Merge pull request #22 from Comfy-Org/feat/encode-nested-autogrow-dicts

fix(encoder): recurse into nested AUTOGROW / DYNAMIC_COMBO runtime dicts

client.py

@@ -30,13 +30,79 @@ log = logging.getLogger("comfy_remote_nodes.client")
 
 # Capability tokens advertised on every outbound request. Currently
 # covers the input/output encodings the registered providers (LTXV,
-# Gemini) need, plus the ``schema:v3`` marker so descriptors come back
-# in the V3 ``get_v1_info()`` shape.
+# Gemini, Ideogram, ...) need, plus the ``schema:v3`` marker so
+# descriptors come back in the V3 ``get_v1_info()`` shape. Servers use
+# this list to negotiate per-request: e.g. a provider only emits the
+# multi-frame ``image:png_base64_batch`` envelope when this list
+# contains :data:`Capability.IMAGE_PNG_BASE64_BATCH`, otherwise it
+# falls back to the single-frame ``image:png_base64`` shape (with a
+# server-side truncation warning logged when extras are dropped).
 CLIENT_CAPABILITIES = [
     Capability.SCHEMA_V3,
     Capability.VIDEO_MP4_INLINE,
+    Capability.VIDEO_MP4_BASE64,
     Capability.IMAGE_PNG_BASE64,
+    Capability.IMAGE_PNG_BASE64_BATCH,
+    # Advertises that this client can emit a multi-frame IMAGE
+    # envelope with per-frame presigned-GET URIs instead of inline
+    # base64 frames when the summed batch payload would overflow the
+    # 8 MiB inline cap (see ``serialization.maybe_externalize``).
+    # Opportunistic: the client falls back to it whenever a batch
+    # envelope it produced exceeds the cap, so descriptors don't
+    # need to publish it in their ``input_serialization`` to
+    # benefit — any input that already accepts ``png_base64_batch``
+    # transparently accepts the externalised variant on the server
+    # via the shared ``resolve_image_envelope_pngs`` dispatch.
+    Capability.IMAGE_PNG_BASE64_BATCH_URI,
     Capability.MASK_PNG_BASE64,
+    Capability.AUDIO_MP3_INLINE,
+    Capability.AUDIO_MP3_BASE64,
+    Capability.AUDIO_WAV_BASE64,
+    Capability.AUDIO_WAV_INLINE,
+    # OGG/Opus decode path. PyAV / FFmpeg handle OGG/Opus natively in
+    # ``audio_bytes_to_audio_input`` so the proxy_node side is the
+    # same as MP3/WAV — advertising the token unlocks per-provider
+    # ``output_format`` choices that emit Opus (e.g. ElevenLabs'
+    # ``opus_48000_192``) without requiring servers to label-launder
+    # the upstream container as ``mp3_base64``.
+    Capability.AUDIO_OPUS_BASE64,
+    # Advertises that this client can drive the
+    # ``execute_async`` + poll-loop lifecycle (submit, poll, cancel,
+    # TASK_LOST replay). Servers gate ``async_polling``-mode
+    # descriptors on the presence of this token so an older client
+    # that only knows ``/execute`` doesn't silently submit a
+    # long-poll node into the sync endpoint.
+    Capability.ASYNC_EXECUTE,
+    # Generic opaque pass-through bucket for partner helper-config
+    # custom IO types (OPENAI_INPUT_FILES, OPENAI_CHAT_CONFIG,
+    # GEMINI_INPUT_FILES, RECRAFT_*, future partner helper types).
+    # Servers may gate descriptors that publish ``input_serialization``
+    # / ``output_serialization`` entries with the ``"opaque"`` marker
+    # on the presence of this token; legacy clients that don't know
+    # about IO.Custom fallthrough get filtered out before they try to
+    # parse a descriptor whose io_type strings they wouldn't recognise.
+    Capability.IO_OPAQUE,
+    # 3D MODEL envelope (binary glTF 2.0 — single-mesh GLB). The
+    # decoder in ``serialization.decode_model3d_envelope`` hands raw
+    # GLB bytes to ``comfy_api.latest._util.File3D`` so downstream
+    # nodes (preview3d / SaveGLB / Load3D) see the same value shape
+    # they would from a local Tripo / Rodin / Meshy node. Servers
+    # gate ``MODEL_3D`` provider descriptors on this token so a
+    # legacy client without the decoder surfaces NEGOTIATION_FAILED
+    # at descriptor-load time rather than crashing at execute.
+    Capability.MODEL_3D_GLB_INLINE,
+    # 3D MODEL BUNDLE envelope (multi-file: primary mesh plus N
+    # companion files — textures / .mtl / .bin / etc. — as a single
+    # coherent unit). The decoder in
+    # ``serialization.decode_model3d_envelope`` dispatches on
+    # ``encoding="bundle_inline"`` and returns a ``BundledFile3D``
+    # (a ``File3D`` subclass) whose ``get_source()`` / ``save_to()``
+    # materialise the whole bundle on disk so cross-file refs (.obj
+    # -> .mtl -> textures; .gltf -> .bin + textures) resolve via
+    # the filesystem. Servers gate bundle-emitting descriptors on
+    # this token; legacy clients surface NEGOTIATION_FAILED at
+    # descriptor-load time rather than crashing at execute.
+    Capability.MODEL_3D_BUNDLE_INLINE,
 ]
 
 # Reported in the X-RNP-Client-Version header.

notes/run_batch_uri_followups.py

@@ -0,0 +1,352 @@
+"""Smoke test for the three batch-URI followups (PR-2/PR-3 cleanup):
+
+1. ``serialization.maybe_externalize`` keys batch dispatch off the
+   ``encoding`` string (``png_base64_batch``) — NOT field presence
+   (``frames``).
+2. ``serialization.maybe_externalize`` rejects an empty/missing
+   ``frames`` list on a ``png_base64_batch`` envelope as a hard
+   ``RnpProtocolError`` rather than silently no-op-ing.
+3. ``proxy_node._deserialize_output`` accepts the externalised
+   ``png_base64_batch_uri`` shape on the response path: fetches every
+   ``uri`` (in parallel via ``asyncio.gather``), rewrites to inline
+   ``frames``, and decodes into a ``[B,H,W,C]`` tensor with frame
+   ordering preserved.
+
+Run with any python that has ``torch`` and ``Pillow`` available:
+
+    python notes/run_batch_uri_followups.py
+"""
+from __future__ import annotations
+
+import asyncio
+import base64
+import http.server
+import importlib.util
+import io
+import os
+import socketserver
+import sys
+import threading
+import time
+import types
+
+import numpy as np
+import torch
+from PIL import Image
+
+# ---------------------------------------------------------------------------
+# Locate the worktree under test.
+# ---------------------------------------------------------------------------
+_HERE = os.path.dirname(os.path.abspath(__file__))
+_CLIENT_DIR = os.path.dirname(_HERE)
+
+
+def _load_module(name: str, path: str) -> types.ModuleType:
+    spec = importlib.util.spec_from_file_location(name, path)
+    mod = importlib.util.module_from_spec(spec)
+    sys.modules[name] = mod
+    spec.loader.exec_module(mod)
+    return mod
+
+
+# Stub out client.upload_asset so the dispatch tests don't need a real
+# RNP server — the externalize batch path will call this once per frame.
+_uploads: list[bytes] = []
+
+
+async def _fake_upload_asset(server_url, raw, *, content_type, auth_headers=None):  # noqa: ANN001
+    _uploads.append(bytes(raw))
+    return f"https://fake-storage/{len(_uploads) - 1}"
+
+
+def _install_fake_client() -> None:
+    fake_client = types.ModuleType("comfy_remote_nodes.client")
+    fake_client.upload_asset = _fake_upload_asset
+    sys.modules["comfy_remote_nodes.client"] = fake_client
+
+
+def _png_b64_for_color(color: tuple[int, int, int]) -> str:
+    img = Image.new("RGB", (4, 4), color=color)
+    buf = io.BytesIO()
+    img.save(buf, format="PNG")
+    return base64.b64encode(buf.getvalue()).decode("ascii")
+
+
+def _png_bytes_for_color(color: tuple[int, int, int]) -> bytes:
+    img = Image.new("RGB", (4, 4), color=color)
+    buf = io.BytesIO()
+    img.save(buf, format="PNG")
+    return buf.getvalue()
+
+
+# ---------------------------------------------------------------------------
+# Test 1+2: maybe_externalize dispatch + empty-frames
+# ---------------------------------------------------------------------------
+
+async def _test_maybe_externalize() -> list[str]:
+    failures: list[str] = []
+
+    # Re-import serialization fresh because it caches a `client` ref.
+    sys.modules.pop("comfy_remote_nodes", None)
+    sys.modules.pop("comfy_remote_nodes.serialization", None)
+    sys.modules.pop("comfy_remote_nodes.protocol", None)
+
+    # Build a minimal package shim so `from . import client` works.
+    pkg = types.ModuleType("comfy_remote_nodes")
+    pkg.__path__ = [_CLIENT_DIR]
+    sys.modules["comfy_remote_nodes"] = pkg
+    _install_fake_client()
+    protocol = _load_module(
+        "comfy_remote_nodes.protocol",
+        os.path.join(_CLIENT_DIR, "protocol.py"),
+    )
+    serialization = _load_module(
+        "comfy_remote_nodes.serialization",
+        os.path.join(_CLIENT_DIR, "serialization.py"),
+    )
+
+    # --- Case A: batch envelope with payload bigger than cap → externalises
+    _uploads.clear()
+    big_frames = [_png_b64_for_color((i * 30, 0, 0)) for i in range(4)]
+    env = {
+        "type": "image",
+        "encoding": "png_base64_batch",
+        "frames": big_frames,
+        "shape": [4, 4, 4, 3],
+        "dtype": "uint8",
+        "color_space": "srgb",
+    }
+    out = await serialization.maybe_externalize(
+        env, server_url="http://fake", max_inline_bytes=1, auth_headers={},
+    )
+    if out.get("encoding") != "png_base64_batch_uri":
+        failures.append(
+            f"A: expected encoding=png_base64_batch_uri, got {out.get('encoding')!r}"
+        )
+    if "frames" in out:
+        failures.append("A: 'frames' field still present after externalise")
+    if not isinstance(out.get("uris"), list) or len(out["uris"]) != 4:
+        failures.append(f"A: expected 4 uris, got {out.get('uris')!r}")
+    if len(_uploads) != 4:
+        failures.append(f"A: expected 4 uploads, got {len(_uploads)}")
+
+    # --- Case B: dispatch keys off encoding, NOT field presence.
+    # An envelope with a stray `frames` field but a non-batch encoding
+    # must NOT be treated as a batch (would have under the old
+    # `isinstance(frames, list) and frames` dispatch).
+    _uploads.clear()
+    rogue = {
+        "type": "image",
+        "encoding": "png_base64",  # singleton encoding
+        "data": _png_b64_for_color((10, 20, 30)) * 200,  # > cap
+        "frames": ["bogus-extra-field"],
+    }
+    out_b = await serialization.maybe_externalize(
+        rogue, server_url="http://fake", max_inline_bytes=1, auth_headers={},
+    )
+    if out_b.get("encoding") == "png_base64_batch_uri":
+        failures.append(
+            "B: dispatch took batch branch on a singleton encoding (rogue frames field)"
+        )
+    if "uri" not in out_b:
+        failures.append("B: singleton encoding should have externalised to 'uri'")
+    if len(_uploads) != 1:
+        failures.append(f"B: expected 1 upload (singleton path), got {len(_uploads)}")
+
+    # --- Case C: empty frames on a batch encoding is a hard error.
+    bad = {
+        "type": "image",
+        "encoding": "png_base64_batch",
+        "frames": [],
+        "shape": [0, 4, 4, 3],
+    }
+    try:
+        await serialization.maybe_externalize(
+            bad, server_url="http://fake", max_inline_bytes=1, auth_headers={},
+        )
+    except protocol.RnpProtocolError:
+        pass
+    else:
+        failures.append("C: empty 'frames' on batch encoding did not raise RnpProtocolError")
+
+    # --- Case D: under-cap batch: no upload, envelope unchanged.
+    _uploads.clear()
+    small = {
+        "type": "image",
+        "encoding": "png_base64_batch",
+        "frames": big_frames,
+        "shape": [4, 4, 4, 3],
+    }
+    out_d = await serialization.maybe_externalize(
+        small,
+        server_url="http://fake",
+        max_inline_bytes=10 * 1024 * 1024,
+        auth_headers={},
+    )
+    if out_d is not small and out_d != small:
+        failures.append("D: under-cap batch should be a no-op")
+    if _uploads:
+        failures.append(f"D: under-cap batch should not upload, got {len(_uploads)}")
+
+    return failures
+
+
+# ---------------------------------------------------------------------------
+# Test 3: _deserialize_output handles png_base64_batch_uri on response
+# ---------------------------------------------------------------------------
+
+DELAY_S = 0.4
+N_FRAMES = 4
+
+
+class _SlowPngHandler(http.server.BaseHTTPRequestHandler):
+    def do_GET(self) -> None:  # noqa: N802
+        time.sleep(DELAY_S)
+        try:
+            idx = int(self.path.rsplit("/", 1)[-1])
+        except ValueError:
+            self.send_response(404); self.end_headers(); return
+        # Distinct color per frame so we can verify ordering after decode.
+        body = _png_bytes_for_color((idx * 60 % 256, 80, 120))
+        self.send_response(200)
+        self.send_header("Content-Type", "image/png")
+        self.send_header("Content-Length", str(len(body)))
+        self.end_headers()
+        self.wfile.write(body)
+
+    def log_message(self, *_a, **_kw) -> None:
+        return
+
+
+async def _test_deserialize_output() -> list[str]:
+    failures: list[str] = []
+
+    server = socketserver.ThreadingTCPServer(("127.0.0.1", 0), _SlowPngHandler)
+    host, port = server.server_address
+    th = threading.Thread(target=server.serve_forever, daemon=True)
+    th.start()
+    try:
+        # Stub the proxy_node's _fetch_url_envelope so we don't have to
+        # bring up the whole comfy_api_nodes.util.client stack.
+        import urllib.request as _urlreq
+
+        async def fake_fetch(envelope, policy, *, auth_headers, node_cls):  # noqa: ANN001, ARG001
+            # Real ``_fetch_url_envelope`` is properly async (aiohttp);
+            # the smoke stub uses sync urllib so we hop to a thread to
+            # preserve the parallel-gather semantics under test.
+            uri = envelope["uri"]
+            def _do_fetch() -> bytes:
+                with _urlreq.urlopen(uri, timeout=10.0) as resp:
+                    return resp.read()
+            return await asyncio.to_thread(_do_fetch)
+
+        # Re-import proxy_node helpers via AST to avoid loading ComfyUI.
+        import ast
+
+        with open(os.path.join(_CLIENT_DIR, "proxy_node.py"), "r", encoding="utf-8") as fh:
+            src = fh.read()
+        tree = ast.parse(src)
+        wanted = {"_deserialize_output"}
+        nodes = [n for n in tree.body if isinstance(n, ast.AsyncFunctionDef) and n.name in wanted]
+        if len(nodes) != len(wanted):
+            failures.append(f"missing AST nodes; found {[n.name for n in nodes]}")
+            return failures
+
+        # Re-import the protocol + serialization modules with the package
+        # shim so `from . import` works inside serialization.
+        sys.modules.pop("comfy_remote_nodes.serialization", None)
+        sys.modules.pop("comfy_remote_nodes.protocol", None)
+        pkg = types.ModuleType("comfy_remote_nodes")
+        pkg.__path__ = [_CLIENT_DIR]
+        sys.modules["comfy_remote_nodes"] = pkg
+        _install_fake_client()
+        protocol = _load_module(
+            "comfy_remote_nodes.protocol",
+            os.path.join(_CLIENT_DIR, "protocol.py"),
+        )
+        serialization = _load_module(
+            "comfy_remote_nodes.serialization",
+            os.path.join(_CLIENT_DIR, "serialization.py"),
+        )
+
+        ns: dict = {
+            "Any": object,
+            "RnpProtocolError": protocol.RnpProtocolError,
+            "ErrorCode": protocol.ErrorCode,
+            "is_envelope": protocol.is_envelope,
+            "serialization": serialization,
+            "_fetch_url_envelope": fake_fetch,
+            "log": types.SimpleNamespace(warning=lambda *a, **kw: None),
+        }
+        mod_src = ast.Module(body=nodes, type_ignores=[])
+        exec(compile(mod_src, "<batch_uri_smoke>", "exec"), ns)
+        deserialize = ns["_deserialize_output"]
+
+        # Build the externalised batch envelope on the response path.
+        uris = [f"http://{host}:{port}/frame/{i}" for i in range(N_FRAMES)]
+        envelope = {
+            "type": "image",
+            "encoding": "png_base64_batch_uri",
+            "uris": uris,
+            "shape": [N_FRAMES, 4, 4, 3],
+            "dtype": "uint8",
+            "color_space": "srgb",
+        }
+        url_fetch_policy = {"image": {"url_kind": "presigned", "timeout_s": 10}}
+
+        t0 = time.monotonic()
+        tensor = await deserialize(
+            envelope, meta=None,
+            url_fetch_policy=url_fetch_policy,
+            server_url="http://fake",
+            auth_headers={},
+            node_cls=None,
+        )
+        elapsed = time.monotonic() - t0
+
+        if not isinstance(tensor, torch.Tensor):
+            failures.append(f"expected torch.Tensor, got {type(tensor).__name__}")
+            return failures
+        if tuple(tensor.shape) != (N_FRAMES, 4, 4, 3):
+            failures.append(f"shape {tuple(tensor.shape)} != ({N_FRAMES}, 4, 4, 3)")
+
+        # Each frame's R-channel should match the expected color.
+        for i in range(N_FRAMES):
+            expected_r = (i * 60 % 256) / 255.0
+            actual_r = float(tensor[i, 0, 0, 0])
+            if abs(actual_r - expected_r) > 1e-3:
+                failures.append(
+                    f"frame {i}: R channel {actual_r:.3f} != expected {expected_r:.3f} "
+                    f"(ordering broken or wrong fetch)"
+                )
+
+        # Wall-clock: parallel should be ~DELAY_S, sequential would be ~N*DELAY_S.
+        sequential_lower = N_FRAMES * DELAY_S * 0.7
+        if elapsed >= sequential_lower:
+            failures.append(
+                f"wall clock {elapsed:.2f}s >= sequential lower bound "
+                f"{sequential_lower:.2f}s — gather not running in parallel"
+            )
+
+        return failures
+    finally:
+        server.shutdown(); server.server_close()
+
+
+def main() -> int:
+    failures: list[str] = []
+    failures += asyncio.run(_test_maybe_externalize())
+    failures += asyncio.run(_test_deserialize_output())
+    if failures:
+        print()
+        print("FAIL:")
+        for f in failures:
+            print(" -", f)
+        return 1
+    print("PASS: maybe_externalize dispatches on encoding + rejects empty frames; "
+          "_deserialize_output handles png_base64_batch_uri in parallel")
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())

