skill-auditorlisted

# Skill Auditor Audit the **skill as a product and control plane**, not the domain task the skill performs. Treat a skill package as a system with routing, instructions, references, executable helpers, tests, host metadata, and distribution artifacts. ## Workflow 1. Freeze `skill_id`, version/commit, audit mode, host claims, and package source. 2. Inventory the complete skill topology before judging absence. 3. Audit trigger precision and negative boundaries before instruction prose. 4. Inspect instruction ownership, overlaps, conflicting rules, implicit escalation, and false-completion language. 5. Verify every referenced local file/dependency and every material host/runtime claim. 6. Inspect executable coverage, negative cases, mutation/branch evidence where available, and whether packaged ZIP behavior matches source behavior. 7. Check progressive disclosure: `SKILL.md` stays control-plane sized while optional depth lives in direct references. 8. Produce findings with concrete locators and evidence; separate structural fact from interpretation. 9. If the user wants empirical proof that the skill improves model behavior, hand off to `skill-evaluator`. 10. If changes are requested, hand off to `skill-creator`; never silently self-modify the package. Use LIGHT for a bounded single-skill check, STANDARD by default, DEEP for release/publication hardening, and DELTA only when a real prior version is available. ## Audit surfaces Audit at least the surfaces applicable to the