WCAG
Invoking this skill measures the page you name against WCAG 2.2 Level A and AA and produces an audit report. The user gives a URL, a local file, or a dev-server address. If they give none, first look in the working directory for a dev-server address or built static HTML and offer the candidates; otherwise ask once.
Source inspection alone never settles a criterion, so every report comes from a real browser run.
Path Resolution Rule: All
references/*andscripts/*paths below are relative to this skill's installation directory (orskill://wcag/in OMP), NEVER the target repository's working directory.
Measure
node <skill-dir>/scripts/wcag.mjs <url|file.html>
Needs Node 22 or newer and a Chromium-based browser on PATH; if none is found, set CHROME_BIN to the browser's full path before the first run:
POSIX (macOS): export CHROME_BIN="/Applications/Google Chrome.app/Contents/MacOS/Google Chrome"
POSIX (Linux): export CHROME_BIN=/usr/bin/chromium
Windows (PowerShell): $env:CHROME_BIN='C:\Program Files\Google\Chrome\Application\chrome.exe'
The script runs nine checks in five passes: desktop 1280x800, phone 390x844, dark colour scheme, reduced motion, and forced colours. A run takes 1-10 minutes per page depending on page weight: run it as a background job or with a timeout of at least 600 s and read the per-pass progress lines. If a pass stalls, report the exit code and last progress line; never edit the skill's scripts. It prints a summary on stdout and writes the full JSON to .tmp/wcag/report.json in the working directory. Exit code: 0 no failures, 2 failures found, 1 the run broke.
Several pages: audit one page per job, or a few in parallel, each from its own working directory (the report path is fixed, so parallel runs overwrite each other). Never run them as one long sequential job: at 1-10 minutes a page, a batch hits a job time limit (one was killed at 1,500 s). Size each job's timeout from the per-page time above.
Every non-GET request is blocked during form submits and key presses, so a run changes nothing on the server. For a page behind a login, point the script at a static copy of the rendered page that keeps its CSS and images.
| Check | Measures | Criteria |
|---|---|---|
contrast |
Computed colour pairs with alpha, opacity, and ancestor backgrounds composited. Text over images or gradients is sampled from screenshots taken with the text hidden. Placeholders and field borders | 1.4.3, 1.4.11 |
keyboard |
Real Tab key events. Records every stop, detects traps, compares before and after pixels for a focus indicator, samples whether a sticky element covers the focused one | 2.1.2, 2.4.7, 2.4.11, 2.4.3, 2.4.1 |
targets |
Box geometry of every pointer target, with the 24 px spacing-circle test | 2.5.8 |
reflow |
Re-lays the page at 320 px, at 200% zoom, and with the 1.4.12 text-spacing overrides; finds horizontal scroll and cut-off text; flips orientation | 1.4.10, 1.4.4, 1.4.12, 1.3.4 |
ax-tree |
Chromium's accessibility tree: nameless controls, file-name alt text, vague links, heading skips, landmarks | 4.1.2, 1.1.1, 2.4.4, 1.3.1 |
input-model |
DevTools event listeners: click-only divs, down-event actions, roles without keys, label in name, hover content dismissal and hoverability. In the first pass it also presses single keys with nothing focused to find shortcuts | 2.1.1, 2.5.2, 2.5.3, 1.4.13, 2.1.4 |
dom |
Page facts: lang, title, zoom lock, autoplay, refresh, autocomplete tokens, paste blocking, looping animation | 3.1.1, 2.4.2, 1.4.4, 1.4.2, 1.3.5, 3.3.8 |
axe |
axe-core 4.14 with the wcag2a, wcag2aa, wcag21a, wcag21aa, wcag22aa tags. Contrast and target size are left to the checks above |
many |
Each check runs in each pass unless it is marked pass: "initial". contrast skips the forced-colours pass, because that mode replaces the author's colours.
Read the report
criteria.rows has one row for each of the 55 criteria:
| Status | Meaning |
|---|---|
fail |
A check measured a failure. Report it with its selector and snippet |
review |
A check found a candidate it cannot judge. A person decides |
no-failures-found |
A check measured this criterion fully and found nothing |
partial-no-failures |
Some aspects are automated and clean; the rest needs a person |
manual |
No script settles it. See references/manual-criteria.md |
no-failures-found is not conformance. Never write "passes WCAG" or "AA compliant". Write what ran: "no failures in the automated checks; N criteria need manual review".
Present the report
Report in this shape:
# WCAG 2.2 AA Audit: <url>
## Summary
- Automated checks: N failures across M criteria
- Needs human review: K candidates
- Still requires manual testing: N criteria taken from the report's manual rows (screen reader, captions, etc.)
## Measured Failures (Automated)
For each failure:
- **Criterion**: [e.g. 1.4.3 Contrast (Minimum)]
- **Selector**: `form > button.submit`
- **Measured**: computed 2.8:1 ratio (needed 4.5:1)
- **Code snippet**: `<button class="submit">Send</button>`
- **Recommended fix**: Increase contrast ratio using a compliant foreground/background hex pair.
## Needs Review
Candidates the script flagged that a human must judge, with the specific question to answer.
## Manual Testing Still Required
Screen reader pass, caption accuracy, alt-text quality, consistent navigation across pages, error-message wording.
Boundaries
Audit-only. Reports measured findings and recommends fixes for manual remediation. If the user asks what to fix, rank the findings by impact and confirm each against the source before recommending it.
Facts that stay true
WCAG 2.2 has 86 success criteria; Level A plus AA is 55 (31 A, 24 AA). 4.1.1 Parsing was removed. Large text is 24 px, or 18.66 px bold. Normal text needs 4.5:1, large text and UI components 3:1. Targets are 24 by 24 CSS px unless an exception applies. px font sizes are not failures, and several h1 elements are not failures. This skill targets 2.2.