Merge branch 'feat/serverless-multimodal' into feat/image-list-serverless-base

feat: add explicit image return codecs and quality controls
fix: preserve mixed-mode preflight across execution planner integration
2026-10-14 00:37:48 +00:00 · 2026-10-14 00:37:42 +00:00 · 2026-10-13 23:56:35 +00:00 · 2026-10-13 23:55:26 +00:00 · 2026-10-13 23:53:58 +00:00 · 2026-10-13 22:38:10 +00:00
211 changed files with 46850 additions and 7847 deletions
@@ -1,4 +1,3 @@
 # These are supported funding model platforms

 github: robertvoy
-buy_me_a_coffee: robertvoy
@@ -19,6 +19,6 @@ jobs:
      - name: Check out code
        uses: actions/checkout@v4
      - name: Publish Custom Node
-        uses: Comfy-Org/publish-node-action@v1
+        uses: Comfy-Org/publish-node-action@main
        with:
          personal_access_token: ${{ secrets.REGISTRY_ACCESS_TOKEN }}
@@ -0,0 +1,9 @@
+bin/
+logs/
+gpu_config.json
+__pycache__/
+**/__pycache__/
+*.py[cod]
+node_modules/
+npm-debug.log*
+AGENTS.md
@@ -0,0 +1 @@
+20
@@ -5,7 +5,7 @@
 <a href="/docs/worker-setup-guides.md"><img src="https://img.shields.io/badge/Setup_Guides-grey?style=flat&logo=gitbook&logoColor=white" alt="Setup Guides"></a>
 <a href="/workflows"><img src="https://img.shields.io/badge/Workflows-grey?style=flat&logo=json&logoColor=white" alt="Workflows"></a>
 <a href="https://buymeacoffee.com/robertvoy"><img src="https://img.shields.io/badge/Donation-grey?style=flat&logo=buymeacoffee&logoColor=white" alt="Donation"></a>
-<a href="https://x.com/robertvoy3D"><img src="https://img.shields.io/twitter/follow/robertvoy3D" alt="Twitter"></a>
+<a href="https://x.com/rbw_ai"><img src="https://img.shields.io/twitter/follow/rbw_ai" alt="Twitter"></a>
 <br><br>
 </div>

@@ -25,7 +25,12 @@
 #### Distributed Upscaling
 - Accelerate Ultimate SD Upscale by distributing tiles across GPUs
 - Intelligent distribution
- Handles single images and batches
+- 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
@@ -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)

@@ -51,7 +57,6 @@ ComfyUI Distributed supports three types of workers:
 ## Requirements

 - ComfyUI
-> Note: Desktop app not currently supported
 - Multiple NVIDIA GPUs
 > No additional GPUs? Use [Cloud Workers](https://github.com/robertvoy/ComfyUI-Distributed/blob/main/docs/worker-setup-guides.md#cloud-workers)
 - That's it
@@ -92,8 +97,9 @@ Generate multiple images in the time it takes to generate one. Each worker uses
 1. Open your ComfyUI workflow
 2. Add **Distributed Seed** → connect to sampler's seed
 3. Add **Distributed Collector** → after VAE Decode
-4. Enable workers in the UI
-5. Run the workflow!
+4. Optional: enable `load_balance` on Distributed Collector to run on one least-busy participant
+5. Enable workers in the UI
+6. Run the workflow!

 ### Parallel WAN Generation
 Generate multiple videos in the time it takes to generate one. Each worker uses a different seed.
@@ -116,7 +122,7 @@ Accelerate Ultimate SD Upscaler by distributing tiles across multiple workers, w

 ![Clipboard Image (3)](https://github.com/user-attachments/assets/ffb57a0d-7b75-4497-96d2-875d60865a1a)

-> [Download workflow](/workflows/distributed-txt2img.json)
+> [Download workflow](/workflows/distributed-upscale.json)

 1. Load your image
 2. Upscale with ESRGAN or similar
@@ -142,6 +148,47 @@ Accelerate Ultimate SD Upscaler by distributing video tiles across multiple work

 ---

+## Developer API
+
+Control your distributed cluster programmatically without opening the browser.
+
+* **Endpoint:** `POST /distributed/queue`
+* **Functionality:** Accepts a ComfyUI API-format prompt, dispatches it to the requested reachable workers, and returns the master `prompt_id`.
+* **Documentation:** [See API Examples & Scripts](https://github.com/robertvoy/ComfyUI-Distributed/blob/main/docs/comfyui-distributed-api.md)
+
+> **⚠️ Security Warning:** Do not expose your ComfyUI port to the public internet. If you need remote access, run ComfyUI behind a secure proxy (like Cloudflare or a VPN).
+
+---
+
+## Distributed Value
+
+Use **Distributed Value** when you want per-worker overrides (for example, different prompts/models/settings per worker).
+
+- Output type adapts to the connected input where possible (`STRING`, `INT`, `FLOAT`, `COMBO`).
+- The node shows only currently enabled workers.
+- If worker enablement changes, worker fields update automatically.
+- When disconnected, it resets to default string mode and clears per-worker overrides.
+- On execution, master uses `default_value`; workers use their mapped override with typed coercion fallback to default.
+
+---
+
+## Nodes
+
+| Node | Description |
+|------|-------------|
+| **Distributed Seed** | Generates unique seeds for each worker |
+| **Distributed Collector** | Collects results (image/video frames and optionally audio) from workers back to the master; `load_balance` can route the run to one least-busy participant |
+| **Distributed Value** | Outputs per-worker override values with fallback to default |
+| **Ultimate SD Upscale Distributed** | Distributes upscale tiles across workers |
+| **Image Batch Divider** | Splits image batches for multi-GPU output |
+| **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 |
+
+---
+
 ## FAQ

 <details>
@@ -156,12 +203,12 @@ No, it does not speed up the generation of a single image or video. Instead, it

 <details>
 <summary>Does it work with the ComfyUI desktop app?</summary>
-Currently, it is not compatible with the ComfyUI desktop app.
+Yes, it does now.
 </details>

 <details>
 <summary>Can I combine my RTX 5090 with a GTX 980 to get faster results?</summary>
-Yes, you can combine different GPUs, but performance is optimized when using similar GPUs. A significant performance imbalance between GPUs may cause bottlenecks. For upscaling, setting `static_distribution` to `false` allows the faster GPU to handle more processing, which can mitigate some bottlenecks. Note that this setting only applies to upscaling tasks.
+Yes, you can combine different GPUs, but performance is optimized when using similar GPUs. A significant performance imbalance between GPUs may cause bottlenecks.
 </details>

 <details>
@@ -169,6 +216,11 @@ Yes, you can combine different GPUs, but performance is optimized when using sim
 Yes, it is compatible with cloud providers. Refer to the setup guides for detailed instructions.
 </details>

+<details>
+<summary>Can I use my main machine just to coordinate workers without rendering?</summary>
+Yes. Open the Distributed panel and uncheck the master toggle to run in orchestrator-only mode. The master will distribute work to workers but won't render locally. If all workers become unavailable, the master automatically re-enables to ensure your workflow still runs.
+</details>
+
 <details>
 <summary>Can I make this work with my Docker setup?</summary>
 Yes, it is compatible with Docker setups, but you will need to configure your Docker environment yourself. Unfortunately, assistance with Docker configuration is not provided.
@@ -176,6 +228,26 @@ Yes, it is compatible with Docker setups, but you will need to configure your Do

 ---

+## Community projects
+
+Community-maintained add-ons and forks that extend ComfyUI-Distributed.
+
+### [ComfyUI-H3MC-Distributed](https://github.com/ymiyad/ComfyUI-H3MC-Distributed) by ymiyad
+
+Connects H3 Motion Context with ComfyUI-Distributed to generate multiple candidate takes of the next video clip across your workers. Compare the results, pick your preferred take, and use it as context for the next clip.
+
+Each saved candidate video includes its continuation latent, so you don't need a shared folder or separate latent files. This add-on speeds up iteration through parallel candidates; it does not split a single clip across GPUs.
+
+### [ComfyUI-Deadline-Distributed](https://github.com/doubletwisted/ComfyUI-Deadline-Distributed) by doubletwisted
+
+A fork of ComfyUI-Distributed with Thinkbox/AWS Deadline 10 integration for render farms. Claim Deadline workers from the Distributed panel, register them automatically, and collect their results back into your master workflow.
+
+Requires Deadline 10 and the author's ComfyUI-Deadline-Plugin.
+
+See each repository for installation instructions, requirements, and support.
+
+---
+
 ## Disclaimer

 This software is provided "as is" without any warranties, express or implied, including merchantability, fitness for a particular purpose, or non-infringement. The developers and copyright holders are not liable for any claims, damages, or liabilities arising from the use, modification, or distribution of the software. Users are solely responsible for ensuring compliance with applicable laws and regulations and for securing their networks against unauthorized access, hacking, data breaches, or loss. The developers assume no liability for any damages or incidents resulting from misuse, improper configuration, or external threats.
@@ -195,7 +267,3 @@ Buy me a coffee at: https://buymeacoffee.com/robertvoy



-
-
-
-
@@ -1,65 +1,22 @@
-import os
-import sys
+"""ComfyUI-Distributed's native V3 extension entrypoint."""
+from comfy_api.v0_0_2 import ComfyExtension, io

-# Add the directory to Python path
-sys.path.append(os.path.dirname(__file__))
+from .nodes.v3 import NODES
+from .runtime.bootstrap import initialize

-# Patch ComfyUI execution validation for DistributedVideoCollector
-try:
-    import execution
-    from .distributed import ImageBatchDivider
-    
-    # Store original validate function if it exists
-    if hasattr(execution, 'validate_outputs'):
-        original_validate_outputs = execution.validate_outputs
-        
-        def patched_validate_outputs(executor, node_id, result, node_class):
-            if node_class == ImageBatchDivider:
-                return  # Skip validation for our dynamic output node
-            return original_validate_outputs(executor, node_id, result, node_class)
-        
-        execution.validate_outputs = patched_validate_outputs
-    else:
-        # Fallback: patch the main execute method if validate_outputs doesn't exist
-        if hasattr(execution.PromptExecutor, 'execute'):
-            original_execute = execution.PromptExecutor.execute
-            
-            def patched_execute(self, prompt, prompt_id, extra_data={}, execute_outputs=[]):
-                # This is a more complex patch - for now just call original
-                # The ByPassTypeTuple should handle most validation issues
-                return original_execute(self, prompt, prompt_id, extra_data, execute_outputs)
-            
-            execution.PromptExecutor.execute = patched_execute
-            
-except ImportError:
-    pass  # ComfyUI execution module not available during import
+WEB_DIRECTORY = './web'

-# Import everything needed from the main module
-from .distributed import (
-    NODE_CLASS_MAPPINGS as DISTRIBUTED_CLASS_MAPPINGS, 
-    NODE_DISPLAY_NAME_MAPPINGS as DISTRIBUTED_DISPLAY_NAME_MAPPINGS
-)

-# Import utilities
-from .utils.config import ensure_config_exists, CONFIG_FILE
-from .utils.logging import debug_log
+class DistributedExtension(ComfyExtension):
+    async def on_load(self) -> None:
+        initialize()

-# Import distributed upscale nodes
-from .distributed_upscale import (
-    NODE_CLASS_MAPPINGS as UPSCALE_CLASS_MAPPINGS,
-    NODE_DISPLAY_NAME_MAPPINGS as UPSCALE_DISPLAY_NAME_MAPPINGS
-)
+    async def get_node_list(self) -> list[type[io.ComfyNode]]:
+        return list(NODES)

-WEB_DIRECTORY = "./web"

-ensure_config_exists()
+async def comfy_entrypoint() -> DistributedExtension:
+    return DistributedExtension()

-# Merge node mappings
-NODE_CLASS_MAPPINGS = {**DISTRIBUTED_CLASS_MAPPINGS, **UPSCALE_CLASS_MAPPINGS}
-NODE_DISPLAY_NAME_MAPPINGS = {**DISTRIBUTED_DISPLAY_NAME_MAPPINGS, **UPSCALE_DISPLAY_NAME_MAPPINGS}

