Files
sync_replay_project/docs/ARCHITECTURE.md
T
foxyzandClaude Opus 5.5 d3be13ecda Errors: refuse links the probe can't play (with Put it up anyway), plain error texts, and others see who can't play
report.error / state.members[].error (optional, no protocol bump): an
errored member is out of the ready gate until its player opens the media
or the media changes; others see "Bob can't play this: ..." and a mark
on the avatar; your own pill offers Retry (reload).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-10 21:29:41 +03:00

1188 lines
75 KiB
Markdown
Raw Blame History

This file contains ambiguous Unicode characters
This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.
# Architecture
Plain overview first: [HOW-IT-WORKS.md](HOW-IT-WORKS.md). Wire format:
[PROTOCOL.md](PROTOCOL.md). Plans: [ROADMAP.md](ROADMAP.md).
## Packages
| Path | What | Rule |
|---|---|---|
| `packages/core` | protocol messages, sync engine (`SyncEngine`), constants, config, Range parsing, logging, Jellyfin client, updater | pure Dart, no Flutter |
| `packages/server` | Relic HTTP + WebSocket server (`SyncServer`, `ServerHub`, `Session`, media endpoint, relay buffer, Jellyfin relay), the `serve()` entrypoint, `bin/server.dart` (dedicated server) | pure Dart, no Flutter |
| `apps/desktop` | Flutter app: UI, `SyncClient`, `SessionCoordinator`, `PlaybackController` (media_kit/mpv), `ServerIsolate`, update UI (`UpdateController`) | 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. `{'t':'file', 'path', 'reply'}`
on the same port makes a local file the hosted one and answers its name and
size (`ServerIsolate.setFile`): the host's in-session File swap then sends a
`swap` with that name and size, which the server matches to the file.
- **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; the
shell says so in a dialog (during setup the host attempt fails with it).
Stop (leave, close, cancel during setup, a failed own join)
sends `stop`, waits up to 3 s for exit, then kills the isolate.
- **Solo session** (watching alone): the same hosting path, so nothing
about solo is a second code path or on the wire. Putting something up
while offline (`AppShellState._putUp`) calls `SessionState.startSolo`:
a `ServerIsolate` with `loopbackOnly` (an extra `'loopbackOnly': true`
key next to the `AppConfig` map; `SyncServer` binds 127.0.0.1 instead
of every IPv4 interface), port 0 (a free one), no password, a host key;
the app joins it and `SessionState.solo` is set. Everything a session
has works (Jellyfin progress, next episode, skip pills, the player); a
link goes up direct (no relay mode). Invite friends opens the Host card,
whose submit (`SessionState.host` while solo) remembers the
coordinator's `lastPick` (the last `swapLink` / `swapJellyfin` /
`swapFile` of this connection) and position, leaves and stops the solo
server, hosts on the card's port and password, swaps the same pick and
seeks once a state shows it (`switching` is set meanwhile; the shell
closes the player and reopens it as it was). If hosting fails or is
cancelled, solo starts again with the same media. Stop leaves without
asking.
- **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`,
`--jellyfin <file>`, `--[no-]yt-dlp`, `--yt-dlp-dir`, `--yt-dlp-cookies`, `--config`; default port 8080) into the same map and calls `serve`
(`--upgrade` / `--check-update` run the [updater](#updater) instead);
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 (`members[].host`: the crown, art rights). 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. Nobody removes anyone: members sit each
other out ("Sitting out" below), the host included, since the hosted
server and its `/m/` file serving do not need the host's player.
## 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 and every
in-set member is ready, fresh and has `buffered >= need`. `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 others present everyone must ready.
- **In-set** (who the gate waits for): a member joins it on joining, or
on coming back from stale, while not PLAYING (any media, the one the
server started with included); joining while PLAYING it stays out
(`ready=false`) until it readies: a late joiner never pauses anyone. A
swap puts every fresh member in the set and keeps everyone's ready flag,
so the next video starts once everyone ready has buffered (the client
setting "Unready when the video changes", `AppSettings.unreadyOnSwap`,
default off, sends `ready=false` itself on a new `mediaId`; not for the
first state of a session, and a reconnect does not restore a ready that
a swap meanwhile would have dropped). Stale, leave, sitting out and a
player error remove from the set.
- **Player error** (`EngineMember.error`, `report.error`,
`state.members[].error`): a member whose player cannot open the current
media says why in its reports; the server takes it out of the set at
once (like sitting out, but automatic: `ready`/`force` do not put it
back, its ready flag is kept), so the others start without it. A report
without it (the media opened, after a reload too) or a swap clears it:
back in the set like a join while not PLAYING; while PLAYING from its
next frame if it is ready (the others then hold for its buffer, as for a
stale member coming back), else from its `ready`.
- **Sitting out** (`EngineMember.out`, `state.members[].out`): a member
stays connected (chat, feed, party, nudges) but is never in the set and
never ready: no `ready` or `force` reaches it, a swap leaves it out, its
reports keep it fresh but their `buffered` counts as 0. Anyone can sit
anyone out (`out` with `out:true`, the host included); only the member
itself comes back (`out:false`): it is then in the set like a fresh join
(unless PLAYING) and still unready. The app: an out member's player has
nothing open (no stream or relay pulls; a relay it serves goes on) and
shows the join-back card (cover, title, episode, the server position and
playing/paused). "Join back" or Space sends `out:false` and then
`ready`: at once while HELD (the gate waits for its buffer), while
PLAYING once the player has buffered `need` (at most 15 s), so the others
do not pause for the load. Putting something up while out comes back
first. A reconnect that lands as a new member sits out again.
- **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, except one within 3 s of the end (`finishSlackSecs`), which plays
out to mpv's own stop; at the end the position shows the full duration.
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. The coordinator
applies states one at a time, newest first (a state arriving meanwhile
replaces any still waiting): interleaved applies could land an older
pause after a newer play. 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: the session's seek bar (`ChapterSeekBar`, one `seek` per pointer
move while dragging and one on release, like media_kit's), the arrow
keys, skips, chapter jumps 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 goes back into the set with its next frame if it is ready
(also one stale at a swap) or the server is not PLAYING.
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 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). 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.
With "When idle, sit out instead of just unready"
(`AppSettings.idleSitsOut`, default off) the 10 min step sits out.
- **Nudge**: `nudge` names a member; the server passes it to that member
only (no feed), at most 3 per second and 20 per minute per sender
(`nudgesPerSec`, `nudgesPerMin`; the rest is dropped silently). The
nudged app plays the nudge sound whatever its per-sound switch (the
volume still applies), toasts "Alice nudged you" and asks the desktop
for attention when its window is not focused: the runners' `starsync/window`
channel (`attention`), the urgency hint on Linux (GTK, cleared on focus)
and `FlashWindowEx` until foreground on Windows.
- **Deviations from Syncplay**: clock never follows the slowest watcher; no
pause-on-leave; anyone can force ready/unready (all or one); presence 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
Three sources, same sync. Any of them can also go through the relay
(below).
- **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 (a web page through mpv's yt-dlp hook, see "yt-dlp"); the
server is only the clock. A pasted link that
`parseJellyfinLink` recognises goes up as a Jellyfin ref instead when the
ServerId is known (in the link, else `/System/Info/Public`), so its
`api_key` is never shared (`mediaForLink`, host and swap alike).
