메뉴
Quarterly skill that walks the skill-index against the actual SKILL.md files on disk, reports drift (skills indexed that no longer exist, skills on disk that aren't enrolled, skills whose descriptions have changed since enrollment). Hardens the auto-enrollment pipeline against silent rot.
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
SOC 직업 분류 기준
| name | Caddy Audit |
| description | Quarterly skill that walks the skill-index against the actual SKILL.md files on disk, reports drift (skills indexed that no longer exist, skills on disk that aren't enrolled, skills whose descriptions have changed since enrollment). Hardens the auto-enrollment pipeline against silent rot. |
Auto-enrollment makes adding skills easy. Easy adds without periodic audits accumulate drift. /caddy-audit is the quarterly hygiene pass that keeps the skill index honest.
/caddy-explain shows surprising results that suggest the index doesn't match realityLoad .claude/skill-index.json into memory. Each entry has at minimum: name, triggers, why.
List every SKILL.md file in skills/<name>/SKILL.md and any project-scope .claude/skills/<name>.md. For each, parse the YAML frontmatter for name and description.
For each skill in the index, check:
name matchesdescription is consistent with the indexed why (the indexed why is usually a tighter summary; flag if they're contradictory)For each SKILL.md file on disk, check:
If the skill files store original triggers in their frontmatter (not all do), compare to the indexed triggers. If the SKILL.md has been edited since enrollment AND its trigger frontmatter has diverged from the index, flag as TRIGGER-STALE.
This step is best-effort; not all skills carry trigger frontmatter, and that's fine.
🔍 Caddy Audit ({date})
Index: {N} entries. Disk: {M} SKILL.md files.
✅ HEALTHY — {K} skills are correctly indexed and on disk.
⚠️ INDEX-ORPHANS ({count}) — in index, no matching SKILL.md:
- /skill-a — last seen at skills/skill-a/SKILL.md (deleted? renamed?)
- /skill-b — never had a file?
⚠️ DISK-ORPHANS ({count}) — SKILL.md exists, not in index:
- skills/new-thing/SKILL.md — run /caddy-enroll skills/new-thing/SKILL.md
- skills/experimental/SKILL.md — never enrolled
⚠️ TRIGGER-STALE ({count}, best-effort) — SKILL.md frontmatter diverges from index:
- /skill-c — SKILL.md has 12 triggers, index has 8
Recommended actions:
1. {1-3 specific commands the principal should run to remediate}
If the principal wants longitudinal data:
# Save to vault/memory/AUDITS/caddy-{YYYY-MM-DD}.md
This becomes a track record over time of how clean the catalog stays. Patterns surface — "we always have 3-5 disk orphans because /skills-builder doesn't always run /caddy-enroll" — which informs whether the auto-enrollment pipeline needs hardening.
skill-index.json. The principal should approve that — even if "obvious," the entry might exist for a reason (e.g., a planned skill not yet implemented)./caddy-enroll for the affected skill. The skill file is canonical.Auto-enrollment plus the natural rate of skill experimentation produces drift. Without periodic auditing, the index slowly diverges from reality:
why linesThe 200ms latency budget means Caddy can't afford to do this checking on every invocation. Hence the quarterly audit pattern.
/caddy-enroll — the skill that adds a SKILL.md to the index (this audit's complement)/caddy-explain — the inspector for individual scoring decisions/caddy-audit is a member; this is the maintenance loop for the catalog itselfThe capsule verb, invoked explicitly — reconcile from the live memory, write a resume-ready capsule, then stop. Mid-session checkpoint or completion capsule; a rolling best-effort version of the same write also runs automatically on every Stop event via daemons/stop-capsule-writer.mjs.
The resume verb, invoked explicitly — load the newest valid capsule by created_at, re-ground against live memory, then ACT on waiting_on. Mid-session or on explicit ask; the post-clear boot fires the same procedure automatically via daemons/resume-verb.mjs.
Run the volunteering ledger before declaring work done. Forces surfacing of what changed/didn't, what was noticed-but-not-fixed, what was guessed, and what tradeoffs were made on the principal's behalf. The standard - the principal should never be surprised later by something the agent knew at the time.
Capture a confident agent claim (Phase 1) or resolve one as right/wrong (Phase 2) in vault/memory/TRUST_DECAY.md. Builds the longitudinal record of agent calibration over time. Pairs with /honesty-check (which should reduce Phase 1 frequency) and /diagnose (which surfaces causes that often reveal Phase 2 events).
One evidence-gated pass that sequences /dream, /reconcile, /sweep-now, and /digest over your own vault, with a /system-check gate and a ledger-review pass over the measurement layer (Honesty Ledger, Trust Decay, Failure Modes). Stays exactly as human-gated as the legs it calls -- stages, never auto-applies. Skips a leg gracefully when there is nothing to do; every leg that runs leaves an evidence line.
Dispatch Hestia to run all 3 vault sweeps (DELEGATION_TRACKER stale items, HEAT_INDEX dormant flips, broken wikilinks) and append the result to HESTIA_SWEEP_LOG.