A Kotlin Multiplatform port of AndroidX App Startup.

On Android it is androidx.startup. Initializer, AppInitializer, Context and the class token are typealiases of the AndroidX types, so an initializer written once in commonMain compiles to implements androidx.startup.Initializer, is discovered by the same InitializationProvider, and is instantiated by the same reflection. There is no wrapper type and no adapter anywhere in the Android path.

On the other ten targets the library ships its own runtime. It computes the initialization order with Kahn's algorithm, executes it sequentially on the calling thread, and reports cycles as a trimmed path rather than as a bare "cycle detected".

android, desktop (JVM), iosArm64, iosSimulatorArm64, iosX64, macosArm64, macosX64, linuxX64, mingwX64, js (browser and Node), wasmJs (browser and Node).

macosX64 and iosX64 are compiled and linked but never run: they are disabled on an arm64 Mac and CI has no x86_64 macOS runner, so a green build is not evidence that their test binaries executed.

kotlin {
    sourceSets {
        commonMain.dependencies {
            implementation("io.github.kunal26das:startup:3.0.0")
        }
    }
}

Use api rather than implementation if an iOS host has to see these types, because a framework can only export a dependency its source set declares with api; see From Swift and Objective-C below.

The Android artifact depends on androidx.startup:startup-runtime with api scope, so an Android consumer can implement Initializer without declaring AndroidX itself, and it publishes the same two Android floors that dependency does: minSdk 21 and minCompileSdk 34. Adopting this library never narrows the device range or forces a compileSdk move relative to plain androidx.startup; :startup:checkAndroidFloors fails the build if either floor rises.

Two requirements come from the way the artifacts are compiled:

• Kotlin 2.4 or newer. The library ships metadata version 2.4.0. AGP 9's bundled compiler is older and reports Class ... was compiled with an incompatible version of Kotlin, so an Android consumer has to apply the Kotlin Multiplatform or Kotlin plugin at 2.4.x rather than rely on it.
• JVM target 11 or newer. Every registration function is inline, and Kotlin refuses to inline bytecode built for a newer JVM target than the caller's.

One more requirement comes from the consumer's own code rather than from the artifacts. An application that writes an expect class initializer, which is the shape below for anything touching a platform SDK, needs

kotlin {
    compilerOptions { freeCompilerArgs.add("-Xexpect-actual-classes") }
}

Kotlin still reports expect/actual classes as Beta, so without that flag every such initializer emits a BETA_EXPECT_ACTUAL_CLASSES warning. It is a warning, not an error, and a consumer whose initializers all live in commonMain and extend BaseInitializer needs no flag at all.

class LoggerInitializer : BaseInitializer<Logger>() {
    override fun create(context: StartupContext): Logger = Logger()
}

class NetworkInitializer : Initializer<Network> {
    override fun create(context: StartupContext): Network {
        val logger = Startup.getInstance(context)
            .initializeComponent(initializerKey<LoggerInitializer>())
        return Network(logger)
    }

    override fun dependencies(): List<AnyInitializerKey> =
        listOf(initializerKey<LoggerInitializer>())
}

BaseInitializer exists because neither half of an expect/actual pair may give a member a default body. Extend it whenever an initializer has no dependencies and the dependencies() override disappears.

Use StartupContext rather than Context in shared code. Both name the same type, but Context collides with android.content.Context in any file that imports both.

Starting Crashlytics, a Cocoa reporter and a browser reporter are different calls, so declare the initializer expect and give each platform its own actual. The class keeps one fully qualified name on every target, so a single AndroidManifest <meta-data> entry addresses it whichever actual is compiled in, and it registers in the same manifest as a shared initializer.

expect class CrashReportingInitializer() : Initializer<CrashReporting> {
    override fun create(context: StartupContext): CrashReporting
    override fun dependencies(): List<AnyInitializerKey>
}

actual class CrashReportingInitializer actual constructor() : Initializer<CrashReporting> {
    actual override fun create(context: StartupContext): CrashReporting {
        val logger = Startup.getInstance(context)
            .initializeComponent(initializerKey<LoggerInitializer>())
        return CrashReporting("AndroidCrashReporter(${context.packageName})", logger)
    }

    actual override fun dependencies(): List<AnyInitializerKey> =
        listOf(initializerKey<LoggerInitializer>())
}

The expect class has to redeclare every member it does not inherit a body for. Initializer carries both of its members abstract, so a subclass that declares none is abstract; an expect that omits them fails with "has no corresponding expected declaration" on the actual side.

Extend BaseInitializer and only create has to be redeclared. BaseInitializer supplies a concrete dependencies(), so the expect names one member instead of two. This is the shape to reach for unless the initializer really does declare dependencies:

expect class CrashReportingInitializer() : BaseInitializer<CrashReporting> {
    override fun create(context: StartupContext): CrashReporting
}

actual class CrashReportingInitializer actual constructor() : BaseInitializer<CrashReporting>() {
    actual override fun create(context: StartupContext): CrashReporting =
        CrashReporting("AndroidCrashReporter(${context.packageName})")
}

Redeclaring create on the expect is what makes it an expected member, so the actual spells its override actual override. Leave the actual off and the compiler answers Declaration must be marked with 'actual'.

There is a shorter form, and it is not portable. Drop the body from the expect entirely and create becomes an inherited abstract member rather than an expected one, so the actual overrides it with a plain override and the expect never names it at all:

expect class CrashReportingInitializer() : BaseInitializer<CrashReporting>

That compiles, links and runs on all eleven platform compilations. A metadata compilation rejects it:

e: CrashReportingInitializer.kt:5:8 Class 'CrashReportingInitializer' is not abstract and does not
implement abstract member:
expect fun create(context: Context): T

The task is compileCommonMainKotlinMetadata, and every KMP module with a shared commonMain has one: anything published, anything running KSP in commonMain, anything relying on cinterop commonization. ./gradlew build runs it; compileKotlin<Target>, compileAndroidMain and link<Target> do not, so this shape can pass a target-by-target verification on all eleven and still fail the next build. It is not a BaseInitializer quirk either — a plain abstract class with an unimplemented abstract member behaves identically under an expect class, so nothing this library could ship would change it. Reach for it only in a module you know has no metadata compilation, such as a test source set.

