feat: Add thinking-level benchmarking support

[ScuttleBot]

Clean rebase of #76 (originally from #12 by @jb510) against current main.

Summary

Adds --thinking flag to configure reasoning depth for models that support it.

Changes

• Add VALID_THINKING_LEVELS constant: off, minimal, low, medium, high, xhigh, adaptive
• Add thinking_level parameter to execute_openclaw_task()
• Pass --thinking to openclaw agent command
• Add --thinking argument to benchmark.py with validation
• Update README command reference

Usage

# Run with high reasoning depth
./scripts/run.sh --model openrouter/anthropic/claude-sonnet-4 --thinking high

# Run with minimal thinking for speed
./scripts/run.sh --model openrouter/anthropic/claude-sonnet-4 --thinking minimal

Why a new PR?

The original PR #76 had extensive merge conflicts due to main branch evolution (parallel judging, categories, incremental results, etc.). Rather than resolve 6 conflict blocks across 3 files, this is a clean implementation of the same feature on current main.

Supersedes #76.

Co-authored-by: ForceConstant ForceConstant@users.noreply.github.com
Co-authored-by: jb510 jb510@users.noreply.github.com

[kilo-code-bot]

Code Review Summary

Status: No Issues Found | Recommendation: Merge

Clean implementation of thinking-level support. The thinking_level value is validated against a whitelist before use and passed to subprocess.run as a list argument (not shell-interpolated), so there's no injection risk. The optional parameter flows correctly through the call chain with proper None defaults.

Files Reviewed (3 files)

• scripts/lib_agent.py — VALID_THINKING_LEVELS constant + thinking_level param in execute_openclaw_task()
• scripts/benchmark.py — --thinking CLI arg with validation
• README.md — docs update

---

_{Reviewed by claude-4.6-sonnet-20260217 · 99,364 tokens}