notes/run_dynamic_combo_parser.py

@@ -0,0 +1,526 @@
+"""Smoke test for the DYNAMIC_COMBO + AUTOGROW parser extension.
+
+Verifies the proxy_node parser now constructs real V3
+``IO.DynamicCombo.Input`` and ``IO.Autogrow.Input`` objects from the
+flattened RNP wire shapes that the server already emits, instead of
+falling through to the opaque ``IO.Custom`` bucket.
+
+Checks:
+
+1.  Bria-shaped 2-branch DYNAMIC_COMBO (the "false" branch has empty
+    ``inputs``; the "true" branch has 3 BOOLEAN sub-widgets) parses
+    into an ``IO.DynamicCombo.Input`` whose two ``Option``s carry the
+    right keys + sub-input objects. Proves empty branches are kept and
+    that the outer id is whatever the descriptor key was (here
+    ``"moderation"``, *not* ``"model"``).
+2.  Grok-V2-shaped DYNAMIC_COMBO with three branches, each containing
+    a nested AUTOGROW(IMAGE) plus a COMBO + INT (and an aspect_ratio
+    COMBO on 2 of the 3 branches). Asserts the AUTOGROW becomes a real
+    ``IO.Autogrow.Input`` with a ``TemplatePrefix`` whose underlying
+    input is ``IO.Image.Input``, and that prefix / min / max survive.
+3.  ElevenLabs-shaped DYNAMIC_COMBO whose branches mix
+    ``IO.Custom(<opaque_io_type>).Input`` (voiceN custom-IO) with
+    primitive STRING sub-widgets — proves the recursive opaque
+    fallback path inside a branch still works for partner helper
+    types.
+4.  Synthetic nested DYNAMIC_COMBO-inside-DYNAMIC_COMBO — no provider
+    uses this today, but defensive coverage so the recursive entry
+    point keeps working.
+5.  Malformed DYNAMIC_COMBO / AUTOGROW shapes return ``None`` (so the
+    caller skips the descriptor entirely) instead of silently
+    falling through to the opaque IO.Custom bucket.
+6.  Regression: ``_collect_local_validate`` output is byte-identical
+    before and after the parser change (the rules-extraction path is
+    independent of input construction and must not regress).
+7.  AUTOGROW template that is itself a dynamic input is rejected
+    (upstream Autogrow asserts this).
+8.  Source-level check: the proxy_node docstring + dispatch covers
+    DYNAMIC_COMBO and AUTOGROW (not just the existing primitive set).
+
+Run with the ComfyUI venv (needs torch / comfy_api on PYTHONPATH):
+
+    python notes/run_dynamic_combo_parser.py
+"""
+from __future__ import annotations
+
+import importlib.util
+import os
+import sys
+import types
+
+# ---------------------------------------------------------------------------
+# Locate the worktree under test.
+# ---------------------------------------------------------------------------
+_HERE = os.path.dirname(os.path.abspath(__file__))
+_CLIENT_DIR = os.path.dirname(_HERE)
+_COMFYUI_DIR = os.path.dirname(os.path.dirname(_CLIENT_DIR))
+
+
+def _load_module(name: str, path: str) -> types.ModuleType:
+    spec = importlib.util.spec_from_file_location(name, path)
+    mod = importlib.util.module_from_spec(spec)
+    sys.modules[name] = mod
+    spec.loader.exec_module(mod)
+    return mod
+
+
+# ---------------------------------------------------------------------------
+# Stub all the heavy deps proxy_node + comfy_api/latest/_io.py pull in.
+#
+# We only need three things from the world:
+#
+# * the real V3 IO classes (DynamicCombo / Autogrow / Image / Combo / String /
+#   Int / Boolean / Custom) — load ``_io.py`` standalone so we don't pay the
+#   full ``comfy_api.latest`` package init (PIL, numpy, comfy.cli_args,
+#   comfy_execution, the sync-class generator etc.).
+# * the real proxy_node parser helpers — load ``proxy_node.py`` after the
+#   ``comfy_api.latest`` shim is in place.
+# * the real protocol module (lightweight; for the ErrorCode / Header / etc.
+#   imports proxy_node does up top).
+#
+# Everything else (server.PromptServer, aiohttp, comfy_api_nodes.*) is
+# stubbed because proxy_node only references those at request-time, not at
+# import-time-after-trivial-stub.
+# ---------------------------------------------------------------------------
+
+# 1) Stub ``torch`` — _io.py uses ``torch.Tensor`` only in type annotations
+#    (``Type = torch.Tensor`` class-level field), so a bare-class stub works.
+_torch_stub = types.ModuleType("torch")
+class _Tensor:  # noqa: D401
+    pass
+_torch_stub.Tensor = _Tensor
+sys.modules.setdefault("torch", _torch_stub)
+
+# 2) Stub ``comfy_execution.graph_utils.ExecutionBlocker`` — only used as a
+#    class reference inside Output construction; never instantiated here.
+_comfy_execution = types.ModuleType("comfy_execution")
+_graph_utils = types.ModuleType("comfy_execution.graph_utils")
+class _ExecutionBlocker:  # noqa: D401
+    pass
+_graph_utils.ExecutionBlocker = _ExecutionBlocker
+sys.modules.setdefault("comfy_execution", _comfy_execution)
+sys.modules.setdefault("comfy_execution.graph_utils", _graph_utils)
+
+# 3) Stub ``comfy_api.internal`` — _io.py imports a handful of helper names.
+#    The parser code paths exercised here don't actually invoke any of these;
+#    bare-function / bare-class stubs are sufficient to satisfy the import.
+_comfy_api = types.ModuleType("comfy_api")
+_comfy_api.__path__ = []
+_comfy_api_internal = types.ModuleType("comfy_api.internal")
+class _ComfyNodeInternal:  # noqa: D401
+    pass
+class _NodeOutputInternal:  # noqa: D401
+    pass
+class _classproperty:  # noqa: D401
+    def __init__(self, f):
+        self.f = f
+    def __get__(self, instance, owner):
+        return self.f(owner)
+def _identity(x, *_args, **_kw):
+    return x
+def _first_real_override(*_a, **_kw):
+    return None
+def _is_class(x):
+    return isinstance(x, type)
+def _prune_dict(d):
+    return {k: v for k, v in (d or {}).items() if v is not None}
+_comfy_api_internal._ComfyNodeInternal = _ComfyNodeInternal
+_comfy_api_internal._NodeOutputInternal = _NodeOutputInternal
+_comfy_api_internal.classproperty = _classproperty
+_comfy_api_internal.copy_class = _identity
+_comfy_api_internal.first_real_override = _first_real_override
+_comfy_api_internal.is_class = _is_class
+_comfy_api_internal.prune_dict = _prune_dict
+_comfy_api_internal.shallow_clone_class = _identity
+sys.modules.setdefault("comfy_api", _comfy_api)
+sys.modules.setdefault("comfy_api.internal", _comfy_api_internal)
+
+# 3b) Stub ``comfy_api.input`` for the lazy imports inside _io.py's
+#     Curve / RangeInput class bodies — the parser code never touches
+#     those types, but the *class body* runs at module load.
+_comfy_api_input = types.ModuleType("comfy_api.input")
+class _CurvePoint: pass  # noqa: E701,D401
+class _RangeInput: pass  # noqa: E701,D401
+_comfy_api_input.CurvePoint = _CurvePoint
+_comfy_api_input.RangeInput = _RangeInput
+sys.modules.setdefault("comfy_api.input", _comfy_api_input)
+
+# 4) Stub ``comfy_api.latest._util`` — _io.py imports MESH / VOXEL / SVG /
+#    File3D as type-token classes. We don't construct any here.
+_comfy_api_latest = types.ModuleType("comfy_api.latest")
+_comfy_api_latest.__path__ = []
+_comfy_api_latest_util = types.ModuleType("comfy_api.latest._util")
+class _MESH: pass  # noqa: E701,D401
+class _VOXEL: pass  # noqa: E701,D401
+class _SVG: pass  # noqa: E701,D401
+class _File3D: pass  # noqa: E701,D401
+_comfy_api_latest_util.MESH = _MESH
+_comfy_api_latest_util.VOXEL = _VOXEL
+_comfy_api_latest_util.SVG = _SVG
+_comfy_api_latest_util.File3D = _File3D
+sys.modules.setdefault("comfy_api.latest", _comfy_api_latest)
+sys.modules.setdefault("comfy_api.latest._util", _comfy_api_latest_util)
+
+# 5) Load the real ``_io.py`` as ``comfy_api.latest._io`` — its
+#    relative ``from ._util import ...`` requires the parent-package
+#    setup we just did.
+_io_path = os.path.join(_COMFYUI_DIR, "comfy_api", "latest", "_io.py")
+_io_mod = _load_module("comfy_api.latest._io", _io_path)
+# Expose ``IO`` exactly like the real package does
+# (``from . import _io_public as io; IO = io`` where _io_public is just
+# ``from ._io import *``).
+_comfy_api_latest.IO = _io_mod
+_comfy_api_latest.io = _io_mod
+IO = _io_mod  # local alias for the assertions below
+
+# 6) Stub proxy_node's other heavyweight imports.
+for missing in (
+    "server",
+    "aiohttp",
+    "comfy_api_nodes",
+    "comfy_api_nodes.util",
+    "comfy_api_nodes.util.client",
+    "comfy_api_nodes.util.common_exceptions",
+):
+    if missing not in sys.modules:
+        m = types.ModuleType(missing)
+        sys.modules[missing] = m
+        m.__path__ = []  # type: ignore[attr-defined]
+sys.modules["server"].PromptServer = type("_StubPS", (), {})
+sys.modules["comfy_api_nodes.util.client"].ApiEndpoint = type("_StubAE", (), {})
+sys.modules["comfy_api_nodes.util.client"].poll_op_raw = lambda *a, **k: None
+sys.modules["comfy_api_nodes.util.common_exceptions"].ProcessingInterrupted = type(
+    "_StubPI", (Exception,), {},
+)
+
+# 7) Emulate the package layout for the client so its relative imports
+#    work (``from .protocol import ...`` etc.).
+pkg = types.ModuleType("comfy_remote_nodes_test")
+pkg.__path__ = [_CLIENT_DIR]
+sys.modules["comfy_remote_nodes_test"] = pkg
+sys.modules["comfy_remote_nodes_test.client"] = types.ModuleType(
+    "comfy_remote_nodes_test.client",
+)
+sys.modules["comfy_remote_nodes_test.serialization"] = types.ModuleType(
+    "comfy_remote_nodes_test.serialization",
+)
+_load_module(
+    "comfy_remote_nodes_test.protocol",
+    os.path.join(_CLIENT_DIR, "protocol.py"),
+)
+
+proxy_node = _load_module(
+    "comfy_remote_nodes_test.proxy_node",
+    os.path.join(_CLIENT_DIR, "proxy_node.py"),
+)
+
+
+# ---------------------------------------------------------------------------
+# 1. Bria-shaped 2-branch DYNAMIC_COMBO.
+# ---------------------------------------------------------------------------
+bria_spec = [
+    "DYNAMIC_COMBO",
+    {
+        "options": [
+            {"key": "false", "inputs": []},
+            {
+                "key": "true",
+                "inputs": [
+                    ["prompt_content_moderation", "BOOLEAN", {"default": False}],
+                    ["visual_input_moderation", "BOOLEAN", {"default": False}],
+                    ["visual_output_moderation", "BOOLEAN", {"default": True}],
+                ],
+            },
+        ],
+        "tooltip": "Moderation settings",
+    },
+]
+inp = proxy_node._parse_input_spec("moderation", bria_spec, optional=False)
+assert inp is not None, "Bria DYNAMIC_COMBO returned None"
+assert isinstance(inp, IO.DynamicCombo.Input), (
+    f"Expected IO.DynamicCombo.Input, got {type(inp).__name__}"
+)
+assert inp.id == "moderation", f"outer id should track descriptor key, got {inp.id!r}"
+assert len(inp.options) == 2, f"expected 2 branches, got {len(inp.options)}"
+assert [o.key for o in inp.options] == ["false", "true"], (
+    f"branch keys wrong: {[o.key for o in inp.options]}"
+)
+# Empty branch preserved (UX needs the dropdown entry).
+assert inp.options[0].inputs == [], "empty 'false' branch dropped"
+assert len(inp.options[1].inputs) == 3, (
+    f"'true' branch should have 3 sub-inputs, got {len(inp.options[1].inputs)}"
+)
+for sub in inp.options[1].inputs:
+    assert isinstance(sub, IO.Boolean.Input), (
+        f"branch sub-input should be IO.Boolean.Input, got {type(sub).__name__}"
+    )
+assert [s.id for s in inp.options[1].inputs] == [
+    "prompt_content_moderation",
+    "visual_input_moderation",
+    "visual_output_moderation",
+], "branch sub-input names lost"
+print("PASS: Bria DYNAMIC_COMBO parses with empty + populated branches.")
+
+
+# ---------------------------------------------------------------------------
+# 2. Grok-V2-shaped DYNAMIC_COMBO with nested AUTOGROW(IMAGE) per branch.
+# ---------------------------------------------------------------------------
+def _grok_v2_subinputs(*, max_ref_images: int, with_aspect_ratio: bool) -> list:
+    inputs: list = [
+        [
+            "images",
+            "AUTOGROW",
+            {
+                "template": [
+                    "IMAGE",
+                    {"local_validate": {"image_max_batch": max_ref_images}},
+                ],
+                "prefix": "image",
+                "min": 1,
+                "max": max_ref_images,
+                "tooltip": (
+                    "Reference image to edit."
+                    if max_ref_images == 1
+                    else f"Reference image(s) to edit. Up to {max_ref_images} images."
+                ),
+            },
+        ],
+        ["resolution", "COMBO", {"options": ["768p", "1k"]}],
+        ["number_of_images", "INT", {
+            "default": 1, "min": 1, "max": 10, "step": 1,
+        }],
+    ]
+    if with_aspect_ratio:
+        inputs.append(["aspect_ratio", "COMBO", {
+            "options": ["auto", "1:1", "16:9"], "default": "auto",
+        }])
+    return inputs
+
+grok_spec = [
+    "DYNAMIC_COMBO",
+    {
+        "options": [
+            {
+                "key": "grok-imagine-image-quality",
+                "inputs": _grok_v2_subinputs(max_ref_images=3, with_aspect_ratio=True),
+            },
+            {
+                "key": "grok-imagine-image-pro",
+                "inputs": _grok_v2_subinputs(max_ref_images=1, with_aspect_ratio=False),
+            },
+            {
+                "key": "grok-imagine-image",
+                "inputs": _grok_v2_subinputs(max_ref_images=3, with_aspect_ratio=True),
+            },
+        ],
+        "tooltip": "The model to use for editing.",
+    },
+]
+grok_inp = proxy_node._parse_input_spec("model", grok_spec, optional=False)
+assert grok_inp is not None, "Grok V2 DYNAMIC_COMBO returned None"
+assert isinstance(grok_inp, IO.DynamicCombo.Input), (
+    f"Expected IO.DynamicCombo.Input, got {type(grok_inp).__name__}"
+)
+assert len(grok_inp.options) == 3
+# Pro branch (max_ref_images=1, no aspect_ratio).
+pro_branch = next(o for o in grok_inp.options if o.key == "grok-imagine-image-pro")
+assert len(pro_branch.inputs) == 3, f"pro branch should have 3 sub-inputs, got {len(pro_branch.inputs)}"
+images_inp = pro_branch.inputs[0]
+assert isinstance(images_inp, IO.Autogrow.Input), (
+    f"Expected nested AUTOGROW to parse as IO.Autogrow.Input, got "
+    f"{type(images_inp).__name__}"
+)
+assert images_inp.id == "images"
+assert isinstance(images_inp.template, IO.Autogrow.TemplatePrefix), (
+    f"Expected TemplatePrefix template, got {type(images_inp.template).__name__}"
+)
+assert images_inp.template.prefix == "image"
+assert images_inp.template.min == 1
+assert images_inp.template.max == 1, f"pro branch max should be 1, got {images_inp.template.max}"
+assert isinstance(images_inp.template.input, IO.Image.Input), (
+    f"AUTOGROW template input should be IO.Image.Input, got "
+    f"{type(images_inp.template.input).__name__}"
+)
+# Quality branch (max_ref_images=3 + aspect_ratio).
+qual_branch = next(o for o in grok_inp.options if o.key == "grok-imagine-image-quality")
+assert len(qual_branch.inputs) == 4, f"quality branch should have 4 sub-inputs, got {len(qual_branch.inputs)}"
+qual_images = qual_branch.inputs[0]
+assert isinstance(qual_images, IO.Autogrow.Input)
+assert qual_images.template.max == 3, f"quality branch max should be 3, got {qual_images.template.max}"
+# Primitive widgets after the autogrow.
+assert isinstance(qual_branch.inputs[1], IO.Combo.Input)
+assert isinstance(qual_branch.inputs[2], IO.Int.Input)
+assert isinstance(qual_branch.inputs[3], IO.Combo.Input)
+print("PASS: Grok V2 DYNAMIC_COMBO + nested AUTOGROW parses correctly.")
+
+
+# ---------------------------------------------------------------------------
+# 3. ElevenLabs-shaped branch with custom-IO sub-inputs (voiceN).
+# ---------------------------------------------------------------------------
+el_spec = [
+    "DYNAMIC_COMBO",
+    {
+        "options": [
+            {
+                "key": "2",
+                "inputs": [
+                    ["voice0", "ELEVENLABS_VOICE", {}],
+                    ["text0", "STRING", {"multiline": True, "default": ""}],
+                    ["voice1", "ELEVENLABS_VOICE", {}],
+                    ["text1", "STRING", {"multiline": True, "default": ""}],
+                ],
+            },
+        ],
+        "tooltip": "Number of dialogue turns.",
+    },
+]
+el_inp = proxy_node._parse_input_spec("inputs", el_spec, optional=False)
+assert el_inp is not None, "ElevenLabs DYNAMIC_COMBO returned None"
+assert isinstance(el_inp, IO.DynamicCombo.Input)
+assert el_inp.id == "inputs", (
+    f"outer id should be 'inputs' (not always 'model'), got {el_inp.id!r}"
+)
+branch = el_inp.options[0]
+# IO.Custom("FOO") returns a synthesized ComfyTypeIO subclass — each call
+# creates a *new* class, so isinstance-against-a-fresh-Custom doesn't
+# work. The @comfytype decorator stamps ``io_type`` onto the Input class
+# instead; check that the synthesized custom Input carries the right
+# string (this is how the frontend's connection-validity check chains
+# sockets).
+voice0 = branch.inputs[0]
+assert voice0.get_io_type() == "ELEVENLABS_VOICE", (
+    f"voice0 should be opaque ELEVENLABS_VOICE Input, got "
+    f"{voice0.get_io_type()!r} on {type(voice0).__name__}"
+)
+assert isinstance(branch.inputs[1], IO.String.Input)
+assert isinstance(branch.inputs[3], IO.String.Input)
+print("PASS: ElevenLabs DYNAMIC_COMBO branch mixes custom-IO + primitive.")
+
+
+# ---------------------------------------------------------------------------
+# 4. Synthetic nested DYNAMIC_COMBO inside DYNAMIC_COMBO.
+# ---------------------------------------------------------------------------
+nested = [
+    "DYNAMIC_COMBO",
+    {
+        "options": [
+            {
+                "key": "outer_a",
+                "inputs": [
+                    [
+                        "submodel",
+                        "DYNAMIC_COMBO",
+                        {
+                            "options": [
+                                {"key": "inner_x", "inputs": [
+                                    ["x", "INT", {"default": 1, "min": 0, "max": 10}],
+                                ]},
+                                {"key": "inner_y", "inputs": []},
+                            ],
+                            "tooltip": "Inner branch.",
+                        },
+                    ],
+                ],
+            },
+        ],
+        "tooltip": "Outer branch.",
+    },
+]
+nest_inp = proxy_node._parse_input_spec("model", nested, optional=False)
+assert isinstance(nest_inp, IO.DynamicCombo.Input)
+inner = nest_inp.options[0].inputs[0]
+assert isinstance(inner, IO.DynamicCombo.Input), (
+    f"nested DynamicCombo should be IO.DynamicCombo.Input, got {type(inner).__name__}"
+)
+assert [o.key for o in inner.options] == ["inner_x", "inner_y"]
+assert isinstance(inner.options[0].inputs[0], IO.Int.Input)
+print("PASS: nested DYNAMIC_COMBO-in-DYNAMIC_COMBO parses recursively.")
+
+
+# ---------------------------------------------------------------------------
+# 5. Malformed dynamic specs return None (no opaque fallthrough).
+# ---------------------------------------------------------------------------
+# Missing options list on DYNAMIC_COMBO.
+assert proxy_node._parse_input_spec("x", ["DYNAMIC_COMBO", {}], optional=False) is None
+# Option missing key.
+bad = ["DYNAMIC_COMBO", {"options": [{"inputs": []}]}]
+assert proxy_node._parse_input_spec("x", bad, optional=False) is None
+# Option's inputs entry not a 3-tuple.
+bad2 = ["DYNAMIC_COMBO", {"options": [{"key": "a", "inputs": [["x"]]}]}]
+assert proxy_node._parse_input_spec("x", bad2, optional=False) is None
+# AUTOGROW without template.
+assert proxy_node._parse_input_spec("x", ["AUTOGROW", {"prefix": "p"}], optional=False) is None
+# AUTOGROW without prefix/names.
+assert (
+    proxy_node._parse_input_spec("x", ["AUTOGROW", {"template": ["IMAGE", {}]}], optional=False)
+    is None
+)
+print("PASS: malformed DYNAMIC_COMBO / AUTOGROW specs return None.")
+
+
+# ---------------------------------------------------------------------------
+# 6. Regression: _collect_local_validate output unchanged for these inputs.
+# ---------------------------------------------------------------------------
+desc_grok = {
+    "input": {
+        "required": {"model": grok_spec},
+        "optional": {},
+    },
+}
+rules = proxy_node._collect_local_validate(desc_grok)
+# rules['model'] should carry __branches__ with per-branch maps that each
+# carry an __template__ entry under "images" with image_max_batch.
+assert "model" in rules, f"_collect_local_validate dropped 'model': {rules!r}"
+br = rules["model"].get("__branches__")
+assert isinstance(br, dict) and set(br) == {
+    "grok-imagine-image-quality",
+    "grok-imagine-image-pro",
+    "grok-imagine-image",
+}, f"branches mismatch: {br!r}"
+pro = br["grok-imagine-image-pro"]
+assert "images" in pro, f"pro branch should carry 'images' rule: {pro!r}"
+tmpl_rules = pro["images"].get("__template__")
+assert tmpl_rules == {"image_max_batch": 1}, (
+    f"pro branch template rules wrong: {tmpl_rules!r}"
+)
+qual = br["grok-imagine-image-quality"]
+assert qual["images"].get("__template__") == {"image_max_batch": 3}
+print("PASS: _collect_local_validate output unchanged (branches + template intact).")
+
+
+# ---------------------------------------------------------------------------
+# 7. AUTOGROW with a dynamic template input is rejected.
+# ---------------------------------------------------------------------------
+ag_with_dyn_template = [
+    "AUTOGROW",
+    {
+        "template": [
+            "DYNAMIC_COMBO",
+            {"options": [{"key": "k", "inputs": []}]},
+        ],
+        "prefix": "img",
+        "min": 1,
+        "max": 3,
+    },
+]
+assert (
+    proxy_node._parse_input_spec("bad", ag_with_dyn_template, optional=False) is None
+), "AUTOGROW with DYNAMIC_COMBO template should be rejected"
+print("PASS: AUTOGROW rejects dynamic template input (matches upstream assert).")
+
+
+# ---------------------------------------------------------------------------
+# 8. Source-level: docstring + dispatch mention DYNAMIC_COMBO + AUTOGROW.
+# ---------------------------------------------------------------------------
+proxy_src = open(
+    os.path.join(_CLIENT_DIR, "proxy_node.py"), "r", encoding="utf-8",
+).read()
+assert "DYNAMIC_COMBO" in proxy_src, "proxy_node source missing DYNAMIC_COMBO branch"
+assert "AUTOGROW" in proxy_src, "proxy_node source missing AUTOGROW branch"
+assert "_parse_named_input_spec" in proxy_src, (
+    "proxy_node source missing _parse_named_input_spec helper"
+)
+assert "_build_input" in proxy_src, "proxy_node source missing _build_input helper"
+print("PASS: proxy_node source contains the new dispatch + helpers.")
+
+print("ALL CHECKS PASSED.")

