Software Android Native

Native Android Development

Use this skill for native Android work only. It is the default shared-skill entrypoint for Compose-first Android apps targeting API 28+, bounded rewrites from older codebases, and agent-assisted workflows in Android Studio, Codex, and Claude Code.

Quick Reference

When to Use This Skill

Use this skill to:

• Build new Compose-first screens and features for Android apps targeting API 28+
• Plan and execute bounded rewrites from Views or older Kotlin/Java codebases
• Set up agent-assisted Android workflows in Android Studio, Codex, or Claude Code
• Implement Kotlin Coroutines, Flow, and ViewModel state patterns
• Prepare data safety declarations, target SDK compliance, and release gates
• Review native Android code for architecture, performance, and compliance

Defaults

• New native Android work: prefer Jetpack Compose for new screens and Views interop only where existing mature flows or third-party SDKs require it.
• New observable UI state: prefer ViewModel + StateFlow and keep UI-facing state collected on the main thread with collectAsStateWithLifecycle.
• Async work: prefer Kotlin Coroutines with structured concurrency; use Dispatchers.IO for blocking I/O and Dispatchers.Default for CPU-bound work.
• Dependency injection: prefer Hilt for new projects.
• Local persistence: prefer Room with KSP annotation processing.
• Build system: prefer Gradle KTS (build.gradle.kts) with version catalogs (libs.versions.toml).
• Navigation: prefer type-safe Compose Navigation 2.9+ with @Serializable route classes.
• New unit and integration tests: prefer JUnit 5 with Turbine for Flow assertions.
• UI tests: prefer Compose Testing APIs; keep Espresso for legacy Views screens.
• Release gates: treat target SDK compliance, data safety declarations, ProGuard/R8 rules, Play Integrity, accessibility, and real-device verification as non-optional.

ASCII Flow

Android native task
  -> Confirm app shape: Compose, Views interop, service, or release gate
  -> Prove Gradle, emulator/device, install, and launch reality
  -> Choose architecture: ViewModel, StateFlow, Hilt, Room, Navigation
  -> Implement bounded slice with lifecycle-aware state and tests
  -> Check Kotlin, Compose, R8, billing, and Play-policy traps
  -> Build, install, launch, inspect logs, and report proof

Known Kotlin Traps

These are current headline footguns for Compose-first native Android on Kotlin 2.x. Each is source-backed; re-verify versions against the linked release notes before quoting a fix window.

• CalledFromWrongThreadException / ConcurrentModificationException in SnapshotStateObserver. Compose reads snapshot state on the main thread; mutating _uiState.value from a coroutine body that ran on Dispatchers.IO without switching back surfaces as a main-thread crash or a reentrant-modification race. Fix: keep withContext(Dispatchers.IO) { ... } blocks pure (return a value, do not mutate state inside), assign _uiState.value = ... on the main dispatcher, and collect via collectAsStateWithLifecycle() — never a raw collect inside LaunchedEffect. See references/compose-state-concurrency.md.
• Strong Skipping Mode identity checks. On Kotlin 2.x + Compose 1.8+, Strong Skipping Mode compares unstable params by instance reference. Emitting UI state as a fresh data class per event (_state.value = _state.value.copy(...)) defeats skipping even though the observable values are unchanged — LazyColumn rows recompose on every unrelated update. Fix: hoist derived lists and filtered views to the ViewModel with stateIn(scope, WhileSubscribed(5000), ...); split UI state into @Immutable slices; pass PersistentList<T> from kotlinx-collections-immutable. Source: developer.android.com/develop/ui/compose/performance/stability/strongskipping.
• LazyListScope lambda memoization gap. Strong Skipping's automatic lambda memoization applies inside @Composable functions; it does not apply inside LazyListScope.items { ... } or LazyColumn's content block. Unremembered callbacks captured there produce a new instance per recomposition. Fix: val { { vm.onClick(id) } } at the items-block site, or hoist to a stable () -> Unit reference.
• Compose plugin version skew on Kotlin 2.x. Since Kotlin 2.0 the Compose compiler ships with the Kotlin compiler and is applied via the Gradle plugin kotlin("plugin.compose"). A stale or missing plugin declaration surfaces as Argument type mismatch: actual type 'Function0<Unit>', but '@Composable ComposableFunction0<Unit>' was expected — the transform did not run. Fix: lock plugin.compose to the exact Kotlin version in libs.versions.toml. Source: developer.android.com/jetpack/androidx/releases/compose-kotlin.
• Compose runtime regressions fixed upstream. If you see a crash in pausable composition under LookaheadScope, nested Popup positioning against the screen instead of the parent, or a reentrant-modification crash in SnapshotStateObserver, upgrade to the latest Compose UI patch release before treating the problem as app-level — as of 2026-07-11 the production Compose BOM is in the 2026.06.xx line (Compose runtime/UI ~1.11.4, with 1.12 in beta requiring compileSdk 37 + AGP 9), and the SnapshotStateObserver reentrant-modification guard landed in the 1.10.0-rc01 cycle; a project still pinned below that line should treat this class of crash as a known-fixed upgrade target, not a fresh bug. Verify current at developer.android.com/jetpack/androidx/releases/compose-ui. Route to ../software-android-runtime-debugging/references/compose-debugging.md.
• kotlinx-serialization + R8 full mode. Since kotlinx-serialization 1.9.0, AGP 8.x release builds can fail with SerializationException: Serializer for class 'X' is not found or ExceptionInInitializerError, and the build emits warnings about <1>$* keep rules. Fix: keep the generated $serializer classes explicitly and run a release-variant smoke test in CI that exercises every @Serializable entry point. Route to ../software-android-runtime-debugging/references/proguard-r8-triage.md.

