All Audits

Run the whole audit set over one repo in a single sweep. Thirteen audit skills fan out in parallel, each as its own process; each writes a self-contained report; the reports collect under one folder behind an index.html that links them. The sweep runs to completion on its own: past the index, it grills every report widest-first and lands the run as one spec. Grilling itself still questions the user on each fresh finding, and writes each rejection to the audited repo's ignore file as it's decided; the single confirmation gate is only about the tracker — nothing goes to the spec issue or its tickets before that one approval. It reports and files only — no edit is applied, no PR is opened.

Scope

$ARGUMENTS names the repo path if given, else the current working directory.

Whole-repo by default, and the branch-scope exception — see harness/AUDIT-RUN.md.

The set — thirteen skills

Each audits the whole repo. Most write a findings.jsonl + grouped-summary report.html pair (see harness/findings-schema.md); the three structural audits — ponytail-audit, thermo-nuclear-code-quality-review, and improve-codebase-architecture — render a full HTML report as their record instead, with no separate log file:

• ponytail-audit — over-engineering: what to delete, shrink, or replace with stdlib.
• test-audit — tests that prove nothing or check the wrong thing.
• comment-audit — comments that do not earn their place.
• thermo-nuclear-code-quality-review — abstraction quality, giant files, spaghetti growth.
• improve-codebase-architecture — shallow modules and deepening opportunities.
• audit-instructions — instruction files (CLAUDE.md, SKILL.md) against Anthropic's current guidance.
• dead-code — functions, classes, imports nobody calls.
• duplication — one behavior with two homes, token clones and same-data-two-ways.
• error-handling — swallowed errors against the repo's fail-loud rule.
• docstring-coverage — undocumented public API on a py.typed surface.
• domain-drift — code vocabulary that drifts from the project's domain terms.
• type-tightness — loose Any, unexplained # type: ignore, fake boundaries.
• crap-audit — per-function CRAP score (complexity² × uncovered fraction³ + complexity), real risk hotspots.

skill-audit is not in the set — it scans the global skills directory, not this repo. mutation-audit is not in the set either — it is opt-in and targeted at one module (never a whole-repo sweep); invoke it by name, or run driver.py --mutation a.py,b.py for the scripted per-module form.

Opt-in edits

Every audit skill in the set reports by default. Applying edits is opt-in — the user asks by name — and lands as a reviewable PR on its own branch, never a direct commit to main. Carry this contract into any new audit skill you add here.

The runnable sweep — driver.py

driver.py is the Python orchestrator that runs each guarded audit as its own claude -p "/name" process and collects the reports. driver.py --help is the flag reference — argparse prints it from the module docstring, and this doc doesn't restate it, so the two can't drift apart. A run is three steps: plan_sweep asks the staleness cache what still needs running, execute spawns the audits (the only step that does), and collect assembles the collection and renders the index — --index calls collect alone. audits_data.py holds the audit set — adding an audit is one entry there.

Staleness cache. The two expensive LLM passes — domain-drift and type-tightness — are gated: while the repo is materially unchanged since their last run, they are skipped and their cached report is reused in the index, marked "unchanged since <sha>". should_run() in driver.py owns the skip/run decision; driver.decide/driver.update_cache gather git state, read/write the per-repo record at ~/.cache/all-audits/<repo-key>.json, and persists each gated report to a stable location that survives the run-dir TTL prune (RunDir owns that folder layout and the 3-day prune, for both the sweep and mutation mode). Skip holds only when the tree is clean, fewer than N files (default 10) changed since the last-run SHA, no change touched domain-drift's ground-truth (CONTEXT.md, docs/adr/), and the last run is within the time backstop (default 30 days).

Ignore file. A repo can carry .audit-ignore.md at its root to suppress findings it has already reviewed and rejected, so a sweep stops re-raising them. Opt-in per repo — see harness/IGNORE-FILE.md.

Run

Two front doors, one engine. driver.py runs the whole sweep — it fans each audit out as its own claude -p "/name" process, collects the report folders, and builds index.html with the synthesis lede. Both front doors call it.

The claude -p "/name" path is why the script, not a subagent, runs the audits. Several audit skills set disable-model-invocation, so the Skill tool refuses them from any agent — a subagent fan-out silently loses them. A -p "/name" process is an explicit slash invocation, which the gate honors, so every audit runs the same way.

From a terminal, run driver.py directly for its flags and the staleness cache.

When you are invoked in-session as /all-audits [path]:

1. Resolve the repo. Use $ARGUMENTS if given, else the current directory. If the argument is short (/all-audits short [path]), add --short to run only the three structural audits.

2. Run the script in the background. Run driver.py [--short] <repo>. It does the fan-out, the collection, the index.html, and the lede. It reports only — it applies nothing and opens no PR. Do not fan out subagents yourself and do not invoke any audit through the Skill tool; that is the broken path this replaces.

3. Report the index, then grill. The script prints index: <path> and opens that page itself — do not open it again. Print its absolute path and list the audits, then continue straight into the grill — no separate ask needed — which walks every report widest-first per harness/GRILLING.md.

After the sweep

The sweep never stops at the index waiting to be told to grill — it continues on its own. For the grill's walk of every report — order, per-report grilling, the auto-carry rule, and how the run lands as one spec with its rejections written to .audit-ignore.md as they're decided — see harness/GRILLING.md.