notes/run_encode_nested_dicts.py

@@ -0,0 +1,360 @@
+"""Smoke test for nested dict recursion in ``_encode_one`` /
+``_encode_inputs`` / ``_externalize_nested``.
+
+Background: ComfyUI's V3 ``build_nested_inputs`` rebuilds AUTOGROW and
+DYNAMIC_COMBO runtime kwargs into nested dicts (e.g. ``{"reference_images":
+{"image1": <tensor>, "image2": <tensor>}}`` for a top-level AG, or
+``{"model": {"branch": "wan2.7-r2v", "video_refs": {"video1":
+<VideoInput>}}}`` for AG-inside-DC). Before this fix, ``_encode_one``
+returned a non-tensor non-VideoInput non-audio dict unchanged, which
+meant per-slot tensors / VideoInputs survived past encoding and would
+either crash at ``json.dumps`` time or violate the server's contract
+(comfy_rnp_server's ``_ordered_wan_videoedit_image_envelopes`` etc.
+expect each slot value to already be an RNP envelope).
+
+Asserts:
+
+1. Top-level AG-of-IMAGE dict: each ``image<N>`` tensor becomes an
+   image envelope; the wrapping dict structure is preserved.
+2. Top-level AG-of-VIDEO dict: each ``video<N>`` VideoInput becomes a
+   video envelope.
+3. DC-wrapping-AG: top-level dict carries a branch key (string) AND a
+   nested AG dict of tensors / VideoInputs; encoder recurses one level
+   deeper and encodes the leaves while leaving the branch key untouched.
+4. AUDIO input is NOT swallowed by the dict-recursion branch: the
+   ``{"waveform": ..., "sample_rate": ...}`` shape still routes to
+   ``encode_audio_input``.
+5. Already-encoded envelope dicts pass through unchanged (no double-
+   encoding) — protects against re-running the encoder on a value that
+   the caller pre-encoded.
+6. Plain scalar config dicts (``{"width": 1024, "height": 576}``) pass
+   through unchanged.
+7. ``_externalize_nested`` walks the same shape: a nested AG dict of
+   oversize envelopes uploads each envelope via the stubbed externalize
+   path and the envelope's ``data`` field is swapped for ``uri``;
+   non-envelope siblings are untouched; the wrapping dict structure is
+   preserved.
+8. JSON-serializability post-condition: ``json.dumps`` round-trips the
+   full encoded payload (regression guard for the exact bug fixed).
+9. Single-tensor / single-VideoInput / single-audio top-level inputs
+   still encode the same way they did before (no regression for the
+   existing leaf path).
+
+Run with any python that has ``torch`` available (uses the ComfyUI
+venv on this workstation):
+
+    python notes/run_encode_nested_dicts.py
+
+The test loads ``_encode_one`` / ``_encode_inputs`` / ``_externalize_nested``
+out of ``proxy_node.py`` via AST extraction so we don't pay the cost
+of importing ComfyUI's ``comfy_api`` / ``comfy_api_nodes`` packages
+(which those helpers don't touch). Same standalone pattern as
+``run_image_max_batch.py``.
+"""
+from __future__ import annotations
+
+import ast
+import asyncio
+import importlib.util
+import json
+import os
+import sys
+import types
+
+import torch
+
+# ---------------------------------------------------------------------------
+# Locate the worktree under test (this file's parent's parent).
+# ---------------------------------------------------------------------------
+
+_HERE = os.path.dirname(os.path.abspath(__file__))
+_CLIENT_DIR = os.path.dirname(_HERE)
+
+
+def _load_protocol() -> types.ModuleType:
+    """Load ``protocol.py`` standalone (no relative imports)."""
+    path = os.path.join(_CLIENT_DIR, "protocol.py")
+    spec = importlib.util.spec_from_file_location("rnp_smoke_protocol", path)
+    mod = importlib.util.module_from_spec(spec)
+    sys.modules[spec.name] = mod
+    spec.loader.exec_module(mod)
+    return mod
+
+
+class _FakeVideoInput:
+    """Duck-typed VideoInput stand-in: ``save_to`` + ``get_duration``
+    are the only attrs ``serialization.is_video_input`` checks for."""
+    def __init__(self, tag: str, duration: float = 4.0) -> None:
+        self.tag = tag
+        self._duration = duration
+
+    def save_to(self, *args, **kwargs):  # pragma: no cover — never called by stub
+        return None
+
+    def get_duration(self) -> float:
+        return self._duration
+
+
+def _build_serialization_stub(protocol: types.ModuleType) -> types.ModuleType:
+    """Minimal ``serialization`` shim covering exactly the surface the
+    helpers under test use."""
+    mod = types.ModuleType("rnp_smoke_serialization")
+
+    def _is_torch_tensor(value):  # noqa: ANN001
+        return isinstance(value, torch.Tensor)
+
+    def is_audio_input(value):  # noqa: ANN001
+        return (
+            isinstance(value, dict)
+            and {"waveform", "sample_rate"} <= set(value.keys())
+        )
+
+    def is_video_input(value):  # noqa: ANN001
+        return (
+            callable(getattr(value, "save_to", None))
+            and callable(getattr(value, "get_duration", None))
+        )
+
+    def is_model3d_input(value):  # noqa: ANN001
+        return False  # unused in this test
+
+    def encode_image_tensor(tensor, *, accepts_batch=False):  # noqa: ANN001, ARG001
+        return {
+            "type": "image",
+            "encoding": "png_base64",
+            "data": f"img:{int(tensor.shape[0])}",
+        }
+
+    def encode_mask_tensor(tensor):  # noqa: ANN001, ARG001
+        return {"type": "mask", "encoding": "png_base64", "data": "mask"}
+
+    def encode_audio_input(value):  # noqa: ANN001, ARG001
+        return {"type": "audio", "encoding": "mp3_base64", "data": "aud"}
+
+    def encode_video_input(value):  # noqa: ANN001
+        return {
+            "type": "video",
+            "encoding": "mp4_base64",
+            "data": f"vid:{getattr(value, 'tag', '?')}",
+        }
+
+    def encode_model3d_input(value):  # noqa: ANN001, ARG001  # pragma: no cover
+        return {"type": "model_3d", "encoding": "glb_base64", "data": "m3d"}
+
+    HEAVY_TYPES = {"image", "mask", "audio", "video", "model_3d"}
+
+    def is_envelope(value):  # noqa: ANN001
+        return (
+            isinstance(value, dict)
+            and isinstance(value.get("type"), str)
+            and value.get("type") in HEAVY_TYPES
+            and isinstance(value.get("encoding"), str)
+        )
+
+    # Cap-aware ``maybe_externalize`` stub: swap ``data`` for ``uri`` when
+    # the inline payload would exceed ``max_inline_bytes``. Lets us prove
+    # that ``_externalize_nested`` reaches every per-slot envelope.
+    upload_log: list[str] = []
+
+    async def maybe_externalize(
+        envelope,  # noqa: ANN001
+        *, server_url=None, max_inline_bytes=None, auth_headers=None,  # noqa: ANN001, ARG001
+    ):
+        if not server_url or max_inline_bytes is None:
+            return envelope
+        data = envelope.get("data")
+        if not isinstance(data, str):
+            return envelope
+        if len(data) <= max_inline_bytes:
+            return envelope
+        upload_log.append(f"{envelope.get('type')}:{data}")
+        out = {k: v for k, v in envelope.items() if k != "data"}
+        out["uri"] = f"https://upload/{envelope.get('type')}/{len(upload_log)}"
+        return out
+
+    mod._is_torch_tensor = _is_torch_tensor
+    mod.is_audio_input = is_audio_input
+    mod.is_video_input = is_video_input
+    mod.is_model3d_input = is_model3d_input
+    mod.encode_image_tensor = encode_image_tensor
+    mod.encode_mask_tensor = encode_mask_tensor
+    mod.encode_audio_input = encode_audio_input
+    mod.encode_video_input = encode_video_input
+    mod.encode_model3d_input = encode_model3d_input
+    mod.is_envelope = is_envelope
+    mod.maybe_externalize = maybe_externalize
+    mod._upload_log = upload_log  # exposed for assertions
+    return mod
+
+
+def _build_proxy_node_stub(
+    protocol: types.ModuleType, serialization: types.ModuleType,
+) -> types.ModuleType:
+    """Extract the helpers under test from ``proxy_node.py``."""
+    mod = types.ModuleType("rnp_smoke_proxy_node")
+    mod.__dict__.update({
+        "Any": object,
+        "log": types.SimpleNamespace(warning=lambda *a, **kw: None),
+        "serialization": serialization,
+        "RnpProtocolError": protocol.RnpProtocolError,
+        "ErrorCode": protocol.ErrorCode,
+    })
+    src_path = os.path.join(_CLIENT_DIR, "proxy_node.py")
+    with open(src_path, "r", encoding="utf-8") as fh:
+        src = fh.read()
+    tree = ast.parse(src)
+    wanted = {
+        "_enforce_local_validate",
+        "_check_image_max_batch",
+        "_encode_one",
+        "_encode_inputs",
+        "_externalize_nested",
+    }
+    nodes = [
+        n for n in tree.body
+        if isinstance(n, (ast.AsyncFunctionDef, ast.FunctionDef))
+        and n.name in wanted
+    ]
+    snippet = ast.Module(body=nodes, type_ignores=[])
+    code = compile(snippet, src_path, "exec")
+    exec(code, mod.__dict__)
+    missing = wanted - set(mod.__dict__)
+    if missing:
+        raise RuntimeError(f"proxy_node helpers missing from extract: {missing}")
+    return mod
+
+
+# ---------------------------------------------------------------------------
+# Tests
+# ---------------------------------------------------------------------------
+
+async def main() -> int:
+    protocol = _load_protocol()
+    serialization = _build_serialization_stub(protocol)
+    proxy_node = _build_proxy_node_stub(protocol, serialization)
+
+    img1 = torch.zeros((1, 64, 64, 3), dtype=torch.float32)
+    img2 = torch.zeros((1, 32, 32, 3), dtype=torch.float32)
+    vid1 = _FakeVideoInput("v1")
+    vid2 = _FakeVideoInput("v2")
+    audio = {
+        "waveform": torch.zeros((1, 1, 1000), dtype=torch.float32),
+        "sample_rate": 16000,
+    }
+
+    # ---- 1. Top-level AG-of-IMAGE dict encodes per-slot
+    out = await proxy_node._encode_inputs(
+        {"reference_images": {"image1": img1, "image2": img2}},
+    )
+    refs = out["reference_images"]
+    assert isinstance(refs, dict) and set(refs.keys()) == {"image1", "image2"}, refs
+    assert refs["image1"]["type"] == "image" and refs["image1"]["data"] == "img:1"
+    assert refs["image2"]["type"] == "image" and refs["image2"]["data"] == "img:1"
+    print("ok: top-level AG-of-IMAGE encoded per slot")
+
+    # ---- 2. Top-level AG-of-VIDEO dict encodes per-slot
+    out = await proxy_node._encode_inputs(
+        {"video_refs": {"video1": vid1, "video2": vid2}},
+    )
+    vrefs = out["video_refs"]
+    assert isinstance(vrefs, dict) and set(vrefs.keys()) == {"video1", "video2"}, vrefs
+    assert vrefs["video1"]["type"] == "video" and vrefs["video1"]["data"] == "vid:v1"
+    assert vrefs["video2"]["type"] == "video" and vrefs["video2"]["data"] == "vid:v2"
+    print("ok: top-level AG-of-VIDEO encoded per slot")
+
+    # ---- 3. DC-wrapping-AG: branch key untouched, nested AG dict encoded
+    dc_value = {
+        "model": "wan2.7-r2v",
+        "video_refs": {"video1": vid1},
+        "reference_images": {"image1": img1, "image2": img2},
+        "duration": 5,
+    }
+    out = await proxy_node._encode_inputs({"model": dc_value})
+    enc = out["model"]
+    assert enc["model"] == "wan2.7-r2v", enc
+    assert enc["duration"] == 5, enc
+    assert enc["video_refs"]["video1"]["type"] == "video"
+    assert enc["reference_images"]["image1"]["type"] == "image"
+    assert enc["reference_images"]["image2"]["type"] == "image"
+    print("ok: DC-wrapping-AG encoded leaves at every level")
+
+    # ---- 4. AUDIO dict is encoded, NOT recursed into
+    out = await proxy_node._encode_inputs({"audio": audio})
+    assert out["audio"] == {"type": "audio", "encoding": "mp3_base64", "data": "aud"}
+    print("ok: AUDIO dict still routes to encode_audio_input")
+
+    # ---- 5. Already-encoded envelope passes through (no double-encode)
+    pre_encoded = {"type": "image", "encoding": "png_base64", "data": "preimg"}
+    out = await proxy_node._encode_inputs({"images": pre_encoded})
+    assert out["images"] == pre_encoded, out
+    print("ok: pre-encoded envelope passed through unchanged")
+
+    # ---- 6. Plain scalar config dict passes through unchanged
+    out = await proxy_node._encode_inputs(
+        {"config": {"width": 1024, "height": 576, "name": "x"}},
+    )
+    assert out["config"] == {"width": 1024, "height": 576, "name": "x"}, out
+    print("ok: scalar config dict passed through unchanged")
+
+    # ---- 7. _externalize_nested walks per-slot envelopes
+    serialization._upload_log.clear()
+    big_img = torch.zeros((1, 64, 64, 3), dtype=torch.float32)
+    # encode_image_tensor stub emits ``data="img:1"`` (5 chars). Cap at 3
+    # to force every slot to externalize.
+    out = await proxy_node._encode_inputs(
+        {"reference_images": {"image1": big_img, "image2": big_img}},
+        server_url="https://fake.server",
+        max_inline_bytes=3,
+    )
+    refs = out["reference_images"]
+    assert refs["image1"].get("uri", "").startswith("https://upload/image/"), refs
+    assert refs["image2"].get("uri", "").startswith("https://upload/image/"), refs
+    assert "data" not in refs["image1"] and "data" not in refs["image2"]
+    assert len(serialization._upload_log) == 2, serialization._upload_log
+    print("ok: nested AG envelopes externalized per slot")
+
+    # ---- 7b. _externalize_nested leaves non-envelope siblings alone
+    serialization._upload_log.clear()
+    out = await proxy_node._encode_inputs(
+        {"model": {
+            "model": "wan2.7-r2v",
+            "duration": 5,
+            "reference_images": {"image1": big_img},
+        }},
+        server_url="https://fake.server",
+        max_inline_bytes=3,
+    )
+    enc = out["model"]
+    assert enc["model"] == "wan2.7-r2v" and enc["duration"] == 5, enc
+    assert enc["reference_images"]["image1"]["uri"].startswith("https://upload/image/")
+    print("ok: scalar siblings preserved alongside externalized envelopes")
+
+    # ---- 8. JSON-serializability post-condition
+    payload = await proxy_node._encode_inputs({
+        "reference_images": {"image1": img1, "image2": img2},
+        "video_refs": {"video1": vid1},
+        "audio": audio,
+        "config": {"width": 1024},
+        "scalar": "hello",
+        "n": 42,
+    })
+    json.dumps(payload)  # would raise TypeError pre-fix
+    print("ok: full encoded payload is JSON-serializable")
+
+    # ---- 9. Leaf-path regression: single-tensor / single-video / single-audio
+    out = await proxy_node._encode_inputs({"image": img1})
+    assert out["image"]["type"] == "image", out
+    out = await proxy_node._encode_inputs({"video": vid1})
+    assert out["video"]["type"] == "video", out
+    out = await proxy_node._encode_inputs({"audio": audio})
+    assert out["audio"]["type"] == "audio", out
+    out = await proxy_node._encode_inputs({"prompt": "hi", "seed": 7})
+    assert out == {"prompt": "hi", "seed": 7}, out
+    print("ok: top-level leaf path unchanged")
+
+    print("ALL OK")
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(asyncio.run(main()))

notes/run_image_max_batch.py

@@ -0,0 +1,441 @@
+"""Smoke test for the descriptor-driven ``local_validate.image_max_batch``
+client-side fail-fast.
+
+Exercises the proxy_node's ``_collect_local_validate`` /
+``_enforce_local_validate`` pair against a fake descriptor that mirrors
+the Flux2 wire shape. Asserts that:
+
+1. ``_collect_local_validate`` extracts ``{"images": {"image_max_batch": 2}}``
+   from the descriptor's IMAGE-input options dict.
+2. Encoding a 3-frame tensor for that input raises ``RnpProtocolError``
+   with ``code=INPUT_INVALID`` and ``user_facing=True`` *before* any
+   serialization / HTTP attempt.
+3. A 1-frame tensor passes through cleanly.
+4. A descriptor without ``local_validate`` is a no-op.
+5. ``cap=1`` surfaces the single-frame copy.
+6. A ``local_validate`` block missing the ``image_max_batch`` key is a no-op.
+
+Run with any python that has ``torch`` available:
+
+    python notes/run_image_max_batch.py
+
+The test loads the helper functions out of ``proxy_node.py`` via AST
+extraction so it doesn't pay the cost of importing ComfyUI's
+``comfy_api`` / ``comfy_api_nodes`` packages (which the helper
+functions don't touch). Standalone — no test runner / fixtures needed.
+"""
+from __future__ import annotations
+
+import ast
+import asyncio
+import importlib.util
+import os
+import sys
+import types
+
+import torch
+
+# ---------------------------------------------------------------------------
+# Locate the worktree under test (this file's parent's parent).
+# ---------------------------------------------------------------------------
+
+_HERE = os.path.dirname(os.path.abspath(__file__))
+_CLIENT_DIR = os.path.dirname(_HERE)
+
+
+def _load_protocol() -> types.ModuleType:
+    """Load ``protocol.py`` standalone (no relative imports)."""
+    path = os.path.join(_CLIENT_DIR, "protocol.py")
+    spec = importlib.util.spec_from_file_location("rnp_smoke_protocol", path)
+    mod = importlib.util.module_from_spec(spec)
+    sys.modules[spec.name] = mod
+    spec.loader.exec_module(mod)
+    return mod
+
+
+def _build_serialization_stub(protocol: types.ModuleType) -> types.ModuleType:
+    """Minimal ``serialization`` shim covering exactly the surface
+    ``_encode_one`` / ``_enforce_local_validate`` / ``_encode_inputs`` use:
+
+    * ``_is_torch_tensor(value)`` — duck-type check for torch tensors.
+    * ``is_audio_input(value)`` — always False (no audio in this test).
+    * ``encode_image_tensor(tensor, *, accepts_batch)`` — fake envelope.
+    * ``encode_mask_tensor(tensor)`` — fake envelope.
+    * ``is_envelope(value)`` — False so ``maybe_externalize`` is skipped.
+    """
+    mod = types.ModuleType("rnp_smoke_serialization")
+
+    def _is_torch_tensor(value):  # noqa: ANN001
+        try:
+            import torch as _torch
+        except ImportError:
+            return False
+        return isinstance(value, _torch.Tensor)
+
+    def is_audio_input(value):  # noqa: ANN001
+        return isinstance(value, dict) and {"waveform", "sample_rate"} <= set(value.keys())
+
+    def encode_image_tensor(tensor, *, accepts_batch=False):  # noqa: ANN001, ARG001
+        return {"type": "image", "encoding": "png_base64", "frames": int(tensor.shape[0])}
+
+    def encode_mask_tensor(tensor):  # noqa: ANN001, ARG001
+        return {"type": "mask", "encoding": "png_base64"}
+
+    def encode_audio_input(value):  # noqa: ANN001, ARG001
+        return {"type": "audio", "encoding": "mp3_base64"}
+
+    def is_envelope(value):  # noqa: ANN001
+        return False  # smoke test never externalizes
+
+    mod._is_torch_tensor = _is_torch_tensor
+    mod.is_audio_input = is_audio_input
+    mod.encode_image_tensor = encode_image_tensor
+    mod.encode_mask_tensor = encode_mask_tensor
+    mod.encode_audio_input = encode_audio_input
+    mod.is_envelope = is_envelope
+
+    async def maybe_externalize(*args, **kwargs):  # pragma: no cover — unused
+        return args[0] if args else None
+    mod.maybe_externalize = maybe_externalize
+    return mod
+
+
+def _build_proxy_node_stub(
+    protocol: types.ModuleType, serialization: types.ModuleType,
+) -> types.ModuleType:
+    """Extract the four helper functions under test from ``proxy_node.py``
+    and exec them in a minimal namespace so we don't drag the rest of
+    the module's heavy ComfyUI imports into the test process.
+    """
+    mod = types.ModuleType("rnp_smoke_proxy_node")
+    mod.__dict__.update({
+        "Any": object,
+        "log": types.SimpleNamespace(warning=lambda *a, **kw: None),
+        "serialization": serialization,
+        "RnpProtocolError": protocol.RnpProtocolError,
+        "ErrorCode": protocol.ErrorCode,
+    })
+    src_path = os.path.join(_CLIENT_DIR, "proxy_node.py")
+    with open(src_path, "r", encoding="utf-8") as fh:
+        src = fh.read()
+    tree = ast.parse(src)
+    wanted = {
+        "_collect_local_validate",
+        "_rules_from_input_spec",
+        "_enforce_local_validate",
+        "_check_image_max_batch",
+        "_encode_one",
+        "_encode_inputs",
+    }
+    nodes = [
+        n for n in tree.body
+        if isinstance(n, (ast.AsyncFunctionDef, ast.FunctionDef)) and n.name in wanted
+    ]
+    snippet = ast.Module(body=nodes, type_ignores=[])
+    code = compile(snippet, src_path, "exec")
+    exec(code, mod.__dict__)
+    missing = wanted - set(mod.__dict__)
+    if missing:
+        raise RuntimeError(f"proxy_node helpers missing from extract: {missing}")
+    return mod
+
+
+# ---------------------------------------------------------------------------
+# Test fixtures
+# ---------------------------------------------------------------------------
+
+def _flux2_like_descriptor(cap: int) -> dict:
+    """Mirrors the Flux2 ``images`` input descriptor surface."""
+    return {
+        "name": "FakeFlux2_RNP",
+        "input": {
+            "required": {
+                "prompt": ["STRING", {"multiline": True, "default": ""}],
+            },
+            "optional": {
+                "images": ["IMAGE", {
+                    "tooltip": "Up to N images to be used as references.",
+                    "local_validate": {"image_max_batch": cap},
+                }],
+            },
+            "hidden": {},
+        },
+    }
+
+
+def _no_validate_descriptor() -> dict:
+    return {
+        "name": "FakeNoValidate_RNP",
+        "input": {
+            "required": {},
+            "optional": {
+                "images": ["IMAGE", {"tooltip": "no cap"}],
+            },
+            "hidden": {},
+        },
+    }
+
+
+def _autogrow_descriptor(cap: int, prefix: str = "reference") -> dict:
+    """Mirrors GrokVideoReferenceNode_RNP's top-level Autogrow input
+    after server descriptor unwraps the nested DynamicCombo branch
+    (kept simpler here for direct AUTOGROW coverage)."""
+    return {
+        "name": "FakeAutogrow_RNP",
+        "input": {
+            "required": {
+                "reference_images": ["AUTOGROW", {
+                    "template": ["IMAGE", {
+                        "local_validate": {"image_max_batch": cap},
+                    }],
+                    "prefix": prefix,
+                    "min": 1,
+                    "max": cap,
+                }],
+            },
+            "optional": {},
+            "hidden": {},
+        },
+    }
+
+
+def _dyncombo_with_autogrow_descriptor() -> dict:
+    """Mirrors GrokImageEditNodeV2_RNP's ``model`` DynamicCombo with
+    two branches each carrying an Autogrow ``images`` template — one
+    capped at 3, the other at 1."""
+    def branch(key: str, cap: int) -> dict:
+        return {
+            "key": key,
+            "inputs": [
+                ["images", "AUTOGROW", {
+                    "template": ["IMAGE", {
+                        "local_validate": {"image_max_batch": cap},
+                    }],
+                    "prefix": "image",
+                    "min": 1,
+                    "max": cap,
+                }],
+                ["resolution", "COMBO", {"options": ["1K", "2K"]}],
+            ],
+        }
+    return {
+        "name": "FakeDynComboAutogrow_RNP",
+        "input": {
+            "required": {
+                "model": ["DYNAMIC_COMBO", {
+                    "options": [
+                        branch("grok-imagine-image-quality", 3),
+                        branch("grok-imagine-image-pro", 1),
+                    ],
+                }],
+            },
+            "optional": {},
+            "hidden": {},
+        },
+    }
+
+
+# ---------------------------------------------------------------------------
+# Tests
+# ---------------------------------------------------------------------------
+
+async def main() -> int:
+    protocol = _load_protocol()
+    serialization = _build_serialization_stub(protocol)
+    proxy_node = _build_proxy_node_stub(protocol, serialization)
+
+    # ---- 1. _collect_local_validate extracts the rule
+    desc = _flux2_like_descriptor(cap=2)
+    rules = proxy_node._collect_local_validate(desc)
+    assert rules == {"images": {"image_max_batch": 2}}, rules
+    print("ok: _collect_local_validate extracted", rules)
+
+    # ---- 2. over-cap tensor raises INPUT_INVALID before encoding
+    over_cap = torch.zeros((3, 64, 64, 3), dtype=torch.float32)
+    raised = False
+    try:
+        await proxy_node._encode_inputs(
+            {"images": over_cap},
+            input_serialization={"images": ["png_base64", "png_base64_batch"]},
+            local_validate=rules,
+            node_id="FakeFlux2_RNP",
+        )
+    except protocol.RnpProtocolError as e:
+        raised = True
+        assert e.code == protocol.ErrorCode.INPUT_INVALID, e.code
+        assert e.user_facing is True, e.user_facing
+        msg = str(e)
+        assert "FakeFlux2_RNP" in msg, msg
+        assert "3" in msg and "2" in msg, msg
+        print("ok: over-cap raised INPUT_INVALID:", msg)
+    assert raised, "over-cap tensor did not raise"
+
+    # ---- 3. at-cap tensor passes through (no exception)
+    at_cap = torch.zeros((2, 64, 64, 3), dtype=torch.float32)
+    out = await proxy_node._encode_inputs(
+        {"images": at_cap},
+        input_serialization={"images": ["png_base64", "png_base64_batch"]},
+        local_validate=rules,
+        node_id="FakeFlux2_RNP",
+    )
+    assert "images" in out
+    print("ok: at-cap (B=2) encoded without error")
+
+    # ---- 4. cap=1 / multi-frame raises with the single-frame copy
+    desc1 = _flux2_like_descriptor(cap=1)
+    rules1 = proxy_node._collect_local_validate(desc1)
+    raised = False
+    try:
+        await proxy_node._encode_inputs(
+            {"images": over_cap},
+            local_validate=rules1,
+            node_id="FakeI2V_RNP",
+        )
+    except protocol.RnpProtocolError as e:
+        raised = True
+        assert e.code == protocol.ErrorCode.INPUT_INVALID
+        assert "single image" in str(e), str(e)
+        print("ok: cap=1 single-frame copy fired:", str(e))
+    assert raised
+
+    # ---- 5. descriptor without local_validate is a no-op
+    desc_n = _no_validate_descriptor()
+    rules_n = proxy_node._collect_local_validate(desc_n)
+    assert rules_n == {}, rules_n
+    out = await proxy_node._encode_inputs(
+        {"images": over_cap},
+        input_serialization={"images": ["png_base64", "png_base64_batch"]},
+        local_validate=rules_n,
+        node_id="FakeNoValidate_RNP",
+    )
+    assert "images" in out
+    print("ok: missing local_validate is a no-op")
+
+    # ---- 6. local_validate block without image_max_batch sub-key is a no-op
+    out = await proxy_node._encode_inputs(
+        {"images": over_cap},
+        input_serialization={"images": ["png_base64", "png_base64_batch"]},
+        local_validate={"images": {"some_other_rule": 7}},
+        node_id="FakeOther_RNP",
+    )
+    assert "images" in out
+    print("ok: missing image_max_batch sub-key is a no-op")
+
+    # ---- 7. AUTOGROW: each slot value enforced against the template cap
+    desc_ag = _autogrow_descriptor(cap=2)
+    rules_ag = proxy_node._collect_local_validate(desc_ag)
+    assert "reference_images" in rules_ag, rules_ag
+    assert rules_ag["reference_images"].get("__template__") == {"image_max_batch": 2}, rules_ag
+    print("ok: AUTOGROW collected __template__ rules:", rules_ag)
+
+    one_frame = torch.zeros((1, 64, 64, 3), dtype=torch.float32)
+    raised = False
+    try:
+        # Slot 1 is over-cap (B=3 > cap=2) → should fire on that slot
+        await proxy_node._encode_inputs(
+            {"reference_images": {"reference0": one_frame, "reference1": over_cap}},
+            local_validate=rules_ag,
+            node_id="FakeAutogrow_RNP",
+        )
+    except protocol.RnpProtocolError as e:
+        raised = True
+        assert e.code == protocol.ErrorCode.INPUT_INVALID
+        msg = str(e)
+        # Slot path label should mention reference_images.<slot_key>
+        assert "reference_images" in msg, msg
+        assert "reference1" in msg, msg
+        print("ok: AUTOGROW slot over-cap fired:", msg)
+    assert raised, "AUTOGROW slot over-cap did not raise"
+
+    # All-under-cap autogrow passes (no exception)
+    out = await proxy_node._encode_inputs(
+        {"reference_images": {"reference0": one_frame, "reference1": one_frame, "reference2": None}},
+        local_validate=rules_ag,
+        node_id="FakeAutogrow_RNP",
+    )
+    assert "reference_images" in out
+    print("ok: AUTOGROW all-under-cap (None slots tolerated) passed")
+
+    # ---- 8. DYNAMIC_COMBO with nested AUTOGROW: branch-aware enforcement
+    desc_dc = _dyncombo_with_autogrow_descriptor()
+    rules_dc = proxy_node._collect_local_validate(desc_dc)
+    assert "model" in rules_dc, rules_dc
+    branches = rules_dc["model"].get("__branches__")
+    assert isinstance(branches, dict), rules_dc
+    assert "grok-imagine-image-quality" in branches
+    assert "grok-imagine-image-pro" in branches
+    assert branches["grok-imagine-image-quality"]["images"]["__template__"] == {"image_max_batch": 3}
+    assert branches["grok-imagine-image-pro"]["images"]["__template__"] == {"image_max_batch": 1}
+    print("ok: DYNAMIC_COMBO collected branch-keyed templates")
+
+    # Branch A (cap=3): slot with B=4 should raise
+    over_4 = torch.zeros((4, 64, 64, 3), dtype=torch.float32)
+    raised = False
+    try:
+        await proxy_node._encode_inputs(
+            {"model": {
+                "model": "grok-imagine-image-quality",
+                "resolution": "1K",
+                "images": {"image0": over_4},
+            }},
+            local_validate=rules_dc,
+            node_id="FakeDynComboAutogrow_RNP",
+        )
+    except protocol.RnpProtocolError as e:
+        raised = True
+        assert e.code == protocol.ErrorCode.INPUT_INVALID
+        msg = str(e)
+        assert "image0" in msg and "3" in msg and "4" in msg, msg
+        print("ok: DYNAMIC_COMBO branch-A (cap=3) fired on B=4 slot:", msg)
+    assert raised
+
+    # Branch A (cap=3): all slots within cap → passes
+    out = await proxy_node._encode_inputs(
+        {"model": {
+            "model": "grok-imagine-image-quality",
+            "resolution": "2K",
+            "images": {"image0": one_frame, "image1": one_frame, "image2": one_frame},
+        }},
+        local_validate=rules_dc,
+        node_id="FakeDynComboAutogrow_RNP",
+    )
+    assert "model" in out
+    print("ok: DYNAMIC_COMBO branch-A all-under-cap passed")
+
+    # Branch B (cap=1): a B=3 slot fires the single-frame copy
+    raised = False
+    try:
+        await proxy_node._encode_inputs(
+            {"model": {
+                "model": "grok-imagine-image-pro",
+                "resolution": "2K",
+                "images": {"image0": over_cap},
+            }},
+            local_validate=rules_dc,
+            node_id="FakeDynComboAutogrow_RNP",
+        )
+    except protocol.RnpProtocolError as e:
+        raised = True
+        assert e.code == protocol.ErrorCode.INPUT_INVALID
+        assert "single image" in str(e), str(e)
+        print("ok: DYNAMIC_COMBO branch-B (cap=1) fired single-frame copy:", str(e))
+    assert raised
+
+    # Unknown branch key → no enforcement (defensive)
+    out = await proxy_node._encode_inputs(
+        {"model": {
+            "model": "unknown-branch-key",
+            "images": {"image0": over_cap},
+        }},
+        local_validate=rules_dc,
+        node_id="FakeDynComboAutogrow_RNP",
+    )
+    assert "model" in out
+    print("ok: DYNAMIC_COMBO unknown branch key is a defensive no-op")
+
+    print("PASS: image_max_batch local_validate smoke test (incl. nested traversal)")
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(asyncio.run(main()))

