/audit — security findings that become ready bug plans

Security holes die in chat transcripts. /audit turns an exhaustive code- and-dependency pass into prioritized .plans bug plans so the fleet can close them like any other ready work.

The gap

Teams already have scanners, ad-hoc reviews, and “we should fix that someday” threads. What Anchor was missing was a session that:

1. Covers first-party code and third-party dependencies in one pipeline
2. Maps severity to plan Priority (P1–P3) so /work picks the right order
3. Tags every fix plan with Preferred models: frontier, reasoner so mid and small workers leave security repair alone under normal fit rules
4. Writes plans only — no silent “I fixed it,” no exploit PoCs, no secret dumps

What /audit does

One project, one session. Pipeline (fixed):

resolve project → model-fit gate → inventory
  → dependency audit → first-party code audit
  → dedupe / severity → Priority
  → present findings → confirm (or --write / --dry-run)
  → emit bug plans → stop

It never auto-starts /work on the plans it creates.

Flag	Role
--dry-run	Full findings package; zero plan files
--write	Skip confirm after the package
--to drafts	Private drafts instead of ready bugs/
--force-model	Override the mid/small refuse (risk on you)
--deps-only / --code-only	Narrow buckets

Default writes land as sticky .plans/bugs/sec-<kebab>.local.md so unvetted findings stay gitignored until a human renames them for tracking.

Frontier / reasoner only (by default)

A weak model that “audits” confidently is worse than no audit: false negatives leave holes open; false positives bury the ready queue. /audit refuses mid and small sessions unless the operator passes --force-model. Grok 4.5 is mid-class for this gate — high reasoning effort is a cost dial, not a promotion.

Every emitted fix plan repeats the floor so fleet pickers keep security work on the right tier.

Safety

• No exploit payloads or weaponized “proofs” in chat, Steps, or the repo
• Secrets redacted (path + last 4 / hash)
• Cap of 25 plans per run; residual risk when a scanner is missing
• No public disclosure side-effects from the skill defaults

Where it ships

• Claude: .claude/commands/audit.md
• Grok: .grok/skills/audit/SKILL.md
• Docs: Skills → /audit
• Scaffolded with the other dual-use skills (scripts/anchor.py)

Pair with /work on the plans it writes (on a reasoner/frontier session) and /review when those fixes are ready for human sign-off.