Either shape compiles to implements androidx.startup.Initializer with a public no-argument constructor and AndroidX's own dependencies() signature, so reflection is unaffected. PlatformInitializer and MemberlessInitializer in the library's own commonTest pin both across the eleven platform compilations and AndroidInitializerContractTest pins their bytecode; sample's RuntimeInfoInitializer is what keeps this section honest, because it lives in a commonMain that really is metadata-compiled. Delete its override fun create line and ./gradlew build goes red.

Every actual must be a class with a public no-argument constructor. AndroidX ignores the factory in the manifest object and reflects with getDeclaredConstructor().newInstance(), so an actual object compiles everywhere and throws only on Android, at process start. sample's CrashReportingContractTest pins both halves of that contract with reflection.

Only the part that varies has to be expect. Where the initializer itself is shared and just needs a per-platform value, keep one class in commonMain and put the seam behind a small expect — the sample resolves its SDK name through expect object Platform, with actuals in appleMain, desktopMain, jsMain, wasmJsMain, linuxMain and mingwMain. An expect may be declared in an intermediate source set, so Platform lives in nonAndroidMain and Android never sees it.

val manifest = StartupManifest {
    metaData<AnalyticsInitializer> { AnalyticsInitializer() }
    lazyInitializer<NetworkInitializer> { NetworkInitializer() }
    lazyInitializer<LoggerInitializer> { LoggerInitializer() }
}

metaData registers a component that is initialized eagerly at startup, the equivalent of a <meta-data android:value="androidx.startup" /> entry. lazyInitializer registers one that is created only when something asks for it. remove<T>() is the equivalent of tools:node="remove" and hides an entry an included manifest contributed.

Manifests compose, later entries winning, so a library can ship one that an application overrides:

val applicationManifest = StartupManifest {
    include(libraryManifest)
    remove<LibraryDebugInitializer>()
    metaData<ApplicationInitializer> { ApplicationInitializer() }
}

On Android remove<T>() only suppresses what Startup.install would otherwise start. A component a library contributed through its own AndroidManifest is created by InitializationProvider before any application code runs, so nothing here can reach it; suppressing that needs a real tools:node="remove" entry in the application's AndroidManifest, written by hand.

metaData<T>, lazyInitializer<T> and remove<T>() name the component at compile time. Each also has an overload that takes the key instead, for an initializer the compiler cannot name: one a host application constructed and handed to Kotlin, one discovered from a plugin, or one written in Swift. initializerKey(initializer) builds the key from the instance, and it is the same key initializerKey<T>() would have reified:

fun manifest(supplied: List<Initializer<*>>): StartupManifest = StartupManifest {
    metaData<AnalyticsInitializer> { AnalyticsInitializer() }
    for (initializer in supplied) metaData(initializerKey(initializer)) { initializer }
}

Registered that way, a host-supplied initializer is an ordinary node: it is ordered behind whatever it declares in dependencies(), other components may depend on it, it is created once, and a cycle or a missing registration around it is diagnosed like any other. Running such initializers outside the graph, before Startup.install, gives up all four.

Startup.install(context, manifest)

val analytics = Startup.getInstance(context)
    .initializeComponent(initializerKey<AnalyticsInitializer>())

On Android the AndroidManifest is the primary way to start components: InitializationProvider reads it before any application code runs, exactly as it does for a pure AndroidX app. Declare the eager components there and nothing else is needed at boot. sample/src/androidMain/AndroidManifest.xml is a worked example, and AndroidManifestParityTest fails the build if it drifts from SampleStartup.manifest.

Startup.install remains available as the programmatic equivalent, and is the only route on the other ten targets.

Each eager component is one <meta-data> line inside the provider block. Write those lines by hand, exactly as a plain androidx.startup application does:

<manifest xmlns:android="http://schemas.android.com/apk/res/android"
    xmlns:tools="http://schemas.android.com/tools">
    <application>
        <provider
            android:name="androidx.startup.InitializationProvider"
            android:authorities="${applicationId}.androidx-startup"
            android:exported="false"
            tools:node="merge">
            <meta-data
                android:name="com.example.AnalyticsInitializer"
                android:value="androidx.startup" />
        </provider>
    </application>
</manifest>

InitializationProvider resolves each name with Class.forName, so every one has to be fully qualified. The xmlns:tools declaration is required as soon as a tools:node="remove" entry is in the block.

AndroidX answers isEagerlyInitialized from what InitializationProvider discovered in the AndroidManifest, so a component started only by Startup.install still reports false there. Declaring the component in the manifest is what makes the two agree.

This is the one failure this library can produce that plain androidx.startup cannot, and it is worth naming. On Android the StartupManifest factories are never called: InitializationProvider reads the AndroidManifest and nothing else. So a component registered with metaData<T> and left out of the XML runs correctly on ten targets and, in an app that relies on InitializationProvider alone, silently never runs on the eleventh. Startup.install does start it on Android, but isEagerlyInitialized still reports false for it. There is no exception, no log and no lint check, because from AndroidX's point of view nothing is wrong.

Until 1.1.0 the library answered that itself, with verifyAndroidManifest(context), androidManifestDrift(context) and androidManifestDrift(declared). 2.0.0 removes all three: androidx.startup has no counterpart for any of them, and mirroring androidx.startup is this library's whole contract. The problem has not gone away with them, so read the rest of this section rather than treating the removal as a fix. The answer from 2.0.0 is the one a plain androidx.startup application already uses, and it is two things.

