Express WCAG 2.2 audit of a URL or an HTML string: 20+ automated signals with a weighted 0-100 score — images without alt (1.1.1), low-quality alt text (filenames, generics, placeholders; also 1.1.1), controls without accessible names (4.1.2), form fields without labels (3.3.2), missing autocomplete on user-data fields (1.3.5), click handlers on non-interactive elements (2.1.1), unknown ARIA roles and broken aria-labelledby (4.1.2), duplicated unnamed landmarks, timed meta refresh (2.2.1), missing skip mechanism (2.4.1), lang/title (3.1.1, 2.4.2), heading structure (1.3.1), blocked zoom (1.4.4), captions (1.2.2), autoplay audio (1.4.2), generic/duplicated link text (2.4.4), target=_blank without warning (3.2.5), positive tabindex (2.4.3), aria-hidden on focusable elements, tables without th, duplicate ids, duplicate accesskeys. Each finding includes concrete remediation. Filter, not verdict: automation covers ~1/3 of WCAG; query a11y_criterion for what a criterion means. — Prefer a11y_audit_dom when JS renders the content or contrast/2.5.8/focus matter; use pages for light multi-page sampling. Static CANNOT measure class/var() CSS contrast (1.4.3): the report's not_measured says so — a 100 is never 'contrast is fine'. Returns JSON: {mode, url|file, score 0-100, findings [{criterion, severity, signal, issue, remediation}], summary {high, medium, low}, limits, score_note, overlay_widgets? (detected third-party overlay scripts — data, not a WCAG finding)}.

Find an element by its accessible name and get its deterministic route selector (the same tag:nth-of-type identity annotate() uses): controls, headings, images with alt, iframes with title, fields with an explicit name. High confidence only on exact unique matches; when several elements share the name it refuses to pick one and returns the full candidate list (ambiguous) — a match is never invented. Feeds agents that edit code by selector and panel deep-links.

Sealed patch-set of the deterministic safe fixes: build runs a11y_autofix once and records WHICH fixes were applied to WHICH page (SHA-256-keyed, format a11y-patches/1, engine version stamped); verify re-checks a set against a page — the same page must reproduce the same fixes, a different page is stale (sets never blind-apply). The artifact an agent or CI applies upstream: fixing the original web is the client job, the engine hands over the evidence.

Deep RENDERED WCAG audit via local Playwright/Chromium: real computed text contrast against effective backgrounds with alpha compositing (1.4.3), pixel-sampled measurement of text over images/gradients (background captured with the text transparent), minimum target size 24×24 (2.5.8, new in WCAG 2.2), visible focus indicator heuristic (2.4.7), plus rendered versions of the static checks (alt, accessible names, labels, headings, lang/title, tabindex, aria-hidden, captions, tables). Findings include remediation. Requires playwright: pip install playwright && playwright install chromium. — Use instead of a11y_audit_url on JS-heavy pages; needs local Playwright. Returns JSON: {mode, url, score 0-100, findings [{criterion, severity, signal, issue, remediation}], summary, score_note} — same shape as the static audit.

Exact WCAG contrast ratio for a color pair: per-criterion verdicts 1.4.3 (AA), 1.4.6 (AAA), 1.4.11 (non-text). Accepts #hex, rgb(), hsl(), CSS color names; rgba/hsl with alpha is composited over the background. If AA fails, suggests the nearest passing color. — For flat pairs; text over images needs a11y_contrast_image. Returns JSON: {text, background, ratio, verdicts [{criterion, level, threshold, passes}], aa_suggestion?}.

TEXT OVER IMAGE contrast: pixel-level sampling of the real background behind the text box → worst/median/p95 ratio, % of area passing AA, and automatic hostile-zone detection on a 3×3 grid (worst_zone). What pair-only checkers cannot do. region="x,y,w,h" recommended. — Use instead of a11y_contrast_pair whenever the background is a photo/gradient. Returns JSON: {worst_ratio, median_ratio, p95_ratio, area_passes_aa_normal_text_4_5, worst_zone, sampled_pixels}.