-__all__ = ['NODE_CLASS_MAPPINGS', 'NODE_DISPLAY_NAME_MAPPINGS']
-
-debug_log("Loaded Distributed nodes.")
-debug_log(f"Config file: {CONFIG_FILE}")
-debug_log(f"Available nodes: {list(NODE_CLASS_MAPPINGS.keys())}")
+__all__ = ['comfy_entrypoint', 'WEB_DIRECTORY']
@@ -0,0 +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
@@ -0,0 +1,590 @@
+import json
+from contextlib import asynccontextmanager
+from typing import Any
+
+from aiohttp import web
+import server
+
+try:
+    from ..utils.config import config_transaction, load_config, save_config
+except ImportError:
+    from ..utils.config import load_config
+
+    try:
+        from ..utils.config import save_config
+    except ImportError:
+        def save_config(_config):
+            return True
+
+    @asynccontextmanager
+    async def config_transaction():
+        config = load_config()
+        original_snapshot = json.dumps(config, sort_keys=True)
+        yield config
+        if json.dumps(config, sort_keys=True) != original_snapshot:
+            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):
+    return value > 0
+
+
+CONFIG_SCHEMA = {
+    "workers": (list, None),
+    "master": (dict, None),
+    "settings": (dict, None),
+    "tunnel": (dict, None),
+    "managed_processes": (dict, None),
+    "worker_timeout_seconds": (int, _positive_int),
+    "debug": (bool, None),
+    "auto_launch_workers": (bool, None),
+    "stop_workers_on_master_exit": (bool, None),
+    "master_delegate_only": (bool, None),
+    "websocket_orchestration": (bool, None),
+    "has_auto_populated_workers": (bool, None),
+}
+
+_SETTINGS_FIELDS = {
+    "worker_timeout_seconds",
+    "debug",
+    "auto_launch_workers",
+    "stop_workers_on_master_exit",
+    "master_delegate_only",
+    "websocket_orchestration",
+    "has_auto_populated_workers",
+}
+
+_WORKER_FIELDS = [
+    ("enabled", None, False),
+    ("name", None, False),
+    ("port", None, False),
+    ("host", normalize_host, True),
+    ("cuda_device", None, True),
+    ("extra_args", None, True),
+    ("type", None, False),
+]
+
+_MASTER_FIELDS = [
+    ("name", None, False),
+    ("host", normalize_host, True),
+    ("port", None, False),
+    ("cuda_device", None, False),
+    ("extra_args", None, False),
+]
+
+
+def _apply_field_patch(target: dict, data: dict, field_rules: list) -> None:
+    """Apply a partial update to a target dict based on field rules."""
+    for key, normalizer, remove_on_none in field_rules:
+        if key not in data:
+            continue
+        value = data[key]
+        if value is None and remove_on_none:
+            target.pop(key, None)
+        else:
+            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):
+    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 and credential redaction."""
+    try:
+        data = await request.json()
+    except Exception:
+        return _error_response("Invalid JSON payload", 400)
+
+    if not isinstance(data, dict):
+        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 = {}
+    errors = []
+
+    for key, value in data.items():
+        if key not in CONFIG_SCHEMA:
+            errors.append(f"Unknown field: {key}")
+            continue
+
+        expected_type, validator = CONFIG_SCHEMA[key]
+        if not isinstance(value, expected_type):
+            errors.append(f"{key}: expected {expected_type.__name__}")
+            continue
+
+        if validator and not validator(value):
+            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:
+            validated_root[key] = value
+
+    if errors:
+        return web.json_response({
+            "status": "error",
+            "error": errors,
+            "message": "; ".join(errors),
+        }, status=400)
+
+    store = None
+    response_config = None
+    try:
+        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}")
+async def queue_status_endpoint(request):
+    """Check if a job queue is initialized."""
+    try:
+        job_id = request.match_info['job_id']
+        
+        # Import to ensure initialization
+        from ..upscale.job_store import ensure_tile_jobs_initialized
+        prompt_server = ensure_tile_jobs_initialized()
+        
+        async with prompt_server.distributed_tile_jobs_lock:
+            exists = job_id in prompt_server.distributed_pending_tile_jobs
+        
+        debug_log(f"Queue status check for job {job_id}: {'exists' if exists else 'not found'}")
+        return web.json_response({"exists": exists, "job_id": job_id})
+    except Exception as e:
+        return await handle_api_error(request, e, 500)
+
+@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:
+        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 _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:
+            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")
+
+            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)
+
+                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:
+                    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)
+
+
+@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:
+        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 _error_response("Missing worker_id", 400)
+
+        result = None
+        async with config_transaction() as config:
+            workers = config.get("workers", [])
+            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": "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):
+    """Updates a specific key in the settings object."""
+    try:
+        data = await request.json()
+        key = data.get("key")
+        value = data.get("value")
+
+        if not key or value is None:
+            return await handle_api_error(request, "Missing 'key' or 'value' in request", 400)
+        if key not in _SETTINGS_FIELDS:
+            return await handle_api_error(request, f"Unknown setting: {key}", 400)
+
+        async with config_transaction() as config:
+            if 'settings' not in config:
+                config['settings'] = {}
+
+            config['settings'][key] = value
+
+            return web.json_response({"status": "success", "message": f"Setting '{key}' updated."})
+    except Exception as e:
+        return await handle_api_error(request, e, 400)
+
+@server.PromptServer.instance.routes.post("/distributed/config/update_master")
+async def update_master_endpoint(request):
+    """Updates master configuration."""
+    try:
+        data = await request.json()
+        
+        async with config_transaction() as config:
+            if 'master' not in config:
+                config['master'] = {}
+            _apply_field_patch(config['master'], data, _MASTER_FIELDS)
+
+            return web.json_response({"status": "success", "message": "Master configuration updated."})
+    except Exception as e:
+        return await handle_api_error(request, e, 400)
@@ -0,0 +1,90 @@
+"""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,
+        "protocol_version": PROTOCOL_VERSION,
+        "node_ids": ["DistributedImageListStart", "DistributedImageListEnd"],
+        "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")
+        store_worker_item(data.get("token"), item)
+        return web.json_response({"status": "stored", "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)
+
+
+@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)
@@ -0,0 +1,479 @@
+import json
+import asyncio
+import io
+import os
+import base64
+import binascii
+import time
+
+from aiohttp import web
+import server
+import torch
+from PIL import Image
+
+from ..utils.logging import debug_log
+from ..utils.image import pil_to_tensor, ensure_contiguous
+from ..utils.network import handle_api_error
+from ..utils.constants import JOB_INIT_GRACE_PERIOD, MEMORY_CLEAR_DELAY
+from ..utils import transfer_codec
+try:
+    from .queue_orchestration import ensure_distributed_state, orchestrate_distributed_execution
+except ImportError:
+    from .queue_orchestration import orchestrate_distributed_execution
+
+    def ensure_distributed_state():
+        return None
+from .queue_request import parse_queue_request_payload
+
+prompt_server = server.PromptServer.instance
+
+# Worker result envelopes accepted by POST /distributed/job_complete:
+#
+# application/json (legacy, one frame per request):
+#   { "job_id": str, "worker_id": str, "batch_idx": int, "image": <base64 PNG>, "is_last": bool,
+#     "audio"?: {...} }
+#
+# multipart/form-data (compressed, many frames per request):
+#   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
+
+
+def _decode_image_sync(image_path):
+    """Decode image/video file and compute hash in a threadpool worker."""
+    import base64
+    import hashlib
+    import folder_paths
+
+    full_path = folder_paths.get_annotated_filepath(image_path)
+    if not os.path.exists(full_path):
+        raise FileNotFoundError(image_path)
+
+    hash_md5 = hashlib.md5()
+    with open(full_path, 'rb') as f:
+        for chunk in iter(lambda: f.read(4096), b""):
+            hash_md5.update(chunk)
+    file_hash = hash_md5.hexdigest()
+
+    video_extensions = {'.mp4', '.avi', '.mov', '.mkv', '.webm'}
+    file_ext = os.path.splitext(full_path)[1].lower()
+
+    if file_ext in video_extensions:
+        with open(full_path, 'rb') as f:
+            file_data = f.read()
+        mime_types = {
+            '.mp4': 'video/mp4',
+            '.avi': 'video/x-msvideo',
+            '.mov': 'video/quicktime',
+            '.mkv': 'video/x-matroska',
+            '.webm': 'video/webm'
+        }
+        mime_type = mime_types.get(file_ext, 'video/mp4')
+        image_data = f"data:{mime_type};base64,{base64.b64encode(file_data).decode('utf-8')}"
+    else:
+        with Image.open(full_path) as img:
+            if img.mode not in ('RGB', 'RGBA'):
+                img = img.convert('RGB')
+            buffer = io.BytesIO()
+            img.save(buffer, format='PNG', compress_level=1)
+            image_data = f"data:image/png;base64,{base64.b64encode(buffer.getvalue()).decode('utf-8')}"
+
+    return {
+        "status": "success",
+        "image_data": image_data,
+        "hash": file_hash,
+    }
+
+
+def _check_file_sync(filename, expected_hash):
+    """Check file presence and hash in a threadpool worker."""
+    import hashlib
+    import folder_paths
+
+    full_path = folder_paths.get_annotated_filepath(filename)
+    if not os.path.exists(full_path):
+        return {
+            "status": "success",
+            "exists": False,
+        }
+
+    hash_md5 = hashlib.md5()
+    with open(full_path, 'rb') as f:
+        for chunk in iter(lambda: f.read(4096), b""):
+            hash_md5.update(chunk)
+    file_hash = hash_md5.hexdigest()
+
+    return {
+        "status": "success",
+        "exists": True,
+        "hash_matches": file_hash == expected_hash,
+    }
+
+
+def _decode_canonical_png_tensor(image_payload):
+    """Decode canonical base64 PNG payload into a contiguous IMAGE tensor."""
+    if not isinstance(image_payload, str) or not image_payload.strip():
+        raise ValueError("Field 'image' must be a non-empty base64 PNG string.")
+
+    encoded = image_payload.strip()
+    if encoded.startswith("data:"):
+        header, sep, data_part = encoded.partition(",")
+        if not sep:
+            raise ValueError("Field 'image' data URL is malformed.")
+        if not header.lower().startswith("data:image/png;base64"):
+            raise ValueError("Field 'image' must be a PNG data URL when using data:* format.")
+        encoded = data_part
+
+    try:
+        png_bytes = base64.b64decode(encoded, validate=True)
+    except (binascii.Error, ValueError) as exc:
+        raise ValueError("Field 'image' is not valid base64 PNG data.") from exc
+
+    if not png_bytes:
+        raise ValueError("Field 'image' decoded to empty PNG data.")
+
+    try:
+        with Image.open(io.BytesIO(png_bytes)) as img:
+            img = img.convert("RGB")
+            tensor = pil_to_tensor(img)
+        return ensure_contiguous(tensor)
+    except Exception as exc:
+        raise ValueError(f"Failed to decode PNG image payload: {exc}") from exc
+
+
+def _decode_audio_payload(audio_payload):
+    """Decode canonical audio payload into an AUDIO dict."""
+    from ..utils.audio_payload import decode_audio_payload
+
+    return decode_audio_payload(audio_payload)
+
+
+@server.PromptServer.instance.routes.post("/distributed/prepare_job")
+async def prepare_job_endpoint(request):
+    try:
+        data = await request.json()
+        multi_job_id = data.get('multi_job_id')
+        if not multi_job_id:
+            return await handle_api_error(request, "Missing multi_job_id", 400)
+
+        ensure_distributed_state()
+        async with prompt_server.distributed_jobs_lock:
+            if multi_job_id not in prompt_server.distributed_pending_jobs:
+                prompt_server.distributed_pending_jobs[multi_job_id] = asyncio.Queue()
+        
+        debug_log(f"Prepared queue for job {multi_job_id}")
+        return web.json_response({"status": "success"})
+    except Exception as e:
+        return await handle_api_error(request, e)
+
+@server.PromptServer.instance.routes.post("/distributed/clear_memory")
+async def clear_memory_endpoint(request):
+    debug_log("Received request to clear VRAM.")
+    try:
+        # Use ComfyUI's prompt server queue system like the /free endpoint does
+        if hasattr(server.PromptServer.instance, 'prompt_queue'):
+            server.PromptServer.instance.prompt_queue.set_flag("unload_models", True)
+            server.PromptServer.instance.prompt_queue.set_flag("free_memory", True)
+            debug_log("Set queue flags for memory clearing.")
+        
+        # Wait a bit for the queue to process
+        await asyncio.sleep(MEMORY_CLEAR_DELAY)
+        
+        # Also do direct cleanup as backup, but with error handling
+        import gc
+        import comfy.model_management as mm
+        
+        try:
+            mm.unload_all_models()
+        except AttributeError as e:
+            debug_log(f"Warning during model unload: {e}")
+        
+        try:
+            mm.soft_empty_cache()
+        except Exception as e:
+            debug_log(f"Warning during cache clear: {e}")
+        
+        for _ in range(3):
+            gc.collect()
+        
+        if torch.cuda.is_available():
+            torch.cuda.empty_cache()
+            torch.cuda.ipc_collect()
+        
+        debug_log("VRAM cleared successfully.")
+        return web.json_response({"status": "success", "message": "GPU memory cleared."})
+    except Exception as e:
+        # Even if there's an error, try to do basic cleanup
+        import gc
+        gc.collect()
+        if torch.cuda.is_available():
+            torch.cuda.empty_cache()
+        debug_log(f"Partial VRAM clear completed with warning: {e}")
+        return web.json_response({"status": "success", "message": "GPU memory cleared (with warnings)"})
+
+
+@server.PromptServer.instance.routes.post("/distributed/queue")
+async def distributed_queue_endpoint(request):
+    """Queue a distributed workflow, mirroring the UI orchestration pipeline."""
+    try:
+        raw_payload = await request.json()
+    except Exception as exc:
+        return await handle_api_error(request, f"Invalid JSON payload: {exc}", 400)
+
+    try:
+        payload = parse_queue_request_payload(raw_payload)
+    except ValueError as exc:
+        return await handle_api_error(request, exc, 400)
+
+    try:
+        prompt_id, prompt_number, worker_count, node_errors = await orchestrate_distributed_execution(
+            payload.prompt,
+            payload.workflow_meta,
+            payload.client_id,
+            enabled_worker_ids=payload.enabled_worker_ids,
+            delegate_master=payload.delegate_master,
+            trace_execution_id=payload.trace_execution_id,
+        )
+        return web.json_response({
+            "prompt_id": prompt_id,
+            "number": prompt_number,
+            "node_errors": node_errors,
+            "worker_count": worker_count,
+            "auto_prepare_supported": True,
+        })
+    except Exception as exc:
+        return await handle_api_error(request, exc, 500)
+
+@server.PromptServer.instance.routes.post("/distributed/load_image")
+async def load_image_endpoint(request):
+    """Load an image or video file and return it as base64 data with hash."""
+    try:
+        data = await request.json()
+        image_path = data.get("image_path")
+        
+        if not image_path:
+            return await handle_api_error(request, "Missing image_path", 400)
+        loop = asyncio.get_running_loop()
+        payload = await loop.run_in_executor(None, _decode_image_sync, image_path)
+        return web.json_response(payload)
+    except FileNotFoundError:
+        return await handle_api_error(request, f"File not found: {image_path}", 404)
+    except Exception as e:
+        return await handle_api_error(request, e, 500)
+
+@server.PromptServer.instance.routes.post("/distributed/check_file")
+async def check_file_endpoint(request):
+    """Check if a file exists and matches the given hash."""
+    try:
+        data = await request.json()
+        filename = data.get("filename")
+        expected_hash = data.get("hash")
+        
+        if not filename or not expected_hash:
+            return await handle_api_error(request, "Missing filename or hash", 400)
+        loop = asyncio.get_running_loop()
+        payload = await loop.run_in_executor(None, _check_file_sync, filename, expected_hash)
+        return web.json_response(payload)
+        
+    except Exception as e:
+        return await handle_api_error(request, e, 500)
+
+
+async def _enqueue_worker_items(multi_job_id, items):
+    """Put decoded items on the job queue, waiting briefly for the master to create it.
+
+    Returns the queue size after enqueueing, or None if the job never appeared.
+    """
+    deadline = time.monotonic() + float(JOB_INIT_GRACE_PERIOD)
+    while True:
+        async with prompt_server.distributed_jobs_lock:
+            pending = prompt_server.distributed_pending_jobs.get(multi_job_id)
+            if pending is not None:
+                for item in items:
+                    await pending.put(item)
+                return pending.qsize()
+        if time.monotonic() > deadline:
+            return None
+        await asyncio.sleep(0.05)
+
+
+def _decode_compressed_chunk_sync(envelope, parts):
+    """Decode a compressed chunk into per-frame IMAGE tensors (threadpool)."""
+    frames = transfer_codec.decode_chunk(envelope, parts)
+    return [
+        ensure_contiguous(torch.from_numpy(transfer_codec.uint8_frame_to_tensor_array(frame)))
+        for frame in frames
+    ]
+
+
+def _validate_compressed_envelope(envelope):
+    errors = []
+    if not isinstance(envelope, dict):
+        return ["envelope: expected JSON object"]
+    for key in ("job_id", "worker_id"):
+        value = envelope.get(key)
+        if not isinstance(value, str) or not value.strip():
+            errors.append(f"{key}: expected non-empty string")
+    start_index = envelope.get("start_index")
+    if not isinstance(start_index, int) or isinstance(start_index, bool) or start_index < 0:
+        errors.append("start_index: expected non-negative integer")
+    if not isinstance(envelope.get("is_last"), bool):
+        errors.append("is_last: expected boolean")
+    try:
+        if transfer_codec.normalize_format(envelope.get("format")) == transfer_codec.LEGACY_FORMAT:
+            errors.append("format: legacy_png must be sent as JSON")
+    except transfer_codec.TransferCodecError as exc:
+        errors.append(f"format: {exc}")
+    audio_payload = envelope.get("audio")
+    if audio_payload is not None and not isinstance(audio_payload, dict):
+        errors.append("audio: expected object when provided")
+    return errors
+
+
+async def _job_complete_compressed(request):
+    try:
+        form = await request.post()
+    except Exception as exc:
+        too_large = getattr(web, "HTTPRequestEntityTooLarge", None)
+        if too_large is not None and isinstance(exc, too_large):
+            return await handle_api_error(
+                request, f"Request body exceeds the master's upload limit (--max-upload-size): {exc}", 413)
+        return await handle_api_error(request, f"Invalid multipart payload: {exc}", 400)
+
+    raw_envelope = form.get("envelope")
+    if isinstance(raw_envelope, (bytes, bytearray)):
+        raw_envelope = raw_envelope.decode("utf-8")
+    elif hasattr(raw_envelope, "file"):
+        raw_envelope = raw_envelope.file.read().decode("utf-8")
+    try:
+        envelope = json.loads(raw_envelope) if isinstance(raw_envelope, str) else None
+    except ValueError as exc:
+        return await handle_api_error(request, f"envelope: invalid JSON ({exc})", 400)
+    if envelope is None:
+        return await handle_api_error(request, "envelope: missing field", 400)
+
+    errors = _validate_compressed_envelope(envelope)
+    if errors:
+        return await handle_api_error(request, errors, 400)
+
+    parts = []
+    for field in form.getall("media", []):
+        if hasattr(field, "file"):
+            parts.append(field.file.read())
+        elif isinstance(field, (bytes, bytearray)):
+            parts.append(bytes(field))
+        else:
+            return await handle_api_error(request, "media: expected binary file parts", 400)
+
+    loop = asyncio.get_running_loop()
+    try:
+        tensors = await loop.run_in_executor(None, _decode_compressed_chunk_sync, envelope, parts)
+        decoded_audio = _decode_audio_payload(envelope.get("audio")) if envelope.get("audio") is not None else None
+    except ValueError as exc:
+        return await handle_api_error(request, exc, 400)
+
+    multi_job_id = envelope["job_id"].strip()
+    worker_id = envelope["worker_id"].strip()
+    start_index = envelope["start_index"]
+    is_last = envelope["is_last"]
+    last_pos = len(tensors) - 1
+    items = [
+        {
+            "tensor": tensor,
+            "worker_id": worker_id,
+            "image_index": start_index + pos,
+            "is_last": bool(is_last and pos == last_pos),
+            "audio": decoded_audio if pos == last_pos else None,
+        }
+        for pos, tensor in enumerate(tensors)
+    ]
+
+    queue_size = await _enqueue_worker_items(multi_job_id, items)
+    if queue_size is None:
+        return await handle_api_error(request, "job not initialized", 404)
+
+    debug_log(
+        f"job_complete received {envelope['format']} chunk - job_id: {multi_job_id}, worker: {worker_id}, "
+        f"frames: {start_index}..{start_index + len(items) - 1}, "
+        f"bytes: {sum(len(p) for p in parts)}, is_last: {is_last}, queue_size: {queue_size}"
+    )
+    return web.json_response({"status": "success", "frames": len(items)})
+
+
+@server.PromptServer.instance.routes.post("/distributed/job_complete")
+async def job_complete_endpoint(request):
+    content_type = str(getattr(request, "content_type", "") or "").lower()
+    if content_type.startswith("multipart/"):
+        try:
+            return await _job_complete_compressed(request)
+        except Exception as e:
+            return await handle_api_error(request, e)
+
+    try:
+        data = await request.json()
+    except Exception as exc:
+        return await handle_api_error(request, f"Invalid JSON payload: {exc}", 400)
+
+    if not isinstance(data, dict):
+        return await handle_api_error(request, "Expected a JSON object body", 400)
+
+    try:
+        job_id = data.get("job_id")
+        worker_id = data.get("worker_id")
+        batch_idx = data.get("batch_idx")
+        image_payload = data.get("image")
+        audio_payload = data.get("audio")
+        is_last = data.get("is_last")
+
+        errors = []
+        if not isinstance(job_id, str) or not job_id.strip():
+            errors.append("job_id: expected non-empty string")
+        if not isinstance(worker_id, str) or not worker_id.strip():
+            errors.append("worker_id: expected non-empty string")
+        if not isinstance(batch_idx, int) or batch_idx < 0:
+            errors.append("batch_idx: expected non-negative integer")
+        image_provided = image_payload is not None
+        audio_provided = audio_payload is not None
+        if not image_provided and not audio_provided:
+            errors.append("expected at least one of image or audio")
+        if image_provided and (not isinstance(image_payload, str) or not image_payload.strip()):
+            errors.append("image: expected non-empty base64 PNG string")
+        if audio_provided and not isinstance(audio_payload, dict):
+            errors.append("audio: expected object when provided")
+        if not isinstance(is_last, bool):
+            errors.append("is_last: expected boolean")
+        if errors:
+            return await handle_api_error(request, errors, 400)
+
+        try:
+            tensor = _decode_canonical_png_tensor(image_payload) if image_provided else None
+            decoded_audio = _decode_audio_payload(audio_payload) if audio_provided else None
+        except ValueError as exc:
+            return await handle_api_error(request, exc, 400)
+        multi_job_id = job_id.strip()
+        worker_id = worker_id.strip()
+
+        queue_size = await _enqueue_worker_items(
+            multi_job_id,
+            [
+                {
+                    "tensor": tensor,
+                    "worker_id": worker_id,
+                    "image_index": int(batch_idx),
+                    "is_last": is_last,
+                    "audio": decoded_audio,
+                }
+            ],
+        )
+        if queue_size is None:
+            return await handle_api_error(request, "job not initialized", 404)
+
+        debug_log(
+            f"job_complete received canonical envelope - job_id: {multi_job_id}, "
+            f"worker: {worker_id}, batch_idx: {batch_idx}, is_last: {is_last}, "
+            f"queue_size: {queue_size}"
+        )
+
+        return web.json_response({"status": "success"})
+    except Exception as e:
+        return await handle_api_error(request, e)
@@ -0,0 +1,4 @@
+# Orchestration helpers split out from queue_orchestration.py:
+# - prompt_transform.py: graph pruning + hidden input overrides
+# - media_sync.py: remote media/path normalization
+# - dispatch.py: worker probe + prompt dispatch
@@ -0,0 +1,405 @@
+import asyncio
+import json
+import uuid
+
+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:
+    def trace_debug(*_args, **_kwargs):
+        return None
+
+    def trace_info(*_args, **_kwargs):
+        return None
+
+try:
+    from ..schemas import parse_positive_int
+except ImportError:
+    def parse_positive_int(value, default):
+        try:
+            parsed = int(value)
+            return parsed if parsed > 0 else default
+        except (TypeError, ValueError):
+            return default
+
+_least_busy_rr_index = 0
+
+
+async def worker_is_active(worker):
+    """Ping worker's /prompt endpoint to confirm it's reachable."""
+    url = build_worker_url(worker)
+    return await probe_worker(url, timeout=3.0) is not None
+
+
+async def worker_ws_is_active(worker):
+    """Ping worker's websocket endpoint to confirm it's reachable."""
+    session = await get_client_session()
+    url = build_worker_url(worker, "/distributed/worker_ws")
+    try:
+        ws = await session.ws_connect(url, heartbeat=20, timeout=3)
+        await ws.close()
+        return True
+    except asyncio.TimeoutError:
+        debug_log(f"[Distributed] Worker WS probe timed out: {url}")
+        return False
+    except aiohttp.ClientConnectorError:
+        debug_log(f"[Distributed] Worker WS unreachable: {url}")
+        return False
+    except Exception as e:
+        debug_log(f"[Distributed] Worker WS probe unexpected error: {e}")
+        return False
+
+
+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
+
+
+async def _dispatch_via_websocket(worker_url, payload, client_id, timeout=60.0):
+    """Open a fresh worker websocket, dispatch one prompt, wait for ack, then close."""
+    request_id = uuid.uuid4().hex
+    ws_payload = {
+        "type": "dispatch_prompt",
+        "request_id": request_id,
+        "prompt": payload.get("prompt"),
+        "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
+
+    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(
+    worker,
+    prompt_obj,
+    workflow_meta,
+    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
+    if extra_data:
+        payload["extra_data"] = extra_data
+
+    if use_websocket:
+        try:
+            prompt_id = await _dispatch_via_websocket(
+                worker_url,
+                {
+                    "prompt": prompt_obj,
+                    "workflow": workflow_meta,
+                    **target_payload,
+                },
+                client_id,
+            )
+            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:
+                trace_info(trace_execution_id, f"Websocket dispatch failed for worker {worker_id}: {exc}")
+            else:
+                log(f"[Distributed] Websocket dispatch failed for worker {worker_id}: {exc}")
+            raise
+
+    session = await get_client_session()
+    async with session.post(
+        url,
+        json=payload,
+        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 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"
+                    )
+
+                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(
+    workers,
+    use_websocket,
+    delegate_master,
+    trace_execution_id=None,
+    probe_concurrency=8,
+):
+    """Probe workers and return (active_workers, updated_delegate_master)."""
+    probe_limit = parse_positive_int(probe_concurrency, 8)
+    probe_semaphore = asyncio.Semaphore(probe_limit)
+
+    if trace_execution_id and workers:
+        trace_debug(
+            trace_execution_id,
+            f"Probing {len(workers)} workers with probe_concurrency={probe_limit}",
+        )
+
+    probe_results = await asyncio.gather(
+        *[
+            _probe_worker_active(worker, use_websocket, probe_semaphore)
+            for worker in workers
+        ]
+    )
+
+    active_workers = []
+    for worker, is_active in probe_results:
+        if is_active:
+            active_workers.append(worker)
+        else:
+            if trace_execution_id:
+                trace_info(trace_execution_id, f"Worker {worker['name']} ({worker['id']}) is offline, skipping.")
+            else:
+                log(f"[Distributed] Worker {worker['name']} ({worker['id']}) is offline, skipping.")
+
+    if trace_execution_id and workers:
+        trace_debug(
+            trace_execution_id,
+            f"Worker probe complete: active={len(active_workers)}/{len(workers)}",
+        )
+
+    if not active_workers and delegate_master:
+        if trace_execution_id:
+            trace_debug(trace_execution_id, "All workers offline while delegate-only requested; enabling master participation.")
+        else:
+            debug_log("All workers offline while delegate-only requested; enabling master participation.")
+        delegate_master = False
+
+    return active_workers, delegate_master
+
+
+def _extract_queue_remaining(payload):
+    if not isinstance(payload, dict):
+        return 0
+    try:
+        queue_remaining = int(payload.get("exec_info", {}).get("queue_remaining", 0))
+    except (TypeError, ValueError):
+        queue_remaining = 0
+    return max(queue_remaining, 0)
+
+
+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:
+            return None
+        return {
+            "worker": worker,
+            "queue_remaining": _extract_queue_remaining(payload),
+        }
+
+
+def _select_idle_round_robin(statuses):
+    global _least_busy_rr_index
+    if not statuses:
+        return None
+    index = _least_busy_rr_index % len(statuses)
+    _least_busy_rr_index += 1
+    return statuses[index]
+
+
+async def select_least_busy_worker(
+    workers,
+    trace_execution_id=None,
+    probe_concurrency=8,
+    probe_timeout=3.0,
+):
+    """Select one worker by queue depth, round-robin among idle workers."""
+    if not workers:
+        return None
+
+    probe_limit = parse_positive_int(probe_concurrency, 8)
+    probe_semaphore = asyncio.Semaphore(probe_limit)
+    statuses = await asyncio.gather(
+        *[
+            _probe_worker_queue(worker, probe_semaphore, probe_timeout)
+            for worker in workers
+        ]
+    )
+    statuses = [status for status in statuses if status is not None]
+    if not statuses:
+        if trace_execution_id:
+            trace_info(trace_execution_id, "Least-busy selection failed: no worker queue probes succeeded.")
+        else:
+            log("[Distributed] Least-busy selection failed: no worker queue probes succeeded.")
+        return None
+
+    idle_statuses = [status for status in statuses if status["queue_remaining"] == 0]
+    if idle_statuses:
+        selected = _select_idle_round_robin(idle_statuses)
+    else:
+        selected = min(statuses, key=lambda status: status["queue_remaining"])
+
+    worker = selected["worker"]
+    queue_remaining = selected["queue_remaining"]
+    if trace_execution_id:
+        trace_debug(
+            trace_execution_id,
+            f"Least-busy worker selected: {worker.get('name')} ({worker.get('id')}), queue_remaining={queue_remaining}",
+        )
+    else:
+        debug_log(
+            f"Least-busy worker selected: {worker.get('name')} ({worker.get('id')}), queue_remaining={queue_remaining}"
+        )
+    return worker
@@ -0,0 +1,291 @@
+import asyncio
+import hashlib
+import mimetypes
+import os
+import re
+
+import aiohttp
+
+from ...utils.logging import debug_log, log
+from ...utils.network import build_worker_url, get_client_session
+from ...utils.trace_logger import trace_debug, trace_info
+
+
+LIKELY_FILENAME_RE = re.compile(
+    r"\.(ckpt|safetensors|pt|pth|bin|yaml|json|png|jpg|jpeg|webp|gif|bmp|mp4|avi|mov|mkv|webm|"
+    r"wav|mp3|flac|m4a|aac|ogg|opus|aiff|aif|wma|latent|txt|vae|lora|embedding)"
+    r"(\s*\[\w+\])?$",
+    re.IGNORECASE,
+)
+MEDIA_FILE_RE = re.compile(
+    r"\.(png|jpg|jpeg|webp|gif|bmp|mp4|avi|mov|mkv|webm|wav|mp3|flac|m4a|aac|ogg|opus|aiff|aif|wma)(\s*\[\w+\])?$",
+    re.IGNORECASE,
+)
+
+
+def _normalize_media_reference(value):
+    """Normalize one media string value to a path-like reference or None."""
+    if not isinstance(value, str):
+        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
+
+
+def convert_paths_for_platform(obj, target_separator):
+    """Recursively normalize likely file paths for the worker platform separator."""
+    if target_separator not in ("/", "\\"):
+        return obj
+
+    def _convert(value):
+        if isinstance(value, str):
+            if ("/" in value or "\\" in value) and LIKELY_FILENAME_RE.search(value):
+                trimmed = value.strip()
+                has_drive = bool(re.match(r"^[A-Za-z]:(\\\\|/)", trimmed))
+                is_absolute = trimmed.startswith("/") or trimmed.startswith("\\\\")
+                has_protocol = bool(re.match(r"^\w+://", trimmed))
+
+                # URLs are not local paths and should never be separator-normalized.
+                if has_protocol:
+                    return trimmed
+
+                # Keep relative media-style paths in forward-slash form (Comfy-style annotated paths).
+                if not has_drive and not is_absolute and not has_protocol and MEDIA_FILE_RE.search(trimmed):
+                    return re.sub(r"[\\]+", "/", trimmed)
+
+                if target_separator == "\\":
+                    return re.sub(r"[\\/]+", r"\\", trimmed)
+                return re.sub(r"[\\/]+", "/", trimmed)
+            return value
+        if isinstance(value, list):
+            return [_convert(item) for item in value]
+        if isinstance(value, dict):
+            return {key: _convert(item) for key, item in value.items()}
+        return value
+
+    return _convert(obj)
+
+
+def _find_media_references(prompt_obj):
+    """Find media file references in image/video/audio/file inputs used by worker prompts."""
+    media_refs = set()
+    for node in prompt_obj.values():
+        if not isinstance(node, dict):
+            continue
+        inputs = node.get("inputs", {})
+        for key in ("image", "video", "audio", "file"):
+            cleaned = _normalize_media_reference(inputs.get(key))
+            if cleaned:
+                media_refs.add(cleaned)
+    return sorted(media_refs)
+
+
+def _rewrite_prompt_media_inputs(prompt_obj, worker_media_paths):
+    """Rewrite media string inputs to worker-local uploaded paths."""
+    if not isinstance(worker_media_paths, dict) or not worker_media_paths:
+        return
+
+    for node in prompt_obj.values():
+        if not isinstance(node, dict):
+            continue
+        inputs = node.get("inputs", {})
+        if not isinstance(inputs, dict):
+            continue
+        for key in ("image", "video", "audio", "file"):
+            value = inputs.get(key)
+            cleaned = _normalize_media_reference(value)
+            if not cleaned:
+                continue
+            worker_path = worker_media_paths.get(cleaned)
+            if worker_path:
+                inputs[key] = worker_path
+
+
+def _load_media_file_sync(filename):
+    """Load local media bytes and hash for worker upload sync."""
+    import folder_paths
+
+    full_path = folder_paths.get_annotated_filepath(filename)
+    if not os.path.exists(full_path):
+        raise FileNotFoundError(filename)
+
+    with open(full_path, "rb") as f:
+        file_bytes = f.read()
+
+    file_hash = hashlib.md5(file_bytes).hexdigest()
+    mime_type = mimetypes.guess_type(full_path)[0]
+    if not mime_type:
+        ext = os.path.splitext(full_path)[1].lower()
+        if ext in {".mp4", ".avi", ".mov", ".mkv", ".webm"}:
+            mime_type = "video/mp4"
+        else:
+            mime_type = "image/png"
+    return file_bytes, file_hash, mime_type
+
+
+async def fetch_worker_path_separator(worker, trace_execution_id=None):
+    """Best-effort fetch of a worker's path separator from /distributed/system_info."""
+    url = build_worker_url(worker, "/distributed/system_info")
+    session = await get_client_session()
+    try:
+        async with session.get(url, timeout=aiohttp.ClientTimeout(total=5)) as resp:
+            if resp.status != 200:
+                return None
+            payload = await resp.json()
+            separator = ((payload or {}).get("platform") or {}).get("path_separator")
+            return separator if separator in ("/", "\\") else None
+    except Exception as exc:
+        if trace_execution_id:
+            trace_debug(trace_execution_id, f"Failed to fetch worker system info ({worker.get('id')}): {exc}")
+        else:
+            debug_log(f"[Distributed] Failed to fetch worker system info ({worker.get('id')}): {exc}")
+        return None
+
+
+async def fetch_worker_has_node(worker, node_class, trace_execution_id=None):
+    """Best-effort check that a worker has ``node_class`` registered.
+
+    Uses ComfyUI's ``GET /object_info/{node_class}``, which returns ``{}`` for
+    unknown classes. Any failure counts as "not available" so callers fall
+    back to nodes every ComfyUI install has.
+    """
+    url = build_worker_url(worker, f"/object_info/{node_class}")
+    session = await get_client_session()
+    try:
+        async with session.get(url, timeout=aiohttp.ClientTimeout(total=5)) as resp:
+            if resp.status != 200:
+                return False
+            payload = await resp.json()
+            return isinstance(payload, dict) and node_class in payload
+    except Exception as exc:
+        message = f"Failed to probe node '{node_class}' on worker {worker.get('id')}: {exc}"
+        if trace_execution_id:
+            trace_debug(trace_execution_id, message)
+        else:
+            debug_log(f"[Distributed] {message}")
+        return False
+
+
+async def _upload_media_to_worker(worker, filename, file_bytes, file_hash, mime_type, trace_execution_id=None):
+    """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:
+        async with session.post(
+            check_url,
+            json={"filename": normalized, "hash": file_hash},
+            timeout=aiohttp.ClientTimeout(total=6),
+        ) as resp:
+            if resp.status == 200:
+                payload = await resp.json()
+                if payload.get("exists") and payload.get("hash_matches"):
+                    return False, normalized
+    except Exception as exc:
+        if trace_execution_id:
+            trace_debug(trace_execution_id, f"Media check failed for '{normalized}' on worker {worker.get('id')}: {exc}")
+        else:
+            debug_log(f"[Distributed] Media check failed for '{normalized}' on worker {worker.get('id')}: {exc}")
+
+    parts = normalized.split("/")
+    clean_name = parts[-1]
+    subfolder = "/".join(parts[:-1])
+
+    form = aiohttp.FormData()
+    form.add_field("image", file_bytes, filename=clean_name, content_type=mime_type)
+    form.add_field("type", "input")
+    form.add_field("subfolder", subfolder)
+    form.add_field("overwrite", "true")
+
+    upload_url = build_worker_url(worker, "/upload/image")
+    async with session.post(
+        upload_url,
+        data=form,
+        timeout=aiohttp.ClientTimeout(total=30),
+    ) as resp:
+        resp.raise_for_status()
+        try:
+            payload = await resp.json()
+        except Exception:
+            payload = {}
+
+    name = str((payload or {}).get("name") or clean_name).strip()
+    subfolder = str((payload or {}).get("subfolder") or "").strip().replace("\\", "/").strip("/")
+    worker_path = f"{subfolder}/{name}" if subfolder else name
+    return True, worker_path
+
+
+async def sync_worker_media(worker, prompt_obj, trace_execution_id=None):
+    """Sync referenced media files from master to a remote worker before dispatch."""
+    media_refs = _find_media_references(prompt_obj)
+    if not media_refs:
+        return
+
+    loop = asyncio.get_running_loop()
+    uploaded = 0
+    skipped = 0
+    missing = 0
+    worker_media_paths = {}
+    for filename in media_refs:
+        try:
+            file_bytes, file_hash, mime_type = await loop.run_in_executor(
+                None, _load_media_file_sync, filename
+            )
+        except FileNotFoundError:
+            missing += 1
+            if trace_execution_id:
+                trace_info(trace_execution_id, f"Media file '{filename}' not found on master; worker may fail to load it.")
+            else:
+                log(f"[Distributed] Media file '{filename}' not found on master; worker may fail to load it.")
+            continue
+        except Exception as exc:
+            if trace_execution_id:
+                trace_info(trace_execution_id, f"Failed to load media '{filename}' for worker sync: {exc}")
+            else:
+                log(f"[Distributed] Failed to load media '{filename}' for worker sync: {exc}")
+            continue
+
+        try:
+            changed, worker_path = await _upload_media_to_worker(
+                worker,
+                filename,
+                file_bytes,
+                file_hash,
+                mime_type,
+                trace_execution_id=trace_execution_id,
+            )
+            if worker_path:
+                worker_media_paths[filename] = worker_path
+            if changed:
+                uploaded += 1
+            else:
+                skipped += 1
+        except Exception as exc:
+            if trace_execution_id:
+                trace_info(trace_execution_id, f"Failed to upload media '{filename}' to worker {worker.get('id')}: {exc}")
+            else:
+                log(f"[Distributed] Failed to upload media '{filename}' to worker {worker.get('id')}: {exc}")
+
+    _rewrite_prompt_media_inputs(prompt_obj, worker_media_paths)
+
+    summary = (
+        f"Media sync for worker {worker.get('id')}: "
+        f"uploaded={uploaded}, skipped={skipped}, missing={missing}, referenced={len(media_refs)}"
+    )
+    if trace_execution_id:
+        trace_debug(trace_execution_id, summary)
+    else:
+        debug_log(f"[Distributed] {summary}")
@@ -0,0 +1,687 @@
+import json
+from collections import deque
+
+from ...utils.logging import debug_log
+
+
+class PromptIndex:
+    """Cache prompt metadata for faster worker/master prompt preparation."""
+
+    def __init__(self, prompt_obj):
+        self._prompt_json = json.dumps(prompt_obj)
+        self.nodes_by_class = {}
+        self.class_by_node = {}
+        self.inputs_by_node = {}
+        for node_id, node in _iter_prompt_nodes(prompt_obj):
+            class_type = node.get("class_type")
+            node_id_str = str(node_id)
+            if class_type:
+                self.nodes_by_class.setdefault(class_type, []).append(node_id_str)
+            self.class_by_node[node_id_str] = class_type
+            self.inputs_by_node[node_id_str] = node.get("inputs", {})
+        self._upstream_cache = {}
+
+    def copy_prompt(self):
+        return json.loads(self._prompt_json)
+
+    def nodes_for_class(self, class_name):
+        return self.nodes_by_class.get(class_name, [])
+
+    def has_upstream(self, start_node_id, target_class):
+        cache_key = (str(start_node_id), target_class)
+        if cache_key in self._upstream_cache:
+            return self._upstream_cache[cache_key]
+
+        visited = set()
+        stack = [str(start_node_id)]
+        while stack:
+            node_id = stack.pop()
+            if node_id in visited:
+                continue
+            visited.add(node_id)
+            inputs = self.inputs_by_node.get(node_id, {})
+            for value in inputs.values():
+                if isinstance(value, list) and len(value) == 2:
+                    upstream_id = str(value[0])
+                    if self.class_by_node.get(upstream_id) == target_class:
+                        self._upstream_cache[cache_key] = True
+                        return True
+                    if upstream_id in self.inputs_by_node:
+                        stack.append(upstream_id)
+
+        self._upstream_cache[cache_key] = False
+        return False
+
+
+def _iter_prompt_nodes(prompt_obj):
+    for node_id, node in prompt_obj.items():
+        if isinstance(node, dict):
+            yield str(node_id), node
+
+
+def find_nodes_by_class(prompt_obj, class_name):
+    nodes = []
+    for node_id, node in _iter_prompt_nodes(prompt_obj):
+        if node.get("class_type") == class_name:
+            nodes.append(node_id)
+    return nodes
+
+
+def _find_downstream_nodes(prompt_obj, start_ids):
+    """Return all nodes reachable downstream from the provided IDs."""
+    adjacency = {}
+    for node_id, node in _iter_prompt_nodes(prompt_obj):
+        inputs = node.get("inputs", {})
+        for value in inputs.values():
+            if isinstance(value, list) and len(value) == 2:
+                source_id = str(value[0])
+                adjacency.setdefault(source_id, set()).add(str(node_id))
+
+    connected = set(start_ids)
+    queue = deque(start_ids)
+    while queue:
+        current = queue.popleft()
+        for dependent in adjacency.get(current, ()):  # pragma: no branch - simple iteration
+            if dependent not in connected:
+                connected.add(dependent)
+                queue.append(dependent)
+    return connected
+
+
+def _create_numeric_id_generator(prompt_obj):
+    """Return a closure that yields new numeric string IDs."""
+    max_id = 0
+    for node_id in prompt_obj.keys():
+        try:
+            numeric = int(node_id)
+        except (TypeError, ValueError):
+            continue
+        max_id = max(max_id, numeric)
+
+    counter = max_id
+
+    def _next_id():
+        nonlocal counter
+        counter += 1
+        return str(counter)
+
+    return _next_id
+
+
+def _find_upstream_nodes(prompt_obj, start_ids):
+    """Return all nodes reachable upstream from start_ids, including start nodes."""
+    connected = set(str(node_id) for node_id in start_ids)
+    queue = deque(connected)
+    while queue:
+        node_id = queue.popleft()
+        node = prompt_obj.get(node_id) or {}
+        inputs = node.get("inputs", {})
+        for value in inputs.values():
+            if isinstance(value, list) and len(value) == 2:
+                source_id = str(value[0])
+                if source_id in prompt_obj and source_id not in connected:
+                    connected.add(source_id)
+                    queue.append(source_id)
+    return connected
+
+
+_DELEGATE_MASTER_RETAINED_UPSTREAM_CLASSES = {
+    "PrimitiveBoolean",
+    "PrimitiveFloat",
+    "PrimitiveInt",
+    "PrimitiveNode",
+    "PrimitiveString",
+}
+
+_DELEGATE_MASTER_ALWAYS_RETAINED_UPSTREAM_CLASSES = {
+    "LoadImage",
+}
+
+_DELEGATE_MASTER_SAFE_SCALAR_TYPES = {"BOOLEAN", "FLOAT", "INT", "STRING"}
+_DELEGATE_MASTER_SAFE_LIST_TYPES = {"LIST"}
+
+# ComfyUI 0.23 exposes CreateList via the newer schema API rather than the
+# legacy RETURN_TYPES/INPUT_TYPES attributes. Treat it as a safe config utility
+# only after its connected inputs recursively prove safe.
+_DELEGATE_MASTER_SAFE_DYNAMIC_CONFIG_OUTPUT_CLASSES = {"CreateList"}
+_DELEGATE_MASTER_SAFE_DYNAMIC_CONFIG_INPUT_PREFIXES = {
+    "CreateList": ("inputs.",),
+}
+
+# Test hook. At runtime this stays None and the ComfyUI node registry is loaded lazily.
+_DELEGATE_MASTER_NODE_CLASS_MAPPINGS = None
+
+
+def _get_delegate_master_node_class_mappings():
+    """Return ComfyUI node-class mappings when available."""
+    if _DELEGATE_MASTER_NODE_CLASS_MAPPINGS is not None:
+        return _DELEGATE_MASTER_NODE_CLASS_MAPPINGS
+    try:
+        import nodes as comfy_nodes  # type: ignore
+    except Exception:  # pragma: no cover - depends on ComfyUI runtime imports
+        return {}
+    return getattr(comfy_nodes, "NODE_CLASS_MAPPINGS", {}) or {}
+
+
+def _get_delegate_master_node_class(class_type):
+    mappings = _get_delegate_master_node_class_mappings()
+    return mappings.get(class_type) if isinstance(mappings, dict) else None
+
+
+def _normalize_delegate_master_return_type(return_type):
+    if return_type is None:
+        return ""
+    return str(return_type).strip().upper()
+
+
+def _delegate_master_type_is_safe_scalar(type_name):
+    return type_name in _DELEGATE_MASTER_SAFE_SCALAR_TYPES
+
+
+def _delegate_master_type_is_safe_config(type_name):
+    return _delegate_master_type_is_safe_scalar(type_name) or type_name in _DELEGATE_MASTER_SAFE_LIST_TYPES
+
+
+def _delegate_master_output_is_safe_scalar(class_type, output_index):
+    """Return True when a registered node output is lightweight config data."""
+    if class_type in _DELEGATE_MASTER_SAFE_DYNAMIC_CONFIG_OUTPUT_CLASSES:
+        return True
+    node_class = _get_delegate_master_node_class(class_type)
+    return_types = getattr(node_class, "RETURN_TYPES", ()) if node_class is not None else ()
+    try:
+        output_type = return_types[int(output_index)]
+    except (IndexError, TypeError, ValueError):
+        return False
+    return _delegate_master_type_is_safe_config(_normalize_delegate_master_return_type(output_type))
+
+
+def _get_delegate_master_input_types(class_type):
+    node_class = _get_delegate_master_node_class(class_type)
+    input_types = getattr(node_class, "INPUT_TYPES", None) if node_class is not None else None
+    if callable(input_types):
+        try:
+            input_types = input_types()
+        except TypeError:
+            return {}
+    return input_types if isinstance(input_types, dict) else {}
+
+
+def _normalize_delegate_master_input_type(input_spec):
+    if isinstance(input_spec, (list, tuple)) and input_spec:
+        return _normalize_delegate_master_return_type(input_spec[0])
+    return _normalize_delegate_master_return_type(input_spec)
+
+
+def _delegate_master_input_is_safe_scalar(class_type, input_name):
+    """Return True when a registered downstream input expects config data."""
+    for prefix in _DELEGATE_MASTER_SAFE_DYNAMIC_CONFIG_INPUT_PREFIXES.get(class_type, ()):
+        if input_name.startswith(prefix):
+            return True
+    input_types = _get_delegate_master_input_types(class_type)
+    for section_name in ("required", "optional"):
+        section = input_types.get(section_name, {})
+        if isinstance(section, dict) and input_name in section:
+            input_type = _normalize_delegate_master_input_type(section[input_name])
+            return _delegate_master_type_is_safe_config(input_type)
+    return False
+
+
+def _is_delegate_master_always_retained_upstream_node(node):
+    if not isinstance(node, dict):
+        return False
+    class_type = node.get("class_type")
+    return isinstance(class_type, str) and class_type in _DELEGATE_MASTER_ALWAYS_RETAINED_UPSTREAM_CLASSES
+
+
+def _is_delegate_master_retained_upstream_node(node, output_index=0):
+    """Return True for lightweight upstream nodes safe to keep on the master."""
+    if not isinstance(node, dict):
+        return False
+    class_type = node.get("class_type")
+    if not isinstance(class_type, str):
+        return False
+    return (
+        class_type in _DELEGATE_MASTER_RETAINED_UPSTREAM_CLASSES
+        or class_type.startswith("Primitive")
+        or _delegate_master_output_is_safe_scalar(class_type, output_index)
+    )
+
+
+def _collect_delegate_master_retained_upstream_branch(
+    prompt_obj,
+    node_id,
+    output_index,
+    memo,
+    visiting,
+):
+    """Return safe retained branch nodes, or None when the branch is not safe."""
+    node_id = str(node_id)
+    cache_key = (node_id, output_index)
+    if cache_key in memo:
+        cached = memo[cache_key]
+        return None if cached is None else set(cached)
+    if cache_key in visiting:
+        memo[cache_key] = None
+        return None
+
+    node = prompt_obj.get(node_id)
+    if not _is_delegate_master_retained_upstream_node(node, output_index):
+        memo[cache_key] = None
+        return None
+
+    visiting.add(cache_key)
+    retained = {node_id}
+    inputs = node.get("inputs", {}) if isinstance(node, dict) else {}
+    class_type = node.get("class_type") if isinstance(node, dict) else None
+    for input_name, value in inputs.items():
+        if not (isinstance(value, list) and len(value) == 2):
+            continue
+        if not _delegate_master_input_is_safe_scalar(class_type, input_name):
+            visiting.remove(cache_key)
+            memo[cache_key] = None
+            return None
+        source_id = str(value[0])
+        branch = _collect_delegate_master_retained_upstream_branch(
+            prompt_obj,
+            source_id,
+            value[1],
+            memo,
+            visiting,
+        )
+        if branch is None:
+            visiting.remove(cache_key)
+            memo[cache_key] = None
+            return None
+        retained.update(branch)
+
+    visiting.remove(cache_key)
+    memo[cache_key] = frozenset(retained)
+    return retained
+
+
+def _find_delegate_master_retained_upstream_nodes(prompt_obj, start_ids):
+    """Return lightweight upstream nodes needed by kept delegate-master nodes."""
+    connected = set()
+    memo = {}
+    for node_id in start_ids:
+        node = prompt_obj.get(str(node_id)) or {}
+        inputs = node.get("inputs", {})
+        class_type = node.get("class_type") if isinstance(node, dict) else None
+        for input_name, value in inputs.items():
+            if not (isinstance(value, list) and len(value) == 2):
+                continue
+            source_node = prompt_obj.get(str(value[0]))
+            if _is_delegate_master_always_retained_upstream_node(source_node):
+                connected.add(str(value[0]))
+                continue
+            if not _delegate_master_input_is_safe_scalar(class_type, input_name):
+                continue
+            branch = _collect_delegate_master_retained_upstream_branch(
+                prompt_obj,
+                value[0],
+                value[1],
+                memo,
+                set(),
+            )
+            if branch is not None:
+                connected.update(branch)
+    return connected
+
+
+WORKER_SINK_CLASS = "DistributedWorkerSink"
+
+
+def prune_prompt_for_worker(prompt_obj, use_worker_sink=False):
+    """Prune worker prompt to distributed nodes and their upstream dependencies.
+
+    Removing downstream nodes can leave the worker prompt without an output
+    node. One is re-added after each such distributed node: the hidden
+    ``DistributedWorkerSink`` when ``use_worker_sink`` is set (the worker must
+    have it registered), otherwise PreviewImage/PreviewAudio for workers
+    running older versions. The preview nodes write every frame to the
+    worker's temp directory, which the sink avoids.
+    """
+    collector_ids = find_nodes_by_class(prompt_obj, "DistributedCollector")
+    upscale_ids = find_nodes_by_class(prompt_obj, "UltimateSDUpscaleDistributed")
+    distributed_ids = collector_ids + upscale_ids
+    if not distributed_ids:
+        return prompt_obj
+
+    connected = _find_upstream_nodes(prompt_obj, distributed_ids)
+    pruned_prompt = {}
+    for node_id in connected:
+        node = prompt_obj.get(node_id)
+        if node is not None:
+            pruned_prompt[node_id] = json.loads(json.dumps(node))
+
+    # Generate IDs from the original prompt so we never reuse IDs from pruned downstream nodes.
+    next_id = _create_numeric_id_generator(prompt_obj)
+    for dist_id in distributed_ids:
+        if dist_id not in pruned_prompt:
+            continue
+        downstream = _find_downstream_nodes(prompt_obj, [dist_id])
+        has_removed_downstream = any(node_id != dist_id for node_id in downstream)
+        if has_removed_downstream:
+            original_node = prompt_obj.get(str(dist_id), {})
+            class_type = original_node.get("class_type")
+            inputs = original_node.get("inputs", {})
+            image_connected = class_type != "DistributedCollector" or (
+                isinstance(inputs.get("images"), list)
+                and len(inputs["images"]) == 2
+            )
+            audio_connected = (
+                class_type == "DistributedCollector"
+                and isinstance(inputs.get("audio"), list)
+                and len(inputs["audio"]) == 2
+            )
+
+            if use_worker_sink and (image_connected or audio_connected):
+                pruned_prompt[next_id()] = {
+                    "inputs": {"value": [dist_id, 0 if image_connected else 1]},
+                    "class_type": WORKER_SINK_CLASS,
+                    "_meta": {"title": "Distributed Worker Sink (auto-added)"},
+                }
+            elif image_connected:
+                preview_id = next_id()
+                pruned_prompt[preview_id] = {
+                    "inputs": {"images": [dist_id, 0]},
+                    "class_type": "PreviewImage",
+                    "_meta": {"title": "Preview Image (auto-added)"},
+                }
+            elif audio_connected:
+                preview_id = next_id()
+                pruned_prompt[preview_id] = {
+                    "inputs": {"audio": [dist_id, 1]},
+                    "class_type": "PreviewAudio",
+                    "_meta": {"title": "Preview Audio (auto-added)"},
+                }
+
+    return pruned_prompt
+
+
+def prepare_delegate_master_prompt(prompt_obj, collector_ids):
+    """Prune master prompt so it only executes post-collector nodes in delegate mode."""
+    # Typed collectors need no fabricated IMAGE placeholder. Sever all media
+    # links, including STRING scalar branches, before retained-config analysis.
+    # Their expected_media control was frozen by participant overrides.
+    if any((prompt_obj.get(cid) or {}).get("class_type") == "DistributedMediaCollector"
+           for cid in collector_ids):
+        from ...serverless_bridge.contract import SOCKET_KINDS
+        prompt_obj = json.loads(json.dumps(prompt_obj))
+        for cid in collector_ids:
+            node = prompt_obj.get(cid) or {}
+            if node.get("class_type") == "DistributedMediaCollector":
+                inputs = node.setdefault("inputs", {})
+                inputs.setdefault("expected_media", json.dumps([name for name in SOCKET_KINDS if name in inputs]))
+                for name in SOCKET_KINDS:
+                    inputs.pop(name, None)
+    downstream = _find_downstream_nodes(prompt_obj, collector_ids)
+    nodes_to_keep = set(collector_ids)
+    nodes_to_keep.update(downstream)
+    nodes_to_keep.update(
+        _find_delegate_master_retained_upstream_nodes(prompt_obj, nodes_to_keep)
+    )
+
+    pruned_prompt = {}
+    for node_id in nodes_to_keep:
+        node = prompt_obj.get(node_id)
+        if node is not None:
+            pruned_prompt[node_id] = json.loads(json.dumps(node))
+
+    pruned_ids = set(pruned_prompt.keys())
+    for node_id, node in pruned_prompt.items():
+        inputs = node.get("inputs")
+        if not inputs:
+            continue
+        for input_name, input_value in list(inputs.items()):
+            if isinstance(input_value, list) and len(input_value) == 2:
+                source_id = str(input_value[0])
+                if source_id not in pruned_ids:
+                    inputs.pop(input_name, None)
+                    debug_log(
+                        f"Removed upstream reference '{input_name}' from node {node_id} for delegate-only master prompt."
+                    )
+
+    # Generate IDs from the original prompt to avoid ID collisions with pruned nodes.
+    next_id = _create_numeric_id_generator(prompt_obj)
+    for collector_id in collector_ids:
+        collector_entry = pruned_prompt.get(collector_id)
+        if not collector_entry:
+            continue
+        original_inputs = (prompt_obj.get(collector_id) or {}).get("inputs", {})
+        original_images = original_inputs.get("images")
+        if not (isinstance(original_images, list) and len(original_images) == 2):
+            continue
+        placeholder_id = next_id()
+        pruned_prompt[placeholder_id] = {
+            "class_type": "DistributedEmptyImage",
+            "inputs": {
+                "height": 64,
+                "width": 64,
+                "channels": 3,
+            },
+            "_meta": {
+                "title": "Distributed Empty Image (auto-added)",
+            },
+        }
+        collector_entry.setdefault("inputs", {})["images"] = [placeholder_id, 0]
+        debug_log(
+            f"Inserted placeholder node {placeholder_id} for collector {collector_id} in delegate-only master prompt."
+        )
+
+    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("DistributedMediaCollector")
+        + 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}"
+    return job_map
+
+
+def _override_seed_nodes(prompt_copy, prompt_index, is_master, participant_id, worker_index_map):
+    """Configure DistributedSeed nodes for master or worker role."""
+    for node_id in prompt_index.nodes_for_class("DistributedSeed"):
+        node = prompt_copy.get(node_id)
+        if not isinstance(node, dict):
+            continue
+        inputs = node.setdefault("inputs", {})
+        inputs["is_worker"] = not is_master
+        if is_master:
+            inputs["worker_id"] = ""
+        else:
+            inputs["worker_id"] = f"worker_{worker_index_map.get(participant_id, 0)}"
+
+
+def _override_collector_nodes(
+    prompt_copy,
+    prompt_index,
+    is_master,
+    participant_id,
+    job_id_map,
+    master_url,
+    enabled_json,
+    delegate_master,
+):
+    """Configure legacy and typed collectors without changing their media."""
+    collector_ids = (prompt_index.nodes_for_class("DistributedCollector")
+                     + prompt_index.nodes_for_class("DistributedMediaCollector"))
+    for node_id in collector_ids:
+        node = prompt_copy.get(node_id)
+        if not isinstance(node, dict):
+            continue
+
+        if prompt_index.has_upstream(node_id, "UltimateSDUpscaleDistributed"):
+            node.setdefault("inputs", {})["pass_through"] = True
+            continue
+
+        inputs = node.setdefault("inputs", {})
+        inputs["multi_job_id"] = job_id_map.get(node_id, node_id)
+        inputs["is_worker"] = not is_master
+        inputs["enabled_worker_ids"] = enabled_json
+        if node.get("class_type") == "DistributedMediaCollector":
+            from ...serverless_bridge.contract import SOCKET_KINDS
+            inputs["expected_media"] = json.dumps([socket for socket in SOCKET_KINDS if socket in inputs])
+        if is_master:
+            inputs["delegate_only"] = bool(delegate_master)
+            inputs.pop("master_url", None)
+            inputs.pop("worker_id", None)
+        else:
+            inputs["master_url"] = master_url
+            inputs["worker_id"] = participant_id
+            inputs["delegate_only"] = False
+
+
+def _override_upscale_nodes(
+    prompt_copy,
+    prompt_index,
+    is_master,
+    participant_id,
+    job_id_map,
+    master_url,
+    enabled_json,
+):
+    """Configure UltimateSDUpscaleDistributed nodes for master or worker role."""
+    for node_id in prompt_index.nodes_for_class("UltimateSDUpscaleDistributed"):
+        node = prompt_copy.get(node_id)
+        if not isinstance(node, dict):
+            continue
+        inputs = node.setdefault("inputs", {})
+        inputs["multi_job_id"] = job_id_map.get(node_id, node_id)
+        inputs["is_worker"] = not is_master
+        inputs["enabled_worker_ids"] = enabled_json
+        if is_master:
+            inputs.pop("master_url", None)
+            inputs.pop("worker_id", None)
+        else:
+            inputs["master_url"] = master_url
+            inputs["worker_id"] = participant_id
+
+
+def _override_value_nodes(prompt_copy, prompt_index, is_master, participant_id, worker_index_map):
+    """Configure DistributedValue nodes for master or worker role."""
+    for node_id in prompt_index.nodes_for_class("DistributedValue"):
+        node = prompt_copy.get(node_id)
+        if not isinstance(node, dict):
+            continue
+        inputs = node.setdefault("inputs", {})
+        inputs["is_worker"] = not is_master
+        if is_master:
+            inputs["worker_id"] = ""
+        else:
+            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,
+    enabled_worker_ids,
+    job_id_map,
+    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"
+    worker_index_map = {wid: idx for idx, wid in enumerate(enabled_worker_ids)}
+    enabled_json = json.dumps(enabled_worker_ids)
+
+    _override_seed_nodes(prompt_copy, prompt_index, is_master, participant_id, worker_index_map)
+    _override_value_nodes(prompt_copy, prompt_index, is_master, participant_id, worker_index_map)
+    _override_collector_nodes(
+        prompt_copy,
+        prompt_index,
+        is_master,
+        participant_id,
+        job_id_map,
+        master_url,
+        enabled_json,
+        delegate_master,
+    )
+    _override_upscale_nodes(
+        prompt_copy,
+        prompt_index,
+        is_master,
+        participant_id,
+        job_id_map,
+        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
@@ -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)}")
@@ -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])
@@ -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)
@@ -0,0 +1,199 @@
+"""Pure compilation of prepared typed-media collectors for serverless workers.
+
+The legacy image compiler remains separate. This compiler emits one MediaExportV1
+execution target per connected socket, 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, DEFAULT_QUALITY, validate_options
+from .serverless_prompt import (
+    _PRIMITIVE_NODES,
+    _REMOVED_SCALAR_NODES,
+    _is_link,
+    _is_unsupported_distributed,
+    _make_scalar_resolver,
+    _prune_graph,
+)
+
+
+_COLLECTOR = "DistributedMediaCollector"
+
+
+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 class_type == "DistributedCollector" or _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 DistributedMediaCollector 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(SOCKET_KINDS):
+                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", {})
+        validate_options(inputs.get("transfer_format", DEFAULT_FORMAT),
+                         inputs.get("transfer_quality", DEFAULT_QUALITY))
+        for field in ("multi_job_id", "worker_id"):
+            _validate_identifier(inputs.get(field), f"collector {collector_id!r} {field}")
+        connected = [socket for socket in SOCKET_KINDS 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, kind in SOCKET_KINDS.items():
+            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 = validate_options(inputs.get("transfer_format", DEFAULT_FORMAT),
+                                                inputs.get("transfer_quality", DEFAULT_QUALITY))
+                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 DistributedMediaCollector must have prepared literal ``multi_job_id``
+    and ``worker_id`` identities and at least one linked media input. Present
+    media inputs (including texts) must be API links; literal text belongs in a
+    PrimitiveString node. Nonmedia collector 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
@@ -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
@@ -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)]]
@@ -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)
@@ -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
@@ -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
@@ -0,0 +1,640 @@
+import asyncio
+import copy
+import time
+import uuid
+
+import server
+
+from ..utils.async_helpers import queue_prompt_payload
+from ..utils.config import load_config
+from ..utils.constants import (
+    ORCHESTRATION_MEDIA_SYNC_CONCURRENCY,
+    ORCHESTRATION_MEDIA_SYNC_TIMEOUT,
+    ORCHESTRATION_WORKER_PROBE_CONCURRENCY,
+    ORCHESTRATION_WORKER_PREP_CONCURRENCY,
+)
+from ..utils.logging import debug_log, log
+from ..utils.network import build_master_url, build_master_callback_url
+from ..utils.trace_logger import trace_debug
+from .schemas import parse_positive_float, parse_positive_int
+from .orchestration.dispatch import (
+    dispatch_worker_prompt,
+    select_active_workers,
+    select_least_busy_worker,
+)
+from .orchestration.media_sync import (
+    convert_paths_for_platform,
+    fetch_worker_has_node,
+    fetch_worker_path_separator,
+    sync_worker_media,
+)
+from .orchestration.prompt_transform import (
+    PromptIndex,
+    apply_participant_overrides,
+    find_nodes_by_class,
+    generate_job_id_map,
+    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
+
+
+def _generate_execution_trace_id():
+    return f"exec_{int(time.time() * 1000)}_{uuid.uuid4().hex[:6]}"
+
+
+def ensure_distributed_state(server_instance=None):
+    """Ensure prompt_server has the state used by distributed queue orchestration."""
+    ps = server_instance or prompt_server
+    if not hasattr(ps, "distributed_pending_jobs"):
+        ps.distributed_pending_jobs = {}
+    if not hasattr(ps, "distributed_jobs_lock"):
+        ps.distributed_jobs_lock = asyncio.Lock()
+
+
+# Initialize top-level distributed queue state at module import time.
+ensure_distributed_state()
+
+
+async def _ensure_distributed_queue(job_id):
+    """Ensure a queue exists for the given distributed job ID."""
+    ensure_distributed_state()
+    async with prompt_server.distributed_jobs_lock:
+        if job_id not in prompt_server.distributed_pending_jobs:
+            prompt_server.distributed_pending_jobs[job_id] = asyncio.Queue()
+
+
+def _resolve_enabled_workers(config, requested_ids=None):
+    """Return a list of worker configs that should participate."""
+    workers = []
+    for worker in config.get("workers", []):
+        worker_id = str(worker.get("id") or "").strip()
+        if not worker_id:
+            continue
+
+        if requested_ids is not None:
+            if worker_id not in requested_ids:
+                continue
+        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)
+        except (TypeError, ValueError):
+            log(f"[Distributed] Invalid port '{raw_port}' for worker {worker_id}; defaulting to 8188.")
+            port = 8188
+
+        workers.append(
+            {
+                "id": worker_id,
+                "name": worker.get("name", worker_id),
+                "host": worker.get("host"),
+                "port": port,
+                "type": worker.get("type", "local"),
+            }
+        )
+    return workers
+
+
+def _resolve_orchestration_limits(config):
+    """Resolve bounded concurrency/timeouts for worker preparation pipeline."""
+    settings = (config or {}).get("settings", {}) or {}
+    worker_probe_concurrency = parse_positive_int(
+        settings.get("worker_probe_concurrency"),
+        ORCHESTRATION_WORKER_PROBE_CONCURRENCY,
+    )
+    worker_prep_concurrency = parse_positive_int(
+        settings.get("worker_prep_concurrency"),
+        ORCHESTRATION_WORKER_PREP_CONCURRENCY,
+    )
+    media_sync_concurrency = parse_positive_int(
+        settings.get("media_sync_concurrency"),
+        ORCHESTRATION_MEDIA_SYNC_CONCURRENCY,
+    )
+    media_sync_timeout_seconds = parse_positive_float(
+        settings.get("media_sync_timeout_seconds"),
+        ORCHESTRATION_MEDIA_SYNC_TIMEOUT,
+    )
+    return (
+        worker_probe_concurrency,
+        worker_prep_concurrency,
+        media_sync_concurrency,
+        media_sync_timeout_seconds,
+    )
+
+
+def _is_load_balance_enabled(value):
+    if isinstance(value, bool):
+        return value
+    if isinstance(value, (int, float)):
+        return bool(value)
+    if isinstance(value, str):
+        return value.strip().lower() in {"1", "true", "yes", "on"}
+    return False
+
+
+def _prompt_requests_load_balance(prompt_index):
+    collector_ids = (prompt_index.nodes_for_class("DistributedCollector")
+                     + prompt_index.nodes_for_class("DistributedMediaCollector"))
+    for node_id in collector_ids:
+        inputs = prompt_index.inputs_by_node.get(node_id, {})
+        if _is_load_balance_enabled(inputs.get("load_balance", False)):
+            return True
+    return False
+
+
+def _validate_media_workers(prompt_index, workers):
+    """Reject incompatible selections before probing, pruning or dispatching.
+
+    Typed collectors currently have only a master/serverless transport. Check
+    all selected workers, not just reachable or load-balanced participants, so
+    a bad selection cannot silently turn into a different execution mode.
+    """
+    typed = bool(prompt_index.nodes_for_class("DistributedMediaCollector"))
+    if typed and (prompt_index.nodes_for_class("DistributedCollector")
+                  or prompt_index.nodes_for_class("UltimateSDUpscaleDistributed")):
+        raise ValueError(
+            "DistributedMediaCollector cannot be mixed with DistributedCollector or "
+            "UltimateSDUpscaleDistributed in one prompt"
+        )
+    for worker in workers:
+        if worker.get("type") != "serverless":
+            if typed:
+                raise ValueError(
+                    "DistributedMediaCollector does not support selected direct workers; "
+                    "use master execution or comfy-media-v1 serverless pools"
+                )
+            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 typed != (profile == "comfy-media-v1"):
+            raise ValueError(
+                "DistributedMediaCollector requires comfy-media-v1 serverless pools; "
+                "legacy collectors require legacy-image-v1"
+            )
+
+
+async def _prepare_worker_payload(
+    worker,
+    prompt_index,
+    enabled_ids,
+    job_id_map,
+    master_url,