The AndroidManifest is the source of truth on Android. Write its <meta-data> entries by hand, declare there exactly what should start eagerly, and treat the StartupManifest as the registry for the other ten targets. Nothing in the library reconciles the two for you.

Keep a parity test of your own. It is a dozen lines against the API that is staying, and it goes in an Android source set because that is where a key is a java.lang.Class and can name its component the way Class.forName needs it named:

class AndroidManifestParityTest {

    @Test
    fun theManifestDeclaresExactlyTheEagerComponents() {
        val xml = File(System.getProperty("myapp.androidManifest")!!).readText()
        val declared = Regex("<meta-data[^>]*?android:name=\"([^\"]+)\"", RegexOption.DOT_MATCHES_ALL)
            .findAll(xml)
            .map { it.groupValues[1] }
            .toSet()
        assertEquals(manifest.eagerComponents.map { it.name }.toSet(), declared)
    }
}

Both directions matter. A component eager in the StartupManifest and absent from the XML never runs on Android; one the XML declares while the StartupManifest keeps it lazy or removed runs eagerly on Android alone. A name the XML declares that the StartupManifest has never heard of is your call rather than the library's: a mixed application is free to declare initializers written directly against androidx.startup beside these, and a test that reports those is useless in exactly the applications that need one most. The set comparison above treats such a name as a failure, so subtract them, or compare only the names the StartupManifest knows.

Put the XML on the test task's inputs, or the check stops running. A test that reads a file Gradle does not know about stays UP-TO-DATE when only that file changes, so deleting a <meta-data> line leaves the build green and the drift undetected — which is the very failure the check exists to catch:

tasks.withType<Test>().configureEach {
    val manifest = layout.projectDirectory.file("src/androidMain/AndroidManifest.xml")
    inputs.file(manifest)
        .withPropertyName("androidMainManifest")
        .withPathSensitivity(PathSensitivity.RELATIVE)
    systemProperty("myapp.androidManifest", manifest.asFile.absolutePath)
}

sample's own AndroidManifestParityTest is exactly the test above, run against a real AndroidManifest.xml and a real StartupManifest on every build. Copy it. Its negative control is the whole point: delete a <meta-data> line from sample/src/androidMain/AndroidManifest.xml and ./gradlew :sample:testAndroidHostTest fails, naming the component that would have stopped running on Android alone.

Android-only source sets can keep using the verbatim AndroidX spelling against the very same components:

androidx.startup.AppInitializer.getInstance(context)
    .initializeComponent(AnalyticsInitializer::class.java)

There is no manifest merger and no ContentProvider, so Startup.install is the only registration step. Pass DefaultContext when your initializers need nothing from the platform:

Startup.install(DefaultContext, manifest)

Initialization runs sequentially on the calling thread, and every entry point is serialized behind one reentrant lock, which is what AndroidX gets from synchronized (sLock): a component is created exactly once however many threads ask for it, and an Initializer.create may call back into initializeComponent without deadlocking. The engine holds no coroutine of its own, because runBlocking does not exist on Kotlin/JS or Kotlin/Wasm; a component whose own work suspends says so with CoroutineInitializer, see Initializing something that suspends below. StartupPlan.waves exposes the Kahn levels as data, and Startup.install(context, manifest, runner) hands each level to a WaveRunner of your choosing — see Running a wave concurrently below for what a task may not do.

AppInitializer.initializeComponentOrNull(key) is the read for a component whose key is an AnyInitializerKey — the element type of dependencies(), what initializerKey(initializer) returns, and the only key a host that discovered an initializer at run time can build — and for one whose create returned null. initializeComponent takes neither: its key is InitializerKey<out Initializer<T>> and its T is bound to Any.

AppInitializer is the same two members here that it is on Android. Until 1.1.0 it carried three more off Android — isInitialized(component), initializationOrder() and manifest() — and 2.0.0 removes them. On Android AppInitializer is androidx.startup.AppInitializer, which exposes neither the order it created things in nor whether a given component exists, and keeps no accessible state to derive either from, so androidx.startup had no counterpart to mirror and those three were the only members of the API that ten targets had and the eleventh did not.

Record what you need from inside your own create, which is what sample's SampleReport does:

class NetworkInitializer : Initializer<Network> {
    override fun create(context: StartupContext): Network {
        val logger = Startup.getInstance(context)
            .initializeComponent(initializerKey<LoggerInitializer>())
        return Network(logger).also { logger.ready("network") }
    }

    override fun dependencies(): List<AnyInitializerKey> = listOf(initializerKey<LoggerInitializer>())
}

That reads the same on all eleven targets, which the removed members never could. For the manifest, hold on to the StartupManifest you passed to Startup.install: it is an ordinary value, and components, eagerComponents, isEager and in all still answer from it. See Upgrading from 1.x below.

Export the library from your framework first. implementation("io.github.kunal26das:startup") is enough for Kotlin and is not enough for Swift. A dependency module that a framework does not export has its module name mangled into every class the header does carry, and the declarations that appear in no exported signature are dropped from it altogether. Add both lines below, or the Swift snippets in this section name types that do not exist:

kotlin {
    sourceSets.commonMain.dependencies {
        api("io.github.kunal26das:startup:3.0.0")
    }
    listOf(iosArm64(), iosSimulatorArm64(), iosX64()).forEach {
        it.binaries.framework {
            baseName = "Shared"
            export("io.github.kunal26das:startup:3.0.0")
        }
    }
}

A build whose targets are declared elsewhere, in a convention plugin, reaches the same frameworks without naming them:

import org.jetbrains.kotlin.gradle.plugin.mpp.Framework
import org.jetbrains.kotlin.gradle.plugin.mpp.KotlinNativeTarget

kotlin {
    targets.withType<KotlinNativeTarget>().configureEach {
        binaries.withType<Framework>().configureEach {
            export("io.github.kunal26das:startup:3.0.0")
        }
    }
}

export requires api. Left on implementation, the build fails with dependencies exported in the framework are not specified as API dependencies of a corresponding source set.

