Repository navigation
DeepSeek V4 Pro via zen/go: reasoning_content inconsistency causes "must be passed back to the API" error in multi-turn tool calls #25000
Description
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_contentfrom 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:
- Ensure
reasoning_contentis consistently included in ALL assistant messages when thinking mode is active, including tool call messages, OR - Strip
reasoning_contentfrom 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
Activity
github-actions commentedon Apr 29, 2026
This issue might be a duplicate of existing issues. Please check:
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 commentedon Apr 29, 2026
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 commentedon Apr 29, 2026
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 commentedon Apr 30, 2026
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 commentedon May 1, 2026
Ok u r right. Maybe hard to find a more elegant solution.
Mingran23 commentedon Aug 14, 2026
I can reproduce a very similar issue with OpenCode Go and DeepSeek V4 Flash.
Environment:
Reproduction:
This happens every time in my setup (100% reproducible).
When I disable thinking, the same workflow works normally.
I subscribed to OpenCode Go today (2026-08-14), so this is a current reproduction on the Go subscription path.
Unfortunately, the Operit integration I'm using does not expose the upstream 400 response body. The only error surfaced on my side is:
NotFoundError: User cancelled the requestDevice() chooser.rekram1-node commentedon Aug 14, 2026
Are u using opencode? If not then it is because your harness has a bug.