review-a11y — accessibility review of changed code

The review counterpart of the ultra11y skill: where ultra11y produces full audits and compliance deliverables, this skill reviews exactly the code under change — a diff, a branch, a PR — and reports like a code reviewer. Same division of labour: the deterministic, install-free engine bundled with this skill does the mechanical work (machine-checkable WCAG 2.2 AA non-conformities, each tied to a success criterion, cited file:line), and the AI agent adjudicates the judgment — refute false positives, decide alt relevance/link purpose/focus logic statically from the code, and route what needs a rendered DOM to scan. A human is at most optional oversight.

Engine resolution: every command below is node scripts/ultra11y.mjs … where scripts/ultra11y.mjs lives next to this SKILL.md (the engine ships inside the skill; no install, no key). Run it with the reviewed project as the working directory. In the ultra11y repository itself, the repo root's scripts/ultra11y.mjs is the same file. On Claude Code that path is ${CLAUDESKILLDIR}/scripts/ultra11y.mjs; on Codex and OpenCode it is the skill directory the plugin installed. If neither resolves — a harness with no skill system — npx -y ultra11y <command> is the same engine. Per-harness deltas: references/harnesses.md in the ultra11y skill.

Core rules (shared with ultra11y):
1. Never invent a non-conformity: report only what the engine found or what you can
see and cite in the code under review.
2. Review the change, not the world: findings outside the reviewed scope are noted in
one line at most ("pre-existing, out of scope"), never mixed into the verdict.
3. Residual is explicit, never silently conforming: the rendering criteria (contrast,
focus visibility, zoom/reflow, content-on-hover) are named as residual risks for the changed
components — never asserted from source, resolved in the rendered page (scan). The
judgment criteria (alt relevance, link purpose, reading order) the AI agent adjudicates
from the code itself, recording a reason — no status without a justified verdict.
4. Language: write the review in the conversation's language, but technical tokens
stay in English even in French prose — aria-live stays aria-live (never
« région live »), same for tabindex, alt, role="alert", landmark role names.

1. Scope — review what changed

Pick the FIRST that applies (ask only when genuinely ambiguous):

node scripts/ultra11y.mjs audit <files the user named> --jsx --json     # explicit files
node scripts/ultra11y.mjs audit --staged --graph --json                 # staged snapshot (pre-commit review)
node scripts/ultra11y.mjs audit --changed --graph --json                # working diff
node scripts/ultra11y.mjs audit --since origin/main --graph --json     # the whole branch (PR review) — use the repo's default branch

• Always pass --graph for JSX/TSX (real AST + cross-file rules: icon-only component used

without a name, anchor target defined in another file…); it also suppresses a class of single-file false positives.

• Pass --lang matching the conversation so the engine's message/remediation strings are

in the review's language (unset, the CLI auto-detects — a fallback, not a substitute).

• Whole-repo review is an audit, not a review — offer the ultra11y skill instead.
• Test/story markup (.test., .stories., tests/) is excluded by default; name

such a file explicitly to include it, or pass --no-default-excludes.

2. Judge — the engine's findings are candidates, not verdicts

For each finding (--json gives criteria, severity, file:line, message, remediation):

• Confirm the occurrence: open the cited line; the offending element must really be

there in the change under review.

• preliminary: true findings (.vue/.svelte/.astro source, library components):

the template parser cannot see slots, conditionals or provide/inject wiring. Confirm against the rendered DOM if captures exist (.ultra11y/captures, from the ultra11y skill's render --setup harvester), otherwise report them as provisional with that caveat — or refute them with evidence from the code.

• Library-rendered JSX (DSFR, MUI…): a clean source review can still ship broken HTML

(false negatives) — say so explicitly when the change touches such components, and recommend a rendered-DOM check (scan, or the ultra11y capture pipeline).

• The judgment criteria are the AI agent's to adjudicate: alt text relevance, link purpose

in context, visible label vs accessible name, keyboard/focus logic of the changed component (read the full component source, not just the flagged line), recording a reason for each call.

• Refuting a finding requires cited evidence (the code that disproves it); silently

dropping findings is forbidden.

3. Report — a review, not a compliance document

Render in the conversation's language (Core rule 4), grouped by severity:

## Accessibility review — <scope, e.g. branch feat/x vs main, 12 files>

### 🔴 Blocking
- `src/components/Form.tsx:42` — **3.3.2 Labels or Instructions** — input has no
  label and placeholder is its only hint → associate a real <label for>, keep the
  placeholder as an example only.

### 🟠 Major / 🟡 Minor
- …same shape…

### Residual risks (rendering-dependent — resolve in the rendered page via scan)
- contrast of the new button variants (1.4.3), focus visibility on the custom menu (2.4.7)

### Verdict
1 blocking, 2 major to fix before merge; 2 provisional SFC findings to confirm on the
rendered DOM; residual risks listed above.

• Keep the engine's remediation text as the base of each fix — it is already localized

and criterion-grounded.

• Pre-existing issues spotted in passing go in ONE final line, out of the verdict.

Running as a subagent. This skill is often dispatched rather than typed: by the plugin hook on a pending git commit / git push / gh pr create, or by the ultra11y skill once its audit has produced fixes. When it is, the report above is the return value — it becomes a tool result for the caller, not a message to a human. So return it whole and nothing else: no preamble, no "I ran the engine and found…", no offer to fix. A caller that receives a summary has to guess at the findings, which is the one thing this skill exists to stop. Everything in §4 stays opt-in and belongs to the caller, not to you.

4. After the review (opt-in)

• Safe mechanical fixes: fix --staged --write --safe (auto-applies the safe codemods

to staged files and re-stages them) or node scripts/ultra11y.mjs fix <files> --write; dry-run without --write.

• Formal deliverables (dated report, auditor conformance blocks, PRD backlog, GitHub

issues, RGAA/country-standard vocabulary): that is the ultra11y skill — same engine, same findings (report, prd, check/verify gates).

• Make it automatic: installed as a Claude Code plugin, a PreToolUse hook stops a

pending git commit / git push / gh pr create that carries findings and hands them over — that is what invokes this skill without anyone asking. It never fires twice for the same findings, so a retry lands. For a gate that does not need an agent at all, the ultra11y skill's init --hook installs a mechanical pre-commit run of this same staged audit.

Do not

• Invent a non-conformity the engine did not find and you cannot cite.
• Assert contrast/focus/zoom from source — they are residual risks until rendered (resolve via scan).
• Mix pre-existing repo issues into the verdict on the change.
• Translate technical tokens in French prose (aria-live, never « région live »).