openclaw - 💡(How to fix) Fix Kimi K2.5 (anthropic-messages) multi-turn tool calling fails: reasoning_content missing in assistant tool_use messages [1 participants]

Bug Description

When using Kimi K2.5 (or any Moonshot model) via the anthropic-messages API format with thinking/reasoning enabled, multi-turn tool calling fails with a 400 Bad Request error on the second and subsequent turns.

Error Message

thinking is enabled but reasoning_content is missing in assistant tool call message at index N

Root Cause

Moonshot/Kimi's Anthropic-compatible endpoint has a strict requirement: when thinking is enabled, every assistant message containing tool_use blocks must also include a reasoning_content field in the conversation history. Even an empty string is rejected; a minimum of one space is required.

The current convertMessages function in anthropic.js does not populate reasoning_content for assistant messages with tool calls when replaying conversation history.

Steps to Reproduce

1. Configure OpenClaw with Kimi K2.5 via anthropic-messages API format:

   model: kimi/k2p5
   api: anthropic-messages
   thinking: true

2. Send a message that triggers a tool call (e.g., use web_search or any custom tool).
3. The first request succeeds — Kimi returns thinking blocks + tool_use blocks.
4. Send a follow-up message in the same conversation.
5. OpenClaw replays the assistant message (with tool_use) back to Kimi, but without reasoning_content.
6. Kimi API returns 400 Bad Request with the error above.
7. OpenClaw falls back to the next model in the fallback chain (e.g., GPT-5.4).

Expected Behavior

Multi-turn tool calling should work correctly with Kimi K2.5 when thinking is enabled, preserving reasoning_content in assistant messages that contain tool_use.

Environment

• OpenClaw version: 2026.4.21
• Node.js: v22.22.2
• Model: kimi/k2p5 (also reproducible with moonshot-v1-128k)
• API format: anthropic-messages

Related Issues

• #57573 — Same bug in the Kimi web search provider path (fixed by PR #59356)
• #16952 — Earlier report of the same error string, possibly related

Suggested Fix

In convertMessages within anthropic.js, when the provider is Kimi/Moonshot and the assistant message contains tool_use blocks, collect any thinking blocks into reasoning_content and always include it (using " " as fallback if empty):

const isKimiAnthropic = model.provider === "kimi" || model.provider === "moonshot" || 
                        model.baseUrl?.includes("moonshot.cn") || model.baseUrl?.includes("kimi.com");

// ... when building assistant message ...
if (isKimiAnthropic && hasToolUse) {
    const rc = reasoningChunks.join("\n\n");
    assistantMessage.reasoning_content = rc.length > 0 ? rc : " ";
}

This is the same category of fix that was applied to the web search provider in PR #59356, but needs to be applied to the main model path as well.

Additional Context

This bug is not unique to OpenClaw — the same root cause has been reported in multiple projects:

• CoPaw #388 — fixed in PR #707
• nanobot #390
• LiteLLM #21672

All of these share the same pattern: Moonshot's stateless API requires clients to resend reasoning_content in assistant messages during multi-step tool calling, but the client library's message serialization omits this field.