You may already have it without writing the line. A framework with transitiveExport = true that exports a module which declares api("io.github.kunal26das:startup:3.0.0") exports this library too, which is what a convention plugin that exports a shared core module typically produces. Check the generated header for swift_name("InitializerKeyKt") before adding anything: if it is there, the export is already in place.

This is what a consumer's own framework header contains either way. It is measured rather than predicted: :sample:checkConsumerObjCExport links one framework each way and greps both.

The last two rows are the ones that bite. A top-level function facade appears in no exported signature, so InitializerKeyKt never reaches the header at all: a Swift class can conform to StartupInitializer and can never build a key, which leaves its dependencies() able to return only []. Swift reports cannot find type 'InitializerKey' in scope and says nothing about export.

With the export in place, Startup, StartupManifest, StartupManifestBuilder, AppInitializer, Initializer, BaseInitializer, InitializerKey, InitializerKeyKt, StartupContext and DefaultContext all arrive under those names, so an iOS host can build a manifest and implement an initializer in Swift. Two things follow from the fact that a reified type argument cannot cross that boundary:

• The reified registration functions are hidden from the header. metaData<T>, lazyInitializer<T>, remove<T>() and initializerKey<T>() carry @HiddenFromObjC, because Kotlin/Native would otherwise export their non-inline bodies with T collapsed to Initializer<*> — four Swift-callable methods that compile, run, and register every call site under one key.
• Swift uses the key overloads instead. metaData(component:factory:), lazyInitializer(component:factory:), remove(component:) and initializerKey(initializer:) are exported and do what they say:

func boot(_ application: UIApplication) throws {
    let manifest = StartupManifest.companion.invoke { builder in
        let lifecycle = ViewControllerLifecycleInitializer()
        builder.metaData(component: InitializerKeyKt.initializerKey(initializer: lifecycle)) { lifecycle }
    }
    try Startup.shared.install(context: DefaultContext.shared, manifest: manifest)
}

install is try from 3.0.0; see Every entry point carries @Throws below.

The Swift name of Context is StartupContext. A Kotlin typealias does not survive the Objective-C export, so the class itself carries @ObjCName("StartupContext"). That keeps the Swift name equal to the alias this README already tells Kotlin authors to prefer, and it keeps a bare Context out of the framework's namespace — which matters, because UIViewControllerRepresentable declares a Context of its own and a framework-level one shadows it in the iOS host file every Compose Multiplatform app has, with a does not conform to protocol error that never mentions the name.

Name any initializer from Swift with initializerKey(initializer:). It is the only key overload Swift can reach in practice — a reified type argument cannot cross the boundary, and although initializerKey(kClass:) is exported, a KotlinKClass is not obtainable from Swift to pass it. It works the same whether the initializer was written in Kotlin or in Swift:

final class HostInitializer: NSObject, Initializer {
    func create(context: StartupContext) -> Any? { ... }

    func dependencies() -> [InitializerKey<Initializer>] {
        [InitializerKeyKt.initializerKey(initializer: KoinInitializer())]
    }
}

KoinInitializer there is a Kotlin class, and naming it this way constructs one. Be clear-eyed about that: initializerKey(objCClass:), removed in 2.0.0, took the class object and constructed nothing, so a Swift dependencies() that names three Kotlin components now runs three constructors that the returned keys then throw away, once per call. It is free for an initializer that holds nothing in its constructor and does its work in create, which is what this library asks of every initializer anyway: AndroidX builds each one reflectively through getDeclaredConstructor().newInstance(), at a moment the author does not choose. It is not free for one whose constructor has a side effect, and such an initializer should not have one. If the cost is real for you, hoist the keys into a let computed once rather than rebuilding them per dependencies() call. See Upgrading from 1.x below.

Every entry point carries @Throws(StartupException::class). Without it Kotlin/Native does not propagate an exception to Swift as an NSError: it prints Function doesn't have or inherit @Throws annotation and thus exception isn't propagated and terminates the process, so a diagnosable launch failure arrived as an abort. Startup.install, Startup.getInstance, AppInitializer.initializeComponent, AppInitializer.initializeComponentOrNull, AppInitializer.isEagerlyInitialized, StartupTask.invoke, StartupPlanner.plan and StartupPlanner.validate are all try in Swift now. :startup:checkObjCExport asserts the NSError parameter on every one of them, so a dropped annotation fails the build rather than the app.

A Swift WaveRunner has one spelling that compiles, because the task list crosses as an array of Objective-C objects:

final class ConcurrentWaveRunner: NSObject, WaveRunner {
    func run(wave: [StartupTask]) {
        DispatchQueue.concurrentPerform(iterations: wave.count) { index in
            do { try wave[index].invoke() } catch { }
        }
    }
}

A Swift runner cannot rethrow, and does not have to. run is exported without an error parameter, so a throws conformance does not compile — Swift reports candidate throws, but protocol does not allow it. Catching and dropping is correct here: the task recorded the failure before it reached you, and the engine re-raises it as a StartupException naming the component once run returns. That is why the contract below says a runner must let a failure out where the language allows.

concurrentPerform runs some iterations on the calling thread and some on workers, which is worth knowing: an iteration that lands on the calling thread may resolve another component and one that lands on a worker may not, so a runner written this way must not touch AppInitializer at all.

:startup's checkObjCExport task links Startup.framework and asserts on the generated header, and :sample's checkConsumerObjCExport does the same for the two frameworks a consumer gets, so neither the export shape nor the recipe above can regress unnoticed.

Everything in the API mapping below behaves the same on all eleven targets. Six things do not, and the first four are cases where code written and tested on Android would misbehave elsewhere.

