Summary

When using OpenRouter models that support reasoning tokens (e.g. x-ai/grok-4.1-fast, OpenAI o-series, Gemini thinking models), OpenClaw's /think directives (off/low/medium/high) are not mapped to OpenRouter's reasoning.effort API parameter.

Current Behavior

• /think:high sets thinkingLevel internally but this doesn't get passed through to the OpenRouter API body
• OpenRouter expects a reasoning object in the request body:

{
  "reasoning": {
    "effort": "high"
  }
}

• Supported models: Grok (effort levels), Anthropic (max_tokens), Gemini (max_tokens), OpenAI o-series (effort levels)

Expected Behavior

When the provider is openrouter, map OpenClaw's thinking levels to OpenRouter's reasoning parameter:

For Anthropic models via OpenRouter, could also map to reasoning.max_tokens.

Reference

• OpenRouter reasoning docs: https://openrouter.ai/docs/guides/best-practices/reasoning-tokens
• OpenRouter normalizes reasoning.effort across providers (Grok, OpenAI, Gemini, Anthropic)
• The reasoning.exclude option could also be useful (model reasons internally but doesn't return reasoning tokens)

Use Case

Using openrouter/x-ai/grok-4.1-fast as heartbeat model — want reasoning enabled at the API level, not just as a prompt hint. Grok 4.1 Fast is extremely cheap ($0.20/$0.50 per M tokens) so reasoning tokens are very cost-effective.

Environment

• OpenClaw version: 2026.2.9
• Provider: openrouter
• Model: openrouter/x-ai/grok-4.1-fast