Ultra11y

ultra11y — audit WCAG 2.2 AA and write accessible markup

On accessibility, an automated tool only sees part of the problem. ultra11y owns that with a division of labour: the deterministic, install-free engine (node scripts/ultra11y.mjs <command> — no npm install, no key; the JSX/TSX parser is embedded in the bundle) does the mechanical work — detect the machine-checkable non-conformities and tie each to the right WCAG success criterion — and the active coding agent adjudicates the judgment criteria itself — alt relevance, link purpose in context, reading order — statically, from the evidence the engine harvests (verify --manual), each verdict gated by verify/check. Only the truly rendered-DOM criteria (computed contrast, visible focus, zoom/reflow, content-on-hover) fall to the scan tier (axe-core in a real browser); a human is at most optional oversight. Gates stop any hallucinated non-conformity from surviving, and nothing is ever silently "conforming".

Do not confuse rules with criteria. The bundle ships 93 static checks, WCAG AA has 55 success criteria, and RGAA has 106 criteria. Several rules may evidence one criterion, one rule may contribute to several pack criteria, and most criteria remain open until a scan or an adjudication. The only honest run-level coverage is the grid's live provenance tally (moteur + scan + agent, plus any named undecidable residue), never a fixed promise such as “48 agent criteria”. GitHub summaries therefore show decided/total, not a percentage over the small decided subset.

WCAG 2.2 Level AA is the worldwide core. Country standards — France's RGAA, the US Section 508, the EU EN 301 549 — are pluggable standards packs that map their criteria onto WCAG. Add --standard rgaa to re-key every command and every output — audit included, findings and all; plug an external pack at runtime with --pack ./pack.json (or a .ultra11yrc.json), no rebuild; or contribute your country (see references/standards.md). Packs (and their concrete implementation guidance — the RGAA SocialGouv/etalab good/bad patterns) can be AI-ingested and gated by pack check so a fabricated mapping never passes — see references/packs.md and references/guidance.md.

Most of the work is yours. Of the 55 WCAG 2.2 AA success criteria, the engine can decide 3 outright; 14 need a rendered page (scan) and 38 are judgment calls the agent adjudicates. So the engine's clean run is a starting point, never a verdict — and references/adjudication.md is the page you actually work from.

Core rules:

1. Never invent a non-conformity: every NC cites a real, resolvable element (check verifies it).
2. Native HTML first, ARIA last; never duplicate implicit semantics.
3. Never invent a conformity either: a C is cited exactly like an NC — it names in citations[] the harvested evidence it cleared, each anchor resolvable and drawn from the evidence that criterion was actually shown. A criterion presented with no evidence cannot be C at all; it stays manual. Neither can one whose reading was INCOMPLETE (evidenceComplete: false): a criterion you were shown part of can still fail on what you saw, but it can never be cleared over what you did not. And an adjudicated C is never merged with an engine-decided one: reports list it separately and keep it out of the automatic pass rate. 3b. What you are shown is the whole population, said once per distinct thing. Evidence is one representative per CONTENT CLASS, carrying occurrences, the other anchors in alsoAt, and the pages it appears on — 887 links across 38 captured pages are 97 distinct (text, href) pairs. Ruling on the representative rules on all of them, which is what makes an honest C reachable; and a citation naming ANY occurrence of a class the criterion carries is accepted, so cite the one you actually opened.
4. Residual is explicit, never silently conforming: the AI agent adjudicates every judgment criterion itself (verify --manual, gated), and the rendering criteria go to scan — which now CONCLUDES C (decidedBy: "scan") when every rule carrying a criterion ran on every page in scope and raised nothing, and stays silent when even one page's signals were incomplete. Any criterion still unproven stays "to assess manually" — no status without a recorded, justified verdict. To make that a gate rather than a hope, check --in <audit.json> --require-decided fails while anything is still to assess, and --allow-undecided <file> takes the criteria you genuinely cannot decide, each with its reason (a named list, never a threshold). --require-decided=pages holds EVERY page's own grid to the same bar: a criterion failing on one route is settled for the run and may still be nobody's verdict on the routes it never fired on, which is what a per-page deliverable is actually judged on. The emitted worklist is authoritative. It may deliberately contain a criterion whose run-wide status is already NC, because that failure did not decide the criterion on the other pages. Such an item is not redundant: return one grounded verdict for it (or an explicit manual reason), never leave it blank because the global tally looks decided.
5. Look the criterion up; never recall it. criteria [--standard <pack>] <id> — or ultra11y_criteria over MCP — returns the criterion's own wording, its numbered tests, and the glossary definitions those tests cite. RGAA 8.3 is not "the page needs a lang attribute"; it is test 8.3.1 with two alternative conditions, and the glossary decides what its terms mean. An auditor block that cites 8.3.1 after reading it is grounded; one that cites it from memory is a guess wearing a reference. Before auditing against a country standard, run the plan (ultra11y_method): it tells you how many of its criteria that still need adjudication to earn C — for RGAA, 104 of 106. See references/mcp.md.
6. The FINAL rendered semantic HTML must be correct. The engine sees only source; a component library (DSFR/MUI…) or .vue/.svelte/.astro SFC hides the real markup, so a green source audit is not proof. Verify the produced semantic HTML — install the zero-touch capture harvester (render --setup) so every component your tests render is serialized to .ultra11y/captures and audited, with audit --require-captures gating the blind spots. See references/automation.md / rendered.md.
7. Language: ALWAYS pass --lang matching the language of your conversation with the user; ask the user when ambiguous. Without the flag the CLI auto-detects (repo <html lang> → the active standard's default locale → English) — a scripted/CI fallback, not a substitute for passing --lang yourself.
8. Technical tokens stay in English, even in French prose. In any French deliverable you write (report commentary, PRD, tracker tickets, judgment verdicts), attribute/ element/role names and their values are code, not prose — never translate them: aria-live stays aria-live (never « région live »), same for tabindex, alt, role="alert", landmark role names. The engine's own fr catalog follows this; match it. Normative standard vocabulary (RGAA wording such as « lien d'évitement ») keeps its official French.

The reference library

Domain knowledge first, then the tooling. Read the one that matches the question in front of you.