- **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.
- **Jellyfin item**: the media is a ref (`{serverId, itemId, ...}`, wire in
[PROTOCOL.md](PROTOCOL.md) "Jellyfin media"); the server never stores or
sends a URL or token for it. Each app resolves it with its own account and
direct-plays the static stream; guests use the server's own login when
it has one ("Jellyfin relay"), else the member who put it up ("Relay").
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, unknown session | 401, 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=5xx,reconnect_delay_max=4,reconnect_max_retries=3`
(comma form; this mpv build rejects colon-joined): a dropped connection
or a 5xx is retried, at most 3 times per read and 4 s apart at most; a
4xx (a CDN refusing a request) fails at once instead of being retried
(older ffmpeg ignores `reconnect_max_retries`). A load that fails (mpv's
fatal "No video or audio streams selected", or an EDL track that cannot
be opened) before a duration is known is reported once per open
(`onLoadError`), with ffmpeg's HTTP status when this player gets
ffmpeg's log (ffmpeg logs through one mpv instance per process, the
first one alive). That, a yt-dlp hook error and a failed `openUrl` become
the coordinator's `playerError` in plain words (`playErrorText`, core
`link_probe.dart`: 412/precondition/login/cookies/sign in → "This site
needs cookies — Settings → Player → yt-dlp → Cookies", 403 → "Not allowed
(403)", 404/410 → "Not found (404)", "Unsupported URL" → "Not a video
page yt-dlp knows", network errors → "Couldn't reach the site", anything
else as is). It rides in `report.error` (see "Player error" in the sync
model) until the media opens (a new duration) or the server's `mediaId`
moves on; the pill shows "Can't play this: ..." with Retry (Reload
video). 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` while the controller is created: 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). Otherwise the session player logs at `warn`
(`PlaybackController.logLevel`, switched with `mpv_request_log_messages`
on media_kit's handle): media_kit drains mpv's log on the UI thread, one
message at a time, so a chatty level, or ffmpeg logging every retry of a
refused request, could starve the UI. Shader and yt-dlp hook errors are
`error` lines and still arrive. The thumbnail and sound players keep
media_kit's default (`error`) and nothing listens to their log. 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.
Chapters, segments, tracks (`PlaybackController` signals, values from
`playback/media_marks.dart`, no widgets; any controls UI reads them):
- `chapters`: mpv `chapter-list` (observed through the same dynamic
`observeProperty` and read as JSON): `MediaChapter` index, title, start;
end = next start, the last one's = `duration`.
- `segments`: the Jellyfin item's MediaSegments (`GET /MediaSegments/{id}`,
10.10+: Intro, Outro, Recap, Preview, Commercial), fetched by the
coordinator with `OwnPlay` only; relay and blocked viewers cannot call
Jellyfin and use chapters alone. Empty without a segment provider plugin
(the case on the owner's 10.11 server).
- `tracks`: audio and subtitle `MpvTrack`s from mpv `track-list` (id,
language, title, codec, channels, default / forced / hearing-impaired /
external, `selected`), re-read after every selection.
- `position` (250 ms steps), `duration`, `buffered` (end of the cached
range: mpv `demuxer-cache-time`, media_kit's `buffer`, 250 ms steps);
computed `opening`, `ending` (`SkipRange`) and `currentChapter`.
Skip rules (pure, `media_marks.dart`):
- Opening = first Intro segment, else the first chapter whose title has a
whole word OP / Opening / Intro (any case, digits allowed: "OP1").
Ending = first Outro segment, else ED / Ending / Credits / Outro.
- `skipOfferAt` (the coordinator's `skipOffer`, null at `finished`): inside
the opening (up to its last second) "Skip OP" to its end; inside the
ending "Skip ED" to its end only (no guessing about previews: the viewer
decides). Without any opening, without chapters (a chaptered file
without an OP chapter has none on purpose) and with media over 20 min,
"Skip 1:30" during the first 6 min.
- +1:30 (`skipForward`: the pill, the controls' "1:30" button,
Shift+Right) is a `nudge`, so repeated presses add up.
- Chapter steps (PgUp / PgDn, `chapterStepTarget`): forward to the next
start; back to the current start, or the previous one within 5 s of it
(mpv's `chapter-seek-threshold`).
- Every skip and chapter jump is a user seek (debounced, everyone follows).
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](ROADMAP.md) "Epic: Jellyfin".
- `JellyfinClient`: one server + one user. Auth (Quick Connect, password,
`/Users/Me` check, logout), browsing (views, Resume, NextUp, Latest, paged
`/Items` with `DateCreated`, search, seasons, episodes, recently played),
image URLs (tag + maxWidth, no token
needed), PlaybackInfo, stream URLs, MediaSegments (a 404 = none),
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 (the server relay
sends exactly its credential's 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.
## Jellyfin relay
For viewers without an account on the item's Jellyfin
(`packages/server/lib/src/jellyfin_relay.dart`). The server may hold one
`JellyfinCredential` (config key `jellyfin`: `baseUrl`, `serverId`, `token`,
`client`, `device`, `deviceId`, `version`; primitives, so it crosses the
isolate boundary in the `serve()` map). The hosting app fills it with its
own app-obtained token and that token's device fields; the dedicated server
reads it from `--jellyfin <file>` (YAML/JSON with those keys, keeps the
token off the command line) or a `jellyfin:` map in `--config`. An empty
`serverId` is looked up at start (`/System/Info/Public`); if that fails the
relay stays off. The token is never logged (`toString` and config parse
errors leave it out).
- Available (`state.relay`) only while the current media is a ref whose
`serverId` equals the credential's. Only that item is relayed, only to a
valid viewer token: no open proxy.
- Each request opens `/Videos/{id}/stream?static=true&mediaSourceId=` with
`Authorization: MediaBrowser` built from exactly the configured fields
(Jellyfin rewrites a token's device fields from the header), forwards
`Range` untouched and `Accept-Encoding: identity`, and returns Jellyfin's
status and range headers. Bytes stream through (no buffering) and count
toward `bytesServed` as sent. A viewer disconnect cancels the upstream
response, which closes that connection (tested: `HttpClientRequest.abort`
does nothing once the response has arrived; cancelling the response
subscription does).
- No progress reporting in the server: watched state is per client.
## Relay
A member's media through the server's memory, for viewers who cannot
reach it themselves (`packages/core/lib/src/relay.dart` for the wire and
the source, `relay_pieces.dart` for segmented streams,
`packages/server/lib/src/relay_buffer.dart`, `piece_buffer.dart`,
`relay_routes.dart`; wire in [PROTOCOL.md](PROTOCOL.md) "Relay").
