langgraph 1.2.2__tar.gz → 1.2.3__tar.gz

{langgraph-1.2.2 → langgraph-1.2.3}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: langgraph
-Version: 1.2.2
+Version: 1.2.3
 Summary: Building stateful, multi-actor applications with LLMs
 Project-URL: Homepage, https://docs.langchain.com/oss/python/langgraph/overview
 Project-URL: Documentation, https://reference.langchain.com/python/langgraph/
@@ -25,7 +25,7 @@ Requires-Python: >=3.10
 Requires-Dist: langchain-core<2,>=1.4.0
 Requires-Dist: langgraph-checkpoint<5.0.0,>=4.1.0
 Requires-Dist: langgraph-prebuilt<1.2.0,>=1.1.0
-Requires-Dist: langgraph-sdk<0.4.0,>=0.3.0
+Requires-Dist: langgraph-sdk<0.5.0,>=0.4.2
 Requires-Dist: pydantic>=2.7.4
 Requires-Dist: xxhash>=3.5.0
 Description-Content-Type: text/markdown

{langgraph-1.2.2 → langgraph-1.2.3}/langgraph/_internal/_config.py

@@ -79,6 +79,34 @@ def patch_checkpoint_map(
         return config
 
 
+def _merge_callbacks(base: Callbacks, new: Callbacks) -> Callbacks:
+    """Merge two callbacks values (None / list / BaseCallbackManager).
+
+    Six cases total (3 base types x 2 non-None new types).
+    """
+    if new is None:
+        return base
+    if base is None:
+        return new.copy() if isinstance(new, (list, BaseCallbackManager)) else new
+    if isinstance(new, list):
+        if isinstance(base, list):
+            return base + new
+        if isinstance(base, BaseCallbackManager):
+            mngr = base.copy()
+            for cb in new:
+                mngr.add_handler(cb, inherit=True)
+            return mngr
+    elif isinstance(new, BaseCallbackManager):
+        if isinstance(base, list):
+            mngr = new.copy()
+            for cb in base:
+                mngr.add_handler(cb, inherit=True)
+            return mngr
+        if isinstance(base, BaseCallbackManager):
+            return base.merge(new)
+    raise NotImplementedError(f"Unsupported callback types: {type(base)}, {type(new)}")
+
+
 def merge_configs(*configs: RunnableConfig | None) -> RunnableConfig:
     """Merge multiple configs into one.
 