Nearest opaque color (RGB distance) to fg that reaches the target ratio against bg (4.5 default). Returns the color, its ratio and whether it lightens or darkens. — The fixer companion to a11y_contrast_pair failures. Returns JSON: {color, ratio, action} — the nearest opaque color reaching the target.

Generates an Accessibility Statement in HTML: art. 10 RD 1112/2018 (Spanish public sector, framework="rd1112") or European Accessibility Act wording (Directive (EU) 2019/882 / Ley 11/2023, framework="eaa"; EN output uses Directive (EU) 2016/2102 / EAA wording). The generated document is itself accessible. Saves to output_path when given. — Legal statement generator; drive inaccessible_content from audit findings and a11y_evidence. Returns {html, summary} — the accessible legal statement; writes a file only when output_path is given.

Accessibility snapshot of a URL: interactive elements (tag, role, accessible name, href) in DOM order, the REAL tab focus order, and when Playwright ≥1.49 is available the computed ACCESSIBILITY TREE (aria snapshot — what a screen reader announces). Save it before a deploy and compare after with a11y_diff. Requires local Playwright. — Capture half of the watch loop; compare with a11y_diff (or a11y_diff_urls for two live URLs). Returns JSON: {url, ts, elements [{tag, role, name, ref…}], focus_order}.

Regression diff between two accessibility snapshots (before/after a deploy): added/removed/renamed interactives and focus-order changes. ok=false means a regression to review. Accepts inline JSON (starting with "{") or file paths. — Compares two existing snapshots; a11y_diff_urls snapshots both for you. Returns JSON: {ok, added, removed, renamed, focus_order_changed, first_differences} — ok:false means a regression to review.

Snapshot two URLs and diff them in a single call (e.g. staging vs production). Requires local Playwright. — Convenience for staging-vs-production; manual control is snapshot + a11y_diff. Returns the same diff shape as a11y_diff, for two live URLs.

Returns injectable JavaScript for an aria-live announcement monitor (bookmarklet or page.evaluate): logs every dynamic-region announcement with time, politeness, role and text — what a screen reader would say, visible on screen. — Diagnostics aid to inject in a browser; not an audit by itself. Returns the injectable JavaScript as text (bookmarklet or page.evaluate).

Returns an HONEST accessibility badge as accessible SVG: score, date and scope (automated screening ≈ 1/3 of WCAG), color-coded by score. Deliberately does NOT say "conformant" — the honest seal. Embed it in audited sites or statements. — Scope: the honest SVG seal for audited sites; the machine-readable bundle is a11y_evidence. Returns the honest SVG badge as text content (the agent writes the file).

DETERMINISTIC safe auto-fixes applied to HTML — the honest anti-overlay: a short closed list of fixes where the correct answer is unique (unblock viewport zoom 1.4.4, add the exact autocomplete token 1.3.5, fill missing html lang and empty title when provided). Everything requiring judgment (alt text, contrast, accessible names) is NOT touched — it returns not_applied with the reason and remediation instead. Returns fixed_html + applied + not_applied. — Only the closed allowlist of provably safe fixes; judgment fixes come back as not_applied with remediation. Returns JSON: {fixed_html, applied [{fix, …}], not_applied [{issue, remediation}]} as text content (the agent writes the file).

Reflow check at 320px — criterion 1.4.10 (AA), the check axe and Lighthouse do not automate. Loads the URL at 1280px, then at 320px, and reports real horizontal scroll + the overflowing elements. Method note: browser zoom RE-LAYS OUT at 320 CSS px (that is the standard's own equivalence: 1280 @ 400% zoom = 320px), so the correct measurement is a 320px viewport. Requires local Playwright. — Scope: 320px reflow only (1.4.10); for the full rendered pass use a11y_audit_dom. Returns JSON: {url, viewport, horizontal_scroll, offenders [{selector, width}]}. Run once per template, not per page: fixed-width layouts are a property of the template.

