fix(cli): include build environment hash in Docker cache keys

libs/checkpoint-conformance/uv.lock

@@ -279,7 +279,7 @@ wheels = [
 
 [[package]]
 name = "langgraph-checkpoint"
-version = "4.3.1"
+version = "4.3.0"
 source = { editable = "../checkpoint" }
 dependencies = [
     { name = "langchain-core" },

libs/checkpoint-postgres/uv.lock

@@ -278,7 +278,7 @@ wheels = [
 
 [[package]]
 name = "langgraph-checkpoint"
-version = "4.3.1"
+version = "4.3.0"
 source = { editable = "../checkpoint" }
 dependencies = [
     { name = "langchain-core" },

libs/checkpoint-sqlite/uv.lock

@@ -285,7 +285,7 @@ wheels = [
 
 [[package]]
 name = "langgraph-checkpoint"
-version = "4.3.1"
+version = "4.3.0"
 source = { editable = "../checkpoint" }
 dependencies = [
     { name = "langchain-core" },

libs/checkpoint/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "hatchling.build"
 
 [project]
 name = "langgraph-checkpoint"
-version = "4.3.1"
+version = "4.3.0"
 description = "Library with base interfaces for LangGraph checkpoint savers."
 authors = []
 requires-python = ">=3.10"

libs/checkpoint/uv.lock

@@ -301,7 +301,7 @@ wheels = [
 
 [[package]]
 name = "langgraph-checkpoint"
-version = "4.3.1"
+version = "4.3.0"
 source = { editable = "." }
 dependencies = [
     { name = "langchain-core" },

libs/cli/langgraph_cli/config.py

@@ -1589,6 +1589,8 @@ ADD {relpath} /deps/{name}
     docker_file_contents.extend(
         [
             f"FROM {image_str}",
+            # BuildKit secrets do not invalidate RUN caches; hosted builds supply this hash.
+            "ARG USER_ENV_VARS_HASH",
             "",
             os.linesep.join(config["dockerfile_lines"]),
             "",
@@ -1679,6 +1681,8 @@ def node_config_to_docker(
 
     docker_file_contents = [
         f"FROM {image_str}",
+        # BuildKit secrets do not invalidate RUN caches; hosted builds supply this hash.
+        "ARG USER_ENV_VARS_HASH",
         "",
         os.linesep.join(config["dockerfile_lines"]),
         "",

libs/cli/langgraph_cli/uv_lock.py

@@ -955,6 +955,8 @@ def python_config_to_docker_uv_lock(
         docker_plan.add_blank()
 
     docker_plan.add_instruction("FROM", image_str)
+    # BuildKit secrets do not invalidate RUN caches; hosted builds supply this hash.
+    docker_plan.add_instruction("ARG", "USER_ENV_VARS_HASH")
     docker_plan.add_blank()
     docker_plan.extend_nonempty(config["dockerfile_lines"])
     if config["dockerfile_lines"]:

libs/cli/tests/unit_tests/cli/test_cli.py

@@ -148,6 +148,7 @@ services:
             dockerfile_inline: |
                 # syntax=docker/dockerfile:1.4
                 FROM langchain/langgraph-api:3.11
+                ARG USER_ENV_VARS_HASH
                 # -- Adding local package . --
                 ADD . /deps/cli
                 # -- End of local package . --

libs/cli/tests/unit_tests/test_config.py

@@ -910,6 +910,7 @@ def test_config_to_docker_simple():
     expected_docker_stdin = f"""\
 # syntax=docker/dockerfile:1.4
 FROM langchain/langgraph-api:3.11
+ARG USER_ENV_VARS_HASH
 # -- Installing local requirements --
 COPY --from=outer-requirements.txt requirements.txt /deps/outer-graphs_reqs_a/graphs_reqs_a/requirements.txt
 RUN PYTHONDONTWRITEBYTECODE=1 uv pip install --system --no-cache-dir -c /api/constraints.txt -r /deps/outer-graphs_reqs_a/graphs_reqs_a/requirements.txt
@@ -974,6 +975,7 @@ def test_config_to_docker_outside_path():
         """\
 # syntax=docker/dockerfile:1.4
 FROM langchain/langgraph-api:3.11
+ARG USER_ENV_VARS_HASH
 # -- Adding non-package dependency unit_tests --
 ADD . /deps/outer-unit_tests/unit_tests
 RUN set -ex && \\
@@ -1034,6 +1036,7 @@ def test_config_to_docker_pipconfig():
     expected_docker_stdin = (
         """\
 FROM langchain/langgraph-api:3.11
+ARG USER_ENV_VARS_HASH
 ADD pipconfig.txt /pipconfig.txt
 # -- Adding non-package dependency unit_tests --
 ADD . /deps/outer-unit_tests/unit_tests
@@ -1097,6 +1100,7 @@ def test_config_to_docker_local_deps():
     )
     expected_docker_stdin = f"""\
 FROM langchain/langgraph-api-custom:3.11
+ARG USER_ENV_VARS_HASH
 # -- Adding non-package dependency graphs --
 ADD ./graphs /deps/outer-graphs/src
 RUN set -ex && \\
@@ -1144,6 +1148,7 @@ dependencies = ["langchain"]"""
     os.remove(pyproject_path)
     expected_docker_stdin = (
         """FROM langchain/langgraph-api:3.11
+ARG USER_ENV_VARS_HASH
 # -- Adding local package . --
 ADD . /deps/unit_tests
 # -- End of local package . --
@@ -1177,6 +1182,7 @@ def test_config_to_docker_end_to_end():
         base_image="langchain/langgraph-api",
     )
     expected_docker_stdin = f"""FROM langchain/langgraph-api:3.12
+ARG USER_ENV_VARS_HASH
 ARG meow
 ARG foo
 ADD pipconfig.txt /pipconfig.txt
@@ -1222,6 +1228,7 @@ def test_config_to_docker_nodejs():
         base_image="langchain/langgraphjs-api",
     )
     expected_docker_stdin = """FROM langchain/langgraphjs-api:20
+ARG USER_ENV_VARS_HASH
 ARG meow
 ARG foo
 ADD . /deps/unit_tests
@@ -1413,6 +1420,7 @@ def test_config_to_docker_nodejs_internal_docker_tag():
         base_image="langchain/langgraphjs-api",
     )
     expected_docker_stdin = """FROM langchain/langgraphjs-api:my-tag
+ARG USER_ENV_VARS_HASH
 ARG meow
 ARG foo
 ADD . /deps/unit_tests
@@ -1524,6 +1532,7 @@ def test_config_to_docker_gen_ui_python():
     )
 
     expected_docker_stdin = f"""FROM langchain/langgraph-api:3.11
+ARG USER_ENV_VARS_HASH
 RUN /storage/install-node.sh
 # -- Adding non-package dependency unit_tests --
 ADD . /deps/outer-unit_tests/unit_tests
@@ -1571,6 +1580,7 @@ def test_config_to_docker_multiplatform():
     )
 
     expected_docker_stdin = f"""FROM langchain/langgraph-api:3.11
+ARG USER_ENV_VARS_HASH
 RUN /storage/install-node.sh
 # -- Adding non-package dependency unit_tests --
 ADD . /deps/outer-unit_tests/unit_tests
@@ -2843,6 +2853,7 @@ def test_config_to_compose_simple_config():
             context: .
             dockerfile_inline: |
                 FROM langchain/langgraph-api:3.11
+                ARG USER_ENV_VARS_HASH
                 # -- Adding non-package dependency unit_tests --
                 ADD . /deps/outer-unit_tests/unit_tests
                 RUN set -ex && \\
@@ -2884,6 +2895,7 @@ def test_config_to_compose_env_vars():
             context: .
             dockerfile_inline: |
                 FROM langchain/langgraph-api-custom:3.11
+                ARG USER_ENV_VARS_HASH
                 # -- Adding non-package dependency unit_tests --
                 ADD . /deps/outer-unit_tests/unit_tests
                 RUN set -ex && \\
@@ -2929,6 +2941,7 @@ def test_config_to_compose_env_file():
             context: .
             dockerfile_inline: |
                 FROM langchain/langgraph-api:3.11
+                ARG USER_ENV_VARS_HASH
                 # -- Adding non-package dependency unit_tests --
                 ADD . /deps/outer-unit_tests/unit_tests
                 RUN set -ex && \\
@@ -2967,6 +2980,7 @@ def test_config_to_compose_watch():
             context: .
             dockerfile_inline: |
                 FROM langchain/langgraph-api:3.11
+                ARG USER_ENV_VARS_HASH
                 # -- Adding non-package dependency unit_tests --
                 ADD . /deps/outer-unit_tests/unit_tests
                 RUN set -ex && \\
@@ -3014,6 +3028,7 @@ def test_config_to_compose_end_to_end():
             context: .
             dockerfile_inline: |
                 FROM langchain/langgraph-api:3.11
+                ARG USER_ENV_VARS_HASH
                 # -- Adding non-package dependency unit_tests --
                 ADD . /deps/outer-unit_tests/unit_tests
                 RUN set -ex && \\

