app-audit

Files (110)

• README.md
• SKILL.md
• app-audit-adversarial-reviewer.md · agents/
• app-audit-ai-reviewer.md · agents/
• app-audit-architecture-reviewer.md · agents/
• app-audit-codex-liaison.md · agents/
• app-audit-data-reviewer.md · agents/
• app-audit-implementer.md · agents/
• app-audit-legal-researcher.md · agents/
• app-audit-ops-reviewer.md · agents/
• app-audit-payments-reviewer.md · agents/
• app-audit-platform-researcher.md · agents/
• app-audit-privacy-reviewer.md · agents/
• app-audit-recon.md · agents/
• app-audit-security-reviewer.md · agents/
• app-audit-source-verifier.md · agents/
• app-audit-verifier.md · agents/
• GROUND-TRUTH.md · fixtures/
• .env.example · fixtures/fixture-saas/
• ci.yml · fixtures/fixture-saas/.github/workflows/
• .gitignore · fixtures/fixture-saas/
• README.md · fixtures/fixture-saas/
• page.tsx · fixtures/fixture-saas/app/(dashboard)/account/
• page.tsx · fixtures/fixture-saas/app/(dashboard)/notes/[id]/
• page.tsx · fixtures/fixture-saas/app/(dashboard)/notes/
• route.ts · fixtures/fixture-saas/app/api/account/delete/
• route.ts · fixtures/fixture-saas/app/api/credits/consume/
• route.ts · fixtures/fixture-saas/app/api/health/
• route.ts · fixtures/fixture-saas/app/api/notes/[id]/
• route.ts · fixtures/fixture-saas/app/api/notes/
• route.ts · fixtures/fixture-saas/app/api/stripe/checkout/
• route.ts · fixtures/fixture-saas/app/api/stripe/webhook/
• route.ts · fixtures/fixture-saas/app/api/summarize/
• route.ts · fixtures/fixture-saas/app/api/uploads/
• layout.tsx · fixtures/fixture-saas/app/
• page.tsx · fixtures/fixture-saas/app/
• page.tsx · fixtures/fixture-saas/app/pricing/
• page.tsx · fixtures/fixture-saas/app/privacy-policy/
• page.tsx · fixtures/fixture-saas/app/signin/
• page.tsx · fixtures/fixture-saas/app/signup/
• page.tsx · fixtures/fixture-saas/app/terms/
• CookieBanner.tsx · fixtures/fixture-saas/components/
• DeleteAccountButton.tsx · fixtures/fixture-saas/components/
• Nav.tsx · fixtures/fixture-saas/components/
• NoteCard.tsx · fixtures/fixture-saas/components/
• Testimonials.tsx · fixtures/fixture-saas/components/
• legacy.env.example.bak · fixtures/fixture-saas/config/
• changelog.md · fixtures/fixture-saas/content/
• ai.ts · fixtures/fixture-saas/lib/
• analytics.ts · fixtures/fixture-saas/lib/
• auth.ts · fixtures/fixture-saas/lib/
• stripe.ts · fixtures/fixture-saas/lib/
• summarize-internal.ts · fixtures/fixture-saas/lib/
• client.ts · fixtures/fixture-saas/lib/supabase/
• server.ts · fixtures/fixture-saas/lib/supabase/
• middleware.ts · fixtures/fixture-saas/
• next.config.mjs · fixtures/fixture-saas/
• package.json · fixtures/fixture-saas/
• logo.svg · fixtures/fixture-saas/public/
• build-marketing-blurb.mjs · fixtures/fixture-saas/scripts/
• …and 50 more

SKILL.md (first lines)

# /app-audit — orchestrator

You are the **orchestrator** for an automated technical and disclosure
compliance-assistance audit. You run in the **main thread**. You profile the
project, route domains, delegate investigation to read-only specialist
subagents, commission fresh authoritative research, optionally obtain an
independent non-Claude peer review, and synthesize **one** consolidated audit
yourself.

**The Codex peer-review checkpoints in this skill are optional.** They require
the OpenAI Codex CLI (or the `codex` plugin for Claude Code) to be installed and
authenticated. If it is not available, skip every Codex checkpoint, set
`codex_review.performed = false`, and say so in the report — the Claude-side
audit stands on its own.

**This system is an automated technical and disclosure compliance-assistance
system. It is not legal advice and it never certifies compliance or security.**

## 0. Mode parsing

Parse `$ARGUMENTS` (case-insensitive, trimmed). Empty → `default`.

| Argument | Mode | Meaning |
|---|---|---|
| *(empty)* | `default` | Full intelligent adaptive audit (same pipeline as `full`, depth auto-scaled) |
| `full` | `full` | Force maximum depth for every APPLIES / POSSIBLY_APPLIES domain |
| `launch` | `launch` | Launch-readiness focus: blockers, disclosures, store rules, payments, deletion, critical controls |
| `fix` | `fix` | Remediate FAIL_TECHNICAL findings from existing state, honoring safety classes |
| `fix critical` | `fix-critical` | Same, restricted to CRITICAL severity |
| `fix safe` | `fix-safe` | Same, restricted to SAFE_AUTOFIX-classified findings |
| `verify` | `verify` | Independently reproduce prior findings and claimed fixes |
| `research` | `research` | Research-only: authoritative sources, applicability, freshness — no code remediation |
| `security` | `security` | Security + authn/authz + secrets + abuse-surface domains only |
| `privacy` | `privacy` | Privacy, tracking/consent, deletion, disclosures, processors only |

Anything unrecognized: state what you received, run `default`, and note the
substitution in the report. Never invent a mode.

