rabbyhub Rabby Mobile Testing
Choose the narrowest test layer that can detect the intended regression without replacing the behavior under test with mocks. Read references/test-classification.md before adding or reclassifying coverage.
Evidence Priority
When two layers make claims about the same behavior, use this authority order:
- a complete real-device App scenario;
- a real Hermes runtime scenario;
- a Node/JS integration test;
- a unit or component test.
This is an evidence hierarchy, not a reason to run every change only at the most expensive layer. Lower layers remain the fastest way to localize a fault and protect a deterministic contract. However, a passing lower-layer test can never overrule a failure above it. Treat that disagreement as a missing or incorrect lower-layer assertion, preserve the higher-fidelity failure, and add coverage at the narrowest layer that can reproduce it.
Classify by the boundary actually exercised. A typed scenario running in a real App on a physical device may count as real-device evidence; invoking the Hermes engine without the native lifecycle, storage, navigation host, and visible product result counts only as Hermes evidence. Never describe a Node bootstrap integration test as a verified App launch.
Quick Decision
| Question | Test layer |
|---|---|
| Is one pure module or reducer enough? | Unit |
| Must real internal Stores, Providers, Hooks, or Services cooperate? | JS integration |
| Must Hermes, MMKV, SQLite, a native bridge, or app lifecycle be real? | Hermes device integration |
| Must the complete user journey and package behavior be proven? | E2E |
Do not promote a test merely because it renders React. A component test that mocks its internal Store and Service is still a unit/component test.
When reorganizing an existing mock-heavy suite, preserve focused diagnostic unit tests and add integration contracts for the repository-owned cooperation that those mocks replace. Follow the reclassification rules in references/test-classification.md; mock count alone is not a reason to delete a unit test.
JS Integration Contract
- Name files
*.integration.test.tsor*.integration.test.tsxunderapps/mobile/src. - Compose real repository Stores, Providers, Hooks, Services, registries, and coordinators.
- Mock only true process boundaries such as network, time, keychain, filesystem, device APIs, and navigation when navigation itself is not under test.
- Assert visible output, emitted events, persisted state, or public Store/Service results. Avoid assertions on private implementation calls.
- Clean up subscriptions and unregister shared Service loaders in
finallyor test cleanup. - Keep the lane serial because the mobile app has process-wide registries and singleton state.
For TypeORM and repository contracts, prefer the repository's real Node
in-memory SQLite driver over mocking repositories or query results. Run the
real entity metadata, migrations, repositories, constraints, and transactions
against :memory:. This proves deterministic database semantics in CI, but it
does not prove the native SQLite bridge, file-backed durability, WAL behavior,
multi-connection ordering, force-stop behavior, or device performance.
For startup orchestration, do not mount the complete native App in Node just
to claim broad coverage. Use the real launch task definitions, startup
scheduler, phase registry, Home milestones, Service Registry, and Stores.
Replace only the explicit module-loading or process boundary with a typed,
deterministic catalog. Assert both the manual-unlock and valid-session paths,
and prove post-Home and on-demand work cannot run early.
When startup correctness depends on app state, run the production headless bootstrap seam instead of recreating its decisions in a test-only Store. Use the real app-lock Store, render gate, and Home milestones while substituting only lock/keychain and security-chain I/O. The integration runner executes each cold-start path in a separate Jest process so process-wide state cannot leak between valid-session, manual-unlock, failure, and no-account scenarios. A runtime account-add scenario should continue from the real no-account bootstrap state and exercise the same production transition used by the account event.
Run:
yarn workspace rabby-mobile test:integration:boundaries
yarn workspace rabby-mobile test:integration
The boundary check rejects repository-internal Jest mocks and alternate module lifecycles such as jest.resetModules().
Device And E2E Coverage
When Node Jest cannot represent the risk, use the real non-production app boundary. Read ../mobile-testable-component-boundaries/SKILL.md before adding typed actions, lifecycle observations, or deep-link orchestration.
For database changes, run the non-production native in-memory SQLite contract on a physical device before claiming native compatibility. Then exercise the affected file-backed business path when persistence, scheduling, or performance is part of the change. A passing Node SQLite suite is supporting evidence only; it never overrides a device failure.
Do not turn Hermes device integration into coordinate-heavy E2E when a typed action can invoke the same production handler. Keep full E2E for complete journeys, package upgrade behavior, signing, network integration, and cross-Screen outcomes.
Handoff
- State which layer was added and why lower layers were insufficient.
- Report both integration commands and any device/E2E scenarios run.
- If a true native boundary remains mocked, name the residual risk explicitly.