libs/cli/tests/unit_tests/test_env_cache.py

@@ -0,0 +1,142 @@
+"""Generated Dockerfile cache keys; opt into the real BuildKit test locally."""
+
+import hashlib
+import os
+import pathlib
+import subprocess
+import uuid
+
+import pytest
+
+from langgraph_cli.config import config_to_docker, validate_config
+
+
+def generated_prefix(tmp_path: pathlib.Path, runtime: str) -> str:
+    config = {
+        "graphs": {"agent": "agent:graph"},
+        "dockerfile_lines": [
+            f"RUN --mount=type=secret,id=build_marker echo {tmp_path.name} >/dev/null; "
+            "sha256sum /run/secrets/build_marker | cut -d ' ' -f 1 > /result"
+        ],
+    }
+    if runtime == "node":
+        config["node_version"] = "20"
+        config["graphs"] = {"agent": "./agent.js:graph"}
+    else:
+        config.update(python_version="3.11", dependencies=["requests"])
+        if runtime == "uv":
+            config.pop("dependencies")
+            config["source"] = {"kind": "uv", "root": "."}
+            (tmp_path / "uv.lock").write_text("version = 1\n")
+            (tmp_path / "pyproject.toml").write_text(
+                '[project]\nname = "agent"\nversion = "0.1.0"\n'
+                '[build-system]\nrequires = ["setuptools"]\n'
+                'build-backend = "setuptools.build_meta"\n'
+            )
+    dockerfile, _ = config_to_docker(
+        tmp_path / "langgraph.json",
+        validate_config(config),
+        base_image="langchain/langgraph-api",
+    )
+    # Keep the generated stage header and custom RUN exactly as emitted. Avoid
+    # installing real graph dependencies: this test exercises the cache boundary.
+    return dockerfile.split(" > /result", 1)[0] + " > /result\n"
+
+
+@pytest.mark.parametrize("runtime", ["python", "uv", "node"])
+def test_env_hash_in_build_stage_before_custom_run(tmp_path, runtime):
+    prefix = generated_prefix(tmp_path, runtime)
+    assert prefix.index("FROM ") < prefix.index("ARG USER_ENV_VARS_HASH")
+    assert prefix.index("ARG USER_ENV_VARS_HASH") < prefix.index("RUN ")
+
+
+@pytest.mark.skipif(
+    os.environ.get("LANGGRAPH_TEST_BUILDKIT") != "1",
+    reason="Set LANGGRAPH_TEST_BUILDKIT=1 with a local Docker BuildKit daemon",
+)
+@pytest.mark.parametrize("runtime", ["python", "uv", "node"])
+def test_secret_hash_invalidates_real_buildkit_cache(tmp_path, runtime, request):
+    prefix = generated_prefix(tmp_path, runtime)
+    lines = prefix.splitlines()
+    lines[0] = "FROM alpine:3.21"
+    prefix = "\n".join(lines) + "\n"
+    dockerfile = tmp_path / "Dockerfile"
+    marker = tmp_path / "build-marker"
+    (tmp_path / ".dockerignore").write_text("*\n!Dockerfile\n")
+    output_counter = 0
+    cache_image = "langgraph-cli-cache-test:" + uuid.uuid4().hex
+    request.addfinalizer(
+        lambda: subprocess.run(
+            ["docker", "image", "rm", cache_image], capture_output=True, check=False
+        )
+    )
+
+    def build(contents, value, *, pass_hash=True, export_image=False):
+        nonlocal output_counter
+        output_counter += 1
+        marker.write_text(value)
+        dockerfile.write_text(
+            contents + "FROM scratch\nCOPY --from=0 /result /result\n"
+        )
+        output = tmp_path / f"output-{output_counter}"
+        args = [
+            "docker",
+            "buildx",
+            "build",
+            "--progress=plain",
+            "--secret",
+            f"id=build_marker,src={marker}",
+            "--output",
+            f"type=local,dest={output}",
+            "-f",
+            str(dockerfile),
+            str(tmp_path),
+        ]
+        if pass_hash:
+            args.extend(
+                [
+                    "--build-arg",
+                    f"USER_ENV_VARS_HASH={hashlib.sha256(value.encode()).hexdigest()}",
+                ]
+            )
+        if output_counter > 1:
+            args.extend(["--cache-from", cache_image])
+        if export_image:
+            # Export a legacy image with inline cache, then explicitly import it.
+            image_args = args.copy()
+            output_index = image_args.index("--output")
+            del image_args[output_index : output_index + 2]
+            image_args.extend(
+                ["--load", "--tag", cache_image, "--cache-to", "type=inline"]
+            )
+            subprocess.run(image_args, capture_output=True, text=True, check=True)
+        # List arguments only; all inputs are local public test markers.
+        result = subprocess.run(args, capture_output=True, text=True, check=True)
+        return (output / "result").read_text().strip(), result.stderr
+
+    first = f"{runtime}-{tmp_path.name}-first"
+    second = f"{runtime}-{tmp_path.name}-second"
+    legacy = prefix.replace("ARG USER_ENV_VARS_HASH\n", "")
+    legacy_output, _ = build(legacy, first, export_image=True)
+    assert legacy_output == hashlib.sha256(first.encode()).hexdigest()
+    stale_output, _ = build(legacy, second)
+    assert stale_output == legacy_output  # Secret contents alone do not bust cache.
+
+    fixed_output, _ = build(prefix, second)
+    assert fixed_output == hashlib.sha256(second.encode()).hexdigest()
+    assert fixed_output != stale_output  # Old cache is safe when adopting the ARG.
+    repeated_output, repeated_log = build(prefix, second)
+    assert repeated_output == fixed_output
+    run_vertex = next(
+        line.split()[0]
+        for line in repeated_log.splitlines()
+        if "] RUN " in line and "sha256sum" in line
+    )
+    assert f"{run_vertex} CACHED" in repeated_log
+    changed_output, _ = build(prefix, first)
+    assert changed_output == hashlib.sha256(first.encode()).hexdigest()
+    assert changed_output != repeated_output
+
+    # Ordinary CLI callers can still build without supplying the optional hash.
+    unset_output, _ = build(prefix, first, pass_hash=False)
+    assert unset_output == hashlib.sha256(first.encode()).hexdigest()

libs/langgraph/langgraph/graph/message.py

@@ -22,7 +22,7 @@ from langchain_core.messages import (
 )
 from typing_extensions import TypedDict, deprecated
 
-from langgraph._internal._constants import CONF, CONFIG_KEY_SEND
+from langgraph._internal._constants import CONF, CONFIG_KEY_SEND, NS_SEP
 from langgraph.graph.state import StateGraph
 from langgraph.pregel._messages import (
     _messages_delta_reducer as _messages_delta_reducer,
@@ -343,7 +343,7 @@ def push_message(
     )
 
     from langgraph.config import get_config
-    from langgraph.pregel._messages import StreamMessagesHandler, graph_namespace
+    from langgraph.pregel._messages import StreamMessagesHandler
 
     config = get_config()
     message = next(x for x in convert_to_messages([message]))
@@ -359,12 +359,15 @@ def push_message(
     ):
         handlers = config["callbacks"]
 
-    metadata = config["metadata"]
-    for handler in handlers:
-        if isinstance(handler, StreamMessagesHandler):
-            ns = graph_namespace(metadata)
-            if handler.subgraphs or not ns or ns == handler.parent_ns:
-                handler._emit((ns, metadata), message, dedupe=False)
+    if stream_handler := next(
+        (x for x in handlers if isinstance(x, StreamMessagesHandler)), None
+    ):
+        metadata = config["metadata"]
+        message_meta = (
+            tuple(cast(str, metadata["langgraph_checkpoint_ns"]).split(NS_SEP)),
+            metadata,
+        )
+        stream_handler._emit(message_meta, message, dedupe=False)
 
     if state_key:
         config[CONF][CONFIG_KEY_SEND]([(state_key, message)])

libs/langgraph/langgraph/pregel/_draw.py

@@ -228,10 +228,6 @@ def draw_graph(
             metadata["__interrupt"] = "before"
         elif name in interrupt_after_nodes:
             metadata["__interrupt"] = "after"
-        # Duck-typed so core doesn't import ToolNode from langgraph-prebuilt.
-        tools_by_name = getattr(node.bound, "tools_by_name", None)
-        if isinstance(tools_by_name, Mapping) and tools_by_name:
-            metadata["__tools"] = list(tools_by_name)
         graph.add_node(node.bound, name, metadata=metadata or None)
     # add start node
     if START not in nodes:

libs/langgraph/langgraph/pregel/_messages.py

@@ -23,7 +23,7 @@ from langchain_core.outputs import ChatGeneration, ChatGenerationChunk, LLMResul
 from pydantic import BaseModel
 
 from langgraph._internal._config import filter_to_user_tags
-from langgraph._internal._constants import NS_SEP, PUSH
+from langgraph._internal._constants import NS_SEP
 from langgraph.channels.binop import _get_overwrite
 from langgraph.channels.delta import DeltaChannel
 from langgraph.constants import TAG_HIDDEN, TAG_NOSTREAM
@@ -55,23 +55,6 @@ def _state_values(obj: Any) -> Sequence[Any]:
     return ()
 
