langchain - 💡(How to fix) Fix ChatOpenAI drops reasoning_content from OSS models (like vLLM/DeepSeek) during AIMessage conversion

Submission checklist

• 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 posted a self-contained, minimal, reproducible example. A maintainer can copy it and run it AS IS.

Package (Required)

• langchain
• langchain-openai
• langchain-anthropic
• langchain-classic
• langchain-core
• langchain-model-profiles
• langchain-tests
• langchain-text-splitters
• langchain-chroma
• langchain-deepseek
• langchain-exa
• langchain-fireworks
• langchain-groq
• langchain-huggingface
• langchain-mistralai
• langchain-nomic
• langchain-ollama
• langchain-openrouter
• langchain-perplexity
• langchain-qdrant
• langchain-xai
• Other / not sure / general

Related Issues / PRs

https://github.com/langchain-ai/langchain/issues/37347

Reproduction Steps / Example Code (Python)

1. Initialize `ChatOpenAI` pointing to a local `vLLM` server serving a reasoning model (e.g., `DeepSeek-R1`).

from langchain_openai import ChatOpenAI

llm = ChatOpenAI(
    model="openai/gpt-oss-20b",
    openai_api_base="http://localhost:8000/v1",
    openai_api_key="EMPTY"
)

response = llm.invoke("What is 10 + 10? Please think step by step.")

2. The raw response from vLLM contains `reasoning_content` in the `choices[0].message` payload.
3. Check `response.additional_kwargs`. The `reasoning_content` is missing, and `response.content` only contains the final output without the reasoning trace.

Description

When using langchain_openai.ChatOpenAI to connect to an OpenAI-compatible endpoint that serves open-source reasoning models (e.g., via vLLM or Ollama), the reasoning_content field returned in the raw JSON response is silently dropped.

Many users serve models like DeepSeek-R1 via vLLM, which utilizes the reasoning_content field for the model's thought process. However, because _convert_dict_to_message in langchain_openai/chat_models/base.py hardcodes the extraction of content, function_call, tool_calls, and audio, any additional non-standard fields like reasoning_content are destroyed before the AIMessage is returned.

Steps to Reproduce

1. Initialize ChatOpenAI pointing to a local vLLM server serving a reasoning model (e.g., DeepSeek-R1).

from langchain_openai import ChatOpenAI

llm = ChatOpenAI(
    model="deepseek-r1-20b",
    openai_api_base="http://localhost:8000/v1",
    openai_api_key="EMPTY"
)

response = llm.invoke("What is 10 + 10? Please think step by step.")

1. The raw response from vLLM contains reasoning_content in the choices[0].message payload.
2. Check response.additional_kwargs. The reasoning_content is missing, and response.content only contains the final output without the reasoning trace.

Expected Behavior

ChatOpenAI should either:

1. Dump unhandled payload fields into additional_kwargs (e.g., additional_kwargs["reasoning_content"] = _dict.get("reasoning_content")) so that users can still access custom fields.
2. Provide a parameter like include_extra=True to preserve extra fields from the raw provider response.

Code snippet where the issue occurs:

In _convert_dict_to_message (langchain_openai/chat_models/base.py):

    if role == "assistant":
        content = _dict.get("content", "") or ""
        additional_kwargs: dict = {}
        
        if function_call := _dict.get("function_call"):
            additional_kwargs["function_call"] = dict(function_call)
            
        tool_calls = []
        invalid_tool_calls = []
        if raw_tool_calls := _dict.get("tool_calls"):
            # ... parses tool calls ...

        if audio := _dict.get("audio"):
            additional_kwargs["audio"] = audio
            
        # reasoning_content is ignored!
            
        return AIMessage(
            content=content,
            additional_kwargs=additional_kwargs,
            name=name,
            id=id_,
            tool_calls=tool_calls,
            invalid_tool_calls=invalid_tool_calls,
        )

Workaround

Currently, users are forced to use ChatDeepSeek (which overrides this method to preserve reasoning_content), even if the model isn't strictly a DeepSeek model, just to preserve the vLLM response format. Passing additional_kwargs dynamically in the base class would resolve this limitation natively.

System Info

• langchain-openai==1.2.2
• langchain-core==1.4.1
• langchain==1.3.4