Автономное улучшение кода через турнирный отбор

★ 8.5 · vibe-coding

self-improve is a Claude Code skill that runs a fully autonomous iterative code-improvement loop with tournament selection. Specialized agents handle each phase: a researcher analyses the codebase and generates hypotheses, a planner converts them into structured plans, an architect and a critic review each plan against six criteria and harness rules (H001/H002/H003), and executors implement changes and run the configured benchmark command. All artifacts are stored under `.omc/self-improve/topics/{slug}/` — iteration history, merge reports, raw scores, and a progress chart. The loop never pauses for confirmation; on agent failure it retries once and continues, while `validate.sh` prevents the benchmark code from being modified. It is aimed at developers who need hands-off performance optimisation of a repository against a measurable target metric.