notes/run_model3d_bundle_decoder.py

@@ -0,0 +1,461 @@
+"""Smoke test for the client-side multi-file MODEL_3D bundle decoder.
+
+Verifies:
+
+1. ``Capability.MODEL_3D_BUNDLE_INLINE`` is intact at the vendored
+   protocol layer and exposed as the well-formed
+   ``model_3d:bundle_inline`` token.
+2. ``is_envelope`` recognises a ``bundle_inline`` envelope (still
+   ``type=model_3d`` — HEAVY_TYPES unchanged).
+3. ``decode_model3d_envelope`` dispatches ``encoding=bundle_inline``
+   to the new ``BundledFile3D`` class, returns a real ``File3D``
+   subclass, and round-trips the primary mesh bytes via
+   ``get_data()`` / ``get_bytes()`` without materialising anything
+   on disk.
+4. ``get_source()`` materialises the whole bundle into a temp
+   directory using the producer-supplied relative paths, so a .obj
+   that references its .mtl by filename (and a texture in
+   ``textures/``) finds them on disk via the filesystem.
+5. ``save_to(dest)`` copies the whole bundle into the destination
+   directory — the primary file under the caller-specified name,
+   sidecars under their ORIGINAL relative paths — so cross-file refs
+   inside the .obj (which point at ``model.mtl``) still resolve when
+   the user saves the model somewhere stable.
+6. Bundle-specific accessors (``primary_path``, ``file_paths``,
+   ``file_bytes``, ``file_role``, ``paths_with_role``) return the
+   expected values; unknown roles round-trip without error.
+7. Malformed bundle envelopes (missing primary_path, missing
+   files, missing data, primary_path not in files) raise
+   ``RnpProtocolError`` with ``code=INTERNAL`` — no opaque
+   fallthrough.
+8. End-to-end against the server's ``make_model3d_bundle_envelope``
+   helper (imported from the comfy-rnp-server worktree) — the bytes
+   the server packs into each file are exactly the bytes the
+   client unpacks, and the path sanitization happens server-side so
+   the client never sees a malicious path.
+9. ``CLIENT_CAPABILITIES`` advertises ``model_3d:bundle_inline``
+   so server-side capability-gated descriptors negotiate cleanly.
+
+Run with the ComfyUI venv (needs torch / comfy_api on PYTHONPATH):
+
+    python notes/run_model3d_bundle_decoder.py
+"""
+from __future__ import annotations
+
+import base64
+import importlib.util
+import os
+import shutil
+import sys
+import tempfile
+import types
+
+# ---------------------------------------------------------------------------
+# Locate the worktree under test.
+# ---------------------------------------------------------------------------
+_HERE = os.path.dirname(os.path.abspath(__file__))
+_CLIENT_DIR = os.path.dirname(_HERE)
+_COMFYUI_DIR = os.path.dirname(os.path.dirname(_CLIENT_DIR))
+_SERVER_DIR = os.path.normpath(
+    os.path.join(_COMFYUI_DIR, "..", "comfy-rnp-server")
+)
+
+
+def _load_module(name: str, path: str) -> types.ModuleType:
+    spec = importlib.util.spec_from_file_location(name, path)
+    mod = importlib.util.module_from_spec(spec)
+    sys.modules[name] = mod
+    spec.loader.exec_module(mod)
+    return mod
+
+
+# Make ComfyUI importable so ``comfy_api.latest._util.geometry_types.File3D``
+# resolves at first BundledFile3D construction.
+if _COMFYUI_DIR not in sys.path:
+    sys.path.insert(0, _COMFYUI_DIR)
+if _SERVER_DIR not in sys.path:
+    sys.path.insert(0, _SERVER_DIR)
+
+# ComfyUI's full ``comfy_api`` tree pulls torch / PIL / comfy_execution at
+# import time. Stub just what BundledFile3D needs — the leaf
+# ``comfy_api.latest._util.geometry_types.File3D`` class with a faithful
+# subclass-friendly implementation.
+class _StubFile3D:
+    """Faithful subset of upstream ``comfy_api.latest._util.geometry_types.File3D``.
+
+    Implements the same parent-class shape (``_source`` / ``_format``
+    attrs + ``get_source`` / ``get_data`` / ``get_bytes`` / ``save_to``
+    methods) so subclassing it produces an object indistinguishable
+    from the real thing for the decoder's purposes. The smoke test
+    deliberately stubs torch + the heavy ``comfy_api.latest._util``
+    package init to keep the test runnable in the rnp venv.
+    """
+    def __init__(self, source, file_format: str = "") -> None:
+        self._source = source
+        self._format = (file_format or "").lstrip(".").lower()
+
+    @property
+    def format(self) -> str:
+        return self._format
+
+    @format.setter
+    def format(self, value: str) -> None:
+        self._format = (value or "").lstrip(".").lower()
+
+
+# Stub ``comfy_api.latest._util`` and the nested ``geometry_types`` module
+# so ``from comfy_api.latest._util.geometry_types import File3D`` works
+# without pulling torch / PIL / etc.
+_comfy_api = types.ModuleType("comfy_api")
+_comfy_api.__path__ = []  # type: ignore[attr-defined]
+_comfy_api_latest = types.ModuleType("comfy_api.latest")
+_comfy_api_latest.__path__ = []  # type: ignore[attr-defined]
+_comfy_api_latest_util = types.ModuleType("comfy_api.latest._util")
+_comfy_api_latest_util.__path__ = []  # type: ignore[attr-defined]
+_comfy_api_latest_util.File3D = _StubFile3D
+_comfy_api_latest_util_geom = types.ModuleType("comfy_api.latest._util.geometry_types")
+_comfy_api_latest_util_geom.File3D = _StubFile3D
+sys.modules.setdefault("comfy_api", _comfy_api)
+sys.modules.setdefault("comfy_api.latest", _comfy_api_latest)
+sys.modules.setdefault("comfy_api.latest._util", _comfy_api_latest_util)
+sys.modules.setdefault("comfy_api.latest._util.geometry_types", _comfy_api_latest_util_geom)
+
+# Emulate the client package layout for the relative imports.
+pkg = types.ModuleType("comfy_remote_nodes_test")
+pkg.__path__ = [_CLIENT_DIR]
+sys.modules["comfy_remote_nodes_test"] = pkg
+
+protocol = _load_module(
+    "comfy_remote_nodes_test.protocol",
+    os.path.join(_CLIENT_DIR, "protocol.py"),
+)
+# serialization imports ``client`` for ``rnp_client``; the bundle decoder
+# doesn't use it, so a tiny stub suffices.
+sys.modules["comfy_remote_nodes_test.client"] = types.ModuleType(
+    "comfy_remote_nodes_test.client",
+)
+serialization = _load_module(
+    "comfy_remote_nodes_test.serialization",
+    os.path.join(_CLIENT_DIR, "serialization.py"),
+)
+
+
+# ---------------------------------------------------------------------------
+# Build a minimal Wavefront-OBJ + MTL + PNG-stub bundle for the round-trip.
+# These bytes don't need to be loadable by an actual OBJ parser — the
+# decoder is bytes-in / bytes-out — but using realistic content makes
+# debugging easier if a future regression breaks the cross-file refs.
+# ---------------------------------------------------------------------------
+def _make_minimal_obj() -> bytes:
+    return (
+        b"# Minimal OBJ for RNP bundle smoke test.\n"
+        b"mtllib model.mtl\n"
+        b"v 0.0 0.0 0.0\nv 1.0 0.0 0.0\nv 0.0 1.0 0.0\n"
+        b"usemtl test_material\n"
+        b"f 1 2 3\n"
+    )
+
+
+def _make_minimal_mtl() -> bytes:
+    return (
+        b"# Minimal MTL for RNP bundle smoke test.\n"
+        b"newmtl test_material\n"
+        b"Ka 1.0 1.0 1.0\n"
+        b"Kd 1.0 1.0 1.0\n"
+        b"map_Kd textures/albedo.png\n"
+        b"map_bump textures/normal.png\n"
+    )
+
+
+def _make_stub_png(tag: bytes) -> bytes:
+    # Not a real PNG — the decoder doesn't parse it. Tag distinguishes
+    # each "texture" so the round-trip assertions can pin which bytes
+    # went where.
+    return b"\x89PNG\r\n\x1a\n" + tag
+
+
+obj_bytes = _make_minimal_obj()
+mtl_bytes = _make_minimal_mtl()
+albedo_bytes = _make_stub_png(b"ALBEDO_STUB")
+normal_bytes = _make_stub_png(b"NORMAL_STUB")
+roughness_bytes = _make_stub_png(b"ROUGHNESS_STUB")
+
+
+# ---------------------------------------------------------------------------
+# 1. Capability constant intact at the vendored protocol layer.
+# ---------------------------------------------------------------------------
+assert protocol.Capability.MODEL_3D_BUNDLE_INLINE == "model_3d:bundle_inline", (
+    f"Capability.MODEL_3D_BUNDLE_INLINE wrong value: "
+    f"{protocol.Capability.MODEL_3D_BUNDLE_INLINE!r}"
+)
+print("PASS: Capability.MODEL_3D_BUNDLE_INLINE = 'model_3d:bundle_inline'.")
+
+
+# ---------------------------------------------------------------------------
+# Build a hand-crafted bundle envelope for the next several checks.
+# ---------------------------------------------------------------------------
+hand_envelope = {
+    "type":         "model_3d",
+    "encoding":     "bundle_inline",
+    "format":       "obj",
+    "primary_path": "model.obj",
+    "files": [
+        {"path": "model.obj",            "format": "obj", "role": "mesh",
+         "data": base64.b64encode(obj_bytes).decode("ascii"),
+         "byte_size": len(obj_bytes)},
+        {"path": "model.mtl",            "format": "mtl", "role": "material",
+         "data": base64.b64encode(mtl_bytes).decode("ascii")},
+        {"path": "textures/albedo.png",  "format": "png", "role": "texture_diffuse",
+         "data": base64.b64encode(albedo_bytes).decode("ascii")},
+        {"path": "textures/normal.png",  "format": "png", "role": "texture_normal",
+         "data": base64.b64encode(normal_bytes).decode("ascii")},
+        {"path": "textures/roughness.png","format": "png","role": "texture_roughness",
+         "data": base64.b64encode(roughness_bytes).decode("ascii")},
+        # Unknown role — must round-trip without error.
+        {"path": "extras/notes.txt",     "role": "producer_notes",
+         "data": base64.b64encode(b"hello").decode("ascii")},
+    ],
+}
+
+
+# ---------------------------------------------------------------------------
+# 2. is_envelope recognises a bundle envelope.
+# ---------------------------------------------------------------------------
+assert protocol.is_envelope(hand_envelope), (
+    "is_envelope returned False for a well-formed bundle_inline envelope; "
+    "HEAVY_TYPES update somehow regressed."
+)
+print("PASS: is_envelope recognises bundle_inline envelopes.")
+
+
+# ---------------------------------------------------------------------------
+# 3. decode_model3d_envelope dispatches + round-trips primary bytes.
+# ---------------------------------------------------------------------------
+bundle = serialization.decode_model3d_envelope(hand_envelope)
+bundle_cls = serialization._bundled_file3d_class()
+assert isinstance(bundle, bundle_cls), (
+    f"decode_model3d_envelope should return BundledFile3D, got "
+    f"{type(bundle).__name__}"
+)
+# It must ALSO be a File3D (parent class) — that's the whole point of
+# the subclass approach.
+assert isinstance(bundle, _StubFile3D), (
+    f"BundledFile3D must subclass File3D — got {type(bundle).__mro__}"
+)
+# Primary file bytes round-trip via get_data / get_bytes (no disk I/O).
+assert bundle.get_bytes() == obj_bytes, (
+    f"primary OBJ bytes lost: {len(bundle.get_bytes())} vs {len(obj_bytes)}"
+)
+src = bundle.get_data()
+src.seek(0)
+assert src.read() == obj_bytes, "primary OBJ get_data() didn't round-trip"
+assert bundle.format == "obj", f"primary format wrong: {bundle.format!r}"
+print("PASS: decode_model3d_envelope returns BundledFile3D + primary bytes round-trip.")
+
+
+# ---------------------------------------------------------------------------
+# 4. get_source() materialises the whole bundle on disk.
+# ---------------------------------------------------------------------------
+disk_path = bundle.get_source()
+assert os.path.isfile(disk_path), f"get_source() should return existing path: {disk_path}"
+assert disk_path.endswith("model.obj"), f"primary path wrong: {disk_path}"
+with open(disk_path, "rb") as f:
+    assert f.read() == obj_bytes, "disk-materialised primary OBJ bytes mismatched"
+# Sidecars must be at the SAME relative paths so the .obj's mtllib /
+# map_Kd refs resolve.
+temp_root = bundle.materialize_to_temp_dir()
+assert os.path.isfile(os.path.join(temp_root, "model.mtl")), (
+    "model.mtl missing from materialised bundle"
+)
+assert os.path.isfile(os.path.join(temp_root, "textures", "albedo.png")), (
+    "textures/albedo.png missing from materialised bundle"
+)
+assert os.path.isfile(os.path.join(temp_root, "textures", "normal.png"))
+assert os.path.isfile(os.path.join(temp_root, "textures", "roughness.png"))
+assert os.path.isfile(os.path.join(temp_root, "extras", "notes.txt"))
+# Sidecar bytes match.
+with open(os.path.join(temp_root, "model.mtl"), "rb") as f:
+    assert f.read() == mtl_bytes
+with open(os.path.join(temp_root, "textures", "albedo.png"), "rb") as f:
+    assert f.read() == albedo_bytes
+print("PASS: get_source() materialises whole bundle on disk at producer-supplied paths.")
+
+
+# ---------------------------------------------------------------------------
+# 5. save_to(dest) copies the whole bundle into the destination directory.
+# ---------------------------------------------------------------------------
+save_dir = tempfile.mkdtemp(prefix="rnp_save_to_test_")
+try:
+    out_path = bundle.save_to(os.path.join(save_dir, "renamed_scene.obj"))
+    assert os.path.isfile(out_path)
+    with open(out_path, "rb") as f:
+        assert f.read() == obj_bytes, "primary OBJ bytes wrong after save_to"
+    # Sidecars keep ORIGINAL filenames so cross-file refs inside the
+    # OBJ (which point at "model.mtl") still resolve relative to the
+    # renamed primary file's directory.
+    assert os.path.isfile(os.path.join(save_dir, "model.mtl")), (
+        "save_to should copy model.mtl alongside the renamed primary"
+    )
+    assert os.path.isfile(os.path.join(save_dir, "textures", "albedo.png"))
+    assert os.path.isfile(os.path.join(save_dir, "textures", "normal.png"))
+    with open(os.path.join(save_dir, "textures", "normal.png"), "rb") as f:
+        assert f.read() == normal_bytes
+    print("PASS: save_to copies whole bundle, primary renamed, sidecars preserve filenames.")
+finally:
+    shutil.rmtree(save_dir, ignore_errors=True)
+
+
+# ---------------------------------------------------------------------------
+# 6. Bundle-specific accessors.
+# ---------------------------------------------------------------------------
+assert bundle.primary_path == "model.obj"
+assert set(bundle.file_paths) == {
+    "model.obj", "model.mtl",
+    "textures/albedo.png", "textures/normal.png", "textures/roughness.png",
+    "extras/notes.txt",
+}
+assert bundle.file_bytes("textures/albedo.png") == albedo_bytes
+# Optional roles.
+assert bundle.file_role("model.obj") == "mesh"
+assert bundle.file_role("textures/normal.png") == "texture_normal"
+# Unknown roles round-trip — no enum check.
+assert bundle.file_role("extras/notes.txt") == "producer_notes"
+# Missing role (none supplied) returns None.
+assert bundle.file_role("nonexistent.png") is None
+# Convenience lookup.
+assert bundle.paths_with_role("texture_normal") == ["textures/normal.png"]
+assert bundle.paths_with_role("nonexistent_role") == []
+print("PASS: bundle accessors (primary_path / file_paths / file_role / paths_with_role).")
+
+
+# ---------------------------------------------------------------------------
+# 7. Malformed bundle envelopes raise RnpProtocolError with code=INTERNAL.
+# ---------------------------------------------------------------------------
+def _expect_protocol_error(bad_env: dict, needle: str) -> None:
+    try:
+        serialization.decode_model3d_envelope(bad_env)
+    except protocol.RnpProtocolError as e:
+        assert e.code == protocol.ErrorCode.INTERNAL, f"wrong error code: {e.code}"
+        assert needle in str(e), f"error msg doesn't mention {needle!r}: {e}"
+    else:
+        raise AssertionError(f"expected RnpProtocolError for {needle!r}")
+
+_expect_protocol_error(
+    {"type": "model_3d", "encoding": "bundle_inline", "files": [
+        {"path": "x.obj", "data": "AA=="},
+    ]},
+    "primary_path",
+)
+_expect_protocol_error(
+    {"type": "model_3d", "encoding": "bundle_inline", "primary_path": "x.obj"},
+    "files",
+)
+_expect_protocol_error(
+    {"type": "model_3d", "encoding": "bundle_inline", "primary_path": "x.obj",
+     "files": [{"path": "y.obj", "data": "AA=="}]},
+    "primary_path 'x.obj' not in",
+)
+_expect_protocol_error(
+    {"type": "model_3d", "encoding": "bundle_inline", "primary_path": "x.obj",
+     "files": [{"path": "x.obj"}]},  # missing data
+    "missing inline 'data'",
+)
+_expect_protocol_error(
+    {"type": "model_3d", "encoding": "wat_inline", "data": "AA=="},
+    "Unsupported model_3d encoding",
+)
+print("PASS: malformed bundle envelopes raise RnpProtocolError(INTERNAL).")
+
+
+# ---------------------------------------------------------------------------
+# 8. End-to-end: server make_model3d_bundle_envelope -> client decode.
+#    Skips when the server worktree isn't on disk.
+# ---------------------------------------------------------------------------
+server_protocol_path = os.path.join(
+    _SERVER_DIR, "comfy_rnp_protocol", "envelopes.py",
+)
+if os.path.exists(server_protocol_path):
+    server_pkg = types.ModuleType("comfy_rnp_protocol_e2e_bundle")
+    server_pkg.__path__ = [os.path.join(_SERVER_DIR, "comfy_rnp_protocol")]
+    sys.modules["comfy_rnp_protocol_e2e_bundle"] = server_pkg
+    _load_module(
+        "comfy_rnp_protocol_e2e_bundle.constants",
+        os.path.join(_SERVER_DIR, "comfy_rnp_protocol", "constants.py"),
+    )
+    server_envelopes = _load_module(
+        "comfy_rnp_protocol_e2e_bundle.envelopes",
+        os.path.join(_SERVER_DIR, "comfy_rnp_protocol", "envelopes.py"),
+    )
+    assert hasattr(server_envelopes, "make_model3d_bundle_envelope"), (
+        "Server envelopes module missing make_model3d_bundle_envelope — "
+        "server-side PR #38 should have added it."
+    )
+    server_env = server_envelopes.make_model3d_bundle_envelope(
+        files=[
+            {"path": "model.obj",            "format": "obj", "role": "mesh",
+             "data": base64.b64encode(obj_bytes).decode("ascii"),
+             "byte_size": len(obj_bytes)},
+            {"path": "model.mtl",            "format": "mtl", "role": "material",
+             "data": base64.b64encode(mtl_bytes).decode("ascii")},
+            {"path": "textures/albedo.png",  "format": "png", "role": "texture_diffuse",
+             "data": base64.b64encode(albedo_bytes).decode("ascii")},
+        ],
+        primary_path="model.obj",
+        format="obj",
+    )
+    assert server_env["type"] == "model_3d", server_env
+    assert server_env["encoding"] == "bundle_inline", server_env
+    assert server_env["primary_path"] == "model.obj"
+    assert len(server_env["files"]) == 3
+    # Path sanitization happens server-side. Confirm a malicious path
+    # would be rejected at envelope build time (defence-in-depth — the
+    # client decoder doesn't see this).
+    try:
+        server_envelopes.make_model3d_bundle_envelope(
+            files=[{"path": "../etc/passwd", "data": "AA=="}],
+            primary_path="../etc/passwd",
+        )
+    except ValueError as e:
+        assert "parent-dir" in str(e) or "absolute" in str(e), (
+            f"path sanitization mistake: {e!r}"
+        )
+    else:
+        raise AssertionError("server should reject '..' path at envelope build time")
+    # Pipe through the client decoder.
+    server_bundle = serialization.decode_model3d_envelope(server_env)
+    assert isinstance(server_bundle, bundle_cls), (
+        f"server-built envelope decoded to wrong type: {type(server_bundle)}"
+    )
+    assert server_bundle.get_bytes() == obj_bytes
+    assert server_bundle.file_bytes("textures/albedo.png") == albedo_bytes
+    assert server_bundle.file_role("model.obj") == "mesh"
+    print(
+        "PASS: end-to-end server make_model3d_bundle_envelope -> client "
+        "decode_model3d_envelope preserves all files + roles + bytes."
+    )
+else:
+    print(
+        f"SKIP end-to-end: server worktree not found at {_SERVER_DIR!r}"
+    )
+
+
+# ---------------------------------------------------------------------------
+# 9. CLIENT_CAPABILITIES advertises model_3d:bundle_inline.
+#    client.py imports aiohttp / comfy_api_nodes — same source-level
+#    contract check we used for the GLB cap in PR #1.
+# ---------------------------------------------------------------------------
+client_src = open(
+    os.path.join(_CLIENT_DIR, "client.py"), "r", encoding="utf-8",
+).read()
+assert "CLIENT_CAPABILITIES = [" in client_src, (
+    "client.py missing CLIENT_CAPABILITIES list"
+)
+cc_start = client_src.index("CLIENT_CAPABILITIES = [")
+cc_end = client_src.index("]", cc_start)
+cc_block = client_src[cc_start:cc_end]
+assert "Capability.MODEL_3D_BUNDLE_INLINE" in cc_block, (
+    "Capability.MODEL_3D_BUNDLE_INLINE not advertised in CLIENT_CAPABILITIES"
+)
+print("PASS: CLIENT_CAPABILITIES advertises Capability.MODEL_3D_BUNDLE_INLINE.")
+
+print("ALL CHECKS PASSED.")

