncio.run(media.resolve_artifact(artifact))
+    artifact["transport"]["data_base64"] = "!!!!"
+    with pytest.raises(ValueError, match="base64"):
+        asyncio.run(media.resolve_artifact(artifact))
+
+
+def test_empty_text_is_a_value():
+    artifact = media.artifact_from_payload("empty", "STRING", {"encoding": "utf8-v1"}, b"")
+    assert asyncio.run(media.resolve_artifact(artifact)) == b""
+
+
+@pytest.mark.parametrize("url,expires", [("http://example.com/a", "2099-01-01T00:00:00Z"),
+    ("https://user:secret@example.com/a", "2099-01-01T00:00:00Z"),
+    ("https://example.com/a", "2000-01-01T00:00:00Z"),
+    ("https://example.com/a", "2099-01-01"),
+    ("https://example.com:1234/a", "2099-01-01T00:00:00Z")])
+def test_invalid_url_envelopes(url, expires):
+    with pytest.raises(ValueError):
+        media.artifact_from_payload("a", "STRING", {"encoding": "utf8-v1"}, b"a",
+            transport={"type": "url", "url": url, "expires_at": expires})
+
+
+def test_url_resolution_is_bounded_and_hash_verified():
+    artifact = media.artifact_from_payload("a", "STRING", {"encoding": "utf8-v1"}, b"hello",
+        transport={"type": "url", "url": "https://example.com/a", "expires_at": "2099-01-01T00:00:00Z"})
+    calls = []
+    async def download(url, *, max_bytes):
+        calls.append((url, max_bytes))
+        return b"hello"
+    assert asyncio.run(media.resolve_artifact(artifact, download=download)) == b"hello"
+    assert calls == [("https://example.com/a", 5)]
+
+
+def test_descriptor_depth_and_nonfinite_rejected():
+    value = {}
+    cursor = value
+    for _ in range(15):
+        cursor["next"] = {}
+        cursor = cursor["next"]
+    for descriptor in [value, {"x": float("nan")}, {"x": object()}]:
+        with pytest.raises(ValueError):
+            media.artifact_from_payload("a", "STRING", descriptor, b"")

tests/test_media_end_to_end.py

@@ -0,0 +1,48 @@
+"""Opt-in CPU acceptance of Collector IMAGE/AUDIO and the five-kind bridge.
+
+COMFYUI_SOURCE_ROOT must name an existing compatible checkout; run with that
+checkout's Python environment. No installation, normal instance, or provider is
+used. The child owns its loopback server and every writable Comfy directory.
+"""
+import os
+from pathlib import Path
+import subprocess
+import sys
+import tempfile
+
+import pytest
+
+ROOT = Path(__file__).resolve().parents[1]
+
+
+def test_typed_media_end_to_end_cpu():
+    configured = os.environ.get("COMFYUI_SOURCE_ROOT")
+    if configured is None:
+        pytest.skip("Set COMFYUI_SOURCE_ROOT for real CPU typed-media end-to-end acceptance")
+    core = Path(configured).expanduser().resolve()
+    assert configured and all((core / name).is_file() for name in (
+        "execution.py", "server.py", "comfy/cli_args.py",
+        "comfy_api/latest/_input_impl/video_types.py",
+    )), "Configured COMFYUI_SOURCE_ROOT must provide PromptExecutor and native V3 VIDEO"
+    scratch = Path(os.environ.get("TMPDIR", str(Path.home() / ".hermes/cache/scratch")))
+    scratch.mkdir(parents=True, exist_ok=True)
+    env = dict(os.environ)
+    for name in ("PYTHONPATH", "COMFYUI_SOURCE_ROOT"):
+        env.pop(name, None)
+    env.update(CUDA_VISIBLE_DEVICES="", HIP_VISIBLE_DEVICES="", ROCR_VISIBLE_DEVICES="",
+               PYTHONDONTWRITEBYTECODE="1", PYTHONHASHSEED="0", COMFYUI_IS_WORKER="1",
+               OMP_NUM_THREADS="1", MKL_NUM_THREADS="1")
+    with tempfile.TemporaryDirectory(prefix="media-e2e-", dir=scratch) as run_root:
+        env["TMPDIR"] = run_root
+        result = subprocess.run(
+            [sys.executable, str(ROOT / "tests/helpers/media_end_to_end_check.py"),
+             str(core), run_root], cwd=run_root, env=env,
+            capture_output=True, text=True, timeout=150,
+        )
+    assert result.returncode == 0, result.stdout + result.stderr
+    for marker in ("NATIVE_INPUTS_AND_DIRECT_EXPORT_REFS_OK", "REAL_HTTP_BRIDGE_RESULTS_OK",
+                   "MASTER_AND_DELEGATE_ORDER_OK", "ATOMIC_DECODE_REJECTION_OK",
+                   "IMAGE_TRANSFER_HTTP_OK",
+                   "MEDIA_END_TO_END_CPU_OK"):
+        assert marker in result.stdout, result.stdout + result.stderr
+    print(result.stdout)

tests/test_media_handler.py

@@ -0,0 +1,877 @@
+"""CPU-only endpoint adapter tests against an actual loopback Comfy HTTP fake."""
+import asyncio
+import base64
+import copy
+import hashlib
+import importlib
+import json
+import sys
+import types
+from contextlib import asynccontextmanager
+from pathlib import Path
+
+import aiohttp
+from aiohttp import web
+import pytest
+
+ROOT = Path(__file__).resolve().parents[1]
+PACKAGE = "media_handler_testpkg"
+package = types.ModuleType(PACKAGE)
+package.__path__ = [str(ROOT / "serverless_bridge")]
+sys.modules[PACKAGE] = package
+handler = importlib.import_module(PACKAGE + ".handler")
+contract = importlib.import_module(PACKAGE + ".contract")
+
+
+def request(kind="STRING", *, incoming=True):
+    socket = {value: key for key, value in contract.SOCKET_KINDS.items()}[kind]
+    binding = dict(binding_id="binding:1", collector_id="collector:1", job_id="job/42",
+                   participant_id="worker:0", node_id="export:1", socket=socket, kind=kind)
+    artifact = contract.artifact_from_payload("input:1", kind, {"encoding": "test"}, b"input bytes")
+    return {"workflow": {
+        "source:1": {"class_type": "Source", "inputs": {"value": "literal"}},
+        "export:1": {"class_type": "MediaExportV1", "inputs": {
+            "request_id": "request:1", "binding_json": json.dumps(binding), socket: ["source:1", 0]}},
+    }, "media": {"version": 1, "profile": contract.PROFILE, "request_id": "request:1",
+                  "artifacts": [artifact] if incoming else [],
+                  "inputs": [{"node_id": "source:1", "input_name": "value", "artifact_id": "input:1"}] if incoming else [],
+                  "outputs": [binding]}}
+
+
+class Comfy:
+    def __init__(self, root, data=None, *, payload=b"result bytes"):
+        self.root = root
+        self.data = data or request()
+        self.payload = payload
+        self.submitted = []
+        self.views = []
+        self.deletes = []
+        self.interrupts = []
+        self.polls = 0
+        self.active = False
+        self.mode = "success"
+        self.history_failures = 0
+        self.record_mutation = None
+        self.prompt_seen = asyncio.Event()
+        self.terminal_seen = False
+
+    def staged(self):
+        return list(self.root.glob("distributed-media-v1/*/*.bin"))
+
+    def record(self):
+        binding = self.data["media"]["outputs"][0]
+        aid = "a" * 32
+        artifact = contract.artifact_from_payload(aid, binding["kind"], {"encoding": "test"}, self.payload)
+        artifact["transport"] = {"type": "comfy-file", "filename": aid + ".bin",
+                                 "subfolder": "distributed-media-v1/" + hashlib.sha256(b"request:1").hexdigest(),
+                                 "folder_type": "temp"}
+        value = {"request_id": "request:1", "binding": dict(binding, count=1), "artifacts": [artifact],
+                 "items": [{"binding_id": binding["binding_id"], "item_index": 0, "artifact_id": aid}]}
+        if self.record_mutation:
+            self.record_mutation(value)
+        return value
+
+    async def submit(self, req):
+        data = await req.json()
+        self.submitted.append(data)
+        self.active = True
+        self.prompt_seen.set()
+        if self.data["media"]["artifacts"]:
+            assert self.staged()
+            link = data["prompt"]["source:1"]["inputs"]["value"]
+            imported = data["prompt"][link[0]]
+            assert imported["class_type"] == "MediaImportV1"
+            metadata = json.loads(imported["inputs"]["artifact_json"])
+            assert "transport" not in metadata
+            assert (self.root / imported["inputs"]["filename"]).read_bytes() == b"input bytes"
+        if self.mode == "ambiguous":
+            return web.Response(status=500)
+        if self.mode == "partial":
+            return web.json_response({"prompt_id": "prompt:1", "node_errors": {"export:1": {"errors": ["bad"]}}})
+        return web.json_response({"prompt_id": "prompt:1", "node_errors": {}})
+
+    async def history(self, req):
+        self.polls += 1
+        if self.history_failures:
+            self.history_failures -= 1
+            return web.Response(status=503)
+        if self.mode in {"queued", "running", "unrelated", "ambiguous"}:
+            return web.json_response({})
+        if self.data["media"]["artifacts"]:
+            assert self.staged(), "staging must survive until terminal confirmation"
+        self.active = False
+        self.terminal_seen = True
+        if self.mode in {"error", "interrupted"}:
+            status = {"completed": False, "status_str": "error", "messages": [["execution_interrupted", {}]]}
+        else:
+            status = {"completed": True, "status_str": "success", "messages": []}
+        return web.json_response({"prompt:1": {"status": status, "outputs": {} if self.mode == "missing" else {
+            "export:1": {"media_exports": [self.record()]}}}})
+
+    async def view(self, req):
+        self.views.append(dict(req.query))
+        assert self.terminal_seen
+        payload = self.payload if self.mode != "corrupt" else b"X" * len(self.payload)
+        if self.mode == "oversized":
+            payload += b"X"
+        return web.Response(body=payload)
+
+    def entry(self, pid="prompt:1"):
+        extra = dict(self.submitted[0].get("extra_data", {}), client_id=self.submitted[0]["client_id"])
+        return [1, pid, self.submitted[0]["prompt"], extra, ["export:1"]]
+
+    async def queue(self, req):
+        if req.method == "POST":
+            body = await req.json()
+            assert body == {"delete": ["prompt:1"]}
+            self.deletes.append(body)
+            assert self.staged()
+            self.mode = "deleted"
+            self.active = False
+            return web.Response(status=200)
+        running = [self.entry()] if self.mode == "running" else []
+        if self.mode == "unrelated":
+            running = [self.entry("not-our-prompt")]
+        pending = [self.entry()] if self.mode in {"queued", "ambiguous"} else []
+        return web.json_response({"queue_running": running, "queue_pending": pending})
+
+    async def interrupt(self, req):
+        assert self.mode == "running"
+        assert self.staged()
+        self.interrupts.append(await req.json())
+        self.mode = "interrupted"
+        return web.Response(status=200)
+
+
+@asynccontextmanager
+async def serve(comfy):
+    app = web.Application()
+    app.add_routes([web.post("/prompt", comfy.submit), web.get("/history/{prompt_id}", comfy.history),
+                    web.get("/view", comfy.view), web.get("/queue", comfy.queue),
+                    web.post("/queue", comfy.queue), web.post("/interrupt", comfy.interrupt)])
+    runner = web.AppRunner(app)
+    await runner.setup()
+    site = web.TCPSite(runner, "127.0.0.1", 0)
+    await site.start()
+    port = site._server.sockets[0].getsockname()[1]
+    try:
+        yield f"http://127.0.0.1:{port}"
+    finally:
+        await runner.cleanup()
+
+
+def run(coro):
+    return asyncio.run(coro)
+
+
+@pytest.mark.parametrize("kind,index", [("IMAGE", 0), ("AUDIO", 1), ("MASK", 2), ("STRING", 3), ("VIDEO", 4)])
+def test_loopback_submit_poll_view_cleanup_and_caller_ownership(tmp_path, kind, index):
+    async def check():
+        data = request(kind)
+        original = copy.deepcopy(data)
+        comfy = Comfy(tmp_path, data)
+        async with serve(comfy) as url, aiohttp.ClientSession() as session:
+            result = await handler.execute_media_request(data, comfy_url=url, input_dir=tmp_path, session=session)
+            assert not session.closed
+        contract.validate_response(result["media"], request_id="request:1", bindings=data["media"]["outputs"])
+        assert await contract.resolve_artifact(result["media"]["artifacts"][0]) == comfy.payload
+        assert len(comfy.submitted) == 1
+        assert comfy.submitted[0]["partial_execution_targets"] == ["export:1"]
+        assert comfy.submitted[0]["prompt"]["source:1"]["inputs"]["value"][1] == index
+        assert comfy.views == [{"filename": "a" * 32 + ".bin", "subfolder": "distributed-media-v1/" + hashlib.sha256(b"request:1").hexdigest(), "type": "temp"}]
+        assert data == original
+        assert not comfy.staged()
+    run(check())
+
+
+@pytest.mark.parametrize("mode,match", [("corrupt", "checksum"), ("missing", "Missing"), ("error", "execution"), ("oversized", "length")])
+def test_terminal_failures_clean_only_after_terminal(tmp_path, mode, match):
+    async def check():
+        comfy = Comfy(tmp_path)
+        comfy.mode = mode
+        async with serve(comfy) as url:
+            with pytest.raises((ValueError, RuntimeError), match=match):
+                await handler.execute_media_request(comfy.data, comfy_url=url, input_dir=tmp_path)
+        assert len(comfy.submitted) == 1
+        assert not comfy.staged()
+    run(check())
+
+
+@pytest.mark.parametrize("mutation", [
+    lambda d: d["media"].update(version=True),
+    lambda d: d["media"].update(profile="legacy"),
+    lambda d: d["media"].update(request_id="../escape"),
+    lambda d: d["media"]["outputs"][0].update(node_id="missing"),
+    lambda d: d["workflow"]["export:1"].update(class_type="SaveImage"),
+    lambda d: d["workflow"]["export:1"]["inputs"].update(request_id="other"),
+    lambda d: d["workflow"]["export:1"]["inputs"].update(binding_json="{}"),
+    lambda d: d["workflow"]["export:1"]["inputs"].update(images=["source:1", 0]),
+    lambda d: d["workflow"]["export:1"]["inputs"].update(texts="literal"),
+    lambda d: d["workflow"].update(unrelated={"class_type": "SaveImage", "inputs": {}}),
+    lambda d: d["workflow"].update(unsafe={"class_type": "MediaImportV1", "inputs": {}}),
+    lambda d: d["media"]["inputs"].append(copy.deepcopy(d["media"]["inputs"][0])),
+    lambda d: d["media"]["inputs"][0].update(node_id="missing"),
+    lambda d: d["media"]["inputs"][0].update(input_name="missing"),
+    lambda d: d["media"].update(inputs=[]),
+    lambda d: d["media"]["artifacts"][0].update(sha256="0" * 64),
+])
+def test_rejection_before_submit(tmp_path, mutation):
+    async def check():
+        data = request()
+        mutation(data)
+        comfy = Comfy(tmp_path, data)
+        async with serve(comfy) as url:
+            with pytest.raises(ValueError):
+                await handler.execute_media_request(data, comfy_url=url, input_dir=tmp_path)
+        assert not comfy.submitted
+        assert not comfy.staged()
+    run(check())
+
+
+@pytest.mark.parametrize("mutation", [
+    lambda r: r.update(request_id="other"),
+    lambda r: r["binding"].update(count=2),
+    lambda r: r["binding"].update(collector_id="other"),
+    lambda r: r["items"][0].update(item_index=1),
+    lambda r: r["artifacts"][0]["transport"].update(filename="../escape"),
+    lambda r: r["artifacts"][0]["transport"].update(subfolder="distributed-media-v1/other"),
+    lambda r: r["artifacts"][0]["transport"].update(folder_type="input"),
+    lambda r: r["artifacts"][0].update(byte_length=contract.MAX_ARTIFACT_BYTES + 1),
+])
+def test_bad_result_manifest_rejected_before_any_file_read(tmp_path, mutation):
+    async def check():
+        comfy = Comfy(tmp_path)
+        comfy.record_mutation = mutation
+        async with serve(comfy) as url:
+            with pytest.raises(ValueError):
+                await handler.execute_media_request(comfy.data, comfy_url=url, input_dir=tmp_path)
+        assert not comfy.views
+        assert not comfy.staged()
+    run(check())
+
+
+def test_idempotent_history_retry_without_resubmission(tmp_path):
+    async def check():
+        comfy = Comfy(tmp_path)
+        comfy.history_failures = 2
+        async with serve(comfy) as url:
+            await handler.execute_media_request(comfy.data, comfy_url=url, input_dir=tmp_path)
+        assert comfy.polls == 3
+        assert len(comfy.submitted) == 1
+    run(check())
+
+
+@pytest.mark.parametrize("mode", ["queued"])
+def test_deadline_cancels_only_own_prompt_before_cleanup(tmp_path, monkeypatch, mode):
+    monkeypatch.setattr(handler, "POLL_INTERVAL_SECONDS", 0.005)
+    async def check():
+        comfy = Comfy(tmp_path)
+        comfy.mode = mode
+        async with serve(comfy) as url:
+            with pytest.raises(TimeoutError):
+                await handler.execute_media_request(comfy.data, comfy_url=url, input_dir=tmp_path, timeout_seconds=0.04)
+        assert len(comfy.submitted) == 1
+        assert bool(comfy.deletes) == (mode == "queued")
+        assert bool(comfy.interrupts) == (mode == "running")
+        assert not comfy.staged()
+    run(check())
+
+
+def test_unconfirmed_terminal_retains_files_and_never_interrupts_unrelated(tmp_path, monkeypatch):
+    monkeypatch.setattr(handler, "POLL_INTERVAL_SECONDS", 0.005)
+    monkeypatch.setattr(handler, "CANCEL_GRACE_SECONDS", 0.06)
+    async def check():
+        comfy = Comfy(tmp_path)
+        comfy.mode = "unrelated"
+        async with serve(comfy) as url:
+            with pytest.raises(RuntimeError, match="retained"):
+                await handler.execute_media_request(comfy.data, comfy_url=url, input_dir=tmp_path, timeout_seconds=0.04)
+        assert comfy.staged()
+        assert not comfy.interrupts
+        assert not comfy.deletes
+        assert len(comfy.submitted) == 1
+    run(check())
+
+
+def test_ambiguous_post_is_not_retried_and_own_queue_entry_is_reconciled(tmp_path):
+    async def check():
+        comfy = Comfy(tmp_path)
+        comfy.mode = "ambiguous"
+        async with serve(comfy) as url:
+            with pytest.raises(RuntimeError):
+                await handler.execute_media_request(comfy.data, comfy_url=url, input_dir=tmp_path)
+        assert len(comfy.submitted) == 1
+        assert len(comfy.deletes) == 1
+        assert not comfy.staged()
+    run(check())
+
+
+def test_task_cancellation_fails_closed_and_preserves_caller_session(tmp_path):
+    async def check():
+        comfy = Comfy(tmp_path)
+        comfy.mode = "running"
+        async with serve(comfy) as url, aiohttp.ClientSession() as session:
+            task = asyncio.create_task(handler.execute_media_request(comfy.data, comfy_url=url, input_dir=tmp_path, session=session))
+            await comfy.prompt_seen.wait()
+            await asyncio.sleep(0.01)
+            task.cancel()
+            with pytest.raises(handler.MediaExecutionUncertain, match="prompt-scoped interruption unavailable"):
+                await task
+            assert not session.closed
+        assert not comfy.interrupts
+        assert comfy.staged()
+    run(check())
+
+
+def test_legacy_images_are_owned_and_rewritten_without_overwriting_user_files(tmp_path):
+    async def check():
+        data = request(incoming=False)
+        data["workflow"]["source:1"] = {"class_type": "LoadImage", "inputs": {"image": "same.png"}}
+        data["images"] = [{"name": "same.png", "image": base64.b64encode(b"new image").decode()}]
+        (tmp_path / "same.png").write_bytes(b"user image")
+        comfy = Comfy(tmp_path, data)
+        async with serve(comfy) as url:
+            await handler.execute_media_request(data, comfy_url=url, input_dir=tmp_path)
+        filename = comfy.submitted[0]["prompt"]["source:1"]["inputs"]["image"]
+        assert filename.startswith("distributed-media-v1/")
+        assert filename.endswith(".bin")
+        assert (tmp_path / "same.png").read_bytes() == b"user image"
+        assert not comfy.staged()
+    run(check())
+
+
+def test_url_input_uses_bounded_injected_downloader(tmp_path):
+    async def check():
+        data = request()
+        data["media"]["artifacts"][0]["transport"] = {"type": "url", "url": "https://example.org/a", "expires_at": "2099-01-01T00:00:00Z"}
+        calls = []
+        async def download(url, *, max_bytes):
+            calls.append((url, max_bytes))
+            return b"input bytes"
+        comfy = Comfy(tmp_path, data)
+        async with serve(comfy) as url:
+            await handler.execute_media_request(data, comfy_url=url, input_dir=tmp_path, download=download)
+        assert calls == [("https://example.org/a", len(b"input bytes"))]
+    run(check())
+
+
+def test_large_result_requires_publisher_and_validates_returned_transport(tmp_path, monkeypatch):
+    monkeypatch.setattr(handler, "MAX_RESPONSE_JSON_BYTES", 100_000)
+    async def check():
+        comfy = Comfy(tmp_path, payload=b"Z" * 100_000)
+        async with serve(comfy) as url:
+            with pytest.raises(ValueError, match="publisher"):
+                await handler.execute_media_request(comfy.data, comfy_url=url, input_dir=tmp_path)
+        assert not comfy.views
+        assert not comfy.staged()
+        published = []
+        async def publish(metadata, payload):
+            assert "transport" not in metadata
+            assert payload == comfy.payload
+            published.append(metadata)
+            return {"type": "url", "url": "https://example.org/result", "expires_at": "2099-01-01T00:00:00Z"}
+        async with serve(comfy) as url:
+            result = await handler.execute_media_request(comfy.data, comfy_url=url, input_dir=tmp_path, publish=publish)
+        assert len(published) == 1
+        assert result["media"]["artifacts"][0]["transport"]["type"] == "url"
+        assert len(json.dumps(result).encode()) <= 100_000
+        async def bad_publish(metadata, payload):
+            return {"type": "url", "url": "http://example.org/result", "expires_at": "2099-01-01T00:00:00Z"}
+        async with serve(comfy) as url:
+            with pytest.raises(ValueError):
+                await handler.execute_media_request(comfy.data, comfy_url=url, input_dir=tmp_path, publish=bad_publish)
+    run(check())
+
+
+@pytest.mark.parametrize("name", ["../photo.png", "folder/photo.png", "folder\\photo.png", ".", "..", "/photo.png"])
+def test_legacy_basename_rejected_without_submit(tmp_path, name):
+    async def check():
+        data = request(incoming=False)
+        data["images"] = [{"name": name, "image": "YQ=="}]
+        comfy = Comfy(tmp_path, data)
+        async with serve(comfy) as url:
+            with pytest.raises(ValueError):
+                await handler.execute_media_request(data, comfy_url=url, input_dir=tmp_path)
+        assert not comfy.submitted
+    run(check())
+
+
+def test_staging_rejects_symlink_without_writing_or_submit(tmp_path):
+    async def check():
+        outside = tmp_path / "unrelated"
+        outside.mkdir()
+        (tmp_path / "distributed-media-v1").symlink_to(outside, target_is_directory=True)
+        comfy = Comfy(tmp_path)
+        async with serve(comfy) as url:
+            with pytest.raises(ValueError, match="symlink"):
+                await handler.execute_media_request(comfy.data, comfy_url=url, input_dir=tmp_path)
+        assert not comfy.submitted
+        assert not list(outside.iterdir())
+    run(check())
+
+
+def test_cleanup_preserves_other_runs_in_same_request_namespace(tmp_path):
+    async def check():
+        directory = tmp_path / "distributed-media-v1" / hashlib.sha256(b"request:1").hexdigest()
+        directory.mkdir(parents=True)
+        sibling = directory / ("b" * 32 + ".bin")
+        sibling.write_bytes(b"other run")
+        comfy = Comfy(tmp_path)
+        async with serve(comfy) as url:
+            await handler.execute_media_request(comfy.data, comfy_url=url, input_dir=tmp_path)
+        assert sibling.read_bytes() == b"other run"
+        assert comfy.staged() == [sibling]
+    run(check())
+
+
+def test_input_total_limit_checked_before_downloader_or_submit(tmp_path):
+    async def check():
+        data = request()
+        data["media"]["artifacts"] = []
+        data["media"]["inputs"] = []
+        for index in range(3):
+            artifact = contract.artifact_from_payload(f"input:{index}", "STRING", {}, b"",
+                transport={"type": "url", "url": "https://example.org/input", "expires_at": "2099-01-01T00:00:00Z"})
+            artifact["byte_length"] = contract.MAX_ARTIFACT_BYTES
+            data["media"]["artifacts"].append(artifact)
+            name = f"value:{index}"
+            data["workflow"]["source:1"]["inputs"][name] = "literal"
+            data["media"]["inputs"].append({"node_id": "source:1", "input_name": name, "artifact_id": artifact["id"]})
+        comfy = Comfy(tmp_path, data)
+        async def forbidden_download(*args, **kwargs):
+            pytest.fail("oversized request reached downloader")
+        async with serve(comfy) as url:
+            with pytest.raises(ValueError, match="total byte"):
+                await handler.execute_media_request(data, comfy_url=url, input_dir=tmp_path, download=forbidden_download)
+        assert not comfy.submitted
+    run(check())
+
+
+def test_output_total_limit_checked_before_fetch(tmp_path):
+    def excessive(record):
+        first = record["artifacts"][0]
+        record["artifacts"] = []
+        record["items"] = []
+        record["binding"]["count"] = 3
+        for index, char in enumerate("abc"):
+            artifact = copy.deepcopy(first)
+            artifact["id"] = char * 32
+            artifact["byte_length"] = contract.MAX_ARTIFACT_BYTES
+            artifact["transport"]["filename"] = artifact["id"] + ".bin"
+            record["artifacts"].append(artifact)
+            record["items"].append({"binding_id": "binding:1", "item_index": index, "artifact_id": artifact["id"]})
+    async def check():
+        comfy = Comfy(tmp_path)
+        comfy.record_mutation = excessive
+        async with serve(comfy) as url:
+            with pytest.raises(ValueError, match="total byte"):
+                await handler.execute_media_request(comfy.data, comfy_url=url, input_dir=tmp_path)
+        assert not comfy.views
+        assert not comfy.staged()
+    run(check())
+
+
+def test_partial_target_acceptance_is_not_reported_as_success(tmp_path):
+    async def check():
+        comfy = Comfy(tmp_path)
+        comfy.mode = "partial"
+        async with serve(comfy) as url:
+            with pytest.raises(RuntimeError, match="incomplete"):
+                await handler.execute_media_request(comfy.data, comfy_url=url, input_dir=tmp_path)
+        assert len(comfy.submitted) == 1
+        assert not comfy.views
+        assert not comfy.staged()
+    run(check())
+
+
+# Native routes are compiled from the pinned source, without importing ComfyUI
+# (which would initialize the GPU). All dependencies and HTTP are CPU-only.
+@pytest.mark.parametrize("current,requested,race,expected", [
+    ("own", "own", False, ["own"]),
+    ("other", "own", False, []),
+    ("own", "own", True, ["successor"]),
+])
+def test_pinned_native_interrupt_route_has_snapshot_race(current, requested, race, expected):
+    import ast
+    import logging
+    source = Path("/home/robert/ComfyUI_dev/server.py")
+    if not source.is_file():
+        pytest.skip("Pinned ComfyUI source unavailable")
+    tree = ast.parse(source.read_text())
+    routes = [node for node in ast.walk(tree) if isinstance(node, ast.AsyncFunctionDef)
+              and any("/interrupt" in ast.unparse(d) for d in node.decorator_list)]
+    assert len(routes) == 1
+    route = routes[0]
+    route.decorator_list = []
+    interrupts = []
+    class Queue:
+        running = current
+        def get_current_queue(self):
+            snapshot = [[1, self.running, {}, {}, []]]
+            if race:
+                self.running = "successor"
+            return snapshot, []
+    queue = Queue()
+    scope = {"json": json, "logging": logging, "web": web,
+             "self": types.SimpleNamespace(prompt_queue=queue),
+             "nodes": types.SimpleNamespace(interrupt_processing=lambda: interrupts.append(queue.running))}
+    exec(compile(ast.Module(body=[route], type_ignores=[]), str(source), "exec"), scope)
+    async def check():
+        async def body():
+            return {"prompt_id": requested}
+        response = await scope[route.name](types.SimpleNamespace(json=body))
+        assert response.status == 200
+    run(check())
+    assert interrupts == expected
+
+
+def exporter_file(root, *, name="a" * 32 + ".bin", payload=b"result bytes"):
+    directory = root / "distributed-media-v1" / hashlib.sha256(b"request:1").hexdigest()
+    directory.mkdir(parents=True, exist_ok=True)
+    path = directory / name
+    path.write_bytes(payload)
+    return path
+
+
+@pytest.mark.parametrize("mode", ["success", "corrupt", "oversized", "error", "interrupted", "partial"])
+def test_owned_export_cleanup_after_terminal_success_or_failure(tmp_path, mode):
+    async def check():
+        temp = tmp_path / "temp"
+        owned = exporter_file(temp)
+        sibling = exporter_file(temp, name="b" * 32 + ".bin", payload=b"unrelated")
+        ordinary = exporter_file(temp, name="keep.txt", payload=b"ordinary")
+        comfy = Comfy(tmp_path / "input")
+        comfy.mode = mode
+        view = comfy.view
+        async def reading(req):
+            assert owned.exists(), "cleanup raced artifact collection"
+            return await view(req)
+        comfy.view = reading
+        async with serve(comfy) as url:
+            if mode == "success":
+                await handler.execute_media_request(comfy.data, comfy_url=url, input_dir=comfy.root, temp_dir=temp)
+            else:
+                with pytest.raises((ValueError, RuntimeError)):
+                    await handler.execute_media_request(comfy.data, comfy_url=url, input_dir=comfy.root, temp_dir=temp)
+        assert comfy.terminal_seen
+        assert not owned.exists()
+        assert sibling.read_bytes() == b"unrelated"
+        assert ordinary.read_bytes() == b"ordinary"
+        assert not comfy.staged()
+    run(check())
+
+
+def test_omitted_temp_root_leaves_exporter_lifecycle_ownership(tmp_path):
+    async def check():
+        owned = exporter_file(tmp_path / "temp")
+        comfy = Comfy(tmp_path / "input")
+        async with serve(comfy) as url:
+            await handler.execute_media_request(comfy.data, comfy_url=url, input_dir=comfy.root)
+        assert owned.read_bytes() == comfy.payload
+    run(check())
+
+
+@pytest.mark.parametrize("mode", ["running", "unrelated"])
+def test_uncertain_exports_and_inputs_retained_without_interrupt(tmp_path, monkeypatch, mode):
+    monkeypatch.setattr(handler, "POLL_INTERVAL_SECONDS", 0.005)
+    monkeypatch.setattr(handler, "CANCEL_GRACE_SECONDS", 0.04)
+    async def check():
+        temp = tmp_path / "temp"
+        owned = exporter_file(temp)
+        comfy = Comfy(tmp_path / "input")
+        comfy.mode = mode
+        async with serve(comfy) as url:
+            with pytest.raises(handler.MediaExecutionUncertain):
+                await handler.execute_media_request(comfy.data, comfy_url=url, input_dir=comfy.root,
+                                                    temp_dir=temp, timeout_seconds=0.03)
+        assert owned.exists()
+        assert comfy.staged()
+        assert not comfy.interrupts
+        assert not comfy.deletes
+    run(check())
+
+
+@pytest.mark.parametrize("component", ["root", "namespace", "request", "file"])
+def test_export_cleanup_never_traverses_symlinks(tmp_path, component):
+    async def check():
+        temp = tmp_path / "temp"
+        owned = exporter_file(temp)
+        selected = {"root": temp, "namespace": owned.parent.parent,
+                    "request": owned.parent, "file": owned}[component]
+        outside = tmp_path / "outside"
+        selected.rename(outside)
+        selected.symlink_to(outside, target_is_directory=component != "file")
+        comfy = Comfy(tmp_path / "input")
+        async with serve(comfy) as url:
+            await handler.execute_media_request(comfy.data, comfy_url=url, input_dir=comfy.root, temp_dir=temp)
+        assert selected.is_symlink()
+        assert owned.read_bytes() == comfy.payload
+    run(check())
+
+
+@pytest.mark.parametrize("mutation", [
+    lambda r: r.update(request_id="other"),
+    lambda r: r["binding"].update(collector_id="other"),
+    lambda r: r["artifacts"][0]["transport"].update(filename="../escape"),
+    lambda r: r["artifacts"][0]["transport"].update(subfolder="distributed-media-v1/other"),
+])
+def test_unidentified_exports_are_not_deleted(tmp_path, mutation):
+    async def check():
+        temp = tmp_path / "temp"
+        owned = exporter_file(temp)
+        comfy = Comfy(tmp_path / "input")
+        comfy.record_mutation = mutation
+        async with serve(comfy) as url:
+            with pytest.raises(ValueError):
+                await handler.execute_media_request(comfy.data, comfy_url=url, input_dir=comfy.root, temp_dir=temp)
+        assert owned.exists()
+        assert not comfy.views
+    run(check())
+
+
+@pytest.mark.parametrize("failure", ["required", "publish"])
+def test_publication_failures_cleanup_identified_exports(tmp_path, monkeypatch, failure):
+    monkeypatch.setattr(handler, "MAX_RESPONSE_JSON_BYTES", 100_000)
+    async def check():
+        temp = tmp_path / "temp"
+        comfy = Comfy(tmp_path / "input", payload=b"Z" * 100_000)
+        owned = exporter_file(temp, payload=comfy.payload)
+        async def publish(metadata, payload):
+            assert owned.exists()
+            raise ValueError("publisher failed")
+        async with serve(comfy) as url:
+            with pytest.raises(ValueError, match="publisher"):
+                await handler.execute_media_request(comfy.data, comfy_url=url, input_dir=comfy.root,
+                                                    temp_dir=temp, publish=publish if failure == "publish" else None)
+        assert not owned.exists()
+    run(check())
+
+
+def test_repeated_cancel_drains_collection_before_cleanup(tmp_path, monkeypatch):
+    monkeypatch.setattr(handler, "MAX_RESPONSE_JSON_BYTES", 100_000)
+    async def check():
+        temp = tmp_path / "temp"
+        comfy = Comfy(tmp_path / "input", payload=b"Z" * 100_000)
+        owned = exporter_file(temp, payload=comfy.payload)
+        started, draining, release = asyncio.Event(), asyncio.Event(), asyncio.Event()
+        async def publish(metadata, payload):
+            started.set()
+            try:
+                await asyncio.Future()
+            finally:
+                draining.set()
+                await release.wait()
+                assert owned.exists(), "cleanup while publisher still reading"
+        async with serve(comfy) as url:
+            task = asyncio.create_task(handler.execute_media_request(comfy.data, comfy_url=url,
+                                       input_dir=comfy.root, temp_dir=temp, publish=publish))
+            await asyncio.wait_for(started.wait(), 1)
+            task.cancel()
+            await asyncio.wait_for(draining.wait(), 1)
+            task.cancel()
+            await asyncio.sleep(0)
+            assert owned.exists()
+            release.set()
+            with pytest.raises(asyncio.CancelledError):
+                await task
+        assert not owned.exists()
+    run(check())
+
+
+@pytest.mark.parametrize("cancel_caller", [False, True])
+def test_stubborn_publisher_teardown_is_bounded_and_retains_files(tmp_path, monkeypatch, cancel_caller):
+    monkeypatch.setattr(handler, "MAX_RESPONSE_JSON_BYTES", 100_000)
+    monkeypatch.setattr(handler, "DRAIN_GRACE_SECONDS", 0.03)
+    async def check():
+        temp = tmp_path / "temp"
+        comfy = Comfy(tmp_path / "input", payload=b"Z" * 100_000)
+        owned = exporter_file(temp, payload=comfy.payload)
+        started, draining, release = asyncio.Event(), asyncio.Event(), asyncio.Event()
+        async def publish(metadata, payload):
+            started.set()
+            try:
+                await asyncio.Future()
+            finally:
+                draining.set()
+                await release.wait()
+                assert owned.exists()
+        async with serve(comfy) as url:
+            task = asyncio.create_task(handler.execute_media_request(comfy.data, comfy_url=url,
+                        input_dir=comfy.root, temp_dir=temp, publish=publish,
+                        timeout_seconds=5 if cancel_caller else 0.15))
+            try:
+                await asyncio.wait_for(started.wait(), 1)
+                if cancel_caller:
+                    task.cancel()
+                await asyncio.wait_for(draining.wait(), 1)
+                if cancel_caller:
+                    task.cancel()  # must not reset the grace or recancel the reader
+                done, _ = await asyncio.wait({task}, timeout=0.5)
+                assert task in done, "callback suppressed cancellation indefinitely"
+                with pytest.raises(handler.MediaExecutionUncertain, match="teardown"):
+                    task.result()
+                assert comfy.terminal_seen and owned.exists() and comfy.staged()
+                assert handler._UNCERTAIN_WORK
+                pending = list(handler._UNCERTAIN_WORK)
+                release.set()
+                await asyncio.gather(*pending, return_exceptions=True)
+                await asyncio.sleep(0)
+                assert not handler._UNCERTAIN_WORK
+                assert owned.exists() and comfy.staged(), "late callback must not retire files"
+            finally:
+                release.set()
+                if not task.done():
+                    task.cancel()
+                await asyncio.gather(task, return_exceptions=True)
+    run(check())
+
+
+def test_export_replaced_during_collection_is_preserved(tmp_path):
+    async def check():
+        temp = tmp_path / "temp"
+        owned = exporter_file(temp)
+        comfy = Comfy(tmp_path / "input")
+        view = comfy.view
+        async def replace(req):
+            replacement = owned.with_suffix(".replacement")
+            replacement.write_bytes(b"unrelated replacement")
+            replacement.replace(owned)
+            return await view(req)
+        comfy.view = replace
+        async with serve(comfy) as url:
+            await handler.execute_media_request(comfy.data, comfy_url=url, input_dir=comfy.root, temp_dir=temp)
+        assert owned.read_bytes() == b"unrelated replacement"
+    run(check())
+
+
+
+@pytest.mark.parametrize("status", ["success", "error"])
+def test_partial_history_cleans_only_valid_identified_exporters(tmp_path, status):
+    async def check():
+        temp = tmp_path / "temp"
+        owned = exporter_file(temp)
+        unidentified = exporter_file(temp, name="c" * 32 + ".bin")
+        data = request()
+        second = dict(data["media"]["outputs"][0], binding_id="binding:2", node_id="export:2", collector_id="collector:2")
+        data["media"]["outputs"].append(second)
+        data["workflow"]["export:2"] = {"class_type": "MediaExportV1", "inputs": {
+            "request_id": "request:1", "binding_json": json.dumps(second), "texts": ["source:1", 0]}}
+        comfy = Comfy(tmp_path / "input", data)
+        comfy.mode = status
+        async with serve(comfy) as url:
+            with pytest.raises((ValueError, RuntimeError)):
+                await handler.execute_media_request(data, comfy_url=url, input_dir=comfy.root, temp_dir=temp)
+        assert not owned.exists()
+        assert unidentified.exists()
+        assert not comfy.views
+    run(check())
+
+
+def test_missing_history_exports_remain_lifecycle_owned(tmp_path):
+    async def check():
+        temp = tmp_path / "temp"
+        owned = exporter_file(temp)
+        comfy = Comfy(tmp_path / "input")
+        comfy.mode = "missing"
+        async with serve(comfy) as url:
+            with pytest.raises(ValueError, match="Missing"):
+                await handler.execute_media_request(comfy.data, comfy_url=url, input_dir=comfy.root, temp_dir=temp)
+        assert owned.exists()
+        assert not comfy.staged()
+    run(check())
+
+
+@pytest.mark.parametrize("replacement", ["symlink", "directory", "hardlink"])
+def test_nonregular_or_linked_export_files_are_preserved(tmp_path, replacement):
+    async def check():
+        temp = tmp_path / "temp"
+        owned = exporter_file(temp)
+        outside = tmp_path / "outside"
+        outside.write_bytes(b"unrelated")
+        if replacement == "hardlink":
+            alias = tmp_path / "alias"
+            alias.hardlink_to(owned)
+        else:
+            owned.unlink()
+            if replacement == "symlink":
+                owned.symlink_to(outside)
+            else:
+                owned.mkdir()
+        comfy = Comfy(tmp_path / "input")
+        async with serve(comfy) as url:
+            await handler.execute_media_request(comfy.data, comfy_url=url, input_dir=comfy.root, temp_dir=temp)
+        assert owned.exists()
+        assert outside.read_bytes() == b"unrelated"
+    run(check())
+
+
+
+@pytest.mark.parametrize("shared", [False, True])
+def test_direct_exporter_input_prunes_only_obsolete_loader_ancestors(tmp_path, shared):
+    async def check():
+        data = request("VIDEO")
+        data["workflow"]["source:1"] = {"class_type": "LoadVideo", "inputs": {"file": "unavailable.mp4"}}
+        data["media"]["inputs"] = [{"node_id": "export:1", "input_name": "video", "artifact_id": "input:1"}]
+        if shared:
+            second = dict(data["media"]["outputs"][0], binding_id="binding:2", node_id="export:2", collector_id="collector:2")
+            data["media"]["outputs"].append(second)
+            data["workflow"]["export:2"] = {"class_type": "MediaExportV1", "inputs": {
+                "request_id": "request:1", "binding_json": json.dumps(second), "video": ["source:1", 0]}}
+        fake_data = copy.deepcopy(data)
+        fake_data["media"].update(artifacts=[], inputs=[])
+        comfy = Comfy(tmp_path, fake_data)
+        history = comfy.history
+        async def full_history(req):
+            response = await history(req)
+            body = json.loads(response.text)
+            if shared:
+                record = comfy.record()
+                record["binding"] = dict(second, count=1)
+                record["artifacts"][0]["id"] = "b" * 32
+                record["artifacts"][0]["transport"]["filename"] = "b" * 32 + ".bin"
+                record["items"] = [{"binding_id": second["binding_id"], "item_index": 0, "artifact_id": "b" * 32}]
+                body["prompt:1"]["outputs"]["export:2"] = {"media_exports": [record]}
+            return web.json_response(body)
+        comfy.history = full_history
+        async with serve(comfy) as url:
+            result = await handler.execute_media_request(data, comfy_url=url, input_dir=tmp_path)
+        contract.validate_response(result["media"], request_id="request:1", bindings=data["media"]["outputs"])
+        submitted = comfy.submitted[0]["prompt"]
+        imported, slot = submitted["export:1"]["inputs"]["video"]
+        assert slot == 4
+        assert submitted[imported]["class_type"] == "MediaImportV1"
+        assert json.loads(submitted[imported]["inputs"]["artifact_json"])["kind"] == "VIDEO"
+        assert ("source:1" in submitted) is shared
+        assert set(submitted) == ({"export:1", imported, "source:1", "export:2"} if shared else {"export:1", imported})
+        assert not comfy.staged()
+    run(check())
+
+
+@pytest.mark.parametrize("target,kind", [("request_id", "VIDEO"), ("binding_json", "VIDEO"), ("video", "STRING")])
+def test_direct_exporter_input_rejects_controls_and_wrong_kind(tmp_path, target, kind):
+    async def check():
+        data = request("VIDEO")
+        data["media"]["inputs"] = [{"node_id": "export:1", "input_name": target, "artifact_id": "input:1"}]
+        data["media"]["artifacts"][0]["kind"] = kind
+        comfy = Comfy(tmp_path, data)
+        async with serve(comfy) as url:
+            with pytest.raises(ValueError):
+                await handler.execute_media_request(data, comfy_url=url, input_dir=tmp_path)
+        assert not comfy.submitted
+        assert not comfy.staged()
+    run(check())
+
+
+def test_slow_download_times_out_before_any_prompt(tmp_path):
+    async def check():
+        data = request()
+        data["media"]["artifacts"][0]["transport"] = {"type": "url", "url": "https://example.org/input", "expires_at": "2099-01-01T00:00:00Z"}
+        async def download(*args, **kwargs):
+            await asyncio.sleep(60)
+        comfy = Comfy(tmp_path, data)
+        async with serve(comfy) as url:
+            with pytest.raises(TimeoutError):
+                await handler.execute_media_request(data, comfy_url=url, input_dir=tmp_path, download=download, timeout_seconds=0.02)
+        assert not comfy.submitted
+        assert not comfy.staged()
+    run(check())

