Downloads asked the platform for an unmetered network and nothing more, so
there was no way to run them on mobile. A setting under Downloads now decides,
defaulting to Wi-Fi only, and it also governs roaming: the platform permits
roaming by default, which would have contradicted the Wi-Fi-only choice.
The flag rides the enqueue request like the download quality does, cached in the
source repository so the poller does not read the preference per row.
The song context menus gain Info, which lists every field we hold for the song
and, when it is downloaded, the file's format, bitrate, sample rate, channels
and size read from the file itself. For the song currently playing it takes the
encoding from the media3 stream instead, so a transcoded stream reports what is
actually being played. The menu is shared, so this reaches the album, playlist,
now playing and queue menus at once.
The downloads manager shows the same format and bitrate beside the size on each
song row, read lazily for the rows on screen rather than for the whole tree.
Downloads always fetched the original file. The setting takes the same choices
as the stream quality, and when it transcodes the download uses the stream
endpoint with those parameters instead, so the stored file is already the
wanted encoding rather than the original that would be transcoded on each play.
A preference change applied before the source is ready falls back to original,
and the URI is built from the same cached source the stream URL uses.
A Subsonic server answers a refused download with HTTP 200 and an XML error
document, which the platform downloader saves like any payload, so the row went
Completed and playback then tried to decode the error page. No audio container
starts with markup, so the leading bytes tell them apart: a Completed download
that starts with < or { is failed instead, and a stored Completed row is
re-checked on reconcile so existing bad files get cleaned up too.
The cross-source reconcile went too far: a source switched away from can have
waiting rows that promote will never hand over, since it resolves only while
active and downloads are only ever started for the active source. Those rows
kept the loop alive forever, so it scanned the whole table and ran every second
for the life of the process, and the notification sat there claiming to be
downloading songs that were not moving.
Both now count a row as in flight only when the engine owns it or it belongs to
the active source. Reconcile still covers every source, so a download running
for a source you switched away from keeps updating.
The downloaded-filter DAO test gains a second artist that must be excluded (its
first version passed whichever artist clause you wrote), and the preference test
covers the legacy three-field value and a downloaded round trip.
One toggle per tab, persisted with the rest of the list query, narrowing to
entities that have at least one completed download. It rides the existing filter
SQL: every album, artist and playlist paging query takes a flag and the shared
filter constant adds an EXISTS over song_downloads, so it composes with the
sort, search and starred filter rather than being a separate mode.
The filters sheet now renders on every tab, since playlists support Downloaded
but not Starred.
A source switched away from kept downloading on the platform while our poller
ignored it, so its rows froze as Running, the notification vanished and the only
Cancel went with it. The poll now reconciles every source's in-flight rows (the
engine reports status without a source) and the notification reads all of them,
which is also what cancelActive already acted on.
The grouped tab queries drop entities with nothing downloaded, since the UI
treats a missing entry and an all-zero one the same.
The indicator's two-value enum was used as a boolean, so it is one; and the
library screenshot now asserts the ring is drawn, rather than capturing an
image that would stay green if it disappeared.
The platform posts one notification per download, so queueing an album buries
the UI in them and they can be swiped away. Hiding them (VISIBILITY_HIDDEN,
which needs DOWNLOAD_WITHOUT_NOTIFICATION) and posting one aggregated
notification instead gives a single ongoing entry with the overall progress and
a Cancel action that stops everything in flight.
It is refreshed from the download state the poller already maintains, so it
tracks the engine without extra queries, and it is tied to this process: if we
are killed the downloads continue with no notification, which is the trade for
not writing a service.
A finished download no longer draws anything on album covers, artist rows or
playlist rows: the badge is the progress ring alone, so a fully downloaded tile
looks like any other. The ring takes the theme's primary for the arc over a
theme track, with the one soft shadow, instead of the white glyph and its
outline.
The tree read the track artist off each song, so a downloaded album split into
one group per guest performer and every song without a track artist landed in
"Unknown artist". It reads the album's artist now, which is also the artist the
node's delete resolves through, so the two agree.
The badges are white glyphs with a dark outline and a progress ring with a
visible track, after a look at the rendered result; larger and clearer than the
first pass, which the art swallowed.
The split bar read badly; a partly downloaded list goes back to stacked
Download and Delete items.
The badge showed progress for any list with a downloaded song in it, which
made a mostly-downloaded album look like it was fetching. It now shows only
when that list is complete or has a download actually in flight; a partly
downloaded idle list shows nothing.
