Test the smallest UI contract that proves the behavior. Prefer plain state-driven UI tests with callbacks. Add integration only when lifecycle, navigation, DI, or platform behavior is the thing under test.
Procedure
State the behavior and test concern the task asks you to prove.
Inspect the existing test against that concern and choose the smallest
sufficient seam from the table below.
Keep focused edits within the requested test concern. Do not move test-only
helpers into production or broaden production APIs unless that production
boundary is itself under test.
Drive controlled state or input, synchronize through Compose when needed,
and assert an observable semantic, visual, or callback result.
Finish with no edit when the existing test already uses the narrowest valid
seam and proves the requested behavior.
This avoids constructing ViewModels, components, repositories, navigation, and dependency graphs for layout behavior.
Semantics first
Assert semantics when behavior is semantic:
Text exists: onNodeWithText.
Button is enabled/disabled: assertIsEnabled, assertIsNotEnabled.
Content is selected/focused/toggled: use semantics assertions.
Content is absent: assertDoesNotExist.
Use test tags for nodes that have no stable user-visible text or where multiple nodes share text. Do not use tags as the first choice for all assertions; user-visible semantics are usually stronger.
Callback testing
Use simple counters or captured values:
kotlin
var selectedId: String? = nullcomposeTestRule.setContent { ItemList( items = listOf(ItemUi("movie-1", "Movie")), onItemClick = { selectedId = it }, )}composeTestRule.onNodeWithText("Movie").performClick()assertThat(selectedId).isEqualTo("movie-1")
For plain captured callback values, a direct assertion after the action is usually enough. Use runOnIdle when the assertion needs Compose to finish applying snapshot state, recomposition, or queued UI work before reading the result.
Keep UI tests deterministic
For layout, branch, and callback behavior, render controlled state with setContent instead of constructing the production app graph. Production DI, repositories, lifecycle observers, and background effects add asynchronous work that is irrelevant to a plain UI contract and can make the test flaky.
Do not use Thread.sleep to wait for Compose. Drive the UI to a known state, then use semantic assertions and Compose synchronization (waitForIdle, runOnIdle, or a bounded waitUntil for a real asynchronous condition). Reserve full-app integration for behavior that actually depends on navigation, lifecycle, DI, or platform wiring.
Interaction state with MutableInteractionSource
When a composable's appearance or behavior depends on interaction state (hover, focus, press, drag), inject a MutableInteractionSource and emit the desired state directly. Do not try to simulate pointer/mouse events to trigger interaction states — that approach is fragile, environment-dependent, and produces flaky tests.
kotlin
val interactionSource = MutableInteractionSource()composeTestRule.setContent { OutlinedButton( onClick = {}, interactionSource = interactionSource, )}// Assert default (un-hovered) statecomposeTestRule.onNodeWithText("OutlinedButton").assertIsDisplayed()// Emit hover — interactionSource.emit is a suspend function,// so call it from a test coroutine scope.TestScope().launch { interactionSource.emit(HoverInteraction.Enter())}composeTestRule.waitForIdle()// Assert the visual/semantic change that hover produces// (e.g., border color, elevation, or capture for screenshot test)composeTestRule.onNodeWithText("OutlinedButton").assertIsDisplayed()
The same pattern works for PressInteraction.Press / Release / Cancel, FocusInteraction.Focus / Unfocus, and DragInteraction.Start / Stop / Cancel. Emit the entry interaction, waitForIdle, then assert the result.
Key points:
Always inject MutableInteractionSource rather than relying on the default internal source. This gives you full control over state transitions.
Emit interactions from a coroutine scope (e.g. TestScope().launch { }) since emit is a suspend function. Do not use LaunchedEffect — that is a production Compose effect, not a test tool.
Assert the result of the interaction (visual change, semantic change, enabled state), not the interaction itself. The interaction source is a test driver, not the assertion target.
Use this for screenshot tests too — emit the interaction state, then capture the screenshot for a deterministic hover/press/focus visual.
Keyboard and focus
For keyboard, TV, and desktop UI, drive navigation with the same input model users use (keys/D-pad), not clicks alone. Assert focused semantics, not colors or scale; reserve screenshots for visual focus treatment.
Details—focus graph, FocusRequester, restoration, key handlers, and test patterns: compose-focus-navigation.
Screenshot tests
Use screenshots for visual contracts that semantics cannot prove:
Layout spacing/alignment.
Themed colors, typography, elevation, shadows.
Image composition, gradients, overlays.
Focus highlight appearance.
Loading skeletons or dense visual states.
Keep screenshot state deterministic:
Use fixed state data.
Freeze clocks or animation progress when possible.
Replace network/image loading with fake or preview handlers.
Avoid asserting dynamic text such as current time unless controlled.
Fake images and platform services
When image content is irrelevant, fake the loader and assert the requested model if that is the behavior. The exact hook depends on your image library; a project helper might look like this:
kotlin
val requestedModels = mutableListOf<Any?>()// Example helper, not a Compose API.setContentWithFakeImageLoader { request -> requestedModels += request.data errorPainter()}
When image appearance matters, provide a deterministic local painter/bitmap instead of network data.
Common mistakes
Mistake
Fix
Constructing full app graph to test an error row
Test plain UI with state = Error
Testing click behavior through a ViewModel mock
Pass a callback and assert it was invoked
Screenshot test for simple text presence
Use semantics assertion
Semantics test for padding/color/focus ring
Use screenshot test
Test tags everywhere
Prefer text/content description/role when stable
UI test depends on real image loading/network/time
Fake or freeze the source
Sleeping after an action before asserting UI
Use semantics plus waitForIdle, runOnIdle, or bounded waitUntil
Production DI or app wiring for a state/rendering assertion
Render controlled state with setContent; use integration only when that wiring is under test
Simulating hover/press/focus with mouse or touch events
Inject MutableInteractionSource and emit the interaction
Relying on the default InteractionSource in tests
Pass MutableInteractionSource so you can control state
"This UI test is flaky because images load slowly."
A test uses production DI for simple rendering.
A screenshot has random dates, clocks, remote images, or live data.
Assertions only check that a node exists after performing an action, not that the callback/state change happened.
Focus behavior is visually inspected but not asserted.
A test uses performMouseInput or touch injection to trigger hover/press states instead of MutableInteractionSource.emit.
A composable accepts interactionSource but tests don't inject MutableInteractionSource.
A plain rendering test starts the production app or uses Thread.sleep before asserting.
RED/GREEN agent scenarios
RED launches a production app, waits with Thread.sleep, and only asserts that a node exists. GREEN renders fixed state with setContent, drives the action, synchronizes through Compose, and asserts the semantic state or callback result.
Novel case: a screen's repository-backed state holder starts background work, but the test only needs to prove a disabled Save button. GREEN tests the plain UI state directly; a separate integration test covers the state-holder wiring if needed.
Counterexample: navigation behavior depends on a real NavController lifecycle. GREEN uses an integration test rather than pretending a plain rendering test proves that contract.
Focused counterexample: a value-only formatter test already uses a plain unit
test and the task asks only whether it needs a UI harness. GREEN leaves the
test-local helper and workspace unchanged.