Nightly pa11y-ci runs (configured in
.github/workflows/a11y-pa11y-nightly.yml) write their JSON output here as
pa11y-ci-<UTC-date>.json. Each run also uploads the same file as a workflow
artifact named pa11y-ci-nightly-<UTC-date> retained for 90 days.
Layout#
infra/a11y-reports/
├── README.md # this file
├── .gitignore # ignores generated reports
├── pa11y-ci-2026-05-05.json # generated daily (not tracked in git)
├── pa11y-ci-2026-05-06.json
└── …
The reports themselves are excluded from version control via the .gitignore
file; the directory exists so the workflow has a stable path to write into and
so contributors browsing the repo can find the index of where reports live. The
canonical historical record is the GitHub Actions artifact stream.
Configuration#
- pa11y-ci config:
tools/a11y/.pa11yci.json— defaults (WCAG2AA, 60s timeout, 1.5s wait, 1280x800 viewport, headless Chrome with--no-sandbox) plus the per-app URL list. - Catalog dependency:
pa11y-ci: ^4.0.1inpnpm-workspace.yaml. - Workflow:
.github/workflows/a11y-pa11y-nightly.yml— cron-scheduled (08:00 UTC daily) plus manualworkflow_dispatch. Builds + starts the apps on the configured ports, runspa11y-ci --json --config tools/a11y/.pa11yci.json, writes the JSON to this directory, uploads it as an artifact, and posts a summary to the workflow's GitHub step summary.
Reading a report#
Each entry has documentTitle, pageUrl, and per-issue code, type,
message, selector, and context. To filter for serious issues only:
jq '
to_entries
| map({
url: .key,
errors: [.value.issues[] | select(.type == "error")]
})
| map(select(.errors | length > 0))
' infra/a11y-reports/pa11y-ci-<date>.json
To compare two nights:
diff \
<(jq -S 'to_entries | map({url:.key, count:(.value.issues|length)})' infra/a11y-reports/pa11y-ci-<date1>.json) \
<(jq -S 'to_entries | map({url:.key, count:(.value.issues|length)})' infra/a11y-reports/pa11y-ci-<date2>.json)
Why nightly, not per-PR?#
The per-PR a11y signal is the V1-P2-3500 axe auto-fixture (every Playwright spec
runs expectNoSeriousA11yViolations(page) via axeAfterEach). pa11y-ci runs
against fully-built apps with realistic loading states and complements the
unit-level axe checks with a slower, broader sweep that catches issues visible
only after the page settles in production-like state. Running it nightly keeps
the PR feedback loop fast while still capturing regressions within ~24h.
Per-surface pass/fail verdicts are tracked under V1-P2-3510..3519 (per-app WCAG 2.1 AA verification); V1-P2-3501 owns the tooling and report pipeline, not the green-build assertion.