`langchain-deepseek` drops `reasoning_content` when building the request → DeepSeek thinking mode returns 400 on multi-turn tool-call histories (still reproduces on v1.1.0)

[chake111]

Repro: DeepSeek thinking mode returns 400 — reasoning_content must be passed back

Title (suggestion)

langchain-deepseek drops reasoning_content when building the request → DeepSeek thinking mode returns 400 on multi-turn tool-call histories (still reproduces on v1.1.0)

Environment

• OS: Windows 11
• Python: 3.12.13
• langchain: 1.2.12
• langchain-core: 1.2.18
• langchain-deepseek: 1.1.0 (latest, reproduces) / 1.0.1 (also reproduces)
• langchain-openai: 1.1.11
• openai: 2.26.0
• Model: deepseek-v4-flash (thinking mode, enabled by default)
• Base URL: https://api.deepseek.com

Steps to reproduce

pip install -U langchain-deepseek==1.1.0 langchain-openai

import os
from dotenv import load_dotenv
from langchain.chat_models import init_chat_model
from langchain_core.messages import HumanMessage, AIMessage, ToolMessage

load_dotenv()  # provides DEEPSEEK_BASE_URL / DEEPSEEK_API_KEY / DEEPSEEK_MODEL

llm = init_chat_model(
    base_url=os.getenv("DEEPSEEK_BASE_URL"),
    api_key=os.getenv("DEEPSEEK_API_KEY"),
    model=os.getenv("DEEPSEEK_MODEL"),  # deepseek-v4-flash
)

# Hand-crafted multi-turn history: assistant called a tool, tool replied.
# This is the standard shape an agent loop produces after a tool call.
messages = [
    HumanMessage("今天北京天气如何？"),
    AIMessage(
        content="",
        tool_calls=[{
            "name": "get_weather",
            "args": {"location": "北京"},
            "id": "call_00_nUD2NC9QRN5Cg1GaoIkBJQ4s",
        }],
    ),
    ToolMessage(
        content="今天北京天气晴朗，万里无云~",
        tool_call_id="call_00_nUD2NC9QRN5Cg1GaoIkBJQ4s",
    ),
]

llm.invoke(messages)

Actual result

openai.BadRequestError: Error code: 400 - {'error': {'message': 'The `reasoning_content` in the thinking mode must be passed back to the API.', 'type': 'invalid_request_error', 'param': None, 'code': 'invalid_request_error'}}

Expected result

The request should be accepted. The hand-crafted history above is a legitimate tool-call turn; a real agent loop produces exactly this shape and should not hard-fail.

Root cause analysis

langchain_openai._convert_message_to_dict builds the outbound assistant message from content + tool_calls and never merges AIMessage.additional_kwargs (verified against the source of langchain-openai 1.4.2). So even if reasoning_content is present on the message — e.g. captured from a previous DeepSeek response via msg.additional_kwargs["reasoning_content"] — it is dropped before the request is sent.

DeepSeek's official Thinking Mode docs state that for tool-call turns, reasoning_content must participate in context concatenation and must be passed back in all subsequent turns. With thinking mode on, the API rejects the request when it is missing.

Workaround that confirms the diagnosis — disabling thinking mode makes the exact same history pass:

llm = init_chat_model(
    base_url=os.getenv("DEEPSEEK_BASE_URL"),
    api_key=os.getenv("DEEPSEEK_API_KEY"),
    model=os.getenv("DEEPSEEK_MODEL"),
    extra_body={"thinking": {"type": "disabled"}},
)

Related, previously-closed PRs

The same fix was attempted multiple times but every PR was closed without merging:

• langchain_openai: round-trip reasoning_content for OpenAI-compatible providers (DeepSeek fix) #37065 langchain_openai: round-trip reasoning_content for OpenAI-compatible providers (DeepSeek fix) — auto-closed (contributor not assigned to issue)
• fix(deepseek): pass reasoning_content back in multi-turn requests #37179 fix(deepseek): pass reasoning_content back in multi-turn requests — auto-closed (same reason)
• feat(deepseek): add reasoning_content support for v3.2 thinking mode tool calls #34199 feat(deepseek): add reasoning_content support for v3.2 thinking mode tool calls — closed as duplicate of feat(deepseek): support interleaved reasoning in multi-turn tool call loop #34177
• feat(deepseek): support interleaved reasoning in multi-turn tool call loop #34177 feat(deepseek): support interleaved reasoning in multi-turn tool call loop — closed 2026-06-03 by maintainer
• fix(deepseek): preserve reasoning_content in multi-turn conversations #34516 / fix(deepseek): preserve reasoning_content for AIMessages with tool_calls #34438 / fix(deepseek): deepseek-reasoner model cannot return reasoning_content with tools #34092 — same fix, all closed 2026-06-03

The closing comment claimed the failure "no longer reproduces on langchain-deepseek 1.1.0" and that DeepSeek changed the contract so reasoning_content should be excluded from input. That claim is contradicted by:

1. The linked docs page in the closing comment now 404s, and the current Thinking Mode docs still state reasoning_content must be passed back after tool calls.
2. The exact 400 reproduces here on langchain-deepseek 1.1.0 (the version cited in the closing comment) against the live API on 2026-08-10.
3. langchain-deepseek 1.1.0 contains no request-side reasoning_content handling — it only reads the field from responses into additional_kwargs; langchain-openai still drops additional_kwargs on outbound conversion.

Suggested fix

In _convert_message_to_dict (or a DeepSeek-specific subclass), forward AIMessage.additional_kwargs["reasoning_content"] into the outbound assistant message dict (empty string is accepted by the API as a placeholder when no reasoning text exists). This is exactly what the closed PRs above implemented.

[langchain-oss-automated-triage]

This issue was automatically closed because it appears to have been submitted programmatically — issue types are automatically set when using the GitHub web interface, and this issue has none.

We do not allow automated issue submission at this time.

Please use one of the issue templates.