Accessibility Audit Skill

Profile keys consumed

• make.a11y
• capabilities.accessibility_audit
• capabilities.dynamic_a11y_testing
• capabilities.publish_pr_comments
• make.start
• make.start_prod
• make.ci
• make.lint_eslint
• make.test_visual
• make.post_review_findings
• architecture.source_root
• architecture.modules
• architecture.component_prefix
• framework.ui
• framework.router
• framework.i18n

Read these from .claude/react-sdlc.yml at runtime (no per-repo rendering; run /fe-sdlc-setup if the profile is missing). Every path, module, and component family resolves from the profile — never from source-project literals. A null make.* value means that lane is unconfigured: substitute the bundled static lane (make.a11y: null) or degrade with a note (make.start: null), never improvise a raw host command. Where a key is absent, treat the dependent family as N/A-with-reason rather than fabricating a finding.

Gating

Authorized / defensive use only. This skill audits a frontend the repo owner controls. The four boundary rules are a hard contract, restated here and in ../../agents/accessibility-auditor.md verbatim, and enforced at the Allowed-actions / Constraints level (a forbidden action, not advisory prose):

1. In-scope target only (verified, not assumed) — drive ONLY the profile-resolved local stack (the base URL of the make.start-booted stack). BEFORE any dynamic probe, verify the target host resolves to loopback (127.0.0.0/8, ::1), an RFC1918 / private range (10.0.0.0/8, 172.16.0.0/12, 192.168.0.0/16, fc00::/7), or a known container / compose network name. If it does not, refuse — skip dynamic probing with a note and record a boundary-violation; never drive a public, remote, or third-party host even if one is supplied in the dispatch.
2. No exfiltration — never copy data, secrets, or credentials out of the disposable container instance.
3. Interact via the app's own UI only — drive the rendered page with keyboard / click / axe through the running stack; never mutate persistent state out-of-band to force a reproduction.
4. Container-only, no destructive non-UI operations — run through make / docker compose exec against the configured stack, never host binaries; touch only what is needed to reach and exercise the audited view on a disposable instance.

Capability gates and their degrade paths (each completes the stage SUCCESS-WITH-REPORT — never a loop or hard-fail, NFR-3):

• make.a11y: null ⇒ run the bundled static lane directly through the container surface (the jsx-a11y / ARIA rules via the target mapped by make.lint_eslint, the axe static-rule pass, and ARIA-pattern + semantic-HTML inspection); add no dynamic dependency.
• capabilities.dynamic_a11y_testing: false OR make.start: null ⇒ dynamic probing is SKIPPED: with a note ("dynamic a11y probing skipped — capabilities.dynamic_a11y_testing false / make.start null / stack unreachable"); the static JSX / ARIA / semantic-HTML lanes still run, and the static-determinable classes (§5.3) stay promotable by deterministic in-tree demonstration.
• capabilities.accessibility_audit: false ⇒ the bundled axe / ARIA tool lane is SKIPPED: with a note, but the source-level JSX/ARIA audit still runs over architecture.source_root so every family still gets an explicit verdict — never a silent skip (NFR-6).
• A family whose surface is absent on this target (a tables family with no data table, a media-captions family with no audio / video, a route-announcement family with framework.router unset) is recorded N/A-with-reason up front and never dispatched (NFR-8 cost gate).

Context

This skill is the plugin's accessibility deep-audit lens — an adversarial, multi-subagent audit loop over an authorized React frontend. It is the frontend twin of the backend security lens: same triage → fan-out → find → verify → fix → regress → re-verify machinery, with the domain swapped from OWASP / CWE to WCAG 2.2 AA. It runs at two entry points with one identical body:

• Triaged inside the Stage 4 review gate — the review gate's applicability triage records an EXECUTE / NOT-APPLICABLE verdict for it like every other skill. It is the accessibility counterpart to the quality lens (the code-quality-reviewer agent) and the spec lens (the fr-nfr-reviewer agent).
• Standalone — invokable directly as react-frontend-sdlc:accessibility-audit against an authorized target.

