agent-audit

Skill

Owner: Fury · Team: Fury · Source: ~/.openclaw/workspace-fury/skills/agent-audit/SKILL.md

Weekly ecosystem audit — per-agent checks for memory hygiene, schedule health, spec drift, eval regression, security flags, cost, and docs sync. Use for the Sunday audit run or any on-demand oversight question.

---

Playbook (mirrored from disk)

Agent Audit

Run per agent listed in ~/.openclaw/platform/agent-registry.md. Evidence for every finding.

Checklist (per agent)

1. Memory hygiene — MEMORY.md contradictions vs reality (decommissioned projects listed
   as live, duplicated sections); dailies piling up undistilled (>14 days since last curation).
2. Schedule health — openclaw cron status + crontab -l: failing jobs (consecutiveErrors),
   double-booking (two jobs same slot/purpose), orphaned jobs pointing at dead agents/scripts.
3. Spec-vs-reality drift — AGENTS.md vs the scripts/cron prompts that actually drive the
   agent (canonical example: analyzer AGENTS.md vs analyzer-codex-worker.sh, June 2026).
4. Eval regression — run sets under ~/.openclaw/platform/evals/ where feasible; compare
   to each set’s threshold; flag drops.
5. Security flags — grep workspaces for secret patterns (sk-, token , password,
   Bearer ); unscanned new skills (no jarvise-scan record); report leak LOCATIONS only.
6. Cost/usage — openclaw models status usage lines; flag anomalies vs last week.
7. Docs sync — Quartz pages under ~/quartz/content/teams/ vs agent-registry.md; on drift:
   update pages + run ~/quartz-rebuild.sh + verify HTTP 200.

Report format

# Agent Audit — YYYY-MM-DD
| agent | memory | schedule | drift | evals | security | notes |
|-------|--------|----------|-------|-------|----------|-------|
| ...   | ✅/⚠️/❌ | ...      | ...   | ...   | ...      | ...   |

## Top fixes (max 5, ranked by risk)
1. <fix> — <evidence> — <proposed owner>

## Summary
<one paragraph>

File at reports/audit-YYYY-MM-DD.md; final output = the report (cron delivers it).