What problem does it solve? Code reviews answer whether code is correct, but they do not answer whether a PR is safe to merge: how expensive it is to undo, what it reaches if it is wrong, and how quickly a failure would surface. This Skill prices being wrong by grading every effect of a PR on a durability ladder and sizing its exposed population. ## Core Features & Use Cases - Reversibility grading: Enumerates each effect of the change (writes, events, outbound calls, deploy-time actions like migrations and terraform applies) and grades it on a D0–D5 durability ladder, producing a two-way or one-way door verdict with an undo procedure and residue list. - Blast radius analysis: Runs six lenses (code fan-out, deploy surface, traffic, failure loudness, detection time, coupling) and quantifies exposed population with labeled bases — measured, derived, or estimated. - Signal scanning and skeptic pass: Ships scripts that fetch PR metadata via gh and scan the diff for durability and radius signals, then spawns adversarial subagents to attack the draft verdict from both sides. - Use Case: Before merging a PR that adds a database migration and changes an SMS-sending path, run the assessment to learn the migration is a one-way door, the SMS path can reach 40,000 customers before detection, and the smallest merge conditions (a feature flag plus an alarm) that make shipping boring. ## Quick Start Ask the assistant to assess merge risk for PR #123 in the current repository at standard depth.