+    config,
+    delegate_master,
+    trace_execution_id,
+    worker_prep_semaphore,
+    media_sync_semaphore,
+    media_sync_timeout_seconds,
+):
+    """Prepare one worker prompt payload with bounded concurrency and media-sync timeout."""
+    async with worker_prep_semaphore:
+        _validate_media_workers(prompt_index, [worker])
+        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 legacy pruner only recognizes Collector/Upscale roots.
+                # The typed compiler validates the complete private graph and
+                # prunes it itself, retaining every connected typed 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,
+            prompt_server_instance=prompt_server,
+        )
+
+        worker_type = str(worker.get("type") or "local").strip().lower()
+        is_remote_like = bool(worker.get("host")) and worker_type != "local"
+        # Older workers lack the sink node; they keep the PreviewImage terminator.
+        has_sink_task = asyncio.ensure_future(
+            fetch_worker_has_node(worker, WORKER_SINK_CLASS, trace_execution_id=trace_execution_id)
+        )
+        try:
+            if is_remote_like:
+                path_separator = await fetch_worker_path_separator(worker, trace_execution_id=trace_execution_id)
+                if path_separator:
+                    worker_prompt = convert_paths_for_platform(worker_prompt, path_separator)
+            use_worker_sink = await has_sink_task
+        finally:
+            if not has_sink_task.done():
+                has_sink_task.cancel()
+
+        worker_prompt = prune_prompt_for_worker(worker_prompt, use_worker_sink=use_worker_sink)
+        worker_prompt = apply_participant_overrides(
+            worker_prompt,
+            worker["id"],
+            enabled_ids,
+            job_id_map,
+            worker_master_url,
+            delegate_master,
+            prompt_index,
+        )
+
+        if is_remote_like:
+            async with media_sync_semaphore:
+                try:
+                    await asyncio.wait_for(
+                        sync_worker_media(worker, worker_prompt, trace_execution_id=trace_execution_id),
+                        timeout=media_sync_timeout_seconds,
+                    )
+                except asyncio.TimeoutError:
+                    trace_debug(
+                        trace_execution_id,
+                        (
+                            f"Media sync timed out after {media_sync_timeout_seconds:.1f}s "
+                            f"for worker {worker.get('name')} ({worker.get('id')}); continuing dispatch."
+                        ),
+                    )
+
+        return worker, worker_prompt
+
+
+async def orchestrate_distributed_execution(
+    prompt_obj,
+    workflow_meta,
+    client_id,
+    enabled_worker_ids=None,
+    delegate_master=None,
+    trace_execution_id=None,
+):
+    """Core orchestration logic for the /distributed/queue endpoint.
+
+    Returns:
+        tuple[str, int, int, dict]: (prompt_id, number, worker_count, node_errors)
+    """
+    ensure_distributed_state()
+    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)
+    (
+        worker_probe_concurrency,
+        worker_prep_concurrency,
+        media_sync_concurrency,
+        media_sync_timeout_seconds,
+    ) = _resolve_orchestration_limits(config)
+    requested_ids = enabled_worker_ids if enabled_worker_ids is not None else None
+    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:
+        if any(isinstance(node, dict) and node.get("class_type") in {
+                "DistributedCollector", "DistributedMediaCollector", "UltimateSDUpscaleDistributed"
+        } for node in prompt_obj.values()):
+            raise ValueError("Distributed Image List cannot be mixed with other distributed collectors")
+        prompt_obj, outer_targets = prepare_image_list_execution(prompt_obj)
+        target_kwargs["partial_execution_targets"] = outer_targets
+    prompt_index = PromptIndex(prompt_obj)
+    typed_media_mode = bool(prompt_index.nodes_for_class("DistributedMediaCollector"))
+    if typed_media_mode:
+        from ..serverless_bridge.contract import identifier
+        # PromptIndex's JSON copy would otherwise turn integer keys into strings.
+        for node_id in prompt_obj:
+            identifier(node_id)
+    image_list_start_ids = prompt_index.nodes_for_class("DistributedImageListStart")
+    image_list_mode = bool(image_list_start_ids)
+    if image_list_mode:
+        # Validate before probing or queueing anything so malformed boundaries and
+        # old worker releases fail before any upstream image is materialized.
+        other_distributed = (
+            prompt_index.nodes_for_class("DistributedCollector")
+            + prompt_index.nodes_for_class("DistributedMediaCollector")
+            + prompt_index.nodes_for_class("UltimateSDUpscaleDistributed")
+        )
+        if other_distributed:
+            raise ValueError(
+                "Distributed Image List v1 cannot be mixed with Distributed Collector, "
+                "DistributedMediaCollector or Ultimate SD Upscale nodes in one prompt"
+            )
+        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:
+        if typed_media_mode and requested_ids is not None:
+            configured_ids = {worker.get("id") for worker in config.get("workers", [])}
+            if any(worker_id not in configured_ids for worker_id in requested_ids):
+                raise ValueError("Unknown explicitly selected media worker ID")
+        workers = _resolve_enabled_workers(config, requested_ids)
+        _validate_media_workers(prompt_index, workers)
+        if typed_media_mode and delegate_master and not workers:
+            raise RuntimeError("Delegate-only DistributedMediaCollector requires a selected comfy-media-v1 pool")
+    load_balance_requested = _prompt_requests_load_balance(prompt_index)
+    trace_debug(
+        execution_trace_id,
+        (
+            f"Orchestration start: requested_workers={len(workers)}, "
+            f"requested_ids={requested_ids if requested_ids is not None else 'enabled_only'}, "
+            f"websocket={use_websocket}, "
+            f"probe_concurrency={worker_probe_concurrency}, "
+            f"prep_concurrency={worker_prep_concurrency}, "
+            f"media_sync_concurrency={media_sync_concurrency}, "
+            f"media_sync_timeout={media_sync_timeout_seconds:.1f}s, "
+            f"load_balance={load_balance_requested}"
+        ),
+    )
+
+    # Keep the legacy Collector fallback. Image-list and typed-media delegate
+    # modes must not silently enable master participation.
+    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, selected_delegate_master = await select_active_workers(
+        workers,
+        use_websocket,
+        False if image_list_mode else delegate_master,
+        trace_execution_id=execution_trace_id,
+        probe_concurrency=worker_probe_concurrency,
+    )
+    if typed_media_mode and delegate_master and not active_workers:
+        raise RuntimeError("No selected comfy-media-v1 worker is reachable for delegate-only DistributedMediaCollector")
+    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"
+            )
+        idle_ids = (await preflight_image_list_workers(active_workers, require_idle=delegate_master)
+                    if active_workers else {})
+        active_workers = [worker for worker in active_workers if worker["id"] in idle_ids]
+        if not active_workers and delegate_master:
+            raise RuntimeError("No feature-compatible direct worker is idle in orchestrator-only mode")
+    else:
+        delegate_master = selected_delegate_master
+
+    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.
+            candidate_workers.append(
+                {
+                    "id": "master",
+                    "name": "Master",
+                    "host": master_url,
+                    "type": "local",
+                }
+            )
+
+        selected_worker = None
+        if candidate_workers:
+            selected_worker = await select_least_busy_worker(
+                candidate_workers,
+                trace_execution_id=execution_trace_id,
+                probe_concurrency=worker_probe_concurrency,
+            )
+        if selected_worker is None and candidate_workers:
+            trace_debug(
+                execution_trace_id,
+                "Load-balance selection probe failed; using first available candidate.",
+            )
+            selected_worker = candidate_workers[0]
+
+        if selected_worker is not None and str(selected_worker.get("id")) == "master":
+            # Master selected as least busy; run master workload only.
+            active_workers = []
+            delegate_master = False
+            trace_debug(
+                execution_trace_id,
+                "Load-balance selected master for execution (workers skipped).",
+            )
+        elif selected_worker is not None:
+            active_workers = [selected_worker]
+            # Worker selected as least busy; keep master orchestrator-only for this run.
+            delegate_master = True
+            trace_debug(
+                execution_trace_id,
+                f"Load-balance selected worker {selected_worker.get('id')} (master set to delegate-only).",
+            )
+        else:
+            trace_debug(
+                execution_trace_id,
+                "Load-balance requested but no execution candidates were available.",
+            )
+            active_workers = []
+            delegate_master = False
+
+    enabled_ids = [worker["id"] for worker in active_workers]
+
+    discovery_prefix = f"exec_{int(time.time() * 1000)}_{uuid.uuid4().hex[:6]}"
+    job_id_map = generate_job_id_map(prompt_index, discovery_prefix)
+
+    if not job_id_map:
+        trace_debug(execution_trace_id, "No distributed nodes detected; queueing prompt on master only.")
+        queue_result = await queue_prompt_payload(
+            prompt_obj,
+            workflow_meta,
+            client_id,
+            include_queue_metadata=True,
+            **target_kwargs,
+        )
+        return (
+            queue_result["prompt_id"],
+            queue_result["number"],
+            0,
+            queue_result.get("node_errors", {}),
+        )
+
+    master_prompt = prompt_index.copy_prompt()
+    master_prompt = apply_participant_overrides(
+        master_prompt,
+        "master",
+        enabled_ids,
+        job_id_map,
+        master_url,
+        delegate_master,
+        prompt_index,
+        workflow_meta=workflow_meta,
+    )
+
+    if delegate_master and not image_list_mode:
+        collector_ids = (find_nodes_by_class(master_prompt, "DistributedCollector")
+                         + find_nodes_by_class(master_prompt, "DistributedMediaCollector"))
+        upscale_nodes = find_nodes_by_class(master_prompt, "UltimateSDUpscaleDistributed")
+        if upscale_nodes:
+            debug_log(
+                "Delegate-only master mode currently does not support UltimateSDUpscaleDistributed nodes; running full prompt on master."
+            )
+        elif not collector_ids:
+            debug_log(
+                "Delegate-only master mode requested but no collectors found in master prompt. Running full prompt on master."
+            )
+        else:
+            master_prompt = prepare_delegate_master_prompt(master_prompt, collector_ids)
+
+    if active_workers:
+        trace_debug(
+            execution_trace_id,
+            "Active distributed workers: "
+            + ", ".join(f"{worker['name']} ({worker['id']})" for worker in active_workers),
+        )
+    worker_payloads = []
+    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(
+            *[
+                _prepare_worker_payload(
+                    worker,
+                    prompt_index,
+                    enabled_ids,
+                    job_id_map,
+                    master_url,
+                    config,
+                    delegate_master,
+                    execution_trace_id,
+                    worker_prep_semaphore,
+                    media_sync_semaphore,
+                    media_sync_timeout_seconds,
+                )
+                for worker in active_workers
+            ]
+        )
+
+    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(
+                    worker,
+                    wprompt,
+                    workflow_meta,
+                    client_id,
+                    use_websocket=use_websocket,
+                    trace_execution_id=execution_trace_id,
+                )
+                for worker, wprompt in regular_payloads
+            ]
+        )
+
+    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(active_workers) if image_list_mode else len(worker_payloads)), node_errors
@@ -0,0 +1,79 @@
+from dataclasses import dataclass
+from typing import Any, Dict, List, Optional
+
+
+@dataclass(frozen=True)
+class QueueRequestPayload:
+    prompt: Dict[str, Any]
+    workflow_meta: Any
+    client_id: str
+    delegate_master: Optional[bool]
+    enabled_worker_ids: List[str]
+    auto_prepare: bool
+    trace_execution_id: Optional[str]
+
+
+def parse_queue_request_payload(data: Any) -> QueueRequestPayload:
+    """Parse and validate /distributed/queue payload into a normalized shape."""
+    if not isinstance(data, dict):
+        raise ValueError("Expected a JSON object body")
+
+    auto_prepare_raw = data.get("auto_prepare", True)
+    if not isinstance(auto_prepare_raw, bool):
+        raise ValueError("auto_prepare must be a boolean when provided")
+    auto_prepare = auto_prepare_raw
+
+    prompt = data.get("prompt")
+    # Auto-prepare is always on server-side; keep the field for wire compatibility.
+    if prompt is None:
+        workflow_payload = data.get("workflow")
+        if isinstance(workflow_payload, dict):
+            candidate_prompt = workflow_payload.get("prompt")
+            if isinstance(candidate_prompt, dict):
+                prompt = candidate_prompt
+
+    if not isinstance(prompt, dict):
+        raise ValueError("Field 'prompt' must be an object")
+
+    enabled_ids_raw = data.get("enabled_worker_ids")
+    workers_field = data.get("workers")
+    if enabled_ids_raw is None and workers_field is not None:
+        if not isinstance(workers_field, list):
+            raise ValueError("Field 'workers' must be a list when provided")
+        enabled_ids_raw = []
+        for entry in workers_field:
+            worker_id = entry.get("id") if isinstance(entry, dict) else entry
+            if worker_id is not None:
+                enabled_ids_raw.append(str(worker_id))
+
+    if enabled_ids_raw is None:
+        raise ValueError("enabled_worker_ids required")
+    else:
+        if not isinstance(enabled_ids_raw, list):
+            raise ValueError("enabled_worker_ids must be a list of worker IDs")
+        enabled_ids = [str(worker_id).strip() for worker_id in enabled_ids_raw if str(worker_id).strip()]
+
+    delegate_master = data.get("delegate_master")
+    if delegate_master is not None and not isinstance(delegate_master, bool):
+        raise ValueError("delegate_master must be a boolean when provided")
+
+    client_id = data.get("client_id")
+    if not isinstance(client_id, str) or not client_id.strip():
+        raise ValueError("client_id required")
+    client_id = client_id.strip()
+
+    trace_execution_id = data.get("trace_execution_id")
+    if trace_execution_id is not None:
+        if not isinstance(trace_execution_id, str):
+            raise ValueError("trace_execution_id must be a string when provided")
+        trace_execution_id = trace_execution_id.strip() or None
+
+    return QueueRequestPayload(
+        prompt=prompt,
+        workflow_meta=data.get("workflow"),
+        client_id=client_id,
+        delegate_master=delegate_master,
+        enabled_worker_ids=enabled_ids,
+        auto_prepare=auto_prepare,
+        trace_execution_id=trace_execution_id,
+    )
@@ -0,0 +1,54 @@
+def require_fields(data: dict, *fields) -> list[str]:
+    """Return field names that are missing or empty in a JSON object."""
+    if not isinstance(data, dict):
+        return list(fields)
+
+    missing = []
+    for field in fields:
+        if field not in data:
+            missing.append(field)
+            continue
+        value = data.get(field)
+        if value is None:
+            missing.append(field)
+            continue
+        if isinstance(value, str) and not value.strip():
+            missing.append(field)
+
+    return missing
+
+
+def validate_worker_id(worker_id: str, config: dict) -> bool:
+    """Return True when worker_id exists in config['workers']."""
+    worker_id_str = str(worker_id)
+    workers = (config or {}).get("workers", [])
+    return any(str(worker.get("id")) == worker_id_str for worker in workers)
+
+
+def validate_positive_int(value, field_name: str) -> str | None:
+    """Validate positive integers and return an error string when invalid."""
+    try:
+        parsed = int(value)
+    except (TypeError, ValueError):
+        return f"Field '{field_name}' must be a positive integer."
+    if parsed <= 0:
+        return f"Field '{field_name}' must be a positive integer."
+    return None
+
+
+def parse_positive_int(value, default: int) -> int:
+    """Parse value as positive int, returning default on failure."""
+    try:
+        parsed = int(value)
+    except (TypeError, ValueError):
+        return max(1, int(default))
+    return max(1, parsed)
+
+
+def parse_positive_float(value, default: float) -> float:
+    """Parse value as positive float, returning default on failure."""
+    try:
+        parsed = float(value)
+    except (TypeError, ValueError):
+        return max(0.0, float(default))
+    return max(0.0, parsed)
@@ -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)
@@ -0,0 +1,51 @@
+from aiohttp import web
+import server
+
+from ..utils.cloudflare import cloudflare_tunnel_manager
+from ..utils.config import load_config
+from ..utils.logging import debug_log, log
+from ..utils.network import handle_api_error
+
+
+@server.PromptServer.instance.routes.get("/distributed/tunnel/status")
+async def tunnel_status_endpoint(request):
+    """Return Cloudflare tunnel status and last known details."""
+    try:
+        status = cloudflare_tunnel_manager.get_status()
+        config = load_config()
+        master_host = (config.get("master") or {}).get("host")
+        return web.json_response({
+            "status": "success",
+            "tunnel": status,
+            "master_host": master_host
+        })
+    except Exception as e:
+        return await handle_api_error(request, e, 500)
+
+@server.PromptServer.instance.routes.post("/distributed/tunnel/start")
+async def tunnel_start_endpoint(request):
+    """Start a Cloudflare tunnel pointing at the current ComfyUI server."""
+    try:
+        result = await cloudflare_tunnel_manager.start_tunnel()
+        config = load_config()
+        return web.json_response({
+            "status": "success",
+            "tunnel": result,
+            "master_host": (config.get("master") or {}).get("host")
+        })
+    except Exception as e:
+        return await handle_api_error(request, e, 500)
+
+@server.PromptServer.instance.routes.post("/distributed/tunnel/stop")
+async def tunnel_stop_endpoint(request):
+    """Stop the managed Cloudflare tunnel if running."""
+    try:
+        result = await cloudflare_tunnel_manager.stop_tunnel()
+        config = load_config()
+        return web.json_response({
+            "status": "success",
+            "tunnel": result,
+            "master_host": (config.get("master") or {}).get("host")
+        })
+    except Exception as e:
+        return await handle_api_error(request, e, 500)
@@ -0,0 +1,228 @@
+import asyncio
+import io
+import time
+
+from aiohttp import web
+from PIL import Image
+import server
+
+from ..upscale.job_models import BaseJobState, ImageJobState, TileJobState
+from ..upscale.job_store import MAX_PAYLOAD_SIZE, ensure_tile_jobs_initialized
+from ..upscale.payload_parsers import _parse_tiles_from_form
+from ..utils.logging import debug_log
+from ..utils.network import handle_api_error
+
+
+@server.PromptServer.instance.routes.post("/distributed/heartbeat")
+async def heartbeat_endpoint(request):
+    try:
+        data = await request.json()
+        worker_id = data.get('worker_id')
+        multi_job_id = data.get('multi_job_id')
+
+        if not worker_id or not multi_job_id:
+            return await handle_api_error(request, "Missing worker_id or multi_job_id", 400)
+
+        prompt_server = ensure_tile_jobs_initialized()
+        async with prompt_server.distributed_tile_jobs_lock:
+            if multi_job_id in prompt_server.distributed_pending_tile_jobs:
+                job_data = prompt_server.distributed_pending_tile_jobs[multi_job_id]
+                if isinstance(job_data, BaseJobState):
+                    job_data.worker_status[worker_id] = time.time()
+                    debug_log(f"Heartbeat from worker {worker_id}")
+                    return web.json_response({"status": "success"})
+                return await handle_api_error(request, "Worker status tracking not available", 400)
+            return await handle_api_error(request, "Job not found", 404)
+    except Exception as e:
+        return await handle_api_error(request, e, 500)
+
+
+@server.PromptServer.instance.routes.post("/distributed/submit_tiles")
+async def submit_tiles_endpoint(request):
+    """Endpoint for workers to submit processed tiles in static mode."""
+    try:
+        content_length = request.headers.get('content-length')
+        if content_length and int(content_length) > MAX_PAYLOAD_SIZE:
+            return await handle_api_error(request, f"Payload too large: {content_length} bytes", 413)
+
+        data = await request.post()
+        multi_job_id = data.get('multi_job_id')
+        worker_id = data.get('worker_id')
+        is_last = data.get('is_last', 'False').lower() == 'true'
+
+        if multi_job_id is None or worker_id is None:
+            return await handle_api_error(request, "Missing multi_job_id or worker_id", 400)
+
+        prompt_server = ensure_tile_jobs_initialized()
+
+        batch_size = int(data.get('batch_size', 0))
+
+        # Handle completion signal
+        if batch_size == 0 and is_last:
+            async with prompt_server.distributed_tile_jobs_lock:
+                if multi_job_id in prompt_server.distributed_pending_tile_jobs:
+                    job_data = prompt_server.distributed_pending_tile_jobs[multi_job_id]
+                    if not isinstance(job_data, TileJobState):
+                        return await handle_api_error(request, "Job not configured for tile submissions", 400)
+                    await job_data.queue.put({
+                        'worker_id': worker_id,
+                        'is_last': True,
+                        'tiles': [],
+                    })
+                    debug_log(f"Received completion signal from worker {worker_id}")
+                    return web.json_response({"status": "success"})
+
+        try:
+            tiles = _parse_tiles_from_form(data)
+        except ValueError as e:
+            return await handle_api_error(request, str(e), 400)
+
+        # Submit tiles to queue
+        async with prompt_server.distributed_tile_jobs_lock:
+            if multi_job_id in prompt_server.distributed_pending_tile_jobs:
+                job_data = prompt_server.distributed_pending_tile_jobs[multi_job_id]
+                if not isinstance(job_data, TileJobState):
+                    return await handle_api_error(request, "Job not configured for tile submissions", 400)
+
+                q = job_data.queue
+                if batch_size > 0 or len(tiles) > 0:
+                    await q.put({
+                        'worker_id': worker_id,
+                        'tiles': tiles,
+                        'is_last': is_last,
+                    })
+                    debug_log(f"Received {len(tiles)} tiles from worker {worker_id} (is_last={is_last})")
+                else:
+                    await q.put({
+                        'worker_id': worker_id,
+                        'is_last': True,
+                        'tiles': [],
+                    })
+
+                return web.json_response({"status": "success"})
+            return await handle_api_error(request, "Job not found", 404)
+    except Exception as e:
+        return await handle_api_error(request, e, 500)
+
+
+@server.PromptServer.instance.routes.post("/distributed/submit_image")
+async def submit_image_endpoint(request):
+    """Endpoint for workers to submit processed images in dynamic mode."""
+    try:
+        content_length = request.headers.get('content-length')
+        if content_length and int(content_length) > MAX_PAYLOAD_SIZE:
+            return await handle_api_error(request, f"Payload too large: {content_length} bytes", 413)
+
+        data = await request.post()
+        multi_job_id = data.get('multi_job_id')
+        worker_id = data.get('worker_id')
+        is_last = data.get('is_last', 'False').lower() == 'true'
+
+        if multi_job_id is None or worker_id is None:
+            return await handle_api_error(request, "Missing multi_job_id or worker_id", 400)
+
+        prompt_server = ensure_tile_jobs_initialized()
+
+        # Handle image submission
+        if 'full_image' in data and 'image_idx' in data:
+            image_idx = int(data.get('image_idx'))
+            img_data = data['full_image'].file.read()
+            img = Image.open(io.BytesIO(img_data)).convert("RGB")
+
+            debug_log(f"Received full image {image_idx} from worker {worker_id}")
+
+            async with prompt_server.distributed_tile_jobs_lock:
+                if multi_job_id in prompt_server.distributed_pending_tile_jobs:
+                    job_data = prompt_server.distributed_pending_tile_jobs[multi_job_id]
+                    if not isinstance(job_data, ImageJobState):
+                        return await handle_api_error(request, "Job not configured for image submissions", 400)
+                    await job_data.queue.put({
+                        'worker_id': worker_id,
+                        'image_idx': image_idx,
+                        'image': img,
+                        'is_last': is_last,
+                    })
+                    return web.json_response({"status": "success"})
+
+        # Handle completion signal (no image data)
+        elif is_last:
+            async with prompt_server.distributed_tile_jobs_lock:
+                if multi_job_id in prompt_server.distributed_pending_tile_jobs:
+                    job_data = prompt_server.distributed_pending_tile_jobs[multi_job_id]
+                    if not isinstance(job_data, ImageJobState):
+                        return await handle_api_error(request, "Job not configured for image submissions", 400)
+                    await job_data.queue.put({
+                        'worker_id': worker_id,
+                        'is_last': True,
+                    })
+                    debug_log(f"Received completion signal from worker {worker_id}")
+                    return web.json_response({"status": "success"})
+        else:
+            return await handle_api_error(request, "Missing image data or invalid request", 400)
+
+        return await handle_api_error(request, "Job not found", 404)
+    except Exception as e:
+        return await handle_api_error(request, e, 500)
+
+
+@server.PromptServer.instance.routes.post("/distributed/request_image")
+async def request_image_endpoint(request):
+    """Endpoint for workers to request tasks (images in dynamic mode, tiles in static mode)."""
+    try:
+        data = await request.json()
+        worker_id = data.get('worker_id')
+        multi_job_id = data.get('multi_job_id')
+
+        if not worker_id or not multi_job_id:
+            return await handle_api_error(request, "Missing worker_id or multi_job_id", 400)
+
+        prompt_server = ensure_tile_jobs_initialized()
+        async with prompt_server.distributed_tile_jobs_lock:
+            if multi_job_id in prompt_server.distributed_pending_tile_jobs:
+                job_data = prompt_server.distributed_pending_tile_jobs[multi_job_id]
+                if not isinstance(job_data, BaseJobState):
+                    return await handle_api_error(request, "Invalid job data structure", 500)
+
+                mode = job_data.mode
+                if isinstance(job_data, ImageJobState):
+                    pending_queue = job_data.pending_images
+                elif isinstance(job_data, TileJobState):
+                    pending_queue = job_data.pending_tasks
+                else:
+                    return await handle_api_error(request, "Invalid job configuration", 400)
+
+                try:
+                    task_idx = await asyncio.wait_for(pending_queue.get(), timeout=0.1)
+                    job_data.assigned_to_workers.setdefault(worker_id, []).append(task_idx)
+                    job_data.worker_status[worker_id] = time.time()
+                    remaining = pending_queue.qsize()
+
+                    if mode == 'dynamic':
+                        debug_log(f"UltimateSDUpscale API - Assigned image {task_idx} to worker {worker_id}")
+                        return web.json_response({"image_idx": task_idx, "estimated_remaining": remaining})
+                    debug_log(f"UltimateSDUpscale API - Assigned tile {task_idx} to worker {worker_id}")
+                    return web.json_response({
+                        "tile_idx": task_idx,
+                        "estimated_remaining": remaining,
+                        "batched_static": job_data.batched_static,
+                    })
+                except asyncio.TimeoutError:
+                    if mode == 'dynamic':
+                        return web.json_response({"image_idx": None})
+                    return web.json_response({"tile_idx": None})
+            return await handle_api_error(request, "Job not found", 404)
+    except Exception as e:
+        return await handle_api_error(request, e, 500)
+
+
+@server.PromptServer.instance.routes.get("/distributed/job_status")
+async def job_status_endpoint(request):
+    """Endpoint to check if a job is ready."""
+    multi_job_id = request.query.get('multi_job_id')
+    if not multi_job_id:
+        return web.json_response({"ready": False})
+    prompt_server = ensure_tile_jobs_initialized()
+    async with prompt_server.distributed_tile_jobs_lock:
+        job_data = prompt_server.distributed_pending_tile_jobs.get(multi_job_id)
+        ready = bool(isinstance(job_data, BaseJobState) and job_data.queue is not None)
+        return web.json_response({"ready": ready})
@@ -0,0 +1,727 @@
+import json
+import asyncio
+import os
+import time
+import platform
+import subprocess
+import socket
+
+import torch
+import aiohttp
+from aiohttp import web
+import server
+
+from ..utils.config import load_config
+from ..utils.logging import debug_log, log
+from ..utils.network import (
+    build_worker_url,
+    get_client_session,
+    get_server_port,
+    handle_api_error,
+    normalize_host,
+    probe_worker,
+)
+from ..utils.constants import CHUNK_SIZE
+from ..workers import get_worker_manager
+from ..workers.ports import allocate_worker_ports
+from .schemas import require_fields, validate_worker_id
+from ..workers.detection import (
+    get_machine_id,
+    is_docker_environment,
+    is_runpod_environment,
+)
+try:
+    from ..utils.async_helpers import PromptValidationError, queue_prompt_payload
+except ImportError:
+    from ..utils.async_helpers import queue_prompt_payload
+
+    class PromptValidationError(RuntimeError):
+        def __init__(self, message, validation_error=None, node_errors=None):
+            super().__init__(str(message))
+            self.validation_error = validation_error if isinstance(validation_error, dict) else {}
+            self.node_errors = node_errors if isinstance(node_errors, dict) else {}
+
+
+@server.PromptServer.instance.routes.get("/distributed/worker_ws")
+async def worker_ws_endpoint(request):
+    """WebSocket endpoint for worker prompt dispatch."""
+    ws = web.WebSocketResponse(heartbeat=30)
+    await ws.prepare(request)
+
+    async for msg in ws:
+        if msg.type == aiohttp.WSMsgType.TEXT:
+            try:
+                data = json.loads(msg.data or "{}")
+            except json.JSONDecodeError:
+                await ws.send_json({
+                    "type": "dispatch_ack",
+                    "request_id": None,
+                    "ok": False,
+                    "error": "Invalid JSON payload.",
+                })
+                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",
+                    "request_id": data.get("request_id"),
+                    "ok": False,
+                    "error": "Unsupported websocket message type.",
+                })
+                continue
+
+            prompt = data.get("prompt")
+            if not isinstance(prompt, dict):
+                await ws.send_json({
+                    "type": "dispatch_ack",
+                    "request_id": data.get("request_id"),
+                    "ok": False,
+                    "error": "Field 'prompt' must be an object.",
+                })
+                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",
+                    "request_id": data.get("request_id"),
+                    "ok": True,
+                    "prompt_id": prompt_id,
+                })
+            except PromptValidationError as exc:
+                await ws.send_json({
+                    "type": "dispatch_ack",
+                    "request_id": data.get("request_id"),
+                    "ok": False,
+                    "error": str(exc),
+                    "validation_error": exc.validation_error,
+                    "node_errors": exc.node_errors,
+                })
+            except Exception as exc:
+                await ws.send_json({
+                    "type": "dispatch_ack",
+                    "request_id": data.get("request_id"),
+                    "ok": False,
+                    "error": str(exc),
+                })
+        elif msg.type == aiohttp.WSMsgType.ERROR:
+            log(f"[Distributed] Worker websocket error: {ws.exception()}")
+
+    return ws
+
+
+@server.PromptServer.instance.routes.post("/distributed/worker/clear_launching")
+async def clear_launching_state(request):
+    """Clear the launching flag when worker is confirmed running."""
+    try:
+        wm = get_worker_manager()
+        data = await request.json()
+        missing = require_fields(data, "worker_id")
+        if missing:
+            return await handle_api_error(request, f"Missing required field(s): {', '.join(missing)}", 400)
+
+        worker_id = str(data.get("worker_id")).strip()
+        config = load_config()
+        if not validate_worker_id(worker_id, config):
+            return await handle_api_error(request, f"Worker {worker_id} not found", 404)
+        
+        # Clear launching flag in managed processes
+        if worker_id in wm.processes:
+            if 'launching' in wm.processes[worker_id]:
+                del wm.processes[worker_id]['launching']
+                wm.save_processes()
+                debug_log(f"Cleared launching state for worker {worker_id}")
+        
+        return web.json_response({"status": "success"})
+    except Exception as e:
+        return await handle_api_error(request, e, 500)
+
+
+def get_network_ips():
+    """Get all network IPs, trying multiple methods."""
+    ips = []
+    hostname = socket.gethostname()
+
+    # Method 1: Try socket.getaddrinfo
+    try:
+        addr_info = socket.getaddrinfo(hostname, None)
+        for info in addr_info:
+            ip = info[4][0]
+            if ip and ip not in ips and not ip.startswith('::'):  # Skip IPv6 for now
+                ips.append(ip)
+    except (socket.gaierror, OSError):
+        pass
+
+    # Method 2: Try to connect to external server and get local IP
+    try:
+        s = socket.socket(socket.AF_INET, socket.SOCK_DGRAM)
+        s.connect(("8.8.8.8", 80))  # Google DNS
+        local_ip = s.getsockname()[0]
+        s.close()
+        if local_ip not in ips:
+            ips.append(local_ip)
+    except (OSError, socket.error):
+        pass
+
+    # Method 3: Platform-specific commands
+    try:
+        if platform.system() == "Windows":
+            # Windows ipconfig
+            result = subprocess.run(["ipconfig"], capture_output=True, text=True)
+            lines = result.stdout.split('\n')
+            for i, line in enumerate(lines):
+                if 'IPv4' in line and i + 1 < len(lines):
+                    ip = lines[i].split(':')[-1].strip()
+                    if ip and ip not in ips:
+                        ips.append(ip)
+        else:
+            # Unix/Linux/Mac ifconfig or ip addr
+            try:
+                result = subprocess.run(["ip", "addr"], capture_output=True, text=True)
+            except (FileNotFoundError, OSError):
+                try:
+                    result = subprocess.run(["ifconfig"], capture_output=True, text=True)
+                except (FileNotFoundError, OSError):
+                    result = None
+
+            import re
+            ip_pattern = re.compile(r'inet\s+(\d+\.\d+\.\d+\.\d+)')
+            if result is not None:
+                for match in ip_pattern.finditer(result.stdout):
+                    ip = match.group(1)
+                    if ip and ip not in ips:
+                        ips.append(ip)
+    except (OSError, subprocess.SubprocessError):
+        pass
+
+    return ips
+
+
+def get_recommended_ip(ips):
+    """Choose the best IP for master-worker communication."""
+    # Priority order:
+    # 1. Private network ranges (192.168.x.x, 10.x.x.x, 172.16-31.x.x)
+    # 2. Other non-localhost IPs
+    # 3. Localhost as last resort
+
+    private_ips = []
+    public_ips = []
+
+    for ip in ips:
+        if ip.startswith('127.') or ip == 'localhost':
+            continue
+        elif (ip.startswith('192.168.')
+                or ip.startswith('10.')
+                or (ip.startswith('172.') and 16 <= int(ip.split('.')[1]) <= 31)):
+            private_ips.append(ip)
+        else:
+            public_ips.append(ip)
+
+    # Prefer private IPs
+    if private_ips:
+        # Prefer 192.168 range as it's most common
+        for ip in private_ips:
+            if ip.startswith('192.168.'):
+                return ip
+        return private_ips[0]
+    elif public_ips:
+        return public_ips[0]
+    elif ips:
+        return ips[0]
+    else:
+        return None
+
+
+def _get_cuda_info():
+    """Detect CUDA device index and total physical GPU count.
+
+    Returns (cuda_device, cuda_device_count, physical_device_count).
+    All three are 0/None if CUDA is unavailable.
+    """
+    if not torch.cuda.is_available():
+        return None, 0, 0
+    try:
+        cuda_device_count = torch.cuda.device_count()
+        cuda_visible = os.environ.get('CUDA_VISIBLE_DEVICES', '')
+        if cuda_visible and cuda_visible.strip():
+            visible_devices = [int(d.strip()) for d in cuda_visible.split(',') if d.strip().isdigit()]
+            if visible_devices:
+                cuda_device = visible_devices[0]
+                try:
+                    result = subprocess.run(
+                        ['nvidia-smi', '--query-gpu=name', '--format=csv,noheader'],
+                        capture_output=True,
+                        text=True,
+                        timeout=5,
+                    )
+                    physical_device_count = (
+                        len(result.stdout.strip().split('\n'))
+                        if result.returncode == 0
+                        else max(visible_devices) + 1
+                    )
+                except (FileNotFoundError, OSError, subprocess.SubprocessError):
+                    physical_device_count = max(visible_devices) + 1
+                return cuda_device, cuda_device_count, physical_device_count
+            else:
+                return 0, cuda_device_count, cuda_device_count
+        else:
+            cuda_device = torch.cuda.current_device()
+            return cuda_device, cuda_device_count, cuda_device_count
+    except Exception as e:
+        debug_log(f"CUDA detection error: {e}")
+        return None, 0, 0
+
+
+def _collect_network_info_sync():
+    """Collect network/cuda info in a worker thread to avoid blocking route handlers."""
+    cuda_device, cuda_device_count, physical_device_count = _get_cuda_info()
+    device_count = physical_device_count if physical_device_count > 0 else cuda_device_count
+    master_port = get_server_port()
+    config = load_config()
+    worker_ports = []
+    if not config.get("settings", {}).get("has_auto_populated_workers") and not config.get("workers"):
+        worker_count = device_count - (1 if cuda_device is not None and 0 <= cuda_device < device_count else 0)
+        worker_ports = allocate_worker_ports(master_port, config.get("workers", []), worker_count)
+    hostname = socket.gethostname()
+    all_ips = get_network_ips()
+    recommended_ip = get_recommended_ip(all_ips)
+    return {
+        "hostname": hostname,
+        "all_ips": all_ips,
+        "recommended_ip": recommended_ip,
+        "cuda_device": cuda_device,
+        "cuda_device_count": device_count,
+        "master_port": master_port,
+        "local_worker_ports": worker_ports,
+    }
+
+
+def _read_worker_log_sync(log_file, lines_to_read):
+    """Read worker log content from disk in a threadpool worker."""
+    file_size = os.path.getsize(log_file)
+
+    with open(log_file, 'r', encoding='utf-8', errors='replace') as f:
+        if lines_to_read > 0 and file_size > 1024 * 1024:
+            # Read last N lines efficiently from end of file.
+            lines = []
+            f.seek(0, 2)
+            file_length = f.tell()
+            chunk_size = min(CHUNK_SIZE, file_length)
+
+            while len(lines) < lines_to_read and f.tell() > 0:
+                current_pos = max(0, f.tell() - chunk_size)
+                f.seek(current_pos)
+                chunk = f.read(chunk_size)
+                chunk_lines = chunk.splitlines()
+                if current_pos > 0:
+                    chunk_lines = chunk_lines[1:]
+                lines = chunk_lines + lines
+                f.seek(current_pos)
+
+            content = '\n'.join(lines[-lines_to_read:])
+            truncated = len(lines) > lines_to_read
+        else:
+            content = f.read()
+            truncated = False
+
+    return {
+        "content": content,
+        "file_size": file_size,
+        "truncated": truncated,
+        "lines_shown": lines_to_read if truncated else content.count('\n') + 1,
+    }
+
+
+def _parse_positive_int_query(value, default, minimum=1, maximum=10000):
+    """Parse bounded positive integer query params with sane fallback."""
+    try:
+        parsed = int(value)
+    except (TypeError, ValueError):
+        return default
+    parsed = max(minimum, parsed)
+    if maximum is not None:
+        parsed = min(maximum, parsed)
+    return parsed
+
+
+def _find_worker_by_id(config, worker_id):
+    worker_id_str = str(worker_id).strip()
+    for worker in config.get("workers", []):
+        if str(worker.get("id")).strip() == worker_id_str:
+            return worker
+    return None
+
+
+@server.PromptServer.instance.routes.get("/distributed/local_log")
+async def get_local_log_endpoint(request):
+    """Return this instance's in-memory ComfyUI log buffer."""
+    try:
+        from app.logger import get_logs
+    except Exception as e:
+        return await handle_api_error(request, f"Failed to import app.logger: {e}", 500)
+
+    try:
+        lines_to_read = _parse_positive_int_query(request.query.get("lines"), default=300, maximum=3000)
+        logs = get_logs()
+        if logs is None:
+            return web.json_response(
+                {
+                    "status": "success",
+                    "content": "",
+                    "entries": 0,
+                    "source": "memory",
+                    "truncated": False,
+                    "lines_shown": 0,
+                }
+            )
+
+        entries = list(logs)
+        selected_entries = entries[-lines_to_read:]
+        content = "".join(
+            entry.get("m", "") if isinstance(entry, dict) else str(entry)
+            for entry in selected_entries
+        )
+        lines_shown = content.count("\n") + (1 if content else 0)
+
+        return web.json_response(
+            {
+                "status": "success",
+                "content": content,
+                "entries": len(selected_entries),
+                "source": "memory",
+                "truncated": len(entries) > len(selected_entries),
+                "lines_shown": lines_shown,
+            }
+        )
+    except Exception as e:
+        return await handle_api_error(request, e, 500)
+
+
+@server.PromptServer.instance.routes.get("/distributed/network_info")
+async def get_network_info_endpoint(request):
+    """Get network interfaces and recommend best IP for master."""
+    try:
+        loop = asyncio.get_running_loop()
+        info = await loop.run_in_executor(None, _collect_network_info_sync)
+        
+        return web.json_response({
+            "status": "success",
+            **info,
+            "message": "Auto-detected network configuration"
+        })
+    except Exception as e:
+        return await handle_api_error(request, e, 500)
+
+@server.PromptServer.instance.routes.get("/distributed/system_info")
+async def get_system_info_endpoint(request):
+    """Get system information including machine ID for local worker detection."""
+    try:
+        import socket
+        
+        return web.json_response({
+            "status": "success",
+            "hostname": socket.gethostname(),
+            "machine_id": get_machine_id(),
+            "platform": {
+                "system": platform.system(),
+                "machine": platform.machine(),
+                "node": platform.node(),
+                "path_separator": os.sep,  # Add path separator
+                "os_name": os.name  # Add OS name (posix, nt, etc.)
+            },
+            "is_docker": is_docker_environment(),
+            "is_runpod": is_runpod_environment(),
+            "runpod_pod_id": os.environ.get('RUNPOD_POD_ID')
+        })
+    except Exception as e:
+        return await handle_api_error(request, e, 500)
+
+@server.PromptServer.instance.routes.post("/distributed/launch_worker")
+async def launch_worker_endpoint(request):
+    """Launch a worker process from the UI."""
+    try:
+        wm = get_worker_manager()
+        data = await request.json()
+        missing = require_fields(data, "worker_id")
+        if missing:
+            return await handle_api_error(request, f"Missing required field(s): {', '.join(missing)}", 400)
+
+        worker_id = str(data.get("worker_id")).strip()
+        
+        # Find worker config
+        config = load_config()
+        if not validate_worker_id(worker_id, config):
+            return await handle_api_error(request, f"Worker {worker_id} not found", 404)
+        worker = next((w for w in config.get("workers", []) if str(w.get("id")) == worker_id), None)
+        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
+        
+        # Check if already running (managed by this instance)
+        if worker_id_str in wm.processes:
+            proc_info = wm.processes[worker_id_str]
+            process = proc_info.get('process')
+            
+            # Check if still running
+            is_running = False
+            if process:
+                is_running = process.poll() is None
+            else:
+                # Restored process without subprocess object
+                is_running = wm._is_process_running(proc_info['pid'])
+            
+            if is_running:
+                return await handle_api_error(request, "Worker already running (managed by UI)", 409)
+            else:
+                # Process is dead, remove it
+                del wm.processes[worker_id_str]
+                wm.save_processes()
+        
+        # Launch the worker
+        try:
+            loop = asyncio.get_running_loop()
+            pid = await loop.run_in_executor(None, wm.launch_worker, worker)
+            log_file = wm.processes[worker_id_str].get('log_file')
+            return web.json_response({
+                "status": "success",
+                "pid": pid,
+                "message": f"Worker {worker['name']} launched",
+                "log_file": log_file
+            })
+        except ValueError as e:
+            return await handle_api_error(request, f"Failed to launch worker: {str(e)}", 400)
+        except Exception as e:
+            return await handle_api_error(request, f"Failed to launch worker: {str(e)}", 500)
+            
+    except Exception as e:
+        return await handle_api_error(request, e, 400)
+
+
+@server.PromptServer.instance.routes.post("/distributed/stop_worker")
+async def stop_worker_endpoint(request):
+    """Stop a worker process that was launched from the UI."""
+    try:
+        wm = get_worker_manager()
+        data = await request.json()
+        missing = require_fields(data, "worker_id")
+        if missing:
+            return await handle_api_error(request, f"Missing required field(s): {', '.join(missing)}", 400)
+
+        worker_id = str(data.get("worker_id")).strip()
+        config = load_config()
+        if not validate_worker_id(worker_id, config):
+            return await handle_api_error(request, f"Worker {worker_id} not found", 404)
+
+        success, message = wm.stop_worker(worker_id)
+        
+        if success:
+            return web.json_response({"status": "success", "message": message})
+        else:
+            return await handle_api_error(
+                request,
+                message,
+                404 if "not managed" in message else 409,
+            )
+            
+    except Exception as e:
+        return await handle_api_error(request, e, 400)
+
+
+@server.PromptServer.instance.routes.get("/distributed/managed_workers")
+async def get_managed_workers_endpoint(request):
+    """Get list of workers managed by this UI instance."""
+    try:
+        managed = get_worker_manager().get_managed_workers()
+        return web.json_response({
+            "status": "success",
+            "managed_workers": managed
+        })
+    except Exception as e:
+        return await handle_api_error(request, e, 500)
+
+
+@server.PromptServer.instance.routes.get("/distributed/local-worker-status")
+async def get_local_worker_status_endpoint(request):
+    """Check status of all local workers (localhost/no host specified)."""
+    try:
+        config = load_config()
+        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"]:
+                worker_id = worker["id"]
+                port = worker["port"]
+                
+                # Check if worker is enabled
+                if not worker.get("enabled", False):
+                    worker_statuses[worker_id] = {
+                        "online": False,
+                        "enabled": False,
+                        "processing": False,
+                        "queue_count": 0
+                    }
+                    continue
+                
+                # Try to connect to worker
+                try:
+                    worker_url = build_worker_url(worker)
+                    data = await probe_worker(worker_url, timeout=2.0)
+                    if data is None:
+                        worker_statuses[worker_id] = {
+                            "online": False,
+                            "enabled": True,
+                            "processing": False,
+                            "queue_count": 0,
+                            "error": "Unavailable",
+                        }
+                        continue
+                    queue_remaining = data.get("exec_info", {}).get("queue_remaining", 0)
+                    worker_statuses[worker_id] = {
+                        "online": True,
+                        "enabled": True,
+                        "processing": queue_remaining > 0,
+                        "queue_count": queue_remaining
+                    }
+                except asyncio.TimeoutError:
+                    worker_statuses[worker_id] = {
+                        "online": False,
+                        "enabled": True,
+                        "processing": False,
+                        "queue_count": 0,
+                        "error": "Timeout"
+                    }
+                except Exception as e:
+                    worker_statuses[worker_id] = {
+                        "online": False,
+                        "enabled": True,
+                        "processing": False,
+                        "queue_count": 0,
+                        "error": str(e)
+                    }
+        
+        return web.json_response({
+            "status": "success",
+            "worker_statuses": worker_statuses
+        })
+    except Exception as e:
+        debug_log(f"Error checking local worker status: {e}")
+        return await handle_api_error(request, e, 500)
+
+
+@server.PromptServer.instance.routes.get("/distributed/worker_log/{worker_id}")
+async def get_worker_log_endpoint(request):
+    """Get log content for a specific worker."""
+    try:
+        wm = get_worker_manager()
+        worker_id = request.match_info['worker_id']
+        
+        # Ensure worker_id is string
+        worker_id = str(worker_id)
+        
+        # Check if we manage this worker
+        if worker_id not in wm.processes:
+            return await handle_api_error(request, f"Worker {worker_id} not managed by UI", 404)
+        
+        proc_info = wm.processes[worker_id]
+        log_file = proc_info.get('log_file')
+        
+        if not log_file or not os.path.exists(log_file):
+            return await handle_api_error(request, "Log file not found", 404)
+        
+        # Read last N lines (or full file if small)
+        lines_to_read = _parse_positive_int_query(request.query.get('lines'), default=1000)
+        
+        try:
+            loop = asyncio.get_running_loop()
+            payload = await loop.run_in_executor(None, _read_worker_log_sync, log_file, lines_to_read)
+            
+            return web.json_response({
+                "status": "success",
+                "content": payload["content"],
+                "log_file": log_file,
+                "file_size": payload["file_size"],
+                "truncated": payload["truncated"],
+                "lines_shown": payload["lines_shown"],
+            })
+            
+        except Exception as e:
+            return await handle_api_error(request, f"Error reading log file: {str(e)}", 500)
+            
+    except Exception as e:
+        return await handle_api_error(request, e, 500)
+
+
+@server.PromptServer.instance.routes.get("/distributed/remote_worker_log/{worker_id}")
+async def get_remote_worker_log_endpoint(request):
+    """Proxy a remote worker log request to the worker's local in-memory log endpoint."""
+    try:
+        worker_id = str(request.match_info["worker_id"]).strip()
+        config = load_config()
+        worker = _find_worker_by_id(config, worker_id)
+        if not worker:
+            return await handle_api_error(request, f"Worker {worker_id} not found", 404)
+
+        # Remote log proxy is only meaningful for remote/cloud workers.
+        host = normalize_host(worker.get("host")) or ""
+        if not host:
+            return await handle_api_error(
+                request,
+                f"Worker {worker_id} is local; use /distributed/worker_log/{worker_id} instead.",
+                400,
+            )
+
+        lines_to_read = _parse_positive_int_query(request.query.get("lines"), default=300, maximum=3000)
+        worker_url = build_worker_url(worker, "/distributed/local_log")
+        session = await get_client_session()
+        async with session.get(
+            worker_url,
+            params={"lines": str(lines_to_read)},
+            timeout=aiohttp.ClientTimeout(total=5),
+        ) as resp:
+            if resp.status >= 400:
+                body = await resp.text()
+                return await handle_api_error(
+                    request,
+                    f"Remote worker {worker_id} returned HTTP {resp.status}: {body[:400]}",
+                    resp.status,
+                )
+
+            try:
+                data = await resp.json()
+            except Exception as e:
+                return await handle_api_error(
+                    request,
+                    f"Remote worker {worker_id} returned invalid JSON: {e}",
+                    502,
+                )
+
+        return web.json_response(data)
+    except Exception as e:
+        return await handle_api_error(request, e, 500)
@@ -0,0 +1,31 @@
+# conftest.py — project-level pytest configuration.
+#
+# Problem: custom_nodes/ComfyUI-Distributed/__init__.py uses relative imports
+# (from .nodes.v3 import ...) that fail when pytest tries to import it as a
+# standalone module during Package.setup() for the root package node.
+#
+# Fix: patch Package.setup() to skip the root-package's __init__.py import.
+# All actual package context is provided by each test module via
+# importlib.util.spec_from_file_location with synthetic stub packages.
+
+from _pytest.python import Package
+
+_orig_pkg_setup = Package.setup
+
+
+def _patched_pkg_setup(self) -> None:
+    # Skip the root package setup — its __init__.py uses relative imports
+    # that require a parent package (ComfyUI's plugin loader) which is not
+    # available in the test environment.
+    if self.path == self.config.rootpath:
+        return
+    _orig_pkg_setup(self)
+
+
+Package.setup = _patched_pkg_setup
+
+collect_ignore = [
+    "__init__.py",
+    "distributed.py",
+    "distributed_upscale.py",
+]
@@ -0,0 +1,289 @@
+# ComfyUI-Distributed API (Experimental)
+
+This document describes the **public HTTP API** added to ComfyUI-Distributed to allow queueing *distributed* workflows from external tools (scripts, services, CI jobs, render farms, etc.) without using the ComfyUI web UI.
+
+## Demo
+
+- Video walkthrough: https://youtu.be/yiQlPd0MzLk
+
+## Examples Repository
+
+- Examples repo: https://github.com/umanets/ComfyUI-Distributed-API-examples.git
+
+---
+
+## Overview
+
+### What this adds
+
+- `POST /distributed/queue` — queues a workflow using the same distributed orchestration rules as the UI:
+  - Detects distributed nodes in the prompt (`DistributedCollector`, `UltimateSDUpscaleDistributed`).
+  - Resolves enabled/selected workers.
+  - Probes and dispatches workers through `/distributed/worker_ws` by default.
+  - If `settings.websocket_orchestration=false`, probes with `GET /prompt` and dispatches with `POST /prompt` instead.
+  - Queues the master workflow in ComfyUI’s prompt queue.
+  - If any `DistributedCollector` has `load_balance=true`, selects one least-busy participant for this run.
+
+### What it does *not* add
+
+- Authentication/authorization.
+- A separate “job status” API for distributed results (you still use ComfyUI’s normal prompt history / websocket flow, and the existing `/distributed/queue_status/{job_id}` behavior for collector queues).
+
+---
+
+## Endpoint: `POST /distributed/queue`
+
+Queue a workflow for distributed execution.
+
+### URL
+
+- `http://<master-host>:<master-port>/distributed/queue`
+
+### Headers
+
+- `Content-Type: application/json`
+
+### Request Body
+
+```json
+{
+  "prompt": { "<node_id>": { "class_type": "...", "inputs": { } } },
+  "workflow": { },
+  "client_id": "external-client",
+  "delegate_master": false,
+  "enabled_worker_ids": ["1", "2"],
+  "workers": ["1", "2"],
+  "auto_prepare": true,
+  "trace_execution_id": "exec_1700000000_ab12cd"
+}
+```
+
+#### Fields
+
+- `prompt` (required unless `workflow.prompt` is present, object)
+  - A complete ComfyUI API-format prompt graph, using the same shape as `POST /prompt`.
+  - This is not the normal visual workflow export from the ComfyUI editor.
+- `workflow` (optional, object)
+  - Workflow metadata that ComfyUI normally stores in `extra_pnginfo.workflow`.
+  - If you don’t care about UI metadata, you can omit it.
+- `client_id` (required, string)
+  - Passed through as `extra_data.client_id` (useful if you consume ComfyUI websocket events).
+- `delegate_master` (optional, boolean)
+  - If `true`, attempts “workers-only” execution for workflows based on `DistributedCollector`.
+  - Current limitation: delegate-only mode **does not support** `UltimateSDUpscaleDistributed` and will fall back to running the full prompt on master.
+- `enabled_worker_ids` (required, array of strings)
+  - The explicit worker IDs to consider for this run.
+- `workers` (optional, array of strings or objects with `id`)
+  - Transitional alias for `enabled_worker_ids` used by older clients.
+- `auto_prepare` (optional, boolean)
+  - Kept for wire compatibility.
+  - Backend orchestration always runs with auto-prepare semantics.
+  - If top-level `prompt` is omitted, backend will attempt `workflow.prompt`.
+- `trace_execution_id` (optional, string)
+  - Passed through to orchestration logs.
+  - Server log lines include the marker as `[exec:<trace_execution_id>]`.
+
+##### How to get `enabled_worker_ids`
+
+Worker IDs come from the plugin config (`GET /distributed/config`) under `workers[].id`.
+
+Example (bash + `jq`):
+
+```bash
+curl -s "http://127.0.0.1:8188/distributed/config" \
+  | jq -r '.workers[] | "id=\(.id)\tname=\(.name)\tenabled=\(.enabled)\thost=\(.host)\tport=\(.port)\ttype=\(.type)"'
+```
+
+Example (PowerShell):
+
+```powershell
+$cfg = Invoke-RestMethod "http://127.0.0.1:8188/distributed/config"
+$cfg.workers | Select-Object id,name,enabled,host,port,type | Format-Table -AutoSize
+```
+
+### Response Body
+
+```json
+{
+  "prompt_id": "<uuid>",
+  "worker_count": 2,
+  "auto_prepare_supported": true
+}
+```
+
+- `prompt_id` — the master prompt id queued into ComfyUI.
+- `worker_count` — number of workers that received a dispatched prompt (only those that passed the health check).
+
+### Status Codes
+
+- `200` — queued.
+- `400` — invalid JSON or invalid body.
+- `500` — orchestration/dispatch failure (see server logs for details).
+
+---
+
+## Worker requirements (important)
+
+For a worker to participate, it must be reachable from the master. By default:
+
+- WebSocket probe and dispatch: `<worker-base>/distributed/worker_ws` must accept the connection.
+
+If `settings.websocket_orchestration=false`:
+
+- Health check: `GET <worker-base>/prompt` must return HTTP 200.
+- Dispatch: `POST <worker-base>/prompt` must accept the prompt.
+
+Also, for collector-based flows:
+
+- Workers will send results back to the master via `POST /distributed/job_complete` (that route must be reachable from workers).
+
+---
+
+## Endpoint: `POST /distributed/job_complete`
+
+Submit one completed worker image back to the master collector queue.
+
+### URL
+
+- `http://<master-host>:<master-port>/distributed/job_complete`
+
+### Request Body
+
+```json
+{
+  "job_id": "exec_1234567890_17",
+  "worker_id": "worker-1",
+  "batch_idx": 0,
+  "image": "data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAA...",
+  "is_last": false
+}
+```
+
+### Canonical envelope (required fields)
+
+- `job_id` (string, required)
+- `worker_id` (string, required)
+- `batch_idx` (integer >= 0, required)
+- `image` (string, required)
+  - 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.
+
+### CORS note
+
+If you call the API from a browser (not from a backend), ensure the master ComfyUI is started with `--enable-cors-header`.
+
+---
+
+## Log Endpoints
+
+### `GET /distributed/worker_log/{worker_id}`
+
+Read log files for workers launched locally by the master UI process manager.
+
+- Intended for managed local workers.
+- Query param: `lines` (optional, default `1000`).
+
+### `GET /distributed/local_log`
+
+Read this ComfyUI instance's in-memory runtime log buffer.
+
+- Available on any ComfyUI-Distributed instance (master or worker).
+- Query param: `lines` (optional, default `300`, max `3000`).
+
+### `GET /distributed/remote_worker_log/{worker_id}`
+
+Proxy endpoint on master that fetches logs from a configured remote/cloud worker's
+`/distributed/local_log`.
+
+- Intended for remote/cloud workers in master config.
+- Query param: `lines` (optional, default `300`, max `3000`).
+
+---
+
+## Examples
+
+### 1) Minimal `curl` request envelope
+
+```bash
+curl -X POST "http://127.0.0.1:8188/distributed/queue" \
+  -H "Content-Type: application/json" \
+  -d @payload.json
+```
+
+`payload.json` must contain a complete ComfyUI API-format prompt. The abbreviated envelope below illustrates the request shape but is not directly executable:
+
+```json
+{
+  "prompt": {
+    "<node_id>": {
+      "class_type": "<node class>",
+      "inputs": {"<required_input>": "<value or connection>"}
+    }
+  },
+  "enabled_worker_ids": [],
+  "client_id": "external-client"
+}
+```
+
+Export or construct a valid API-format prompt with all required node inputs and at least one output node before submitting it.
+
+### 2) Python (`requests`)
+
+```python
+import requests
+
+url = "http://127.0.0.1:8188/distributed/queue"
+payload = {
+    "prompt": {...},
+    "workflow": {...},
+    "client_id": "external-client",
+    "delegate_master": False,
+    "enabled_worker_ids": ["1", "2"],
+}
+
+r = requests.post(url, json=payload, timeout=60)
+r.raise_for_status()
+print(r.json())
+```
+
+### 3) JavaScript (`fetch`)
+
+```js
+const url = "http://127.0.0.1:8188/distributed/queue";
+
+const payload = {
+  prompt: {/* ... */},
+  workflow: {/* ... */},
+  client_id: "external-client",
+  delegate_master: false,
+  enabled_worker_ids: ["1", "2"],
+};
+
+const resp = await fetch(url, {
+  method: "POST",
+  headers: { "Content-Type": "application/json" },
+  body: JSON.stringify(payload),
+});
+
+if (!resp.ok) throw new Error(await resp.text());
+console.log(await resp.json());
+```
+
+---
+
+## Operational notes / gotchas
+
+- If the workflow contains **no distributed nodes**, the endpoint falls back to normal master queueing and returns `worker_count: 0`.
+- Worker selection is “best-effort”: offline workers are skipped.
+- For public URLs/tunnels: prefer configuring `master.host` with an explicit scheme (`https://...`) to avoid ambiguity.
+
+---
+
+## Changelog (this feature)
+
+- Added `POST /distributed/queue` endpoint.
+- Added orchestration module used by the endpoint.
@@ -0,0 +1,38 @@
+# Compressed 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).
+
+| `transfer_format` | Type | Notes |
+|---|---|---|
+| `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. |
+
+`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).
+
+Measured on a 121-frame 1280×720 batch (one worker's result):
+
+| Format | Sent | vs legacy | PSNR |
+|---|---|---|---|
+| `legacy_png` | 446 MB | 1× | lossless |
+| `png` | 163 MB | 2.7× | lossless |
+| `ffv1` | 128 MB | 3.5× | lossless |
+| `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 |
+| `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)`.
@@ -0,0 +1,101 @@
+# 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`. It returns `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 outputs contain `None` entries.
+
+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.
+
+## 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.
+- Worker eligibility is determined before execution and checked again before dispatch. Busy or transiently unreachable workers are excluded; healthy candidates can continue. Malformed capability/queue responses and incompatible versions fail explicitly. Workers joining later are not added to an already-running region.
+- 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.
@@ -0,0 +1,158 @@
+# Typed serverless media (PR2)
+
+`comfy-media-v1` adds typed list results, exact by default, for `IMAGE`, `AUDIO`, `MASK`,
+`STRING`, and native ComfyUI `VIDEO`. It is an **explicit endpoint profile**, not
+an automatically detected capability. Omitting `media_profile` still selects
+`legacy-image-v1`; existing image collectors and handlers keep their old contract.
+See [legacy serverless setup](serverless-workers.md) for credentials and recovery.
+
+## Use a typed collector
+
+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 Media Collector** (`DistributedMediaCollector`), connect at
+   least one typed input, and connect its outputs to the appropriate downstream
+   nodes. For literal text, connect a `PrimitiveString` node to `texts`.
+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 has five optional inputs and five **list outputs**, in this order:
+
+| Socket / output index | Native type | Each list item |
+| --- | --- | --- |
+| `images` / 0 | `IMAGE` | One BHWC tensor, including its batch dimension |
+| `audio` / 1 | `AUDIO` | `{ "waveform": BCT_tensor, "sample_rate": positive_int }` |
+| `masks` / 2 | `MASK` | One BHW tensor |
+| `texts` / 3 | `STRING` | Exact text, including whitespace or an empty string |
+| `video` / 4 | `VIDEO` | One supported native video object |
+
+For each connected socket, order is **master first, then the prepared participant
+order, then that participant's item order**—never completion order. Delegate-only
+mode omits the master. An unconnected socket returns `[]`. A connected remote
+socket must return 1–64 items; an empty string is an item, an empty list is not.
+Lists are not concatenated into one tensor, resampled, resized, or aligned across
+sockets. Different items may have different valid shapes or audio sample rates.
+The entire remote response is validated and decoded before any of its results
+are delivered; missing, duplicate, mismatched, or corrupt results fail collection.
+
+## Image return format and quality
+
+`transfer_format` defaults to **`exact`** and `transfer_quality` to **90**.
+These controls affect **remote IMAGE returns only**, not master-local images,
+uploaded inputs, MASK, AUDIO, STRING or native VIDEO values.
+
+- `exact`: preserve floating-point dtype and values, with no image quantization.
+- `legacy_png`, `png`, `ffv1`: lossless **8-bit image pixels**, not exact float
+  tensors. Image values are clipped to [0,1] and quantized as in the existing
+  Collector. `legacy_png` uses PNG without compression; `png` compresses it.
+- `jpeg`, `webp`: lossy still-image compression.
+- `h264`, `h265`, `av1`: lossy video encoding of each IMAGE tensor's frame batch.
+  This does not transcode the native `video` socket or combine separate list items.
+- 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; 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). The original Collector's
+settings, default and fallback behavior are unchanged.
+
+## 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`).
+
+### Native VIDEO is not 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.
+- Remote typed transport is **serverless only**. Direct-worker typed transport
+  is not implemented. Do not mix a typed collector with legacy image collectors,
+  unsupported distributed/batch nodes, or collector-to-collector dependencies.
+  Downstream processing stays on the master. Image-list scheduling is separate
+  PR3 work, not part of this contract.
+
+## 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 and its offline tests; they do not claim
+live Runpod/provider or GPU end-to-end validation.
@@ -0,0 +1,173 @@
+# Runpod serverless workers (preview)
+
+A **Serverless (Runpod)** worker is a queue-based endpoint, not a permanent
+ComfyUI machine. Runpod chooses the actual GPU workers. Distributed submits one
+normal ComfyUI API workflow per requested image and collects the returned images.
+The endpoint workers do not need ComfyUI-Distributed 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. Use an image **Distributed Collector** and queue through Distributed as usual.
+   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.
+
+## First-version limits
+
+- Image collectors only. Audio/video transfer and audio collectors are rejected.
+- 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.
+
+This feature is verified with offline provider/transport tests. A successful
+health test is not a paid end-to-end generation test; test your deployed handler
+and model inventory with one image per run before increasing it.
@@ -0,0 +1,53 @@
+# Native V3 node API
+
+The package registers one `ComfyExtension` through `comfy_entrypoint()` and
+imports the versioned `comfy_api.v0_0_2` API. Use a ComfyUI version that provides
+that API; there is no V1 registration fallback.
+
+## Compatibility
+
+- All eight node IDs, display names, categories, visible input order, defaults
+  and output order are retained. Existing execution algorithms remain in the
+  private collector, utilities and upscale modules.
+- The image/audio dividers explicitly declare ten typed outputs. Their existing
+  frontend extensions still show only the selected number of outputs. This
+  replaces the V1 `ByPassTypeTuple` indexing workaround without changing saved
+  IMAGE/AUDIO socket indices or the ten returned values.
+- Standard hidden context uses `cls.hidden`. Worker/orchestration metadata keeps
+  its existing prompt-input names and Python defaults via `accept_all_inputs`;
+  it does not become visible widgets.
+- The collector retains list-input handling. Upscale retains its always-changing
+  fingerprint and creates a private runtime helper for each V3 execution, so
+  mutable helper state is not attached to sanitized V3 class clones.
+- Routes, distributed state and the existing worker startup/shutdown hooks are
+  initialized by `runtime/bootstrap.py` during `ComfyExtension.on_load()`.
+  Worker mode still suppresses automatic worker launch. `distributed.py` is no
+  longer an entrypoint.
+
+## Verification
+
+Run ordinary unit tests from this repository:
+
+```bash
+python -m pytest tests -q -o addopts=
+```
+
+Opt into real-framework acceptance using a ComfyUI checkout and its interpreter:
+
+```bash
+COMFYUI_SOURCE_ROOT=/path/to/ComfyUI \
+  /path/to/ComfyUI/.venv/bin/python -m pytest tests -q -o addopts=
+```
+
+The acceptance subprocess uses CPU mode and the real ComfyUI loader, input
+parser, V3 class preparation, prompt validation and `PromptExecutor`. It compares
+all node schemas with the V1 fixture, checks all five bundled workflows' node
+IDs and socket/link contracts, exercises injected worker metadata, image/audio
+lists, collector aggregation and divider outputs, rejects invalid upscale enums,
+and decodes an actual preview PNG referenced by executor history.
+
+The upscale GPU/model boundary is mocked to check argument forwarding and
+per-execution helper isolation. Full checkpoint inference, browser canvas
+acceptance and multi-host HTTP transport are not exercised. No HTTP listener,
+workers or model downloads are started; preview files use a temporary scratch
+directory. The test does not install this branch into a live custom-node folder.
@@ -15,5 +15,6 @@
 7. In ComfyUI, open the GPU panel on the left.
 > If you set SAGE_ATTENTION to true, add "--use-sage-attention" to Extra Args on the workers.
 8. Launch the workers.
-9. Upload video, add prompt and run workflow.
-10. Right-click the Video Combine node and click Save Preview to save the video.
+9. [Load the workflow.](https://github.com/robertvoy/ComfyUI-Distributed/blob/main/workflows/distributed-upscale-video.json)
+10. Upload video, add prompt and run workflow.
+11. Right-click the Video Combine node and click Save Preview to save the video.
@@ -6,6 +6,14 @@

