What problem does it solve? Work products, PRs, and merged decisions often ship with hidden defects, contradictions, and omissions because self-review is biased toward confirmation. This Skill enforces a structured adversarial review loop that assumes failure, validates each critique before acting on it, and re-verifies the final state before completion is reported. ## Core Features & Use Cases - Attack and validate separation: Runs distinct modes (attack, validate-critique, refine-approved-findings, regression-recheck, decision-report) so criticisms are fact-checked for validity, scope, and cost before any change is made. - Bounded full-scope review loops: Executes exactly two complete full-scope review rounds per approved work lineage, then limits further work to targeted correction and regression verification, preventing endless audit churn. - Post-merge and repository-wide audits: Re-reads the new main HEAD, confirmed decisions, untouched consumers, and derivative drift after merges, and routes repository-wide findings to the correct owning Skill. - Use Case: After merging a PR that updates a canonical policy, run this Skill to verify no stale references, untouched consumer templates, or superseded decisions remain active, and to classify every finding as MUST_FIX, SHOULD_FIX, DEFER, or REJECTED_CRITIQUE before reporting completion. ## Quick Start Run an adversarial review of the latest merged PR against the current main HEAD and report validated findings with their severity decisions.
