Rebar Wiki

playwright-e2e-harness-prepitch

2 min read

Pattern: Playwright E2E harness for PrePitch

Source: raw/eval-2026-04-19-prepitch-playwright.md (CON-124 evaluator PASS, 7/9 strict). Scope: end-to-end browser + backend-WS tests for the PrePitch roleplay app.

Structure

apps/prepitch/
├── playwright.config.ts           # single chromium project, workers=1, fullyParallel=false
└── e2e/playwright/
    ├── tests/
    │   ├── text-mode.spec.ts
    │   ├── voice-legacy.spec.ts
    │   ├── voice-convai.spec.ts
    │   ├── backend-ws.spec.ts
    │   ├── backend-convai-sse.spec.ts
    │   ├── agent-vs-agent-text.spec.ts
    │   └── agent-vs-agent-voice.spec.ts
    └── fixtures/
        ├── stubs/speech-recognition.js     # MUST be .js — addInitScript never transpiles .ts
        ├── stubs/convai-websocket.js       # stubs raw window.WebSocket at wss://mock-convai.test/*
        └── audio/{tts,stt}.ts              # real OpenAI TTS + Whisper wrappers for agent-vs-agent

Rules that bit us

  1. Playwright addInitScript requires plain JS. .ts stubs never execute — the injected file is taken as-is. Symptom: window.__fakeSR never appears, SR start never replaces the global. Fix: ship fixtures/stubs/*.js.
  2. Default per-test timeout (60s) is too short for PrePitch flows. Real Claude first-chunk is 3–6s; full buyer_response is 10–20s; debrief 15–90s. Use testInfo.setTimeout(180_000..480_000) inside individual tests instead of a blunt global.
  3. fullyParallel must be false — sessions share a single Prisma DB and the in-memory activeSessions map.
  4. /personas vs /personas?template=true are different lists. Browser specs MUST use /personas so selectOption(...) matches an actual <option value>. Backend-direct WS specs can use either.
  5. Convai stub layer: stubbing raw window.WebSocket at the wss URL is more robust than stubbing the @11labs/client module surface. Document the deviation if your spec names elevenlabs-client.ts.
  6. Fake audio for agent-vs-agent-voice: pass --use-fake-ui-for-media-stream + --use-fake-device-for-media-stream to chromium.launch. Gate the real-API path behind PREPITCH_E2E_RUN_A2A_VOICE=1 — 480s runs can still time out on headless WSL hosts.
  7. No dedicated e2e/playwright/tsconfig.json exists. Root bunx tsc --noEmit covers the tree, but any spec command that references -p e2e/playwright/tsconfig.json will fail until the file is added.

Run commands

cd apps/prepitch
bunx playwright test --list         # sanity: expect 8 tests across 7 files
bunx playwright test                # full run, dev servers auto-start via webServer config
PREPITCH_E2E_RUN_A2A_VOICE=1 bunx playwright test agent-vs-agent-voice
  • Spec: apps/prepitch/specs/CON-playwright-phase1.md
  • Eval report: raw/processed/eval-2026-04-19-prepitch-playwright.md
  • Evaluator log: system/evaluator-log.md (2026-04-19 CON-124 entry)
  • Source issues: CON-120, CON-123, CON-124

Graph View

Backlinks