 <img width="600" src="https://github.com/user-attachments/assets/609c42aa-8a1c-4a3f-939e-f3552fa1d54f" />

+### Master participation modes
+
+The master can either contribute GPU work or stay in **orchestrator-only** mode:
+
+- **Participating**: Master renders alongside workers, useful when you want every available GPU.
+- **Orchestrator-only**: Master sends jobs to selected workers but skips local rendering. Enable this by opening the Distributed panel and unchecking the master toggle. The master card will display *“Master disabled: running as orchestrator only.”*
+- **Fallback**: If orchestrator-only is enabled but no workers remain selected, the master automatically re-enables execution to guarantee the workflow still runs. The UI shows a green *“Master fallback execution active”* badge so you know work is executing locally again.
+
 ### Types of Workers

 - **Local workers**: Additional GPUs on the same machine as the master
@@ -66,12 +74,14 @@

 > ComfyUI instances running on a cloud service like Runpod. 

+For faster result transfers and lower bandwidth usage, see [Compressed Result Transfer](compressed-result-transfer.md).
+
 ### Deploy Cloud Worker on Runpod

 📺 [Watch Tutorial](https://www.youtube.com/watch?v=wxKKWMQhYTk)

 **On Runpod:**
-> If using your own template, make sure you launch ComfyUI with the `--enable-cors-header` argument and you `git clone ComfyUI-Distributed` into custom_nodes. ⚠️ **Required!**
+> If using your own template, launch ComfyUI with `--listen --enable-cors-header` and clone `ComfyUI-Distributed` into `custom_nodes`. ⚠️ **Required!**