Keyboard-trap detection (2.1.2) with REAL Tab walking in Chromium: up to 60 real tab stops, cycle detection (the modal pattern), then the decisive test — does ESCAPE release the cycle? A modal that cycles and releases on Escape is correct and NOT reported; a cycle Escape cannot leave is a trap (high severity). Returns the full tab stop list too. Requires local Playwright. — Scope: keyboard traps (2.1.2); focus visibility is part of a11y_audit_dom. Returns JSON: {tab_stops, trap_detected, escape_releases, findings}. Run on pages with modals, dialogs or embedded apps; a plain content page has nothing to trap.

Infinite-scroll accessibility audit — the documented disaster nobody automates (Deque guidance + ARIA APG Feed pattern; criteria 2.4.3, 4.1.3, 2.2.2). Real scrolling batches in Chromium with a live-region observer: does the focused element SURVIVE each batch (re-render destroys it — the documented failure)? Is new content ANNOUNCED (aria-live/status receives text, role=feed)? Does the feed END or offer a load-more alternative (footer reachability)? APG feed pattern as positive signal. Honest guard: if no standard items are detected (login walls, non-standard markup) the unfounded signals are skipped with a note. Requires local Playwright. — For paginated feeds only; general page audit is a11y_audit_url/dom. Returns JSON: {mode, url, score, findings [{criterion, severity, signal, issue, remediation}], summary} — same shape as the audits.

Builds the verifiable evidence pack: the machine→human handoff object for WCAG conformance work. Takes one or more audit reports (any mode — static, rendered, reflow, keyboard, scroll) and returns: the full 55-criterion A/AA matrix (automated-fail / automated-review / not-flagged — NOT pass / not-run / manual-only), each row priced with its human-effort class (MIN 1-3 / MED 5-10 / MAX 15-30 min of human review remaining) and total remaining minutes every artifact SHA-256-hashed with timestamps, an empty signature block (name, credential, date) whose statement must reference the pack's own sha256, and the tamper-evidence rule stated. Vendor-neutral: any qualified human can review it. Evidence, never conformance. — The tier-3 handoff object; feed it every audit report you have. Returns the pack JSON (a11y-evidence-pack/2) as text content (the agent writes the file). Run after the audits you trust, and pass verified codes from the manual checklist — agent-verified never erases an automated-fail.

Semi-automatic applicability scoping — the layer between the audit and the conformance claim (the pattern of the Government of Canada CIO wizard for EN 301 549, informed by the audit instead of 118 hand-answered questions). Reads the features block of the reports (media elements, form controls — positive evidence of what the audit SAW) and proposes the structural exclusions: no video anywhere → 1.2.x proposed-excluded, no form controls → 3.3.x proposed-excluded. The machine PROPOSES, the human CONFIRMS via answers {"1.2.4": {"status": "excluded", "reason": "…"}} — only confirmed exclusions may become not-applicable downstream (a11y_evidence, a11y_generate_vpat, a11y_generate_openacr); a proposal never excludes. standard=en301549 adds the EN 301 549 V4.1.1 (2026) clause to every row (9.x web / 10.x documents / 11.x software — numbering verified 55/55 A/AA) and the EN-only families (5, 6, 7, 8, 12, 13) as honest applicability rows this toolkit does not evaluate. Live counter: N of 55 criteria in scope. — The applicability layer before conformance: run after the audits, confirm the machine proposals via answers, then feed the scope here — a11y_evidence/vpat/openacr turn confirmed exclusions into not-applicable with their reason. Returns JSON: {format: a11y-scope/1, standard, target, criteria [{criterion, level, status, clause?, evidence?, reason?}], counters, en301549_families?, sha256} — statuses in-scope|proposed-excluded|excluded; only human-confirmed exclusions count downstream. Run after the audits, before any conformance artifact: confirm the machine proposals via answers, then pass the scope to a11y_evidence — only human-confirmed exclusions become not-applicable, and each exclusion carries its reason.

