openclaw - 💡(How to fix) Fix reasoning_content not preserved for non-DeepSeek OpenAI-compatible providers (e.g. Xiaomi MiMo)

Bug Description

OpenAI-compatible providers that require reasoning_content to be passed back in multi-turn conversations (e.g. Xiaomi MiMo) fail with a 400 error when tool calls are involved. This works for DeepSeek because OpenClaw has dedicated handling (createDeepSeekV4OpenAICompatibleThinkingWrapper + ensureDeepSeekV4AssistantReasoningContent), but other providers with the same API requirement are not covered.

Steps to Reproduce

1. Configure an OpenAI-compatible provider with a reasoning model (e.g. Xiaomi MiMo mimo-v2.5 with reasoning: true)
2. Send a message that triggers a tool call
3. The model returns reasoning_content + tool_calls
4. OpenClaw sends the tool result back, but does not include the previous reasoning_content in the assistant message
5. API returns: 400 Param Incorrect — "The reasoning_content in the thinking mode must be passed back to the API."

Expected Behavior

OpenClaw should preserve reasoning_content in assistant messages for all OpenAI-compatible providers that return it, not just DeepSeek.

Error Logs

[model-fallback/decision] model fallback decision: decision=candidate_failed requested=xiaomi/mimo-v2.5 candidate=xiaomi/mimo-v2.5 reason=format next=none detail=400 Param Incorrect
[agent/embedded] embedded run agent end: isError=true model=mimo-v2.5 provider=xiaomi error=LLM request failed: provider rejected the request schema or tool payload. rawError=400 Param Incorrect

Root Cause

In provider-stream-shared-BUAWmRcw.js:

• ensureDeepSeekV4AssistantReasoningContent adds reasoning_content: "" to assistant messages missing it
• createDeepSeekV4OpenAICompatibleThinkingWrapper applies this only when shouldPatchModel returns true (DeepSeek V4 models only)

Other providers (Xiaomi, potentially Kimi, etc.) with the same reasoning_content requirement are not covered.

Suggested Fix

Generalize the reasoning_content preservation logic so it applies to any OpenAI-compatible provider/model that returns reasoning_content in its response, rather than gating it behind a DeepSeek-specific model check.

Workaround

Set reasoning: false for the model in openclaw.json to disable thinking mode and avoid the reasoning_content requirement.

Environment

• OpenClaw version: 2026.5.7
• Provider: Xiaomi MiMo (token-plan-cn.xiaomimimo.com/v1)
• Model: mimo-v2.5
• Platform: macOS