tests/test_media_image_codec.py

@@ -0,0 +1,330 @@
+"""Real CPU-only IMAGE transfers; no ComfyUI/bootstrap/provider imports."""
+import importlib
+import sys
+import types
+from pathlib import Path
+
+import numpy as np
+import pytest
+import torch
+from PIL import Image
+
+# Load a private package, not Distributed's root __init__.
+ROOT = Path(__file__).resolve().parents[1] / "serverless_bridge"
+PACKAGE = "image_codec_tests_bridge"
+pkg = types.ModuleType(PACKAGE)
+pkg.__path__ = [str(ROOT)]
+sys.modules[PACKAGE] = pkg
+codec = importlib.import_module(PACKAGE + ".codec")
+images = importlib.import_module(PACKAGE + ".image_codec")
+FORMATS = ("legacy_png", "png", "jpeg", "webp", "ffv1", "h264", "h265", "av1")
+LOSSLESS = ("legacy_png", "png", "ffv1")
+
+
+@pytest.fixture(autouse=True)
+def no_gpu(monkeypatch):
+    def forbidden(*a, **kw):
+        pytest.fail("GPU query or initialization is forbidden")
+    for name in ("is_available", "init", "_lazy_init", "device_count"):
+        monkeypatch.setattr(torch.cuda, name, forbidden)
+
+
+def available(fmt):
+    if fmt in images._VIDEO_CODECS:
+        av = pytest.importorskip("av")
+        name = images._VIDEO_CODECS[fmt][0]
+        if name not in av.codecs_available:
+            pytest.skip("Software encoder unavailable: " + name)
+
+
+def sample(channels=3, dtype=torch.float32, shape=(3, 65, 67)):
+    g = torch.Generator(device="cpu").manual_seed(42)
+    return torch.rand((*shape, channels), generator=g, device="cpu", dtype=dtype)
+
+
+def quantized(value):
+    return torch.from_numpy(np.clip(value.float().numpy() * 255, 0, 255).astype(np.uint8)).float() / 255
+
+
+@pytest.mark.parametrize("dtype", [torch.float16, torch.bfloat16, torch.float32, torch.float64])
+def test_exact_unchanged(dtype):
+    value = sample(4, dtype, (2, 3, 5))
+    value.flatten()[0] = -0.0
+    expected = codec.encode_payload("IMAGE", value)
+    assert images.encode_image(value) == expected
+    result = images.decode_image(*expected)
+    assert torch.equal(result.view(torch.uint8), value.view(torch.uint8))
+
+
+@pytest.mark.parametrize("fmt", FORMATS)
+@pytest.mark.parametrize("channels", [1, 3, 4])
+def test_real_roundtrips_shape_order_quantization_alpha(fmt, channels):
+    available(fmt)
+    value = sample(channels)
+    # Large inter-frame difference also verifies frame ordering.
+    value[0] *= 0.2
+    value[2] = value[2] * 0.2 + 0.8
+    if channels == 4:
+        value[..., 3] = torch.linspace(-0.1, 1.1, value[..., 3].numel()).reshape(value[..., 3].shape)
+    descriptor, payload = images.encode_image(value, transfer_format=fmt)
+    assert descriptor["encoding"] == "image-transfer-v1"
+    result = codec.decode_payload("IMAGE", descriptor, payload)
+    assert result.shape == value.shape
+    assert result.device.type == "cpu" and result.is_contiguous()
+    color = 3 if channels == 4 else channels
+    expected = quantized(value[..., :color])
+    if fmt in LOSSLESS:
+        assert torch.equal(result[..., :color], expected)
+    else:
+        # Random RGB noise loses chroma detail in 4:2:0 even at high quality.
+        assert (result[..., :color] - expected).abs().mean() < 0.12
+        assert result[0, ..., :color].mean() < result[2, ..., :color].mean()
+    if channels == 4:
+        assert torch.equal(result[..., 3], value[..., 3])
+
+
+@pytest.mark.parametrize("fmt", ["png", "ffv1", "jpeg", "webp", "h264", "h265", "av1"])
+def test_float64_alpha_is_exact(fmt):
+    available(fmt)
+    value = sample(4, torch.float64, (1, 65, 67))
+    descriptor, payload = images.encode_image(value, transfer_format=fmt)
+    out = images.decode_image(descriptor, payload)
+    assert out.dtype == torch.float64
+    assert torch.equal(out[..., 3].contiguous().view(torch.uint8), value[..., 3].contiguous().view(torch.uint8))
+
+
+@pytest.mark.parametrize("fmt", ["jpeg", "webp", "h264", "h265", "av1"])
+def test_lossy_quality_changes_final_payload_size(fmt):
+    available(fmt)
+    value = sample(shape=(2, 96, 96))
+    low = images.encode_image(value, transfer_format=fmt, transfer_quality=15)[1]
+    high = images.encode_image(value, transfer_format=fmt, transfer_quality=95)[1]
+    assert len(high) > len(low)
+
+
+def test_truncation_clipping_grayscale_noncontiguous():
+    value = torch.tensor([-0.1, 0.5, 1.1, 0.12345], device="cpu").reshape(1, 2, 2, 1).transpose(1, 2)
+    out = images.decode_image(*images.encode_image(value, transfer_format="png"))
+    assert torch.equal(out, quantized(value))
+    assert out[0, 0, 1, 0] == 1
+
+
+@pytest.mark.parametrize("value", [None, [], torch.ones(2, 3, 3), torch.ones(1, 2, 3, 2), torch.ones(1, 2, 3, 0), torch.ones(1, 2, 3, 3, dtype=torch.uint8), torch.ones(1, 2, 3, 3, device="meta"), torch.ones(1, 2, 3, 3).to_sparse()])
+def test_invalid_values(value):
+    with pytest.raises(ValueError):
+        images.encode_image(value, transfer_format="png")
+
+
+@pytest.mark.parametrize("bad", [float("nan"), float("inf"), -float("inf")])
+def test_nonfinite_rejected(bad):
+    with pytest.raises(ValueError, match="finite"):
+        images.encode_image(torch.full((1, 1, 1, 3), bad), transfer_format="png")
+
+
+@pytest.mark.parametrize("options", [{"transfer_format": "PNG"}, {"transfer_format": []}, {"transfer_quality": True}, {"transfer_quality": 0}, {"transfer_quality": 101}, {"transfer_quality": 1.5}, {"max_bytes": True}, {"max_bytes": -1}])
+def test_strict_options(options):
+    with pytest.raises(ValueError):
+        images.encode_image(sample(shape=(1, 1, 1)), **options)
+
+
+@pytest.mark.parametrize("change", [{"extra": 1}, {"format": "exact"}, {"quality": True}, {"layout": "BCHW"}, {"shape": [1, 2**60, 1, 3]}, {"shape": [True, 1, 1, 3]}, {"parts": [0]}, {"parts": [True]}, {"parts": [2**60]}, {"alpha": {"extra": 0}}])
+def test_bad_descriptors(change):
+    descriptor, payload = images.encode_image(sample(shape=(1, 3, 5)), transfer_format="png")
+    descriptor.update(change)
+    with pytest.raises(ValueError):
+        images.decode_image(descriptor, payload)
+
+
+def test_truncated_appended_and_mismatched_parts():
+    d, p = images.encode_image(sample(shape=(2, 3, 5)), transfer_format="png")
+    for payload in (p[:-1], p + b"x"):
+        with pytest.raises(ValueError):
+            images.decode_image(d, payload)
+    for change in ({"shape": [1, 3, 5, 3]}, {"parts": [len(p)]}, {"shape": [2, 2, 5, 3]}):
+        bad = dict(d, **change)
+        with pytest.raises(ValueError):
+            images.decode_image(bad, p)
+
+
+@pytest.mark.parametrize("fmt", ["png", "jpeg", "webp", "ffv1", "h264", "h265", "av1"])
+def test_payload_and_decoded_budget(fmt):
+    available(fmt)
+    value = sample(shape=(1, 65, 67))
+    d, p = images.encode_image(value, transfer_format=fmt)
+    with pytest.raises(ValueError, match="limit|budget"):
+        images.decode_image(d, p, max_bytes=100)
+    with pytest.raises(ValueError, match="limit|budget"):
+        images.encode_image(value, transfer_format=fmt, max_bytes=100)
+
+
+def test_oversized_shape_before_materialization(monkeypatch):
+    value = torch.ones(1).expand(1, 8192, 8192, 4)
+    def forbidden(*a, **kw):
+        pytest.fail("Materialized an oversized tensor")
+    monkeypatch.setattr(torch.Tensor, "contiguous", forbidden)
+    monkeypatch.setattr(torch, "isfinite", forbidden)
+    with pytest.raises(ValueError):
+        images.encode_image(value, transfer_format="png")
+
+
+def test_header_dimensions_checked_before_pixels(monkeypatch):
+    d, p = images.encode_image(sample(shape=(1, 9, 11)), transfer_format="png")
+    d["shape"] = [1, 1, 1, 3]
+    def forbidden(*a, **kw):
+        pytest.fail("Pixels decoded before header dimensions were checked")
+    monkeypatch.setattr(Image.Image, "load", forbidden)
+    with pytest.raises(ValueError):
+        images.decode_image(d, p)
+
+
+def test_other_kinds_do_not_dispatch():
+    d, p = images.encode_image(sample(shape=(1, 3, 5)), transfer_format="png")
+    for kind in ("MASK", "AUDIO", "STRING"):
+        with pytest.raises(ValueError):
+            codec.decode_payload(kind, d, p)
+
+
+def test_missing_video_encoder_is_explicit(monkeypatch):
+    monkeypatch.setitem(images._VIDEO_CODECS, "h264", ("missing_software_encoder", 51, "yuv420p"))
+    with pytest.raises(ValueError, match="encoder|unavailable"):
+        images.encode_image(sample(shape=(1, 64, 64)), transfer_format="h264")
+
+
+def test_bounded_sink_catches_final_flush_bytes():
+    sink = images._BoundedIO(4)
+    sink.write(b"1234")
+    with pytest.raises(ValueError, match="limit"):
+        sink.write(b"5")
+    assert sink.getvalue() == b"1234"
+
+
+def test_crf_mapping():
+    assert images._quality_to_crf(90, 51) == 13
+    assert images._quality_to_crf(90, 63) == 16
+    assert images._quality_to_crf(100, 63) == 0
+
+
+@pytest.mark.parametrize("fmt", FORMATS)
+@pytest.mark.parametrize("dtype", [torch.float16, torch.bfloat16, torch.float32, torch.float64])
+def test_compressed_preserves_source_dtype_and_exact_alpha(fmt, dtype):
+    available(fmt)
+    value = sample(4, dtype, (1, 17, 19))
+    value[..., 3].flatten()[0] = -0.0
+    d, p = images.encode_image(value, transfer_format=fmt)
+    with torch.device("meta"):
+        result = images.decode_image(d, p)
+    assert result.device.type == "cpu" and result.dtype == dtype
+    assert d["dtype"] == str(dtype).removeprefix("torch.")
+    assert torch.equal(result[..., 3].contiguous().view(torch.uint8), value[..., 3].contiguous().view(torch.uint8))
+    if fmt in LOSSLESS:
+        assert torch.equal(result[..., :3], quantized(value[..., :3]).to(dtype))
+
+
+@pytest.mark.parametrize("fmt", ["ffv1", "h264", "h265", "av1"])
+def test_video_count_and_header_fail_before_native_decoder(fmt, monkeypatch):
+    available(fmt)
+    d, p = images.encode_image(sample(shape=(2, 65, 67)), transfer_format=fmt)
+    def forbidden(*a, **kw):
+        pytest.fail("Native decoder reached before header validation")
+    monkeypatch.setattr(images, "_decode_video", forbidden)
+    for shape in ([1, 65, 67, 3], [3, 65, 67, 3], [2, 32, 32, 3]):
+        bad = dict(d, shape=shape)
+        with pytest.raises(ValueError):
+            images.decode_image(bad, p)
+    for payload in (p[:-1], p[:40], p + b"extra"):
+        bad = dict(d, parts=[len(payload)])
+        with pytest.raises(ValueError):
+            images.decode_image(bad, payload)
+
+
+@pytest.mark.parametrize("fmt", ["ffv1", "h264", "h265", "av1"])
+def test_video_does_not_use_native_container_probe(fmt, monkeypatch):
+    available(fmt)
+    d, p = images.encode_image(sample(shape=(2, 65, 67)), transfer_format=fmt)
+    av = images._import_av()
+    def forbidden(*a, **kw):
+        pytest.fail("Unbounded native stream probing is forbidden")
+    monkeypatch.setattr(av, "open", forbidden)
+    assert images.decode_image(d, p).shape == (2, 65, 67, 3)
+
+
+@pytest.mark.parametrize("fmt", ["ffv1", "h264", "h265", "av1"])
+def test_real_video_flush_bound(fmt):
+    available(fmt)
+    frames = [np.zeros((64, 64, 3), dtype=np.uint8)]
+    good = images._encode_video(iter(frames), fmt, 90, 64, 64, 100000)
+    with pytest.raises(ValueError, match="limit"):
+        images._encode_video(iter(frames), fmt, 90, 64, 64, len(good) - 1)
+
+
+def test_alpha_invalid_dtype_nonfinite_and_extra_fields():
+    import copy
+    d, p = images.encode_image(sample(4, shape=(1, 3, 5)), transfer_format="png")
+    for field, value in (("dtype", "float64"), ("shape", [1, 3, 6, 1]), ("extra", 1)):
+        bad = copy.deepcopy(d)
+        bad["alpha"]["descriptor"][field] = value
+        with pytest.raises(ValueError):
+            images.decode_image(bad, p)
+    import struct
+    bad = p[:-4] + struct.pack("<f", float("nan"))
+    with pytest.raises(ValueError, match="finite"):
+        images.decode_image(d, bad)
+
+
+def test_descriptor_structure_bound_and_frame_limit():
+    d, p = images.encode_image(sample(shape=(1, 3, 5)), transfer_format="png")
+    for bad in (dict(d, quality="x" * 70000), dict(d, parts=[1] * 241), dict(d, shape=[241, 1, 1, 3]), dict(d, dtype="x")):
+        with pytest.raises(ValueError):
+            images.decode_image(bad, p)
+    with pytest.raises(ValueError):
+        images.encode_image(torch.zeros(241, 1, 1, 3), transfer_format="png")
+
+
+@pytest.mark.parametrize("fmt", ["png", "jpeg", "webp"])
+def test_all_still_headers_checked_before_pillow_open(fmt, monkeypatch):
+    d, p = images.encode_image(sample(shape=(1, 31, 33)), transfer_format=fmt)
+    d["shape"] = [1, 2, 2, 3]
+    def forbidden(*a, **kw):
+        pytest.fail("Pillow opened a mismatched header")
+    monkeypatch.setattr(Image, "open", forbidden)
+    with pytest.raises(ValueError, match="dimensions"):
+        images.decode_image(d, p)
+
+
+@pytest.mark.parametrize("fmt", ["ffv1", "h264", "h265", "av1"])
+def test_video_longer_batch_order_and_coded_allocation_bound(fmt):
+    available(fmt)
+    value = torch.linspace(0, 1, 16).reshape(16, 1, 1, 1).expand(16, 64, 64, 3)
+    d, p = images.encode_image(value, transfer_format=fmt)
+    header = images._video_header(p, fmt, 64, 64, 16)
+    coded = images._coded_frame_bound(p, fmt, header)
+    assert 16 <= coded <= images.MAX_FRAMES * 4
+    result = images.decode_image(d, p)
+    assert torch.all((result.mean(dim=(1, 2, 3)) - value.mean(dim=(1, 2, 3))).abs() < 0.03)
+
+
+@pytest.mark.parametrize("fmt", ["h264", "h265", "av1"])
+def test_packet_unit_bounds_before_native_decode(fmt):
+    available(fmt)
+    d, p = images.encode_image(sample(shape=(1, 64, 64)), transfer_format=fmt)
+    config, bounds = images._video_header(p, fmt, 64, 64, 1)
+    start, end = bounds[0]
+    packet = p[start:end]
+    with pytest.raises(ValueError):
+        images._coded_frame_bound(packet[:-1], fmt, (config, [(0, len(packet) - 1)]))
+    # An adversarial single block can contain many coded frames: bound it too.
+    enormous = packet * (images.MAX_FRAMES * 4 + 1)
+    with pytest.raises(ValueError, match="limit"):
+        images._coded_frame_bound(enormous, fmt, (config, [(0, len(enormous))]))
+
+
+@pytest.mark.parametrize("fmt", ["h264", "h265", "av1"])
+def test_bitstream_resolution_cannot_bypass_native_pixel_limit(fmt, monkeypatch):
+    available(fmt)
+    d, p = images.encode_image(sample(shape=(1, 256, 256)), transfer_format=fmt)
+    header = images._video_header(p, fmt, 256, 256, 1)
+    av = images._import_av()
+    # A lying container is not enough to disable max_pixels in the decoder.
+    # The bitstream still contains 256x256 pixels but the budget permits 64x64.
+    with pytest.raises((ValueError, av.error.FFmpegError)):
+        list(images._decode_video(p, fmt, 64, 64, 1, header))

tests/test_media_prompt_transform.py

@@ -0,0 +1,83 @@
+"""Participant identity, typed compilation and legacy Collector pruning."""
+import copy
+import importlib
+import json
+import sys
+import types
+from pathlib import Path
+
+import pytest
+
+ROOT = Path(__file__).resolve().parents[1]
+PACKAGE = "media_transform_testpkg"
+for suffix in ("", ".api", ".api.orchestration", ".utils", ".serverless_bridge"):
+    module = types.ModuleType(PACKAGE + suffix)
+    module.__path__ = [str(ROOT.joinpath(*suffix.strip(".").split(".")))] if suffix else [str(ROOT)]
+    sys.modules[module.__name__] = module
+pt = importlib.import_module(PACKAGE + ".api.orchestration.prompt_transform")
+compiler = importlib.import_module(PACKAGE + ".api.orchestration.serverless_media_prompt")
+SOCKETS = ("images", "audio")
+
+
+def prompt(sockets=SOCKETS):
+    nodes = {
+        "source": {"class_type": "MediaSource", "inputs": {}},
+        "collector": {"class_type": "DistributedCollector", "inputs": {
+            socket: ["source", SOCKETS.index(socket)] for socket in sockets}},
+    }
+    for socket in sockets:
+        nodes["sink:" + socket] = {"class_type": "Capture", "inputs": {"value": ["collector", SOCKETS.index(socket)]}}
+    return nodes
+
+
+def prepare(graph, participant, delegate):
+    index = pt.PromptIndex(graph)
+    jobs = pt.generate_job_id_map(index, "run")
+    assert jobs == {"collector": "run_collector"}
+    return pt.apply_participant_overrides(index.copy_prompt(), participant,
+        ["serverless:pool:1", "serverless:pool:2"], jobs, "http://master", delegate, index)
+
+
+def test_typed_worker_preparation_compiles_both_bindings_without_source_mutation():
+    graph = prompt()
+    before = copy.deepcopy(graph)
+    worker = prepare(graph, "serverless:pool:2", True)
+    compiled, bindings = compiler.compile_media_prompt(worker, "request:2")
+    assert len(bindings) == 2
+    assert {row["participant_id"] for row in bindings} == {"serverless:pool:2"}
+    assert {row["job_id"] for row in bindings} == {"run_collector"}
+    assert {row["socket"] for row in bindings} == set(SOCKETS)
+    assert set(compiled) == {"source"} | {row["node_id"] for row in bindings}
+    assert graph == before
+
+
+@pytest.mark.parametrize("sockets", [("images",), ("audio",), SOCKETS])
+def test_delegate_master_keeps_collector_socket_contract_and_legacy_placeholder(sockets):
+    graph = prompt(sockets)
+    prepared = prepare(graph, "master", True)
+    before = copy.deepcopy(prepared)
+    delegated = pt.prepare_delegate_master_prompt(prepared, ["collector"])
+    assert "source" not in delegated
+    assert {"collector"} | {"sink:" + s for s in sockets} <= set(delegated)
+    inputs = delegated["collector"]["inputs"]
+    assert "audio" not in inputs and "expected_media" not in inputs
+    placeholders = [key for key, node in delegated.items() if node['class_type'] == 'DistributedEmptyImage']
+    if "images" in sockets:
+        assert len(placeholders) == 1
+        assert inputs['images'] == [placeholders[0], 0]
+    else:
+        assert not placeholders and 'images' not in inputs
+    assert inputs["delegate_only"] is True
+    assert json.loads(inputs["enabled_worker_ids"]) == ["serverless:pool:1", "serverless:pool:2"]
+    assert prepared == before
+
+
+@pytest.mark.parametrize("socket", SOCKETS)
+def test_participating_master_retains_exact_media_links(socket):
+    graph = prompt()
+    original = copy.deepcopy(graph)
+    master = prepare(graph, "master", False)
+    assert master["collector"]["inputs"][socket] == graph["collector"]["inputs"][socket]
+    assert master["collector"]["inputs"]["delegate_only"] is False
+    assert "worker_id" not in master["collector"]["inputs"]
+    assert graph == original

tests/test_media_queue_orchestration.py

@@ -0,0 +1,552 @@
+"""Offline queue integration: real compilers, packers and wire validators.
+
+Only host dispatch/probing/queueing and the manager lifecycle are mocked. No
+provider requests, services or GPU work are permitted by these tests.
+"""
+import asyncio
+import base64
+import copy
+import importlib
+import json
+import os
+from pathlib import Path
+import socket
+import subprocess
+import sys
+import types
+import uuid
+import wave
+from unittest.mock import AsyncMock, Mock, patch
+
+import aiohttp
+import pytest
+# Preload before the shared integration loader snapshots sys.modules: extension
+# modules cannot safely be unloaded and imported again between fixture contexts.
+import torch
+from PIL import Image
+
+from test_serverless_integration import load_modules, module, PKG
+
+
+PROFILE = "comfy-media-v1"
+
+
+def node(class_type, **inputs):
+    return {"class_type": class_type, "inputs": inputs}
+
+
+def collector_prompt():
+    return {
+        "source": node("EmptyImage", width=32, height=32, batch_size=1, color=0),
+        "collector": node("DistributedCollector", images=["source", 0], load_balance=False),
+        "save": node("SaveImage", images=["collector", 0], filename_prefix="local"),
+    }
+
+
+def legacy_prompt():
+    return {
+        "source": node("EmptyImage", width=32, height=32, batch_size=1, color=0),
+        "collector": node("DistributedCollector", images=["source", 0]),
+        "save": node("SaveImage", images=["collector", 0], filename_prefix="local"),
+    }
+
+
+@pytest.fixture
+def harness(monkeypatch):
+    q, modules, ps = load_modules()
+    with patch.dict(sys.modules, modules):
+        pool = {"id": "pool", "name": "Media", "type": "serverless", "provider": "runpod",
+                "endpoint_id": "offline-endpoint", "enabled": True, "jobs_per_run": 2,
+                "timeout_seconds": 30, "media_profile": PROFILE}
+        config = {"workers": [pool], "settings": {"master_delegate_only": True}}
+        q.load_config = lambda: config
+        q.select_active_workers = AsyncMock(side_effect=lambda workers, _ws, delegate, **kw: (workers, delegate))
+        q.queue_prompt_payload = AsyncMock(return_value={"prompt_id": "master-prompt", "number": 7, "node_errors": {}})
+        manager = types.SimpleNamespace(start=AsyncMock(), cancel=AsyncMock(), ensure_capacity=Mock())
+        sys.modules[PKG + ".api.orchestration.serverless_runtime"] = module("runtime", get_manager=lambda: manager)
+        # These are the actual modules, not permissive serializer/packing stubs.
+        client = importlib.import_module(PKG + ".api.orchestration.runpod_client")
+        compiler = importlib.import_module(PKG + ".api.orchestration.serverless_media_prompt")
+        packer = importlib.import_module(PKG + ".api.orchestration.serverless_typed_inputs")
+        validator = importlib.import_module(PKG + ".api.orchestration.serverless_typed_results")
+        handler = importlib.import_module(PKG + ".serverless_bridge.handler")
+        def forbidden(*args, **kwargs):
+            raise AssertionError("Offline queue tests must not access a network or GPU")
+        monkeypatch.setattr(aiohttp, "ClientSession", forbidden)
+        monkeypatch.setattr(socket, "create_connection", forbidden)
+        monkeypatch.setattr(socket.socket, "connect", forbidden)
+        monkeypatch.setattr(socket.socket, "connect_ex", forbidden)
+        monkeypatch.setattr(torch.cuda, "init", forbidden)
+        monkeypatch.setattr(torch.cuda, "current_device", forbidden)
+        h = types.SimpleNamespace(q=q, ps=ps, pool=pool, config=config, manager=manager,
+                                  client=client, compiler=compiler, packer=packer,
+                                  validator=validator, handler=handler)
+        yield h
+
+
+@pytest.fixture
+def media_paths(tmp_path, monkeypatch):
+    paths = types.SimpleNamespace(get_annotated_filepath=lambda name: str(tmp_path / name))
+    for kind in ("input", "output", "temp"):
+        setattr(paths, "get_" + kind + "_directory", lambda: str(tmp_path))
+    monkeypatch.setitem(sys.modules, "folder_paths", paths)
+    return tmp_path
+
+
+def run(h, prompt=None, **kwargs):
+    return asyncio.run(h.q.orchestrate_distributed_execution(
+        prompt if prompt is not None else collector_prompt(), None, "client", **kwargs))
+
+
+def assert_no_dispatch(h, *, no_probe=False):
+    h.q.queue_prompt_payload.assert_not_awaited()
+    h.q.dispatch_worker_prompt.assert_not_awaited()
+    h.manager.start.assert_not_awaited()
+    h.manager.cancel.assert_not_awaited()
+    assert h.ps.distributed_pending_jobs == {}
+    if no_probe:
+        h.q.select_active_workers.assert_not_awaited()
+        h.q.select_least_busy_worker.assert_not_awaited()
+        h.q.fetch_worker_has_node.assert_not_awaited()
+        h.q.sync_worker_media.assert_not_awaited()
+
+
+@pytest.mark.parametrize("delegate", [True, False])
+def test_real_compile_pack_serialize_and_canonical_bindings(harness, media_paths, delegate):
+    h = harness
+    image_path = media_paths / "picture.png"
+    Image.new("RGBA", (2, 3), color=(17, 29, 41, 53)).save(image_path)
+    image_bytes = image_path.read_bytes()
+    prompt = collector_prompt()
+    prompt["image"] = node("LoadImage", image="picture.png")
+    prompt["audio"] = node("GeneratedAudio")
+    prompt["collector"]["inputs"].update(images=["image", 0], audio=["audio", 0])
+    prompt["collector2"] = node("DistributedCollector", audio=["audio", 0])
+    prompt["unused"] = node("PreviewImage", images=["image", 0])
+    source = copy.deepcopy(prompt)
+    with patch.object(h.compiler, "compile_media_prompt", wraps=h.compiler.compile_media_prompt) as compile_call, \
+         patch.object(h.packer, "prepare_media_inputs", wraps=h.packer.prepare_media_inputs) as pack_call, \
+         patch.object(h.client, "serialize_request", wraps=h.client.serialize_request) as serialize_call, \
+         patch.object(h.q, "prune_prompt_for_worker", side_effect=AssertionError("typed graph must bypass legacy pruning")):
+        result = run(h, prompt, delegate_master=delegate)
+    assert result == ("master-prompt", 7, 2, {})
+    assert compile_call.call_count == pack_call.await_count == 2
+    assert serialize_call.call_count >= 2
+    assert prompt == source
+    h.q.dispatch_worker_prompt.assert_not_awaited()
+    jobs = h.manager.start.call_args.args[2]
+    assert [j["worker"]["id"] for j in jobs] == ["serverless:pool:1", "serverless:pool:2"]
+    assert len({j["media"]["request_id"] for j in jobs}) == 2
+    master = h.q.queue_prompt_payload.call_args.args[0]
+    for job in jobs:
+        media = job["media"]
+        assert str(uuid.UUID(media["request_id"])) == media["request_id"]
+        assert uuid.UUID(media["request_id"]).version == 4
+        assert media["profile"] == PROFILE and media["version"] == 1
+        assert media["outputs"] == job["manifest"]
+        assert media["outputs"] is not job["manifest"]
+        assert media["artifacts"] == media["inputs"] == []
+        assert {b["socket"] for b in job["manifest"]} == {"images", "audio"}
+        assert len(job["manifest"]) == 3
+        assert base64.b64decode(job["images"][0]["image"]) == image_bytes
+        assert job["workflow"]["image"]["inputs"]["image"] == job["images"][0]["name"]
+        assert not ({"save", "unused", "collector", "collector2"} & set(job["workflow"]))
+        body = h.client.serialize_request(job["workflow"], job["images"], 30, media=media)
+        wire = json.loads(body)
+        assert wire["input"] == {"workflow": job["workflow"], "images": job["images"], "media": media}
+        assert wire["policy"] == {"executionTimeout": 30000, "ttl": 30000}
+        _, _, _, targets = h.handler._validate_request(wire["input"])
+        assert targets == [b["node_id"] for b in job["manifest"]]
+        assert h.validator.validate_prepared_job(job) is True
+        for binding in job["manifest"]:
+            assert binding["participant_id"] == job["worker"]["id"]
+            assert binding["job_id"] == master[binding["collector_id"]]["inputs"]["multi_job_id"]
+            controls = job["workflow"][binding["node_id"]]["inputs"]
+            assert json.loads(controls["binding_json"]) == binding
+            assert controls["request_id"] == media["request_id"]
+    assert master["collector"]["inputs"]["delegate_only"] is delegate
+    assert "expected_media" not in master["collector"]["inputs"]
+    if delegate:
+        assert not {"source", "audio"} & set(master)
+        # LoadImage is an existing explicitly retained lightweight source; the
+        # Collector is nevertheless detached from it and uses the placeholder.
+        assert master["image"] == prompt["image"]
+        assert "audio" not in master["collector"]["inputs"]
+        placeholder = master["collector"]["inputs"]["images"][0]
+        assert master[placeholder]["class_type"] == "DistributedEmptyImage"
+    else:
+        assert master["collector"]["inputs"]["audio"] == ["audio", 0]
+        assert "audio" in master and "image" in master
+
+
+@pytest.mark.parametrize("audio_only", [False, True])
+def test_delegate_prunes_image_audio_sources_and_audio_only_needs_no_placeholder(harness, audio_only):
+    prompt = collector_prompt()
+    prompt["audio"] = node("GeneratedAudio")
+    prompt["collector"]["inputs"]["audio"] = ["audio", 0]
+    if audio_only:
+        del prompt["collector"]["inputs"]["images"]
+        prompt["save"] = node("AudioSink", audio=["collector", 1])
+    original = copy.deepcopy(prompt)
+    run(harness, prompt)
+    master = harness.q.queue_prompt_payload.call_args.args[0]
+    inputs = master["collector"]["inputs"]
+    assert "audio" not in inputs and "expected_media" not in inputs
+    assert "source" not in master and "audio" not in master
+    if audio_only:
+        assert "images" not in inputs and set(master) == {"collector", "save"}
+    else:
+        assert master[inputs["images"][0]]["class_type"] == "DistributedEmptyImage"
+    for job in harness.manager.start.call_args.args[2]:
+        assert {b["socket"] for b in job["manifest"]} == ({"audio"} if audio_only else {"images", "audio"})
+        assert "audio" in job["workflow"]
+        assert ("source" in job["workflow"]) is not audio_only
+        assert harness.validator.validate_prepared_job(job) is True
+    assert prompt == original
+
+
+@pytest.mark.parametrize("selected", ["master", "worker"])
+def test_typed_load_balance_selects_master_or_one_serverless_slot(harness, selected):
+    h = harness
+    prompt = collector_prompt()
+    prompt["collector"]["inputs"]["load_balance"] = "true"
+    h.q.select_least_busy_worker.side_effect = lambda candidates, **kw: next(
+        w for w in candidates if (w["id"] == "master") == (selected == "master"))
+    result = run(h, prompt, delegate_master=False)
+    h.q.select_least_busy_worker.assert_awaited_once()
+    master = h.q.queue_prompt_payload.call_args.args[0]
+    if selected == "master":
+        assert result[2] == 0
+        h.manager.start.assert_not_awaited()
+        assert master["collector"]["inputs"]["delegate_only"] is False
+        assert "source" in master
+    else:
+        assert result[2] == 1
+        assert len(h.manager.start.call_args.args[2]) == 1
+        assert master["collector"]["inputs"]["delegate_only"] is True
+        assert "source" not in master
+
+
+@pytest.mark.parametrize("profile", [None, "legacy-image-v1"])
+def test_exact_rejects_legacy_pool_before_probe(harness, profile):
+    h = harness
+    prompt = collector_prompt()
+    prompt["collector"]["inputs"]["transfer_format"] = "exact"
+    if profile is None:
+        h.pool.pop("media_profile")
+    else:
+        h.pool["media_profile"] = profile
+    with pytest.raises(ValueError, match="exact requires comfy-media-v1"):
+        run(h, prompt)
+    assert_no_dispatch(h, no_probe=True)
+
+
+@pytest.mark.parametrize("profile", [None, "auto", "comfy-media-v2", False])
+def test_invalid_explicit_profiles_fail_closed(harness, profile):
+    harness.pool["media_profile"] = profile
+    with pytest.raises(ValueError, match="media_profile"):
+        run(harness)
+    assert_no_dispatch(harness, no_probe=True)
+
+
+@pytest.mark.parametrize("direct_type", ["local", "remote", "cloud"])
+@pytest.mark.parametrize("explicit", [False, True])
+def test_direct_workers_participate_alongside_typed_pools(harness, direct_type, explicit):
+    h = harness
+    h.config["workers"].append({"id": "direct", "type": direct_type, "host": "offline.invalid",
+                                "enabled": not explicit, "port": 8188})
+    kwargs = {"enabled_worker_ids": ["pool", "direct"]} if explicit else {}
+    prompt = collector_prompt()
+    original = copy.deepcopy(prompt)
+    assert run(h, prompt, delegate_master=False, **kwargs)[2] == 3
+    assert prompt == original
+    h.q.dispatch_worker_prompt.assert_awaited_once()
+    master = h.q.queue_prompt_payload.call_args.args[0]
+    assert json.loads(master["collector"]["inputs"]["enabled_worker_ids"]) == [
+        "serverless:pool:1", "serverless:pool:2", "direct"]
+    jobs = h.manager.start.call_args.args[2]
+    assert len(jobs) == 2 and all(j["media"]["profile"] == PROFILE for j in jobs)
+    if direct_type != "local":
+        h.q.sync_worker_media.assert_awaited_once()
+
+
+def test_disabled_or_unselected_workers_do_not_participate(harness):
+    h = harness
+    h.config["workers"] += [
+        {"id": "direct", "type": "local", "enabled": False},
+        dict(h.pool, id="legacy", media_profile="legacy-image-v1", enabled=False),
+    ]
+    assert run(h)[2] == 2
+    h.pool["enabled"] = False
+    assert run(h, enabled_worker_ids=["pool"])[2] == 2
+
+
+@pytest.mark.parametrize("workers", [True, False])
+def test_image_list_cannot_mix_with_collector_even_master_only(harness, workers):
+    if not workers:
+        harness.config["workers"] = []
+    prompt = collector_prompt()
+    prompt["incompatible"] = node("DistributedImageListStart")
+    with pytest.raises(ValueError, match="cannot be mixed"):
+        run(harness, prompt, delegate_master=False)
+    assert_no_dispatch(harness, no_probe=True)
+
+
+def test_mixed_direct_typed_and_legacy_profiles_keep_each_wire_contract(harness):
+    h = harness
+    h.config["workers"] += [dict(h.pool, id="legacy", media_profile="legacy-image-v1"),
+                             {"id": "direct", "type": "local", "enabled": True}]
+    assert run(h, collector_prompt())[2] == 5
+    h.q.dispatch_worker_prompt.assert_awaited_once()
+    jobs = h.manager.start.call_args.args[2]
+    assert len(jobs) == 4
+    for job in jobs:
+        if job["worker"]["media_profile"] == PROFILE:
+            assert job["media"]["outputs"] == job["manifest"]
+            assert h.validator.validate_prepared_job(job) is True
+        else:
+            assert "media" not in job
+            assert any(n["class_type"] == "SaveImage" for n in job["workflow"].values())
+
+
+def test_typed_master_only_is_supported_without_serverless_work(harness):
+    h = harness
+    h.config["workers"] = []
+    assert run(h, delegate_master=False)[2] == 0
+    master = h.q.queue_prompt_payload.call_args.args[0]
+    assert master["collector"]["inputs"]["enabled_worker_ids"] == "[]"
+    assert master["collector"]["inputs"]["images"] == ["source", 0]
+    h.manager.start.assert_not_awaited()
+    h.q.dispatch_worker_prompt.assert_not_awaited()
+
+
+@pytest.mark.parametrize("offline", [False, True])
+def test_collector_delegate_without_workers_preserves_master_fallback(harness, offline):
+    h = harness
+    if offline:
+        h.q.select_active_workers.side_effect = lambda *a, **kw: ([], False)
+    else:
+        h.config["workers"] = []
+    assert run(h)[2] == 0
+    master = h.q.queue_prompt_payload.call_args.args[0]
+    assert master["collector"]["inputs"]["delegate_only"] is False
+    assert master["collector"]["inputs"]["images"] == ["source", 0]
+    h.manager.start.assert_not_awaited()
+
+
+@pytest.mark.parametrize("bad", [" pool ", "pool white", "pool\n"])
+def test_pool_media_identity_is_not_repaired(harness, bad):
+    harness.pool["id"] = bad
+    with pytest.raises(ValueError, match="identifier"):
+        run(harness)
+    assert_no_dispatch(harness, no_probe=True)
+
+
+def test_integer_source_node_identity_is_not_normalized(harness):
+    prompt = collector_prompt()
+    prompt[1] = prompt.pop("source")
+    prompt["collector"]["inputs"]["images"] = [1, 0]
+    with pytest.raises(ValueError, match="identifier"):
+        run(harness, prompt)
+    assert_no_dispatch(harness, no_probe=True)
+
+
+@pytest.mark.parametrize("broken", ["dangling-unused", "literal", "missing-file", "bad-loader-controls", "oversize"])
+def test_compilation_packing_and_size_errors_precede_all_dispatch(harness, media_paths, broken):
+    h = harness
+    prompt = collector_prompt()
+    if broken == "dangling-unused":
+        prompt["unused"] = node("PreviewImage", images=["missing", 0])
+    elif broken == "literal":
+        prompt["collector"]["inputs"]["images"] = "must use a linked image node"
+    elif broken in ("missing-file", "bad-loader-controls"):
+        prompt["audio"] = node("LoadAudio", audio="missing.wav")
+        prompt["collector"]["inputs"]["audio"] = ["audio", 0]
+        if broken == "bad-loader-controls":
+            prompt["audio"]["inputs"]["duration"] = 5
+    else:
+        prompt["source"]["inputs"]["label"] = "雪" * 1000
+    original = copy.deepcopy(prompt)
+    with patch.object(h.client, "MAX_REQUEST_BYTES", 100 if broken == "oversize" else h.client.MAX_REQUEST_BYTES):
+        with pytest.raises((ValueError, OSError)):
+            run(h, prompt)
+    assert prompt == original
+    assert_no_dispatch(h)
+
+
+@pytest.mark.parametrize("corruption", ["manifest", "participant", "request-control", "binding-control", "profile", "input-ref"])
+def test_whole_batch_identity_preflight_before_master_and_manager(harness, corruption):
+    h = harness
+    real_prepare = h.q._prepare_worker_payload
+    async def corrupt_later(*args, **kwargs):
+        worker, prepared = await real_prepare(*args, **kwargs)
+        if worker["id"].endswith(":2"):
+            binding = prepared["manifest"][0]
+            if corruption == "manifest":
+                binding["job_id"] = "wrong-job"
+            elif corruption == "participant":
+                prepared["worker"] = dict(worker, id="different-participant")
+            elif corruption == "request-control":
+                prepared["workflow"][binding["node_id"]]["inputs"]["request_id"] = "different-request"
+            elif corruption == "binding-control":
+                prepared["workflow"][binding["node_id"]]["inputs"]["binding_json"] = "{}"
+            elif corruption == "profile":
+                prepared["worker"] = dict(worker, media_profile="legacy-image-v1")
+            else:
+                prepared["media"]["inputs"] = [{"node_id": "missing", "input_name": "texts", "artifact_id": "unknown"}]
+        return worker, prepared
+    with patch.object(h.q, "_prepare_worker_payload", side_effect=corrupt_later):
+        with pytest.raises(ValueError):
+            run(h)
+    h.manager.ensure_capacity.assert_not_called()
+    assert_no_dispatch(h)
+
+
+def test_batch_preflight_also_precedes_regular_direct_dispatch(harness):
+    h = harness
+    h.pool.pop("media_profile")
+    h.config["workers"].append({"id": "direct", "type": "local", "enabled": True})
+    real_prepare = h.q._prepare_worker_payload
+    async def corrupt_later(*args, **kwargs):
+        worker, prepared = await real_prepare(*args, **kwargs)
+        if worker["id"] == "serverless:pool:2":
+            prepared["worker"] = dict(worker, media_profile=PROFILE)
+        return worker, prepared
+    with patch.object(h.q, "_prepare_worker_payload", side_effect=corrupt_later):
+        with pytest.raises(ValueError, match="mismatch"):
+            run(h, legacy_prompt())
+    assert_no_dispatch(h)
+
+
+def test_capacity_preflight_precedes_collector_state_and_master(harness):
+    h = harness
+    h.manager.ensure_capacity.side_effect = ValueError("capacity exhausted")
+    with pytest.raises(ValueError, match="capacity"):
+        run(h)
+    assert_no_dispatch(h)
+
+
+def test_master_queue_failure_cleans_queues_and_never_starts_manager(harness):
+    h = harness
+    h.q.queue_prompt_payload.side_effect = ValueError("invalid master graph")
+    with pytest.raises(ValueError, match="invalid master"):
+        run(h)
+    assert h.ps.distributed_pending_jobs == {}
+    h.manager.start.assert_not_awaited()
+    h.q.dispatch_worker_prompt.assert_not_awaited()
+
+
+def test_manager_setup_failure_emits_one_terminal_per_collector_participant(harness):
+    h = harness
+    prompt = collector_prompt()
+    prompt["audio"] = node("GeneratedAudio")
+    prompt["collector"]["inputs"]["audio"] = ["audio", 0]
+    h.manager.start.side_effect = ValueError("journal failure")
+    with pytest.raises(ValueError, match="journal failure"):
+        run(h, prompt)
+    h.manager.cancel.assert_awaited_once()
+    queued = next(iter(h.ps.distributed_pending_jobs.values()))
+    assert queued.qsize() == 2  # two participants, not four per-socket completions
+    messages = [queued.get_nowait(), queued.get_nowait()]
+    assert {m["worker_id"] for m in messages} == {"serverless:pool:1", "serverless:pool:2"}
+    assert all(m["is_last"] is True and m["serverless_error"] for m in messages)
+
+
+@pytest.mark.parametrize("profile", [None, "legacy-image-v1"])
+def test_legacy_envelope_and_compiler_are_unchanged(harness, profile):
+    h = harness
+    if profile is None:
+        h.pool.pop("media_profile")
+    else:
+        h.pool["media_profile"] = profile
+    prompt = legacy_prompt()
+    original = copy.deepcopy(prompt)
+    with patch.object(h.compiler, "compile_media_prompt", side_effect=AssertionError("legacy must not use typed compiler")), \
+         patch.object(h.packer, "prepare_media_inputs", side_effect=AssertionError("legacy must not use typed packer")):
+        assert run(h, prompt)[2] == 2
+    for job in h.manager.start.call_args.args[2]:
+        assert "media" not in job
+        assert any(n["class_type"] == "SaveImage" for n in job["workflow"].values())
+        expected = {"input": {"workflow": job["workflow"], "images": job["images"]},
+                    "policy": {"executionTimeout": 30000, "ttl": 30000}}
+        assert h.client.serialize_request(job["workflow"], job["images"], 30) == json.dumps(
+            expected, allow_nan=False, ensure_ascii=False, separators=(",", ":")).encode()
+    assert prompt == original
+
+
+def test_legacy_no_worker_delegate_fallback_is_preserved(harness):
+    h = harness
+    h.config["workers"] = []
+    assert run(h, legacy_prompt())[2] == 0
+    assert h.q.queue_prompt_payload.call_args.args[0]["collector"]["inputs"]["delegate_only"] is False
+
+
+def test_native_audio_compile_pack_serialized_input_cpu(harness, media_paths):
+    """Exercise real PyAV + core PCM conversion, not a mocked packing result."""
+    if os.environ.get("_MEDIA_QUEUE_CORE_TEST") != "1":
+        core = os.environ.get("COMFYUI_SOURCE_ROOT")
+        if not core:
+            pytest.skip("Set COMFYUI_SOURCE_ROOT for CPU native audio queue integration")
+        env = dict(os.environ)
+        env.pop("PYTHONPATH", None)
+        env.pop("COMFYUI_SOURCE_ROOT", None)
+        scratch = Path.home() / ".hermes/cache/scratch"
+        scratch.mkdir(parents=True, exist_ok=True)
+        env.update(_MEDIA_QUEUE_CORE_TEST="1", CUDA_VISIBLE_DEVICES="", HIP_VISIBLE_DEVICES="",
+                   PYTHONDONTWRITEBYTECODE="1", TMPDIR=str(scratch))
+        bootstrap = """
+import sys
+sys.path.insert(0, sys.argv[1])
+import comfy.cli_args
+comfy.cli_args.args.cpu = True
+import torch
+
+
+def forbidden(*args, **kwargs):
+    raise AssertionError('CPU audio integration may not initialize a GPU')
+torch.cuda.is_available = lambda: False
+torch.cuda.device_count = lambda: 0
+torch.cuda.init = forbidden
+torch.cuda.current_device = forbidden
+torch.cuda._lazy_init = forbidden
+import pytest
+raise SystemExit(pytest.main([sys.argv[2], '-q', '-p', 'no:cacheprovider', '-k',
+                             'test_native_audio_compile_pack_serialized_input_cpu']))
+"""
+        result = subprocess.run([sys.executable, "-c", bootstrap, core, str(Path(__file__).resolve())],
+                                env=env, capture_output=True, text=True, timeout=120)
+        assert result.returncode == 0, result.stdout + result.stderr
+        assert "1 passed" in result.stdout
+        return
+    h = harness
+    h.pool["jobs_per_run"] = 1
+    path = media_paths / "tone.wav"
+    pcm = torch.tensor([-100, 0, 100, 32767, -32768, 7], dtype=torch.int16)
+    with wave.open(str(path), "wb") as stream:
+        stream.setnchannels(1)
+        stream.setsampwidth(2)
+        stream.setframerate(8000)
+        stream.writeframes(pcm.numpy().tobytes())
+    original_bytes = path.read_bytes()
+    prompt = collector_prompt()
+    prompt["audio"] = node("LoadAudio", audio="tone.wav")
+    prompt["collector"]["inputs"]["audio"] = ["audio", 0]
+    original = copy.deepcopy(prompt)
+    run(h, prompt)
+    job = h.manager.start.call_args.args[2][0]
+    wire = json.loads(h.client.serialize_request(job["workflow"], job["images"], 30, media=job["media"]))
+    artifacts = wire["input"]["media"]["artifacts"]
+    refs = wire["input"]["media"]["inputs"]
+    assert len(artifacts) == len(refs) == 1
+    binding = next(b for b in job["manifest"] if b["socket"] == "audio")
+    assert refs == [{"node_id": binding["node_id"], "input_name": "audio", "artifact_id": artifacts[0]["id"]}]
+    contract = importlib.import_module(PKG + ".serverless_bridge.contract")
+    codec = importlib.import_module(PKG + ".serverless_bridge.codec")
+    payload = asyncio.run(contract.resolve_artifact(artifacts[0]))
+    value = codec.decode_payload("AUDIO", artifacts[0]["descriptor"], payload)
+    assert value["sample_rate"] == 8000
+    assert value["waveform"].shape == (1, 1, 6)
+    assert torch.equal(value["waveform"], (pcm.to(torch.float32) / 32768).reshape(1, 1, -1))
+    assert path.read_bytes() == original_bytes
+    assert prompt == original
+    assert h.validator.validate_prepared_job(job) is True

