create_agent did not remove the reasoning from the message list

[angular-moon]

Checked other resources

• This is a bug, not a usage question.
• I added a clear and descriptive title that summarizes this issue.
• I used the GitHub search to find a similar question and didn't find it.
• I am sure that this is a bug in LangChain rather than my code.
• The bug is not resolved by updating to the latest stable version of LangChain (or the specific integration package).
• This is not related to the langchain-community package.
• I read what a minimal reproducible example is (https://stackoverflow.com/help/minimal-reproducible-example).
• I posted a self-contained, minimal, reproducible example. A maintainer can copy it and run it AS IS.

Example Code

# enable thinking
init_chat_model(
      openai_api_key=Config.MODEL_API_KEY,
      model=Config.MODEL_NAME or "doubao-seed-1-6-251015",
      model_provider="openai",
      base_url=Config.MODEL_API_BASE_URL,
      temperature=0,
      extra_body={"thinking": {"type": "enabled"}},
      stream_usage=True,
      use_responses_api=True,
  )

def get_weather(city: str) -> str:
    """Get weather for a given city."""
    return f"It's always sunny in {city}!"


weather_agent = create_agent(
    model=llm,
    tools=[get_weather],
    system_prompt="You are a helpful assistant",
)

"""
BadRequestError: Error code: 400 - {'error': {'code': 'InvalidParameter', 'message': 'The parameter `input` specified in the request are not valid: item reasoning is not supported. Request id: 02176234456560087423b8d6cd1ec64c1b9b82c48fdb45880e407', 'param': 'input', 'type': 'BadRequest'}}
During task with name 'model' and id '50decdd0-1d8c-d2ba-bd63-ba6d7fa8df3b'
"""
result = weather_agent.invoke(
    input={"messages": [HumanMessage(content="What's the weather like in San Francisco?")]}
)

for msg in result["messages"]:
    print(msg.type, msg.content_blocks)
    print("\n")

Error Message and Stack Trace (if applicable)

Cell In[9], line 25
11 weather_agent = create_agent(
12     model=llm,
13     tools=[get_weather],
14     system_prompt="You are a helpful assistant",
15 )
17 # result = llm.invoke("What's the weather like in San Francisco?")
18
19 # print(result);
(...) 23
24 # Run the agent
---> 25 result = weather_agent.invoke(
26     input={"messages": [HumanMessage(content="What's the weather like in San Francisco?")]}
27 )
29 for msg in result["messages"]:
30     print(msg.type, msg.content_blocks)

File c:\Users\Administrator\Desktop\langchain_demo.venv\Lib\site-packages\langgraph\pregel\main.py:3094, in Pregel.invoke(self, input, config, context, stream_mode, print_mode, output_keys, interrupt_before, interrupt_after, durability,** kwargs)
3091 chunks: list[dict[str, Any] | Any] = []
3092 interrupts: list[Interrupt] = []
-> 3094 for chunk in self.stream(
3095     input,
3096     config,
...
1049 break
1051 assert response is not None, "could not resolve response (should never happen)"

BadRequestError: Error code: 400 - {
    'error': {
        'code': 'InvalidParameter',
        'message': 'The parameter input specified in the request are not valid: item reasoning is not supported. Request id: 02176234456560087423b8d6cd1ec64c1b9b82c48fdb45880e407',
        'param': 'input',
        'type': 'BadRequest'
    }
}
During task with name 'model' and id '50decdd0-1d8c-d2ba-bd63-ba6d7fa8df3b'

Description

The agent created with create_agent reports an error when invoke.
Expectation: Properly handle messages and automatically remove reasoning when calling the llm api.

System Info

System Information

OS: Windows
OS Version: 10.0.26200
Python Version: 3.13.1 (main, Jan 14 2025, 22:47:35) [MSC v.1942 64 bit (AMD64)]

Package Information

langchain_core: 1.0.2
langchain: 1.0.3
langsmith: 0.4.38
langchain_openai: 1.0.1
langgraph_sdk: 0.2.9