• The order of independent components. Both runtimes always create a dependency before the component that declares it. For components with no edge between them, AndroidX walks each eager root depth first while StartupPlanner emits Kahn levels, so the two pick different valid topological orders. Anything that must run before something else has to say so in dependencies().
• Components nobody registered. AndroidX resolves any class reflectively and never consults a manifest, so on Android initializeComponent succeeds for a component that no StartupManifest registers and for one registered with remove<T>(). Off Android the manifest is the only registry there is, and both throw StartupException.
• Where a failure comes from. Off Android the planner raises this library's StartupException with a components path. On Android the failure comes out of AndroidX as androidx.startup.StartupException, with the message Cannot initialize <FQCN>. Cycle detected. and no path. StartupPlanner.validate(manifest) is the way to get this library's diagnostics on Android too.
• A factory that builds something else. A factory has to produce exactly the class its key names. The check is commonMain, so it fires on every target where the graph is planned — but Startup.install on Android never calls a factory, so there it fires only under StartupPlanner.validate(manifest). An Android app that never validates keeps running, on the class AndroidX reflected rather than the one the factory would have built.
• A component that produces nothing. Off Android Initializer<T> leaves T unbounded and the Objective-C export gives a Swift author Any? to return, so create may hand back null; the engine stores it, initializeComponentOrNull returns it, and initializeComponent names the component rather than failing a cast. On Android androidx.startup.Initializer.create is @NonNull, so such an initializer does not compile at all — an Android-compatible component with no product is an Initializer<Unit> that returns Unit. It is the same constraint that bounds CoroutineInitializer<T : Any>.
• Which registry decides. Off Android the StartupManifest is the whole registry. On Android it is the AndroidManifest, and the factories in the StartupManifest are never called. That is the one failure mode adopting this library adds. The AndroidManifest is the source of truth there and a parity test of your own is the answer to it; see Keep the two Android registries in step above.

By default both runtimes create one component at a time on the calling thread. Everything in a StartupPlan wave depends only on earlier waves, though, so a wave is safe to run all at once, and Startup.install(context, manifest, runner) hands each one to a WaveRunner in turn:

Startup.install(context, manifest) { wave ->
    runBlocking { coroutineScope { wave.map { async(startupDispatcher) { it() } }.awaitAll() } }
}

The dispatcher there is the whole point. A task is an ordinary blocking call rather than a suspending one, so async { } without a dispatcher inherits runBlocking's single-threaded event loop and runs the wave one task after another on the calling thread — which is exactly what Startup.install(context, manifest) already does, at the cost of a runner that looks concurrent.

Give it a dispatcher of its own rather than reaching for Dispatchers.Default. A task blocks the thread it runs on for as long as its component takes, so a wave of CoroutineInitializers dispatched onto the pool their own createAsync resumes on starves that pool and hangs the install; see Initializing something that suspends. Dispatchers.IO is elastic on both the JVM and Kotlin/Native and is the safe default here.

The library keeps the ordering, the cycle detection, the deduplication and the created components; the concurrency is yours. install waits for each wave before planning the next, so run must invoke every task exactly once, must let a task's failure out rather than swallowing it, and must not return until all of them have finished. A second invocation of a task is refused at the call, by StartupTask.invoke itself; the other two are checked once run returns. Either way the violation is a StartupException naming the components it applies to rather than a component quietly filed as null.

A task is a StartupTask, and it names the component it will create. task.component is the key and task.toString() is the component's name, so a runner can route a wave rather than merely run it — the one component that has to stay on the calling thread dispatched differently from the rest — and can attribute a slow or failed wave to a component. It is also where a host hangs its own tracing, because the engine has no hook of its own:

Startup.install(context, manifest) { wave ->
    runBlocking {
        coroutineScope {
            wave.map { task ->
                async(if (task.component == mainThreadOnly) Dispatchers.Main else Dispatchers.IO) {
                    trace(task.toString()) { task() }
                }
            }.awaitAll()
        }
    }
}

A task may call AppInitializer.initializeComponent only from the thread that called install, and only for what an earlier wave already created. The lock is held across the whole install and it is reentrant, so a runner that stays on the calling thread — which is every runner on Kotlin/JS and Kotlin/Wasm — may read a component an earlier wave built. From any other thread that call can never be served, and it now fails immediately with a StartupException saying so rather than waiting on a lock the installing thread cannot release until run returns.

A component of the wave being run right now is refused too, on every thread. Nothing a wave creates is written back until run returns, so a sibling is neither created nor creatable from inside one — and it is refused by name rather than reported as a cycle, because two components share a wave precisely when neither declares the other, so there is no cycle to draw. A component that asks for itself inside a wave gets the same refusal, which says so; without a runner that one is still reported as the cycle it is. Declaring the edge in dependencies() is what makes the call safe from anywhere, because that is what puts the dependency in a strictly earlier wave. That is what makes sample's own manifest safe under a runner: NetworkInitializer and AnalyticsInitializer resolve what they need imperatively and declare the same edges, which is the AndroidX-documented pattern, so each of them only ever reads back a component an earlier wave already built.

On Android the runner is ignored — androidx.startup creates each component itself, depth first on the calling thread, and offers no seam to change it. For ordinary Initializers a runner is therefore a performance decision on the other ten targets and never a correctness one, so anything that must run before something else still has to say so in dependencies().

For a CoroutineInitializer it is a correctness decision. create blocks the thread it is called on until createAsync finishes, and the runner is what picks that thread: on Android it is always InitializationProvider's, which is the main thread, while off Android a runner can move it to a worker or leave it on whoever called install. Which thread that is decides whether the component's own dispatching can make progress — see Initializing something that suspends.

You can also skip AppInitializer for the concurrent part entirely: plan with StartupPlanner.plan(manifest, roots, satisfied), read plan.waves, construct your own initializers — you wrote the factories — call create(context) on a level in parallel, and hold the results yourself. That is a fork in the road rather than a seam, and it is worth measuring first: an initializer that hands its work to a background scheduler and returns immediately costs the same either way.

