What problem does it solve? Plugin quality fixes often address only the surface symptom, leaving deeper defect layers to resurface across slow file-fix-reopen cycles, and self-closed fixes may never be verified against real silicon. This Skill closes that loop by drilling each systematic gap to a fully-converged state in a sandbox worktree and auditing every core-closed issue on the actual benchmark. ## Core Features & Use Cases - Deep-resolution before handoff: Prototype chip-agnostic fixes in a throwaway git worktree, peel every failing layer until aggregate probes report zero residual across two clean-room rounds, then hand the gatekeeper one verified bundle (candidate.patch plus per-layer regression tests). - Deterministic closed-audit rule: At every cron tick, run check_closed_for_field_audit.sh to find CLOSED issues labeled core-closed but lacking field-verified, then verify each on the real benchmark and reopen with counter-evidence when inadequate. - Artifact-first verification: Classify each fix with fix_surface_classify.py so consumer-only changes are verified in seconds against persisted reports instead of triggering 40-minute phase3 re-runs. - Use Case: A maintainer schedules a cron against a benchmark IC folder; each tick the loop audits closed fixes, drills newly found gaps to convergence, files sanitized ORGANIC backlog issues, and stops only when no new gaps and no un-audited closures remain. ## Quick Start Run the field-agent loop as a cron against the benchmark IC folder at benchmark-data/ic/<target> with an LLM-review prompt to audit closed issues and file converged fix bundles.