tests/test_media_transfer_options.py

@@ -0,0 +1,96 @@
+"""Collector controls -> compiler -> endpoint validation, no network or GPU."""
+import copy
+import json
+
+import pytest
+
+from test_serverless_media_prompt import compiler, _prompt
+from test_media_handler import handler
+from test_collector_typed_media import collector_module, native_media, assert_exact
+
+FORMATS = ('legacy_png', 'png', 'ffv1', 'jpeg', 'webp', 'h264', 'h265', 'av1', 'exact')
+
+
+def compile_request(fmt='legacy_png', quality=90):
+    prompt = _prompt()
+    prompt['source:1'] = {'class_type': 'Source', 'inputs': {}}
+    prompt['collector:1']['inputs'].update(transfer_format=fmt, transfer_quality=quality,
+                                          audio=['source:1', 1])
+    original = copy.deepcopy(prompt)
+    workflow, outputs = compiler.compile_media_prompt(prompt, 'request:1')
+    assert prompt == original
+    return {'workflow': workflow, 'images': [], 'media': {
+        'version': 1, 'profile': 'comfy-media-v1', 'request_id': 'request:1',
+        'artifacts': [], 'inputs': [], 'outputs': outputs}}
+
+
+@pytest.mark.parametrize('fmt', FORMATS)
+def test_compiler_and_handler_preserve_image_policy_only(fmt):
+    data = compile_request(fmt, 73)
+    assert {b['socket'] for b in data['media']['outputs']} == {'images', 'audio'}
+    for binding in data['media']['outputs']:
+        inputs = data['workflow'][binding['node_id']]['inputs']
+        if binding['kind'] == 'IMAGE' and fmt != 'exact':
+            assert inputs['transfer_format'] == ('png' if fmt == 'legacy_png' else fmt)
+            assert inputs['transfer_quality'] == 73
+        else:
+            assert 'transfer_format' not in inputs and 'transfer_quality' not in inputs
+    handler._validate_request(data)
+
+
+def test_saved_collector_defaults_compile_to_png_not_exact():
+    workflow, bindings = compiler.compile_media_prompt(_prompt(), 'request:1')
+    inputs = workflow[bindings[0]['node_id']]['inputs']
+    assert (inputs['transfer_format'], inputs['transfer_quality']) == ('png', 90)
+
+
+@pytest.mark.parametrize('fmt,quality', [('JPEG',90),('',90),('jpeg',True),('jpeg',0),
+                                          ('jpeg',101),('jpeg','90'),(['source:1',0],90)])
+def test_compiler_rejects_invalid_or_linked_controls(fmt, quality):
+    with pytest.raises(ValueError):
+        compile_request(fmt, quality)
+
+
+@pytest.mark.parametrize('field,value', [('transfer_format','unknown'),('transfer_quality',0),
+                                        ('transfer_quality',True),('transfer_quality','90')])
+def test_handler_rejects_tampered_policy(field, value):
+    data = compile_request('jpeg')
+    binding = next(b for b in data['media']['outputs'] if b['kind'] == 'IMAGE')
+    data['workflow'][binding['node_id']]['inputs'][field] = value
+    with pytest.raises(ValueError):
+        handler._validate_request(data)
+
+
+@pytest.mark.parametrize('socket,kind', [('audio', 'AUDIO'), ('masks', 'MASK'), ('texts', 'STRING'), ('video', 'VIDEO')])
+def test_handler_rejects_image_controls_on_nonimage_bridge_exports(socket, kind):
+    # The bridge still supports all five kinds; these are direct bridge exports,
+    # not nonexistent sockets on DistributedCollector.
+    data = compile_request('exact')
+    binding = next(b for b in data['media']['outputs'] if b['kind'] == 'AUDIO')
+    binding.update(socket=socket, kind=kind)
+    inputs = data['workflow'][binding['node_id']]['inputs']
+    inputs.pop('audio')
+    inputs.update({socket: ['source:1', 1], 'binding_json': json.dumps(binding)})
+    handler._validate_request(data)
+    inputs['transfer_format'] = 'jpeg'
+    with pytest.raises(ValueError):
+        handler._validate_request(data)
+
+
+@pytest.mark.parametrize('fmt', FORMATS)
+def test_master_passthrough_stays_native_for_every_format(collector_module, fmt):
+    media = native_media()
+    actual = collector_module.DistributedCollectorNode().run(
+        images=[media['images']], audio=[media['audio']], pass_through=[True],
+        transfer_format=[fmt], transfer_quality=[73])
+    assert actual[0] is media['images']
+    assert actual[1] is media['audio']
+    assert_exact(actual[0], media['images'])
+
+
+def test_collector_widget_contract_keeps_saved_defaults_and_appends_exact(collector_module):
+    optional = collector_module.DistributedCollectorNode.INPUT_TYPES()['optional']
+    assert list(optional) == ['images', 'audio', 'transfer_format', 'transfer_quality']
+    assert tuple(optional['transfer_format'][0]) == FORMATS
+    assert optional['transfer_format'][1]['default'] == 'legacy_png'
+    assert optional['transfer_quality'][1]['default'] == 90

tests/test_media_video_codec.py

@@ -0,0 +1,562 @@
+"""Native ComfyUI VIDEO tests, isolated from the custom-node import graph.
+
+The outer pytest process starts one bounded CPU-only subprocess. Inside it the
+real pinned core is imported *after* cli_args.args.cpu=True; no fake VideoInput
+classes, media transcodes, providers, server, or GPU queries are needed.
+"""
+import os
+from pathlib import Path
+import subprocess
+import sys
+
+import pytest
+
+
+if os.environ.get("_DISTRIBUTED_VIDEO_CODEC_CORE_TEST") != "1":
+    def test_native_video_codec_in_cpu_subprocess():
+        configured = os.environ.get("COMFYUI_SOURCE_ROOT", os.environ.get("COMFYUI_VIDEO_TEST_ROOT"))
+        if configured is None:
+            pytest.skip("Set COMFYUI_SOURCE_ROOT for CPU native VIDEO acceptance")
+        core = Path(configured).expanduser().resolve()
+        assert configured and (core / "comfy_api/latest/_input_impl/video_types.py").is_file(), (
+            "Configured COMFYUI_SOURCE_ROOT (or COMFYUI_VIDEO_TEST_ROOT) must provide native video classes"
+        )
+        env = dict(os.environ)
+        scratch = Path(os.environ.get("TMPDIR", str(Path.home() / ".hermes/cache/scratch")))
+        scratch.mkdir(parents=True, exist_ok=True)
+        env["TMPDIR"] = str(scratch)
+        for name in ("PYTHONPATH", "COMFYUI_SOURCE_ROOT"):
+            env.pop(name, None)
+        env.update(CUDA_VISIBLE_DEVICES="", HIP_VISIBLE_DEVICES="",
+                   _DISTRIBUTED_VIDEO_CODEC_CORE_TEST="1", PYTHONDONTWRITEBYTECODE="1")
+        bootstrap = """
+import sys
+sys.path.insert(0, sys.argv[1])
+import comfy.cli_args
+comfy.cli_args.args.cpu = True
+import torch
+def forbidden(*args, **kwargs):
+    raise AssertionError('VIDEO tests must not query or initialize a GPU')
+# Third-party core imports probe availability even in CPU mode: answer without
+# executing a GPU query, then forbid even these probes during codec execution.
+torch.cuda.is_available = lambda: False
+torch.cuda.device_count = lambda: 0
+for name in ('init', '_lazy_init', 'current_device'):
+    setattr(torch.cuda, name, forbidden)
+import comfy_api.latest._input_impl.video_types
+for name in ('is_available', 'device_count'):
+    setattr(torch.cuda, name, forbidden)
+import pytest
+raise SystemExit(pytest.main([sys.argv[2], '-q', '-p', 'no:cacheprovider']))
+"""
+        result = subprocess.run(
+            [sys.executable, "-c", bootstrap, str(core), str(Path(__file__).resolve())],
+            env=env, text=True, capture_output=True, timeout=150,
+            cwd=Path(__file__).resolve().parents[1],
+        )
+        print(result.stdout)
+        assert result.returncode == 0, result.stdout + result.stderr
+else:
+    import copy
+    from fractions import Fraction
+    import importlib
+    import io
+    import json
+    import math
+    import types
+
+    import av
+    import numpy as np
+    import torch
+    from comfy_api.latest._input import VideoInput
+    from comfy_api.latest._input_impl.video_types import (
+        VideoFromComponents, VideoFromFile, VideoFromList,
+    )
+    from comfy_api.latest._util import VideoComponents
+
+    # Only this sibling package is synthetic; all native media types are real.
+    package_name = "_distributed_video_codec_test"
+    package = types.ModuleType(package_name)
+    package.__path__ = [str(Path(__file__).resolve().parents[1] / "serverless_bridge")]
+    sys.modules[package_name] = package
+    codec = importlib.import_module(package_name + ".video_codec")
+
+    def audio(dtype=torch.float64):
+        return {"waveform": torch.tensor([[[0.123456789, -0.0, -1.5, 2.0]]], dtype=dtype, device="cpu"),
+                "sample_rate": 48000}
+
+    def components(dtype=torch.float64, *, metadata=None, alpha=True, sound=True):
+        image = torch.arange(36, device="cpu", dtype=torch.float64).reshape(2, 2, 3, 3) / 997
+        image[0, 0, 0, 0] = -0.0
+        return VideoFromComponents(VideoComponents(
+            images=image.to(dtype), frame_rate=Fraction(30000, 1001),
+            audio=audio(dtype) if sound else None,
+            alpha=torch.full((2, 2, 3), 0.123456789, dtype=dtype, device="cpu") if alpha else None,
+            metadata=metadata,
+        ), bit_depth=10, color_space="HDR")
+
+    def exact_tensor(actual, expected):
+        assert actual.device.type == "cpu"
+        assert actual.dtype == expected.dtype
+        assert actual.shape == expected.shape
+        assert torch.equal(actual.contiguous().view(torch.uint8), expected.contiguous().view(torch.uint8))
+
+    def file_value(data=b"opaque original media bytes", **kwargs):
+        return VideoFromFile(io.BytesIO(data), **kwargs)
+
+    def forbid(*args, **kwargs):
+        raise AssertionError("Unexpected materialization/transcoding/allocation")
+
+    @pytest.mark.parametrize("dtype", [torch.float16, torch.bfloat16, torch.float32, torch.float64])
+    @pytest.mark.parametrize("alpha,sound", [(False, False), (True, False), (False, True), (True, True)])
+    def test_components_exact_roundtrip(dtype, alpha, sound, monkeypatch):
+        metadata = {"title": "雪 🎞", "flags": [True, None, -4, 0.125], "nested": {"empty": []}}
+        source = components(dtype, metadata=metadata, alpha=alpha, sound=sound)
+        expected = source.get_components()
+        monkeypatch.setattr(VideoFromComponents, "get_components", forbid)
+        monkeypatch.setattr(VideoFromComponents, "save_to", forbid)
+        descriptor, payload = codec.encode_video(source)
+        assert descriptor["encoding"] == "native-video-v1"
+        assert descriptor["representation"] == "components"
+        assert descriptor["frame_rate"] == {"numerator": 30000, "denominator": 1001}
+        # This is the actual JSON wire descriptor, not a Python-only shortcut.
+        descriptor = json.loads(json.dumps(descriptor, allow_nan=False))
+        result = codec.decode_video(descriptor, payload)
+        assert type(result) is VideoFromComponents
+        assert isinstance(result, VideoInput)
+        actual = vars(result)["_VideoFromComponents__components"]
+        exact_tensor(actual.images, expected.images)
+        assert actual.frame_rate == expected.frame_rate
+        assert type(actual.frame_rate) is Fraction
+        assert actual.metadata == metadata
+        assert actual.metadata is not metadata
+        assert vars(result)["_VideoFromComponents__bit_depth"] == 10
+        assert vars(result)["_VideoFromComponents__color_space"] == "HDR"
+        if alpha:
+            exact_tensor(actual.alpha, expected.alpha)
+        else:
+            assert actual.alpha is None
+        if sound:
+            exact_tensor(actual.audio["waveform"], expected.audio["waveform"])
+            assert actual.audio["sample_rate"] == expected.audio["sample_rate"]
+        else:
+            assert actual.audio is None
+        actual.images.fill_(99)
+        assert not torch.equal(actual.images, expected.images)
+        actual.metadata["nested"]["empty"].append("new")
+        assert descriptor["metadata"]["nested"]["empty"] == []
+        assert metadata["nested"]["empty"] == []
+
+    @pytest.mark.parametrize("dtype", [torch.float16, torch.bfloat16, torch.float32, torch.float64])
+    def test_native_bhw1_alpha_shape_is_preserved(dtype):
+        source = components(dtype)
+        original = source.get_components()
+        vars(source)["_VideoFromComponents__components"].alpha = original.alpha.unsqueeze(-1)
+        descriptor, payload = codec.encode_video(source)
+        result = codec.decode_video(descriptor, payload).get_components()
+        exact_tensor(result.alpha, original.alpha.unsqueeze(-1))
+        assert descriptor["alpha"]["descriptor"]["shape"] == [2, 2, 3, 1]
+
+    @pytest.mark.parametrize("start,duration,crop", [
+        (-1.25, -2.5, (-3, 5, 7, -9)), (0, 0, None), (-0.0, 0.0, (0, 0, 2, 2)),
+    ])
+    def test_file_preserves_bytes_cursor_signed_trim_crop_and_ownership(start, duration, crop, monkeypatch):
+        source = file_value(start_time=start, duration=duration, crop=crop)
+        stream = vars(source)["_VideoFromFile__file"]
+        stream.seek(7)
+        monkeypatch.setattr(VideoFromFile, "get_components", forbid)
+        monkeypatch.setattr(VideoFromFile, "save_to", forbid)
+        descriptor, payload = codec.encode_video(source)
+        assert stream.tell() == 7
+        assert payload == stream.getvalue()
+        assert descriptor["start_time"] == start
+        assert descriptor["duration"] == duration
+        result = codec.decode_video(json.loads(json.dumps(descriptor)), payload)
+        second = codec.decode_video(descriptor, payload)
+        attrs = vars(result)
+        assert type(result) is VideoFromFile
+        assert attrs["_VideoFromFile__start_time"] == start
+        assert type(attrs["_VideoFromFile__start_time"]) is type(start)
+        assert attrs["_VideoFromFile__duration"] == duration
+        assert attrs["_VideoFromFile__crop"] == crop
+        if type(start) is float:
+            assert math.copysign(1, attrs["_VideoFromFile__start_time"]) == math.copysign(1, start)
+        decoded = attrs["_VideoFromFile__file"]
+        assert type(decoded) is io.BytesIO
+        assert decoded is not stream
+        stream.close()
+        assert decoded.getvalue() == payload
+        decoded.write(b"changed")
+        assert vars(second)["_VideoFromFile__file"].getvalue() == payload
+
+    def test_real_tiny_pyav_video_is_still_readable_after_source_closed():
+        buffer = io.BytesIO()
+        with av.open(buffer, "w", format="matroska") as container:
+            stream = container.add_stream("ffv1", rate=Fraction(24, 1))
+            # Width 32 avoids a native fillborders filter limitation on tiny
+            # unaligned YUV frames; this is still a tiny CPU-only fixture.
+            stream.width = stream.height = 32
+            stream.pix_fmt = "yuv420p"
+            for value in (16, 64, 128):
+                frame = av.VideoFrame.from_ndarray(np.full((32, 32, 3), value, dtype=np.uint8), format="rgb24")
+                for packet in stream.encode(frame):
+                    container.mux(packet)
+            for packet in stream.encode():
+                container.mux(packet)
+        data = buffer.getvalue()
+        source = VideoFromFile(buffer, crop=(0, 0, 2, 2))
+        descriptor, payload = codec.encode_video(source)
+        expected = source.get_components()
+        result = codec.decode_video(descriptor, payload)
+        buffer.close()
+        assert payload == data
+        actual = result.get_components()
+        exact_tensor(actual.images, expected.images)
+        assert actual.frame_rate == expected.frame_rate
+        assert tuple(actual.images.shape) == (3, 2, 2, 3)
+
+    def test_segments_order_complete_audio_no_transcode_and_independent_buffers(monkeypatch):
+        values = [file_value(b"first", start_time=-2), file_value(b"second", duration=-4, crop=(1, 2, 3, 4))]
+        source = VideoFromList([VideoFromList(values[:1]), values[1]], complete_audio=audio())
+        monkeypatch.setattr(VideoFromFile, "get_components", forbid)
+        monkeypatch.setattr(VideoFromFile, "save_to", forbid)
+        monkeypatch.setattr(VideoFromList, "get_components", forbid)
+        monkeypatch.setattr(VideoFromList, "save_to", forbid)
+        descriptor, payload = codec.encode_video(source)
+        assert descriptor["representation"] == "segments"
+        assert payload.startswith(b"firstsecond")
+        assert len(descriptor["segments"]) == 2
+        decoded = codec.decode_video(json.loads(json.dumps(descriptor)), payload)
+        assert type(decoded) is VideoFromList
+        assert [vars(v)["_VideoFromFile__file"].getvalue() for v in decoded.videos] == [b"first", b"second"]
+        exact_tensor(decoded.complete_audio["waveform"], source.complete_audio["waveform"])
+        for old, new in zip(source.videos, decoded.videos):
+            assert old is not new
+            vars(old)["_VideoFromFile__file"].close()
+            assert not vars(new)["_VideoFromFile__file"].closed
+        assert vars(decoded.videos[0])["_VideoFromFile__start_time"] == -2
+        assert vars(decoded.videos[1])["_VideoFromFile__crop"] == (1, 2, 3, 4)
+
+    def test_path_roots_symlinks_and_no_path_in_wire(tmp_path, monkeypatch):
+        allowed = tmp_path / "allowed"
+        allowed.mkdir()
+        media = allowed / "private-name.mkv"
+        media.write_bytes(b"media bytes")
+        sibling = tmp_path / "allowed-sibling"
+        sibling.mkdir()
+        secret = sibling / "secret.mkv"
+        secret.write_bytes(b"private")
+        (allowed / "escape.mkv").symlink_to(secret)
+        (allowed / "inside.mkv").symlink_to(media)
+        descriptor, payload = codec.encode_video(VideoFromFile(str(media)), allowed_roots=[allowed])
+        assert payload == b"media bytes"
+        assert "private-name" not in json.dumps(descriptor)
+        assert codec.encode_video(VideoFromFile(str(allowed / "inside.mkv")), allowed_roots=[allowed])[1] == payload
+        for path in (secret, allowed / "escape.mkv", allowed, allowed / "missing"):
+            with pytest.raises(ValueError):
+                codec.encode_video(VideoFromFile(str(path)), allowed_roots=[allowed])
+        with pytest.raises(ValueError):
+            codec.encode_video(VideoFromFile(str(media)), allowed_roots=[])
+        import folder_paths
+        for name in ("get_input_directory", "get_output_directory", "get_temp_directory"):
+            monkeypatch.setattr(folder_paths, name, lambda: str(allowed))
+        assert codec.encode_video(VideoFromFile(str(media)))[1] == payload
+        with pytest.raises(ValueError):
+            codec.encode_video(VideoFromFile(str(secret)))
+        result = codec.decode_video(descriptor, payload)
+        media.unlink()
+        assert vars(result)["_VideoFromFile__file"].getvalue() == b"media bytes"
+
+    def test_path_special_files_and_symlink_swap_are_rejected(tmp_path, monkeypatch):
+        fifo = tmp_path / "fifo"
+        os.mkfifo(fifo)
+        with pytest.raises(ValueError):
+            codec.encode_video(VideoFromFile(str(fifo)), allowed_roots=[tmp_path])
+        allowed = tmp_path / "allowed"
+        allowed.mkdir()
+        actual = allowed / "subdir"
+        actual.mkdir()
+        media = actual / "clip.mkv"
+        media.write_bytes(b"allowed")
+        outside = tmp_path / "outside"
+        outside.mkdir()
+        (outside / "clip.mkv").write_bytes(b"forbidden")
+        real_resolve = codec.Path.resolve
+        swapped = False
+
+        def swap_after_resolution(path, *args, **kwargs):
+            nonlocal swapped
+            resolved = real_resolve(path, *args, **kwargs)
+            if path == media and not swapped:
+                swapped = True
+                actual.rename(allowed / "old")
+                actual.symlink_to(outside, target_is_directory=True)
+            return resolved
+
+        monkeypatch.setattr(codec.Path, "resolve", swap_after_resolution)
+        with pytest.raises(ValueError):
+            codec.encode_video(VideoFromFile(str(media)), allowed_roots=[allowed])
+        assert swapped
+
+    def test_file_size_budget_checked_before_read(tmp_path, monkeypatch):
+        path = tmp_path / "large.mkv"
+        path.write_bytes(b"12345")
+        monkeypatch.setattr(codec.os, "read", forbid)
+        with pytest.raises(ValueError, match="limit|budget"):
+            codec.encode_video(VideoFromFile(str(path)), max_bytes=4, allowed_roots=[tmp_path])
+
+    @pytest.mark.parametrize("limit", [True, False, -1, 1.5, None])
+    def test_invalid_limits(limit):
+        with pytest.raises(ValueError):
+            codec.encode_video(file_value(), max_bytes=limit)
+        with pytest.raises(ValueError):
+            codec.decode_video({}, b"", max_bytes=limit)
+
+    def test_aggregate_budget_and_exact_boundary():
+        value = VideoFromList([file_value(b"12"), file_value(b"345")])
+        descriptor, payload = codec.encode_video(value, max_bytes=5)
+        assert payload == b"12345"
+        codec.decode_video(descriptor, payload, max_bytes=5)
+        with pytest.raises(ValueError):
+            codec.encode_video(value, max_bytes=4)
+        with pytest.raises(ValueError):
+            codec.decode_video(descriptor, payload, max_bytes=4)
+        stream = vars(value.videos[0])["_VideoFromFile__file"]
+        stream.seek(1)
+        with pytest.raises(ValueError):
+            codec.encode_video(value.videos[0], max_bytes=1)
+        assert stream.tell() == 1
+
+    @pytest.mark.parametrize("kind", ["components", "file", "segments"])
+    def test_unknown_subclasses_never_materialized(kind):
+        bases = {"components": VideoFromComponents, "file": VideoFromFile, "segments": VideoFromList}
+        cls = type("UnknownVideo", (bases[kind],), {"get_components": forbid, "save_to": forbid})
+        value = object.__new__(cls)
+        with pytest.raises(ValueError, match="class|type|unsupported|Unsupported"):
+            codec.encode_video(value)
+        with pytest.raises(ValueError):
+            codec.encode_video(object())
+
+    @pytest.mark.parametrize("kind", ["components", "file", "segments"])
+    def test_bridge_rejects_subclasses_without_materializing(kind, monkeypatch):
+        nodes = importlib.import_module(package_name + ".nodes")
+        bases = {"components": VideoFromComponents, "file": VideoFromFile, "segments": VideoFromList}
+        cls = type("UnknownVideo", (bases[kind],), {"get_components": forbid, "save_to": forbid})
+        monkeypatch.setattr(torch.Tensor, "cpu", forbid)
+        with pytest.raises(ValueError, match="Unsupported native VIDEO class"):
+            nodes.encode_value("VIDEO", object.__new__(cls))
+
+    @pytest.mark.parametrize("dtype", [torch.float16, torch.bfloat16, torch.float32, torch.float64])
+    @pytest.mark.parametrize("kind", ["components", "segments", "file"])
+    def test_bridge_stages_cpu_tensors_without_materialization(dtype, kind, monkeypatch):
+        nodes = importlib.import_module(package_name + ".nodes")
+        if kind == "components":
+            source = components(dtype, metadata={"nested": ["雪", None, True]})
+            state = vars(source)["_VideoFromComponents__components"]
+            state.alpha = state.alpha.unsqueeze(-1)
+            tensors = [state.images, state.alpha, state.audio["waveform"]]
+        elif kind == "segments":
+            source = VideoFromList([file_value(start_time=-1.5, duration=-2, crop=(-1, 2, 3, 4))],
+                                   complete_audio=audio(dtype))
+            tensors = [source.complete_audio["waveform"]]
+        else:
+            source = file_value(start_time=-0.0, duration=-2, crop=(-1, 2, 3, 4))
+            tensors = []
+        for tensor in tensors:
+            tensor.requires_grad_(True)
+        expected = codec.encode_video(source)
+        detached, staged = [], []
+        real_detach, real_cpu = torch.Tensor.detach, torch.Tensor.cpu
+
+        def detach(tensor):
+            detached.append(id(tensor))
+            return real_detach(tensor)
+
+        def cpu(tensor, *args, **kwargs):
+            assert not tensor.requires_grad
+            staged.append(tensor)
+            return real_cpu(tensor, *args, **kwargs)
+
+        for cls in (VideoFromComponents, VideoFromFile, VideoFromList):
+            monkeypatch.setattr(cls, "get_components", forbid)
+            monkeypatch.setattr(cls, "save_to", forbid)
+        monkeypatch.setattr(torch.Tensor, "detach", detach)
+        monkeypatch.setattr(torch.Tensor, "cpu", cpu)
+        actual = nodes.encode_value("VIDEO", source)
+        assert len(staged) == len(tensors)
+        assert all(id(tensor) in detached for tensor in tensors)
+        assert all(tensor.requires_grad for tensor in tensors)
+        assert actual == expected
+        decoded = nodes.decode_value("VIDEO", *actual)
+        assert type(decoded) is type(source)
+        assert codec.encode_video(decoded) == expected
+        assert len(staged) == len(tensors)  # Pure codec never stages implicitly.
+
+    @pytest.mark.parametrize("field", ["images", "alpha", "audio", "complete_audio"])
+    def test_pure_video_rejects_non_cpu_storage_without_transfer(field, monkeypatch):
+        if field == "complete_audio":
+            source = VideoFromList([file_value()], complete_audio=audio())
+            target = source.complete_audio["waveform"]
+        else:
+            source = components()
+            target = getattr(vars(source)["_VideoFromComponents__components"], field)
+            if field == "audio":
+                target = target["waveform"]
+        # A property spy on real CPU tensors exercises the guard without a GPU
+        # allocation, subclass, device discovery or CUDA initialization.
+        real_device = torch.Tensor.device
+        monkeypatch.setattr(torch.Tensor, "device", property(
+            lambda tensor: types.SimpleNamespace(type="cuda") if tensor is target
+            else real_device.__get__(tensor)))
+        monkeypatch.setattr(torch.Tensor, "cpu", forbid)
+        with pytest.raises(ValueError, match="must be on CPU"):
+            codec.encode_video(source)
+
+    @pytest.mark.parametrize("field,bad", [
+        ("_VideoFromFile__start_time", True), ("_VideoFromFile__start_time", float("nan")),
+        ("_VideoFromFile__duration", float("inf")), ("_VideoFromFile__crop", (1, 2, True, 4)),
+        ("_VideoFromFile__crop", [1, 2, 3, 4]), ("_VideoFromFile__crop", (1, 2, 3)),
+        ("_VideoFromFile__file", b"not a BytesIO"),
+    ])
+    def test_invalid_native_file_attributes(field, bad):
+        value = file_value()
+        setattr(value, field, bad)
+        with pytest.raises(ValueError):
+            codec.encode_video(value)
+
+    def test_missing_attributes_and_invalid_component_state():
+        value = file_value()
+        del vars(value)["_VideoFromFile__duration"]
+        with pytest.raises(ValueError):
+            codec.encode_video(value)
+        for field, bad in [("frame_rate", 24.0), ("frame_rate", Fraction(0)),
+                           ("alpha", torch.ones(1, 1, 1, device="cpu"))]:
+            value = components()
+            setattr(vars(value)["_VideoFromComponents__components"], field, bad)
+            with pytest.raises(ValueError):
+                codec.encode_video(value)
+
+    @pytest.mark.parametrize("metadata", [
+        {"value": float("nan")}, {"value": float("inf")}, {1: "non-string key"},
+        {"value": (1, 2)}, {"value": object()}, {"large": "x" * 65536},
+    ])
+    def test_invalid_metadata_rejected_both_directions(metadata):
+        with pytest.raises(ValueError):
+            codec.encode_video(components(metadata=metadata))
+        descriptor, payload = codec.encode_video(components())
+        descriptor["metadata"] = metadata
+        with pytest.raises(ValueError):
+            codec.decode_video(descriptor, payload)
+
+    def test_metadata_depth_cycles_and_exact_size():
+        metadata = {}
+        current = metadata
+        for _ in range(40):
+            current["child"] = {}
+            current = current["child"]
+        for value in (metadata,):
+            with pytest.raises(ValueError):
+                codec.encode_video(components(metadata=value))
+        cyclic = {}
+        cyclic["self"] = cyclic
+        with pytest.raises(ValueError):
+            codec.encode_video(components(metadata=cyclic))
+        exact = {"x": "a" * (65536 - 8)}
+        assert len(json.dumps(exact, separators=(",", ":")).encode()) == 65536
+        descriptor, payload = codec.encode_video(components(metadata=exact))
+        assert codec.decode_video(descriptor, payload).get_components().metadata == exact
+
+    @pytest.mark.parametrize("field,bad", [
+        ("encoding", "other"), ("representation", "other"), ("byte_length", True),
+        ("byte_length", -1), ("byte_length", 12345), ("start_time", False),
+        ("duration", float("inf")), ("crop", [0, 0, True, 4]), ("crop", [1, 2, 3]),
+    ])
+    def test_malformed_file_descriptors(field, bad):
+        descriptor, payload = codec.encode_video(file_value())
+        descriptor[field] = bad
+        with pytest.raises(ValueError):
+            codec.decode_video(descriptor, payload)
+
+    @pytest.mark.parametrize("field,bad", [("offset", True), ("offset", -1), ("offset", 1),
+                                            ("length", True), ("length", -1), ("length", 2**80)])
+    def test_invalid_offsets_lengths(field, bad):
+        descriptor, payload = codec.encode_video(file_value())
+        descriptor["data"][field] = bad
+        with pytest.raises(ValueError):
+            codec.decode_video(descriptor, payload)
+
+    def test_truncation_trailing_gaps_overlap_and_unknown_fields():
+        value = VideoFromList([file_value(b"123"), file_value(b"456")])
+        original, payload = codec.encode_video(value)
+        for wire in (payload[:-1], payload + b"x"):
+            with pytest.raises(ValueError):
+                codec.decode_video(original, wire)
+        for offset in (2, 4):
+            descriptor = copy.deepcopy(original)
+            descriptor["segments"][1]["offset"] = offset
+            with pytest.raises(ValueError):
+                codec.decode_video(descriptor, payload)
+        descriptor = copy.deepcopy(original)
+        descriptor["surprise"] = "bad"
+        with pytest.raises(ValueError):
+            codec.decode_video(descriptor, payload)
+        descriptor = copy.deepcopy(original)
+        descriptor["segments"][0]["descriptor"]["byte_length"] += 1
+        with pytest.raises(ValueError):
+            codec.decode_video(descriptor, payload)
+        descriptor = copy.deepcopy(original)
+        descriptor["segments"][0]["descriptor"] = descriptor
+        with pytest.raises(ValueError):
+            codec.decode_video(descriptor, payload)
+
+    def test_all_descriptors_validated_before_tensor_allocation(monkeypatch):
+        original, payload = codec.encode_video(components())
+        monkeypatch.setattr(codec, "decode_payload", forbid)
+        cases = []
+        for key, value in [("shape", [2**80, 2, 3]), ("shape", [True, 2, 3]),
+                           ("dtype", "pickle"), ("layout", "BHWC")]:
+            descriptor = copy.deepcopy(original)
+            descriptor["alpha"]["descriptor"][key] = value
+            cases.append(descriptor)
+        descriptor = copy.deepcopy(original)
+        descriptor["audio"]["descriptor"]["sample_rate"] = True
+        cases.append(descriptor)
+        descriptor = copy.deepcopy(original)
+        descriptor["alpha"]["offset"] += 1
+        cases.append(descriptor)
+        for descriptor in cases:
+            with pytest.raises(ValueError):
+                codec.decode_video(descriptor, payload)
+
+    @pytest.mark.parametrize("field,bad", [("numerator", True), ("denominator", True),
+                                            ("denominator", 0), ("numerator", -3),
+                                            ("denominator", -1001)])
+    def test_invalid_fraction_fields(field, bad):
+        descriptor, payload = codec.encode_video(components())
+        descriptor["frame_rate"][field] = bad
+        with pytest.raises(ValueError):
+            codec.decode_video(descriptor, payload)
+
+    def test_segment_limit_and_exact_maximum(monkeypatch):
+        value = VideoFromList([file_value(b"x") for _ in range(64)])
+        descriptor, payload = codec.encode_video(value)
+        result = codec.decode_video(descriptor, payload)
+        assert len(result.videos) == 64
+        value.videos.append(file_value(b"x"))
+        with pytest.raises(ValueError):
+            codec.encode_video(value)
+        descriptor["segments"].append(copy.deepcopy(descriptor["segments"][-1]))
+        monkeypatch.setattr(codec, "decode_payload", forbid)
+        with pytest.raises(ValueError):
+            codec.decode_video(descriptor, payload)
+        value.videos = []
+        with pytest.raises(ValueError):
+            codec.encode_video(value)
+        value.videos = [components()]
+        with pytest.raises(ValueError):
+            codec.encode_video(value)
+
+    @pytest.mark.parametrize("descriptor,payload", [(None, b""), ([], b""), ({}, bytearray()), ({}, "")])
+    def test_invalid_top_level_types(descriptor, payload):
+        with pytest.raises(ValueError):
+            codec.decode_video(descriptor, payload)