 
-def graph_namespace(metadata: Mapping[str, Any]) -> tuple[str, ...]:
-    """Namespace of the graph a task runs in, from the task's run metadata.
-
-    Drops the task's own segment, plus one per enclosing `@task` call: a call
-    runs in its caller's graph but nests its segment under the caller's. A
-    call's `langgraph_path` is `(PUSH, <caller's path>, <index>, True)`, while
-    a `Send` task's ends in `False`.
-    """
-    ns = tuple(cast(str, metadata["langgraph_checkpoint_ns"]).split(NS_SEP))
-    path = metadata.get("langgraph_path", ())
-    drop = 1
-    while len(path) == 4 and path[0] == PUSH and path[-1] is True:
-        drop += 1
-        path = path[1]
-    return ns[:-drop]
-
-
 class StreamMessagesHandler(BaseCallbackHandler, _StreamingCallbackHandler):
     """A callback handler that implements stream_mode=messages.
 
@@ -165,7 +148,9 @@ class StreamMessagesHandler(BaseCallbackHandler, _StreamingCallbackHandler):
         **kwargs: Any,
     ) -> Any:
         if metadata and (not tags or (TAG_NOSTREAM not in tags)):
-            ns = graph_namespace(metadata)
+            ns = tuple(cast(str, metadata["langgraph_checkpoint_ns"]).split(NS_SEP))[
+                :-1
+            ]
             if not self.subgraphs and len(ns) > 0 and ns != self.parent_ns:
                 return
             if (filtered_tags := filter_to_user_tags(tags)) is not None:
@@ -228,8 +213,10 @@ class StreamMessagesHandler(BaseCallbackHandler, _StreamingCallbackHandler):
             and kwargs.get("name") == metadata.get("langgraph_node")
             and (not tags or TAG_HIDDEN not in tags)
         ):
-            ns = graph_namespace(metadata)
-            if not self.subgraphs and len(ns) > 0 and ns != self.parent_ns:
+            ns = tuple(cast(str, metadata["langgraph_checkpoint_ns"]).split(NS_SEP))[
+                :-1
+            ]
+            if not self.subgraphs and len(ns) > 0:
                 return
             self.metadata[run_id] = (ns, metadata)
             for value in _state_values(inputs):

libs/langgraph/langgraph/pregel/_retry.py

@@ -34,7 +34,6 @@ from langgraph._internal._constants import (
     CONFIG_KEY_THREAD_ID,
     CONFIG_KEY_TIMED_ATTEMPT_OBSERVER,
     NS_SEP,
-    RETURN,
 )
 from langgraph._internal._runnable import create_task_in_config_context
 from langgraph._internal._timeout import sync_timeout_unsupported
@@ -734,14 +733,15 @@ async def arun_with_retry(
                 if stream:
                     async for _ in task.proc.astream(task.input, config):
                         pass
-                    return _return_value(task)
+                    break
                 return await task.proc.ainvoke(task.input, config)
             result = await _arun_with_timeout(
                 task, config, resolved_timeout, attempt_ctx, stream=stream
             )
             _finish_timed_attempt(config, attempt_ctx)
             if stream:
-                return _return_value(task)
+                # if successful, end
+                break
             return result
         except ParentCommand as exc:
             ns: str = config[CONF][CONFIG_KEY_CHECKPOINT_NS]
@@ -831,12 +831,6 @@ async def arun_with_retry(
             config = patch_configurable(config, {CONFIG_KEY_RESUMING: True})
 
 
-def _return_value(task: PregelExecutableTask) -> Any:
-    # Streaming drains the task instead of returning its result, so a caller
-    # awaiting a `@task` needs the value the task wrote.
-    return next((value for channel, value in task.writes if channel == RETURN), None)
-
-
 def _should_retry_on(retry_policy: RetryPolicy, exc: Exception) -> bool:
     """Check if the given exception should be retried based on the retry policy."""
     if isinstance(retry_policy.retry_on, Sequence):

libs/langgraph/langgraph/pregel/_tools.py

@@ -7,9 +7,8 @@ from uuid import UUID
 
 from langchain_core.callbacks import BaseCallbackHandler
 
+from langgraph._internal._constants import NS_SEP
 from langgraph.constants import TAG_NOSTREAM
-from langgraph.errors import GraphDrained, GraphInterrupt
-from langgraph.pregel._messages import graph_namespace
 from langgraph.pregel.protocol import StreamChunk
 
 try:
@@ -92,9 +91,12 @@ class StreamToolCallHandler(BaseCallbackHandler, _StreamingCallbackHandler):
     ) -> tuple[str, ...] | None:
         """Resolve the namespace this tool call should emit at, or `None` to skip.
 
-        Mirrors `StreamMessagesHandler.on_chat_model_start`: the namespace
-        is that of the graph the tool runs in (`graph_namespace`). Returns
-        `None` when the call should be silently suppressed:
+        Mirrors `StreamMessagesHandler.on_chat_model_start`'s namespace
+        derivation: parses `langgraph_checkpoint_ns` (which ends with
+        the `node_name:task_id` of the calling node), drops that
+        trailing segment, and returns the containing subgraph's own
+        namespace. Returns `None` when the call should be silently
+        suppressed:
 
         - `metadata` is missing — handler is attached to a context
           without Pregel routing info.
@@ -107,9 +109,11 @@ class StreamToolCallHandler(BaseCallbackHandler, _StreamingCallbackHandler):
             return None
         if tags and TAG_NOSTREAM in tags:
             return None
-        ns = (
-            graph_namespace(metadata) if metadata.get("langgraph_checkpoint_ns") else ()
-        )
+        nskey = metadata.get("langgraph_checkpoint_ns")
+        if not nskey:
+            ns: tuple[str, ...] = ()
+        else:
+            ns = tuple(cast(str, nskey).split(NS_SEP))[:-1]
         if not self.subgraphs and len(ns) > 0 and ns != self.parent_ns:
             return None
         return ns
@@ -184,10 +188,6 @@ class StreamToolCallHandler(BaseCallbackHandler, _StreamingCallbackHandler):
             return
         ns, tool_call_id, token = info
         self._reset_writer(token)
-        # An interrupted or drained tool is paused, not failed: it runs again
-        # on resume and reports its start and end then.
-        if isinstance(error, (GraphInterrupt, GraphDrained)):
-            return
         self.stream(
             (
                 ns,

libs/langgraph/langgraph/pregel/main.py

@@ -2880,18 +2880,6 @@ class Pregel(
                     for h in callback_manager.inheritable_handlers
                     if not isinstance(h, StreamMessagesHandlerV2)
                 ]
-            if CONFIG_KEY_TASK_ID not in config[CONF]:
-                # Not run as a task (a child invoked with its own thread, say),
-                # this graph's namespace starts at the root, so a caller's
-                # handler would take this graph's events for its own graph's.
-                callback_manager.handlers = [
-                    h for h in callback_manager.handlers if not _is_stream_handler(h)
-                ]
-                callback_manager.inheritable_handlers = [
-                    h
-                    for h in callback_manager.inheritable_handlers
-                    if not _is_stream_handler(h)
-                ]
             if "ls_integration" not in callback_manager.metadata:
                 callback_manager.add_metadata({"ls_integration": "langgraph"})
             run_manager = callback_manager.on_chain_start(
@@ -2910,9 +2898,9 @@ class Pregel(
                 )
             # set up subgraph checkpointing
             config = self._own_checkpoint_config(config)
-            stream_handlers: list[StreamMessagesHandler | StreamToolCallHandler] = []
             # set up messages stream mode
             if "messages" in stream_modes:
+                ns_ = cast(str | None, config[CONF].get(CONFIG_KEY_CHECKPOINT_NS))
                 use_stream_messages_v2 = bool(
                     version == "v2" and config[CONF].get(CONFIG_KEY_STREAM_MESSAGES_V2)
                 )
@@ -2921,12 +2909,24 @@ class Pregel(
                     if use_stream_messages_v2
                     else StreamMessagesHandler
                 )
-                stream_handlers.append(messages_handler_cls(stream.put, subgraphs))
+                run_manager.inheritable_handlers.append(
+                    messages_handler_cls(
+                        stream.put,
+                        subgraphs,
+                        parent_ns=tuple(ns_.split(NS_SEP)) if ns_ else None,
+                    )
+                )
 