@@ -113,34 +141,9 @@ def merge_configs(*configs: RunnableConfig | None) -> RunnableConfig:
                 else:
                     base[key] = value
             elif key == "callbacks":
-                base_callbacks = base.get("callbacks")
-                # callbacks can be either None, list[handler] or manager
-                # so merging two callbacks values has 6 cases
-                if isinstance(value, list):
-                    if base_callbacks is None:
-                        base["callbacks"] = value.copy()
-                    elif isinstance(base_callbacks, list):
-                        base["callbacks"] = base_callbacks + value
-                    else:
-                        # base_callbacks is a manager
-                        mngr = base_callbacks.copy()
-                        for callback in value:
-                            mngr.add_handler(callback, inherit=True)
-                        base["callbacks"] = mngr
-                elif isinstance(value, BaseCallbackManager):
-                    # value is a manager
-                    if base_callbacks is None:
-                        base["callbacks"] = value.copy()
-                    elif isinstance(base_callbacks, list):
-                        mngr = value.copy()
-                        for callback in base_callbacks:
-                            mngr.add_handler(callback, inherit=True)
-                        base["callbacks"] = mngr
-                    else:
-                        # base_callbacks is also a manager
-                        base["callbacks"] = base_callbacks.merge(value)
-                else:
-                    raise NotImplementedError
+                base["callbacks"] = _merge_callbacks(
+                    base.get("callbacks"), cast(Callbacks, value)
+                )
             elif key == "recursion_limit":
                 if config["recursion_limit"] != DEFAULT_RECURSION_LIMIT:
                     base["recursion_limit"] = config["recursion_limit"]
@@ -309,7 +312,40 @@ def ensure_config(*configs: RunnableConfig | None) -> RunnableConfig:
         for k, v in config.items():
             if _is_not_empty(v) and k in CONFIG_KEYS:
                 if k == CONF:
-                    empty[k] = cast(dict, v).copy()
+                    # Shallow-merge configurable dicts across configs so values
+                    # bound via with_config(...) (e.g. ls_agent_type) are
+                    # preserved when later configs (e.g. invoke-time) only
+                    # specify a subset of keys like thread_id.
+                    existing = empty.get(k)
+                    empty[k] = (
+                        {**cast(dict, existing), **cast(dict, v)}
+                        if existing
+                        else cast(dict, v).copy()
+                    )
+                elif k == "callbacks":
+                    empty["callbacks"] = _merge_callbacks(
+                        empty.get("callbacks"), cast(Callbacks, v)
+                    )
+                elif k == "metadata":
+                    # Shallow-merge metadata dicts across configs so values
+                    # bound via with_config(...) (e.g. user_id) are preserved
+                    # when later configs supply other metadata keys.
+                    existing = empty.get("metadata")
+                    empty["metadata"] = (
+                        {**cast(dict, existing), **cast(dict, v)}
+                        if existing
+                        else cast(dict, v).copy()
+                    )
+                elif k == "tags":
+                    # Concatenate tags across configs so values bound via
+                    # with_config(...) are preserved when later configs
+                    # supply additional tags. Matches merge_configs.
+                    existing_tags: list[str] | None = empty.get("tags")
+                    empty["tags"] = (
+                        [*existing_tags, *cast(list, v)]
+                        if existing_tags
+                        else list(cast(list, v))
+                    )
                 else:
                     empty[k] = v  # type: ignore[literal-required]
         for k, v in config.items():
@@ -366,3 +402,17 @@ _PROPAGATE_TO_METADATA = frozenset(
         "graph_id",
     )
 )
+
+
+def filter_to_user_tags(tags: Sequence[str] | None) -> list[str] | None:
+    """Drop langgraph's internal `seq:step:*` bookkeeping tags.
+
+    `seq:step:N` tags are added internally to mark sequence steps; everything
+    else (user-supplied tags and any other framework tags) is kept. Returns the
+    surviving tags, or `None` if none remain. Shared by the `messages` and
+    `tasks` stream handlers so both surface the same tag set on their metadata.
+    """
+    if not tags:
+        return None
+    filtered = [t for t in tags if not t.startswith("seq:step")]
+    return filtered or None

{langgraph-1.2.2 → langgraph-1.2.3}/langgraph/errors.py

@@ -21,6 +21,7 @@ __all__ = (
     "InvalidUpdateError",
     "GraphBubbleUp",
     "GraphInterrupt",
+    "NodeCancelledError",
     "NodeError",
     "NodeInterrupt",
     "NodeTimeoutError",
@@ -164,6 +165,28 @@ class NodeError:
     """Exception raised by the failed node."""
 
 
+class NodeCancelledError(Exception):
+    """Raised when a node body raises ``asyncio.CancelledError`` itself.
+
+    ``asyncio.CancelledError`` is a ``BaseException`` and the pregel runner
+    treats cancelled task futures as silent tear-down (e.g. when it stops
+    sibling tasks after a peer fails). That is the correct behaviour for
+    *framework-initiated* cancellation, but a user node that raises
+    ``asyncio.CancelledError`` from its own body should surface as a node
+    failure, the same way any other exception would.
+
+    The retry layer converts user-raised ``asyncio.CancelledError`` into this
+    type so it flows through the normal error path and the run reports as
+    ``error`` instead of silently succeeding.
+    """
+
+    node: str
+
+    def __init__(self, node: str, message: str | None = None) -> None:
+        super().__init__(message or f"Node {node!r} raised asyncio.CancelledError")
+        self.node = node
+
+
 class NodeTimeoutError(Exception):
     """Raised when a node invocation exceeds one of its configured timeouts.
 

