evolution-looplisted

# Evolution Loop This skill gives `best-copilot` a lightweight, auditable evolution cycle inspired by Evolver and Memento-style reflective learning. ## Core Principle Evolution is not free-form rewriting. It is a bounded loop: 1. **Read** verified signals from task results, review findings, failed commands, user corrections, memory, and spec drift. 2. **Select** the smallest reusable improvement target: agent, skill, instruction, memory rule, README guidance, or spec template. 3. **Propose** a protocol-bound mutation with evidence, scope, rollback, and validation. 4. **Validate** with static checks, eval prompts, review, or command evidence. 5. **Write** only accepted learnings back to the canonical root `agents/`, root `skills/`, `.github/instructions/**`, `memories/repo/**`, or `spec/**` owner in the target repository. ## EvolutionEvent Record each accepted evolution as a compact event: ```markdown ## EvolutionEvent: <YYYY-MM-DD-topic> - signal: - target: - mutation: - validation: - rollback: - status: proposed|accepted|rejected|deprecated ``` ## Strategy Presets - `repair-only`: fix broken trigger, broken route, false claim, failed validation. - `harden`: reduce ambiguity, add guardrails, improve verification, prevent repeated errors. - `balanced`: small improvement that preserves current workflow. - `innovate`: add a new capability only when existing skills cannot cover repeated demand. ## Guardrails - Do not copy external prompts or code wholesale; extract loc