The album and playlist detail screens' more-menus had no download entries at
all, leaving delete reachable only via the hero's long press. They now take
the same status-driven items as everywhere else.
A Downloads screen reachable from Settings shows the active source's downloads
as an artist > album > song tree with per-node sizes and song counts, each node
collapsible. Artist and album nodes delete with a confirmation showing what they
free; song rows delete straight away, and a summary row offers Delete all.
Sizes come from the files on disk rather than the engine's reported total, so
they agree with the byte count the delete dialog already quotes. The tree is
built from one query joining the download rows to song, album and artist names.
Album tiles, artist rows and playlist rows now carry the badge (lower right on
tiles, trailing end on rows) and their long-press menus take the same bulk
Download / Cancel / Delete entries the hero uses, driven by the shared actions
rule so a partly downloaded item offers both rows and an in-flight one offers
Cancel. The artist page's album cells get the badge and album menu too.
The three list callbacks the artist menu used (and its artist-specific
view-model calls) collapse into one bulkDownload action over a MenuTarget, since
the same menu code now serves album, playlist and artist.
The detail heroes ask for one entity's status at a time. The list surfaces
need every status on screen at once, so add one grouped flow per tab that
returns a lookup of entity id to counts, which the tiles and rows read the way
song rows read the per-song map.
ChoiceGroup emitted its header and option rows as siblings straight into the AlertDialog text slot, and they drew on top of each other, leaving only one radio visible with garbled labels. Wrap the group in its own Column so the rows stack, and cover it with a test that asserts the presets occupy separate rows.
Third review round:
- `promote` read every source's rows, so a single inactive source with waiting
rows could fill the window and starve the source in use: its songs stayed
queued with no engine id, for good, through every tick. It now scopes to the
active source, which is also the only one that can resolve a URL here.
- An artist with some songs downloaded offered only Download, because the more
menu rendered `action()` rather than the actions the state offers. It renders
the whole list now, so a partial artist gets Download and Delete.
- The hero's long press is only advertised when there is something to delete,
instead of announcing a Delete it would ignore, and it declares a button role
again after losing `IconButton`.
- The delete dialog uses the platform size formatter rather than a hand-rolled
one, which also fixes 999_999 reading as "1000.0 kB".
- The poll warms artwork through the same helper reconcile uses.
- Cancelling a big playlist did per-row work: a binder cancel, two queries and a
delete each, thousands of times, under the lock. The platform takes a list of
ids in one call, so the cancel is now one binder call, one delete per chunk of
rows and one sweep, which is both the several seconds and the stale button.
- Partial files are covered by the same sweep, which already deletes any file in
a source directory that no row names; a test now proves a stray .part goes with
the cancel.
- The hero's control was a plain clickable Box after the long-press work, so its
splash was square; it clips to a circle the way an IconButton does.
The hero's control now carries the whole rule: a tap does what the state offers
(Download, Cancel, Delete) and a long press asks to delete when there is anything
downloaded to delete, so a partially downloaded album can be emptied without
starting the rest first. Nothing downloaded means the long press does nothing.
The rule itself returns the actions rather than one of them, so a partially
downloaded list offers both Download and Delete; the tap takes the first. That
list is what the library tabs' context menus will read.
Only a single song deletes without asking; the artist's delete now opens the same
dialog as the album and playlist, naming the artist and the space it frees. The
artist's download action stays in its more-menu rather than getting a visible
control, deliberately, so bulk downloading a whole artist is a considered
action.
Two things a 3k-song playlist turned up, plus the hero's downloaded state.
- The platform runs about eight downloads at a time, but a bulk action handed it
every song at once. Three thousand pending requests meant a notification flood,
and a poll tick that rewrote every row's bytes/total, so the whole list churned
ten times a second. Now a bulk action only records the songs and hands over at
most IN_FLIGHT_LIMIT, oldest first, promoting the next as each finishes, so our
row state mirrors what is actually running and a cancel only has to reach what
is in flight.
- Cancelling was therefore unusable at that size. It now touches at most the
in-flight requests instead of thousands, and the rows wait behind it.
- A queued row with no engine id is waiting its turn, not lost, so reconcile no
longer re-enqueues the whole list on every pass; a row whose engine entry
vanished goes back to waiting and the next promote picks it up. A song that
cannot be handed over fails instead of being retried every tick.
- Requests follow the order the screen shows: artist songs resolve through the
same albums ordering the artist page uses, so they were the one list that did
not.
- The hero's fully-downloaded state shows the downloaded glyph rather than a bin,
and tapping it asks before deleting, naming the album or playlist and the space
it frees.