notes/run_model3d_decoder.py

@@ -0,0 +1,316 @@
+"""Smoke test for the client-side MODEL_3D decoder.
+
+Verifies:
+
+1. The vendored ``protocol.HEAVY_TYPES`` now recognises ``model_3d`` so
+   ``is_envelope`` returns True for an inbound MODEL_3D envelope.
+2. ``serialization.decode_model3d_envelope`` round-trips an inline
+   ``glb_inline`` envelope into a ``comfy_api.latest._util.File3D``
+   whose ``BytesIO`` payload matches the original GLB bytes.
+3. ``decode_envelope`` dispatches ``type=model_3d`` to the new decoder.
+4. The proxy_node ``_OUTPUT_CLASSES`` now maps ``"MODEL_3D"`` to
+   ``IO.File3DGLB.Output`` so a descriptor with ``output=["MODEL_3D"]``
+   parses without falling through to the opaque IO.Custom branch.
+5. ``client.CLIENT_CAPABILITIES`` advertises
+   ``model_3d:glb_inline`` on every outbound request so descriptors
+   gated on the token negotiate cleanly.
+6. End-to-end against the server's ``make_model3d_envelope`` helper
+   (imported from the comfy-rnp-server worktree) — the bytes the
+   server packs are exactly the bytes the client decodes.
+
+Run with the ComfyUI venv (needs torch / comfy_api on PYTHONPATH):
+
+    python notes/run_model3d_decoder.py
+"""
+from __future__ import annotations
+
+import base64
+import importlib.util
+import io
+import json
+import os
+import sys
+import types
+
+# ---------------------------------------------------------------------------
+# Locate the worktree under test.
+# ---------------------------------------------------------------------------
+_HERE = os.path.dirname(os.path.abspath(__file__))
+_CLIENT_DIR = os.path.dirname(_HERE)
+_COMFYUI_DIR = os.path.dirname(os.path.dirname(_CLIENT_DIR))
+_SERVER_DIR = os.path.normpath(
+    os.path.join(_COMFYUI_DIR, "..", "comfy-rnp-server")
+)
+
+
+def _load_module(name: str, path: str) -> types.ModuleType:
+    spec = importlib.util.spec_from_file_location(name, path)
+    mod = importlib.util.module_from_spec(spec)
+    sys.modules[name] = mod
+    spec.loader.exec_module(mod)
+    return mod
+
+
+# Make ComfyUI importable so ``comfy_api.latest._util.File3D`` resolves.
+if _COMFYUI_DIR not in sys.path:
+    sys.path.insert(0, _COMFYUI_DIR)
+# And the server worktree so we can import the ENCODER side end-to-end.
+if _SERVER_DIR not in sys.path:
+    sys.path.insert(0, _SERVER_DIR)
+
+# ComfyUI's full ``comfy_api`` tree pulls torch / PIL / comfy /
+# comfy_execution at import time — way too heavy for a decoder
+# round-trip check. Stub just the leaf module our decoder imports
+# (``comfy_api.latest._util``) with a faithful pure-Python ``File3D``.
+class _StubFile3D:
+    def __init__(self, source, file_format: str = "") -> None:
+        self._source = source
+        self._format = file_format
+
+_comfy_api = types.ModuleType("comfy_api")
+_comfy_api_latest = types.ModuleType("comfy_api.latest")
+_comfy_api_latest_util = types.ModuleType("comfy_api.latest._util")
+_comfy_api_latest_util.File3D = _StubFile3D
+_comfy_api.latest = _comfy_api_latest
+_comfy_api_latest._util = _comfy_api_latest_util
+sys.modules.setdefault("comfy_api", _comfy_api)
+sys.modules.setdefault("comfy_api.latest", _comfy_api_latest)
+sys.modules.setdefault("comfy_api.latest._util", _comfy_api_latest_util)
+
+# Load the client modules as if they were a package — the ``protocol``
+# module imports from ``.protocol`` relative to its package, so we
+# emulate the package layout under the name ``comfy_remote_nodes``.
+pkg = types.ModuleType("comfy_remote_nodes_test")
+pkg.__path__ = [_CLIENT_DIR]
+sys.modules["comfy_remote_nodes_test"] = pkg
+
+protocol = _load_module(
+    "comfy_remote_nodes_test.protocol",
+    os.path.join(_CLIENT_DIR, "protocol.py"),
+)
+# serialization imports ``client`` for ``rnp_client`` but we don't need
+# the HTTP client for the decoder path; provide a tiny stub so the
+# import succeeds.
+client_stub = types.ModuleType("comfy_remote_nodes_test.client")
+sys.modules["comfy_remote_nodes_test.client"] = client_stub
+serialization = _load_module(
+    "comfy_remote_nodes_test.serialization",
+    os.path.join(_CLIENT_DIR, "serialization.py"),
+)
+
+
+# ---------------------------------------------------------------------------
+# Build a minimal valid binary glTF 2.0 (GLB) payload for the round-trip.
+# Format: 12-byte header + N chunks. Header: magic(4) + version(4) + length(4).
+# Magic must be b"glTF", version=2. We include exactly one JSON chunk that
+# describes an empty asset (enough to satisfy any GLB-magic check; downstream
+# nodes do their own validation).
+# ---------------------------------------------------------------------------
+def _make_minimal_glb() -> bytes:
+    json_payload = b'{"asset":{"version":"2.0"}}'
+    # Chunks must be 4-byte aligned; pad JSON with spaces (0x20).
+    pad = (4 - (len(json_payload) % 4)) % 4
+    json_payload = json_payload + (b" " * pad)
+    json_chunk = (
+        len(json_payload).to_bytes(4, "little")  # chunk length
+        + b"JSON"                                  # chunk type
+        + json_payload                             # chunk data
+    )
+    total = 12 + len(json_chunk)
+    header = (
+        b"glTF"
+        + (2).to_bytes(4, "little")
+        + total.to_bytes(4, "little")
+    )
+    return header + json_chunk
+
+
+# ---------------------------------------------------------------------------
+# 1. HEAVY_TYPES recognises model_3d.
+# ---------------------------------------------------------------------------
+assert "model_3d" in protocol.HEAVY_TYPES, (
+    f"HEAVY_TYPES missing 'model_3d': {sorted(protocol.HEAVY_TYPES)}"
+)
+assert protocol.Capability.MODEL_3D_GLB_INLINE == "model_3d:glb_inline", (
+    f"Capability.MODEL_3D_GLB_INLINE wrong value: "
+    f"{protocol.Capability.MODEL_3D_GLB_INLINE!r}"
+)
+print("PASS: HEAVY_TYPES includes 'model_3d'; Capability constant intact.")
+
+
+# ---------------------------------------------------------------------------
+# 2. is_envelope returns True for a MODEL_3D envelope.
+# ---------------------------------------------------------------------------
+glb_bytes = _make_minimal_glb()
+envelope = {
+    "type":     "model_3d",
+    "encoding": "glb_inline",
+    "data":     base64.b64encode(glb_bytes).decode("ascii"),
+    "format":   "glb",
+    "byte_size": len(glb_bytes),
+}
+assert protocol.is_envelope(envelope), (
+    "is_envelope returned False for a well-formed MODEL_3D envelope; "
+    "HEAVY_TYPES update did not take effect."
+)
+print("PASS: is_envelope recognises model_3d envelopes.")
+
+
+# ---------------------------------------------------------------------------
+# 3. decode_model3d_envelope round-trips inline bytes.
+# ---------------------------------------------------------------------------
+file3d = serialization.decode_model3d_envelope(envelope)
+from comfy_api.latest._util import File3D as _File3D_for_check  # noqa: E402
+assert isinstance(file3d, _File3D_for_check), (
+    f"Expected File3D, got {type(file3d).__name__}"
+)
+# File3D exposes get_data() / get_source(); the public API on this
+# version returns a BytesIO when constructed with one. Drain bytes:
+src = file3d._source  # internal, but stable across this code path
+assert hasattr(src, "read"), f"File3D source not BytesIO-like: {type(src)}"
+src.seek(0)
+round_trip = src.read()
+assert round_trip == glb_bytes, (
+    f"GLB byte mismatch after decode: input={len(glb_bytes)} bytes, "
+    f"output={len(round_trip)} bytes"
+)
+assert round_trip[:4] == b"glTF", "GLB magic bytes lost in round-trip"
+print(f"PASS: decode_model3d_envelope round-trips {len(glb_bytes)} GLB bytes.")
+
+
+# ---------------------------------------------------------------------------
+# 4. decode_envelope dispatches model_3d to the new decoder.
+# ---------------------------------------------------------------------------
+file3d2 = serialization.decode_envelope(envelope)
+assert isinstance(file3d2, _File3D_for_check), (
+    f"decode_envelope dispatch failed for type=model_3d: got "
+    f"{type(file3d2).__name__}"
+)
+print("PASS: decode_envelope dispatches type='model_3d' correctly.")
+
+
+# ---------------------------------------------------------------------------
+# 5. Unsupported encoding raises a hard error (defence-in-depth so a
+#    future encoding never silently downgrades to garbage bytes).
+# ---------------------------------------------------------------------------
+bad = dict(envelope, encoding="obj_inline")
+try:
+    serialization.decode_model3d_envelope(bad)
+except protocol.RnpProtocolError as e:
+    assert "Unsupported model_3d encoding" in str(e), (
+        f"Wrong error message for unsupported encoding: {e!r}"
+    )
+    assert e.code == protocol.ErrorCode.INTERNAL
+    print("PASS: unsupported model_3d encoding raises RnpProtocolError.")
+else:
+    raise AssertionError(
+        "decode_model3d_envelope accepted an unsupported encoding"
+    )
+
+
+# ---------------------------------------------------------------------------
+# 6. End-to-end: server's make_model3d_envelope -> client decode.
+#    Skips when the server worktree isn't on disk (e.g. CI running
+#    against just the client repo).
+# ---------------------------------------------------------------------------
+server_protocol_path = os.path.join(
+    _SERVER_DIR, "comfy_rnp_protocol", "envelopes.py",
+)
+if os.path.exists(server_protocol_path):
+    # Import the server-side envelopes module standalone (it has no
+    # ComfyUI deps and uses its own sibling ``constants`` import).
+    server_pkg = types.ModuleType("comfy_rnp_protocol_e2e")
+    server_pkg.__path__ = [os.path.join(_SERVER_DIR, "comfy_rnp_protocol")]
+    sys.modules["comfy_rnp_protocol_e2e"] = server_pkg
+    server_constants = _load_module(
+        "comfy_rnp_protocol_e2e.constants",
+        os.path.join(_SERVER_DIR, "comfy_rnp_protocol", "constants.py"),
+    )
+    server_envelopes = _load_module(
+        "comfy_rnp_protocol_e2e.envelopes",
+        os.path.join(_SERVER_DIR, "comfy_rnp_protocol", "envelopes.py"),
+    )
+    # Build using the server helper exactly as a provider would.
+    server_env = server_envelopes.make_model3d_envelope(
+        encoding="glb_inline",
+        data=base64.b64encode(glb_bytes).decode("ascii"),
+        format="glb",
+        byte_size=len(glb_bytes),
+    )
+    assert server_env["type"] == "model_3d", server_env
+    assert server_env["encoding"] == "glb_inline", server_env
+    assert "model_3d" in server_envelopes.HEAVY_TYPES, (
+        "Server HEAVY_TYPES missing 'model_3d' — server PR #37 should "
+        "have added it; reinstall comfy_rnp_protocol if this fires."
+    )
+    # Pipe through the client decoder.
+    file3d3 = serialization.decode_envelope(server_env)
+    src = file3d3._source
+    src.seek(0)
+    assert src.read() == glb_bytes, "End-to-end byte mismatch"
+    print(
+        "PASS: end-to-end server make_model3d_envelope -> client "
+        "decode_envelope preserves bytes."
+    )
+else:
+    print(
+        f"SKIP end-to-end: server worktree not found at {_SERVER_DIR!r}"
+    )
+
+
+# ---------------------------------------------------------------------------
+# 7. proxy_node._OUTPUT_CLASSES maps "MODEL_3D" -> IO.File3DGLB.Output.
+#    proxy_node pulls in the full comfy_api / server / comfy_api_nodes
+#    tree at import time (too heavy to stub here), so we verify the
+#    mapping with a source-level check instead. The pattern is exact:
+#    a quoted ``"MODEL_3D":`` key whose value is the
+#    ``IO.File3DGLB.Output`` class reference inside the
+#    ``_OUTPUT_CLASSES`` dict literal. The same source check would
+#    catch a typo (e.g. ``IO.File3DAny.Output``) that a smoke test
+#    importing the live module would miss too — the dict literal is
+#    the contract.
+# ---------------------------------------------------------------------------
+proxy_src = open(
+    os.path.join(_CLIENT_DIR, "proxy_node.py"), "r", encoding="utf-8",
+).read()
+assert "_OUTPUT_CLASSES = {" in proxy_src, "proxy_node missing _OUTPUT_CLASSES dict"
+assert '"MODEL_3D":' in proxy_src and "IO.File3DGLB.Output" in proxy_src, (
+    "proxy_node._OUTPUT_CLASSES does not appear to map "
+    "'MODEL_3D' -> IO.File3DGLB.Output"
+)
+# Verify the mapping is in the same dict literal (not somewhere else).
+oc_start = proxy_src.index("_OUTPUT_CLASSES = {")
+oc_end = proxy_src.index("}", oc_start)
+oc_block = proxy_src[oc_start:oc_end]
+assert '"MODEL_3D":' in oc_block, (
+    "'MODEL_3D' key not inside the _OUTPUT_CLASSES dict literal"
+)
+assert "IO.File3DGLB.Output" in oc_block, (
+    "IO.File3DGLB.Output value not inside the _OUTPUT_CLASSES dict literal"
+)
+print(
+    "PASS: proxy_node._OUTPUT_CLASSES dict literal contains "
+    "'MODEL_3D' -> IO.File3DGLB.Output."
+)
+
+
+# ---------------------------------------------------------------------------
+# 8. CLIENT_CAPABILITIES advertises model_3d:glb_inline.
+#    client.py imports aiohttp / comfy_api_nodes too — same source-level
+#    contract check.
+# ---------------------------------------------------------------------------
+client_src = open(
+    os.path.join(_CLIENT_DIR, "client.py"), "r", encoding="utf-8",
+).read()
+assert "CLIENT_CAPABILITIES = [" in client_src, (
+    "client.py missing CLIENT_CAPABILITIES list"
+)
+cc_start = client_src.index("CLIENT_CAPABILITIES = [")
+cc_end = client_src.index("]", cc_start)
+cc_block = client_src[cc_start:cc_end]
+assert "Capability.MODEL_3D_GLB_INLINE" in cc_block, (
+    "Capability.MODEL_3D_GLB_INLINE not advertised in CLIENT_CAPABILITIES"
+)
+print("PASS: CLIENT_CAPABILITIES advertises Capability.MODEL_3D_GLB_INLINE.")
+
+print("ALL CHECKS PASSED.")

protocol.py

@@ -21,7 +21,18 @@ DEFAULT_POLL_INTERVAL_S = 1.5
 # Values mirror comfy_api_nodes.util.client.poll_op_raw so RNP-routed
 # nodes behave identically to a direct upstream call when the descriptor
 # leaves a field unset.
-DEFAULT_MAX_POLL_ATTEMPTS = 160
+#
+# Lifetime budget design: the wire format declares ``soft_timeout_s``
+# (server-side stuck threshold) and ``hard_timeout_s`` (server-side
+# hard kill) as durations. The client's poll-loop cap is a derived
+# value — ``max_poll_attempts = ceil((hard_timeout_s + DEFAULT_POLL_GRACE_S)
+# / poll_interval_s)`` — not a separate wire field, so changing the
+# poll cadence can never silently shrink or stretch the contract.
+# Industry precedent: OAuth2 device flow §3.5, Kubernetes
+# ``--timeout``, gRPC deadlines all express bounds as durations.
+DEFAULT_SOFT_TIMEOUT_S = 240.0
+DEFAULT_HARD_TIMEOUT_S = 300.0
+DEFAULT_POLL_GRACE_S = 30.0
 DEFAULT_TIMEOUT_PER_POLL_S = 120.0
 DEFAULT_MAX_RETRIES_PER_POLL = 10
 DEFAULT_RETRY_DELAY_PER_POLL_S = 1.0
@@ -93,14 +104,130 @@ class Capability:
     IMAGE_PNG_BASE64 = "image:png_base64"
     IMAGE_JPEG_BASE64 = "image:jpeg_base64"
     IMAGE_RAW_CHW_UINT8 = "image:raw_b64_chw_uint8"
+    # Multi-frame IMAGE envelope: a single value carries N PNG-base64
+    # frames sharing the same H/W/C, decoded into a [B,H,W,C] tensor on
+    # this client. Mirrors the AUDIO sweep precedent (parallel encodings
+    # rather than extending ``png_base64``) so legacy server emits still
+    # decode correctly when the new token isn't advertised.
+    IMAGE_PNG_BASE64_BATCH = "image:png_base64_batch"
+    # Externalised counterpart of ``image:png_base64_batch``: same
+    # multi-frame IMAGE envelope shape, but every frame's PNG bytes
+    # are uploaded out-of-band via the presigned-PUT path (see
+    # ``serialization.maybe_externalize``) and the envelope carries
+    # a parallel ``uris`` list instead of inline ``frames``. Required
+    # because the original ``maybe_externalize`` only inspected the
+    # singleton ``data`` field, so a batch of N PNGs whose summed
+    # base64 payload crossed the 8 MiB inline cap silently rode the
+    # wire as a 30+ MiB POST body. Per-frame URIs (NOT a tar/zip
+    # blob) keep this composable with the existing per-frame fetch
+    # helpers and let the server fetch in parallel. Decoder dispatch
+    # keys off the ``encoding`` string — ``frames`` vs ``uris`` is a
+    # payload field, not a shape selector.
+    IMAGE_PNG_BASE64_BATCH_URI = "image:png_base64_batch_uri"
     VIDEO_MP4_INLINE = "video:mp4_inline"
+    VIDEO_MP4_BASE64 = "video:mp4_base64"
     AUDIO_MP3_INLINE = "audio:mp3_inline"
+    AUDIO_MP3_BASE64 = "audio:mp3_base64"
+    AUDIO_WAV_BASE64 = "audio:wav_base64"
+    AUDIO_WAV_INLINE = "audio:wav_inline"
+    # OGG/Opus container as base64 in ``data``. Mirrors the parallel-
+    # encoding negotiation pattern already used for the AUDIO sweep
+    # (mp3 vs. wav) — servers only emit ``audio:opus_base64`` when
+    # this client advertises the token, otherwise they fall back to
+    # ``mp3_base64``. PyAV/FFmpeg decodes OGG/Opus natively so
+    # ``decode_audio_envelope`` ships the bytes through
+    # ``audio_bytes_to_audio_input`` the same way it ships MP3/WAV.
+    AUDIO_OPUS_BASE64 = "audio:opus_base64"
     MASK_PNG_BASE64 = "mask:png_base64"
+    SVG_XML_BASE64 = "svg:svg_xml_base64"
     SCHEMA_V3 = "schema:v3"
     ASYNC_EXECUTE = "execute:async"
+    # Pass-through custom-IO bucket: when an input/output type string
+    # is not in the standard primitive/media set, the proxy_node parser
+    # falls back to ``IO.Custom(<type>)`` so partner helper-config nodes
+    # (OpenAIInputFiles, OpenAIChatConfig, GeminiInputFiles, Recraft
+    # style/color/controls helpers, future partner helper-config types)
+    # round-trip without per-type encoder plumbing. The descriptor's
+    # original io_type string is preserved on the V3 input so node-to-
+    # node connections only chain when the strings match. Encoder
+    # dispatch passes opaque dict/list/string values through
+    # ``_encode_one`` unchanged — they're already JSON-serialisable.
+    IO_OPAQUE = "io:opaque"
+    # 3D MODEL envelope: a single-mesh GLB (binary glTF 2.0, magic
+    # b"glTF") wrapped opaquely in ``data`` (base64). Mirrors the
+    # parallel-encoding pattern already used for VIDEO_MP4_INLINE — the
+    # server emits this token only when the inbound request advertised
+    # it; legacy clients without the token will surface
+    # NEGOTIATION_FAILED at descriptor-load time. The decoder on the
+    # client side hands raw GLB bytes straight to ComfyUI's
+    # ``Types.File3D`` / ``IO.File3DGLB`` machinery (no in-server GLB
+    # parse — the magic-byte check at offset 0 is the only validity
+    # guard the server applies).
+    MODEL_3D_GLB_INLINE = "model_3d:glb_inline"
+    # 3D MODEL BUNDLE envelope: a primary mesh file plus N companion
+    # files (textures / .mtl / .bin / other PBR maps) carried inline as
+    # a single coherent unit. Wire shape:
+    #
+    #   {
+    #     "type":         "model_3d",
+    #     "encoding":     "bundle_inline",
+    #     "format":       "obj",
+    #     "primary_path": "model.obj",
+    #     "files": [
+    #       {"path": "model.obj",            "format": "obj", "role": "mesh",
+    #        "data": "<b64>", "byte_size": 12345},
+    #       {"path": "model.mtl",            "format": "mtl", "role": "material",
+    #        "data": "<b64>"},
+    #       {"path": "textures/albedo.png",  "format": "png", "role": "texture_diffuse",
+    #        "data": "<b64>"}
+    #     ]
+    #   }
+    #
+    # ``path`` is the authoritative cross-file reference key — a POSIX-
+    # relative path (no ``..``, no absolute prefix, no Windows drive
+    # letter, no backslashes, no case-insensitive collisions). The
+    # client decoder materialises files into a temp directory using
+    # exactly these paths so a ``.obj`` that references its ``.mtl``
+    # via filename (or a ``.gltf`` that references its ``.bin`` /
+    # textures via relative URI) resolves on disk.
+    #
+    # ``role`` is an OPTIONAL producer hint (NOT a closed enum) —
+    # well-known values include ``mesh``, ``material``, ``buffer``,
+    # ``texture_diffuse``, ``texture_metallic``, ``texture_normal``,
+    # ``texture_roughness``, ``texture_ao``, ``texture_emissive``,
+    # ``texture_height``. Unknown roles round-trip without error;
+    # consumers may use ``role`` for convenience but ``path`` is the
+    # primary identifier.
+    #
+    # ``primary_path`` MUST appear in ``files`` exactly once and pins
+    # the entry-point file (the one passed to ComfyUI's
+    # ``Types.File3D`` constructor by the client decoder). The
+    # top-level ``format`` (if set) MUST match that primary file's
+    # ``format``.
+    #
+    # The client decoder returns a ``BundledFile3D`` (a thin
+    # ``File3D``-compatible wrapper) — existing downstream loaders
+    # that call ``get_source()`` / ``get_data()`` / ``save_to()`` get
+    # the bundle's primary file with all sidecars materialised on
+    # disk first so cross-file refs resolve correctly. Override of
+    # ``save_to()`` copies the *whole bundle* into the destination
+    # directory, not just the primary file.
+    #
+    # Negotiation: the server emits this encoding ONLY when the
+    # inbound request advertised this token; legacy clients without
+    # the bundle decoder will surface NEGOTIATION_FAILED at
+    # descriptor-load time. Whole-descriptor gating — if a node has
+    # both bundle and non-bundle outputs and the cap is absent, the
+    # whole descriptor is skipped rather than synthesising a
+    # partial-socket variant.
+    #
+    # A future sibling ``Capability.MODEL_3D_BUNDLE_URI`` would carry
+    # ``uri`` per file for URL-fetch-out-of-band semantics; not in
+    # this PR.
+    MODEL_3D_BUNDLE_INLINE = "model_3d:bundle_inline"
 
 
-HEAVY_TYPES = frozenset({"image", "video", "audio", "mask"})
+HEAVY_TYPES = frozenset({"image", "video", "audio", "mask", "model_3d"})
 
 
 def is_envelope(value: Any) -> bool:

proxy_node.py

@@ -3,9 +3,17 @@
 The descriptor wire format is V3 ``Schema.get_v1_info()`` verbatim, so
 per-input dicts are handed straight to V3's IO classes without
 reinterpretation. ``_parse_input_spec`` knows STRING / INT / FLOAT /
-BOOLEAN / COMBO / IMAGE / VIDEO / AUDIO / MASK; an input outside that
-set causes the whole node to be skipped with a warning rather than
-half-built.
+BOOLEAN / COMBO / IMAGE / VIDEO / AUDIO / MASK plus the two V3
+dynamic input types DYNAMIC_COMBO (branch-selector dropdown with
+per-branch sub-widgets) and AUTOGROW (variable-arity slot template);
+any other io_type string falls through to ``IO.Custom(io_type)`` (the
+§B opaque pass-through bucket — partner helper-config types like
+RECRAFT_*, OPENAI_INPUT_FILES, OPENAI_CHAT_CONFIG, GEMINI_INPUT_FILES
+round-trip as raw JSON blobs with the original io_type preserved on
+the V3 socket, so node-to-node connections only chain when the strings
+match). Only a non-string / empty io_type causes a skip. Malformed
+DYNAMIC_COMBO / AUTOGROW specs skip the descriptor (return ``None``)
+rather than degrading to a single opaque socket.
 