             # set up tools stream mode
             if "tools" in stream_modes:
-                stream_handlers.append(StreamToolCallHandler(stream.put, subgraphs))
-            run_manager.inheritable_handlers.extend(stream_handlers)
+                ns_tools = cast(str | None, config[CONF].get(CONFIG_KEY_CHECKPOINT_NS))
+                run_manager.inheritable_handlers.append(
+                    StreamToolCallHandler(
+                        stream.put,
+                        subgraphs,
+                        parent_ns=tuple(ns_tools.split(NS_SEP)) if ns_tools else None,
+                    )
+                )
 
             # set up custom stream mode
             if "custom" in stream_modes:
@@ -3009,10 +3009,6 @@ class Pregel(
                 cache_policy=self.cache_policy,
                 has_graph_lifecycle_callbacks=bool(graph_callback_manager.handlers),
             ) as loop:
-                # Known only now: the loop numbers the namespace of every
-                # subgraph a task runs after its first.
-                for handler in stream_handlers:
-                    handler.parent_ns = loop.checkpoint_ns
                 emit_graph_lifecycle_events(loop)
                 # create runner
                 runner = PregelRunner(
@@ -3295,18 +3291,6 @@ class Pregel(
                     for h in callback_manager.inheritable_handlers
                     if not isinstance(h, StreamMessagesHandlerV2)
                 ]
-            if CONFIG_KEY_TASK_ID not in config[CONF]:
-                # Not run as a task (a child invoked with its own thread, say),
-                # this graph's namespace starts at the root, so a caller's
-                # handler would take this graph's events for its own graph's.
-                callback_manager.handlers = [
-                    h for h in callback_manager.handlers if not _is_stream_handler(h)
-                ]
-                callback_manager.inheritable_handlers = [
-                    h
-                    for h in callback_manager.inheritable_handlers
-                    if not _is_stream_handler(h)
-                ]
             if "ls_integration" not in callback_manager.metadata:
                 callback_manager.add_metadata({"ls_integration": "langgraph"})
             run_manager = await callback_manager.on_chain_start(
@@ -3326,9 +3310,7 @@ class Pregel(
                         True
                         for h in run_manager.handlers
                         if isinstance(h, _StreamingCallbackHandler)
-                        and not isinstance(
-                            h, (StreamMessagesHandler, StreamToolCallHandler)
-                        )
+                        and not isinstance(h, StreamMessagesHandler)
                     ),
                     False,
                 )
@@ -3341,9 +3323,10 @@ class Pregel(
                 )
             # set up subgraph checkpointing
             config = self._own_checkpoint_config(config)
-            stream_handlers: list[StreamMessagesHandler | StreamToolCallHandler] = []
             # set up messages stream mode
             if "messages" in stream_modes:
+                # namespace can be None in a root level graph?
+                ns_ = cast(str | None, config[CONF].get(CONFIG_KEY_CHECKPOINT_NS))
                 use_stream_messages_v2 = bool(
                     version == "v2" and config[CONF].get(CONFIG_KEY_STREAM_MESSAGES_V2)
                 )
@@ -3352,12 +3335,24 @@ class Pregel(
                     if use_stream_messages_v2
                     else StreamMessagesHandler
                 )
-                stream_handlers.append(messages_handler_cls(stream_put, subgraphs))
+                run_manager.inheritable_handlers.append(
+                    messages_handler_cls(
+                        stream_put,
+                        subgraphs,
+                        parent_ns=tuple(ns_.split(NS_SEP)) if ns_ else None,
+                    )
+                )
 
             # set up tools stream mode
             if "tools" in stream_modes:
-                stream_handlers.append(StreamToolCallHandler(stream_put, subgraphs))
-            run_manager.inheritable_handlers.extend(stream_handlers)
+                ns_tools = cast(str | None, config[CONF].get(CONFIG_KEY_CHECKPOINT_NS))
+                run_manager.inheritable_handlers.append(
+                    StreamToolCallHandler(
+                        stream_put,
+                        subgraphs,
+                        parent_ns=tuple(ns_tools.split(NS_SEP)) if ns_tools else None,
+                    )
+                )
 
             # set up custom stream mode
             def stream_writer(c: Any) -> None:
@@ -3466,10 +3461,6 @@ class Pregel(
                 cache_policy=self.cache_policy,
                 has_graph_lifecycle_callbacks=bool(graph_callback_manager.handlers),
             ) as loop:
-                # Known only now: the loop numbers the namespace of every
-                # subgraph a task runs after its first.
-                for handler in stream_handlers:
-                    handler.parent_ns = loop.checkpoint_ns
                 await aemit_graph_lifecycle_events(loop)
                 # create runner
                 runner = PregelRunner(
@@ -4467,10 +4458,6 @@ def _coerce_checkpoint_values(payload: Any, mapper: Callable[[Any], Any]) -> Non
         payload["values"] = mapper(payload["values"])
 
 
-def _is_stream_handler(handler: object) -> bool:
-    return isinstance(handler, (StreamMessagesHandler, StreamToolCallHandler))
-
-
 def _resolve_parent_ns(
     graph_config: RunnableConfig | None, call_config: RunnableConfig | None
 ) -> tuple[str, ...]:

libs/langgraph/langgraph/pregel/remote.py

@@ -847,11 +847,6 @@ class RemoteGraph(PregelProtocol):
             # filter for what was actually requested
             if mode not in requested:
                 continue
-            # subgraph events may have been requested only for the parent stream.
-            # messages are exempt: the server dedupes them across namespaces, so
-            # dropping them here would also drop the root-level message
-            if not subgraphs and NS_SEP in chunk.event and mode != "messages":
-                continue
 
             if chunk.event.startswith("messages"):
                 chunk = chunk._replace(data=tuple(chunk.data))
@@ -1002,11 +997,6 @@ class RemoteGraph(PregelProtocol):
             # filter for what was actually requested
             if mode not in requested:
                 continue
-            # subgraph events may have been requested only for the parent stream.
-            # messages are exempt: the server dedupes them across namespaces, so
-            # dropping them here would also drop the root-level message
-            if not subgraphs and NS_SEP in chunk.event and mode != "messages":
-                continue
 
             if chunk.event.startswith("messages"):
                 chunk = chunk._replace(data=tuple(chunk.data))

libs/langgraph/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "hatchling.build"
 
 [project]
 name = "langgraph"
-version = "1.2.18"
+version = "1.2.17"
 description = "Building stateful, multi-actor applications with LLMs"
 authors = []
 requires-python = ">=3.10"

libs/langgraph/tests/__snapshots__/test_large_cases.ambr

@@ -133,11 +133,6 @@
             "ToolNode"
           ],
           "name": "tools"
-        },
-        "metadata": {
-          "__tools": [
-            "search_api"
-          ]
         }
       },
       {
@@ -226,11 +221,6 @@
             "ToolNode"
           ],
           "name": "tools"
-        },
-        "metadata": {
-          "__tools": [
-            "search_api"
-          ]
         }
       },
       {

libs/langgraph/tests/test_pregel.py

@@ -8950,39 +8950,6 @@ def test_subgraph_streaming_sync() -> None:
     assert result["num_chunks"] == 9
 
 
-def test_subgraph_streaming_node_output_messages_sync() -> None:
-    """Messages returned by a node (not streamed by an LLM) are emitted when a
-    graph is explicitly streamed with `stream_mode="messages"` inside a node."""
-
-    def respond(state: MessagesState) -> dict:
-        return {"messages": [AIMessage("hi there", id="ai1")]}
-
-    subgraph = (
-        StateGraph(MessagesState)
-        .add_node("respond", respond)
-        .add_edge(START, "respond")
-        .compile()
-    )
-    msgs = {"messages": [("user", "hello")]}
-
-    top_level = [msg for msg, _ in subgraph.stream(msgs, stream_mode="messages")]
-
-    def parent_node(state: MessagesState, config: RunnableConfig) -> dict:
-        nonlocal in_node
-        in_node = [
-            msg for msg, _ in subgraph.stream(msgs, config, stream_mode="messages")
-        ]
-        return {}
-
-    in_node: list = []
-    StateGraph(MessagesState).add_node("parent", parent_node).add_edge(
-        START, "parent"
-    ).compile().invoke(msgs)
-
-    assert [m.content for m in top_level] == ["hi there"]
-    assert [m.content for m in in_node] == ["hi there"]
-
-
 def test_get_graph_nonterminal_last_step_source(snapshot: SnapshotAssertion) -> None:
     class State(TypedDict):
         messages: list[str]

libs/langgraph/tests/test_pregel_async.py

@@ -6834,31 +6834,6 @@ async def test_multiple_updates() -> None:
     ]
 
