KMP Migration Analyzer

This skill helps plan the migration of a native Android Kotlin codebase (or a slice of it) to Kotlin Multiplatform. It does not perform the migration. Its single deliverable is a sequential migration report that tells the user what must change, in what order, how to batch it, and where the safe checkpoints are — the points at which the app still compiles and runs even though the migration isn't finished.

The one principle everything serves: behavior-identical, 1:1

The migration must change nothing about how the app behaves. The whole point is to isolate any breakage to "caused by the migration" rather than "caused by us changing the code." So:

• Never propose refactors, redesigns, or "improvements." Translate what exists, as it exists.
• Never hand off to the user's coding-practice skills (data-layer, MVI, presentation, etc.). Applying preferred patterns would re-architect code and reintroduce exactly the behavioral risk we're trying to eliminate. The only skill this one references is kmp-module-structure, and only for Gradle-9 quirks and convention-plugin / module structure on the KMP side.
• Compose Multiplatform is the UI target — not SwiftUI or native iOS views — unless the user explicitly says otherwise.
• You are read-only on the target codebase. You inspect and report; you never edit source. The only files you write are the migration report(s) under docs/migration/.

The two-phase migration model (this drives ALL ordering)

Every recommendation is sequenced around two phases. Understanding why this order exists is more important than memorizing it: we want each step to be verifiable on a working app before the next step, so a human can build and test after every move and know nothing silently broke.

• Phase A — In-Android prep (project is still native Android, fully verifiable on Android). Replace anything that would not be allowed in KMP common code while the project is still a normal Android project. Because the app still builds and runs as Android, each replacement can be tested against existing behavior before anything structural happens. Examples: Groovy Gradle → .kts, Retrofit → Ktor, RxJava → coroutines/Flow, Views/Fragments → Compose, Mockito → MockK, Moshi/Gson → kotlinx.serialization, SharedPreferences → DataStore, Glide → Coil, Hilt → Koin, JUnit → kotlin.test. After each Phase-A step, the Android app still compiles & runs — that's a checkpoint.
• Phase B — Structural move into KMP. Convert modules to KMP modules and move code into commonMain, replicating the existing module structure (see references/ordering-rules.md). Per-module rule: migrate that module's tests first, then the module, so behavior is verified as code moves. Each module/batch that compiles & runs after the move is a checkpoint. For the :composeApp / :androidApp split, the Gradle-9 AGP plugin changes, and convention plugins, consult the kmp-module-structure skill.

How to run an analysis: two stages inside this one skill

The skill always starts with a cheap whole-project Survey, then does Deep analysis per batch. On a small/single-module project the two collapse into one combined report. On a large project the Survey produces a roadmap and the Deep stage runs per batch, possibly across sessions.

Stage 1 — Survey (always first; cheap; whole-project)

Read only the structural surface — do not read every source file here, or a large project won't fit in one pass:

• settings.gradle(.kts) → the module list and project structure.
• each module's build.gradle(.kts) → applied plugins and dependencies.
• the version catalog (gradle/libs.versions.toml) → libraries, versions, and alpha versions.
• gradle.properties, AndroidManifest.xml files, and a rough inventory of source areas (e.g. counts of *.kt files, presence of res/, jni//cpp/ for NDK, *.java files).

Then produce the roadmap (docs/migration/00-roadmap.md, template in references/report-templates.md) containing:

1. Module graph + dependency edges; classify as single-module vs multi-module (per-feature / per-layer). Flag pure-Kotlin/JVM modules — they must become explicit KMP modules.
2. Readiness heat-map per module: already-KMP-clean vs. carries blockers, with blocker tags (Groovy, Retrofit, Views, RxJava, Mockito, Hilt, Java java.time, Moshi/Gson, etc.). Detect blockers by matching dependencies and imports against references/library-map.md.
3. Global migration order from references/ordering-rules.md (Phase-A items first; Groovy→kts before everything; leaf/independent modules before their dependents; tests before their module).
4. Proposed batches — which modules to migrate together in one working session, size-bounded so each batch reaches a checkpoint. This is the user's "how do we batch this" decision support.
5. Checkpoint plan — where the app is expected to compile & run (references/checkpoints.md).
6. Decisions to Resolve (global) — project-wide clarifications (NDK present? alpha versions? unknown libraries?). See ambiguity handling below.

If the project is a single module (or small enough to analyze fully in one pass), continue straight into Stage 2 for the one batch and emit a single combined report instead of stopping.