Kotlin Anti-Patterns

These are behaviors to actively refuse in new code; they are not compile errors but each has caused measurable harm in production Kotlin/Android codebases.

Route deeper pattern material through references/compose-state-concurrency.md.

Architecture Judgment Calls

Decisions that need a rationale, not just a default pick. As of 2026-07-11, verify each version-specific claim at the linked source before quoting it.

• Compose vs Views in 2026. Compose is the default for all new screens; there is no scenario in a greenfield API 28+ app where Views is the right starting point. Keep Views only for: (1) a third-party SDK that ships a View-based render surface with no Compose wrapper (some map, ad, or video SDKs), (2) a legacy screen mid-migration where the cost of a full rewrite outweighs the interop tax, or (3) SurfaceView/TextureView-backed continuous rendering (camera preview, custom video) where Compose's AndroidView bridge is the right embedding, not a reason to avoid Compose for the rest of the screen. Do not accept "Views is faster" as a reason in 2026 — Compose's skip/restart model with Strong Skipping enabled is on par with or ahead of View-based RecyclerView diffing for list-heavy UI when state discipline (K6, K7) is followed.
• Hilt vs Koin. Hilt (compile-time, annotation-processor-based, built on Dagger) remains the repo default for new API 28+ apps: it fails at compile time on a broken graph, has first-class @HiltViewModel / WorkManager / Compose Navigation integration, and is what most enterprise Android codebases already standardize on. Prefer Koin instead only when: the team explicitly wants to avoid annotation processing and Gradle plugin overhead (KSP-free build), the project is small enough that compile-time graph validation matters less than iteration speed, or the codebase is a Kotlin Multiplatform module where Hilt cannot run (Hilt is Android/JVM-only; Koin runs on all KMP targets). Do not switch an existing Hilt codebase to Koin mid-project without a concrete, named pain point — DI framework churn has a high cost for a marginal ergonomics gain.
• Kotlin Multiplatform (KMP). KMP has been stable since November 2023 and Compose Multiplatform for iOS reached stable with 1.8.0; Jetpack libraries including Room, DataStore, and ViewModel now ship commonMain artifacts. This is a real option for sharing business logic (networking, persistence, ViewModel state) across Android and iOS — but it is a product/architecture decision, not a default for this skill. If the task is "should we share code with iOS," route to software-mobile (or the software-mobile-architect advisor, where available) for the cross-platform tradeoff call before writing shared-module code; this skill assumes the Android-native side once that call is made. Verify current KMP/Compose Multiplatform stability status at kotlinlang.org/docs/multiplatform/supported-platforms.html.
• When NOT to go native. If the actual question is "should this feature be a native Android screen at all" (vs. a cross-platform framework, a web view, or a KMP-shared module), that decision belongs to software-mobile (or the software-mobile-architect advisor) — do not let this skill's Compose-first defaults silently answer a platform-choice question it was not asked.
• Coroutines vs Flow failure modes. A suspend fun returns one value and is the wrong tool for anything that emits more than once (search-as-you-type, connectivity state, DB observation) — using a polling suspend loop instead of Flow produces stale reads and duplicate work. Conversely, wrapping a one-shot operation (a single network POST) in a Flow that a caller collects once adds Flow's cancellation/backpressure machinery for no benefit — a plain suspend fun is simpler and equally cancellable via structured concurrency. Rule of thumb: one value now -> suspend fun; zero-to-many values over time -> Flow; a single ViewModel-to-UI event stream that should not replay -> SharedFlow with replay = 0, not StateFlow.
• Process death and SavedStateHandle. ViewModel survives configuration change but not process death under memory pressure. Anything the user would be upset to lose on a background-kill-and-restore (form input mid-fill, scroll position, in-progress multi-step flow state) must go through SavedStateHandle (@HiltViewModel constructor-injects it automatically), not just ViewModel field state. Test this with adb shell am kill <package> while backgrounded, not just rotation — rotation alone never exercises the process-death path and gives false confidence.

