agent-auditlisted

# Agent Audit ## Purpose Spot patterns of agent misuse from the dispatch-log so the operator can either tighten frontmatter (e.g., add `runtime-fallback`) or update the lead's discipline (e.g., the lead is dispatching the wrong specialist for a task type). Read-only — produces a report, makes no changes. ## Scope Audit-only. The skill produces a markdown report listing: 1. **Lead-did-specialist-work** — sessions where the lead's `Bash` calls touched code (git commit, npm install, build commands). Indicates the lead bypassed dispatch. 2. **Wrong-runtime patterns** — agents whose `runtime:` doesn't match the bulk of what they were actually dispatched on. E.g. an agent declared `runtime: codex` but most calls used `--runtime claude` overrides — frontmatter is stale. 3. **Budget violations** — `budget_violation` events from dispatch-log.ndjson (added v0.8): real cost exceeded `max-cost-per-task`. 4. **Repeated failures** — agents with >2 consecutive `exit_code != 0` calls. Likely a misconfigured agent body or a runtime regression. 5. **Unused agents** — agents present in `<project>/.claude/agents/` that haven't been dispatched in the audit window. Either dead code or under-utilized. ## When to use - Weekly / monthly retrospective on yakOS usage in a project. - After a stretch of "feels like agents aren't pulling their weight" to get evidence. - Before a refactor of project agents — see what's actually used. ## When NOT to use - For a singl