 
-@NEEDS_CONTEXTVARS
-async def test_task_in_subgraph_node_returns_its_value_with_tools_mode() -> None:
-    class State(TypedDict):
-        out: int | None
-
-    @task
-    async def double(x: int) -> int:
-        return x * 2
-
-    async def node(state: State) -> State:
-        return {"out": await double(21)}
-
-    inner = StateGraph(State).add_node("node", node).add_edge(START, "node").compile()
-    graph = (
-        StateGraph(State).add_node("inner", inner).add_edge(START, "inner").compile()
-    )
-
-    chunks = [
-        chunk
-        async for chunk in graph.astream({"out": None}, stream_mode=["values", "tools"])
-    ]
-
-    assert chunks[-1] == ("values", {"out": 42})
-
-
 @NEEDS_CONTEXTVARS
 async def test_falsy_return_from_task(async_checkpointer: BaseCheckpointSaver) -> None:
     """Test with a falsy return from a task."""
@@ -7623,40 +7598,6 @@ async def test_entrypoint_from_async_generator() -> None:
             yield "b"
 
 
-@NEEDS_CONTEXTVARS
-@pytest.mark.parametrize("timeout", [None, 30])
-@pytest.mark.parametrize("version", ["v1", "v2"])
-async def test_awaited_task_returns_its_value_under_astream_events(
-    version: Literal["v1", "v2"], timeout: int | None
-) -> None:
-    class State(TypedDict):
-        doubled: int
-        zeroed: int
-
-    @task(timeout=timeout)
-    async def double(x: int) -> int:
-        return x * 2
-
-    @task(timeout=timeout)
-    async def zero() -> int:
-        return 0
-
-    async def node(state: State) -> State:
-        return {"doubled": await double(21), "zeroed": await zero()}
-
-    graph = StateGraph(State).add_node("node", node).add_edge(START, "node").compile()
-
-    outputs = [
-        event["data"]["output"]
-        async for event in graph.astream_events(
-            {"doubled": 0, "zeroed": 1}, version=version
-        )
-        if event["event"] == "on_chain_end" and event["name"] == "node"
-    ]
-
-    assert outputs == [{"doubled": 42, "zeroed": 0}]
-
-
 @NEEDS_CONTEXTVARS
 async def test_named_tasks_functional() -> None:
     class Foo:
@@ -9503,44 +9444,6 @@ async def test_subgraph_streaming_async() -> None:
     assert result["num_chunks"] == 9
 
 
-async def test_subgraph_streaming_node_output_messages_async() -> None:
-    """Messages returned by a node (not streamed by an LLM) are emitted when a
-    graph is explicitly streamed with `stream_mode="messages"` inside a node."""
-
-    async def respond(state: MessagesState) -> dict:
-        return {"messages": [AIMessage("hi there", id="ai1")]}
-
-    subgraph = (
-        StateGraph(MessagesState)
-        .add_node("respond", respond)
-        .add_edge(START, "respond")
-        .compile()
-    )
-    msgs = {"messages": [("user", "hello")]}
-
-    top_level = [msg async for msg, _ in subgraph.astream(msgs, stream_mode="messages")]
-
-    async def parent_node(state: MessagesState, config: RunnableConfig) -> dict:
-        nonlocal in_node
-        in_node = [
-            msg
-            async for msg, _ in subgraph.astream(msgs, config, stream_mode="messages")
-        ]
-        return {}
-
-    in_node: list = []
-    await (
-        StateGraph(MessagesState)
-        .add_node("parent", parent_node)
-        .add_edge(START, "parent")
-        .compile()
-        .ainvoke(msgs)
-    )
-
-    assert [m.content for m in top_level] == ["hi there"]
-    assert [m.content for m in in_node] == ["hi there"]
-
-
 @NEEDS_CONTEXTVARS
 async def test_null_resume_disallowed_with_multiple_interrupts(
     async_checkpointer: BaseCheckpointSaver,

libs/langgraph/tests/test_remote_graph.py

@@ -16,11 +16,9 @@ from langgraph_sdk.schema import StreamPart
 from pydantic import BaseModel
 from typing_extensions import TypedDict
 
-from langgraph._internal._constants import CONFIG_KEY_STREAM
 from langgraph.errors import GraphInterrupt
 from langgraph.graph import END, START, MessagesState, StateGraph, add_messages
 from langgraph.pregel import Pregel
-from langgraph.pregel.protocol import StreamProtocol
 from langgraph.pregel.remote import RemoteGraph
 from langgraph.types import Interrupt, StateSnapshot
 from tests.any_str import AnyStr
@@ -865,87 +863,6 @@ async def test_astream_subgraphs_keeps_caller_namespace():
     assert stream_parts == _CALLER_PREFIXED
 
 
-# Subgraph events requested only on behalf of a streaming parent graph
-_PARENT_STREAMED_PARTS = [
-    StreamPart(event="updates|sub:abc", data={"inner": {"chunk": "inner"}}),
-    StreamPart(event="messages|sub:abc", data=[{"content": "token"}, {}]),
-    StreamPart(event="updates", data={"sub": {"chunk": "root"}}),
-]
-
-
-def _parent_streamed_config(parent_parts: list) -> RunnableConfig:
-    return {
-        "configurable": {
-            "thread_id": "thread_1",
-            "checkpoint_ns": "outer:123",
-            CONFIG_KEY_STREAM: StreamProtocol(
-                parent_parts.append, {"updates", "messages"}
-            ),
-        }
-    }
-
-
-def test_stream_without_subgraphs_skips_parent_requested_subgraph_events():
-    mock_sync_client = MagicMock()
-    mock_sync_client.runs.stream.return_value = _PARENT_STREAMED_PARTS
-    remote_pregel = RemoteGraph("test_graph_id", sync_client=mock_sync_client)
-
-    for stream_mode, expected in [
-        ("updates", [{"sub": {"chunk": "root"}}]),
-        # messages are not filtered: the server dedupes them across namespaces
-        ("messages", [({"content": "token"}, {})]),
-    ]:
-        parent_parts: list = []
-        stream_parts = list(
-            remote_pregel.stream(
-                {"input": "data"},
-                _parent_streamed_config(parent_parts),
-                stream_mode=stream_mode,
-            )
-        )
-
-        assert stream_parts == expected
-        # subgraph events are still requested and forwarded to the parent
-        assert mock_sync_client.runs.stream.call_args.kwargs["stream_subgraphs"]
-        assert [ns for ns, _, _ in parent_parts] == [
-            ("outer:123", "sub:abc"),
-            ("outer:123", "sub:abc"),
-            ("outer:123",),
-        ]
-
-
-async def test_astream_without_subgraphs_skips_parent_requested_subgraph_events():
-    mock_async_client = MagicMock()
-    async_iter = MagicMock()
-    async_iter.__aiter__.return_value = _PARENT_STREAMED_PARTS
-    mock_async_client.runs.stream.return_value = async_iter
-    remote_pregel = RemoteGraph("test_graph_id", client=mock_async_client)
-
-    for stream_mode, expected in [
-        ("updates", [{"sub": {"chunk": "root"}}]),
-        # messages are not filtered: the server dedupes them across namespaces
-        ("messages", [({"content": "token"}, {})]),
-    ]:
-        parent_parts: list = []
-        stream_parts = [
-            part
-            async for part in remote_pregel.astream(
-                {"input": "data"},
-                _parent_streamed_config(parent_parts),
-                stream_mode=stream_mode,
-            )
-        ]
-
-        assert stream_parts == expected
-        # subgraph events are still requested and forwarded to the parent
-        assert mock_async_client.runs.stream.call_args.kwargs["stream_subgraphs"]
-        assert [ns for ns, _, _ in parent_parts] == [
-            ("outer:123", "sub:abc"),
-            ("outer:123", "sub:abc"),
-            ("outer:123",),
-        ]
-
-
 def test_invoke():
     # set up test
     mock_sync_client = MagicMock()

