Audit — Unified Audit Command

Auto-detect and execute the appropriate audit workflow.

Methodology Tracking

Every audit report MUST include a methodology field (JSON) or Methodology section (Markdown) documenting:

• Files Scanned — every file actually read
• Knowledge Tiers — which levels (L1–L4) were activated and whether each was available
• Standards Source — the skill files, checklists, and external standards used as rule authority
• Coverage Gaps — anything unavailable (RAG down, checklist not found, no platform rules loaded)

Track these as you work. Never leave them empty or with placeholder values.

Knowledge Retrieval (Pre-Audit)

Before executing any audit mode, activate knowledge-retrieval to load relevant context:

• L1 docs/ — existing conventions, past findings, ADRs for the files under review
• L2 RAG — component implementations, token definitions, usage patterns
• L4 Grep/Glob — fallback if RAG unavailable (search packages/, src/ directly)
• L5 Context7 — library API verification for external dependency usage

RAG unavailable? Skip L2, go directly to L4 Grep/Glob. Never block the audit waiting for RAG.

Subagent Constraint

Subagents cannot spawn further subagents — neither Agent tool nor Task tool is available in subagent context. Therefore, this skill runs inline in the main conversation (no context: fork). The main conversation is the orchestrator — it dispatches specialist agents and merges their results.

Step 0 — Flag Override + Mode Selection

If $ARGUMENTS contains --poc, --beta, or --stable: extract the maturity tier and pass it through to references/ui-workflow.md workflow (Step 0.6). These flags combine with --ui — they are not standalone modes.

If $ARGUMENTS starts with --ui and no maturity tier flag (--poc/--beta/--stable) is present: Ask the developer to confirm maturity tier before dispatching:

"What's the maturity stage of {ComponentName}?

• --poc — prototype / proof-of-concept (relaxed rules, phased roadmap)
• --beta — in active development (moderate strictness)
• --stable — production-ready (full strictness)

Reply with the flag or just poc / beta / stable." Wait for reply, then set the maturity tier and proceed.

If $ARGUMENTS starts with --ui: dispatch epost-muji via Agent tool. Pass component name + platform flags + maturity tier (if present) + references/ui-workflow.md workflow. If $ARGUMENTS starts with --a11y: dispatch epost-a11y-specialist via Agent tool. Pass references/a11y-workflow.md + platform hint. If $ARGUMENTS starts with --close --ui: load references/ui-close.md and execute inline. If $ARGUMENTS starts with --close: load references/a11y-close.md and execute inline. If $ARGUMENTS starts with --code: dispatch epost-code-reviewer via Agent tool. If auto-detected as hybrid (see Hybrid Detection below): run Hybrid Orchestration. Otherwise: continue to Auto-Detection.

Hybrid Detection

Trigger hybrid mode when ALL conditions met:

• Target contains klara-theme files (path contains libs/klara-theme/ or libs/common/)
• File count >= 20 OR multiple subdirectories in scope
• No explicit --ui or --code flag (those force single-agent mode)

Hybrid Orchestration (main context)

This runs in the main conversation, NOT in a subagent. The main context has Agent tool available.

session_folder = reports/{YYMMDD-HHMM}-{slug}-audit/

1. Create session folder: Bash("mkdir -p {session_folder}")
2. Dispatch epost-muji via Agent tool with Template A+ from references/delegation-templates.md:
   • Fill: Scope, Component(s), Mode: library, Platform, Output path: {session_folder}/muji-ui-audit.md
   • WAIT for muji to complete
3. Read muji report at {session_folder}/muji-ui-audit.md. Extract:
   • finding_locations: Set of file:line flagged by muji
   • verdict: muji's overall verdict
   • a11y_findings: contents of ## A11Y Findings section (if present)
4. If a11y findings exist AND maturity tier is NOT poc: dispatch epost-a11y-specialist via Agent tool (Template B):
   • Output path: {session_folder}/a11y-audit.md
   • WAIT for completion
   • POC exception: If --poc, skip a11y dispatch — A11Y findings are already advisory-only in muji's report (no dedicated a11y pass needed until beta)
5. Dispatch epost-code-reviewer via Agent tool:
   • Pass: file list, {session_folder}/muji-ui-audit.md path (for dedup), SEC/PERF/TS/ARCH/STATE/LOGIC scope
   • Output path: {session_folder}/code-review-findings.md
   • WAIT for completion
6. Merge reports into {session_folder}/report.md:
   • Executive Summary with overall verdict
   • ## UI Audit — muji verdict, finding count, link to muji-ui-audit.md
   • ## A11Y Audit (if ran) — link to a11y-audit.md
   • ## Code Review — code-reviewer findings inline
   • Methodology section 6.5. Run build verification:
   bashCopy

   node .claude/hooks/lib/build-gate.cjs
   

   Append ## Build Verification section to {session_folder}/report.md:
   • Exit 0: Build verification: ✓ PASS ({platform}, {duration_ms}ms)
   • Exit 1: Build verification: ✗ FAIL — {error excerpt} (advisory — does not block report)
   • Exit 0 (no command): Build verification: skipped (no build command detected)
7. Write session.json per references/session-json-schema.md
8. Update reports/index.json per core/references/index-protocol.md

Verdict = max(muji, a11y, code-reviewer) where REDESIGN > FIX-AND-REAUDIT > APPROVE.

Single-Agent Delegation Protocol

For non-hybrid dispatches (--ui, --code, --a11y):

1. Create session folder per references/output-contract.md
2. Select template from references/delegation-templates.md
3. Fill all {placeholders} — include Output path: {session_folder}/{filename}
4. Dispatch via Agent tool to the specialist agent
5. Wait for specialist report
6. Run build verification: node .claude/hooks/lib/build-gate.cjs — append ## Build Verification to report (advisory)
7. Write session.json and update reports/index.json

Output contract: references/output-contract.md is the single source of truth for paths and responsibilities.

Aspect Files

Auto-Detection

Analyze $ARGUMENTS keywords and context:

Platform Detection (--ui mode)

When delegating to epost-muji, detect target platforms:

• Explicit --platform web|ios|android|all in args → pass through
• .swift context → --platform ios
• .kt/.kts context → --platform android
• .tsx/.jsx/.ts context → --platform web
• No context → --platform all

Variant Summary

Examples

• /audit --ui EpostButton → muji audits EpostButton across all platforms
• /audit --ui EpostCard --platform web → muji audits web-only
• /audit --ui SmartLetterComposer --poc → organism audit with poc maturity tier, phased roadmap verdict
• /audit --ui SmartLetterComposer --platform web --beta → organism audit with beta maturity tier
• /audit --a11y → a11y specialist audits staged changes
• /audit --code → reviewer audits staged code changes
• /audit --close --ui 3 → mark UI finding ID 3 as resolved
• /audit EpostInput → auto-detected as UI audit → delegates to muji