{langgraph-1.2.2 → langgraph-1.2.3}/langgraph/pregel/_messages.py

@@ -15,6 +15,7 @@ from langchain_core.messages.utils import convert_to_messages
 from langchain_core.outputs import ChatGeneration, ChatGenerationChunk, LLMResult
 from pydantic import BaseModel
 
+from langgraph._internal._config import filter_to_user_tags
 from langgraph._internal._constants import NS_SEP
 from langgraph.constants import TAG_HIDDEN, TAG_NOSTREAM
 from langgraph.pregel.protocol import StreamChunk
@@ -143,9 +144,8 @@ class StreamMessagesHandler(BaseCallbackHandler, _StreamingCallbackHandler):
             ]
             if not self.subgraphs and len(ns) > 0 and ns != self.parent_ns:
                 return
-            if tags:
-                if filtered_tags := [t for t in tags if not t.startswith("seq:step")]:
-                    metadata["tags"] = filtered_tags
+            if (filtered_tags := filter_to_user_tags(tags)) is not None:
+                metadata["tags"] = filtered_tags
             self.metadata[run_id] = (ns, metadata)
 
     def on_llm_new_token(

langgraph-1.2.3/langgraph/pregel/_remote_run_stream.py

@@ -0,0 +1,374 @@
+# libs/langgraph/langgraph/pregel/_remote_run_stream.py
+from __future__ import annotations
+
+import logging
+import sys
+from collections.abc import AsyncIterator, Iterator, Mapping
+from types import TracebackType
+from typing import Any, cast
+
+from langchain_core.runnables import RunnableConfig
+from langgraph_sdk._async.stream import AsyncThreadStream
+from langgraph_sdk._sync.stream import SyncThreadStream
+from langgraph_sdk.client import LangGraphClient, SyncLangGraphClient
+from langgraph_sdk.stream.decoders import DataDecoder
+
+from langgraph.types import Command
+
+logger = logging.getLogger(__name__)
+
+
+def _translate_command_input(input: Any) -> Any:
+    """Translate a local `Command` into the v3 wire `input`, else passthrough.
+
+    The v3 server decides start-vs-resume from thread state (an interrupted
+    run or pending interrupts) and, on resume, wraps the whole `input` as
+    `{"resume": input}` itself. So a resume `Command` must surface its raw
+    `resume` value as the wire `input` (not the serialized dataclass, which
+    the server would double-wrap). The v3 `run.start` path has no `goto` /
+    `update` channel, so those are rejected.
+
+    `langgraph_sdk` is upstream of `langgraph`, so this `Command`-aware
+    marshalling lives here on the adapter (langgraph) side of the boundary.
+    """
+    if isinstance(input, Command):
+        if input.goto or input.update:
+            raise NotImplementedError(
+                "RemoteGraph v3 streaming supports `Command(resume=...)` only; "
+                "`goto` / `update` are not supported by the v3 `run.start` path."
+            )
+        return input.resume
+    return input
+
+
+class _ChannelProjection:
+    """Decoded projection for a wire channel the SDK doesn't type natively.
+
+    Subscribes to `channel` and decodes each event's `params["data"]` through the
+    SDK's `DataDecoder` — the same decoder the SDK's own plain-payload projections
+    (`values` / `updates` / `checkpoints` / `tasks`) use, which yields the item
+    shape that local's `UpdatesTransformer` / `CheckpointsTransformer` /
+    `TasksTransformer` / `CustomTransformer` push, so iterating this matches the
+    corresponding local projection. Iterate with `for` against a sync stream and
+    `async for` against an async stream (matching the underlying SDK). Opening
+    the subscription requires the stream to be entered (`with` / `async with`).
+    """
+
+    def __init__(self, sdk: AsyncThreadStream | SyncThreadStream, channel: str) -> None:
+        self._sdk = sdk
+        self._channel = channel
+
+    def __iter__(self) -> Iterator[Any]:
+        # Sync lane: the sync adapter's SDK returns a sync iterator here.