libs/langgraph/tests/test_stream_namespaces.py

@@ -1,308 +0,0 @@
-import sys
-from typing import Any
-
-import pytest
-from langchain_core.language_models.fake_chat_models import GenericFakeChatModel
-from langchain_core.messages import AIMessage
-from langchain_core.tools import tool
-
-from langgraph.func import entrypoint, task
-from langgraph.graph import START, MessagesState, StateGraph
-from langgraph.graph.message import push_message
-from langgraph.graph.state import CompiledStateGraph
-
-pytestmark = pytest.mark.anyio
-
-NEEDS_CONTEXTVARS = pytest.mark.skipif(
-    sys.version_info < (3, 11),
-    reason="Python 3.11+ is required for async contextvars support",
-)
-
-
-@tool
-def echo(text: str) -> str:
-    """Echo the text."""
-    return text
-
-
-def _one_node_graph(node: Any) -> CompiledStateGraph:
-    return (
-        StateGraph(MessagesState)
-        .add_node(node)
-        .add_edge(START, node.__name__)
-        .compile()
-    )
-
-
-def _chat_and_echo(state: MessagesState) -> dict:
-    echo.invoke({"text": "done"})
-    return {"messages": [GenericFakeChatModel(messages=iter(["one two"])).invoke("hi")]}
-
-
-async def _achat_and_echo(state: MessagesState) -> dict:
-    await echo.ainvoke({"text": "done"})
-    model = GenericFakeChatModel(messages=iter(["one two"]))
-    return {"messages": [await model.ainvoke("hi")]}
-
-
-def _segment_names(ns: tuple[str, ...]) -> tuple[str, ...]:
-    return tuple(segment.split(":")[0] for segment in ns)
-
-
-def _summary(part: dict) -> tuple[tuple[str, ...], str]:
-    if part["type"] == "messages":
-        return _segment_names(part["ns"]), part["data"][0].content
-    return _segment_names(part["ns"]), part["data"]["event"]
-
-
-_CHAT_AND_ECHO_EVENTS = [
-    ((), "tool-started"),
-    ((), "tool-finished"),
-    ((), "one"),
-    ((), " "),
-    ((), "two"),
-]
-
-_CHILD_CONFIGS = pytest.mark.parametrize(
-    "child_config",
-    [
-        None,
-        {"configurable": {"thread_id": "child"}},
-        {"configurable": {"thread_id": "parent"}},
-    ],
-    ids=["no_config", "own_thread", "caller_thread"],
-)
-
-
-_TASK_CALLERS = pytest.mark.parametrize(
-    "nested", [False, True], ids=["from_entrypoint", "from_task_in_task"]
-)
-
-
-@_TASK_CALLERS
-@pytest.mark.parametrize("subgraphs", [False, True])
-def test_task_events_stream_under_the_graph_namespace(
-    nested: bool, subgraphs: bool
-) -> None:
-    @task
-    def chat_and_echo() -> None:
-        _chat_and_echo({"messages": []})
-
-    @task
-    def outer() -> None:
-        chat_and_echo().result()
-
-    @entrypoint()
-    def workflow(inputs: dict) -> None:
-        (outer() if nested else chat_and_echo()).result()
-
-    parts = workflow.stream(
-        {}, stream_mode=["messages", "tools"], subgraphs=subgraphs, version="v2"
-    )
-
-    assert [_summary(part) for part in parts] == _CHAT_AND_ECHO_EVENTS
-
-
-@NEEDS_CONTEXTVARS
-@_TASK_CALLERS
-@pytest.mark.parametrize("subgraphs", [False, True])
-async def test_task_events_stream_under_the_graph_namespace_async(
-    nested: bool, subgraphs: bool
-) -> None:
-    @task
-    async def chat_and_echo() -> None:
-        await _achat_and_echo({"messages": []})
-
-    @task
-    async def outer() -> None:
-        await chat_and_echo()
-
-    @entrypoint()
-    async def workflow(inputs: dict) -> None:
-        await (outer() if nested else chat_and_echo())
-
-    parts = workflow.astream(
-        {}, stream_mode=["messages", "tools"], subgraphs=subgraphs, version="v2"
-    )
-
-    assert [_summary(part) async for part in parts] == _CHAT_AND_ECHO_EVENTS
-
-
-def test_task_events_stream_under_the_namespace_of_their_subgraph() -> None:
-    @task
-    def chat_and_echo() -> None:
-        _chat_and_echo({"messages": []})
-
-    def node(state: MessagesState) -> dict:
-        chat_and_echo().result()
-        return {}
-
-    sub = _one_node_graph(node)
-    parent = StateGraph(MessagesState).add_node("sub", sub).add_edge(START, "sub")
-
-    parts = parent.compile().stream(
-        {"messages": []},
-        stream_mode=["messages", "tools"],
-        subgraphs=True,
-        version="v2",
-    )
-
-    assert [_summary(part) for part in parts] == [
-        (("sub",), event) for _, event in _CHAT_AND_ECHO_EVENTS
-    ]
-
-
-@_CHILD_CONFIGS
-def test_child_graph_streams_only_to_the_node_streaming_it(
-    child_config: dict | None,
-) -> None:
-    child = _one_node_graph(_chat_and_echo)
-    received = []
-
-    def node(state: MessagesState) -> dict:
-        received.extend(
-            child.stream(
-                {"messages": []}, child_config, stream_mode=["messages", "tools"]
-            )
-        )
-        return {}
-
-    parts = _one_node_graph(node).stream(
-        {"messages": []},
-        {"configurable": {"thread_id": "parent"}},
-        stream_mode=["messages", "tools"],
-    )
-
-    assert list(parts) == []
-    assert len(received) == len(_CHAT_AND_ECHO_EVENTS)
-
-
-@NEEDS_CONTEXTVARS
-@_CHILD_CONFIGS
-async def test_child_graph_streams_only_to_the_node_streaming_it_async(
-    child_config: dict | None,
-) -> None:
-    child = _one_node_graph(_achat_and_echo)
-    received = []
-
-    async def node(state: MessagesState) -> dict:
-        stream = child.astream(
-            {"messages": []}, child_config, stream_mode=["messages", "tools"]
-        )
-        received.extend([part async for part in stream])
-        return {}
-
-    parts = _one_node_graph(node).astream(
-        {"messages": []},
-        {"configurable": {"thread_id": "parent"}},
-        stream_mode=["messages", "tools"],
-    )
-
-    assert [part async for part in parts] == []
-    assert len(received) == len(_CHAT_AND_ECHO_EVENTS)
-
-
-@pytest.mark.parametrize(
-    "child_config",
-    [
-        {"configurable": {"thread_id": "child"}},
-        {"configurable": {"thread_id": "parent"}},
-    ],
-    ids=["own_thread", "caller_thread"],
-)
-def test_child_graph_on_its_own_thread_skips_a_caller_streaming_subgraphs(
-    child_config: dict,
-) -> None:
-    child = _one_node_graph(_chat_and_echo)
-
-    def node(state: MessagesState) -> dict:
-        child.invoke({"messages": []}, child_config)
-        return {}
-
-    parts = _one_node_graph(node).stream(
-        {"messages": []},
-        {"configurable": {"thread_id": "parent"}},
-        stream_mode=["messages", "tools"],
-        subgraphs=True,
-    )
-
-    assert list(parts) == []
-
-
-def test_node_gets_the_events_of_every_subgraph_it_streams() -> None:
-    sub = _one_node_graph(_chat_and_echo)
-    counts = []
-
-    def node(state: MessagesState) -> dict:
-        for _ in range(2):
-            stream = sub.stream({"messages": []}, stream_mode=["messages", "tools"])
-            counts.append(len(list(stream)))
-        return {}
-
-    _one_node_graph(node).invoke({"messages": []})
-
-    assert counts == [len(_CHAT_AND_ECHO_EVENTS)] * 2
-
-
-@NEEDS_CONTEXTVARS
-async def test_node_gets_the_events_of_every_subgraph_it_streams_async() -> None:
-    sub = _one_node_graph(_achat_and_echo)
-    counts = []
-
-    async def node(state: MessagesState) -> dict:
-        for _ in range(2):
-            stream = sub.astream({"messages": []}, stream_mode=["messages", "tools"])
-            counts.append(len([part async for part in stream]))
-        return {}
-
-    await _one_node_graph(node).ainvoke({"messages": []})
-
-    assert counts == [len(_CHAT_AND_ECHO_EVENTS)] * 2
-
-
-def _push(state: MessagesState) -> dict:
-    push_message(AIMessage(id="pushed", content="hi"), state_key=None)
-    return {}
-
-
-def _namespaces(graph: CompiledStateGraph, subgraphs: bool) -> list[tuple[str, ...]]:
-    parts = graph.stream(
-        {"messages": []}, stream_mode="messages", subgraphs=subgraphs, version="v2"
-    )
-    return [_segment_names(part["ns"]) for part in parts]
-
-
-def _streams_pushing_graph(received: list) -> Any:
-    pushing = _one_node_graph(_push)
-
-    def node(state: MessagesState) -> dict:
-        for message, _ in pushing.stream({"messages": []}, stream_mode="messages"):
-            received.append(message.id)
-        return {}
-
-    return node
-
-
-def test_pushed_message_streams_under_the_namespace_of_its_graph() -> None:
-    pushing = _one_node_graph(_push)
-    parent = StateGraph(MessagesState).add_node("sub", pushing).add_edge(START, "sub")
-
-    assert _namespaces(pushing, subgraphs=True) == [()]
-    assert _namespaces(parent.compile(), subgraphs=False) == []
-    assert _namespaces(parent.compile(), subgraphs=True) == [("sub",)]
-
-
-def test_pushed_message_streams_to_the_node_that_streams_its_graph() -> None:
-    received: list[str] = []
-    graph = _one_node_graph(_streams_pushing_graph(received))
-
-    assert _namespaces(graph, subgraphs=False) == []
-    assert received == ["pushed"]
-
-
-def test_pushed_message_streams_to_every_stream_that_takes_it() -> None:
-    received: list[str] = []
-    graph = _one_node_graph(_streams_pushing_graph(received))
-
-    assert _namespaces(graph, subgraphs=True) == [("node",)]
-    assert received == ["pushed"]