What gets relayed (`SessionCoordinator.swapPicked`):
| Put up | Others get it |
|---|---|
| File, by the host | `/m` from the hosting app (no relay) |
| File, by anyone else | relayed from that member's computer |
| Link, "Everyone fetches it" (default) | each app fetches it |
| Link, "Server fetches it" | the server reads it once into the buffer; the link is never broadcast; refused on an open server (it would fetch any address it can reach for anyone) |
| Link, "Through me" | relayed from that member's computer; for links only they can open |
| Jellyfin item (browser pick or pasted link) | own account first, then the server's login (`/j`), then relayed from the member who put it up (with their sign-in, or the pasted link's key, which stays on their machine) |
Relayed links are range-read: a media file as is, a web page as yt-dlp's
formats, resolved where the bytes are fetched (see "yt-dlp"); the swap
carries only the page URL and the mode. A page's formats are one file, or
a video-only and an audio-only file (YouTube has only those): a media of
two **tracks**, each relayed as its own byte stream with its own buffer
and URL, played together by viewers as an mpv EDL
(`edl://!no_clip;%n%<video>;!new_stream;!no_clip;%n%<audio>`,
`relayViewerUrl` / `mpvEdl`). media_kit opens an `edl://` URL unchanged
(it writes it to a playlist and `loadlist`s it) and mpv keeps both
streams in sync across seeks; one URL keeps the app's "one media URL"
path (sync, swap, reopen) as is. The source member plays the page
itself (yt-dlp hook) like any other link.
A page with only segmented streams (HLS `m3u8_native`, DASH
`http_dash_segments`, e.g. X's HLS, generic `.m3u8`/`.mpd`) is relayed as
**pieces** tracks (`"as":"pieces"`, one muxed playlist or video + audio):
the fetcher parses the upstream media playlist (`parseHlsPlaylist`), or
yt-dlp's DASH `fragments` (`dashPieces`: a first fragment without a
duration is the init section, so no MPD parser), into a
`RelayPieceList`: each segment, `EXT-X-MAP` init section and AES-128 key
is a piece, one upstream GET, with `EXT-X-BYTERANGE` folded into the
piece. The server serves a VOD media playlist it writes itself from that
list (`RelayPieceList.playlist`: durations, media sequence,
discontinuities, MAP, AES-128 keys with their IVs; every URI its own
`p/<i>`), so a source can never point viewers anywhere else. One URL per
piece also avoids mpv's seek hang on byte-range fMP4 playlists (Apple's
`bipbop` hung on the first seek read directly). Refused: DRM (SAMPLE-AES,
`skd:`, any non-identity key format, `EXT-X-SESSION-KEY`, yt-dlp's
`has_drm`), a live playlist (no `EXT-X-ENDLIST`; yt-dlp's `is_live`
first), a master or I-frame playlist. A remux on the fetcher (ffmpeg
`-c copy` into one pipe) and yt-dlp's own download + merge were rejected:
the size is unknown until the end, a seek past the written part cannot be
served, and it needs ffmpeg or disk.
How bytes move:
- The source dials `/r/.../source` (outbound, so NAT/CGNAT are fine) and
answers the server's pulls from the file or the URL (`RelaySource` over
a `RelayMedia`: its `RelayTrack`s, each a `RelayBytes`:
`FileRelayBytes`, `HttpRelayBytes`, `LinkRelayMedia` / `PageRelayMedia`
for links; a Jellyfin source sends its own `Authorization` header from
this machine). Viewers only ever talk HTTP to the server, so the source
never sees mpv's request pattern. The tracks of a media share the
channel and its flow control (they compete for the same uplink).
- Upstream GETs (`HttpRelayBytes.read`) are split into back-to-back
ranges of at most yt-dlp's `http_chunk_size` for the format, else
10 MiB (`relayUpstreamChunkBytes`); a buffer pull stays one read.
YouTube serves a range up to 10 MiB at full speed but throttles a longer
or open-ended one to about twice the bitrate (measured 0.8 MB/s against
8.6 to 12.5 MB/s); chunked, a 1080p relay seeks in under 0.8 s (4 to
6.6 s unchunked) and even beats mpv reading YouTube directly, since the
buffer reads ahead at full speed.
- The server keeps one buffer per track of a relayed media (one for a
file), a `RelayBuffer` for bytes: 1 MiB slices in
memory, filled by one sequential pull at a time. Kept: the first and
last 2 MiB (container indexes; mpv reads both on open and on seeks),
64 MiB behind the slowest viewer and 256 MiB past the furthest, the
same as each viewer's mpv buffer (`PlaybackController`), so a skip
inside the viewers' buffer window is served from memory. Evicted by
distance from the viewers, never LRU. With 2+ viewers the source sends
each byte about once. Several tracks split the ahead and behind budgets
by size (one duration, so by bitrate; `relayBudgets`), at least 8 MiB
each, which keeps both windows at the same playback distance (1080p
video + aac: about 246 + 10 MiB ahead); each pins its own head and
tail. A viewer is its token in every track's buffer.
- A pieces track gets a `PieceBuffer`: whole pieces by index, each one
pull (a piece names its index on the channel), served whole with Range
inside a piece. A viewer's position is its newest requested segment
(opening the playlist puts a new viewer at the first one, so its first
pieces are fetched before it asks). Filled: a blocked viewer's piece
(oldest wait first; a pull nobody waits on gives way at once), the init
and key pieces the viewers' segments use, then segments ahead of each
viewer up to the ahead budget, slowest first. Init and key pieces are
kept; segments are evicted by distance (behind costs 4x), sizes
estimated from the bytes per second of the pieces in so far (before
that 4 Mbit/s video, 160 kbit/s audio, which also splits the budget).
A failed pull is retried after 1 s. Measured (mpv 0.41 through the
dedicated server, loopback, ~30 MB/s upstream): Mux's TS HLS (1080p,
one muxed playlist) opened in 0.8 s, seeks 0.07 to 1.1 s; Akamai's DASH
(1080p + aac, two playlists, EDL) 1.3 s, seeks 0.03 to 1.4 s; an X
video's HLS video + audio 0.3 to 0.75 s, seeks 0.03 to 0.4 s; A/V sync
within 0.2 ms after every seek; a second viewer of the same pieces
opened in 0.02 s from the buffer.
- Filling a byte buffer: a viewer blocked on a slice first (oldest wait
first, handing the source over at slice boundaries), then the pinned
slices, then read-ahead from the slowest viewer. Nothing is fetched while nobody
watches through the relay (a Jellyfin source with only account holders
uploads nothing).
- mpv leaves superseded requests open but unread after a seek: only a
viewer's newest request moves its position or asks the source for
anything; older ones still get bytes that are already buffered.
- Flow control on the source (`RelayWindow`): one in-flight budget per
connection, 2 x delivery rate x min RTT (rate max-filtered over 2 s),
clamped 256 KiB to 8 MiB; the server acks bytes as they arrive. A
cancelled pull's unacked bytes are written off, and a cancel that lands
while a frame is being read stops it before it is sent.
- Cleanup: the buffer is memory only. Any swap disposes it (its URLs then
404, the source is closed with 4001 and stops), and so does server
stop. A crash frees it with the process.
- Source away: viewers keep playing from the buffer; a request waits up
to 60 s for missing bytes. The source reconnects with backoff (same
token after a resume); a different size (or piece count and roles)
drops that track's buffer.
`state.media.relay.live` follows the channel, and viewers show "waiting
for <name> to pass the video on" while it is down.