ANR and Frame Budget Arithmetic

Use these thresholds when diagnosing jank or ANR reports; re-derive the math rather than quoting a remembered number.

• Input dispatch ANR: 5 seconds. If the main thread does not return from handling a touch or key event within 5s, the system raises Input dispatching timed out. This is the most common production ANR class and is almost always a synchronous DB/network/disk call on the main thread, not an actual 5-second-long computation.
• Foreground service start ANR: 5 seconds. startForegroundService() must reach startForeground() within 5s or the system raises ForegroundServiceDidNotStartInTimeException territory (see S3 below for the related ForegroundServiceStartNotAllowedException/MissingForegroundServiceTypeException cases).
• Broadcast receiver ANR: 10s foreground / 60s background. onReceive() running past this window on the relevant app-state timer raises an ANR; move any real work off onReceive() into WorkManager or a coroutine launched from a longer-lived scope.
• Service execution ANR: 20s foreground / 200s background. onCreate()/onStartCommand()/onBind() blocking past this window raises an ANR; verify current thresholds at developer.android.com/topic/performance/anrs/diagnose-and-fix-anrs since these have shifted across Android versions.
• Frame budget at 60Hz: 1000ms / 60 = 16.666...ms per frame. A composable recomposition, layout, and draw pass that together exceed ~16.67ms drops a frame; two consecutive misses read as visible jank.
• Frame budget at 120Hz: 1000ms / 120 = 8.333...ms per frame. High-refresh-rate devices (now the majority of mid-range and flagship Android hardware) halve the available budget — a composable that was "fine" at 60Hz can visibly jank at 120Hz. Do not assume a perf fix validated at 60Hz on an emulator holds on a 120Hz physical device; re-verify with adb shell dumpsys gfxinfo <package> framestats or Android Studio's Frame Profiler on the actual refresh rate.

Runtime Truth And Prompting

Use a proof-first execution loop for native Android work:

• verify tool reality first: Android Studio Gemini, Gradle CLI, ADB, and emulator or device availability
• prove build, install, and launch before UI diagnosis
• keep repo memory lean and fact-only
• require bounded slices with explicit proof artifacts

Load references/runtime-proof-and-prompts.md for:

• AI-agent defaults
• proof-first and token-discipline rules
• the execution loop
• high-value prompt shape

Rewrite Workflow

1. Lock the baseline: existing app behavior, minimum API level, device classes, external integrations, and non-goals.
2. Choose the target defaults: Compose-first, API 28+, ViewModel + StateFlow, Hilt, Room + KSP, Kotlin Coroutines, JUnit 5, Compose Testing.
3. Slice the rewrite into bounded vertical features: app shell (Application class, Hilt setup, navigation graph), auth/session, core navigation, feature flows, integrations, release surfaces.
   • If the project uses multi-module Gradle, verify module dependencies and build order before adding new modules.
   • If migrating from Java to Kotlin, convert one file at a time using Android Studio's converter, then review and fix idiom issues. Do not bulk-convert entire packages without validation.
   • If migrating from Views to Compose, use ComposeView in existing XML layouts as a bridge. Do not rewrite an entire Activity/Fragment hierarchy in one pass.
4. For each slice, require evidence: build success, install and launch success, targeted tests, parity notes, and known gaps.
5. Keep release-only concerns visible throughout: data safety declarations, target SDK compliance, ProGuard/R8 rules, Play Integrity, push/deep-link behavior, store metadata.
6. End every batch with a handoff: changed behavior, validation performed, residual risk, next slice.
7. When a backend change eliminates an error class (e.g., unifying two API paths into one), immediately remove the now-impossible error types, decoders, and UI states from the Android client. Dead error handling misleads future developers about what can actually happen and inflates the codebase.

Specialized Patterns

Load references/ui-and-integration-patterns.md when the work involves:

• Compose, adaptive-layout, or foldable-specific implementation patterns
• native/backend integration gotchas for auth, onboarding, or Supabase or Firebase-backed flows
• Google Play Billing rules and RTDN constraints

When NOT to Use This Skill

Use a different skill when:

• Cross-platform or platform-choice decisions -> software-mobile
• Android test execution, device matrix, Espresso deep dives -> qa-testing-android
• Web UI or browser app implementation -> software-frontend
• General architecture without Android-specific constraints -> software-architecture-design
• Backend platform selection (Supabase, Firebase, Convex) -> software-baas-platforms

Scenarios

Recipes keyed to symptoms or migration moments. Each lists the shortest path to resolution using patterns above.