The graph's promise is that a dependency is created before the component that declares it. For a component whose real work is a suspend call — which is nearly every mobile SDK — a plain Initializer can only keep half of it: create has nowhere to await, so the idiomatic escape is to launch the work and return, and the graph then orders the launches rather than the completions. A dependency edge that does not wait is not a dependency edge.

CoroutineInitializer is where a component says its work suspends:

class FirebaseInitializer : CoroutineInitializer<Unit> {
    override suspend fun createAsync(context: StartupContext) = Firebase.start(context)
}

create is inherited and blocks the calling thread until createAsync returns, so anything declaring this component in dependencies() starts after it has finished rather than after it has begun. Four consequences follow, and every one of them is yours to accept:

• It blocks. On Android that thread is whichever one InitializationProvider runs on, which is the main thread; elsewhere it is whoever called Startup.install, unless a WaveRunner moved the wave. A component that must not block startup should still launch and return, and say so by staying an ordinary Initializer.
• It must not need the thread it is blocking. A createAsync that dispatches to the main dispatcher, from the main thread, deadlocks. That is the ordinary runBlocking hazard and this type does not change it.
• It must not be dispatched onto the pool it will resume on. This is the same hazard one step out, and it is the one a runner walks into. Each blocked create holds a worker of the pool the runner sent it to; if createAsync then needs a worker of that same pool to resume, a wave with as many such components as the pool has parallelism starves. Measured: twenty components whose createAsync does withContext(Dispatchers.Default), run by a Dispatchers.Default runner on an 18-core machine, entered 18 bodies, resumed none, and never returned. Three of the same components finish; so do twenty that only delay. Give the runner a dispatcher the components do not use.
• The thread guard does not reach past createAsync. A wave task that calls AppInitializer.initializeComponent is refused at once, but only on the thread the task body runs on. createAsync switching dispatchers moves the call to a thread the guard does not know about, where it waits for a lock the install cannot release — with or without a runner. Declare the edge in dependencies(); resolving a component from inside createAsync is not supported.

CoroutineInitializer is a Kotlin-side type. Kotlin interface default bodies do not become Objective-C protocol defaults, so a Swift class conforming to it inherits nothing and must write the blocking create(context:) itself. A Swift initializer that has to await should implement Initializer directly and do its own waiting.

CoroutineInitializer<T : Any> bounds T to Any, because androidx.startup.Initializer declares create @NonNull and this interface implements it on Android like any other.

Kotlin/JS and Kotlin/Wasm have one thread and no way to park it, so there is nothing for create to block and it throws a StartupException that says exactly that. Those two targets run every ordinary Initializer as before; it is only the blocking bridge that has nowhere to stand.

StartupException is deliberately our own type. AndroidX annotates its exception @RestrictTo(LIBRARY), so aliasing it would make every consumer's catch clause fail lint's error-severity RestrictedApi check. Failures raised by AndroidX itself still arrive as AndroidX's own type on Android, exactly as they do in an app that uses androidx.startup directly.

There is no expect companion object on AppInitializer because a Java static has no member for one to match, which is why Startup exists.

Until 1.1.0 this table carried eight more rows, each of them a declaration androidx.startup genuinely had no counterpart for, and each of them a platform asymmetry: four existed for Android's manifest and did nothing useful off it, three existed on the other ten targets and not on Android, one existed on the Apple targets alone. 2.0.0 removes all eight, so the table now has one shape on all eleven targets. See Upgrading from 1.x below.

One no equivalent row survives, StartupManifest { metaData(key) { it } }, and it was never in that group. It is the registration androidx.startup performs by name in XML, expressed as a key, which makes it closer to androidx.startup than the reified overload beside it, and it is the only way Swift or a plugin host can register an initializer the compiler cannot name.

2.1.0 tightens one rule. A factory must now build exactly the class its key names; 2.0.0 accepted a subclass and filed it under the supertype. That never worked on Android, where AndroidX ignores the factory and reflects the key, so metaData<Base> { Derived() } started Base on Android and Derived everywhere else from one manifest. Register the subclass under its own key. Because the covariance of Kotlin function types lets the reified overloads express the mismatch too, this affects metaData<T> { ... } and lazyInitializer<T> { ... } as well as the key-taking overloads.

The rejection lives in commonMain and fires wherever the graph is planned, Android included: Startup.install never calls a factory there, so it does not surface it, but StartupPlanner.validate(manifest) does, on all eleven targets.

2.1.0 also adds Startup.install(context, manifest, runner) and the WaveRunner it takes, which run each wave of the plan however the host wants rather than on the calling thread; see Running a wave concurrently. It is ignored on Android, and nothing else published in 2.0.0 changed.

3.0.0 changes what a WaveRunner receives. run(wave: List<() -> Unit>) is now run(wave: List<StartupTask>). A task is still invoked the same way — it() — so a runner written as wave.map { async { it() } } needs no edit; one that named the type of its parameter does. What the change buys is that a task now names its component, which is what makes routing, tracing and an attributable failure possible at all.

3.0.0 enforces the WaveRunner contract it always documented. A runner that runs a task twice is refused at the second call; one that skips a task or catches a task's failure and returns anyway is a StartupException naming the components once run returns, instead of a component filed as null that fails much later somewhere else. A wave that fails now also keeps the components that succeeded beside the failure, which is what the sequential path always did.

3.0.0 fails fast instead of waiting for a lock that cannot be released. A wave task that calls AppInitializer.initializeComponent from a thread other than the installing one now throws immediately. It used to park on the JVM and spin without yielding on Kotlin/Native, where it cost a core for as long as the install ran and never ended. A task on the installing thread still works for anything an earlier wave created, as it always did — the flat prohibition in the 2.x README was wrong about that half — but not for a component of the wave in flight, which is refused by name. A thread that is not running a task is unaffected: its wait for the install to finish really does end, so it still waits.

