What problem does it solve? Autonomous implementation runs report tasks as done based on executed evidence, but a passing test is neither necessary nor sufficient proof that a feature works. Work can be left uncommitted, a suite can be green but broken, and tasks can be scattered across commits — failure modes the evidence gate cannot see. This Skill independently verifies what a run produced before you trust its result or open the PR. ## Core Features & Use Cases - Independent run verification: Spawns the read-only run-verifier agent against the run's worktree to return a structured verdict of { clean, findings, learnings_path }. - Severity-ranked findings: Distinguishes blocking findings (uncommitted-work, unproven-suite) from cosmetic warnings (scattered-task), each with the exact file, symbol, or commit and the fix. - Optional remediation with --fix: Commits uncommitted work using repo conventions, re-runs the suite and records a fresh receipt, or re-splits a collapsed commit, then re-verifies. - Use Case: After an unattended implement run reports all tasks closed, run verification on the same worktree to confirm the commits actually contain the work and the suite genuinely passes before opening the pull request. ## Quick Start Ask the assistant to verify the autonomous implementation run on the current worktree, optionally passing a branch name or --fix to remediate mechanical findings.
