What problem does it solve? Hardening a local agent harness against unauthorized repository writes requires systematic adversarial testing of write gates, confirm/reason contracts, kill switches, and clone jails — ad-hoc review easily misses bypasses or accidentally loosens asserts to make tests green. ## Core Features & Use Cases - Corpus-driven redteam loop: Runs existing Rust integration tests (real_repo_loop, agentic_foundations, shim_and_agent_routes, invariant_guard) as the living adversarial suite against the write surface. - Finding classification: Buckets results into new_bypasses, known_gaps, fixed_findings, and false_greens so regressions are fixed before anything else and verification failures are detected. - Minimal gate fixes with regression tests: Maps each finding to a family (gate order, confirm+reason, kill switch, argv boundary, clone jail, approval binding) and closes it with the smallest refuse path plus a durable assert. - Use Case: After modifying the writer or git approval flow in src/agentic, run this skill to verify every hostile-argv, jail-escape, and missing-confirm attack still fails closed with exit code 4 before merging. ## Quick Start Run the write-policy redteam against my recent changes to the writer and git approval gates and report any new bypasses.
