DeepSeek V4 Pro via zen/go: reasoning_content inconsistency causes "must be passed back to the API" error in multi-turn tool calls

[WhiteGiverMa]

Description

When using DeepSeek V4 Pro via the opencode.ai/zen/go/v1 endpoint, multi-turn conversations with tool calls fail intermittently with:

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

Root Cause

DeepSeek V4 Pro's thinking mode is all-or-nothing: once ANY assistant message in the conversation history includes reasoning_content, ALL subsequent assistant messages MUST also include it. If even one message lacks it (e.g., a tool_calls message where reasoning was omitted), the API rejects the entire request.

Reproduction (minimal)

FAILS — first assistant has reasoning_content, 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
}

→ Returns: "The reasoning_content in the thinking mode must be passed back to the API."

PASSES — ALL assistant messages consistently have reasoning_content:

{
  "model": "deepseek-v4-pro",
  "messages": [
    {"role": "user", "content": "读SOUL.md"},
    {"role": "assistant", "content": null, "tool_calls": [{"id":"c1","type":"function","function":{"name":"read","arguments":"{\"path\":\"SOUL.md\"}"}}], "reasoning_content": "需要读文件，调用read工具。"},
    {"role": "tool", "tool_call_id": "c1", "content": "# SOUL\n我是时崎狂三"},
    {"role": "user", "content": "总结"}
  ],
  "tools": [{"type":"function","function":{"name":"read","description":"read","parameters":{"type":"object","properties":{"path":{"type":"string"}}}}}
],
  "max_tokens": 30
}

→ Works fine (200 OK).

Impact

• Any multi-turn conversation using DeepSeek V4 Pro with tool calls (e.g., OpenClaw, Claude Code, Cursor, etc.) will fail after the first tool call
• The zen/go proxy appears to sometimes omit reasoning_content from tool_calls responses, creating the inconsistency
• This is the typical pattern for coding agents where model generates text → calls tools → continues conversation

Suggested Fix

The zen/go proxy should either:

1. Ensure reasoning_content is consistently included in ALL assistant messages when thinking mode is active, including tool call messages, OR
2. Strip reasoning_content from ALL messages in the history before forwarding to DeepSeek if consistency cannot be guaranteed

Environment

• Endpoint: https://opencode.ai/zen/go/v1/chat/completions
• Model: deepseek-v4-pro
• Timestamp: 2026-04-30 03:30 UTC

[github-actions]

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

• DeepSeek thinking mode: reasoning_content not passed back for tool call turns, causing 400 errors #24722: DeepSeek thinking mode: reasoning_content not passed back for tool call turns, causing 400 errors (nearly identical issue)
• Bug: DeepSeek reasoning_content not passed back in multi-turn conversations #24261: Bug: DeepSeek reasoning_content not passed back in multi-turn conversations
• [DeepSeek V4] reasoning_content not round-tripped — 400 error on multi-turn tool calls #24190: [DeepSeek V4] reasoning_content not round-tripped — 400 error on multi-turn tool calls
• [Bug] DeepSeek V4 fails with 'reasoning_content must be passed back' error in multi-turn conversations #24124: [Bug] DeepSeek V4 fails with 'reasoning_content must be passed back' error in multi-turn conversations
• DeepSeek thinking mode: reasoning_content must be passed back to API on conversation continuation #24104: DeepSeek thinking mode: reasoning_content must be passed back to API on conversation continuation

Please check if your issue is already covered by one of these, and consider adding any new details (e.g., specific to the zen/go proxy) as a comment on the most relevant existing issue instead.

[WhiteGiverMa]

Duplicate of #24722 and others. Our contribution: confirmed the same bug affects the zen/go proxy endpoint and provided minimal reproduction with exact error message. Adding detailed findings on #24722.

[WhiteGiverMa]

Reopening — this bug is still unresolved. Keeping this open for visibility. Our zen/go proxy repro adds value: minimal curl reproduction with exact error message. Full context also posted on #24722.

[rekram1-node]

Well deepseeks own endpoints behave this exact same way so this isnt a bug, we may inject the reasoning_content: "" into streams to prevent errors for consumers tho

[WhiteGiverMa]

Ok u r right. Maybe hard to find a more elegant solution.