playwright-testing（Playwright 测试）

Frontend Testing

Unlock reliable confidence fast: enable safe refactors by choosing the right test layer, making the app observable, and eliminating nondeterminism so failures are actionable.

Philosophy: Confidence Per Minute

Frontend tests fail for two reasons: the product is broken, or the test is lying. Your job is to maximize signal and minimize "test is lying".

Before writing a test, ask:

• What user risk am I covering (money, progression, auth, data loss, crashes)?
• What's the narrowest layer that catches this bug class (pure logic vs UI vs full browser)?
• What nondeterminism exists (time, RNG, async loading, network, animations, fonts, GPU)?
• What "ready" signal can I wait on besides setTimeout?
• What should a failure print/screenshot so it's diagnosable in CI?

Core principles:

1. Test the contract, not the implementation: assert stable user-meaningful outcomes and public seams.
2. Prefer determinism over retries: make time/RNG/network controllable; remove flake at the source.
3. Observe like a debugger: console errors, network failures, screenshots, and state dumps on failure.
4. One critical flow first: a reliable smoke test beats 50 flaky tests.

Test Layer Decision Tree

Pick the cheapest layer that provides needed confidence:

Quick Start: First Smoke Test

1. Define 1 critical flow: "page loads → user can start → one key action works"
2. Add a test seam to the app (see below)
3. Choose runner: Playwright MCP for E2E, unit tests for logic
4. Fail loudly: treat console errors and failed requests as test failures
5. Stabilize: seed RNG, freeze time, fix viewport, disable animations

Concrete MCP Workflow: Testing a Game

Step-by-step sequence for testing a Phaser/canvas game:

1. mcp__playwright__browser_navigate
   → http://localhost:3000?test=1&seed=42

2. mcp__playwright__browser_evaluate
   → () => new Promise(r => { const c = () => window.__TEST__?.ready ? r(true) : setTimeout(c, 100); c(); })
   (Wait for game ready)

3. mcp__playwright__browser_console_messages
   → level: "error"
   (Fail if any errors)

4. mcp__playwright__browser_snapshot
   → Get UI state and refs

5. mcp__playwright__browser_click
   → element: "Start Button", ref: [from snapshot]

6. mcp__playwright__browser_evaluate
   → () => window.__TEST__.state()
   (Assert game state is correct)

7. mcp__playwright__browser_press_key
   → key: "ArrowRight" (or WASD for movement)

8. mcp__playwright__browser_evaluate
   → () => window.__TEST__.state().player.x
   (Verify movement happened)

9. mcp__playwright__browser_take_screenshot
   → filename: "gameplay-state.png"
   (Visual evidence after deterministic setup)

Recommended Test Seams

Add to the app for testability (read-only, stable, minimal):

window.__TEST__ = {
  ready: false,           // true after first interactive frame
  seed: null,             // current RNG seed
  sceneKey: null,         // current scene/route
  state: () => ({         // JSON-serializable snapshot
    scene: this.sceneKey,
    player: { x, y, hp },
    score: gameState.score,
    entities: entities.map(e => ({ id: e.id, type: e.type, x: e.x, y: e.y }))
  }),
  commands: {             // optional mutation commands
    reset: () => {},
    seed: (n) => {},
    skipIntro: () => {}
  }
};

Rule: Expose IDs + essential fields, not raw Phaser/engine objects.

Anti-Patterns to Avoid

❌ Testing the wrong layer: E2E tests for pure logic *Why tempting*: "Let's just test everything through the browser" *Better*: Unit tests for logic; reserve E2E for integration contracts

❌ Testing implementation details: Asserting DOM structure/classnames *Why tempting*: Easy to assert what you can see in DevTools *Better*: Assert user-meaningful outputs (text, score, HP changes)

❌ Sleep-driven tests: wait 2s then click *Why tempting*: Simple and "works on my machine" *Better*: Wait on explicit readiness (DOM marker, window.__TEST__.ready)

❌ Uncontrolled randomness: RNG/time in assertions *Why tempting*: "The game uses random, so the test should too" *Better*: Seed RNG (?seed=42), freeze time, assert stable invariants

❌ Pixel snapshots without determinism: Canvas screenshots that flake *Why tempting*: "I'll catch visual bugs automatically" *Better*: Deterministic mode first; then screenshot at known stable frames

❌ Retries as a strategy: "Just bump retries to 3" *Why tempting*: Quick fix that makes CI green *Better*: Fix the flake source; retries hide real problems

Debugging Failed Tests

When a test fails, gather evidence in this order:

1. Console errors: mcp__playwright__browser_console_messages({level: "error"})
2. Network failures: mcp__playwright__browser_network_requests() → check for non-2xx
3. Screenshot: mcp__playwright__browser_take_screenshot() → visual state at failure
4. App state: mcp__playwright__browser_evaluate({function: "() => window.__TEST__.state()"})
5. Classify the flake (see references/flake-reduction.md):

- Readiness? → add explicit wait - Timing? → control animation/physics - Environment? → lock viewport/DPR - Data? → isolate test data

Graduation Criteria: When Is Testing "Enough"?

Minimum viable test suite:

• 1 smoke test that proves the app loads and primary action works
• Test seam exists (window.__TEST__ with ready flag and state)
• Deterministic mode for canvas/games (?test=1 enables seeding)
• Console errors fail tests (no silent failures)
• CI runs tests on every push

Level up when:

• Critical paths (auth, payment, save/load) have dedicated E2E
• Unit tests cover complex logic (pathfinding, damage calc, state machines)
• Visual regression on key screens (menu, HUD) with locked determinism

Visual Regression with imgdiff.py

For pixel comparison of screenshots:

# Compare baseline to current
python scripts/imgdiff.py baseline.png current.png --out diff.png

# Allow small tolerance (anti-aliasing differences)
python scripts/imgdiff.py baseline.png current.png --max-rms 2.0

Exit codes: 0 = identical, 1 = different, 2 = error

UI Slicing Regressions (Nine-Slice / Ribbons / Bars)

Canvas UI issues (panel seams, segmented ribbons, invisible HUD fills) are best caught with a dedicated UI harness instead of the full gameplay flow.

1. Build a simple test.html/scene that loads *only* the UI assets.
2. Render raw slices next to assembled panels (multi-size), and include ribbon/bars with both “raw crop + scale” and “stitched multi-slice” views.
3. Expose window.__TEST__ with .commands.showTest(n) so Playwright can toggle each mode deterministically.
4. Capture targeted screenshots (panels, ribbons, bars) and diff them in CI.

See references/phaser-canvas-testing.md for the deterministic setup + screenshot workflow.

Variation Guidance

Adapt approach based on context:

• DOM app: Standard Playwright selectors, wait for text/elements
• Canvas game: Test seams mandatory, wait via window.__TEST__.ready
• Hybrid: DOM for menus, test seams for gameplay
• CI-only GPU: May need software rendering flags or skip visual tests
• UI slicing regressions: For nine-slice/ribbon/bar artifacts, prefer a small UI harness scene/page with deterministic modes and targeted screenshots (references/phaser-canvas-testing.md).

Bundled Resources

Read these when needed:

• references/playwright-mcp-cheatsheet.md: Detailed MCP tool patterns
• references/phaser-canvas-testing.md: Deterministic mode for Phaser games
• references/flake-reduction.md: Flake classification and fixes

Remember

You can make almost any frontend (including canvas/WebGL games) testable by adding a tiny, stable seam for readiness + state. One reliable smoke test is the foundation. Aim for tests that are boring to maintain: deterministic, explicit about readiness, and rich in failure evidence. The goal is confidence, not coverage numbers.