What problem does it solve? It verifies whether a packaged Keyhac skill bundle is truly self-contained by having an isolated AI session write a real Keyhac action using only the bundle, recording every question the bundle could not answer as evidence of documentation gaps. ## Core Features & Use Cases - Isolated room construction: Builds a clean-room directory containing only the skill bundles, RULES.md, TASK.md, and QUESTIONS.md, with a canary CLAUDE.md that detects contaminated sessions. - Non-interactive session driving: Launches a restricted Claude session with scoped permissions, a single Keyhac MCP endpoint, and no Bash, so the room cannot reach the repository or ask the operator questions. - Transcript auditing and scoring: Audits the session transcript for rule breaches (reading the checkout, installed package, or other authors' actions), runs mechanical checks, and reports QUESTIONS.md as the actual result. - Use Case: After editing the keyhac-action-authoring skill, run make cleanroom CASE=1 to confirm the rebuilt bundle teaches a fresh session everything it needs, then fix the skill for every question recorded. ## Quick Start Ask the assistant to clean-room test the skill bundle with case 1 by running make cleanroom CASE=1 after enabling Keyhac's MCP Server switch.