tests/test_runpod_client.py

@@ -0,0 +1,167 @@
+import importlib.util
+import json
+import sys
+import unittest
+from pathlib import Path
+
+ROOT = Path(__file__).resolve().parents[1]
+
+def load_module():
+    spec = importlib.util.spec_from_file_location("dist_runpod_client", ROOT / "api/orchestration/runpod_client.py")
+    module = importlib.util.module_from_spec(spec)
+    sys.modules[spec.name] = module
+    spec.loader.exec_module(module)
+    return module
+
+client = load_module()
+
+
+def workflow_for_request_size(target_size, timeout_seconds):
+    """Build an ASCII workflow whose compact UTF-8 request body has exact size."""
+    workflow = {"1": {"class_type": "Text", "inputs": {"value": ""}}}
+
+    def body_size(candidate):
+        payload = {
+            "input": {"workflow": candidate, "images": []},
+            "policy": {
+                "executionTimeout": timeout_seconds * 1000,
+                "ttl": timeout_seconds * 1000,
+            },
+        }
+        return len(json.dumps(
+            payload, allow_nan=False, ensure_ascii=False, separators=(",", ":")
+        ).encode("utf-8"))
+
+    padding = target_size - body_size(workflow)
+    if padding < 0:
+        raise AssertionError("target is smaller than the request fixture")
+    workflow["1"]["inputs"]["value"] = "x" * padding
+    if body_size(workflow) != target_size:
+        raise AssertionError("request fixture did not reach the exact target size")
+    return workflow
+
+class Response:
+    def __init__(self, status=200, payload=None):
+        self.status, self.payload = status, payload
+    async def __aenter__(self):
+        return self
+    async def __aexit__(self, *args):
+        return False
+    async def json(self):
+        return self.payload
+
+class Session:
+    def __init__(self, responses):
+        self.responses, self.calls = list(responses), []
+    def request(self, method, url, **kwargs):
+        self.calls.append((method, url, kwargs))
+        response = self.responses.pop(0)
+        if isinstance(response, Exception):
+            raise response
+        return response
+
+class ClientTests(unittest.IsolatedAsyncioTestCase):
+    async def test_submit_exact_generic_contract_and_once(self):
+        session = Session([Response(payload={"id": "job-1", "status": "IN_QUEUE"})])
+        api = client.RunpodClient(session, "endpoint123", "secret-value")
+        result = await api.submit({"1": {"class_type": "EmptyImage", "inputs": {}}}, [], 1800)
+        self.assertEqual(result, "job-1")
+        method, url, kwargs = session.calls[0]
+        self.assertEqual((method, url), ("POST", "https://api.runpod.ai/v2/endpoint123/run"))
+        self.assertIsInstance(kwargs["data"], bytes)
+        sent_payload = json.loads(kwargs["data"])
+        self.assertEqual(set(sent_payload["input"]), {"workflow", "images"})
+        self.assertEqual(sent_payload["policy"]["ttl"], 1800000)
+        self.assertFalse(kwargs["allow_redirects"])
+        self.assertEqual(kwargs["headers"]["Authorization"], "Bearer secret-value")
+
+    async def test_submit_timeout_is_ambiguous_and_never_retried(self):
+        session = Session([TimeoutError("secret-value in unsafe error")])
+        api = client.RunpodClient(session, "endpoint123", "secret-value")
+        with self.assertRaises(client.AmbiguousSubmission) as ctx:
+            await api.submit({}, [], 30)
+        self.assertNotIn("secret-value", str(ctx.exception))
+        self.assertEqual(len(session.calls), 1)
+
+    async def test_http_submit_failure_not_retried_or_leaked(self):
+        for status in [401, 429, 500, 302]:
+            session = Session([Response(status, {"error": "secret-value"})])
+            api = client.RunpodClient(session, "endpoint123", "secret-value")
+            with self.assertRaises(client.RunpodError) as ctx:
+                await api.submit({}, [], 30)
+            self.assertEqual(len(session.calls), 1)
+            self.assertNotIn("secret-value", str(ctx.exception))
+
+    async def test_health_zero_workers_is_valid(self):
+        session = Session([Response(payload={"workers": {"idle": 0}, "jobs": {"inQueue": 0}})])
+        api = client.RunpodClient(session, "endpoint123", "key")
+        self.assertEqual((await api.health())["workers"]["idle"], 0)
+
+    async def test_invalid_provider_id_no_network(self):
+        session = Session([])
+        with self.assertRaises(ValueError):
+            client.RunpodClient(session, "https://attacker.invalid", "key")
+        api = client.RunpodClient(session, "endpoint123", "key")
+        with self.assertRaises(ValueError):
+            await api.status("../other")
+        self.assertFalse(session.calls)
+
+    async def test_status_and_cancel_only_target_exact_job(self):
+        session = Session([Response(payload={"status": "IN_PROGRESS"}), Response(payload={"status": "CANCELLED"})])
+        api = client.RunpodClient(session, "endpoint123", "key")
+        await api.status("job-1")
+        await api.cancel("job-1")
+        self.assertEqual([c[:2] for c in session.calls], [("GET", "https://api.runpod.ai/v2/endpoint123/status/job-1"), ("POST", "https://api.runpod.ai/v2/endpoint123/cancel/job-1")])
+
+    async def test_missing_id_is_ambiguous(self):
+        api = client.RunpodClient(Session([Response(payload={"status": "IN_QUEUE"})]), "endpoint123", "key")
+        with self.assertRaises(client.AmbiguousSubmission):
+            await api.submit({}, [], 30)
+
+    async def test_large_payload_rejected_before_submit(self):
+        session = Session([])
+        api = client.RunpodClient(session, "endpoint123", "key")
+        with self.assertRaises(ValueError):
+            await api.submit({}, [{"image": "x" * client.MAX_REQUEST_BYTES}], 30)
+        self.assertFalse(session.calls)
+
+    async def test_exact_10mb_serialized_request_is_sent_as_validated_bytes(self):
+        session = Session([Response(payload={"id": "boundary-job"})])
+        api = client.RunpodClient(session, "endpoint123", "key")
+        timeout_seconds = 37
+        workflow = workflow_for_request_size(10_000_000, timeout_seconds)
+
+        result = await api.submit(workflow, [], timeout_seconds)
+
+        self.assertEqual(result, "boundary-job")
+        sent = session.calls[0][2]["data"]
+        self.assertIsInstance(sent, bytes)
+        self.assertEqual(len(sent), 10_000_000)
+        self.assertEqual(json.loads(sent)["policy"]["executionTimeout"], 37_000)
+
+    async def test_one_byte_over_10mb_is_rejected_without_request_or_payload_leak(self):
+        session = Session([Response(payload={"id": "must-not-be-used"})])
+        api = client.RunpodClient(session, "endpoint123", "key")
+        workflow = workflow_for_request_size(10_000_001, 37)
+        workflow["1"]["inputs"]["value"] = "private graph value" + workflow["1"]["inputs"]["value"][19:]
+
+        with self.assertRaisesRegex(ValueError, "10,000,001") as ctx:
+            await api.submit(workflow, [], 37)
+
+        self.assertNotIn("private graph value", str(ctx.exception))
+        self.assertFalse(session.calls)
+
+    async def test_wire_size_counts_utf8_graph_bytes_and_policy_fields(self):
+        session = Session([Response(payload={"id": "unicode-job"})])
+        api = client.RunpodClient(session, "endpoint123", "key")
+        workflow = {"1": {"class_type": "Text", "inputs": {"value": "café雪"}}}
+
+        await api.submit(workflow, [], 41)
+
+        sent = session.calls[0][2]["data"]
+        expected = json.dumps({
+            "input": {"workflow": workflow, "images": []},
+            "policy": {"executionTimeout": 41_000, "ttl": 41_000},
+        }, allow_nan=False, ensure_ascii=False, separators=(",", ":")).encode("utf-8")
+        self.assertEqual(sent, expected)
+        self.assertGreater(len(sent), len(sent.decode("utf-8")))

tests/test_serverless_collector.py

@@ -0,0 +1,202 @@
+import asyncio
+
+import pytest
+import torch
+
+from test_collector_list_inputs import _load_collector_module
+
+
+class _FakeManager:
+    def __init__(self, *, pending=False, owned=True):
+        self.pending = pending
+        self.owned = owned
+        self.pending_calls = []
+        self.owned_calls = []
+        self.cancel_calls = []
+
+    def is_pending(self, collector_job_id, participant_id):
+        self.pending_calls.append((collector_job_id, participant_id))
+        return self.pending
+
+    def owns(self, collector_job_id, participant_id):
+        self.owned_calls.append((collector_job_id, participant_id))
+        return self.owned
+
+    async def cancel(self, worker_id=None, execution_id=None, collector_job_id=None):
+        self.cancel_calls.append({
+            "worker_id": worker_id,
+            "execution_id": execution_id,
+            "collector_job_id": collector_job_id,
+        })
+        return 1
+
+
+def _set_up_collector(module, job_id, queue, manager=None):
+    module.prompt_server.distributed_jobs_lock = asyncio.Lock()
+    module.prompt_server.distributed_pending_jobs = {job_id: queue}
+    if manager is not None:
+        module.prompt_server.distributed_serverless = manager
+    return module.DistributedCollectorNode()
+
+
+def test_serverless_pending_waits_past_normal_worker_timeout():
+    module = _load_collector_module()
+    job_id = "collector-pending"
+    participant = "serverless:pool:1"
+    queue = asyncio.Queue()
+    manager = _FakeManager(pending=True)
+    collector = _set_up_collector(module, job_id, queue, manager)
+    module.get_worker_timeout_seconds = lambda: 0.01
+    module.HEARTBEAT_INTERVAL = 0.1
+    probe_calls = []
+    module.probe_worker = lambda *args, **kwargs: probe_calls.append(args)
+
+    async def deliver_after_provider_work():
+        await asyncio.sleep(0.14)
+        queue.put_nowait({
+            "worker_id": participant,
+            "image_index": 0,
+            "tensor": torch.ones(1, 2, 2, 2),
+            "is_last": True,
+        })
+
+    async def run_collection():
+        producer = asyncio.create_task(deliver_after_provider_work())
+        result = await collector.execute(
+            images=None,
+            audio=None,
+            multi_job_id=job_id,
+            enabled_worker_ids=f'["{participant}"]',
+            delegate_only=True,
+        )
+        await producer
+        return result
+
+    images, _audio = asyncio.run(run_collection())
+
+    assert tuple(images.shape) == (1, 2, 2, 2)
+    assert manager.pending_calls
+    assert probe_calls == []
+    assert job_id not in module.prompt_server.distributed_pending_jobs
+
+
+def test_serverless_queue_error_raises_and_cancels_only_collector_binding():
+    module = _load_collector_module()
+    job_id = "collector-error"
+    queue = asyncio.Queue()
+    queue.put_nowait({
+        "worker_id": "serverless:pool:1",
+        "serverless_error": "Provider execution failed safely",
+        "is_last": True,
+    })
+    manager = _FakeManager()
+    collector = _set_up_collector(module, job_id, queue, manager)
+
+    with pytest.raises(RuntimeError, match="serverless.*Provider execution failed safely"):
+        asyncio.run(collector.execute(
+            images=None,
+            audio=None,
+            multi_job_id=job_id,
+            enabled_worker_ids='["serverless:pool:1"]',
+            delegate_only=True,
+        ))
+
+    assert manager.cancel_calls == [{
+        "worker_id": None,
+        "execution_id": None,
+        "collector_job_id": job_id,
+    }]
+    assert job_id not in module.prompt_server.distributed_pending_jobs
+
+
+def test_owned_serverless_slot_without_pending_result_fails_instead_of_partial_success():
+    module = _load_collector_module()
+    job_id = "collector-stopped"
+    queue = asyncio.Queue()
+    manager = _FakeManager(pending=False, owned=True)
+    collector = _set_up_collector(module, job_id, queue, manager)
+    module.get_worker_timeout_seconds = lambda: 0.01
+    module.HEARTBEAT_INTERVAL = 0.1
+    module.probe_worker = lambda *_args, **_kwargs: pytest.fail("serverless slots are not worker URLs")
+
+    with pytest.raises(RuntimeError, match="serverless.*stopped.*result|serverless.*result.*stopped"):
+        asyncio.run(collector.execute(
+            images=None,
+            audio=None,
+            multi_job_id=job_id,
+            enabled_worker_ids='["serverless:pool:1"]',
+            delegate_only=True,
+        ))
+
+    assert manager.owned_calls
+    assert manager.cancel_calls[0]["collector_job_id"] == job_id
+    assert job_id not in module.prompt_server.distributed_pending_jobs
+
+
+def test_interruption_cancels_serverless_job_and_removes_collector_queue():
+    module = _load_collector_module()
+    job_id = "collector-interrupted"
+    queue = asyncio.Queue()
+    manager = _FakeManager(pending=True)
+    collector = _set_up_collector(module, job_id, queue, manager)
+    interrupted = module.comfy.model_management.InterruptProcessingException("stop")
+    module.comfy.model_management.throw_exception_if_processing_interrupted = lambda: (_ for _ in ()).throw(interrupted)
+
+    with pytest.raises(module.comfy.model_management.InterruptProcessingException, match="stop"):
+        asyncio.run(collector.execute(
+            images=None,
+            audio=None,
+            multi_job_id=job_id,
+            enabled_worker_ids='["serverless:pool:1"]',
+            delegate_only=True,
+        ))
+
+    assert manager.cancel_calls[0]["collector_job_id"] == job_id
+    assert job_id not in module.prompt_server.distributed_pending_jobs
+
+
+def test_serverless_tensor_shape_mismatch_raises_instead_of_silent_fallback():
+    module = _load_collector_module()
+    job_id = "collector-shape-mismatch"
+    queue = asyncio.Queue()
+    queue.put_nowait({
+        "worker_id": "serverless:pool:1",
+        "image_index": 0,
+        "tensor": torch.ones(1, 3, 2, 2),
+        "is_last": True,
+    })
+    manager = _FakeManager()
+    collector = _set_up_collector(module, job_id, queue, manager)
+    local_images = torch.zeros(1, 2, 2, 2)
+
+    with pytest.raises(RuntimeError, match="serverless.*shape|shape.*serverless"):
+        asyncio.run(collector.execute(
+            images=local_images,
+            audio=None,
+            multi_job_id=job_id,
+            enabled_worker_ids='["serverless:pool:1"]',
+        ))
+
+    assert manager.cancel_calls[0]["collector_job_id"] == job_id
+    assert job_id not in module.prompt_server.distributed_pending_jobs
+
+
+def test_regular_worker_timeout_keeps_partial_behavior_without_creating_manager():
+    module = _load_collector_module()
+    job_id = "collector-regular-timeout"
+    queue = asyncio.Queue()
+    collector = _set_up_collector(module, job_id, queue)
+    module.get_worker_timeout_seconds = lambda: 0.01
+    module.HEARTBEAT_INTERVAL = 0.1
+    local_images = torch.zeros(1, 2, 2, 2)
+    collector._get_serverless_manager = lambda *_args, **_kwargs: pytest.fail("regular collection must not create manager")
+
+    images, _audio = asyncio.run(collector.execute(
+        images=local_images,
+        audio=None,
+        multi_job_id=job_id,
+        enabled_worker_ids='["worker-a"]',
+    ))
+
+    assert torch.equal(images, local_images)
+    assert job_id not in module.prompt_server.distributed_pending_jobs

tests/test_serverless_config.py

@@ -0,0 +1,436 @@
+import importlib.util
+import json
+import os
+import stat
+import sys
+import tempfile
+import threading
+import types
+import unittest
+from pathlib import Path
+from unittest import mock
+
+
+def _load_serverless_config_module(*, without_fcntl=False):
+    module_path = Path(__file__).resolve().parents[1] / "utils" / "serverless_config.py"
+    module_name = (
+        "dist_serverless_config_no_fcntl_test_module"
+        if without_fcntl
+        else "dist_serverless_config_test_module"
+    )
+    package_name = module_name + "_pkg"
+    package = types.ModuleType(package_name)
+    package.__path__ = [str(module_path.parent)]
+    sys.modules[package_name] = package
+    module_name = package_name + ".serverless_config"
+    sys.modules.pop(module_name, None)
+    spec = importlib.util.spec_from_file_location(module_name, module_path)
+    assert spec is not None and spec.loader is not None
+    module = importlib.util.module_from_spec(spec)
+    sys.modules[module_name] = module
+    if without_fcntl:
+        with mock.patch.dict(sys.modules, {"fcntl": None}):
+            spec.loader.exec_module(module)
+    else:
+        spec.loader.exec_module(module)
+    return module
+
+
+serverless_config = _load_serverless_config_module()
+
+
+class ValidateServerlessWorkerTests(unittest.TestCase):
+    def test_returns_copy_with_runpod_defaults(self):
+        worker = {"id": "worker-1", "endpoint_id": "ep_123-abc", "name": "GPU"}
+
+        result = serverless_config.validate_serverless_worker(worker)
+
+        self.assertEqual(
+            result,
+            {
+                "id": "worker-1",
+                "endpoint_id": "ep_123-abc",
+                "name": "GPU",
+                "type": "serverless",
+                "provider": "runpod",
+                "jobs_per_run": 1,
+                "timeout_seconds": 1800,
+            },
+        )
+        self.assertIsNot(result, worker)
+        self.assertEqual(worker, {"id": "worker-1", "endpoint_id": "ep_123-abc", "name": "GPU"})
+
+    def test_omitted_media_profile_remains_absent_and_means_legacy(self):
+        result = serverless_config.validate_serverless_worker(
+            {"id": "worker-1", "endpoint_id": "ep-123"}
+        )
+        self.assertNotIn("media_profile", result)
+        self.assertEqual(result.get("media_profile", "legacy-image-v1"), "legacy-image-v1")
+
+    def test_explicit_media_profiles_are_retained(self):
+        for profile in ("legacy-image-v1", "comfy-media-v1"):
+            with self.subTest(profile=profile):
+                result = serverless_config.validate_serverless_worker(
+                    {"id": "worker-1", "endpoint_id": "ep-123", "media_profile": profile}
+                )
+                self.assertEqual(result["media_profile"], profile)
+
+    def test_media_profile_rejects_unknown_nonstring_and_null_values(self):
+        for profile in ("", "auto", "comfy-media-v2", " comfy-media-v1", "COMFY-MEDIA-V1", None, True, 1, [], {}):
+            with self.subTest(profile=profile):
+                with self.assertRaisesRegex(ValueError, "media_profile"):
+                    serverless_config.validate_serverless_worker(
+                        {"id": "worker-1", "endpoint_id": "ep-123", "media_profile": profile}
+                    )
+
+    def test_accepts_inclusive_numeric_bounds(self):
+        for jobs, timeout in ((1, 30), (16, 86400)):
+            with self.subTest(jobs=jobs, timeout=timeout):
+                result = serverless_config.validate_serverless_worker(
+                    {
+                        "id": "worker-1",
+                        "endpoint_id": "ep-123",
+                        "jobs_per_run": jobs,
+                        "timeout_seconds": timeout,
+                    }
+                )
+                self.assertEqual(result["jobs_per_run"], jobs)
+                self.assertEqual(result["timeout_seconds"], timeout)
+
+    def test_legacy_concurrency_is_accepted_and_dropped(self):
+        # Configs saved before Images per run replaced the separate setting.
+        for legacy in (1, 2, 16, 17, True, "3"):
+            with self.subTest(legacy=legacy):
+                result = serverless_config.validate_serverless_worker(
+                    {"id": "worker-1", "endpoint_id": "ep-123", "jobs_per_run": 2, "concurrency": legacy}
+                )
+                self.assertNotIn("concurrency", result)
+                self.assertEqual(result["jobs_per_run"], 2)
+
+    def test_endpoint_url_is_rejected_with_a_plain_message(self):
+        with self.assertRaisesRegex(ValueError, "not a URL"):
+            serverless_config.validate_serverless_worker(
+                {"id": "worker-1", "endpoint_id": "https://api.runpod.ai/v2/abc123/run"}
+            )
+
+    def test_rejects_missing_or_empty_id_and_endpoint(self):
+        base: dict[str, object] = {"id": "worker-1", "endpoint_id": "ep-123"}
+        for field, value in (("id", None), ("id", "  "), ("endpoint_id", ""), ("endpoint_id", "https://api.runpod.io/ep-123")):
+            with self.subTest(field=field, value=value):
+                worker = dict(base)
+                worker[field] = value
+                with self.assertRaises(ValueError):
+                    serverless_config.validate_serverless_worker(worker)
+
+        worker = dict(base)
+        del worker["id"]
+        with self.assertRaises(ValueError):
+            serverless_config.validate_serverless_worker(worker)
+
+    def test_endpoint_id_must_be_ascii_opaque_identifier(self):
+        for endpoint_id in ("ep/123", "ep 123", "ép-123", "ep:123", "https://example.test"):
+            with self.subTest(endpoint_id=endpoint_id):
+                with self.assertRaises(ValueError):
+                    serverless_config.validate_serverless_worker(
+                        {"id": "worker-1", "endpoint_id": endpoint_id}
+                    )
+
+    def test_endpoint_id_matches_runpod_client_length_limit(self):
+        valid_id = "e" * 200
+        result = serverless_config.validate_serverless_worker(
+            {"id": "worker-1", "endpoint_id": valid_id}
+        )
+        self.assertEqual(result["endpoint_id"], valid_id)
+        with self.assertRaises(ValueError):
+            serverless_config.validate_serverless_worker(
+                {"id": "worker-1", "endpoint_id": "e" * 201}
+            )
+
+    def test_enabled_must_be_boolean_when_supplied(self):
+        base = {"id": "worker-1", "endpoint_id": "ep-123"}
+        for enabled in (True, False):
+            with self.subTest(enabled=enabled):
+                self.assertIs(
+                    serverless_config.validate_serverless_worker(
+                        {**base, "enabled": enabled}
+                    )["enabled"],
+                    enabled,
+                )
+        for enabled in ("false", 0, 1, None):
+            with self.subTest(enabled=enabled):
+                with self.assertRaises(ValueError):
+                    serverless_config.validate_serverless_worker(
+                        {**base, "enabled": enabled}
+                    )
+
+    def test_rejects_unknown_provider_and_non_serverless_type(self):
+        base = {"id": "worker-1", "endpoint_id": "ep-123"}
+        for extra in ({"provider": "other"}, {"type": "local"}):
+            with self.subTest(extra=extra):
+                with self.assertRaises(ValueError):
+                    serverless_config.validate_serverless_worker({**base, **extra})
+
+    def test_numeric_fields_reject_bool_float_and_out_of_range(self):
+        base = {"id": "worker-1", "endpoint_id": "ep-123"}
+        invalid_values = {
+            "jobs_per_run": (True, 1.0, 0, 17),
+            "timeout_seconds": (True, 30.0, 29, 86401),
+        }
+        for field, values in invalid_values.items():
+            for value in values:
+                with self.subTest(field=field, value=value):
+                    with self.assertRaises(ValueError):
+                        serverless_config.validate_serverless_worker({**base, field: value})
+
+    def test_requires_worker_object(self):
+        for worker in (None, [], "worker"):
+            with self.subTest(worker=worker):
+                with self.assertRaises(ValueError):
+                    serverless_config.validate_serverless_worker(worker)
+
+
+class SecretStoreTests(unittest.TestCase):
+    def setUp(self):
+        self.tempdir = tempfile.TemporaryDirectory()
+        self.root = Path(self.tempdir.name)
+        self.path = self.root / "private" / "runpod-secrets.json"
+
+    def tearDown(self):
+        self.tempdir.cleanup()
+
+    def test_set_get_has_delete_and_persists_as_json_mapping(self):
+        store = serverless_config.SecretStore(self.path)
+        store.set("worker-1", "secret-value")
+
+        self.assertEqual(store.get("worker-1"), "secret-value")
+        self.assertTrue(store.has("worker-1"))
+        self.assertIsNone(store.get("missing"))
+        self.assertFalse(store.has("missing"))
+        self.assertEqual(json.loads(self.path.read_text(encoding="utf-8")), {"worker-1": "secret-value"})
+        self.assertEqual(serverless_config.SecretStore(self.path).get("worker-1"), "secret-value")
+
+        store.delete("worker-1")
+        self.assertIsNone(store.get("worker-1"))
+        self.assertFalse(store.has("worker-1"))
+
+    def test_rejects_api_keys_with_newlines_or_over_4096_characters(self):
+        store = serverless_config.SecretStore(self.path)
+        for key in ("key\nwith-line-feed", "key\rwith-carriage-return", "k" * 4097):
+            with self.subTest(key_length=len(key)):
+                with self.assertRaises(serverless_config.SecretStoreError):
+                    store.set("worker-1", key)
+        store.set("worker-1", "k" * 4096)
+        self.assertEqual(store.get("worker-1"), "k" * 4096)
+
+    def test_new_secret_file_and_directory_have_restrictive_modes(self):
+        store = serverless_config.SecretStore(self.path)
+        store.set("worker-1", "secret-value")
+
+        self.assertEqual(stat.S_IMODE(self.path.stat().st_mode), 0o600)
+        self.assertEqual(stat.S_IMODE(self.path.parent.stat().st_mode), 0o700)
+
+    def test_missing_store_reads_as_empty_without_creating_files(self):
+        store = serverless_config.SecretStore(self.path)
+
+        self.assertIsNone(store.get("worker-1"))
+        self.assertFalse(store.has("worker-1"))
+        self.assertFalse(self.path.parent.exists())
+
+    def test_corrupt_store_fails_closed_without_overwriting_or_leaking_contents(self):
+        self.path.parent.mkdir()
+        corrupt = b'{"partial": "NEVER-LEAK-THIS-secret'
+        self.path.write_bytes(corrupt)
+        store = serverless_config.SecretStore(self.path)
+
+        for operation in (
+            lambda: store.get("worker-1"),
+            lambda: store.set("worker-1", "replacement-secret"),
+            lambda: store.delete("worker-1"),
+        ):
+            with self.subTest(operation=operation):
+                with self.assertRaises(Exception) as raised:
+                    operation()
+                self.assertNotIn("NEVER-LEAK-THIS", str(raised.exception))
+                self.assertNotIn("replacement-secret", str(raised.exception))
+                self.assertEqual(self.path.read_bytes(), corrupt)
+
+    def test_rejects_non_mapping_or_non_string_store_values_without_leaking_data(self):
+        for corrupt in (b'["opaque-stored-value"]', b'{"worker": 23}', b'{"": "opaque-stored-value"}'):
+            with self.subTest(corrupt=corrupt):
+                self.path.parent.mkdir(exist_ok=True)
+                self.path.write_bytes(corrupt)
+                with self.assertRaises(Exception) as raised:
+                    serverless_config.SecretStore(self.path).get("worker")
+                self.assertNotIn("opaque-stored-value", str(raised.exception))
+
+    def test_rejects_symlink_file_and_preserves_link_target(self):
+        target = self.root / "target.json"
+        target.write_text('{"worker-1": "target-secret"}', encoding="utf-8")
+        self.path.parent.mkdir()
+        self.path.symlink_to(target)
+        store = serverless_config.SecretStore(self.path)
+
+        with self.assertRaises(Exception):
+            store.get("worker-1")
+        with self.assertRaises(Exception):
+            store.set("worker-1", "replacement-secret")
+        self.assertEqual(target.read_text(encoding="utf-8"), '{"worker-1": "target-secret"}')
+
+    def test_rejects_symlinked_parent_directory(self):
+        real_dir = self.root / "real"
+        real_dir.mkdir()
+        link_dir = self.root / "linked"
+        link_dir.symlink_to(real_dir, target_is_directory=True)
+        store = serverless_config.SecretStore(link_dir / "secrets.json")
+
+        with self.assertRaises(Exception):
+            store.set("worker-1", "secret-value")
+        self.assertFalse((real_dir / "secrets.json").exists())
+
+    def test_rejects_parent_traversal_in_store_path(self):
+        store = serverless_config.SecretStore(self.root / "nested" / ".." / "secrets.json")
+
+        with self.assertRaises(Exception):
+            store.set("worker-1", "secret-value")
+        self.assertFalse((self.root / "secrets.json").exists())
+
+    def test_threaded_updates_from_independent_instances_are_not_lost(self):
+        workers = 12
+        barrier = threading.Barrier(workers)
+        errors = []
+
+        def write_one(index):
+            try:
+                barrier.wait(timeout=5)
+                serverless_config.SecretStore(self.path).set(f"worker-{index}", f"key-{index}")
+            except Exception as exc:  # collected to fail in the test thread
+                errors.append(exc)
+
+        threads = [threading.Thread(target=write_one, args=(i,)) for i in range(workers)]
+        for thread in threads:
+            thread.start()
+        for thread in threads:
+            thread.join(timeout=10)
+
+        self.assertFalse(any(thread.is_alive() for thread in threads), "writer thread did not finish")
+        self.assertEqual(errors, [])
+        saved = json.loads(self.path.read_text(encoding="utf-8"))
+        self.assertEqual(saved, {f"worker-{i}": f"key-{i}" for i in range(workers)})
+
+
+class PublicConfigTests(unittest.TestCase):
+    def test_redacts_credentials_preserves_other_fields_and_marks_only_serverless(self):
+        with tempfile.TemporaryDirectory() as tempdir:
+            store = serverless_config.SecretStore(Path(tempdir) / "private" / "keys.json")
+            store.set("serverless-1", "stored-api-key")
+            config = {
+                "workers": [
+                    {
+                        "id": "serverless-1",
+                        "type": "serverless",
+                        "provider": "runpod",
+                        "endpoint_id": "ep-123",
+                        "media_profile": "comfy-media-v1",
+                        "api_key": "inline-secret",
+                        "clear_api_key": "another-secret",
+                        "credential": "credential-secret",
+                        "nested": {"runpod_credential": "nested-secret", "safe": "kept"},
+                        "name": "remote GPU",
+                    },
+                    {"id": "serverless-2", "type": "serverless", "name": "not configured"},
+                    {"id": "local-1", "type": "local", "api_key": "local-secret"},
+                ],
+                "settings": {"mode": "preserved", "credentials_note": "remove-me"},
+            }
+            original = json.loads(json.dumps(config))
+
+            result = serverless_config.public_config(config, store)
+
+        first, second, local = result["workers"]
+        self.assertEqual(first["api_key_configured"], True)
+        self.assertEqual(second["api_key_configured"], False)
+        self.assertNotIn("api_key_configured", local)
+        self.assertNotIn("api_key", first)
+        self.assertNotIn("clear_api_key", first)
+        self.assertNotIn("credential", first)
+        self.assertNotIn("runpod_credential", first["nested"])
+        self.assertEqual(first["nested"]["safe"], "kept")
+        self.assertNotIn("api_key", local)
+        self.assertNotIn("credentials_note", result["settings"])
+        self.assertEqual(first["name"], "remote GPU")
+        self.assertEqual(first["media_profile"], "comfy-media-v1")
+        self.assertNotIn("media_profile", second)
+        self.assertEqual(config, original)
+        self.assertIsNot(result, config)
+        self.assertIsNot(result["workers"], config["workers"])
+
+    def test_redacts_normalized_secret_field_aliases(self):
+        config = {
+            "workers": [{
+                "id": "serverless-1",
+                "type": "serverless",
+                "apiKey": "camel-api-key",
+                "APIKEY": "upper-api-key",
+                "clearApiKey": "camel-clear-key",
+                "CLEAR_API_KEY": "upper-clear-key",
+                "runpod_CREDENTIAL": "provider-credential",
+                "nested": {"credential_note": "nested-credential"},
+                "safe": "preserved",
+            }]
+        }
+
+        result = serverless_config.public_config(
+            config, serverless_config.SecretStore(self.root_path())
+        )
+
+        worker = result["workers"][0]
+        self.assertEqual(worker, {
+            "id": "serverless-1",
+            "type": "serverless",
+            "nested": {},
+            "safe": "preserved",
+            "api_key_configured": False,
+        })
+
+    def test_handles_non_dict_config_without_mutating_it(self):
+        config = ["safe", {"api_key": "secret"}]
+        result = serverless_config.public_config(config, serverless_config.SecretStore(self.root_path()))
+        self.assertEqual(result, ["safe", {}])
+        self.assertEqual(config, ["safe", {"api_key": "secret"}])
+
+    def test_module_import_and_non_serverless_redaction_when_fcntl_unavailable(self):
+        module = _load_serverless_config_module(without_fcntl=True)
+        store = module.SecretStore(self.root_path())
+        config = {
+            "workers": [
+                {"id": "local-1", "type": "local", "api_key": "inline-secret"}
+            ]
+        }
+
+        result = module.public_config(config, store)
+
+        self.assertEqual(result, {"workers": [{"id": "local-1", "type": "local"}]})
+        self.assertEqual(config["workers"][0]["api_key"], "inline-secret")
+
+    def test_secret_operations_fail_clearly_without_fcntl(self):
+        module = _load_serverless_config_module(without_fcntl=True)
+        store = module.SecretStore(self.root_path())
+        unsupported_message = "Serverless credential storage is currently supported on POSIX masters only"
+
+        for operation in (
+            lambda: store.get("worker-1"),
+            lambda: store.set("worker-1", "api-key"),
+            lambda: store.delete("worker-1"),
+            lambda: store.has("worker-1"),
+        ):
+            with self.subTest(operation=operation):
+                with self.assertRaisesRegex(module.SecretStoreError, unsupported_message):
+                    operation()
+
+    @staticmethod
+    def root_path():
+        return Path(tempfile.gettempdir()) / "not-created-serverless-secret-store" / "secrets.json"
+
+
+if __name__ == "__main__":
+    unittest.main()

tests/test_serverless_integration.py

@@ -0,0 +1,183 @@
+"""Offline integration: real graph preparation + orchestration, fake provider."""
+import asyncio
+import aiohttp as real_aiohttp
+import importlib.util
+import json
+import sys
+import types
+import unittest
+from pathlib import Path
+from unittest.mock import AsyncMock, patch
+
+ROOT = Path(__file__).resolve().parents[1]
+PKG = "dist_serverless_integration"
+
+
+def module(name, **attrs):
+    value = types.ModuleType(name)
+    value.__dict__.update(attrs)
+    return value
+
+
+def load_modules():
+    modules = {}
+    for part, directory in [("", ROOT), (".api", ROOT / "api"), (".api.orchestration", ROOT / "api/orchestration"), (".utils", ROOT / "utils")]:
+        modules[PKG + part] = module(PKG + part, __path__=[str(directory)])
+    ps = types.SimpleNamespace(routes=types.SimpleNamespace(get=lambda _: lambda fn: fn, post=lambda _: lambda fn: fn))
+    modules["server"] = module("server", PromptServer=types.SimpleNamespace(instance=ps))
+    modules[PKG + ".utils.logging"] = module("logging", debug_log=lambda *a: None, log=lambda *a: None)
+    modules[PKG + ".utils.trace_logger"] = module("trace_logger", trace_debug=lambda *a: None)
+    modules[PKG + ".utils.config"] = module("config", load_config=lambda: {})
+    modules[PKG + ".utils.async_helpers"] = module("async_helpers", queue_prompt_payload=AsyncMock())
+    modules[PKG + ".utils.network"] = module("network", build_master_url=lambda **k: "http://master", build_master_callback_url=lambda *a, **k: "http://master")
+    modules[PKG + ".api.orchestration.dispatch"] = module("dispatch", dispatch_worker_prompt=AsyncMock(), select_active_workers=AsyncMock(), select_least_busy_worker=AsyncMock())
+    modules[PKG + ".api.orchestration.media_sync"] = module("media_sync", convert_paths_for_platform=lambda p, _: p, fetch_worker_path_separator=AsyncMock(), fetch_worker_has_node=AsyncMock(return_value=False), sync_worker_media=AsyncMock())
+    with patch.dict(sys.modules, modules):
+        spec = importlib.util.spec_from_file_location(PKG + ".api.queue_orchestration", ROOT / "api/queue_orchestration.py")
+        q = importlib.util.module_from_spec(spec)
+        spec.loader.exec_module(q)
+        # Keep imported package namespace only; restore shared Comfy runtime stub.
+        modules.update({k: v for k, v in sys.modules.items() if k.startswith(PKG)})
+    return q, modules, ps
+
+
+class IntegrationTests(unittest.IsolatedAsyncioTestCase):
+    async def asyncSetUp(self):
+        self.q, modules, self.ps = load_modules()
+        self.patch = patch.dict(sys.modules, modules)
+        self.patch.start()
+        self.addCleanup(self.patch.stop)
+        self.pool = {"id": "pool", "name": "Runpod", "type": "serverless", "provider": "runpod", "endpoint_id": "endpoint", "enabled": True, "jobs_per_run": 2, "concurrency": 1, "timeout_seconds": 30}
+        self.q.load_config = lambda: {"workers": [self.pool], "settings": {"master_delegate_only": True}}
+        self.q.select_active_workers = AsyncMock(side_effect=lambda workers, _ws, delegate, **kw: (workers, delegate))
+        self.q.queue_prompt_payload = AsyncMock(return_value={"prompt_id": "master-prompt", "number": 1})
+        self.manager = types.SimpleNamespace(start=AsyncMock(), cancel=AsyncMock(), ensure_capacity=lambda _: None)
+        sys.modules[PKG + ".api.orchestration.serverless_runtime"] = module("runtime", get_manager=lambda: self.manager)
+        sys.modules[PKG + ".api.orchestration.serverless_media"] = module("media", prepare_images=AsyncMock(return_value=[]))
+        self.prompt = {
+            "1": {"class_type": "EmptyImage", "inputs": {"width": 32, "height": 32, "batch_size": 1, "color": 0}},
+            "2": {"class_type": "DistributedCollector", "inputs": {"images": ["1", 0], "load_balance": False}},
+            "3": {"class_type": "SaveImage", "inputs": {"images": ["2", 0], "filename_prefix": "local"}},
+        }
+
+    async def test_real_compiler_slots_to_master_collector_without_callbacks(self):
+        result = await self.q.orchestrate_distributed_execution(self.prompt, None, "client")
+        self.assertEqual(result[2], 2)
+        self.q.dispatch_worker_prompt.assert_not_awaited()
+        self.manager.start.assert_awaited_once()
+        jobs = self.manager.start.call_args.args[2]
+        self.assertEqual([job["worker"]["id"] for job in jobs], ["serverless:pool:1", "serverless:pool:2"])
+        self.assertEqual(len({job["manifest"][0]["prefix"] for job in jobs}), 2)
+        for job in jobs:
+            graph = job["workflow"]
+            self.assertTrue(any(n["class_type"] == "SaveImage" for n in graph.values()))
+            self.assertFalse(any("Distributed" in n["class_type"] for n in graph.values()))
+            self.assertNotIn("http://master", json.dumps(graph))
+        master_graph = self.q.queue_prompt_payload.call_args.args[0]
+        self.assertEqual(json.loads(master_graph["2"]["inputs"]["enabled_worker_ids"]), ["serverless:pool:1", "serverless:pool:2"])
+
+    async def test_compile_pack_provider_decode_and_collect_real_components(self):
+        """Exercise the whole image path, mocking only the provider and Comfy queue."""
+        import base64
+        import io
+        import tempfile
+        import torch
+        from PIL import Image
+        from test_collector_list_inputs import _load_collector_module
+
+        sys.modules["aiohttp"] = real_aiohttp
+        for leaf in ("serverless_media", "serverless_runtime"):
+            name = PKG + ".api.orchestration." + leaf
+            spec = importlib.util.spec_from_file_location(name, ROOT / "api/orchestration" / (leaf + ".py"))
+            real_module = importlib.util.module_from_spec(spec)
+            sys.modules[name] = real_module
+            spec.loader.exec_module(real_module)
+        runtime = sys.modules[PKG + ".api.orchestration.serverless_runtime"]
+        with tempfile.TemporaryDirectory() as directory:
+            root = Path(directory)
+            image_buffer = io.BytesIO()
+            Image.new("RGB", (2, 3), color=(255, 0, 128)).save(image_buffer, format="PNG")
+            image_bytes = image_buffer.getvalue()
+            (root / "source.png").write_bytes(image_bytes)
+            paths = module("folder_paths", get_input_directory=lambda: directory,
+                           get_output_directory=lambda: directory, get_temp_directory=lambda: directory,
+                           get_annotated_filepath=lambda name: str(root / name))
+            submissions = []
+
+            class FakeProvider:
+                async def submit(self, graph, images, timeout_seconds):
+                    self.prefix = next(n["inputs"]["filename_prefix"] for n in graph.values() if n["class_type"] == "SaveImage")
+                    submissions.append((graph, images))
+                    self_image = next(n["inputs"]["image"] for n in graph.values() if n["class_type"] == "LoadImage")
+                    assert self_image == images[0]["name"]
+                    assert base64.b64decode(images[0]["image"]) == image_bytes
+                    return "provider-job-" + str(len(submissions))
+
+                async def status(self, _job_id):
+                    filename = self.prefix + "_00001_.png"
+                    return {"status": "COMPLETED", "output": {"images": [{
+                        "filename": filename, "artifact_path": filename, "artifact_type": "output",
+                        "type": "base64", "data": base64.b64encode(image_bytes).decode(),
+                    }]}}
+
+            manager = runtime.ServerlessManager(self.ps, None, root / "state" / "jobs.json", None)
+            manager._client = AsyncMock(side_effect=lambda _: FakeProvider())
+            self.ps.distributed_serverless = manager
+            self.prompt["1"] = {"class_type": "LoadImage", "inputs": {"image": "source.png"}}
+            collector_module = _load_collector_module()
+            collector_module.prompt_server = self.ps
+            sys.modules["aiohttp"] = real_aiohttp
+            with patch.object(runtime, "get_manager", return_value=manager), patch.dict(sys.modules, {"folder_paths": paths}):
+                result = await self.q.orchestrate_distributed_execution(self.prompt, None, "client")
+                master_graph = self.q.queue_prompt_payload.call_args.args[0]
+                inputs = master_graph["2"]["inputs"]
+                images, _audio = await asyncio.wait_for(collector_module.DistributedCollectorNode().execute(
+                    images=None, audio=None, multi_job_id=inputs["multi_job_id"],
+                    enabled_worker_ids=inputs["enabled_worker_ids"], delegate_only=True,
+                ), timeout=3)
+                await asyncio.gather(*manager.tasks.values())
+            self.assertEqual(result[2], 2)
+            self.assertEqual(len(submissions), 2)
+            self.assertEqual(tuple(images.shape), (2, 3, 2, 3))
+            self.assertEqual(images.dtype, torch.float32)
+            self.assertTrue(torch.equal(images[0], images[1]))
+            self.assertEqual(float(images[0, 0, 0, 0]), 1.0)
+            self.assertEqual([record["status"] for record in manager.ledger.all()], ["COMPLETED", "COMPLETED"])
+            self.assertEqual(self.ps.distributed_pending_jobs, {})
+            self.q.dispatch_worker_prompt.assert_not_awaited()
+
+    async def test_audio_rejected_before_any_dispatch_or_master_queue(self):
+        self.prompt["2"]["inputs"]["audio"] = ["9", 0]
+        self.prompt["9"] = {"class_type": "LoadAudio", "inputs": {"audio": "test.wav"}}
+        with self.assertRaises(ValueError):
+            await self.q.orchestrate_distributed_execution(self.prompt, None, "client")
+        self.manager.start.assert_not_awaited()
+        self.q.dispatch_worker_prompt.assert_not_awaited()
+        self.q.queue_prompt_payload.assert_not_awaited()
+
+    async def test_disabled_pool_excluded(self):
+        self.pool["enabled"] = False
+        self.assertEqual(self.q._resolve_enabled_workers({"workers": [self.pool]}), [])
+
+    async def test_master_validation_failure_submits_no_paid_job(self):
+        self.q.queue_prompt_payload.side_effect = ValueError("invalid local graph")
+        with self.assertRaisesRegex(ValueError, "invalid local graph"):
+            await self.q.orchestrate_distributed_execution(self.prompt, None, "client")
+        self.manager.start.assert_not_awaited()
+
+    async def test_oversized_serverless_request_fails_before_dispatch_or_master_queue(self):
+        media_module = sys.modules[PKG + ".api.orchestration.serverless_media"]
+        media_module.prepare_images.return_value = [{
+            "name": "private-image-name.png",
+            "image": "x" * 10_000_001,
+        }]
+
+        with self.assertRaisesRegex(ValueError, "10,000,") as ctx:
+            await self.q.orchestrate_distributed_execution(self.prompt, None, "client")
+
+        self.assertNotIn("private-image-name.png", str(ctx.exception))
+        self.assertNotIn("x" * 100, str(ctx.exception))
+        self.q.dispatch_worker_prompt.assert_not_awaited()
+        self.q.queue_prompt_payload.assert_not_awaited()
+        self.manager.start.assert_not_awaited()
+        self.assertEqual(self.ps.distributed_pending_jobs, {})