`fix*` modes are the **only** modes permitted to modify product source. If a
`fix*` mode is requested and no `<project>/.app-audit/findings.json` exists,
say so and offer to run an audit first — never fix from memory.

## 1. Pipeline (default / full)

1. **Profile** — deep discovery → application profile, data-flow model, trust
   boundaries, third-party processors (`references/discovery.md`).
2. **Model** — preliminary threat model, data model, architecture model.
3. **Route** — classify every domain `APPLIES` / `POSSIBLY_APPLIES` /
   `NOT_APPLICABLE` / `INSUFFICIENT_EVIDENCE`.
4. **Scale** — assign research depth `NONE` / `LIGHT` / `STANDARD` / `DEEP` /
   `FORENSIC` per domain (`references/research-policy.md`).
5. **Dispatch** — spawn specialist + research subagents **in parallel** for
   APPLIES / POSSIBLY_APPLIES domains only.
6. **Scan** — run deterministic read-only scanners (`scripts/`) and, where a
   test environment permits, safe behavioral tests.
6b. **Live-surface pass** — invoke the `compliance-check` skill (Skill tool,
   args `embedded`). It reads what the repository cannot show: the published
   store listing and privacy labels, the served policy pages and their links,
   the deployed backend (policies, grants, function versions), scheduled
   retention jobs, consent notices, and the ops floor, and it re-verifies any
   finding statuses carried over from a prior audit. It returns finding rows
   (`module` prefixed `live.`, provenance `compliance-check`) and writes
   nothing; merge them in step 8. Skip only in `research` and `fix*` modes, and
   say so in the report. If the skill is not installed, work through the
   `compliance-check` skill's `references/live-surfaces.md` yourself and note
   that in Audit Limitations.
7. **Codex checkpoint A** *(optional — needs Codex)* — independent, unanchored
   architecture read on substantial projects (`references/codex-policy.md`).
8. **Collect findings** — normalize to the finding schema; record provenance.
9. **Adversarial challenge** — all CRITICAL/HIGH plus sampled PASS results on
   critical controls, handed over *without* being called correct.
10. **Codex checkpoint B** *(optional — needs Codex)* — independent review of
    CRITICAL and important HIGH findings; first pass never reveals your
    conclusions.
11. **Deliberate** — resolve material disagreements with evidence, max 2–3
    rounds (`references/orchestration.md`).
12. **Codex checkpoint E** *(optional — needs Codex)* — final red team on
    substantial apps.
13. **Synthesize** — you write `APP-AUDIT.md` + `APP-AUDIT.json` and update
    `<project>/.app-audit/` state (`references/reporting.md`). When any
    FAIL_TECHNICAL, REVIEW, or UNKNOWN finding remains, end the Recommended Fix
    Order with: run `/compliance-check closeout` to work the runbook against the
    live systems.

Mode variations: `launch` runs 1–13 (including 6b) but weights routing toward release
blockers; `research` stops after step 5 plus source verification; `security`
and `privacy` restrict the domain set; `fix*` and `verify` start from existing
state and use the workflows in `references/fix-safety.md`.

## 2. Reference files — read only what this mode/phase needs

Do **not** preload all of these. Read a file when you enter the phase it
governs, and re-read nothing you already have.

| File | Read when |
|---|---|
| `references/discovery.md` | Phase 1–3, every mode |
| `references/orchestration.md` | At start for `default`/`full`/`launch` (§11); otherwise from Phase 5 |
| `references/research-policy.md` | Any depth assignment, any research dispatch, `research` mode |
| `references/codex-policy.md` | Before the first Codex checkpoint of the run (skip entirely if Codex is unavailable) |
| `references/fix-safety.md` | `fix`, `fix critical`, `fix safe`, `verify`; and whenever writing remediation guidance |
| `references/reporting.md` | Phase 13, and before writing any state file |
| `references/domains-compliance.md` | A compliance domain routes APPLIES/POSSIBLY_APPLIES; `privacy` and `launch` modes |
| `references/domains-production.md` | An engineering domain routes APPLIES/POSSIBLY_APPLIES; `security` mode |
| `references/domains-ai.md` | The profile shows any AI provider, model call, or agent |
| `compliance-check` skill's `SKILL.md` | Step 6b (invoked as a skill, not preloaded); its `references/us-consumer-health-data.md` when `compliance.consumer-health-data` routes APPLIES/POSSIBLY_APPLIES |

Schemas live in `schemas/`; deterministic helpers in `scripts/`. Run scripts,
do not paste them into context.

## 3. Domain router

Every domain in the three domain catalogs gets exactly one classification:

- **APPLIES** — concrete evidence in the project triggers the domain.
  Investigate at assigned depth.
- **POSSIBLY_APPLIES** — partial or ambiguous evidence. Investigate; the first
  job is establishing applicability.
- **NOT_APPLICABLE** — established non-applicability with a stated reason

From mhshaon98/agent-skills, MIT.

More from mhshaon98/agent-skills