test: split the Android platform test aggregation and share the scripted adb stub

[thymikee]

Summary

Companion to #1102, for the Android offender AGENTS.md names directly ("platform index.test.ts … shrink opportunistically"). The 2,735-line src/platforms/android/__tests__/index.test.ts splits along its already well-factored source modules — every test moved verbatim, 92 tests before and after:

File	Tests	Mirrors
ui-hierarchy.test.ts	22	ui-hierarchy.ts
app-lifecycle-install.test.ts	13	app-lifecycle.ts (install/resolve/infer/component parsing)
app-lifecycle-open.test.ts	19	app-lifecycle.ts (open/close, deep links, launch args, TV, fallback resolve-activity)
input-actions.test.ts	11	input-actions.ts (type/fill/swipe/scroll/rotate)
settings.test.ts	14	settings.ts
notifications.test.ts / app-parsers.test.ts	2 / 1	notifications.ts / app-parsers.ts
appended to existing device-input-state.test.ts	10	device-input-state.ts (keyboard state/dismiss)

adb-mock consistency fix folded in: the monolith carried a local withMockedAdb fork because it needs scripted per-subcommand adb responses, which the shared arg-recorder helper can't express. That fork now lives in src/__tests__/test-utils/mocked-binaries.ts as withScriptedAdb, next to the recorder withMockedAdb, and hands each callback a fresh copy of the shared ANDROID_EMULATOR fixture (which the local fork had duplicated field-for-field).

The fresh copy is load-bearing, and the split caught a real latent hazard: the Android TV test mutated the callback's device (device.target = 'tv'). The old per-call object literal absorbed that silently; with a shared fixture the mutation leaked into the following test and flipped its launch to LEANBACK. The helper now clones per call and the TV test builds { ...device, target: 'tv' } instead of mutating.

Per the discussion: no new e2e scenario — the real-emulator Android workflow (settings replay + snapshot-helper smoke) stays the e2e layer, and these integration tests keep covering install/fill flows against scripted stubs.

Touched files: 10 (1 deleted, 7 created, device-input-state.test.ts extended, mocked-binaries.ts).

Validation

All 92 relocated tests pass in their new homes (including the previously-leaking LEANBACK/fallback pair, which now passes in both orderings), and the full suites are green: 349 test files / 3248 unit tests, smoke 13/13, check:quick (oxlint + tsgo) clean, fallow audit vs origin/main clean. Blocks were moved verbatim except the one withMockedAdb → withScriptedAdb rename and the TV-test mutation fix described above.

🤖 Generated with Claude Code

https://claude.ai/code/session_01FqeW8sA2ZnvnftdvpCqFMS

---

Generated by Claude Code

[github-actions]

Size Report

Metric	Base	Current	Diff
JS raw	1.5 MB	1.5 MB	0 B
JS gzip	491.3 kB	491.3 kB	0 B
npm tarball	590.3 kB	590.3 kB	+9 B
npm unpacked	2.1 MB	2.1 MB	+54 B

Startup median (7 runs, lower is better):

Scenario	Base	Current	Diff
CLI --version	25.9 ms	25.6 ms	-0.3 ms
CLI --help	49.4 ms	49.4 ms	+0.0 ms

Top changed chunks: no changes in the largest emitted chunks.

[thymikee]

Blocking review finding: the split lets the scripted-adb tests run in parallel across files while they still mutate global process state (PATH and AGENT_DEVICE_TEST_ARGS_FILE). That makes the new files interfere with each other when run together.

Repro from the PR branch:

pnpm exec vitest run src/platforms/android/__tests__/app-lifecycle-install.test.ts src/platforms/android/__tests__/app-lifecycle-open.test.ts src/platforms/android/__tests__/app-parsers.test.ts src/platforms/android/__tests__/device-input-state.test.ts src/platforms/android/__tests__/input-actions.test.ts src/platforms/android/__tests__/notifications.test.ts src/platforms/android/__tests__/settings.test.ts src/platforms/android/__tests__/ui-hierarchy.test.ts

That run failed 10 tests with adb exit/timeouts. Running one of the failing files alone passes:

pnpm exec vitest run src/platforms/android/__tests__/app-lifecycle-open.test.ts

So the monolith split changed execution semantics: the old single file effectively serialized the env-mutating adb stubs, but the new file split allows concurrent mutation. Please make the scripted-adb helper/tests safe under Vitest's file parallelism, or explicitly serialize this group if that is the intended contract, then rerun the grouped command above.

