mobile-stability-runner - SKILL.md Agent Skill

name: mobile-stability-runner description: Mobile crash & stability testing for a UI increment on iOS/Android. Expo-aware — auto-detects Expo (managed/bare) and picks the runner (Expo → Maestro on Expo Go / dev-client, bare React Native → Detox). Runs a crash-focused battery beyond functional E2E — cold-start smoke, background/foreground, low-memory, deep links, permission denial, network loss mid-flow, rotation, rapid navigation — and detects native crashes / JS redbox / ANR, surfacing anything captured by a configured crash reporter (Sentry / Crashlytics / Expo). Writes mobile-stability-report.md. Graceful-degrades to a local-run breadcrumb when no simulator/device is available. Use in TESTING when the platform is mobile. disable-model-invocation: false allowed-tools: Read Write Bash(mkdir ) Bash(jq ) Bash(ls ) Bash(npx ) Bash(npm ) Bash(yarn ) Bash(pnpm ) Bash(expo ) Bash(eas ) Bash(maestro ) Bash(detox ) Bash(xcrun ) Bash(adb ) Bash(curl )

Mobile Stability Runner

Functional E2E asks "does the flow work?"; this asks "does the app crash?" — the failure mode that actually bites mobile, and especially Expo, where the crash is usually not in the happy path but at cold start (a native-module / config mismatch) or on a lifecycle event. It's the mobile sibling of the front-end battery.

Phase Contract

Runs in TESTING, mobile only. Read vibeflow.config.json → platform. If it is not ios / android / all, emit "not a mobile platform — skipping" and stop. Writes test/flow files into the project + .vibeflow/reports/mobile-stability-report.md.

Step 1: Detect the stack + runner (Expo-aware)

Fingerprint the mobile project — never assume:

Expo vs bare RN — Expo if app.json / app.config.{js,ts} exists and expo is a dependency. Note managed vs bare (expo prebuild / an ios/android dir present).
Runner (auto-detect):
- Expo → Maestro (works against Expo Go or a dev-client build; the smoothest path for managed Expo). Use a dev-client / EAS Build when the increment uses native modules Expo Go can't host.
- bare React Native → Detox (existing e2e-test-writer infra).
- Record the choice + why in the report.
Crash reporter — detect Sentry (@sentry/react-native), Crashlytics (@react-native-firebase/crashlytics), or Expo error reporting; if present, you'll pull crashes captured during the run.
Device / simulator availability — xcrun simctl list devices booted (iOS) / adb devices (Android) / EAS. If none is available, do NOT fail hard: write the report with runner: <chosen>, environment: unavailable, and a ▶ Next: breadcrumb for the operator to run locally (maestro test … / eas build --profile development / npx expo start), then stop with verdict NEEDS_REVISION (not BLOCKED — it's an env gap, not a crash).

Step 2: The crash-focused battery

Beyond the functional scenarios (those belong to e2e-test-writer), exercise the crash-prone surface. Generate a flow per item in the runner's format (Maestro .yaml flows / Detox specs):

Cold-start smoke — launch the app fresh; assert it reaches the first screen with no native crash and no redbox. (Expo's #1 crash: a native module / app.config / env mismatch that only shows at startup.)
Background → foreground — send to background, wait, resume; no crash, state intact.
Low memory — simulate memory pressure (xcrun simctl … / adb shell), resume; no crash.
Deep links — open each registered scheme/universal link cold and warm; no crash, correct route.
Permission denial — deny camera/location/notifications and proceed; graceful, no crash.
Network loss mid-flow — drop connectivity during a key request; no crash, error surfaced.
Rotation + rapid navigation — rotate, push/pop fast; no crash, no leaked state.
JS unhandled rejection / fatal — assert the global handler catches it (no silent white screen).

Weight cold-start + permission + network higher for financial (a crash on a payment/auth screen is a hard fail).

Step 3: Run + collect crashes

Run the flows via the chosen runner. For each, detect:

Native crash — the app process exits / a .crash / tombstone appears (xcrun simctl … diagnose / adb logcat *:F).
JS fatal / redbox — a fatal exception or red-box error.
ANR (Android) — main-thread block.

Pull any crashes the configured reporter captured during the window (Sentry / Crashlytics / Expo). Attach the stack + the minimal repro (the flow + step).

Step 4: Report

Write .vibeflow/reports/mobile-stability-report.md:

Runner + environment + Expo-mode used.
A scenario × outcome table: passed / crashed (with crash type + stack) / not-applicable (reason) / skipped-env-unavailable.
Per finding { finding, why, impact, confidence } + a repro.
Verdict: PASS (every applicable scenario ran with no crash) / NEEDS_REVISION (env unavailable, or a non-P0 soft issue) / BLOCKED (any crash on cold-start or a P0 flow — non-negotiable).

Final Step: Auto-Consensus Marker (only on PASS — Sprint 43 arm-on-pass)

Write .vibeflow/state/consensus-needed.json (primaryArtifact .vibeflow/reports/mobile-stability-report.md, createdBy mobile-stability-runner) only if VERDICT == "PASS". On NEEDS_REVISION/BLOCKED, do not write the marker — emit the crash/env advisory + a ▶ Next: (fix the crash, or run locally with the named command), and stop, so consensus-gate doesn't block the fix.

Output

runner flow files (Maestro .yaml / Detox specs) in the project
.vibeflow/reports/mobile-stability-report.md (scenario × crash outcome + verdict)