 1. Register a [Runpod](https://get.runpod.io/0bw29uf3ug0p) account.
 2. On Runpod, go to Storage > New Network Volume and create a volume that will store the models you need. Start with 40 GB, you can always add more later. Learn more [about Network Volumes](https://docs.runpod.io/pods/storage/create-network-volumes).
@@ -84,7 +94,7 @@
 	- SAGE_ATTENTION: optional optimisation (set to true/false)
 5. Deploy your pod.
 6. Connect to your pod using JupyterLabs. This gives us access to the pod's file system.
-7. Download models into /workspaces/ComfyUI/models/ (these will remain on your network drive even after you terminate the pod). Example commands below:
+7. Download models into `/workspace/ComfyUI/models/` (these will remain on your network drive even after you terminate the pod). Example commands below:
 ```
 # Download from CivitAI
 comfy model download --url https://civitai.com/api/download/models/1759168 --relative-path /workspace/ComfyUI/models/checkpoints --set-civitai-api-token $CIVITAI_API_TOKEN
@@ -0,0 +1 @@
+"""Dynamic per-image image-list execution internals."""
@@ -0,0 +1,70 @@
+"""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,
+    _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, ("images", "masks", "texts"))
+        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
@@ -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)
@@ -0,0 +1,184 @@
+"""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()
+
+
+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._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):
+        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)
+            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:
+            raise ImageListCancelledError("Local image-list execution cancelled") from exc
+        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 = execution.HierarchicalCache(execution.CacheKeySetInputSignature)
+                    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)
@@ -0,0 +1 @@
+"""Private execution helpers; public registration lives in nodes.v3."""
@@ -0,0 +1,776 @@
+import torch
+import io
+import json
+import asyncio
+import time
+import base64
+
+import aiohttp
+import server as _server
+import comfy.model_management
+from comfy.utils import ProgressBar
+
+from ..utils.logging import debug_log, log
+from ..utils.config import get_worker_timeout_seconds, load_config, is_master_delegate_only
+from ..utils.constants import HEARTBEAT_INTERVAL, TRANSFER_MAX_CHUNK_BYTES, TRANSFER_MAX_REQUEST_BYTES
+from ..utils.image import tensor_to_pil, pil_to_tensor, ensure_contiguous
+from ..utils.network import build_worker_url, get_client_session, probe_worker
+from ..utils.audio_payload import encode_audio_payload
+from ..utils.async_helpers import run_async_in_server_loop
+from ..utils import transfer_codec
+
+prompt_server = _server.PromptServer.instance
+
+
+def _encode_all_chunks(frames, fmt, quality, max_chunk_bytes, max_frame_bytes):
+    """Materialize every transfer chunk (runs in an executor)."""
+    return list(transfer_codec.encode_chunks(frames, fmt, quality, max_chunk_bytes, max_frame_bytes))
+
+
+class DistributedCollectorNode:
+    INPUT_IS_LIST = True
+    EMPTY_AUDIO = {"waveform": torch.zeros(1, 2, 1), "sample_rate": 44100}
+
+    @classmethod
+    def INPUT_TYPES(s):
+        return {
+            "required": {
+                "load_balance": (
+                    "BOOLEAN",
+                    {
+                        "default": False,
+                        "tooltip": "Run this workflow on one least-busy participant (master included when participating).",
+                    },
+                ),
+            },
+            "optional": {
+                "images": ("IMAGE",),
+                "audio": ("AUDIO",),
+                "transfer_format": (
+                    list(transfer_codec.TRANSFER_FORMATS),
+                    {
+                        "default": transfer_codec.DEFAULT_TRANSFER_FORMAT,
+                        "tooltip": transfer_codec.TRANSFER_FORMAT_TOOLTIP,
+                    },
+                ),
+                "transfer_quality": (
+                    "INT",
+                    {
+                        "default": transfer_codec.DEFAULT_TRANSFER_QUALITY,
+                        "min": 1,
+                        "max": 100,
+                        "step": 1,
+                        "tooltip": transfer_codec.TRANSFER_QUALITY_TOOLTIP,
+                    },
+                ),
+            },
+            "hidden": {
+                "multi_job_id": ("STRING", {"default": ""}),
+                "is_worker": ("BOOLEAN", {"default": False}),
+                "master_url": ("STRING", {"default": ""}),
+                "enabled_worker_ids": ("STRING", {"default": "[]"}),
+                "worker_batch_size": ("INT", {"default": 1, "min": 1, "max": 1024}),
+                "worker_id": ("STRING", {"default": ""}),
+                "pass_through": ("BOOLEAN", {"default": False}),
+                "delegate_only": ("BOOLEAN", {"default": False}),
+            },
+        }
+
+    RETURN_TYPES = ("IMAGE", "AUDIO")
+    RETURN_NAMES = ("images", "audio")
+    FUNCTION = "run"
+    CATEGORY = "image"
+    
+    @staticmethod
+    def _unwrap_list_input(value):
+        """Unwrap scalar inputs when ComfyUI passes them via INPUT_IS_LIST."""
+        if isinstance(value, (list, tuple)) and len(value) == 1:
+            return value[0]
+        return value
+
+    def _normalize_images_input(self, images):
+        """Collapse ComfyUI list IMAGE inputs into a normal batched IMAGE tensor."""
+        if isinstance(images, (list, tuple)):
+            if not images:
+                raise ValueError("Collector received an empty image list")
+            if not all(isinstance(image, torch.Tensor) for image in images):
+                raise TypeError("Collector expected IMAGE list items to be torch.Tensor instances")
+            if len(images) == 1:
+                return ensure_contiguous(images[0])
+            return ensure_contiguous(torch.cat([ensure_contiguous(image) for image in images], dim=0))
+        return ensure_contiguous(images)
+
+    def _normalize_audio_input(self, audio):
+        """Collapse ComfyUI list AUDIO inputs into a single AUDIO payload when present."""
+        if not isinstance(audio, (list, tuple)):
+            return audio
+
+        audio_items = [item for item in audio if item is not None]
+        if not audio_items:
+            return None
+        if len(audio_items) == 1:
+            return audio_items[0]
+
+        waveforms = []
+        sample_rate = 44100
+        for item in audio_items:
+            if not isinstance(item, dict):
+                raise TypeError("Collector expected AUDIO list items to be dictionaries")
+            waveform = item.get("waveform")
+            if waveform is None or waveform.numel() == 0:
+                continue
+            waveforms.append(waveform)
+            if sample_rate == 44100:
+                sample_rate = item.get("sample_rate", 44100)
+
+        if not waveforms:
+            return None
+        return {"waveform": torch.cat(waveforms, dim=-1), "sample_rate": sample_rate}
+
+    def run(self, images=None, load_balance=False, audio=None, multi_job_id="", is_worker=False, master_url="", enabled_worker_ids="[]", worker_batch_size=1, worker_id="", pass_through=False, delegate_only=False, transfer_format=transfer_codec.DEFAULT_TRANSFER_FORMAT, transfer_quality=transfer_codec.DEFAULT_TRANSFER_QUALITY):
+        if images is not None:
+            images = self._normalize_images_input(images)
+        audio = self._normalize_audio_input(audio)
+        load_balance = self._unwrap_list_input(load_balance)
+        multi_job_id = self._unwrap_list_input(multi_job_id)
+        is_worker = self._unwrap_list_input(is_worker)
+        master_url = self._unwrap_list_input(master_url)
+        enabled_worker_ids = self._unwrap_list_input(enabled_worker_ids)
+        worker_batch_size = self._unwrap_list_input(worker_batch_size)
+        worker_id = self._unwrap_list_input(worker_id)
+        pass_through = self._unwrap_list_input(pass_through)
+        delegate_only = self._unwrap_list_input(delegate_only)
+        transfer_format = self._unwrap_list_input(transfer_format)
+        transfer_quality = self._unwrap_list_input(transfer_quality)
+
+        remote_only_master = (
+            bool(multi_job_id)
+            and not is_worker
+            and (delegate_only or is_master_delegate_only())
+        )
+        if images is None and audio is None and not remote_only_master:
+            raise ValueError("DistributedCollector requires at least one image or audio input")
+
+        # Create empty audio if not provided
+        empty_audio = {"waveform": torch.zeros(1, 2, 1), "sample_rate": 44100}
+
+        if not multi_job_id or pass_through:
+            if pass_through:
+                debug_log("Collector: pass-through mode enabled, returning images unchanged")
+            return (images, audio if audio is not None else empty_audio)
+
+        # Use async helper to run in server loop
+        result = run_async_in_server_loop(
+            self.execute(
+                images,
+                audio,
+                load_balance,
+                multi_job_id,
+                is_worker,
+                master_url,
+                enabled_worker_ids,
+                worker_batch_size,
+                worker_id,
+                delegate_only,
+                transfer_format,
+                transfer_quality,
+            )
+        )
+        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,
+        audio,
+        multi_job_id,
+        master_url,
+        worker_id,
+        transfer_format=transfer_codec.DEFAULT_TRANSFER_FORMAT,
+        transfer_quality=transfer_codec.DEFAULT_TRANSFER_QUALITY,
+    ):
+        """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.LEGACY_FORMAT:
+            sent = await self._send_compressed_batch(
+                image_batch, audio, multi_job_id, master_url, worker_id, fmt, transfer_quality
+            )
+            if sent:
+                return
+            log(
+                "Worker - Master did not accept compressed transfer "
+                "(older ComfyUI-Distributed on the master?). Falling back to legacy_png."
+            )
+        await self._send_legacy_batch(image_batch, audio, multi_job_id, master_url, worker_id)
+
+    @staticmethod
+    def _resolve_transfer_format(transfer_format):
+        try:
+            fmt = transfer_codec.normalize_format(transfer_format)
+        except transfer_codec.TransferCodecError as exc:
+            log(f"Worker - {exc}; using legacy_png.")
+            return transfer_codec.LEGACY_FORMAT
+        if fmt not in transfer_codec.VIDEO_FORMATS or transfer_codec.video_codec_available(fmt):
+            return fmt
+        fallback = "h264" if fmt != "ffv1" and transfer_codec.video_codec_available("h264") else "png"
+        log(f"Worker - Encoder for '{fmt}' is not available in this PyAV build; using '{fallback}'.")
+        return fallback
+
+    @staticmethod
+    def _build_multipart(envelope, parts):
+        form = aiohttp.FormData()
+        form.add_field("envelope", json.dumps(envelope), content_type="application/json")
+        for i, data in enumerate(parts):
+            form.add_field("media", data, filename=f"{i}.bin", content_type="application/octet-stream")
+        return form
+
+    # Room left in each request for the multipart envelope (ids, frame
+    # metadata, part headers). Audio is budgeted separately.
+    TRANSFER_ENVELOPE_RESERVE_BYTES = 256 * 1024
+    # Body of a 400 from a master without compressed-transfer support: it
+    # tries request.json() on the multipart body and fails.
+    _OLD_MASTER_REJECTION = "Invalid JSON payload"
+
+    async def _send_compressed_batch(self, image_batch, audio, multi_job_id, master_url, worker_id, fmt, quality):
+        """Send the batch as compressed multipart chunks.
+
+        All chunks are encoded before anything is sent, so a frame that cannot
+        fit in one request fails the transfer without delivering a partial
+        batch. Every request stays within TRANSFER_MAX_REQUEST_BYTES.
+
+        Returns False (without having delivered anything) when the master
+        rejects the first chunk because it predates compressed transfer, so
+        the caller can retry with legacy_png. Any other rejection raises.
+        """
+        encoded_audio = encode_audio_payload(audio)
+        session = await get_client_session()
+        url = f"{master_url}/distributed/job_complete"
+        loop = asyncio.get_running_loop()
+
+        frames = await loop.run_in_executor(None, transfer_codec.tensor_batch_to_uint8, image_batch)
+        raw_bytes = int(frames.nbytes)
+        total = int(frames.shape[0])
+        media_limit = max(1, TRANSFER_MAX_REQUEST_BYTES - self.TRANSFER_ENVELOPE_RESERVE_BYTES)
+        started = time.time()
+        # Encoding is CPU-bound (seconds for video); keep the server loop responsive.
+        chunks = await loop.run_in_executor(
+            None, _encode_all_chunks, frames, fmt, quality, min(TRANSFER_MAX_CHUNK_BYTES, media_limit), media_limit,
+        )
+        del frames
+
+        # Attach audio to the final chunk only if the request stays in bounds;
+        # otherwise finish with a separate audio-only completion.
+        audio_bytes = len(json.dumps(encoded_audio)) if encoded_audio is not None else 0
+        last_media = sum(len(p) for p in chunks[-1]["parts"])
+        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
+
+        sent_bytes = 0
+        for index, chunk in enumerate(chunks):
+            parts = chunk.pop("parts")
+            is_final_chunk = index == len(chunks) - 1
+            envelope = {
+                "job_id": str(multi_job_id),
+                "worker_id": str(worker_id),
+                "is_last": bool(is_final_chunk and not audio_separate),
+                **chunk,
+            }
+            if is_final_chunk and audio_inline:
+                envelope["audio"] = encoded_audio
+            chunk_bytes = sum(len(p) for p in parts)
+            timeout_seconds = min(600, 60 + chunk_bytes // (256 * 1024))
+            try:
+                async with session.post(
+                    url,
+                    data=self._build_multipart(envelope, parts),
+                    timeout=aiohttp.ClientTimeout(total=timeout_seconds),
+                ) as response:
+                    if response.status >= 400:
+                        body = await response.text()
+                        if index == 0 and response.status == 400 and self._OLD_MASTER_REJECTION in body:
+                            debug_log(f"Worker - Master rejected compressed chunk: {body[:300]}")
+                            return False
+                        raise RuntimeError(
+                            f"Master rejected {fmt} chunk {index + 1}/{len(chunks)} "
+                            f"({chunk_bytes} bytes) with HTTP {response.status}: {body[:300]}. "
+                            "If this is a size limit, make COMFYUI_DISTRIBUTED_TRANSFER_MAX_REQUEST_BYTES "
+                            "on the worker smaller than the master's --max-upload-size."
+                        )
+            except Exception as e:
+                log(f"Worker - Failed to send {fmt} chunk to master: {e}")
+                debug_log(f"Worker - Full error details: URL={url}")
+                raise
+            sent_bytes += chunk_bytes
+
+        requests = len(chunks)
+        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 {fmt} in {requests} request(s): "
+            f"{sent_bytes / 1e6:.2f} MB (raw {raw_bytes / 1e6:.2f} MB, "
+            f"{transfer_codec.describe_ratio(sent_bytes, raw_bytes)} smaller) in {time.time() - started:.1f}s"
+        )
+        return True
+
+    @staticmethod
+    async def _post_json_completion(session, url, payload):
+        timeout_seconds = 60 if "image" in payload else 600
+        try:
+            async with session.post(
+                url,
+                json=payload,
+                timeout=aiohttp.ClientTimeout(total=timeout_seconds),
+            ) as response:
+                response.raise_for_status()
+        except Exception as e:
+            media_type = "image/audio" if "image" in payload else "audio-only"
+            log(f"Worker - Failed to send canonical {media_type} envelope to master: {e}")
+            debug_log(f"Worker - Full error details: URL={url}")
+            raise
+
+    async def _send_legacy_batch(self, image_batch, audio, multi_job_id, master_url, worker_id):
+        """Original transfer: one JSON request per frame with a base64 PNG."""
+        encoded_audio = encode_audio_payload(audio)
+        session = await get_client_session()
+        url = f"{master_url}/distributed/job_complete"
+
+        payloads = []
+        batch_size = 0 if image_batch is None else image_batch.shape[0]
+        if batch_size == 0:
+            if encoded_audio is None:
+                raise ValueError("Worker completion requires image or audio data")
+            payloads.append(
+                {
+                    "job_id": str(multi_job_id),
+                    "worker_id": str(worker_id),
+                    "batch_idx": 0,
+                    "audio": encoded_audio,
+                    "is_last": True,
+                }
+            )
+        else:
+            for batch_idx in range(batch_size):
+                img = tensor_to_pil(image_batch[batch_idx:batch_idx+1], 0)
+                byte_io = io.BytesIO()
+                img.save(byte_io, format='PNG', compress_level=0)
+                encoded_image = base64.b64encode(byte_io.getvalue()).decode('utf-8')
+                payload = {
+                    "job_id": str(multi_job_id),
+                    "worker_id": str(worker_id),
+                    "batch_idx": int(batch_idx),
+                    "image": f"data:image/png;base64,{encoded_image}",
+                    "is_last": bool(batch_idx == batch_size - 1),
+                }
+                if payload["is_last"] and encoded_audio is not None:
+                    payload["audio"] = encoded_audio
+                payloads.append(payload)
+
+        for payload in payloads:
+            await self._post_json_completion(session, url, payload)
+
+    def _combine_audio(self, master_audio, worker_audio, empty_audio, worker_order=None):
+        """Combine audio from master and workers into a single audio output.
+
+        Ordering: master first, then workers in `worker_order` (if provided),
+        then any unexpected worker ids in sorted order.
+        """
+        audio_pieces = []
+        sample_rate = 44100
+
+        # Add master audio first if present
+        if master_audio is not None:
+            waveform = master_audio.get("waveform")
+            if waveform is not None and waveform.numel() > 0:
+                audio_pieces.append(waveform)
+                sample_rate = master_audio.get("sample_rate", 44100)
+
+        # Add worker audio in configured enabled-worker order first.
+        ordered_worker_ids = [str(worker_id) for worker_id in (worker_order or [])]
+        seen = set()
+        for worker_id_str in ordered_worker_ids:
+            seen.add(worker_id_str)
+            w_audio = worker_audio.get(worker_id_str)
+            if w_audio is not None:
+                waveform = w_audio.get("waveform")
+                if waveform is not None and waveform.numel() > 0:
+                    audio_pieces.append(waveform)
+                    # Use first available sample rate
+                    if sample_rate == 44100:
+                        sample_rate = w_audio.get("sample_rate", 44100)
+
+        # Append any audio from unexpected worker ids deterministically.
+        for worker_id_str in sorted(worker_audio.keys()):
+            if worker_id_str in seen:
+                continue
+            w_audio = worker_audio[worker_id_str]
+            if w_audio is not None:
+                waveform = w_audio.get("waveform")
+                if waveform is not None and waveform.numel() > 0:
+                    audio_pieces.append(waveform)
+                    if sample_rate == 44100:
+                        sample_rate = w_audio.get("sample_rate", 44100)
+
+        if not audio_pieces:
+            return empty_audio
+
+        try:
+            # Concatenate along the samples dimension (dim=-1)
+            # Ensure all pieces have same batch and channel dimensions
+            combined_waveform = torch.cat(audio_pieces, dim=-1)
+            debug_log(f"Master - Combined audio: {len(audio_pieces)} pieces, final shape={combined_waveform.shape}")
+            return {"waveform": combined_waveform, "sample_rate": sample_rate}
+        except Exception as e:
+            log(f"[Distributed] Master - Audio combination failed, returning silence: {e}")
+            return empty_audio
+
+    def _store_worker_result(self, worker_images: dict, item: dict) -> int:
+        """Store one canonical queue item in worker_images in-place.
+
+        Canonical format:
+        - item has 'worker_id', 'image_index', and 'tensor'
+        Returns 1 when stored, otherwise 0.
+        """
+        worker_id = item['worker_id']
+        tensor = item.get('tensor')
+        image_index = item.get('image_index')
+        if tensor is None or image_index is None:
+            return 0
+
+        worker_images.setdefault(worker_id, {})
+        worker_images[worker_id][image_index] = tensor
+        return 1
+
+    def _reorder_and_combine_tensors(
+        self,
+        worker_images: dict,
+        worker_order: list,
+        master_batch_size: int,
+        images_on_cpu,
+        delegate_mode: bool,
+        fallback_images,
+    ):
+        """Assemble final tensor, or return None when the job contains only audio."""
+        ordered_tensors = []
+        if not delegate_mode and images_on_cpu is not None:
+            for i in range(master_batch_size):
+                ordered_tensors.append(images_on_cpu[i:i+1])
+
+        ordered_worker_ids = [str(worker_id) for worker_id in (worker_order or [])]
+        seen = set()
+        for worker_id_str in ordered_worker_ids:
+            seen.add(worker_id_str)
+            if worker_id_str not in worker_images:
+                continue
+            for idx in sorted(worker_images[worker_id_str].keys()):
+                ordered_tensors.append(worker_images[worker_id_str][idx])
+
+        # Append any unexpected worker ids deterministically.
+        for worker_id_str in sorted(worker_images.keys()):
+            if worker_id_str in seen:
+                continue
+            for idx in sorted(worker_images[worker_id_str].keys()):
+                ordered_tensors.append(worker_images[worker_id_str][idx])
+
+        cpu_tensors = []
+        for t in ordered_tensors:
+            if t.is_cuda:
+                t = t.cpu()
+            t = ensure_contiguous(t)
+            cpu_tensors.append(t)
+
+        if cpu_tensors:
+            return ensure_contiguous(torch.cat(cpu_tensors, dim=0))
+        if fallback_images is not None:
+            return ensure_contiguous(fallback_images)
+        return None
+
+    async def execute(self, images, audio, load_balance=False, multi_job_id="", is_worker=False, master_url="", enabled_worker_ids="[]", worker_batch_size=1, worker_id="", delegate_only=False, transfer_format=transfer_codec.DEFAULT_TRANSFER_FORMAT, transfer_quality=transfer_codec.DEFAULT_TRANSFER_QUALITY):
+        if is_worker:
+            # Worker mode: send images and audio to master in a single batch
+            image_count = 0 if images is None else images.shape[0]
+            debug_log(f"Worker - Job {multi_job_id} complete. Sending {image_count} image(s) to master")
+            await self.send_batch_to_master(
+                images, audio, multi_job_id, master_url, worker_id, transfer_format, transfer_quality
+            )
+            return (images, audio if audio is not None else self.EMPTY_AUDIO)
+        else:
+            delegate_mode = delegate_only or is_master_delegate_only()
+            # Master mode: collect images and audio from workers
+            enabled_workers_raw = json.loads(enabled_worker_ids)
+            enabled_workers = []
+            seen_enabled = set()
+            for worker_id in enabled_workers_raw:
+                worker_id_str = str(worker_id)
+                if worker_id_str in seen_enabled:
+                    continue
+                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
+
+            num_workers = len(expected_workers)
+            if num_workers == 0:
+                return (images, audio if audio is not None else self.EMPTY_AUDIO)
+
+            # Create the queue before any expensive local work to avoid job_complete race.
+            async with prompt_server.distributed_jobs_lock:
+                if multi_job_id not in prompt_server.distributed_pending_jobs:
+                    prompt_server.distributed_pending_jobs[multi_job_id] = asyncio.Queue()
+                    debug_log(f"Master - Initialized queue early for job {multi_job_id}")
+                else:
+                    existing_size = prompt_server.distributed_pending_jobs[multi_job_id].qsize()
+                    debug_log(f"Master - Using existing queue for job {multi_job_id} (current size: {existing_size})")
+
+            if delegate_mode:
+                master_batch_size = 0
+                images_on_cpu = None
+                master_audio = None
+                debug_log(f"Master - Job {multi_job_id}: Delegate-only mode enabled, collecting exclusively from {num_workers} workers")
+            else:
+                if images is None:
+                    images_on_cpu = None
+                    master_batch_size = 0
+                else:
+                    images_on_cpu = ensure_contiguous(images.cpu())
+                    master_batch_size = images.shape[0]
+                master_audio = audio  # Keep master's audio for later
+                debug_log(f"Master - Job {multi_job_id}: Master has {master_batch_size} images, collecting from {num_workers} workers...")
+
+
+            # Initialize storage for collected images and audio
+            worker_images = {}  # Dict to store images by worker_id and index
+            worker_audio = {}   # Dict to store audio by worker_id
+            
+            # Collect images until all workers report they're done
+            collected_count = 0
+            workers_done = set()
+            
+            # Use unified worker timeout from config/UI with simple sliced waits
+            base_timeout = float(get_worker_timeout_seconds())
+            slice_timeout = min(max(0.1, HEARTBEAT_INTERVAL / 20.0), base_timeout)
+            last_activity = time.time()
+            
+            
+            # Get queue size before starting
+            async with prompt_server.distributed_jobs_lock:
+                q = prompt_server.distributed_pending_jobs[multi_job_id]
+                initial_size = q.qsize()
+
+            # NEW: Initialize progress bar for workers (total = num_workers)
+            p = ProgressBar(num_workers)
+
+            def mark_worker_done(done_worker_id):
+                done_worker_id = str(done_worker_id)
+                if done_worker_id not in expected_workers:
+                    debug_log(
+                        f"Master - Ignoring completion from unexpected worker {done_worker_id} for job {multi_job_id}"
+                    )
+                    return
+                if done_worker_id in workers_done:
+                    debug_log(
+                        f"Master - Ignoring duplicate completion from worker {done_worker_id} for job {multi_job_id}"
+                    )
+                    return
+                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"])
+                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.
+                    comfy.model_management.throw_exception_if_processing_interrupted()
+                    try:
+                        async with prompt_server.distributed_jobs_lock:
+                            q = prompt_server.distributed_pending_jobs[multi_job_id]
+                        result = await asyncio.wait_for(q.get(), timeout=slice_timeout)
+                        consume_result(result)
+
+                        # Real results refresh the regular-worker timeout baseline.
+                        last_activity = time.time()
+                        base_timeout = float(get_worker_timeout_seconds())
+
+                    except asyncio.TimeoutError:
+                        if (time.time() - last_activity) < base_timeout:
+                            comfy.model_management.throw_exception_if_processing_interrupted()
+                            continue
+                        comfy.model_management.throw_exception_if_processing_interrupted()
+                        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_regular):
+                            log(
+                                "Master - Heartbeat timeout: "
+                                f"worker={missing_worker_id}, 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 only regular worker /prompt endpoints.
+                        any_busy = False
+                        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}"
+                                        )
+                                        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:
+                            last_activity = time.time()
+                            base_timeout = float(get_worker_timeout_seconds())
+                            continue
+
+                        async with prompt_server.distributed_jobs_lock:
+                            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:
+                                        remaining_items.append(final_q.get_nowait())
+                                    except asyncio.QueueEmpty:
+                                        break
+                        for item in remaining_items:
+                            consume_result(item)
+                        break
+
+                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:
+                    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:
+                        raise RuntimeError(
+                            "Serverless result batch shape mismatch while combining participants "
+                            f"{sorted(serverless_results)}: {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
@@ -0,0 +1,270 @@
+import json
+import math
+from functools import wraps
+
+import comfy.samplers
+
+from ..utils.logging import debug_log, log
+from ..utils.async_helpers import run_async_in_server_loop
+from ..upscale.job_store import ensure_tile_jobs_initialized
+
+from ..upscale.tile_ops import TileOpsMixin
+from ..upscale.result_collector import ResultCollectorMixin
+from ..upscale.worker_comms import WorkerCommsMixin
+from ..upscale.job_state import JobStateMixin
+from ..upscale.modes.single_gpu import SingleGpuModeMixin
+from ..upscale.modes.static import StaticModeMixin
+from ..upscale.modes.dynamic import DynamicModeMixin
+
+def sync_wrapper(async_func):
+    """Decorator to wrap async methods for synchronous execution."""
+    @wraps(async_func)
+    def sync_func(self, *args, **kwargs):
+        # Use run_async_in_server_loop for ComfyUI compatibility
+        return run_async_in_server_loop(
+            async_func(self, *args, **kwargs),
+            timeout=600.0  # 10 minute timeout for long operations
+        )
+    return sync_func
+
+def _parse_enabled_worker_ids(enabled_worker_ids):
+    """Parse enabled worker IDs from either JSON or list input."""
+    if isinstance(enabled_worker_ids, list):
+        return [str(worker_id) for worker_id in enabled_worker_ids]
+    if not enabled_worker_ids:
+        return []
+    if isinstance(enabled_worker_ids, str):
+        try:
+            parsed = json.loads(enabled_worker_ids)
+        except json.JSONDecodeError:
+            log("USDU Dist: Invalid enabled_worker_ids JSON; defaulting to no workers.")
+            return []
+        if isinstance(parsed, list):
+            return [str(wid) for wid in parsed]
+    return []
+
+class UltimateSDUpscaleDistributed(
+    DynamicModeMixin,
+    StaticModeMixin,
+    SingleGpuModeMixin,
+    ResultCollectorMixin,
+    WorkerCommsMixin,
+    JobStateMixin,
+    TileOpsMixin,
+):
+
+    """
+    Distributed version of Ultimate SD Upscale (No Upscale).
+    
+    Supports two currently selected processing modes:
+    1. Single GPU: No workers available, process everything locally
+    2. Distributed tile queue: Workers pull tiles from a shared queue
+    
+    Features:
+    - Tile-based batch handling for video/image upscaling
+    - Tiled VAE support for memory efficiency
+    - Shared work queue so faster workers can process more tiles
+    - Backward compatible with single-image workflows
+    
+    Environment Variables:
+    - COMFYUI_MAX_BATCH: Chunk size for tile sending (default 20)
+    - COMFYUI_MAX_PAYLOAD_SIZE: Max API payload bytes (default 50MB)
+    
+    The hidden dynamic_threshold input is retained for workflow compatibility but
+    does not affect the current mode-selection policy.
+    """
+
+    def __init__(self):
+        """Initialize the node and ensure persistent storage exists."""
+        # Pre-initialize the persistent storage on node creation
+        ensure_tile_jobs_initialized()
+        debug_log("UltimateSDUpscaleDistributed - Node initialized")
+
+    @classmethod
+    def INPUT_TYPES(s):
+        return {
+            "required": {
+                "upscaled_image": ("IMAGE",),
+                "model": ("MODEL",),
+                "positive": ("CONDITIONING",),
+                "negative": ("CONDITIONING",),
+                "vae": ("VAE",),
+                "seed": ("INT", {"default": 0, "min": 0, "max": 0xffffffffffffffff}),
+                "steps": ("INT", {"default": 20, "min": 1, "max": 10000}),
+                "cfg": ("FLOAT", {"default": 8.0, "min": 0.0, "max": 100.0}),
+                "sampler_name": (comfy.samplers.KSampler.SAMPLERS,),
+                "scheduler": (comfy.samplers.KSampler.SCHEDULERS,),
+                "denoise": ("FLOAT", {"default": 0.5, "min": 0.0, "max": 1.0, "step": 0.01}),
+                "tile_width": ("INT", {"default": 512, "min": 64, "max": 2048, "step": 8}),
+                "tile_height": ("INT", {"default": 512, "min": 64, "max": 2048, "step": 8}),
+                "padding": ("INT", {"default": 32, "min": 0, "max": 256, "step": 8}),
+                "mask_blur": ("INT", {"default": 8, "min": 0, "max": 256}),
+                "force_uniform_tiles": ("BOOLEAN", {"default": True}),
+                "tiled_decode": ("BOOLEAN", {"default": False}),
+            },
+            "hidden": {
+                "multi_job_id": ("STRING", {"default": ""}),
+                "is_worker": ("BOOLEAN", {"default": False}),
+                "master_url": ("STRING", {"default": ""}),
+                "enabled_worker_ids": ("STRING", {"default": "[]"}),
+                "worker_id": ("STRING", {"default": ""}),
+                "tile_indices": ("STRING", {"default": ""}),  # Unused - kept for compatibility
+                "dynamic_threshold": ("INT", {"default": 8, "min": 1, "max": 64}),
+            },
+        }
+
+    RETURN_TYPES = ("IMAGE",)
+    FUNCTION = "run"
+    CATEGORY = "image/upscaling"
+
+    @classmethod
+    def IS_CHANGED(cls, **kwargs):
+        """Force re-execution."""
+        return float("nan")  # Always re-execute
+
+    def run(self, upscaled_image, model, positive, negative, vae, seed, steps, cfg, 
+            sampler_name, scheduler, denoise, tile_width, tile_height, padding, 
+            mask_blur, force_uniform_tiles, tiled_decode,
+            multi_job_id="", is_worker=False, master_url="", enabled_worker_ids="[]", 
+            worker_id="", tile_indices="", dynamic_threshold=8):
+        """Entry point - runs SYNCHRONOUSLY like Ultimate SD Upscaler."""
+        # Strict WAN/FLOW batching: error if batch is not 4n+1 (except allow 1)
+        try:
+            batch_size = int(getattr(upscaled_image, 'shape', [1])[0])
+        except Exception:
+            batch_size = 1
+        # Enforce 4n+1 batches globally for any model when batch > 1 (master only)
+        if not is_worker and batch_size != 1 and (batch_size % 4 != 1):
+            raise ValueError(
+                f"Batch size {batch_size} is not of the form 4n+1. "
+                "This node requires batch sizes of 1 or 4n+1 (1, 5, 9, 13, ...). "
+                "Please adjust the batch size."
+            )
+        if not multi_job_id:
+            # No distributed processing, run single GPU version
+            return self.process_single_gpu(upscaled_image, model, positive, negative, vae,
+                                          seed, steps, cfg, sampler_name, scheduler, denoise,
+                                          tile_width, tile_height, padding, mask_blur, force_uniform_tiles, tiled_decode)
+        
+        if is_worker:
+            # Worker mode: process tiles synchronously
+            return self.process_worker(upscaled_image, model, positive, negative, vae,
+                                      seed, steps, cfg, sampler_name, scheduler, denoise,
+                                      tile_width, tile_height, padding, mask_blur,
+                                      force_uniform_tiles, tiled_decode, multi_job_id, master_url,
+                                      worker_id, enabled_worker_ids, dynamic_threshold)
+        else:
+            # Master mode: distribute and collect synchronously
+            return self.process_master(upscaled_image, model, positive, negative, vae,
+                                     seed, steps, cfg, sampler_name, scheduler, denoise,
+                                     tile_width, tile_height, padding, mask_blur,
+                                     force_uniform_tiles, tiled_decode, multi_job_id, enabled_worker_ids, 
+                                     dynamic_threshold)
+
+    def process_worker(self, upscaled_image, model, positive, negative, vae,
+                      seed, steps, cfg, sampler_name, scheduler, denoise,
+                      tile_width, tile_height, padding, mask_blur,
+                      force_uniform_tiles, tiled_decode, multi_job_id, master_url,
+                      worker_id, enabled_worker_ids, dynamic_threshold):
+        """Unified worker processing - handles both static and dynamic modes."""
+        # Get batch size to determine mode
+        batch_size = upscaled_image.shape[0]
+        
+        # Ensure mode consistency across master/workers via shared threshold
+        # Determine mode (must match master's logic)
+        enabled_workers = json.loads(enabled_worker_ids)
+        num_workers = len(enabled_workers)
+        # Compute number of tiles for this image to decide if tile distribution makes sense
+        _, height, width, _ = upscaled_image.shape
+        all_tiles = self.calculate_tiles(width, height, self.round_to_multiple(tile_width), self.round_to_multiple(tile_height), force_uniform_tiles)
+        num_tiles_per_image = len(all_tiles)
+
+        mode = self._determine_processing_mode(batch_size, num_workers, dynamic_threshold)
+        # For USDU-style processing, we want tile distribution whenever workers are available
+        # and there is more than one tile to process, even if batch == 1.
+        if num_workers > 0 and num_tiles_per_image > 1:
+            mode = "static"
+            
+        debug_log(f"USDU Dist Worker - Batch size {batch_size}")
+        
+        if mode == "dynamic":
+            return self.process_worker_dynamic(upscaled_image, model, positive, negative, vae,
+                                             seed, steps, cfg, sampler_name, scheduler, denoise,
+                                             tile_width, tile_height, padding, mask_blur,
+                                             force_uniform_tiles, tiled_decode, multi_job_id, master_url,
+                                             worker_id, enabled_worker_ids, dynamic_threshold)
+        
+        # Static mode - enhanced with health monitoring and retry logic
+        return self._process_worker_static_sync(upscaled_image, model, positive, negative, vae,
+                                               seed, steps, cfg, sampler_name, scheduler, denoise,
+                                               tile_width, tile_height, padding, mask_blur,
+                                               force_uniform_tiles, tiled_decode, multi_job_id, master_url,
+                                               worker_id, enabled_workers)
+
+    def process_master(self, upscaled_image, model, positive, negative, vae,
+                      seed, steps, cfg, sampler_name, scheduler, denoise,
+                      tile_width, tile_height, padding, mask_blur,
+                      force_uniform_tiles, tiled_decode, multi_job_id, enabled_worker_ids, 
+                      dynamic_threshold):
+        """Unified master processing with enhanced monitoring and failure handling."""
+        # Round tile dimensions
+        tile_width = self.round_to_multiple(tile_width)
+        tile_height = self.round_to_multiple(tile_height)
+        
+        # Get image dimensions and batch size
+        batch_size, height, width, _ = upscaled_image.shape
+        
+        # Calculate all tiles and grid
+        all_tiles = self.calculate_tiles(width, height, tile_width, tile_height, force_uniform_tiles)
+        num_tiles_per_image = len(all_tiles)
+        rows = math.ceil(height / tile_height)
+        cols = math.ceil(width / tile_width)
+        log(
+            f"USDU Dist: Canvas {width}x{height} | Tile {tile_width}x{tile_height} | Grid {rows}x{cols} ({num_tiles_per_image} tiles/image) | Batch {batch_size}"
+        )
+        
+        # Parse enabled workers
+        enabled_workers = json.loads(enabled_worker_ids)
+        num_workers = len(enabled_workers)
+        
+        # Determine processing mode
+        mode = self._determine_processing_mode(batch_size, num_workers, dynamic_threshold)
+        # Prefer tile-based static distribution when workers are available and there are multiple tiles,
+        # even for batch == 1, to spread tiles across GPUs like the legacy dynamic tile queue.
+        if num_workers > 0 and num_tiles_per_image > 1:
+            mode = "static"
+        
+        log(f"USDU Dist: Workers {num_workers} | Mode {mode} | Threshold {dynamic_threshold}")
+
+        if mode == "single_gpu":
+            # No workers, process all tiles locally
+            return self.process_single_gpu(upscaled_image, model, positive, negative, vae,
+                                         seed, steps, cfg, sampler_name, scheduler, denoise,
+                                         tile_width, tile_height, padding, mask_blur, force_uniform_tiles, tiled_decode)
+        
+        elif mode == "dynamic":
+            # Dynamic mode for large batches
+            return self.process_master_dynamic(upscaled_image, model, positive, negative, vae,
+                                             seed, steps, cfg, sampler_name, scheduler, denoise,
+                                             tile_width, tile_height, padding, mask_blur,
+                                             force_uniform_tiles, tiled_decode, multi_job_id, enabled_workers)
+        
+        # Static mode - enhanced with unified job management
+        return self._process_master_static_sync(upscaled_image, model, positive, negative, vae,
+                                               seed, steps, cfg, sampler_name, scheduler, denoise,
+                                               tile_width, tile_height, padding, mask_blur,
+                                               force_uniform_tiles, tiled_decode, multi_job_id, enabled_workers,
+                                               all_tiles, num_tiles_per_image)
+
+    def _determine_processing_mode(self, batch_size: int, num_workers: int, dynamic_threshold: int) -> str:
+        """Determines processing mode per requested policy:
+        - any workers     => prefer static (tile-based) for USDU
+        - no workers      => single_gpu
+        """
+        if num_workers == 0:
+            return "single_gpu"
+        # Default to static when distributed; master/worker may still override if special cases arise
+        return "static"
+
+# Ensure initialization before registering routes
+ensure_tile_jobs_initialized()
@@ -0,0 +1,258 @@
+"""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,
+    create_master_job,
+    normalize_image_list_inputs,
+    normalize_worker_result,
+    run_image_list_job,
+    encode_result,
+    encode_request_payload,
+    unwrap_single,
+)
+
+
+# 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_token = unwrap_single(worker_item_token)
+        worker_context_id = unwrap_single(worker_context_id)
+        image_list_workflow_meta = unwrap_single(image_list_workflow_meta)
+        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, and text lists.",
+            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"),
+            ],
+            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),
+            ],
+            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)):
+                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_result_token = unwrap_single(worker_result_token)
+        master_callback_url = unwrap_single(master_callback_url)
+        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")
+            result = normalize_worker_result(images, masks, texts)
+            if unwrap_single(image_list_local_execution):
+                from ..image_list.runtime import set_master_result
+                set_master_result(context_values[0], 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_values[0], int(unwrap_single(worker_item_index)), payload), timeout=60,
+                )
+            return io.NodeOutput([result["image"]], [result.get("mask")], [result.get("text")])
+
+        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))
+        except ImageListCancelledError:
+            import comfy.model_management
+            raise comfy.model_management.InterruptProcessingException()
+        images_out, masks_out, texts_out = result_lists
+        if len(images_out) != len(masks_out) or len(images_out) != len(texts_out):
+            raise RuntimeError("Image-list collector received mismatched image, mask, or text result counts")
+        return io.NodeOutput(images_out, masks_out, texts_out)
+
+
+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}")
@@ -0,0 +1,229 @@
+"""Typed, lossless list collection for serverless media participants.
+
+The master compiler supplies expected_media before pruning delegate inputs.
+The provider manager owns remote deadlines and delivers one atomic native-media
+message per participant. This node never serializes, concatenates or resamples.
+"""
+import asyncio
+import json
+import time
+
+import comfy.model_management
+from comfy_api.v0_0_2 import io
+import server
+
+from ..serverless_bridge.contract import MAX_ITEMS, SOCKET_KINDS, identifier
+from ..serverless_bridge.image_transfer import (FORMATS, DEFAULT_FORMAT, DEFAULT_QUALITY,
+                                               FORMAT_TOOLTIP, QUALITY_TOOLTIP, validate_options)
+from ..utils.async_helpers import run_async_in_server_loop
+
+POLL_SECONDS = 0.1
+START_GRACE_SECONDS = 2.0
+
+
+def _control(value, name, kind):
+    # Only execution controls are scalars wrapped by ComfyUI's list mapper.
+    if isinstance(value, list):
+        if len(value) != 1:
+            raise ValueError(f'{name} must be a singleton control')
+        value = value[0]
+    if type(value) is not kind:
+        raise ValueError(f'Invalid {name} control')
+    return value
+
+
+def _names(value, name):
+    try:
+        result = json.loads(_control(value, name, str))
+    except (ValueError, TypeError):
+        raise ValueError(f'{name} must be a JSON string list') from None
+    if (not isinstance(result, list) or any(not isinstance(item, str) for item in result)
+            or len(result) != len(set(result))):
+        raise ValueError(f'{name} must contain unique strings')
+    return result
+
+
+def _validate_media(media, expected):
+    if not isinstance(media, dict) or set(media) != set(expected):
+        raise ValueError('Media sockets do not match expected_media')
+    for socket, items in media.items():
+        if (socket not in SOCKET_KINDS or not isinstance(items, list)
+                or not 1 <= len(items) <= MAX_ITEMS or any(item is None for item in items)):
+            raise ValueError('Connected media sockets require nonempty bounded item lists')
+
+
+def _prepare(media, controls):
+    worker = _control(controls['is_worker'], 'is_worker', bool)
+    delegate = _control(controls['delegate_only'], 'delegate_only', bool)
+    passthrough = _control(controls['pass_through'], 'pass_through', bool)
+    _control(controls['load_balance'], 'load_balance', bool)  # Routing is master-owned.
+    if worker:
+        raise ValueError('DistributedMediaCollector does not support direct-worker execution')
+    participants = _names(controls['enabled_worker_ids'], 'enabled_worker_ids')
+    for participant in participants:
+        identifier(participant)
+        if not participant.startswith('serverless:') or not participant[len('serverless:'):]:
+            raise ValueError('DistributedMediaCollector requires serverless participant IDs')
+    expected = _names(controls['expected_media'], 'expected_media')
+    if any(socket not in SOCKET_KINDS for socket in expected):
+        raise ValueError('Unknown expected_media socket')
+    local = {socket: items for socket, items in media.items() if items is not None}
+    _validate_media(local, local)
+    expected = expected or list(local)
+    if not expected:
+        raise ValueError('At least one connected media socket is required')
+    if not delegate or passthrough:
+        _validate_media(local, expected)
+    if delegate and not passthrough and not participants:
+        raise ValueError('Delegate-only collection requires serverless participants')
+    return local, participants, expected, delegate, passthrough
+
+
+def _outputs(media):
+    return tuple(media.get(socket, []) for socket in SOCKET_KINDS)
+
+
+def _consume(message, participants, expected, received):
+    if not isinstance(message, dict):
+        raise RuntimeError('Invalid serverless media delivery')
+    participant = message.get('worker_id')
+    if not isinstance(participant, str) or participant not in participants:
+        raise RuntimeError('Unexpected serverless participant')
+    if participant in received:
+        raise RuntimeError('Duplicate serverless participant delivery')
+    if message.get('is_last') is not True:
+        raise RuntimeError('Serverless media delivery must be atomic and final')
+    if 'serverless_error' in message:
+        if (set(message) != {'worker_id', 'serverless_error', 'is_last'}
+                or not isinstance(message['serverless_error'], str) or not message['serverless_error']):
+            raise RuntimeError('Invalid serverless error delivery')
+        # Do not echo provider-controlled text, credentials or signed URLs.
+        raise RuntimeError(f'Serverless participant {participant} failed to produce media')
+    if set(message) != {'worker_id', 'media', 'is_last'}:
+        raise RuntimeError('Invalid serverless media delivery fields')
+    _validate_media(message['media'], expected)
+    received[participant] = message['media']
+
+
+async def _collect(job, media, controls):
+    ps = server.PromptServer.instance
+    queue = None
+    try:
+        # Reuse the master's precreated queue, including early completions.
+        async with ps.distributed_jobs_lock:
+            queue = ps.distributed_pending_jobs.setdefault(job, asyncio.Queue())
+        local, participants, expected, delegate, passthrough = _prepare(media, controls)
+        if passthrough or not participants:
+            return _outputs(local)
+        received = {}
+        started = time.monotonic()
+        while len(received) < len(participants):
+            comfy.model_management.throw_exception_if_processing_interrupted()
+            async with ps.distributed_jobs_lock:
+                if ps.distributed_pending_jobs.get(job) is not queue:
+                    raise RuntimeError('Serverless collector queue was removed or replaced')
+            try:
+                message = await asyncio.wait_for(queue.get(), timeout=POLL_SECONDS)
+            except asyncio.TimeoutError:
+                comfy.model_management.throw_exception_if_processing_interrupted()
+                # wait_for cancellation may race an atomic manager delivery.
+                if not queue.empty():
+                    continue
+                manager = getattr(ps, 'distributed_serverless', None)
+                for participant in participants:
+                    if participant in received:
+                        continue
+                    if manager is not None and manager.is_pending(job, participant):
+                        continue  # The manager enforces its remote absolute deadline.
+                    if manager is not None and manager.owns(job, participant):
+                        raise RuntimeError('Serverless participant stopped without delivering media')
+                    # The master queues its prompt before manager.start journals slots.
+                    if time.monotonic() - started >= START_GRACE_SECONDS:
+                        raise RuntimeError('Serverless manager did not start the expected participant')
+                continue
+            _consume(message, participants, expected, received)
+
+        comfy.model_management.throw_exception_if_processing_interrupted()
+        # Drain and close atomically with the manager's delivery lock. A queued
+        # duplicate after the final expected message must not yield success.
+        async with ps.distributed_jobs_lock:
+            if ps.distributed_pending_jobs.get(job) is not queue:
+                raise RuntimeError('Serverless collector queue was removed or replaced')
+            while not queue.empty():
+                _consume(queue.get_nowait(), participants, expected, received)
+            del ps.distributed_pending_jobs[job]
+        ordered = [] if delegate else [local]
+        ordered.extend(received[participant] for participant in participants)
+        return tuple([item for source in ordered for item in source.get(socket, [])]
+                     for socket in SOCKET_KINDS)
+    except BaseException:
+        # Remove only our queue before cancellation can trigger final deliveries.
+        async with ps.distributed_jobs_lock:
+            if queue is not None and ps.distributed_pending_jobs.get(job) is queue:
+                del ps.distributed_pending_jobs[job]
+        manager = getattr(ps, 'distributed_serverless', None)
+        if manager is not None:
+            try:
+                await manager.cancel(collector_job_id=job)
+            except BaseException:
+                pass  # Preserve the original failure/interruption, not cleanup's.
+        raise
+    finally:
+        async with ps.distributed_jobs_lock:
+            if queue is not None and ps.distributed_pending_jobs.get(job) is queue:
+                del ps.distributed_pending_jobs[job]
+
+
+async def _bridge_result(coro):
+    # The legacy sync bridge catches Exception, not asyncio.CancelledError.
+    # Carry BaseException as data so cancellation is re-raised on the caller.
+    try:
+        return await coro, None
+    except BaseException as error:
+        return None, error
+
+
+class DistributedMediaCollector(io.ComfyNode):
+    @classmethod
+    def define_schema(cls):
+        types = (io.Image, io.Audio, io.Mask, io.String, io.Video)
+        return io.Schema(
+            node_id='DistributedMediaCollector', display_name='Distributed Media Collector',
+            category='utils',
+            inputs=[io.Boolean.Input('load_balance', default=False),
+                    # Native non-widget inputs expose forceInput via extra_dict.
+                    *[kind.Input(socket, optional=True,
+                                 **({'force_input': True} if socket == 'texts'
+                                    else {'extra_dict': {'forceInput': True}}))
+                      for socket, kind in zip(SOCKET_KINDS, types)],
+                    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=[kind.Output(display_name=socket, is_output_list=True)
+                     for socket, kind in zip(SOCKET_KINDS, types)],
+            is_input_list=True, accept_all_inputs=True,
+        )
+
+    @classmethod
+    def execute(cls, images=None, audio=None, masks=None, texts=None, video=None,
+                load_balance=False, multi_job_id='', is_worker=False,
+                enabled_worker_ids='[]', delegate_only=False, pass_through=False,
+                expected_media='[]', transfer_format=DEFAULT_FORMAT, transfer_quality=DEFAULT_QUALITY):
+        validate_options(_control(transfer_format, 'transfer_format', str),
+                         _control(transfer_quality, 'transfer_quality', int))
+        media = dict(images=images, audio=audio, masks=masks, texts=texts, video=video)
+        controls = dict(load_balance=load_balance, is_worker=is_worker,
+                        enabled_worker_ids=enabled_worker_ids, delegate_only=delegate_only,
+                        pass_through=pass_through, expected_media=expected_media)
+        job = _control(multi_job_id, 'multi_job_id', str)
+        if not job:
+            local, participants, _, _, passthrough = _prepare(media, controls)
+            if participants and not passthrough:
+                raise ValueError('Serverless collection requires multi_job_id')
+            return io.NodeOutput(*_outputs(local))
+        identifier(job)
+        result, error = run_async_in_server_loop(_bridge_result(_collect(job, media, controls)))
+        if error is not None:
+            raise error
+        return io.NodeOutput(*result)
@@ -0,0 +1,354 @@
+import torch
+import json
+
+from ..utils.logging import debug_log, log
+
+
+def _chunk_bounds(total_items: int, n_splits: int) -> list[tuple[int, int]]:
+    """Return contiguous [start, end) bounds for n_splits chunks."""
+    split_count = max(1, int(n_splits))
+    total = max(0, int(total_items))
+    base, remainder = divmod(total, split_count)
+
+    bounds: list[tuple[int, int]] = []
+    start = 0
+    for idx in range(split_count):
+        size = base + (1 if idx < remainder else 0)
+        end = start + size
+        bounds.append((start, end))
+        start = end
+    return bounds
+
+
+class DistributedSeed:
+    """
+    Distributes seed values across multiple GPUs.
+    On master: passes through the original seed.
+    On workers: adds offset based on worker ID.
+    """
+    
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {
+            "required": {
+                "seed": ("INT", {
+                    "default": 1125899906842, 
+                    "min": 0,
+                    "max": 1125899906842624,
+                    "forceInput": False  # Widget by default, can be converted to input
+                }),
+            },
+            "hidden": {
+                "is_worker": ("BOOLEAN", {"default": False}),
+                "worker_id": ("STRING", {"default": ""}),
+            },
+        }
+    
+    RETURN_TYPES = ("INT",)
+    RETURN_NAMES = ("seed",)
+    FUNCTION = "distribute"
+    CATEGORY = "utils"
+    
+    def distribute(self, seed, is_worker=False, worker_id=""):
+        if not is_worker:
+            # Master node: pass through original values
+            debug_log(f"Distributor - Master: seed={seed}")
+            return (seed,)
+        else:
+            # Worker node: apply offset based on worker index
+            # Find worker index from enabled_worker_ids
+            try:
+                # Worker IDs are passed as "worker_0", "worker_1", etc.
+                if worker_id.startswith("worker_"):
+                    worker_index = int(worker_id.split("_")[1])
+                else:
+                    # Fallback: try to parse as direct index
+                    worker_index = int(worker_id)
+                
+                offset = worker_index + 1
+                new_seed = seed + offset
+                debug_log(f"Distributor - Worker {worker_index}: seed={seed} → {new_seed}")
+                return (new_seed,)
+            except (ValueError, IndexError) as e:
+                debug_log(f"Distributor - Error parsing worker_id '{worker_id}': {e}")
+                # Fallback: return original seed
+                return (seed,)
+
+
+# Define ByPassTypeTuple for flexible return types
+class AnyType(str):
+    def __ne__(self, __value: object) -> bool:
+        return False
+
+any_type = AnyType("*")
+
+
+class DistributedValue:
+    """
+    Outputs a different value per worker.
+    On master: returns default_value.
+    On workers: looks up the worker-specific value from a JSON map,
+    falling back to default_value if not set.
+    """
+
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {
+            "required": {
+                "default_value": ("STRING", {"default": ""}),
+                "worker_values": ("STRING", {"default": "{}"}),
+            },
+            "hidden": {
+                "is_worker": ("BOOLEAN", {"default": False}),
+                "worker_id": ("STRING", {"default": ""}),
+            },
+        }
+
+    RETURN_TYPES = (any_type,)
+    RETURN_NAMES = ("value",)
+    FUNCTION = "distribute"
+    CATEGORY = "utils"
+
+    @staticmethod
+    def _coerce(value, value_type):
+        """Convert a string value to the requested type."""
+        if value_type == "INT":
+            return int(float(value))
+        if value_type == "FLOAT":
+            return float(value)
+        return value  # STRING and COMBO stay as strings
+
+    @staticmethod
+    def _coerce_safe(value, value_type):
+        """Best-effort coercion with graceful fallback to original value."""
+        try:
+            return DistributedValue._coerce(value, value_type)
+        except (TypeError, ValueError):
+            return value
+
+    def distribute(self, default_value, worker_values="{}", is_worker=False, worker_id=""):
+        values = {}
+        value_type = "STRING"
+
+        try:
+            values = json.loads(worker_values) if isinstance(worker_values, str) else worker_values
+            if not isinstance(values, dict):
+                values = {}
+        except json.JSONDecodeError as e:
+            debug_log(f"DistributedValue - Error parsing worker_values: {e}")
+            values = {}
+
+        value_type = values.get("_type", "STRING")
+        coerced_default = self._coerce_safe(default_value, value_type)
+
+        if not is_worker:
+            debug_log(f"DistributedValue - Master: returning default '{coerced_default}'")
+            return (coerced_default,)
+
+        try:
+            if worker_id.startswith("worker_"):
+                idx = int(worker_id.split("_")[1])
+            else:
+                idx = int(worker_id)
+            key = str(idx + 1)  # worker_0 → key "1" (1-indexed)
+            raw = values.get(key, "")
+            if raw:
+                coerced = self._coerce(raw, value_type)
+                debug_log(f"DistributedValue - Worker {idx}: returning '{coerced}'")
+                return (coerced,)
+        except (ValueError, IndexError) as e:
+            debug_log(f"DistributedValue - Error: {e}")
+        debug_log(f"DistributedValue - Worker fallback: returning default '{coerced_default}'")
+        return (coerced_default,)
+
+class DistributedModelName:
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {
+            "required": {
+                "text": ("STRING", {"default": ""}),
+            },
+            "hidden": {
+                "unique_id": "UNIQUE_ID",
+                "extra_pnginfo": "EXTRA_PNGINFO",
+            },
+        }
+
+    RETURN_TYPES = (any_type,)
+    RETURN_NAMES = ("output",)
+    FUNCTION = "log_input"
+    OUTPUT_NODE = True
+    CATEGORY = "utils"
+
+    def _stringify(self, value):
+        if isinstance(value, str):
+            return value
+        if isinstance(value, (int, float, bool)):
+            return str(value)
+        try:
+            return json.dumps(value, indent=4)
+        except Exception:
+            return str(value)
+
+    def _update_workflow(self, extra_pnginfo, unique_id, values):
+        if not extra_pnginfo:
+            return
+        info = extra_pnginfo[0] if isinstance(extra_pnginfo, list) else extra_pnginfo
+        if not isinstance(info, dict) or "workflow" not in info:
+            return
+        node_id = None
+        if isinstance(unique_id, list) and unique_id:
+            node_id = str(unique_id[0])
+        elif unique_id is not None:
+            node_id = str(unique_id)
+        if not node_id:
+            return
+        workflow = info["workflow"]
+        node = next((x for x in workflow["nodes"] if str(x.get("id")) == node_id), None)
+        if node:
+            node["widgets_values"] = [values]
+
+    def log_input(self, text, unique_id=None, extra_pnginfo=None):
+        values = []
+        if isinstance(text, list):
+            for val in text:
+                values.append(self._stringify(val))
+        else:
+            values.append(self._stringify(text))
+
+        # Keep widget display in workflow metadata if available.
+        self._update_workflow(extra_pnginfo, unique_id, values)
+
+        if isinstance(values, list) and len(values) == 1:
+            return {"ui": {"text": values}, "result": (values[0],)}
+        return {"ui": {"text": values}, "result": (values,)}
+
+class ByPassTypeTuple(tuple):
+    def __getitem__(self, index):
+        if index > 0:
+            index = 0
+        item = super().__getitem__(index)
+        if isinstance(item, str):
+            return any_type
+        return item
+
+class ImageBatchDivider:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {
+            "required": {
+                "images": ("IMAGE",),
+                "divide_by": ("INT", {
+                    "default": 2, 
+                    "min": 1, 
+                    "max": 10, 
+                    "step": 1,
+                    "display": "number",
+                    "tooltip": "Number of parts to divide the batch into"
+                }),
+            }
+        }
+    
+    RETURN_TYPES = ByPassTypeTuple(("IMAGE", ))  # Flexible for variable outputs
+    RETURN_NAMES = ByPassTypeTuple(tuple([f"batch_{i+1}" for i in range(10)]))
+    FUNCTION = "divide_batch"
+    OUTPUT_NODE = True
+    CATEGORY = "image"
+    
+    def divide_batch(self, images, divide_by):
+        total_splits = max(1, min(int(divide_by), 10))
+        total_frames = images.shape[0]
+        empty_tensor = images[:0]
+        bounds = _chunk_bounds(total_frames, total_splits)
+        outputs = [images[start:end] if end > start else empty_tensor for start, end in bounds]
+
+        while len(outputs) < 10:
+            outputs.append(empty_tensor)
+
+        return tuple(outputs[:10])
+
+
+class AudioBatchDivider:
+    """Divides an audio waveform into sequential segments along the time dimension."""
+
+    @classmethod
+    def INPUT_TYPES(s):
+        return {
+            "required": {
+                "audio": ("AUDIO",),
+                "divide_by": ("INT", {
+                    "default": 2,
+                    "min": 1,
+                    "max": 10,
+                    "step": 1,
+                    "display": "number",
+                    "tooltip": "Number of sequential time segments to create"
+                }),
+            }
+        }
+
+    RETURN_TYPES = ByPassTypeTuple(("AUDIO",))  # Flexible for variable outputs
+    RETURN_NAMES = ByPassTypeTuple(tuple([f"audio_{i+1}" for i in range(10)]))
+    FUNCTION = "divide_audio"
+    OUTPUT_NODE = True
+    CATEGORY = "audio"
+
+    def divide_audio(self, audio, divide_by):
+        import torch
+
+        waveform = audio.get("waveform")
+        sample_rate = audio.get("sample_rate", 44100)
+
+        if waveform is None or waveform.numel() == 0:
+            # Return empty audio for all outputs
+            empty_audio = {"waveform": torch.zeros(1, 2, 1), "sample_rate": sample_rate}
+            return tuple([empty_audio] * 10)
+
+        total_splits = max(1, min(int(divide_by), 10))
+        total_samples = int(waveform.shape[-1])
+        bounds = _chunk_bounds(total_samples, total_splits)
+
+        outputs = []
+        empty_waveform = waveform[..., :0]
+        for start, end in bounds:
+            split_waveform = waveform[..., start:end] if end > start else empty_waveform
+            outputs.append({
+                "waveform": split_waveform,
+                "sample_rate": sample_rate
+            })
+
+        # Pad with empty audio up to max (10) to match RETURN_TYPES length
+        empty_audio = {
+            "waveform": empty_waveform,
+            "sample_rate": sample_rate
+        }
+
+        while len(outputs) < 10:
+            outputs.append(empty_audio)
+
+        return tuple(outputs)
+
+
+class DistributedEmptyImage:
+    """Produces an empty IMAGE batch used when the master delegates all work."""
+
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {
+            "required": {
+                "height": ("INT", {"default": 64, "min": 1, "max": 4096, "step": 1}),
+                "width": ("INT", {"default": 64, "min": 1, "max": 4096, "step": 1}),
+                "channels": ("INT", {"default": 3, "min": 1, "max": 4, "step": 1}),
+            }
+        }
+
+    RETURN_TYPES = ("IMAGE",)
+    FUNCTION = "create"
+    CATEGORY = "image"
+
+    def create(self, height, width, channels):
+        import torch
+
+        shape = (0, height, width, channels)
+        tensor = torch.zeros(shape, dtype=torch.float32)
+        return (tensor,)
@@ -0,0 +1,216 @@
+"""Native V3 schemas with the existing execution algorithms kept intact.
+
+Runtime objects are private, per-execution helpers, not registered V1 nodes.
+This avoids sharing mutable instance state through V3's sanitized class clones.
+"""
+import comfy.samplers
+from comfy_api.v0_0_2 import io
+
+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
+from .media_collector import DistributedMediaCollector
+
+
+class DistributedSeed(io.ComfyNode):
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id='DistributedSeed', display_name='Distributed Seed', category='utils',
+            inputs=[io.Int.Input('seed', default=1125899906842, min=0,
+                                 max=1125899906842624, force_input=False)],
+            outputs=[io.Int.Output(display_name='seed')],
+            accept_all_inputs=True,
+        )
+
+    @classmethod
+    def execute(cls, seed, is_worker=False, worker_id=''):
+        return io.NodeOutput(*_utilities.DistributedSeed().distribute(seed, is_worker, worker_id))
+
+
+class DistributedValue(io.ComfyNode):
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id='DistributedValue', display_name='Distributed Value', category='utils',
