DeepSeek thinking mode: reasoning_content not passed back for tool call turns, causing 400 errors

[moe5445]

Description

When using DeepSeek models (e.g. deepseek-v4-pro, deepseek-v4-flash) with thinking mode enabled, the API requires that reasoning_content from assistant messages that contained tool calls be included in all subsequent requests. If it's omitted, the API returns a 400 error.

From DeepSeek's docs:

▎ "For turns that do perform tool calls, the reasoning_content must be fully passed back to the API in all subsequent requests. If your code does not ▎ correctly pass back reasoning_content, the API will return a 400 error."

Reproduction:

1. Configure OpenCode to use a DeepSeek model with thinking mode enabled 2. Ask a question that triggers a tool call (e.g. weather lookup, date check, file read)
2. The first turn works, but subsequent turns fail with a 400 error

Root cause:
When serializing assistant messages into the conversation context, OpenCode likely drops reasoning_content from the message object. For non-tool-call
turns DeepSeek ignores missing reasoning_content, but for tool call turns it's mandatory.
Expected behavior:
reasoning_content must be preserved in assistant messages that contain tool_calls when sending the conversation history to DeepSeek's API.

Plugins

No response

OpenCode version

No response

Steps to reproduce

No response

Screenshot and/or share link

No response

Operating System

No response

Terminal

No response

[github-actions]

This issue might be a duplicate of existing issues. Please check:

• DeepSeek V4 Pro (OpenRouter) reasoning_content error - v1.14.27 #24569: DeepSeek V4 Pro (OpenRouter) reasoning_content error — same root cause, reasoning_content not passed back in thinking mode
• deepseek v4 pro开启前思考模式的时候api层丢弃reasoning_content导致接口调用失败 #24714: DeepSeek v4 pro reasoning_content dropped at API layer causing 400 errors — identical symptom and error message
• Thinking blocks in assistant messages are modified/dropped, causing API error on multi-turn conversations #18254: Thinking blocks in assistant messages modified/dropped, causing API errors on multi-turn conversations — same underlying issue

Please review these issues and consider closing this one as a duplicate if the problem is already being tracked.

[keydexofficial]

you may give a try to langcli.com 's Langcli which supports deepseek v4 flash and v4 pro perfectly.
It fixed this reasoning_content problem.
100% compatible with Claude Code.

[karuboniru]

please check this session and related OpenAI API call payload

[WhiteGiverMa]

Additional data point from the zen/go proxy (opencode.ai/zen/go/v1):

Confirmed: same bug affects zen/go endpoint

Minimal zen/go reproduction:

Request with mixed reasoning_content (first assistant has it, second tool_calls does not):

{
  "model": "deepseek-v4-pro",
  "messages": [
    {"role": "user", "content": "帮我查下文件"},
    {"role": "assistant", "content": "好的，我来帮你查。", "reasoning_content": "用户需要查文件，我应该用工具。"},
    {"role": "user", "content": "读SOUL.md"},
    {"role": "assistant", "content": null, "tool_calls": [{"id":"c1","type":"function","function":{"name":"read","arguments":"{\"path\":\"SOUL.md\"}"}}]},
    {"role": "tool", "tool_call_id": "c1", "content": "SOUL内容"},
    {"role": "user", "content": "写个总结"}
  ],
  "tools": [{"type":"function","function":{"name":"read","description":"read","parameters":{"type":"object","properties":{"path":{"type":"string"}}}}}
],
  "max_tokens": 20
}

→ "Error from provider (DeepSeek): The reasoning_content in the thinking mode must be passed back to the API."

Same repro, if ALL assistant messages consistently include reasoning_content (including the tool_calls message), the request succeeds.

Observed behavior

The zen/go proxy sometimes does not return reasoning_content on tool_calls responses, creating the inconsistency that triggers the DeepSeek API rejection.

Suggested fix for zen/go

Either:

1. Ensure reasoning_content is always included in every assistant message (including tool_calls) when thinking mode is active, OR
2. Strip all reasoning_content from messages before forwarding if consistency cannot be guaranteed

This was originally filed as #25000 (now closed as duplicate).