What problem does it solve? When an AI code review flags an issue, you cannot tell whether it is a real defect or a false positive without an independent check, and rubber-stamping dismissals risks shipping real bugs. This Skill sends the finding to a fresh Codex thread for blind verification, then maps the verdict through severity-graded policy thresholds so dismissals are auditable and abuse-resistant. ## Core Features & Use Cases - Three verification intents: dismiss (false-positive check), confirm (does the issue exist), and clarify (what is the actual impact), each with its own audit token ([DISMISS_VERDICT] or [SEEK_VERDICT]). - Anti-anchoring protocol: Claude's hypothesis is recorded locally but never sent to Codex, and a fresh thread is always used so the verdict is genuinely independent. - Graduated policy mapping: P0/P1 dismissals produce only DISMISS_CANDIDATE and require human confirmation, while P2/Nit dismissals can be automated at lower confidence thresholds. - Anti-abuse guards: three consecutive dismissals trigger heightened thresholds, and confirm/clarify are capped at one per finding per commit. - Use Case: A review flags a P1 race condition you believe is a false positive. Run seek-verdict with intent=dismiss; Codex independently inspects the diff and returns NON_ACTIONABLE at 0.92 confidence, producing a DISMISS_CANDIDATE that a human confirms before the finding is closed. ## Quick Start Ask the agent to verify a review finding with seek-verdict, for example: "Use seek-verdict to check whether the P1 finding in auth/session.go is a false positive."
