QA Find Bugs Mobile

Mobile QA bug hunt

Read the shared QA contract first, then mobile/CLAUDE.md, especially its production-write and simulator instructions. Use Expo MCP on iPhone 17e unless the user names another device. A mobile viewport in Playwright is not a substitute for exercising the native app.

Arguments: optional server URL, journey names, device name, wN, SHIP=1, DRY_RUN=1. Default to the app's selected server and report mode; inspect the selected server before signing in or writing anything.

Boot the exact device and client

1. Record xcrun simctl list devices available -j, identify the iPhone 17e UDID, and record its runtime, appearance, and initial boot state. Do not use the first booted device as a proxy for the requested one.

2. Work inside mobile/. Install existing deps with yarn install --frozen-lockfile if needed. Reuse a matching Metro server or start plain npx expo start --localhost. Do not set EXPO_UNSTABLE_MCP_SERVER=1; the local binary needs no remote MCP tunnel.

3. Use the native development build io.beancount.ios. Check it is installed on the target UDID and launch it with xcrun simctl launch <udid> io.beancount.ios. If necessary, build for that device with SENTRY_DISABLE_AUTO_UPLOAD=true npx expo run:ios --device <udid>. Do not count Expo Go or an old native binary's missing module/permission as a product regression. Reusing a compatible build does not authorize copying credentials or another simulator's data container.

4. Start the local MCP server:

   node node_modules/expo-mcp/bin/expo-mcp.mjs \
     --dev-server-url http://localhost:8081 --root /absolute/path/to/mobile
   

   Connect through an MCP client or JSON-RPC stdio: initialize, notifications/initialized, tools/list, then tools/call. If the tools are absent from the agent's tool list, use this installed local server; no new dependency or hosted Expo authentication is needed. Discover its actual schemas rather than inventing a device argument.

5. The currently pinned Expo MCP chooses the sole booted iOS simulator and rejects multiple devices. Record other booted simulators, temporarily shut them down when appropriate for this QA session, and restore them afterward. If another session needs one running, report the conflict instead of silently tapping it. Verify the screenshot is from iPhone 17e before the first action.

Drive and authenticate

• Use automation_tap with {projectRoot, platform:"ios", testID} or {x,y}, automation_find_view, automation_take_screenshot, and collect_app_logs. Keep screenshots under mobile/tmp/qa-<date>/; decode image content to a file, not a base64 dump in the transcript. Bound individual log captures to 10s.
• Prefer testID and inspected view bounds. Taps use logical points. Expo MCP resizes screenshots, so do not blindly divide its image coordinates by 3. Determine the native dimensions and capture scale, or use view bounds.
• Load QA_EMAIL / QA_PASSWORD from mobile/.env, unless the user supplies another path. Pass --env-file mobile/.env to the shared helper from the repo root (or --env-file .env when working inside mobile/). Do not reuse the dashboard file implicitly; the two packages may use different QA accounts. Keep the values inside the credential helper. Tap the native Sign In action, complete its browser OAuth/PKCE authorization, and verify that control returns to the app and survives a relaunch. Never inject a fabricated token/session into reactive variables to bypass the flow.
• The pinned MCP has no typing or scrolling tool. For ordinary text, the local Simulator hardware keyboard can be driven through osascript when macOS Accessibility access is available. Credential typing must stay in a helper and target a verified field. Do not assume an automation text tool exists.
• Reach otherwise inaccessible screens with inspected Expo Router deep links: xcrun simctl openurl <udid> 'beancount:///(app)/<route>?param=value'. Record that this proves the destination, not the navigation/scroll journey. Do not use malformed or fabricated ledger/transaction IDs as a valid control.
• A development warning banner is not shipped UI. Inspect its underlying log, dismiss it if needed, and recheck. Keep genuine runtime errors distinct from instrumentation limits. Simulator haptics, camera hardware and Android remain unverified unless separately exercised.

Sweep whole journeys

Wait for each journey to settle, capture evidence and relevant logs, and relaunch before retaining a failure. Do not flip config.features.agentChat; its disabled state is deliberate. Investigate source under mobile/app/ (thin routes), mobile/src/screens/ (behavior), src/common/ (Apollo, OAuth, vars, selectors), and src/components/. Generated GraphQL code is evidence, not an edit target. Trace server failures under the relevant backend package's guidance.

Apply the shared triage, root-cause, dedupe and handoff contract. State which native journeys actually ran on iPhone 17e, which were only deep-linked, and which need credentials, write authorization, physical hardware, or another OS.