An Android-first chess repertoire teacher built with native Jetpack Compose and Kotlin Multiplatform. Ashva (अश्व, “horse”) takes its name from the chess knight.
Early alpha · Android 0.14.0. An offline opening-course and original-game teacher, not an exhaustive opening book or independently expert-reviewed curriculum. No Play Store release or playable iOS client exists. Game browsing, per-move coaching and separate analyzed branches are verified; the broader roadmap remains unfinished.
| Learn | Study a line | Identify |
|---|---|---|
 |
 |
 |
Screenshots are from the Android 0.3.1 alpha on an isolated API 33 emulator, using synthetic starter-lesson state. Any demo percentages are explicitly labeled.
- Study and practice all 149 catalog families as White or Black, with rules-derived move explanations and original Ashva plans. The main Learn/Explore flow opens courses, not just starter demos.
- See the Full idea, plans and complete move list; replay, jump, pause and change speed.
- Expand Understand this position during study for both-color pawn structures/break candidates, threats, tactical geometry and conditional middlegame/endgame plans. Board facts, original plans and engine output stay distinct.
- Choose a named variation and return to its exact branch point.
- Get the next-move hint automatically after an incorrect attempt. Legal off-line moves are not automatically blunders; the board stays unchanged until a correct move or confirmed switch.
- Search all 3,815 named source routes plus 78 authored study continuations in 47 families. The bundled opening pack installs locally on first launch; no internet or account is required. Raw source lessons remain under Sourced; original demos remain under Starter.
- Build a family-scoped repertoire for either color: choose preferred moves, include opponent replies, inspect unanswered branches, and practice only recorded routes fitting those choices. Find saved choices under My repertoires; earlier revisions remain available for bookmarked lessons.
- Save named multi-opening repertoires for either color, with exact family revisions and an interleaved practice queue. Check shared-position conflicts and unavailable members before practice; explicitly edit membership to adopt newer revisions. No choices are silently overwritten or unsupported move orders stitched together.
- See original-move counts from installed broadcasts at the editor's current position, including replies outside your family snapshot. January 2020 adds 857 accepted scores to the retained 79-score April fixture; both together supply 936 score records. Exact sources, rejected-record counts and first-visit rules are visible—not representative master popularity, engine rankings or win odds.
- Identify positions using the starter book plus installed source endpoints, including transpositions and explicit ambiguous/unknown/last-known results.
- Paste standard PGN/FEN into the identifier; imports currently open the final position, not the guided game player.
- Browse/search/filter installed full scores under Players & GM games, explicitly follow players, import a private PGN and replay original games from either color with exact-source cold bookmarks. Each move has a checked board-change explanation and conditional principle; user commentary and reported identities/titles remain unverified.
- Compare the next recorded move with bounded offline engine candidates, explore a clearly hypothetical continuation and return to the unchanged original game. Explanations and engine-score perspective work from either POV; previews never rewrite the score/result or prove mastery.
- Resume a saved lesson after cold relaunch, preserving color, cursor, branch history and hint context.
- Android0.15 adds actual chosen-scope recall: practice a route, a saved family policy or a pinned multi-opening set to enroll its learner decisions; Review asks one due position at a time. Profile separates unaided/assisted/unsuccessful answers, Study views and retained ungraded history. Wait for saved status; temporary failed saves have an explicit, idempotent retry.
- Install the three reviewed, bundled source packs into local Room/SQLite without internet or an account.
- Analyze alternatives offline with a separate Stockfish 19 engine, compare the lesson/attempted move and replay hypothetical continuations without changing the lesson.
| Content | Delivered | Limits |
|---|---|---|
| Opening courses | 149 families / 3,893 routes | 3,815 source routes plus 78 authored continuations (15–33 half-moves) in 47 families; conditional Ashva plans and board-derived facts, not independently reviewed theory |
| Opening taxonomy pack | 3,815 sequences / 3,174 names | Immutable CC0 routes, 1–36 half-moves; raw source presentation remains separate from Ashva guidance |
| Legacy starter demos | 7 openings / 13 short lines | Retained under Starter and for existing bookmarks/policies, not the default course catalog |
| Broadcast packs | 857 January + 79 April 2020 accepted scores | 82,679 half-moves, browsable/replayable in Players & GM games; January attempts952 inputs/quarantines95. Not complete careers, authenticated identities or representative master popularity |
The main Ruy Lopez course has 235 named source routes plus 9 study continuations; Sicilian has 391 plus 7, London System 4 plus 4, French 212 plus 5, and Caro-Kann 110 plus 4. All families have move explanations and both-color positional guidance; uncommon families explicitly use general guidance, not claimed specialist theory. Authored study continuations are illustrative, not historical scores or engine-certified best lines. In the trainer, Build / edit my repertoire saves preferred moves and included replies separately from the broad course explorer. Individual policies remain family/color/snapshot-bound; existing seed/source policies and bookmarks retain their original identities and content. My repertoires also supports named, revision-pinned multi-family practice queues. Android0.15 replaces demo Review/Profile/Home progress with actual chosen-scope decisions and an original deterministic spaced schedule. Hints, successful exposed analysis and recent Study count as assistance, not unaided recall. Established means three separated due successes and not overdue; it is not a chess rating or empirically calibrated mastery. See recall identities, intervals and limits.
