Harness GC (Garbage Collection)

harness-gc applies OpenAI's 'Harness Engineering' garbage-collection model to your codebase. Three audit agents run in parallel — doc-auditor (doc freshness vs. reality), arch-inspector (layer drift, cycles, dead code), quality-scorer (domain quality A–F + 12-principle maturity L1–L5) — then gc-synthesizer consolidates findings into a single report with auto-fix, semi-auto, and manual buckets. Maturity scoring follows the 12 harness principles (Entry Point, Map-not-Manual, Invariant Enforcement, Convention, Progressive Disclosure, Layered Arch + Coverage, GC Automation, Observability, Knowledge in Repo, Reproducibility, Modularity, Self-Documentation). Each principle is anchored with 0-10 few-shot rubrics; the evaluator stays skeptical by default ('prove it's good' instead of 'prove it's bad') and adds rather than subtracts from zero. Knip strict is the authoritative source for dead code so the report doesn't disagree with the linter. The `--remediate` mode closes the loop: quality-scorer hands the fixer agent concrete diagnostic findings, the two negotiate a Sprint Contract with expected score deltas, Auto/Semi actions are applied, and the suite re-audits. If actual deltas underperform expectations twice in a row, the loop reports `approach-exhausted` rather than wasting another iteration on a strategy that isn't working. A harness self-audit (Part F) flags agents that haven't found a real issue in three consecutive runs — a `💡 SIMPLIFY` tag suggesting that part of the harness may have outlived its usefulness as the underlying model improves.