Using it in CI

Install the CLI and its browser, then run an audit as a build step and gate on the exit code:

npm i -g a11y-loop
a11y-loop install-browser --only-shell   # the headless shell is enough for audits
a11y-loop doctor                         # exits 2, with the reason, if an audit cannot run
a11y-loop audit http://localhost:3000 --json --out report.json

To cache the browser between runs, set PLAYWRIGHT_BROWSERS_PATH to a cached directory for both install-browser and the audit, and key the cache on a11y-loop's Playwright version. Every command is non-interactive: none of them prompts, and install-browser ignores stdin.

A composite GitHub Action (action.yml) ships in release 0.4.0 and later: uses: ChanMeng666/a11y-loop@v0.4.1, listed on the GitHub Marketplace. On 0.3.2 or earlier, call the CLI from a run: step as above. The overview lists its inputs and outputs.

Exit codes are the loop's contract

Code	Meaning
0	No violations, threshold met, no regression, the plan carries accessibility content (or is not UI work, or is too short to assess), or (doctor) an audit can run
1	Violations found, a regression was introduced, or a UI plan names no accessibility decisions
2	Tool or usage error (e.g. a missing browser), or (doctor) an audit cannot run here

That means a CI pipeline can branch on the result directly, with no prose parsing:

a11y-loop audit http://localhost:3000 --json --out report.json || exit 1

Regression gating with diff

a11y-loop diff --before base.json --after head.json

in a PR check turns the audit into a regression gate: it fails only on genuinely NEW violations, so a PR that fixes ten violations and introduces none still passes, even though the raw count changed.

A typical pipeline runs audit against the main branch's deployed preview to produce base.json, runs audit again against the PR's preview to produce head.json, then runs diff between them as the merge gate.

Gating the plan with plan-check

a11y-loop plan-check --file plan.md

checks the step before any of that: it exits 1 when a plan that changes UI work names no accessibility decisions. It needs no browser and keeps no state, so it can run on a plan file committed with the PR, or by an agent before it builds. It checks that the questions were asked, not that the answers are any good.