| name | audit |
| description | Self-audit this repo's skill-first shape - hook suite, word budgets, cross-file duplication, dangling brief paths, version mirrors - and report a scorecard with top fixes |
| disable-model-invocation | true |
Self-audit
Run the deterministic checks below and report a scorecard. Every check is a command whose output decides — no judgment calls, no prose scoring.
Checks
- Hook suite & release gate —
python3 -m pytest adapters/claude-code/hooks/tests -q from the repo root. Green = pass; a failure names its own fix.
- Word budgets — every
skills/**/*.md, adapters/*/skills/**/*.md, and .claude/skills/**/*.md fits one read: ~865-word target, 2,000-word hard ceiling. wc -w each file; over ceiling = fail, over target = warn. List offenders with their counts.
- Doctrine hygiene — no instruction line duplicated verbatim across skill files (the enforcement half of CLAUDE.md's three-check meta-rule). Normalize lines (trim whitespace; drop blanks, headings, fences, and table rows), keep lines ≥ 40 characters, and flag any line appearing in more than one file. Duplication = fail, naming both files.
- Dangling brief paths — every brief a SKILL.md or brief references as a
NAME.md / ../dir/NAME.md sibling path exists on disk, resolved relative to the referencing file. A missing target = fail.
- Version mirrors —
python3 scripts/bump-version.py --check (also pinned by the pytest gate). Report the line.
Report
- Overall — categories passed / total.
- Categories — one line each: pass / warn / fail, with the evidence count (files over budget, duplicated lines, missing paths).
- Top fixes — the concrete fix per failing category, worst first. If nothing fails, state that all categories are clean.
