Walk through how you'd debug a Playwright test that flakes only in CI.

Playwright's trace viewer makes CI-only flake dramatically easier than other frameworks. The systematic approach:

1. Get the trace. Configure CI to upload test-results/ as artifacts on failure. Download the relevant trace zip locally.

npx playwright show-trace test-results/login-test-chromium/trace.zip

The trace shows everything: timeline, DOM at each step, network log, screenshots, console. Most CI-only failures are diagnosable from the trace alone.

2. Reproduce the environment. Run locally in the exact CI Docker image:

docker run -it -v $(pwd):/work -w /work \
  mcr.microsoft.com/playwright:v1.45.0-jammy \
  npx playwright test login.spec.ts

If the flake reproduces in the container, the cause is environmental (font, locale, OS, Playwright/browser version). If it doesn't, look for timing or load patterns.

3. Common CI-only causes:

• Timing: CI runners are usually slower than dev laptops; auto-wait works but assertions can hit timeout sooner. Bump per-test timeout where justified, or wait for a specific event (waitForResponse) instead of relying on Locator auto-wait.
• Missing system fonts: text wraps differently, layout shifts, screenshots diverge. Install fonts in the Docker image.
• Locale / time zone: CI often runs in UTC while local is local time. Pin TZ and LANG env vars.
• Smaller viewport: headless default is sometimes 1280×720 but config defaults to a larger size. Pin viewport in config.
• Network speed / third-party services: a real Stripe / Auth0 call that's slow in CI. Stub at the network layer.
• Concurrent test pressure: shared backend state from parallel workers. Ensure each worker has isolated test data.

4. Increase observability first. Before adding retries, add:

• More trace: 'on' (capture trace on every retry, not just first).
• video: 'retain-on-failure' to see post-failure state.
• Custom console.log markers in the test.
• A pre-failure dump of page.content() for HTML diffing.

5. Then consider retries. retries: 2 in CI is reasonable; chronic 3+ retries is a code smell.

The senior signal: trace-first mentality, environmental matching, observability over retries.