The player gains a queue sheet showing the current track and every upcoming track, with move up/down and remove; changes are persisted in the playback snapshot. Music and longform are now separate: library sync splits songs by the user's folder assignments, so longform no longer leaks into search, Shuffle all or moods. The new Longform tab groups audiobooks and podcasts by artist, offers a generated read-only playlist per artist, "Shuffle artist" for random tracks, separate longform search, and a Continue listening list with progress bars and resume. Positions are saved locally and to the server while playing, on pause and on track changes. Recommendation cards now show download progress (queued, downloading, failed with error and retry) and the mood refreshes while downloads are running.
AudioManager Client
Kotlin Multiplatform client for AudioManager — a mood-first, self-hosted music server.
The client and server are designed together. There is no OpenSubsonic compatibility layer: the app speaks the native /api/v1 API, where playback is a server-driven radio session and "likes" are replaced by explicit add to mood actions.
Targets
| Target | UI | Playback | Offline downloads |
|---|---|---|---|
| Android | Compose Multiplatform | Media3 / ExoPlayer | Yes |
| Desktop (JVM) | Compose Multiplatform | MP3SPI, decoded to PCM + SourceDataLine |
Yes |
| Web (Wasm) | Compose Multiplatform | HTML5 Audio |
Browser-managed cache only |
The shared module contains the API client, session/cache storage, radio queue, app state and all screens. Platform code is limited to HTTP engines, storage, audio and back handling.
Features
- Login with a server URL, username and password. The server returns a revocable bearer token.
- Mood grid with live counts; create, rename, delete and edit moods.
- Seed management: search the owner's music, add or remove anchors.
- Server-driven radio: the app keeps a small queue and asks for more from the same mood session.
- Playback events are reported automatically (
started,completed,skipped) so the server learns. - Two add-to-mood actions in the player:
- Add to current mood (the mood of the active radio session)
- Choose another mood (picker)
- YT Music recommendations in the mood editor: refresh on demand, accept or reject candidates, thumbnails, and a clear state for library matches (
Seed added), accepted-for-download tracks (Queued,Downloading,Download failedwith retry) and new picks. Accepting a track the library already has adds it as a seed immediately; the rest are downloaded by the server and promoted automatically. - Library search and playback, plus Shuffle all to queue the whole library without starting a mood.
- A queue sheet from the player: the current track plus every upcoming track, with move up/down and remove. Changes are persisted locally.
- Longform tab for audiobooks and podcasts: per-artist read-only playlists, “Shuffle artist” for random episodes, a resume list with progress, and a separate search scope. Long-form never enters music search, Shuffle all, moods or radio.
- Sleep timer (15/30/45/60/90 minutes) that pauses playback. On Android the pause is scheduled inside the media service, so it still fires when the app is backgrounded or the task is swiped away.
- Offline library cache: moods, songs and the last sync revision are stored locally. When the server is unreachable, downloaded tracks still play.
- The play queue and current track are saved locally, so a restart reopens the player paused and resumes on the next tap.
- Reconnects automatically with exponential backoff (15s to 5min) while offline.
- Downloads retry up to three times before reporting an error; covers and YT thumbnails are cached for the app's lifetime.
- Original MP3 downloads with progress state, removal and storage usage.
- Session persistence: the login survives app restarts until the token is revoked or expires.
- Android background playback through a Media3
MediaSessionService: notification, lock-screen and headset controls, audio focus, and playback that survives the app being backgrounded or the task being swiped away.
Project layout
composeApp/
|-- src/commonMain/kotlin/win/taptap/musicmanager/client/
| |-- api/ # Ktor client and DTOs mirroring /api/v1
| |-- data/ # AppStorage, LibraryCache, session persistence
| |-- playback/ # AudioPlayer abstraction
| |-- store/ # AppStore: state, radio queue, downloads, editing
| `-- ui/ # Compose screens and theme
|-- src/androidMain/ # Activity, Media3 player, filesystem storage
|-- src/desktopMain/ # Main window, MP3SPI player, filesystem storage
|-- src/wasmJsMain/ # Compose viewport, HTML audio, localStorage
|-- src/commonTest/ # API, cache and store tests
|-- src/desktopTest/ # MP3 decoding and PCM conversion tests
`-- ci/ # Android SDK CI image
Build
Requirements: JDK 21, Android SDK 35 for the app target, Node (Kotlin downloads its own for web).