Measured (mpv 0.41 viewers, real server and source, a shaped link
between source and server; `tool/relay_e2e.dart`, 2-minute 1080p test
file averaging 10 Mbit/s with a 15 Mbit/s busy section):
| Source uplink | Open | Seeks (incl. +90 s skip) | Stalls in 50 s |
|---|---|---|---|
| 400 Mbit/s, 20 ms | 0.8 s | 0.01 to 0.03 s | 0 |
| 40 Mbit/s, 40 ms | 2.1 s | 1.1 s first, then 0.03 s | 0 |
| 20 Mbit/s, 150 ms | 5.4 s | 1.1 to 5.3 s | 3 |
| 12 Mbit/s, 40 ms | 5.4 s | 0.5 to 2.9 s | 3 |
3 viewers skipping together on 40 Mbit/s: no stalls, the source sent
101 MiB for 287 MiB served. A late joiner opened in 0.04 s from the
buffer. A 5 s source outage: no stall; 30 s: one stall once the buffer ran
out, then it recovered. Freezes of 0.3 to 0.8 s in 30% of seconds: no
stall. So the source's upload must beat the media's average bitrate
(peaks are covered by read-ahead once it has some lead), and comfortably
so for quick seeks. Rejected along the way: a fixed per-request window
(seeks of 14 s on slow links, data for abandoned positions clogging the
pipe), one shared window across requests (deadlocked on mpv's unread
requests), Dart `http2` (connection window stuck at 64 KiB).
## yt-dlp
Web pages with a video (YouTube, Twitter/X, ...) play through yt-dlp in
all three link modes. Nothing is bundled or installed by hand.
- **Manager** (`YtDlp`, `packages/core/lib/src/yt_dlp.dart`, used by the
app and the server): the managed copy in its folder (`yt-dlp` /
`yt-dlp.exe`) first; if missing it downloads the official standalone
build for the platform (`ytDlpAssetName`: `yt-dlp_linux`,
`yt-dlp_linux_aarch64`, musl variants, `yt-dlp.exe`, ...) from the tag
`releases/latest` redirects to, checks it against that release's
`SHA2-256SUMS`, makes it executable (not on Windows) and renames it into
place. If the download fails, a `yt-dlp` on PATH is the fallback. After
that, at most once a day (`last-check` stamp in the folder) and in the
background, a newer tag is downloaded the same way; nothing waits for
it. `version`, `lastError`, `busy` and `changes` feed Settings → Player
→ yt-dlp ("Update now" forces the check). Every failure is non-fatal.
Folders: the app's is `<app support>/yt-dlp` (`AppYtDlp`, warmed up at
start), shared with its hosted server; the dedicated server's is
`--yt-dlp-dir`, else `$STATE_DIRECTORY/yt-dlp` (systemd), else
`yt-dlp` next to the binary if writable, else
`~/.cache/watch-together/yt-dlp`.
- **Cookies** (`YtDlpCookies`, `cookies.dart`): a yt-dlp folder holds at
most one source, `cookies.txt` (a private Netscape copy, mode 0600,
written to a `.part` that is made private before anything goes into
it) or `cookies-browser` (`BROWSER[:PROFILE]`). Every resolve adds
`--cookies <copy>` or `--cookies-from-browser <spec>` (the update and
`--version` runs do not); yt-dlp may write the copy back. Imports
(`parseCookies`) take a Netscape cookies.txt or a cookie extension's
JSON array (`name`, `value`, `domain`, `path`, `expirationDate` /
`expires`, `hostOnly`, `secure`, `httpOnly`, `session`; host-only
cookies keep their bare host with `FALSE`, the rest get a leading '.';
session cookies expire `0`), grouped by site (`cookieSite`: the
registrable domain by a short rule, with google.com filed under
youtube.com and bilibili's CDNs under bilibili.com). The app: Settings
→ Player → yt-dlp → Cookies (a browser and profile, or a file picked or
dropped, kept whole: every site in it; "Remove cookies"; the shell's
own drop target is off on Settings so only this box takes the file); its hosted server shares the folder,
so it uses them too. mpv's hook gets the same as `ytdl-raw-options`
(`cookies=%N%<path>` / `cookies-from-browser=%N%<spec>`, length-quoted),
re-set before each external link when it changed. The dedicated
server: `--yt-dlp-cookies <file>` / `ytDlpCookies` converted into its
folder at start; without it the copy and browser setting are removed.
Values are never logged and never leave the machine.
- **Resolve** (`YtDlp.resolve`, `resolveArgs`): `yt-dlp -J --no-playlist --no-warnings
--ignore-config` plus the cookies, 25 s timeout: title, uploader, duration, thumbnail,
live or not, and what a relay reads (`pickTracks`), each format with
its URL, `http_headers` and `http_chunk_size`. Only plain http(s)
formats without DRM count:
- the best single file (`pickSingleFile`: video and audio not `none`, a
video container, tallest up to 1080p, else the shortest above), unless
- a video-only + audio-only pair is taller (`pickSplit`): video tallest
up to 1080p (else the shortest above), HEVC before AV1 before VP9
before h264 (the most compact first: relayed bytes cost every viewer
and the server buffer, and mpv decodes them all), then bitrate; audio the original language, not a
`-drc` variant unless alone, m4a/aac before opus, then bitrate.
YouTube has only these (299 + 140 for a 1080p60 video).
- Neither: segmented streams (`pickPieces`: `m3u8_native`/`m3u8` with a
media playlist URL, or `http_dash_segments` with http(s) `fragments`;
not `has_drm`): one muxed playlist (video and audio not known to be
absent) unless a video-only + audio-only pair is taller, with the
same height, codec and audio rules. Relayed as pieces (see "Relay").
- A live stream, or nothing of the above: nothing to relay.
- **Probe** (`probeLink`, `link_probe.dart`): a `Range: bytes=0-0` GET; a
video/audio type, or a generic one on a media file name, is a file
(`looksLikeMediaFile`); anything else goes to yt-dlp, and when yt-dlp
cannot read it the page's `og:title` / `<title>` is kept with yt-dlp's
reason and the GET's HTTP status. `LinkProbe.problem`: why it will not
play, in plain words (`playErrorText` of yt-dlp's reason, else of a 4xx
status), null for a file or a page yt-dlp read. The app probes on paste
(`AppYtDlp.probe`, remembered 10 minutes; media file names too, not a
Jellyfin item link) for the
Link preview and the swap: a page is named by its title, never its URL
tail, and its thumbnail goes up as the banner (no poster). A link with a
problem, or one the probe cannot reach ("Couldn't reach the site"), is
not put up in any mode: the Link page shows the reason (preview and
under the button) and offers "Put it up anyway" (`LinkMedia.anyway`),
for when the server or a friend with cookies may still open it;
`swapLink` refuses it the same way for any other caller.
- **Everyone fetches it**: the swap sends the page URL. Before opening an
external link every app points mpv's hook at its managed copy
(`script-opts=ytdl_hook-ytdl_path=...`, `ytdl-format=bv*[height<=1080]+ba/b`,
once per player), waiting for the download unless the URL names a media
file. A hook error is a player error (see the mpv settings under "Media"). The hook (mpv
0.41) gives mpv the first requested format's `http_headers` (User-Agent
as `user-agent`; Referer, Cookie, X-Forwarded-For as file-local
`http-header-fields`, which win over the empty list media_kit sets on
load) and its `cookies` (`stream-lavf-o`), for every track of its EDL:
Bilibili's CDN answers 403 without the Referer. The app never opens a
resolved format URL in mpv itself (the putter of "Through me" plays the
page through the hook too); relays read them with yt-dlp's headers.