StartupTask's constructor is public, so a runner can still be exercised against a wave the test built rather than only through Startup.install with a real manifest.

3.0.0 adds AppInitializer.initializeComponentOrNull(key), the read for an AnyInitializerKey and for a component whose create returned null. Both were previously unreadable: the key type did not fit initializeComponent, and a null product failed its cast with a bare NullPointerException raised inside the library. That failure is now a StartupException naming the component.

3.0.0 adds CoroutineInitializer, and with it a dependency on kotlinx-coroutines-core for the Android, JVM and native artifacts. See Initializing something that suspends.

3.0.0 annotates the public API with @Throws(StartupException::class), which is source-breaking for Swift: a call that could fail now needs try. Kotlin callers are unaffected.

2.0.0 removes the eight declarations 1.1.0 deprecated, and nothing else. Each carried DeprecationLevel.WARNING in 1.1.0, so a consumer that took the warnings has nothing left to do; one upgrading straight from 1.0.0 gets errors instead, and the table below is the whole list.

The reason is the same for all eight. None of them had a counterpart in androidx.startup, and mirroring androidx.startup is this library's whole contract — for Android that contract is literal, because Initializer, AppInitializer, Context and the key are typealiases of the AndroidX types. They were also the only platform asymmetry in the API mapping table: three existed on the ten non-Android targets and not on Android, four existed for Android's manifest and did nothing useful off it, and one existed on the Apple targets alone.

The first three. sample's SampleReport is the worked example. The initialization order it prints is a list each component appends to from inside its own create, which is exactly why the report reads the same on all eleven targets instead of on ten. isInitialized has no direct replacement and does not need one: asking for a component that already exists returns it without running create again, so there is nothing to guard. manifest() handed back the value you passed to Startup.install — keep it in a val, and components, eagerComponents, isEager and in all still answer from it.

The Android four. The problem they addressed has not gone away. The StartupManifest and the AndroidManifest really are two registries and only Android reads the second, so a component in one and missing from the other still misbehaves on exactly one platform. What changes is who owns the answer: from 2.0.0 the AndroidManifest is the source of truth on Android, its <meta-data> entries are written by hand as a plain androidx.startup application writes them, and a consumer that wants the two held in step keeps its own test. Keep the two Android registries in step above has one, in a dozen lines, against the API that is staying, and sample's AndroidManifestParityTest runs exactly it on every build. The generated <meta-data> block is also gone, so paste the lines once from that test's failure message or write them out; they are three lines for three eager components and they change about as often as the components do.

The Apple one. initializerKey(objCClass:) was the only Apple-only declaration in the API. initializerKey(initializer:) needs an instance, so naming a component from Swift now constructs a throwaway one and runs its constructor — a cost the class-object overload did not have, and the one genuine regression in this release. It bites only an initializer whose constructor does something, and none should: AndroidX builds every initializer reflectively through getDeclaredConstructor().newInstance(), at a moment the author does not choose, so the work belongs in create.

Nothing else changes. Every other declaration published in 1.1.0 stays exactly as it was, the key-taking registration overloads included — metaData(component, factory), lazyInitializer(component, factory), remove(component), initializerKey(initializer) and initializerKey(kClass). They are not deprecated and are not going anywhere: androidx.startup's own registration is by name in XML, so a key-taking overload is closer to it than a reified one, and without them Swift cannot register a host-supplied initializer at all.

Register factories, never instances, and let every factory construct its initializer with no arguments:

val manifest = StartupManifest {
    metaData<NetworkInitializer> { NetworkInitializer() }
}

A factory that passes constructor arguments compiles, and works on all ten non-Android targets:

val manifest = StartupManifest {
    metaData<NetworkInitializer> { NetworkInitializer(httpClient) }
}

On Android it throws. AndroidX never calls the factory: InitializationProvider and AppInitializer build every initializer with getDeclaredConstructor().newInstance(), so a class without a public no-argument constructor fails at runtime with StartupException(NoSuchMethodException). Take what an initializer needs from create(context), or from another component resolved through initializeComponent.

For the same reason the factory is never used to construct the initializer twice: it is called at most once per run, and only off Android. StartupPlanner.plan and StartupPlanner.validate also call it, on every target including Android, because reading dependencies() needs an instance.

These diagnostics come from StartupPlanner, which is what runs on the ten non-Android targets. On Android the initialization itself is AndroidX's, so a failure there arrives as androidx.startup.StartupException instead. StartupPlanner.validate(manifest) is commonMain code and runs everywhere, including in an Android unit test, so it is how to get the diagnostics below on every platform.

A cycle names the component the walk re-entered and prints the cycle itself, not the acyclic path that led to it. For Entry -> LoopHead -> LoopTail -> LoopHead:

Cannot initialize LoopHead. Cycle detected: LoopHead -> LoopTail -> LoopHead

A path longer than twelve components is elided in the middle. The same path is available as data on StartupException.components, first element repeated last, so a test can assert on it instead of matching text:

val exception = assertFailsWith<StartupException> { StartupPlanner.validate(manifest) }
assertEquals(
    listOf(
        initializerKey<LoopHeadInitializer>(),
        initializerKey<LoopTailInitializer>(),
        initializerKey<LoopHeadInitializer>(),
    ),
    exception.components,
)

A dependency nobody registered names both ends and the remedy:

Cannot initialize Orphan. No initializer is registered for it, required by OrphanDependent.
Register it in a StartupManifest with metaData or lazyInitializer, then install that manifest
with Startup.install(context, manifest).

