accessibility-audit-49cdb5— agent skill

Configuration (fill in before first run)

Replace these placeholders: - [DEV_URL] - e.g. http://localhost:3000 - [SITEMAP_OR_ROUTE_LIST] - e.g. docs/sitemap.md - [APP_SOURCE_GLOB] - e.g. app/**/page.tsx + components/**/*.tsx, src/**/*.vue, templates/**/*.html - [DEV_COMMAND] - e.g. npm run dev, python manage.py runserver, swift run

>

Then fill in the APCA probe selectors in Step 3 (see Stack adaptation section at end).

Single source of truth for accessibility checks across the project. Combines static grep-based checks (run in all modes) with live browser checks (axe-core WCAG scan + APCA contrast measurement) when a dev server is available.

Scope boundary:

• Owns: static a11y patterns (aria, tabindex, labels, focus, keyboard), axe-core WCAG 2.2 scan, APCA contrast computation.
• Not here: design token compliance and component adoption → /ui-audit. Typography and aesthetic polish → /visual-audit. Viewport-specific a11y (WCAG 1.4.4 resize, 1.4.10 reflow, tap targets) → /responsive-audit.

Two execution modes:

• Static mode (dev server not running): Steps 0-2, 5 only. Can run concurrently with browser-based skills per pipeline.md.
• Full mode (dev server running): all steps including APCA contrast + axe-core scan. Sequential - shares the browser MCP session with other live audits.
• WCAG mode (wcag): same as full, plus axe-core best-practice tag.

Step 0 - Target resolution + mode selection

Parse $ARGUMENTS for an optional mode keyword and a target: token.

STRICT PARSING - mandatory: derive mode and target ONLY from explicit text in $ARGUMENTS. Do NOT infer from conversation context, recent blocks, or memory. Unknown tokens → treat as absent and announce ignored args in the banner.

Announce: Running accessibility-audit in [STATIC | FULL | WCAG] mode - scope: [FULL | target: <resolved>]

Step 1 - Mode detection (if mode is auto)

Run:

curl -s -o /dev/null -w "%{http_code}" [DEV_URL]

• Response 200 or 302 → proceed in full mode.
• Anything else → degrade to static mode and record in the report: ⚠️ Dev server not reachable at [DEV_URL] - APCA + axe-core checks skipped. Re-run after starting [DEV_COMMAND] for full coverage.

If mode was forced (static/full/wcag), skip this step - but if full/wcag was forced and the server is unreachable, fail explicitly: ❌ Mode [full|wcag] requested but dev server not running. Start [DEV_COMMAND] or re-run with "static".

Step 2 - Static a11y checks (A1-A8)

Read [SITEMAP_OR_ROUTE_LIST] (if present) to build the target file list - page files + component files. If no sitemap, fall back to [APP_SOURCE_GLOB] filtered by target: scope.

Read ${CLAUDE_SKILL_DIR}/CHECKS.md for severity classification and WCAG references.

Launch a single Explore subagent (model: haiku) with the full file list and the check definitions below. Pass file paths explicitly - do not ask the agent to discover them.

Ripgrep note: patterns are written for ripgrep (rg). Use | for alternation. Character classes work as-is.

Instructions for the Explore agent

"Run all 8 checks below. For each check: report total match count, list every match as file:line - excerpt, and state PASS (0 matches) or FAIL (N matches). If 0 matches, explicitly state '0 matches - PASS'. Do not skip any check.

File scope: use ONLY the files provided.

A1 - Icon-only buttons missing aria-label [Severity: Critical - WCAG 4.1.2] Find interactive elements that render only an icon (no visible text content) and lack aria-label or aria-labelledby. Use the grep pattern from [A1_ICON_BUTTON_PATTERN] in Stack adaptation. Filter out lines already containing aria-label. Expected: 0 matches. Icon-only interactive elements must have an accessible name.