Choose by situation

• "Produce a detailed audit/report in Claude Code" → use the active Claude session as the auditor, close the rendered and judgment criteria, challenge its claims, then publish the dated Markdown report, printable HTML and evidenced per-page dossiers. Read references/claude-code-report.md. This is the interactive publication profile; finish with its full deliverable set. The compact summary is the CI branch below.
• "Audit / compliance report" → node scripts/ultra11y.mjs audit … --json, then report (synthesis table + one auditor conformance block per NC criterion — same block prd and tickets use), then check; read references/audit.md.
• "Code rendered by a library (DSFR, MUI…) or a .vue/.svelte/.astro SFC / avoid false negatives" → audit the produced HTML, not the source template. Easiest: install the zero-touch capture harvester (render --setup) so your tests serialize every rendered component to .ultra11y/captures (auto-ingested, findings attributed to the source component; render --coverage and audit --require-captures track blind spots). Or render (build→audit recipe or SSR snapshot --scaffold) then audit on the output, and scan for computed rendering. SFC-source findings are flagged preliminary (a scope.sourceTemplate caveat); read references/rendered.md.
• "A finding looks wrong / false positive on a component" → the engine auto-suppresses most component false positives (slot/prop-injected names, component children, dynamic bindings, conditional headings) and marks SFC/library-source findings preliminary; confirm or refute the rest with verify --apply; read references/false-positives.md.
• "Large repo / audit smartly" → focus: --changed (git diff), template prioritization, dedup, --max-files; read references/scale.md.
• "Cross-file analysis (tree + dependencies), JSX/TSX as a real AST" → audit --graph resolves imports and applies cross-file rules (an icon-only component used without a name, an anchor target in another file…), no browser; read references/cross-file.md.
• "Generate the fix markdown / PRDs" → prd (the SAME auditor conformance block report's NC section renders — theme/criterion/test/WCAG+level/ finding/expected/verification in the active standard's vocabulary — as a backlog); --split criterion, --format doc for a product-requirements doc, --format remediation for the legacy dev backlog). It writes markdown only; read references/prd.md.
• "Open tickets / issues for this (GitHub · GitLab · Jira)" → tickets, a separate command that reads the audit and files it, writing no markdown. --grain picks what one ticket is: per criterion (default), per page, per page+criterion, per file, or one consolidated. De-dupe is by exact title, so re-running never duplicates. Always --dry-run first; read references/tickets.md.
• "Plug or author a standards pack (RGAA & beyond), AI-ingest external rules" → --pack/.ultra11yrc.json to load at runtime, pack check to gate it (the anti-hallucination guardrail), pack scaffold to start one; concrete before/after implementation guidance attaches to findings/PRD; read references/packs.md and references/guidance.md.
• "Adjudicate the judgment criteria (judgment phase)" → verify --manual --in audit.json emits an ADJUDICATION worklist (ADJUDICATE.todo.json + ADJUDICATE.md), one item per residual criterion, pre-loaded with the engine's harvested evidence (every alt, link text + context, literal colour pairs, control labels, heading outline, ARIA state, tabindex, lang-of-parts); the AI agent fills each verdict — C/NA (with a justification), NC (with a groundable finding), or manual (with a reason) — then verify --apply … --in audit.json folds them back FAIL-CLOSED; read references/judgment.md.
• "Many items to adjudicate/verify (fan the judgment out to subagents)" → orchestrate --run <dir> emits, from the run's CURRENT worklists, one launchable multi-agent workflow per ready phase + the agents/<role>.md dispatch contracts + a sequential RUNBOOK.md — the default execution path on a subagent-capable harness; see Orchestration — route by harness below.
• "Focus, keyboard & interaction logic (the interaction-logic part)" → the engine marks focus order/visible/trap and on-focus/on-input criteria as residual risks; the AI agent reads the full component source and adjudicates the keyboard/focus behaviour (visible-focus and the other rendered criteria go to scan); read references/focus-and-logic.md.
• "Put the fixes in place" → fix (dry-run by default, --write applies the safe codemods, proposes the rest without inventing anything); read references/fix.md.
• "Fix by priority, no regressions (correction phase)" → fix (--write, --iterate) + the prd backlog, blocking→major→minor; read references/correction.md.
• "Automatic repo gate (hook / CI)" → init --hook writes a git pre-commit gate over the strict staged snapshot (audits the exact index blobs, auto-applies safe fixes and re-stages them, blocks only on judgment issues); init --baseline/--ci is the opt-in "block only NEW non-conformities" variant. Installed as a Claude Code plugin, a third gate runs the review-a11y skill (not just the engine) on a pending commit/push/PR — that one is shipped by the plugin, not by init. For library/SFC code, commit rendered captures (render --setup) and stage them so the real semantic HTML is what's checked (audit --require-captures); read references/automation.md.
• "Show me the problems while I build the page" → dev --next writes a one-line overlay component and dev starts a loopback side-car: a floating panel lists the current page's non-conformities, each opening its file:line in your editor, and http://127.0.0.1:4111 accumulates the per-page grid as you browse. Renders nothing outside development; read references/devtools.md.
• "Check a page during our E2E tests" → render --e2e writes Playwright/Cypress fixtures that audit the page as your test left it (logged in, form filled, modal open) — state a separate scan run does not have — and persist each checked page as a snapshot. failOn gates the test; failOn: false records without failing, which is how you adopt it on an existing backlog; read references/e2e.md.