+        decoder = DataDecoder(self._channel)
+        events = cast(Iterator[Any], self._sdk.subscribe([self._channel]))
+        for event in events:
+            yield from decoder.feed(event)
+
+    def __aiter__(self) -> AsyncIterator[Any]:
+        return self._aiter()
+
+    async def _aiter(self) -> AsyncIterator[Any]:
+        # Async lane: the async adapter's SDK returns an async iterator here.
+        decoder = DataDecoder(self._channel)
+        events = cast(AsyncIterator[Any], self._sdk.subscribe([self._channel]))
+        async for event in events:
+            for item in decoder.feed(event):
+                yield item
+
+
+class _ProjectionRegistry(Mapping[str, Any]):
+    """Read-only name -> projection registry mirroring local `GraphRunStream.extensions`.
+
+    Resolution follows the langchain-protocol wire channels, and every entry
+    yields the same decoded item shape local does (`params.data`):
+
+    - `values` / `messages` / `tool_calls` / `subgraphs` resolve to the SDK's
+      decoded typed projections. `tool_calls` is the `tools` channel — tool
+      *execution* events, distinct from the tool-call *inputs* inside `messages`.
+    - `updates` / `checkpoints` / `tasks` / `custom` have no typed SDK
+      projection, so they resolve to a `_ChannelProjection` that subscribes to
+      the channel and yields `params.data` — matching the local transformer
+      output for those channels.
+    - any other name is a specific custom-extension channel
+      (`thread.extensions[name]`, i.e. `custom:<name>`).
+
+    `lifecycle` is intentionally absent: local derives a status payload from it
+    rather than yielding `params.data`, and the SDK consumes it as control-plane
+    (driving `output` / `interrupted`), so its shape can't be matched — it
+    remains reachable via the raw `events` iterator. `debug` is absent too: it
+    is not a v3 wire channel.
+    """
+
+    # Channels the SDK decodes into typed projections.
+    _TYPED = ("values", "messages", "tool_calls", "subgraphs")
+    # Wire channels with no typed SDK projection — decoded here to match local.
+    _DECODED = ("updates", "checkpoints", "tasks", "custom")
+    _NATIVE = _TYPED + _DECODED
+
+    def __init__(self, sdk: AsyncThreadStream | SyncThreadStream) -> None:
+        self._sdk = sdk
+
+    def __getitem__(self, name: str) -> Any:
+        if name in self._TYPED:
+            return getattr(self._sdk, name)
+        if name in self._DECODED:
+            return _ChannelProjection(self._sdk, name)
+        return self._sdk.extensions[name]
+
+    def __iter__(self) -> Iterator[str]:
+        return iter(self._NATIVE)
+
+    def __len__(self) -> int:
+        return len(self._NATIVE)
+
+
+class _RemoteGraphRunStream:
+    """Sync adapter: SyncThreadStream -> GraphRunStream surface."""
+
+    def __init__(
+        self,
+        *,
+        sync_client: SyncLangGraphClient,
+        sdk_thread: SyncThreadStream,
+        input: Any,
+        config: RunnableConfig | None,
+        metadata: dict[str, Any] | None,
+    ) -> None:
+        self._client = sync_client
+        self._sdk = sdk_thread
+        self._start_kwargs: dict[str, Any] = {
+            "input": _translate_command_input(input),
+            "config": config,
+            "metadata": metadata,
+        }
+        self._run_id: str | None = None
+        self._closed = False
+        self._events_iter: Iterator[Any] | None = None
+
+    def __enter__(self) -> _RemoteGraphRunStream:
+        if self._closed:
+            raise RuntimeError("_RemoteGraphRunStream already closed")
+        self._sdk.__enter__()
+        try:
+            result = self._sdk.run.start(**self._start_kwargs)
+        except BaseException:
+            self._sdk.__exit__(*sys.exc_info())
+            raise
+        self._run_id = result["run_id"]