[thymikee]

Fixed in 5a5dcd8 — the serialization contract is now explicit, plus the split had silently invalidated the slow-test ratchet pins, which is likely what your 10 failures actually were:

• Five PINNED_SLOW_UNIT_TESTS entries still keyed index.test.ts :: … (the fillAndroid/installAndroidApp real-time offenders), so after the split those tests lost their exemption and fail at 2× budget under parallel load. The pins now point at the split file names (and the reporter's own pinned-key fixture is updated).
• The six env-mutating scripted-adb files (app-lifecycle-install, app-lifecycle-open, device-input-state, input-actions, notifications, settings) now run in a dedicated android-adb vitest project with singleFork, restoring the monolith's serialized execution semantics as an explicit contract; ui-hierarchy and app-parsers are pure parsing and stay in the parallel unit project. test/test:unit run both projects.

Reran your grouped command: 8 files / 94 tests pass, with the six serialized files visibly running under |android-adb| and the two pure ones under |unit|. Full test:unit (349 files / 3248 tests) + smoke + fallow also green.

---

Generated by Claude Code

[github-actions]

PR Preview Action v1.8.1
Preview removed because the pull request was closed.
2026-07-06 09:23 UTC

[thymikee]

Follow-up after the latest fix: the original grouped-command blocker is resolved for me too, but the new serialization contract is currently expressed through a removed Vitest 4 config path.

The grouped run now starts with:

DEPRECATED  `test.poolOptions` was removed in Vitest 4. All previous `poolOptions` are now top-level options.

This matters because the PR depends on serialization to preserve the old monolithic index.test.ts execution semantics, and vitest.config.ts adds that via poolOptions.forks.singleFork. Please switch the android-adb project to the supported Vitest 4 config surface, for example project-level fileParallelism: false with the default forks pool, then rerun the grouped command. After that I think this is ready.

[thymikee]

Follow-up from the latest review: the Vitest 4 serialization issue is only part of the blocker.

• vitest.config.ts: poolOptions.forks.singleFork is ignored on Vitest 4, so the intended android-adb serialization is not enforced. Please move this to supported Vitest 4 project-level config, such as maxWorkers: 1 / fileParallelism: false as appropriate.
• Keep file isolation unless the Android app-resolution cache is explicitly cleared. With serialization plus isolate: false, androidAppResolutionCache in src/platforms/android/app-lifecycle.ts can leak between the split files; shuffled runs can fail depending on cache population/invalidation order.
• The split also means vitest run --project unit skips the Android ADB files with exit 0. Please update docs/scripts or add a guard so maintainers do not accidentally treat that as full unit coverage.

Keeping this as not merge-ready until serialization and isolation/coverage semantics are explicit.

[thymikee]

Addressed the remaining review blockers in e54d617.

Changes:

• Replaced the ignored Vitest 4 poolOptions.forks.singleFork path with project-level fileParallelism: false, maxWorkers: 1, and explicit isolate: true for android-adb.
• Renamed the broad project from unit to unit-core and updated test / test:unit to run unit-core + android-adb, so stale vitest run --project unit now fails instead of silently skipping the Android ADB files.

Validation:

• pnpm exec vitest run src/platforms/android/__tests__/app-lifecycle-install.test.ts src/platforms/android/__tests__/app-lifecycle-open.test.ts src/platforms/android/__tests__/app-parsers.test.ts src/platforms/android/__tests__/device-input-state.test.ts src/platforms/android/__tests__/input-actions.test.ts src/platforms/android/__tests__/notifications.test.ts src/platforms/android/__tests__/settings.test.ts src/platforms/android/__tests__/ui-hierarchy.test.ts passed: 8 files / 94 tests.
• pnpm exec vitest run --project unit now exits with No projects matched the filter "unit".
• pnpm format passed.
• pnpm check:tooling passed.
• pnpm check:unit passed: 349 Vitest files / 3252 tests, then smoke tests passed with the expected live-web skip.

Remaining risk: none beyond normal CI confirmation.

[thymikee]

Follow-up pushed in be8bd77.

Changes:

• Updated AGENTS.md PR readiness guidance to stop naming stale --project unit and instead point agents to pnpm test:unit or direct vitest run --project unit-core --project android-adb.
• No test split or serialization code changed.

Validation:

• git diff --check passed.
• Docs-only guidance change; no test gate rerun needed.

[thymikee]

Review pass after be8bd77: no remaining actionable blockers found.

The Android ADB split now uses supported Vitest 4 serialization (fileParallelism: false, maxWorkers: 1) with explicit file isolation, test / test:unit run both unit-core and android-adb, and stale --project unit usage now fails instead of silently skipping the serialized files. The follow-up AGENTS.md guidance now points agents at pnpm test:unit or the explicit unit-core + android-adb projects.

Validation reported on the PR covers the grouped Android ADB run, stale-project guard, pnpm format, pnpm check:tooling, and pnpm check:unit; CI is green. Marking this ready for human maintainer judgment.