Form error testing (3.3.1 Error Identification, 3.3.3 Error Suggestion) — the guided flow no competitor automates: fills every validatable field with INVALID data, really submits, and judges the post-submit DOM — are errors identified in text and associated with the field (aria-invalid + aria-describedby, error summary, role=alert), or does the form swallow them? Native browser validation counts as identification (unless the form has novalidate); forms that navigate on submit are honestly noted as not measurable in-page. Requires local Playwright. Returns JSON: per-field error judgments {field, submitted_value_class, error_identified, announced, suggestion_present} plus an overall 3.3.1/3.3.3 verdict. Use on test/staging forms, or forms you own: invalid data is REALLY submitted. Do not run against production forms that trigger real emails, orders or leads.

The W3C's own parser as a toolkit mode: checks a URL or raw HTML against the Nu Html Checker (validator.w3.org/nu) — doctype, encoding, structural validity, plus alt/lang/role issues from the authoritative source, mapped to WCAG criteria where they overlap. PRIVACY: html mode POSTs the document to the W3C service (url mode shares only the URL, like a11y_audit_url); self-hosted vnu instances supported via base_url for sensitive content. Returns JSON: Nu validator messages mapped to WCAG criteria where they overlap.

Content on Hover or Focus (1.4.13): finds tooltip/overlay candidates, hovers each, and tests whether Escape dismisses the result — tooltips that do not dismiss are flagged. Requires local Playwright. — Scope: tooltips only; full audit is a11y_audit_dom, keyboard traps are a11y_keyboard. Returns JSON: {candidates, dismissed_on_escape, findings}.

Screen reader TRANSCRIPT: what a blind user HEARS on this page. Walks the accessibility tree linearly and returns the announcement text with roles, names and states — the linearized reading experience, as prose an agent can READ to understand the page from a blind user's perspective. — Read-only: for structure use a11y_snapshot, for keyboard traps use a11y_keyboard. Requires Playwright. Returns JSON: {url, total, announcements [{time, politeness, role, text}] — the linearized reading order}.

Disprover pattern (from Cloudflare's security-audit-skill): re-runs the audit against the live page and marks each finding confirmed or rejected — findings that don't reproduce are rejected with the reason. Catches false positives, race conditions, and page changes between audit and report. Returns a fresh score over confirmed findings only. — Run this before acting on any audit report. Returns JSON: {confirmed, rejected, score} — a fresh score over confirmed findings only.

Coverage ledger (from Cloudflare's security-audit-skill): persistent record of what has been audited, when, and with what result. Actions: record (add audit result), gaps (what has never been checked on a URL), summary (portfolio overview). Accumulates across runs — second audits show resolved findings. record returns the stored entry; gaps returns criteria never checked; summary returns {entries, active_signals, resolved, mean_score}. record after every audit run; gaps before planning the next pass. The ledger file accumulates across runs — give each project its own ledger_path.

Explains a WCAG 2.2 success criterion in plain language (es/en): what it requires, typical failures, and how to verify it with this toolkit (which tool automates which part). Codes like "1.4.3", "2.5.8", "4.1.2". Give a finding SIGNAL instead (e.g. img_no_alt, kbd_trap — the id in every audit finding) and it resolves to its criterion AND its published ACT rule(s) with official w3.org links. Use it whenever you need to explain WHY a finding matters or what the criterion actually says. — Knowledge lookup; does not fetch or audit anything — includes the official Understanding WCAG 2.2 link for the criterion. Returns JSON: {criterion, level, understanding, toolkit_signals, requires, typical_failures, how_to_check} — signal mode adds {signal, act_rules [{id, name, status, url}]}; {error, …} for unknown codes.

