hook-audit

Repeated corrections/skill improvements: read Poteto.

Audit ~/.claude/hook-metrics/ per REFERENCE.md.

Modes: default/--hooks activity; --retro adds flow/environment; --all also adds latency, skill firing and drift.

Flow

1. Inventory hooks/date range; separate evals from real runs. Cohort by harness version/model; split or exclude mixed-model sessions using model-switches.jsonl.
2. Aggregate blocks, warnings, nudges, denials, sessions, trend; compute P50/P95 and wall time when requested.
3. Compare scripts, observed keys, and enforcement; distinguish true zero-fire candidates, untested hooks, and advisory rules.
4. Retro: follow Session-environment retrospective; add available retro metrics. Missing telemetry does not block transcript findings.
5. All: inspect skill-fires.jsonl; run bash scripts/generate-hook-configs.sh --check.
6. Model-switch policy: use /quantify-impact; success/rework is primary, cache-write cost a guardrail.
7. Rank at most five actions by telemetry/session impact. Recommend only unless edits were requested.

Before deletion, shadow via HOOK_SHADOW_RULES in a representative, version-qualified trial; compare outcomes/violations. Never shadow strict safety/permission.

Done

Report metrics/values, sample size, 7-day trend, next action; name gaps. Below five comparable real sessions is preliminary. Cite source files and exact harness_version + model cohort for prune/severity changes.