• "Audit the real PAGE, not the component" → a page snapshot (.ultra11y/pages/<id>/) is the whole rendered document plus the browser-only signals (computed styles, boxes, a11y tree, screenshot), ingested by audit automatically. Because it is a FULL document, the page-scoped rules run on it — that is where RGAA 8.3 (lang), 8.5/8.6 (title) and 12.6 become decidable at all; read references/pages.md.
• "Decide the rendering criteria without a browser in CI" → once a page snapshot exists, the rendered tier runs inside the ordinary audit: computed-style contrast (RGAA 3.2/10.5), screenshot-measured contrast for text over a gradient/image (the CSSOM blind spot axe-core shares), and colour-only link identification (RGAA 10.6, previously undecidable). Each leaves the criterion manual when it cannot measure — never a guess; read references/pages.md.
• "Give me the RGAA grid, page by page" → pages --in audits/audit-latest.json --standard rgaa (also embedded in report): one row per criterion, one column per page URL, rebuilt from the committed audit.json alone. Two rules hold: a finding is attributed to a page only when something SAYS so (else it is reported as unattributed, never spread), and C by silence is earned only by a page whose real DOM was audited — a source-only page keeps its undecided criteria « à évaluer »; read references/pages.md.
• "I want a report PER PAGE, not a matrix" → pages --in audits/audit-latest.json --standard rgaa --format report --split page --out audits/pages: one dossier per page — identity, its screenshot, its rate, EVERY criterion of the standard with its status on that page, then each non-conformity as the ordinary auditor block. Nothing is re-decided (the same projection as the grid) and no format is invented (the same auditor block as the report), so a page sheet and the compliance report cannot disagree. check gates it against invented criteria like any other report; read references/pages.md.
• "Which pages should I even audit?" → pages discover --crawl http://localhost:3000 --write (or --sitemap <url>) writes the sample.pages block for you, taking each page's name from its served <title> and NEVER overwriting the auth/storageState/notes already declared. Then sample check lints the coverage and scan --sample scans and snapshots it. A client-rendered SPA exposes no routes in its served HTML — use a sitemap there; read references/dynamic.md.
• "Audit these URLs page by page, in CI" → the shipped Action takes urls, sitemap, crawl or sample, snapshots every page it visits and writes every criterion's status under every page both in the job summary and in audits/pages-status.md. Use pages-report: compact for that status-only package: the Markdown, page JSON, source audit and verdict ledger only (no adjudication worklists, remediation report, HTML or crops), or keep the default true for the detailed page dossiers too. Request a pages/full PR comment when the same grid is wanted on the pull request; read references/ci.md.
• "Nobody is here to rule on the judgment criteria" → judge --in audits/audit-latest.json --standard rgaa [--apply] adjudicates them with a model, for a run with no coding agent in the loop (CI, the extension, an E2E run). It is a CALLER, not a second judge: same worklist, same evidence, same prompt, and the verdicts pass the same fail-closed gate yours do. Use --runner api with an API key, or --runner claude|codex with that local CLI's existing subscription login (--runner cli still aliases Claude). Inside an agent, prefer verify --manual and rule the worklist yourself; read references/judgment.md.
• "Show the findings ON the pull request, not just a red job" → --format sarif (upload to code scanning → inline annotations at the right file:line) or --format github (::error:: workflow commands + a $GITHUB_STEP_SUMMARY table) — from report --standard rgaa when you want the pack's criteria rather than WCAG's; read references/ci.md.
• "Make this code accessible / review it" → audit the snippet (audit - < component.html) native-first; read references/authoring.md and references/forbidden-patterns.md.
• "What does criterion X mean" → criteria (e.g. criteria 1.4.3, or criteria --standard rgaa 8.3); see references/criteria.md.
• "How trustworthy is a given check / what does the engine NOT catch" → the engine is scored against the W3C ACT-Rules test corpus (~1 100 third-party examples): which checks are consistent, where recall is partial, which deviations are deliberate, and which statically-decidable rules are still missing — read references/act.md. Use it to calibrate how much of a criterion the engine really settles before you adjudicate.
• "Country standard (RGAA, Section 508, EN 301 549)" → --standard <pack> on report/prd/criteria/check/verify; see references/standards.md and references/methodology.md. For a country-standard audit, PROPOSE the scan by default: such an audit is normatively defined over a declared page sample, and an un-scanned --standard rgaa report is marked partial — say so rather than implying full coverage. The sample workflow (sample check, scan --sample, --merge) is in references/audit.md.
• "High-assurance audit" → verify --report … --semantic; see references/verify.md.
• "Check contrast / rendering (dynamic tier)" → scan <url> --merge … runs axe-core in a headless browser and, on the local runtime, probes focus visibility, 200% zoom, text spacing and content-on-hover, with bounded stateful interactions and authenticated pages; read references/dynamic.md.
• "Close the criteria that came back needs-rendered-dom" → they did not need a better model, they needed a browser. scan persists WHAT IT MEASURED beside each snapshot (probes.json, axe.json), so a page it zoomed, reflowed and tabbed through can come back conforming rather than « à évaluer » — RGAA 3.2/3.3/10.4/10.11/10.12 close on the measurement, and 10.1/12.8 become adjudicable from the real DOM. verify --manual warns before you spend a pass, check --require-rendered refuses, and the crawl is unbounded by default (--max 0 / --depth 0), announcing every page it reaches. Read references/dynamic.md and references/pages.md.

