What problem does it solve? Code reviews only examine what was just delivered, leaving the surrounding perimeter — inherited defaults, sibling writers, incomplete gates, fail-open error paths — unexamined, which is where many real defects live. This Skill runs a fresh-context sweep of an existing subsystem against ten named defect classes and produces a findings document where every finding carries reproducible proof. ## Core Features & Use Cases - Ten defect classes: Systematically checks inherited defaults, sibling writers, incomplete gates, fail-open error paths, silent truncation, stale assertions, identity confusion, time-based inference, unasked authorization, and over-serving the client. - Proof-first findings: Each finding states what breaks, in what circumstance, with a falsifiable proof (failing test, query, or reproduction); unprovable items are downgraded to labelled hypotheses. - Fresh-context execution: Runs in a forked defect-sweeper agent so the sweep never inherits the builder's blind spots, and writes results to .gener8v/sweeps/<subsystem-slug>-sweep.md. - Use Case: After a burst of changes to the billing module, run a sweep before launch to discover that a sibling route writes the same table without the guard, with a grep command proving it. ## Quick Start Ask the AI to sweep a specific subsystem with a named boundary, for example: sweep api/billing/webhooks for defects, excluding the invoice renderer.