- **Through me**: the putter's probe must give tracks, and a pieces
track's playlist must be readable, else the swap is refused ("Live
streams can't be passed on", DRM, or nothing a relay can read) with a
hint to use "Everyone fetches it". The swap names the tracks; its relay
source reads them with yt-dlp's headers (`PageRelayMedia`). On 403/410
(expired format URLs, 6 h on YouTube; signed segment URLs) from any
track it resolves the page again once for all tracks, finds the same
format ids with the same sizes and piece counts and roles (playlists
re-read and re-mapped by index; else "the video changed on the page")
and goes on from where it stopped. Format URLs are bound to the
resolving machine's address, so they never leave it.
- **Server fetches it**: the server probes the link itself
(`LinkRelayMedia`) with its own yt-dlp (off with `--no-yt-dlp` /
`ytDlp: false`) and relays its tracks the same way, setting
`relay.tracks` with `live`. The putter only refuses a page its own
yt-dlp says cannot be relayed (the server may reach what the putter
cannot); when the server cannot read it, the putter gets an `err` with
the reason. Viewers open a fetched
link's `/r/` URL only once `relay.live` is set ("the server is getting
the video…" until then): resolving takes seconds and media_kit's mpv
gives up on a silent request after 5 s (`network-timeout`).
## Jellyfin in a session
Client side (`lib/src/jellyfin/session_media.dart`, `progress.dart`,
`track_prefs.dart`, wired in `SessionCoordinator`):
- **Resolve**: a `state` whose media is a ref is resolved per viewer
(`resolveJellyfinRef`): an account for that ServerId that is not expired
-> PlaybackInfo with its own user (PlaySessionId, media source; `/Items/{id}`
alongside for series and episode numbers), then mpv plays the static
stream with the token in `http-header-fields` (`Authorization:
MediaBrowser` with this install's fields), not in the URL: `OwnPlay`,
`report.jf {serverId: true}`. Else, when `state.relay`, `welcome.relayUrl`
or `/j/default/<token>/<itemId>` on the dialled host: `RelayPlay`,
`jf: false`. Else, when `media.relay` names a source, the `/r/` URL:
`RelayPlay`. Else `BlockedPlay`: the session shows "No Jellyfin account
here for <server>, and nobody is passing it on" with a button that
opens Integrations over the session; linking (or reconnecting) the
account retries by itself. A 401 marks the account expired. The member
list shows "via host" for members whose `jf` is false for the current
item. A reconnect to the same item with `OwnPlay` keeps playing as is.
- **Progress** (`JellyfinProgressReporter`, `OwnPlay` only): Playing once the
file is open and mpv has picked its tracks; Progress every 10 s and at
once on pause, resume, a seek (position more than 3 s from where playback
should be) or a track change; Stopped on swap, leave, sitting out, window close
(bounded to 1.5 s inside the graceful leave) and at `finished` (position =
duration, so Jellyfin marks it played); seeking back from the end sends
Playing again. Reports carry `PositionTicks`, `IsPaused` and mpv's
`aid`/`sid` as Jellyfin stream indices (`jellyfinIndexForMpvTrack`,
`SubtitleStreamIndex -1` = off). They go out in order on one chain;
failures are logged, one quiet toast per session, playback is never
touched; a 401 marks the account expired and ends reporting.
- **Episodes**: with `OwnPlay` on an episode, `adjacentEpisodes` (one
`/Shows/{id}/Episodes` call; `episodeBefore` / `episodeAfter`) gives the
previous and next episode. The controls show previous / next buttons
(also `<` `>`, Shift+P / Shift+N); from `nextPromptStart` (the ending's
start when one is known, see "Skip rules", else the last 2 min of the
runtime) a floating "Next episode" pill shows outside the auto-hiding
controls. Auto-advance still waits for `finished`. Any
of them swaps to that episode's ref with the clicker's account. Guests
(relay, blocked) see none of it.
- **Auto-advance** (Settings, on by default): at `finished` with a known
next episode, every client with the setting on shows a 10 s countdown
card (Cancel / Play now; without the setting, a "Next" card with Play
next). Exactly one client swaps: `autoAdvanceSwapper`, the lowest member
id among non-stale members whose `jf` for that server is true. Only what
every client sees in `state` decides it (who put the item up is not in
`state`; a late joiner never saw that swap), so clients agree without a
wire change. Cancel is local: it stops the countdown on that screen, and
the swapper's Cancel stops the swap; a seek by anyone ends `finished`
and with it every countdown.
- **Tracks**: the controls' Audio and Subtitles buttons list
`PlaybackController.tracks`. A pick is applied, reported, and saved as a `TrackChoice`
(language, title, codec, channels, forced/HI: from the Jellyfin streams
with `OwnPlay`, else from mpv's track list) per series and as the global
fallback in the `track_prefs` Hive box. On the next Jellyfin item it is
matched with `matchTrack` against the source's streams (mapped to mpv ids
with `mpvTrackId`) or mpv's track list, and selected locally. Nothing goes
on the wire.
- **Change video**: the player's "Change video" opens the Library tab;
Link goes through `mediaForLink`, a Jellyfin pick as a ref, a file
through `ServerIsolate.setFile` (host) or the relay (`swapFile`).
## Desktop UI
One window, `AppShell` (`lib/src/shell/`), whatever the connection state
(online = `SessionState.route` is `session`). Motion is shared
(`motion.dart`): one curve (`fastEaseInToSlowEaseOut`, 320 ms), a fade
through for views that replace each other (the old one gone after 35%, the
new one fading and scaling in from 0.985), `showPopover` (cards that fade
and scale in from where they open; they scroll when the window is short)
and `HScroll` (horizontal rows: mouse drag or Shift+wheel; the plain wheel
always scrolls the page). Space never presses a focused button
(`appShortcuts`): it is the ready key.
- **Party strip** (`party_strip.dart`, 60 px): your avatar (profile card; 42 px like the party's),
the Library / Schedule / Settings tabs with a growing underline (an
`HScroll` when narrow), the update indicator, then offline a grey dot,
"Offline", Join a friend and Host (connect cards; icon-less "Join" and
"Host" under 1100 px), solo a "Solo · only you" chip, Invite friends
(the Host card, titled "Invite friends", "Start hosting") and Stop (no
dialog; no avatars or chat), online the party: `AvatarStack` (host first, then
ready; spreads on hover; 5 shown, 3 under 1100 px), the server name and
"ready n/m" (hidden under 1100 px), chat with an unread badge, and Leave
(asks first). The stack opens the members card (status, host / relay
pills, "via host", "sitting out"; per member force ready, Nudge and Sit
out (anyone, the host too), your own row Sit out / Join back; "Copy invite
address": the dialled address when joined; hosting, `inviteAddresses`
of `NetworkInterface.list` (IPv4, no loopback or link-local, private
ranges first) with the Host port, and a note that friends outside the
LAN need a forwarded port or a reachable server). Empty strip space is the window
drag area; the window buttons sit at its end.