Orchestration — route by harness

The judgment phases fan out: ADJUDICATE.todo.json (one item per residual criterion) and VERIFY.todo.json (one entry per NC claim) are independent per-item worklists, and orchestrate --run <dir> emits a launchable workflow, per-role dispatch contracts and a sequential RUNBOOK from whichever worklists currently exist. Fan-out is an optimization, never a requirement: the gates are harness-independent, every phase has a sequential fallback with identical artifacts, and the fail-closed verify --apply fold always stays with you. Read references/orchestration.md for the routing table and the rules.

Command cheat sheet

The full surface is in --help; these are the lines you reach for.

node scripts/ultra11y.mjs audit "src/**/*.tsx" --graph --out audits --json   # the default audit (AST + cross-file, persisted)
node scripts/ultra11y.mjs audit - < component.html                           # a snippet via stdin
node scripts/ultra11y.mjs audit --changed --json                             # only the git diff (large repo)
node scripts/ultra11y.mjs audit --staged --fail-on blocking                  # gate EXACTLY the staged snapshot (pre-commit)
node scripts/ultra11y.mjs audit "dist/**/*.html"                             # the RENDERED HTML (reliable for DSFR/MUI…)
node scripts/ultra11y.mjs render --setup                                     # install the capture harvester (tests → .ultra11y/captures)
node scripts/ultra11y.mjs audit --require-captures                           # gate the components with no rendered capture
node scripts/ultra11y.mjs scan http://localhost:3000 --runtime local --cwd . --merge audits/audit-latest.json
node scripts/ultra11y.mjs check --in audits/audit-latest.json --require-rendered              # gate: a rendering criterion left open by a run that rendered nothing
node scripts/ultra11y.mjs pages --in audits/audit-latest.json --standard rgaa --json --out audits  # the per-page grid, for a machine (audits/pages.json)
node scripts/ultra11y.mjs verify --report audits/wcag-YYYY-MM-DD.md --in audits/audit-latest.json --manual --out audits
node scripts/ultra11y.mjs verify --apply audits/ADJUDICATE.todo.json --in audits/audit-latest.json --out audits
node scripts/ultra11y.mjs orchestrate --run audits                           # fan the judgment phases out (--eco for the sequential path)
node scripts/ultra11y.mjs report --in audits/audit-latest.json --out audits  # → audits/wcag-YYYY-MM-DD.md
node scripts/ultra11y.mjs prd     --in audits/audit-latest.json              # the same auditor block as a backlog
node scripts/ultra11y.mjs tickets --in audits/audit-latest.json --dry-run    # file it: GitHub/GitLab/Jira
node scripts/ultra11y.mjs check  --report audits/wcag-YYYY-MM-DD.md          # integrity gate
node scripts/ultra11y.mjs criteria 1.4.3                                     # one success criterion (--list for all)
node scripts/ultra11y.mjs fix "src/**/*.html" --write --iterate               # apply the safe codemods to a fixpoint
node scripts/ultra11y.mjs init --hook                                        # pre-commit gate (--baseline for the regression variant)
node scripts/ultra11y.mjs pack check ./packs/section508.json                  # gate an (AI-)authored standards pack

