sync_replay_project/docs/ARCHITECTURE.md

Architecture

Plain overview first: HOW-IT-WORKS.md. Wire format: PROTOCOL.md. Plans: ROADMAP.md.

Packages

Path	What	Rule
packages/core	protocol messages, sync engine (SyncEngine), constants, config, Range parsing, logging, Jellyfin client	pure Dart, no Flutter
packages/server	Relic HTTP + WebSocket server (SyncServer, ServerHub, Session, media endpoint), the serve() entrypoint, bin/server.dart (dedicated server)	pure Dart, no Flutter
apps/desktop	Flutter app: UI, SyncClient, SessionCoordinator, PlaybackController (media_kit/mpv), ServerIsolate, updater	ships with watch-together-server
tool	harnesses: range_logger, shaper, fake_viewer, e2e_check	dev only
deploy	systemd unit for the dedicated server

Guard: packages/core/test/flutter_free_test.dart fails if anything under core/lib imports Flutter, dart:ui or signals_flutter; CI runs analyze + tests for core and server on every push. Server has no such test but the same rule; CI checks it with dart compile exe of bin/server.dart. The client WebSocket (web_socket_channel) lives in apps/desktop; Relic has no client.

Server lifecycle

One entrypoint: serve(config, events:) in packages/server. config is an AppConfig.toMap() (primitives only, so it crosses isolates). It starts a SyncServer (bound to all IPv4 interfaces) and reports on the events port: ready (with the port and a stop SendPort), error (start failed), and every log record as log. Sending {'t':'stop'} to the stop port closes all sockets and returns, and the isolate exits.

• Hosting from the app: SessionState.host() spawns an isolate running serve (ServerIsolate), waits for ready, then joins it over 127.0.0.1 like any other client. Server logs go to the app's log with a [server] prefix. The isolate runs with errorsAreFatal: false: an uncaught async error is logged and the server keeps going. If the isolate dies anyway, the session shows "server crashed: ..." and leaves; Home shows it on the Host card (during setup the host attempt fails with it). Stop (leave, close, cancel during setup, a failed own join, being kicked) sends stop, waits up to 3 s for exit, then kills the isolate.
• Dedicated server: bin/server.dart compiled with dart compile exe and shipped next to the app as watch-together-server. It parses flags (--port, --secret, --name, --media, --advertise-host, --config; default port 8080) into the same map and calls serve; SIGINT and SIGTERM stop it. A Flutter binary cannot run without a display, which is why the server is its own build.

SyncServer runs a 1 s heartbeat: stale sweep, closed-socket sweep, silent socket drop, state + feed broadcast. There is one session, id default.

Passwords: the default secret is empty, which means an open server (a warning is logged). Otherwise one shared secret, compared at hello.

Host: when hosting, the app generates a random hostKey per hosting, puts it in the server config and sends it in its own hello. The member whose key matches is the host and cannot be kicked. The address a connection comes from plays no part (so a reverse proxy or tunnel on the same machine does not make everyone the host). The dedicated server has no key, so it has no host. Kick = revoke token, send kicked, close; the person may rejoin at once with a new token.

Sync model

Loosely modeled on Syncplay, with a different readiness design.

