What problem does it solve? Code changes often reach reviewers with defects that a hostile reading would have caught, and each external review gate cycle costs queue time and shared reviewer capacity. This Skill runs a self-review loop whose subagents try to break your diff rather than bless it, so findings die locally before anyone else reviews the change. ## Core Features & Use Cases - Adversarial rounds with budgets: Scope the change with git, announce a round estimate and ceiling from the changed-file count, and launch parallel subagents prompted to refute the code with executed proof. - Verified findings and fixes: Every finding and proposed fix is executed before being applied, fixes are applied at the class level rather than the site, and tests are mutation-proved by reintroducing the defect and requiring red. - Non-convergence exits and journaling: Detects loops auditing their own output, offers three named exits, and records every round in a JSONL journal with commit trailers like Adversarial-Rounds for auditability. - Use Case: Before pushing a feature branch, run adversarial rounds on the diff against origin/main, fix confirmed critical and high findings, then push and let the review gate's verdict close the loop. ## Quick Start Run one adversarial review round on my current uncommitted changes against origin/main and report any critical or high findings with executed proof.