- **Avatars** (`avatar.dart`): `StatusAvatar` draws a ring outside the
picture with a gap in the background colour (both scale with the size):
green ready, amber buffering (with a spinner), grey not ready, dimmed
stale, dimmed with a moon ("away") sitting out (last in the stack); a crown for the host (`members[].host` in the state; your own
app also knows before the first state) and a tower for the member
passing the video on. `memberAvatarImage` is the one picture lookup: yours from the
profile, anyone else's from `coordinator.avatars` by `avatarHash`.
Colours are local: yours is picked in the profile, others' come from
their name.
- **Connect cards** (`connect_cards.dart`): Join lists recent servers as
one-click rows that poll `GET /health` every 10 s while the card is open
(`health_probe.dart`: online, name, viewer count, "needs update" when
`proto` differs from ours), then address, port and password. Host takes
a port and a password (hosted ports as chips) and puts nothing up: what
to watch is picked in the library afterwards. Submit state (busy,
cancel, per-card errors) is `ConnectFlow`, owned by the shell so an
attempt and its error outlive the card; a card closes itself once
in a server that is not solo. Join says that a friend gives the address
and that no account or Jellyfin is needed.
- **Profile** (`profile_card.dart`, `screens/home/profile.dart`): a card
tinted with your colour: picture (hover "Change" opens `CropPicker`:
drag, wheel or slider zoom, previews at 88/42/24 px, exports a 256 px
PNG), name edited in place, colour swatches, links to the Settings pages.
`Profile` stores the name in the connect form and the PNG and colour in
the `connect_form` box (an old picture path is shrunk once and kept as
bytes). Online, a name change (once typing settles) or picture change
goes out through `Profile.push` (`coordinator.updateProfile`).
- **Welcome** (`welcome.dart`): offline the Library opens with
`WelcomeBand` over the source switch, only for newcomers (no Jellyfin
account, no recent server, not dismissed with "Don't show again";
sizes away when online): the app
icon, Starsync and "Watch together, in sync", then three `WayCard`s,
Join a friend first (filled), Host, Watch solo (opens Files, or
Jellyfin when linked); beside the name from 940 px, under it below,
stacked under 520 px. With no Jellyfin server linked the Jellyfin source
is `NoJellyfinPanel`: Files and Link cards ("No account needed") and an
optional Jellyfin card whose "Link a server" opens the sign-in in place.
- **Tabs**: Library (`library_tab.dart`): a source switch (Jellyfin /
Link / Files, fade through) with the Jellyfin library chips beside it.
Its state (source, link field and relay mode, selected Jellyfin card,
last pick) is the shell's `MediaPickerController`
(`screens/home/media_picker.dart`; picks are sealed `PickedMedia`,
`picked_media.dart`), so it survives tab switches; it opens on Jellyfin
with the last put-up selected. Jellyfin cards select on click and open
`JellyfinDetailsPanel` (`jellyfin/details_panel.dart`, 400 px, slides
in, docked from 1000 px library width, floating over it below; content
cross-fades between items): banner, meta, "Put it up · S1E3", mark
watched / unwatched, remove from Continue, overview, the episode list
(click picks the episode, double-click puts it up). A double-click on a
card puts it up (a series: `playableFor`, its next episode). Link
(`library_link.dart`): paste field, a preview (a page's yt-dlp title,
uploader, length and thumbnail, else the file name; see "yt-dlp"), the three relay modes online, and the
last 12 links put up (`recentLinks` in the `connect_form` box). Files
is `FilesPage` (`files/`): watched folders (`FolderStore`) scanned in
an isolate, videos grouped by series; series start collapsed, a header
click opens one and open ones stay open for the app session
(`showFilesFolder` opens a folder from outside). With nobody else
watching (`solo` on the Library, Files and Schedule: offline or solo)
the buttons say Play, Link has no relay modes and Files no relay notice;
offline a put-up starts a solo session. Schedule (`schedule/`) is a week of day columns built
from AniList's airing schedule (`anilist.dart`: pages cached 10 min in
memory and 1 h on disk, stale data shown on 429 or failure) matched
against AnimeFlow (`animeflow.dart`, app API with the Jellyfin token;
`/library` marks shows on Jellyfin it doesn't track, through every AniList
entry of a season (`anilist_map`: a season AniList splits has several)) into a status per
episode (`schedule_model.dart`). An aired episode of such a show reads
"on Jellyfin" only when Jellyfin has it: the season plus AniList
episode + the run's offset, checked against the series' episodes seen
by the linked account (`jellyfinEpisodeNumbers`, one request per series
memoised 3 min; `confirmOnJellyfin` is the pure step); absent reads
"aired · not on Jellyfin yet", unknown (no linked account, lookup
failed) "aired", neither with Put it up. Requests and the owner's inbox are in
`requests_sheet.dart`. It refreshes every minute; mouse drag and
Shift+wheel scroll the board sideways, the plain wheel scrolls a column.
Settings (`settings_tab.dart`) has a left nav with icons
(icon chips under 760 px) over pages of rounded Material cards
(`SettingsPanel`), one help line under every setting: profile, party &
network (Host card port and password in the connect form, recent
servers), player (the shared rows below, custom shader files),
Jellyfin & integrations (linked servers, AnimeFlow's `/me` via
`AnimeFlowClient.fromAccounts`, auto-advance), notifications & sounds
(what pops up; no switches), updates, shortcuts, about.
- **Drop and paste** (`handoff.dart`, `fox_utils`): the shell is one
`DropTarget` (a "Drop to open" overlay while dragging) and a
`HardwareKeyboard` handler for Ctrl+V outside a text field (clipboard
files first, then text). `classifyHandoff` (pure) turns paths or text
into videos (the Files extensions, `kindOf`), a folder, an http(s) link,
an image or nothing; videos win over folders, folders over images.
`HandoffCard` opens as a popover at the pointer: a video goes up through
the shell's put-up (Play when alone; offline that starts solo) or adds
its folder to Files; a folder is added to Files and opened; a link fills
Library → Link (with others the relay mode is picked there). Images are ignored.
Nothing happens while a dialog or popover is over the shell.
- **Sounds** (`sfx/`): the coordinator plays an `SfxCue` for ready,
unready, force-ready, join, leave, swap and others' chat (feed
events), being sat out by someone (the `kick` sound), a nudge (always,
whatever its switch), the idle warning, and connection lost / back (once per drop).
Nothing for what the video already says (play, pause, seeks, the end).
The sounds are synthesized, not sampled: `synth.dart` renders voices
(plucked strings with bends, glass and star bells, a singing bowl,
hollow knocks, wooden clacks, filtered-noise air and stardust, a long
dark space reverb) to quiet WAVs; `sound_bank.dart` voices every event
in three presets (`SoundPreset`: Cosmic, Abyss, Wood & glass) in one
shared key, so mixed presets still agree. `SynthSfx` renders a sound on
first use into the temp dir (cache key: preset, sound, ready index) and
plays it on a pool of three audio-only mpv players, at
`AppSettings.soundVolume` (square curve) unless switched off in
`soundsOff`. Settings → Notifications & sounds has the volume, the
overall sound set (`soundPreset`), and per sound a preview, its own set
(`soundPresetFor`, else the overall one) and a switch; a preview always
re-renders, so an edited spec plays after a hot reload.