tests/test_serverless_lifecycle.py

@@ -0,0 +1,198 @@
+"""Billing-safe lifecycle regressions: local completion is not remote completion."""
+import asyncio
+import tempfile
+import types
+import unittest
+from pathlib import Path
+from unittest.mock import AsyncMock, patch
+
+from test_serverless_runtime import runtime, Secrets
+
+
+class LifecycleTests(unittest.IsolatedAsyncioTestCase):
+    async def asyncSetUp(self):
+        self.temp = tempfile.TemporaryDirectory()
+        self.addCleanup(self.temp.cleanup)
+        self.server = types.SimpleNamespace(distributed_pending_jobs={"collector": asyncio.Queue()}, distributed_jobs_lock=asyncio.Lock())
+        self.path = Path(self.temp.name) / "jobs.json"
+        self.manager = runtime.ServerlessManager(self.server, Secrets(), self.path, AsyncMock(return_value=object()))
+        self.worker = {"id": "serverless:pool:1", "pool_id": "pool", "endpoint_id": "ep", "timeout_seconds": 30, "type": "serverless"}
+        self.prepared = {"worker": self.worker, "workflow": {}, "images": [], "manifest": [{"job_id": "collector", "prefix": "x", "kind": "image"}]}
+        self.api = types.SimpleNamespace(submit=AsyncMock(return_value="job"), status=AsyncMock(), cancel=AsyncMock(return_value={"status": "CANCELLED"}))
+        self.client_patch = patch.object(runtime, "RunpodClient", return_value=self.api)
+        self.client_patch.start()
+        self.addCleanup(self.client_patch.stop)
+        self.decode_patch = patch.object(runtime, "decode_results", AsyncMock(return_value={"collector": ["tensor"]}))
+        self.decode_patch.start()
+        self.addCleanup(self.decode_patch.stop)
+        # These regressions exercise capacity holding, so use a limit of one.
+        self.limit_patch = patch.object(runtime, "PER_WORKER_JOB_LIMIT", 1)
+        self.limit_patch.start()
+        self.addCleanup(self.limit_patch.stop)
+
+    async def asyncTearDown(self):
+        await self.manager.cancel()
+
+    def second(self, **changes):
+        return dict(self.prepared, worker=dict(self.worker, id="serverless:pool:2", **changes))
+
+    def retained(self, status="CANCEL_REQUESTED", job_id: str | None = "old-job"):
+        self.manager.ledger.update("old", worker_id="pool", endpoint_id="ep", job_id=job_id,
+                                   status=status, created_at=1, updated_at=1)
+
+    async def test_capacity_is_held_until_slow_cancellation_is_confirmed(self):
+        cancel_started, release_cancel = asyncio.Event(), asyncio.Event()
+        self.api.submit.side_effect = ["job-1", "job-2"]
+        self.api.status.side_effect = [runtime.RunpodError("poll failed"), {"status": "COMPLETED", "output": {}}]
+        async def cancel(_job_id):
+            cancel_started.set()
+            await release_cancel.wait()
+            return {"status": "CANCELLED"}
+        self.api.cancel.side_effect = cancel
+        await self.manager.start("exec", "prompt", [self.prepared, self.second()])
+        try:
+            await asyncio.wait_for(cancel_started.wait(), 1)
+            await asyncio.sleep(0.01)
+            self.assertEqual(self.api.submit.await_count, 1)
+        finally:
+            release_cancel.set()
+            await self.manager.wait_for_execution("exec")
+        self.assertEqual(self.api.submit.await_count, 2)
+
+    async def test_unconfirmed_cancel_holds_capacity_until_reconciled(self):
+        self.api.status.side_effect = runtime.RunpodError("poll failed")
+        self.api.cancel.return_value = {"status": "IN_PROGRESS"}
+        await self.manager.start("first", "prompt", [self.prepared])
+        await self.manager.wait_for_execution("first")
+        self.assertEqual(self.manager.ledger.all()[0]["status"], "CANCEL_REQUESTED")
+        self.assertTrue(self.manager.pool_is_active("pool"))
+        await self.manager.start("second", "prompt", [self.second()])
+        try:
+            await asyncio.sleep(0.01)
+            self.assertEqual(self.api.submit.await_count, 1)
+            self.api.status.side_effect = None
+            self.api.status.return_value = {"status": "COMPLETED", "output": {}}
+            await self.manager.reconcile(worker_id="pool")
+            await asyncio.wait_for(self.manager.wait_for_execution("second"), 1)
+            self.assertEqual(self.api.submit.await_count, 2)
+        finally:
+            self.api.cancel.return_value = {"status": "CANCELLED"}
+            await self.manager.cancel()
+
+    async def test_restarted_manager_reserves_known_unresolved_jobs(self):
+        self.retained("IN_PROGRESS")
+        self.manager = runtime.ServerlessManager(self.server, Secrets(), self.path, AsyncMock(return_value=object()))
+        self.assertTrue(self.manager.pool_is_active("pool"))
+        self.assertFalse(self.manager.pool_is_active("other"))
+        await self.manager.start("exec", "prompt", [self.second(timeout_seconds=0.02)])
+        await self.manager.wait_for_execution("exec")
+        self.api.submit.assert_not_awaited()
+        self.assertEqual(self.manager.ledger.all()[-1]["status"], "TIMED_OUT")
+
+    async def test_recovery_closes_never_submitted_waiting_records(self):
+        self.retained("WAITING", job_id=None)
+        self.manager = runtime.ServerlessManager(self.server, Secrets(), self.path, AsyncMock(return_value=object()))
+        self.assertFalse(self.manager.pool_is_active("pool"))
+        self.assertEqual(self.manager.ledger.all()[0]["status"], "CANCELLED")
+        self.api.submit.assert_not_awaited()
+
+    async def test_recovery_keeps_interrupted_submission_unresolved(self):
+        self.retained("SUBMITTING", job_id=None)
+        self.manager = runtime.ServerlessManager(self.server, Secrets(), self.path, AsyncMock(return_value=object()))
+        self.assertTrue(self.manager.pool_is_active("pool"))
+        self.assertEqual(self.manager.ledger.all()[0]["status"], "UNKNOWN")
+        await self.manager.reconcile(worker_id="pool")
+        self.api.status.assert_not_awaited()
+        self.api.submit.assert_not_awaited()
+
+    async def test_ambiguous_submission_requires_explicit_recovery_acknowledgment(self):
+        self.api.submit.side_effect = runtime.AmbiguousSubmission("acceptance unknown")
+        await self.manager.start("exec", "prompt", [self.prepared])
+        await self.manager.wait_for_execution("exec")
+        record = self.manager.ledger.all()[0]
+        self.assertTrue(self.manager.pool_is_active("pool"))
+        with self.assertRaises(ValueError):
+            await self.manager.acknowledge_recovery("pool", record["record_id"], provider_checked=False)
+        with self.assertRaises(ValueError):
+            await self.manager.acknowledge_recovery("other", record["record_id"], provider_checked=True)
+        await self.manager.acknowledge_recovery("pool", record["record_id"], provider_checked=True)
+        self.assertFalse(self.manager.pool_is_active("pool"))
+        self.assertEqual(self.manager.ledger.all()[0]["status"], "RESOLVED")
+        self.api.submit.assert_awaited_once()
+
+    async def test_cancel_already_completed_provider_job_is_terminal(self):
+        self.retained()
+        self.api.cancel.return_value = {"status": "COMPLETED"}
+        await self.manager.cancel(worker_id="pool")
+        self.assertEqual(self.manager.ledger.all()[0]["status"], "COMPLETED")
+        self.assertFalse(self.manager.pool_is_active("pool"))
+
+    async def test_reconciliation_failure_never_releases_capacity(self):
+        self.retained()
+        self.api.status.side_effect = runtime.RunpodError("not reachable")
+        await self.manager.reconcile(worker_id="pool")
+        self.assertTrue(self.manager.pool_is_active("pool"))
+        self.assertEqual(self.manager.ledger.all()[0]["status"], "CANCEL_REQUESTED")
+
+    async def test_cancel_before_first_task_step_finalizes_local_record(self):
+        await self.manager.start("exec", "prompt", [self.prepared])
+        await self.manager.cancel(worker_id="pool")
+        self.assertEqual(self.manager.ledger.all()[0]["status"], "CANCELLED")
+        self.assertFalse(self.server.distributed_pending_jobs["collector"].empty())
+        self.assertFalse(self.manager.pool_is_active("pool"))
+        self.api.submit.assert_not_awaited()
+
+    async def test_live_record_cannot_be_acknowledged_or_reconciled(self):
+        polling = asyncio.Event()
+        async def status(_job_id):
+            polling.set()
+            await asyncio.Event().wait()
+        self.api.status.side_effect = status
+        await self.manager.start("exec", "prompt", [self.prepared])
+        await asyncio.wait_for(polling.wait(), 1)
+        record = self.manager.ledger.all()[0]
+        with self.assertRaises(ValueError):
+            await self.manager.acknowledge_recovery("pool", record["record_id"], provider_checked=True)
+        await self.manager.reconcile(worker_id="pool")
+        self.assertEqual(self.api.status.await_count, 1)
+        self.assertTrue(self.manager.pool_is_active("pool"))
+
+    async def test_repeated_cancel_does_not_interrupt_provider_cleanup(self):
+        polling, cancelling, release = asyncio.Event(), asyncio.Event(), asyncio.Event()
+        async def status(_job_id):
+            polling.set()
+            await asyncio.Event().wait()
+        async def cancel(_job_id):
+            cancelling.set()
+            await release.wait()
+            return {"status": "CANCELLED"}
+        self.api.status.side_effect = status
+        self.api.cancel.side_effect = cancel
+        await self.manager.start("exec", "prompt", [self.prepared])
+        await polling.wait()
+        first = asyncio.create_task(self.manager.cancel(worker_id="pool"))
+        await cancelling.wait()
+        second = asyncio.create_task(self.manager.cancel(worker_id="pool"))
+        await asyncio.sleep(0)
+        release.set()
+        await asyncio.gather(first, second)
+        self.api.cancel.assert_awaited_once_with("job")
+        self.assertEqual(self.manager.ledger.all()[0]["status"], "CANCELLED")
+
+    async def test_partial_start_cleanup_does_not_leave_waiting_protection(self):
+        broken = self.second()
+        del broken["worker"]["endpoint_id"]
+        with self.assertRaises(KeyError):
+            await self.manager.start("exec", "prompt", [self.prepared, broken])
+        await self.manager.cancel(execution_id="exec")
+        self.assertFalse(self.manager.pool_is_active("pool"))
+        self.api.submit.assert_not_awaited()
+
+    async def test_public_jobs_never_hides_unresolved_recovery_ids(self):
+        for index in range(22):
+            self.manager.ledger.update(str(index), worker_id="pool", endpoint_id="ep", status="UNKNOWN", updated_at=index)
+        for index in range(5):
+            self.manager.ledger.update("done-" + str(index), worker_id="pool", status="COMPLETED", updated_at=30 + index)
+        records = self.manager.public_jobs("pool")
+        self.assertEqual({record["record_id"] for record in records}, {str(index) for index in range(22)})
+        self.assertTrue(all(record["status"] == "UNKNOWN" for record in records))

tests/test_serverless_media.py

@@ -0,0 +1,447 @@
+import base64
+import importlib.util
+import io
+import sys
+import tempfile
+import types
+import unittest
+from pathlib import Path
+from unittest.mock import AsyncMock, patch
+
+from PIL import Image
+
+ROOT = Path(__file__).resolve().parents[1]
+MODULE_PATH = ROOT / "api" / "orchestration" / "serverless_media.py"
+
+
+def load_module():
+    spec = importlib.util.spec_from_file_location("dist_serverless_media_test", MODULE_PATH)
+    assert spec is not None and spec.loader is not None
+    module = importlib.util.module_from_spec(spec)
+    sys.modules[spec.name] = module
+    spec.loader.exec_module(module)
+    return module
+
+
+media = load_module()
+
+
+def load_runpod_module():
+    spec = importlib.util.spec_from_file_location(
+        "dist_serverless_media_runpod_test", ROOT / "api" / "orchestration" / "runpod_client.py"
+    )
+    assert spec is not None and spec.loader is not None
+    module = importlib.util.module_from_spec(spec)
+    sys.modules[spec.name] = module
+    spec.loader.exec_module(module)
+    return module
+
+
+runpod = load_runpod_module()
+
+
+def png_bytes(mode="RGB", size=(2, 1), color=None):
+    if color is None:
+        color = (10, 20, 30, 40) if mode == "RGBA" else (10, 20, 30)
+    buffer = io.BytesIO()
+    Image.new(mode, size, color).save(buffer, format="PNG")
+    return buffer.getvalue()
+
+
+class FolderPathsFixture:
+    def __init__(self, root):
+        self.input = root / "input"
+        self.output = root / "output"
+        self.temp = root / "temp"
+        for directory in (self.input, self.output, self.temp):
+            directory.mkdir(parents=True, exist_ok=True)
+
+    def get_input_directory(self):
+        return str(self.input)
+
+    def get_output_directory(self):
+        return str(self.output)
+
+    def get_temp_directory(self):
+        return str(self.temp)
+
+    def get_annotated_filepath(self, filename):
+        value = filename.replace("\\", "/")
+        # Match real ComfyUI folder_paths: root annotations are suffixes,
+        # not directory prefixes (output/foo.png is an input subdirectory).
+        for marker, root in (("input", self.input), ("output", self.output), ("temp", self.temp)):
+            suffix = f" [{marker}]"
+            if value.endswith(suffix):
+                return str(root / value[:-len(suffix)])
+        return str(self.input / value)
+
+
+class ServerlessMediaInputTests(unittest.IsolatedAsyncioTestCase):
+    async def asyncSetUp(self):
+        self.tempdir = tempfile.TemporaryDirectory()
+        self.paths = FolderPathsFixture(Path(self.tempdir.name))
+        self.folder_paths = types.ModuleType("folder_paths")
+        self.folder_paths.get_input_directory = self.paths.get_input_directory
+        self.folder_paths.get_output_directory = self.paths.get_output_directory
+        self.folder_paths.get_temp_directory = self.paths.get_temp_directory
+        self.folder_paths.get_annotated_filepath = self.paths.get_annotated_filepath
+        self.folder_patch = patch.dict(sys.modules, {"folder_paths": self.folder_paths})
+        self.folder_patch.start()
+        self.addCleanup(self.folder_patch.stop)
+        self.addCleanup(self.tempdir.cleanup)
+
+    async def test_packs_image_and_rewrites_duplicate_loadimage_inputs(self):
+        payload = png_bytes()
+        (self.paths.input / "cat.png").write_bytes(payload)
+        workflow = {
+            "1": {"class_type": "LoadImage", "inputs": {"image": "cat.png [input]", "other": "unchanged"}},
+            "2": {"class_type": "LoadImage", "inputs": {"image": "cat.png [input]"}},
+            "3": {"class_type": "SaveImage", "inputs": {"filename_prefix": "stable"}},
+        }
+
+        packed = await media.prepare_images(workflow)
+
+        self.assertEqual(len(packed), 1)
+        self.assertEqual(set(packed[0]), {"name", "image"})
+        self.assertRegex(packed[0]["name"], r"^[A-Za-z0-9_.-]+\.png$")
+        self.assertEqual(base64.b64decode(packed[0]["image"], validate=True), payload)
+        self.assertEqual(workflow["1"]["inputs"]["image"], packed[0]["name"])
+        self.assertEqual(workflow["2"]["inputs"]["image"], packed[0]["name"])
+        self.assertEqual(workflow["1"]["inputs"]["other"], "unchanged")
+        self.assertEqual(workflow["3"]["inputs"]["filename_prefix"], "stable")
+
+    async def test_resolves_annotated_input_output_and_temp_sources(self):
+        sources = ((self.paths.input, "one.png"), (self.paths.output, "two.jpg"), (self.paths.temp, "three.webp"))
+        workflow = {}
+        for index, (directory, filename) in enumerate(sources, 1):
+            (directory / filename).write_bytes(png_bytes())
+            root = directory.name
+            workflow[str(index)] = {"class_type": "LoadImage", "inputs": {"image": f"{filename} [{root}]"}}
+
+        packed = await media.prepare_images(workflow)
+
+        self.assertEqual(len(packed), 3)
+        self.assertEqual({base64.b64decode(item["image"]) for item in packed}, {png_bytes()})
+        self.assertEqual({node["inputs"]["image"] for node in workflow.values()}, {item["name"] for item in packed})
+
+    async def test_directory_prefix_is_not_a_root_annotation(self):
+        (self.paths.input / "output").mkdir()
+        payload = png_bytes(color=(100, 0, 0))
+        (self.paths.input / "output" / "same.png").write_bytes(payload)
+        (self.paths.output / "same.png").write_bytes(png_bytes())
+        workflow = {"1": {"class_type": "LoadImage", "inputs": {"image": "output/same.png"}}}
+        packed = await media.prepare_images(workflow)
+        self.assertEqual(base64.b64decode(packed[0]["image"]), payload)
+
+    async def test_rejects_unknown_root_annotation(self):
+        (self.paths.input / "cat.png").write_bytes(png_bytes())
+        workflow = {"1": {"class_type": "LoadImage", "inputs": {"image": "cat.png [abc123]"}}}
+        with self.assertRaisesRegex(media.ServerlessMediaError, "annotation"):
+            await media.prepare_images(workflow)
+
+    async def test_recognizes_custom_file_image_reference(self):
+        (self.paths.input / "source.bmp").write_bytes(png_bytes())
+        workflow = {"1": {"class_type": "CustomImageLoader", "inputs": {"file": "source.bmp"}}}
+
+        packed = await media.prepare_images(workflow)
+
+        self.assertEqual(len(packed), 1)
+        self.assertEqual(workflow["1"]["inputs"]["file"], packed[0]["name"])
+
+    async def test_rejects_missing_and_dynamic_loadimage_values(self):
+        cases = (
+            {"1": {"class_type": "LoadImage", "inputs": {}}},
+            {"1": {"class_type": "LoadImage"}},
+            {"1": {"class_type": "LoadImage", "inputs": None}},
+            {"1": {"class_type": "LoadImage", "inputs": {"image": ["2", 0]}}},
+        )
+        for workflow in cases:
+            with self.subTest(workflow=workflow):
+                with self.assertRaisesRegex(media.ServerlessMediaError, "LoadImage"):
+                    await media.prepare_images(workflow)
+
+    async def test_rejects_missing_source_and_unapproved_extension(self):
+        for value in ("absent.png", "animation.gif", "clip.mp4"):
+            with self.subTest(value=value):
+                workflow = {"1": {"class_type": "LoadImage", "inputs": {"image": value}}}
+                with self.assertRaises(media.ServerlessMediaError):
+                    await media.prepare_images(workflow)
+
+    async def test_rejects_unsupported_audio_or_video_references_without_reading(self):
+        for key, value in (("audio", "private.wav"), ("video", "private.mp4"), ("file", "private.wav")):
+            with self.subTest(key=key):
+                workflow = {"1": {"class_type": "CustomLoader", "inputs": {key: value}}}
+                with self.assertRaisesRegex(media.ServerlessMediaError, "audio|video|media"):
+                    await media.prepare_images(workflow)
+
+    async def test_rejects_parent_traversal_and_absolute_outside_roots(self):
+        outside = Path(self.tempdir.name).parent / "outside.png"
+        outside.write_bytes(png_bytes())
+        self.addCleanup(lambda: outside.unlink(missing_ok=True))
+        for value in ("../outside.png", str(outside)):
+            with self.subTest(value=value):
+                workflow = {"1": {"class_type": "LoadImage", "inputs": {"image": value}}}
+                with self.assertRaises(media.ServerlessMediaError):
+                    await media.prepare_images(workflow)
+
+    async def test_rejects_symlink_escape_from_allowed_root(self):
+        outside = Path(self.tempdir.name).parent / "linked-outside.png"
+        outside.write_bytes(png_bytes())
+        self.addCleanup(lambda: outside.unlink(missing_ok=True))
+        (self.paths.input / "link.png").symlink_to(outside)
+        workflow = {"1": {"class_type": "LoadImage", "inputs": {"image": "link.png"}}}
+
+        with self.assertRaises(media.ServerlessMediaError):
+            await media.prepare_images(workflow)
+
+    async def test_rejects_individual_and_aggregate_input_size_overages(self):
+        huge = self.paths.input / "large.jpg"
+        with huge.open("wb") as handle:
+            handle.truncate(media.MAX_INPUT_IMAGE_BYTES + 1)
+        workflow = {"1": {"class_type": "LoadImage", "inputs": {"image": "large.jpg"}}}
+        with self.assertRaisesRegex(media.ServerlessMediaError, "size"):
+            await media.prepare_images(workflow)
+
+        huge.unlink()
+        first = self.paths.input / "first.png"
+        second = self.paths.input / "second.png"
+        first.write_bytes(b"a" * (media.MAX_INPUT_TOTAL_BYTES // 2 + 1))
+        second.write_bytes(b"b" * (media.MAX_INPUT_TOTAL_BYTES // 2 + 1))
+        workflow = {
+            "1": {"class_type": "LoadImage", "inputs": {"image": "first.png"}},
+            "2": {"class_type": "LoadImage", "inputs": {"image": "second.png"}},
+        }
+        with self.assertRaisesRegex(media.ServerlessMediaError, "size"):
+            await media.prepare_images(workflow)
+
+
+    async def test_accepts_raw_image_over_old_7mib_when_complete_request_fits_10mb(self):
+        raw_size = 7 * 1024 * 1024 + 1
+        source = self.paths.input / "large-input.png"
+        source.write_bytes(b"x" * raw_size)
+        workflow = {"1": {"class_type": "LoadImage", "inputs": {"image": "large-input.png"}}}
+
+        packed = await media.prepare_images(workflow)
+        request_body = runpod.serialize_request(workflow, packed, 30)
+
+        self.assertGreater(raw_size, 7 * 1024 * 1024)
+        self.assertEqual(len(base64.b64decode(packed[0]["image"])), raw_size)
+        self.assertLessEqual(len(request_body), 10_000_000)
+
+
+class ServerlessMediaOutputTests(unittest.IsolatedAsyncioTestCase):
+    def manifest(self, prefix="render"):
+        return [{"collector_id": "collector-1", "job_id": "job-1", "node_id": "12", "kind": "image", "prefix": prefix}]
+
+    def artifact(self, filename, payload=None, **extra):
+        entry = {
+            "filename": filename,
+            "artifact_type": "output",
+            "artifact_path": f"/safe/output/{filename}",
+            "subfolder": "",
+            "type": "base64",
+            "data": base64.b64encode(payload if payload is not None else png_bytes()).decode("ascii"),
+        }
+        entry.update(extra)
+        return entry
+
+    async def test_decodes_rgb_to_normalized_float32_and_orders_by_counter(self):
+        output = {"images": [self.artifact("render_00009_.png"), self.artifact("render_00002_.png")]}
+
+        result = await media.collect_images(output, self.manifest())
+
+        self.assertEqual(list(result), ["job-1"])
+        self.assertEqual(len(result["job-1"]), 2)
+        self.assertTrue(all(image.dtype.name == "float32" for image in result["job-1"]))
+        self.assertEqual(result["job-1"][0].shape, (1, 2, 3))
+        self.assertAlmostEqual(float(result["job-1"][0][0, 0, 0]), 10 / 255)
+
+    async def test_groups_images_under_manifest_job_ids(self):
+        manifest = [
+            self.manifest("first")[0],
+            {"collector_id": "collector-2", "job_id": "job-2", "node_id": "14", "kind": "image", "prefix": "second"},
+        ]
+        output = {"images": [self.artifact("first_00001_.png"), self.artifact("second_00001_.png")]}
+
+        result = await media.collect_images(output, manifest)
+
+        self.assertEqual(set(result), {"job-1", "job-2"})
+        self.assertEqual(len(result["job-1"]), 1)
+        self.assertEqual(len(result["job-2"]), 1)
+
+    async def test_preserves_rgba_alpha(self):
+        output = {"images": [self.artifact("render_00001_.png", png_bytes(mode="RGBA"))]}
+
+        result = await media.collect_images(output, self.manifest())
+
+        self.assertEqual(result["job-1"][0].shape, (1, 2, 4))
+        self.assertAlmostEqual(float(result["job-1"][0][0, 0, 3]), 40 / 255)
+
+    async def test_signed_url_is_downloaded_only_through_injected_downloader(self):
+        output = {"images": [{
+            "filename": "render_00001_.png", "artifact_type": "output", "type": "gcs_url",
+            "url": "https://storage.example.test/signed",
+        }]}
+        fetch = AsyncMock(return_value=png_bytes())
+
+        result = await media.collect_images(output, self.manifest(), download=fetch)
+
+        self.assertEqual(result["job-1"][0].shape, (1, 2, 3))
+        fetch.assert_awaited_once_with("https://storage.example.test/signed")
+
+    async def test_ignores_unrelated_artifacts_but_requires_every_manifest_collector(self):
+        output = {"images": [
+            self.artifact("preview_00001_.png"),
+            self.artifact("render_00001_.png"),
+            {"filename": "mesh.obj", "type": "url", "url": "not a fetch target"},
+        ]}
+
+        result = await media.collect_images(output, self.manifest())
+
+        self.assertEqual(len(result["job-1"]), 1)
+        with self.assertRaisesRegex(media.ServerlessMediaError, "expected|missing"):
+            await media.collect_images({"images": [self.artifact("preview_00001_.png")]}, self.manifest())
+
+    async def test_accepts_compiler_length_flat_safe_prefix(self):
+        prefix = "serverless_" + "j" * 24 + "_" + "w" * 24 + "_" + "c" * 24 + "_image_" + "a" * 64
+        manifest = self.manifest(prefix)
+        output = {"images": [self.artifact(f"{prefix}_00001_.png")]}
+
+        result = await media.collect_images(output, manifest)
+
+        self.assertEqual(len(result["job-1"]), 1)
+
+    async def test_rejects_duplicate_manifest_prefix_and_unsafe_prefix(self):
+        duplicate = self.manifest() * 2
+        with self.assertRaises(media.ServerlessMediaError):
+            await media.collect_images({"images": []}, duplicate)
+        unsafe = self.manifest("../render")
+        with self.assertRaises(media.ServerlessMediaError):
+            await media.collect_images({"images": []}, unsafe)
+
+    async def test_rejects_duplicate_canonical_artifact_identity(self):
+        item = self.artifact("render_00001_.png")
+        with self.assertRaisesRegex(media.ServerlessMediaError, "[Dd]uplicate"):
+            await media.collect_images({"images": [item, dict(item)]}, self.manifest())
+
+    async def test_rejects_matching_non_output_artifact_and_non_numeric_counter(self):
+        wrong_type = self.artifact("render_00001_.png", artifact_type="temp")
+        with self.assertRaises(media.ServerlessMediaError):
+            await media.collect_images({"images": [wrong_type]}, self.manifest())
+        with self.assertRaises(media.ServerlessMediaError):
+            await media.collect_images({"images": [self.artifact("render_final.png")]}, self.manifest())
+
+    async def test_rejects_provider_errors_even_when_job_completed(self):
+        for output in (
+            {"error": "raw provider secret", "images": []},
+            {"errors": ["raw provider secret"], "images": []},
+            {"status": "COMPLETED", "output": {"error": "raw provider secret", "images": []}},
+        ):
+            with self.subTest(output=output):
+                with self.assertRaises(media.ServerlessMediaError) as ctx:
+                    await media.collect_images(output, self.manifest())
+                self.assertNotIn("raw provider secret", str(ctx.exception))
+
+    async def test_rejects_invalid_base64_and_malformed_image(self):
+        for item in (
+            self.artifact("render_00001_.png", data="!not base64!"),
+            self.artifact("render_00001_.png", data=base64.b64encode(b"not an image").decode()),
+        ):
+            with self.subTest(item=item):
+                with self.assertRaises(media.ServerlessMediaError):
+                    await media.collect_images({"images": [item]}, self.manifest())
+
+    async def test_validates_shape_consistency_within_collector(self):
+        output = {"images": [
+            self.artifact("render_00001_.png", png_bytes(size=(2, 1))),
+            self.artifact("render_00002_.png", png_bytes(size=(3, 1))),
+        ]}
+        with self.assertRaisesRegex(media.ServerlessMediaError, "shape|dimensions"):
+            await media.collect_images(output, self.manifest())
+
+    async def test_bounds_output_file_count_and_bytes(self):
+        too_many = {"images": [{}] * (media.MAX_OUTPUT_FILES + 1)}
+        with self.assertRaisesRegex(media.ServerlessMediaError, "files"):
+            await media.collect_images(too_many, self.manifest())
+
+        with patch.object(media, "MAX_OUTPUT_IMAGE_BYTES", 4):
+            output = {"images": [self.artifact("render_00001_.png", data=base64.b64encode(b"12345").decode())]}
+            with self.assertRaisesRegex(media.ServerlessMediaError, "size"):
+                await media.collect_images(output, self.manifest())
+
+    async def test_aggregate_payload_limit_is_checked_as_artifacts_are_read(self):
+        output = {"images": [
+            self.artifact("render_00001_.png", data=base64.b64encode(b"12").decode()),
+            self.artifact("render_00002_.png", data=base64.b64encode(b"34").decode()),
+        ]}
+        with patch.object(media, "MAX_OUTPUT_TOTAL_BYTES", 3):
+            with self.assertRaisesRegex(media.ServerlessMediaError, "size"):
+                await media.collect_images(output, self.manifest())
+
+    async def test_rejects_non_public_or_non_https_artifact_urls_before_downloading(self):
+        for url in ("http://example.test/a", "https://127.0.0.1/a", "https://user:pass@example.test/a"):
+            with self.subTest(url=url):
+                fetch = AsyncMock(return_value=png_bytes())
+                output = {"images": [{"filename": "render_00001_.png", "type": "url", "url": url}]}
+                with self.assertRaises(media.ServerlessMediaError):
+                    await media.collect_images(output, self.manifest(), download=fetch)
+                fetch.assert_not_awaited()
+
+    async def test_public_resolver_rejects_non_global_dns_answers(self):
+        resolver = media._PublicResolver()
+        resolver._getaddrinfo = AsyncMock(return_value=[(2, 1, 6, "", ("8.8.8.8", 443))])
+        self.assertEqual(len(await resolver.resolve("cdn.example.test", 443)), 1)
+        resolver._getaddrinfo = AsyncMock(return_value=[(2, 1, 6, "", ("10.0.0.5", 443))])
+        with self.assertRaises(OSError):
+            await resolver.resolve("cdn.example.test", 443)
+
+    async def test_default_download_revalidates_redirects_and_disables_environment_proxies(self):
+        class Content:
+            def __init__(self, chunks):
+                self.chunks = iter(chunks)
+            async def iter_chunked(self, _size):
+                for chunk in self.chunks:
+                    yield chunk
+
+        class Response:
+            def __init__(self, status, headers=None, chunks=()):
+                self.status, self.headers = status, headers or {}
+                self.content = Content(chunks)
+            async def __aenter__(self):
+                return self
+            async def __aexit__(self, *_args):
+                return False
+
+        class Session:
+            def __init__(self, responses):
+                self.responses, self.calls = list(responses), []
+            async def __aenter__(self):
+                return self
+            async def __aexit__(self, *_args):
+                return False
+            def get(self, url, **kwargs):
+                self.calls.append((url, kwargs))
+                return self.responses.pop(0)
+
+        session = Session([Response(302, {"Location": "https://cdn.example.test/final"}), Response(200, chunks=[b"abc", b"def"])])
+        with patch.object(media.aiohttp, "ClientSession", return_value=session) as create_session:
+            payload = await media.download_public_artifact("https://storage.example.test/signed")
+
+        self.assertEqual(payload, b"abcdef")
+        self.assertEqual([call[0] for call in session.calls], ["https://storage.example.test/signed", "https://cdn.example.test/final"])
+        self.assertTrue(all(call[1]["allow_redirects"] is False for call in session.calls))
+        self.assertFalse(create_session.call_args.kwargs["trust_env"])
+        self.assertIsInstance(create_session.call_args.kwargs["connector"]._resolver, media._PublicResolver)
+        self.assertNotIn("Authorization", session.calls[0][1].get("headers", {}))
+        await create_session.call_args.kwargs["connector"].close()
+
+        blocked = Session([Response(302, {"Location": "https://127.0.0.1/metadata"})])
+        with patch.object(media.aiohttp, "ClientSession", return_value=blocked) as blocked_session:
+            with self.assertRaises(media.ServerlessMediaError):
+                await media.download_public_artifact("https://storage.example.test/signed")
+        self.assertEqual(len(blocked.calls), 1)
+        await blocked_session.call_args.kwargs["connector"].close()
+
+
+if __name__ == "__main__":
+    unittest.main()

tests/test_serverless_media_prompt.py

@@ -0,0 +1,426 @@
+"""CPU-only graph/compiler tests; no ComfyUI package bootstrap is needed."""
+
+import copy
+import importlib
+import json
+import subprocess
+import sys
+import types
+from pathlib import Path
+
+import pytest
+
+
+ROOT = Path(__file__).resolve().parents[1]
+PACKAGE = "serverless_media_prompt_testpkg"
+# Real paths let sibling imports use production code, without running the
+# extension's ComfyUI-dependent __init__.py or API registration side effects.
+for suffix in ("", ".api", ".api.orchestration", ".serverless_bridge"):
+    package = types.ModuleType(PACKAGE + suffix)
+    package.__path__ = [str(ROOT.joinpath(*suffix.strip(".").split(".")))] if suffix else [str(ROOT)]
+    sys.modules[package.__name__] = package
+
+compiler = importlib.import_module(f"{PACKAGE}.api.orchestration.serverless_media_prompt")
+legacy = importlib.import_module(f"{PACKAGE}.api.orchestration.serverless_prompt")
+contract = importlib.import_module(f"{PACKAGE}.serverless_bridge.contract")
+compile_media_prompt = compiler.compile_media_prompt
+SOCKETS = {"images": "IMAGE", "audio": "AUDIO"}
+
+
+def _node(class_type, **inputs):
+    return {"class_type": class_type, "inputs": inputs}
+
+
+def _collector(**sockets):
+    return _node(
+        "DistributedCollector", multi_job_id="job/42", worker_id="worker_0", **sockets
+    )
+
+
+def _prompt():
+    return {
+        "source:1": _node("LoadImage", image="example.png"),
+        "collector:1": _collector(images=["source:1", 0]),
+    }
+
+
+def _run(prompt):
+    return compile_media_prompt(prompt, "request:1")
+
+
+@pytest.mark.parametrize("socket,kind", SOCKETS.items())
+def test_one_export_for_each_typed_socket_with_exact_canonical_binding(socket, kind):
+    prompt = {"source": _node("Source"), "collector": _collector(**{socket: ["source", 0]})}
+    graph, bindings = _run(prompt)
+    assert len(bindings) == 1
+    binding = bindings[0]
+    assert set(binding) == contract.BINDING_FIELDS
+    assert binding == {
+        "binding_id": binding["binding_id"],
+        "collector_id": "collector",
+        "job_id": "job/42",
+        "participant_id": "worker_0",
+        "node_id": binding["node_id"],
+        "socket": socket,
+        "kind": kind,
+    }
+    assert all(isinstance(value, str) for value in binding.values())
+    assert contract.validate_bindings(bindings) == {binding["binding_id"]: binding}
+    assert graph[binding["node_id"]] == {
+        "class_type": "MediaExportV1",
+        "inputs": {
+            "request_id": "request:1",
+            "binding_json": json.dumps(binding),
+            socket: ["source", 0],
+            **({"transfer_format": "png", "transfer_quality": 90} if kind == "IMAGE" else {}),
+        },
+    }
+    assert set(graph) == {"source", binding["node_id"]}
+
+
+def test_multiple_collectors_have_independent_image_audio_targets():
+    prompt = {
+        "shared": _node("MultiMediaSource"),
+        "collector:z": _collector(audio=["shared", 1]),
+        "collector:a": _collector(images=["shared", 0], audio=["shared", 1]),
+    }
+    prompt["collector:z"]["inputs"].update(multi_job_id="job:z", worker_id="participant:z")
+    graph, bindings = _run(prompt)
+    assert [(b["collector_id"], b["socket"]) for b in bindings] == [
+        ("collector:a", "images"), ("collector:a", "audio"), ("collector:z", "audio"),
+    ]
+    assert len({b["binding_id"] for b in bindings}) == 3
+    targets = {b["node_id"] for b in bindings}
+    assert len(targets) == 3
+    assert set(graph) == {"shared"} | targets
+    for binding in bindings:
+        export = graph[binding["node_id"]]
+        inputs = prompt[binding["collector_id"]]["inputs"]
+        assert export["inputs"][binding["socket"]] == inputs[binding["socket"]]
+        assert binding["job_id"] == inputs["multi_job_id"]
+        assert binding["participant_id"] == inputs["worker_id"]
+
+
+def test_prunes_descendants_and_unrelated_outputs_preserving_required_output_utility():
+    prompt = _prompt()
+    prompt.update({
+        "required-output": _node("RequiredOutputUtility", images=["source:1", 0]),
+        "consumer": _node("ImageTransform", image=["required-output", 0]),
+        "descendant": _node("ImageTransform", image=["collector:1", 0]),
+        "after": _node("SaveImage", images=["descendant", 0]),
+        "unused-output": _node("SaveImage", images=["source:1", 0]),
+        "unrelated": _node("LoadAudio"),
+    })
+    prompt["required-output"]["_meta"] = {"title": "Required OUTPUT_NODE utility"}
+    prompt["collector:1"]["inputs"].update(images=["consumer", 0], load_balance=True, pass_through=False)
+    graph, bindings = _run(prompt)
+    assert set(graph) == {"source:1", "required-output", "consumer", bindings[0]["node_id"]}
+    assert graph["required-output"] == prompt["required-output"]
+    assert not {"load_balance", "pass_through"} & graph[bindings[0]["node_id"]]["inputs"].keys()
+
+
+def test_upstream_primitive_string_preserves_empty_and_link_looking_text():
+    for text in ("", '["source:1", 0]', "hello\nworld"):
+        prompt = {"text": _node("PrimitiveString", value=text),
+                  "source": _node("TextToImage", text=["text", 0]),
+                  "collector": _collector(images=["source", 0])}
+        graph, bindings = _run(prompt)
+        assert graph["text"] == prompt["text"]
+        assert graph["source"]["inputs"]["text"] == ["text", 0]
+        assert graph[bindings[0]["node_id"]]["inputs"]["images"] == ["source", 0]
+
+
+@pytest.mark.parametrize("class_type,inputs", [
+    ("DistributedSeed", {}),
+    ("DistributedSeed", {"seed": 10, "is_worker": True, "worker_id": "worker_1"}),
+    ("DistributedSeed", {"seed": 10, "is_worker": False, "worker_id": "worker_1"}),
+    ("DistributedSeed", {"seed": 10, "is_worker": True, "worker_id": "bad"}),
+    ("DistributedValue", {"default_value": "3.9", "worker_values": '{"_type":"INT","1":"8.7"}', "is_worker": True, "worker_id": "worker_0"}),
+    ("DistributedValue", {"default_value": "3.9", "worker_values": '{"_type":"INT","2":""}', "is_worker": True, "worker_id": "worker_1"}),
+    ("DistributedValue", {"default_value": "fallback", "worker_values": "malformed"}),
+    ("DistributedValue", {"default_value": 1.25, "worker_values": '{"_type":"FLOAT","1":"bad"}', "is_worker": True, "worker_id": "0"}),
+    ("DistributedModelName", {"text": "model.safetensors"}),
+    ("DistributedModelName", {"text": {"name": "checkpoint"}}),
+    ("DistributedModelName", {"text": ["a", "b", "c"]}),
+])
+def test_upstream_scalar_values_match_legacy_compiler(class_type, inputs):
+    prompt = _prompt()
+    prompt["scalar"] = _node(class_type, **inputs)
+    prompt["source:1"]["inputs"]["parameter"] = ["scalar", 0]
+    old_prompt = copy.deepcopy(prompt)
+    old_prompt["collector:1"]["class_type"] = "DistributedCollector"
+    expected, _ = legacy.compile_serverless_prompt(old_prompt)
+    graph, _ = _run(prompt)
+    assert graph["source:1"] == expected["source:1"]
+    assert "scalar" not in graph
+
+
+def test_imports_same_scalar_resolver_and_resolves_primitive_chains():
+    assert compiler._make_scalar_resolver is legacy._make_scalar_resolver
+    prompt = _prompt()
+    prompt.update({
+        "int": _node("PrimitiveInt", value=25),
+        "int-alias": _node("PrimitiveInt", value=["int", 0]),
+        "float": _node("PrimitiveFloat", value=1.25),
+        "bool": _node("PrimitiveBoolean", value=True),
+        "string": _node("PrimitiveString", value="model.safetensors"),
+        "seed": _node("DistributedSeed", seed=["int-alias", 0], is_worker=True, worker_id="worker_0"),
+        "value": _node("DistributedValue", default_value=["float", 0], worker_values='{"_type":"FLOAT","1":"2.75"}', is_worker=True, worker_id="worker_0"),
+        "enabled": _node("DistributedValue", default_value=["bool", 0]),
+        "model": _node("DistributedModelName", text=["string", 0]),
+    })
+    prompt["source:1"]["inputs"].update(seed=["seed", 0], cfg=["value", 0], enabled=["enabled", 0], model=["model", 0])
+    graph, bindings = _run(prompt)
+    assert graph["source:1"]["inputs"] == {"image": "example.png", "seed": 26, "cfg": 2.75, "enabled": True, "model": "model.safetensors"}
+    assert set(graph) == {"source:1", bindings[0]["node_id"]}
+
+
+def test_distributed_text_value_can_be_wrapped_by_primitive_string():
+    prompt = {
+        "value": _node("DistributedModelName", text="model.safetensors"),
+        "text": _node("PrimitiveString", value=["value", 0]),
+        "source": _node("TextToImage", text=["text", 0]),
+        "collector": _collector(images=["source", 0]),
+    }
+    graph, bindings = _run(prompt)
+    assert graph["text"]["inputs"]["value"] == "model.safetensors"
+    assert graph["source"]["inputs"]["text"] == ["text", 0]
+    assert graph[bindings[0]["node_id"]]["inputs"]["images"] == ["source", 0]
+    assert "value" not in graph
+
+
+@pytest.mark.parametrize("class_type,input_name", [
+    ("DistributedSeed", "seed"), ("DistributedValue", "default_value"), ("DistributedModelName", "text"),
+])
+def test_dynamic_scalar_resolution_stays_fail_closed(class_type, input_name):
+    prompt = _prompt()
+    prompt["dynamic"] = _node("ArbitraryValue")
+    prompt["scalar"] = _node(class_type, **{input_name: ["dynamic", 0]})
+    prompt["source:1"]["inputs"]["parameter"] = ["scalar", 0]
+    with pytest.raises(ValueError, match="literal or supported Primitive"):
+        _run(prompt)
+
+
+def test_namespaced_export_ids_avoid_existing_ids_even_on_pruned_nodes():
+    prompt = _prompt()
+    _, first = _run(prompt)
+    original_id = first[0]["node_id"]
+    assert ":" in original_id and not original_id.isdigit()
+    prompt[original_id] = _node("ExistingUtility")
+    _, second = _run(prompt)
+    assert second[0]["node_id"] not in prompt
+    assert second[0]["binding_id"] == first[0]["binding_id"]
+    prompt[second[0]["node_id"]] = _node("AnotherUtility")
+    graph, third = _run(prompt)
+    assert third[0]["node_id"] not in prompt
+    assert third == _run(dict(reversed(list(prompt.items()))))[1]
+    assert set(graph) == {"source:1", third[0]["node_id"]}
+
+
+def test_long_valid_identifiers_remain_exact_and_generated_ids_fit_contract():
+    collector_id = "namespace/" + "c" * 190
+    prompt = {"source:1": _node("Source"), collector_id: _collector(images=["source:1", 0])}
+    prompt[collector_id]["inputs"].update(multi_job_id="j" * 200, worker_id="p" * 200)
+    _, bindings = _run(prompt)
+    assert bindings[0]["collector_id"] == collector_id
+    assert bindings[0]["job_id"] == "j" * 200
+    assert bindings[0]["participant_id"] == "p" * 200
+    contract.validate_bindings(bindings)
+
+
+def test_deterministic_independent_copies_and_request_identity():
+    prompt = _prompt()
+    prompt["source:1"]["_meta"] = {"nested": {"values": [1, 2, 3]}}
+    original = copy.deepcopy(prompt)
+    graph, bindings = _run(prompt)
+    assert prompt == original
+    assert (graph, bindings) == _run(prompt)
+    assert (graph, bindings) == _run(dict(reversed(list(prompt.items()))))
+    other_graph, other_bindings = compile_media_prompt(prompt, "request:2")
+    assert other_bindings == bindings
+    target = bindings[0]["node_id"]
+    assert other_graph[target]["inputs"]["request_id"] == "request:2"
+    graph["source:1"]["_meta"]["nested"]["values"].append(4)
+    graph[target]["inputs"]["images"][0] = "changed"
+    bindings[0]["collector_id"] = "changed"
+    assert prompt == original
+    assert json.loads(graph[target]["inputs"]["binding_json"])["collector_id"] == "collector:1"
+    assert other_graph["source:1"] == original["source:1"]
+
+
+@pytest.mark.parametrize("bad", [None, [], {}, {"x": None}, {"x": []}, {"x": {}},
+    {"x": {"class_type": None}}, {"x": {"class_type": ""}},
+    {"x": {"class_type": "Source", "inputs": []}},
+    {"x": {"class_type": "Source", "inputs": None}},
+    {"x": {"class_type": "Source", "inputs": {1: "literal"}}},
+])
+def test_rejects_malformed_prompt_or_missing_collectors(bad):
+    with pytest.raises(ValueError):
+        _run(bad)
+
+
+@pytest.mark.parametrize("bad", [None, 1, True, "", " bad", "bad ", "bad\\id", "bad\nid", "x" * 201])
+def test_rejects_malformed_ids_without_normalizing(bad):
+    prompt = _prompt()
+    prompt[bad] = _node("Source")
+    with pytest.raises(ValueError):
+        _run(prompt)
+
+
+@pytest.mark.parametrize("bad", [None, 1, True, "", " bad", "bad ", "bad\\id", "x" * 201])
+@pytest.mark.parametrize("field", ["multi_job_id", "worker_id", "request_id"])
+def test_rejects_malformed_binding_and_request_identity(field, bad):
+    prompt = _prompt()
+    request_id = "request:1"
+    if field == "request_id":
+        request_id = bad
+    else:
+        prompt["collector:1"]["inputs"][field] = bad
+    with pytest.raises(ValueError):
+        compile_media_prompt(prompt, request_id)
+
+
+@pytest.mark.parametrize("field", ["multi_job_id", "worker_id"])
+def test_requires_prepared_literal_hidden_identity(field):
+    prompt = _prompt()
+    del prompt["collector:1"]["inputs"][field]
+    with pytest.raises(ValueError):
+        _run(prompt)
+    prompt["collector:1"]["inputs"][field] = ["source:1", 0]
+    with pytest.raises(ValueError):
+        _run(prompt)
+
+
+@pytest.mark.parametrize("socket", SOCKETS)
+@pytest.mark.parametrize("value", [None, "literal", 5, True, [], ["source:1"], ["source:1", 0, 1], ("source:1", 0)])
+def test_every_present_media_socket_requires_an_api_link(socket, value):
+    prompt = _prompt()
+    prompt["collector:1"]["inputs"][socket] = value
+    with pytest.raises(ValueError, match="link"):
+        _run(prompt)
+
+
+def test_each_collector_requires_at_least_one_media_socket():
+    prompt = _prompt()
+    prompt["empty"] = _collector(load_balance=True, pass_through=False)
+    with pytest.raises(ValueError, match="connected|linked"):
+        _run(prompt)
+
+
+@pytest.mark.parametrize("link", [["missing", 0], ["source:1", -1], ["source:1", 0.0],
+    ["source:1", "0"], ["source:1", True], ["source:1", False], [1, 0],
+    [True, 0], [None, 0], [[], 0], [{}, 0], [" source:1", 0],
+])
+@pytest.mark.parametrize("location", ["collector", "ordinary", "unrelated"])
+def test_rejects_invalid_links_everywhere_before_pruning(location, link):
+    prompt = _prompt()
+    prompt["1"] = _node("Source")  # An integer/bool source must not be stringified to this ID.
+    if location == "collector":
+        prompt["collector:1"]["inputs"]["images"] = link
+    elif location == "ordinary":
+        prompt["source:1"]["inputs"]["parameter"] = link
+    else:
+        prompt["unrelated"] = _node("Utility", value=link)
+    with pytest.raises(ValueError):
+        _run(prompt)
+
+
+@pytest.mark.parametrize("class_type", [
+    "DistributedMediaCollector", "DistributedEmptyImage", "UltimateSDUpscaleDistributed",
+    "DistributedImageListStart", "DistributedImageListEnd", "DistributedWorkerSink",
+    "ImageBatchDivider", "AudioBatchDivider", "CustomDistributedNode", "custombatchdivider",
+])
+def test_rejects_unsupported_distributed_families_even_if_unused(class_type):
+    prompt = _prompt()
+    prompt["unsupported"] = _node(class_type)
+    with pytest.raises(ValueError, match="unsupported"):
+        _run(prompt)
+
+
+@pytest.mark.parametrize("class_type", ["DistributedSeed", "DistributedValue", "DistributedModelName", "PrimitiveInt", "PrimitiveString"])
+def test_rejects_invalid_known_scalar_output_slots(class_type):
+    prompt = _prompt()
+    prompt["scalar"] = _node(class_type)
+    prompt["source:1"]["inputs"]["parameter"] = ["scalar", 1]
+    with pytest.raises(ValueError, match="slot"):
+        _run(prompt)
+
+
+@pytest.mark.parametrize("socket", SOCKETS)
+def test_rejects_direct_distributed_scalar_media_links(socket):
+    prompt = {"scalar": _node("DistributedValue", default_value="text"), "collector": _collector(**{socket: ["scalar", 0]})}
+    with pytest.raises(ValueError, match="scalar"):
+        _run(prompt)
+
+
+@pytest.mark.parametrize("indirect", [False, True])
+def test_rejects_collector_chains_across_different_media(indirect):
+    prompt = _prompt()
+    source = "collector:1"
+    if indirect:
+        prompt["processing"] = _node("ImageToAudio", images=[source, 0])
+        source = "processing"
+    prompt["collector:2"] = _collector(audio=[source, 0])
+    with pytest.raises(ValueError, match="collector-to-collector"):
+        _run(prompt)
+
+
+@pytest.mark.parametrize("mode", ["ordinary", "self", "unrelated", "primitive", "collector"])
+def test_rejects_cycles_in_entire_prompt(mode):
+    prompt = _prompt()
+    if mode == "ordinary":
+        prompt["source:1"]["inputs"]["other"] = ["other", 0]
+        prompt["other"] = _node("Utility", value=["source:1", 0])
+    elif mode == "self":
+        prompt["source:1"]["inputs"]["other"] = ["source:1", 0]
+    elif mode == "unrelated":
+        prompt["unused"] = _node("Utility", value=["unused", 0])
+    elif mode == "primitive":
+        prompt["p1"] = _node("PrimitiveInt", value=["p2", 0])
+        prompt["p2"] = _node("PrimitiveInt", value=["p1", 0])
+    else:
+        prompt["source:1"]["inputs"]["other"] = ["collector:1", 0]
+    with pytest.raises(ValueError, match="cycle"):
+        _run(prompt)
+
+
+def test_deep_acyclic_graph_does_not_depend_on_python_recursion_limit():
+    prompt = {"n:0": _node("Source")}
+    for index in range(1, 1500):
+        prompt[f"n:{index}"] = _node("Utility", value=[f"n:{index - 1}", 0])
+    prompt["collector"] = _collector(images=["n:1499", 0])
+    graph, bindings = _run(prompt)
+    assert len(graph) == 1501
+    assert graph[bindings[0]["node_id"]]["inputs"]["images"] == ["n:1499", 0]
+
+
+def test_binding_count_respects_bridge_contract():
+    prompt = {"source": _node("Source")}
+    for index in range(contract.MAX_ARTIFACTS + 1):
+        prompt[f"collector:{index}"] = _collector(images=["source", 0])
+    with pytest.raises(ValueError, match="count"):
+        _run(prompt)
+
+
+def test_import_and_compile_do_not_import_comfy_or_tensor_libraries():
+    script = '''
+import builtins, importlib, pathlib, sys, types
+root = pathlib.Path(sys.argv[1])
+name = "isolated_media_compile"
+for suffix in ("", ".api", ".api.orchestration", ".serverless_bridge"):
+    package = types.ModuleType(name + suffix)
+    package.__path__ = [str(root.joinpath(*suffix.strip(".").split(".")))] if suffix else [str(root)]
+    sys.modules[package.__name__] = package
+original_import = builtins.__import__
+def guarded_import(name, *args, **kwargs):
+    if name.split(".")[0] in {"comfy", "comfy_api", "nodes", "server", "torch", "numpy"}:
+        raise AssertionError("unexpected runtime import: " + name)
+    return original_import(name, *args, **kwargs)
+builtins.__import__ = guarded_import
+module = importlib.import_module(name + ".api.orchestration.serverless_media_prompt")
+prompt = {"s": {"class_type": "Source", "inputs": {}}, "c": {"class_type": "DistributedCollector", "inputs": {"images": ["s", 0], "multi_job_id": "job", "worker_id": "worker"}}}
+graph, bindings = module.compile_media_prompt(prompt, "request")
+assert graph[bindings[0]["node_id"]]["class_type"] == "MediaExportV1"
+'''
+    result = subprocess.run([sys.executable, "-c", script, str(ROOT)], text=True, capture_output=True)
+    assert result.returncode == 0, result.stderr