+        return self
+
+    def __exit__(
+        self,
+        exc_type: type[BaseException] | None,
+        exc: BaseException | None,
+        tb: TracebackType | None,
+    ) -> None:
+        if self._closed:
+            return
+        self._closed = True
+        self._sdk.__exit__(exc_type, exc, tb)
+
+    @property
+    def output(self) -> Any:
+        return self._sdk.output
+
+    @property
+    def interrupted(self) -> bool:
+        """Whether the remote run is currently paused at an interrupt.
+
+        Reads the SDK's current value without blocking. This differs from
+        local `GraphRunStream.interrupted`, which drives the run to terminal
+        before returning the flag. Sync callers needing a wait-for-interrupt
+        pattern should switch to the async API and drain a projection.
+        """
+        return self._sdk.interrupted
+
+    @property
+    def interrupts(self) -> list[Any]:
+        """Current outstanding interrupt payloads (non-blocking snapshot)."""
+        return list(self._sdk.interrupts)
+
+    @property
+    def values(self) -> Any:
+        """Live state-snapshot projection (mirrors local `run.values`)."""
+        return self._sdk.values
+
+    @property
+    def messages(self) -> Any:
+        """Live message-stream projection (mirrors local `run.messages`)."""
+        return self._sdk.messages
+
+    @property
+    def subgraphs(self) -> Any:
+        """Subgraph-handle projection (mirrors local `run.subgraphs`)."""
+        return self._sdk.subgraphs
+
+    @property
+    def tool_calls(self) -> Any:
+        """Tool-execution projection (the `tools` channel).
+
+        These are tool *execution* events (started / output / finished),
+        distinct from the tool-call *inputs* carried inside `messages`.
+        """
+        return self._sdk.tool_calls
+
+    @property
+    def extensions(self) -> Mapping[str, Any]:
+        """Name -> projection registry (mirrors local `run.extensions`)."""
+        return _ProjectionRegistry(self._sdk)
+
+    def abort(self) -> None:
+        if self._closed:
+            return
+        self._closed = True
+        if self._run_id is not None:
+            try:
+                self._client.runs.cancel(self._sdk.thread_id, self._run_id, wait=False)
+            except Exception:
+                logger.debug("abort: runs.cancel failed", exc_info=True)
+        try:
+            self._sdk.close()
+        except Exception:
+            logger.debug("abort: sdk.close failed", exc_info=True)
+
+    def __iter__(self) -> Iterator[Any]:
+        if self._events_iter is None:
+            self._events_iter = iter(self._sdk.events)
+        return self._events_iter
+
+    def interleave(self, *names: str) -> Iterator[tuple[str, Any]]:
+        yield from self._sdk.interleave_projections(list(names))
+
+
+class _AsyncRemoteGraphRunStream:
+    """Async adapter: AsyncThreadStream -> AsyncGraphRunStream surface."""
+
+    def __init__(
+        self,
+        *,
+        client: LangGraphClient,
+        sdk_thread: AsyncThreadStream,
+        input: Any,
+        config: RunnableConfig | None,
+        metadata: dict[str, Any] | None,
+    ) -> None:
+        self._client = client
+        self._sdk = sdk_thread
+        self._start_kwargs: dict[str, Any] = {
+            "input": _translate_command_input(input),
+            "config": config,
+            "metadata": metadata,
+        }
+        self._run_id: str | None = None
+        self._closed = False
+        self._events_aiter: AsyncIterator[Any] | None = None
+
+    async def __aenter__(self) -> _AsyncRemoteGraphRunStream:
+        if self._closed:
+            raise RuntimeError("_AsyncRemoteGraphRunStream already closed")
+        await self._sdk.__aenter__()
+        try:
+            result = await self._sdk.run.start(**self._start_kwargs)