A2 - Positive tabindex [Severity: Critical - WCAG 2.4.3] Pattern: tabIndex=[1-9]|tabindex="[1-9] Exclude: lines with //, {/*. Expected: 0 matches. Positive tabindex breaks natural tab order. Only tabIndex={0} or tabIndex={-1} are valid.

A3 - Focus outline suppression without compensation [Severity: High - WCAG 2.4.7] Pattern: lines containing outline-none or outline: none that do NOT also contain a focus-visible ring or outline restoration. Method: grep for outline suppression, then filter out lines that also contain focus restoration patterns (e.g. focus-visible:ring, focus:ring, :focus-visible { outline). Expected: 0 matches. Suppressing the focus outline without providing an alternative indicator breaks keyboard navigation in high-contrast mode.

A4 - Native `<img>` without alt [Severity: Medium - WCAG 1.1.1] Pattern: <img\s Exclude: lines containing //, {/*, alt=, .svg, data:, role="presentation". Expected: 0 matches. Every raster image must have an alt attribute (empty string allowed for decorative use with role="presentation").

A5 - Form inputs without accessible labels [Severity: Critical - WCAG 1.3.1, 3.3.2, 4.1.2] Find form controls (input, textarea, select - including framework-specific component names) without an associated label. Association methods: for/htmlFor matching input id, aria-label, or aria-labelledby. Use [A5_FORM_CONTROL_PATTERN] from Stack adaptation for framework-specific selectors. Check within ±5 lines for label association. Exclude: inputs inside form field wrapper components with a sibling label component (framework form libraries handle labeling internally). Expected: 0 matches. Every form control must have an accessible name.

A6 - Focus indicator too thin [Severity: High - WCAG 1.4.11] Pattern: focus ring/outline declarations that rely on a default width without explicit sizing (e.g. bare focus-visible:ring without a size like ring-2, or outline-width under 2px). Expected: 0 matches on interactive elements. Focus indicators must meet WCAG 1.4.11 (3:1 non-text contrast) - a 1px ring is typically insufficient.

A7 - onClick on non-interactive elements [Severity: Critical - WCAG 2.1.1] Find non-interactive elements (div, span, li, p) with click/tap event handlers that lack keyboard accessibility. Use [A7_CLICK_HANDLER_PATTERN] from Stack adaptation for framework-specific event binding syntax. For each match, verify: (1) role="button" or equivalent interactive role present, AND (2) keyboard handler (onKeyDown, @keydown, on:keydown, etc.) present. Missing either = keyboard users cannot activate it. Exclude: comment lines; headless UI composition wrappers that delegate to child interactive elements. Expected: 0 matches missing keyboard support.

A8 - Navigation trigger keyboard accessibility [Severity: High - WCAG 2.1.1] Scope: main layout entry point and navigation components. Verify: the primary navigation trigger is keyboard-reachable at ALL breakpoints - not hidden at certain screen sizes without an alternative. Use [A8_NAV_TRIGGER_PATTERN] and [A8_RESPONSIVE_HIDE_PATTERN] from Stack adaptation. Flag: nav trigger wrapped only in a responsive-hide class with no matching trigger at other breakpoints. Expected: every layout exposes a keyboard-reachable nav trigger at every breakpoint."

Step 3 - APCA contrast measurement (full/wcag mode only - live)

Pick 3 representative pages from the target scope:

1. The dashboard or root route (/)
2. One data-list page (first list route in scope)
3. One form/write page (first form or wizard route in scope)

For each page:

1. mcp__playwright__browser_navigate → [DEV_URL]/<route>
2. Ensure logged in with the role from the route's sitemap entry (if auth required).
3. browser_wait_for network idle (2000ms max).
4. For each theme (light, dark):
   - Apply `[THEME_TOGGLE_ACTION]` from Stack adaptation (see Step 5d of `/visual-audit` for detection snippet).
   - browser_wait_for 500ms.
   - Run the APCA probe below via browser_evaluate.

APCA probe

Adapt selectors below to match your project's design system tokens. The defaults target common patterns (muted text, card surface, primary CTA, border). See Stack adaptation section.