+            inputs=[io.String.Input('default_value', default=''),
+                    io.String.Input('worker_values', default='{}')],
+            outputs=[io.AnyType.Output(display_name='value')],
+            accept_all_inputs=True,
+        )
+
+    @classmethod
+    def execute(cls, default_value, worker_values='{}', is_worker=False, worker_id=''):
+        return io.NodeOutput(*_utilities.DistributedValue().distribute(
+            default_value, worker_values, is_worker, worker_id))
+
+
+class DistributedModelName(io.ComfyNode):
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id='DistributedModelName', display_name='Distributed Model Name', category='utils',
+            inputs=[io.String.Input('text', default='')],
+            outputs=[io.AnyType.Output(display_name='output')],
+            hidden=[io.Hidden.unique_id, io.Hidden.extra_pnginfo], is_output_node=True,
+        )
+
+    @classmethod
+    def execute(cls, text):
+        result = _utilities.DistributedModelName().log_input(
+            text, unique_id=cls.hidden.unique_id, extra_pnginfo=cls.hidden.extra_pnginfo)
+        return io.NodeOutput(*result['result'], ui=result['ui'])
+
+
+class ImageBatchDivider(io.ComfyNode):
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id='ImageBatchDivider', display_name='Image Batch Divider', category='image',
+            inputs=[io.Image.Input('images'),
+                    io.Int.Input('divide_by', default=2, min=1, max=10, step=1,
+                                 display_mode=io.NumberDisplay.number,
+                                 tooltip='Number of parts to divide the batch into')],
+            # The existing frontend still displays only divide_by sockets.
+            outputs=[io.Image.Output(display_name=f'batch_{index + 1}') for index in range(10)],
+            is_output_node=True,
+        )
+
+    @classmethod
+    def execute(cls, images, divide_by):
+        return io.NodeOutput(*_utilities.ImageBatchDivider().divide_batch(images, divide_by))
+
+
+class AudioBatchDivider(io.ComfyNode):
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id='AudioBatchDivider', display_name='Audio Segment Divider', category='audio',
+            inputs=[io.Audio.Input('audio'),
+                    io.Int.Input('divide_by', default=2, min=1, max=10, step=1,
+                                 display_mode=io.NumberDisplay.number,
+                                 tooltip='Number of sequential time segments to create')],
+            outputs=[io.Audio.Output(display_name=f'audio_{index + 1}') for index in range(10)],
+            is_output_node=True,
+        )
+
+    @classmethod
+    def execute(cls, audio, divide_by):
+        return io.NodeOutput(*_utilities.AudioBatchDivider().divide_audio(audio, divide_by))
+
+
+class DistributedEmptyImage(io.ComfyNode):
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id='DistributedEmptyImage', display_name='Distributed Empty Image', category='image',
+            inputs=[io.Int.Input('height', default=64, min=1, max=4096, step=1),
+                    io.Int.Input('width', default=64, min=1, max=4096, step=1),
+                    io.Int.Input('channels', default=3, min=1, max=4, step=1)],
+            outputs=[io.Image.Output()],
+        )
+
+    @classmethod
+    def execute(cls, height, width, channels):
+        return io.NodeOutput(*_utilities.DistributedEmptyImage().create(height, width, channels))
+
+
+class DistributedCollector(io.ComfyNode):
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id='DistributedCollector', display_name='Distributed Collector', category='image',
+            inputs=[io.Boolean.Input('load_balance', default=False,
+                                     tooltip='Run this workflow on one least-busy participant (master included when participating).'),
+                    io.Image.Input('images', optional=True), io.Audio.Input('audio', optional=True),
+                    io.Combo.Input('transfer_format', options=list(_transfer_codec.TRANSFER_FORMATS),
+                                   default=_transfer_codec.DEFAULT_TRANSFER_FORMAT, optional=True,
+                                   tooltip=_transfer_codec.TRANSFER_FORMAT_TOOLTIP),
+                    io.Int.Input('transfer_quality', default=_transfer_codec.DEFAULT_TRANSFER_QUALITY,
+                                 min=1, max=100, step=1, optional=True,
+                                 tooltip=_transfer_codec.TRANSFER_QUALITY_TOOLTIP)],
+            outputs=[io.Image.Output(display_name='images'), io.Audio.Output(display_name='audio')],
+            is_input_list=True, accept_all_inputs=True,
+        )
+
+    @classmethod
+    def execute(cls, images=None, load_balance=False, audio=None, multi_job_id='',
+                is_worker=False, master_url='', enabled_worker_ids='[]', worker_batch_size=1,
+                worker_id='', pass_through=False, delegate_only=False,
+                transfer_format=_transfer_codec.DEFAULT_TRANSFER_FORMAT,
+                transfer_quality=_transfer_codec.DEFAULT_TRANSFER_QUALITY):
+        return io.NodeOutput(*_CollectorRuntime().run(
+            images=images, load_balance=load_balance, audio=audio, multi_job_id=multi_job_id,
+            is_worker=is_worker, master_url=master_url, enabled_worker_ids=enabled_worker_ids,
+            worker_batch_size=worker_batch_size, worker_id=worker_id,
+            pass_through=pass_through, delegate_only=delegate_only,
+            transfer_format=transfer_format, transfer_quality=transfer_quality))
+
+
+class UltimateSDUpscaleDistributed(io.ComfyNode):
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id='UltimateSDUpscaleDistributed',
+            display_name='Ultimate SD Upscale Distributed (No Upscale)', category='image/upscaling',
+            inputs=[
+                io.Image.Input('upscaled_image'), io.Model.Input('model'),
+                io.Conditioning.Input('positive'), io.Conditioning.Input('negative'), io.Vae.Input('vae'),
+                io.Int.Input('seed', default=0, min=0, max=0xffffffffffffffff),
+                io.Int.Input('steps', default=20, min=1, max=10000),
+                io.Float.Input('cfg', default=8.0, min=0.0, max=100.0),
+                io.Combo.Input('sampler_name', options=comfy.samplers.KSampler.SAMPLERS),
+                io.Combo.Input('scheduler', options=comfy.samplers.KSampler.SCHEDULERS),
+                io.Float.Input('denoise', default=0.5, min=0.0, max=1.0, step=0.01),
+                io.Int.Input('tile_width', default=512, min=64, max=2048, step=8),
+                io.Int.Input('tile_height', default=512, min=64, max=2048, step=8),
+                io.Int.Input('padding', default=32, min=0, max=256, step=8),
+                io.Int.Input('mask_blur', default=8, min=0, max=256),
+                io.Boolean.Input('force_uniform_tiles', default=True),
+                io.Boolean.Input('tiled_decode', default=False),
+            ], outputs=[io.Image.Output()], accept_all_inputs=True,
+        )
+
+    @classmethod
+    def fingerprint_inputs(cls, **kwargs):
+        return _UpscaleRuntime.IS_CHANGED(**kwargs)
+
+    @classmethod
+    def execute(cls, upscaled_image, model, positive, negative, vae, seed, steps, cfg,
+                sampler_name, scheduler, denoise, tile_width, tile_height, padding,
+                mask_blur, force_uniform_tiles, tiled_decode, multi_job_id='', is_worker=False,
+                master_url='', enabled_worker_ids='[]', worker_id='', tile_indices='', dynamic_threshold=8):
+        return io.NodeOutput(*_UpscaleRuntime().run(
+            upscaled_image, model, positive, negative, vae, seed, steps, cfg,
+            sampler_name, scheduler, denoise, tile_width, tile_height, padding,
+            mask_blur, force_uniform_tiles, tiled_decode, multi_job_id, is_worker,
+            master_url, enabled_worker_ids, worker_id, tile_indices, dynamic_threshold))
+
+
+class DistributedWorkerSink(io.ComfyNode):
+    """Internal output node that terminates pruned worker prompts.
+
+    Worker prompts drop everything downstream of the distributed node, which
+    leaves no output node for ComfyUI to execute towards. This sink provides
+    one without writing preview files (PreviewImage wrote every frame to the
+    worker's temp directory). Hidden from search/menus; inserted by the master.
+    """
+
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id='DistributedWorkerSink', display_name='Distributed Worker Sink (internal)', category='utils',
+            inputs=[io.AnyType.Input('value')],
+            is_output_node=True, is_dev_only=True,
+        )
+
+    @classmethod
+    def execute(cls, value):
+        return io.NodeOutput()
+
+
+NODES = [DistributedCollector, DistributedSeed, DistributedModelName, DistributedValue,
+         ImageBatchDivider, AudioBatchDivider, DistributedEmptyImage, UltimateSDUpscaleDistributed,
+         DistributedWorkerSink, DistributedImageListStart, DistributedImageListEnd, DistributedMediaCollector]
@@ -0,0 +1,12 @@
+{
+  "name": "comfyui-distributed-web-tests",
+  "private": true,
+  "type": "module",
+  "scripts": {
+    "test:web": "bash ./scripts/test-web.sh",
+    "test:web:watch": "bash ./scripts/test-web.sh --watch"
+  },
+  "devDependencies": {
+    "vitest": "^2.1.9"
+  }
+}
@@ -1,7 +1,7 @@
 [project]
 name = "ComfyUI-Distributed"
 description = "ComfyUI extension that enables multi-GPU processing locally, remotely and in the cloud"
-version = "1.1.0"
+version = "1.5.0"
 license = {file = "LICENSE"}
 dependencies  = []

@@ -13,3 +13,7 @@ Repository = "https://github.com/robertvoy/ComfyUI-Distributed"
 PublisherId = "robertvoy"
 DisplayName = "ComfyUI-Distributed"
 Icon = "https://raw.githubusercontent.com/robertvoy/ComfyUI-Distributed/refs/heads/main/web/distributed-logo-icon.png"
+
+[tool.pytest.ini_options]
+testpaths = ["tests"]
+pythonpath = ["."]
@@ -0,0 +1 @@
+"""Internal extension lifecycle support (not a node provider)."""
@@ -0,0 +1,38 @@
+"""Initialize routes, distributed state and the existing worker lifecycle."""
+import atexit
+import os
+
+import server
+
+from ..utils.config import CONFIG_FILE, ensure_config_exists
+from ..utils.logging import debug_log
+from ..workers.startup import delayed_auto_launch, register_async_signals, sync_cleanup
+from ..upscale.job_store import ensure_tile_jobs_initialized
+
+_initialized = False
+
+
+def initialize():
+    """Called by ComfyExtension.on_load; initialize once per loaded package."""
+    global _initialized
+    if _initialized:
+        return
+
+    ensure_config_exists()
+    from .. import api  # noqa: F401 - registers the existing @routes.* handlers
+    from ..api.queue_orchestration import ensure_distributed_state
+
+    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)
+        delayed_auto_launch()
+        register_async_signals()
+
+    _initialized = True
+    debug_log('Loaded Distributed nodes.')
+    debug_log(f'Config file: {CONFIG_FILE}')
@@ -0,0 +1,29 @@
+#!/usr/bin/env bash
+set -euo pipefail
+
+SCRIPT_DIR="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")" && pwd)"
+REPO_ROOT="$(cd -- "${SCRIPT_DIR}/.." && pwd)"
+
+export NVM_DIR="${NVM_DIR:-$HOME/.nvm}"
+if [[ -s "${NVM_DIR}/nvm.sh" ]]; then
+    # shellcheck source=/dev/null
+    . "${NVM_DIR}/nvm.sh"
+fi
+
+if ! command -v node >/dev/null 2>&1 || ! command -v npm >/dev/null 2>&1; then
+    echo "[test-web] node/npm are not available." >&2
+    echo "[test-web] Install nvm and Node, or ensure node/npm are on PATH." >&2
+    exit 1
+fi
+
+if [[ -f "${REPO_ROOT}/.nvmrc" ]] && command -v nvm >/dev/null 2>&1; then
+    nvm use >/dev/null
+fi
+
+cd "${REPO_ROOT}"
+
+if [[ "${1:-}" == "--watch" ]]; then
+    exec npx vitest web/tests
+fi
+
+exec npx vitest run web/tests
@@ -0,0 +1,251 @@
+# 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.
+
+For collector setup, list ordering, supported native VIDEO states, and PR2
+limitations, see [typed serverless media](../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 text request builder (no files or model required by this graph):
+
+```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. Defaults are `exact` and `90`; old exact requests
+omit the controls and retain their original representation. Other media exports
+must not carry image controls. The master compiler supplies them from the typed
+Collector's widgets; callers should not alter binding identity metadata.
+
+Formats: `exact`, `legacy_png`, `png`, `ffv1`, `jpeg`, `webp`, `h264`, `h265`, `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 still use their original exact codecs.
+
+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.
+
+Verification scope: source and offline tests. No live provider/GPU validation is
+claimed, and this directory is not a tested turnkey provider deployment.
@@ -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()
@@ -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
@@ -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
@@ -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()
@@ -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
@@ -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
@@ -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]})
@@ -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())
@@ -0,0 +1,147 @@
+import copy
+import importlib.util
+import sys
+import types
+import unittest
+from pathlib import Path
+from unittest.mock import patch
+
+
+class _FakeResponse:
+    def __init__(self, payload, status=200):
+        self.payload = payload
+        self.status = status
+
+
+class _FakeRequest:
+    def __init__(self, payload=None):
+        self._payload = payload
+
+    async def json(self):
+        return self._payload
+
+
+def _load_config_routes_module():
+    module_path = Path(__file__).resolve().parents[2] / "api" / "config_routes.py"
+    package_name = "dist_api_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__ = []
+    sys.modules[package_name] = root_pkg
+
+    api_pkg = types.ModuleType(f"{package_name}.api")
+    api_pkg.__path__ = []
+    sys.modules[f"{package_name}.api"] = api_pkg
+
+    utils_pkg = types.ModuleType(f"{package_name}.utils")
+    utils_pkg.__path__ = [str(module_path.parents[1] / "utils")]
+    sys.modules[f"{package_name}.utils"] = utils_pkg
+
+    import aiohttp  # Load the real dependency; isolate the fake response at the tested module.
+
+    class _Routes:
+        def get(self, _path):
+            def _decorator(fn):
+                return fn
+            return _decorator
+
+        def post(self, _path):
+            def _decorator(fn):
+                return fn
+            return _decorator
+
+    server_module = types.ModuleType("server")
+    server_module.PromptServer = types.SimpleNamespace(instance=types.SimpleNamespace(routes=_Routes()))
+    sys.modules["server"] = server_module
+
+    logging_module = types.ModuleType(f"{package_name}.utils.logging")
+    logging_module.debug_log = lambda *_args, **_kwargs: None
+    logging_module.log = lambda *_args, **_kwargs: None
+    sys.modules[f"{package_name}.utils.logging"] = logging_module
+
+    network_module = types.ModuleType(f"{package_name}.utils.network")
+
+    async def _handle_api_error(_request, error, status=500):
+        return _FakeResponse({"status": "error", "message": str(error)}, status=status)
+
+    network_module.handle_api_error = _handle_api_error
+    network_module.normalize_host = lambda value: value
+    sys.modules[f"{package_name}.utils.network"] = network_module
+
+    default_config = {
+        "workers": [],
+        "master": {"host": ""},
+        "settings": {"debug": False},
+        "tunnel": {},
+    }
+
+    config_module = types.ModuleType(f"{package_name}.utils.config")
+    config_module.load_config = lambda: copy.deepcopy(default_config)
+    config_module.save_config = lambda _cfg: True
+    sys.modules[f"{package_name}.utils.config"] = config_module
+
+    spec = importlib.util.spec_from_file_location(f"{package_name}.api.config_routes", 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)
+
+    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
+
+
+config_routes = _load_config_routes_module()
+
+
+class ConfigRoutesTests(unittest.IsolatedAsyncioTestCase):
+    async def test_get_config_returns_core_sections(self):
+        cfg = {"workers": [], "master": {}, "settings": {}, "tunnel": {}}
+        with patch.object(config_routes, "load_config", return_value=cfg):
+            response = await config_routes.get_config_endpoint(_FakeRequest())
+
+        self.assertEqual(response.status, 200)
+        self.assertIn("workers", response.payload)
+        self.assertIn("master", response.payload)
+        self.assertIn("settings", response.payload)
+
+    async def test_update_config_valid_field_persists(self):
+        cfg = {"workers": [], "master": {}, "settings": {"debug": False}, "tunnel": {}}
+        with patch.object(config_routes, "load_config", return_value=cfg), patch.object(
+            config_routes, "save_config", return_value=True
+        ):
+            response = await config_routes.update_config_endpoint(_FakeRequest({"debug": True}))
+
+        self.assertEqual(response.status, 200)
+        self.assertEqual(response.payload["status"], "success")
+        self.assertTrue(response.payload["config"]["settings"]["debug"])
+
+    async def test_update_config_unknown_field_returns_400(self):
+        cfg = {"workers": [], "master": {}, "settings": {"debug": False}, "tunnel": {}}
+        with patch.object(config_routes, "load_config", return_value=cfg):
+            response = await config_routes.update_config_endpoint(_FakeRequest({"unknown_field": 1}))
+
+        self.assertEqual(response.status, 400)
+        self.assertIn("unknown_field", " ".join(response.payload.get("error", [])).lower())
+
+    async def test_update_config_wrong_type_returns_400(self):
+        cfg = {"workers": [], "master": {}, "settings": {"debug": False}, "tunnel": {}}
+        with patch.object(config_routes, "load_config", return_value=cfg):
+            response = await config_routes.update_config_endpoint(_FakeRequest({"debug": "true"}))
+
+        self.assertEqual(response.status, 400)
+        self.assertIn("debug", " ".join(response.payload.get("error", [])).lower())
+
+
+if __name__ == "__main__":
+    unittest.main()
@@ -0,0 +1,555 @@
+import importlib.util
+import sys
+import types
+import unittest
+import asyncio
+import base64
+import io
+import json
+from dataclasses import dataclass
+from pathlib import Path
+from unittest.mock import AsyncMock, patch
+
+import numpy as np
+import torch
+
+
+class _FakeResponse:
+    def __init__(self, payload, status=200):
+        self.payload = payload
+        self.status = status
+
+
+class _FakeRequest:
+    def __init__(self, payload):
+        self._payload = payload
+
+    async def json(self):
+        return self._payload
+
+
+def _register_real_transfer_codec(package_name, utils_pkg):
+    """Load the real utils/transfer_codec.py into the synthetic test package."""
+    codec_path = Path(__file__).resolve().parents[2] / "utils" / "transfer_codec.py"
+    codec_spec = importlib.util.spec_from_file_location(f"{package_name}.utils.transfer_codec", codec_path)
+    codec_module = importlib.util.module_from_spec(codec_spec)
+    sys.modules[f"{package_name}.utils.transfer_codec"] = codec_module
+    codec_spec.loader.exec_module(codec_module)
+    utils_pkg.transfer_codec = codec_module
+    return codec_module
+
+
+def _load_job_routes_module():
+    module_path = Path(__file__).resolve().parents[2] / "api" / "job_routes.py"
+    package_name = "dist_api_queue_testpkg"
+
+    # Reset package namespace to avoid stale module state across test runs.
+    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__ = []
+    sys.modules[package_name] = root_pkg
+
+    api_pkg = types.ModuleType(f"{package_name}.api")
+    api_pkg.__path__ = []
+    sys.modules[f"{package_name}.api"] = api_pkg
+
+    utils_pkg = types.ModuleType(f"{package_name}.utils")
+    utils_pkg.__path__ = []
+    sys.modules[f"{package_name}.utils"] = utils_pkg
+    _register_real_transfer_codec(package_name, utils_pkg)
+
+    # aiohttp.web stub
+    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
+
+    # server module stub with route decorators
+    class _Routes:
+        def get(self, _path):
+            def _decorator(fn):
+                return fn
+            return _decorator
+
+        def post(self, _path):
+            def _decorator(fn):
+                return fn
+            return _decorator
+
+    prompt_server_instance = types.SimpleNamespace(
+        routes=_Routes(),
+        distributed_jobs_lock=None,
+        distributed_pending_jobs={},
+    )
+    server_module = types.ModuleType("server")
+    server_module.PromptServer = types.SimpleNamespace(instance=prompt_server_instance)
+    sys.modules["server"] = server_module
+
+    # torch stub (only needed to satisfy import)
+    created_torch_stub = False
+    if "torch" not in sys.modules:
+        created_torch_stub = True
+        torch_module = types.ModuleType("torch")
+        torch_module.cuda = types.SimpleNamespace(
+            is_available=lambda: False,
+            empty_cache=lambda: None,
+            ipc_collect=lambda: None,
+        )
+        sys.modules["torch"] = torch_module
+
+    # PIL stub (only needed to satisfy import)
+    created_pil_stub = False
+    if "PIL" not in sys.modules:
+        created_pil_stub = True
+        pil_module = types.ModuleType("PIL")
+        image_module = types.ModuleType("PIL.Image")
+        pil_module.Image = image_module
+        sys.modules["PIL"] = pil_module
+        sys.modules["PIL.Image"] = image_module
+
+    logging_module = types.ModuleType(f"{package_name}.utils.logging")
+    logging_module.debug_log = lambda *_args, **_kwargs: None
+    logging_module.log = lambda *_args, **_kwargs: None
+    sys.modules[f"{package_name}.utils.logging"] = logging_module
+
+    image_module = types.ModuleType(f"{package_name}.utils.image")
+    image_module.pil_to_tensor = lambda *_args, **_kwargs: None
+    image_module.ensure_contiguous = lambda tensor: tensor
+    sys.modules[f"{package_name}.utils.image"] = image_module
+
+    audio_payload_module = types.ModuleType(f"{package_name}.utils.audio_payload")
+
+    def _decode_audio_payload(payload):
+        if payload is None:
+            return None
+        if not isinstance(payload, dict):
+            raise ValueError("Field 'audio' must be an object when provided.")
+
+        encoded = payload.get("data")
+        shape = payload.get("shape")
+        dtype = payload.get("dtype", "float32")
+        sample_rate = payload.get("sample_rate", 44100)
+        if not isinstance(encoded, str) or not encoded.strip():
+            raise ValueError("Field 'audio.data' must be a non-empty base64 string.")
+        if not isinstance(shape, list) or len(shape) != 3:
+            raise ValueError("Field 'audio.shape' must be a 3-item list.")
+        if dtype != "float32":
+            raise ValueError("Field 'audio.dtype' must be 'float32'.")
+        try:
+            shape_tuple = tuple(int(dim) for dim in shape)
+        except Exception as exc:
+            raise ValueError("Field 'audio.shape' must contain integers.") from exc
+
+        raw = base64.b64decode(encoded, validate=True)
+        expected_bytes = int(np.prod(shape_tuple, dtype=np.int64)) * 4
+        if len(raw) != expected_bytes:
+            raise ValueError("Field 'audio.data' byte size mismatch.")
+
+        waveform = torch.from_numpy(np.frombuffer(raw, dtype=np.float32).reshape(shape_tuple).copy())
+        return {"waveform": waveform, "sample_rate": int(sample_rate)}
+
+    audio_payload_module.decode_audio_payload = _decode_audio_payload
+    sys.modules[f"{package_name}.utils.audio_payload"] = audio_payload_module
+
+    network_module = types.ModuleType(f"{package_name}.utils.network")
+
+    async def _handle_api_error(_request, error, status=500):
+        return _FakeResponse({"status": "error", "message": str(error)}, status=status)
+
+    network_module.handle_api_error = _handle_api_error
+    sys.modules[f"{package_name}.utils.network"] = network_module
+
+    constants_module = types.ModuleType(f"{package_name}.utils.constants")
+    constants_module.MEMORY_CLEAR_DELAY = 0.0
+    constants_module.JOB_INIT_GRACE_PERIOD = 10.0
+    sys.modules[f"{package_name}.utils.constants"] = constants_module
+
+    async_helpers_module = types.ModuleType(f"{package_name}.utils.async_helpers")
+    async_helpers_module.queue_prompt_payload = AsyncMock(return_value="prompt_local")
+    sys.modules[f"{package_name}.utils.async_helpers"] = async_helpers_module
+
+    queue_orchestration_module = types.ModuleType(f"{package_name}.api.queue_orchestration")
+    queue_orchestration_module.orchestrate_distributed_execution = AsyncMock(return_value=("prompt_dist", 7, 1, {}))
+    sys.modules[f"{package_name}.api.queue_orchestration"] = queue_orchestration_module
+
+    @dataclass(frozen=True)
+    class _QueuePayload:
+        prompt: dict
+        workflow_meta: object
+        client_id: str
+        delegate_master: object
+        enabled_worker_ids: list
+        auto_prepare: bool
+        trace_execution_id: object
+
+    def _parse_queue_request_payload(data):
+        if not isinstance(data, dict):
+            raise ValueError("Expected a JSON object body")
+        prompt = data.get("prompt")
+        if not isinstance(prompt, dict):
+            raise ValueError("Field 'prompt' must be an object")
+        enabled = data.get("enabled_worker_ids")
+        if not isinstance(enabled, list):
+            raise ValueError("enabled_worker_ids required")
+        client_id = data.get("client_id")
+        if not isinstance(client_id, str) or not client_id.strip():
+            raise ValueError("client_id required")
+        return _QueuePayload(
+            prompt=prompt,
+            workflow_meta=data.get("workflow"),
+            client_id=client_id,
+            delegate_master=data.get("delegate_master"),
+            enabled_worker_ids=enabled,
+            auto_prepare=bool(data.get("auto_prepare", True)),
+            trace_execution_id=data.get("trace_execution_id"),
+        )
+
+    queue_request_module = types.ModuleType(f"{package_name}.api.queue_request")
+    queue_request_module.parse_queue_request_payload = _parse_queue_request_payload
+    sys.modules[f"{package_name}.api.queue_request"] = queue_request_module
+
+    spec = importlib.util.spec_from_file_location(f"{package_name}.api.job_routes", 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)
+    # 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)
+    if created_torch_stub:
+        sys.modules.pop("torch", None)
+    if created_pil_stub:
+        sys.modules.pop("PIL.Image", None)
+        sys.modules.pop("PIL", None)
+
+    return module
+
+
+job_routes = _load_job_routes_module()
+
+
+class DistributedQueueEndpointTests(unittest.IsolatedAsyncioTestCase):
+    async def test_distributed_queue_happy_path_returns_prompt_metadata(self):
+        request = _FakeRequest(
+            {
+                "prompt": {"1": {"class_type": "Node"}},
+                "enabled_worker_ids": ["w1"],
+                "client_id": "client-1",
+                "auto_prepare": True,
+            }
+        )
+        with patch.object(
+            job_routes,
+            "orchestrate_distributed_execution",
+            new=AsyncMock(return_value=("prompt_123", 42, 2, {})),
+        ):
+            response = await job_routes.distributed_queue_endpoint(request)
+
+        self.assertEqual(response.status, 200)
+        self.assertEqual(response.payload.get("prompt_id"), "prompt_123")
+        self.assertEqual(response.payload.get("number"), 42)
+        self.assertEqual(response.payload.get("node_errors"), {})
+        self.assertTrue(response.payload.get("auto_prepare_supported"))
+
+    async def test_distributed_queue_missing_prompt_returns_400(self):
+        request = _FakeRequest(
+            {
+                "enabled_worker_ids": ["w1"],
+                "client_id": "client-1",
+            }
+        )
+        response = await job_routes.distributed_queue_endpoint(request)
+        self.assertEqual(response.status, 400)
+        self.assertIn("prompt", response.payload.get("message", "").lower())
+
+    async def test_distributed_queue_missing_enabled_worker_ids_returns_400(self):
+        request = _FakeRequest(
+            {
+                "prompt": {"1": {"class_type": "Node"}},
+                "client_id": "client-1",
+            }
+        )
+        response = await job_routes.distributed_queue_endpoint(request)
+        self.assertEqual(response.status, 400)
+        self.assertIn("enabled_worker_ids", response.payload.get("message", "").lower())
+
+
+class JobCompleteAudioPayloadTests(unittest.IsolatedAsyncioTestCase):
+    def _encoded_audio_payload(self):
+        waveform = np.arange(8, dtype=np.float32).reshape(1, 2, 4)
+        return {
+            "sample_rate": 44100,
+            "shape": [1, 2, 4],
+            "dtype": "float32",
+            "data": base64.b64encode(waveform.tobytes()).decode("ascii"),
+        }
+
+    async def test_job_complete_accepts_audio_payload(self):
+        queue = asyncio.Queue()
+        job_routes.prompt_server.distributed_jobs_lock = asyncio.Lock()
+        job_routes.prompt_server.distributed_pending_jobs = {"job-1": queue}
+        request = _FakeRequest(
+            {
+                "job_id": "job-1",
+                "worker_id": "worker-1",
+                "batch_idx": 0,
+                "image": "data:image/png;base64,AAAA",
+                "audio": self._encoded_audio_payload(),
+                "is_last": True,
+            }
+        )
+
+        with patch.object(job_routes, "_decode_canonical_png_tensor", return_value="tensor-data"):
+            response = await job_routes.job_complete_endpoint(request)
+
+        self.assertEqual(response.status, 200)
+        queued = await queue.get()
+        self.assertEqual(queued["worker_id"], "worker-1")
+        self.assertTrue(queued["is_last"])
+        self.assertIsNotNone(queued["audio"])
+        self.assertEqual(queued["audio"]["sample_rate"], 44100)
+        self.assertEqual(tuple(queued["audio"]["waveform"].shape), (1, 2, 4))
+
+    async def test_job_complete_accepts_audio_without_image(self):
+        queue = asyncio.Queue()
+        job_routes.prompt_server.distributed_jobs_lock = asyncio.Lock()
+        job_routes.prompt_server.distributed_pending_jobs = {"audio-only-job": queue}
+        request = _FakeRequest(
+            {
+                "job_id": "audio-only-job",
+                "worker_id": "worker-1",
+                "batch_idx": 0,
+                "audio": self._encoded_audio_payload(),
+                "is_last": True,
+            }
+        )
+
+        with patch.object(job_routes, "_decode_canonical_png_tensor") as decode_image:
+            response = await job_routes.job_complete_endpoint(request)
+
+        self.assertEqual(response.status, 200)
+        decode_image.assert_not_called()
+        queued = await queue.get()
+        self.assertIsNone(queued["tensor"])
+        self.assertEqual(queued["audio"]["sample_rate"], 44100)
+
+    async def test_job_complete_rejects_payload_without_image_or_audio(self):
+        request = _FakeRequest(
+            {
+                "job_id": "job-1",
+                "worker_id": "worker-1",
+                "batch_idx": 0,
+                "is_last": True,
+            }
+        )
+
+        response = await job_routes.job_complete_endpoint(request)
+
+        self.assertEqual(response.status, 400)
+        self.assertIn("image or audio", response.payload.get("message", "").lower())
+
+    async def test_job_complete_rejects_invalid_image_even_when_audio_is_present(self):
+        request = _FakeRequest(
+            {
+                "job_id": "job-1",
+                "worker_id": "worker-1",
+                "batch_idx": 0,
+                "image": 123,
+                "audio": self._encoded_audio_payload(),
+                "is_last": True,
+            }
+        )
+
+        response = await job_routes.job_complete_endpoint(request)
+
+        self.assertEqual(response.status, 400)
+        self.assertIn("image", response.payload.get("message", "").lower())
+
+    async def test_job_complete_returns_400_for_malformed_audio(self):
+        request = _FakeRequest(
+            {
+                "job_id": "job-1",
+                "worker_id": "worker-1",
+                "batch_idx": 0,
+                "audio": {"data": "AAAA", "shape": [1, 2], "dtype": "float32"},
+                "is_last": True,
+            }
+        )
+
+        response = await job_routes.job_complete_endpoint(request)
+
+        self.assertEqual(response.status, 400)
+        self.assertIn("audio.shape", response.payload.get("message", "").lower())
+
+    def test_decode_audio_payload_rejects_bad_shape(self):
+        bad = {
+            "sample_rate": 44100,
+            "shape": [1, 2],
+            "dtype": "float32",
+            "data": base64.b64encode(b"\x00\x00\x00\x00").decode("ascii"),
+        }
+        with self.assertRaises(ValueError):
+            job_routes._decode_audio_payload(bad)
+
+    def test_decode_audio_payload_rejects_bad_dtype(self):
+        payload = {
+            "sample_rate": 44100,
+            "shape": [1, 2, 4],
+            "dtype": "float16",
+            "data": base64.b64encode((np.zeros((1, 2, 4), dtype=np.float32)).tobytes()).decode("ascii"),
+        }
+        with self.assertRaises(ValueError):
+            job_routes._decode_audio_payload(payload)
+
+
+class _FakeFileField:
+    def __init__(self, data):
+        self.file = io.BytesIO(data)
+
+
+class _FakeMultipartRequest:
+    content_type = "multipart/form-data"
+
+    def __init__(self, envelope, parts):
+        from multidict import MultiDict
+
+        fields = [("envelope", envelope if isinstance(envelope, str) else json.dumps(envelope))]
+        fields += [("media", _FakeFileField(part)) for part in parts]
+        self._form = MultiDict(fields)
+
+    async def post(self):
+        return self._form
+
+    async def json(self):
+        raise ValueError("not JSON")
+
+
+class JobCompleteCompressedPayloadTests(unittest.IsolatedAsyncioTestCase):
+    codec = job_routes.transfer_codec
+
+    def setUp(self):
+        self.queue = asyncio.Queue()
+        job_routes.prompt_server.distributed_jobs_lock = asyncio.Lock()
+        job_routes.prompt_server.distributed_pending_jobs = {"job-z": self.queue}
+
+    def _frames(self, count=4, height=20, width=30):
+        rng = np.random.default_rng(1)
+        return rng.integers(0, 256, (count, height, width, 3), dtype=np.uint8)
+
+    def _request_for(self, frames, fmt, *, start_offset=0, is_last=True, audio=None, **overrides):
+        chunk = next(self.codec.encode_chunks(frames, fmt, 90, 10**9))
+        parts = chunk.pop("parts")
+        envelope = {
+            "job_id": "job-z",
+            "worker_id": "worker-1",
+            "is_last": is_last,
+            **chunk,
+            "start_index": chunk["start_index"] + start_offset,
+        }
+        if audio is not None:
+            envelope["audio"] = audio
+        envelope.update(overrides)
+        return _FakeMultipartRequest(envelope, parts)
+
+    async def _drain(self):
+        items = []
+        while not self.queue.empty():
+            items.append(await self.queue.get())
+        return items
+
+    async def test_lossless_chunk_enqueues_one_item_per_frame_bit_exact(self):
+        frames = self._frames()
+        response = await job_routes.job_complete_endpoint(self._request_for(frames, "ffv1", start_offset=10))
+
+        self.assertEqual(response.status, 200)
+        items = await self._drain()
+        self.assertEqual([item["image_index"] for item in items], [10, 11, 12, 13])
+        self.assertEqual([item["is_last"] for item in items], [False, False, False, True])
+        self.assertEqual({item["worker_id"] for item in items}, {"worker-1"})
+        for item, frame in zip(items, frames):
+            self.assertEqual(tuple(item["tensor"].shape), (1, 20, 30, 3))
+            self.assertEqual(item["tensor"].dtype, torch.float32)
+            expected = frame.astype(np.float32) / 255.0
+            self.assertTrue(np.array_equal(item["tensor"][0].numpy(), expected))
+
+    async def test_non_final_chunk_never_marks_worker_done(self):
+        response = await job_routes.job_complete_endpoint(
+            self._request_for(self._frames(count=2), "png", is_last=False)
+        )
+
+        self.assertEqual(response.status, 200)
+        self.assertFalse(any(item["is_last"] for item in await self._drain()))
+
+    async def test_audio_is_attached_only_to_final_frame(self):
+        audio = JobCompleteAudioPayloadTests._encoded_audio_payload(None)
+        response = await job_routes.job_complete_endpoint(
+            self._request_for(self._frames(count=3), "jpeg", audio=audio)
+        )
+
+        self.assertEqual(response.status, 200)
+        items = await self._drain()
+        self.assertEqual([item["audio"] is not None for item in items], [False, False, True])
+        self.assertEqual(tuple(items[-1]["audio"]["waveform"].shape), (1, 2, 4))
+
+    async def test_rejects_frame_count_mismatch(self):
+        response = await job_routes.job_complete_endpoint(
+            self._request_for(self._frames(count=2), "ffv1", count=5)
+        )
+
+        self.assertEqual(response.status, 400)
+        self.assertIn("frames", response.payload["message"])
+        self.assertTrue(self.queue.empty())
+
+    async def test_rejects_bad_envelope(self):
+        response = await job_routes.job_complete_endpoint(
+            self._request_for(self._frames(count=1), "png", format="gif", is_last="yes")
+        )
+
+        self.assertEqual(response.status, 400)
+        self.assertIn("format", response.payload["message"])
+        self.assertIn("is_last", response.payload["message"])
+
+    async def test_rejects_invalid_envelope_json(self):
+        response = await job_routes.job_complete_endpoint(_FakeMultipartRequest("{not json", [b"x"]))
+
+        self.assertEqual(response.status, 400)
+        self.assertIn("envelope", response.payload["message"])
+
+    async def test_unknown_job_returns_404(self):
+        job_routes.prompt_server.distributed_pending_jobs = {}
+        with patch.object(job_routes, "JOB_INIT_GRACE_PERIOD", 0.0):
+            response = await job_routes.job_complete_endpoint(self._request_for(self._frames(count=1), "png"))
+
+        self.assertEqual(response.status, 404)
+
+    async def test_legacy_json_payload_still_accepted(self):
+        request = _FakeRequest(
+            {
+                "job_id": "job-z",
+                "worker_id": "worker-1",
+                "batch_idx": 3,
+                "image": "data:image/png;base64,AAAA",
+                "is_last": True,
+            }
+        )
+        with patch.object(job_routes, "_decode_canonical_png_tensor", return_value="tensor-data"):
+            response = await job_routes.job_complete_endpoint(request)
+
+        self.assertEqual(response.status, 200)
+        items = await self._drain()
+        self.assertEqual(len(items), 1)
+        self.assertEqual((items[0]["image_index"], items[0]["tensor"]), (3, "tensor-data"))
+
+
+if __name__ == "__main__":
+    unittest.main()
@@ -0,0 +1,363 @@
+import importlib.util
+import sys
+import types
+import unittest
+from pathlib import Path
+
+
+def _load_media_sync_module():
+    module_path = Path(__file__).resolve().parents[2] / "api" / "orchestration" / "media_sync.py"
+    package_name = "dist_ms_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__ = []
+    sys.modules[package_name] = root_pkg
+
+    api_pkg = types.ModuleType(f"{package_name}.api")
+    api_pkg.__path__ = []
+    sys.modules[f"{package_name}.api"] = api_pkg
+
+    orch_pkg = types.ModuleType(f"{package_name}.api.orchestration")
+    orch_pkg.__path__ = []
+    sys.modules[f"{package_name}.api.orchestration"] = orch_pkg
+
+    utils_pkg = types.ModuleType(f"{package_name}.utils")
+    utils_pkg.__path__ = []
+    sys.modules[f"{package_name}.utils"] = utils_pkg
+
+    logging_module = types.ModuleType(f"{package_name}.utils.logging")
+    logging_module.debug_log = lambda *_args, **_kwargs: None
+    logging_module.log = lambda *_args, **_kwargs: None
+    sys.modules[f"{package_name}.utils.logging"] = logging_module
+
+    network_module = types.ModuleType(f"{package_name}.utils.network")
+    network_module.build_worker_url = lambda worker, endpoint="": f"http://localhost{endpoint}"
+
+    async def _fake_session():
+        raise RuntimeError("network calls not used in pure-function tests")
+
+    network_module.get_client_session = _fake_session
+    sys.modules[f"{package_name}.utils.network"] = network_module
+
+    trace_module = types.ModuleType(f"{package_name}.utils.trace_logger")
+    trace_module.trace_debug = lambda *_args, **_kwargs: None
+    trace_module.trace_info = lambda *_args, **_kwargs: None
+    sys.modules[f"{package_name}.utils.trace_logger"] = trace_module
+
+    created_aiohttp_stub = False
+    if "aiohttp" not in sys.modules:
+        created_aiohttp_stub = True
+        aiohttp_module = types.ModuleType("aiohttp")
+
+        class _ClientTimeout:
+            def __init__(self, total=None):
+                pass
+
+        class _FormData:
+            def add_field(self, *args, **kwargs):
+                pass
+
+        aiohttp_module.ClientTimeout = _ClientTimeout
+        aiohttp_module.FormData = _FormData
+        sys.modules["aiohttp"] = aiohttp_module
+
+    spec = importlib.util.spec_from_file_location(
+        f"{package_name}.api.orchestration.media_sync",
+        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)
+
+    if created_aiohttp_stub:
+        sys.modules.pop("aiohttp", None)
+
+    return module
+
+
+ms = _load_media_sync_module()
+
+
+# ---------------------------------------------------------------------------
+# convert_paths_for_platform
+# ---------------------------------------------------------------------------
+
+class ConvertPathsForPlatformTests(unittest.TestCase):
+    def test_forward_slash_target_normalises_backslashes(self):
+        obj = {"ckpt_name": "C:\\Models\\model.safetensors"}
+        result = ms.convert_paths_for_platform(obj, "/")
+        self.assertEqual(result["ckpt_name"], "C:/Models/model.safetensors")
+
+    def test_backslash_target_normalises_forward_slashes(self):
+        obj = {"ckpt_name": "/models/checkpoints/model.safetensors"}
+        result = ms.convert_paths_for_platform(obj, "\\")
+        self.assertIn("\\", result["ckpt_name"])
+        self.assertNotIn("/", result["ckpt_name"])
+
+    def test_relative_media_paths_always_stay_forward_slash(self):
+        """Relative image/video/audio paths (Comfy annotated style) must not be backslash-ified."""
+        obj = {"image": "subfolder/my_photo.png"}
+        result = ms.convert_paths_for_platform(obj, "\\")
+        self.assertEqual(result["image"], "subfolder/my_photo.png")
+
+    def test_relative_audio_paths_stay_forward_slash(self):
+        obj = {"audio": "subfolder/my_track.wav"}
+        result = ms.convert_paths_for_platform(obj, "\\")
+        self.assertEqual(result["audio"], "subfolder/my_track.wav")
+
+    def test_annotated_relative_media_path_stays_forward_slash(self):
+        obj = {"image": "input/frame.jpg [abc123]"}
+        result = ms.convert_paths_for_platform(obj, "\\")
+        self.assertIn("/", result["image"])
+        self.assertNotIn("\\", result["image"].split("[")[0])
+
+    def test_non_filename_strings_are_untouched(self):
+        obj = {"prompt": "a beautiful cat", "count": 5}
+        result = ms.convert_paths_for_platform(obj, "\\")
+        self.assertEqual(result["prompt"], "a beautiful cat")
+        self.assertEqual(result["count"], 5)
+
+    def test_url_strings_are_untouched(self):
+        obj = {"url": "https://example.com/model.safetensors"}
+        result = ms.convert_paths_for_platform(obj, "\\")
+        self.assertEqual(result["url"], "https://example.com/model.safetensors")
+
+    def test_invalid_separator_returns_obj_unchanged(self):
+        obj = {"ckpt_name": "/models/model.safetensors"}
+        result = ms.convert_paths_for_platform(obj, "|")
+        self.assertEqual(result, obj)
+
+    def test_nested_dict_is_processed_recursively(self):
+        obj = {"node": {"ckpt_name": "C:\\Models\\model.safetensors"}}
+        result = ms.convert_paths_for_platform(obj, "/")
+        self.assertEqual(result["node"]["ckpt_name"], "C:/Models/model.safetensors")
+
+    def test_list_items_are_processed_recursively(self):
+        obj = [{"ckpt_name": "C:\\Models\\model.safetensors"}, "plain string"]
+        result = ms.convert_paths_for_platform(obj, "/")
+        self.assertEqual(result[0]["ckpt_name"], "C:/Models/model.safetensors")
+        self.assertEqual(result[1], "plain string")
+
+    def test_non_string_scalar_values_are_untouched(self):
+        obj = {"seed": 42, "enabled": True, "ratio": 1.5}
+        result = ms.convert_paths_for_platform(obj, "/")
+        self.assertEqual(result["seed"], 42)
+        self.assertTrue(result["enabled"])
+
+    def test_absolute_unix_path_to_windows(self):
+        obj = {"lora": "/home/user/loras/my_lora.safetensors"}
+        result = ms.convert_paths_for_platform(obj, "\\")
+        self.assertNotIn("/", result["lora"])
+
+    def test_already_normalised_path_is_idempotent(self):
+        obj = {"ckpt": "C:/Models/model.safetensors"}
+        result = ms.convert_paths_for_platform(obj, "/")
+        self.assertEqual(result["ckpt"], "C:/Models/model.safetensors")
+
+
+# ---------------------------------------------------------------------------
+# _find_media_references
+# ---------------------------------------------------------------------------
+
+class FindMediaReferencesTests(unittest.TestCase):
+    def test_finds_file_input_for_load_video(self):
+        prompt = {"1": {"class_type": "LoadVideo", "inputs": {"file": "1 - Copy.mp4"}}}
+        refs = ms._find_media_references(prompt)
+        self.assertIn("1 - Copy.mp4", refs)
+
+    def test_strips_annotation_suffix(self):
+        prompt = {"1": {"class_type": "LoadImage", "inputs": {"image": "photo.jpg [abc123]"}}}
+        refs = ms._find_media_references(prompt)
+        self.assertIn("photo.jpg", refs)
+        self.assertFalse(any("[" in r for r in refs))
+
+    def test_normalises_backslashes_in_path(self):
+        prompt = {"1": {"class_type": "LoadImage", "inputs": {"image": "sub\\img.png"}}}
+        refs = ms._find_media_references(prompt)
+        self.assertIn("sub/img.png", refs)
+
+    def test_ignores_non_media_text_inputs(self):
+        prompt = {"1": {"class_type": "CLIPTextEncode", "inputs": {"text": "a cat"}}}
+        refs = ms._find_media_references(prompt)
+        self.assertEqual(refs, [])
+
+    def test_ignores_node_link_values(self):
+        """Inputs that are [node_id, slot] lists should be ignored."""
+        prompt = {"1": {"class_type": "Anything", "inputs": {"image": ["2", 0]}}}
+        refs = ms._find_media_references(prompt)
+        self.assertEqual(refs, [])
+
+    def test_deduplicates_same_file_across_nodes(self):
+        prompt = {
+            "1": {"class_type": "LoadImage", "inputs": {"image": "cat.png"}},
+            "2": {"class_type": "LoadImage", "inputs": {"image": "cat.png"}},
+        }
+        refs = ms._find_media_references(prompt)
+        self.assertEqual(len(refs), 1)
+
+    def test_returns_sorted_list(self):
+        prompt = {
+            "1": {"class_type": "LoadImage", "inputs": {"image": "z_image.png"}},
+            "2": {"class_type": "LoadImage", "inputs": {"image": "a_image.jpg"}},
+        }
+        refs = ms._find_media_references(prompt)
+        self.assertEqual(refs, sorted(refs))
+
+    def test_ignores_non_dict_nodes(self):
+        prompt = {"1": "not a node dict", "2": {"class_type": "LoadImage", "inputs": {"image": "img.png"}}}
+        refs = ms._find_media_references(prompt)
+        self.assertIn("img.png", refs)
+
+    def test_empty_prompt_returns_empty_list(self):
+        self.assertEqual(ms._find_media_references({}), [])
+
+    def test_multiple_media_types_all_found(self):
+        prompt = {
+            "1": {"class_type": "LoadImage", "inputs": {"image": "frame.png"}},
+            "2": {"class_type": "LoadVideo", "inputs": {"video": "clip.mp4"}},
+            "3": {"class_type": "LoadAudio", "inputs": {"audio": "track.wav"}},
+        }
+        refs = ms._find_media_references(prompt)
+        self.assertIn("frame.png", refs)
+        self.assertIn("clip.mp4", refs)
+        self.assertIn("track.wav", refs)
+
+
+class RewritePromptMediaInputsTests(unittest.TestCase):
+    def test_rewrites_video_file_input_to_worker_path(self):
+        prompt = {
+            "79": {"class_type": "LoadVideo", "inputs": {"file": "1 - Copy.mp4"}},
+        }
+        ms._rewrite_prompt_media_inputs(prompt, {"1 - Copy.mp4": "videos/1 - Copy.mp4"})
+        self.assertEqual(prompt["79"]["inputs"]["file"], "videos/1 - Copy.mp4")
+
+    def test_rewrites_audio_input_and_strips_annotation_when_matching(self):
+        prompt = {
+            "1": {"class_type": "LoadAudio", "inputs": {"audio": "song.wav [input]"}},
+        }
+        ms._rewrite_prompt_media_inputs(prompt, {"song.wav": "song.wav"})
+        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()
+
+
+# ---------------------------------------------------------------------------
+# fetch_worker_has_node
+# ---------------------------------------------------------------------------
+
+class _FakeResponse:
+    def __init__(self, status, payload):
+        self.status = status
+        self._payload = payload
+
+    async def json(self):
+        if isinstance(self._payload, Exception):
+            raise self._payload
+        return self._payload
+
+    async def __aenter__(self):
+        return self
+
+    async def __aexit__(self, *exc):
+        return False
+
+
+class _FakeSession:
+    def __init__(self, response=None, error=None):
+        self.response = response
+        self.error = error
+        self.urls = []
+
+    def get(self, url, timeout=None):
+        self.urls.append(url)
+        if self.error:
+            raise self.error
+        return self.response
+
+
+class FetchWorkerHasNodeTests(unittest.TestCase):
+    def _probe(self, session):
+        import asyncio
+
+        async def _get_session():
+            return session
+
+        original = ms.get_client_session
+        ms.get_client_session = _get_session
+        try:
+            return asyncio.run(ms.fetch_worker_has_node({"id": "w1"}, "DistributedWorkerSink"))
+        finally:
+            ms.get_client_session = original
+
+    def test_registered_node_is_reported(self):
+        session = _FakeSession(_FakeResponse(200, {"DistributedWorkerSink": {"name": "DistributedWorkerSink"}}))
+        self.assertTrue(self._probe(session))
+        self.assertEqual(session.urls, ["http://localhost/object_info/DistributedWorkerSink"])
+
+    def test_unknown_node_returns_empty_object(self):
+        # ComfyUI answers 200 {} for classes that are not registered.
+        self.assertFalse(self._probe(_FakeSession(_FakeResponse(200, {}))))
+
+    def test_http_error_counts_as_missing(self):
+        self.assertFalse(self._probe(_FakeSession(_FakeResponse(404, None))))
+
+    def test_bad_json_counts_as_missing(self):
+        self.assertFalse(self._probe(_FakeSession(_FakeResponse(200, ValueError("bad json")))))
+
+    def test_connection_error_counts_as_missing(self):
+        self.assertFalse(self._probe(_FakeSession(error=OSError("unreachable"))))
@@ -0,0 +1,219 @@
+"""End-to-end: real worker send path -> real aiohttp server -> job_complete endpoint."""
+import asyncio
+import importlib.util
+import sys
+from pathlib import Path
+
+import numpy as np
+import pytest
+import torch
+
+_TESTS_DIR = Path(__file__).resolve().parents[1]
+
+
+def _load_from(path, name):
+    spec = importlib.util.spec_from_file_location(name, path)
+    module = importlib.util.module_from_spec(spec)
+    spec.loader.exec_module(module)
+    return module
+
+
+def _load_modules():
+    real_aiohttp = sys.modules["aiohttp"]  # imported inside the isolated test
+    queue_tests = _load_from(_TESTS_DIR / "api" / "test_distributed_queue.py", "dist_transfer_it_queue")
+    collector_tests = _load_from(_TESTS_DIR / "test_collector_list_inputs.py", "dist_transfer_it_collector")
+    job_routes = queue_tests._load_job_routes_module()
+    collector = collector_tests._load_collector_module()
+
+    # Error responses must be real HTTP responses here (the unit-test harness
+    # returns plain objects); mirror utils.network.handle_api_error.
+    from aiohttp import web
+
+    async def handle_api_error(_request, error, status=500):
+        if isinstance(error, list):
+            return web.json_response({"errors": [str(item) for item in error]}, status=status)
+        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
+    return job_routes, collector
+
+
+@pytest.fixture(autouse=True)
+def _isolate_sys_modules():
+    """The loaders below install stub modules (server, comfy, aiohttp...); undo that."""
+    saved = dict(sys.modules)
+    # Earlier test modules may have left a stub `aiohttp` behind; drop it so the
+    # real package is imported for this test (restored afterwards).
+    for name in [n for n in sys.modules if n == "aiohttp" or n.startswith("aiohttp.")]:
+        if getattr(sys.modules[name], "__file__", None) is None:
+            del sys.modules[name]
+    yield
+    sys.modules.clear()
+    sys.modules.update(saved)
+
+
+async def _run_transfer(job_routes, collector_module, images, fmt, *, client_max_size,
+                        request_limit, chunk_target, audio=None):
+    """Send `images` from a real worker node to the real completion route.
+
+    Returns (queued items, sizes of every request body the server received,
+    exception raised by the worker or None).
+    """
+    import aiohttp
+    from aiohttp import web
+
+    queue = asyncio.Queue()
+    job_routes.prompt_server.distributed_jobs_lock = asyncio.Lock()
+    job_routes.prompt_server.distributed_pending_jobs = {"job-it": queue}
+    body_sizes = []
+
+    @web.middleware
+    async def record_size(request, handler):
+        body_sizes.append(request.content_length)
+        return await handler(request)
+
+    app = web.Application(client_max_size=client_max_size, middlewares=[record_size])
+    app.router.add_post("/distributed/job_complete", job_routes.job_complete_endpoint)
+    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]
+
+    error = None
+    try:
+        async with aiohttp.ClientSession() as session:
+            async def _get_session():
+                return session
+
+            collector_module.get_client_session = _get_session
+            collector_module.TRANSFER_MAX_REQUEST_BYTES = request_limit
+            collector_module.TRANSFER_MAX_CHUNK_BYTES = chunk_target
+            node = collector_module.DistributedCollectorNode()
+            try:
+                await node.send_batch_to_master(
+                    images, audio, "job-it", f"http://127.0.0.1:{port}", "worker-1", fmt, 90
+                )
+            except Exception as exc:  # surfaced to the caller for assertions
+                error = exc
+    finally:
+        await runner.cleanup()
+
+    items = []
+    while not queue.empty():
+        items.append(queue.get_nowait())
+    return items, body_sizes, error
+
+
+def _noise_images(count, height=360, width=640, seed=97):
+    return torch.from_numpy(np.random.default_rng(seed).random((count, height, width, 3), dtype=np.float32))
+
+
+@pytest.mark.parametrize("fmt", ["h264", "ffv1"])
+def test_high_entropy_video_stays_under_server_upload_limit(fmt):
+    """Encoder flush used to push finalized chunks past the limit (HTTP 400)."""
+    pytest.importorskip("aiohttp")
+    job_routes, collector_module = _load_modules()
+    images = _noise_images(24)
+    limit = 2 * 1024 * 1024
+    reserve = collector_module.DistributedCollectorNode.TRANSFER_ENVELOPE_RESERVE_BYTES
+
+    items, body_sizes, error = asyncio.run(_run_transfer(
+        job_routes, collector_module, images, fmt,
+        client_max_size=limit, request_limit=limit, chunk_target=limit - reserve))
+
+    assert error is None, error
+    assert len(body_sizes) > 1 and max(body_sizes) <= limit
+    assert [item["image_index"] for item in items] == list(range(24))
+    assert [item["is_last"] for item in items] == [False] * 23 + [True]
+
+
+def test_server_rejection_for_size_is_reported_not_retried_as_legacy():
+    """A worker configured above the master's limit fails loudly instead of
+    silently resending the whole batch as uncompressed PNG."""
+    pytest.importorskip("aiohttp")
+    job_routes, collector_module = _load_modules()
+    images = _noise_images(8)
+
+    items, body_sizes, error = asyncio.run(_run_transfer(
+        job_routes, collector_module, images, "ffv1",
+        client_max_size=1024 * 1024, request_limit=8 * 1024 * 1024, chunk_target=4 * 1024 * 1024))
+
+    assert isinstance(error, RuntimeError)
+    assert "HTTP 413" in str(error) and "--max-upload-size" in str(error)
+    assert len(body_sizes) == 1
+    assert items == []
+
+
+def test_frame_that_cannot_fit_sends_nothing():
+    pytest.importorskip("aiohttp")
+    job_routes, collector_module = _load_modules()
+    images = _noise_images(3)
+    reserve = collector_module.DistributedCollectorNode.TRANSFER_ENVELOPE_RESERVE_BYTES
+
+    items, body_sizes, error = asyncio.run(_run_transfer(
+        job_routes, collector_module, images, "png",
+        client_max_size=100 * 1024 * 1024, request_limit=reserve + 64 * 1024, chunk_target=64 * 1024))
+
+    assert isinstance(error, collector_module.transfer_codec.FrameTooLargeError)
+    assert body_sizes == []
+    assert items == []
+
+
+@pytest.mark.parametrize("fmt", ["png", "ffv1", "h264", "jpeg"])
+def test_worker_to_master_over_real_http(fmt):
+    # Import here, not at module level: other test harnesses only install their
+    # aiohttp stubs when the real package has not been imported yet.
+    aiohttp = pytest.importorskip("aiohttp")
+    from aiohttp import web
+
+    job_routes, collector_module = _load_modules()
+    images = torch.rand(5, 24, 36, 3)
+
+    async def scenario():
+        queue = asyncio.Queue()
+        job_routes.prompt_server.distributed_jobs_lock = asyncio.Lock()
+        job_routes.prompt_server.distributed_pending_jobs = {"job-it": queue}
+
+        app = web.Application()
+        app.router.add_post("/distributed/job_complete", job_routes.job_complete_endpoint)
+        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]
+
+        async with aiohttp.ClientSession() as session:
+            async def _get_session():
+                return session
+
+            collector_module.get_client_session = _get_session
+            collector_module.TRANSFER_MAX_CHUNK_BYTES = 2000  # several requests for still formats
+            node = collector_module.DistributedCollectorNode()
+            await node.send_batch_to_master(
+                images, None, "job-it", f"http://127.0.0.1:{port}", "worker-1", fmt, 90
+            )
+        await runner.cleanup()
+
+        items = []
+        while not queue.empty():
+            items.append(queue.get_nowait())
+        return items
+
+    items = asyncio.run(scenario())
+
+    assert [item["image_index"] for item in items] == [0, 1, 2, 3, 4]
+    assert [item["is_last"] for item in items] == [False, False, False, False, True]
+    expected = np.clip(255.0 * images.numpy(), 0, 255).astype(np.uint8).astype(np.float32) / 255.0
+    got = torch.cat([item["tensor"] for item in items]).numpy()
+    assert got.shape == expected.shape
+    if fmt in ("png", "ffv1"):
+        assert np.array_equal(got, expected)
+    else:
+        assert np.abs(got - expected).mean() < 0.25  # random noise: only sanity-check lossy codecs
@@ -0,0 +1,304 @@
+import asyncio
+import importlib.util
+import io
+import sys
+import types
+import unittest
+from dataclasses import dataclass, field
+from pathlib import Path
+
+from PIL import Image
+
+
+class _FakeResponse:
+    def __init__(self, payload, status=200):
+        self.payload = payload
+        self.status = status
+
+
+class _FakeRequest:
+    def __init__(self, json_payload=None, post_payload=None, headers=None, query=None):
+        self._json_payload = json_payload
+        self._post_payload = post_payload or {}
+        self.headers = headers or {}
+        self.query = query or {}
+
+    async def json(self):
+        return self._json_payload
+
+    async def post(self):
+        return self._post_payload
+
+
+class _Routes:
+    def post(self, _path):
+        def _decorator(fn):
+            return fn
+
+        return _decorator
+
+    def get(self, _path):
+        def _decorator(fn):
+            return fn
+
+        return _decorator
+
+
+def _load_usdu_routes_module():
+    module_path = Path(__file__).resolve().parents[2] / "api" / "usdu_routes.py"
+    package_name = "dist_api_usdu_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__ = []
+    sys.modules[package_name] = root_pkg
+
+    api_pkg = types.ModuleType(f"{package_name}.api")
+    api_pkg.__path__ = []
+    sys.modules[f"{package_name}.api"] = api_pkg
+
+    upscale_pkg = types.ModuleType(f"{package_name}.upscale")
+    upscale_pkg.__path__ = []
+    sys.modules[f"{package_name}.upscale"] = upscale_pkg
+
+    utils_pkg = types.ModuleType(f"{package_name}.utils")
+    utils_pkg.__path__ = []
+    sys.modules[f"{package_name}.utils"] = utils_pkg
+
+    prompt_server_holder = {
+        "value": types.SimpleNamespace(
+            distributed_tile_jobs_lock=asyncio.Lock(),
+            distributed_pending_tile_jobs={},
+        )
+    }
+
+    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
+
+    server_module = types.ModuleType("server")
+    server_module.PromptServer = types.SimpleNamespace(instance=types.SimpleNamespace(routes=_Routes()))
+    sys.modules["server"] = server_module
+
+    logging_module = types.ModuleType(f"{package_name}.utils.logging")
+    logging_module.debug_log = lambda *_args, **_kwargs: None
+    sys.modules[f"{package_name}.utils.logging"] = logging_module
+
+    network_module = types.ModuleType(f"{package_name}.utils.network")
+
+    async def _handle_api_error(_request, error, status=500):
+        return _FakeResponse({"status": "error", "message": str(error)}, status=status)
+
+    network_module.handle_api_error = _handle_api_error
+    sys.modules[f"{package_name}.utils.network"] = network_module
+
+    job_store_module = types.ModuleType(f"{package_name}.upscale.job_store")
+    job_store_module.MAX_PAYLOAD_SIZE = 1024
+    job_store_module.ensure_tile_jobs_initialized = lambda: prompt_server_holder["value"]
+    sys.modules[f"{package_name}.upscale.job_store"] = job_store_module
+
+    job_models_module = types.ModuleType(f"{package_name}.upscale.job_models")
+
+    class BaseJobState:
+        pass
+
+    @dataclass
+    class TileJobState(BaseJobState):
+        multi_job_id: str
+        mode: str = field(default="static", init=False)
+        queue: asyncio.Queue = field(default_factory=asyncio.Queue)
+        pending_tasks: asyncio.Queue = field(default_factory=asyncio.Queue)
+        completed_tasks: dict = field(default_factory=dict)
+        worker_status: dict = field(default_factory=dict)
+        assigned_to_workers: dict = field(default_factory=dict)
+        batch_size: int = 0
+        num_tiles_per_image: int = 0
+        batched_static: bool = False
+
+    @dataclass
+    class ImageJobState(BaseJobState):
+        multi_job_id: str
+        mode: str = field(default="dynamic", init=False)
+        queue: asyncio.Queue = field(default_factory=asyncio.Queue)
+        pending_images: asyncio.Queue = field(default_factory=asyncio.Queue)
+        completed_images: dict = field(default_factory=dict)
+        worker_status: dict = field(default_factory=dict)
+        assigned_to_workers: dict = field(default_factory=dict)
+        batch_size: int = 0
+        num_tiles_per_image: int = 0
+        batched_static: bool = False
+
+        @property
+        def pending_tasks(self):
+            return self.pending_images
+
+        @property
+        def completed_tasks(self):
+            return self.completed_images
+
+    job_models_module.BaseJobState = BaseJobState
+    job_models_module.TileJobState = TileJobState
+    job_models_module.ImageJobState = ImageJobState
+    sys.modules[f"{package_name}.upscale.job_models"] = job_models_module
+
+    parsers_module = types.ModuleType(f"{package_name}.upscale.payload_parsers")
+    parsers_module._parse_tiles_from_form = lambda _data: []
+    sys.modules[f"{package_name}.upscale.payload_parsers"] = parsers_module
+
+    spec = importlib.util.spec_from_file_location(f"{package_name}.api.usdu_routes", 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)
+
+    if created_aiohttp_stub:
+        sys.modules.pop("aiohttp", None)
+
+    module.web = types.SimpleNamespace(
+        json_response=lambda payload, status=200: _FakeResponse(payload, status=status)
+    )
+
+    module._prompt_server_holder = prompt_server_holder
+    module._TileJobState = TileJobState
+    module._ImageJobState = ImageJobState
+    return module
+
+
+usdu_routes = _load_usdu_routes_module()
+
+
+class _UploadField:
+    def __init__(self, data):
+        self.file = io.BytesIO(data)
+
+
+def _tiny_png_bytes():
+    image = Image.new("RGB", (1, 1), (255, 0, 0))
+    buf = io.BytesIO()
+    image.save(buf, format="PNG")
+    return buf.getvalue()
+
+
+class USDURoutesTests(unittest.IsolatedAsyncioTestCase):
+    async def asyncSetUp(self):
+        usdu_routes._prompt_server_holder["value"] = types.SimpleNamespace(
+            distributed_tile_jobs_lock=asyncio.Lock(),
+            distributed_pending_tile_jobs={},
+        )
+
+    async def test_heartbeat_updates_worker_status(self):
+        prompt_server = usdu_routes._prompt_server_holder["value"]
+        job_data = usdu_routes._TileJobState("job-1")
+        prompt_server.distributed_pending_tile_jobs["job-1"] = job_data
+
+        request = _FakeRequest(json_payload={"worker_id": "worker-a", "multi_job_id": "job-1"})
+        response = await usdu_routes.heartbeat_endpoint(request)
+
+        self.assertEqual(response.status, 200)
+        self.assertEqual(response.payload.get("status"), "success")
+        self.assertIn("worker-a", job_data.worker_status)
+
+    async def test_heartbeat_missing_fields_returns_400(self):
+        request = _FakeRequest(json_payload={"worker_id": "worker-a"})
+        response = await usdu_routes.heartbeat_endpoint(request)
+
+        self.assertEqual(response.status, 400)
+        self.assertIn("missing", response.payload.get("message", "").lower())
+
+    async def test_request_image_dynamic_assigns_next_image(self):
+        prompt_server = usdu_routes._prompt_server_holder["value"]
+        job_data = usdu_routes._ImageJobState("job-2")
+        await job_data.pending_images.put(7)
+        prompt_server.distributed_pending_tile_jobs["job-2"] = job_data
+
+        request = _FakeRequest(json_payload={"worker_id": "worker-a", "multi_job_id": "job-2"})
+        response = await usdu_routes.request_image_endpoint(request)
+
+        self.assertEqual(response.status, 200)
+        self.assertEqual(response.payload.get("image_idx"), 7)
+        self.assertEqual(response.payload.get("estimated_remaining"), 0)
+        self.assertEqual(job_data.assigned_to_workers["worker-a"], [7])
+        self.assertIn("worker-a", job_data.worker_status)
+
+    async def test_request_image_static_assigns_tile_and_batched_flag(self):
+        prompt_server = usdu_routes._prompt_server_holder["value"]
+        job_data = usdu_routes._TileJobState("job-3")
+        job_data.batched_static = True
+        await job_data.pending_tasks.put(4)
+        prompt_server.distributed_pending_tile_jobs["job-3"] = job_data
+
+        request = _FakeRequest(json_payload={"worker_id": "worker-a", "multi_job_id": "job-3"})
+        response = await usdu_routes.request_image_endpoint(request)
+
+        self.assertEqual(response.status, 200)
+        self.assertEqual(response.payload.get("tile_idx"), 4)
+        self.assertTrue(response.payload.get("batched_static"))
+        self.assertEqual(job_data.assigned_to_workers["worker-a"], [4])
+
+    async def test_submit_tiles_completion_signal_enqueues_last_marker(self):
+        prompt_server = usdu_routes._prompt_server_holder["value"]
+        job_data = usdu_routes._TileJobState("job-4")
+        prompt_server.distributed_pending_tile_jobs["job-4"] = job_data
+
+        request = _FakeRequest(
+            post_payload={
+                "multi_job_id": "job-4",
+                "worker_id": "worker-a",
+                "batch_size": "0",
+                "is_last": "true",
+            },
+            headers={"content-length": "128"},
+        )
+
+        response = await usdu_routes.submit_tiles_endpoint(request)
+
+        self.assertEqual(response.status, 200)
+        queued = await job_data.queue.get()
+        self.assertEqual(queued["worker_id"], "worker-a")
+        self.assertTrue(queued["is_last"])
+        self.assertEqual(queued["tiles"], [])
+
+    async def test_submit_image_enqueues_processed_image_payload(self):
+        prompt_server = usdu_routes._prompt_server_holder["value"]
+        job_data = usdu_routes._ImageJobState("job-5")
+        prompt_server.distributed_pending_tile_jobs["job-5"] = job_data
+
+        request = _FakeRequest(
+            post_payload={
+                "multi_job_id": "job-5",
+                "worker_id": "worker-a",
+                "image_idx": "2",
+                "full_image": _UploadField(_tiny_png_bytes()),
+                "is_last": "false",
+            },
+            headers={"content-length": "256"},
+        )
+
+        response = await usdu_routes.submit_image_endpoint(request)
+
+        self.assertEqual(response.status, 200)
+        queued = await job_data.queue.get()
+        self.assertEqual(queued["worker_id"], "worker-a")
+        self.assertEqual(queued["image_idx"], 2)
+        self.assertIn("image", queued)
+
+    async def test_job_status_endpoint_reports_ready(self):
+        prompt_server = usdu_routes._prompt_server_holder["value"]
+        prompt_server.distributed_pending_tile_jobs["job-6"] = usdu_routes._TileJobState("job-6")
+
+        request = _FakeRequest(query={"multi_job_id": "job-6"})
+        response = await usdu_routes.job_status_endpoint(request)
+
+        self.assertEqual(response.status, 200)
+        self.assertTrue(response.payload.get("ready"))
+
+
+if __name__ == "__main__":
+    unittest.main()
@@ -0,0 +1,505 @@
+import importlib.util
+import os
+import sys
+import tempfile
+import types
+import unittest
+from collections import deque
+from pathlib import Path
+from unittest.mock import patch
+
+
+class _FakeResponse:
+    def __init__(self, payload, status=200):
+        self.payload = payload
+        self.status = status
+
+
+class _FakeRequest:
+    def __init__(self, payload=None, match_info=None, query=None):
+        self._payload = payload
+        self.match_info = match_info or {}
+        self.query = query or {}
+
+    async def json(self):
+        return self._payload
+
+
+class _FakeHTTPClientResponse:
+    def __init__(self, payload, status=200):