- **Player** (`session_player.dart`): `SessionPlayer` holds the one
`Video` of the current player (a new `GlobalKey` per player generation).
The shell keeps it in one layer of the body at one place in the element
tree: full it fills the body, mini it is 384x216 in a corner (drag it, it
snaps to the nearest corner with the shared curve). A mode change fades
the layer out, moves it while invisible and fades it in (with the tabs
when the full player comes or goes), so the video is never mounted twice
nor remounted. A tab sends the full player to mini; × on the mini sits
out (the player unloads and shows the join-back card, `SitOutCard`, over
the still mounted `Video`); × on that card closes the player, the strip
brings it back, and joining back with it closed shows the mini. Fullscreen
is `windowManager.setFullScreen` with the strip hidden (party avatars in
a corner while paused); F, Esc and double-click on the picture toggle it.
The player keys (Settings → Keyboard shortcuts) work while the player is
full and nothing is in front of it; ↑ / ↓ step the volume. Space (ready)
and Tab (chat) are app-wide: `AppKeys` (`widgets/app_keys.dart`, around
the shell) takes them first while in a server, unless a text field has
focus or a dialog or menu is in front. Tab reaches the shell's chat
toggle through `ChatKey` (the strip while shown, the player in
fullscreen). The app's shortcuts map Tab / Shift+Tab to nothing, so focus
never lands on a button Space would press. Something coming up while the player
is closed opens it: full right after joining or after putting it up
here, else mini. Both control designs show `NowPlayingTitle` top left
(`now_playing_title.dart`: 2:3 cover, title, episode line from
`coordinator.nowPlaying`, art through `artBytes`; it fades through when
the media changes); the mini shows title and episode without the cover.
Until playback first starts on a media (server not playing, position
under 1 s) `StartBanner` (`start_banner.dart`) lies under the controls:
the banner (else a blur of the cover, else dark), darkened, with cover,
big title and episode; it fades out on play and stays gone for that
media; under 320 px tall (the mini) it drops the cover and shrinks the
title. One status pill (`player_status.dart`) sits bottom centre above
the controls: `playerStatusFor` (pure) picks the most pressing of
removed, reconnecting (Cancel), a warn/error health message (Retry),
connecting (only after 2 s), finished, someone loading (below
`StateBroadcast.bufferNeed`, the gate's own need at the current
position: less near the end, none at finished; their avatar with the progress round it), you not
ready (a click marks you ready) and who everyone waits for (avatars,
"2/3 ready"); nothing otherwise, info never. Your own `playerError`
comes right after reconnecting ("Can't play this: ...", Retry reloads
the video); members with `members[].error` show as "Bob can't play
this: ..." (before your ready prompt, which then rides in the detail
line) and are never in "waiting for". A member's new error also toasts
once (the pill is hidden while playing); avatars carry a red error mark
and the members card says "can't play this: ...". 34 px avatars, a main line
and a muted detail line. Frosted glass, colour only for problems; it fades through
between kinds. The mini shows its short line in its hover bar. ⚙ in both
control designs opens the player settings card (`controls/
player_settings.dart`, `showPopover`): player design, progress line
when hidden, upscaling shader, hardware decoding (set live on mpv's
`hwdec`), subtitle and audio track. The rows (`playerSettingRows`) are
the same widgets Settings → Player shows; all apply live. The progress
line (2 px, `AppSettings.progressLine`) is the collapsed uosc timeline;
for the standard design `ControlsHoverProbe` mirrors media_kit's
visibility rule (up on mount and mouse movement, down 3 s later or on
exit), as media_kit does not expose it. Fullscreen adds the
`WindowButtons` to the top bar while the controls are up (restore
leaves fullscreen).
- **Chat** (`chat_panel.dart`): a 340 px panel sliding in at the right
(strip button or Tab): chat with avatars; the "Activity" chip (off by
default, `showActivity`, kept for the app's run only) adds the other
feed lines between them. Opening focuses the field; Tab or Esc in it
close the panel and focus goes back where it was. Unread chat badges
the strip's and both player designs' chat buttons; in fullscreen with
the controls down a glass chip (`PlayerGlass`) top right shows it.
**Toasts** (`toasts.dart`): the coordinator's toasts bottom left, with
the avatar of their `ToastItem.actor` (feed and chat toasts; the app's
own have none); each
slides in, stays 5 s, then slides out and collapses its gap. A click on
a chat toast opens the chat; the others let clicks through.
Settings persist in two Hive boxes: `connect_form` (last form, recent
servers and links, the profile picture and colour; saved even when a join
fails) and
`settings`. Link media names are a web page's title (see "yt-dlp"), else
derived from the URL path, never containing a query string.
Jellyfin on the desktop (`lib/src/jellyfin/`, over the core client):
- `accounts.dart`: `JellyfinAccounts` (singleton `instance`), one
`JellyfinAccount` per ServerId (address list, user, token) in the
`jellyfin` Hive box with a random per-install DeviceId; Client = app name,
Version = app version, Device = hostname. Only tokens obtained by
`sign_in.dart` (Quick Connect or password) are stored. Tokens are checked
on start (`/Users/Me`, each known address in turn): online, offline, or
expired (kept, shown as Reconnect; never dropped silently). A 401 while
browsing also marks it expired. `clientFor(serverId)` gives the shared
signed-in client; `cached()` memoises browse calls for 3 min.
- `sign_in.dart` + `sign_in_panel.dart`: address -> `/System/Info/Public`
-> Quick Connect (code, poll, cancel) or password; reconnect must hit the
same ServerId. `accounts_panel.dart` is the Integrations card.
- `browser.dart`: back stack of `BrowseRoute`s (Home rows, library grid,
series) kept alive under `Visibility(maintainState)` so back keeps
scroll and loaded pages; search (300 ms debounce) overlays them. Where
it is (server, stack, a data version that moves after watched-state
changes so rows reload in place, the Continue keys) is a
`BrowseController` shared with the header chips and the details panel.
Cards (`cards.dart`) fill the width with whole columns (`fitColumns`);
rows are `HScroll`s, rows and grids lazy builders; library grids page
100 at a time with skeletons up to the total count. Images
(`images.dart`) are `Image.network` at width steps (160..2560) with the
image tag in the URL, so the image cache key changes only when the
image does.
- Home (`home_page.dart`, rules in `home_rows.dart`): one Continue row,
`mergeContinue` over Resume, NextUp and recently played: one card per
series, ordered by when it was last played, the resume item if that was
the last thing played in it, else next up; series untouched for 60 days
drop out unless nothing else is left; "Remove from Continue" clears a
resume point and hides the series (`dismissed:<serverId>` in the
`jellyfin` box) until it is played again. "Latest in X" loads single
items (`groupItems=false`) and `latestCards` makes one series card per
series (its poster, "S1E1" or "N new"); with `hideReReleases` (setting,
row menu, default on) items added over 30 days after they aired are left
out. "Show all" opens the library sorted by date added, newest first.