With Nix, everything is provided by shell.nix (JDK 21, Gradle 8, Node 22, Android SDK 35 with build-tools 35.0.0, and the desktop runtime libraries):
nix-shell
The shell accepts the unfree Android SDK license for you; see the comment at the top of shell.nix. Without Nix, point ANDROID_HOME at an SDK containing platform 35 and build-tools 35.0.0.
export ANDROID_HOME=/opt/android-sdk
# Unit tests (JVM)
gradle :composeApp:desktopTest
# Android debug APK
gradle :composeApp:assembleDebug
# composeApp/build/outputs/apk/debug/composeApp-debug.apk
# Desktop application (run)
gradle :composeApp:run
# Desktop fat jar
gradle :composeApp:packageUberJarForCurrentOS
# Web bundle
gradle :composeApp:wasmJsBrowserDistribution
# composeApp/build/dist/wasmJs/productionExecutable
Serve the web bundle from any static host (for example python3 -m http.server inside the production directory). The server allows cross-origin API calls; enter its URL on the login screen. When the bundle is hosted on the same origin as the server, the URL is prefilled.
Android cleartext HTTP is enabled for local development. Use HTTPS in production.
App state and offline behaviour
AppStore is a plain Kotlin class with a StateFlow<AppState>; Compose observes it. It owns:
- Auth — restore token, refresh token, logout.
- Library sync — first launch pages through
/library/songs; later launches use thesincedelta and tombstones. Songs are split into music and longform using the folders assigned at login, so the two never mix client-side. - Mood cache — mood details are stored so the app can open them offline.
- Radio — one session per mood, batches of 12, prefetch when the queue is short, no repeats within a session.
- Downloads — stored under the platform app directory, keyed by deterministic song ID.
- Playback snapshot — the session, upcoming queue and current track are persisted on every queue change and restored on the next launch.
- Longform progress — position is saved locally and to
/longform/progresswhile playing, on pause and on track changes; the resume list is fed from the server when online. - Offline retry — a background loop retries the library sync until the server responds again.
On startup the app is usable before the network call completes: cached moods and library render immediately, then refresh in the background.
Radio and add-to-mood
open mood ──► POST /moods/{id}/sessions ──► queue (12 tracks)
│
├─ track start ──► POST /sessions/{id}/events {type=started}
├─ track end ──► {type=completed} ──► next batch
├─ skip ──► {type=skipped} ──► next batch
└─ add ──► POST /sessions/{id}/add-to-mood
├─ no moodId: current mood
└─ moodId: chosen mood
Adding a song makes it a permanent seed, so it is always eligible in future radio and visible in the mood editor.
CI and releases
.gitea/workflows/ci.yaml runs on Gitea Actions using gitea.tap-tap.win/taptap/audiomanager-client-ci:3. The image has the Android SDK, Node, Yarn and a warmed Gradle dependency cache, so jobs do not download them on every run.
- main / PR: run
desktopTest, build the Android APK and the development Wasm bundle. The optimized Wasm build is CPU-bound and reserved for releases. v*tag: additionally build the desktop fat jar and the optimized Wasm bundle, then attach all three to the Gitea release.
Rebuild the CI image after changing ci/Dockerfile or dependency versions:
scripts/build-ci-image.sh 3
The image build performs one warm-up compile (kotlinNodeJsSetup, kotlinYarnSetup, desktopTest, assembleDebug, compileKotlinWasmJs) and then removes the sources, keeping only /root/.gradle.
Known limitations
- The media notification exposes play/pause and seeking. Queue actions (skip, add to mood) stay in the app because the mood session and its events belong to the server.
- Desktop playback cannot seek (sequential decoder); pause/resume and skipping work. Longform resume therefore starts from the beginning on desktop. MP3s are decoded to 16-bit PCM before the line is opened, because Java Sound has no line for the compressed format.
- Web downloads are disabled because the browser owns its storage; the app still streams and caches metadata.
- Accepted recommendations are downloaded by the server into
<music>/AudioManager Imports, scanned and added as seeds. The server needs outbound network access for this. - YT Music discovery runs server-side through
yt-dlp; downloading accepted tracks that are not in the library is still roadmap. - No iOS target yet. The shared module is ready for one.