Optional packages not installed

langserve

Other Dependencies

claude-agent-sdk: Installed. No version info available.
httpx: 0.28.1
jsonpatch: 1.33
langchain-anthropic: Installed. No version info available.
langchain-aws: Installed. No version info available.
langchain-community: Installed. No version info available.
langchain-deepseek: Installed. No version info available.
langchain-fireworks: Installed. No version info available.
langchain-google-genai: Installed. No version info available.
langchain-google-vertexai: Installed. No version info available.
langchain-groq: Installed. No version info available.
langchain-huggingface: Installed. No version info available.
langchain-mistralai: Installed. No version info available.
langchain-ollama: Installed. No version info available.
langchain-perplexity: Installed. No version info available.
langchain-together: Installed. No version info available.
langchain-xai: Installed. No version info available.
langgraph: 1.0.2
langsmith-pyo3: Installed. No version info available.
openai: 2.6.1
openai-agents: Installed. No version info available.
opentelemetry-api: Installed. No version info available.
opentelemetry-exporter-otlp-proto-http: Installed. No version info available.
opentelemetry-sdk: Installed. No version info available.
orjson: 3.11.4
packaging: 25.0
pydantic: 2.12.3
pytest: Installed. No version info available.
pyyaml: 6.0.3
requests: 2.32.5
requests-toolbelt: 1.0.0
rich: Installed. No version info available.
tenacity: 9.1.2
tiktoken: 0.12.0
typing-extensions: 4.15.0
vcrpy: Installed. No version info available.
zstandard: 0.25.0

[AmazingcatAndrew]

I wrote a test which proves that LangChain actually correctly preserved the field name "thinking" rather than incorrectly converted to "reasoning" or something else:

import os
import sys
import json
import httpx

# ==== Intercept all httpx.Client.send calls ====
def mock_send(self, request, *args, **kwargs):
    print("\n================= [Mocked HTTPX Request] =================")
    print("URL:", request.url)
    try:
        payload = json.loads(request.content.decode()) if request.content else {}
        print("Request Body:\n", json.dumps(payload, indent=2, ensure_ascii=False))
    except Exception:
        print("(No JSON body detected)")
    print("==========================================================")

    # Construct a fake response matching the OpenAI Chat API structure
    response_json = {
        "id": "mock-response",
        "object": "chat.completion",
        "choices": [
            {"message": {"role": "assistant", "content": "(mock) Hello, this is a simulated response."}, "finish_reason": "stop"}
        ]
    }
    return httpx.Response(200, json=response_json, request=request)

httpx.Client.send = mock_send

# ==== Force fix package search path ====
base_dir = os.path.dirname(os.path.dirname(__file__))
sys.path.insert(0, base_dir)

# ==== Import LangChain OpenAI ====
from langchain_openai import ChatOpenAI

# ==== Initialize with fake key/base_url ====
model = ChatOpenAI(
    model="gpt-4o-mini",
    openai_api_key="fake-key",
    base_url="https://mock.endpoint",
    extra_body={"thinking": {"type": "enabled"}},
)

# ==== Invoke and print result ====
print("\nCalling LangChain -> ChatOpenAI.invoke:")
response = model.invoke("Please briefly explain the theory of relativity.")

print("\n================= [MOCK OUTPUT] =================")
print(response)
print("=================================================")

result:

================= [Mocked HTTPX Request] =================
URL: https://mock.endpoint/chat/completions
Request Body:
 {
  "messages": [
    {
      "content": "Please briefly explain the theory of relativity.",
      "role": "user"
    }
  ],
  "model": "gpt-4o-mini",
  "stream": false,
  "thinking": {
    "type": "enabled"
  }
}
==========================================================

This indicates that the problem might be in the server-side parameter parsing. I am willing to take on this issue, please give me 1 week to handle.

[AmazingcatAndrew]

This problem likely doesn't originate from Langchain. You can try using Volcengine's OpenAI-compatible SDK directly. If that also fails, then the problem is that the Doubao API server or its compatibility layer doesn't fully support that field yet, especially since doubao-seed-1-6-251015 is a very new model. Also, if you remove the "thinking" field from the request body, you'll be calling the non-thinking mode of doubao-seed-1-6-251015, as the model supports both thinking and non-thinking modes. @angular-moon

[angular-moon]

@AmazingcatAndrew, thank you for your reply.
After careful analysis, I found that OpenAI currently encourages incorporating reasoning into the context. However, many model providers compatible with the OpenAI API (that do not offer libraries integrated with LangChain) do not support embedding reasoning in the context. Could create_agent provide a parameter to control whether the output of reasoning is included in the context, or offer a built-in middleware? This is actually a quite common requirement.

[AmazingcatAndrew]

I think you are talking about another issue, you can check issue #33672 and PR #33836 (just submitted by me) for more details.

[angular-moon]

Thank you very much. The issue I want to solve is indeed very similar to #33672, but there are some differences. The volcengine provides a responses API that is fully compatible with OpenAI's interface, but the two platforms handle reasoning content within context differently, which leads to the issue I mentioned.