Accessibility checks
Sisyphus finds elements the same way screen readers do. So when it can’t name a button, neither can a blind user. Every test run checks for that, at no extra cost.
What it checks
Section titled “What it checks”After a test’s steps, and at every screenshot: step, Sisyphus looks at the screen for:
| Rule | Finds |
|---|---|
unnamed-control |
A button, link, field, checkbox, tab, menu item or similar with no accessible name. A screen reader just says “button”. |
hidden-control |
A control inside an element that hides its children from assistive tech (inside an option, a button, an image, or aria-hidden), so it can’t be reached. Web and Electron. |
Findings are warnings: printed by sis run, listed in Studio’s Runs, included in results for AI agents.
✓ Rows match by text whatever their role 0.1s ⚠ button unnamed-control · this button has no accessible name: screen readers announce it as just "button"Make them fail the build
Section titled “Make them fail the build”defaults: a11y: strict # warn (default) | strict | offWith strict, a finding fails the test with A11Y_VIOLATION. Set it per test (a11y: strict), or for one run (sis run --a11y strict). A good path: start with warnings, fix them, then switch to strict so new ones can’t sneak in.
Fixing them
Section titled “Fixing them”- Web: give icon-only buttons an
aria-label, label every field with<label>, don’t put buttons inside other buttons or list options. - Native: set accessibility labels on custom controls.
Every fix also gives your tests a readable selector: button "Delete Milk" instead of #btn-3.
Limits
Section titled “Limits”This isn’t a full accessibility audit. It sees the screens your tests visit (at the end, and at each screenshot), up to 20 findings per test. On macOS, WebKit-based apps (Safari, Tauri) don’t expose enough to detect hidden-control; test the same UI in a browser to catch those.
