audit-full

Orchestrate the full adversarial audit suite: run each audit prompt in its own fresh-context worker when the harness supports isolated workers, validate every report, then synthesize one deduplicated fixes backlog. You are the conductor -- you perform no audit yourself, and you write nothing but the backlog.

Control-plane boundary

This spec governs only the task it defines. A newer human instruction or higher-priority runtime policy overrides this spec. STOP ends autonomous work under this spec; it never prevents the agent from honoring that control plane.

Where the audit specs live

The sibling audit specs are audit-codebase.md, audit-docs.md, audit-process.md, audit-security.md, audit-ux.md -- plus the diff-scoped audit-change.md, which needs a target and joins a suite run only when the user supplies a base ref at preflight. Derive the specs directory from this file's path, and the suite root as its parent; the validator lives at /dev/scripts/validate-report.mjs. If this spec was injected and its file path is unavailable, ask the user for the installed suite path rather than guessing. A spec whose file is missing is simply unavailable -- list it as such in preflight.

Phase 0 -- Preflight (the only planned interactive gate)

1. Discover: which sibling prompt files exist; the repo root; existing reports in docs/audits/.
2. Measure: approximate file count and size of the repo. Each audit reads it in full, so the suite costs roughly one full read per audit selected -- state this.
3. Snapshot: record the commit SHA. Require a clean tree by default -- a SHA plus a "dirty" flag does not identify what was audited, and two workers can otherwise audit different content under the same label. When determining dirty state, exclude UNTRACKED files under docs/audits/ -- the suite writes its own (untracked) reports there, so prior-run artifacts and this run's own writes must not trip this gate; changes to tracked files, including tracked prior reports under docs/audits/ that later audits read as input, still count as dirty. If dirty, stop and offer: (a) the user commits/stashes first, (b) explicit opt-in to proceed dirty -- then also record a checksum of git diff output plus the untracked-file list so drift stays detectable, (c) abort. Derive the run directory from the snapshot: docs/audits/-/ (short SHA = git rev-parse --short HEAD; if the directory already exists -- same day, same SHA -- suffix -2, -3...). Every artifact of this run (reports, sidecars, backlog) lives inside it; flat docs/audits/ remains the home of standalone single-audit runs.
4. Ask ONE question set before dispatching (use the harness's interactive-question mechanism where available; otherwise a plain numbered prompt), presenting the measurement from step 2 as a cost statement:
   • Which audits to run (default: all available whole-repo audits). Offer the diff-scoped change audit only with a base ref; suggest the repo's default branch when the current branch differs from it. Pre-validate the target yourself before offering: the ref resolves (git rev-parse) and the diff is non-empty -- never dispatch a change audit that will fail fast on arrival.
   • Worker model/tier only when the harness exposes choices and the user explicitly selects one; otherwise retain the current runtime selection.
   • Reasoning effort only when the harness exposes it as a session-level setting; never assume it can be dispatched per worker.
   • Parallel or sequential dispatch (parallel is faster; sequential spreads token burn and rate-limit pressure). No further interaction except the explicit fallbacks below (unsupported model/effort override, mid-suite drift opt-in): after this gate, run to completion or stop per the rules below.