Android Testing

Android Testing

Android-specific testing on a test-first foundation. This reference focuses on the test-first discipline plus the Compose-test and KMP traps that are easy to get wrong, not the basics of the three tiers, fakes over mocks, runTest, or Given-When-Then naming — except where a name breaks one target's compile.

Test-first (the foundation)

• No production code without a failing test first. Write the test, watch it fail (RED), write the minimal code to pass (GREEN), then refactor. A behaviour you never watched fail isn't proven.
• A bug isn't fixed until a test that was red because of the bug is green.

This layers on top of any dedicated TDD discipline skill (superpowers:test-driven-development, ace:test-driven-development) but requires none. For bootstrapping the test stack from scratch (test DI, JUnit/Robolectric/Roborazzi/Paparazzi selection, the instrumented runner, Compose Preview Screenshot Testing, UI Automator, Jacoco), see Google's official testing-setup skill (android skills add testing-setup).

commonTest names must survive Kotlin/Native, not just the JVM

commonTest is compiled once per declared target, so a green jvmTest proves nothing about the iOS compile. Backticks themselves are fine on every target — they are escaping syntax, not part of the name — but Kotlin/Native's frontend rejects 24 characters inside them: . ; , ( ) [ ] { } / < > : \ $ & ~ * ? # | § % @ (FirNativeIdentifierChecker — an unconditional error since Kotlin 1.7; the set also reserves the IR mangler's own characters). Kotlin/JVM forbids only . ; [ ] / < > : \. The comma is the gap between the two lists; a space is in neither, so a Given-When-Then convention with commas fails every native target while CI stays green:

// commonTest — compiles for jvm; fails compileTestKotlinIosSimulatorArm64 with
//   error: name contains illegal characters: ","
@Test fun `given an empty cart, when an item is added, then the total updates`() { … }

// commonTest — same sentence, commas dropped. Backticks and spaces compile everywhere.
@Test fun `given an empty cart when an item is added then the total updates`() { … }

Two rules follow. Keep , . ( ) : ? out of every commonTest declaration name — the check covers classes, properties, type aliases, value parameters and enum entries, not just funs. And verify a KMP test with allTests or <target>Test, never jvmTest, which depends only on compileTestKotlinJvm; a device-only target has no test task at all, so compile its test binary explicitly (linkDebugTestIosArm64).

Don't confuse this with the Android rule Kotlin's coding conventions cite: D8 rejects spaces below minSdk 30, which bites instrumented tests only — and never commas, which the DEX SimpleName grammar has never allowed at any API level.

Compose test v2 APIs run on StandardTestDispatcher

The v2 entry points — androidx.compose.ui.test.junit4.v2.createComposeRule / createAndroidComposeRule and androidx.compose.ui.test.v2.runComposeUiTest — use StandardTestDispatcher (matching kotlinx.coroutines.test.runTest). The same-named v1 functions in androidx.compose.ui.test.junit4 / androidx.compose.ui.test are deprecated and stay on UnconfinedTestDispatcher, so the import decides the behaviour. Use v2 for new tests. When moving a test to v2, a LaunchedEffect that ran eagerly under v1 may need mainClock.advanceTimeBy(0) / runCurrent() to drain queued work. Check which packages the project's Compose version ships before relying on either.

Two-schedulers trap (the one coroutine-test gotcha worth stating): a MainDispatcherRule's TestDispatcher and the dispatcher runTest { } creates have separate TestCoroutineSchedulers. Pass mainRule.dispatcher into runTest(mainRule.dispatcher) so Dispatchers.Main and the test body share one — otherwise advanceUntilIdle() flushes only one and assertions race the ViewModel.

Semantics first, testTag as fallback

Prefer user-visible semantics over testTag — real users and screen readers see semantics; testTag is invisible to everyone except tests. Selector priority: (1) onNodeWithText; (2) onNodeWithContentDescription; (3) role/state matchers (hasClickAction(), isSelected(), isFocused(), isEnabled()); (4) onNodeWithTag only when there's no stable user-visible text or it's duplicated/ambiguous (lists of identical rows, per-locale copy, multiple instances). A text assertion survives refactors and exercises accessibility; a testTag assertion breaks the moment the tag changes and misses the user-facing regression.

(Counterview worth knowing: skydoves/android-testing-skills argues tag-first for i18n robustness, backed by androidx/material3's own 1825 : 424 : 46 testTag : onNodeWithText : onNodeWithContentDescription ratio. Defensible if you have separate accessibility coverage; this skill defaults to semantics-first because it catches a class of bugs testTag never can.)

Callbacks as test surfaces

A composable's contract is "render state, emit callbacks" — test exactly that; don't route the assertion through a ViewModel mock.

@Test fun `tapping article row invokes onArticleClick with id`() {
    var clickedId: String? = null
    composeTestRule.setContent { ArticleRow(Article(id = "42", title = "Hello"), onArticleClick = { clickedId = it }) }
    composeTestRule.onNodeWithText("Hello").performClick()
    assertEquals("42", clickedId)
}

Choosing the test shape — the smallest contract that proves the behaviour

Proving	Shape
Text rendered, conditional content, loading/error branches, callback wiring	Plain UI Compose test (state + callbacks, no graph)
Focus navigation, keyboard, TV/D-pad	Compose test with performKeyInput + assertIsFocused() (see compose/references/focus-navigation.md)
Visual contract semantics can't prove — spacing, themed colour, typography, elevation, gradients, skeletons	Screenshot test, one per meaningful state
State holder updates UI correctly	State-holder unit test + ONE wiring smoke test
Lifecycle, navigation, or DI integration itself	Integration test (createAndroidComposeRule, Hilt rule, real graph)

Test clock vs wall clock: for any Compose-state-observable condition use mainClock.advanceTimeUntil(ms) { state.value == Done } (deterministic, fast); use rule.waitUntil(ms) { … } only for non-Compose conditions (Job.isCompleted, an external counter). Mixing the two in one test is a common flake source. Inside waitUntil { } return a Boolean (onNode(…).isDisplayed()), or use waitUntilExactlyOneExists / waitUntilDoesNotExist — a throwing assertIsDisplayed() fails the test on the first poll instead of waiting.

mainClock.advanceTimeBy(ms) rounds up to a whole frame (16ms) — advanceTimeBy(50) advances 64ms. Pass ignoreFrameDuration = true when an assertion depends on the exact elapsed time.

Animation tests need mainClock.autoAdvance = false set before setContent — otherwise the framework's InfiniteAnimationPolicy throws CancellationException on indeterminate animations, and finite animations finish in one auto-advanced burst with no observable intermediate state. After pausing, drive frames with advanceTimeByFrame() (kick-off) then advanceTimeBy(durationMillis).

Turbine's timeout is wall-clock — the default 3 seconds ignores runTest's virtual time, so advanceTimeBy never trips or avoids it. A slow CI machine can time out a test that virtual time says is instant; pass test(timeout = …) when that bites.

Screenshot determinism: fixed state data (no current time, random seeds, or remote URLs in the screenshot path); frozen clocks (Clock.fixed(...)) and animation progress; a fake image loader for image-heavy screens (Coil 3's test double: LocalImageLoader provides FakeImageLoader(LocalContext.current), or setContentWithFakeImageLoader { … }); one shot per meaningful state (loading / error / success / empty), not one per UI element.