What problem does it solve? Automated tests cannot verify behavior that only appears in a real install — a real server, a real global tool, a real data bank. This Skill structures that hand-run verification so every checklist answer is backed by the command that was run and the output that was read, making a filled checklist distinguishable from a plausible-sounding one. ## Core Features & Use Cases - Derived facts, not pinned ones: scripts/derive-facts.py reads the version, MCP tool count, and prompt count from the source tree at run time and fails loudly on zero, so the checklist compares the running binary against what the tree actually says. - Evidence-required template: templates/checklist-template.json forces every item to record the exact command, verbatim evidence, observed result, a tri-state status (pass/fail/skipped/substituted), and an explicit acceptance decision. - Anti-rot design: documents six ways predecessor checklists silently drifted — pinned versions, steps for deleted features, inherited verdict reasons — and the defenses against each. - Use Case: Before shipping an AiRaccoon release, run the checklist to force-update the global tool, verify the server lifecycle on a non-default port, exercise the write/read/noise-filter paths against a scratch data root, and record the filled checklist under docs/work/checklist/. ## Quick Start Ask the agent to run the AiRaccoon manual checklist against the current repository and write the filled results to docs/work/checklist/.