Stage 2 — Deep analysis (per batch / per user-given slice)

For one batch (or the slice the user hands you), read the relevant source and produce docs/migration/batch-NN-<name>.md (template in references/report-templates.md), linking back to the roadmap. For each module in the batch, enumerate:

• Every blocker occurrence, mapped through references/library-map.md to its category + action
  • exact target coordinate(s). Cite file_path:line so the user can find each one.
• Library version lookups. For each migration target, look up the latest KMP-compatible version on Maven Central (see references/library-map.md for how). For libraries not in the map, search Maven Central / the web to determine whether a KMP build exists; if it's ambiguous, raise it as a Decision rather than guessing.
• expect/actual candidates. Android SDK / system APIs with no common equivalent stay on Android and need an iOS actual. Where the iOS behavior is non-obvious (AlarmManager scheduling, sensors, GPS/location), flag it as a Decision and note that the iOS mechanism must be explored.
• Resources. Android res/ → Compose Multiplatform resources (same structure, different folder). List what moves; no behavior change.
• Build-config fields. Any buildConfigField → BuildKonfig + a convention plugin.
• Tests. Classify each test's framework. Migrate only the Java-based ones that won't work in KMP (Mockito → MockK, JUnit → kotlin.test, Compose/Espresso UI & E2E → shared Compose UI testing, Java assertions → AssertK, flow tests → Turbine). Leave Kotlin-friendly tests that already run in KMP unchanged. Migrate a module's tests before the module. Never change test logic — only swap the framework adapter, so the tests still assert the same behavior.
• Ordered task list with checkpoints for the batch (Phase-A items → tests → structural move), each checkpoint stating the concrete verification command.
• Decisions to Resolve (batch-local).

Detection: what to look for and what it means

The full mapping lives in references/library-map.md (read it before classifying anything). It has three categories:

1. Direct migration (Phase A): there IS a KMP-ready replacement — e.g. Retrofit→Ktor, SharedPreferences→DataStore, Moshi/Gson→kotlinx.serialization, RxJava→coroutines/Flow, Glide→Coil, Hilt→Koin, JUnit→kotlin.test, Mockito→MockK, java.time→kotlinx-datetime, permissions→Moko, WindowSizeClass→CMP adaptive, BuildConfig→BuildKonfig, Groovy→.kts, Views→Compose. Report the exact target coordinate(s) the user uses.
2. Keep on Android + add iOS via expect/actual: Android-specific SDKs with no common equivalent stay as-is on Android; you note an iOS actual is needed. Known equivalents: notifications → KMPNotifier; Splash Screen API → keep Android MainActivity, add an iOS storyboard. Jetpack dependencies that KMP already supports stay as-is, no action.
3. Always clarify with the user: NDK usage; alpha versions in the catalog; libraries usable in KMP common but with no obvious equivalent; system APIs whose iOS behavior differs. See references/clarify-cases.md.

Ambiguity handling (hybrid: gather, then batch-ask)

Run as autonomously as possible. Gather all findings first, using web/Maven research to resolve uncertainty so you ask as few questions as possible. Then:

• Identify the subset of blocking decisions — the ones that change ordering or batching (typically: NDK strategy, an alpha version that turns out not to be KMP-supported, or how to handle a library with no equivalent). Ask those in a single batched AskUserQuestion so the order can be finalized.
• Put everything else in the report's Decisions to Resolve section, each with options and a recommendation, so the user resolves them by reading rather than being interrupted.

Report rules

• Reports go in the analyzed project's docs/migration/ folder (00-roadmap.md + batch-NN-<name>.md). They are sequential — read top to bottom and worked through in order.
• Mark every safe point with ✅ CHECKPOINT — Android app should compile & run here (Phase A) or ✅ CHECKPOINT — <module/batch> migrated; app compiles & runs (Phase B), and include the concrete verification command, e.g. ./gradlew :app:assembleDebug plus running the existing test suite.
• Cite file_path:line for every concrete finding so the user can act on it directly.
• Read references/report-templates.md for the exact section order of each report.

Reference files (read these as needed)

• references/library-map.md — detection signals → category → action → exact target coordinates + Maven Central lookup. Read before classifying any dependency.
• references/ordering-rules.md — two-phase model details, module/Gradle structure rules, dependency ordering, and batching heuristics.
• references/checkpoints.md — the catalog of "compiles & runs" gates and how to phrase the verification step for each.
• references/clarify-cases.md — the always-ask-the-user catalog (NDK, alpha versions, unknown libs, divergent system APIs) and how to frame each question.
• references/report-templates.md — the exact skeletons for 00-roadmap.md and batch-NN-*.md.