Kimi for Coding (k2p5) fails with tool calls when thinking enabled: reasoning_content is missing in assistant tool call message

Description

Environment

OpenCode: v1.1.39
OS: Windows (running via WSL2 Ubuntu 22.04)
Provider: kimi-for-coding (via opencode auth login)
Model: k2p5 (Kimi K2.5)
API endpoint (from logs): https://api.kimi.com/coding/v1/messages
Problem
When using Kimi for Coding with thinking enabled, any session that triggers tool calls (e.g. Grep/Read) fails with HTTP 400.
Actual
OpenCode errors:
error, status code: 400, message: thinking is enabled but reasoning_content is missing in assistant tool call message at index 101
Sometimes the index changes as the conversation grows, e.g.:
... missing in assistant tool call message at index 104
Expected
Tool calls should work while thinking is enabled, or OpenCode should disable thinking / handle reasoning_content correctly for tool call messages (or provide a documented toggle/workaround).
Reproduction steps

opencode auth login → choose Kimi for Coding
Start a session using model k2p5 (Kimi K2.5)
Ask the agent to do something that triggers tools, e.g. “grep X in file Y” / “read file”
Observe error 400 above.
Logs (excerpt)
From ~/.local/share/opencode/log/.log:

Request:
- url=https://api.kimi.com/coding/v1/messages
- model=k2p5
- thinking={"type":"enabled","budget_tokens":31999}
Response:
- 400 thinking is enabled but reasoning_content is missing in assistant tool call message at index ...
  Notes / workaround
  Disabling thinking (thinking: disabled) avoids the crash, but removes visible “thinking”.

version'."},"query":{"type":"string","description":"The question or task you need help with. Be specific and include relevant details. Good: 'How to set up authentication with JWT in Express.js' or 'React useEffect cleanup function examples'. Bad: 'auth' or 'hooks'. IMPORTANT: Do not include any sensitive or confidential information such as API keys, passwords, credentials, or personal data in your query."}},"required":["libraryId","query"],"additionalProperties":false,"$schema":"http://json-schema.org/draft-07/schema#"}}],"tool_choice":{"type":"auto"},"stream":true},"statusCode":400,"responseHeaders":{"cf-cache-status":"DYNAMIC","cf-ray":"9c503405bf4c1f6e-SIN","connection":"keep-alive","content-length":"200","content-type":"application/json; charset=utf-8","date":"Wed, 28 Jan 2026 11:38:23 GMT","server":"cloudflare","strict-transport-security":"max-age=31536000; includeSubDomains"},"responseBody":"{\"error\":{\"type\":\"invalid_request_error\",\"message\":\"error, status code: 400, message: thinking is enabled but reasoning_content is missing in assistant tool call message at index 101\"},\"type\":\"error\"}","isRetryable":false,"data":{"type":"error","error":{"type":"invalid_request_error","message":"error, status code: 400, message: thinking is enabled but reasoning_content is missing in assistant tool call message at index 101"}}}} stream error
ERROR 2026-01-28T11:38:22 +3ms service=session.processor error=error, status code: 400, message: thinking is enabled but reasoning_content is missing in assistant tool call message at index 101 stack="AI_APICallError: error, status code: 400, message: thinking is enabled but reasoning_content is missing in assistant tool call message at index 101\n    at <anonymous> (../../node_modules/.bun/@ai-sdk+provider-utils@3.0.20+d6123d32214422cb/node_modules/@ai-sdk/provider-utils/dist/index.mjs:936:18)\n    at processTicksAndRejections (native:7:39)" process
INFO  2026-01-28T11:38:22 +0ms service=bus type=session.error publishing
INFO  2026-01-28T11:38:22 +2ms service=bus type=message.updated publishing

Plugins

dcp, auth antigravity

OpenCode version

v1.1.39

Steps to reproduce

Run opencode auth login and select Kimi for Coding
Start a new session using model k2p5
Ask the agent to perform an action that triggers tool calls, e.g.:
- “Grep subject_anchor in windows-app/story_presets.json”
Observe the request fails with HTTP 400:
thinking is enabled but reasoning_content is missing in assistant tool call message at index ...

Workaround
If thinking is disabled, tool calls work.

Screenshot and/or share link

Operating System

windows 11 pro, but i use WSL full opencode

Terminal

wezterm

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

can not use deepseek-api in opencode #10788: DeepSeek API fails with missing reasoning_content field in assistant tool call messages - same root cause with different provider
Deepseek-chat:options:thinking:type:enabled thinking tools bug #9397: DeepSeek with thinking enabled fails on tool calls - same error pattern and root cause
DeepSeek 3.2: Missing reasoning_content #4947: DeepSeek 3.2 missing reasoning_content - similar historical issue
Claude Extended Thinking: assistant message content order causes API error #9364: Claude Extended Thinking fails with tool calls - same thinking + tool call issue on different provider
Kimi-K2 via OpenRouter fails with tool call tokens inside thinking block #8851: Kimi-K2 via OpenRouter tool call tokens in thinking block - related Kimi provider issue

This appears to be a regression affecting multiple providers (Kimi, DeepSeek, Claude) when thinking/reasoning is enabled and tool calls are attempted. Previous issues suggest this was fixed before but the solution may have been reverted.

Feel free to ignore if this is a distinct case from the above.

I'm seeing the same thing, seems likely to be a setting we need to tweak

Hm it should be using the anthropic messages endpoint, not openai, reasoning_content makes it seem like it's using the openai one?

This issue was resolved because of the dcp plugin.

Hm it should be using the anthropic messages endpoint, not openai, reasoning_content makes it seem like it's using the openai one?

This issue was resolved because of the dcp plugin.

@rekram1-node

Hm it should be using the anthropic messages endpoint, not openai, reasoning_content makes it seem like it's using the openai one?

im will close this issue, thnk u so much @rekram1-node

I ran into this problem too. Could you please tell me how was it resolved afterwards?I tried using anthropic messages endpoint but it didnt work.

@willy-scr

This issue was resolved because of the dcp plugin.

Hm it should be using the anthropic messages endpoint, not openai, reasoning_content makes it seem like it's using the openai one?

This issue was resolved because of the dcp plugin.

I am using https://github.com/Opencode-DCP/opencode-dynamic-context-pruning that @willy-scr recommended but I agree with @zero-zrc0

I ran into this problem too. Could you please tell me how was it resolved afterwards?I tried using anthropic messages endpoint but it didnt work.

Using this DCP plugin seems more like a workaround instead of actually fixing the provider. Unless it means we need a LiteLLM Proxy to properly seperate thinking tags and tool calls.

Might be a different stack, but I was facing the same issue with ai-sdk. Here's what solved it for me

// reasoning_content — Kimi's API requires this field on every assistant
// tool-call message when thinking mode is enabled.
const kimiProvider = createOpenAICompatible({
  name: 'moonshot',
  baseURL: 'https://api.moonshot.ai/v1',
  apiKey: KIMI_API_KEY,
  includeUsage: true,
  transformRequestBody: (body: Record<string, unknown>) => {
    const messages = body.messages
    if (!Array.isArray(messages)) return body
    for (const msg of messages) {
      if (
        msg.role === 'assistant' &&
        Array.isArray(msg.tool_calls) &&
        msg.tool_calls.length > 0 &&
        !msg.reasoning_content
      ) {
        msg.reasoning_content = '.'
      }
    }
    return body
  },
})

fix: add KIMI_THINKING_DISABLED env var + revert reasoning_content pa…