Audit Command

<!-- managed by @austencloud/claude-skills — do not edit manually, run: npx @austencloud/claude-skills sync -->

---
description: Use when assessing code quality of a feature or component across 8 dimensions
---

# Audit Command

**Args:** `$ARGUMENTS` (optional: target path, or "list", "targets", "stats", "--auto-claim")

## Quick Commands (pass-through)

If the user passes a tracker command directly, just run it:

```bash
npx -p @austencloud/code-quality ac-audit $ARGUMENTS
```

This handles: `list`, `targets`, `stats`, `status <target>`, `resolve-issue <target> <index>`

## Pipeline Workflow

For actual audits (no args, or a target path), run the **three-phase pipeline**:

```
ac-audit --auto-claim → ac-evidence → audit-evaluator agent → record → present
```

### Role Separation (CRITICAL)

| Role | Can read source? | Can grade? | Can fix? |
|------|-----------------|------------|----------|
| Evidence collector (script) | Yes | No | No |
| Evaluator agent | Evidence JSON + source | Yes | No |
| Fixer agent | Cited files only | No | Yes |
| This orchestrator (you) | Coordinates all | No | No |

**You do NOT grade code.** The evaluator agent does. You orchestrate the pipeline.

---

### Phase 1: Claim Target

**No args (auto-select):**

```bash
npx -p @austencloud/code-quality ac-audit --auto-claim
```

Parse `CLAIMED_TARGET:` from output. If `AUTO_CLAIM_TARGET:` appears, the claim succeeded.

**With target specified** (e.g., `/audit src/lib/features/shop`):

```bash
npx -p @austencloud/code-quality ac-audit claim "$ARGUMENTS"
```

**Large module protection:** If the claim command shows "MODULE TOO LARGE", pick a sub-feature from the list and claim that instead.

---

### Phase 2: Collect Evidence

Run the deterministic evidence collector:

```bash
npx -p @austencloud/code-quality ac-evidence "<target>" --out .audit-evidence.json
```

This produces structured JSON with per-dimension findings. No LLM involved. Takes ~10 seconds.

---

### Phase 3: Evaluate

Spawn the **audit-evaluator** agent (`subagent_type: "audit-evaluator"`) with the evidence file path and target scope. It returns a scorecard, `GRADES_JSON`, and `ISSUES_JSON` blocks.

---

### Phase 4: Record

Parse the evaluator's `GRADES_JSON` and `ISSUES_JSON` from its response. Record to tracker:

```bash
npx -p @austencloud/code-quality ac-audit record "<target>" --grades "<A+,A,A,B,A,A,A+,A>" --issues-json '<json>'
```

Grade order: Architecture, Code Quality, Svelte 5, Accessibility, UX States, UI Consistency, Performance, Security.

---

### Phase 5: Present to User

Show the evaluator's scorecard and issues to the user. Then ask:

1. **Fix now** - Spawn fixer agent for all issues
2. **Fix critical only** - Spawn fixer for critical/serious issues
3. **Skip** - Leave issues for later

---

### Phase 6 (Optional): Fix

Spawn the **audit-fixer** agent (`subagent_type: "audit-fixer"`) with the `ISSUES_JSON`. After fixes, re-collect evidence, re-evaluate, re-record, and present before/after.

## Post-Audit

1. Recording auto-releases the claim
2. If fixes were made, offer to commit
3. Show grades, issues fixed/deferred, and next recommendation

## Key Rules

- Evidence is deterministic. Thresholds are mechanical. 0 violations = A+.
- Evaluator cannot fix code. Re-evaluation uses fresh evidence, not the fixer's word.
- Claims expire after 4 hours. Large modules (>30 files) audit as sub-features.