The browser tests had not started

The browser job spent fourteen minutes and fifty-seven seconds installing Playwright and its system dependencies. Then that step was cancelled. The browser tests were skipped. The job had stopped before it reached the work it was named for.

That is what I found in one recorded CI attempt while reviewing a workflow repair. It was not the explanation for every earlier failure. A different run had completed installation and failed at the test step. What failed in that run is outside this note. The distinction matters: the job name describes its purpose, not necessarily the work it reached.

The repair removed this command from the job:

npx playwright install --with-deps chromium webkit

In its place, the job used a prebuilt Playwright browser image, version 1.55.1, pinned to a content digest. Browser and operating-system dependency setup moved into the supplied image. Application dependencies still needed installing; the test command and fifteen-minute job timeout stayed. The exact repair commit changed only the workflow, although the larger pull request also included site changes.

There was a small trap in matching the image to the runner. The lockfile contained a top-level playwright-core at version 1.62.1. That was not the core dependency belonging to this test runner. Its recorded chain was:

@playwright/test 1.55.1
  -> playwright 1.55.1
     -> playwright-core 1.55.1

The last package was nested beneath playwright. Reading the first familiar package name would have suggested the wrong image version. Follow the runner’s dependency chain, rather than assuming every package with a similar name describes the same installation.

The subsequent Linux job ran the browser tests: 58 passed and two were skipped. I inspected the diff, lockfile, job records and saved test output; I did not perform the repair. The selected before-job metadata does not identify the precise cause of the installation cancellation. These runs are not a controlled speed comparison or a promise of future reliability.

The useful change was making more of the test environment a versioned input. When a browser job goes red, first locate the boundary it failed to cross.