Blindfold journey verdict — the capability no other accessibility tool has: the agent attempted a REAL task (sign up, checkout, password reset) perceiving only the accessibility tree and acting by accessible name and keyboard; perceive via your browser accessibility snapshot (state-preserving). This turns its friction log into a deterministic scored verdict. Pass = task completable non-visually by the agent; friction items map to WCAG criteria and flow into the evidence pack (report:journey). Use with the blindfold-task prompt: act by accessible name, log friction per step, then call this. Task usability evidence, never conformance. Returns JSON: {verdict pass|partial|fail|blocked, score 0-100, findings [{criterion, severity, signal, issue, remediation, step}], summary, rules} — writes a file only when output_path is given. Run only after a real blindfold attempt (the blindfold-task prompt): the verdict prices the walk, it does not replace it.

Generates a VPAT 2.5 (WCAG edition) accessibility conformance report in accessible HTML from an a11y_evidence pack (or raw audit reports): the full 55-criterion WCAG 2.2 A/AA table with Remarks auto-filled from the pack (status, measured values, examples, remaining human effort) and every Conformance Level starting at Not Evaluated — automated fails force Does Not Support, review-level signals stay Not Evaluated with the evidence in Remarks, and a qualified human sets the rest via overrides (supports, partially_supports, does_not_support, not_applicable, not_evaluated). The document is a DRAFT until the signature block, which must reference the pack sha256, is completed. VPAT is the US/ADA counterpart of a11y_generate_declaration (EU/EAA): the machine fills Remarks, the human signs the claim. — The US/ADA counterpart of a11y_generate_declaration; feed it the a11y_evidence pack — levels stay Not Evaluated until a qualified human sets them. Returns {html, summary {level: count}, levels {code: level}, pack_sha256, draft} — the VPAT draft; writes a file only when output_path is given. Run after the audits you trust; a qualified human must set the final levels (overrides) and sign before publishing — the machine fills Remarks, never the claim.

Generates an OpenACR (the GSA open standard, VPAT-derived) accessibility conformance report in machine-readable YAML from an a11y_evidence pack (or raw audit reports): the full 55-criterion WCAG 2.2 A/AA matrix as chapters with one web component per criterion, adherence levels starting at not-evaluated (automated fails force does-not-support; review-level signals stay not-evaluated with the evidence in notes) and notes auto-filled from the pack (status, measured values, remaining human effort, sha256). Same honesty rule as a11y_generate_vpat: the machine fills notes, a qualified human sets levels via overrides (supports, partially_supports, …; hyphenated OpenACR forms accepted). OpenACR is what US federal procurement and GSA buyers parse — emit it from the same pack as the human-readable VPAT. — The machine-readable (GSA OpenACR YAML) sibling of a11y_generate_vpat — emit both from the same pack: humans read the VPAT, procurement systems parse the OpenACR. Returns {yaml, summary {level: count}, levels {code: level}, pack_sha256, format, draft} — the OpenACR YAML draft; writes a file only when output_path is given. Run after the audits you trust; validate and render in the GSA OpenACR editor (gsa.github.io/openacr-editor) — a qualified human must set the final adherence levels before publishing.

REAL contrast of an element against whatever is behind it (image, gradient, video, photo): loads the page in Chromium, screenshots the element matched by selector with its text turned transparent — so only the real background is sampled, no text pixels, no antialiasing edges — and contrasts the pixels with the text color (computed from the element when not given). The automated version of a11y_contrast_image: no manual screenshot, no region math. Requires local Playwright. — Use instead of a11y_contrast_image when the text lives in a live page (captures and measures it automatically); for flat pairs use a11y_contrast_pair. Returns JSON: {url, selector, text_color, worst_ratio, median_ratio, p95_ratio, area_passes_aa_normal_text_4_5, worst_zone?, sampled_pixels, interpretation}.

WebVTT caption QUALITY (1.2.2) — the substance, not the presence. Everyone stops at " exists"; this downloads and parses the actual .vtt and judges what viewers get: malformed timing lines, empty cues, overlaps, impossible reading speed (>21 chars/second, BBC/Netflix standard) and coverage vs the media duration when given. Pass the .vtt URL (same http/https guard as the audits). Not a substitute for human review of accuracy or speaker identification. Returns JSON: {mode: "vtt", cues, characters, captioned_seconds, findings [{criterion, severity, signal, issue, remediation}], summary, limits}.