• Transport is playing, held or finished. The server owns the clock: positionAt(now) = position + (playing ? now - serverTime : 0). Reports from clients never move it. Only seek, swap, gate transitions and the end of media write position/serverTime and bump eventId.
• Clock offset: serverTime is on the server's clock. Each client sends a ping with its clock every second and estimates the offset NTP style (ClockSync: offset = s - (t0 + t3) / 2 of the sample with the smallest round trip among the last 16). All extrapolation uses SyncClient.serverNow() (local clock plus offset), so a client whose clock is off still lands on the server position.
• Reports: clients send position, buffered, eventId every 1 s, plus the media duration and its mediaId once the player knows it. The server stores buffered and liveness, and takes the first reported duration for the current media (swap resets it); a report with eventId older than current has its position ignored.
• Ready gate (HELD to PLAYING): the in-set is non-empty, every in-set member is ready, fresh and has buffered >= need, and quorum holds: in-set size >= 2, unless at most one member is fresh. need is 4 s (minBufferSecs), or near the end what is left minus 2 s (endSlackSecs, floor 0), since mpv cannot buffer past the end. So a lone viewer plays alone; with a second viewer present one Space waits for the other. Joining members start with inSet=false, ready=false and never block. Disconnect removes from the set; swap clears everyone.
• Pause: while PLAYING, any in-set member who becomes not ready, under need buffered or stale flips the server to HELD (feed action pause by server). Recovery resumes by itself when the gate passes.
• End: while PLAYING with a known duration, reaching it sets FINISHED (feed finished by server, position = duration). Clients pause where they are; FINISHED never auto-resumes. A seek turns it into HELD at the new position (resuming at once if the gate passes); a swap resets to HELD.
• Seek: allowed from any member in any transport; the server writes the position, bumps eventId and records who sent it (seekBy, seekSeq in state for that event); flags are untouched. Clients diff each state broadcast by eventId (StateDiff): new event means seek to the computed server position, transport change means play/pause. Between events: drift over 5 s for more than 2 s seeks, ahead by more than 1.5 s sets rate 0.95 until level, otherwise nothing.
• User seeks: the player is a SyncPlayer whose public seek is the user's: media_kit's seek bar (one seek per pointer move while dragging), the arrow keys and seekTo all become explicit events. The player moves at once (scrub preview); UserSeekGate sends one seek intent with a fresh seq after 300 ms without another request (seekDebounceMs). While the user is seeking, and until the echo of the last sent seq arrives (3 s at most), server seeks never touch the player; the echo of our own seek is not applied either (the player is already there). A server seek that has not landed after 1.5 s is re-applied once (_checkSeekLanded), to the server position recomputed now, and only while its event is still current and the user is not seeking.
• Liveness: WS ping every 5 s; silent over 15 s means stale (out of the gate, stale feed); over 30 s the socket is closed and the member removed. A stale member that was ready goes back into the set with its next frame. Closed sockets are polled with isClosed every second: Relic's server-side onDone does not fire for dead peers, so presence never relies on it.
• Resume: a reconnecting client sends its previous token as hello.resume. If the server still has that member (an unclean drop it has not noticed), the new socket takes it over: same token, member id, ready flag and media URL (the player keeps streaming), no join/leave feed, and the dead socket is closed without removing the member. An unknown or kicked token joins as a new member.
• Client policies: no state broadcast for 5 s means connection lost, clear the server state and reconnect every 1.5 s until cancelled (an epoch counter makes a cancelled attempt stale) or kicked. After a reconnect the ready flag is restored from the server's state. Cancel during setup aborts the join. Closing the window leaves gracefully (2 s timeout).
• Idle (IdleTracker): measured from the later of the last input and the moment playback stopped. While HELD and self ready: warning at 8 min, auto-unready at 10 min, and never less than 2 min after the warning.
• Deviations from Syncplay: clock never follows the slowest watcher; no pause-on-leave; anyone can force ready/unready (all or one); quorum gate; the clock is server time (clients correct for their offset to it).

Additive wire changes only; bump proto and PROTOCOL.md in the same commit.

Media

Two sources, same sync.

• Link: the media field holds an http(s) URL. It travels as media.url in swap and state, and as mediaUrl in welcome. Each app plays it directly; the server is only the clock.
• Host local file: SyncServer.setLocalFile / mediaPath makes the file the session media. GET /m/<sid>/<token>/<name> serves it with a Range table, gated by the viewer's token. The server builds the URL per viewer (http://<host>:<port>/m/default/<token>/<name>), so tokens never reach other viewers. <host> is the authority the client dialled; advertiseHost (default empty) only overrides it when set explicitly. When a state broadcast has no media.url, the client builds the URL itself from the host it connected to. Bytes stream file to socket and are counted as streamed (bytesServed); nothing is stored. The host plays through loopback like everyone else.

Range table (resolveMediaRequest, tested byte-exact):

Request	Response
no Range	200, Content-Length = size
bytes=a-b	206, Content-Range: bytes a-b/size
bytes=a-	206 to EOF
bytes=-n	206 last n bytes
unsatisfiable	416, Content-Range: bytes */size
host has no file	503 + Retry-After: 2
bad / unknown token, kicked, unknown session	401, 403, 410

Always Accept-Ranges: bytes; Content-Type from the extension (contentTypeForFile). Relic hands the Range over as a typed RangeHeader; the endpoint takes the first range and keeps the unit, so non-bytes units still get 416.

What mpv does against it (recorded with tool/range_logger.dart): mp4 opens once with no Range then streams bytes=0- to the end; mkv opens bytes=0- twice (probe, playback), reads a small tail at size-n (cues, expressed open-ended), then every seek is a fresh bytes=N- connection. A server outage of a few seconds is ridden out from mpv's cache with no error. Longer outages exhaust the cache and rely on reconnect (not chaos-tested yet).

mpv settings (PlaybackController): Player with bufferSize 256 MiB (media_kit's bufferSize sets both demuxer-max-bytes and demuxer-max-back-bytes), libass on; pinned via runtime property: demuxer-readahead-secs=150, cache-secs=150 (about two minutes of read-ahead: rides out short drops, intro skips are instant), demuxer-max-bytes=268435456 (150 s up to ~14 Mbit/s; a 1.4 GB / 24 min file is ~8), demuxer-max-back-bytes=67108864, stream-lavf-o=reconnect_streamed=1,reconnect_on_http_error=1,reconnect_delay_max=4 (comma form; this mpv build rejects colon-joined). The cache is in RAM: media_kit sets cache-on-disk without a cache-dir, and mpv falls back. mpv ignores invalid values silently, so every pin is read back and a mismatch is logged and kept in pinMismatches. Measured against the /m endpoint (readahead_e2e_test): a paused player buffers ~147 s ahead (10 s before). media_kit has no public setProperty/getProperty; the code calls them through a guarded dynamic call on player.platform. Hardware decoding comes from Settings (auto, off, custom) and is passed to the VideoController; env SYNC_RELAY_HWDEC (e.g. no) overrides it. It is read per player, so a change applies on the next join or Reload video.

