`ToolMessage` is not msgpack serializable when tool call result contains `non_text_contents` (artifact)

[styukovs]

Hello,

When the tool response contains an artifact, it has the type

from mcp.types import EmbeddedResource, ImageContent

NonTextContent = ImageContent | EmbeddedResource

and appears in the ToolMessage as something like:

ToolMessage(
    content=...,
    name=...,
    id=...,
    tool_call_id=...,
    artifact=[
        EmbeddedResource(
            type="resource",
            resource=TextResourceContents(
                uri=AnyUrl("https://example.com/"),
                mimeType=None,
                meta=None,
                text=...
            )
        )
    ]
)

The problem occurs when I use this tool inside a ReAct agent together with a checkpointer.
In this case, I get the following error:

TypeError: Type is not msgpack serializable: ToolMessage

It seems this happens because of the required uri field, which is of type AnyUrl, and msgpack cannot serialize it directly.

[eyurtsev]

Are you creating this ToolMessage manually?

[styukovs]

No, I receive this ToolMessage after invoking a StructuredTool, which wraps an MCPTool.

The EmbeddedResource is an artifact that is returned by the MCP server as part of its response.

[baliyanvinay]

I also faced the same issue when the tool call response content has a resource "type". Then the serialization breaks.
Not sure if it is a langchain problem or a langgraph state problem @styukovs
The response format is correct as "content_and_artifact"

{
    "content": [
        {
            "type": "text",
            "text": "successfully downloaded text file (SHA: df8252babaab7f5c7154b53e11c0c37459a061c5)"
        },
        {
            "type": "resource",
            "resource": {
                "uri": "repo://baliyanvinay/Data-Structures-and-Algorithms/contents/README.md",
                "mimeType": "text/plain; charset=utf-8",
                "text": "# Data-Structure\n"
            }
        }
    ]
}

Related issue: langchain-ai/langgraph#5054

[Rainlv]

ToolMessage with Pandas Dataframe artifact,

Work fine with InMemorySaver

serializer = JsonPlusSerializer(pickle_fallback=True)

checkpointer = InMemorySaver(serde=serializer)

but fail with AsyncSqliteSaver

conn = aiosqlite.connect(APP_CONFIG.LANGGRAPH_CHECKPOINT_URI)
checkpointer = AsyncSqliteSaver(
            conn=conn, serde=serializer
        )

version:

langchain                   0.3.27
langchain-community         0.3.21
langchain-core              0.3.76
langchain-openai            0.3.30
langchain-text-splitters    0.3.9
langgraph                   0.6.5
langgraph-checkpoint        2.1.1
langgraph-checkpoint-sqlite 2.0.11
langgraph-prebuilt          0.6.4
langgraph-sdk               0.2.2

log:

2025-09-22 15:54:30 | ERROR | __main__:run_agent_and_stream:199 - An error occurred during prediction stream: cannot pickle '_thread.RLock' object
2025-09-22 15:54:30 | ERROR | __main__:run_agent_and_stream:200 - Traceback (most recent call last):
  File "D:\Users\wangxh64\Documents\projects\fcst_agent\.venv\Lib\site-packages\langgraph\checkpoint\serde\jsonplus.py", line 216, in dumps_typed
    return "msgpack", _msgpack_enc(obj)
                      ^^^^^^^^^^^^^^^^^
  File "D:\Users\wangxh64\Documents\projects\fcst_agent\.venv\Lib\site-packages\langgraph\checkpoint\serde\jsonplus.py", line 676, in _msgpack_enc
    return ormsgpack.packb(data, default=_msgpack_default, option=_option)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
TypeError: Type is not msgpack serializable: ToolMessage

During handling of the above exception, another exception occurred:

Traceback (most recent call last):
  File "D:\Users\wangxh64\Documents\projects\fcst_agent\web\app.py", line 151, in run_agent_and_stream
    async for event_json in agent.run(
  File "D:\Users\wangxh64\Documents\projects\fcst_agent\src\output\json_response.py", line 14, in __aiter__
    async for event in self.stream_events:
  File "D:\Users\wangxh64\Documents\projects\fcst_agent\.venv\Lib\site-packages\langchain_core\runnables\base.py", line 1540, in astream_events
    async for event in event_stream:
  File "D:\Users\wangxh64\Documents\projects\fcst_agent\.venv\Lib\site-packages\langchain_core\tracers\event_stream.py", line 1034, in _astream_events_implementation_v2
    await task
  File "D:\Users\wangxh64\Documents\projects\fcst_agent\.venv\Lib\site-packages\langchain_core\tracers\event_stream.py", line 989, in consume_astream
    async for _ in event_streamer.tap_output_aiter(run_id, stream):
  File "D:\Users\wangxh64\Documents\projects\fcst_agent\.venv\Lib\site-packages\langchain_core\tracers\event_stream.py", line 213, in tap_output_aiter
    async for chunk in output:
  File "D:\Users\wangxh64\Documents\projects\fcst_agent\.venv\Lib\site-packages\langgraph\pregel\main.py", line 2883, in astream
    async with AsyncPregelLoop(
  File "D:\Users\wangxh64\Documents\projects\fcst_agent\.venv\Lib\site-packages\langgraph\pregel\_loop.py", line 1245, in __aexit__
    return await exit_task
           ^^^^^^^^^^^^^^^
  File "C:\Users\wangxh64\AppData\Roaming\uv\python\cpython-3.11.13-windows-x86_64-none\Lib\contextlib.py", line 745, in __aexit__
    raise exc_details[1]
  File "C:\Users\wangxh64\AppData\Roaming\uv\python\cpython-3.11.13-windows-x86_64-none\Lib\contextlib.py", line 728, in __aexit__
    cb_suppress = await cb(*exc_details)
                  ^^^^^^^^^^^^^^^^^^^^^^
  File "D:\Users\wangxh64\Documents\projects\fcst_agent\.venv\Lib\site-packages\langgraph\pregel\_executor.py", line 210, in __aexit__
    raise exc
  File "D:\Users\wangxh64\Documents\projects\fcst_agent\.venv\Lib\site-packages\langgraph\pregel\_loop.py", line 1132, in _checkpointer_put_after_previous
    await prev
  File "D:\Users\wangxh64\Documents\projects\fcst_agent\.venv\Lib\site-packages\langgraph\pregel\_loop.py", line 1134, in _checkpointer_put_after_previous
    await cast(BaseCheckpointSaver, self.checkpointer).aput(
  File "D:\Users\wangxh64\Documents\projects\fcst_agent\.venv\Lib\site-packages\langgraph\checkpoint\sqlite\aio.py", line 505, in aput
    type_, serialized_checkpoint = self.serde.dumps_typed(checkpoint)
                                   ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "D:\Users\wangxh64\Documents\projects\fcst_agent\.venv\Lib\site-packages\langgraph\checkpoint\serde\jsonplus.py", line 221, in dumps_typed
    return "pickle", pickle.dumps(obj)
                     ^^^^^^^^^^^^^^^^^
TypeError: cannot pickle '_thread.RLock' object

[shaunaa126]

Also can confirm this fails for AsyncPostgresSaver as well, I am using langchain v1 and deep_agents. Has anyone found a workaround for this?

  "langgraph (==1.0.0a4)",
  "langgraph-sdk (==0.2.9)", 
  "langgraph-supervisor (==0.0.28)",
  "langgraph-cli[inmem] (==0.3.7)",
  "langgraph-checkpoint-postgres (==2.0.25)",
  "langchain-core (==1.0.0a8)",
  "langchain-openai (==1.0.0a4)",
  "langchain-ollama (==1.0.0a1)",
  "langchain (==1.0.0a13)",  
  "deepagents (==0.0.5)",
  "langchain-mcp-adapters (==0.1.11)",

[klauswong]

I was stuck for a long time too and seems like adding pickle_fallback=True to the serializer saved my problem

def get_async_postgres_saver() -> AsyncPostgresSaver:
    async_postgres_saver = AsyncPostgresSaver.from_conn_string(
        settings.checkpointer_database_url, serde=JsonPlusSerializer(pickle_fallback=True)
    )
    return async_postgres_saver

[styukovs]

@eyurtsev, could you please review this PR langchain-ai/langgraph#6271?

[sydney-runkle]

Confirmed this is a limitation on the LG side right now.