--standard rgaa goes on audit too, and then everything speaks RGAA. It is honoured by audit/scan/fix as well as report/prd/tickets/criteria/check/verify/judge, and .ultra11yrc.json { "standard": "rgaa" } makes it the default for all of them. Under it, audit prints an RGAA-titled summary tabulated by thématique with findings tagged [8.4] rather than [3.1.1], in the pack's own language; --json and --out write a pack-keyed document (106 criteria, 13 themes) carrying the WCAG core inside core for the pipeline; and CI annotations, SARIF, the report, the PRD and the per-page sheets name RGAA criteria and nothing else — no WCAG cross-reference anywhere in a pack deliverable. The engine still keys on success criteria internally, because a pack criterion is DEFINED as a projection of them; that is plumbing, and no reader sees it. Add --pack ./pack.json to load a standard at runtime, and --json anywhere for machine output. --lang follows the conversation (pass it explicitly — Core rule 7).

The loop: audit → render → judge → fix → re-audit

To converge on conformance (not a single pass), chain the steps, letting the agent drive the judgment and content stages:

1. Audit the source (audit … --graph) for a first map; on library-rendered code, audit the render (render → build/SSR → audit) for reliable verdicts (otherwise the scope-risk note reminds you). 1b. RENDER BEFORE YOU ADJUDICATE — this step is not optional when a rendering criterion is open. scan <url|file> --runtime local --merge audits/audit-latest.json --out audits (or scan --sample) drives a real browser, persists each page to .ultra11y/pages/<id>/ with what it measured (probes.json, axe.json), and folds it in. Skipping it does not leave those criteria merely open — it makes them undecidable by anyone: no reading of the source settles computed contrast, 200 % zoom, 320 px reflow or text spacing, so every adjudication pass over them costs a model and returns needs-rendered-dom. Measured on one RGAA run over a two-file fixture: 80 criteria to adjudicate from source alone, 41 once a single page was scanned — and of the seven a three-pass, $24.90 cascade was left holding, six closed on the measurement. verify --manual now says so before you spend anything, and check --in <audit.json> --require-rendered makes it a gate.

2. Adjudicate & refute with verify, two worklists. (a) verify --manual --in audit.json emits ADJUDICATE.todo.json — one item per residual judgment criterion, pre-loaded with the engine's harvested evidence — which the AI agent rules on (C/NC/NA, or manual with a reason when it truly needs-rendered-dom), each verdict carrying a justification or a groundable finding; verify --apply … --in audit.json folds them back FAIL-CLOSED (agent NCs become real agent:<sc> findings that re-render in §2). (b) verify --report … [--semantic] builds VERIFY.todo.json, which attacks both directions. It refutes any preliminary/SFC/library-source non-conformity the rendered DOM disproves, and — when a verdict ledger exists, which is the default — it also puts the ledger's agent-adjudicated conformities on trial, one item per citation, question inverted: does this evidence establish the criterion, or only show that its subject exists? Nothing used to challenge a C, so a criterion cleared because its subject was PRESENT rather than RIGHT shipped as a conformance claim. Same verdicts (supported/partial/refuted/unsupported); verify --apply drops the refuted/unsupported non-conformities and sends a refuted conformity's criterion back to « à évaluer » — never to NC, because refuting a conformity proves nothing against the criterion. --no-conformities opts out. This includes focus & interaction logic (read the full component source: keyboard operability, focus order/visibility, traps, on-focus/on-input changes; see references/focus-and-logic.md) and the per-rule traps in references/false-positives.md. Both worklists fan out (orchestrate --run <dir> --phase adjudicate|verify-report — see Orchestration — route by harness); the --apply fold always stays with you. Rule each criterion against its OWN text, not your memory of it. verify --manual also writes one small brief per criterion (adjudicate/<criteriaId>.md) carrying the criterion's official wording, its numbered tests and — under a country standard — that standard's own test methodology (RGAA publishes one for all 258 of its tests), plus its glossary terms, technical note and particular cases. Read that brief; under --standard rgaa it is RGAA that decides, never the WCAG success criterion behind it. Each brief also cites the criterion's official page: if a wording stays ambiguous and you have a web tool you may go read it — never to contradict the vendored text, and a web page is never an acceptable normativeRef (--no-web drops the offer; CI drops it by default).

