Add --thinking flag to orchestrator for reasoning depth control

[ScuttleBot]

Summary

Adds --thinking flag to orchestrate_vultr.py so users can set reasoning depth when benchmarking models that support it (e.g., mercury-2).

How it works

1. Orchestrator: --thinking medium writes the level to /root/benchmark_thinking.txt on each Vultr instance
2. Bench runner: reads the file and passes --thinking <level> to benchmark.py
3. Benchmark: already supports --thinking and passes it through to OpenClaw

Usage

# Single model with medium reasoning
uv run orchestrate_vultr.py --models openrouter/inception/mercury-2 --thinking medium

# Batch with high reasoning
uv run orchestrate_vultr.py --models model1 model2 --count 2 --thinking high

Valid levels

off, minimal, low, medium, high, xhigh, adaptive

Notes

• The benchmark.py script already had --thinking support — this just wires it through the orchestrator
• If --thinking is omitted, behavior is unchanged (uses model default)
• Follows the same file-based pattern as --no-fail-fast and --official-key

[kilo-code-bot]

Code Review Summary

Status: No Issues Found | Recommendation: Merge

Solid implementation. The file-based handoff pattern is consistent with existing --no-fail-fast and --official-key handling. Shell escaping with single-quote doubling ('\'') is correct, the atomic write (.tmp + mv) prevents partial reads, and the bash array expansion "${THINKING_ARG[@]}" is idiomatic and safe.

Files Reviewed (2 files)

• bench_runner.sh — thinking file read and arg forwarding to benchmark.py
• orchestrate_vultr.py — --thinking CLI arg, SSH file write with atomic rename
• docs/benchmark-observability-plan.md — documentation only, no code issues

---

_{Reviewed by claude-4.6-sonnet-20260217 · 97,911 tokens}