+        except BaseException:
+            await self._sdk.__aexit__(*sys.exc_info())
+            raise
+        self._run_id = result["run_id"]
+        return self
+
+    async def __aexit__(
+        self,
+        exc_type: type[BaseException] | None,
+        exc: BaseException | None,
+        tb: TracebackType | None,
+    ) -> None:
+        if self._closed:
+            return
+        self._closed = True
+        await self._sdk.__aexit__(exc_type, exc, tb)
+
+    async def output(self) -> Any:
+        """Drive the remote run to completion and return the final state.
+
+        Awaits the SDK's terminal-state awaitable, matching local
+        `AsyncGraphRunStream.output()` (a method, not a property, so
+        `run.output` without `await` fails at type-check time rather than
+        silently yielding a coroutine).
+        """
+        return await self._sdk.output
+
+    async def interrupted(self) -> bool:
+        """Whether the remote run is currently paused at an interrupt.
+
+        Reads the SDK's current value without blocking. This differs from
+        local `AsyncGraphRunStream.interrupted()`, which drives the run to
+        terminal before returning the flag. Callers that need a
+        wait-for-interrupt pattern should drain a projection (e.g.,
+        `async for snap in stream._sdk.values`) until the SDK's paused
+        sentinel fires, then call this method.
+        """
+        return self._sdk.interrupted
+
+    async def interrupts(self) -> list[Any]:
+        """Current outstanding interrupt payloads.
+
+        Non-blocking; reads the SDK's current snapshot. See `interrupted`
+        for the divergence from local v3 semantics.
+        """
+        return list(self._sdk.interrupts)
+
+    @property
+    def values(self) -> Any:
+        """Live state-snapshot projection (mirrors local `run.values`)."""
+        return self._sdk.values
+
+    @property
+    def messages(self) -> Any:
+        """Live message-stream projection (mirrors local `run.messages`)."""
+        return self._sdk.messages
+
+    @property
+    def subgraphs(self) -> Any:
+        """Subgraph-handle projection (mirrors local `run.subgraphs`)."""
+        return self._sdk.subgraphs
+
+    @property
+    def tool_calls(self) -> Any:
+        """Tool-execution projection (the `tools` channel).
+
+        These are tool *execution* events (started / output / finished),
+        distinct from the tool-call *inputs* carried inside `messages`.
+        """
+        return self._sdk.tool_calls
+
+    @property
+    def extensions(self) -> Mapping[str, Any]:
+        """Name -> projection registry (mirrors local `run.extensions`)."""
+        return _ProjectionRegistry(self._sdk)
+
+    async def abort(self) -> None:
+        if self._closed:
+            return
+        self._closed = True
+        if self._run_id is not None:
+            try:
+                await self._client.runs.cancel(
+                    self._sdk.thread_id, self._run_id, wait=False
+                )
+            except Exception:
+                logger.debug("abort: runs.cancel failed", exc_info=True)
+        try:
+            await self._sdk.close()
+        except Exception:
+            logger.debug("abort: sdk.close failed", exc_info=True)
+
+    def __aiter__(self) -> AsyncIterator[Any]:
+        if self._events_aiter is None:
+            self._events_aiter = self._sdk.events.__aiter__()
+        return self._events_aiter
+
+    # Note: deliberately no `interleave()` on the async adapter. Local
+    # `AsyncGraphRunStream` doesn't have one either (async callers compose
+    # with `asyncio.gather` / `asyncio.as_completed`). The sync adapter
+    # provides `interleave()` because sync callers have no comparable
+    # primitive for iterating multiple iterators concurrently.

