Add --reasoning flag for models that support thinking/reasoning levels

[ScuttleBot]

Adds a --reasoning CLI argument to benchmark.py that gets passed through to the OpenClaw agent's models.json configuration. This enables benchmarking models like inception/mercury-2 with different reasoning levels (low, medium, high).

Usage

# Run mercury-2 with medium reasoning
python3 scripts/benchmark.py --model openrouter/inception/mercury-2 --reasoning medium

Changes

• benchmark.py: add --reasoning argument, pass to ensure_agent_exists()
• lib_agent.py: accept reasoning param, apply to model config for both custom endpoints and standard OpenRouter flow

The reasoning value is set on the model entry in models.json and passed through to the provider API (e.g. OpenRouter's reasoning parameter).

[kilo-code-bot]

Code Review Summary

Status: No Issues Found | Recommendation: Merge

The new commit adds _apply_reasoning_to_model and _set_agent_thinking_default helpers, plus wires the reasoning param through both the custom-endpoint and OpenRouter paths. The implementation is solid:

• Error handling is consistent throughout (try/except with logger.warning fallbacks)
• The custom-endpoint path sets reasoning directly on the model entry dict
• The OpenRouter path correctly calls _apply_reasoning_to_model with fallback model-entry creation
• _set_agent_thinking_default safely reads and writes the OpenClaw global config using setdefault to avoid KeyErrors

Files Reviewed (2 files)

• scripts/benchmark.py
• scripts/lib_agent.py

---

_{Reviewed by claude-4.6-sonnet-20260217 · 116,618 tokens}

[ScuttleBot]

Closing this PR — PinchBench already has a --thinking flag that correctly sets thinking levels via OpenClaw's --thinking CLI option. Verified: running with --thinking medium produces transcripts with thinkingLevel: medium. No additional --reasoning flag is needed.