What problem does it solve? Single-pass code reviews often miss counterexamples, false claims, and edge cases because one reviewer cannot challenge its own assumptions. This Skill runs a structured adversarial review where multiple named agents independently analyze targets, rebut each other's findings across rounds, and converge on a verified verdict. ## Core Features & Use Cases - Multi-round adversarial debate: Round 1 collects independent analyses as claims; later rounds force participants to rebut or concede each other's specific points via a shared whiteboard (claim/rebuttal/consensus/verdict). - Dedicated counterexample and fact-check lenses: Assigns a participant to break proposed detection logic with concrete inputs, and requires fact-checkers to verify behavior across states and cross-check claims against embedded screenshots. - Guardrails and fallback: Handles unresponsive participants, circuit-breaker stops, and silent-free fallback to a legacy subprocess runner or fan-out evaluation, with post-hoc detection of unauthorized edits by participants. - Use Case: When reviewing a new quality-gate script, launch a discussion where one agent designs the logic, another crafts counterexample inputs that defeat it, and a fact-checker validates state-dependent behavior, producing a verdict JSON listing only issues that survived cross-examination. ## Quick Start Ask the AI to run a discussion review by forming a specialist team to debate and review the specified target files using the discussion-review workflow.