libs/langgraph/tests/test_tool_stream_handler.py

@@ -13,17 +13,13 @@ from typing import Annotated, Any
 import pytest
 from langchain_core.messages import AIMessage
 from langchain_core.tools import tool
-from langgraph.checkpoint.memory import InMemorySaver
 from langgraph.prebuilt import ToolNode, ToolRuntime
 from typing_extensions import TypedDict
 
 from langgraph.constants import END, START
-from langgraph.errors import GraphDrained
 from langgraph.graph import StateGraph
 from langgraph.graph.message import add_messages
 from langgraph.pregel._tools import _tool_call_writer
-from langgraph.runtime import RunControl
-from langgraph.types import Command, interrupt
 
 
 class _State(TypedDict):
@@ -58,14 +54,14 @@ def _caller_async(tool_name: str, tool_args: dict[str, Any], tc_id: str = "tc1")
     return caller
 
 
-def _build_graph(caller, tools, checkpointer=None) -> Any:
+def _build_graph(caller, tools) -> Any:
     sg = StateGraph(_State)
     sg.add_node("caller", caller)
     sg.add_node("tools", ToolNode(tools))
     sg.add_edge(START, "caller")
     sg.add_edge("caller", "tools")
     sg.add_edge("tools", END)
-    return sg.compile(checkpointer=checkpointer)
+    return sg.compile()
 
 
 def _tool_events(stream) -> list[tuple[tuple[str, ...], dict]]:
@@ -150,105 +146,6 @@ class TestSyncGraphSyncTool:
         assert kinds == ["tool-started", "tool-error"]
         assert events[1][1]["message"] == "nope"
 
-    def test_interrupt_in_tool_is_not_a_tool_error(self) -> None:
-        @tool
-        def ask() -> str:
-            """asks."""
-            return interrupt("ok?")
-
-        graph = _build_graph(_caller_sync("ask", {}), [ask], InMemorySaver())
-        config = {"configurable": {"thread_id": "1"}}
-        paused = _tool_events(
-            graph.stream(
-                {"messages": []}, config, stream_mode=["tools"], subgraphs=True
-            )
-        )
-        resumed = _tool_events(
-            graph.stream(
-                Command(resume="yes"), config, stream_mode=["tools"], subgraphs=True
-            )
-        )
-
-        assert [p["event"] for _, p in paused] == ["tool-started"]
-        assert [(p["event"], p["tool_call_id"]) for _, p in resumed] == [
-            ("tool-started", "tc1"),
-            ("tool-finished", "tc1"),
-        ]
-
-    def test_tool_handing_off_to_the_parent_still_ends_its_call(self) -> None:
-        class State(TypedDict):
-            messages: Annotated[list, add_messages]
-            done: bool
-
-        def hand_off(state: State) -> Command:
-            return Command(graph=Command.PARENT, goto="finish", update={"done": True})
-
-        inner = StateGraph(State).add_node("hand_off", hand_off)
-        inner = inner.add_edge(START, "hand_off").compile()
-
-        @tool
-        def handoff() -> str:
-            """hands off."""
-            inner.invoke({"messages": [], "done": False})
-            return "unreachable"
-
-        graph = (
-            StateGraph(State)
-            .add_node("caller", _caller_sync("handoff", {}))
-            .add_node("tools", ToolNode([handoff]))
-            .add_node("finish", lambda state: {"done": True})
-            .add_edge(START, "caller")
-            .add_edge("caller", "tools")
-            .add_edge("tools", END)
-            .add_edge("finish", END)
-            .compile()
-        )
-        events = [
-            payload["event"]
-            for _, payload in _tool_events(
-                graph.stream(
-                    {"messages": [], "done": False},
-                    stream_mode=["tools"],
-                    subgraphs=True,
-                )
-            )
-        ]
-
-        assert events[0] == "tool-started"
-        assert events[-1] == "tool-error", events
-
-    def test_drained_tool_is_not_a_tool_error(self) -> None:
-        control = RunControl()
-
-        def request_drain(state: _State) -> dict:
-            control.request_drain()
-            return {}
-
-        inner = (
-            StateGraph(_State)
-            .add_node("request_drain", request_drain)
-            .add_node("after", lambda state: {})
-            .add_edge(START, "request_drain")
-            .add_edge("request_drain", "after")
-            .compile()
-        )
-
-        @tool
-        def drained() -> str:
-            """drains."""
-            inner.invoke({"messages": []}, control=control)
-            return "unreachable"
-
-        graph = _build_graph(_caller_sync("drained", {}), [drained])
-        chunks: list = []
-        with pytest.raises(GraphDrained):
-            for chunk in graph.stream(
-                {"messages": []}, stream_mode=["tools"], subgraphs=True
-            ):
-                chunks.append(chunk)
-
-        assert [p["event"] for _, p in _tool_events(chunks)] == ["tool-started"]
-
     def test_writer_unset_outside_tool(self) -> None:
         # Outside any tool body the ContextVar that ToolRuntime reads
         # is unset — emitting from there would be a no-op.
@@ -299,25 +196,6 @@ class TestAsyncGraphAsyncTool:
         assert kinds == ["tool-started", "tool-output-delta", "tool-finished"]
         assert events[1][1]["delta"] == "hi"
 
-    @pytest.mark.anyio
-    async def test_interrupt_in_async_tool_is_not_a_tool_error(self) -> None:
-        @tool
-        async def aask() -> str:
-            """asks."""
-            return interrupt("ok?")
-
-        graph = _build_graph(_caller_async("aask", {}), [aask], InMemorySaver())
-        config = {"configurable": {"thread_id": "1"}}
-        events = []
-        for graph_input in ({"messages": []}, Command(resume="yes")):
-            async for ns, mode, payload in graph.astream(
-                graph_input, config, stream_mode=["tools"], subgraphs=True
-            ):
-                if mode == "tools":
-                    events.append(payload["event"])
-
-        assert events == ["tool-started", "tool-started", "tool-finished"]
-
 
 class TestConcurrentToolCalls:
     def test_parallel_tool_calls_do_not_bleed(self) -> None:

libs/langgraph/uv.lock

@@ -1439,7 +1439,7 @@ wheels = [
 
 [[package]]
 name = "langgraph"
-version = "1.2.18"
+version = "1.2.17"
 source = { editable = "." }
 dependencies = [
     { name = "langchain-core" },
@@ -1625,7 +1625,7 @@ wheels = [
 
 [[package]]
 name = "langgraph-checkpoint"
-version = "4.3.1"
+version = "4.3.0"
 source = { editable = "../checkpoint" }
 dependencies = [
     { name = "langchain-core" },

libs/prebuilt/uv.lock

@@ -286,7 +286,7 @@ wheels = [
 
 [[package]]
 name = "langgraph"
-version = "1.2.18"
+version = "1.2.17"
 source = { editable = "../langgraph" }
 dependencies = [
     { name = "langchain-core" },
@@ -370,7 +370,7 @@ test = [
 
 [[package]]
 name = "langgraph-checkpoint"
-version = "4.3.1"
+version = "4.3.0"
 source = { editable = "../checkpoint" }
 dependencies = [
     { name = "langchain-core" },

libs/sdk-py/langgraph_sdk/__init__.py

@@ -5,7 +5,7 @@ from langgraph_sdk.encryption.types import DecryptResult, EncryptionContext
 from langgraph_sdk.priority import Priority
 from langgraph_sdk.priority.types import PriorityContext, PrioritySettings
 
-__version__ = "0.4.10"
+__version__ = "0.4.9"
 