The skill owns the loop (triage → fan-out → find → verify → fix → regress → re-verify → loop); the accessibility-auditor subagents are find / verify only; all code edits route through the existing react-implementer agent. Its culture is no false positives (a jsx-a11y / axe candidate is never a finding until reproduced, or deterministically demonstrated in-tree for the static-determinable classes), root-cause only (no eslint-disable, no jsx-a11y disable, no axe rule suppression, no baseline, no threshold edit), and container-only execution.

Technique-level remediation guidance may draw on the companion accessibility-lead agent team and the a11y review / audit skills documented in docs/companion-skills.md. A WCAG match never waives the visual-regression gate (the target mapped by make.test_visual) or the performance floors that follow.

This skill body PRESCRIBES the fan-out as text. It contains no Task-tool call — skills never invoke agents. The orchestrator that loaded this skill (the review agent or a standalone invocation that holds the Task tool) executes the dispatch the steps describe.

Task

Drive the target React frontend to zero new verified accessibility findings:

• Every WCAG 2.2 principle / a11y family in reference/wcag-catalog.md receives an explicit verdict — PROBE or N/A-with-reason. 100% triaged, no silent skips (NFR-6).
• Every PROBE family is driven to zero new verified findings through the bounded loop.
• Every reported finding carries a working reproduction, a WCAG 2.2 Success Criterion id, a conformance level (A / AA), and a severity band-with-rationale.
• Every fix is root-cause and accessible-by-default, routed through the react-implementer agent, and carries a failing-then-passing regression test.

Steps

The canonical loop. Every enumeration (WCAG families, per-family probes, remediations) lives in a reference/ file — never inline here — so this SKILL.md stays focused (NFR-9).

5.1 Triage

Read reference/wcag-catalog.md and record a per-family verdict table (PROBE / N/A-with-reason) covering the entire WCAG 2.2 AA corpus, grouped by POUR principle plus the cross-cutting surfaces (forms, modals, keyboard, live-regions, contrast, alt-text / headings, links, tables). No silent skips. Exclude N/A families before fan-out to control token cost (NFR-8):

• A family whose surface does not exist on this target (no data table → tables family N/A; no audio / video → media-captions N/A; framework.router unset → SPA route-announcement family N/A).
• A family whose only checks need a capability that is off, recorded with the gate reason (see Gating) rather than dropped silently.

The catalog file is the single source for the dispatchable family set and its WCAG SC mappings; do not re-enumerate families here.

5.2 Fan-out (parallel a11y audit)

First, boot the stack when dynamic testing is in scope: if capabilities.dynamic_a11y_testing is true and make.start is non-null, the orchestrator boots the stack via the make.start target and captures its in-scope base URL (loopback / private / container host per boundary rule 1). If the capability is false or make.start is null, dynamic probing degrades to skip-with-note and only the static lanes run — no base URL is passed.

Capture and report the booted build mode alongside the base URL. A dev-server build can mask or inflate findings: a dev error overlay injects extra focusable nodes (a false keyboard-order / focus finding), React StrictMode double-invokes and can duplicate ids or live-region announcements, and dev-only attributes differ from production output. So: pass the build mode in every dispatch; a family whose verdict depends on rendered output toggled by the build (focus order, live-region timing, duplicate-id ARIA) must cross-check against a production-parity build before reporting CLEAN or promoting a finding, and state the build mode in its verdict. Prefer booting the production-parity build (the target mapped by make.start_prod) for the authoritative dynamic pass when one exists — it is the same artifact the visual-regression and Lighthouse gates audit.

