sb-audit — drift survey + decision board

The periodic gate. Surveys the catalog, seeds the ledger, names consolidation candidates.

Run it

SKILL=${CLAUDE_PLUGIN_ROOT:-${CLAUDE_PLUGIN_ROOT}}
CORE=${CLAUDE_PLUGIN_ROOT:-${CLAUDE_PLUGIN_ROOT}}
"$SKILL/scripts/audit-drift.sh"                 # naming drift across stories
"$SKILL/scripts/audit-archived.sh"              # L1/L2/L3 preservation surveillance (HEAVY warning)
"$CORE/scripts/find-stories-by-tag.sh" decision:pending   # stale decisions needing a nudge
"$CORE/scripts/extract-component-usage.sh"      # REFRESH real prop-usage → component-usage.json (auto, no hand-edits)
"$SKILL/scripts/audit-controls.sh" .            # Controls-panel coverage: component stories missing argTypes + manager showPanel
"$CORE/scripts/refresh-usage.sh" --docs .       # ONE-COMMAND refresh: components + tokens + routes + health JSON + re-stamp per-component usage MDX

# App flow + component lists: re-run the repo-local app-graph extractor if present, then VALIDATE it
[ -f .storybook/scripts/extract-app-graph.mjs ] && node .storybook/scripts/extract-app-graph.mjs   # refresh app-graph.json + component-pages.json
[ -f .storybook/app-graph.json ] && node -e 'const g=require("./.storybook/app-graph.json"); const by={}; for(const e of g.edges) by[e.attribution]=(by[e.attribution]||0)+1; console.log("flow edges by attribution:", by, "| unresolved:", g.unresolvedEdges.length); for (const u of g.unresolvedEdges) console.log("  UNRESOLVED", u.provenance, "→", u.to, "—", u.reason)'

Sequence — don't run all six cold. Start with audit-drift.sh (cheap, names consolidation candidates). The HEAVY audit-archived.sh L1/L2/L3 scan only earns its cost once the catalog is sizable — skip it under ~20 chosen stories (nothing to prune yet). Run audit-controls.sh + extract-component-usage.sh when you're checking coverage; refresh-usage.sh --docs is the all-in-one rendered refresh — run it last, not alongside the individual extractors it supersedes.

One-command usage refresh. refresh-usage.sh runs all four extractors (inventory/token map, component prop-usage, routes, design-system health — every rendered file; see CONTEXT.md §STORAGE MAP) and — with --docs — re-stamps every <Name>.usage.mdx. The autodocs import the JSON, so a Storybook rebuild after this reflects current reality with no hand-editing. Trigger it three ways: per-script (here, or call one extractor directly), every audit (this step), or in CI before storybook build. The usage prop on StateGrid/StateMatrix reads the same JSON to badge each variant cell with its real call-site count (declared-but-unused → struck through).

Auto-refresh usage (no human authoring). Re-running extract-component-usage.sh here keeps component-usage.json current as the app changes; the usage-profile.mdx autodocs read that JSON at build time, so the "real usage in this app" tables update themselves on the next Storybook build. Wire the same one-liner into CI for repos that want it always fresh. If usage shifted (a variant that was declaredButUnused is now shipped, or vice-versa), flag the affected stories' usage:unused tags as a drift finding.

Decide (ask yourself)

• Drift — 17 names for "empty state" → consolidation candidate; 1 cluster → noise, ignore.
• Archive — audit-archived.sh flags >20 chosen stories → prune to the durable ledger seeded from templates/design-decisions.md ($SKILL/scripts/prune-to-ledger.sh); otherwise leave it.
• Pending decisions — stale decision:pending → remind stakeholders.
• Controls coverage — audit-controls.sh flags component stories that pass args but wire neither argTypes nor a controls-disable (so they expose no usable Controls panel), and warns if .storybook/manager.ts doesn't showPanel. Fix by wiring argTypes (unions → select, flags → boolean, group via table.category, hide non-serializable props) — or disabling controls on render-only stories — per the Controls pattern in sb-stories (its without-mcp reference, §13).
• Flow / app-graph drift — the validate step above prints edges-by-attribution + every UNRESOLVED. If unresolvedEdges grew, or a new entry's reason is "dynamic dispatch" / a nav idiom the extractor doesn't parse, that's a real edge being dropped — the app added a navigation pattern the script misses. Work the under-extraction loop in sb-flows (flow-capture.md), extend the repo-local extract-app-graph.mjs, re-validate. Never accept a silently shrinking map: an edge with no source is unresolved, not absent. (Same JSON also feeds the component lists — a component that lost all parents/pages is a dead-component or import-trace finding.)

Load references/galleries-and-tags.md when scaffolding a TagGallery, references/lifecycle-tags.md for archive/deprecation, references/propagate-workflow.md when a prune candidate appears, references/anti-patterns.md when reviewing AI-generated stories (the full list — mega-stories, inline mocks, CSF2, pseudo-class misuse) — pick based on what surfaced, don't load all four.

Ledger (durable memory you steer)

Append findings to .storybook/audit/findings.md (timestamped, append-only). Log-only rule: a real prod issue out of scope for the story pass (unassociated <label>, sub-AA contrast, headless crash) gets a LOG-ONLY: marker — record it, keep moving, don't touch app code unless asked. After seeding the ledger, prompt to commit it (git add .storybook/audit/) — an uncommitted ledger is one git clean from gone.

Next

/sb-hub → usually a sb-stories consolidation pass or a sb-ship for a graduated Explore.