+        self._payload = payload
+        self.status = status
+
+    async def __aenter__(self):
+        return self
+
+    async def __aexit__(self, _exc_type, _exc, _tb):
+        return False
+
+    async def json(self):
+        return self._payload
+
+    async def text(self):
+        return str(self._payload)
+
+
+class _FakeHTTPClientSession:
+    def __init__(self, payload, status=200):
+        self._payload = payload
+        self._status = status
+        self.calls = []
+
+    def get(self, url, params=None, timeout=None):
+        self.calls.append({"url": url, "params": params, "timeout": timeout})
+        return _FakeHTTPClientResponse(self._payload, status=self._status)
+
+
+class _DummyWorkerManager:
+    def __init__(self):
+        self.processes = {}
+
+    def launch_worker(self, worker):
+        worker_id = str(worker["id"])
+        self.processes[worker_id] = {
+            "pid": 12345,
+            "log_file": f"/tmp/distributed_worker_{worker_id}.log",
+            "process": None,
+        }
+        return 12345
+
+    def _is_process_running(self, _pid):
+        return False
+
+    def save_processes(self):
+        return None
+
+    def stop_worker(self, _worker_id):
+        return True, "Stopped"
+
+    def get_managed_workers(self):
+        return []
+
+
+class _ImmediateLoop:
+    async def run_in_executor(self, _executor, func, *args):
+        return func(*args)
+
+
+def _load_worker_routes_module():
+    module_path = Path(__file__).resolve().parents[2] / "api" / "worker_routes.py"
+    package_name = "dist_api_worker_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__ = []
+    sys.modules[package_name] = root_pkg
+
+    api_pkg = types.ModuleType(f"{package_name}.api")
+    api_pkg.__path__ = []
+    sys.modules[f"{package_name}.api"] = api_pkg
+
+    utils_pkg = types.ModuleType(f"{package_name}.utils")
+    utils_pkg.__path__ = []
+    sys.modules[f"{package_name}.utils"] = utils_pkg
+
+    workers_pkg = types.ModuleType(f"{package_name}.workers")
+    workers_pkg.__path__ = [str(module_path.parents[1] / "workers")]
+    workers_pkg.get_worker_manager = lambda: _DummyWorkerManager()
+    sys.modules[f"{package_name}.workers"] = workers_pkg
+
+    detection_module = types.ModuleType(f"{package_name}.workers.detection")
+    detection_module.is_local_worker = lambda *_args, **_kwargs: True
+    detection_module.is_same_physical_host = lambda *_args, **_kwargs: True
+    detection_module.get_machine_id = lambda: "machine-id"
+    detection_module.is_docker_environment = lambda: False
+    detection_module.is_runpod_environment = lambda: False
+    detection_module.get_comms_channel = lambda *_args, **_kwargs: "lan"
+    sys.modules[f"{package_name}.workers.detection"] = detection_module
+
+    created_aiohttp_stub = False
+    if "aiohttp" not in sys.modules:
+        created_aiohttp_stub = True
+        aiohttp_module = types.ModuleType("aiohttp")
+
+        class _ClientTimeout:
+            def __init__(self, total=None):
+                self.total = total
+
+        class _WSMsgType:
+            TEXT = "TEXT"
+            ERROR = "ERROR"
+            CLOSED = "CLOSED"
+
+        class _WebSocketResponse:
+            def __init__(self, *args, **kwargs):
+                self.args = args
+                self.kwargs = kwargs
+
+            async def prepare(self, _request):
+                return None
+
+            async def send_json(self, _payload):
+                return None
+
+            def __aiter__(self):
+                async def _empty():
+                    if False:
+                        yield None
+                return _empty()
+
+        aiohttp_module.ClientTimeout = _ClientTimeout
+        aiohttp_module.WSMsgType = _WSMsgType
+        aiohttp_module.web = types.SimpleNamespace(
+            json_response=lambda payload, status=200: _FakeResponse(payload, status=status),
+            WebSocketResponse=_WebSocketResponse,
+        )
+        sys.modules["aiohttp"] = aiohttp_module
+
+    class _Routes:
+        def get(self, _path):
+            def _decorator(fn):
+                return fn
+            return _decorator
+
+        def post(self, _path):
+            def _decorator(fn):
+                return fn
+            return _decorator
+
+    server_module = types.ModuleType("server")
+    server_module.PromptServer = types.SimpleNamespace(instance=types.SimpleNamespace(routes=_Routes()))
+    sys.modules["server"] = server_module
+
+    created_torch_stub = False
+    if "torch" not in sys.modules:
+        created_torch_stub = True
+        torch_module = types.ModuleType("torch")
+        torch_module.cuda = types.SimpleNamespace(
+            is_available=lambda: False,
+            empty_cache=lambda: None,
+            ipc_collect=lambda: None,
+            current_device=lambda: 0,
+            device_count=lambda: 0,
+        )
+        sys.modules["torch"] = torch_module
+
+    logging_module = types.ModuleType(f"{package_name}.utils.logging")
+    logging_module.debug_log = lambda *_args, **_kwargs: None
+    logging_module.log = lambda *_args, **_kwargs: None
+    sys.modules[f"{package_name}.utils.logging"] = logging_module
+
+    config_module = types.ModuleType(f"{package_name}.utils.config")
+    config_module.load_config = lambda: {"workers": []}
+    sys.modules[f"{package_name}.utils.config"] = config_module
+
+    network_module = types.ModuleType(f"{package_name}.utils.network")
+
+    async def _handle_api_error(_request, error, status=500):
+        return _FakeResponse({"status": "error", "message": str(error)}, status=status)
+
+    network_module.handle_api_error = _handle_api_error
+    network_module.normalize_host = lambda value: value
+    network_module.build_worker_url = lambda worker, endpoint="": f"http://localhost:{worker.get('port', 8188)}{endpoint}"
+
+    async def _probe_worker(*_args, **_kwargs):
+        return None
+
+    network_module.probe_worker = _probe_worker
+
+    async def _get_client_session():
+        raise RuntimeError("not used in these tests")
+
+    network_module.get_client_session = _get_client_session
+    network_module.get_server_port = lambda: 8189
+    sys.modules[f"{package_name}.utils.network"] = network_module
+
+    constants_module = types.ModuleType(f"{package_name}.utils.constants")
+    constants_module.CHUNK_SIZE = 8192
+    sys.modules[f"{package_name}.utils.constants"] = constants_module
+
+    async_helpers_module = types.ModuleType(f"{package_name}.utils.async_helpers")
+
+    async def _queue_prompt_payload(*_args, **_kwargs):
+        return "prompt-id"
+
+    async_helpers_module.queue_prompt_payload = _queue_prompt_payload
+
+    class _PromptValidationError(RuntimeError):
+        def __init__(self, message="invalid prompt", validation_error=None, node_errors=None):
+            super().__init__(message)
+            self.validation_error = validation_error if isinstance(validation_error, dict) else {}
+            self.node_errors = node_errors if isinstance(node_errors, dict) else {}
+
+    async_helpers_module.PromptValidationError = _PromptValidationError
+    sys.modules[f"{package_name}.utils.async_helpers"] = async_helpers_module
+
+    schemas_module = types.ModuleType(f"{package_name}.api.schemas")
+
+    def _require_fields(data, *fields):
+        missing = []
+        for field in fields:
+            value = data.get(field) if isinstance(data, dict) else None
+            if value is None or (isinstance(value, str) and not value.strip()):
+                missing.append(field)
+        return missing
+
+    def _validate_worker_id(worker_id, config):
+        return any(str(worker.get("id")) == str(worker_id) for worker in config.get("workers", []))
+
+    schemas_module.require_fields = _require_fields
+    schemas_module.validate_worker_id = _validate_worker_id
+    sys.modules[f"{package_name}.api.schemas"] = schemas_module
+
+    spec = importlib.util.spec_from_file_location(f"{package_name}.api.worker_routes", 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)
+    # 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)
+    if created_torch_stub:
+        sys.modules.pop("torch", None)
+
+    return 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)), \
+             patch.object(worker_routes, "get_network_ips", return_value=["127.0.0.1"]), \
+             patch.object(worker_routes, "load_config", return_value={"workers": []}), \
+             patch.object(worker_routes, "get_server_port", return_value=8189), \
+             patch.object(worker_routes, "allocate_worker_ports", return_value=[8191, 8192, 8193]) as allocate:
+            response = await worker_routes.get_network_info_endpoint(_FakeRequest())
+        self.assertEqual(response.payload["master_port"], 8189)
+        self.assertEqual(response.payload["local_worker_ports"], [8191, 8192, 8193])
+        self.assertEqual(response.payload["cuda_device_count"], 4)
+        allocate.assert_called_once_with(8189, [], 3)
+
+    async def test_network_info_does_not_reallocate_existing_configuration(self):
+        with patch.object(worker_routes, "_get_cuda_info", return_value=(0, 4, 4)), \
+             patch.object(worker_routes, "get_network_ips", return_value=[]), \
+             patch.object(worker_routes, "load_config", return_value={"workers": [{"id": "existing", "port": 8189}]}), \
+             patch.object(worker_routes, "allocate_worker_ports") as allocate:
+            response = await worker_routes.get_network_info_endpoint(_FakeRequest())
+        self.assertEqual(response.payload["local_worker_ports"], [])
+        allocate.assert_not_called()
+
+    async def test_launch_conflict_is_a_clear_client_error(self):
+        manager = _DummyWorkerManager()
+        config = {"workers": [{"id": "worker-a", "name": "Worker A", "port": 8189}]}
+        with patch.object(worker_routes, "get_worker_manager", return_value=manager), \
+             patch.object(worker_routes, "load_config", return_value=config), \
+             patch.object(manager, "launch_worker", side_effect=ValueError("Worker port 8189 conflicts with the master port 8189")):
+            response = await worker_routes.launch_worker_endpoint(_FakeRequest({"worker_id": "worker-a"}))
+        self.assertEqual(response.status, 400)
+        self.assertIn("master port 8189", response.payload["message"])
+        self.assertEqual(manager.processes, {})
+
+    async def test_launch_worker_valid_id_returns_200(self):
+        manager = _DummyWorkerManager()
+        config = {"workers": [{"id": "worker-a", "name": "Worker A", "port": 8188}]}
+        request = _FakeRequest({"worker_id": "worker-a"})
+
+        with patch.object(worker_routes, "get_worker_manager", return_value=manager), patch.object(
+            worker_routes, "load_config", return_value=config
+        ), patch.object(
+            worker_routes.asyncio, "get_running_loop", return_value=_ImmediateLoop()
+        ):
+            response = await worker_routes.launch_worker_endpoint(request)
+
+        self.assertEqual(response.status, 200)
+        self.assertEqual(response.payload.get("status"), "success")
+        self.assertEqual(response.payload.get("pid"), 12345)
+
+    async def test_launch_worker_unknown_id_returns_404(self):
+        manager = _DummyWorkerManager()
+        config = {"workers": [{"id": "worker-a", "name": "Worker A", "port": 8188}]}
+        request = _FakeRequest({"worker_id": "missing-worker"})
+
+        with patch.object(worker_routes, "get_worker_manager", return_value=manager), patch.object(
+            worker_routes, "load_config", return_value=config
+        ):
+            response = await worker_routes.launch_worker_endpoint(request)
+
+        self.assertEqual(response.status, 404)
+        self.assertIn("not found", response.payload.get("message", "").lower())
+
+    async def test_worker_log_returns_content_json(self):
+        manager = _DummyWorkerManager()
+        with tempfile.NamedTemporaryFile("w", delete=False, encoding="utf-8") as handle:
+            handle.write("line-1\nline-2\nline-3\n")
+            log_path = handle.name
+
+        manager.processes["worker-a"] = {
+            "pid": 9999,
+            "log_file": log_path,
+            "process": None,
+        }
+
+        request = _FakeRequest(match_info={"worker_id": "worker-a"}, query={"lines": "2"})
+        try:
+            with patch.object(worker_routes, "get_worker_manager", return_value=manager), patch.object(
+                worker_routes.asyncio, "get_running_loop", return_value=_ImmediateLoop()
+            ):
+                response = await worker_routes.get_worker_log_endpoint(request)
+        finally:
+            if os.path.exists(log_path):
+                os.remove(log_path)
+
+        self.assertEqual(response.status, 200)
+        self.assertEqual(response.payload.get("status"), "success")
+        self.assertIn("content", response.payload)
+        self.assertIn("line-3", response.payload["content"])
+
+    async def test_local_log_reads_memory_buffer(self):
+        request = _FakeRequest(query={"lines": "2"})
+        fake_logs = deque(
+            [
+                {"m": "line-1\n"},
+                {"m": "line-2\n"},
+                {"m": "line-3\n"},
+            ],
+            maxlen=300,
+        )
+        app_module = types.ModuleType("app")
+        app_module.__path__ = []
+        logger_module = types.ModuleType("app.logger")
+        logger_module.get_logs = lambda: fake_logs
+        app_module.logger = logger_module
+
+        with patch.dict(sys.modules, {"app": app_module, "app.logger": logger_module}):
+            response = await worker_routes.get_local_log_endpoint(request)
+
+        self.assertEqual(response.status, 200)
+        self.assertEqual(response.payload.get("status"), "success")
+        self.assertEqual(response.payload.get("source"), "memory")
+        self.assertEqual(response.payload.get("entries"), 2)
+        self.assertIn("line-3", response.payload.get("content", ""))
+
+    async def test_remote_worker_log_proxies_to_worker_local_log_endpoint(self):
+        config = {
+            "workers": [
+                {
+                    "id": "worker-remote",
+                    "name": "Remote Worker",
+                    "host": "worker.example.com",
+                    "port": 8188,
+                    "type": "remote",
+                }
+            ]
+        }
+        request = _FakeRequest(match_info={"worker_id": "worker-remote"}, query={"lines": "120"})
+        proxied_payload = {
+            "status": "success",
+            "content": "remote-log-content\n",
+            "entries": 1,
+            "source": "memory",
+            "truncated": False,
+            "lines_shown": 1,
+        }
+        fake_session = _FakeHTTPClientSession(proxied_payload)
+
+        async def _fake_get_client_session():
+            return fake_session
+
+        with patch.object(worker_routes, "load_config", return_value=config), patch.object(
+            worker_routes, "get_client_session", side_effect=_fake_get_client_session
+        ):
+            response = await worker_routes.get_remote_worker_log_endpoint(request)
+
+        self.assertEqual(response.status, 200)
+        self.assertEqual(response.payload.get("content"), "remote-log-content\n")
+        self.assertEqual(len(fake_session.calls), 1)
+        self.assertEqual(fake_session.calls[0]["params"], {"lines": "120"})
+        self.assertTrue(fake_session.calls[0]["url"].endswith("/distributed/local_log"))
+
+    async def test_remote_worker_log_rejects_local_workers(self):
+        config = {"workers": [{"id": "worker-local", "name": "Local Worker", "port": 8188}]}
+        request = _FakeRequest(match_info={"worker_id": "worker-local"})
+
+        with patch.object(worker_routes, "load_config", return_value=config):
+            response = await worker_routes.get_remote_worker_log_endpoint(request)
+
+        self.assertEqual(response.status, 400)
+        self.assertIn("local", response.payload.get("message", "").lower())
+
+
+if __name__ == "__main__":
+    unittest.main()
@@ -0,0 +1,2 @@
+# This conftest.py marks the tests/ directory as the pytest collection root,
+# preventing pytest from traversing into the parent package's __init__.py.
@@ -0,0 +1,691 @@
+{
+  "DistributedCollector": {
+    "input": {
+      "required": {
+        "load_balance": [
+          "BOOLEAN",
+          {
+            "default": false,
+            "tooltip": "Run this workflow on one least-busy participant (master included when participating)."
+          }
+        ]
+      },
+      "optional": {
+        "images": [
+          "IMAGE"
+        ],
+        "audio": [
+          "AUDIO"
+        ]
+      },
+      "hidden": {
+        "multi_job_id": [
+          "STRING",
+          {
+            "default": ""
+          }
+        ],
+        "is_worker": [
+          "BOOLEAN",
+          {
+            "default": false
+          }
+        ],
+        "master_url": [
+          "STRING",
+          {
+            "default": ""
+          }
+        ],
+        "enabled_worker_ids": [
+          "STRING",
+          {
+            "default": "[]"
+          }
+        ],
+        "worker_batch_size": [
+          "INT",
+          {
+            "default": 1,
+            "min": 1,
+            "max": 1024
+          }
+        ],
+        "worker_id": [
+          "STRING",
+          {
+            "default": ""
+          }
+        ],
+        "pass_through": [
+          "BOOLEAN",
+          {
+            "default": false
+          }
+        ],
+        "delegate_only": [
+          "BOOLEAN",
+          {
+            "default": false
+          }
+        ]
+      }
+    },
+    "input_order": {
+      "required": [
+        "load_balance"
+      ],
+      "optional": [
+        "images",
+        "audio"
+      ],
+      "hidden": [
+        "multi_job_id",
+        "is_worker",
+        "master_url",
+        "enabled_worker_ids",
+        "worker_batch_size",
+        "worker_id",
+        "pass_through",
+        "delegate_only"
+      ]
+    },
+    "output": [
+      "IMAGE",
+      "AUDIO"
+    ],
+    "output_name": [
+      "images",
+      "audio"
+    ],
+    "output_is_list": [
+      false,
+      false
+    ],
+    "is_input_list": true,
+    "output_node": false,
+    "category": "image",
+    "display_name": "Distributed Collector"
+  },
+  "DistributedSeed": {
+    "input": {
+      "required": {
+        "seed": [
+          "INT",
+          {
+            "default": 1125899906842,
+            "min": 0,
+            "max": 1125899906842624,
+            "forceInput": false
+          }
+        ]
+      },
+      "hidden": {
+        "is_worker": [
+          "BOOLEAN",
+          {
+            "default": false
+          }
+        ],
+        "worker_id": [
+          "STRING",
+          {
+            "default": ""
+          }
+        ]
+      }
+    },
+    "input_order": {
+      "required": [
+        "seed"
+      ],
+      "hidden": [
+        "is_worker",
+        "worker_id"
+      ]
+    },
+    "output": [
+      "INT"
+    ],
+    "output_name": [
+      "seed"
+    ],
+    "output_is_list": [
+      false
+    ],
+    "is_input_list": false,
+    "output_node": false,
+    "category": "utils",
+    "display_name": "Distributed Seed"
+  },
+  "DistributedModelName": {
+    "input": {
+      "required": {
+        "text": [
+          "STRING",
+          {
+            "default": ""
+          }
+        ]
+      },
+      "hidden": {
+        "unique_id": "UNIQUE_ID",
+        "extra_pnginfo": "EXTRA_PNGINFO"
+      }
+    },
+    "input_order": {
+      "required": [
+        "text"
+      ],
+      "hidden": [
+        "unique_id",
+        "extra_pnginfo"
+      ]
+    },
+    "output": [
+      "*"
+    ],
+    "output_name": [
+      "output"
+    ],
+    "output_is_list": [
+      false
+    ],
+    "is_input_list": false,
+    "output_node": true,
+    "category": "utils",
+    "display_name": "Distributed Model Name"
+  },
+  "DistributedValue": {
+    "input": {
+      "required": {
+        "default_value": [
+          "STRING",
+          {
+            "default": ""
+          }
+        ],
+        "worker_values": [
+          "STRING",
+          {
+            "default": "{}"
+          }
+        ]
+      },
+      "hidden": {
+        "is_worker": [
+          "BOOLEAN",
+          {
+            "default": false
+          }
+        ],
+        "worker_id": [
+          "STRING",
+          {
+            "default": ""
+          }
+        ]
+      }
+    },
+    "input_order": {
+      "required": [
+        "default_value",
+        "worker_values"
+      ],
+      "hidden": [
+        "is_worker",
+        "worker_id"
+      ]
+    },
+    "output": [
+      "*"
+    ],
+    "output_name": [
+      "value"
+    ],
+    "output_is_list": [
+      false
+    ],
+    "is_input_list": false,
+    "output_node": false,
+    "category": "utils",
+    "display_name": "Distributed Value"
+  },
+  "ImageBatchDivider": {
+    "input": {
+      "required": {
+        "images": [
+          "IMAGE"
+        ],
+        "divide_by": [
+          "INT",
+          {
+            "default": 2,
+            "min": 1,
+            "max": 10,
+            "step": 1,
+            "display": "number",
+            "tooltip": "Number of parts to divide the batch into"
+          }
+        ]
+      }
+    },
+    "input_order": {
+      "required": [
+        "images",
+        "divide_by"
+      ]
+    },
+    "output": [
+      "*",
+      "*",
+      "*",
+      "*",
+      "*",
+      "*",
+      "*",
+      "*",
+      "*",
+      "*"
+    ],
+    "output_name": [
+      "batch_1",
+      "batch_2",
+      "batch_3",
+      "batch_4",
+      "batch_5",
+      "batch_6",
+      "batch_7",
+      "batch_8",
+      "batch_9",
+      "batch_10"
+    ],
+    "output_is_list": [
+      false,
+      false,
+      false,
+      false,
+      false,
+      false,
+      false,
+      false,
+      false,
+      false
+    ],
+    "is_input_list": false,
+    "output_node": true,
+    "category": "image",
+    "display_name": "Image Batch Divider"
+  },
+  "AudioBatchDivider": {
+    "input": {
+      "required": {
+        "audio": [
+          "AUDIO"
+        ],
+        "divide_by": [
+          "INT",
+          {
+            "default": 2,
+            "min": 1,
+            "max": 10,
+            "step": 1,
+            "display": "number",
+            "tooltip": "Number of sequential time segments to create"
+          }
+        ]
+      }
+    },
+    "input_order": {
+      "required": [
+        "audio",
+        "divide_by"
+      ]
+    },
+    "output": [
+      "*",
+      "*",
+      "*",
+      "*",
+      "*",
+      "*",
+      "*",
+      "*",
+      "*",
+      "*"
+    ],
+    "output_name": [
+      "audio_1",
+      "audio_2",
+      "audio_3",
+      "audio_4",
+      "audio_5",
+      "audio_6",
+      "audio_7",
+      "audio_8",
+      "audio_9",
+      "audio_10"
+    ],
+    "output_is_list": [
+      false,
+      false,
+      false,
+      false,
+      false,
+      false,
+      false,
+      false,
+      false,
+      false
+    ],
+    "is_input_list": false,
+    "output_node": true,
+    "category": "audio",
+    "display_name": "Audio Segment Divider"
+  },
+  "DistributedEmptyImage": {
+    "input": {
+      "required": {
+        "height": [
+          "INT",
+          {
+            "default": 64,
+            "min": 1,
+            "max": 4096,
+            "step": 1
+          }
+        ],
+        "width": [
+          "INT",
+          {
+            "default": 64,
+            "min": 1,
+            "max": 4096,
+            "step": 1
+          }
+        ],
+        "channels": [
+          "INT",
+          {
+            "default": 3,
+            "min": 1,
+            "max": 4,
+            "step": 1
+          }
+        ]
+      }
+    },
+    "input_order": {
+      "required": [
+        "height",
+        "width",
+        "channels"
+      ]
+    },
+    "output": [
+      "IMAGE"
+    ],
+    "output_name": [
+      "IMAGE"
+    ],
+    "output_is_list": [
+      false
+    ],
+    "is_input_list": false,
+    "output_node": false,
+    "category": "image",
+    "display_name": "Distributed Empty Image"
+  },
+  "UltimateSDUpscaleDistributed": {
+    "input": {
+      "required": {
+        "upscaled_image": [
+          "IMAGE"
+        ],
+        "model": [
+          "MODEL"
+        ],
+        "positive": [
+          "CONDITIONING"
+        ],
+        "negative": [
+          "CONDITIONING"
+        ],
+        "vae": [
+          "VAE"
+        ],
+        "seed": [
+          "INT",
+          {
+            "default": 0,
+            "min": 0,
+            "max": 18446744073709551615
+          }
+        ],
+        "steps": [
+          "INT",
+          {
+            "default": 20,
+            "min": 1,
+            "max": 10000
+          }
+        ],
+        "cfg": [
+          "FLOAT",
+          {
+            "default": 8.0,
+            "min": 0.0,
+            "max": 100.0
+          }
+        ],
+        "sampler_name": [
+          [
+            "euler",
+            "euler_cfg_pp",
+            "euler_ancestral",
+            "euler_ancestral_cfg_pp",
+            "heun",
+            "heunpp2",
+            "exp_heun_2_x0",
+            "exp_heun_2_x0_sde",
+            "dpm_2",
+            "dpm_2_ancestral",
+            "lms",
+            "dpm_fast",
+            "dpm_adaptive",
+            "dpmpp_2s_ancestral",
+            "dpmpp_2s_ancestral_cfg_pp",
+            "dpmpp_sde",
+            "dpmpp_sde_gpu",
+            "dpmpp_2m",
+            "dpmpp_2m_cfg_pp",
+            "dpmpp_2m_sde",
+            "dpmpp_2m_sde_gpu",
+            "dpmpp_2m_sde_heun",
+            "dpmpp_2m_sde_heun_gpu",
+            "dpmpp_3m_sde",
+            "dpmpp_3m_sde_gpu",
+            "ddpm",
+            "lcm",
+            "ipndm",
+            "ipndm_v",
+            "deis",
+            "cfgpp_ud10_ab",
+            "res_multistep",
+            "res_multistep_cfg_pp",
+            "res_multistep_ancestral",
+            "res_multistep_ancestral_cfg_pp",
+            "gradient_estimation",
+            "gradient_estimation_cfg_pp",
+            "er_sde",
+            "seeds_2",
+            "seeds_3",
+            "sa_solver",
+            "sa_solver_pece",
+            "ddim",
+            "uni_pc",
+            "uni_pc_bh2"
+          ]
+        ],
+        "scheduler": [
+          [
+            "simple",
+            "sgm_uniform",
+            "karras",
+            "exponential",
+            "ddim_uniform",
+            "beta",
+            "normal",
+            "linear_quadratic",
+            "kl_optimal"
+          ]
+        ],
+        "denoise": [
+          "FLOAT",
+          {
+            "default": 0.5,
+            "min": 0.0,
+            "max": 1.0,
+            "step": 0.01
+          }
+        ],
+        "tile_width": [
+          "INT",
+          {
+            "default": 512,
+            "min": 64,
+            "max": 2048,
+            "step": 8
+          }
+        ],
+        "tile_height": [
+          "INT",
+          {
+            "default": 512,
+            "min": 64,
+            "max": 2048,
+            "step": 8
+          }
+        ],
+        "padding": [
+          "INT",
+          {
+            "default": 32,
+            "min": 0,
+            "max": 256,
+            "step": 8
+          }
+        ],
+        "mask_blur": [
+          "INT",
+          {
+            "default": 8,
+            "min": 0,
+            "max": 256
+          }
+        ],
+        "force_uniform_tiles": [
+          "BOOLEAN",
+          {
+            "default": true
+          }
+        ],
+        "tiled_decode": [
+          "BOOLEAN",
+          {
+            "default": false
+          }
+        ]
+      },
+      "hidden": {
+        "multi_job_id": [
+          "STRING",
+          {
+            "default": ""
+          }
+        ],
+        "is_worker": [
+          "BOOLEAN",
+          {
+            "default": false
+          }
+        ],
+        "master_url": [
+          "STRING",
+          {
+            "default": ""
+          }
+        ],
+        "enabled_worker_ids": [
+          "STRING",
+          {
+            "default": "[]"
+          }
+        ],
+        "worker_id": [
+          "STRING",
+          {
+            "default": ""
+          }
+        ],
+        "tile_indices": [
+          "STRING",
+          {
+            "default": ""
+          }
+        ],
+        "dynamic_threshold": [
+          "INT",
+          {
+            "default": 8,
+            "min": 1,
+            "max": 64
+          }
+        ]
+      }
+    },
+    "input_order": {
+      "required": [
+        "upscaled_image",
+        "model",
+        "positive",
+        "negative",
+        "vae",
+        "seed",
+        "steps",
+        "cfg",
+        "sampler_name",
+        "scheduler",
+        "denoise",
+        "tile_width",
+        "tile_height",
+        "padding",
+        "mask_blur",
+        "force_uniform_tiles",
+        "tiled_decode"
+      ],
+      "hidden": [
+        "multi_job_id",
+        "is_worker",
+        "master_url",
+        "enabled_worker_ids",
+        "worker_id",
+        "tile_indices",
+        "dynamic_threshold"
+      ]
+    },
+    "output": [
+      "IMAGE"
+    ],
+    "output_name": [
+      "IMAGE"
+    ],
+    "output_is_list": [
+      false
+    ],
+    "is_input_list": false,
+    "output_node": false,
+    "category": "image/upscaling",
+    "display_name": "Ultimate SD Upscale Distributed (No Upscale)"
+  }
+}
@@ -0,0 +1,595 @@
+"""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 = [], []
+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((1, 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['2']['inputs'].get('worker_item_token'), 'per-item node ran in outer graph'
+    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 TestCapture:
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {'required': {'images': ('IMAGE',), 'masks': ('MASK',), 'texts': ('STRING', {'forceInput': True})}}
+    INPUT_IS_LIST = True
+    OUTPUT_NODE = True
+    RETURN_TYPES = ()
+    FUNCTION = 'run'
+    def run(self, images, masks, texts):
+        assert not WORKER, 'post-End capture leaked into a worker prompt'
+        CAPTURED.append((images, masks, texts))
+        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_')]
+    nodes.NODE_CLASS_MAPPINGS.update(TestImageSource=TestImageSource,
+                                   TestImageOperation=TestImageOperation, TestCapture=TestCapture,
+                                   TestCachedModel=TestCachedModel, TestSharedFlag=TestSharedFlag,
+                                   TestRequiredOutputUtility=TestRequiredOutputUtility,
+                                   TestDiscardedPreview=TestDiscardedPreview)
+    # 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],
+            '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]}},
+        '5': {'class_type': 'TestCapture', 'inputs': {'images': ['4', 0], 'masks': ['4', 1], 'texts': ['4', 2]}},
+        '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]}},
+    }
+
+
+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
+            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}}}
+            previous = '1'
+            for start, end in [('2', '3'), ('4', '5'), ('6', '7')]:
+                chained[start] = {'class_type': 'DistributedImageListStart', 'inputs': {
+                    'images': [previous, 0]}}
+                chained[end] = {'class_type': 'DistributedImageListEnd', 'inputs': {
+                    'images': [start, 0], 'context': [start, 3]}}
+                previous = end
+            chained['8'] = {'class_type': 'TestCapture', 'inputs': {
+                'images': ['7', 0], 'masks': ['7', 1], 'texts': ['7', 2]}}
+            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 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', [])]:
+                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)
+                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)}
+            print('IMAGE_LIST_MASTER_OK ' + json.dumps(participation_counts))
+            assert len(outer_executions) == 4 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.
+            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, source_rows(), ['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))
+                    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 = source_rows()
+            bad_rows[0]['seed'] = -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), 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())
@@ -0,0 +1,469 @@
+"""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.media_collector").DistributedMediaCollector
+    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)
+    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)
+                # Collector must preserve native objects and ordered item lists,
+                # not merge batches, resample, convert types, or encode again.
+                assert all(a is e for a, e in zip(media[socket], expected[socket])), (tag, socket)
+            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 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 + variant}
+        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(22050)
+        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("DistributedMediaCollector", load_balance=False,
+                         multi_job_id="job-main", worker_id=participant,
+                         **{socket: ["source", index] for index, socket in enumerate(sockets)}),
+                     "text-collector": node("DistributedMediaCollector", load_balance=False,
+                         multi_job_id="job-text", worker_id=participant, texts=["source", 3])}
+            if key == "worker-a":
+                graph.update({"wav": node("LoadAudio", audio=wav.name), "clip": node("LoadVideo", file=clip.name)})
+                graph["collector"]["inputs"].update(audio=["wav", 0], video=["clip", 0])
+            before = copy.deepcopy(graph)
+            workflow, manifest = compiler.compile_media_prompt(graph, request_id)
+            assert graph == before
+            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-text"}
+            assert set(grouped["job-main"]) == set(sockets)
+            assert grouped["job-text"]["texts"] == expectations[key]["texts"]
+            for socket, kind in zip(sockets, kinds):
+                values, expected = grouped["job-main"][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-main"]["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 ("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("DistributedMediaCollector", transfer_format=fmt,
+                         transfer_quality=73, multi_job_id="job-" + key,
+                         worker_id="serverless:" + key,
+                         **{socket: ["source", i] for i, socket in enumerate(sockets)})}
+            original = copy.deepcopy(graph)
+            workflow, manifest = compiler.compile_media_prompt(graph, request_id)
+            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]
+            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 == "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 png jpeg h264; alpha and other media exact")
+        print("REAL_HTTP_BRIDGE_RESULTS_OK all five kinds; 6 bindings/request; native VIDEO remains readable after cleanup")
+        print("ATOMIC_DECODE_REJECTION_OK corrupt final artifact rejected before delivery")
+
+    participants = list(grouped_results)
+    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] = {socket: [item for values in ordered for item in values[socket]] for socket in sockets}
+        graph = {"collector": node("DistributedMediaCollector", load_balance=False, multi_job_id=job,
+                    enabled_worker_ids=json.dumps(participants), expected_media=json.dumps(sockets), delegate_only=delegate),
+                 "assert": node("AcceptanceMediaAssert", tag=tag,
+                    **{socket: ["collector", index] for index, socket in enumerate(sockets)})}
+        if not delegate:
+            graph["source"] = node("AcceptanceMediaSource", key="local")
+            graph["collector"]["inputs"].update({socket: ["source", index] for index, socket in enumerate(sockets)})
+        # The second group from the very same complete response must reach its
+        # own Collector, rather than being folded into the first collector job.
+        text_job, text_tag = job + "-text", tag + "-text"
+        text_queue = asyncio.Queue()
+        for participant in reversed(participants):
+            await text_queue.put({"worker_id": participant,
+                "media": grouped_results[participant]["job-text"], "is_last": True})
+        ps.distributed_pending_jobs[text_job] = text_queue
+        text_ordered = ([] if delegate else [fixtures["local"]]) + [grouped_results[p]["job-text"] for p in participants]
+        expected_sinks[text_tag] = {"texts": [item for values in text_ordered for item in values["texts"]]}
+        graph["text-collector"] = node("DistributedMediaCollector", load_balance=False, multi_job_id=text_job,
+            enabled_worker_ids=json.dumps(participants), expected_media='["texts"]', delegate_only=delegate)
+        graph["text-assert"] = node("AcceptanceMediaAssert", tag=text_tag, texts=["text-collector", 3])
+        if not delegate:
+            graph["text-collector"]["inputs"]["texts"] = ["source", 3]
+        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"]["text-assert"]["accepted"] == [seen_sinks[text_tag]]
+        assert job not in ps.distributed_pending_jobs and text_job not in ps.distributed_pending_jobs
+    assert source_calls == ["worker-a", "worker-b", "encoded-png", "encoded-jpeg", "encoded-h264", "local"], source_calls
+    assert set(seen_sinks) == {"master", "delegate", "master-text", "delegate-text"}
+    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))
@@ -0,0 +1,484 @@
+"""Exercise the actual loader, schemas and executor in a fresh CPU process.
+
+No HTTP listener, workers, model downloads or live installation changes.
+The original contracts were captured from 32ac027 using the same core.
+"""
+import asyncio
+import inspect
+import json
+import os
+from pathlib import Path
+import sys
+import tempfile
+from unittest.mock import patch
+
+from PIL import Image
+
+ROOT = Path(__file__).resolve().parents[2]
+COMFY_ROOT = Path(sys.argv[1]).resolve()
+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
+from comfy_api.v0_0_2 import io
+import execution
+import nodes
+import server
+import torch
+
+BASELINE = json.loads((ROOT / 'tests/fixtures/v1_node_contracts.json').read_text())
+
+# Inputs intentionally added after the V1 baseline was captured. They must be
+# optional, appended after the baseline inputs, and keep legacy behaviour as
+# their default so saved workflows are unaffected.
+ADDED_OPTIONAL_INPUTS = {
+    'DistributedCollector': {
+        'transfer_format': [
+            ['legacy_png', 'png', 'ffv1', 'jpeg', 'webp', 'h264', 'h265', 'av1'],
+            {'default': 'legacy_png'},
+        ],
+        'transfer_quality': ['INT', {'default': 90, 'min': 1, 'max': 100, 'step': 1}],
+    },
+}
+
+# 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', 'DistributedImageListStart', 'DistributedImageListEnd', 'DistributedMediaCollector'}
+
+
+def normalized_input(value):
+    kind = value[0]
+    opts = dict(value[1]) if len(value) > 1 else {}
+    if kind == 'STRING':
+        # V3 serializes the same default single-line widget explicitly.
+        opts.setdefault('multiline', False)
+    if kind == 'COMBO':
+        kind = opts['options']
+        opts = {key: val for key, val in opts.items() if key != 'options'}
+    if isinstance(kind, list):
+        # Single-selection is the V1 dropdown default as well.
+        opts.setdefault('multiselect', False)
+    return [kind, opts]
+
+
+async def map_node(cls, values, extra=None):
+    """Use the actual input parser and sanitized V3 executor clones."""
+    prepared, missing, hidden = execution.get_input_data(values, cls, unique_id='probe', extra_data=extra or {})
+    assert not missing, missing
+    # Use a worker thread for synchronous nodes: collector/upscale bridge back
+    # to PromptServer's running loop, just as ComfyUI's prompt worker does.
+    def run():
+        return asyncio.run(execution._async_map_node_over_list(
+            'v3-acceptance', 'probe', cls, prepared, cls.FUNCTION, v3_data=hidden))
+    return await asyncio.to_thread(run)
+
+
+def check_saved_workflows(mapping):
+    checked = 0
+    occurrences = 0
+    for path in sorted((ROOT / 'workflows').glob('*.json')):
+        workflow = json.loads(path.read_text())
+        local_nodes = {node['id']: node for node in workflow['nodes']}
+        for saved in workflow['nodes']:
+            if saved['type'] not in mapping:
+                continue
+            occurrences += 1
+            info = mapping[saved['type']].GET_NODE_INFO_V1()
+            inputs = {**info['input'].get('required', {}), **info['input'].get('optional', {})}
+            for socket in saved.get('inputs', []):
+                assert socket['name'] in inputs, (path.name, saved['id'], socket)
+                assert socket['type'] == inputs[socket['name']][0], (path.name, socket)
+            for index, output in enumerate(saved.get('outputs', [])):
+                assert output['type'] == info['output'][index], (path.name, index)
+            for link in workflow['links']:
+                if link[1] == saved['id']:
+                    assert link[2] < len(info['output']), (path.name, link)
+                    target = local_nodes[link[3]]['inputs'][link[4]]
+                    assert target['type'] == info['output'][link[2]], (path.name, link)
+        checked += 1
+    assert checked == 5 and occurrences == 10, (checked, occurrences)
+    print('SAVED_WORKFLOW_CONTRACTS_OK', checked, occurrences)
+
+
+async def check_execution(mapping, module, prompt_server, asset_manager):
+    result = await map_node(mapping['DistributedSeed'],
+                            {'seed': 123, 'is_worker': True, 'worker_id': 'worker_2'})
+    assert result[0].result == (126,)
+    values = json.dumps({'_type': 'INT', '3': '17'})
+    result = await map_node(mapping['DistributedValue'],
+                            {'default_value': '4', 'worker_values': values,
+                             'is_worker': True, 'worker_id': 'worker_2'})
+    assert result[0].result == (17,)
+    metadata = {'workflow': {'nodes': [{'id': 'probe', 'widgets_values': []}]}}
+    result = await map_node(mapping['DistributedModelName'], {'text': 'model.ckpt'},
+                            {'extra_pnginfo': metadata})
+    assert result[0].result == ('model.ckpt',)
+    assert result[0].ui == {'text': ['model.ckpt']}
+    assert metadata['workflow']['nodes'][0]['widgets_values'] == [['model.ckpt']]
+    result = await map_node(mapping['DistributedEmptyImage'],
+                            {'height': 8, 'width': 8, 'channels': 3})
+    empty = result[0].result[0]
+    assert empty.shape == (0, 8, 8, 3) and empty.numel() == 0
+    images = torch.arange(10 * 8 * 8 * 3, dtype=torch.float32).reshape(10, 8, 8, 3)
+    for count in (1, 3, 10):
+        result = await map_node(mapping['ImageBatchDivider'], {'images': images, 'divide_by': count})
+        assert len(result[0].result) == 10
+        assert torch.equal(torch.cat(result[0].result[:count]), images)
+    audio = {'waveform': torch.arange(33, dtype=torch.float32).reshape(1, 1, 33), 'sample_rate': 24000}
+    result = await map_node(mapping['AudioBatchDivider'], {'audio': audio, 'divide_by': 3})
+    assert len(result[0].result) == 10
+    assert torch.equal(torch.cat([item['waveform'] for item in result[0].result[:3]], dim=-1), audio['waveform'])
+    # INPUT_IS_LIST must preserve all images/audio but unwrap transport scalars.
+    collector = mapping['DistributedCollector']
+    prepared, missing, hidden = execution.get_input_data(
+        {'load_balance': False, 'multi_job_id': '', 'pass_through': True}, collector, 'collector-list')
+    assert not missing
+    prepared['images'] = [images[:2], images[2:]]
+    prepared['audio'] = [audio, audio]
+    result = await execution._async_map_node_over_list(
+        'v3-acceptance', 'collector-list', collector, prepared, collector.FUNCTION, v3_data=hidden)
+    assert torch.equal(result[0].result[0], images)
+    assert result[0].result[1]['waveform'].shape[-1] == 66
+    # Exercise real async aggregation without an HTTP endpoint or worker.
+    queue = asyncio.Queue()
+    await queue.put({'worker_id': 'worker_2', 'tensor': images[2:3], 'image_index': 0, 'is_last': True})
+    prompt_server.distributed_pending_jobs['v3-aggregate'] = queue
+    result = await map_node(collector,
+                            {'images': images[:2], 'load_balance': False,
+                             'multi_job_id': 'v3-aggregate', 'enabled_worker_ids': '["worker_2"]'})
+    assert torch.equal(result[0].result[0], images[:3])
+    assert 'v3-aggregate' not in prompt_server.distributed_pending_jobs
+
+    # The registered node must forward the transfer options to the runtime
+    # helper, and default to the legacy transfer when they are not set.
+    collector_runtime = sys.modules[module.__name__ + '.nodes.collector'].DistributedCollectorNode
+    forwarded = []
+    def fake_run(self, **kwargs):
+        forwarded.append(kwargs)
+        return (kwargs['images'], kwargs['audio'])
+    with patch.object(collector_runtime, 'run', fake_run):
+        await map_node(collector, {'images': images[:1], 'load_balance': False, 'pass_through': True,
+                                   'transfer_format': 'av1', 'transfer_quality': 70})
+        await map_node(collector, {'images': images[:1], 'load_balance': False, 'pass_through': True})
+    # Set values arrive list-wrapped (is_input_list); unset ones use the
+    # execute() defaults. The runtime helper unwraps both forms.
+    assert [(kw['transfer_format'], kw['transfer_quality']) for kw in forwarded] == [
+        (['av1'], [70]), ('legacy_png', 90)], forwarded
+
+    # Prove V3's per-execution class clones do not leak helper instance state;
+    # replace only the GPU/model boundary, not the input parser or executor.
+    upscale_runtime = sys.modules[module.__name__ + '.nodes.distributed_upscale'].UltimateSDUpscaleDistributed
+    seen = []
+    def fake_upscale(self, *args):
+        assert not hasattr(self, 'acceptance_marker')
+        self.acceptance_marker = True
+        seen.append(args)
+        return (args[0],)
+    inputs = {'upscaled_image': images, 'model': object(), 'positive': [], 'negative': [],
+              'vae': object(), 'seed': 1, 'steps': 1, 'cfg': 1.0,
+              'sampler_name': 'euler', 'scheduler': 'normal', 'denoise': 0.5,
+              'tile_width': 64, 'tile_height': 64, 'padding': 0, 'mask_blur': 0,
+              'force_uniform_tiles': True, 'tiled_decode': False,
+              'multi_job_id': 'tile-job', 'is_worker': True, 'master_url': 'http://master.invalid',
+              'enabled_worker_ids': '["worker_2"]', 'worker_id': 'worker_2',
+              'tile_indices': '[2,3]', 'dynamic_threshold': 9}
+    with patch.object(upscale_runtime, 'run', fake_upscale):
+        for _ in range(2):
+            result = await map_node(mapping['UltimateSDUpscaleDistributed'], inputs)
+            assert result[0].result[0] is images
+    assert len(seen) == 2 and seen[0][-7:] == tuple(inputs[name] for name in (
+        'multi_job_id', 'is_worker', 'master_url', 'enabled_worker_ids', 'worker_id',
+        'tile_indices', 'dynamic_threshold')), seen[0]
+    import math
+    assert math.isnan(mapping['UltimateSDUpscaleDistributed'].fingerprint_inputs(multi_job_id='tile-job'))
+    assert math.isnan(mapping['UltimateSDUpscaleDistributed'].fingerprint_inputs(multi_job_id=''))
+
+    graph = {
+        '1': {'class_type': 'EmptyImage', 'inputs': {'height': 8, 'width': 8, 'batch_size': 10, 'color': 0}},
+        '2': {'class_type': 'DistributedCollector', 'inputs': {'images': ['1', 0], 'load_balance': False,
+                                                              'pass_through': True}},
+        '3': {'class_type': 'ImageBatchDivider', 'inputs': {'images': ['2', 0], 'divide_by': 10}},
+        '4': {'class_type': 'PreviewImage', 'inputs': {'images': ['3', 9]}},
+    }
+    valid = await execution.validate_prompt('v3-graph', graph, None)
+    assert valid[0], valid
+    invalid = {'1': {'class_type': 'UltimateSDUpscaleDistributed',
+                     'inputs': {**inputs, 'sampler_name': 'INVALID_ENUM', 'scheduler': 'INVALID_ENUM'}},
+               '2': {'class_type': 'PreviewImage', 'inputs': {'images': ['1', 0]}}}
+    rejected = await execution.validate_prompt('v3-invalid', invalid, None)
+    assert not rejected[0]
+    errors = [error for entry in rejected[3].values() for error in entry['errors']]
+    enum_errors = [error['extra_info']['input_name'] for error in errors if error['type'] == 'value_not_in_list']
+    assert {'sampler_name', 'scheduler'} <= set(enum_errors), errors
+    import folder_paths
+    scratch = Path(os.environ.get('TMPDIR', Path.home() / '.hermes/cache/scratch'))
+    with tempfile.TemporaryDirectory(prefix='v3-preview-', dir=scratch) as temp:
+        with patch.object(folder_paths, 'temp_directory', temp):
+            executor = execution.PromptExecutor(
+                prompt_server, cache_args={'ram': 0, 'ram_inactive': 0}, asset_manager=asset_manager)
+            await asyncio.to_thread(executor.execute, graph, 'v3-graph', {}, valid[2])
+            assert executor.success, executor.status_messages
+            history = executor.history_result
+            record = history['outputs']['4']['images'][0]
+            preview = Path(temp) / record.get('subfolder', '') / record['filename']
+            with Image.open(preview) as image:
+                assert image.size == (8, 8) and image.mode == 'RGB'
+                assert image.getextrema() == ((0, 0), (0, 0), (0, 0))
+
+    # A worker prompt terminated by the sink validates and runs through the
+    # real executor without writing anything to the temp directory.
+    sink_graph = {
+        '1': {'class_type': 'EmptyImage', 'inputs': {'height': 8, 'width': 8, 'batch_size': 4, 'color': 0}},
+        '2': {'class_type': 'DistributedCollector', 'inputs': {'images': ['1', 0], 'load_balance': False,
+                                                              'pass_through': True}},
+        '3': {'class_type': 'DistributedWorkerSink', 'inputs': {'value': ['2', 0]}},
+    }
+    valid = await execution.validate_prompt('v3-sink', sink_graph, None)
+    assert valid[0], valid
+    assert valid[2] == ['3'], valid[2]
+    collected = []
+    def record_run(self, **kwargs):
+        collected.append(kwargs['images'])
+        return (kwargs['images'][0], None)
+    with tempfile.TemporaryDirectory(prefix='v3-sink-', dir=scratch) as temp:
+        with patch.object(folder_paths, 'temp_directory', temp), patch.object(collector_runtime, 'run', record_run):
+            executor = execution.PromptExecutor(
+                prompt_server, cache_args={'ram': 0, 'ram_inactive': 0}, asset_manager=asset_manager)
+            await asyncio.to_thread(executor.execute, sink_graph, 'v3-sink', {}, valid[2])
+            assert executor.success, executor.status_messages
+            assert len(collected) == 1 and collected[0][0].shape == (4, 8, 8, 3), collected
+            assert list(Path(temp).rglob('*')) == [], list(Path(temp).rglob('*'))
+            assert executor.history_result['outputs'].get('3', {}) == {}, executor.history_result
+    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):
+        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)
+    assert not (ROOT / 'distributed.py').exists(), 'obsolete root bootstrap remains'
+    assert await nodes.load_custom_node(str(ROOT)), 'ComfyUI loader rejected the pack'
+    module = sys.modules[str(ROOT).replace('.', '_x_')]
+    assert not hasattr(module, 'NODE_CLASS_MAPPINGS'), 'V1 map shadows V3 entrypoint'
+    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)
+    assert set(mapping) == set(BASELINE) | ADDED_NODES
+    for node_id in ADDED_NODES:
+        assert nodes.NODE_CLASS_MAPPINGS[node_id] is mapping[node_id]
+    media_info = mapping['DistributedMediaCollector'].GET_NODE_INFO_V1()
+    assert media_info['output'] == ['IMAGE', 'AUDIO', 'MASK', 'STRING', 'VIDEO'], media_info
+    assert all(media_info['output_is_list']), media_info
+    assert media_info['input']['optional']['texts'][1]['forceInput'] is True
+    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():
+        if node_id in ADDED_NODES:
+            continue
+        assert issubclass(cls, io.ComfyNode)
+        assert nodes.NODE_CLASS_MAPPINGS[node_id] is cls
+        old = dict(BASELINE[node_id])
+        if node_id in ('ImageBatchDivider', 'AudioBatchDivider'):
+            # V1's ByPassTypeTuple advertises '*' when indexed, while its
+            # underlying tuple and existing frontend declare IMAGE/AUDIO.
+            # Native V3 declares all ten existing typed sockets explicitly.
+            assert old['output'] == ['*'] * 10
+            old['output'] = ['IMAGE' if node_id == 'ImageBatchDivider' else 'AUDIO'] * 10
+        new = json.loads(json.dumps(cls.GET_NODE_INFO_V1()))
+        for group in ('required', 'optional'):
+            old_inputs = old['input'].get(group, {})
+            new_inputs = dict(new['input'].get(group, {}))
+            if group == 'optional':
+                for name, expected in ADDED_OPTIONAL_INPUTS.get(node_id, {}).items():
+                    assert name in new_inputs, (node_id, name, 'added input missing')
+                    actual = normalized_input(new_inputs.pop(name))
+                    kind, opts = normalized_input(expected)
+                    assert actual[0] == kind, (node_id, name, actual)
+                    for key, value in opts.items():
+                        assert actual[1].get(key) == value, (node_id, name, key, actual)
+            assert list(old_inputs) == list(new_inputs), (node_id, group, 'input order')
+            for name, original in old_inputs.items():
+                assert normalized_input(original) == normalized_input(new_inputs[name]), (node_id, name, original, new_inputs[name])
+        for key in ('output', 'output_name', 'output_is_list', 'is_input_list', 'output_node', 'category', 'display_name'):
+            assert old[key] == new[key], (node_id, key, old[key], new[key])
+        # Standard context lives in cls.hidden; orchestrator metadata remains
+        # accepted by its original kwarg name, without creating new widgets.
+        signature = inspect.signature(cls.execute)
+        for name, field in old['input'].get('hidden', {}).items():
+            if isinstance(field, list):
+                assert cls.GET_SCHEMA().accept_all_inputs, node_id
+                assert name in signature.parameters, (node_id, name)
+                assert signature.parameters[name].default == field[1]['default'], (node_id, name)
+            else:
+                assert name in new['input']['hidden'], (node_id, name)
+        expected_hidden = []
+        if node_id == 'DistributedModelName':
+            expected_hidden.extend(['unique_id', 'extra_pnginfo'])
+        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'
+    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 [out.io_type for out in end_schema.outputs] == ['IMAGE', 'MASK', 'STRING']
+    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')
+
+
+asyncio.run(main())
@@ -0,0 +1,146 @@
+import importlib.util
+import sys
+import types
+import unittest
+from pathlib import Path
+from unittest.mock import AsyncMock, patch
+
+
+class _PromptQueue:
+    def __init__(self):
+        self.items = []
+
+    def put(self, item):
+        self.items.append(item)
+
+
+def _load_async_helpers_module():
+    module_path = Path(__file__).resolve().parents[1] / "utils" / "async_helpers.py"
+    package_name = "dist_async_helpers_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__ = []
+    sys.modules[package_name] = root_pkg
+
+    utils_pkg = types.ModuleType(f"{package_name}.utils")
+    utils_pkg.__path__ = []
+    sys.modules[f"{package_name}.utils"] = utils_pkg
+
+    execution_module = types.ModuleType("execution")
+
+    async def _validate_prompt(prompt_id, prompt, partial_execution_targets):
+        return (True, None, ["9"], {})
+
+    execution_module.validate_prompt = _validate_prompt
+    execution_module.SENSITIVE_EXTRA_DATA_KEYS = []
+    sys.modules["execution"] = execution_module
+
+    prompt_server = types.SimpleNamespace(
+        trigger_on_prompt=lambda payload: payload,
+        number=12,
+        prompt_queue=_PromptQueue(),
+    )
+    server_module = types.ModuleType("server")
+    server_module.PromptServer = types.SimpleNamespace(instance=prompt_server)
+    sys.modules["server"] = server_module
+
+    network_module = types.ModuleType(f"{package_name}.utils.network")
+    network_module.get_server_loop = lambda: None
+    sys.modules[f"{package_name}.utils.network"] = network_module
+
+    spec = importlib.util.spec_from_file_location(f"{package_name}.utils.async_helpers", 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, prompt_server
+
+
+async_helpers, prompt_server = _load_async_helpers_module()
+
+
+class QueuePromptPayloadTests(unittest.IsolatedAsyncioTestCase):
+    async def test_queue_prompt_payload_includes_create_time_and_client_metadata(self):
+        result = await async_helpers.queue_prompt_payload(
+            {"1": {"class_type": "Node"}},
+            workflow_meta={"id": "workflow-1"},
+            client_id="client-1",
+            include_queue_metadata=True,
+        )
+
+        self.assertIsInstance(result["prompt_id"], str)
+        self.assertTrue(result["prompt_id"])
+        self.assertEqual(result["number"], 12)
+        self.assertEqual(result["node_errors"], {})
+
+        self.assertEqual(prompt_server.number, 13)
+        self.assertEqual(len(prompt_server.prompt_queue.items), 1)
+        queued_item = prompt_server.prompt_queue.items[0]
+        self.assertEqual(queued_item[0], 12)
+        extra_data = queued_item[3]
+        self.assertEqual(extra_data["client_id"], "client-1")
+        self.assertIn("create_time", extra_data)
+        self.assertIsInstance(extra_data["create_time"], int)
+        self.assertGreater(extra_data["create_time"], 0)
+        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()
@@ -0,0 +1,102 @@
+import importlib.util
+import sys
+import types
+import unittest
+from pathlib import Path
+
+import torch
+
+
+def _load_utilities_module():
+    module_path = Path(__file__).resolve().parents[1] / "nodes" / "utilities.py"
+    package_name = "dist_divider_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__ = []
+    sys.modules[package_name] = root_pkg
+
+    nodes_pkg = types.ModuleType(f"{package_name}.nodes")
+    nodes_pkg.__path__ = []
+    sys.modules[f"{package_name}.nodes"] = nodes_pkg
+
+    utils_pkg = types.ModuleType(f"{package_name}.utils")
+    utils_pkg.__path__ = []
+    sys.modules[f"{package_name}.utils"] = utils_pkg
+
+    logging_module = types.ModuleType(f"{package_name}.utils.logging")
+    logging_module.debug_log = lambda *_args, **_kwargs: None
+    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}.nodes.utilities",
+        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
+
+
+utils = _load_utilities_module()
+
+
+class ImageBatchDividerTests(unittest.TestCase):
+    def test_divides_images_into_contiguous_chunks(self):
+        divider = utils.ImageBatchDivider()
+        images = torch.arange(10, dtype=torch.float32).reshape(10, 1, 1, 1)
+
+        outputs = divider.divide_batch(images, 3)
+
+        self.assertEqual(outputs[0].shape[0], 4)
+        self.assertEqual(outputs[1].shape[0], 3)
+        self.assertEqual(outputs[2].shape[0], 3)
+        self.assertEqual(outputs[0][:, 0, 0, 0].tolist(), [0.0, 1.0, 2.0, 3.0])
+        self.assertEqual(outputs[1][:, 0, 0, 0].tolist(), [4.0, 5.0, 6.0])
+        self.assertEqual(outputs[2][:, 0, 0, 0].tolist(), [7.0, 8.0, 9.0])
+
+    def test_unused_image_outputs_are_empty(self):
+        divider = utils.ImageBatchDivider()
+        images = torch.arange(4, dtype=torch.float32).reshape(4, 1, 1, 1)
+
+        outputs = divider.divide_batch(images, 2)
+
+        self.assertEqual(len(outputs), 10)
+        for idx in range(2, 10):
+            self.assertEqual(outputs[idx].shape[0], 0)
+
+
+class AudioBatchDividerTests(unittest.TestCase):
+    def test_divides_audio_samples_into_contiguous_chunks(self):
+        divider = utils.AudioBatchDivider()
+        audio = {
+            "waveform": torch.arange(10, dtype=torch.float32).reshape(1, 1, 10),
+            "sample_rate": 24000,
+        }
+
+        outputs = divider.divide_audio(audio, 3)
+
+        self.assertEqual(outputs[0]["waveform"][0, 0].tolist(), [0.0, 1.0, 2.0, 3.0])
+        self.assertEqual(outputs[1]["waveform"][0, 0].tolist(), [4.0, 5.0, 6.0])
+        self.assertEqual(outputs[2]["waveform"][0, 0].tolist(), [7.0, 8.0, 9.0])
+
+    def test_unused_audio_outputs_are_empty(self):
+        divider = utils.AudioBatchDivider()
+        audio = {
+            "waveform": torch.arange(8, dtype=torch.float32).reshape(1, 1, 8),
+            "sample_rate": 24000,
+        }
+
+        outputs = divider.divide_audio(audio, 2)
+
+        self.assertEqual(len(outputs), 10)
+        for idx in range(2, 10):
+            self.assertEqual(outputs[idx]["waveform"].shape[-1], 0)
+
+
+if __name__ == "__main__":
+    unittest.main()
@@ -0,0 +1,638 @@
+import importlib.util
+import sys
+import types
+import asyncio
+from pathlib import Path
+
+import pytest
+import torch
+
+
+def _register_real_transfer_codec(package_name, utils_pkg):
+    """Load the real utils/transfer_codec.py into the synthetic test package."""
+    codec_path = Path(__file__).resolve().parents[1] / "utils" / "transfer_codec.py"
+    codec_spec = importlib.util.spec_from_file_location(f"{package_name}.utils.transfer_codec", codec_path)
+    codec_module = importlib.util.module_from_spec(codec_spec)
+    sys.modules[f"{package_name}.utils.transfer_codec"] = codec_module
+    codec_spec.loader.exec_module(codec_module)
+    utils_pkg.transfer_codec = codec_module
+    return codec_module
+
+
+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"
+
+    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__ = []
+    sys.modules[package_name] = root_pkg
+
+    nodes_pkg = types.ModuleType(f"{package_name}.nodes")
+    nodes_pkg.__path__ = []
+    sys.modules[f"{package_name}.nodes"] = nodes_pkg
+
+    utils_pkg = types.ModuleType(f"{package_name}.utils")
+    utils_pkg.__path__ = []
+    sys.modules[f"{package_name}.utils"] = utils_pkg
+    _register_real_transfer_codec(package_name, utils_pkg)
+
+    class _Routes:
+        def post(self, _path):
+            return lambda fn: fn
+
+        def get(self, _path):
+            return lambda fn: fn
+
+    prompt_server = types.SimpleNamespace(
+        routes=_Routes(),
+        distributed_jobs_lock=None,
+        distributed_pending_jobs={},
+    )
+    server_module = types.ModuleType("server")
+    server_module.PromptServer = types.SimpleNamespace(instance=prompt_server)
+    monkeypatch.setitem(sys.modules, "server", server_module)
+
+    comfy_module = types.ModuleType("comfy")
+    model_management = types.ModuleType("comfy.model_management")
+
+    class InterruptProcessingException(Exception):
+        pass
+
+    model_management.InterruptProcessingException = InterruptProcessingException
+    model_management.throw_exception_if_processing_interrupted = lambda: None
+    comfy_module.model_management = model_management
+
+    comfy_utils = types.ModuleType("comfy.utils")
+
+    class ProgressBar:
+        def __init__(self, _total):
+            self.total = _total
+            self.updates = []
+
+        def update(self, value):
+            self.updates.append(value)
+
+    comfy_utils.ProgressBar = ProgressBar
+    comfy_module.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)
+    monkeypatch.setitem(sys.modules, "aiohttp", aiohttp_module)
+
+    logging_module = types.ModuleType(f"{package_name}.utils.logging")
+    logging_module.debug_log = lambda *_args, **_kwargs: None
+    logging_module.log = lambda *_args, **_kwargs: None
+    sys.modules[f"{package_name}.utils.logging"] = logging_module
+
+    config_module = types.ModuleType(f"{package_name}.utils.config")
+    config_module.get_worker_timeout_seconds = lambda: 0.1
+    config_module.load_config = lambda: {"workers": []}
+    config_module.is_master_delegate_only = lambda: False
+    sys.modules[f"{package_name}.utils.config"] = config_module
+
+    constants_module = types.ModuleType(f"{package_name}.utils.constants")
+    constants_module.HEARTBEAT_INTERVAL = 1.0
+    constants_module.TRANSFER_MAX_CHUNK_BYTES = 32 * 1024 * 1024
+    constants_module.TRANSFER_MAX_REQUEST_BYTES = 96 * 1024 * 1024
+    sys.modules[f"{package_name}.utils.constants"] = constants_module
+
+    image_module = types.ModuleType(f"{package_name}.utils.image")
+    def _ensure_contiguous(tensor):
+        return tensor.contiguous() if hasattr(tensor, "contiguous") else tensor
+
+    image_module.ensure_contiguous = _ensure_contiguous
+    image_module.tensor_to_pil = lambda *_args, **_kwargs: None
+    image_module.pil_to_tensor = lambda value: value
+    sys.modules[f"{package_name}.utils.image"] = image_module
+
+    network_module = types.ModuleType(f"{package_name}.utils.network")
+    network_module.build_worker_url = lambda worker: "http://worker"
+    network_module.get_client_session = lambda: None
+    network_module.probe_worker = lambda *_args, **_kwargs: None
+    sys.modules[f"{package_name}.utils.network"] = network_module
+
+    audio_payload_module = types.ModuleType(f"{package_name}.utils.audio_payload")
+    audio_payload_module.encode_audio_payload = lambda audio: audio
+    sys.modules[f"{package_name}.utils.audio_payload"] = audio_payload_module
+
+    async_helpers_module = types.ModuleType(f"{package_name}.utils.async_helpers")
+    async_helpers_module.run_async_in_server_loop = lambda coro: asyncio.run(coro)
+    sys.modules[f"{package_name}.utils.async_helpers"] = async_helpers_module
+
+    spec = importlib.util.spec_from_file_location(f"{package_name}.nodes.collector", 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
+
+
+def test_collector_opts_into_comfyui_list_inputs():
+    collector = _load_collector_module().DistributedCollectorNode
+
+    assert collector.INPUT_IS_LIST is True
+
+
+def test_collector_exposes_images_as_optional_input():
+    input_types = _load_collector_module().DistributedCollectorNode.INPUT_TYPES()
+
+    assert "images" not in input_types["required"]
+    assert input_types["optional"]["images"] == ("IMAGE",)
+
+
+def test_audio_only_pass_through_returns_no_images_and_preserves_audio():
+    collector = _load_collector_module().DistributedCollectorNode()
+    audio = {"waveform": torch.ones(1, 2, 4), "sample_rate": 48000}
+
+    images, returned_audio = collector.run(images=None, audio=[audio])
+
+    assert images is None
+    assert returned_audio is audio
+
+
+def test_collector_rejects_missing_images_and_audio():
+    collector = _load_collector_module().DistributedCollectorNode()
+
+    try:
+        collector.run(images=None, audio=None)
+    except ValueError as exc:
+        assert "image or audio" in str(exc).lower()
+    else:
+        raise AssertionError("Expected collector to reject a run with no media input")
+
+
+def test_delegate_only_master_allows_no_local_media_input():
+    collector = _load_collector_module().DistributedCollectorNode()
+
+    images, audio = collector.run(
+        images=None,
+        audio=None,
+        multi_job_id=["delegate-audio-job"],
+        delegate_only=[True],
+        enabled_worker_ids=["[]"],
+    )
+
+    assert images is None
+    assert tuple(audio["waveform"].shape) == (1, 2, 1)
+
+
+def test_pass_through_collapses_comfyui_image_list_to_batch_and_unwraps_hidden_inputs():
+    collector = _load_collector_module().DistributedCollectorNode()
+    first = torch.zeros(1, 2, 2, 3)
+    second = torch.ones(1, 2, 2, 3)
+
+    images, audio = collector.run(
+        images=[first, second],
+        load_balance=[False],
+        audio=[None],
+        multi_job_id=[""],
+        is_worker=[False],
+        master_url=[""],
+        enabled_worker_ids=["[]"],
+        worker_batch_size=[1],
+        worker_id=[""],
+        pass_through=[False],
+        delegate_only=[False],
+    )
+
+    assert tuple(images.shape) == (2, 2, 2, 3)
+    assert torch.equal(images[0:1], first)
+    assert torch.equal(images[1:2], second)
+    assert tuple(audio["waveform"].shape) == (1, 2, 1)
+
+
+def test_worker_list_input_sends_one_completion_sequence_with_last_only_on_final_item():
+    module = _load_collector_module()
+    collector = module.DistributedCollectorNode()
+    first = torch.zeros(1, 2, 2, 3)
+    second = torch.ones(1, 2, 2, 3)
+    posted_payloads = []
+
+    class _FakeImage:
+        def save(self, fp, format=None, compress_level=None):
+            fp.write(b"png-bytes")
+
+    class _FakeResponse:
+        async def __aenter__(self):
+            return self
+
+        async def __aexit__(self, exc_type, exc, tb):
+            return False
+
+        def raise_for_status(self):
+            return None
+
+    class _FakeSession:
+        def post(self, url, json, timeout):
+            posted_payloads.append(json)
+            return _FakeResponse()
+
+    async def _fake_get_client_session():
+        return _FakeSession()
+
+    module.tensor_to_pil = lambda *_args, **_kwargs: _FakeImage()
+    module.get_client_session = _fake_get_client_session
+    module.encode_audio_payload = lambda _audio: None
+
+    images, audio = collector.run(
+        images=[first, second],
+        load_balance=[False],
+        audio=[None],
+        multi_job_id=["job-list-1"],
+        is_worker=[True],
+        master_url=["http://master"],
+        enabled_worker_ids=["[]"],
+        worker_batch_size=[1],
+        worker_id=["worker-a"],
+        pass_through=[False],
+        delegate_only=[False],
+    )
+
+    assert tuple(images.shape) == (2, 2, 2, 3)
+    assert tuple(audio["waveform"].shape) == (1, 2, 1)
+    assert len(posted_payloads) == 2
+    assert [payload["batch_idx"] for payload in posted_payloads] == [0, 1]
+    assert [payload["is_last"] for payload in posted_payloads] == [False, True]
+    assert {payload["job_id"] for payload in posted_payloads} == {"job-list-1"}
+    assert {payload["worker_id"] for payload in posted_payloads} == {"worker-a"}
+
+
+def test_audio_only_worker_sends_one_completion_without_image():
+    module = _load_collector_module()
+    collector = module.DistributedCollectorNode()
+    audio = {"waveform": torch.ones(1, 2, 4), "sample_rate": 48000}
+    posted = []
+
+    class _FakeResponse:
+        async def __aenter__(self):
+            return self
+
+        async def __aexit__(self, exc_type, exc, tb):
+            return False
+
+        def raise_for_status(self):