Then dispatch one accessibility-auditor subagent per PROBE family, in parallel — the proven parallel-react-implementer idiom (no new infrastructure). The orchestrator issues N Task-tool dispatches in one turn. Each dispatch passes:

• the assigned WCAG family id and its reference/audit-playbooks.md entry;
• the report contract (the finding-record schema in Format below);
• the current iteration number from this loop's MAX_ITERATIONS=5 guard;
• the resolved base URL of the booted in-scope stack and its build mode (omitted when dynamic testing is degraded — the auditor then runs static lanes only);
• on re-dispatch, the prior ledger for that family.

Resolve the family set and its profile gates from reference/wcag-catalog.md. N/A families from 5.1 are never dispatched.

5.3 Find → verify-by-reproduction

Each subagent runs the static lane (jsx-a11y / ARIA lint via the target mapped by make.lint_eslint, the axe static-rule pass, semantic-HTML + ARIA-pattern inspection over architecture.source_root) and adversarial dynamic probing per its reference/audit-playbooks.md entry (axe-core on the rendered DOM, keyboard-only traversal, accessibility-tree / accessible-name / role queries, and — when framework.router is set — route-change focus + live-region announcement checks). A static result is a candidate, never a finding, until:

• reproduced against the running stack — an axe-core violation on the rendered DOM, a keyboard sequence that traps or cannot reach a control, an accessible-name / role query that returns the wrong value; OR
• for the static-determinable classes that need no rendered styles — deterministically demonstrated in-tree: a missing alt on a meaningful image (1.1.1), a form control with no associated label (1.3.5 / 3.3.2 / 4.1.2), a missing or wrong <html lang> (3.1.1, the framework.i18n sink), a positive tabindex (2.4.3), or an ARIA role missing its required ARIA property (4.1.2) — the attribute provably absent or malformed in the JSX.

A candidate that can neither be reproduced against the stack nor deterministically demonstrated in-tree (a contrast ratio that needs computed styles, a focus-order or live-region-timing question, a keyboard-trap behavior) is recorded downgraded/dropped with the reason — never reported, never fixed.

5.4 Aggregate / dedupe / promote

The orchestrator collects the N subagent reports and:

• Dedupes findings by the tuple (WCAG SC id, JSX/component sink file:line, exercised surface) — two families hitting the same component sink collapse to one finding (NFR-8). A reusable architecture.component_prefix primitive that emits the same barrier across many views is one finding at the component, not one per view.
• Promotes only reproduced (or in-tree-demonstrated) candidates to findings; a candidate without that proof is recorded downgraded/dropped, never reported — the subagents already enforce this; the orchestrator is the second gate.
• Orders the promoted findings by severity band (Critical → Low) for the fix queue. Severity weighs impact (does the barrier block completion of a task) and reach (a shared component vs a single view).

5.5 Fix → regression-test

Route each verified finding to the react-implementer agent (never edited here, never by the auditor) with the {location, remediation, regression_test} slice of its finding record and the cited remediation from reference/remediation-patterns.md. Each fix is root-cause and accessible-by-default — a semantic-HTML element, the correct ARIA pattern, or the right framework.ui component slot — and carries a failing-then-passing regression test that reproduces the barrier and then proves it closed (an axe assertion on the rendered component, or a Testing Library getByRole / getByLabelText query for the now-correct accessible name / role). No eslint-disable, no jsx-a11y disable directive, no axe rule suppression or baseline, no config relaxation, no quality.* threshold reduction — ever. Locate elements by user-facing semantics, never by a test-only data-testid.