 __all__ = [
     "Auth",

libs/sdk-py/langgraph_sdk/_async/http.py

@@ -217,12 +217,6 @@ class HttpClient:
             current_headers = dict(
                 request_headers if reconnect_path is None else reconnect_headers
             )
-            if last_event_id is not None:
-                current_headers = {
-                    key: value
-                    for key, value in current_headers.items()
-                    if key.lower() != "last-event-id"
-                }
             if last_event_id:
                 current_headers["Last-Event-ID"] = last_event_id
 

libs/sdk-py/langgraph_sdk/_sync/http.py

@@ -219,12 +219,6 @@ class SyncHttpClient:
             current_headers = dict(
                 request_headers if reconnect_path is None else reconnect_headers
             )
-            if last_event_id is not None:
-                current_headers = {
-                    key: value
-                    for key, value in current_headers.items()
-                    if key.lower() != "last-event-id"
-                }
             if last_event_id:
                 current_headers["Last-Event-ID"] = last_event_id
 

libs/sdk-py/langgraph_sdk/auth/__init__.py

@@ -811,7 +811,7 @@ class _On:
     def __call__(
         self,
         *,
-        resources: str | Sequence[str] | None = None,
+        resources: str | Sequence[str],
         actions: str | Sequence[str] | None = None,
     ) -> Callable[[AHO], AHO]: ...
 

libs/sdk-py/langgraph_sdk/sse.py

@@ -104,7 +104,7 @@ class SSEDecoder:
             if (
                 not self._event
                 and not self._data
-                and self._last_event_id is None
+                and not self._last_event_id
                 and self._retry is None
             ):
                 return None

libs/sdk-py/tests/test_auth.py

@@ -1,49 +1,8 @@
-import ast
-import inspect
-
 import pytest
 
 from langgraph_sdk import Auth
 
 
-@pytest.mark.parametrize(
-    "kwargs",
-    [{}, {"actions": "read"}, {"resources": None}],
-)
-def test_global_handler_factory_accepts_optional_resources(kwargs) -> None:
-    auth = Auth()
-    tree = ast.parse(inspect.getsource(type(auth.on)))
-    cls = tree.body[0]
-    assert isinstance(cls, ast.ClassDef)
-    factory_overload = next(
-        node
-        for node in cls.body
-        if isinstance(node, ast.FunctionDef)
-        and node.name == "__call__"
-        and node.decorator_list
-        and any(arg.arg == "resources" for arg in node.args.kwonlyargs)
-    )
-    resources = next(
-        arg for arg in factory_overload.args.kwonlyargs if arg.arg == "resources"
-    )
-    assert resources.annotation is not None
-    assert ast.unparse(resources.annotation) == "str | Sequence[str] | None"
-    default = factory_overload.args.kw_defaults[0]
-    assert isinstance(default, ast.Constant) and default.value is None
-
-    @auth.on(**kwargs)
-    async def handler(ctx, value):
-        del ctx, value
-        return None
-
-    if kwargs.get("actions"):
-        assert auth._handlers == {("*", "read"): [handler]}
-        assert auth._global_handlers == []
-    else:
-        assert auth._global_handlers == [handler]
-        assert auth._handlers == {}
-
-
 def test_handler_multiple_resources_and_actions() -> None:
     auth = Auth()
 

libs/sdk-py/tests/test_client_stream.py

@@ -312,14 +312,18 @@ async def test_http_client_stream_recovers_after_disconnect():
     ]
 
 
-def _drop_after(chunks: list[bytes], stream_cls: type, reconnects: list[httpx.Request]):
+def _drop_after_an_empty_id(stream_cls: type, reconnects: list[httpx.Request]):
     def handler(request: httpx.Request) -> httpx.Response:
         if request.method == "POST":
             return httpx.Response(
                 200,
                 headers={"Content-Type": "text/event-stream", "Location": "/reconnect"},
                 stream=stream_cls(
-                    chunks, httpx.RemoteProtocolError("incomplete chunked read")
+                    [
+                        b"id: 1\nevent: values\ndata: 1\n\n",
+                        b"id:\nevent: values\ndata: 2\n\n",
+                    ],
+                    httpx.RemoteProtocolError("incomplete chunked read"),
                 ),
             )
         reconnects.append(request)
@@ -332,99 +336,26 @@ def _drop_after(chunks: list[bytes], stream_cls: type, reconnects: list[httpx.Re
     return handler
 
 
-def _sync_reconnect(
-    chunks: list[bytes], headers: dict[str, str] | None
-) -> httpx.Request:
+def test_sync_stream_reconnects_without_a_cursor_after_an_empty_id():
     reconnects: list[httpx.Request] = []
-    transport = httpx.MockTransport(_drop_after(chunks, ListByteStream, reconnects))
+    transport = httpx.MockTransport(_drop_after_an_empty_id(ListByteStream, reconnects))
     with httpx.Client(transport=transport, base_url="https://example.com") as client:
-        list(SyncHttpClient(client).stream("/stream", "POST", json={}, headers=headers))
-    return reconnects[0]
+        list(SyncHttpClient(client).stream("/stream", "POST", json={}))
+
+    assert "last-event-id" not in reconnects[0].headers
 
 
-async def _async_reconnect(
-    chunks: list[bytes], headers: dict[str, str] | None
-) -> httpx.Request:
+async def test_async_stream_reconnects_without_a_cursor_after_an_empty_id():
     reconnects: list[httpx.Request] = []
     transport = httpx.MockTransport(
-        _drop_after(chunks, AsyncListByteStream, reconnects)
+        _drop_after_an_empty_id(AsyncListByteStream, reconnects)
     )
     async with httpx.AsyncClient(
         transport=transport, base_url="https://example.com"
     ) as client:
-        [
-            part
-            async for part in HttpClient(client).stream(
-                "/stream", "POST", json={}, headers=headers
-            )
-        ]
-    return reconnects[0]
+        [part async for part in HttpClient(client).stream("/stream", "POST", json={})]
 
-
-_EMPTY_IDS = pytest.mark.parametrize(
-    "empty_id",
-    [b"id:\nevent: values\ndata: 2\n\n", b"id:\n\n"],
-    ids=["with-an-event", "alone"],
-)
-_CALLER_CURSORS = pytest.mark.parametrize(
-    "headers",
-    [None, {"Last-Event-ID": "5"}, {"last-event-id": "5"}],
-    ids=["no-caller-cursor", "caller-cursor", "lowercase-caller-cursor"],
-)
-
-
-@_EMPTY_IDS
-@_CALLER_CURSORS
-def test_sync_stream_reconnects_without_a_cursor_after_an_empty_id(
-    empty_id: bytes, headers: dict[str, str] | None
-):
-    reconnect = _sync_reconnect(
-        [b"id: 1\nevent: values\ndata: 1\n\n", empty_id], headers
-    )
-
-    assert "last-event-id" not in reconnect.headers
-
-
-@_EMPTY_IDS
-@_CALLER_CURSORS
-async def test_async_stream_reconnects_without_a_cursor_after_an_empty_id(
-    empty_id: bytes, headers: dict[str, str] | None
-):
-    reconnect = await _async_reconnect(
-        [b"id: 1\nevent: values\ndata: 1\n\n", empty_id], headers
-    )
-
-    assert "last-event-id" not in reconnect.headers
-
-
-def test_sync_stream_id_replaces_a_lowercase_caller_cursor():
-    reconnect = _sync_reconnect(
-        [b"id: 1\nevent: values\ndata: 1\n\n"], {"last-event-id": "5"}
-    )
-
-    assert reconnect.headers.get_list("last-event-id") == ["1"]
-
-
-async def test_async_stream_id_replaces_a_lowercase_caller_cursor():
-    reconnect = await _async_reconnect(
-        [b"id: 1\nevent: values\ndata: 1\n\n"], {"last-event-id": "5"}
-    )
-
-    assert reconnect.headers.get_list("last-event-id") == ["1"]
-
-
-def test_sync_stream_keeps_the_caller_cursor_until_the_stream_sends_an_id():
-    reconnect = _sync_reconnect([b"event: values\ndata: 1\n\n"], {"Last-Event-ID": "5"})
-
-    assert reconnect.headers.get_list("last-event-id") == ["5"]
-
-
-async def test_async_stream_keeps_the_caller_cursor_until_the_stream_sends_an_id():
-    reconnect = await _async_reconnect(
-        [b"event: values\ndata: 1\n\n"], {"Last-Event-ID": "5"}
-    )
-
-    assert reconnect.headers.get_list("last-event-id") == ["5"]
+    assert "last-event-id" not in reconnects[0].headers
 
 
 # --- _sse_to_v2_dict conversion ---

libs/sdk-py/uv.lock

@@ -301,7 +301,7 @@ wheels = [
 
 [[package]]
 name = "langgraph"
-version = "1.2.18"
+version = "1.2.17"
 source = { editable = "../langgraph" }
 dependencies = [
     { name = "langchain-core" },
@@ -385,7 +385,7 @@ test = [
 
 [[package]]
 name = "langgraph-checkpoint"
-version = "4.3.1"
+version = "4.3.0"
 source = { editable = "../checkpoint" }
 dependencies = [
     { name = "langchain-core" },