A component a remove<T>() entry hides says so, rather than claiming nobody registered it, so the remedy on offer is not to re-register the entry the application took out:

Cannot initialize Alpha. A remove() entry hides it, and Beta still declares it as a dependency.
Drop that dependencies() entry, or stop removing the component. Startup.install on Android reads
dependencies() reflectively without consulting a StartupManifest, so it creates it there anyway.

A factory registered under a key it does not build names both classes and the remedy:

Cannot initialize Alpha. Its factory produced a Beta instead. A factory has to build the class its
key names: the product would be filed under the registered key here, while Startup.install on
Android ignores the factory and reflects the key, so one manifest would build two different
graphs. Register it under its own key.

Component names are fully qualified on Android, where the key is a java.lang.Class, and simple elsewhere, because KClass.qualifiedName does not compile on Kotlin/JS.

StartupPlanner.validate(manifest) walks the whole graph without calling Initializer.create, which makes a cycle or a missing registration a test failure rather than a launch failure. It does construct every registered initializer, because reading dependencies() needs one.

A component that re-enters the runtime from inside its own create for something that leads back to it is caught the same way, at the point of re-entry, rather than recursing until the stack dies.

sample is a real application on every target, not a compilation unit that only has to type check. It boots the shared graph and prints what happened: which components are registered and how eagerly, the order they were actually created in, the platform SDK CrashReportingInitializer started, and the shared logger after an analytics.track("launch") call. Those lines come from SampleReport in commonMain, so every entry point is one loop over the same list.

Every row above was executed on one Apple-silicon Mac. macosX64 runs through Rosetta. The Linux and Windows binaries run inside a Linux VM, which is worth doing before a release because it is the only local proof those two targets do more than link:

colima start --vm-type=vz --vz-rosetta
docker run --rm --platform linux/amd64 -v "$PWD":/w -w /w ubuntu:24.04 \
  ./sample/build/bin/linuxX64/debugExecutable/sample.kexe
docker run --rm --platform linux/amd64 -v "$PWD":/w -w /w -e HOME=/tmp -e WINEDEBUG=-all \
  debian:bookworm bash -c 'apt-get update -qq && apt-get install -y -qq wine && \
  wine sample/build/bin/mingwX64/debugExecutable/sample.exe'

Two targets link here but cannot be executed on this host, for reasons outside the project:

• iosX64 needs an x86-64 simulator runtime. The installed iOS 26.5 runtime ships dyld_sim as arm64 only, so a Rosetta spawn aborts with could not use 'dyld_sim' because it is not a compatible arch.
• iosArm64 needs a physical device attached; a paired but disconnected iPhone reports transport: None and cannot be targeted.

iOS is a real app, installed and launched like any other. :sample:iosSimulatorApp links the Kotlin/Native binary, lays out SampleApp.app around the Info.plist in sample/iosApp, and ad-hoc signs it:

./gradlew :sample:iosSimulatorApp
xcrun simctl boot "iPhone 17 Pro Max"
xcrun simctl install booted sample/build/iosApp/SampleApp.app
xcrun simctl launch --console booted io.github.kunal26das.startup.sample.app

The app shows the report in a scrollable monospaced view and prints the same lines, so --console gives the desktop output while the simulator shows the screen. Its entry point is the main in iosMain, which hands control to UIApplicationMain; the other seven console targets share the main in consoleMain, which prints and exits.

The two browser tasks start a webpack dev server on http://localhost:8080/ and never exit; stop them with Ctrl-C. They bind the same port, so run one at a time.

Not runnable on a macOS host. linuxX64 and mingwX64 link there and nowhere else: the artifacts are an x86-64 ELF binary and a PE32+ executable, and running either on macOS fails with exec format error. Link them locally, run them on Linux and Windows, which is what CI does. macosX64, iosX64 and iosArm64 link as well and are not run either, for want of an x86-64 host and a physical device.

androidApp is the one module that exists purely to be launched. sample stays a Kotlin Multiplatform library, because its own AndroidManifest.xml and AndroidManifestParityTest are what prove the shared initializers are declarable the AndroidX way; androidApp applies com.android.application, depends on sample, and does nothing but display the report and mirror it to Logcat under the StartupSample tag. It is not published.

Nothing in androidApp calls Startup.install. The components are created by androidx.startup.InitializationProvider at process start, from the <meta-data> entries sample contributes to the merged manifest, and the first line of the report is AndroidX's own isEagerlyInitialized answering for that:

started by androidx.startup.InitializationProvider: true

Run the same sample on two platforms and the difference documented above is visible in the output: Android creates the graph in AndroidX's depth-first order and everything else in Kahn levels.

./gradlew build
./gradlew testAndroidHostTest desktopTest macosArm64Test iosSimulatorArm64Test jsNodeTest \
          wasmJsNodeTest linkDebugTestLinuxX64 linkDebugTestMingwX64

build also links a debug executable for every Kotlin/Native target and assembles the Android sample app, because sample declares binaries.executable() and androidApp is a real application module. Three verification tasks run beside the tests:

• :startup:checkObjCExport links Startup.framework for iosSimulatorArm64 and asserts on the generated Objective-C header — the reified registration functions absent, the key-taking overloads present, Context exported as StartupContext.
• :sample:checkConsumerObjCExport links two frameworks from sample, which is a consumer of the library: one exports :startup and one does not. It asserts that the first carries the names the Swift snippets above use and the second carries the Startup-prefixed ones with no InitializerKeyKt, so the difference export(...) makes stays true.
• :startup:checkAndroidFloors unzips the published AAR and fails if minSdkVersion rises above 21 or minCompileSdk above 34.

The first two need a macOS host and skip elsewhere; the third runs anywhere.

linuxX64Test, mingwX64Test, macosX64Test and iosX64Test are disabled on an arm64 Mac, so linking their test binaries is the local proof. CI runs the first two for real on ubuntu-latest and windows-latest; macosX64Test and iosX64Test run nowhere.

Apache-2.0. See LICENSE.