 Hidden inputs (``auth_token_comfy_org`` / ``api_key_comfy_org`` /
 ``unique_id`` / ``prompt`` / ``extra_pnginfo`` / ``dynprompt``) are
@@ -22,6 +30,7 @@ from __future__ import annotations
 import asyncio
 import contextlib
 import logging
+import math
 import time
 import uuid
 from typing import Any
@@ -36,8 +45,9 @@ from . import client as rnp_client
 from . import serialization
 from .protocol import (
     DEFAULT_CANCEL_TIMEOUT_S,
-    DEFAULT_MAX_POLL_ATTEMPTS,
+    DEFAULT_HARD_TIMEOUT_S,
     DEFAULT_MAX_RETRIES_PER_POLL,
+    DEFAULT_POLL_GRACE_S,
     DEFAULT_POLL_INTERVAL_S,
     DEFAULT_RETRY_BACKOFF_PER_POLL,
     DEFAULT_RETRY_DELAY_PER_POLL_S,
@@ -146,19 +156,32 @@ def _extract_execution_policy(execution: dict[str, Any]) -> dict[str, Any]:
     node with an empty ``execution`` block behaves identically to a
     direct upstream call. The schema is the wire contract from
     ``comfy_rnp_protocol.constants``: ``poll_interval_s``,
-    ``max_poll_attempts``, ``timeout_per_poll_s``, ``cancel_timeout_s``,
-    ``max_retries_per_poll``, ``retry_delay_per_poll_s``,
-    ``retry_backoff_per_poll``, ``retry: {max, delay_s, backoff,
-    retry_on}``, ``idempotency``.
+    ``soft_timeout_s``, ``hard_timeout_s``, ``timeout_per_poll_s``,
+    ``cancel_timeout_s``, ``max_retries_per_poll``,
+    ``retry_delay_per_poll_s``, ``retry_backoff_per_poll``,
+    ``retry: {max, delay_s, backoff, retry_on}``, ``idempotency``.
+
+    The client's poll-loop cap (``max_poll_attempts``) is **derived**
+    from ``hard_timeout_s + DEFAULT_POLL_GRACE_S`` and
+    ``poll_interval_s`` — there is no wire field for it. Coupling a
+    contract bound to a cadence value silently breaks every time you
+    tune the cadence; expressing the bound as a duration keeps the
+    contract invariant under cadence changes (see protocol.py
+    "Lifetime budget design" comment).
     """
     retry_block = execution.get("retry") if isinstance(execution.get("retry"), dict) else {}
+    poll_interval_s = _coerce_pos(
+        execution.get("poll_interval_s"), DEFAULT_POLL_INTERVAL_S,
+    )
+    hard_timeout_s = _coerce_pos(
+        execution.get("hard_timeout_s"),
+        _coerce_pos(execution.get("soft_timeout_s"), DEFAULT_HARD_TIMEOUT_S),
+    )
+    poll_budget_s = hard_timeout_s + DEFAULT_POLL_GRACE_S
+    max_poll_attempts = max(1, math.ceil(poll_budget_s / poll_interval_s))
     return {
-        "poll_interval_s": _coerce_pos(
-            execution.get("poll_interval_s"), DEFAULT_POLL_INTERVAL_S,
-        ),
-        "max_poll_attempts": _coerce_pos_int(
-            execution.get("max_poll_attempts"), DEFAULT_MAX_POLL_ATTEMPTS,
-        ),
+        "poll_interval_s": poll_interval_s,
+        "max_poll_attempts": max_poll_attempts,
         "timeout_per_poll_s": _coerce_pos(
             execution.get("timeout_per_poll_s"), DEFAULT_TIMEOUT_PER_POLL_S,
         ),
@@ -290,6 +313,32 @@ def build_node_class(
     schema_hash = remote.get("schema_hash")
     execute_path = execute_endpoint["path"]
     execution = remote.get("execution") or {}
+    # ``descriptor.remote.input_serialization`` advertises the wire
+    # encodings each input accepts. The wire shape is
+    # ``{input_name: <str | list[str]>}`` — single string for a
+    # single-encoding input (e.g. ``"png_base64"``), list when the
+    # input accepts multiple encodings (e.g.
+    # ``["png_base64", "png_base64_batch"]`` for a multi-frame IMAGE).
+    # Normalize to ``dict[str, list[str]]`` so the encoder can do a
+    # single ``in`` check per input without re-parsing per call.
+    raw_input_serialization = (
+        remote.get("input_serialization") if isinstance(remote.get("input_serialization"), dict) else {}
+    )
+    input_serialization: dict[str, list[str]] = {}
+    for name, val in raw_input_serialization.items():
+        if isinstance(val, str):
+            input_serialization[name] = [val]
+        elif isinstance(val, list):
+            input_serialization[name] = [v for v in val if isinstance(v, str)]
+
+    # ``descriptor.input.<section>.<name>[1].local_validate`` carries
+    # declarative client-side validation rules (e.g.
+    # ``{"image_max_batch": 9}`` for a Flux2 reference-image input).
+    # The encoder consults this map and raises ``INPUT_INVALID``
+    # locally before any HTTP round trip when an over-cap batch is
+    # wired in. Lookup is defensive: missing ``local_validate`` →
+    # no validation; missing per-rule key → no validation.
+    local_validate = _collect_local_validate(descriptor)
     timeout_s = float(
         execution.get("hard_timeout_s")
         or execution.get("soft_timeout_s")
@@ -370,6 +419,8 @@ def build_node_class(
                 execution_mode=execution_mode,
                 execution_policy=execution_policy,
                 url_fetch_policy=url_fetch_policy,
+                input_serialization=input_serialization,
+                local_validate=local_validate,
             )
 
     RemoteProxyNode.__name__ = node_id
@@ -395,6 +446,111 @@ _HIDDEN_MARKERS = {
 }
 
 
+def _collect_local_validate(descriptor: dict[str, Any]) -> dict[str, dict[str, Any]]:
+    """Pull ``local_validate`` rule dicts off every required/optional input
+    spec in the descriptor, recursing into AUTOGROW templates and
+    DYNAMIC_COMBO branches.
+
+    Returns ``{input_name: rules_dict}`` for every input that carries a
+    non-empty ``local_validate`` block (or whose nested template /
+    branches do). The returned ``rules_dict`` may carry the literal
+    ``local_validate`` keys (e.g. ``{"image_max_batch": 9}``) plus two
+    private composite keys consumed by ``_enforce_local_validate``:
+
+    * ``__template__`` — for AUTOGROW inputs, the rules harvested from
+      the template's options dict. The enforcer applies these to every
+      slot value in the runtime Autogrow dict
+      (``{"image0": <env>, "image1": <env>, ...}``).
+    * ``__branches__`` — for DYNAMIC_COMBO inputs, a
+      ``{branch_key: {sub_name: rules}}`` map. The enforcer reads the
+      selected branch from ``value[<name>]`` and walks the matching
+      branch's sub-inputs (used by GPT Image V2 / Grok ImageEditV2 /
+      Grok VideoReference, where the cap lives on a nested Autogrow
+      template inside a DynamicCombo branch).
+
+    Inputs whose entire subtree is empty are omitted so the encoder can
+    use a single ``in`` check per call.
+    """
+    out: dict[str, dict[str, Any]] = {}
+    input_def = descriptor.get("input") or {}
+    for kind in ("required", "optional"):
+        defs = input_def.get(kind) or {}
+        for name, spec in defs.items():
+            rules = _rules_from_input_spec(spec)
+            if rules:
+                out[name] = rules
+    return out
+
+
+def _rules_from_input_spec(spec: Any) -> dict[str, Any]:
+    """Extract ``local_validate`` rules from one V3 input spec.
+
+    Accepts both wire shapes:
+
+    * ``[io_type, options_dict]`` — top-level inputs (the descriptor's
+      ``input.required[name]`` value, or an AUTOGROW template).
+    * ``[name, io_type, options_dict]`` — DynamicCombo branch sub-inputs
+      (each entry of ``options.options[i].inputs``).
+
+    For ``AUTOGROW`` specs, recurses into ``options["template"]`` and
+    stashes the result under ``__template__``. For ``DYNAMIC_COMBO``,
+    walks each branch's sub-inputs and stashes the resulting
+    ``{branch_key: {sub_name: sub_rules}}`` map under ``__branches__``.
+    """
+    if not isinstance(spec, (list, tuple)) or len(spec) < 2:
+        return {}
+    if (
+        len(spec) >= 3
+        and isinstance(spec[1], str)
+        and isinstance(spec[2], dict)
+    ):
+        # [name, io_type, options]
+        io_type, options = spec[1], spec[2]
+    elif isinstance(spec[1], dict):
+        # [io_type, options]
+        io_type, options = spec[0], spec[1]
+    else:
+        return {}
+    rules: dict[str, Any] = {}
+    direct = options.get("local_validate")
+    if isinstance(direct, dict) and direct:
+        rules.update(direct)
+    if io_type == "AUTOGROW":
+        tmpl_spec = options.get("template")
+        tmpl_rules = (
+            _rules_from_input_spec(tmpl_spec)
+            if isinstance(tmpl_spec, (list, tuple))
+            else {}
+        )
+        if tmpl_rules:
+            rules["__template__"] = tmpl_rules
+    elif io_type == "DYNAMIC_COMBO":
+        branches: dict[str, dict[str, Any]] = {}
+        for opt in (options.get("options") or []):
+            if not isinstance(opt, dict):
+                continue
+            key = opt.get("key")
+            if not isinstance(key, str):
+                continue
+            sub_rules: dict[str, dict[str, Any]] = {}
+            for sub_spec in (opt.get("inputs") or []):
+                if (
+                    not isinstance(sub_spec, (list, tuple))
+                    or len(sub_spec) < 1
+                    or not isinstance(sub_spec[0], str)
+                ):
+                    continue
+                sub_name = sub_spec[0]
+                r = _rules_from_input_spec(sub_spec)
+                if r:
+                    sub_rules[sub_name] = r
+            if sub_rules:
+                branches[key] = sub_rules
+        if branches:
+            rules["__branches__"] = branches
+    return rules
+
+
 def _parse_inputs(descriptor: dict[str, Any]) -> tuple[list[Any] | None, list[str]]:
     """Return ``(input_objs, hidden_input_names)``.
 
@@ -424,34 +580,90 @@ def _parse_inputs(descriptor: dict[str, Any]) -> tuple[list[Any] | None, list[st
 
 
 def _parse_input_spec(name: str, spec: list[Any], optional: bool) -> Any | None:
-    """Map one V3-shaped ``[io_type, options_dict]`` entry to a V3 IO Input."""
+    """Map one V3-shaped ``[io_type, options_dict]`` entry to a V3 IO Input.
+
+    This is the top-level entry used by ``_parse_inputs`` (where the
+    name comes from the descriptor's ``input.required[name]`` /
+    ``input.optional[name]`` key) and recursively as the AUTOGROW
+    template entry (which also lacks a leading name field — the slot
+    name is provided by the surrounding AUTOGROW). Branch sub-inputs
+    inside a DYNAMIC_COMBO option use the 3-tuple ``[name, io_type,
+    options]`` form instead — see :func:`_parse_named_input_spec`.
+    """
     if not isinstance(spec, (list, tuple)) or len(spec) < 1:
         return None
     io_type = spec[0]
     options: dict[str, Any] = spec[1] if len(spec) > 1 and isinstance(spec[1], dict) else {}
+    return _build_input(name, io_type, options, optional)
+
+
+def _parse_named_input_spec(spec: list[Any], optional: bool) -> Any | None:
+    """Map one V3-shaped ``[name, io_type, options_dict]`` entry to a V3 IO Input.
+
+    This is the shape used inside ``DYNAMIC_COMBO`` options' ``inputs``
+    list (each branch sub-input carries its own name as ``spec[0]``).
+    Top-level descriptor inputs use the 2-tuple form handled by
+    :func:`_parse_input_spec`.
+    """
+    if (
+        not isinstance(spec, (list, tuple))
+        or len(spec) < 2
+        or not isinstance(spec[0], str)
+    ):
+        return None
+    name = spec[0]
+    io_type = spec[1]
+    options: dict[str, Any] = spec[2] if len(spec) > 2 and isinstance(spec[2], dict) else {}
+    return _build_input(name, io_type, options, optional)
+
+
+def _build_input(name: str, io_type: Any, options: dict[str, Any], optional: bool) -> Any | None:
+    """Shared dispatch: construct one V3 IO Input from a parsed
+    ``(name, io_type, options)`` triple.
+
+    Recurses for the two V3 dynamic input types:
+
+    * ``DYNAMIC_COMBO`` — selector dropdown whose per-branch sub-inputs
+      live in ``options.options[i].inputs`` (each entry is the 3-tuple
+      ``[name, io_type, options]`` form).
+    * ``AUTOGROW`` — variable-arity slot template. The RNP wire dialect
+      flattens upstream's ``Autogrow.Input.as_dict`` shape: ``template``
+      is a 2-tuple ``[io_type, options]`` entry directly, with
+      ``prefix`` / ``min`` / ``max`` (or ``names`` / ``min``) as siblings
+      of ``template`` rather than nested under ``template.template``.
+
+    Malformed dynamic specs return ``None`` (caller logs + skips the
+    whole descriptor) rather than falling through to the opaque
+    ``IO.Custom`` bucket — that would mask a protocol/schema bug as
+    "looks fine but renders as opaque socket".
+    """
     common = {
         "tooltip": options.get("tooltip"),
         "optional": optional,
     }
+    # ``advanced`` is only accepted by WidgetInput subclasses — the
+    # dynamic constructors (DynamicCombo.Input / Autogrow.Input) do
+    # not take it. Build a separate kwargs bag for primitives.
+    primitive_common = dict(common)
     if options.get("advanced") is not None:
-        common["advanced"] = bool(options.get("advanced"))
+        primitive_common["advanced"] = bool(options.get("advanced"))
 
     if isinstance(io_type, list):
         # Legacy V1-style combo: io_type *is* the options list.
         return IO.Combo.Input(name, options=list(io_type),
-                              default=options.get("default"), **common)
+                              default=options.get("default"), **primitive_common)
     if io_type == "COMBO":
         opts = options.get("options")
         if not isinstance(opts, (list, tuple)):
             return None
         return IO.Combo.Input(name, options=list(opts),
-                              default=options.get("default"), **common)
+                              default=options.get("default"), **primitive_common)
     if io_type == "STRING":
         return IO.String.Input(
             name,
             default=options.get("default", ""),
             multiline=bool(options.get("multiline", False)),
-            **common,
+            **primitive_common,
         )
     if io_type == "INT":
         return IO.Int.Input(
@@ -460,7 +672,7 @@ def _parse_input_spec(name: str, spec: list[Any], optional: bool) -> Any | None:
             min=options.get("min", 0),
             max=options.get("max", 2147483647),
             step=options.get("step", 1),
-            **common,
+            **primitive_common,
         )
     if io_type == "FLOAT":
         return IO.Float.Input(
@@ -469,20 +681,160 @@ def _parse_input_spec(name: str, spec: list[Any], optional: bool) -> Any | None:
             min=options.get("min", 0.0),
             max=options.get("max", 1.0),
             step=options.get("step", 0.01),
-            **common,
+            **primitive_common,
         )
     if io_type == "BOOLEAN":
         return IO.Boolean.Input(
-            name, default=bool(options.get("default", False)), **common,
+            name, default=bool(options.get("default", False)), **primitive_common,
         )
     if io_type == "IMAGE":
-        return IO.Image.Input(name, **common)
+        return IO.Image.Input(name, **primitive_common)
     if io_type == "VIDEO":
-        return IO.Video.Input(name, **common)
+        return IO.Video.Input(name, **primitive_common)
     if io_type == "AUDIO":
-        return IO.Audio.Input(name, **common)
+        return IO.Audio.Input(name, **primitive_common)
     if io_type == "MASK":
-        return IO.Mask.Input(name, **common)
+        return IO.Mask.Input(name, **primitive_common)
+    if io_type == "MODEL_3D":
+        # 3D MODEL input: symmetric counterpart of the ``"MODEL_3D"``
+        # entry in ``_OUTPUT_CLASSES`` above. The Tencent Hunyuan3D
+        # MODEL_3D-input nodes (``TencentModelTo3DUVNode`` /
+        # ``Tencent3DTextureEditNode`` / ``Tencent3DPartNode`` /
+        # ``TencentSmartTopologyNode``) all declare their model_3d
+        # input as ``IO.MultiType.Input(types=[IO.File3DGLB,
+        # IO.File3DOBJ, IO.File3DFBX, IO.File3DAny])`` so the socket
+        # accepts any upstream 3D source: ``Load3D``'s ``File3DAny``
+        # output, Meshy / Hunyuan3D / Tripo / Rodin's per-format
+        # outputs (``FILE_3D_GLB`` / ``FILE_3D_OBJ`` / ``FILE_3D_FBX``),
+        # etc. Mirror that union here so the proxy socket accepts the
+        # same set of comfytypes upstream does. Per-node format
+        # validation (e.g. 3DTextureEdit's FBX-only enforcement) is
+        # done server-side in the provider's ``execute()`` against
+        # the inline envelope's ``format`` extra; this socket-level
+        # filter is purely about frontend connection validity.
+        return IO.MultiType.Input(
+            name,
+            types=[
+                IO.File3DGLB, IO.File3DOBJ, IO.File3DFBX, IO.File3DAny,
+            ],
+            **primitive_common,
+        )
+    if io_type == "DYNAMIC_COMBO":
+        # Wire shape: ``["DYNAMIC_COMBO", {"options": [{"key": str,
+        # "inputs": [[name, io_type, opts], ...]}, ...], "tooltip": ...}]``.
+        # Empty branch input lists are valid (e.g. Bria moderation's
+        # "false" branch with ``"inputs": []``); the branch dropdown
+        # still needs to render. Malformed shape -> None (no opaque
+        # fallthrough — would mask a schema bug).
+        opts = options.get("options")
+        if not isinstance(opts, (list, tuple)):
+            return None
+        parsed_options: list[Any] = []
+        for opt in opts:
+            if not isinstance(opt, dict):
+                return None
+            key = opt.get("key")
+            if not isinstance(key, str):
+                return None
+            sub_specs = opt.get("inputs") or []
+            if not isinstance(sub_specs, (list, tuple)):
+                return None
+            sub_inputs: list[Any] = []
+            for sub_spec in sub_specs:
+                # Branch sub-inputs are always treated as required at
+                # schema-construction time — the wire dialect doesn't
+                # express per-branch optionality. Server-side runtime
+                # validation enforces presence via ``__branches__``.
+                sub = _parse_named_input_spec(sub_spec, optional=False)
+                if sub is None:
+                    return None
+                sub_inputs.append(sub)
+            parsed_options.append(IO.DynamicCombo.Option(key=key, inputs=sub_inputs))
+        return IO.DynamicCombo.Input(
+            name,
+            options=parsed_options,
+            tooltip=options.get("tooltip"),
+            optional=optional,
+        )
+    if io_type == "AUTOGROW":
+        # Wire shape (RNP-flattened):
+        # ``["AUTOGROW", {"template": [io_type, options],
+        #                  "prefix": str, "min": int, "max": int,
+        #                  "tooltip": ...}]``
+        # or with ``names`` instead of ``prefix`` / ``max`` for the
+        # TemplateNames variant.
+        tmpl_spec = options.get("template")
+        if not isinstance(tmpl_spec, (list, tuple)) or len(tmpl_spec) < 1:
+            return None
+        # Reuse top-level parser for the template — the slot name is
+        # cosmetic since TemplatePrefix / TemplateNames rename each
+        # cached copy per-slot.
+        template_input = _parse_input_spec(name, tmpl_spec, optional=False)
+        if template_input is None:
+            return None
+        # Upstream Autogrow asserts the template input is NOT itself a
+        # DynamicInput (DynamicCombo.Input / Autogrow.Input). Mirror
+        # that check at parse time so a malformed descriptor surfaces
+        # as "skipped" rather than blowing up later in the IO machinery.
+        if isinstance(
+            template_input,
+            (IO.DynamicCombo.Input, IO.Autogrow.Input),
+        ):
+            return None
+        if "names" in options:
+            names = options.get("names")
+            if not isinstance(names, list) or not all(isinstance(n, str) for n in names):
+                return None
+            try:
+                template = IO.Autogrow.TemplateNames(
+                    template_input,
+                    names=names,
+                    min=int(options.get("min", 1)),
+                )
+            except AssertionError:
+                return None
+        elif "prefix" in options:
+            prefix = options.get("prefix")
+            if not isinstance(prefix, str):
+                return None
+            try:
+                template = IO.Autogrow.TemplatePrefix(
+                    template_input,
+                    prefix=prefix,
+                    min=int(options.get("min", 1)),
+                    max=int(options.get("max", 10)),
+                )
+            except AssertionError:
+                return None
+        else:
+            return None
+        return IO.Autogrow.Input(
+            name,
+            template=template,
+            tooltip=options.get("tooltip"),
+            optional=optional,
+        )
+    # Opaque custom-IO fallthrough — any unknown io_type string becomes
+    # an ``IO.Custom(io_type)`` socket so partner helper-config nodes
+    # (OpenAIInputFiles → OPENAI_INPUT_FILES, OpenAIChatConfig →
+    # OPENAI_CHAT_CONFIG, GeminiInputFiles → GEMINI_INPUT_FILES, the
+    # already-shipped Recraft RECRAFT_* helpers, future partner types)
+    # round-trip without per-type encoder plumbing. The descriptor's
+    # original ``io_type`` string is preserved on ``IO.Custom``'s
+    # ComfyType.io_type via the V3 ``@comfytype`` decorator, so the
+    # frontend's connection-validity check only chains sockets when the
+    # strings match — an OPENAI_INPUT_FILES output cannot connect to a
+    # GEMINI_INPUT_FILES input. Encoder dispatch in
+    # ``serialization`` already passes non-tensor / non-audio values
+    # through ``_encode_one`` unchanged, and the deserializer treats
+    # non-envelope values as already-native Python objects.
+    #
+    # NOTE: ``DYNAMIC_COMBO`` and ``AUTOGROW`` are handled above and
+    # never reach this fallthrough — malformed dynamic specs return
+    # ``None`` instead, so the caller logs + skips the descriptor
+    # rather than silently degrading to a single opaque socket.
+    if isinstance(io_type, str) and io_type:
+        return IO.Custom(io_type).Input(name, **common)
     return None
 
 
@@ -567,23 +919,47 @@ def _parse_outputs(descriptor: dict[str, Any]) -> list[Any] | None:
         tip = output_tooltips[i] if i < len(output_tooltips) else None
         cls = _OUTPUT_CLASSES.get(io_type)
         if cls is None:
-            log.warning(
-                "Skipping descriptor: unsupported output type %r", io_type,
-            )
-            return None
+            # Opaque custom-IO fallthrough (mirror of
+            # ``_parse_input_spec``) — any unknown output io_type
+            # string becomes ``IO.Custom(io_type).Output`` so partner
+            # helper-config nodes can publish their own opaque
+            # outputs (RECRAFT_*, OPENAI_INPUT_FILES, GEMINI_INPUT_FILES,
+            # OPENAI_CHAT_CONFIG, …). The original io_type string
+            # rides on the resulting Output via V3's ``@comfytype``
+            # decorator, so the frontend only chains output→input
+            # sockets when the strings match.
+            if isinstance(io_type, str) and io_type:
+                cls = IO.Custom(io_type).Output
+            else:
+                log.warning(
+                    "Skipping descriptor: unsupported output type %r",
+                    io_type,
+                )
+                return None
         out.append(cls(display_name=name, is_output_list=is_list, tooltip=tip))
     return out
 
 
 _OUTPUT_CLASSES = {
-    "VIDEO":   IO.Video.Output,
-    "IMAGE":   IO.Image.Output,
-    "AUDIO":   IO.Audio.Output,
-    "MASK":    IO.Mask.Output,
-    "STRING":  IO.String.Output,
-    "INT":     IO.Int.Output,
-    "FLOAT":   IO.Float.Output,
-    "BOOLEAN": IO.Boolean.Output,
+    "VIDEO":    IO.Video.Output,
+    "IMAGE":    IO.Image.Output,
+    "AUDIO":    IO.Audio.Output,
+    "MASK":     IO.Mask.Output,
+    "STRING":   IO.String.Output,
+    "INT":      IO.Int.Output,
+    "FLOAT":    IO.Float.Output,
+    "BOOLEAN":  IO.Boolean.Output,
+    # 3D MODEL output: the Tripo / Rodin / Meshy partner endpoints all
+    # return a single GLB file; the server emits a ``model_3d:glb_inline``
+    # envelope which ``serialization.decode_model3d_envelope`` turns
+    # back into a ``File3D`` so downstream nodes (preview3d / SaveGLB /
+    # Load3D) chain off the output socket unchanged. The descriptor
+    # io_type string is the logical name ``"MODEL_3D"`` (consistent
+    # across vendors) — the resulting V3 socket carries the upstream
+    # ``"FILE_3D_GLB"`` comfytype tag from ``IO.File3DGLB`` so
+    # frontend-side input/output matching works against any GLB-aware
+    # node.
+    "MODEL_3D": IO.File3DGLB.Output,
 }
 