tests/test_serverless_prompt.py

@@ -0,0 +1,392 @@
+import importlib.util
+import json
+import sys
+import types
+from pathlib import Path
+
+import pytest
+
+
+MODULE_PATH = (
+    Path(__file__).resolve().parents[1]
+    / "api"
+    / "orchestration"
+    / "serverless_prompt.py"
+)
+_spec = importlib.util.spec_from_file_location("serverless_prompt_under_test", MODULE_PATH)
+assert _spec is not None and _spec.loader is not None
+_serverless_prompt = importlib.util.module_from_spec(_spec)
+_spec.loader.exec_module(_serverless_prompt)
+compile_serverless_prompt = _serverless_prompt.compile_serverless_prompt
+
+
+def _prepared_prompt():
+    return {
+        "1": {"class_type": "CheckpointLoaderSimple", "inputs": {}, "_meta": {"title": "source"}},
+        "2": {"class_type": "KSampler", "inputs": {"model": ["1", 0], "seed": ["3", 0]}},
+        "3": {
+            "class_type": "DistributedSeed",
+            "inputs": {"seed": 40, "is_worker": True, "worker_id": "worker_0"},
+        },
+        "4": {
+            "class_type": "DistributedCollector",
+            "inputs": {"images": ["2", 0], "multi_job_id": "job/42", "worker_id": "worker_0", "is_worker": True},
+        },
+        "5": {"class_type": "VAEDecode", "inputs": {"samples": ["4", 0]}},
+        "6": {"class_type": "SaveImage", "inputs": {"images": ["5", 0]}},
+        "7": {"class_type": "PreviewImage", "inputs": {"images": ["4", 0]}, "_meta": {"title": "Preview Image (auto-added)"}},
+        "8": {"class_type": "UnrelatedNode", "inputs": {}},
+    }
+
+
+def _run(prompt):
+    return compile_serverless_prompt(prompt)
+
+
+def _load_prompt_transform():
+    module_path = MODULE_PATH.with_name("prompt_transform.py")
+    package_name = "serverless_prompt_transform_testpkg"
+    root_pkg = types.ModuleType(package_name)
+    root_pkg.__path__ = []
+    sys.modules[package_name] = root_pkg
+    for suffix in ("api", "api.orchestration", "utils"):
+        pkg = types.ModuleType(f"{package_name}.{suffix}")
+        pkg.__path__ = []
+        sys.modules[f"{package_name}.{suffix}"] = pkg
+    logging_module = types.ModuleType(f"{package_name}.utils.logging")
+    setattr(logging_module, "debug_log", lambda *_args, **_kwargs: None)
+    setattr(logging_module, "log", lambda *_args, **_kwargs: None)
+    sys.modules[f"{package_name}.utils.logging"] = logging_module
+    spec = importlib.util.spec_from_file_location(
+        f"{package_name}.api.orchestration.prompt_transform", module_path
+    )
+    assert spec is not None and spec.loader is not None
+    module = importlib.util.module_from_spec(spec)
+    spec.loader.exec_module(module)
+    return module
+
+
+def test_compiles_prompt_prepared_by_existing_worker_helpers():
+    prompt_transform = _load_prompt_transform()
+    original = {
+        "1": {"class_type": "CheckpointLoaderSimple", "inputs": {}},
+        "2": {"class_type": "KSampler", "inputs": {"model": ["1", 0], "seed": ["3", 0]}},
+        "3": {"class_type": "DistributedSeed", "inputs": {"seed": 100}},
+        "4": {"class_type": "DistributedCollector", "inputs": {"images": ["2", 0]}},
+        "5": {"class_type": "SaveImage", "inputs": {"images": ["4", 0]}},
+    }
+    pruned = prompt_transform.prune_prompt_for_worker(original)
+    prompt_index = prompt_transform.PromptIndex(pruned)
+    prepared = prompt_transform.apply_participant_overrides(
+        pruned,
+        participant_id="worker-a",
+        enabled_worker_ids=["worker-a"],
+        job_id_map=prompt_transform.generate_job_id_map(prompt_index, "run"),
+        master_url="http://master.invalid",
+        delegate_master=False,
+        prompt_index=prompt_index,
+    )
+
+    graph, manifest = _run(prepared)
+
+    assert graph["2"]["inputs"]["seed"] == 101
+    assert set(graph) == {"1", "2", "7"}
+    assert graph["7"]["class_type"] == "SaveImage"
+    assert manifest[0]["job_id"] == "run_4"
+    assert manifest[0]["collector_id"] == "4"
+
+
+def test_compiles_collector_input_to_flat_saveimage_and_prunes_downstream():
+    original = _prepared_prompt()
+
+    graph, manifest = _run(original)
+
+    assert manifest == [
+        {
+            "collector_id": "4",
+            "job_id": "job/42",
+            "node_id": "9",
+            "kind": "image",
+            "prefix": graph["9"]["inputs"]["filename_prefix"],
+        }
+    ]
+    prefix = manifest[0]["prefix"]
+    assert graph["9"] == {
+        "class_type": "SaveImage",
+        "inputs": {"images": ["2", 0], "filename_prefix": prefix},
+        "_meta": {"title": "Serverless output: collector 4 (image)"},
+    }
+    assert "/" not in prefix and "\\" not in prefix
+    assert graph["2"]["inputs"]["seed"] == 41
+    assert set(graph) == {"1", "2", "9"}
+    assert all("Distributed" not in node["class_type"] for node in graph.values())
+    assert original["3"]["inputs"]["seed"] == 40
+    assert original["4"]["inputs"]["images"] == ["2", 0]
+
+
+def test_manifest_and_prefix_are_repeatable_and_safe_for_untrusted_identity_parts():
+    prompt = _prepared_prompt()
+    prompt["4"]["inputs"].update(multi_job_id="../run/a", worker_id="worker/unsafe")
+
+    first_graph, first_manifest = _run(prompt)
+    second_graph, second_manifest = _run(prompt)
+
+    assert first_manifest == second_manifest
+    assert first_graph == second_graph
+    prefix = first_manifest[0]["prefix"]
+    assert prefix == first_graph[first_manifest[0]["node_id"]]["inputs"]["filename_prefix"]
+    assert not any(char in prefix for char in ("/", "\\", ":"))
+    assert first_manifest[0]["job_id"] == "../run/a"
+
+
+def test_compiles_multiple_disconnected_collectors_in_stable_order():
+    prompt = {
+        "10": {"class_type": "LoadImage", "inputs": {}},
+        "20": {"class_type": "LoadImage", "inputs": {}},
+        "2": {"class_type": "DistributedCollector", "inputs": {"images": ["20", 0], "multi_job_id": "b", "worker_id": "worker_1"}},
+        "1": {"class_type": "DistributedCollector", "inputs": {"images": ["10", 0], "multi_job_id": "a", "worker_id": "worker_0"}},
+        "30": {"class_type": "PreviewImage", "inputs": {"images": ["1", 0]}},
+    }
+
+    graph, manifest = _run(prompt)
+
+    assert [item["collector_id"] for item in manifest] == ["1", "2"]
+    assert [item["node_id"] for item in manifest] == ["31", "32"]
+    assert graph["31"]["inputs"]["images"] == ["10", 0]
+    assert graph["32"]["inputs"]["images"] == ["20", 0]
+    assert set(graph) == {"10", "20", "31", "32"}
+
+
+def test_rejects_audio_input_instead_of_emitting_audio_saver():
+    prompt = {
+        "1": {"class_type": "LoadAudio", "inputs": {}},
+        "2": {"class_type": "DistributedCollector", "inputs": {"audio": ["1", 0]}},
+    }
+
+    with pytest.raises(ValueError, match="audio.*unsupported"):
+        _run(prompt)
+
+
+def test_rejects_collector_without_image_input():
+    with pytest.raises(ValueError, match="collector.*image"):
+        _run({"1": {"class_type": "DistributedCollector", "inputs": {}}})
+
+
+@pytest.mark.parametrize(
+    "class_type",
+    ["UltimateSDUpscaleDistributed", "DistributedEmptyImage", "ImageBatchDivider", "AudioBatchDivider"],
+)
+def test_rejects_unsupported_distributed_and_batch_nodes(class_type):
+    prompt = {
+        "1": {"class_type": class_type, "inputs": {}},
+        "2": {"class_type": "DistributedCollector", "inputs": {"images": ["3", 0]}},
+        "3": {"class_type": "LoadImage", "inputs": {}},
+    }
+
+    with pytest.raises(ValueError, match="unsupported"):
+        _run(prompt)
+
+
+def test_rejects_collector_to_collector_dependency_through_processing_node():
+    prompt = {
+        "1": {"class_type": "LoadImage", "inputs": {}},
+        "2": {"class_type": "DistributedCollector", "inputs": {"images": ["1", 0]}},
+        "3": {"class_type": "ImageScale", "inputs": {"image": ["2", 0]}},
+        "4": {"class_type": "DistributedCollector", "inputs": {"images": ["3", 0]}},
+    }
+
+    with pytest.raises(ValueError, match="collector-to-collector"):
+        _run(prompt)
+
+
+@pytest.mark.parametrize(
+    "link",
+    [["missing", 0], ["1", -1], ["1", "0"], ["1", True]],
+)
+def test_rejects_malformed_api_links(link):
+    prompt = {
+        "1": {"class_type": "LoadImage", "inputs": {}},
+        "2": {"class_type": "DistributedCollector", "inputs": {"images": link}},
+    }
+
+    with pytest.raises(ValueError, match="link"):
+        _run(prompt)
+
+
+def test_rejects_unexpected_output_slot_from_removed_distributed_node():
+    prompt = {
+        "1": {"class_type": "DistributedSeed", "inputs": {"seed": 1}},
+        "2": {"class_type": "KSampler", "inputs": {"seed": ["1", 1]}},
+        "3": {"class_type": "DistributedCollector", "inputs": {"images": ["2", 0]}},
+    }
+
+    with pytest.raises(ValueError, match="output slot"):
+        _run(prompt)
+
+
+def test_distributed_seed_matches_worker_offset_and_master_passthrough():
+    prompt = {
+        "1": {"class_type": "DistributedSeed", "inputs": {"seed": 10, "is_worker": True, "worker_id": "worker_1"}},
+        "2": {"class_type": "DistributedSeed", "inputs": {"seed": 7, "is_worker": False}},
+        "3": {"class_type": "KSampler", "inputs": {"seed": ["1", 0], "steps": ["2", 0]}},
+        "4": {"class_type": "DistributedCollector", "inputs": {"images": ["3", 0]}},
+    }
+
+    graph, _manifest = _run(prompt)
+
+    assert graph["3"]["inputs"] == {"seed": 12, "steps": 7}
+    assert "1" not in graph and "2" not in graph
+
+
+def test_distributed_value_uses_one_based_worker_values_and_type_coercion():
+    prompt = {
+        "1": {"class_type": "DistributedValue", "inputs": {"default_value": "3.9", "worker_values": json.dumps({"_type": "INT", "1": "8.7"}), "is_worker": True, "worker_id": "worker_0"}},
+        "2": {"class_type": "DistributedValue", "inputs": {"default_value": "3.9", "worker_values": json.dumps({"_type": "INT", "2": ""}), "is_worker": True, "worker_id": "worker_1"}},
+        "3": {"class_type": "KSampler", "inputs": {"steps": ["1", 0], "seed": ["2", 0]}},
+        "4": {"class_type": "DistributedCollector", "inputs": {"images": ["3", 0]}},
+    }
+
+    graph, _manifest = _run(prompt)
+
+    assert graph["3"]["inputs"] == {"steps": 8, "seed": 3}
+    assert "1" not in graph and "2" not in graph
+
+
+def test_resolves_stock_primitive_links_for_distributed_seed_and_value():
+    prompt = {
+        "1": {"class_type": "PrimitiveInt", "inputs": {"value": 25}},
+        "2": {"class_type": "DistributedSeed", "inputs": {"seed": ["1", 0], "is_worker": True, "worker_id": "worker_0"}},
+        "3": {"class_type": "PrimitiveString", "inputs": {"value": "4.8"}},
+        "4": {"class_type": "PrimitiveString", "inputs": {"value": '{"_type":"FLOAT","1":"2.75"}'}},
+        "5": {"class_type": "DistributedValue", "inputs": {"default_value": ["3", 0], "worker_values": ["4", 0], "is_worker": True, "worker_id": "worker_0"}},
+        "6": {"class_type": "KSampler", "inputs": {"seed": ["2", 0], "cfg": ["5", 0]}},
+        "7": {"class_type": "DistributedCollector", "inputs": {"images": ["6", 0]}},
+    }
+
+    graph, _manifest = _run(prompt)
+
+    assert graph["6"]["inputs"] == {"seed": 26, "cfg": 2.75}
+    assert not {"1", "2", "3", "4", "5"}.intersection(graph)
+
+
+def test_rejects_dynamic_linked_seed_or_value_without_executing_it():
+    prompt = {
+        "1": {"class_type": "ArbitraryNode", "inputs": {}},
+        "2": {"class_type": "DistributedSeed", "inputs": {"seed": ["1", 0], "is_worker": True, "worker_id": "worker_0"}},
+        "3": {"class_type": "KSampler", "inputs": {"seed": ["2", 0]}},
+        "4": {"class_type": "DistributedCollector", "inputs": {"images": ["3", 0]}},
+    }
+
+    with pytest.raises(ValueError, match="literal or supported Primitive"):
+        _run(prompt)
+
+
+def test_primitive_resolution_detects_cycles():
+    prompt = {
+        "1": {"class_type": "PrimitiveInt", "inputs": {"value": ["2", 0]}},
+        "2": {"class_type": "PrimitiveInt", "inputs": {"value": ["1", 0]}},
+        "3": {"class_type": "DistributedSeed", "inputs": {"seed": ["1", 0]}},
+        "4": {"class_type": "KSampler", "inputs": {"seed": ["3", 0]}},
+        "5": {"class_type": "DistributedCollector", "inputs": {"images": ["4", 0]}},
+    }
+
+    with pytest.raises(ValueError, match="cycle"):
+        _run(prompt)
+
+
+def test_model_name_trivial_text_passthrough_is_inlined():
+    prompt = {
+        "1": {"class_type": "DistributedModelName", "inputs": {"text": "model.safetensors"}},
+        "2": {"class_type": "KSampler", "inputs": {"model_name": ["1", 0]}},
+        "3": {"class_type": "DistributedCollector", "inputs": {"images": ["2", 0]}},
+    }
+
+    graph, _manifest = _run(prompt)
+
+    assert graph["2"]["inputs"]["model_name"] == "model.safetensors"
+    assert "1" not in graph
+
+
+def test_rejects_dynamic_linked_model_name_text():
+    prompt = {
+        "1": {"class_type": "ArbitraryNode", "inputs": {}},
+        "2": {"class_type": "DistributedModelName", "inputs": {"text": ["1", 0]}},
+        "3": {"class_type": "KSampler", "inputs": {"model_name": ["2", 0]}},
+        "4": {"class_type": "DistributedCollector", "inputs": {"images": ["3", 0]}},
+    }
+
+    with pytest.raises(ValueError, match="DistributedModelName.*literal or supported Primitive"):
+        _run(prompt)
+
+
+def test_resolves_boolean_and_float_primitive_links():
+    prompt = {
+        "1": {"class_type": "PrimitiveBoolean", "inputs": {"value": True}},
+        "2": {"class_type": "PrimitiveFloat", "inputs": {"value": 1.25}},
+        "3": {"class_type": "DistributedValue", "inputs": {"default_value": ["1", 0], "is_worker": False}},
+        "4": {"class_type": "DistributedValue", "inputs": {"default_value": ["2", 0], "worker_values": '{"_type":"FLOAT","1":"2.5"}', "is_worker": True, "worker_id": "worker_0"}},
+        "5": {"class_type": "KSampler", "inputs": {"enabled": ["3", 0], "cfg": ["4", 0]}},
+        "6": {"class_type": "DistributedCollector", "inputs": {"images": ["5", 0]}},
+    }
+
+    graph, _manifest = _run(prompt)
+
+    assert graph["5"]["inputs"] == {"enabled": True, "cfg": 2.5}
+    assert not {"1", "2", "3", "4"}.intersection(graph)
+
+
+def test_compiled_graph_matches_golden_prompt_and_manifest():
+    prompt = _prepared_prompt()
+    prompt["3"]["inputs"]["seed"] = ["10", 0]
+    prompt["10"] = {"class_type": "PrimitiveInt", "inputs": {"value": 40}}
+    prompt["11"] = {"class_type": "PrimitiveFloat", "inputs": {"value": 1.25}}
+    prompt["12"] = {
+        "class_type": "DistributedValue",
+        "inputs": {
+            "default_value": ["11", 0],
+            "worker_values": json.dumps({"_type": "FLOAT", "1": "2.5"}),
+            "is_worker": True,
+            "worker_id": "worker_0",
+        },
+    }
+    prompt["13"] = {"class_type": "PrimitiveString", "inputs": {"value": "model.safetensors"}}
+    prompt["14"] = {"class_type": "DistributedModelName", "inputs": {"text": ["13", 0]}}
+    prompt["2"]["inputs"].update(cfg=["12", 0], model_name=["14", 0])
+
+    graph, manifest = _run(prompt)
+
+    expected_prefix = (
+        "serverless_job_42_worker_0_4_image_"
+        "1c81da108a2eb60282bd436512acfc59903aba73e49b19b6f47bbdcafd30d244"
+    )
+    assert list(graph) == ["1", "2", "15"]
+    assert graph == {
+        "1": {
+            "class_type": "CheckpointLoaderSimple",
+            "inputs": {},
+            "_meta": {"title": "source"},
+        },
+        "2": {
+            "class_type": "KSampler",
+            "inputs": {
+                "model": ["1", 0],
+                "seed": 41,
+                "cfg": 2.5,
+                "model_name": "model.safetensors",
+            },
+        },
+        "15": {
+            "class_type": "SaveImage",
+            "inputs": {"images": ["2", 0], "filename_prefix": expected_prefix},
+            "_meta": {"title": "Serverless output: collector 4 (image)"},
+        },
+    }
+    assert manifest == [
+        {
+            "collector_id": "4",
+            "job_id": "job/42",
+            "node_id": "15",
+            "kind": "image",
+            "prefix": expected_prefix,
+        }
+    ]