{langgraph-1.2.2 → langgraph-1.2.3}/langgraph/pregel/_retry.py

@@ -37,13 +37,23 @@ from langgraph._internal._constants import (
 )
 from langgraph._internal._runnable import create_task_in_config_context
 from langgraph._internal._timeout import sync_timeout_unsupported
-from langgraph.errors import GraphBubbleUp, NodeTimeoutError, ParentCommand
+from langgraph.errors import (
+    GraphBubbleUp,
+    NodeCancelledError,
+    NodeTimeoutError,
+    ParentCommand,
+)
 from langgraph.pregel.protocol import StreamProtocol
 from langgraph.runtime import ExecutionInfo, Runtime
 from langgraph.types import Command, PregelExecutableTask, RetryPolicy, TimeoutPolicy
 
 logger = logging.getLogger(__name__)
 SUPPORTS_EXC_NOTES = sys.version_info >= (3, 11)
+# `asyncio.Task.cancelling()` was added in Python 3.11. It reports the number of
+# pending cancel requests on the task: ``0`` means no external code asked us to
+# cancel — so a ``CancelledError`` observed here was raised by the task body
+# itself (the user's node) rather than by pregel cancelling sibling tasks.
+SUPPORTS_TASK_CANCELLING = sys.version_info >= (3, 11)
 
 
 def _timeout_secs(value: float | timedelta) -> float:
@@ -302,6 +312,28 @@ class _IdleProgressCallbackHandler(BaseCallbackHandler):
     on_custom_event = _touch
 
 
+def _is_user_raised_cancelled() -> bool:
+    """Return True if the in-flight ``CancelledError`` came from the task body.
+
+    Pregel cancels sibling tasks via ``task.cancel()`` when a peer fails, which
+    increments ``asyncio.Task.cancelling()`` on the target before the cancel
+    actually fires. A user node that calls ``raise asyncio.CancelledError()``
+    from inside its own body raises while ``cancelling() == 0``, which is the
+    signal we use to convert the exception into a regular
+    :class:`NodeCancelledError`.
+
+    Returns ``False`` when we can't tell (``cancelling()`` unavailable, or no
+    current task — neither should happen in practice from ``arun_with_retry``)
+    so framework-initiated cancellation continues to propagate unchanged.
+    """
+    if not SUPPORTS_TASK_CANCELLING:
+        return False
+    current = asyncio.current_task()
+    if current is None:
+        return False
+    return current.cancelling() == 0
+
+
 def _drain_cancelled(task: asyncio.Task[Any]) -> None:
     # Mark the abandoned task's exception as retrieved so asyncio doesn't log it.
     with suppress(asyncio.CancelledError):
@@ -600,6 +632,12 @@ def run_with_retry(
         except GraphBubbleUp:
             # if interrupted, end
             raise
+        except asyncio.CancelledError as exc:
+            # A sync node has no asyncio context, so any ``CancelledError`` that
+            # reaches here was raised by the node body itself. Surface it as a
+            # regular exception so the pregel runner panics the run instead of
+            # treating the task as a silent tear-down (LSD-1507).
+            raise NodeCancelledError(task.name) from exc
         except Exception as exc:
             if SUPPORTS_EXC_NOTES:
                 exc.add_note(f"During task with name '{task.name}' and id '{task.id}'")
@@ -736,6 +774,24 @@ async def arun_with_retry(
             # if interrupted, end
             _finish_timed_attempt(config, attempt_ctx)
             raise
+        except asyncio.CancelledError as exc:
+            # ``CancelledError`` reaches us in two very different shapes:
+            #   1. Pregel cancelled this task because a sibling failed
+            #      (``asyncio.Task.cancelling() >= 1``). The framework already
+            #      knows the run is failing and we must let cancellation
+            #      propagate so the watchdog/cleanup code in the runner sees a
+            #      cancelled future.
+            #   2. The node body itself raised ``asyncio.CancelledError`` (
+            #      ``cancelling() == 0``). The runner would otherwise treat
+            #      this as silent tear-down and the run would report
+            #      ``success`` even though the node failed (LSD-1507). Convert
+            #      it into :class:`NodeCancelledError` so it follows the same
+            #      path as any other node failure.
+            if _is_user_raised_cancelled():
+                _finish_timed_attempt(config, attempt_ctx, exc)
+                raise NodeCancelledError(task.name) from exc
+            _finish_timed_attempt(config, attempt_ctx, exc)
+            raise
         except Exception as exc:
             _finish_timed_attempt(config, attempt_ctx, exc)
             if SUPPORTS_EXC_NOTES: