What problem does it solve? Long coding sessions accumulate context that quietly turns assumptions into facts, and confident answers are not always correct ones. This Skill catches wrong directions early by forcing every non-trivial decision through a fresh-context adversarial review while course-correction is still cheap. ## Core Features & Use Cases - Bounded doubt cycle: A five-step loop (CLAIM, EXTRACT, DOUBT, RECONCILE, STOP) that names the decision, isolates the smallest reviewable artifact, invokes an adversarial fresh-context reviewer, classifies findings, and stops after at most three cycles. - Adversarial reviewer prompt: A fixed prompt that instructs reviewers to find issues rather than validate, receiving only the artifact and contract so the orchestrator's conclusions cannot bias the review. - Cross-model escalation: In interactive sessions the user is always offered a second opinion from an external CLI (Gemini or Codex) with read-only sandboxing and stdin piping to prevent shell injection; in non-interactive contexts the skip is announced. - Use Case: Before committing a caching layer claimed to be thread-safe, the orchestrator extracts the diff and the concurrency contract, spawns a fresh reviewer biased to disprove the claim, reconciles the findings, and fixes a real race condition before merge. ## Quick Start Ask the agent to apply doubt-driven development to the architectural decision or non-trivial code change you are about to commit.
