Repository navigation
[BUG] DeepSeek error: 'reasoning_content' must be passed back to the API #201
Description
Describe the bug
DeepSeek error: 'reasoning_content' must be passed back to the API
To Reproduce
Steps to reproduce the behavior:
- Create a DeepSeek provider with deepseek-v4-pro and your API.
- Set Model Reasoning Effort to High.
- Run a task.
Expected behavior
I expected it to run smoothly, but I get this:
Provider Error · 400
The provider couldn't process the request as made. Stop the task and try a different approach.
Details: DeepSeek completion error: 400 The reasoning_content in the thinking mode must be passed back to the API.
What version of zoo are you running
i.e. 3.54.1
Activity
leosdad commentedon May 19, 2026
I see now that this issue and anomalyco/opencode#24104 are probably the same. Sorry for the oversight.,
edelauna commentedon May 23, 2026
Root cause identified.
The 400 error occurs because DeepSeek's thinking-mode API requires
reasoning_contentto be present on every assistant message in the conversation history, not just the one that immediately precedes the current request.Why it happens:
When the conversation history contains an assistant message that lacks
reasoning_content— either from a prior turn where DeepSeek generated no thinking output, or from a message generated by a different model (e.g. if you have different models configured for different modes/agents) — DeepSeek can't distinguish it from one of its own thinking-mode responses and rejects the request.The specific code path:
convertToR1Formatinr1-format.tsonly setsreasoning_contenton assistant messages where it can extract a reasoning block from history. Messages without reasoning blocks get noreasoning_contentfield, which triggers the 400.Proposed fix:
After
convertToR1Format, when thinking mode is active, normalize all assistant messages that are missingreasoning_contentby setting it to"". An empty string signals "no thinking this turn" rather than a missing field, which DeepSeek misreads as dropped thinking content.This is confirmed to reproduce with mixed-model setups (e.g. DeepSeek for code mode + another model for debug mode) and in long single-model sessions where some turns produce no reasoning output.
feigou6545456 commentedon Jun 30, 2026
啥时候修复？