tests/test_serverless_routes.py

@@ -0,0 +1,720 @@
+import copy
+import importlib.util
+import json
+import sys
+import tempfile
+import types
+import unittest
+from pathlib import Path
+from unittest.mock import patch
+
+
+class _FakeResponse:
+    def __init__(self, payload, status=200):
+        self.payload = copy.deepcopy(payload)
+        self.status = status
+
+
+class _FakeRequest:
+    def __init__(self, payload=None, error=None):
+        self.payload = payload
+        self.error = error
+
+    async def json(self):
+        if self.error is not None:
+            raise self.error
+        return self.payload
+
+
+class _Routes:
+    def get(self, _path):
+        return lambda handler: handler
+
+    def post(self, _path):
+        return lambda handler: handler
+
+
+def _load_modules():
+    repo = Path(__file__).resolve().parents[1]
+    package_name = "dist_serverless_routes_testpkg"
+    for name in list(sys.modules):
+        if name == package_name or name.startswith(f"{package_name}."):
+            del sys.modules[name]
+
+    root = types.ModuleType(package_name)
+    root.__path__ = [str(repo)]
+    sys.modules[package_name] = root
+    api_pkg = types.ModuleType(f"{package_name}.api")
+    api_pkg.__path__ = [str(repo / "api")]
+    sys.modules[f"{package_name}.api"] = api_pkg
+    utils_pkg = types.ModuleType(f"{package_name}.utils")
+    utils_pkg.__path__ = [str(repo / "utils")]
+    sys.modules[f"{package_name}.utils"] = utils_pkg
+    orchestration_pkg = types.ModuleType(f"{package_name}.api.orchestration")
+    orchestration_pkg.__path__ = [str(repo / "api" / "orchestration")]
+    sys.modules[f"{package_name}.api.orchestration"] = orchestration_pkg
+
+    logging_module = types.ModuleType(f"{package_name}.utils.logging")
+    logging_module.log = lambda *_args, **_kwargs: None
+    logging_module.debug_log = lambda *_args, **_kwargs: None
+    sys.modules[logging_module.__name__] = logging_module
+    constants_module = types.ModuleType(f"{package_name}.utils.constants")
+    constants_module.HEARTBEAT_TIMEOUT = 30
+    sys.modules[constants_module.__name__] = constants_module
+
+    def load_file(module_name, path):
+        spec = importlib.util.spec_from_file_location(module_name, path)
+        assert spec is not None and spec.loader is not None
+        module = importlib.util.module_from_spec(spec)
+        sys.modules[module_name] = module
+        spec.loader.exec_module(module)
+        return module
+
+    config_module = load_file(f"{package_name}.utils.config", repo / "utils" / "config.py")
+    serverless_config = load_file(
+        f"{package_name}.utils.serverless_config", repo / "utils" / "serverless_config.py"
+    )
+
+    runtime = types.ModuleType(f"{package_name}.api.orchestration.serverless_runtime")
+    runtime.secret_store = None
+    runtime.manager = types.SimpleNamespace(tasks={}, runs={}, pool_is_active=lambda _: False)
+    runtime.get_secret_store = lambda: runtime.secret_store
+    runtime.get_manager = lambda: runtime.manager
+    sys.modules[runtime.__name__] = runtime
+
+    import aiohttp  # Ensure the real dependency is loaded; never replace global aiohttp.
+    previous_server = sys.modules.get("server")
+    server = types.ModuleType("server")
+    server.PromptServer = types.SimpleNamespace(instance=types.SimpleNamespace(routes=_Routes()))
+    sys.modules["server"] = server
+
+    network = types.ModuleType(f"{package_name}.utils.network")
+
+    async def handle_api_error(_request, error, status=500):
+        return _FakeResponse({"status": "error", "message": str(error)}, status)
+
+    network.handle_api_error = handle_api_error
+    network.normalize_host = lambda value: value
+    sys.modules[network.__name__] = network
+
+    routes = load_file(f"{package_name}.api.config_routes", repo / "api" / "config_routes.py")
+    routes.web = types.SimpleNamespace(
+        json_response=lambda payload, status=200: _FakeResponse(payload, status)
+    )
+
+    if previous_server is None:
+        sys.modules.pop("server", None)
+    else:
+        sys.modules["server"] = previous_server
+    return config_module, serverless_config, runtime, routes
+
+
+config_module, serverless_config, runtime_module, config_routes = _load_modules()
+
+
+class _RunningTask:
+    def done(self):
+        return False
+
+
+class _ConfigRouteFixture(unittest.IsolatedAsyncioTestCase):
+    def setUp(self):
+        self.tempdir = tempfile.TemporaryDirectory()
+        self.root = Path(self.tempdir.name)
+        config_module.CONFIG_FILE = str(self.root / "gpu_config.json")
+        config_module.invalidate_config_cache()
+        self.config = config_module.get_default_config()
+        self.config["workers"] = []
+        self.assertTrue(config_module.save_config(self.config))
+        self.secrets = serverless_config.SecretStore(self.root / "private" / "credentials.json")
+        runtime_module.secret_store = self.secrets
+        runtime_module.manager = types.SimpleNamespace(tasks={}, runs={}, pool_is_active=lambda _: False)
+
+    def tearDown(self):
+        config_module.invalidate_config_cache()
+        self.tempdir.cleanup()
+
+    def read_config_file(self):
+        return json.loads(Path(config_module.CONFIG_FILE).read_text(encoding="utf-8"))
+
+
+class ServerlessConfigRouteTests(_ConfigRouteFixture):
+    async def test_get_config_redacts_keys_and_reports_only_configured_boolean(self):
+        self.secrets.set("pool-1", "private-api-key")
+        self.config["workers"] = [
+            {"id": "pool-1", "type": "serverless", "provider": "runpod", "endpoint_id": "ep-1",
+             "api_key": "inline-api-key", "name": "GPU"}
+        ]
+        Path(config_module.CONFIG_FILE).write_text(json.dumps(self.config), encoding="utf-8")
+        config_module.invalidate_config_cache()
+
+        response = await config_routes.get_config_endpoint(_FakeRequest())
+
+        self.assertEqual(response.status, 200)
+        worker = response.payload["workers"][0]
+        self.assertTrue(worker["api_key_configured"])
+        self.assertNotIn("api_key", worker)
+        self.assertNotIn("private-api-key", json.dumps(response.payload))
+        self.assertNotIn("inline-api-key", json.dumps(response.payload))
+
+    async def test_get_config_redacts_normalized_secret_field_aliases(self):
+        aliases = {
+            "apiKey": "camel-api-key",
+            "APIKEY": "upper-api-key",
+            "clearApiKey": "camel-clear-key",
+            "CLEAR_API_KEY": "upper-clear-key",
+            "runpod_CREDENTIAL": "provider-credential",
+        }
+        self.config["workers"] = [{
+            "id": "pool-1", "type": "serverless", "endpoint_id": "ep-123",
+            **aliases,
+            "nested": {"credential_note": "nested-credential"},
+        }]
+        Path(config_module.CONFIG_FILE).write_text(json.dumps(self.config), encoding="utf-8")
+        config_module.invalidate_config_cache()
+
+        response = await config_routes.get_config_endpoint(_FakeRequest())
+
+        self.assertEqual(response.status, 200)
+        worker = response.payload["workers"][0]
+        for alias, value in aliases.items():
+            self.assertNotIn(alias, worker)
+            self.assertNotIn(value, json.dumps(response.payload))
+        self.assertNotIn("credential_note", worker["nested"])
+        self.assertNotIn("nested-credential", json.dumps(response.payload))
+
+    async def test_bulk_config_rejects_normalized_secret_field_aliases(self):
+        for secret_field in ("apiKey", "APIKEY", "clearApiKey", "RUNPOD_CREDENTIAL"):
+            with self.subTest(secret_field=secret_field):
+                secret = f"private-{secret_field}"
+                response = await config_routes.update_config_endpoint(_FakeRequest({
+                    "workers": [{"id": "pool-1", "type": "local", secret_field: secret}]
+                }))
+                self.assertEqual(response.status, 400)
+                self.assertNotIn(secret, json.dumps(response.payload))
+        self.assertEqual(self.read_config_file()["workers"], [])
+
+    async def test_create_serverless_worker_without_local_fields_and_store_key_privately(self):
+        secret = "runpod-secret-never-returned"
+        payload = {
+            "worker_id": "pool-1", "name": "GPU pool", "type": "serverless", "provider": "runpod",
+            "endpoint_id": "ep-123", "jobs_per_run": 3, "concurrency": 2,
+            "timeout_seconds": 600, "enabled": True, "api_key": secret,
+        }
+
+        response = await config_routes.update_worker_endpoint(_FakeRequest(payload))
+
+        self.assertEqual(response.status, 200)
+        self.assertTrue(response.payload["api_key_configured"])
+        self.assertNotIn(secret, json.dumps(response.payload))
+        saved = self.read_config_file()["workers"][0]
+        self.assertEqual(saved["endpoint_id"], "ep-123")
+        self.assertEqual(saved["jobs_per_run"], 3)
+        # Legacy field from older browser tabs is accepted but not persisted.
+        self.assertNotIn("concurrency", saved)
+        self.assertEqual(saved["timeout_seconds"], 600)
+        for obsolete in ("host", "port", "cuda_device", "extra_args", "api_key", "clear_api_key"):
+            self.assertNotIn(obsolete, saved)
+        self.assertEqual(self.secrets.get("pool-1"), secret)
+
+    async def test_media_profile_persists_and_is_public_without_changing_credentials(self):
+        for profile in ("legacy-image-v1", "comfy-media-v1"):
+            with self.subTest(profile=profile):
+                response = await config_routes.update_worker_endpoint(_FakeRequest({
+                    "worker_id": "pool-1", "type": "serverless", "endpoint_id": "ep-123",
+                    "media_profile": profile, "api_key": "saved-key", "enabled": False,
+                }))
+                self.assertEqual(response.status, 200)
+                self.assertEqual(self.read_config_file()["workers"][0]["media_profile"], profile)
+                config_module.invalidate_config_cache()
+                public = await config_routes.get_config_endpoint(_FakeRequest())
+                self.assertEqual(public.status, 200)
+                self.assertEqual(public.payload["workers"][0]["media_profile"], profile)
+                self.assertTrue(public.payload["workers"][0]["api_key_configured"])
+                self.assertNotIn("saved-key", json.dumps(public.payload))
+                self.assertEqual(self.secrets.get("pool-1"), "saved-key")
+
+    async def test_old_clients_omit_media_profile_without_downgrading_saved_profile(self):
+        for stored in ({}, {"media_profile": "comfy-media-v1"}):
+            with self.subTest(stored=stored):
+                worker = {"id": "pool-1", "type": "serverless", "endpoint_id": "ep-123", **stored}
+                self.config["workers"] = [worker]
+                self.assertTrue(config_module.save_config(self.config))
+                response = await config_routes.update_worker_endpoint(_FakeRequest({
+                    "worker_id": "pool-1", "name": "Old client save", "jobs_per_run": 2,
+                }))
+                self.assertEqual(response.status, 200)
+                saved = self.read_config_file()["workers"][0]
+                self.assertEqual({k: v for k, v in saved.items() if k == "media_profile"}, stored)
+                # The bulk route is also used by older clients with no profile field.
+                response = await config_routes.update_config_endpoint(_FakeRequest({"workers": [{
+                    "id": "pool-1", "type": "serverless", "endpoint_id": "ep-123", "jobs_per_run": 3,
+                }]}))
+                self.assertEqual(response.status, 200)
+                saved = self.read_config_file()["workers"][0]
+                self.assertEqual({k: v for k, v in saved.items() if k == "media_profile"}, stored)
+
+    async def test_invalid_media_profile_is_rejected_by_worker_and_bulk_routes(self):
+        for profile in ("auto", "comfy-media-v2", None, False, 1, [], {}):
+            with self.subTest(profile=profile):
+                response = await config_routes.update_worker_endpoint(_FakeRequest({
+                    "worker_id": "pool-1", "type": "serverless", "endpoint_id": "ep-123",
+                    "media_profile": profile, "api_key": "never-save",
+                }))
+                self.assertEqual(response.status, 400)
+                response = await config_routes.update_config_endpoint(_FakeRequest({"workers": [{
+                    "id": "pool-1", "type": "serverless", "endpoint_id": "ep-123", "media_profile": profile,
+                }]}))
+                self.assertEqual(response.status, 400)
+                self.assertEqual(self.read_config_file()["workers"], [])
+                self.assertIsNone(self.secrets.get("pool-1"))
+
+    async def test_active_pool_blocks_media_profile_changes_in_both_directions(self):
+        for old_profile, new_profile in ((None, "comfy-media-v1"), ("comfy-media-v1", "legacy-image-v1")):
+            with self.subTest(old_profile=old_profile):
+                worker = {"id": "pool-1", "type": "serverless", "endpoint_id": "ep-123"}
+                if old_profile is not None:
+                    worker["media_profile"] = old_profile
+                self.config["workers"] = [worker]
+                self.assertTrue(config_module.save_config(self.config))
+                original = self.read_config_file()
+                with patch.object(config_routes, "_serverless_pool_is_active", return_value=True):
+                    response = await config_routes.update_worker_endpoint(_FakeRequest({
+                        "worker_id": "pool-1", "media_profile": new_profile,
+                    }))
+                    self.assertEqual(response.status, 409)
+                    response = await config_routes.update_config_endpoint(_FakeRequest({
+                        "workers": [dict(worker, media_profile=new_profile)],
+                    }))
+                    self.assertEqual(response.status, 409)
+                self.assertEqual(self.read_config_file(), original)
+
+    async def test_enable_only_update_preserves_serverless_settings_and_saved_key(self):
+        self.secrets.set("pool-1", "saved-secret")
+        self.config["workers"] = [{
+            "id": "pool-1", "type": "serverless", "provider": "runpod", "name": "Pool",
+            "endpoint_id": "ep-123", "jobs_per_run": 4, "concurrency": 3,
+            "timeout_seconds": 900, "enabled": False,
+        }]
+        self.assertTrue(config_module.save_config(self.config))
+
+        response = await config_routes.update_worker_endpoint(
+            _FakeRequest({"worker_id": "pool-1", "enabled": True})
+        )
+
+        worker = self.read_config_file()["workers"][0]
+        self.assertTrue(worker["enabled"])
+        self.assertEqual(worker["endpoint_id"], "ep-123")
+        self.assertEqual(worker["jobs_per_run"], 4)
+        self.assertNotIn("concurrency", worker)
+        self.assertEqual(worker["timeout_seconds"], 900)
+        self.assertEqual(self.secrets.get("pool-1"), "saved-secret")
+        self.assertTrue(response.payload["api_key_configured"])
+
+    async def test_invalid_serverless_record_is_not_persisted_or_secret_written(self):
+        secret = "must-not-leak"
+        response = await config_routes.update_worker_endpoint(_FakeRequest({
+            "worker_id": "pool-1", "type": "serverless", "provider": "runpod",
+            "endpoint_id": "ep-123", "jobs_per_run": True, "api_key": secret,
+        }))
+
+        self.assertEqual(response.status, 400)
+        self.assertNotIn(secret, json.dumps(response.payload))
+        self.assertEqual(self.read_config_file()["workers"], [])
+        self.assertIsNone(self.secrets.get("pool-1"))
+
+    async def test_json_decode_error_is_generic_and_does_not_echo_body(self):
+        secret = "body-secret-value"
+        response = await config_routes.update_worker_endpoint(
+            _FakeRequest(error=ValueError(f"bad json near {secret}"))
+        )
+        self.assertEqual(response.status, 400)
+        self.assertNotIn(secret, json.dumps(response.payload))
+
+    async def test_bulk_config_rejects_api_keys_at_any_depth_without_persisting(self):
+        secret = "bulk-secret-never-write"
+        response = await config_routes.update_config_endpoint(_FakeRequest({
+            "workers": [{
+                "id": "pool-1", "type": "serverless", "provider": "runpod",
+                "endpoint_id": "ep-123", "nested": {"api_key": secret},
+            }]
+        }))
+
+        self.assertEqual(response.status, 400)
+        self.assertNotIn(secret, json.dumps(response.payload))
+        self.assertEqual(self.read_config_file()["workers"], [])
+        self.assertFalse(self.secrets.path.exists())
+
+    async def test_bulk_update_validates_and_returns_redacted_serverless_config(self):
+        self.secrets.set("pool-1", "saved-secret")
+        response = await config_routes.update_config_endpoint(_FakeRequest({
+            "workers": [{
+                "id": "pool-1", "type": "serverless", "provider": "runpod",
+                "endpoint_id": "ep-123", "name": "Pool", "enabled": True,
+                "host": "obsolete-host", "port": 8188, "cuda_device": 0,
+            }]
+        }))
+
+        self.assertEqual(response.status, 200)
+        self.assertTrue(response.payload["config"]["workers"][0]["api_key_configured"])
+        saved = self.read_config_file()["workers"][0]
+        for obsolete in ("host", "port", "cuda_device"):
+            self.assertNotIn(obsolete, saved)
+        self.assertNotIn("api_key", response.payload["config"]["workers"][0])
+        self.assertEqual(self.secrets.get("pool-1"), "saved-secret")
+
+    async def test_bulk_type_conversion_clears_secret_and_serverless_fields(self):
+        self.secrets.set("pool-1", "bulk-clear-key")
+        self.config["workers"] = [{
+            "id": "pool-1", "type": "serverless", "provider": "runpod", "endpoint_id": "ep-123",
+        }]
+        self.assertTrue(config_module.save_config(self.config))
+
+        response = await config_routes.update_config_endpoint(_FakeRequest({
+            "workers": [{
+                "id": "pool-1", "type": "local", "port": 8188,
+                "provider": "runpod", "endpoint_id": "legacy-ep", "api_key_configured": True,
+                "media_profile": "comfy-media-v1",
+            }]
+        }))
+
+        self.assertEqual(response.status, 200)
+        self.assertIsNone(self.secrets.get("pool-1"))
+        saved = self.read_config_file()["workers"][0]
+        for field in ("provider", "endpoint_id", "api_key_configured", "media_profile"):
+            self.assertNotIn(field, saved)
+
+    async def test_bulk_invalid_serverless_record_does_not_replace_configuration(self):
+        existing = {"id": "local-1", "type": "local", "port": 8188}
+        self.config["workers"] = [existing]
+        self.assertTrue(config_module.save_config(self.config))
+
+        response = await config_routes.update_config_endpoint(_FakeRequest({
+            "workers": [{"id": "pool-1", "type": "serverless", "endpoint_id": "bad/id"}]
+        }))
+
+        self.assertEqual(response.status, 400)
+        self.assertEqual(self.read_config_file()["workers"], [existing])
+
+    async def test_delete_serverless_worker_clears_saved_key(self):
+        self.secrets.set("pool-1", "delete-me")
+        self.config["workers"] = [{
+            "id": "pool-1", "type": "serverless", "provider": "runpod", "endpoint_id": "ep-123"
+        }]
+        self.assertTrue(config_module.save_config(self.config))
+
+        response = await config_routes.delete_worker_endpoint(_FakeRequest({"worker_id": "pool-1"}))
+
+        self.assertEqual(response.status, 200)
+        self.assertEqual(self.read_config_file()["workers"], [])
+        self.assertIsNone(self.secrets.get("pool-1"))
+
+    async def test_type_conversion_clears_secret_and_removes_serverless_fields(self):
+        self.secrets.set("pool-1", "old-key")
+        self.config["workers"] = [{
+            "id": "pool-1", "name": "Pool", "type": "serverless", "provider": "runpod",
+            "endpoint_id": "ep-123", "jobs_per_run": 2, "concurrency": 2, "timeout_seconds": 600,
+            "media_profile": "comfy-media-v1",
+        }]
+        self.assertTrue(config_module.save_config(self.config))
+
+        response = await config_routes.update_worker_endpoint(_FakeRequest({
+            "worker_id": "pool-1", "type": "local", "name": "Local", "port": 8188,
+            "cuda_device": 0, "enabled": False,
+        }))
+
+        self.assertEqual(response.status, 200)
+        self.assertIsNone(self.secrets.get("pool-1"))
+        worker = self.read_config_file()["workers"][0]
+        self.assertEqual(worker["type"], "local")
+        for serverless_field in ("provider", "endpoint_id", "jobs_per_run", "concurrency", "timeout_seconds", "media_profile"):
+            self.assertNotIn(serverless_field, worker)
+
+    async def test_enabled_toggle_allowed_during_active_run_but_endpoint_change_is_rejected(self):
+        self.secrets.set("pool-1", "saved-key")
+        self.config["workers"] = [{
+            "id": "pool-1", "type": "serverless", "provider": "runpod", "endpoint_id": "ep-123",
+            "enabled": False,
+        }]
+        self.assertTrue(config_module.save_config(self.config))
+        runtime_module.manager = types.SimpleNamespace(
+            tasks={"run-1": _RunningTask()},
+            runs={"run-1": {"worker": {"id": "serverless:pool-1:1", "pool_id": "pool-1"}}},
+            pool_is_active=lambda worker_id: worker_id == "pool-1",
+        )
+
+        enabled = await config_routes.update_worker_endpoint(
+            _FakeRequest({"worker_id": "pool-1", "enabled": True})
+        )
+        blocked = await config_routes.update_worker_endpoint(_FakeRequest({
+            "worker_id": "pool-1", "endpoint_id": "ep-456",
+        }))
+
+        self.assertEqual(enabled.status, 200)
+        self.assertEqual(blocked.status, 409)
+        self.assertEqual(self.read_config_file()["workers"][0]["endpoint_id"], "ep-123")
+
+    async def test_blank_api_key_keeps_saved_key_and_clear_api_key_deletes_it(self):
+        self.secrets.set("pool-1", "saved-key")
+        self.config["workers"] = [{
+            "id": "pool-1", "type": "serverless", "provider": "runpod", "endpoint_id": "ep-123",
+        }]
+        self.assertTrue(config_module.save_config(self.config))
+
+        keep = await config_routes.update_worker_endpoint(_FakeRequest({
+            "worker_id": "pool-1", "api_key": "",
+        }))
+        clear = await config_routes.update_worker_endpoint(_FakeRequest({
+            "worker_id": "pool-1", "clear_api_key": True,
+        }))
+
+        self.assertTrue(keep.payload["api_key_configured"])
+        self.assertFalse(clear.payload["api_key_configured"])
+        self.assertIsNone(self.secrets.get("pool-1"))
+
+    async def test_config_save_refuses_inline_api_keys_and_invalid_serverless_records(self):
+        with self.subTest(case="inline key"):
+            config = config_module.get_default_config()
+            config["workers"] = [{"id": "pool-1", "type": "local", "api_key": "private"}]
+            self.assertFalse(config_module.save_config(config))
+        with self.subTest(case="invalid serverless record"):
+            config = config_module.get_default_config()
+            config["workers"] = [{"id": "pool-1", "type": "serverless", "endpoint_id": "bad/id"}]
+            self.assertFalse(config_module.save_config(config))
+
+    async def test_active_pool_blocks_credential_changes_type_conversion_and_deletion(self):
+        self.secrets.set("pool-1", "saved-key")
+        self.config["workers"] = [{
+            "id": "pool-1", "type": "serverless", "provider": "runpod", "endpoint_id": "ep-123",
+        }]
+        self.assertTrue(config_module.save_config(self.config))
+        runtime_module.manager = types.SimpleNamespace(
+            tasks={"run-1": _RunningTask()},
+            runs={"run-1": {"worker": {"id": "serverless:pool-1:1", "pool_id": "pool-1"}}},
+            pool_is_active=lambda worker_id: worker_id == "pool-1",
+        )
+
+        credential_change = await config_routes.update_worker_endpoint(_FakeRequest({
+            "worker_id": "pool-1", "api_key": "new-key",
+        }))
+        type_change = await config_routes.update_worker_endpoint(_FakeRequest({
+            "worker_id": "pool-1", "type": "local", "port": 8188, "cuda_device": 0,
+        }))
+        deletion = await config_routes.delete_worker_endpoint(_FakeRequest({"worker_id": "pool-1"}))
+
+        self.assertEqual(credential_change.status, 409)
+        self.assertEqual(type_change.status, 409)
+        self.assertEqual(deletion.status, 409)
+        self.assertEqual(self.secrets.get("pool-1"), "saved-key")
+        self.assertEqual(self.read_config_file()["workers"][0]["type"], "serverless")
+
+    async def test_secret_write_is_rolled_back_when_config_persistence_fails(self):
+        secret = "write-then-rollback"
+        with patch.object(config_module, "save_config", return_value=False):
+            response = await config_routes.update_worker_endpoint(_FakeRequest({
+                "worker_id": "pool-1", "type": "serverless", "endpoint_id": "ep-123",
+                "api_key": secret,
+            }))
+
+        self.assertEqual(response.status, 500)
+        self.assertIsNone(self.secrets.get("pool-1"))
+        self.assertEqual(self.read_config_file()["workers"], [])
+        self.assertEqual(config_module.load_config()["workers"], [])
+
+    async def test_bulk_secret_delete_is_rolled_back_when_config_persistence_fails(self):
+        self.secrets.set("pool-1", "bulk-restore-key")
+        self.config["workers"] = [{
+            "id": "pool-1", "type": "serverless", "endpoint_id": "ep-123",
+        }]
+        self.assertTrue(config_module.save_config(self.config))
+
+        with patch.object(config_module, "save_config", return_value=False):
+            response = await config_routes.update_config_endpoint(_FakeRequest({
+                "workers": [{"id": "pool-1", "type": "local", "port": 8188}]
+            }))
+
+        self.assertEqual(response.status, 500)
+        self.assertEqual(self.secrets.get("pool-1"), "bulk-restore-key")
+        self.assertEqual(self.read_config_file()["workers"][0]["type"], "serverless")
+
+    async def test_secret_replacement_is_rolled_back_when_config_persistence_fails(self):
+        self.secrets.set("pool-1", "prior-key")
+        self.config["workers"] = [{
+            "id": "pool-1", "type": "serverless", "endpoint_id": "ep-123",
+        }]
+        self.assertTrue(config_module.save_config(self.config))
+
+        with patch.object(config_module, "save_config", return_value=False):
+            response = await config_routes.update_worker_endpoint(_FakeRequest({
+                "worker_id": "pool-1", "api_key": "replacement-key",
+            }))
+
+        self.assertEqual(response.status, 500)
+        self.assertEqual(self.secrets.get("pool-1"), "prior-key")
+        self.assertEqual(self.read_config_file()["workers"][0]["type"], "serverless")
+
+    async def test_delete_secret_is_rolled_back_when_config_persistence_fails(self):
+        self.secrets.set("pool-1", "delete-restore-key")
+        self.config["workers"] = [{
+            "id": "pool-1", "type": "serverless", "endpoint_id": "ep-123",
+        }]
+        self.assertTrue(config_module.save_config(self.config))
+
+        with patch.object(config_module, "save_config", return_value=False):
+            response = await config_routes.delete_worker_endpoint(_FakeRequest({"worker_id": "pool-1"}))
+
+        self.assertEqual(response.status, 500)
+        self.assertEqual(self.secrets.get("pool-1"), "delete-restore-key")
+        self.assertEqual(self.read_config_file()["workers"][0]["id"], "pool-1")
+
+    async def test_regular_worker_delete_does_not_require_posix_secret_storage(self):
+        self.config["workers"] = [{"id": "local-1", "type": "local", "port": 8188}]
+        self.assertTrue(config_module.save_config(self.config))
+        with patch.object(config_routes, "_serverless_secret_store", side_effect=RuntimeError("POSIX only")):
+            response = await config_routes.delete_worker_endpoint(_FakeRequest({"worker_id": "local-1"}))
+        self.assertEqual(response.status, 200)
+        self.assertEqual(self.read_config_file()["workers"], [])
+
+    async def test_enabling_serverless_without_key_is_rejected(self):
+        response = await config_routes.update_worker_endpoint(_FakeRequest({
+            "worker_id": "pool-1", "type": "serverless", "endpoint_id": "ep-123", "enabled": True,
+        }))
+        self.assertEqual(response.status, 400)
+        self.assertEqual(self.read_config_file()["workers"], [])
+
+    async def test_bulk_endpoint_change_is_blocked_during_active_run(self):
+        self.config["workers"] = [{"id": "pool-1", "type": "serverless", "endpoint_id": "ep-123"}]
+        self.assertTrue(config_module.save_config(self.config))
+        runtime_module.manager = types.SimpleNamespace(
+            tasks={"run-1": _RunningTask()},
+            runs={"run-1": {"worker": {"id": "serverless:pool-1:1", "pool_id": "pool-1"}}},
+            pool_is_active=lambda worker_id: worker_id == "pool-1",
+        )
+        updated = copy.deepcopy(self.config["workers"])
+        updated[0]["endpoint_id"] = "ep-other"
+        response = await config_routes.update_config_endpoint(_FakeRequest({"workers": updated}))
+        self.assertEqual(response.status, 409)
+        self.assertEqual(self.read_config_file()["workers"][0]["endpoint_id"], "ep-123")
+
+    async def test_non_serverless_update_behavior_is_preserved(self):
+        self.config["workers"] = [{"id": "local-1", "name": "Local", "type": "local", "port": 8188}]
+        self.assertTrue(config_module.save_config(self.config))
+
+        response = await config_routes.update_worker_endpoint(_FakeRequest({
+            "worker_id": "local-1", "name": "Renamed", "enabled": True,
+        }))
+
+        self.assertEqual(response.status, 200)
+        worker = self.read_config_file()["workers"][0]
+        self.assertEqual(worker["name"], "Renamed")
+        self.assertTrue(worker["enabled"])
+
+    async def test_config_normalization_removes_legacy_local_fields_from_serverless_records(self):
+        raw = {
+            "workers": [{
+                "id": "pool-1", "type": "serverless", "endpoint_id": "ep-123",
+                "host": "localhost", "port": 8188, "cuda_device": 2, "extra_args": "--old",
+            }]
+        }
+        normalized = config_module._merge_with_defaults(raw, config_module.get_default_config())
+        worker = normalized["workers"][0]
+        for obsolete in ("host", "port", "cuda_device", "extra_args"):
+            self.assertNotIn(obsolete, worker)
+
+
+class PersistedPoolGuardTests(_ConfigRouteFixture):
+    """Use the actual manager/ledger across restart, not a mocked idle result."""
+
+    def setUp(self):
+        super().setUp()
+        from test_serverless_runtime import runtime
+        from unittest.mock import AsyncMock
+        self.runtime = runtime
+        self.session_factory = AsyncMock()
+        self.ledger_path = self.root / "jobs.json"
+        self.manager = runtime.ServerlessManager(types.SimpleNamespace(), self.secrets, self.ledger_path, self.session_factory)
+        runtime_module.manager = self.manager
+        self.secrets.set("pool-1", "saved-key")
+        self.config["workers"] = [{"id": "pool-1", "type": "serverless", "endpoint_id": "ep-123", "concurrency": 1}]
+        self.assertTrue(config_module.save_config(self.config))
+
+    async def test_retained_jobs_block_destructive_changes_after_restart(self):
+        original = self.read_config_file()
+        for status in ("UNKNOWN", "CANCEL_REQUESTED", "IN_QUEUE", "IN_PROGRESS", "DOWNLOADING", "SUBMITTING"):
+            for restart in (False, True):
+                with self.subTest(status=status, restart=restart):
+                    self.manager.ledger.update("retained", worker_id="pool-1", endpoint_id="ep-123", status=status, job_id=None if status in {"UNKNOWN", "SUBMITTING"} else "owned-id")
+                    if restart:
+                        self.manager = self.runtime.ServerlessManager(types.SimpleNamespace(), self.secrets, self.ledger_path, self.session_factory)
+                        runtime_module.manager = self.manager
+                    else:
+                        self.manager.tasks["retained"] = types.SimpleNamespace(done=lambda: True)
+                    for fields in ({"api_key": "new-key"}, {"clear_api_key": True}, {"endpoint_id": "ep-other"}, {"type": "local", "port": 8188}, {"media_profile": "comfy-media-v1"}):
+                        response = await config_routes.update_worker_endpoint(_FakeRequest(dict(fields, worker_id="pool-1")))
+                        self.assertEqual(response.status, 409, fields)
+                    response = await config_routes.delete_worker_endpoint(_FakeRequest({"worker_id": "pool-1"}))
+                    self.assertEqual(response.status, 409)
+                    worker = original["workers"][0]
+                    for workers in ([], [dict(worker, type="local")], [dict(worker, endpoint_id="ep-other")], [dict(worker, media_profile="comfy-media-v1")]):
+                        response = await config_routes.update_config_endpoint(_FakeRequest({"workers": workers}))
+                        self.assertEqual(response.status, 409, workers)
+                    self.assertEqual(self.read_config_file(), original)
+                    self.assertEqual(self.secrets.get("pool-1"), "saved-key")
+        self.session_factory.assert_not_awaited()
+
+    async def test_retained_jobs_allow_same_or_omitted_media_profile(self):
+        self.manager.ledger.update("retained", worker_id="pool-1", endpoint_id="ep-123", status="UNKNOWN")
+        for profile in ("legacy-image-v1", "comfy-media-v1"):
+            with self.subTest(profile=profile):
+                worker = {"id": "pool-1", "type": "serverless", "endpoint_id": "ep-123"}
+                # An absent legacy field and an explicit legacy selection are equivalent.
+                if profile != "legacy-image-v1":
+                    worker["media_profile"] = profile
+                self.config["workers"] = [worker]
+                self.assertTrue(config_module.save_config(self.config))
+                for fields in ({"media_profile": profile}, {}):
+                    response = await config_routes.update_worker_endpoint(_FakeRequest({
+                        "worker_id": "pool-1", "name": "Same profile", **fields,
+                    }))
+                    self.assertEqual(response.status, 200)
+                    response = await config_routes.update_config_endpoint(_FakeRequest({"workers": [{
+                        "id": "pool-1", "type": "serverless", "endpoint_id": "ep-123", **fields,
+                    }]}))
+                    self.assertEqual(response.status, 200)
+                    self.assertEqual(self.read_config_file()["workers"][0]["media_profile"], profile)
+        self.session_factory.assert_not_awaited()
+
+    async def test_retained_jobs_allow_images_per_run_and_legacy_concurrency_saves(self):
+        # Upgrading from a config with concurrency=1 must not trip the busy guard,
+        # and the image count stays editable while old jobs are unresolved.
+        self.manager.ledger.update("retained", worker_id="pool-1", endpoint_id="ep-123", status="UNKNOWN")
+        response = await config_routes.update_worker_endpoint(_FakeRequest({
+            "worker_id": "pool-1", "endpoint_id": "ep-123", "jobs_per_run": 2, "concurrency": 2,
+        }))
+        self.assertEqual(response.status, 200)
+        saved = self.read_config_file()["workers"][0]
+        self.assertEqual(saved["jobs_per_run"], 2)
+        self.assertNotIn("concurrency", saved)
+        response = await config_routes.update_config_endpoint(_FakeRequest({"workers": [dict(saved, jobs_per_run=3, concurrency=5)]}))
+        self.assertEqual(response.status, 200)
+        self.assertEqual(self.read_config_file()["workers"][0]["jobs_per_run"], 3)
+
+    async def test_retained_job_allows_disable_and_rename_but_not_key_clear(self):
+        self.manager.ledger.update("retained", worker_id="pool-1", status="UNKNOWN")
+        response = await config_routes.update_worker_endpoint(_FakeRequest({"worker_id": "pool-1", "enabled": False, "name": "Paused"}))
+        self.assertEqual(response.status, 200)
+        self.assertEqual(self.read_config_file()["workers"][0]["name"], "Paused")
+        response = await config_routes.update_worker_endpoint(_FakeRequest({"worker_id": "pool-1", "clear_api_key": True}))
+        self.assertEqual(response.status, 409)
+        self.assertEqual(self.secrets.get("pool-1"), "saved-key")
+
+    async def test_terminal_or_other_pool_records_do_not_block_deletion(self):
+        self.manager.ledger.update("done", worker_id="pool-1", status="RESOLVED")
+        self.manager.ledger.update("other", worker_id="other", status="UNKNOWN")
+        response = await config_routes.delete_worker_endpoint(_FakeRequest({"worker_id": "pool-1"}))
+        self.assertEqual(response.status, 200)
+        self.assertIsNone(self.secrets.get("pool-1"))
+
+
+if __name__ == "__main__":
+    unittest.main()

tests/test_serverless_runtime.py

@@ -0,0 +1,170 @@
+import asyncio
+import importlib.util
+import sys
+import tempfile
+import types
+import unittest
+from pathlib import Path
+from unittest.mock import AsyncMock, patch
+
+ROOT = Path(__file__).resolve().parents[1]
+PACKAGE = "dist_serverless_runtime_tests"
+for name, path in [(PACKAGE, ROOT), (PACKAGE + ".api", ROOT / "api"), (PACKAGE + ".api.orchestration", ROOT / "api/orchestration")]:
+    module = types.ModuleType(name)
+    module.__path__ = [str(path)]
+    sys.modules[name] = module
+
+def load(name):
+    spec = importlib.util.spec_from_file_location(PACKAGE + ".api.orchestration." + name, ROOT / "api/orchestration" / (name + ".py"))
+    assert spec and spec.loader
+    module = importlib.util.module_from_spec(spec)
+    sys.modules[spec.name] = module
+    spec.loader.exec_module(module)
+    return module
+
+runtime = load("serverless_runtime")
+
+class Secrets:
+    def get(self, worker_id):
+        return "private-test-key"
+
+class RuntimeTests(unittest.IsolatedAsyncioTestCase):
+    async def asyncSetUp(self):
+        self.temp = tempfile.TemporaryDirectory()
+        self.server = types.SimpleNamespace(distributed_pending_jobs={"collector-job": asyncio.Queue()}, distributed_jobs_lock=asyncio.Lock())
+        self.manager = runtime.ServerlessManager(self.server, Secrets(), Path(self.temp.name) / "jobs.json", AsyncMock(return_value=object()))
+        self.manager.poll_seconds = 0
+        self.worker = {"id": "pool:1", "pool_id": "pool", "endpoint_id": "ep1", "timeout_seconds": 30, "type": "serverless"}
+        self.prepared = {"worker": self.worker, "workflow": {"1": {}}, "images": [], "manifest": [{"job_id": "collector-job", "prefix": "dist_x", "kind": "image"}]}
+
+    async def asyncTearDown(self):
+        await self.manager.cancel()
+        self.temp.cleanup()
+
+    async def test_job_id_persisted_before_status_and_results_injected(self):
+        api = types.SimpleNamespace(submit=AsyncMock(return_value="provider-job"), cancel=AsyncMock())
+        async def status(job_id):
+            self.assertEqual(self.manager.ledger.all()[0]["job_id"], "provider-job")
+            return {"status": "COMPLETED", "output": {"images": []}}
+        api.status = status
+        with patch.object(runtime, "RunpodClient", return_value=api), patch.object(runtime, "decode_results", AsyncMock(return_value={"collector-job": ["tensor"]})):
+            await self.manager.start("execution", "prompt", [self.prepared])
+            await self.manager.wait_for_execution("execution")
+        self.assertEqual(api.submit.await_count, 1)
+        item = self.server.distributed_pending_jobs["collector-job"].get_nowait()
+        self.assertEqual(item["worker_id"], "pool:1")
+        self.assertEqual(item["tensor"], "tensor")
+        self.assertTrue(item["is_last"])
+        self.assertEqual(self.manager.ledger.all()[0]["status"], "COMPLETED")
+
+    async def test_output_error_marks_failed_and_delivers_explicit_error(self):
+        api = types.SimpleNamespace(submit=AsyncMock(return_value="provider-job"), status=AsyncMock(return_value={"status": "COMPLETED", "output": {"error": "private-test-key"}}), cancel=AsyncMock())
+        with patch.object(runtime, "RunpodClient", return_value=api):
+            await self.manager.start("execution", "prompt", [self.prepared])
+            await self.manager.wait_for_execution("execution")
+        record = self.manager.ledger.all()[0]
+        self.assertEqual(record["status"], "FAILED")
+        item = self.server.distributed_pending_jobs["collector-job"].get_nowait()
+        self.assertIn("serverless_error", item)
+        self.assertNotIn("private-test-key", str(record) + str(item))
+
+    async def test_ambiguous_submit_never_retries(self):
+        api = types.SimpleNamespace(submit=AsyncMock(side_effect=runtime.AmbiguousSubmission("Unknown submission")), cancel=AsyncMock())
+        with patch.object(runtime, "RunpodClient", return_value=api):
+            await self.manager.start("execution", "prompt", [self.prepared])
+            await self.manager.wait_for_execution("execution")
+        self.assertEqual(api.submit.await_count, 1)
+        self.assertEqual(self.manager.ledger.all()[0]["status"], "UNKNOWN")
+
+    async def test_cancel_targets_owned_job_only(self):
+        polled = asyncio.Event()
+        async def status(_job):
+            polled.set()
+            await asyncio.Event().wait()
+        api = types.SimpleNamespace(submit=AsyncMock(return_value="owned-job"), status=status, cancel=AsyncMock(return_value={"status": "CANCELLED"}))
+        with patch.object(runtime, "RunpodClient", return_value=api):
+            await self.manager.start("execution", "prompt", [self.prepared])
+            await polled.wait()
+            await self.manager.cancel(worker_id="other-pool")
+            api.cancel.assert_not_awaited()
+            await self.manager.cancel(worker_id="pool")
+        api.cancel.assert_awaited_once_with("owned-job")
+
+    async def test_all_images_of_a_run_submit_at_once_ignoring_legacy_concurrency(self):
+        self.assertEqual(runtime.PER_WORKER_JOB_LIMIT, 16)
+        gate = asyncio.Event()
+        submitted = []
+        async def submit(*args):
+            submitted.append(args)
+            return f"provider-{len(submitted)}"
+        async def status(_job):
+            await gate.wait()
+            return {"status": "COMPLETED", "output": {"images": []}}
+        api = types.SimpleNamespace(submit=submit, status=status, cancel=AsyncMock())
+        # A config saved before the setting was removed still carries concurrency=1.
+        legacy = dict(self.worker, concurrency=1)
+        jobs = [dict(self.prepared, worker=dict(legacy, id=f"pool:{i}")) for i in range(1, 4)]
+        with patch.object(runtime, "RunpodClient", return_value=api), patch.object(runtime, "decode_results", AsyncMock(return_value={"collector-job": ["tensor"]})):
+            await self.manager.start("execution", "prompt", jobs)
+            for _ in range(50):
+                if len(submitted) == 3:
+                    break
+                await asyncio.sleep(0)
+            self.assertEqual(len(submitted), 3)
+            gate.set()
+            await self.manager.wait_for_execution("execution")
+
+    async def test_per_pool_job_limit_bounds_submission(self):
+        patcher = patch.object(runtime, "PER_WORKER_JOB_LIMIT", 1)
+        patcher.start()
+        self.addCleanup(patcher.stop)
+        gate = asyncio.Event()
+        reached = asyncio.Event()
+        count = 0
+        async def submit(*args):
+            nonlocal count
+            count += 1
+            reached.set()
+            return f"provider-{count}"
+        async def status(_job):
+            await gate.wait()
+            return {"status": "COMPLETED", "output": {"images": []}}
+        api = types.SimpleNamespace(submit=submit, status=status, cancel=AsyncMock())
+        second = dict(self.prepared, worker=dict(self.worker, id="pool:2"))
+        with patch.object(runtime, "RunpodClient", return_value=api), patch.object(runtime, "decode_results", AsyncMock(return_value={"collector-job": ["tensor"]})):
+            await self.manager.start("execution", "prompt", [self.prepared, second])
+            await reached.wait()
+            await asyncio.sleep(0)
+            self.assertEqual(count, 1)
+            gate.set()
+            await self.manager.wait_for_execution("execution")
+        self.assertEqual(count, 2)
+
+    async def test_cancel_persisted_job_after_master_restart(self):
+        self.manager.ledger.update("old", worker_id="pool", endpoint_id="old-endpoint", job_id="old-job", status="IN_PROGRESS", created_at=1, updated_at=1)
+        api = types.SimpleNamespace(cancel=AsyncMock(return_value={"status": "CANCELLED"}))
+        with patch.object(runtime, "RunpodClient", return_value=api) as client:
+            await self.manager.cancel(worker_id="pool")
+        api.cancel.assert_awaited_once_with("old-job")
+        self.assertEqual(client.call_args.args[1], "old-endpoint")
+        self.assertEqual(self.manager.ledger.all()[0]["status"], "CANCELLED")
+
+    async def test_master_failure_cancels_job_without_partial_success(self):
+        self.server.prompt_queue = types.SimpleNamespace(get_history=lambda _id: {"prompt": {"status": {"status_str": "error"}}})
+        api = types.SimpleNamespace(submit=AsyncMock(return_value="job"), cancel=AsyncMock(return_value={"status": "CANCELLED"}))
+        with patch.object(runtime, "RunpodClient", return_value=api):
+            await self.manager.start("execution", "prompt", [self.prepared])
+            await self.manager.wait_for_execution("execution")
+        api.submit.assert_not_awaited()
+        self.assertEqual(self.manager.ledger.all()[0]["status"], "FAILED")
+
+    async def test_deadline_includes_local_slot_wait(self):
+        self.prepared["worker"] = dict(self.worker, timeout_seconds=0.01)
+        api = types.SimpleNamespace(submit=AsyncMock(return_value="job"), status=AsyncMock(side_effect=lambda _: asyncio.sleep(5)), cancel=AsyncMock(return_value={"status": "CANCELLED"}))
+        # A blocked slot expires without a billable submission.
+        self.manager.pool_limits["pool"] = asyncio.Semaphore(0)
+        with patch.object(runtime, "RunpodClient", return_value=api):
+            await self.manager.start("execution", "prompt", [self.prepared])
+            await self.manager.wait_for_execution("execution")
+        api.submit.assert_not_awaited()
+        self.assertEqual(self.manager.ledger.all()[0]["status"], "TIMED_OUT")

tests/test_serverless_status.py

@@ -0,0 +1,98 @@
+import importlib.util
+import json
+import sys
+import types
+import unittest
+from pathlib import Path
+from unittest.mock import AsyncMock, patch
+
+ROOT = Path(__file__).resolve().parents[1]
+PKG = "dist_serverless_status_tests"
+
+
+class StatusTests(unittest.IsolatedAsyncioTestCase):
+    async def asyncSetUp(self):
+        def mod(name, **kwargs):
+            result = types.ModuleType(name)
+            result.__dict__.update(kwargs)
+            return result
+        self.manager = types.SimpleNamespace(health=AsyncMock(return_value={"ok": True, "queueRemaining": 0, "message": "Ready (scales from zero)"}), public_jobs=lambda _: [], cancel=AsyncMock(return_value=1))
+        self.pool = {"id": "pool", "type": "serverless", "endpoint_id": "ep"}
+        modules = {PKG: mod(PKG, __path__=[str(ROOT)]), PKG + ".api": mod("api", __path__=[str(ROOT/'api')]),
+                   PKG + ".utils": mod("utils", __path__=[]),
+                   PKG + ".utils.config": mod("config", load_config=lambda: {"workers": [self.pool]}),
+                   PKG + ".api.orchestration": mod("orchestration", __path__=[str(ROOT/'api/orchestration')]),
+                   PKG + ".api.orchestration.serverless_runtime": mod("runtime", get_manager=lambda: self.manager),
+                   "server": mod("server", PromptServer=types.SimpleNamespace(instance=types.SimpleNamespace(routes=types.SimpleNamespace(get=lambda _: lambda f: f, post=lambda _: lambda f: f))))}
+        with patch.dict(sys.modules, modules):
+            spec = importlib.util.spec_from_file_location(PKG+'.api.serverless_routes', ROOT/'api/serverless_routes.py')
+            self.routes = importlib.util.module_from_spec(spec)
+            spec.loader.exec_module(self.routes)
+
+    async def test_health_only_never_submits_and_supports_scale_to_zero(self):
+        request = types.SimpleNamespace(json=AsyncMock(return_value={"worker_id": "pool"}))
+        result = await self.routes.test_serverless_endpoint(request)
+        self.assertEqual(result.status, 200)
+        self.assertTrue(json.loads(result.text)["ok"])
+        self.manager.health.assert_awaited_once_with(self.pool, fresh=True)
+        self.manager.cancel.assert_not_awaited()
+
+    async def test_unknown_pool_cannot_cancel_other_jobs(self):
+        request = types.SimpleNamespace(json=AsyncMock(return_value={"worker_id": "other"}))
+        result = await self.routes.cancel_serverless_endpoint(request)
+        self.assertEqual(result.status, 404)
+        self.manager.cancel.assert_not_awaited()
+
+    async def test_transport_error_does_not_leak_credentials(self):
+        self.manager.health.side_effect = RuntimeError("Bearer secret-token https://signed.invalid/?token=value")
+        request = types.SimpleNamespace(match_info={"worker_id": "pool"})
+        result = await self.routes.serverless_status_endpoint(request)
+        self.assertFalse(json.loads(result.text)["ok"])
+        self.assertNotIn("secret-token", result.text)
+        self.assertNotIn("signed.invalid", result.text)
+
+    async def test_cancel_uses_only_requested_pool(self):
+        request = types.SimpleNamespace(json=AsyncMock(return_value={"worker_id": "pool"}))
+        result = await self.routes.cancel_serverless_endpoint(request)
+        self.assertEqual(result.status, 200)
+        self.manager.cancel.assert_awaited_once_with(worker_id="pool")
+
+    async def test_reconcile_only_reads_the_requested_pools_jobs(self):
+        self.manager.reconcile = AsyncMock(return_value=2)
+        request = types.SimpleNamespace(json=AsyncMock(return_value={"worker_id": "pool"}))
+        result = await self.routes.reconcile_serverless_endpoint(request)
+        self.assertEqual(result.status, 200)
+        self.assertEqual(json.loads(result.text)["resolved"], 2)
+        self.manager.reconcile.assert_awaited_once_with(worker_id="pool")
+        self.manager.cancel.assert_not_awaited()
+
+    async def test_resolve_requires_explicit_boolean_confirmation_and_record_id(self):
+        self.manager.acknowledge_recovery = AsyncMock()
+        for payload in ({"worker_id": "pool", "record_id": "record"},
+                        {"worker_id": "pool", "record_id": "record", "provider_checked": "true"},
+                        {"worker_id": "pool", "provider_checked": True}):
+            with self.subTest(payload=payload):
+                result = await self.routes.resolve_serverless_endpoint(types.SimpleNamespace(json=AsyncMock(return_value=payload)))
+                self.assertEqual(result.status, 400)
+        self.manager.acknowledge_recovery.assert_not_awaited()
+
+    async def test_resolve_is_scoped_and_does_not_claim_provider_cancellation(self):
+        self.manager.acknowledge_recovery = AsyncMock()
+        request = types.SimpleNamespace(json=AsyncMock(return_value={"worker_id": "pool", "record_id": "record", "provider_checked": True}))
+        result = await self.routes.resolve_serverless_endpoint(request)
+        self.assertEqual(result.status, 200)
+        self.manager.acknowledge_recovery.assert_awaited_once_with("pool", "record", provider_checked=True)
+        self.assertIn("not provider-confirmed", json.loads(result.text)["message"])
+        self.manager.cancel.assert_not_awaited()
+
+    async def test_live_or_mismatched_record_refuses_manual_resolution(self):
+        self.manager.acknowledge_recovery = AsyncMock(side_effect=ValueError("Unresolved inactive job not found for this pool"))
+        request = types.SimpleNamespace(json=AsyncMock(return_value={"worker_id": "pool", "record_id": "record", "provider_checked": True}))
+        result = await self.routes.resolve_serverless_endpoint(request)
+        self.assertEqual(result.status, 409)
+
+    async def test_failed_health_still_exposes_recovery_ids(self):
+        self.manager.health.side_effect = RuntimeError("private transport error")
+        self.manager.public_jobs = lambda _: [{"record_id": "record", "status": "UNKNOWN"}]
+        result = await self.routes.serverless_status_endpoint(types.SimpleNamespace(match_info={"worker_id": "pool"}))
+        self.assertEqual(json.loads(result.text)["jobs"][0]["record_id"], "record")