Shaders (per viewer, local only, no wire change): mpv glsl-shaders, set through the same dynamic setProperty and read back (PlaybackController.applyShaders). The value is the platform path list (: joined, ; on Windows). One saved choice in AppSettings (shaderPreset, shaderQuality, customShaders): applied on every player creation (join, Reload video) and live when changed from the session (the control bar's shader menu, Ctrl+0..6, both also save it). The Anime4K v4.0.1 files live in apps/desktop/assets/shaders/anime4k/ (only the 14 the presets use, plus its MIT LICENSE); mpv needs real paths, so extractAnime4k copies them once per version to <app support>/shaders/anime4k-<version>/ (a .complete marker is written last). anime4kFiles maps each mode to the file list of Anime4K's mpv templates: HQ = the high-end GPU set (VL + M networks), Fast = the low-end set (M + S). A missing file, a path containing the list separator, a read-back mismatch, or an mpv libmpv_render error while shaders are on turns shaders off (saved) with a toast.

Render mode: media_kit's Linux/Windows video output creates mpv's render context (OpenGL, else its software renderer) when the VideoController is created, and nothing in its API says which. The player therefore requests mpv logs at v: OpenGL init logs GL_VERSION under libmpv_render (RenderMode.gpu; "suspected software renderer" means llvmpipe, softwareGl); none within 1.5 s of the controller being ready means the software path (software, where shaders do nothing; the shader toast, the menu and Settings say so). After startup the v level adds no steady traffic (measured: 0 messages over 5 s of playback). The hardware path is fragile: media_kit borrows whatever EGL context GDK left current on the main thread (upstream media-kit#1404), so an X11 session or a Flutter/GDK change can silently fall back to software.

Player lifecycle: the Player and its VideoController live for one session: created on join, disposed on leave, replaced by recreate() for "Reload video" (button or Ctrl+R; volume kept). openUrl opens with play: false; play/pause/seek/rate come only from StateDiff actions, the play gate (localPlayAllowed: self in-set and ready) or the user's own seeks (see Sync model). A server seek within 0.5 s of the current position is skipped. leave() stops the player (otherwise audio keeps playing). On media change (mediaId bump) the coordinator opens the new URL and StateDiff seeks to 0, holds and resets rate.

Frozen picture: FrameWatchdog samples about once a second. It hashes a small capture of the compositor's centre (ShownFrameProbe) and suspects a freeze when position advances while that hash stays constant for 3 s. Two differing mpv screenshot-raw frames show the decoder is still producing pictures, which confirms the freeze (a still scene is dismissed). A confirmed freeze reloads the video with a toast. It fires once until the picture moves again for 10 s, so it cannot loop. Seeks, pause, buffering and a hidden window disarm it.

Never unmount GPU-textured widgets (avatar images) on a transport-gated branch: teardown races mpv's frame upload and the window goes black. Keep them mounted and skip painting (Visibility with maintainState).

Jellyfin client

packages/core/lib/src/jellyfin/, pure Dart over dart:io HttpClient behind a JellyfinHttp interface (tests inject a fake). Plans: ROADMAP.md "Epic: Jellyfin".

• JellyfinClient: one server + one user. Auth (Quick Connect, password, /Users/Me check, logout), browsing (views, Resume, NextUp, Latest, paged /Items, search, seasons, episodes), image URLs (tag + maxWidth, no token needed), PlaybackInfo, stream URLs, Playing / Progress / Stopped reports, played / unplayed / resume position.
• normalizeServerUrl, parseJellyfinLink: pasted addresses and item links -> (baseUrl, itemId); accounts are keyed by ServerId (/System/Info/Public).
• episodeAfter: next episode is the successor in /Shows/{id}/Episodes (crosses seasons, skips other copies of the same episode); NextUp means first unwatched, not next-after-X.
• matchTrack / matchTracks: carry audio/subtitle choice to the next episode (language required, then title words, forced/HI flags, codec, channels).

