Playwright

Playwright drives a real browser, which is what the journey tests need: the composition, the navigation, the focus behaviour and the timing are the subject of the test, and none of them exist outside a browser.

Select the way a user finds things

The single practice that decides whether this level is worth its cost.

// What a person looks for.
page.getByRole('button', { name: 'Apply voucher' });
page.getByLabel('Voucher code');
page.getByRole('alert');

// What a test framework can see, and a user cannot.
page.locator('[data-testid="voucher-btn"]');
page.locator('.form__field > div:nth-child(2) > button');

The first group has a property the second lacks: if the selector breaks, something a real user depends on has changed. A button that lost its accessible name breaks the first and passes the second — and it is broken for every screen reader user in production.

It also means the journey test doubles as a continuous check that the interface remains navigable by role and name, which is most of what accessibility means mechanically.

Wait for conditions, never for time

Every fixed wait is a race that will be lost on a slower machine, and flakiness is what kills this level — not by deletion, but by tests being skipped one at a time, each with a good reason.

Playwright’s assertions retry by default, which is the mechanism to lean on: assert the end state and let it converge, rather than sleeping and hoping.

Run the accessibility scan in the same pass

The page is already loaded, in a real browser, at a meaningful state. Scanning there costs almost nothing and catches the class of violation that only exists in composition — heading order across a whole page, landmark nesting, focus after a route change, a modal opened from a control that focus never returns to.

Component-level scans cannot see any of those, by construction.

What to keep out of it

Component behaviour. If it can be asserted against a story, it should be: same coverage, a fraction of the runtime, and a failure that names a component instead of a step.

Third parties. Payment providers, map tiles, analytics. Slow, rate-limited, non-deterministic, and not what the test is about. Stub at the network boundary.

Visual comparison. Playwright can take screenshots, which makes it tempting to grow a page-snapshot suite here. See visual regression for why page-level snapshots are the version of that idea which stops working first.

Where it connects

• Journey tests is what to write in it.
• axe is what runs alongside it.
• Example mapping supplies the scenarios and their concrete values.