3. Fix by priority: fix --write --iterate for the mechanical part (anti-regression gate), then hand-apply the judgment/content fixes (alt, labels, structure) guided by references/correction.md.

4. Re-audit (on the render where relevant) and repeat.

5. Deliver the auditor block. report (compliance doc: synthesis + one auditor conformance block — theme/criterion/test/WCAG+level/finding/expected/verification — per NC criterion) and prd (the same blocks as an actionable backlog; tickets filing one GitHub issue per criterion with that identical block) are two views of the ONE building block, in the language of this conversation (pass --lang explicitly — Core rule 7).

6. Hand the change over to review-a11y, in a subagent. Once the deliverable is out, the code you just fixed still has to be reviewed as a change, and that is the other skill's job — see Handing over to review-a11y below.

Stop when check and verify --apply are green again and only explicitly-named residual risks remain. (To automate the outer cadence, the harness /loop command can re-run this cycle.)

Handing over to review-a11y

This skill does the analysis end to end: audit, render, adjudicate, fix, re-audit, deliver. When it has produced fixes, dispatch review-a11y on the code under change as a subagent, and report what it returns.

Use the host's native subagent capability: Codex dispatches a subagent with the prompt below; Claude Code can use Agent(...). If the host exposes neither, follow the sequential route below.

Use the review-a11y skill on the working diff. Return its report verbatim.

Two skills, two scopes, and the split is not cosmetic:

Why a subagent rather than an inline pass. A full audit fills a context with a 3 MB engine's output, dozens of reference files and a page sample; a review has to read the diff closely with none of that in the way. Handing it over keeps the review's judgment independent of the audit that just ran — the same reason verify refutes findings instead of trusting them. If the harness has no subagent tool, invoke review-a11y directly: the routing table in references/orchestration.md names the fallbacks, and the review is identical either way.

Do not paraphrase its verdict. It returns a severity-ranked review with file:line and a one-line verdict; that is the deliverable, not a summary of it.

Combining engine, judgment and residual risk

audit classes each criterion C/NC/NA for the static subset and manual for the rest, each carrying an automatability class that says WHY it is manual and how to close it: static (the engine decides), needs-rendering (scan), judgment (you, from the harvested evidence and the decision protocol). Engine NCs are confirmed candidates with a cited file:line; a preliminary: true finding (SFC/library source) is provisional — confirm it against the render or refute it. A criterion is never silently marked conforming: the report is complete only when every applicable criterion is a justified C/NC/NA and every residual risk is named. See references/methodology.md.

Advisory recommendations are a distinct class. A good-practice signal with NO failing normative test renders under a dedicated section and can never flip a criterion to NC nor enter conformancePct. An NC needs a normativeRef; a recommendation does not. Do not promote one to the other — see references/false-positives.md.

Do not

• Invent a non-conformity the engine did not find and you cannot see (contrast on inline literal colours is decided statically; computed contrast — external CSS, variables — goes through scan (Docker tier) or is verified at render before being declared).
• Add ARIA that duplicates native semantics.
• Mark a rendering/judgment criterion "conforming" without a recorded, gated justification (agent adjudication via verify --manual, or scan evidence).
• Hand-edit references/criteria.md (generated from the WCAG dataset via criteria --generate).

Scope

Static engine: offline, deterministic, install-free; inputs are HTML + JSX/TSX (real AST, cross-file analysis via --graph) + stdin. The rendering criteria (computed contrast, reflow) are covered by the optional scan tier (axe-core, Docker or --runtime local). The local runtime additionally probes focus visibility (2.4.7), 200% text zoom (1.4.4), text spacing (1.4.12) and content-on-hover (1.4.13) — observed in the rendered page, raised as NC only when the failure is seen (a clean probe leaves the SC manual, never silently conforming); reading order and alt relevance are the AI agent's judgment, adjudicated from the harvested evidence and gated (verify --manual). Data: WCAG 2.2 © W3C (W3C Document License); the RGAA pack is RGAA 4.1.2 © DINUM, Licence Ouverte / Etalab 2.0 (see NOTICE).