[Bug] Moonshot kimi-k2.5 reasoning_content missing in tool calls leads to API Error 400

[q107580018]

Quick Summary

When using the Moonshot AI model kimi-k2.5 (a reasoning model) via the picoclaw-gateway, the API returns a 400 Bad Request error during multi-turn conversations or tool calls.

Error Log
Plaintext Error processing message: LLM call failed after retries: API request failed: Status: 400 Body: {"error":{"message":"thinking is enabled but reasoning_content is missing in assistant tool call message at index 4","type":"invalid_request_error"}}
Environment
Model: moonshot/kimi-k2.5

PicoClaw Version: Latest (main branch)

[nikolasdehor]

This is the same class of bug as the general reasoning content handling gap discussed in #645, but specific to Moonshot's kimi-k2.5.

The root cause: when thinking mode is enabled, the Moonshot API requires that every assistant message in the conversation history includes a reasoning_content field — including tool call messages. PicoClaw stores conversation history without preserving reasoning_content, so on subsequent turns the API rejects the request because the field is missing from historical assistant messages.

The fix needs to happen at two levels:

1. Provider response parsing — when parsing the Moonshot/OpenAI-compat response, preserve reasoning_content in the stored message struct, not just the final content. This means the Message type in pkg/agent/ needs a ReasoningContent field.

2. Request serialization — when building the messages array for the next API call, include reasoning_content (even if empty string) on assistant messages if the model uses thinking mode.

type Message struct {
    Role             string      `json:"role"`
    Content          string      `json:"content"`
    ReasoningContent string      `json:"reasoning_content,omitempty"`
    ToolCalls        []ToolCall  `json:"tool_calls,omitempty"`
    // ...
}

This is related to #645 (reasoning content silently dropped). Fixing the underlying data model to preserve reasoning content would solve both issues.

Workaround until fixed: disable thinking mode for kimi-k2.5 if the API supports it, or avoid multi-turn tool-calling conversations.