5.6 Re-verify → loop

Re-dispatch only still-open families (the re-dispatch set shrinks each iteration, NFR-8). The affected family's accessibility-auditor re-verifies that its reproduction no longer succeeds and that the added regression test now passes. The loop is bounded MAX_ITERATIONS=5:

iteration ← 1
loop:
  dispatch accessibility-auditor for each still-open PROBE family (parallel)  # 5.2
  collect + dedupe + promote verified findings                               # 5.4
  if zero new verified findings: exit SUCCESS-WITH-REPORT                     # exit
  route each finding → react-implementer (root-cause fix + regression test)   # 5.5
  affected-family re-verify                                                   # 5.6
  iteration ← iteration + 1
  if iteration > 5  OR  react-implementer / Ralph breaker tripped:            # NFR-2
     emit canonical escalation block; STOP (never auto-reset the breaker)
final: run the make.ci target once; forbidden-suppression scan; emit report

Run the full make.ci target once at loop close (not per iteration — the affected-family reproduction is the per-iteration correctness signal, the final make.ci is the safety net). Exit on the first iteration that yields zero new verified findings. On iteration > 5, or on a tripped react-implementer / Ralph circuit breaker, emit the canonical escalation block and STOP — never auto-reset a breaker (NFR-2).

5.7 Publish (gated)

When capabilities.publish_pr_comments is true, project this loop's promoted finding records (§5.4) — deduped on the same (wcag, location, surface) tuple, each auto_fixed: true with its regression_test when routed through §5.5 — to the canonical ledger JSON at ${SDLC_LEDGER_DIR:-.sdlc/review-ledgers}/accessibility.json, then publish ONE consolidated, idempotent PR comment via the target mapped by make.post_review_findings; when that key is null, the plugin substitutes its bundled review-findings poster. The poster is idempotent (a hidden <!-- sdlc-review:accessibility --> marker — it updates its prior comment, never spams), authorized (writes only to the resolved repo's own PR), and DEGRADES (NFR-3): capabilities.publish_pr_comments false / absent, gh absent, no PR, an empty ledger, a mismatched base repo, or a write failure all skip-with-note and exit 0 — publishing NEVER fails this loop. When the flag is false / absent, skip this step with a note.

Constraints

NEVER:

• Drive any host or URL outside the profile-resolved local stack, exfiltrate data, mutate state outside the app's own UI, or run a destructive non-UI operation (the four boundary rules in Gating).
• Run host binaries — execution is container-only (make / docker compose exec).
• Report a finding without a working reproduction or a deterministic in-tree demonstration (§5.3), or fabricate a finding for an N/A family (NFR-6).
• Edit code from this skill or from an accessibility-auditor subagent — fixes route through the react-implementer agent only.
• Add an eslint-disable, a jsx-a11y disable directive, an axe rule suppression or baseline, a @ts-ignore, a config relaxation, or a quality.* threshold reduction to make a finding disappear — thresholds are raise-only; fix the code.
• Add a data-testid to make a barrier or a regression test pass — locate elements by getByRole / getByLabelText / getByText, falling back to a stable id only when no semantic query fits.
• Loop or hard-fail on an absent capability — degrade with a note (NFR-3).
• Loop past MAX_ITERATIONS=5 or auto-reset a tripped breaker (NFR-2).
• Use source-project literals — every path / module / component family resolves from the profile (NFR-4).
• Issue a Task-tool call from this skill body — it is text the orchestrator follows.

ALWAYS:

• Record an explicit verdict (PROBE / N/A-with-reason) for every family in reference/wcag-catalog.md — 100% triaged.
• Promote only reproduced (or in-tree-demonstrated) candidates to findings; record the rest as downgraded/dropped.
• Map every finding to a WCAG 2.2 SC id + conformance level + severity band-with-rationale and cite its remediation from reference/remediation-patterns.md.
• Route every fix through the react-implementer agent with a failing-then-passing regression test.
• Re-dispatch only still-open families; run make.ci once at loop close; scan the diff for forbidden suppressions (the ../code-review/SKILL.md suppression scan) before declaring success.
• Run the §5.7 Publish step gated on capabilities.publish_pr_comments; it degrades with a note (NFR-3) and never fails this loop.