// Returns computed color + background for the three C-checks.
// APCA Lc is NOT computed here - we capture the raw rgb pairs and flag qualitatively.
// The skill's report narrates the pair against Lc thresholds documented below.
const byClass = (sel) => document.querySelector(sel);

// Adapt these selectors to your design system (see Stack adaptation section)
const muted = byClass('[MUTED_TEXT_SELECTOR]');
const card = byClass('[CARD_SURFACE_SELECTOR]');
const cta = byClass('[PRIMARY_CTA_SELECTOR]');
const border = byClass('[BORDER_SELECTOR]');

const style = (el) => el ? getComputedStyle(el) : null;
return {
  theme: document.documentElement.classList.contains('dark') ? 'dark' : 'light',
  mutedOnCard: {
    color: muted ? style(muted).color : null,
    background: card ? style(card).backgroundColor : null
  },
  ctaOnBrand: {
    color: cta ? style(cta).color : null,
    background: cta ? style(cta).backgroundColor : null
  },
  borderOnBg: {
    color: border ? style(border).borderColor : null,
    background: card ? style(card).backgroundColor : null
  }
};

See ${CLAUDE_SKILL_DIR}/CHECKS.md for APCA thresholds (Lc 75/60/45/15) and C1-C3 severity classification.

Record each rgb pair. The report narrates against thresholds qualitatively (getComputedStyle returns resolved rgb, not Lc - final verdict combines the rgb with visible contrast in the screenshot when available).

Step 4 - axe-core browser scan (full/wcag mode only - live)

For each of the 3 pages from Step 3:

// Inject axe-core from CDN
await new Promise((resolve, reject) => {
  const s = document.createElement('script');
  s.src = '[AXE_CORE_CDN_URL]'; // default: https://cdnjs.cloudflare.com/ajax/libs/axe-core/4.10.2/axe.min.js - update in Stack adaptation
  s.onload = resolve;
  s.onerror = reject;
  document.head.appendChild(s);
});

// WCAG 2a / 2aa / 2.1aa / 2.2aa (+ best-practice in wcag mode)
const tags = ['wcag2a', 'wcag2aa', 'wcag21aa', 'wcag22aa'];
// In wcag mode, also push 'best-practice'

const results = await axe.run(document, {
  runOnly: { type: 'tag', values: tags },
  reporter: 'v2'
});

return {
  violations: results.violations.map(v => ({
    id: v.id,
    impact: v.impact,
    description: v.description,
    nodes: v.nodes.length,
    example: v.nodes[0]?.html?.slice(0, 120)
  })),
  passes: results.passes.length,
  incomplete: results.incomplete.length
};

See ${CLAUDE_SKILL_DIR}/CHECKS.md for axe severity mapping (X1-X3) and known false positive suppression guidance.

Step 5 - Report and backlog decision gate

Output format

## Accessibility Audit - [DATE] - [MODE] - [TARGET]

### Executive summary
[2-5 bullets - Critical and High findings only. Write concrete facts: file:line, check ID, impact.
If nothing Critical/High: "No Critical or High findings - a11y posture is clean for the audited scope."
Examples:
- "A5 Critical: 3 form inputs in [settings page file] have no accessible label"
- "C1 High: secondary text on surface background rgb(…) on dark theme - likely below Lc 45"
- "X1 Critical: 'button-name' axe violation on /dashboard - 2 nodes"]

### A11y maturity assessment
| Dimension | Rating | Notes |
|---|---|---|
| Keyboard access | strong / adequate / weak | [A2, A3, A6, A7, A8 results] |
| Accessible naming | strong / adequate / weak | [A1, A4, A5 results] |
| Contrast posture | strong / adequate / weak | [C1, C2, C3 - or "skipped, static mode"] |
| WCAG conformance | strong / adequate / weak | [axe-core X1/X2/X3 counts, or "skipped"] |
| Coverage | full / partial / static-only | [which steps ran] |