Gotchas (Jellyfin 10.11):

• Every request sends Authorization: MediaBrowser Client, Device, DeviceId, Version[, Token] (values percent-encoded). Unauthenticated calls missing a field get 400 with no body.
• A POST must carry Content-Length even when empty: behind nginx a body-less POST without it gets 400 (this was the Quick Connect Initiate failure).
• Jellyfin overwrites a token's stored device name and app version with the header's. Only use tokens this app obtained itself; a relay using a viewer's token must send that viewer's client fields.
• Static streams (/Videos/{id}/stream?static=true) serve the original file with byte ranges. The share URL has no credentials; each side adds its own api_key (authorizeUrl).
• Ticks: 10,000,000 per second.

Desktop UI

One HomeScreen (screens/home/) with a single top bar (top_bar.dart): profile chip, Join and Host tabs, the update indicator, Settings / Integrations / About buttons, and the window drag area and buttons. Only the selected view is mounted; form state lives in HomeShell. Join (join_tab.dart): one compact address/port/password row, then up to 6 Recent server tiles in a grid that adds columns as the window widens. While the Join tab is visible each tile polls the server's GET /health every 10 s (health_probe.dart) for online, name, viewer count and a "needs update" badge when proto differs from ours. Host (host_tab.dart): the media picker (media_picker.dart: Jellyfin placeholder, Link, File; a reusable widget driven by a MediaPickerController) fills the main area; port, password and Host sit in a 340 px column on the right, below the picker under 860 px. Text fields have fixed or capped widths (720 px max) so they stay compact on ultrawide screens. Submit state (busy, cancel, per-tab errors, server crash) is ConnectFlow. Settings (profile, hardware decoding, updates, video shaders, shortcuts), Integrations (Jellyfin "coming soon") and About open in the same body. Settings persist in two Hive boxes: connect_form (last form and recents; saved even when a join fails) and settings. Link media names are derived from the URL path and never contain a query string.

Window chrome is frameless (window_manager): every screen brings its own drag area and min/max/close buttons (Home's top bar; the session uses widgets/window_chrome_bar.dart; both use window_buttons.dart); the session drag layer is window_drag_layer.dart.

Health reporting

Every failure must show in the UI with a cause and a next step; nothing stalls silently. The app has a HealthStatus signal (severity + message) for connecting, reconnecting, wrong password, version mismatch, kicked, host offline, and starvation ("waiting for X - N.N s buffered"). Errors are logged with a stack trace at the boundary, then turned into a displayable state. The server's GET /health returns JSON (ok, name, proto, viewers, bytesServed, uptimeSecs); it is unauthenticated, so it carries nothing sensitive. Home's recent-server tiles poll it for live status.

Updater

apps/desktop/lib/src/update/updater.dart (no Flutter imports, unit tested). Reads a manifest (version, protocol, url, optional sha256, notes) from updateManifestUrl, compares versions, downloads the tarball, checks sha256 when present, then extracts into a staging dir inside the install folder (same filesystem, so moves are renames; the folder must be user-writable). Each top-level entry is moved aside and the new one moved in, with rollback on failure (writing the running executable fails with ETXTBSY; rename works). A packaged install is detected by data/ and lib/ next to the binary; anything else exits with "update manually". A manifest whose protocol is newer than the local one is a required update. Windows apply (*.old rename, swept on next start) exists but is untested. The version is stamped with --dart-define=APP_VERSION (dev runs report 0.0.0-dev and skip checks). watch_together --upgrade (or the shipped update.sh) runs it from the CLI; UpdateController runs the same code for Home's update indicator, which opens the update card, and restarts the app ("Update and restart"). Settings has Check now and an auto-check toggle. The manifest is updates/manifest.json in this repo, written by tool/release.sh.

Decisions

• Desktop + mpv only, no web clients: exists because Jellyfin SyncPlay is bad.
• No transcoding: bandwidth is plentiful and the Jellyfin box should not be loaded.
• No roles, friends trust model: simplest thing that fits the use.
• Blank password = open server: the owner chooses between open and private.
• The host cannot be kicked: the server dies with the host anyway.
• Server runs inside the app on an isolate, and ships as a separate compiled server: a Flutter binary cannot run headless, so a dedicated server needs its own build of packages/server. An isolate keeps the UI responsive and shares the process (sockets, file paths) without a second executable.
• Host by key, not by address: loopback detection broke behind a same-machine reverse proxy or tunnel.
• Server owns the clock; reports never move it: avoids ping-pong between clients.
• Quorum gate: a lone viewer must be able to watch; two or more must start together.
• Relic for the server, web_socket_channel for the client; packages are chosen on merit, not locked.
• Plain ws://, no TLS: friends use LAN, VPN or a forwarded port; TLS (e.g. a reverse proxy) is a documentation item.
• One session per server, id default: the map is keyed for later but unused.
• No server storage: config and logs only; media never touches the server disk.
• Avatars ride in the 1 Hz state broadcast (cap 256 KiB base64 per avatar): simple, costs host uplink; hash-based serving is on the roadmap.