Offline Stockfish supplies bounded candidates; Ashva explains their checked board changes and offers position-specific conditional ideas, not independent strategic certification or historical intent. Original-game coaching and optional local service validation are implemented;0.15 recall is undergoing its final full regression gates. Board/device quality, cloud sync/deployment and a playable iOS client remain separate work. iOS implementation is owner-deferred while Android is completed. No finite database covers every possible continuation or guarantees wins. See game coaching, grounded teaching and engine provenance/limits.
See delivered coverage, game library, content provenance, offline storage and learning design.
Install Android Studio/SDK (platform 37, build tools 37.0.0), NDK 30.0.16248370, JDK 21 and Node 22. Java/Kotlin bytecode targets 17. Engine preparation supports macOS/Linux; macOS with Xcode also supports shared iOS compilation. Set ANDROID_HOME to your SDK or configure an ignored local.properties; do not commit machine-specific paths.
From the project root:
export ASHVA_NDK_DIR="$ANDROID_HOME/ndk/30.0.16248370"
node scripts/prepare-stockfish.mjs
./gradlew :androidApp:assembleDebug
adb devices -l
adb -s <your-device-serial> install -r androidApp/build/outputs/apk/debug/androidApp-debug.apk
adb -s <your-device-serial> shell am start -W -n com.openinglab.app/.MainActivityReplace the serial placeholder with the exact device shown by adb. The launcher name is Ashva; the existing com.openinglab.app package and database name are retained to preserve upgrades and learner data. Do not uninstall or clear app data to update.
Step-by-step user testing covers phone installation, emulator setup, both colors, hints, branch replay, offline packs and cold resume.
Native Compose UI only—no XML screens, Views, Fragments or WebView board. XML is limited to Android platform resources, including launcher/theme/backup policy.
- Kotlin 2.4.20 / AGP 9.4.1 / Gradle 9.7.1.
- Compose BOM 2026.09.00 / Material 3 / Navigation 3 1.1.7 / Lifecycle 2.11.0.
- API 37, minimum Android 8.0 (API 26).
- Room KMP 2.8.5 / SQLite 2.7.1 / KSP 2.3.12 / coroutines 1.11.0.
- Serialization 1.11.0; shared Android, JVM and iOS compilation targets.
Versions are centrally pinned in the version catalog, not a claim that every dependency is always the newest. Upgrades use official stable releases, compatibility checks and tests; experimental channels need an explicit decision.
androidApp owns Compose rendering, Navigation 3 and lifecycle-aware immutable UI state. shared owns legal rules, notation/PGN, position identity, lesson graphs/replay and Room repository contracts. contentTools imports reviewed data locally using those same rules. Analysis/import work stays off the UI thread.
Optional contentService provides local Ktor/PostgreSQL search, source-attributed versioned pack downloads/deltas and a bounded durable analysis/import queue. It is not deployed or connected to Android; no learner uploads, new providers or accounts are enabled. See local service setup, API and checks.
node --test scripts/*.test.mjs
node scripts/public-audit.mjs
node scripts/run-secret-scan.mjs
./gradlew :shared:jvmTest :shared:testAndroidHostTest :contentTools:test
./gradlew :contentService:test
# Opt-in local Docker/host-UCI service checks, after reviewed engine preparation:
node scripts/prepare-service-engine.mjs
node scripts/test-content-service.mjs
./gradlew :contentTools:run --args=verify
./gradlew :androidApp:assembleDebug :androidApp:assembleRelease :androidApp:lintDebug
# Only on a separate test device; connected tests can uninstall target data.
ANDROID_SERIAL=<isolated-test-serial> ./gradlew :androidApp:connectedDebugAndroidTest
# On macOS:
./gradlew :shared:compileKotlinIosSimulatorArm64 :shared:compileKotlinIosArm64The scanner downloads a checksum-pinned official Gitleaks binary into a temporary directory, prints no matched values, and removes only its own temporary directory. Automatic download supports Apple Silicon macOS and Linux x64; other platforms can run Gitleaks 8.30.1 manually with .gitleaks.toml. It scans directories, not Git history. Neither audit is a guarantee that no vulnerability or secret exists.
CI runs build, domain/importer tests, pack verification, lint, secret/privacy checks, isolated Compose tests and shared iOS compilation. Actions are immutable-pinned with read-only repository permissions; no publishing, signing, cloud deployment or paid service is configured. Hosted CI remains unverified until the owner pushes and sees it pass. Its read-only Relay project-memory check uses node scripts/check-relay-context.mjs without requiring a locally installed skill; agents still follow AGENTS.md and the local Relay claim/doctor protocol.
Read CONTRIBUTING, SECURITY, privacy and public-readiness notes. Reports should use synthetic games and redact device logs.
Original Ashva code and project-authored documentation/assets are licensed under Apache-2.0, Copyright 2026 Shishir Jha. See NOTICE for attribution and license scope for exclusions. This does not relicense chess data or third-party components. Opening taxonomy is CC0-1.0; broadcast exports and derived game data are CC-BY-SA-4.0. Keep source URLs, revisions, changes and attribution with redistributed data. See third-party notices and each immutable pack's manifest/ATTRIBUTION.txt.
Ashva is independent of Lichess, Chess.com and the people named in historical data; no endorsement is implied.