my-optimization

To automate continuous code quality improvement, this skill generates multiple candidate pull requests, runs them in isolated worktrees to avoid cross-contamination, and uses CI/CD and review-based multi-objective rewards to select the best change. It repeats this sampling and selection loop until convergence criteria are met.

Evaluating pull request diffs with change penalties and acceptance thresholds is handled through a PPO-like reward selection process. It computes multi-objective rewards from CI results and reviews, penalizes overly risky diffs using configurable KL/change sensitivity, and safely merges the best-performing PR.

Worktree isolation prevents cross-contamination by promoting generated optimization ideas into separate tasks that execute in isolated Git worktrees. This allows parallel rollouts of multiple candidate changes without affecting the main codebase until reward-based selection identifies the best PR for a safe merge.

Yes, this code evolution workflow supports security hardening and performance optimization by iteratively exploring multiple candidate changes. It applies a convergent evolution loop that samples ideas, executes them in isolated worktrees, computes rewards, and merges the best change under change penalties and acceptance thresholds.

You configure multi-objective reward weighting for CI-based pull request selection by setting your acceptance thresholds and KL/change sensitivity in the task context. The skill uses this PPO-like configuration to penalize risky diffs and compute rewards from CI results and reviews to select the best PR.

After the reward selection process completes, rejected worktrees undergo safe cleanup. The skill ensures that only the best-performing pull request is merged, while the isolated environments for rejected candidate changes are cleaned up to maintain repository hygiene during the continuous evolution loop.