| name | post-cold-start-verification |
| description | Audit every ai/ file for gaps, stale placeholders, and inconsistencies after cold-start. Produces a prioritized findings report. |
| disable-model-invocation | true |
Audit the AI knowledge layer. Read-only with one exception: you may write ONE report.
Step 0 — use the mechanical verifier's output, don't redo its job
If ai/analysis/audit-reports/VERIFICATION_MANIFEST.json exists, read its summary
and the non-confirmed claims first: path-existence checking is already done
deterministically by the kit's verify command (run from the kit checkout:
node install.mjs verify <repo> — no LLM
involved, so its statuses are facts). Spend your pass on what a script cannot
judge (checks 1, 3, 4 below, and the semantic half of check 2). If the manifest
is absent or older than the docs it covers, recommend regenerating it in your
report instead of re-deriving path checks by hand.
Checks
- Placeholders: find every
<fill in>, ? Stability, and {{...}} leftover
across ai/ and the entry files (CLAUDE.md, AGENTS.md).
- Internal consistency: every path mentioned in
ai/guide/* exists on disk
(covered by the manifest when present — see Step 0); MODULE_MAP rows correspond
to real directories; FEATURE_MAP entries point at real files; diagrams name
real modules.
- Profile consistency: build/test commands in CLAUDE.md/AGENTS.md match
ai/repo-profile.json; flag divergence (don't silently fix).
- Provenance hygiene: no
[verified] tag lacking a date; nothing you can tell
was agent-written carrying [verified].
Output
Write ai/analysis/audit-reports/<YYYY-MM-DD>-post-cold-start.md: findings grouped
P1 (agent-blocking) / P2 (misleading) / P3 (cosmetic), each with location and a
one-line suggested fix. Tag the report itself [inferred]. Do not modify the files
you are auditing.