Accessibility Audit

Phase CRITIQUE. Measure against WCAG 2.2 level AA and report values, not opinions.

TRIGGER

• "is this accessible", "check accessibility", "WCAG", "screen reader"
• any landed change to a rendered surface, before reporting it done

• general visual or UX quality -> ui-review
• vendor platform conventions -> platform-conformance
• error and empty-state wording -> ui-microcopy

Workflow

1. Route by job, in this order. -> the chosen route is named in the report header.
   • PRIMARY measurement, any web surface: the accessibility-scanner MCP server, which
     this package already declares. It runs the axe-core WCAG 2.2 engine IN-PROCESS, so
     there is no ChromeDriver anywhere in the path and therefore no version-skew failure
     mode. It scrolls before scanning so lazy content is covered, resolves contrast over
     gradients, and returns a selector, a criterion, and a fix link.
   • web surface, for the criteria coverage and the audit's substance: upstream skill
     accessibility.
   • native surface: the matching ehmo platform skill via skill://platform-conformance,
     which carries each platform's own accessibility guidance.
   • FALLBACK, and only for a multi-URL CI gate:
     npx --yes --package=@axe-core/cli axe "<url>" --stdout --exit. It accepts more than one
     quoted URL, JSON node targets, --include and --exclude, and exits non-zero,
     which the MCP scanner cannot do because it takes one page. Two caveats bind it:
     it drives a real Chrome through ChromeDriver, so run
     npx --yes browser-driver-manager install chrome, then eval "$(npx --yes browser-driver-manager which)", and pass --chrome-path "$CHROME_TEST_PATH" with
     --chromedriver-path "$CHROMEDRIVER_TEST_PATH". Installing alone is not enough: it
     downloads the pair and puts neither on axe's path. Measured without them, it exits 2 on
     "This version of
     ChromeDriver only supports Chrome version 152. Current browser version is
     151.0.7922.174", having tested nothing at all.
     And because that environment failure also exits non-zero, a non-zero exit does NOT by
     itself mean an accessibility violation. Read the JSON and judge the findings; the exit
     code alone distinguishes nothing.
     Quote every substituted URL: a query string carries & and ?, which an unquoted
     argument hands to the shell. The bin is axe, not the package name, so pass --package.
2. Check the routed skill name is in your available skills BEFORE loading it. Reading a
   skill:// path that does not exist throws Unknown skill. -> present: LOAD and follow.
   Absent: do NOT improvise. A thin substitute audit is worse than none, because its verdict
   reads exactly like the real one. Report the gap, name the install command from
   skill://accessibility-audit/references/upstream.md, and ASK the user to run it. An
   install applies from the NEXT session, since OMP discovers plugins at startup, so never
   install and retry within this one. The accessibility-scanner MCP server above needs no
   install and, when connected, is the primary measurement route, so measurement continues
   while an install is pending. The upstream's criteria coverage does not: do not improvise it.
   Bound that continuation: both no-install routes are WEB-ONLY and cover only a runnable
   URL. Neither substitutes for source review, for a manual keyboard walkthrough, or for
   native-platform guidance. On a NATIVE surface with the ehmo skill absent, nothing
   substitutes: STOP, report, and ask.
3. Measure on the live surface with skill://ui-review: tab.ariaSnapshot() for roles and
   accessible names, tab.evaluate for computed colors and target boxes, tab.press for
   focus order and traps. -> every finding carries a measured value and a required value.

Rules

MUST Report the measured value AND the required value on every finding.
MUST Give both color values and the computed ratio on a contrast finding. A contrast claim
missing any of the three is not reported at all.
MUST Trust the rendered accessible name over the name inferred from markup.
MUST Report a criterion you could not exercise as untested, naming the blocker.
MUST Read the axe CLI's JSON before calling a non-zero exit a violation. An environment
failure exits non-zero too: measured, a ChromeDriver and Chrome mismatch exits 2 having
tested nothing, which reads exactly like a failing gate.
DEFAULT Collapse one root cause across many elements into one finding with a count.
NOT Pass a surface on markup alone, or cite a criterion number you did not test.
NOT Use the axe CLI for interactive investigation, or the MCP scanner as an exit gate.

OUTPUT
L1 VERDICT: PASS | MINOR | MAJOR -- route used, one sentence why.
Findings -- criterion number, location, measured value, required value, fix.
Untested -- criteria not exercised, with the blocker.
CAP 120w clean · 280w with findings