Fix OpenClaw 400 Reasoning Loop (GPT-5.2 / o-Series)

When you use GPT-5.2, o3, o4-mini, or any OpenAI reasoning model with OpenClaw, an abruptly terminated agent turn can leave a reasoning item (rs_...) in the conversation history without its required follow-up assistant message. OpenAI's API then rejects every subsequent request with a 400, trapping you in an unrecoverable loop.

Identifying the Error

400 Bad Request
{
  "error": {
    "message": "Item rs_abc123... of type `reasoning` was provided without its
                required following item of type `assistant`.",
    "type": "invalid_request_error",
    "code": "invalid_conversation_structure"
  }
}

This error appears in the OpenClaw gateway logs and in the chat interface. The key phrase is of type reasoning was provided without its required following item. Once it appears for one message, it will repeat for every subsequent message in the same session.

Fix 1: Clear Conversation History (Fastest)

The orphaned reasoning item lives in the conversation history that OpenClaw sends to the API. Clearing history removes it immediately.

# In the OpenClaw chat interface, type:
/history clear

# Or via CLI:
openclaw history clear

Fix 2: Switch to a Non-Reasoning Model Temporarily

If you need to preserve the conversation, switch to a model that does not use reasoning tokens (e.g., gpt-4o or claude-sonnet-4-6). The bad history can then be summarised by the non-reasoning model before you switch back.

# openclaw.json — temporarily set a non-reasoning model
{
  "model": "openai/gpt-4o",
  "parameters": {}
}

# After clearing context, switch back to your preferred model

Fix 3: Lower Reasoning Effort to Prevent Future Loops

High reasoning effort (reasoning_effort: "high") produces longer reasoning chains, which are more likely to be orphaned if a run is interrupted. Reduce the effort level:

{
  "model": "openai/gpt-5.2",
  "parameters": {
    "reasoning_effort": "medium"
  }
}

Prevention Tips

• Never force-quit OpenClaw mid-turn. Use Ctrl+C once and wait for the gateway to flush the turn cleanly.
• Keep sessions short. Start a new session for unrelated tasks instead of reusing a long-running context.
• Use "Save summary" before ending long sessions. Ask the agent to summarise the conversation, then start fresh with the summary as context.
• Pin a stable model. Avoid using preview or daily-build model aliases that may change reasoning format between calls.