CI

• ALWAYS validate workflow files with pnpx node-actionlint <file> after any edit. Exit code 0 or it's not valid.
• Playwright in CI: No official setup action (deprecated). Use pnpm exec playwright install --with-deps chromium to install browser + OS-level system dependencies. Only install the browsers you actually need.
• Simulate CI locally before pushing: Run commands with CI=true GITHUB_ACTIONS=true to catch tools that auto-detect CI and change behavior.
• Single validation entry point for CI. Define one script (e.g. validate) that chains lint + typecheck + unit tests + framework-check. CI calls only that.
• Tag-filtered CI test command. For test suites where some cases require a live backend or recorded fixtures unsuitable for CI, define a CI-specific script that filters those out (e.g. via tag filter --tags-filter='!ci-skip'). Keeps CI deterministic.
• Separate concerns by script. test:unit (unit only), test:e2e (browser/e2e only), test (everything concurrently for local), test:ci (the filtered subset). Don't overload one script with multiple modes.
• Check the exact commit SHA that triggered a CI run. Don't carry stale assumptions about what code is running.
• Read ALL errors in CI logs. Download the full log with gh api repos/.../actions/jobs/<id>/logs, then grep for FAIL, Error, unhandled, ECONNREFUSED. CI failures often have MULTIPLE root causes — fix all of them, not just the first one.
• pnpm test -- --flag does NOT work for passing flags through pnpm to the underlying test runner. The -- makes the runner treat flags as positional args. Define a dedicated script (e.g. test:ci) instead of trying to pass flags through.
• continue-on-error pattern: use when you want to collect results from multiple steps and report them together (e.g. PR comment tables) instead of failing fast.
• Adding a build step is NOT free. Estimate build time vs time saved before adding one. A 3min build for marginally faster page serving = net loss.