+            return None
+
+    class _FakeSession:
+        def post(self, url, json, timeout):
+            posted.append((url, json, timeout.total))
+            return _FakeResponse()
+
+    async def _fake_get_client_session():
+        return _FakeSession()
+
+    module.get_client_session = _fake_get_client_session
+    module.encode_audio_payload = lambda value: {"encoded": value is audio}
+
+    images, returned_audio = collector.run(
+        images=None,
+        audio=[audio],
+        multi_job_id=["audio-job"],
+        is_worker=[True],
+        master_url=["http://master"],
+        worker_id=["worker-a"],
+    )
+
+    assert images is None
+    assert returned_audio is audio
+    assert len(posted) == 1
+    assert posted[0][0] == "http://master/distributed/job_complete"
+    assert posted[0][2] == 600
+    assert posted[0][1] == {
+        "job_id": "audio-job",
+        "worker_id": "worker-a",
+        "batch_idx": 0,
+        "audio": {"encoded": True},
+        "is_last": True,
+    }
+
+
+def test_audio_only_master_combines_local_and_worker_audio():
+    module = _load_collector_module()
+    collector = module.DistributedCollectorNode()
+    master_audio = {"waveform": torch.ones(1, 2, 2), "sample_rate": 48000}
+    worker_audio = {"waveform": torch.full((1, 2, 3), 2.0), "sample_rate": 48000}
+    module.prompt_server.distributed_jobs_lock = asyncio.Lock()
+    queue = asyncio.Queue()
+    queue.put_nowait(
+        {
+            "worker_id": "worker-a",
+            "image_index": 0,
+            "tensor": None,
+            "audio": worker_audio,
+            "is_last": True,
+        }
+    )
+    module.prompt_server.distributed_pending_jobs = {"audio-job": queue}
+
+    images, combined_audio = asyncio.run(
+        collector.execute(
+            images=None,
+            audio=master_audio,
+            multi_job_id="audio-job",
+            enabled_worker_ids='["worker-a"]',
+        )
+    )
+
+    assert images is None
+    assert combined_audio["sample_rate"] == 48000
+    assert tuple(combined_audio["waveform"].shape) == (1, 2, 5)
+    assert torch.equal(combined_audio["waveform"][..., :2], master_audio["waveform"])
+    assert torch.equal(combined_audio["waveform"][..., 2:], worker_audio["waveform"])
+
+
+def test_delegate_only_audio_collects_worker_audio_without_placeholder_image():
+    module = _load_collector_module()
+    collector = module.DistributedCollectorNode()
+    worker_audio = {"waveform": torch.full((1, 2, 3), 2.0), "sample_rate": 48000}
+    module.prompt_server.distributed_jobs_lock = asyncio.Lock()
+    queue = asyncio.Queue()
+    queue.put_nowait(
+        {
+            "worker_id": "worker-a",
+            "image_index": 0,
+            "tensor": None,
+            "audio": worker_audio,
+            "is_last": True,
+        }
+    )
+    module.prompt_server.distributed_pending_jobs = {"delegate-audio-job": queue}
+
+    images, combined_audio = asyncio.run(
+        collector.execute(
+            images=None,
+            audio=None,
+            multi_job_id="delegate-audio-job",
+            enabled_worker_ids='["worker-a"]',
+            delegate_only=True,
+        )
+    )
+
+    assert images is None
+    assert combined_audio["sample_rate"] == 48000
+    assert torch.equal(combined_audio["waveform"], worker_audio["waveform"])
+
+
+# --------------------------------------------------------------------------- #
+# Compressed worker -> master transfer
+# --------------------------------------------------------------------------- #
+
+class _RecordingResponse:
+    def __init__(self, status=200, text=""):
+        self.status = status
+        self._text = text
+
+    async def __aenter__(self):
+        return self
+
+    async def __aexit__(self, exc_type, exc, tb):
+        return False
+
+    async def text(self):
+        return self._text
+
+    def raise_for_status(self):
+        if self.status >= 400:
+            raise RuntimeError(f"HTTP {self.status}")
+
+
+class _RecordingSession:
+    def __init__(self, statuses=None, text="Invalid JSON payload"):
+        self.calls = []
+        self._statuses = list(statuses or [])
+        self._text = text
+
+    def post(self, url, timeout, json=None, data=None):
+        self.calls.append({"url": url, "json": json, "data": data, "timeout": timeout.total})
+        status = self._statuses.pop(0) if self._statuses else 200
+        return _RecordingResponse(status=status, text=self._text)
+
+
+def _install_fake_transport(module, session):
+    async def _fake_get_client_session():
+        return session
+
+    module.get_client_session = _fake_get_client_session
+    # Record the multipart content instead of building a real aiohttp.FormData.
+    module.DistributedCollectorNode._build_multipart = staticmethod(
+        lambda envelope, parts: {"envelope": envelope, "parts": list(parts)}
+    )
+
+
+def _run_worker(collector, images, audio=None, **kwargs):
+    return collector.run(
+        images=[images],
+        audio=[audio],
+        multi_job_id=["job-c"],
+        is_worker=[True],
+        master_url=["http://master"],
+        worker_id=["worker-a"],
+        **{key: [value] for key, value in kwargs.items()},
+    )
+
+
+def test_collector_exposes_transfer_widgets_with_legacy_default():
+    module = _load_collector_module()
+    optional = module.DistributedCollectorNode.INPUT_TYPES()["optional"]
+
+    formats, options = optional["transfer_format"]
+    assert "legacy_png" in formats and "h264" in formats and "ffv1" in formats
+    assert options["default"] == "legacy_png"
+    assert optional["transfer_quality"][0] == "INT"
+
+
+def test_worker_sends_video_batch_as_one_multipart_chunk_with_audio_on_last():
+    module = _load_collector_module()
+    collector = module.DistributedCollectorNode()
+    session = _RecordingSession()
+    _install_fake_transport(module, session)
+    module.encode_audio_payload = lambda value: None if value is None else {"encoded": True}
+    images = torch.rand(6, 32, 48, 3)
+
+    _run_worker(
+        collector,
+        images,
+        audio={"waveform": torch.ones(1, 2, 4), "sample_rate": 48000},
+        transfer_format="ffv1",
+        transfer_quality=90,
+    )
+
+    assert len(session.calls) == 1
+    call = session.calls[0]
+    assert call["url"] == "http://master/distributed/job_complete"
+    assert call["json"] is None
+    envelope = call["data"]["envelope"]
+    assert envelope["format"] == "ffv1"
+    assert (envelope["start_index"], envelope["count"]) == (0, 6)
+    assert (envelope["width"], envelope["height"]) == (48, 32)
+    assert envelope["is_last"] is True
+    assert envelope["audio"] == {"encoded": True}
+    assert envelope["job_id"] == "job-c" and envelope["worker_id"] == "worker-a"
+    assert len(call["data"]["parts"]) == 1
+
+
+def test_worker_marks_only_final_chunk_as_last():
+    module = _load_collector_module()
+    collector = module.DistributedCollectorNode()
+    session = _RecordingSession()
+    _install_fake_transport(module, session)
+    module.TRANSFER_MAX_CHUNK_BYTES = 1  # force one frame per request
+
+    _run_worker(collector, torch.rand(3, 8, 8, 3), transfer_format="png")
+
+    envelopes = [call["data"]["envelope"] for call in session.calls]
+    assert [(e["start_index"], e["count"], e["is_last"]) for e in envelopes] == [
+        (0, 1, False),
+        (1, 1, False),
+        (2, 1, True),
+    ]
+    assert "audio" not in envelopes[-1]
+
+
+def test_worker_falls_back_to_legacy_png_when_master_rejects_multipart():
+    from PIL import Image
+    import numpy as np
+
+    module = _load_collector_module()
+    collector = module.DistributedCollectorNode()
+    session = _RecordingSession(statuses=[400])
+    _install_fake_transport(module, session)
+    module.tensor_to_pil = lambda tensor, idx: Image.fromarray(
+        (255 * tensor[idx].cpu().numpy()).astype(np.uint8)
+    )
+
+    _run_worker(collector, torch.rand(2, 8, 8, 3), transfer_format="h264")
+
+    assert session.calls[0]["data"] is not None  # compressed attempt
+    legacy = [call["json"] for call in session.calls[1:]]
+    assert [payload["batch_idx"] for payload in legacy] == [0, 1]
+    assert all(payload["image"].startswith("data:image/png;base64,") for payload in legacy)
+    assert [payload["is_last"] for payload in legacy] == [False, True]
+
+
+def test_worker_uses_legacy_path_for_unknown_transfer_format():
+    from PIL import Image
+    import numpy as np
+
+    module = _load_collector_module()
+    collector = module.DistributedCollectorNode()
+    session = _RecordingSession()
+    _install_fake_transport(module, session)
+    module.tensor_to_pil = lambda tensor, idx: Image.fromarray(
+        (255 * tensor[idx].cpu().numpy()).astype(np.uint8)
+    )
+
+    _run_worker(collector, torch.rand(1, 8, 8, 3), transfer_format="gif")
+
+    assert len(session.calls) == 1
+    assert session.calls[0]["json"]["image"].startswith("data:image/png;base64,")
+
+
+def test_worker_falls_back_when_video_encoder_missing():
+    module = _load_collector_module()
+    collector = module.DistributedCollectorNode()
+    session = _RecordingSession()
+    _install_fake_transport(module, session)
+    codec = module.transfer_codec
+    original = codec.video_codec_available
+    codec.video_codec_available = lambda fmt: fmt == "h264" and original("h264")
+    try:
+        _run_worker(collector, torch.rand(2, 16, 16, 3), transfer_format="av1")
+    finally:
+        codec.video_codec_available = original
+
+    assert session.calls[0]["data"]["envelope"]["format"] == "h264"
+
+
+def test_worker_does_not_fall_back_on_non_legacy_rejection():
+    import pytest
+
+    module = _load_collector_module()
+    collector = module.DistributedCollectorNode()
+    session = _RecordingSession(statuses=[400], text="Invalid multipart payload: Request Entity Too Large")
+    _install_fake_transport(module, session)
+
+    with pytest.raises(RuntimeError, match="MAX_REQUEST_BYTES"):
+        _run_worker(collector, torch.rand(2, 8, 8, 3), transfer_format="png")
+
+    # One compressed attempt, no legacy resend of the batch.
+    assert len(session.calls) == 1
+    assert session.calls[0]["json"] is None
+
+
+def test_worker_sends_nothing_when_a_frame_cannot_fit():
+    import numpy as np
+    import pytest
+
+    module = _load_collector_module()
+    collector = module.DistributedCollectorNode()
+    session = _RecordingSession()
+    _install_fake_transport(module, session)
+    reserve = module.DistributedCollectorNode.TRANSFER_ENVELOPE_RESERVE_BYTES
+    module.TRANSFER_MAX_REQUEST_BYTES = reserve + 32 * 1024
+    module.TRANSFER_MAX_CHUNK_BYTES = 32 * 1024
+    # Frame 0 is tiny; frame 2 is noise that cannot fit in 32 KiB as PNG.
+    images = torch.zeros(3, 128, 128, 3)
+    images[2] = torch.from_numpy(np.random.default_rng(1).random((128, 128, 3), dtype=np.float32))
+
+    with pytest.raises(module.transfer_codec.FrameTooLargeError):
+        _run_worker(collector, images, transfer_format="png")
+
+    assert session.calls == []
+
+
+def test_worker_sends_audio_separately_when_it_would_overflow_last_request():
+    module = _load_collector_module()
+    collector = module.DistributedCollectorNode()
+    session = _RecordingSession()
+    _install_fake_transport(module, session)
+    reserve = module.DistributedCollectorNode.TRANSFER_ENVELOPE_RESERVE_BYTES
+    module.TRANSFER_MAX_REQUEST_BYTES = reserve + 64 * 1024
+    module.TRANSFER_MAX_CHUNK_BYTES = 64 * 1024
+    module.encode_audio_payload = lambda value: None if value is None else {"data": "x" * (128 * 1024)}
+
+    _run_worker(
+        collector,
+        torch.zeros(2, 16, 16, 3),
+        audio={"waveform": torch.ones(1, 2, 4), "sample_rate": 48000},
+        transfer_format="png",
+    )
+
+    multipart = [call["data"]["envelope"] for call in session.calls if call["data"] is not None]
+    assert [e["is_last"] for e in multipart] == [False]
+    assert all("audio" not in e for e in multipart)
+    final = session.calls[-1]["json"]
+    assert final["is_last"] is True and final["audio"]["data"].startswith("x")
+    assert "image" not in final
+
+
+def test_worker_attaches_audio_to_last_chunk_when_it_fits():
+    module = _load_collector_module()
+    collector = module.DistributedCollectorNode()
+    session = _RecordingSession()
+    _install_fake_transport(module, session)
+    module.encode_audio_payload = lambda value: None if value is None else {"data": "small"}
+
+    _run_worker(
+        collector,
+        torch.zeros(2, 16, 16, 3),
+        audio={"waveform": torch.ones(1, 2, 4), "sample_rate": 48000},
+        transfer_format="png",
+    )
+
+    assert len(session.calls) == 1
+    envelope = session.calls[0]["data"]["envelope"]
+    assert envelope["is_last"] is True and envelope["audio"] == {"data": "small"}
@@ -0,0 +1,293 @@
+import importlib.util
+import json
+import os
+import sys
+import tempfile
+import types
+import unittest
+from pathlib import Path
+from unittest.mock import patch
+
+
+def _load_config_module():
+    module_path = Path(__file__).resolve().parents[1] / "utils" / "config.py"
+    package_name = "dist_cfg_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 = _load_config_module()
+
+
+# ---------------------------------------------------------------------------
+# _merge_with_defaults
+# ---------------------------------------------------------------------------
+
+class MergeWithDefaultsTests(unittest.TestCase):
+    def test_non_dict_input_returns_defaults(self):
+        result = config._merge_with_defaults("not a dict", {"key": "default"})
+        self.assertEqual(result, {"key": "default"})
+
+    def test_fills_missing_keys_with_defaults(self):
+        result = config._merge_with_defaults({}, {"a": 1, "b": 2})
+        self.assertEqual(result, {"a": 1, "b": 2})
+
+    def test_loaded_value_overrides_default(self):
+        result = config._merge_with_defaults({"a": 99}, {"a": 1, "b": 2})
+        self.assertEqual(result["a"], 99)
+        self.assertEqual(result["b"], 2)
+
+    def test_nested_dict_merges_recursively(self):
+        defaults = {"settings": {"debug": False, "count": 5}}
+        loaded = {"settings": {"debug": True}}
+        result = config._merge_with_defaults(loaded, defaults)
+        self.assertTrue(result["settings"]["debug"])
+        self.assertEqual(result["settings"]["count"], 5)
+
+    def test_preserves_unknown_keys_for_forward_compatibility(self):
+        result = config._merge_with_defaults({"extra_key": "extra"}, {"a": 1})
+        self.assertEqual(result["extra_key"], "extra")
+
+    def test_none_loaded_value_overrides_default(self):
+        """Explicitly set None in config should override non-None default."""
+        result = config._merge_with_defaults({"a": None}, {"a": "default"})
+        self.assertIsNone(result["a"])
+
+    def test_non_dict_nested_loaded_value_replaces_dict_default(self):
+        """If loaded has a scalar where default has a dict, use the scalar."""
+        defaults = {"settings": {"debug": False}}
+        loaded = {"settings": "flat_string"}
+        result = config._merge_with_defaults(loaded, defaults)
+        self.assertEqual(result["settings"], "flat_string")
+
+
+# ---------------------------------------------------------------------------
+# load_config
+# ---------------------------------------------------------------------------
+
+class LoadConfigTests(unittest.TestCase):
+    def setUp(self):
+        config.invalidate_config_cache()
+
+    def tearDown(self):
+        config.invalidate_config_cache()
+
+    def test_returns_defaults_when_file_missing(self):
+        with patch.object(config, "CONFIG_FILE", "/nonexistent/path/config.json"):
+            cfg = config.load_config()
+        defaults = config.get_default_config()
+        self.assertEqual(cfg["settings"]["debug"], defaults["settings"]["debug"])
+        self.assertIn("workers", cfg)
+
+    def test_loads_valid_json_file(self):
+        data = {
+            "workers": [{"id": "w1"}],
+            "master": {"host": "test.host"},
+            "settings": {},
+            "tunnel": {},
+        }
+        with tempfile.NamedTemporaryFile("w", suffix=".json", delete=False, encoding="utf-8") as f:
+            json.dump(data, f)
+            tmp_path = f.name
+        try:
+            with patch.object(config, "CONFIG_FILE", tmp_path):
+                cfg = config.load_config()
+            self.assertEqual(cfg["master"]["host"], "test.host")
+            self.assertEqual(len(cfg["workers"]), 1)
+        finally:
+            os.unlink(tmp_path)
+
+    def test_merges_loaded_file_with_defaults(self):
+        """Loaded file with partial settings should be filled in from defaults."""
+        data = {"master": {"host": "h"}, "workers": [], "settings": {"debug": True}, "tunnel": {}}
+        with tempfile.NamedTemporaryFile("w", suffix=".json", delete=False, encoding="utf-8") as f:
+            json.dump(data, f)
+            tmp_path = f.name
+        try:
+            with patch.object(config, "CONFIG_FILE", tmp_path):
+                cfg = config.load_config()
+            # debug was set to True
+            self.assertTrue(cfg["settings"]["debug"])
+            # auto_launch_workers is a default key and should be present
+            self.assertIn("auto_launch_workers", cfg["settings"])
+        finally:
+            os.unlink(tmp_path)
+
+    def test_falls_back_to_defaults_on_invalid_json(self):
+        with tempfile.NamedTemporaryFile("w", suffix=".json", delete=False, encoding="utf-8") as f:
+            f.write("{invalid json{{")
+            tmp_path = f.name
+        try:
+            with patch.object(config, "CONFIG_FILE", tmp_path):
+                cfg = config.load_config()
+            self.assertIn("settings", cfg)
+            self.assertIn("workers", cfg)
+        finally:
+            os.unlink(tmp_path)
+
+    def test_second_call_returns_cached_object(self):
+        with tempfile.NamedTemporaryFile("w", suffix=".json", delete=False, encoding="utf-8") as f:
+            json.dump(config.get_default_config(), f)
+            tmp_path = f.name
+        try:
+            with patch.object(config, "CONFIG_FILE", tmp_path):
+                cfg1 = config.load_config()
+                cfg2 = config.load_config()
+            self.assertIs(cfg1, cfg2)
+        finally:
+            os.unlink(tmp_path)
+
+    def test_invalidate_cache_forces_reload(self):
+        data = config.get_default_config()
+        data["master"]["host"] = "first"
+        with tempfile.NamedTemporaryFile("w", suffix=".json", delete=False, encoding="utf-8") as f:
+            json.dump(data, f)
+            tmp_path = f.name
+        try:
+            with patch.object(config, "CONFIG_FILE", tmp_path):
+                cfg1 = config.load_config()
+                config.invalidate_config_cache()
+                data["master"]["host"] = "second"
+                with open(tmp_path, "w", encoding="utf-8") as fh:
+                    json.dump(data, fh)
+                cfg2 = config.load_config()
+            self.assertEqual(cfg1["master"]["host"], "first")
+            self.assertEqual(cfg2["master"]["host"], "second")
+        finally:
+            os.unlink(tmp_path)
+
+
+# ---------------------------------------------------------------------------
+# save_config
+# ---------------------------------------------------------------------------
+
+class SaveConfigTests(unittest.TestCase):
+    def setUp(self):
+        config.invalidate_config_cache()
+
+    def tearDown(self):
+        config.invalidate_config_cache()
+
+    def test_saves_and_reloads_correctly(self):
+        data = config.get_default_config()
+        data["master"]["host"] = "saved.host"
+        with tempfile.TemporaryDirectory() as tmpdir:
+            tmp_path = os.path.join(tmpdir, "config.json")
+            with patch.object(config, "CONFIG_FILE", tmp_path):
+                result = config.save_config(data)
+                self.assertTrue(result)
+                loaded = config.load_config()
+        self.assertEqual(loaded["master"]["host"], "saved.host")
+
+    def test_returns_false_when_path_unwritable(self):
+        with patch.object(config, "CONFIG_FILE", "/nonexistent_dir/config.json"):
+            result = config.save_config({})
+        self.assertFalse(result)
+
+    def test_save_invalidates_cache(self):
+        """After saving, the cache should be cleared so next load re-reads."""
+        with tempfile.TemporaryDirectory() as tmpdir:
+            tmp_path = os.path.join(tmpdir, "config.json")
+            with patch.object(config, "CONFIG_FILE", tmp_path):
+                data = config.get_default_config()
+                config.save_config(data)
+                # Cache is now None; load_config should re-read
+                self.assertIsNone(config._config_cache)
+
+    def test_written_file_is_valid_json(self):
+        data = config.get_default_config()
+        with tempfile.TemporaryDirectory() as tmpdir:
+            tmp_path = os.path.join(tmpdir, "config.json")
+            with patch.object(config, "CONFIG_FILE", tmp_path):
+                config.save_config(data)
+            with open(tmp_path, encoding="utf-8") as fh:
+                parsed = json.load(fh)
+        self.assertEqual(parsed["master"], data["master"])
+
+
+# ---------------------------------------------------------------------------
+# get_worker_timeout_seconds
+# ---------------------------------------------------------------------------
+
+class GetWorkerTimeoutSecondsTests(unittest.TestCase):
+    def test_returns_configured_value(self):
+        cfg = config.get_default_config()
+        cfg["settings"]["worker_timeout_seconds"] = 120
+        with patch.object(config, "load_config", return_value=cfg):
+            self.assertEqual(config.get_worker_timeout_seconds(), 120)
+
+    def test_clamps_zero_to_one(self):
+        cfg = config.get_default_config()
+        cfg["settings"]["worker_timeout_seconds"] = 0
+        with patch.object(config, "load_config", return_value=cfg):
+            self.assertEqual(config.get_worker_timeout_seconds(), 1)
+
+    def test_clamps_negative_to_one(self):
+        cfg = config.get_default_config()
+        cfg["settings"]["worker_timeout_seconds"] = -10
+        with patch.object(config, "load_config", return_value=cfg):
+            self.assertEqual(config.get_worker_timeout_seconds(), 1)
+
+    def test_falls_back_to_provided_default_when_key_missing(self):
+        cfg = config.get_default_config()
+        # worker_timeout_seconds is not present in default config
+        cfg["settings"].pop("worker_timeout_seconds", None)
+        with patch.object(config, "load_config", return_value=cfg):
+            result = config.get_worker_timeout_seconds(default=45)
+        self.assertEqual(result, 45)
+
+    def test_fallback_also_clamped_to_one(self):
+        cfg = config.get_default_config()
+        cfg["settings"].pop("worker_timeout_seconds", None)
+        with patch.object(config, "load_config", return_value=cfg):
+            result = config.get_worker_timeout_seconds(default=0)
+        self.assertEqual(result, 1)
+
+
+# ---------------------------------------------------------------------------
+# is_master_delegate_only
+# ---------------------------------------------------------------------------
+
+class IsMasterDelegateOnlyTests(unittest.TestCase):
+    def test_returns_false_by_default(self):
+        cfg = config.get_default_config()
+        with patch.object(config, "load_config", return_value=cfg):
+            self.assertFalse(config.is_master_delegate_only())
+
+    def test_returns_true_when_enabled(self):
+        cfg = config.get_default_config()
+        cfg["settings"]["master_delegate_only"] = True
+        with patch.object(config, "load_config", return_value=cfg):
+            self.assertTrue(config.is_master_delegate_only())
+
+    def test_returns_false_on_exception(self):
+        def _raise():
+            raise RuntimeError("config exploded")
+
+        with patch.object(config, "load_config", side_effect=RuntimeError("boom")):
+            self.assertFalse(config.is_master_delegate_only())
+
+
+if __name__ == "__main__":
+    unittest.main()
@@ -0,0 +1,196 @@
+import importlib.util
+import sys
+import types
+import unittest
+from pathlib import Path
+from unittest.mock import patch
+
+
+def _load_detection_module():
+    module_path = Path(__file__).resolve().parents[1] / "workers" / "detection.py"
+    package_name = "dist_det_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__ = []
+    sys.modules[package_name] = root_pkg
+
+    workers_pkg = types.ModuleType(f"{package_name}.workers")
+    workers_pkg.__path__ = []
+    sys.modules[f"{package_name}.workers"] = workers_pkg
+
+    utils_pkg = types.ModuleType(f"{package_name}.utils")
+    utils_pkg.__path__ = []
+    sys.modules[f"{package_name}.utils"] = utils_pkg
+
+    logging_module = types.ModuleType(f"{package_name}.utils.logging")
+    logging_module.debug_log = lambda *_args, **_kwargs: None
+    sys.modules[f"{package_name}.utils.logging"] = logging_module
+
+    network_module = types.ModuleType(f"{package_name}.utils.network")
+    network_module.normalize_host = lambda value: value
+
+    async def _fake_session():
+        raise RuntimeError("network calls not used in these tests")
+
+    network_module.get_client_session = _fake_session
+    sys.modules[f"{package_name}.utils.network"] = network_module
+
+    created_aiohttp_stub = False
+    if "aiohttp" not in sys.modules:
+        created_aiohttp_stub = True
+        aiohttp_module = types.ModuleType("aiohttp")
+
+        class _ClientTimeout:
+            def __init__(self, total=None):
+                pass
+
+        aiohttp_module.ClientTimeout = _ClientTimeout
+        sys.modules["aiohttp"] = aiohttp_module
+
+    spec = importlib.util.spec_from_file_location(
+        f"{package_name}.workers.detection",
+        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)
+
+    if created_aiohttp_stub:
+        sys.modules.pop("aiohttp", None)
+
+    return module
+
+
+detection = _load_detection_module()
+
+
+# ---------------------------------------------------------------------------
+# is_docker_environment
+# ---------------------------------------------------------------------------
+
+class IsDockerEnvironmentTests(unittest.TestCase):
+    def test_true_when_dockerenv_file_exists(self):
+        with patch.object(detection.os.path, "exists", return_value=True), \
+             patch.dict(detection.os.environ, {}, clear=True), \
+             patch.object(detection.platform, "node", return_value="my-laptop"):
+            self.assertTrue(detection.is_docker_environment())
+
+    def test_true_when_docker_container_env_var_is_set(self):
+        with patch.object(detection.os.path, "exists", return_value=False), \
+             patch.dict(detection.os.environ, {"DOCKER_CONTAINER": "1"}, clear=True), \
+             patch.object(detection.platform, "node", return_value="my-laptop"):
+            self.assertTrue(detection.is_docker_environment())
+
+    def test_true_when_platform_node_contains_docker(self):
+        with patch.object(detection.os.path, "exists", return_value=False), \
+             patch.dict(detection.os.environ, {}, clear=True), \
+             patch.object(detection.platform, "node", return_value="my-docker-host"):
+            self.assertTrue(detection.is_docker_environment())
+
+    def test_false_when_none_of_the_signals_are_present(self):
+        with patch.object(detection.os.path, "exists", return_value=False), \
+             patch.dict(detection.os.environ, {}, clear=True), \
+             patch.object(detection.platform, "node", return_value="my-laptop"):
+            self.assertFalse(detection.is_docker_environment())
+
+    def test_docker_node_name_is_case_insensitive(self):
+        with patch.object(detection.os.path, "exists", return_value=False), \
+             patch.dict(detection.os.environ, {}, clear=True), \
+             patch.object(detection.platform, "node", return_value="My-Docker-Box"):
+            self.assertTrue(detection.is_docker_environment())
+
+    def test_docker_env_var_empty_string_is_falsy(self):
+        """An empty DOCKER_CONTAINER env var should NOT trigger docker detection."""
+        with patch.object(detection.os.path, "exists", return_value=False), \
+             patch.dict(detection.os.environ, {"DOCKER_CONTAINER": ""}, clear=True), \
+             patch.object(detection.platform, "node", return_value="my-laptop"):
+            self.assertFalse(detection.is_docker_environment())
+
+
+# ---------------------------------------------------------------------------
+# is_runpod_environment
+# ---------------------------------------------------------------------------
+
+class IsRunpodEnvironmentTests(unittest.TestCase):
+    def test_true_when_runpod_pod_id_is_set(self):
+        with patch.dict(detection.os.environ, {"RUNPOD_POD_ID": "pod-abc"}, clear=True):
+            self.assertTrue(detection.is_runpod_environment())
+
+    def test_true_when_runpod_api_key_is_set(self):
+        with patch.dict(detection.os.environ, {"RUNPOD_API_KEY": "key-xyz"}, clear=True):
+            self.assertTrue(detection.is_runpod_environment())
+
+    def test_true_when_both_vars_are_set(self):
+        with patch.dict(
+            detection.os.environ,
+            {"RUNPOD_POD_ID": "pod-abc", "RUNPOD_API_KEY": "key-xyz"},
+            clear=True,
+        ):
+            self.assertTrue(detection.is_runpod_environment())
+
+    def test_false_when_neither_var_is_set(self):
+        with patch.dict(detection.os.environ, {}, clear=True):
+            self.assertFalse(detection.is_runpod_environment())
+
+    def test_true_when_pod_id_is_empty_string(self):
+        """is not None check means even empty string counts as detected."""
+        with patch.dict(detection.os.environ, {"RUNPOD_POD_ID": ""}, clear=True):
+            self.assertTrue(detection.is_runpod_environment())
+
+
+# ---------------------------------------------------------------------------
+# is_local_worker (synchronous paths only)
+# ---------------------------------------------------------------------------
+
+class IsLocalWorkerTests(unittest.IsolatedAsyncioTestCase):
+    async def test_true_for_localhost_host(self):
+        result = await detection.is_local_worker({"host": "localhost", "port": 8188})
+        self.assertTrue(result)
+
+    async def test_true_for_127_0_0_1(self):
+        result = await detection.is_local_worker({"host": "127.0.0.1", "port": 8188})
+        self.assertTrue(result)
+
+    async def test_true_for_0_0_0_0(self):
+        result = await detection.is_local_worker({"host": "0.0.0.0", "port": 8188})
+        self.assertTrue(result)
+
+    async def test_true_when_type_is_local(self):
+        result = await detection.is_local_worker({"type": "local", "host": "remote.example.com"})
+        self.assertTrue(result)
+
+    async def test_false_for_remote_host(self):
+        result = await detection.is_local_worker({"host": "remote.example.com", "port": 8188})
+        self.assertFalse(result)
+
+    async def test_true_when_no_host_key(self):
+        """Missing host defaults to 'localhost'."""
+        result = await detection.is_local_worker({"port": 8188})
+        self.assertTrue(result)
+
+
+# ---------------------------------------------------------------------------
+# get_machine_id
+# ---------------------------------------------------------------------------
+
+class GetMachineIdTests(unittest.TestCase):
+    def test_returns_a_string(self):
+        result = detection.get_machine_id()
+        self.assertIsInstance(result, str)
+
+    def test_returns_non_empty_string(self):
+        result = detection.get_machine_id()
+        self.assertTrue(len(result) > 0)
+
+    def test_stable_across_calls(self):
+        r1 = detection.get_machine_id()
+        r2 = detection.get_machine_id()
+        self.assertEqual(r1, r2)
+
+
+if __name__ == "__main__":
+    unittest.main()
@@ -0,0 +1,574 @@
+import asyncio
+import importlib.util
+import sys
+import types
+import unittest
+from pathlib import Path
+from unittest.mock import AsyncMock, patch
+
+
+def _load_dispatch_module():
+    module_path = Path(__file__).resolve().parents[1] / "api" / "orchestration" / "dispatch.py"
+
+    package_name = "dist_dispatch_testpkg"
+    root_pkg = types.ModuleType(package_name)
+    root_pkg.__path__ = []
+    sys.modules[package_name] = root_pkg
+
+    api_pkg = types.ModuleType(f"{package_name}.api")
+    api_pkg.__path__ = []
+    sys.modules[f"{package_name}.api"] = api_pkg
+
+    orch_pkg = types.ModuleType(f"{package_name}.api.orchestration")
+    orch_pkg.__path__ = []
+    sys.modules[f"{package_name}.api.orchestration"] = orch_pkg
+
+    utils_pkg = types.ModuleType(f"{package_name}.utils")
+    utils_pkg.__path__ = []
+    sys.modules[f"{package_name}.utils"] = utils_pkg
+
+    logging_module = types.ModuleType(f"{package_name}.utils.logging")
+    logging_module.debug_log = lambda *_args, **_kwargs: None
+    logging_module.log = lambda *_args, **_kwargs: None
+    sys.modules[f"{package_name}.utils.logging"] = logging_module
+
+    network_module = types.ModuleType(f"{package_name}.utils.network")
+    network_module.build_worker_url = lambda *_args, **_kwargs: "http://example.invalid"
+
+    async def _probe_worker(*_args, **_kwargs):
+        return None
+
+    network_module.probe_worker = _probe_worker
+
+    async def _fake_session():
+        raise RuntimeError("get_client_session should be mocked in these tests")
+
+    network_module.get_client_session = _fake_session
+    sys.modules[f"{package_name}.utils.network"] = network_module
+
+    created_aiohttp_stub = False
+    if "aiohttp" not in sys.modules:
+        created_aiohttp_stub = True
+        aiohttp_module = types.ModuleType("aiohttp")
+
+        class _ClientTimeout:
+            def __init__(self, total=None):
+                self.total = total
+
+        class _ClientError(Exception):
+            pass
+
+        class _ClientConnectorError(_ClientError):
+            pass
+
+        class _ClientConnectionError(_ClientError):
+            pass
+
+        class _WSMsgType:
+            TEXT = "TEXT"
+            ERROR = "ERROR"
+            CLOSED = "CLOSED"
+
+        class _TCPConnector:
+            def __init__(self, *args, **kwargs):
+                self.args = args
+                self.kwargs = kwargs
+
+        class _ClientSession:
+            def __init__(self, *args, **kwargs):
+                self.closed = False
+
+            async def close(self):
+                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
+        aiohttp_module.web = types.SimpleNamespace(
+            json_response=lambda payload, status=200: {"payload": payload, "status": status}
+        )
+        sys.modules["aiohttp"] = aiohttp_module
+
+    spec = importlib.util.spec_from_file_location(
+        f"{package_name}.api.orchestration.dispatch",
+        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)
+    if created_aiohttp_stub:
+        sys.modules.pop("aiohttp", None)
+    return 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 = [
+            {"id": "w1", "name": "Worker 1"},
+            {"id": "w2", "name": "Worker 2"},
+            {"id": "w3", "name": "Worker 3"},
+        ]
+
+        async def fake_probe(worker):
+            return worker["id"] != "w2"
+
+        with patch.object(dispatch, "worker_is_active", side_effect=fake_probe):
+            active_workers, delegate_master = await dispatch.select_active_workers(
+                workers,
+                use_websocket=False,
+                delegate_master=False,
+                probe_concurrency=3,
+            )
+
+        self.assertEqual([w["id"] for w in active_workers], ["w1", "w3"])
+        self.assertFalse(delegate_master)
+
+    async def test_select_active_workers_disables_delegate_when_all_offline(self):
+        workers = [{"id": "w1", "name": "Worker 1"}]
+
+        async def fake_probe(_worker):
+            return False
+
+        with patch.object(dispatch, "worker_is_active", side_effect=fake_probe):
+            active_workers, delegate_master = await dispatch.select_active_workers(
+                workers,
+                use_websocket=False,
+                delegate_master=True,
+                probe_concurrency=1,
+            )
+
+        self.assertEqual(active_workers, [])
+        self.assertFalse(delegate_master)
+
+    async def test_select_active_workers_uses_websocket_probe_when_enabled(self):
+        workers = [{"id": "w1", "name": "Worker 1"}, {"id": "w2", "name": "Worker 2"}]
+
+        async def fake_http_probe(_worker):
+            return False
+
+        async def fake_ws_probe(_worker):
+            return True
+
+        with patch.object(dispatch, "worker_is_active", side_effect=fake_http_probe) as http_probe, patch.object(
+            dispatch,
+            "worker_ws_is_active",
+            side_effect=fake_ws_probe,
+        ) as ws_probe:
+            active_workers, _ = await dispatch.select_active_workers(
+                workers,
+                use_websocket=True,
+                delegate_master=False,
+                probe_concurrency=4,
+            )
+
+        self.assertEqual([w["id"] for w in active_workers], ["w1", "w2"])
+        self.assertEqual(ws_probe.call_count, 2)
+        self.assertEqual(http_probe.call_count, 0)
+
+    async def test_probe_concurrency_is_bounded(self):
+        workers = [{"id": f"w{i}", "name": f"Worker {i}"} for i in range(6)]
+        state = {"in_flight": 0, "max_in_flight": 0}
+
+        async def fake_probe(_worker):
+            state["in_flight"] += 1
+            state["max_in_flight"] = max(state["max_in_flight"], state["in_flight"])
+            await asyncio.sleep(0.01)
+            state["in_flight"] -= 1
+            return True
+
+        with patch.object(dispatch, "worker_is_active", side_effect=fake_probe):
+            active_workers, _ = await dispatch.select_active_workers(
+                workers,
+                use_websocket=False,
+                delegate_master=False,
+                probe_concurrency=2,
+            )
+
+        self.assertEqual(len(active_workers), len(workers))
+        self.assertLessEqual(state["max_in_flight"], 2)
+        self.assertGreaterEqual(state["max_in_flight"], 2)
+
+    async def test_select_least_busy_worker_round_robins_idle_workers(self):
+        workers = [
+            {"id": "w1", "name": "Worker 1"},
+            {"id": "w2", "name": "Worker 2"},
+            {"id": "w3", "name": "Worker 3"},
+        ]
+        queue_map = {"w1": 0, "w2": 0, "w3": 2}
+
+        async def fake_probe(worker_url, timeout=3.0):
+            worker_id = worker_url.rsplit("/", 1)[-1]
+            return {"exec_info": {"queue_remaining": queue_map[worker_id]}}
+
+        with patch.object(dispatch, "build_worker_url", side_effect=lambda worker: f"http://host/{worker['id']}"), patch.object(
+            dispatch,
+            "probe_worker",
+            side_effect=fake_probe,
+        ):
+            dispatch._least_busy_rr_index = 0
+            selected1 = await dispatch.select_least_busy_worker(workers, probe_concurrency=3)
+            selected2 = await dispatch.select_least_busy_worker(workers, probe_concurrency=3)
+            selected3 = await dispatch.select_least_busy_worker(workers, probe_concurrency=3)
+
+        self.assertEqual(selected1["id"], "w1")
+        self.assertEqual(selected2["id"], "w2")
+        self.assertEqual(selected3["id"], "w1")
+
+    async def test_select_least_busy_worker_chooses_smallest_queue_when_all_busy(self):
+        workers = [
+            {"id": "w1", "name": "Worker 1"},
+            {"id": "w2", "name": "Worker 2"},
+            {"id": "w3", "name": "Worker 3"},
+        ]
+        queue_map = {"w1": 5, "w2": 2, "w3": 4}
+
+        async def fake_probe(worker_url, timeout=3.0):
+            worker_id = worker_url.rsplit("/", 1)[-1]
+            return {"exec_info": {"queue_remaining": queue_map[worker_id]}}
+
+        with patch.object(dispatch, "build_worker_url", side_effect=lambda worker: f"http://host/{worker['id']}"), patch.object(
+            dispatch,
+            "probe_worker",
+            side_effect=fake_probe,
+        ):
+            selected = await dispatch.select_least_busy_worker(workers, probe_concurrency=2)
+
+        self.assertEqual(selected["id"], "w2")
+
+    async def test_select_least_busy_worker_returns_none_when_all_probes_fail(self):
+        workers = [{"id": "w1", "name": "Worker 1"}]
+
+        async def fake_probe(_worker_url, timeout=3.0):
+            return None
+
+        with patch.object(dispatch, "build_worker_url", side_effect=lambda worker: f"http://host/{worker['id']}"), patch.object(
+            dispatch,
+            "probe_worker",
+            side_effect=fake_probe,
+        ):
+            selected = await dispatch.select_least_busy_worker(workers, probe_concurrency=1)
+
+        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_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()
@@ -0,0 +1,208 @@
+import json
+import unittest
+
+
+class DistributedValueTests(unittest.TestCase):
+    """Unit tests for the DistributedValue node's distribute() method."""
+
+    def _make_node(self):
+        # Import inline to avoid plugin-level imports
+        import importlib.util
+        import sys
+        import types
+        from pathlib import Path
+        from unittest.mock import MagicMock
+
+        module_path = Path(__file__).resolve().parents[1] / "nodes" / "utilities.py"
+        pkg_name = "dv_test_pkg"
+
+        for mod_name in list(sys.modules):
+            if mod_name == pkg_name or mod_name.startswith(f"{pkg_name}."):
+                del sys.modules[mod_name]
+
+        # Mock torch if not available
+        if "torch" not in sys.modules:
+            sys.modules["torch"] = MagicMock()
+
+        root_pkg = types.ModuleType(pkg_name)
+        root_pkg.__path__ = []
+        sys.modules[pkg_name] = root_pkg
+
+        utils_pkg = types.ModuleType(f"{pkg_name}.utils")
+        utils_pkg.__path__ = []
+        sys.modules[f"{pkg_name}.utils"] = utils_pkg
+
+        logging_mod = types.ModuleType(f"{pkg_name}.utils.logging")
+        logging_mod.debug_log = lambda *_a, **_k: None
+        logging_mod.log = lambda *_a, **_k: None
+        sys.modules[f"{pkg_name}.utils.logging"] = logging_mod
+
+        spec = importlib.util.spec_from_file_location(
+            f"{pkg_name}.nodes.utilities", module_path
+        )
+        mod = importlib.util.module_from_spec(spec)
+        spec.loader.exec_module(mod)
+        return mod.DistributedValue()
+
+    def setUp(self):
+        self.node = self._make_node()
+
+    def test_master_returns_default(self):
+        result = self.node.distribute(
+            default_value="model_a",
+            worker_values="{}",
+            is_worker=False,
+            worker_id="",
+        )
+        self.assertEqual(result, ("model_a",))
+
+    def test_master_coerces_default_int(self):
+        values = json.dumps({"_type": "INT"})
+        result = self.node.distribute(
+            default_value="42",
+            worker_values=values,
+            is_worker=False,
+            worker_id="",
+        )
+        self.assertEqual(result, (42,))
+        self.assertIsInstance(result[0], int)
+
+    def test_master_coerces_default_float(self):
+        values = json.dumps({"_type": "FLOAT"})
+        result = self.node.distribute(
+            default_value="2.5",
+            worker_values=values,
+            is_worker=False,
+            worker_id="",
+        )
+        self.assertEqual(result, (2.5,))
+        self.assertIsInstance(result[0], float)
+
+    def test_worker_returns_specific_value(self):
+        values = json.dumps({"1": "model_x", "2": "model_y"})
+        result = self.node.distribute(
+            default_value="default",
+            worker_values=values,
+            is_worker=True,
+            worker_id="worker_0",
+        )
+        self.assertEqual(result, ("model_x",))
+
+    def test_worker_second_index(self):
+        values = json.dumps({"1": "model_x", "2": "model_y"})
+        result = self.node.distribute(
+            default_value="default",
+            worker_values=values,
+            is_worker=True,
+            worker_id="worker_1",
+        )
+        self.assertEqual(result, ("model_y",))
+
+    def test_worker_falls_back_to_default_when_key_missing(self):
+        values = json.dumps({"_type": "INT", "1": "3"})
+        result = self.node.distribute(
+            default_value="9",
+            worker_values=values,
+            is_worker=True,
+            worker_id="worker_5",
+        )
+        self.assertEqual(result, (9,))
+        self.assertIsInstance(result[0], int)
+
+    def test_worker_falls_back_to_default_on_empty_value(self):
+        values = json.dumps({"1": ""})
+        result = self.node.distribute(
+            default_value="fallback",
+            worker_values=values,
+            is_worker=True,
+            worker_id="worker_0",
+        )
+        self.assertEqual(result, ("fallback",))
+
+    def test_worker_falls_back_on_invalid_json(self):
+        result = self.node.distribute(
+            default_value="safe",
+            worker_values="not-json",
+            is_worker=True,
+            worker_id="worker_0",
+        )
+        self.assertEqual(result, ("safe",))
+
+    def test_worker_falls_back_on_invalid_worker_id(self):
+        values = json.dumps({"1": "model_x"})
+        result = self.node.distribute(
+            default_value="safe",
+            worker_values=values,
+            is_worker=True,
+            worker_id="bad_id",
+        )
+        self.assertEqual(result, ("safe",))
+
+    def test_worker_id_as_direct_integer(self):
+        values = json.dumps({"1": "model_x"})
+        result = self.node.distribute(
+            default_value="default",
+            worker_values=values,
+            is_worker=True,
+            worker_id="0",
+        )
+        self.assertEqual(result, ("model_x",))
+
+    def test_type_int_coerces_value(self):
+        values = json.dumps({"_type": "INT", "1": "42"})
+        result = self.node.distribute(
+            default_value="0",
+            worker_values=values,
+            is_worker=True,
+            worker_id="worker_0",
+        )
+        self.assertEqual(result, (42,))
+        self.assertIsInstance(result[0], int)
+
+    def test_type_float_coerces_value(self):
+        values = json.dumps({"_type": "FLOAT", "1": "3.14"})
+        result = self.node.distribute(
+            default_value="0",
+            worker_values=values,
+            is_worker=True,
+            worker_id="worker_0",
+        )
+        self.assertAlmostEqual(result[0], 3.14)
+        self.assertIsInstance(result[0], float)
+
+    def test_type_combo_stays_string(self):
+        values = json.dumps({"_type": "COMBO", "1": "model_v2"})
+        result = self.node.distribute(
+            default_value="model_v1",
+            worker_values=values,
+            is_worker=True,
+            worker_id="worker_0",
+        )
+        self.assertEqual(result, ("model_v2",))
+        self.assertIsInstance(result[0], str)
+
+    def test_type_string_default_stays_string(self):
+        values = json.dumps({"1": "hello"})
+        result = self.node.distribute(
+            default_value="default",
+            worker_values=values,
+            is_worker=True,
+            worker_id="worker_0",
+        )
+        self.assertEqual(result, ("hello",))
+        self.assertIsInstance(result[0], str)
+
+    def test_int_coerce_handles_float_string(self):
+        """INT coercion of '3.7' should truncate to 3."""
+        values = json.dumps({"_type": "INT", "1": "3.7"})
+        result = self.node.distribute(
+            default_value="0",
+            worker_values=values,
+            is_worker=True,
+            worker_id="worker_0",
+        )
+        self.assertEqual(result, (3,))
+
+
+if __name__ == "__main__":
+    unittest.main()
@@ -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()
@@ -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"]
@@ -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'])
@@ -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()
@@ -0,0 +1,437 @@
+"""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_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'):
+            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()
@@ -0,0 +1,131 @@
+"""Image-list participation uses the existing master orchestrator setting."""
+import json
+import sys
+import unittest
+from unittest.mock import AsyncMock, 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_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.assertTrue(self.preflight.call_args.kwargs["require_idle"])
+        # 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)
+
+    async def test_busy_workers_do_not_prevent_enabled_master(self):
+        self.preflight.return_value = {}
+        await self.queue()
+        self.assertEqual(json.loads(self.prepared_inputs()["image_list_enabled_worker_ids"]), [])
+
+    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()
@@ -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)
@@ -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))
@@ -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()
@@ -0,0 +1,197 @@
+import asyncio
+import importlib.util
+import sys
+import time
+import types
+import unittest
+from dataclasses import dataclass, field
+from pathlib import Path
+
+
+def _load_job_timeout_module():
+    module_path = Path(__file__).resolve().parents[1] / "upscale" / "job_timeout.py"
+    package_name = "dist_job_timeout_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__ = []
+    sys.modules[package_name] = root_pkg
+
+    upscale_pkg = types.ModuleType(f"{package_name}.upscale")
+    upscale_pkg.__path__ = []
+    sys.modules[f"{package_name}.upscale"] = upscale_pkg
+
+    utils_pkg = types.ModuleType(f"{package_name}.utils")
+    utils_pkg.__path__ = []
+    sys.modules[f"{package_name}.utils"] = utils_pkg
+
+    config_holder = {"value": {"settings": {}, "workers": []}}
+    probe_holder = {"fn": None}
+    prompt_server_holder = {
+        "value": types.SimpleNamespace(
+            distributed_tile_jobs_lock=asyncio.Lock(),
+            distributed_pending_tile_jobs={},
+        )
+    }
+
+    config_module = types.ModuleType(f"{package_name}.utils.config")
+    config_module.load_config = lambda: config_holder["value"]
+    sys.modules[f"{package_name}.utils.config"] = config_module
+
+    constants_module = types.ModuleType(f"{package_name}.utils.constants")
+    constants_module.HEARTBEAT_TIMEOUT = 60
+    sys.modules[f"{package_name}.utils.constants"] = constants_module
+
+    logging_module = types.ModuleType(f"{package_name}.utils.logging")
+    logging_module.debug_log = lambda *_args, **_kwargs: None
+    logging_module.log = lambda *_args, **_kwargs: None
+    sys.modules[f"{package_name}.utils.logging"] = logging_module
+
+    network_module = types.ModuleType(f"{package_name}.utils.network")
+    network_module.build_worker_url = lambda worker: f"http://{worker.get('host', '127.0.0.1')}:{worker.get('port', 8188)}"
+
+    async def _probe_worker(url, timeout=2.0):
+        fn = probe_holder["fn"]
+        if fn is None:
+            return None
+        return await fn(url, timeout)
+
+    network_module.probe_worker = _probe_worker
+    sys.modules[f"{package_name}.utils.network"] = network_module
+
+    job_store_module = types.ModuleType(f"{package_name}.upscale.job_store")
+    job_store_module.ensure_tile_jobs_initialized = lambda: prompt_server_holder["value"]
+    sys.modules[f"{package_name}.upscale.job_store"] = job_store_module
+
+    job_models_module = types.ModuleType(f"{package_name}.upscale.job_models")
+
+    class BaseJobState:
+        pass
+
+    @dataclass
+    class ImageJobState(BaseJobState):
+        multi_job_id: str
+        mode: str = field(default="dynamic", init=False)
+        queue: asyncio.Queue = field(default_factory=asyncio.Queue)
+        pending_images: asyncio.Queue = field(default_factory=asyncio.Queue)
+        completed_images: dict = field(default_factory=dict)
+        worker_status: dict = field(default_factory=dict)
+        assigned_to_workers: dict = field(default_factory=dict)
+        batch_size: int = 0
+        num_tiles_per_image: int = 0
+        batched_static: bool = False
+
+        @property
+        def pending_tasks(self):
+            return self.pending_images
+
+        @property
+        def completed_tasks(self):
+            return self.completed_images
+
+    @dataclass
+    class TileJobState(BaseJobState):
+        multi_job_id: str
+        mode: str = field(default="static", init=False)
+        queue: asyncio.Queue = field(default_factory=asyncio.Queue)
+        pending_tasks: asyncio.Queue = field(default_factory=asyncio.Queue)
+        completed_tasks: dict = field(default_factory=dict)
+        worker_status: dict = field(default_factory=dict)
+        assigned_to_workers: dict = field(default_factory=dict)
+        batch_size: int = 0
+        num_tiles_per_image: int = 0
+        batched_static: bool = False
+
+    job_models_module.BaseJobState = BaseJobState
+    job_models_module.ImageJobState = ImageJobState
+    job_models_module.TileJobState = TileJobState
+    sys.modules[f"{package_name}.upscale.job_models"] = job_models_module
+
+    spec = importlib.util.spec_from_file_location(f"{package_name}.upscale.job_timeout", 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)
+
+    module._config_holder = config_holder
+    module._probe_holder = probe_holder
+    module._prompt_server_holder = prompt_server_holder
+    module._ImageJobState = ImageJobState
+    module._TileJobState = TileJobState
+    return module
+
+
+jt = _load_job_timeout_module()
+
+
+class JobTimeoutRequeueTests(unittest.IsolatedAsyncioTestCase):
+    async def asyncSetUp(self):
+        jt._prompt_server_holder["value"] = types.SimpleNamespace(
+            distributed_tile_jobs_lock=asyncio.Lock(),
+            distributed_pending_tile_jobs={},
+        )
+        jt._config_holder["value"] = {
+            "settings": {"worker_timeout_seconds": 5},
+            "workers": [{"id": "worker-1", "host": "worker.local", "port": 8188}],
+        }
+
+    async def test_requeues_only_incomplete_dynamic_tasks_for_timed_out_worker(self):
+        async def _offline_probe(_url, _timeout):
+            return None
+
+        jt._probe_holder["fn"] = _offline_probe
+        prompt_server = jt._prompt_server_holder["value"]
+        job_data = jt._ImageJobState("job-1")
+        job_data.worker_status["worker-1"] = time.time() - 60.0
+        job_data.assigned_to_workers["worker-1"] = [0, 1]
+        job_data.completed_images[1] = "done"
+        prompt_server.distributed_pending_tile_jobs["job-1"] = job_data
+
+        requeued = await jt._check_and_requeue_timed_out_workers("job-1", total_tasks=2)
+
+        self.assertEqual(requeued, 1)
+        self.assertEqual(await job_data.pending_images.get(), 0)
+        self.assertNotIn("worker-1", job_data.worker_status)
+        self.assertEqual(job_data.assigned_to_workers["worker-1"], [])
+
+    async def test_busy_probe_graces_worker_and_skips_requeue(self):
+        async def _busy_probe(_url, _timeout):
+            return {"exec_info": {"queue_remaining": 3}}
+
+        jt._probe_holder["fn"] = _busy_probe
+        prompt_server = jt._prompt_server_holder["value"]
+        job_data = jt._ImageJobState("job-2")
+        old_heartbeat = time.time() - 60.0
+        job_data.worker_status["worker-1"] = old_heartbeat
+        job_data.assigned_to_workers["worker-1"] = [0]
+        prompt_server.distributed_pending_tile_jobs["job-2"] = job_data
+
+        requeued = await jt._check_and_requeue_timed_out_workers("job-2", total_tasks=1)
+
+        self.assertEqual(requeued, 0)
+        self.assertIn("worker-1", job_data.worker_status)
+        self.assertGreaterEqual(job_data.worker_status["worker-1"], old_heartbeat)
+        self.assertTrue(job_data.pending_images.empty())
+
+    async def test_completed_dynamic_task_is_not_requeued(self):
+        async def _offline_probe(_url, _timeout):
+            return None
+
+        jt._probe_holder["fn"] = _offline_probe
+        prompt_server = jt._prompt_server_holder["value"]
+        job_data = jt._ImageJobState("job-3")
+        job_data.worker_status["worker-1"] = time.time() - 60.0
+        job_data.assigned_to_workers["worker-1"] = [7]
+        job_data.completed_images[7] = "complete"
+        prompt_server.distributed_pending_tile_jobs["job-3"] = job_data
+
+        requeued = await jt._check_and_requeue_timed_out_workers("job-3", total_tasks=1)
+
+        self.assertEqual(requeued, 0)
+        self.assertTrue(job_data.pending_images.empty())
+
+
+if __name__ == "__main__":
+    unittest.main()
@@ -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
@@ -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
@@ -0,0 +1,314 @@
+"""CPU-only typed collection and optional native V3 mapping acceptance."""
+import asyncio
+import importlib.util
+import json
+import os
+from pathlib import Path
+import subprocess
+import sys
+import types
+
+import pytest
+import torch
+
+ROOT = Path(__file__).resolve().parents[1]
+SOCKETS = ('images', 'audio', 'masks', 'texts', 'video')
+WORKERS = ['serverless:pool:1', 'serverless:pool:2']
+
+
+class Manager:
+    def __init__(self, *, pending=False, owned=True):
+        self.pending, self.owned = pending, owned
+        self.cancelled = []
+
+    def owns(self, job, worker):
+        return self.owned
+
+    def is_pending(self, job, worker):
+        return self.pending
+
+    async def cancel(self, **kwargs):
+        self.cancelled.append(kwargs)
+
+
+@pytest.fixture
+def module(monkeypatch):
+    package = '_media_collector_test'
+    for name, path in ((package, ROOT), (package + '.nodes', ROOT / 'nodes'),
+                       (package + '.utils', ROOT / 'utils'),
+                       (package + '.serverless_bridge', ROOT / 'serverless_bridge')):
+        stub = types.ModuleType(name)
+        stub.__path__ = [str(path)]
+        monkeypatch.setitem(sys.modules, name, stub)
+    io = types.SimpleNamespace(ComfyNode=object, NodeOutput=lambda *items: types.SimpleNamespace(result=items))
+    api = types.ModuleType('comfy_api.v0_0_2')
+    api.io = io
+    monkeypatch.setitem(sys.modules, 'comfy_api', types.ModuleType('comfy_api'))
+    monkeypatch.setitem(sys.modules, 'comfy_api.v0_0_2', api)
+    management = types.ModuleType('comfy.model_management')
+    management.throw_exception_if_processing_interrupted = lambda: None
+    comfy = types.ModuleType('comfy')
+    comfy.model_management = management
+    monkeypatch.setitem(sys.modules, 'comfy', comfy)
+    monkeypatch.setitem(sys.modules, 'comfy.model_management', management)
+    ps = types.SimpleNamespace(distributed_pending_jobs={}, distributed_jobs_lock=asyncio.Lock())
+    server = types.ModuleType('server')
+    server.PromptServer = types.SimpleNamespace(instance=ps)
+    monkeypatch.setitem(sys.modules, 'server', server)
+    helper = types.ModuleType(package + '.utils.async_helpers')
+    helper.run_async_in_server_loop = asyncio.run
+    monkeypatch.setitem(sys.modules, helper.__name__, helper)
+    spec = importlib.util.spec_from_file_location(package + '.nodes.media_collector', ROOT / 'nodes/media_collector.py')
+    loaded = importlib.util.module_from_spec(spec)
+    monkeypatch.setitem(sys.modules, spec.name, loaded)
+    spec.loader.exec_module(loaded)
+    monkeypatch.setattr(loaded, 'POLL_SECONDS', 0.005)
+    monkeypatch.setattr(loaded, 'START_GRACE_SECONDS', 0.025)
+    loaded.test_server = ps
+    loaded.test_management = management
+    return loaded
+
+
+def native_media(offset=0):
+    return {
+        'images': [torch.full((1, 2 + offset, 3, 4), offset, dtype=torch.float64),
+                   torch.zeros((2, 1, 1, 3), dtype=torch.float16)],
+        'audio': [{'waveform': torch.zeros((1, 1 + offset, 8), dtype=torch.float64),
+                   'sample_rate': 44100 + offset}],
+        'masks': [torch.zeros((1, 2 + offset, 3), dtype=torch.float64)],
+        'texts': ['', '  café\n🎥\r\n'],
+        'video': [object()],
+    }
+
+
+def setup(module, messages=(), *, manager=None):
+    queue = asyncio.Queue()
+    for message in messages:
+        queue.put_nowait(message)
+    ps = module.test_server
+    ps.distributed_pending_jobs = {'job': queue, 'unrelated': asyncio.Queue()}
+    ps.distributed_serverless = manager if manager is not None else Manager()
+    return ps, queue
+
+
+def message(worker, media):
+    return {'worker_id': worker, 'media': media, 'is_last': True}
+
+
+def collect(module, **kwargs):
+    return module.DistributedMediaCollector.execute(**kwargs).result
+
+
+def remote(module, **kwargs):
+    controls = dict(multi_job_id=['job'], enabled_worker_ids=[json.dumps(WORKERS)],
+                    expected_media=[json.dumps(['texts'])], delegate_only=[True])
+    controls.update(kwargs)
+    return collect(module, **controls)
+
+
+@pytest.mark.parametrize('controls', [{}, {'pass_through': [True]}, {'enabled_worker_ids': ['[]']}])
+def test_all_five_passthrough_preserves_lists_and_native_values(module, controls):
+    values = native_media()
+    result = collect(module, **values, **controls)
+    assert len(result) == 5
+    for index, socket in enumerate(SOCKETS):
+        assert result[index] is values[socket]
+
+
+def test_reversed_arrivals_master_then_workers_then_items(module):
+    master, first, second = native_media(), native_media(1), native_media(2)
+    ps, _ = setup(module, [message(WORKERS[1], second), message(WORKERS[0], first)])
+    result = remote(module, **master, delegate_only=[False], expected_media=[json.dumps(list(SOCKETS))])
+    for index, socket in enumerate(SOCKETS):
+        expected = master[socket] + first[socket] + second[socket]
+        assert len(result[index]) == len(expected)
+        assert all(actual is original for actual, original in zip(result[index], expected))
+    assert list(ps.distributed_pending_jobs) == ['unrelated']
+    assert ps.distributed_serverless.cancelled == []
+
+
+def test_delegate_only_and_empty_string(module):
+    ps, _ = setup(module, [message(WORKERS[1], {'texts': ['b']}), message(WORKERS[0], {'texts': ['']})])
+    result = remote(module)
+    assert result == ([], [], [], ['', 'b'], [])
+    assert 'job' not in ps.distributed_pending_jobs
+
+
+@pytest.mark.parametrize('kwargs', [{}, {'texts': []}, {'texts': ''}, {'texts': ['x'], 'is_worker': [True]},
+    {'texts': ['x'], 'enabled_worker_ids': ['["direct-worker"]']},
+    {'texts': ['x'], 'expected_media': ['["images"]']},
+    {'texts': ['x'], 'pass_through': [True, False]}, {'delegate_only': True},
+    {'texts': ['x'], 'enabled_worker_ids': ['["serverless:pool:1","serverless:pool:1"]']},
+    {'texts': ['x'], 'load_balance': ['false']}])
+def test_rejects_missing_media_or_invalid_controls(module, kwargs):
+    with pytest.raises((ValueError, RuntimeError)):
+        collect(module, **kwargs)
+
+
+@pytest.mark.parametrize('bad', [
+    {'worker_id': 'serverless:other:1', 'media': {'texts': ['x']}, 'is_last': True},
+    {'worker_id': WORKERS[0], 'media': {'texts': ['x']}, 'is_last': False},
+    {'worker_id': WORKERS[0], 'media': {'images': [object()]}, 'is_last': True},
+    {'worker_id': WORKERS[0], 'media': {'texts': []}, 'is_last': True},
+    {'worker_id': WORKERS[0], 'media': {'texts': 'x'}, 'is_last': True},
+    {'worker_id': WORKERS[0], 'media': {'texts': [None]}, 'is_last': True},
+    {'worker_id': WORKERS[0], 'media': {'texts': ['x'] * 65}, 'is_last': True},
+    {'worker_id': WORKERS[0], 'media': {'texts': ['x'], 'unknown': ['x']}, 'is_last': True},
+    {'worker_id': WORKERS[0], 'serverless_error': 'sanitized failure', 'is_last': True},
+])
+def test_rejects_bad_delivery_without_partial_output_and_cleans(module, bad):
+    ps, _ = setup(module, [bad])
+    with pytest.raises((ValueError, RuntimeError)):
+        remote(module)
+    assert list(ps.distributed_pending_jobs) == ['unrelated']
+    assert ps.distributed_serverless.cancelled == [{'collector_job_id': 'job'}]
+
+
+@pytest.mark.parametrize('after_complete', [False, True])
+def test_duplicate_delivery_rejected_including_final_queue_drain(module, after_complete):
+    first = message(WORKERS[0], {'texts': ['a']})
+    second = message(WORKERS[1], {'texts': ['b']})
+    messages = [first, second, first] if after_complete else [first, first, second]
+    ps, _ = setup(module, messages)
+    with pytest.raises(RuntimeError, match='[Dd]uplicate'):
+        remote(module)
+    assert 'job' not in ps.distributed_pending_jobs
+    assert ps.distributed_serverless.cancelled == [{'collector_job_id': 'job'}]
+
+
+@pytest.mark.parametrize('manager', [None, Manager(owned=False), Manager(pending=False)])
+def test_missing_manager_or_finished_participant_fails_bounded(module, manager):
+    ps, _ = setup(module)
+    ps.distributed_serverless = manager
+    with pytest.raises(RuntimeError, match='[Ss]erverless'):
+        remote(module)
+    assert 'job' not in ps.distributed_pending_jobs
+
+
+def test_control_error_also_cleans_exact_job(module):
+    ps, _ = setup(module)
+    with pytest.raises(ValueError):
+        remote(module, expected_media='["unknown"]')
+    assert list(ps.distributed_pending_jobs) == ['unrelated']
+    assert ps.distributed_serverless.cancelled == [{'collector_job_id': 'job'}]
+
+
+def test_launch_race_and_pending_deadline_owned_by_manager(module, monkeypatch):
+    ps, queue = setup(module)
+    ps.distributed_serverless = None
+    async def runner(coro):
+        async def produce():
+            await asyncio.sleep(0.01)
+            ps.distributed_serverless = Manager(pending=True)
+            await asyncio.sleep(0.035)
+            queue.put_nowait(message(WORKERS[1], {'texts': ['b']}))
+            queue.put_nowait(message(WORKERS[0], {'texts': ['a']}))
+        task = asyncio.create_task(produce())
+        result = await coro
+        await task
+        return result
+    monkeypatch.setattr(module, 'run_async_in_server_loop', lambda coro: asyncio.run(runner(coro)))
+    assert remote(module)[3] == ['a', 'b']
+    assert ps.distributed_serverless.cancelled == []
+
+
+@pytest.mark.parametrize('cancel_task', [False, True])
+def test_cancellation_while_blocked_cleans_and_propagates(module, monkeypatch, cancel_task):
+    ps, _ = setup(module, manager=Manager(pending=True))
+    class Interrupted(Exception):
+        pass
+    async def runner(coro):
+        task = asyncio.create_task(coro)
+        await asyncio.sleep(0.012)
+        if cancel_task:
+            task.cancel()
+        else:
+            def interrupt():
+                raise Interrupted('stop')
+            module.test_management.throw_exception_if_processing_interrupted = interrupt
+        return await task
+    monkeypatch.setattr(module, 'run_async_in_server_loop', lambda coro: asyncio.run(runner(coro)))
+    with pytest.raises(asyncio.CancelledError if cancel_task else Interrupted):
+        remote(module)
+    assert list(ps.distributed_pending_jobs) == ['unrelated']
+    assert ps.distributed_serverless.cancelled == [{'collector_job_id': 'job'}]
+
+
+def test_replaced_queue_not_removed(module, monkeypatch):
+    ps, _ = setup(module, manager=Manager(pending=True))
+    replacement = asyncio.Queue()
+    async def runner(coro):
+        task = asyncio.create_task(coro)
+        await asyncio.sleep(0.012)
+        ps.distributed_pending_jobs['job'] = replacement
+        return await task
+    monkeypatch.setattr(module, 'run_async_in_server_loop', lambda coro: asyncio.run(runner(coro)))
+    with pytest.raises(RuntimeError, match='queue'):
+        remote(module)
+    assert ps.distributed_pending_jobs['job'] is replacement
+
+
+def test_native_v3_schema_and_list_mapping_cpu():
+    core = os.environ.get('COMFYUI_SOURCE_ROOT')
+    if not core:
+        pytest.skip('Set COMFYUI_SOURCE_ROOT for CPU-only native V3 mapping')
+    probe = r'''
+import asyncio, importlib.util, sys, types
+from pathlib import Path
+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 torch
+import server
+# No HTTP listener or PromptServer bootstrap is needed for schema/list mapping.
+server.PromptServer.instance = types.SimpleNamespace()
+root = Path(sys.argv[2])
+for name, path in [('typed_collector', root), ('typed_collector.nodes', root/'nodes'),
+                   ('typed_collector.utils', root/'utils'),
+                   ('typed_collector.serverless_bridge', root/'serverless_bridge')]:
+    package = types.ModuleType(name)
+    package.__path__ = [str(path)]
+    sys.modules[name] = package
+spec = importlib.util.spec_from_file_location('typed_collector.nodes.media_collector', root/'nodes/media_collector.py')
+module = importlib.util.module_from_spec(spec)
+sys.modules[spec.name] = module
+spec.loader.exec_module(module)
+cls = module.DistributedMediaCollector
+info = cls.GET_NODE_INFO_V1()
+assert info['output'] == ['IMAGE','AUDIO','MASK','STRING','VIDEO'], info
+assert info['output_is_list'] == [True]*5, info
+assert cls.define_schema().is_input_list and cls.define_schema().accept_all_inputs
+for name, kind in zip(['images','audio','masks','texts','video'], info['output']):
+    field = info['input']['optional'][name]
+    assert field[0] == kind and field[1]['forceInput'] is True, field
+assert info['input']['required']['load_balance'][1]['default'] is False
+async def check():
+    values = {'load_balance':False, 'pass_through':True}
+    prepared, missing, hidden = execution.get_input_data(values, cls, unique_id='probe')
+    assert not missing, missing
+    image = torch.zeros((1,2,3,4), dtype=torch.float64)
+    audio = {'waveform':torch.zeros((1,2,7)), 'sample_rate':48000}
+    mask = torch.ones((1,2,3), dtype=torch.float64)
+    from comfy_api.input_impl import VideoFromComponents
+    from comfy_api.util import VideoComponents
+    from fractions import Fraction
+    video = VideoFromComponents(VideoComponents(images=image[:,:,:,:3].float(), frame_rate=Fraction(24,1)))
+    native = {'images':[image], 'audio':[audio], 'masks':[mask], 'texts':[''], 'video':[video]}
+    prepared.update(native)
+    results = await execution._async_map_node_over_list('cpu','probe',cls,prepared,cls.FUNCTION,v3_data=hidden)
+    results = await execution.resolve_map_node_over_list_results(results)
+    assert len(results) == 1
+    for index, name in enumerate(native):
+        assert results[0].result[index] is native[name]
+    assert results[0].result[4][0] is video
+    print('MEDIA_COLLECTOR_CPU_OK')
+asyncio.run(check())
+'''
+    env = {**os.environ, 'CUDA_VISIBLE_DEVICES': '', 'HIP_VISIBLE_DEVICES': '', 'PYTHONDONTWRITEBYTECODE': '1'}
+    env.pop('PYTHONPATH', None)
+    result = subprocess.run([sys.executable, '-c', probe, core, str(ROOT)], env=env,
+                            capture_output=True, text=True, timeout=90)
+    assert result.returncode == 0, result.stdout + result.stderr
+    assert 'MEDIA_COLLECTOR_CPU_OK' in result.stdout
@@ -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"")
--- a/Show More
+++ b/Show More
				`@@ -0,0 +1 @@`
				`"""Dynamic per-image image-list execution internals."""`
				`@@ -0,0 +1 @@`
				`"""Private execution helpers; public registration lives in nodes.v3."""`
				`@@ -0,0 +1 @@`
				`"""Internal extension lifecycle support (not a node provider)."""`