### Per-check verdicts

Static (all modes):
| # | Check | Matches | Severity | Verdict |
|---|---|---|---|---|
| A1 | Icon-only buttons missing aria-label | N | Critical | ✅/❌ |
| A2 | Positive tabindex | N | Critical | ✅/❌ |
| A3 | outline-none without focus-ring compensation | N | High | ✅/❌ |
| A4 | Native <img> without alt | N | Medium | ✅/❌ |
| A5 | Form inputs without accessible labels | N | Critical | ✅/❌ |
| A6 | Bare focus ring without explicit size | N | High | ✅/❌ |
| A7 | onClick on non-interactive elements | N | Critical | ✅/❌ |
| A8 | Navigation trigger keyboard accessibility | N | High | ✅/❌ |

Live (full/wcag mode only - else "Skipped - static mode"):
| # | Check | Result | Severity | Verdict |
|---|---|---|---|---|
| C1 | Secondary text on surface background (both themes) | [rgb pair / Lc estimate] | High / Critical | ✅/⚠️/❌ |
| C2 | CTA text on brand background | [rgb pair] | High | ✅/❌ |
| C3 | Border / icon vs card bg | [rgb pair] | Medium | ✅/❌ |

axe-core (full/wcag mode only):
| Page | Critical | Serious | Moderate | Minor | Passes |
|---|---|---|---|---|---|
| / | N | N | N | N | N |
| /[list-page] | N | N | N | N | N |
| /[form-page] | N | N | N | N | N |

Top axe violations:
[id - description - N nodes - example HTML]

### Prioritized findings
For each finding with severity Medium or above:
[SEVERITY] [A11Y-?] [check ID] - [file:line or route] - [evidence] - [fix] - [effort: S=<1h / M=half day / L=day+]

### Quick wins
[Findings that meet all three: (a) Medium or High, (b) effort S, (c) single-file fix]
Format: "A11Y-[n]: [one-line description]"
If no quick wins: state explicitly.

Backlog decision gate

Present all findings with severity Medium or above as a numbered decision list, sorted Critical → High → Medium:

Found N a11y findings at Medium or above. Which to add to backlog?

[1] [CRITICAL] A11Y-? - [check ID] - file:line - one-line description
[2] [HIGH]     A11Y-? - [check ID] - file:line - one-line description
[3] [MEDIUM]   A11Y-? - [check ID] - file:line - one-line description
...

Reply with numbers to include (e.g. "1 2 4"), "all", or "none".

Wait for explicit user response before writing anything.

Then write ONLY the approved entries to docs/refactoring-backlog.md:

• Assign ID: A11Y-[n] (next available after existing A11Y entries)
• Add row to priority index
• Add full detail section with: issue, evidence (file:line or rgb pair), check ID, fix suggestion, effort, WCAG reference

Severity guide

See ${CLAUDE_SKILL_DIR}/CHECKS.md → Overall severity classification for the full mapping (A1-A8 static, C1-C3 live, X1-X3 axe-core).

Execution notes

• Do NOT apply a11y fixes during this skill. Audit only.
• The Explore agent in Step 2 handles all static grep work. Do not duplicate searches in the main context.
• Concurrent execution: when invoked from pipeline.md Phase 5d Track A, this skill runs sequentially with /visual-audit → /ux-audit → /responsive-audit - they share the browser MCP session. /ui-audit still launches concurrently first.
• Static mode degradation: if the dev server is unreachable and mode was auto, the report must state explicitly which steps were skipped. Never silently omit checks.
• axe-core false positives: document suppressed false positives explicitly in the report. Do not flag patterns managed by headless UI libraries (inert Dialog children, Portal color-contrast measurements) as violations without manual verification.
• Complementary skills: /accessibility-audit is the single source of truth for accessibility. For design-token compliance and component adoption see /ui-audit; for aesthetic contrast and polish see /visual-audit; for viewport-specific WCAG checks (1.4.4 resize, 1.4.10 reflow, tap targets) see /responsive-audit.

Stack adaptation (fill after scaffold)