fix: do not retry navigations in the expectation loop

[cyppe]

navigate(), refresh(), back() and forward() go through Execution::waitForExpectation(). When the configured timeout is above 1000 ms, that loop sets the client timeout to 1000 ms for each attempt and retries every ExpectationFailedException; after the loop it makes one final call with the restored timeout. Client::execute() turns every Playwright error into that exception, including the timeout of an attempt. So when a page takes more than a second to load, the navigation is sent again while the first load may still be in progress. This PR runs these four methods once, as #269 already does for click(), press() and the other actions.

Example

Playwright::setTimeout(5_000);

Route::get('/', fn (): string => 'home');
Route::get('/slow', function (): string {
    usleep(1_500_000); // longer than one attempt, well within the 5 s timeout

    return 'slow page';
});

$page = visit('/');
$page->navigate('/slow');

On 5.x, the first goto('/slow') times out after one second and the loop calls it again, so /slow is served twice for one navigate() call. In an application's CI this showed up as:

Navigation to "http://127.0.0.1:43981/<product-page>" is interrupted by another navigation to "http://127.0.0.1:43981/<product-page>"

Playwright raises this when a different document commits before the one its goto is waiting for. The retry loop can issue both navigations from one call. A retried back() is worse: in the test below it does not land within the 5 s timeout (Timeout 5000ms exceeded). With this change, /slow is served once.

Why one attempt

Retrying repeats a state-changing command. A repeated goto/reload starts another load, and a repeated goBack/goForward can traverse another history entry. Playwright already waits for the requested load state: goto() and reload() send waitUntil: 'load', and goBack/goForward default to load. That wait happens within the timeout it is given, so a single attempt with the full timeout is the right shape, as for clicks. One behaviour changes on purpose: a navigation that fails is no longer re-sent automatically.

Tests

tests/Browser/Webpage/SlowNavigationTest.php raises the timeout to 5 s inside the file and restores it afterwards. Each slow response sleeps 1.5 s; the initial loads are fast.

• navigate() serves the slow page once and shows it.
• refresh() reloads once.
• back() and forward() each request the page once and land on the expected history entry. The responses send Cache-Control: no-store to avoid HTTP-cache reuse; Playwright's default Chromium launch disables the back/forward cache.

Local results:

• On unchanged 5.x source, the three tests failed in 3 of 3 runs:
  • navigate: Failed asserting that 2 is identical to 1;
  • refresh: 3 is identical to 2;
  • back: Timeout 5000ms exceeded.
• With the fix they passed in 5 of 5 runs, and 3 of 3 with Xdebug in coverage mode.
• Removing any one of the four names from the list makes exactly one test fail.
• Full composer test (exit 0): 349 passed, 27 skipped (the usual platform skips); Rector, Pint, profanity, type coverage (100 %) and PHPStan are clean.
• PHP 8.4, Playwright 1.63.0, Chromium. I did not run PHP 8.5; the CI matrix will.

These are local observations; the 1.5 s sleep leaves headroom within the 5 s timeout, but runner scheduling can never be fully ruled out.

Not covered

withinFrame() still retries its whole callback with a one-second client timeout, so a navigation inside a frame callback can still be repeated with it. That is callback replay, a separate change.

Related PRs

• fix: uploads, bodiless responses, late responses and attempt timeouts under load #279 (closed unmerged by its author) included this change among five fixes, without a test of its own. This PR isolates it with dedicated regression coverage.
• fix: await the matching navigation command response #266 makes goto() and reload() wait for the response to their own command rather than the first matching load event. That is a separate defect in Client::execute(), in a different file. Until it lands, a navigation can still return early on an unrelated load event.
• fix: give each retry attempt the configured timeout, not a fixed second #276 gives each attempt the remaining budget instead of one second. That removes the one-second cutoff for navigations that fit the budget, but failed attempts are still retried. Both PRs edit the same array, so the second to merge may need a small conflict resolution. If fix: give each retry attempt the configured timeout, not a fixed second #276 merges first, these slow-response tests could pass even without this change, so its baseline is worth re-checking.
• fix: exclude mutating actions from the expectation retry loop #249 overlaps the actions fix: don't retry actions that may already have reached the page #269 already excludes and adds withKeyDown; it does not touch navigations.

🤖 Generated with Claude Code