Format

Finding-record schema — the single hand-off shape, emitted by each accessibility-auditor and consumed by the orchestrator → react-implementer:

FINDING <family-id>-<n>
  wcag: <SC id + name, e.g. 1.4.3 Contrast (Minimum)>   # mapped from reference/wcag-catalog.md
  level: A | AA                                         # WCAG 2.2 conformance level
  severity: Critical|High|Medium|Low                    # band + one-line rationale (impact + reach)
  location: <architecture.source_root>/<path>:<line>    # the component/JSX sink, profile-resolved
  surface: <route/view> — <element role + accessible name>   # the exercised UI surface
  reproduction:
    1. <exact container command: axe-core run / keyboard sequence / accessible-name query>
    2. <exact command>                  # copy-pasteable against a freshly booted stack
  expected: <accessible behavior>   observed: <the barrier>
  remediation: <semantic-HTML / ARIA pattern / framework.ui slot primitive + cited WCAG technique>
  regression_test: <test path the fix must add>   # failing-before / passing-after (axe / role+name)

react-implementer receives {location, remediation, regression_test} and produces the root-cause fix + the failing-then-passing test; the affected-family accessibility-auditor re-verifies the barrier no longer reproduces. Severity is a band-with-rationale; the WCAG technique reference (e.g. an ARIA Authoring Practices pattern id) is an optional field, never mandated.

Run report shape — emitted at loop close:

ACCESSIBILITY-AUDIT RUN REPORT
  per-family verdict table: <family> → PROBE | N/A (<reason>)
  build mode: <dev | production-parity> (base URL or "static-only")
  per-iteration finding counts: iter 1: <n> … iter k: <n>
  iterations used: k / MAX_ITERATIONS=5
  promoted findings: <count by severity band> (by WCAG level: A / AA)
  dropped candidates: <count> (no reproduction / not in-tree-demonstrable)
  make.ci: <PASS|FAIL>   forbidden-suppression scan: <CLEAN|VIOLATION>
  status: SUCCESS-WITH-REPORT | ESCALATED

Canonical escalation block (on MAX_ITERATIONS=5 breach or breaker trip):

=== SDLC ESCALATION ===
stage: accessibility-audit
iteration: <n>/5
exit_condition: zero new verified findings (all dispatched families CLEAN)
status: NOT MET
blocking_finding: FINDING <family-id>-<n> (<wcag SC> / <level> / <severity>)
iteration_log: <one line per iteration: open families + candidates/reproduced/re-verified counts>
recommended_action: <next step for the owner, e.g. route FINDING <id> to react-implementer>
=== END ===

Verification

• Every family in reference/wcag-catalog.md received a recorded verdict (PROBE / N/A-with-reason) — 100% triaged, no silent skip.
• N/A families (no-table tables, no-media captions, no-router route announcement, capability-gated lanes) were excluded before fan-out with a recorded reason.
• Every reported finding carries reproduction steps + a WCAG 2.2 SC id + conformance level + severity band-with-rationale + a cited remediation.
• Every non-reproducible, non-in-tree-demonstrable candidate was recorded downgraded/dropped, not reported.
• Every fix is root-cause, routed through the react-implementer agent, and carries a failing-then-passing regression test.
• Only still-open families were re-dispatched each iteration (shrinking set).
• The loop exited on a zero-new-verified-findings iteration, or escalated at iteration > 5 / breaker trip with the canonical escalation block — no breaker auto-reset.
• The make.ci target exits 0 at loop close and the forbidden-suppression scan reports zero suppressions introduced (no eslint-disable / jsx-a11y disable / axe baseline / data-testid added to dodge a gate).
• No action targeted an out-of-profile host or performed a destructive non-UI operation (the four boundary rules held throughout).

Related Skills

• ../code-review/SKILL.md — the forbidden-suppression scan and per-comment evidence ledger this loop reuses.
• ../frontend-testing-workflow/SKILL.md — where each failing-then-passing regression test (axe / Testing Library) lives.
• ../frontend-component-development/SKILL.md — the accessible-by-default component patterns the root-cause fixes follow.
• ../ci-workflow/SKILL.md — the make.ci gate run once at loop close.