What problem does it solve? Improving measurable outcomes like clustering quality, search relevance, build latency, or prompt quality often requires trying many variants and scoring each one consistently. This Skill automates that loop: it defines a measurable goal, builds measurement scaffolding, runs parallel experiments in isolated git worktrees, keeps improvements, and converges toward the best solution without losing state across long runs. ## Core Features & Use Cases - Metric-Driven Experiment Loop: Define a YAML optimization spec with hard metrics or LLM-as-judge scoring, degenerate gates, and diagnostics, then run batched experiments that keep winners and revert losers. - Crash-Safe Persistence: Every experiment result is appended to an on-disk experiment log immediately after measurement, with write verification, per-experiment result markers, and resume support after crashes or context compaction. - Isolated Parallel Execution: Experiments run in dedicated git worktrees (or Codex sandboxes) with a parallelism probe that detects port conflicts, shared SQLite databases, lock files, and GPU contention before dispatch. - Use Case: Improve issue clustering quality across 18k tickets by testing embedding, preprocessing, and algorithm hypotheses, using degenerate gates to reject mega-clusters and an LLM judge to score sampled cluster coherence. ## Quick Start Ask the agent to run ce-optimize with a description of your optimization goal, such as reducing build latency while keeping tests green, and approve the generated spec and baseline before the loop begins.
