What problem does it solve? AI coding agents frequently claim work is done — "tests pass", "bug fixed", "build succeeds" — without actual proof. This Skill acts as a lie detector that blocks unverified claims from reaching commit, catching the most common failure mode in AI-assisted development. ## Core Features & Use Cases - Claim-to-Evidence Matching: Extracts completion claims from agent output and matches each against real tool output (test stdout, build logs, git diffs), issuing CONFIRMED, UNCONFIRMED, or CONTRADICTED verdicts. - Stub & Loop Detection: Scans newly created files for placeholder code (TODO, NotImplementedError, empty returns) and audits tool-call patterns for observation loops that signal a stuck agent. - Plan Diff & Cross-Phase Checks: Diffs actual changes against phase plan files and verifies exports, routes, and end-to-end flows in multi-phase projects. - Use Case: After an agent reports "all 42 tests pass and the feature is implemented", run this gate to confirm the test output exists, the files are not stubs, and every planned task is checked off before allowing a commit. ## Quick Start Ask the agent to validate its completion claims with the completion gate before committing, requiring quoted command output as evidence for every claim.