- Picking an item gives a `JellyfinMedia` (serverId, baseUrl, itemId,
mediaSourceId?, title, subtitle, runtime; no credentials). The library
remembers the last pick in the `jellyfin` box and puts it up as a ref
(`putUpJellyfinMedia`, `swapPicked`); a hosted server gets the relay
credential (`relayCredentialFor`: the account of the
item's server, else the default one, not expired; its token with
exactly `JellyfinAccounts.clientInfo`'s device fields).
- Browser location: every navigation saves the route stack per server
(`browse:<serverId>` in the `jellyfin` box, `encodeBrowseStack`: ids,
names and the fields pages start from), so the library and the next
start open where the browser was. A `BrowseStart` overrides it: "Change
video" opens the library on the series of the episode playing, its
season selected and the episode highlighted.
Session controls (the Standard player design): media_kit's
`MaterialDesktopVideoControls` with our bars (`shell/session_player.dart`, pieces in
`screens/session/player_controls.dart`); its own double-click and keys are
off (its double-click pushes a fullscreen route with a second `Video`), the
picture's double-click is a raw pointer observer (`DoubleClickListener`).
The top bar holds the now-playing title, Change video, chat and back-to-library
(fullscreen: exit, then the window buttons); the bottom bar ends with the
shader button, ⚙ and fullscreen. Its seek bar is off; ours
(`ChapterSeekBar`: chapter diamonds, hover time with the chapter title)
sits in the bottom bar, which is taller by the bar's height (the top bar
keeps its row at the top). A drag sends no hover events, so the bar keeps
the controls up through a guarded dynamic `onHover()` on media_kit's
private controls state (`keepVideoControlsAwake`). Pills (Skip OP / ED /
1:30, Next episode) and cards live outside the auto-hiding controls. The
Next episode pill and the cards stay up; the skip pill (`SkipPillGate`)
shows for 5 s when an offer of a new kind appears, then only while the
controls are up (`ControlsHoverProbe` for this design, `UoscControls.shown`
for uosc) or the mouse is on it. Volume (`controls/volume.dart`, both
designs): mpv's `volume-max` is pinned to `PlaybackController.maxVolume`
(150) per player; `VolumeBar` runs 0..150 with a tick at 100 and the part
above in amber; the wheel and ↑ / ↓ step 5 %, clamped.
The uosc player design (`controls/uosc/`, after mpv's uosc theme;
"Player design" = `AppSettings.mpvControls`, default Standard): the `Video`'s
controls builder wraps both sets in `SessionControlsSwitch`, which follows
the setting live. `UoscControls` is our
own frame (no media_kit controls state): mouse movement shows it, 3 s
without any hides it and the cursor; `hold` keeps it up during a timeline
drag and while a menu route is open, double-click toggles fullscreen.
`UoscTimeline` runs the full width at the bottom: played part filled,
`cacheRect` (position to `buffered`) hatched (`hatchLines`), chapter
diamonds on the top edge, current time and duration in boxes inside it, a
hover tip with the chapter title over the hover time; hidden, it shrinks
to a 2 px line with chapter ticks (faded out when the progress line
setting is off). Above it a row of flat buttons: menu (≡:
change video, chat & activity, reload, mark everyone ready, fullscreen,
leave),
subtitles and audio with track counts, a shader badge ("A HQ", opening the
shared `showShaderMenu`), ready, mark everyone ready, reload; 1:30,
chapters, previous / next episode, change video, volume, ⚙, fullscreen; the
now-playing title top left (`UoscControls.topLeft`, with a top gradient), chat
and back-to-library (fullscreen: exit) top right. Menus (`showUoscMenu`: a popup route above the button, the
current row highlighted) list tracks and chapters with start times. All
actions are the coordinator's (seeks through `player.seek`, the user path;
`selectTrack`, `skipForward`, `seekTo`); keys and the floating pills are
the same in both sets.
Window chrome is frameless (`window_manager`): the party strip is the drag
area and holds the min/max/close buttons (`widgets/window_buttons.dart`);
fullscreen has no drag area and shows the buttons in the player's top bar.
## 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,
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. The Join card's recent servers poll it for live
status.
## App icon and launcher entry
`assets/icon/icon.svg` is the source; `icon_<size>.png` and the Windows
`app_icon.ico` are rendered from it (rsvg-convert, magick). The GTK
application id is `moe.senkochan.watch_together`, which Wayland uses as the
window's `app_id`; compositors take the icon from the `.desktop` file of
that id. A tarball install has none, so on start a Linux release build
(`platform/desktop_entry.dart`, skipped in debug) writes
`~/.local/share/applications/moe.senkochan.watch_together.desktop` (Exec =
the running binary) and the icons into the hicolor theme, only when they
changed. The runner also sets the window icon from the bundle (X11).
## Updater
One implementation, `packages/core/lib/src/updater.dart` (pure Dart, tested
in `packages/core/test/updater_test.dart`), with two front ends: the app and
the dedicated server. Both binaries ship in the same archive, so either one
updates the whole install folder. 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 (app or server, same folder); anything else exits with "update
manually". A manifest whose
protocol is newer than the local one is a required update. Windows apply
renames in-use files to `*.old` (swept on next start); CI tests it against a
running exe on NTFS. The manifest's `assets` maps `linux-x64` /
`windows-x64` to archive URLs; top-level `url` (Linux) is for builds <= 0.2.1. The version is
stamped into both binaries with `APP_VERSION` (`--dart-define` for the app,
`--define` for `dart compile exe`; dev runs report `0.0.0-dev` and the app
skips checks). CLI: `watch-together-server --upgrade` (no display needed;
the shipped `update.sh` / `update.bat` run it), `--check-update` prints only;
it ends with "restart the server to run <version>". `watch_together --upgrade`
does the same from the app binary. In the app, `UpdateController`
(`apps/desktop/lib/src/update/`) runs the same code for the strip's update
indicator, which opens the update card, and restarts the app ("Update and
restart", closing Hive first).
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.
- Sit out instead of kick: friends never need to throw someone out, only to
stop waiting for them; the person stays in chat and can be nudged back.
- 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.
- Presence gate: a lone viewer must be able to watch; everyone present
while the video is stopped must ready before it starts.
- 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. The relay buffer is memory only and sized like a viewer's buffer
(about 320 MiB at most), so a small VPS can carry it.
- The relay is the source's own connection through the server, not a
shared token: a member's Jellyfin sign-in or a link's key never leaves
their machine. Prior art (Plex Relay, Tailscale DERP) relays as a
bandwidth-capped fallback on someone else's servers; here the server is
the owner's, so it carries full quality.
- Avatars by content hash: `state` carries only `avatarHash`; the server keeps
the bytes of current members in memory and serves them at `/a/...`, which
clients fetch once per hash (`AvatarCache`). Uplink per broadcast stays
small whatever the picture size (cap 256 KiB base64 per avatar).
- Media art the same way (`/i/...`, `ArtCache` as `coordinator.art`): the
member who swapped the media in fetches poster and banner once the state
shows its swap landed (`media_art.dart`: Jellyfin series/item images with
its own account, else an AniList title search on the parsed release
name) and sends `art`, so guests without Jellyfin get covers too.
`coordinator.nowPlaying` derives title, episode line and art hashes from
`state.media` (`now_playing_info.dart`).