S1 — Compose recomposition perf bug after Strong Skipping upgrade

1. Enable Compose compiler metrics: add freeCompilerArgs += ["-P", "plugin:androidx.compose.compiler.plugins.kotlin:reportsDestination=..."].
2. Identify composables marked unstable in the report; focus on those receiving the full UI state object.
3. Split the monolithic state into @Immutable slices and hoist derived lists to StateFlow in ViewModel.
4. Replace List<T> params with ImmutableList<T> from kotlinx-collections-immutable.
5. Re-run the metrics; verify the hot composables are now marked skippable.
6. Confirm the fix with a baseline profile trace before and after on a real device.

S2 — R8 stripping kotlinx-serialization classes

1. Reproduce in a release build: run ./gradlew :app:assembleRelease and trigger the failing serialization path.
2. Check the R8 mapping file and logcat for SerializationException: Serializer for class 'X' is not found.
3. Add explicit keep rules for every @Serializable class and its generated $serializer companion; see references/android-release-and-compliance.md.
4. Add a release-variant smoke test in CI that exercises every serialized entry point.
5. Verify the fixed APK with adb install -r and re-run the failing path end-to-end.

S3 — Foreground service crash on API 35

1. Check the crash log for ForegroundServiceStartNotAllowedException or MissingForegroundServiceTypeException.
2. Declare android:foregroundServiceType in the <service> manifest element (e.g. dataSync, mediaPlayback).
3. Pass the matching ServiceInfo.FOREGROUND_SERVICE_TYPE_* flag to startForeground().
4. Add a runtime check: if the app is in the background, use WorkManager instead of starting a foreground service.
5. Test on an API 35 emulator with adb shell am kill to simulate background-state entry before the start call.

S4 — Predictive back gesture migration

1. Set android:enableOnBackInvokedCallback="true" in the manifest <application> element.
2. Replace all onBackPressed() overrides with OnBackPressedCallback registered on onBackPressedDispatcher.
3. For Compose Navigation, confirm NavHost handles OnBackPressedCallback automatically; add explicit callbacks only for custom back logic.
4. Test with gesture navigation enabled on API 33+ device; verify animated back preview renders correctly.
5. Remove any legacy KeyEvent.KEYCODE_BACK handlers that now conflict with the new callback.

S5 — In-app billing v7 entitlement reconciliation

1. Initialize BillingClient as a Hilt singleton; connect in Application.onCreate.
2. On BillingClient.BillingResponseCode.OK after purchase, call the backend to verify the purchase token server-side before granting access.
3. Call acknowledgePurchase() within 3 days; unacknowledged purchases auto-refund.
4. Subscribe to PurchasesUpdatedListener and queryPurchasesAsync(QueryPurchasesParams) on app foreground to catch out-of-band purchases.
5. Handle ITEM_ALREADY_OWNED gracefully by querying existing entitlements rather than surfacing an error.

Navigation

References

Templates

Use the rewrite brief at project start, the feature request per slice, the proof checklist at each verification gate, and the agent handoff at batch boundaries.

Related Skills

Freshness Protocol

Trigger Conditions

• "Latest Android Studio / Compose / Kotlin changes?"
• "Gradle or AGP setup?"
• "Android data safety declaration requirements?"
• "Play Store submission requirements?"
• "Is [Android framework/tool] still the default?"
• "Target SDK deadline?"

How to Freshness-Check

1. Start from data/sources.json for Google/Android docs and release notes.
2. Run a targeted web search for the specific Android Studio, Compose, or Kotlin question.
3. Prefer developer.android.com, kotlinlang.org, and official Android release notes.

What to Report

• Current landscape: stable Jetpack libraries and tooling
• Emerging tools: new Android Studio features, Kotlin evolution, Compose updates
• Deprecated/declining: APIs being sunset, tools losing support
• Recommendation: default choice with rationale

Fact-Checking

• Known bugs, regressions, framework/compiler/runtime footguns, and version-specific crash or workaround guidance must be verified against current primary web sources before being treated as current fact.
• Prefer developer.android.com, kotlinlang.org, and Android release notes for Compose, Kotlin, Hilt, Room, Navigation, WorkManager, Play Billing, and store requirements.
• Prefer official OpenAI and Anthropic docs for Codex / Claude Code memory and workflow behavior.
• Prefer the official Gradle and AGP documentation for build system behavior.
• If a claim is not source-backed or clearly labeled as a repo default, remove it.

Learnings Loop

Before applying this skill on a non-trivial task, read learnings.consolidated.md in this directory (and learnings.md if present).

After applying it, if you encountered a pattern worth remembering, a mistake worth preventing, or a domain fact that surprised you, append one dated bullet to learnings.md via agents-skills-feedback-loop/scripts/append_learning.py. Do not modify SKILL.md itself.