playwright-architect

Architectural lens above the existing tactical playwright-testing. Decides locator philosophy, Page Object boundaries, fixture composition, and the CI-vs-local split before the first test is written. The playwright-testing skill handles concrete assertions, selectors, and visual-regression mechanics once the design is locked.

When to use

• A new Playwright suite is starting and the directory shape, fixture hierarchy, and locator strategy are unsettled.
• An existing suite has flake > 2 % runs, slow CI (> 10 min), or a god-test file mixing unrelated journeys.
• A second app / surface (admin, mobile-web, embedded widget) needs to share fixtures with the main suite.
• German triggers: "Playwright Setup planen", "Page Objects schneiden", "warum flaket der Test?".

Do NOT use when:

• A single test fails and the question is the assertion or selector — route to playwright-testing.
• The platform is mobile-native (Detox, Appium, Maestro) — route to mobile-e2e-strategy.
• Unit / component tests are the question — Playwright is the wrong tool; route to the stack-specific testing skill.

Procedure

1. Inspect the suite, pick the locator philosophy

Review existing tests for current locator patterns; the established strategy wins unless it is the "last resort" tier. One philosophy per suite, written into CONTRIBUTING-tests.md:

Strategy	When
getByRole + accessible name	Default — couples tests to the user contract, not markup
data-testid	Legacy markup, third-party widgets, hash-suffixed CSS modules
Text content (getByText)	Static marketing pages, copy-stable surfaces
CSS / XPath	Last resort — every use is a debt entry

Mixing strategies inside one test file is the smell — pick one and fall back only with comment.

2. Cut Page Object boundaries

A Page Object owns: (a) one URL or one logical surface (modal, drawer), (b) the locators on that surface, (c) the actions a user can perform there. It does NOT own assertions about other surfaces or test setup. Boundary rule: if two POs need to call each other, introduce a flow object (SignupFlow) above them.

3. Compose fixtures, don't inherit

Playwright fixtures stack via test.extend(). Three layers max:

Layer	Owns
Base	Browser, context, storageState, network mocks
Auth	Logged-in user states (admin, member, guest)
Domain	Pre-seeded entities for a journey

Deeper stacks become un-debuggable; flatten by extracting helpers.

4. Plan flake prevention before the first failure

Bake in: auto-retry on network idle, soft assertions for parallel checks, deterministic seed data, time freezing (page.clock), explicit expect.poll() for eventual consistency, no waitForTimeout ever. If a test needs sleep, the fixture is wrong.

5. Split CI vs local execution

Local: headed browser, single worker, slowMo enabled, video off, trace-on-retry. CI: headless, sharded workers (2–8), trace-on-first-retry, video-on-failure, Github reporter + HTML. Document both in playwright.config.ts; do not let local config leak into CI.

Output format

Return:

1. Locator + Page Object plan — chosen strategy, PO list with surface and action count, flow objects when ≥ 2 POs collaborate.
2. Fixture composition — base / auth / domain layer with what each layer sets up.
3. Parallelism + flake budget — worker count, shard strategy, isolation, target flake ceiling, CI-vs-local config delta.

Concrete shape:

Suite:            <name>
Locator strategy: <getByRole | data-testid | text | mixed (justified)>
Page Objects:     <list with surface owned + action count>
Flow objects:     <list — only when ≥ 2 POs collaborate>
Fixture layers:   base / auth / domain — each with what it sets up
Parallelism:      <workers, shards, isolation strategy>
Flake budget:     ≤ 1 % failure on green main; alert threshold
CI-vs-local:      <key config delta, one bullet each>

Gotcha

• page.waitForSelector is almost always wrong — expect(locator).toBeVisible() has built-in retry. The former teaches devs to think "wait, then assert" instead of "auto-retry assertion".
• Storage-state reuse across workers requires cookie domain isolation; the same auth.json across two browsers in parallel shares a session and corrupts state.
• Visual regression in CI requires identical fonts and renderer — pin the Docker image, never run visual tests against host Chromium.
• One Page Object per URL looks clean but creates god-objects for SPA routes; cut by user surface, not by pathname.

Do NOT

• Do NOT cite this skill alongside playwright-testing in the same step — they sit at different tiers; pick one per phase.
• Do NOT design suites for component tests with this skill — Playwright Component Testing has different fixture rules; route there instead.
• Do NOT promise zero flake; budget for ≤ 1 % and instrument the metric. Zero is a goal that disguises silent retries.
• Do NOT push the architecture into the tracker as code AC — output is a design note for refinement, not implementation steps.