core[patch]: check "tool_call_id" existence to prevent KeyError (#27249)

[lumiaspic (lumiaspic)]

Description: Fixes a bug where a KeyError is thrown when tool_call_id is not found, by calling values.get("tool_call_id") instead of values["tool_call_id"]. This error happens when running an agent using a RemoteRunnable from langserve like in this langserve agent with history implementation example.

Issue: Fixes #27249

Dependencies: none

[vercel]

The latest updates on your projects. Learn more about Vercel for Git ↗︎

1 Skipped Deployment

Name	Status	Preview	Comments	Updated (UTC)
langchain	⬜️ Ignored (Inspect)	Visit Preview		Mar 12, 2025 8:49am

[Eugene Yurtsev (eyurtsev)]

@dorian-bucaille

The change does not look correct as tool_call_id is a required field.

It's likely that you're using an older chat model and need to update both langchain-core and the chat model to get tool call ids to appear.

If you take a look at the schema tool_call_id is a required field

[lumiaspic (lumiaspic)]

Hi Eugene Yurtsev (@eyurtsev),

Thanks for your comment. I understand tool_call_id being a required field, and I appreciate your suggestion to update both langchain-core and the chat model to get tool_call_id to appear. I've updated my dependencies to the latest versions of langchain-core and langchain-openai this morning, and I still got the error.

I'm using the langserve agent implementation example with history. I use vLLM to run my models. The problem occurs when the agent tries to formulate a response after calling a tool. If I modify my local version of libs/core/langchain_core/messages/tool.py with the changes from my PR, the error is not blocking, and the agent can continue to stream its response after calling the tool.

Here's some detail about my code:

Server-side (langserve):

llm_from_config = ChatOpenAI(
    openai_api_base=llm_config.get("openai_api_base"),
    openai_api_key=llm_config.get("openai_api_key"),
    model=llm_config.get("model"),
    temperature=llm_config.get("temperature"),
    streaming=True,
    default_headers={
        "Authorization": "Bearer " + llm_config.get("openai_api_key")
    },
)

llm_with_tools = llm_from_config .bind_tools([tool for tool in known_tools])
agent = create_tool_calling_agent(llm_with_tools, known_tools, prompt)
agent_executor = AgentExecutor(agent=agent, tools=known_tools, verbose=True)

app = FastAPI(
    title=config.title,
    version=config.version,
    description=config.description,
)

add_routes(
    app,
    RunnableLambda(_create_agent_executor).with_types(
        input_type=Input, output_type=Output
    ),
)

Client-side:

try:
    async for event in remote_runnable.astream_events(
        {
            "input": prompt,
            "chat_history": chat_history,
            "llm_config": llm_config,
        },
        version="v1",
    ):
        kind = event["event"]
        if kind == "on_chain_start":
            # Événement de démarrage de l'agent
            if event["name"] == "agent":
                [...]
        elif kind == "on_chain_end":
           [...]
        if kind == "on_chat_model_stream":
            [...]
        elif kind == "on_tool_start":
            [...]
        elif kind == "on_tool_end":
            [...]

I've added client-side custom logs to remote_runnable.astream_events and libs/core/langchain_core/messages/tool.py to better understand the issue.

The error message indicates that ToolMessage must have a tool_call_id, but the received values dictionary does not contain this field. Maybe it has to do with the on_prompt_start event? I've included an example below:

{'event': 'on_prompt_start', 'data': {'input': {'input': 'Quelle heure est-il à Moscou ?', 'tools': None, 'llm_config': {'openai_api_base': 'https://iaka-api-dev-2.apps.ocp4.on-prem.innershift.ssghosting.net/llama-3-3-70b/v1', 'openai_api_key': 'ZJUWDkkPOzAPuBfn', 'model': 'llama-3-3-70b', 'temperature': 0.5}, 'chat_history': [], 'intermediate_steps': [[AgentActionMessageLog(tool='get_current_time', tool_input={'timezone': 'Europe/Moscow', 'run_manager': None, 'config': None}, log="\nInvoking: `get_current_time` with `{'timezone': 'Europe/Moscow'}`\n\n\n", message_log=[AIMessageChunk(content='', additional_kwargs={'tool_calls': [{'index': 0, 'id': 'chatcmpl-tool-1f60ee030e91450dabb97a444dec205f', 'function': {'arguments': '{"timezone": "Europe/Moscow"}', 'name': 'get_current_time'}, 'type': 'function'}]}, response_metadata={'finish_reason': 'tool_calls', 'model_name': 'llama-3-3-70b'}, id='run-eeaf8390-687b-466d-b0a2-a10ee953f71d', tool_calls=[{'name': 'get_current_time', 'args': {'timezone': 'Europe/Moscow'}, 'id': 'chatcmpl-tool-1f60ee030e91450dabb97a444dec205f', 'type': 'tool_call'}], tool_call_chunks=[{'name': 'get_current_time', 'args': '{"timezone": "Europe/Moscow"}', 'id': 'chatcmpl-tool-1f60ee030e91450dabb97a444dec205f', 'index': 0, 'type': 'tool_call_chunk'}])]), '{\n  "timezone": "Europe/Moscow",\n  "datetime": "2025-03-13T14:19:08+03:00",\n  "is_dst": false\n}']], 'agent_scratchpad': [AIMessageChunk(content='', additional_kwargs={'tool_calls': [{'index': 0, 'id': 'chatcmpl-tool-1f60ee030e91450dabb97a444dec205f', 'function': {'arguments': '{"timezone": "Europe/Moscow"}', 'name': 'get_current_time'}, 'type': 'function'}]}, response_metadata={'finish_reason': 'tool_calls', 'model_name': 'llama-3-3-70b'}, id='run-eeaf8390-687b-466d-b0a2-a10ee953f71d', tool_calls=[{'name': 'get_current_time', 'args': {'timezone': 'Europe/Moscow'}, 'id': 'chatcmpl-tool-1f60ee030e91450dabb97a444dec205f', 'type': 'tool_call'}], tool_call_chunks=[{'name': 'get_current_time', 'args': '{"timezone": "Europe/Moscow"}', 'id': 'chatcmpl-tool-1f60ee030e91450dabb97a444dec205f', 'index': 0, 'type': 'tool_call_chunk'}]), ToolMessage(content='{\n  "timezone": "Europe/Moscow",\n  "datetime": "2025-03-13T14:19:08+03:00",\n  "is_dst": false\n}', tool_call_id='chatcmpl-tool-1f60ee030e91450dabb97a444dec205f')]}}, 'name': 'ChatPromptTemplate', 'tags': ['seq:step:2'], 'run_id': '079e2a3e-33fa-47e2-b28a-92d2e81af890', 'metadata': {}, 'parent_ids': ['2ffef733-5f39-4bc8-a695-450f16f3a62f', 'a0c872b3-3f24-40df-a495-175552529b34', '59649ecb-7760-4935-ac23-2944723b9fa5']}
ToolMessage must have a tool_call_id. Received:

values={'content': '{\n  "timezone": "Europe/Moscow",\n  "datetime": "2025-03-13T14:19:08+03:00",\n  "is_dst": false\n}', 'additional_kwargs': {'name': 'get_current_time'}, 'response_metadata': {}, 'type': 'tool', 'name': None, 'id': None}

If there's any additional information or context you need from me, please let me know. I'm happy to provide more details or clarify any concerns you may have.

Thank you for your time and consideration!

[Eugene Yurtsev (eyurtsev)]

Hi @dorian-bucaille,

If I modify my local version of libs/core/langchain_core/messages/tool.py with the changes from my PR, the error is not blocking, and the agent can continue to stream its response after calling the tool.

Yes, it makes sense that it will resolve the issue in your case, but it's fixing the problem in the wrong place since by this place in the code the tool call id is required, and we don't want to relax this requirement.

---

Could you include a minimal reproducible example? (i.e., all imports). I should be able to run the example when I copy and paste it. Otherwise it's difficult for me to determine where the issue is.

---

I'd also recommend swapping to langgraph. We primarily recommend langgraph for orchestration these days.

https://langchain-ai.github.io/langgraph/how-tos/create-react-agent/#setup

Tutorial is here:

https://langchain-ai.github.io/langgraph/tutorials/introduction/

[Eugene Yurtsev (eyurtsev)]

We don't want to merge a fix in this part of the code. the underlying issue is happening somewhere else in the code.

[Eugene Yurtsev (eyurtsev)]

Closing this PR for now since the fix is in the wrong place in the code

[lumiaspic (lumiaspic)]

Thank you for your answers. I'll try to provide more detail on the issue later on.