 
@@ -831,6 +1207,8 @@ async def _execute_remote(
     execution_mode: str = ExecutionMode.REQUEST_RESPONSE,
     execution_policy: dict[str, Any] | None = None,
     url_fetch_policy: dict[str, Any] | None = None,
+    input_serialization: dict[str, list[str]] | None = None,
+    local_validate: dict[str, dict[str, Any]] | None = None,
 ) -> IO.NodeOutput:
     policy = execution_policy or _extract_execution_policy({})
     url_fetch_policy = url_fetch_policy or {}
@@ -864,11 +1242,18 @@ async def _execute_remote(
     # Encode heavy-typed inputs (IMAGE/MASK/AUDIO) as RNP value
     # envelopes; scalars pass through. Oversized envelopes are
     # uploaded out-of-band and rewritten to ``{..., uri: <url>}``.
+    # ``input_serialization`` carries the per-input encoding allow-list
+    # from the descriptor — the encoder uses it to decide whether to
+    # ship a multi-frame ``png_base64_batch`` envelope or fall back
+    # to single-frame ``png_base64``.
     payload_inputs = await _encode_inputs(
         inputs,
         server_url=server_url,
         max_inline_bytes=max_inline_bytes,
         auth_headers=auth_headers,
+        input_serialization=input_serialization or {},
+        local_validate=local_validate or {},
+        node_id=node_id,
     )
 
     ticker_node_id = _node_id_for_ticker(cls)
@@ -1043,8 +1428,54 @@ async def _deserialize_output(
         # advisory here, the value already matches the native ComfyUI
         # type for STRING / INT / FLOAT / BOOLEAN.
         return value
-    if "data" not in value and "uri" in value:
-        env_type = value.get("type", "")
+    encoding = value.get("encoding")
+    env_type = value.get("type", "")
+    if encoding == "png_base64_batch_uri":
+        # Externalised IMAGE batch on the response path (mirror of the
+        # request-side ``serialization.maybe_externalize`` behaviour).
+        # No server emits this shape today, but the decoder owns the
+        # symmetry: any envelope shape the client may externalise on
+        # the way out, the client must also accept on the way in.
+        # Per-frame URIs fetched in parallel via ``asyncio.gather`` so
+        # a 9-frame batch import is one round-trip rather than N — the
+        # GET fan-out is the dominant latency for an out-of-band batch.
+        # Frame ordering is preserved (gather returns positional
+        # results) — critical because the IMAGE tensor's batch axis
+        # carries semantic order.
+        policy = url_fetch_policy.get(env_type)
+        if not isinstance(policy, dict):
+            raise RnpProtocolError(
+                f"Server returned a URL batch envelope for output type "
+                f"{env_type!r} but the descriptor declared no "
+                f"remote.url_fetch policy for that type",
+                code=ErrorCode.INTERNAL,
+            )
+        uris = value.get("uris")
+        if not isinstance(uris, list) or not uris:
+            raise RnpProtocolError(
+                "png_base64_batch_uri envelope requires a non-empty 'uris' list",
+                code=ErrorCode.INTERNAL,
+            )
+        import asyncio as _asyncio
+        import base64 as _b64
+        per_frame_envs = [
+            {"type": env_type, "encoding": "png_base64", "uri": u}
+            for u in uris
+        ]
+        raws = await _asyncio.gather(*[
+            _fetch_url_envelope(
+                env, policy,
+                auth_headers=auth_headers,
+                node_cls=node_cls,
+            )
+            for env in per_frame_envs
+        ])
+        # Re-stamp as the inline batch shape so the existing
+        # ``_decode_image_batch_envelope`` path handles it uniformly.
+        value = {k: v for k, v in value.items() if k != "uris"}
+        value["encoding"] = "png_base64_batch"
+        value["frames"] = [_b64.b64encode(r).decode("ascii") for r in raws]
+    elif "data" not in value and "uri" in value:
         policy = url_fetch_policy.get(env_type)
         if not isinstance(policy, dict):
             raise RnpProtocolError(
@@ -1126,6 +1557,9 @@ async def _encode_inputs(
     server_url: str | None = None,
     max_inline_bytes: int | None = None,
     auth_headers: dict[str, str] | None = None,
+    input_serialization: dict[str, list[str]] | None = None,
+    local_validate: dict[str, dict[str, Any]] | None = None,
+    node_id: str | None = None,
 ) -> dict[str, Any]:
     """Convert in-place: replace heavy-typed input values with the
     matching RNP value envelope. Scalars pass through unchanged.
@@ -1136,40 +1570,240 @@ async def _encode_inputs(
     * ``torch.Tensor`` rank 3 with last-dim != 3/4 → mask envelope (B,H,W).
       Rank-2 (H,W) tensors are also treated as masks.
     * ``dict`` with ``waveform`` + ``sample_rate`` → audio envelope.
+    * ``VideoInput`` subclass (duck-typed via ``save_to`` +
+      ``get_duration``) → mp4-base64 video envelope.
     * Everything else passes through as-is (already JSON-serializable).
 
     When ``max_inline_bytes`` is set, any envelope whose inline payload
     exceeds the threshold is externalized via ``POST /rnp/v1/uploads``
     (see ``serialization.maybe_externalize``).
+
+    ``input_serialization`` is the descriptor's per-input encoding
+    allow-list (e.g. ``{"images": ["png_base64", "png_base64_batch"]}``).
+    The encoder consults it to decide whether a multi-frame IMAGE input
+    can ride a single batch envelope vs. falling back to single-frame.
+
+    ``local_validate`` is the per-input declarative-rule map
+    (e.g. ``{"images": {"image_max_batch": 9}}``). For each rule the
+    encoder enforces the constraint locally and raises
+    ``RnpProtocolError(code=INPUT_INVALID, user_facing=True)`` so the
+    proxy_node's friendly-message dispatch surfaces a per-provider
+    error before any paid HTTP round trip. Lookup is defensive: a
+    missing ``local_validate`` entry or missing per-rule key is a
+    no-op.
     """
+    serialization_map = input_serialization or {}
+    validate_map = local_validate or {}
     out: dict[str, Any] = {}
     for name, value in inputs.items():
-        encoded = _encode_one(name, value)
-        # Only envelope-shaped values can be externalized; scalars pass
-        # straight through ``maybe_externalize`` unchanged.
-        if serialization.is_envelope(encoded):
-            encoded = await serialization.maybe_externalize(
-                encoded,
-                server_url=server_url,
-                max_inline_bytes=max_inline_bytes,
-                auth_headers=auth_headers,
-            )
+        _enforce_local_validate(
+            name, value, validate_map.get(name) or {}, node_id=node_id,
+        )
+        encoded = _encode_one(name, value, serialization_map.get(name))
+        # Externalize any envelope payloads, walking into nested dicts so
+        # AUTOGROW slot values and DYNAMIC_COMBO sub-inputs whose encoded
+        # form is a per-slot envelope are uploaded the same way a
+        # top-level singleton envelope is. ``_externalize_nested`` is a
+        # no-op on plain scalars and on dicts that contain no envelopes.
+        encoded = await _externalize_nested(
+            encoded,
+            server_url=server_url,
+            max_inline_bytes=max_inline_bytes,
+            auth_headers=auth_headers,
+        )
         out[name] = encoded
     return out
 
 
-def _encode_one(name: str, value: Any) -> Any:
+async def _externalize_nested(
+    value: Any,
+    *,
+    server_url: str | None,
+    max_inline_bytes: int | None,
+    auth_headers: dict[str, str] | None,
+) -> Any:
+    """Recursively externalize any RNP envelopes inside ``value``.
+
+    Mirrors :func:`_encode_one`'s dict-recursion: ComfyUI hands the
+    proxy_node nested AUTOGROW / DYNAMIC_COMBO runtime dicts
+    (e.g. ``{"reference_images": {"image1": <env>, "image2": <env>}}``
+    or ``{"model": {"branch": "wan2.7-r2v",
+    "video_refs": {"video1": <env>}}}``); without this walk the
+    per-slot envelopes would ship inline regardless of
+    ``max_inline_bytes`` because :func:`_encode_inputs` only inspected
+    the top-level value.
+
+    Stops at envelope leaves (delegates the per-envelope decision to
+    :func:`serialization.maybe_externalize`) and at plain scalars (leaves
+    them unchanged). Lists are intentionally left alone — AG/DC runtime
+    shapes always materialize as dicts, never lists, and recursing into
+    arbitrary lists would touch opaque payloads. Recursion is bounded
+    by the AUTOGROW max slot count (100) and the maximum DC branch
+    nesting depth (1 today), so an in-place ``async def`` walk is fine.
+    """
+    if serialization.is_envelope(value):
+        return await serialization.maybe_externalize(
+            value,
+            server_url=server_url,
+            max_inline_bytes=max_inline_bytes,
+            auth_headers=auth_headers,
+        )
+    if isinstance(value, dict):
+        out: dict[str, Any] = {}
+        for k, v in value.items():
+            out[k] = await _externalize_nested(
+                v,
+                server_url=server_url,
+                max_inline_bytes=max_inline_bytes,
+                auth_headers=auth_headers,
+            )
+        return out
+    return value
+
+
+def _enforce_local_validate(
+    name: str,
+    value: Any,
+    rules: dict[str, Any],
+    *,
+    node_id: str | None,
+) -> None:
+    """Apply each ``local_validate.*`` rule to ``value`` and raise
+    ``RnpProtocolError(INPUT_INVALID)`` on a violation.
+
+    Rules currently supported:
+
+    * ``image_max_batch`` — for IMAGE inputs (rank-4 tensors), assert
+      ``tensor.shape[0] <= cap``. Mirrors the per-call cap raised on
+      the server (e.g. Flux2 raises "The current maximum number of
+      supported images is 9.") so a workflow with N>cap reference
+      frames fails fast on the client before any paid HTTP round trip.
+
+    Two private composite keys are walked transparently:
+
+    * ``__template__`` (AUTOGROW) — ``value`` is expected to be a
+      runtime Autogrow dict (``{"image0": <env>, "image1": <env>, ...}``).
+      The template's rules are applied to each non-``None`` slot value.
+    * ``__branches__`` (DYNAMIC_COMBO) — ``value`` is expected to be the
+      DynamicCombo runtime dict whose ``value[<name>]`` carries the
+      selected branch key. The matching branch's per-sub-input rules are
+      applied recursively. This is what makes GPT Image V2 / Grok
+      ImageEditV2 / Grok VideoReference's nested per-slot caps fire
+      client-side before any HTTP round trip.
+    """
+    if not rules:
+        return
+    _check_image_max_batch(name, value, rules.get("image_max_batch"), node_id=node_id)
+    tmpl = rules.get("__template__")
+    if isinstance(tmpl, dict) and isinstance(value, dict):
+        for slot_key, slot_value in value.items():
+            if slot_value is None:
+                continue
+            _enforce_local_validate(
+                f"{name}.{slot_key}", slot_value, tmpl, node_id=node_id,
+            )
+    branches = rules.get("__branches__")
+    if isinstance(branches, dict) and isinstance(value, dict):
+        branch_key = value.get(name)
+        if isinstance(branch_key, str):
+            sub_rules_map = branches.get(branch_key)
+            if isinstance(sub_rules_map, dict):
+                for sub_name, sub_rules in sub_rules_map.items():
+                    if sub_name in value:
+                        _enforce_local_validate(
+                            f"{name}.{sub_name}", value[sub_name], sub_rules,
+                            node_id=node_id,
+                        )
+
+
+def _check_image_max_batch(
+    name: str,
+    value: Any,
+    cap: Any,
+    *,
+    node_id: str | None,
+) -> None:
+    """Apply the ``image_max_batch`` rule. No-op if cap is missing /
+    non-int / value is not a recognisable image tensor."""
+    if cap is None:
+        return
+    try:
+        cap_int = int(cap)
+    except (TypeError, ValueError):
+        return
+    if not serialization._is_torch_tensor(value):
+        return
+    rank = value.dim()
+    last_dim = int(value.shape[-1]) if rank >= 1 else 0
+    if not (rank == 4 or (rank == 3 and last_dim in (3, 4))):
+        return
+    batch = int(value.shape[0]) if rank == 4 else 1
+    if batch <= cap_int:
+        return
+    label = f" ({node_id})" if node_id else ""
+    if cap_int == 1:
+        msg = (
+            f"Remote node{label}: input {name!r} expects a single image "
+            f"frame, got a batch of {batch}. Wire a single-frame IMAGE "
+            f"into this input."
+        )
+    else:
+        msg = (
+            f"Remote node{label}: input {name!r} accepts at most "
+            f"{cap_int} image frames, got a batch of {batch}. The "
+            f"current maximum number of supported images is "
+            f"{cap_int}."
+        )
+    raise RnpProtocolError(
+        msg,
+        code=ErrorCode.INPUT_INVALID,
+        user_facing=True,
+    )
+
+
+def _encode_one(
+    name: str,
+    value: Any,
+    accepted_encodings: list[str] | None = None,
+) -> Any:
+    # ComfyUI's V3 ``build_nested_inputs`` rebuilds AUTOGROW / DYNAMIC_COMBO
+    # runtime kwargs into nested dicts (``{"reference_images": {"image1":
+    # <tensor>, ...}}`` for top-level AG; ``{"<dc-name>": branch_key,
+    # "<nested-ag>": {"video1": <VideoInput>, ...}}`` for AG-inside-DC).
+    # Recurse into plain dicts so heavy leaves anywhere in that tree get
+    # encoded — without this an AG-of-IMAGE slot would ship the raw
+    # torch.Tensor inside a dict and crash at ``json.dumps`` time. Stops
+    # at envelope dicts (already-encoded payloads) so we don't double-
+    # encode wire-shape values that happen to be dicts.
+    if serialization.is_envelope(value):
+        return value
+    if isinstance(value, dict) and not serialization.is_audio_input(value):
+        return {
+            k: _encode_one(f"{name}.{k}", v, accepted_encodings)
+            for k, v in value.items()
+        }
     if serialization.is_audio_input(value):
         return serialization.encode_audio_input(value)
+    if serialization.is_video_input(value):
+        return serialization.encode_video_input(value)
+    if serialization.is_model3d_input(value):
+        return serialization.encode_model3d_input(value)
     if not serialization._is_torch_tensor(value):
         return value
     rank = value.dim()
     last_dim = int(value.shape[-1]) if rank >= 1 else 0
     if rank == 4 or (rank == 3 and last_dim in (3, 4)):
-        # Per-image tensors: a length-1 batch is the common case; longer
-        # batches send only the first frame (the framework hasn't asked
-        # for batch-of-envelopes semantics yet).
-        return serialization.encode_image_tensor(value)
+        # Per-image tensors: a length-1 batch is the common case.
+        # When the descriptor advertises ``"png_base64_batch"`` for
+        # this input AND the tensor has > 1 frame, ship the multi-
+        # frame envelope; otherwise fall back to single-frame
+        # (``tensor[0]`` for batches >1, full tensor for length-1).
+        accepts_batch = bool(
+            accepted_encodings and "png_base64_batch" in accepted_encodings
+        )
+        return serialization.encode_image_tensor(
+            value, accepts_batch=accepts_batch,
+        )
     if rank in (2, 3):
         return serialization.encode_mask_tensor(value)
     log.warning(

serialization.py

@@ -17,6 +17,7 @@ tensor — no parallel dialect.
 """
 from __future__ import annotations
 
+import asyncio
 import base64
 import logging
 from io import BytesIO
@@ -56,14 +57,64 @@ async def maybe_externalize(
     max_inline_bytes: int | None,
     auth_headers: dict[str, str] | None = None,
 ) -> dict[str, Any]:
-    """If the envelope's inline data exceeds ``max_inline_bytes``,
-    upload the bytes and replace ``data`` with ``uri``.
+    """If the envelope's inline payload exceeds ``max_inline_bytes``,
+    upload it via the presigned-PUT path and rewrite the envelope to
+    reference the uploaded asset(s) instead of inlining base64.
+
+    Two payload shapes are recognised:
+
+    * Singleton ``data`` envelopes (the common single-frame IMAGE /
+      MASK / VIDEO / AUDIO case). Oversize payloads upload as one
+      asset and the envelope's ``data`` field is swapped for ``uri``.
+    * Multi-frame ``frames=[…]`` IMAGE envelopes (encoding
+      ``png_base64_batch``). The original ``maybe_externalize`` only
+      inspected ``data`` and so silently let a 9× 2048² Flux2 batch
+      ride the wire as a 200 MB POST body — defeating the entire
+      reason the cap exists. The new path computes the summed raw
+      byte budget (``sum(len(b64)) * 3 // 4`` — base64 expands 4:3 so
+      this is the exact payload the server would receive) and, if it
+      crosses the cap, uploads each frame as an individual presigned
+      PUT in parallel via :func:`asyncio.gather` then rewrites the
+      envelope to ``encoding=png_base64_batch_uri`` with a parallel
+      ``uris`` list. ``shape`` / ``dtype`` / ``color_space`` ride
+      through unchanged so the server-side decoder still sees the
+      batch dimension. Per-frame URIs (NOT a tar/zip blob) compose
+      with the existing per-frame fetch helpers and let the server
+      fetch in parallel.
 
     No-op when either argument is falsy, the envelope already has a
-    ``uri``, or the inline payload fits under the cap.
+    ``uri`` / ``uris``, or the inline payload fits under the cap.
+
+    Dispatch keys off the ``encoding`` string — NOT field presence
+    (``frames`` vs. ``data``). An encoder that emitted ``frames`` for
+    a non-batch encoding by accident, or shipped both fields together,
+    used to silently take the batch branch and externalise garbage;
+    keying off the canonical encoding catches that as a hard ``frames``
+    check inside the batch branch itself.
     """
     if not server_url or not max_inline_bytes:
         return envelope
+
+    encoding = envelope.get("encoding")
+    if encoding == "png_base64_batch":
+        frames = envelope.get("frames")
+        if not isinstance(frames, list) or not frames:
+            # Mirror the server-side make_image_batch_envelope contract:
+            # an empty/missing frames list is a hard wire bug, NOT a
+            # silent no-op (which used to ship through under the old
+            # ``isinstance(frames, list) and frames`` dispatch and then
+            # blow up on the server's resolve_image_envelope_pngs check).
+            raise RnpProtocolError(
+                "png_base64_batch envelope requires a non-empty 'frames' list",
+                code=ErrorCode.INTERNAL,
+            )
+        return await _maybe_externalize_batch(
+            envelope, frames,
+            server_url=server_url,
+            max_inline_bytes=max_inline_bytes,
+            auth_headers=auth_headers,
+        )
+
     data = envelope.get("data")
     if not isinstance(data, str):
         return envelope
@@ -84,22 +135,122 @@ async def maybe_externalize(
     return out
 
 
+async def _maybe_externalize_batch(
+    envelope: dict[str, Any],
+    frames: list[Any],
+    *,
+    server_url: str,
+    max_inline_bytes: int,
+    auth_headers: dict[str, str] | None,
+) -> dict[str, Any]:
+    """Externalise a multi-frame IMAGE envelope when its summed inline
+    payload overflows ``max_inline_bytes``.
+
+    Per-frame upload (NOT a tar/zip blob): each frame becomes its own
+    presigned-PUT asset so the server can fetch in parallel and the
+    existing per-frame ``resolve_image_envelope_pngs`` helper picks
+    up the URI shape transparently. ``asyncio.gather`` issues every
+    PUT in parallel — the dominant latency for a 9-frame Flux2 batch
+    is the upload, not the server hop, so serialising would multiply
+    end-to-end submit time by N. Frame ordering is preserved (gather
+    returns positional results) — critical because a Flux2 provider
+    maps ``uris[k]`` → ``input_image_{k+1}`` 1:1.
+    """
+    # Cap on raw byte count (base64 expands 4:3, so multiplying by
+    # 3/4 recovers the actual payload the server would receive).
+    total_raw = 0
+    for f in frames:
+        if isinstance(f, str):
+            total_raw += len(f) * 3 // 4
+    if total_raw <= max_inline_bytes:
+        return envelope
+
+    env_type = envelope.get("type", "")
+    content_type = _CONTENT_TYPE_FOR_TYPE.get(env_type, "application/octet-stream")
+
+    async def _upload_one(idx: int, frame_b64: Any) -> str:
+        if not isinstance(frame_b64, str):
+            raise RnpProtocolError(
+                f"image batch frame {idx} is not a base64 string "
+                f"(got {type(frame_b64).__name__})",
+                code=ErrorCode.INTERNAL,
+            )
+        raw = base64.b64decode(frame_b64)
+        return await rnp_client.upload_asset(
+            server_url, raw,
+            content_type=content_type,
+            auth_headers=auth_headers,
+        )
+
+    uris = await asyncio.gather(
+        *(_upload_one(i, f) for i, f in enumerate(frames))
+    )
+    log.info(
+        "RNP: externalized %s batch envelope (%d frames, %d raw bytes) "
+        "-> %d presigned URIs",
+        env_type, len(frames), total_raw, len(uris),
+    )
+
+    # Strip the inline ``frames`` field; preserve ``shape`` / ``dtype``
+    # / ``color_space`` so the server-side decoder still sees the
+    # batch dimension and per-frame metadata. Encoding flips to the
+    # URI variant so dispatch keys cleanly off the string (NOT field
+    # presence — frames vs. uris is a payload field, not a shape
+    # selector).
+    out: dict[str, Any] = {
+        k: v for k, v in envelope.items() if k != "frames"
+    }
+    out["encoding"] = "png_base64_batch_uri"
+    out["uris"] = list(uris)
+    return out
+
+
 # ---------------------------------------------------------------------------
 # Image encode / decode
 # ---------------------------------------------------------------------------
 
-def encode_image_tensor(tensor: Any) -> dict[str, Any]:
-    """Encode a ComfyUI IMAGE tensor (B,H,W,C float32 in 0..1) as a
-    PNG-base64 image envelope.
+def encode_image_tensor(tensor: Any, *, accepts_batch: bool = False) -> dict[str, Any]:
+    """Encode a ComfyUI IMAGE tensor (B,H,W,C float32 in 0..1) as an
+    image envelope.
 
