kimi 有问题 claude调用新版本有错误

[q2467481940]

kimi 有问题 claude调用新版本有错误 Thought for 6s, searched for 1 pattern, listed 1 directory (ctrl+o to expand)
⎿  API Error: 400 thinking is enabled but reasoning_content is missing in assistant tool call message at index 7

[luispater]

kimi建议使用OAuth的方式登录使用，你需要上传request log，来定位具体问题。

[orange-XD]

同样的问题，使用OAuth也是一样，已经回退CPA版本了

[MunMunMiao]

我提交了修复 #3719

[q2467481940]

API Error: 400 thinking is enabled but reasoning_content is missing in assistant tool call message at
index 6

[q2467481940]

kimi建议使用OAuth的方式登录使用，你需要上传request log，来定位具体问题。

API Error: 400 thinking is enabled but reasoning_content is missing in assistant tool call message at
index 6

[MunMunMiao]

@luispater 看下能否合并

[q2467481940]

@luispater 看下能否合并
你测试了吗 这个不会异常消耗token吗 我之前本地修了一下 导致没有缓存了

[luispater]

@luispater 看下能否合并

不能合并，你在claude执行器中增加了第三方兼容的代码，这是不允许的。

claude执行器是为了claude而设计工作的。

正确的做法应该是修改kimi执行器。或者使用插件对请求及响应进行规整。

[sususu98]

Proposed fix direction (aligned with maintainer guidance)

This error is primarily a Kimi upstream protocol issue, not a missing entry in the Claude/GPT/Gemini signature-compatibility matrix:

400 thinking is enabled but reasoning_content is missing in assistant tool call message

When thinking is enabled, Kimi expects assistant messages that contain tool_calls to also carry reasoning_content. A Claude-format history that only has thinking blocks (or empty reasoning) can fail on Kimi even if the request was translated through a Claude-shaped path.

What we should not do

• Do not add third-party provider compatibility into claude_executor. That executor is intended for native Claude targets only (as noted in the thread).

What we should do

1. Kimi-specific request normalization in kimi_executor

   • On the OpenAI chat-completions path, ensure assistant tool_calls messages get a non-empty reasoning_content when thinking is enabled (e.g. backfill from prior assistant reasoning / thinking text where safe).
   • This belongs in normalizeKimiToolMessageLinks() (or a dedicated Kimi helper), not in Claude upstream sanitization.

2. Avoid Claude-target signature sanitization for non-Claude models

   • Some Kimi routes delegate to ClaudeExecutor when source_format=claude. Previously, claude_executor always ran SanitizeClaudeMessagesForClaudeUpstream, which drops incompatible thinking blocks (whole block, not signature-only). That can make Kimi-side failures worse or change history shape unexpectedly.
   • Fix landed locally: gate sanitization so it runs only when the requested model resolves to the Claude provider family (claude-*). Unknown models such as kimi-* / mimo-* bypass the request-side signature matrix on that path.

Verification checklist

• Reproduce with request log: confirm provider=kimi, endpoint (/v1/chat/completions vs /v1/messages), and whether the failing assistant index has tool_calls without reasoning_content.
• After Kimi normalization: same multi-turn tool-call conversation should pass Kimi 400 check.
• After Claude sanitizer gate: Kimi/Mimo models routed through ClaudeExecutor should no longer have thinking blocks stripped solely because Claude-target rules were applied.

Relation to #3719

PR #3719 may address symptoms but should not be merged if it adds third-party compatibility inside claude_executor. The durable fix is Kimi executor (or plugin) + big3-only signature policy.

Happy to open a focused PR for the claude_executor gate + Kimi reasoning_content backfill if maintainers want that split.

[orange-XD]

更新到最新版，实测已经没有这个问题了