tests/test_serverless_typed_inputs.py

@@ -0,0 +1,330 @@
+"""Bounded input packing; native acceptance is opt-in and CPU-only."""
+import asyncio
+import base64
+import copy
+import importlib
+import io
+import os
+from pathlib import Path
+import subprocess
+import sys
+import types
+import wave
+from unittest.mock import Mock
+
+import pytest
+
+ROOT = Path(__file__).resolve().parents[1]
+PACKAGE = "_typed_inputs_tests"
+for suffix in ("", ".api", ".api.orchestration", ".serverless_bridge"):
+    package = types.ModuleType(PACKAGE + suffix)
+    package.__path__ = [str(ROOT.joinpath(*suffix.strip(".").split(".")))] if suffix else [str(ROOT)]
+    sys.modules[package.__name__] = package
+packer = importlib.import_module(PACKAGE + ".api.orchestration.serverless_typed_inputs")
+codec = importlib.import_module(PACKAGE + ".serverless_bridge.codec")
+
+
+def node(kind, **inputs):
+    return {"class_type": kind, "inputs": inputs}
+
+
+def graph(kind="LoadAudio", filename="tone.wav"):
+    key = "audio" if kind == "LoadAudio" else "file"
+    socket = "audio" if kind == "LoadAudio" else "video"
+    return {"source": node(kind, **{key: filename}),
+            "consumer": node("Consumer", **{socket: ["source", 0]})}
+
+
+def run(workflow):
+    return asyncio.run(packer.prepare_media_inputs(workflow))
+
+
+@pytest.fixture
+def paths(tmp_path, monkeypatch):
+    roots = {name: tmp_path / name for name in ("input", "output", "temp")}
+    for path in roots.values():
+        path.mkdir()
+    def resolve(value):
+        for name in roots:
+            marker = " [" + name + "]"
+            if value.endswith(marker):
+                return str(roots[name] / value[:-len(marker)])
+        return str(roots["input"] / value)
+    folders = types.SimpleNamespace(get_annotated_filepath=resolve)
+    for name, path in roots.items():
+        setattr(folders, "get_" + name + "_directory", lambda path=path: str(path))
+    monkeypatch.setitem(sys.modules, "folder_paths", folders)
+    return roots
+
+
+@pytest.mark.parametrize("kind,key", [("LoadAudio", "audio"), ("LoadVideo", "file")])
+@pytest.mark.parametrize("bad", [{}, None, [], {"filename": "a.wav"},
+    {"audio": ["filename", 0]}, {"file": ["filename", 0]}])
+def test_rejects_non_native_or_linked_inputs_before_read(kind, key, bad, monkeypatch):
+    workflow = graph(kind)
+    workflow["source"]["inputs"] = bad
+    read = Mock(side_effect=AssertionError("must not read"))
+    monkeypatch.setattr(packer.video_codec, "_read_path", read)
+    with pytest.raises(ValueError):
+        run(workflow)
+    read.assert_not_called()
+
+
+@pytest.mark.parametrize("kind,key", [("LoadAudio", "audio"), ("LoadVideo", "file")])
+@pytest.mark.parametrize("extra", ["start_time", "duration", "sample_rate", "crop", "upload"])
+def test_rejects_extra_loader_controls(kind, key, extra):
+    workflow = graph(kind)
+    workflow["source"]["inputs"][extra] = 0
+    with pytest.raises(ValueError, match="controls|inputs"):
+        run(workflow)
+
+
+@pytest.mark.parametrize("slot", [1, -1, True, 0.0, "0", None])
+def test_rejects_nonzero_or_noninteger_output_slot(slot):
+    workflow = graph()
+    workflow["consumer"]["inputs"]["audio"] = ["source", slot]
+    with pytest.raises(ValueError, match="output|slot"):
+        run(workflow)
+
+
+def test_rejects_unused_loader_and_malformed_link():
+    workflow = graph()
+    workflow.pop("consumer")
+    with pytest.raises(ValueError, match="consumer"):
+        run(workflow)
+    workflow = graph()
+    workflow["consumer"]["inputs"]["audio"].append(0)
+    with pytest.raises(ValueError):
+        run(workflow)
+
+
+@pytest.mark.parametrize("key,value", [("audio", "private.wav"), ("file", "private.mp4"), ("video", "private.mov")])
+def test_unknown_media_loaders_still_use_strict_scan(key, value):
+    with pytest.raises(ValueError):
+        run({"source": node("CustomLoader", **{key: value})})
+
+
+def test_images_still_rewrite_only_the_private_graph(paths):
+    payload = b"legacy staging preserves the original file bytes"
+    (paths["input"] / "picture.png").write_bytes(payload)
+    source = {"image": node("LoadImage", image="picture.png"),
+              "export": node("MediaExportV1", images=["image", 0])}
+    workflow = copy.deepcopy(source)
+    images, artifacts, refs = run(workflow)
+    assert artifacts == refs == []
+    assert len(images) == 1
+    assert base64.b64decode(images[0]["image"]) == payload
+    assert workflow["image"]["inputs"]["image"] == images[0]["name"]
+    assert workflow["export"] == source["export"]
+    assert source["image"]["inputs"]["image"] == "picture.png"
+
+
+if os.environ.get("_DISTRIBUTED_TYPED_INPUT_CORE_TEST") != "1":
+    def test_native_input_acceptance_cpu_subprocess():
+        configured = os.environ.get("COMFYUI_SOURCE_ROOT")
+        if configured is None:
+            pytest.skip("Set COMFYUI_SOURCE_ROOT for CPU native input acceptance")
+        core = Path(configured).expanduser().resolve()
+        assert configured and (core / "comfy_extras/nodes_audio.py").is_file()
+        env = dict(os.environ)
+        scratch = Path(os.environ.get("TMPDIR", str(Path.home() / ".hermes/cache/scratch")))
+        scratch.mkdir(parents=True, exist_ok=True)
+        env.pop("PYTHONPATH", None)
+        env.pop("COMFYUI_SOURCE_ROOT", None)
+        env.update(TMPDIR=str(scratch), CUDA_VISIBLE_DEVICES="", HIP_VISIBLE_DEVICES="",
+                   PYTHONDONTWRITEBYTECODE="1", _DISTRIBUTED_TYPED_INPUT_CORE_TEST="1")
+        bootstrap = """
+import sys
+sys.path.insert(0, sys.argv[1])
+import comfy.cli_args
+comfy.cli_args.args.cpu = True
+import torch
+def forbidden(*args, **kwargs):
+    raise AssertionError('Input packing must not query or initialize a GPU')
+torch.cuda.is_available = lambda: False
+torch.cuda.device_count = lambda: 0
+for name in ('init', '_lazy_init', 'current_device'):
+    setattr(torch.cuda, name, forbidden)
+import comfy_extras.nodes_audio
+import comfy_api.latest._input_impl.video_types
+for name in ('is_available', 'device_count'):
+    setattr(torch.cuda, name, forbidden)
+import pytest
+raise SystemExit(pytest.main([sys.argv[2], '-q', '-p', 'no:cacheprovider']))
+"""
+        result = subprocess.run([sys.executable, "-c", bootstrap, str(core), str(Path(__file__).resolve())],
+                                cwd=ROOT, env=env, text=True, capture_output=True, timeout=150)
+        print(result.stdout)
+        assert result.returncode == 0, result.stdout + result.stderr
+else:
+    import av
+    import numpy as np
+    import torch
+    from comfy_extras import nodes_audio
+    from comfy_api.latest._input_impl.video_types import VideoFromFile
+
+    def wav_bytes(channels=2, sample_width=2, rate=22050, variant=0):
+        values = np.array([-32768, -12001 + variant, -1, 0, 1, 8193, 32767] * channels,
+                          dtype="<i2" if sample_width == 2 else "<i4")
+        buffer = io.BytesIO()
+        with wave.open(buffer, "wb") as out:
+            out.setnchannels(channels)
+            out.setsampwidth(sample_width)
+            out.setframerate(rate)
+            out.writeframes(values.tobytes())
+        return buffer.getvalue()
+
+    def unpack(artifact):
+        payload = base64.b64decode(artifact["transport"]["data_base64"])
+        assert artifact["byte_length"] == len(payload)
+        return payload
+
+    @pytest.mark.parametrize("root", ["input", "output", "temp"])
+    @pytest.mark.parametrize("channels,width", [(1, 2), (2, 2), (2, 4)])
+    def test_exact_native_audio_and_every_consumer_ref(paths, root, channels, width):
+        source = paths[root] / "tone.wav"
+        source.write_bytes(wav_bytes(channels, width))
+        workflow = graph(filename="tone.wav [" + root + "]")
+        workflow["export"] = node("MediaExportV1", audio=["source", 0])
+        workflow["consumer"]["inputs"]["second_audio"] = ["source", 0]
+        before = copy.deepcopy(workflow)
+        images, artifacts, refs = run(workflow)
+        assert images == [] and len(artifacts) == 1
+        artifact = artifacts[0]
+        expected, rate = nodes_audio.load(str(source))
+        actual = codec.decode_payload("AUDIO", artifact["descriptor"], unpack(artifact))
+        assert actual["sample_rate"] == rate == 22050
+        assert actual["waveform"].dtype == expected.dtype
+        assert torch.equal(actual["waveform"].view(torch.uint8), expected.unsqueeze(0).contiguous().view(torch.uint8))
+        assert refs == [{"node_id": "consumer", "input_name": "audio", "artifact_id": artifact["id"]},
+                        {"node_id": "consumer", "input_name": "second_audio", "artifact_id": artifact["id"]},
+                        {"node_id": "export", "input_name": "audio", "artifact_id": artifact["id"]}]
+        assert workflow == before
+
+    @pytest.mark.parametrize("root", ["input", "output", "temp"])
+    def test_exact_video_bytes_without_container_decode(paths, root, monkeypatch):
+        payload = b"opaque native video bytes\x00\xff"
+        (paths[root] / "clip.mp4").write_bytes(payload)
+        workflow = graph("LoadVideo", "clip.mp4 [" + root + "]")
+        workflow["export"] = node("MediaExportV1", video=["source", 0])
+        before = copy.deepcopy(workflow)
+        def forbidden(*args, **kwargs):
+            raise AssertionError("must not open, decode or transcode video")
+        monkeypatch.setattr(av, "open", forbidden)
+        monkeypatch.setattr(VideoFromFile, "get_components", forbidden)
+        images, artifacts, refs = run(workflow)
+        assert images == [] and len(artifacts) == 1
+        artifact = artifacts[0]
+        assert unpack(artifact) == payload
+        assert artifact["descriptor"] == {"encoding": "native-video-v1", "representation": "file",
+            "byte_length": len(payload), "start_time": 0, "duration": 0, "crop": None,
+            "data": {"offset": 0, "length": len(payload)}}
+        assert refs == [{"node_id": name, "input_name": "video", "artifact_id": artifact["id"]}
+                        for name in ("consumer", "export")]
+        native = packer.video_codec.decode_video(artifact["descriptor"], unpack(artifact))
+        assert vars(native)["_VideoFromFile__file"].getvalue() == payload
+        assert workflow == before
+
+    def test_content_dedupe_kind_separation_and_mixed_images(paths):
+        data = wav_bytes()
+        (paths["input"] / "a.wav").write_bytes(data)
+        (paths["output"] / "b.wav").write_bytes(data)
+        (paths["input"] / "image.png").write_bytes(b"image bytes")
+        workflow = graph(filename="a.wav")
+        workflow["same"] = node("LoadAudio", audio="b.wav [output]")
+        workflow["video"] = node("LoadVideo", file="a.wav")
+        workflow["image"] = node("LoadImage", image="image.png")
+        workflow["export"] = node("MediaExportV1", audio=["same", 0], video=["video", 0], images=["image", 0])
+        before = copy.deepcopy(workflow)
+        images, artifacts, refs = run(workflow)
+        assert len(images) == 1 and len(artifacts) == 2 and len(refs) == 3
+        assert refs[0]["artifact_id"] == refs[1]["artifact_id"] != refs[2]["artifact_id"]
+        assert {a["kind"] for a in artifacts} == {"AUDIO", "VIDEO"}
+        before["image"]["inputs"]["image"] = images[0]["name"]
+        assert workflow == before
+        second = graph(filename="b.wav [output]")
+        assert run(second)[1][0]["id"] == artifacts[0]["id"]
+
+    @pytest.mark.parametrize("kind", ["LoadAudio", "LoadVideo"])
+    def test_path_rejections_and_per_file_bounds(paths, tmp_path, monkeypatch, kind):
+        outside = tmp_path / "outside.wav"
+        outside.write_bytes(wav_bytes())
+        (paths["input"] / "escape.wav").symlink_to(outside)
+        (paths["input"] / "dirlink").symlink_to(tmp_path, target_is_directory=True)
+        os.mkfifo(paths["input"] / "fifo.wav")
+        values = ["../outside.wav", str(outside), "escape.wav", "dirlink/outside.wav", "fifo.wav",
+                  "absent.wav", "tone.wav [unknown]", "https://example.com/a.wav", "input [temp]", "C:\\a.wav"]
+        for value in values:
+            with pytest.raises(ValueError):
+                run(graph(kind, value))
+        (paths["input"] / "large.wav").write_bytes(wav_bytes())
+        monkeypatch.setattr(packer.contract, "MAX_ARTIFACT_BYTES", 8)
+        with pytest.raises(ValueError, match="limit|size"):
+            run(graph(kind, "large.wav"))
+
+    def test_aggregate_decoded_budget_and_duplicate_not_double_charged(paths, monkeypatch):
+        payload = wav_bytes()
+        (paths["input"] / "a.wav").write_bytes(payload)
+        (paths["input"] / "b.wav").write_bytes(wav_bytes(variant=1))
+        workflow = graph(filename="a.wav")
+        workflow["other"] = node("LoadAudio", audio="a.wav")
+        workflow["export"] = node("MediaExportV1", audio=["other", 0])
+        # File source bytes have a separate aggregate bound; use a longer WAV
+        # whose float output dominates its tiny header.
+        with wave.open(io.BytesIO(payload), "rb") as wav:
+            frames = wav.readframes(wav.getnframes())
+        buffer = io.BytesIO()
+        with wave.open(buffer, "wb") as wav:
+            wav.setparams((2, 2, 22050, 0, "NONE", "not compressed"))
+            wav.writeframes(frames * 100)
+        (paths["input"] / "a.wav").write_bytes(buffer.getvalue())
+        artifact = run(graph(filename="a.wav"))[1][0]
+        monkeypatch.setattr(packer.contract, "MAX_TOTAL_BYTES", artifact["byte_length"])
+        assert len(run(workflow)[1]) == 1
+        workflow["other"]["inputs"]["audio"] = "b.wav"
+        with pytest.raises(ValueError, match="limit|size|budget"):
+            run(workflow)
+
+    def test_audio_charged_before_ndarray_cat_or_conversion(monkeypatch):
+        frame = types.SimpleNamespace(samples=100, sample_rate=22050,
+            layout=types.SimpleNamespace(channels=[1, 2]),
+            format=types.SimpleNamespace(name="s16", bytes=2),
+            planes=[types.SimpleNamespace(buffer_size=400)], to_ndarray=Mock())
+        stream = types.SimpleNamespace(index=0, channels=2,
+                                      codec_context=types.SimpleNamespace(sample_rate=22050))
+        container = Mock()
+        container.__enter__ = Mock(return_value=container)
+        container.__exit__ = Mock(return_value=False)
+        container.streams = types.SimpleNamespace(audio=[stream])
+        container.decode = Mock(return_value=iter([frame]))
+        monkeypatch.setattr(av, "open", Mock(return_value=container))
+        cat = Mock(side_effect=AssertionError("no accumulation allowed"))
+        monkeypatch.setattr(torch, "cat", cat)
+        with pytest.raises(ValueError, match="limit|size|budget"):
+            packer._decode_audio(b"input", 799)
+        frame.to_ndarray.assert_not_called()
+        cat.assert_not_called()
+        frame.samples = 1
+        frame.planes[0].buffer_size = 800
+        container.decode.return_value = iter([frame])
+        with pytest.raises(ValueError, match="limit|size|budget"):
+            packer._decode_audio(b"input", 799)
+        frame.to_ndarray.assert_not_called()
+
+    def test_audio_rejects_empty_invalid_and_channel_or_rate_drift(paths, monkeypatch):
+        (paths["input"] / "bad.wav").write_bytes(b"not audio")
+        with pytest.raises(ValueError):
+            run(graph(filename="bad.wav"))
+        stream = types.SimpleNamespace(index=0, channels=2, codec_context=types.SimpleNamespace(sample_rate=22050))
+        for frames in ([], [types.SimpleNamespace(samples=1, sample_rate=44100,
+                     layout=types.SimpleNamespace(channels=[1, 2]))],
+                     [types.SimpleNamespace(samples=1, sample_rate=22050,
+                     layout=types.SimpleNamespace(channels=[1]))]):
+            container = Mock()
+            container.__enter__ = Mock(return_value=container)
+            container.__exit__ = Mock(return_value=False)
+            container.streams = types.SimpleNamespace(audio=[stream])
+            container.decode = Mock(return_value=iter(frames))
+            monkeypatch.setattr(av, "open", Mock(return_value=container))
+            with pytest.raises(ValueError):
+                packer._decode_audio(b"input", 1000)

tests/test_serverless_typed_runtime.py

@@ -0,0 +1,319 @@
+"""Offline typed submission/delivery regressions; no provider or GPU access."""
+import asyncio
+import copy
+import importlib
+import json
+import socket
+import tempfile
+import types
+import unittest
+from pathlib import Path
+from unittest.mock import AsyncMock, patch
+
+# Load once before integration fixtures snapshot/restore sys.modules. PyTorch
+# cannot be re-imported after those fixtures unload its extension-backed modules.
+import torch
+
+from test_runpod_client import Response, Session
+from test_serverless_runtime import runtime, Secrets
+
+client = importlib.import_module(runtime.__package__ + ".runpod_client")
+bridge = runtime.__package__.split(".")[0] + ".serverless_bridge"
+contract = importlib.import_module(bridge + ".contract")
+
+
+def prepared(sockets=("texts",)):
+    worker = {"id": "pool:1", "pool_id": "pool", "endpoint_id": "ep1",
+              "timeout_seconds": 30, "type": "serverless", "media_profile": contract.PROFILE}
+    bindings = [dict(binding_id=f"binding-{i}", collector_id="collector", job_id="collector-job",
+                     participant_id=worker["id"], node_id=f"export-{i}", socket=name,
+                     kind=contract.SOCKET_KINDS[name]) for i, name in enumerate(sockets)]
+    media = dict(version=1, profile=contract.PROFILE, request_id="request-1",
+                 artifacts=[], inputs=[], outputs=copy.deepcopy(bindings))
+    graph = {"source": {"class_type": "Source", "inputs": {"value": "café雪"}}}
+    for binding in bindings:
+        graph[binding["node_id"]] = {"class_type": "MediaExportV1", "inputs": {
+            "request_id": media["request_id"], "binding_json": json.dumps(binding),
+            binding["socket"]: ["source", 0]}}
+    return dict(worker=worker, workflow=graph, images=[], manifest=bindings, media=media)
+
+
+def response(job):
+    envelope = dict(version=1, profile=contract.PROFILE, request_id=job["media"]["request_id"],
+                    bindings=[], artifacts=[], items=[])
+    for index, binding in enumerate(job["manifest"]):
+        envelope["bindings"].append(dict(binding, count=2))
+        for item_index, text in enumerate(("", " \n雪")):
+            aid = f"artifact-{index}-{item_index}"
+            descriptor, payload = {"encoding": "utf8-v1"}, text.encode()
+            if binding["kind"] != "STRING":
+                codec = importlib.import_module(bridge + ".codec")
+                values = {"IMAGE": torch.arange(8, dtype=torch.float64).reshape(1, 1, 2, 4),
+                          "MASK": torch.arange(2, dtype=torch.float16).reshape(1, 1, 2),
+                          "AUDIO": {"waveform": torch.arange(4, dtype=torch.float32).reshape(1, 2, 2), "sample_rate": 48000}}
+                descriptor, payload = codec.encode_payload(binding["kind"], values[binding["kind"]])
+            artifact = contract.artifact_from_payload(aid, binding["kind"], descriptor, payload)
+            envelope["artifacts"].append(artifact)
+            envelope["items"].insert(0, dict(binding_id=binding["binding_id"], item_index=item_index, artifact_id=aid))
+    return {"media": envelope}
+
+
+class TypedClientTests(unittest.IsolatedAsyncioTestCase):
+    async def test_exact_typed_bytes_and_legacy_unchanged(self):
+        job = prepared()
+        session = Session([Response(payload={"id": "job-1"})])
+        api = client.RunpodClient(session, "ep1", "test-only")
+        await api.submit(job["workflow"], [], 30, media=job["media"])
+        expected = {"input": {"workflow": job["workflow"], "images": [], "media": job["media"]},
+                    "policy": {"executionTimeout": 30000, "ttl": 30000}}
+        self.assertEqual(session.calls[0][2]["data"], json.dumps(expected, allow_nan=False,
+                         ensure_ascii=False, separators=(",", ":")).encode())
+        self.assertEqual(len(session.calls), 1)
+        expected["input"].pop("media")
+        legacy = json.dumps(expected, allow_nan=False, ensure_ascii=False, separators=(",", ":")).encode()
+        self.assertEqual(client.serialize_request(job["workflow"], [], 30), legacy)
+        self.assertEqual(client.serialize_request(job["workflow"], [], 30, media=None), legacy)
+
+    async def test_invalid_envelopes_never_post(self):
+        for field, value in [("version", True), ("profile", "legacy-image-v1"),
+                             ("request_id", "../bad"), ("outputs", []), ("inputs", [{}])]:
+            with self.subTest(field=field):
+                job = prepared()
+                job["media"][field] = value
+                session = Session([])
+                with self.assertRaises(ValueError):
+                    await client.RunpodClient(session, "ep1", "test-only").submit(
+                        job["workflow"], [], 30, media=job["media"])
+                self.assertFalse(session.calls)
+        session = Session([])
+        with self.assertRaises(ValueError):
+            await client.RunpodClient(session, "ep1", "test-only").submit({}, [], 30, media={})
+        self.assertFalse(session.calls)
+
+    async def test_typed_size_boundary_uses_exact_body_before_post(self):
+        job = prepared()
+        size = len(client.serialize_request(job["workflow"], [], 30, media=job["media"]))
+        session = Session([Response(payload={"id": "boundary"})])
+        api = client.RunpodClient(session, "ep1", "test-only")
+        with patch.object(client, "MAX_REQUEST_BYTES", size):
+            await api.submit(job["workflow"], [], 30, media=job["media"])
+            job["workflow"]["source"]["inputs"]["value"] += "x"
+            with self.assertRaisesRegex(ValueError, "Serialized serverless request body"):
+                await api.submit(job["workflow"], [], 30, media=job["media"])
+        self.assertEqual(len(session.calls), 1)
+
+
+class TypedRuntimeTests(unittest.IsolatedAsyncioTestCase):
+    async def asyncSetUp(self):
+        self.temp = tempfile.TemporaryDirectory()
+        self.server = types.SimpleNamespace(distributed_pending_jobs={"collector-job": asyncio.Queue()},
+                                            distributed_jobs_lock=asyncio.Lock())
+        self.manager = runtime.ServerlessManager(self.server, Secrets(), Path(self.temp.name) / "jobs.json",
+                                                 AsyncMock(return_value=object()))
+        self.manager.poll_seconds = 0
+        self.job = prepared()
+        self.api = types.SimpleNamespace(submit=AsyncMock(return_value="provider-job"),
+            status=AsyncMock(return_value={"status": "COMPLETED", "output": response(self.job)}),
+            cancel=AsyncMock(return_value={"status": "CANCELLED"}))
+        self.client_patch = patch.object(runtime, "RunpodClient", return_value=self.api)
+        self.client_patch.start()
+
+    async def asyncTearDown(self):
+        await self.manager.cancel()
+        self.client_patch.stop()
+        self.temp.cleanup()
+
+    async def run_job(self):
+        await self.manager.start("execution", "prompt", [self.job])
+        await self.manager.wait_for_execution("execution")
+
+    async def test_preflight_all_jobs_before_any_journal_or_task(self):
+        for mutation in ("profile", "missing_media", "manifest", "participant", "exporter", "empty_media"):
+            with self.subTest(mutation=mutation):
+                bad = prepared()
+                if mutation == "profile":
+                    bad["worker"]["media_profile"] = "legacy-image-v1"
+                elif mutation == "missing_media":
+                    bad.pop("media")
+                elif mutation == "manifest":
+                    bad["manifest"][0]["job_id"] = "other"
+                elif mutation == "participant":
+                    bad["worker"]["id"] = "other-participant"
+                elif mutation == "exporter":
+                    bad["workflow"]["export-0"]["inputs"]["request_id"] = "other"
+                else:
+                    bad["media"] = {}
+                with self.assertRaises(ValueError):
+                    await self.manager.start("execution", "prompt", [self.job, bad])
+                self.assertFalse(self.manager.ledger.all())
+                self.assertFalse(self.manager.tasks)
+                self.assertFalse(self.manager.runs)
+                self.api.submit.assert_not_awaited()
+
+    async def test_text_roundtrip_id_journal_and_payload_cleanup(self):
+        async def status(_job):
+            self.assertEqual(self.manager.ledger.all()[0]["job_id"], "provider-job")
+            return {"status": "COMPLETED", "output": response(self.job)}
+        self.api.status.side_effect = status
+        await self.run_job()
+        self.api.submit.assert_awaited_once_with(self.job["workflow"], [], 30, media=self.job["media"])
+        queue = self.server.distributed_pending_jobs["collector-job"]
+        self.assertEqual(queue.get_nowait(), {"worker_id": "pool:1", "media": {"texts": ["", " \n雪"]}, "is_last": True})
+        self.assertTrue(queue.empty())
+        self.assertEqual(self.manager.ledger.all()[0]["status"], "COMPLETED")
+        run = next(iter(self.manager.runs.values()))
+        for field in ("workflow", "images", "media"):
+            self.assertNotIn(field, run)
+        self.assertNotIn("data_base64", self.manager.ledger.path.read_text())
+        self.api.cancel.assert_not_awaited()
+
+    async def test_all_sockets_grouped_without_native_value_conversion(self):
+        self.job = prepared(tuple(contract.SOCKET_KINDS))
+        native = {binding["binding_id"]: [object(), {"native": object()}] for binding in self.job["manifest"]}
+        with patch.object(contract, "decode_response", AsyncMock(return_value=native)) as decode:
+            await self.run_job()
+        decode.assert_awaited_once()
+        self.assertEqual(decode.call_args.kwargs["request_id"], "request-1")
+        self.assertEqual(decode.call_args.kwargs["bindings"], self.job["manifest"])
+        queue = self.server.distributed_pending_jobs["collector-job"]
+        item = queue.get_nowait()
+        self.assertEqual(set(item), {"worker_id", "media", "is_last"})
+        self.assertEqual(set(item["media"]), set(contract.SOCKET_KINDS))
+        for binding in self.job["manifest"]:
+            for actual, expected in zip(item["media"][binding["socket"]], native[binding["binding_id"]]):
+                self.assertIs(actual, expected)
+        self.assertTrue(queue.empty())
+
+    async def test_identity_and_late_decode_failure_deliver_only_one_error(self):
+        self.job = prepared(("texts", "images"))
+        for mutation in ("request", "binding", "checksum", "descriptor"):
+            with self.subTest(mutation=mutation):
+                output = response(self.job)
+                if mutation == "request":
+                    output["media"]["request_id"] = "other"
+                elif mutation == "binding":
+                    output["media"]["bindings"][-1]["participant_id"] = "other"
+                elif mutation == "checksum":
+                    output["media"]["artifacts"][-1]["sha256"] = "0" * 64
+                else:
+                    output["media"]["artifacts"][-1]["descriptor"] = {}
+                self.api.status.return_value = {"status": "COMPLETED", "output": output}
+                await self.run_job()
+                queue = self.server.distributed_pending_jobs["collector-job"]
+                item = queue.get_nowait()
+                self.assertIn("serverless_error", item)
+                self.assertNotIn("media", item)
+                self.assertTrue(queue.empty())
+        self.api.cancel.assert_not_awaited()
+
+    async def test_closed_queue_is_not_resurrected(self):
+        self.server.distributed_pending_jobs.clear()
+        await self.run_job()
+        self.assertFalse(self.server.distributed_pending_jobs)
+        self.assertEqual(self.manager.ledger.all()[0]["status"], "COMPLETED")
+
+    async def test_real_tensor_audio_and_mask_codecs_preserve_dtype_and_shape(self):
+        self.job = prepared(("texts", "images", "audio", "masks"))
+        self.api.status.return_value = {"status": "COMPLETED", "output": response(self.job)}
+        await self.run_job()
+        result = self.server.distributed_pending_jobs["collector-job"].get_nowait()["media"]
+        self.assertEqual(result["texts"], ["", " \n雪"])
+        self.assertEqual(result["images"][0].dtype, torch.float64)
+        self.assertEqual(result["images"][0].shape, (1, 1, 2, 4))
+        self.assertEqual(result["masks"][0].dtype, torch.float16)
+        self.assertEqual(result["audio"][0]["sample_rate"], 48000)
+        self.assertTrue(torch.equal(result["audio"][0]["waveform"], torch.arange(4, dtype=torch.float32).reshape(1, 2, 2)))
+        self.assertTrue(all(len(values) == 2 for values in result.values()))
+
+    async def test_run_rechecks_profile_before_submission(self):
+        await self.manager.start("execution", "prompt", [self.job])
+        self.job["worker"]["media_profile"] = "legacy-image-v1"
+        await self.manager.wait_for_execution("execution")
+        self.api.submit.assert_not_awaited()
+        self.assertEqual(self.manager.ledger.all()[0]["status"], "FAILED")
+
+    async def test_multiple_collectors_receive_one_message_each(self):
+        self.job = prepared(("texts", "images"))
+        self.job["manifest"][-1]["job_id"] = "second-job"
+        self.job["media"]["outputs"] = copy.deepcopy(self.job["manifest"])
+        self.job["workflow"]["export-1"]["inputs"]["binding_json"] = json.dumps(self.job["manifest"][-1])
+        self.server.distributed_pending_jobs["second-job"] = asyncio.Queue()
+        self.api.status.return_value = {"status": "COMPLETED", "output": response(self.job)}
+        await self.run_job()
+        for job_id, name in (("collector-job", "texts"), ("second-job", "images")):
+            queue = self.server.distributed_pending_jobs[job_id]
+            self.assertEqual(queue.qsize(), 1)
+            item = queue.get_nowait()
+            self.assertEqual(set(item["media"]), {name})
+            self.assertTrue(item["is_last"])
+
+    async def test_prestart_cancel_releases_media_and_deduplicates_errors(self):
+        self.job = prepared(("texts", "images"))
+        await self.manager.start("execution", "prompt", [self.job])
+        await self.manager.cancel(execution_id="execution")
+        self.api.submit.assert_not_awaited()
+        run = next(iter(self.manager.runs.values()))
+        self.assertNotIn("media", run)
+        self.assertEqual(self.server.distributed_pending_jobs["collector-job"].qsize(), 1)
+        self.assertEqual(self.manager.ledger.all()[0]["status"], "CANCELLED")
+
+    async def test_ambiguous_submit_keeps_reservation_without_media(self):
+        self.api.submit.side_effect = runtime.AmbiguousSubmission("Unknown submission")
+        await self.run_job()
+        self.api.submit.assert_awaited_once()
+        self.api.cancel.assert_not_awaited()
+        self.assertEqual(self.manager.ledger.all()[0]["status"], "UNKNOWN")
+        self.assertTrue(self.manager.pool_is_active("pool"))
+        self.assertNotIn("media", next(iter(self.manager.runs.values())))
+
+    async def test_legacy_still_uses_tensor_messages_and_no_media_kwarg(self):
+        self.job.pop("media")
+        self.job["worker"]["media_profile"] = "legacy-image-v1"
+        self.job["manifest"] = [{"job_id": "collector-job", "prefix": "dist_x", "kind": "image"}]
+        with patch.object(runtime, "decode_results", AsyncMock(return_value={"collector-job": ["a", "b"]})):
+            await self.run_job()
+        self.api.submit.assert_awaited_once_with(self.job["workflow"], [], 30)
+        queue = self.server.distributed_pending_jobs["collector-job"]
+        self.assertEqual([queue.get_nowait(), queue.get_nowait()], [
+            {"worker_id": "pool:1", "tensor": "a", "image_index": 0, "is_last": False},
+            {"worker_id": "pool:1", "tensor": "b", "image_index": 1, "is_last": True}])
+
+
+class TypedDownloadTests(unittest.IsolatedAsyncioTestCase):
+    async def test_stream_bound_cap_and_redirect_policy_without_credentials(self):
+        typed = importlib.import_module(runtime.__package__ + ".serverless_typed_results")
+        class DownloadResponse(Response):
+            headers = {}
+            async def chunks(self, _size):
+                yield b"ab"
+                yield b"cd"
+            @property
+            def content(self):
+                return types.SimpleNamespace(iter_chunked=self.chunks)
+        class DownloadSession(Response):
+            def get(self, url, **kwargs):
+                self.calls.append((url, kwargs))
+                return self.response
+        session = DownloadSession()
+        session.calls = []
+        session.response = DownloadResponse()
+        with patch.object(typed.aiohttp, "TCPConnector", return_value=object()), \
+                patch.object(typed.aiohttp, "ClientSession", return_value=session) as factory:
+            self.assertEqual(await typed.download_public_artifact("https://example.com/a", max_bytes=contract.MAX_ARTIFACT_BYTES), b"abcd")
+            self.assertFalse(factory.call_args.kwargs["trust_env"])
+            self.assertNotIn("headers", factory.call_args.kwargs)
+            self.assertFalse(session.calls[0][1]["allow_redirects"])
+            with self.assertRaises(ValueError):
+                await typed.download_public_artifact("https://example.com/a", max_bytes=3)
+            session.response.status = 302
+            session.response.headers = {"Location": "https://127.0.0.1/private"}
+            with self.assertRaises(ValueError):
+                await typed.download_public_artifact("https://example.com/a", max_bytes=4)
+        for url in ("http://example.com/a", "https://127.0.0.1/a", "https://user:pass@example.com/a"):
+            with self.assertRaises(ValueError):
+                await typed.download_public_artifact(url, max_bytes=4)
+        resolver = typed._PublicResolver()
+        with patch.object(resolver, "_getaddrinfo", AsyncMock(return_value=[
+                (socket.AF_INET, socket.SOCK_STREAM, 6, "", ("127.0.0.1", 443))])):
+            with self.assertRaises(OSError):
+                await resolver.resolve("example.com", 443)

tests/test_transfer_codec.py

@@ -133,6 +133,7 @@ def test_quality_to_crf_mapping():
 
 def test_normalize_format_rejects_unknown_values():
     assert tc.normalize_format(" H264 ") == "h264"
+    assert tc.normalize_format(" EXACT ") == "exact"
     assert tc.normalize_format(None) == tc.DEFAULT_TRANSFER_FORMAT
     with pytest.raises(tc.TransferCodecError):
         tc.normalize_format("gif")
@@ -158,6 +159,22 @@ def test_decode_chunk_validates_envelope_against_payload():
         tc.decode_chunk({**video_env, "count": 5}, video["parts"])
 
 
+def test_exact_is_appended_without_changing_saved_format_defaults():
+    assert tc.TRANSFER_FORMATS == (
+        "legacy_png", "png", "ffv1", "jpeg", "webp", "h264", "h265", "av1", "exact",
+    )
+    assert tc.DEFAULT_TRANSFER_FORMAT == "legacy_png"
+    assert tc.EXACT_FORMAT not in tc.IMAGE_FORMATS + tc.VIDEO_FORMATS
+
+
+def test_exact_does_not_enter_uint8_chunk_codec():
+    frames = tc.tensor_batch_to_uint8(_smooth_frames(count=1))
+    with pytest.raises(tc.TransferCodecError, match="exact"):
+        list(tc.encode_chunks(frames, "exact", 90, 10**9))
+    with pytest.raises(tc.TransferCodecError, match="exact"):
+        tc.decode_chunk({"format": "exact", "count": 1, "width": 91, "height": 67}, [b"raw tensor"])
+
+
 def test_legacy_format_is_not_chunk_encoded():
     frames = tc.tensor_batch_to_uint8(_smooth_frames(count=1))
     with pytest.raises(tc.TransferCodecError):

tests/test_v3_collector_transfer.py

@@ -69,6 +69,8 @@ def v3(monkeypatch):
     module(f"{PKG}.nodes.utilities")
     module(f"{PKG}.nodes.distributed_upscale", UltimateSDUpscaleDistributed=object)
     module(f"{PKG}.nodes.collector", DistributedCollectorNode=CollectorRuntime)
+    module(f"{PKG}.nodes.image_list", DistributedImageListStart=object, DistributedImageListEnd=object,
+           DistributedImageListGetValue=object)
     module("comfy", __path__=[])
     module("comfy.samplers", KSampler=types.SimpleNamespace(SAMPLERS=["euler"], SCHEDULERS=["normal"]))
     module("comfy_api", __path__=[])
@@ -96,6 +98,7 @@ def test_registered_collector_schema_exposes_optional_transfer_inputs(v3):
     fmt = inputs["transfer_format"]
     assert fmt.kind == "Combo"
     assert fmt.kwargs["options"] == list(codec.TRANSFER_FORMATS)
+    assert fmt.kwargs["options"][-1] == "exact"
     assert fmt.kwargs["default"] == "legacy_png"
     assert fmt.kwargs["optional"] is True
     quality = inputs["transfer_quality"]
@@ -105,13 +108,14 @@ def test_registered_collector_schema_exposes_optional_transfer_inputs(v3):
     assert schema.is_input_list is True
 
 
-def test_registered_collector_forwards_transfer_options(v3):
+@pytest.mark.parametrize("fmt", ["av1", "exact"])
+def test_registered_collector_forwards_transfer_options(v3, fmt):
     mod, _, calls = v3
 
-    mod.DistributedCollector.execute(images=["img"], audio=[None], transfer_format=["av1"], transfer_quality=[70],
+    mod.DistributedCollector.execute(images=["img"], audio=[None], transfer_format=[fmt], transfer_quality=[70],
                                      multi_job_id=["job"], is_worker=[True])
 
-    assert calls[-1]["transfer_format"] == ["av1"]
+    assert calls[-1]["transfer_format"] == [fmt]
     assert calls[-1]["transfer_quality"] == [70]
     assert calls[-1]["multi_job_id"] == ["job"]
 

tests/test_v3_migration.py

@@ -1,4 +1,5 @@
 """Real-framework acceptance; opt in with COMFYUI_SOURCE_ROOT."""
+import json
 import os
 from pathlib import Path
 import subprocess
@@ -19,3 +20,32 @@ def test_native_v3_runtime():
     )
     assert result.returncode == 0, result.stdout + '\n' + result.stderr
     assert 'V3_ACCEPTANCE_OK' in result.stdout
+
+
+def test_image_list_network_runtime():
+    comfy_root = os.environ.get('COMFYUI_SOURCE_ROOT')
+    if not comfy_root:
+        pytest.skip('Set COMFYUI_SOURCE_ROOT for model-free multi-process HTTP acceptance')
+    helper = Path(__file__).parent / 'helpers' / 'image_list_network_check.py'
+    result = subprocess.run(
+        [sys.executable, str(helper), str(Path(comfy_root).resolve())],
+        capture_output=True, text=True, timeout=180,
+        env={**os.environ, 'COMFYUI_IS_WORKER': '1'},
+    )
+    assert result.returncode == 0, result.stdout + '\n' + result.stderr
+    assert 'IMAGE_LIST_NETWORK_OK' in result.stdout
+    assert 'IMAGE_LIST_VALUES_OK' in result.stdout
+    assert 'IMAGE_LIST_CANCEL_OK' in result.stdout
+    assert 'IMAGE_LIST_MASTER_OK' in result.stdout
+    assert 'IMAGE_LIST_PREFLIGHT_OK' in result.stdout
+    assert 'IMAGE_LIST_SEQUENTIAL_OK' in result.stdout
+    assert 'IMAGE_LIST_FAILURE_OK' in result.stdout
+    reports = [json.loads(line.removeprefix('IMAGE_LIST_TARGETS_OK '))
+               for line in result.stdout.splitlines() if line.startswith('IMAGE_LIST_TARGETS_OK ')]
+    assert len(reports) == 1, result.stdout
+    assert reports[0] == {
+        'modes': ['delegate-only', 'mixed', 'master-only'],
+        'outer_targets': ['4', '5'], 'worker_targets': ['4'],
+        'required_utility_once_per_item': True, 'preview_discarded': True,
+        'saved_prompt_unchanged': True, 'capture_once_per_outer_prompt': True,
+    }

utils/async_helpers.py

@@ -105,24 +105,63 @@ class PromptValidationError(RuntimeError):
         super().__init__(f"Invalid prompt: {merged}")
 
 
+def validate_execution_targets(prompt, targets):
+    """Reject malformed explicit roots instead of falling back to all outputs."""
+    if (not isinstance(targets, list) or not targets
+            or any(not isinstance(node_id, str) or not node_id or node_id not in prompt
+                   for node_id in targets)
+            or len(set(targets)) != len(targets)):
+        raise PromptValidationError({
+            "type": "invalid_partial_execution_targets",
+            "message": "partial_execution_targets must be a non-empty list of unique existing node IDs",
+            "details": "", "extra_info": {},
+        })
+    return list(targets)
+
+
 async def queue_prompt_payload(
     prompt_obj,
     workflow_meta=None,
     client_id=None,
     include_queue_metadata=False,
+    *,
+    partial_execution_targets=None,
 ):
-    """Validate and queue a prompt via ComfyUI's prompt queue."""
+    """Validate and queue a prompt, preserving any explicitly requested roots."""
+    targets = (validate_execution_targets(prompt_obj, partial_execution_targets)
+               if partial_execution_targets is not None else None)
     payload = {"prompt": prompt_obj}
+    if targets is not None:
+        payload["partial_execution_targets"] = list(targets)
     payload = prompt_server.trigger_on_prompt(payload)
     prompt = payload["prompt"]
+    if targets is not None:
+        # Hooks may replace the envelope or mutate its target list. The caller's
+        # requested roots remain authoritative, but must still exist afterwards.
+        validate_execution_targets(prompt, targets)
 
     prompt_id = str(uuid.uuid4())
-    valid = await execution.validate_prompt(prompt_id, prompt, None)
+    valid = await execution.validate_prompt(prompt_id, prompt, targets)
     if not valid[0]:
         error_payload = valid[1] if len(valid) > 1 else "Prompt outputs failed validation"
         node_errors = valid[3] if len(valid) > 3 else {}
         raise PromptValidationError(error_payload, node_errors)
 
+    if targets is not None:
+        node_errors = valid[3] if len(valid) > 3 else {}
+        target_errors = {
+            node_id: entry for node_id, entry in node_errors.items()
+            if node_id in targets or (isinstance(entry, dict)
+                                      and set(entry.get("dependent_outputs", [])) & set(targets))
+        }
+        if set(valid[2]) != set(targets) or target_errors:
+            raise PromptValidationError({
+                "type": "partial_execution_targets_failed_validation",
+                "message": "Not all requested execution targets passed validation",
+                "details": "Refusing to queue a subset or additional output roots",
+                "extra_info": {"requested_targets": targets},
+            }, node_errors)
+
     extra_data = {"create_time": int(time.time() * 1000)}
     if workflow_meta:
         extra_data.setdefault("extra_pnginfo", {})["workflow"] = workflow_meta

utils/config.py

@@ -2,10 +2,13 @@
 Configuration management for ComfyUI-Distributed.
 """
 import asyncio
+import copy
 import os
 import json
+import math
 from contextlib import asynccontextmanager
 from .logging import log
+from .secret_fields import is_secret_field_name as _is_secret_field_name
 
 # Import defaults for timeout fallbacks
 from .constants import HEARTBEAT_TIMEOUT
@@ -15,6 +18,52 @@ _config_cache = None
 _config_mtime = 0.0
 _config_lock = asyncio.Lock()
 
+_SERVERLESS_OBSOLETE_WORKER_FIELDS = {"host", "port", "listen_port", "cuda_device", "extra_args"}
+
+
+def _contains_serverless_secret(value):
+    if isinstance(value, dict):
+        return any(
+            (isinstance(key, str) and _is_secret_field_name(key))
+            or _contains_serverless_secret(item)
+            for key, item in value.items()
+        )
+    if isinstance(value, (list, tuple)):
+        return any(_contains_serverless_secret(item) for item in value)
+    return False
+
+
+def _normalize_serverless_workers(config):
+    """Remove obsolete local-worker fields without adding local defaults."""
+    if not isinstance(config, dict):
+        return config
+    workers = config.get("workers")
+    if isinstance(workers, list):
+        for worker in workers:
+            if isinstance(worker, dict) and worker.get("type") == "serverless":
+                for field in _SERVERLESS_OBSOLETE_WORKER_FIELDS:
+                    worker.pop(field, None)
+    return config
+
+
+def _validated_config_for_save(config):
+    """Return a safe normalized copy ready for gpu_config.json."""
+    if _contains_serverless_secret(config):
+        raise ValueError("API keys must be stored outside the config file")
+    saved = _normalize_serverless_workers(copy.deepcopy(config))
+    workers = saved.get("workers") if isinstance(saved, dict) else None
+    if isinstance(workers, list) and any(
+        isinstance(worker, dict) and worker.get("type") == "serverless" for worker in workers
+    ):
+        from .serverless_config import validate_serverless_worker
+        for index, worker in enumerate(workers):
+            if isinstance(worker, dict) and worker.get("type") == "serverless":
+                saved_worker = validate_serverless_worker(worker)
+                for field in _SERVERLESS_OBSOLETE_WORKER_FIELDS:
+                    saved_worker.pop(field, None)
+                workers[index] = saved_worker
+    return saved
+
 
 def _config_path():
     return CONFIG_FILE
@@ -33,7 +82,8 @@ def get_default_config():
             "worker_probe_concurrency": 8,
             "worker_prep_concurrency": 4,
             "media_sync_concurrency": 2,
-            "media_sync_timeout_seconds": 120
+            "media_sync_timeout_seconds": 120,
+            "image_list_job_timeout_seconds": 0.0
         },
         "tunnel": {
             "status": "stopped",
@@ -62,7 +112,7 @@ def _merge_with_defaults(data, defaults):
         if key not in merged:
             merged[key] = value
 
-    return merged
+    return _normalize_serverless_workers(merged)
 
 
 def invalidate_config_cache():
@@ -97,11 +147,12 @@ def load_config():
     return _config_cache
 
 def save_config(config):
-    """Saves the configuration to file."""
+    """Saves normalized configuration without allowing API keys in the config file."""
     tmp_path = f"{_config_path()}.tmp"
     try:
+        saved_config = _validated_config_for_save(config)
         with open(tmp_path, 'w', encoding='utf-8') as f:
-            json.dump(config, f, indent=2)
+            json.dump(saved_config, f, indent=2)
             f.flush()
             os.fsync(f.fileno())
         os.replace(tmp_path, _config_path())
@@ -121,12 +172,21 @@ async def config_transaction():
     """Acquire config lock, yield loaded config, and save if changed."""
     async with _config_lock:
         config = load_config()
+        original_config = copy.deepcopy(config)
         original_snapshot = json.dumps(config, sort_keys=True)
-        yield config
+        try:
+            yield config
+        except BaseException:
+            if isinstance(config, dict):
+                config.clear()
+                config.update(original_config)
+            raise
         updated_snapshot = json.dumps(config, sort_keys=True)
-        if updated_snapshot != original_snapshot:
-            if not save_config(config):
-                raise RuntimeError("Failed to save config")
+        if updated_snapshot != original_snapshot and not save_config(config):
+            if isinstance(config, dict):
+                config.clear()
+                config.update(original_config)
+            raise RuntimeError("Failed to save config")
 
 def ensure_config_exists():
     """Creates default config file if it doesn't exist. Used by __init__.py"""
@@ -157,6 +217,37 @@ def get_worker_timeout_seconds(default: int = HEARTBEAT_TIMEOUT) -> int:
         return max(1, int(default))
 
 
+def get_image_list_job_timeout_seconds(config=None) -> float:
+    """Return the validated image-list total-job timeout in seconds.
+
+    A value of zero disables the total-job deadline. Missing settings default
+    to zero; configured values must be finite, non-negative numbers.
+    """
+    setting_key = "settings.image_list_job_timeout_seconds"
+    cfg = load_config() if config is None else config
+    if not isinstance(cfg, dict):
+        raise ValueError(f"{setting_key} must be a finite non-negative number")
+
+    settings = cfg.get("settings", {})
+    if not isinstance(settings, dict):
+        raise ValueError(f"{setting_key} must be a finite non-negative number")
+
+    value = settings.get("image_list_job_timeout_seconds", 0.0)
+    if isinstance(value, bool) or not isinstance(value, (int, float, str)):
+        raise ValueError(f"{setting_key} must be a finite non-negative number")
+    if isinstance(value, str) and not value.strip():
+        raise ValueError(f"{setting_key} must be a finite non-negative number")
+
+    try:
+        timeout = float(value)
+    except (OverflowError, ValueError):
+        raise ValueError(f"{setting_key} must be a finite non-negative number") from None
+
+    if not math.isfinite(timeout) or timeout < 0:
+        raise ValueError(f"{setting_key} must be a finite non-negative number")
+    return timeout
+
+
 def is_master_delegate_only() -> bool:
     """Returns True when master should skip local workload and act as orchestrator only."""
     try:

utils/network.py

@@ -87,6 +87,8 @@ def _split_host_and_port(host):
 
 def build_worker_url(worker, endpoint=""):
     """Construct the worker base URL with optional endpoint."""
+    if worker.get("type") == "serverless":
+        raise ValueError("Serverless workers have no ComfyUI worker URL")
     host = (worker.get("host") or "").strip()
     port = int(worker.get("port", worker.get("listen_port", 8188)) or 8188)