-    Multi-image batches send only the first frame for now; the
-    framework hasn't grown batch-of-envelopes semantics yet.
+    ``accepts_batch`` reflects whether the descriptor's
+    ``input_serialization`` for this input lists ``"png_base64_batch"``.
+    When True AND ``tensor.shape[0] > 1``, emit a multi-frame
+    ``png_base64_batch`` envelope (``frames=[<png_b64>, …]``) so the
+    server decodes a real ``[B, H, W, C]`` source. Otherwise emit a
+    single-frame ``png_base64`` envelope for ``tensor[0]`` — preserves
+    backwards compat with descriptors that haven't migrated yet, and
+    matches the wire shape of the existing ``_decode_image_batch_envelope``
+    on the response path so both directions round-trip symmetrically.
     """
     import numpy as np
     from PIL import Image
 
-    # Strip the batch dim if present.
-    if hasattr(tensor, "dim") and tensor.dim() == 4:
+    rank = getattr(tensor, "dim", lambda: 0)()
+    if rank == 4 and accepts_batch and int(tensor.shape[0]) > 1:
+        frames_arr = (tensor.cpu().clamp(0.0, 1.0).numpy() * 255).astype(np.uint8)
+        b, height, width = frames_arr.shape[:3]
+        channels = frames_arr.shape[3] if frames_arr.ndim == 4 else 1
+        pil_mode = {1: "L", 3: "RGB", 4: "RGBA"}.get(channels, "RGB")
+        frames_b64: list[str] = []
+        for i in range(b):
+            img = Image.fromarray(frames_arr[i], mode=pil_mode)
+            buf = BytesIO()
+            img.save(buf, format="PNG")
+            frames_b64.append(base64.b64encode(buf.getvalue()).decode("ascii"))
+        return {
+            "type":        "image",
+            "encoding":    "png_base64_batch",
+            "frames":      frames_b64,
+            "shape":       [b, height, width, channels],
+            "dtype":       "uint8",
+            "color_space": "srgb",
+        }
+
+    # Single-frame fallback: matches the old ``encode_image_tensor``
+    # contract (sends ``tensor[0]`` for a multi-frame batch when the
+    # descriptor doesn't advertise the batch encoding).
+    if rank == 4:
         frame = tensor[0]
     else:
         frame = tensor
@@ -121,8 +272,18 @@ def encode_image_tensor(tensor: Any) -> dict[str, Any]:
 
 
 def decode_image_envelope(envelope: dict[str, Any]) -> Any:
-    """Decode an image envelope into a ComfyUI IMAGE tensor (B,H,W,C float32)."""
+    """Decode an image envelope into a ComfyUI IMAGE tensor.
+
+    Single-frame encodings (``png_base64``, ``jpeg_base64``,
+    ``webp_base64``) decode into ``[1, H, W, C]``. The multi-frame
+    ``png_base64_batch`` encoding (a list of base64-PNG ``frames``)
+    stacks each frame into ``[B, H, W, C]`` — the native ComfyUI
+    image-batch shape — so providers whose upstream returns ``n > 1``
+    surface as a real batched tensor in the workflow runtime.
+    """
     encoding = envelope.get("encoding")
+    if encoding == "png_base64_batch":
+        return _decode_image_batch_envelope(envelope)
     if encoding not in ("png_base64", "jpeg_base64", "webp_base64"):
         raise RnpProtocolError(
             f"Unsupported image encoding: {encoding!r}",
@@ -141,6 +302,69 @@ def decode_image_envelope(envelope: dict[str, Any]) -> Any:
     return torch.from_numpy(arr).unsqueeze(0)
 
 
+def _decode_image_batch_envelope(envelope: dict[str, Any]) -> Any:
+    """Decode a ``png_base64_batch`` envelope into ``[B, H, W, C]``.
+
+    The wire shape is ``{"type":"image", "encoding":"png_base64_batch",
+    "frames":[<png_b64>, ...], "shape":[B,H,W,C], ...}``. Each frame is
+    a PNG-encoded image; all frames must share the same H/W after the
+    RGB conversion (the server-side encoder enforces this). A frame
+    count mismatch against the ``shape[0]`` field surfaces as a hard
+    parse error rather than silently dropping frames — matches the SVG
+    decoder's ``count`` integrity check.
+    """
+    import base64 as _b64
+    import numpy as np
+    import torch
+    from PIL import Image
+
+    frames = envelope.get("frames")
+    if not isinstance(frames, list) or not frames:
+        raise RnpProtocolError(
+            "image batch envelope missing or empty 'frames' list",
+            code=ErrorCode.INTERNAL,
+        )
+    declared_shape = envelope.get("shape")
+    if isinstance(declared_shape, list) and len(declared_shape) >= 1:
+        declared_b = int(declared_shape[0])
+        if declared_b != len(frames):
+            raise RnpProtocolError(
+                f"image batch envelope frame count {len(frames)} does not "
+                f"match declared shape[0]={declared_b}",
+                code=ErrorCode.INTERNAL,
+            )
+
+    arrs: list[Any] = []
+    base_hw: tuple[int, int] | None = None
+    for idx, frame_b64 in enumerate(frames):
+        if not isinstance(frame_b64, str):
+            raise RnpProtocolError(
+                f"image batch frame {idx} is not a base64 string",
+                code=ErrorCode.INTERNAL,
+            )
+        try:
+            raw = _b64.b64decode(frame_b64)
+            img = Image.open(BytesIO(raw)).convert("RGB")
+        except Exception as e:
+            raise RnpProtocolError(
+                f"image batch frame {idx} failed to decode: {e}",
+                code=ErrorCode.INTERNAL,
+            ) from e
+        arr = np.array(img).astype(np.float32) / 255.0
+        if base_hw is None:
+            base_hw = (arr.shape[0], arr.shape[1])
+        elif (arr.shape[0], arr.shape[1]) != base_hw:
+            raise RnpProtocolError(
+                f"image batch frame {idx} size {arr.shape[:2]} does not "
+                f"match frame 0 size {base_hw} (server should have rejected "
+                f"this; likely a wire bug)",
+                code=ErrorCode.INTERNAL,
+            )
+        arrs.append(arr)
+    stacked = np.stack(arrs, axis=0)
+    return torch.from_numpy(stacked)
+
+
 # ---------------------------------------------------------------------------
 # Mask encode / decode
 # ---------------------------------------------------------------------------
@@ -227,7 +451,11 @@ def encode_audio_input(audio: dict[str, Any]) -> dict[str, Any]:
 def decode_audio_envelope(envelope: dict[str, Any]) -> dict[str, Any]:
     """Decode an audio envelope back into a ComfyUI AUDIO dict."""
     encoding = envelope.get("encoding")
-    if encoding not in ("mp3_base64", "mp3_inline", "wav_base64", "wav_inline"):
+    if encoding not in (
+        "mp3_base64", "mp3_inline",
+        "wav_base64", "wav_inline",
+        "opus_base64",
+    ):
         raise RnpProtocolError(
             f"Unsupported audio encoding: {encoding!r}",
             code=ErrorCode.INTERNAL,
@@ -237,9 +465,44 @@ def decode_audio_envelope(envelope: dict[str, Any]) -> dict[str, Any]:
 
 
 # ---------------------------------------------------------------------------
-# Video decode (encode lands when a remote node accepts VIDEO inputs)
+# Video encode / decode
 # ---------------------------------------------------------------------------
 
+def encode_video_input(video: Any) -> dict[str, Any]:
+    """Encode a ComfyUI VIDEO (``VideoInput`` subclass) as an mp4-base64
+    video envelope.
+
+    Mirrors the AUDIO encoder above: re-encodes the Video object to an
+    in-memory mp4/H.264 byte buffer (matches upstream
+    ``comfy_api_nodes/util/conversions.py`` ``video_to_base64_string``)
+    and base64-encodes it. Populates ``duration_s`` metadata when the
+    Video object exposes ``get_duration()`` so server-side providers
+    (e.g. ``GrokVideoExtendProvider._validate_video_duration_envelope``)
+    can range-check without demuxing the MP4.
+    """
+    from comfy_api_nodes.util.conversions import video_to_base64_string
+
+    b64 = video_to_base64_string(video)
+    extra: dict[str, Any] = {}
+    duration_s: float | None = None
+    try:
+        getter = getattr(video, "get_duration", None)
+        if callable(getter):
+            d = getter()
+            if d is not None:
+                duration_s = float(d)
+    except Exception:
+        duration_s = None
+    if duration_s is not None:
+        extra["duration_s"] = duration_s
+    return {
+        "type":     "video",
+        "encoding": "mp4_base64",
+        "data":     b64,
+        **extra,
+    }
+
+
 def decode_video_envelope(envelope: dict[str, Any]) -> Any:
     """Decode a video envelope into a ComfyUI Video object (mp4 inline)."""
     encoding = envelope.get("encoding")
@@ -252,15 +515,391 @@ def decode_video_envelope(envelope: dict[str, Any]) -> Any:
     return InputImpl.VideoFromFile(BytesIO(decode_envelope_data(envelope)))
 
 
+# ---------------------------------------------------------------------------
+# 3D model encode / decode
+# ---------------------------------------------------------------------------
+
+def encode_model3d_input(model_3d: Any) -> dict[str, Any]:
+    """Encode a ComfyUI ``File3D`` (or compatible) as an inline
+    single-file 3D-model envelope.
+
+    Mirrors the AUDIO / VIDEO encoders above: reads the primary mesh
+    bytes (via the ``File3D`` API: ``get_data().read()`` returns a
+    ``BytesIO`` payload; falling back to ``read_bytes()`` for older
+    sources) and base64-encodes them. The envelope's ``encoding`` is
+    ``"glb_inline"`` (the generic single-file inline encoding the
+    server-side decoder dispatches on at :func:`decode_model3d_envelope`
+    — the decoder hands the bytes to ``File3D(BytesIO(...), format)``
+    regardless of whether the actual format is glb / obj / fbx / etc.,
+    so the ``encoding`` name is a logical "inline single-file 3D" tag
+    and the ``format`` extra picks the file extension).
+
+    Multi-file bundles (``BundledFile3D``-style values where every
+    companion file matters — e.g. an OBJ with a sibling .mtl + texture
+    PNGs) are not yet supported on the encode path: this PR ships the
+    minimal wire shape needed for the Tencent Hunyuan3D MODEL_3D-input
+    nodes which all upload a single file via ``upload_3d_model_to_comfyapi``
+    upstream. Bundle encode (``bundle_inline``) lands when a partner
+    node actually needs the companion bytes.
+    """
+    file_format = ""
+    fmt_attr = getattr(model_3d, "format", None)
+    if isinstance(fmt_attr, str):
+        file_format = fmt_attr.lower()
+
+    # ``File3D.get_data()`` returns a ``BytesIO`` (per
+    # ``comfy_api.latest._util.geometry_types.File3D``). Other File3D-
+    # compatible classes may instead expose ``read_bytes`` returning
+    # raw bytes; fall back to that for compatibility.
+    raw: bytes
+    get_data = getattr(model_3d, "get_data", None)
+    if callable(get_data):
+        buf = get_data()
+        if hasattr(buf, "seek"):
+            buf.seek(0)
+        raw = buf.read() if hasattr(buf, "read") else bytes(buf)
+    else:
+        read_bytes = getattr(model_3d, "read_bytes", None)
+        if callable(read_bytes):
+            raw = read_bytes()
+        else:
+            raise RnpProtocolError(
+                f"unsupported 3D model value: {type(model_3d).__name__!r} "
+                f"has no get_data() or read_bytes()",
+                code=ErrorCode.INTERNAL,
+            )
+
+    envelope: dict[str, Any] = {
+        "type":     "model_3d",
+        "encoding": "glb_inline",
+        "data":     base64.b64encode(raw).decode("ascii"),
+    }
+    if file_format:
+        envelope["format"] = file_format
+    return envelope
+
+
+# ---------------------------------------------------------------------------
+# Multi-file 3D bundle: a ``File3D``-compatible class that carries the
+# primary mesh plus N companion files (textures / .mtl / .bin / etc.).
+#
+# Built lazily as a ``File3D`` subclass on first use — importing
+# ``File3D`` at module load pulls torch (via ``comfy_api.latest._util.
+# geometry_types``), which is fine inside ComfyUI but breaks the
+# smoke-test pattern that exercises serialization without the heavy
+# comfy_api tree. The factory below caches the constructed class so
+# every ``BundledFile3D`` instance shares the same dynamic class and
+# ``isinstance(x, File3D)`` always succeeds.
+# ---------------------------------------------------------------------------
+
+_BUNDLED_FILE3D_CLS: Any = None
+
+
+def _bundled_file3d_class() -> type:
+    """Return the cached ``BundledFile3D`` class (a ``File3D`` subclass).
+
+    Constructs it on first call so the ``comfy_api.latest._util``
+    import (which pulls torch) is paid for only when a bundle
+    envelope is actually decoded — keeps the smoke-test stubbing
+    pattern intact for callers that don't exercise this path.
+    """
+    global _BUNDLED_FILE3D_CLS
+    if _BUNDLED_FILE3D_CLS is not None:
+        return _BUNDLED_FILE3D_CLS
+    from comfy_api.latest._util.geometry_types import File3D as _File3D
+
+    class BundledFile3D(_File3D):
+        """``File3D``-compatible wrapper for a multi-file 3D bundle.
+
+        Stores the primary mesh file plus N companion files (textures
+        / .mtl / .bin / etc.) in memory as a ``path -> bytes`` dict;
+        lazily materialises the whole bundle into a temp directory the
+        first time anything cross-file-resolution-sensitive is asked
+        of it (``get_source()`` for an OBJ/GLTF that needs its
+        .mtl/.bin/textures on disk; ``save_to(dest)`` which copies the
+        *whole bundle* into the destination directory).
+
+        Cheap consumers that just want the primary file's bytes
+        (``get_data()`` / ``get_bytes()``) skip the temp-dir
+        materialise entirely.
+
+        The temp directory's lifetime is tied to this object: it is
+        cleaned up on garbage-collect via ``__del__`` (best-effort).
+        For workflows that need stable on-disk paths beyond this
+        object's lifetime, call ``save_to(dest)`` to copy the bundle
+        into a caller-controlled location.
+        """
+
+        def __init__(
+            self,
+            files: dict[str, bytes],
+            primary_path: str,
+            roles: dict[str, str] | None = None,
+            formats: dict[str, str] | None = None,
+        ) -> None:
+            if primary_path not in files:
+                raise ValueError(
+                    f"primary_path {primary_path!r} not in bundle files: "
+                    f"{sorted(files)!r}"
+                )
+            self._files: dict[str, bytes] = dict(files)
+            self._primary_path = primary_path
+            self._roles: dict[str, str] = dict(roles or {})
+            self._formats: dict[str, str] = dict(formats or {})
+            self._temp_dir: str | None = None
+            # Populate parent ``_source`` / ``_format`` so reads of
+            # those plain attributes on downstream consumers that
+            # haven't been updated still work. ``_source`` is a
+            # BytesIO of the primary file; ``_format`` is the primary
+            # file's extension.
+            primary_fmt = (
+                self._formats.get(primary_path)
+                or (
+                    primary_path.rsplit(".", 1)[-1].lower()
+                    if "." in primary_path
+                    else ""
+                )
+            )
+            super().__init__(BytesIO(self._files[primary_path]), primary_fmt)
+
+        # --------------------------------------------------------------
+        # File3D-compatible accessors (overrides)
+        # --------------------------------------------------------------
+
+        def get_source(self) -> str:
+            """Return the on-disk path to the primary file.
+
+            Materialises the whole bundle into a temp directory on
+            first call so cross-file refs (.obj -> .mtl -> texture;
+            .gltf -> .bin + textures) resolve via the filesystem.
+            """
+            self._ensure_materialized()
+            from os.path import join as _join
+            return _join(self._temp_dir, *self._primary_path.split("/"))
+
+        def get_data(self) -> BytesIO:
+            """Return the primary file's bytes (no disk I/O)."""
+            return BytesIO(self._files[self._primary_path])
+
+        def get_bytes(self) -> bytes:
+            """Return the primary file's raw bytes (no disk I/O)."""
+            return self._files[self._primary_path]
+
+        def save_to(self, path: str) -> str:
+            """Copy the WHOLE bundle into the destination directory.
+
+            ``path`` is the destination for the *primary file*;
+            sidecars are copied alongside it preserving their relative
+            paths from the original bundle layout. E.g. for
+
+                primary_path = "model.obj"
+                files = {
+                    "model.obj":            <bytes>,
+                    "model.mtl":            <bytes>,
+                    "textures/albedo.png":  <bytes>,
+                }
+
+            ``save_to("/tmp/out/scene.obj")`` writes:
+
+                /tmp/out/scene.obj
+                /tmp/out/model.mtl
+                /tmp/out/textures/albedo.png
+
+            Sidecars keep their ORIGINAL filenames (so cross-file
+            refs inside the .obj that point at ``model.mtl`` still
+            resolve); only the primary file is renamed to match
+            ``path``. The destination directory is created if needed.
+            """
+            from pathlib import Path
+            dest = Path(path)
+            dest.parent.mkdir(parents=True, exist_ok=True)
+            dest.write_bytes(self._files[self._primary_path])
+            dest_parent_resolved = dest.parent.resolve()
+            for rel_path, raw in self._files.items():
+                if rel_path == self._primary_path:
+                    continue
+                sidecar_dest = dest.parent.joinpath(*rel_path.split("/"))
+                # Path safety — rel_path was sanitized at envelope
+                # build time, but defence-in-depth check on the
+                # resolved path keeps us safe against pathological
+                # inputs that may have slipped through.
+                resolved = sidecar_dest.resolve()
+                if dest_parent_resolved not in resolved.parents and resolved != dest_parent_resolved:
+                    continue
+                sidecar_dest.parent.mkdir(parents=True, exist_ok=True)
+                sidecar_dest.write_bytes(raw)
+            return str(dest)
+
+        # --------------------------------------------------------------
+        # Bundle-specific accessors (additive)
+        # --------------------------------------------------------------
+
+        @property
+        def primary_path(self) -> str:
+            return self._primary_path
+
+        @property
+        def file_paths(self) -> list[str]:
+            return sorted(self._files)
+
+        def file_bytes(self, path: str) -> bytes:
+            """Raw bytes for one file in the bundle by relative path."""
+            return self._files[path]
+
+        def file_role(self, path: str) -> str | None:
+            """Optional producer-supplied role hint (may be ``None``)."""
+            return self._roles.get(path)
+
+        def paths_with_role(self, role: str) -> list[str]:
+            """All file paths whose role hint equals ``role``."""
+            return [p for p, r in self._roles.items() if r == role]
+
+        def materialize_to_temp_dir(self) -> str:
+            """Force-materialise the whole bundle into a temp directory.
+
+            Returns the temp-directory path. Cleanup happens on GC;
+            callers that need stable lifetimes should ``save_to()`` to
+            a caller-controlled location instead.
+            """
+            self._ensure_materialized()
+            return self._temp_dir
+
+        # --------------------------------------------------------------
+        # Internal
+        # --------------------------------------------------------------
+
+        def _ensure_materialized(self) -> None:
+            if self._temp_dir is not None:
+                return
+            import tempfile
+            from pathlib import Path
+            self._temp_dir = tempfile.mkdtemp(prefix="rnp_model3d_bundle_")
+            root = Path(self._temp_dir)
+            for rel_path, raw in self._files.items():
+                # rel_path was sanitized at envelope build time —
+                # split-on-'/' and re-join via Path is safe.
+                dest = root.joinpath(*rel_path.split("/"))
+                dest.parent.mkdir(parents=True, exist_ok=True)
+                dest.write_bytes(raw)
+
+        def __del__(self) -> None:
+            td = getattr(self, "_temp_dir", None)
+            if td is not None:
+                try:
+                    import shutil
+                    shutil.rmtree(td, ignore_errors=True)
+                except Exception:  # noqa: BLE001
+                    pass
+
+        def __repr__(self) -> str:
+            return (
+                f"BundledFile3D(primary={self._primary_path!r}, "
+                f"files={len(self._files)}, format={self._format!r})"
+            )
+
+    _BUNDLED_FILE3D_CLS = BundledFile3D
+    return _BUNDLED_FILE3D_CLS
+
+
+def decode_model3d_envelope(envelope: dict[str, Any]) -> Any:
+    """Decode a 3D-model envelope into a ComfyUI ``File3D`` object.
+
+    Dispatches on the envelope's ``encoding``:
+
+    * ``glb_inline`` — single-file binary glTF 2.0 (magic ``b"glTF"``
+      at offset 0). The bytes are handed to ``File3D`` as a
+      ``BytesIO`` so downstream nodes (preview3d / SaveGLB / Load3D)
+      see exactly the same value shape they would from a local Tripo
+      / Rodin / Meshy node. The envelope's optional ``format`` extra
+      is forwarded as the second constructor arg so the materialiser
+      picks the right file extension when persisting.
+
+    * ``bundle_inline`` — multi-file bundle. The primary mesh file
+      plus N companion files (textures / .mtl / .bin / etc.) are
+      decoded into a :class:`BundledFile3D` (``File3D``-compatible)
+      whose ``get_source()`` / ``save_to()`` materialise the whole
+      bundle on disk so cross-file refs resolve via the filesystem.
+      Path safety (no ``..`` / no absolute prefix / no Windows drive
+      letter / no case-insensitive collisions) was enforced at
+      envelope build time on the server; the client re-asserts the
+      ``primary_path``-present invariant defensively and rejects any
+      file entry that lacks ``data``.
+    """
+    encoding = envelope.get("encoding")
+    if encoding == "glb_inline":
+        from comfy_api.latest._util import File3D
+        file_format = envelope.get("format") or "glb"
+        return File3D(BytesIO(decode_envelope_data(envelope)), str(file_format))
+    if encoding == "bundle_inline":
+        bundle_cls = _bundled_file3d_class()
+        primary_path = envelope.get("primary_path")
+        if not isinstance(primary_path, str) or not primary_path:
+            raise RnpProtocolError(
+                "bundle_inline envelope missing 'primary_path'",
+                code=ErrorCode.INTERNAL,
+            )
+        files_spec = envelope.get("files")
+        if not isinstance(files_spec, list) or not files_spec:
+            raise RnpProtocolError(
+                "bundle_inline envelope missing non-empty 'files' list",
+                code=ErrorCode.INTERNAL,
+            )
+        files: dict[str, bytes] = {}
+        roles: dict[str, str] = {}
+        formats: dict[str, str] = {}
+        for entry in files_spec:
+            if not isinstance(entry, dict):
+                raise RnpProtocolError(
+                    f"bundle_inline file entry not a dict: {entry!r}",
+                    code=ErrorCode.INTERNAL,
+                )
+            path = entry.get("path")
+            if not isinstance(path, str) or not path:
+                raise RnpProtocolError(
+                    f"bundle_inline file entry missing 'path': {entry!r}",
+                    code=ErrorCode.INTERNAL,
+                )
+            data = entry.get("data")
+            if not isinstance(data, str):
+                raise RnpProtocolError(
+                    f"bundle_inline file {path!r} missing inline 'data'",
+                    code=ErrorCode.INTERNAL,
+                )
+            files[path] = base64.b64decode(data)
+            role = entry.get("role")
+            if isinstance(role, str):
+                roles[path] = role
+            fmt = entry.get("format")
+            if isinstance(fmt, str):
+                formats[path] = fmt
+        if primary_path not in files:
+            raise RnpProtocolError(
+                f"bundle_inline primary_path {primary_path!r} not in "
+                f"files list (have: {sorted(files)!r})",
+                code=ErrorCode.INTERNAL,
+            )
+        return bundle_cls(
+            files=files,
+            primary_path=primary_path,
+            roles=roles,
+            formats=formats,
+        )
+    raise RnpProtocolError(
+        f"Unsupported model_3d encoding: {encoding!r}",
+        code=ErrorCode.INTERNAL,
+    )
+
+
 # ---------------------------------------------------------------------------
 # Dispatch by envelope type
 # ---------------------------------------------------------------------------
 
 _DECODERS = {
-    "image": decode_image_envelope,
-    "mask":  decode_mask_envelope,
-    "audio": decode_audio_envelope,
-    "video": decode_video_envelope,
+    "image":    decode_image_envelope,
+    "mask":     decode_mask_envelope,
+    "audio":    decode_audio_envelope,
+    "video":    decode_video_envelope,
+    "model_3d": decode_model3d_envelope,
 }
 
 
@@ -293,6 +932,47 @@ def is_audio_input(value: Any) -> bool:
     )
 
 
+def is_video_input(value: Any) -> bool:
+    """Best-effort check for a ComfyUI ``VideoInput`` subclass.
+
+    Duck-types on the ``save_to`` + ``get_duration`` method pair: both
+    are defined on the ``VideoInput`` abstract base in
+    ``comfy_api.latest._input`` and present on every concrete subclass
+    (``VideoFromFile`` / ``VideoFromComponents``). Avoids importing
+    ``VideoInput`` at module load (which would pull torch via the
+    ``comfy_api`` tree).
+    """
+    if isinstance(value, dict):
+        return False
+    return (
+        callable(getattr(value, "save_to", None))
+        and callable(getattr(value, "get_duration", None))
+    )
+
+
+def is_model3d_input(value: Any) -> bool:
+    """Best-effort check for a ComfyUI ``File3D`` (or compatible).
+
+    Duck-types on the ``format`` string attribute + ``get_data``
+    callable: both are present on the ``File3D`` class in
+    ``comfy_api.latest._util.geometry_types`` and the
+    ``BundledFile3D`` subclass built lazily in
+    :func:`_bundled_file3d_class`. Avoids importing ``File3D`` at
+    module load (which would pull torch via the ``comfy_api`` tree).
+
+    The ``VideoInput`` shape (``save_to`` + ``get_duration``) does not
+    expose a ``format`` string property, so VIDEO and MODEL_3D values
+    don't collide; caller (:func:`_encode_one`) still dispatches
+    AUDIO / VIDEO before MODEL_3D as defense-in-depth.
+    """
+    if isinstance(value, dict):
+        return False
+    return (
+        isinstance(getattr(value, "format", None), str)
+        and callable(getattr(value, "get_data", None))
+    )
+
+
 def _is_torch_tensor(value: Any) -> bool:
     try:
         import torch
@@ -309,10 +989,16 @@ __all__ = [
     "decode_mask_envelope",
     "encode_audio_input",
     "decode_audio_envelope",
+    "encode_video_input",
     "decode_video_envelope",
+    "encode_model3d_input",
+    "decode_model3d_envelope",
+    "_bundled_file3d_class",
     "decode_envelope",
     "is_envelope",
     "is_image_tensor",
     "is_mask_tensor",
     "is_audio_input",
+    "is_video_input",
+    "is_model3d_input",
 ]