---
name: audit
description: Inspect a bounded design artifact against the explicitly selected repository and report evidence, violated rules and unresolved checks. Always read-only; use page or copy separately for authorized fixes.
---

# Evidence-based read-only audit

Read [the shared operating contract](../../references/operating-contract.md) and
[the input contract](../../references/inputs.md) first. Read workspace rules
and design core, the repo's common contract and **only** its `audit` template.
This plugin is stricter than a template permitting repairs: **audit never
edits**. Repairs require a separate explicit page/copy task. Report in
chat/stdout; do not write into the audited repo as a side effect.

Additional required inputs: `target`, exact file/frame/node/version, and
`auditScope`, bounded checks/regions/viewports/states. Use `outputType: report`,
`destination: stdout`. Axes default `5 / 5 / 5` and describe proposed remediation
only; they never alter severity or truth standards.

1. Resolve authorized read scope, target and source roles; load only relevant
   profile/brief/component/layout/token/accessibility entries. The read-only
   [context helper](../../scripts/context.py) validates contained local paths.
2. Inspect actual target evidence with available read-only tools. A snapshot
   does not prove current Figma bindings; a design spec does not prove rendering.
   Date earlier evidence and distinguish hypotheses from observations.
3. Check hierarchy/task clarity, facts, evidence, semantic units/freshness,
   component anatomy, bindings, layout, containment, responsive behavior,
   states, contrast, accessible names, keyboard and focus within scope.
4. Findings include consequence-based severity, exact node/file:line, observed
   behavior/value, violated source rule, smallest compliant correction and
   concrete verification. Unknown component/token mappings stay unresolved.
5. The optional [audit helper](../../scripts/audit.py) checks **limited local
   design-spec JSON** for token references, literal bindings and mismatched
   values. It is supplemental evidence; its exit code/fixture tests are not a
   full visual audit or a host skill invocation.
6. Return the shared envelope, prioritized findings, coverage/evidence table,
   proposed changes and limits. Preserve audited files; apply no corrections.

Results: `pass`, `fail`, `not-run`, `unresolved`, `not-applicable` with reason.
Contrast passes require measured composite pairs, typography, method and
threshold. Never compute compliance percentages. Hard **100% design-system
compliance** is the acceptance rule; partial/static/fixture coverage cannot
establish it.