PDF accessibility screening — the space nobody automated (axe, pa11y, Lighthouse: HTML only). Zero-dep structural first pass on a LOCAL file: tagged tree (/StructTreeRoot — the root failure), title, language, outline, /Alt entries, text layer (scanned detection), encryption, and AcroForm fields without accessible labels (/TU — what a screen reader announces; 3.3.2 inputs, 4.1.2 buttons). Full PDF/UA conformance needs a spec validator (veraPDF); this finds the failures that make full validation pointless. Use for enterprise document pipelines and PDFs linked from the audited site. Returns JSON: {mode: "pdf", pages, tagged, lang, title, images, alt_entries, form_fields, form_labels, findings, summary, limits}.

DOCX accessibility screening — the other half of enterprise paperwork. Zero-dep structural pass on a LOCAL .docx: images with no alt text (docPr without descr), fake headings (bold+big paragraphs with no Heading style — invisible to navigation pane and TOC), document language, tables without marked header row. Full review needs a human; this finds what makes it pointless. Returns JSON: {mode: "docx", language, images, tables, findings, summary, limits}.

Does the site HONOR prefers-reduced-motion? Loads the page normally, counts looping CSS animations (>0.15s infinite), then reloads emulating prefers-reduced-motion: reduce and counts again. If the number does not drop, the site ignores the user's preference (2.2.2; 2.3.3 AAA). Nobody else automates this — axe leaves it as manual review. Covers CSS animations and video[autoplay]; JS/WebGL/scroll-parallax remain manual. Returns JSON: {mode, normal, reduced, respects, score, findings, limits}.

Does content survive forced-colors: active (Windows High Contrast)? Loads the page normally, flags SVGs with hardcoded colored fills (they fall back to inherited and icons can vanish), then emulates forced-colors and compares. Review-grade heuristic — the definitive check is visual on Windows HCM; this prepares it. Heuristic rev 1: SVG fills only, images queued for the next rev. Returns JSON: {mode, normal, forced, score, findings, limits}.

WCAG 3 demo — the view layer nobody ships: takes an audit report and re-groups its WCAG 2.2 findings under the outcomes of the WCAG 3 Working Draft (10 September 2026). Informative only: WCAG 3 is not a completed standard, the W3C has published no official WCAG 2 to WCAG 3 mapping (this one is curated and labeled as such), and NO score, tier or conformance claim is produced — the draft's tier thresholds are still TBD. Use it to show teams where today's 2.2 findings will land tomorrow; keep auditing against WCAG 2.2. Returns JSON: {mode, draft, disclaimer, mapping_note, no_score, conformance_model, outcomes [{outcome, draft_url, findings, severity, signals, criteria_22}], counts, unmapped} — never a score. Run after any audit when the team asks "what about WCAG 3?"; re-baseline on each W3C draft publication (roughly March/September).

The router — start here when unsure which tool solves a task: state the goal in plain language ("gate my PR", "fix the contrast of my hero", "generate the legal statement") and get the ordered tool sequence with the why of each step. Deterministic keyword index over the toolkit map, not an LLM guessing; when nothing matches, the honest fallback route says where everyone begins. Makes 33 specialized tools an O(1) choice — ask once instead of scanning descriptions. Optional sequential-diagnosis mode: pass url or html and the router takes the page's cheap vitals (static pass) and orders the next exams by expected yield — the evidence decides (forms → form error testing, menus → keyboard, video → caption content), severity only breaks ties; the ordered plan never replaces the full catalog. Returns JSON: {route, steps [{step, tool, why}], alternates, prompts, cli, note} — the ordered tool sequence for the goal. Diagnosis mode (url/html passed): {mode: diagnose, vitals {score, summary, signals}, symptoms [{symptom, evidence}], steps [...], guarantee} — evidence orders the exams, severity breaks ties.