Second review round on the bulk downloads:
- The cancel-during-bulk fix had no test, so the invariant it exists for could
regress silently. The fake engine can now delay an enqueue, which holds the
race window open: the queue loop is kept busy, a cancel is fired mid-queue,
and the test asserts the whole list ends up cancelled. Reverting to per-song
locking fails it, which I checked.
- The per-song catch around the queue loop swallowed CancellationException, so
cancelling the caller did not stop the loop. It rethrows now.
- The artwork pass skips songs whose row a cancel has already taken.
- The lock-hold ceiling is recorded where the lock is taken: cancel latency and
the poll tick both scale with the list length, and a per-list cancel
generation would lift that if it ever bites.
- Tests: an empty list says so, one song failing to enqueue leaves the rest of
the list alone, the applyAction test covers Delete as well as Download and
Cancel, and the two playlist fixtures share a helper.
- Cancelling a list while its download loop was still enqueueing missed every
song queued after the cancel looked. The loop now holds the lock for the whole
queue, so a cancel waits for it to finish and then removes all of it. That
needed the enqueue helper split into a locked entry point and a body the loop
can call while already holding the lock — calling the locked one from inside
the lock deadlocked, which is what hung the tests.
- An empty resolved list now says so instead of doing nothing.
- One song failing to enqueue no longer aborts the rest of the batch.
- The artist actions resolve through the albums the artist page lists, matching
what is on screen, rather than through songs.artistId, which for a guest on a
compilation enqueued nothing.
- Bulk actions are one `applyAction` on the repository, so the three view models
and both screens stop repeating the same dispatch.
- Tests: the playlist and artist resolve paths, a Running row counted as
downloading, the action dispatch, and the action rule moved to a plain unit
test.
Third block of #65: bulk downloads and the list-level actions.
- The hero's download button on albums and playlists is real now. It starts the
songs the list is missing, cancels what is in flight, or deletes what is
already downloaded, chosen from the list's own status. The artist has no hero
button, so its bulk actions live in the context menu, which now carries the
same status.
- No new storage: a bulk job is the download rows for the list, and the status is
counted from song_downloads joined to the list's songs, so cancel and delete
stay well defined without anything to keep in sync.
- downloadAll fetches artwork for the whole list under one shared budget, once
per unique cover, and skips songs that are already downloaded so re-tapping
Download resumes rather than restarting.
- make the shuffle a filled pill matching the album/playlist controls
- wrap the artist name with a hanging indent: the first line may run the
full width over the button, while the last line stops short of it
- star/unstar from the context menu fills the icon, then dismisses
- drop the star badge from album art and song rows
- song rows keep a track-number column (natural order fallback), shrink
four-digit numbers, and gain right-hand download and duration columns
- the play indicator always sits in the left column, overlaid on cover art
with an inset primary outline when art is shown
- now playing is left aligned with a star toggle
- the media notification gets a star button backed by a session command
- Add a `SourceRepositoryTest` that sets the preference to 1 and proves the
active source fetches sequentially; removing the `maxConcurrentFetches`
wiring now fails it.
- `playlistSongsRespectsASmallerFetchBound` now holds the first `bound`
fetches and asserts no `(bound + 1)`th starts, instead of timing an
overlap; it fails when the semaphore ignores the bound.
- One `DEFAULT_FETCH_CONCURRENCY` (in `data.source`) backs both the source
default and the preference default, so they cannot drift.
- Clamp the stored preference in `syncConcurrency()` too, so a bad stored
value can't make `SubsonicSource` throw on construction.
The bound was only a constructor default with no production consumer. Add a
"Sync concurrency" preference (Settings -> Network, default 4) and pass it
through SourceRepository to SubsonicSource, so sync parallelism is tunable
without a rebuild. Both the app-scoped source and the on-demand
`activeMusicSource()` used by sync read the preference.
- UserPreferences.syncConcurrency()/setSyncConcurrency() (clamped >= 1)
- SourceRepository combines the preference into toMusicSource
- Settings row + dialog; SettingsViewModel state and setter
- Preferences test for default/round-trip/clamp
- The concurrency bound already existed; this parameter only exists so tests
can force the sequential path, and no production path passes it. Mark it
`@VisibleForTesting` so that is explicit rather than reading as a wired
setting.
- `playlistSongsUseTheSameRowsAtAnyFetchBound` could not fail (bound has no
effect on the rows) — delete it and its `normalized` helper.
- `playlistSongsRespectsASmallerFetchBound` relied on a 300ms overlap
window; use latches so the bound is proven deterministically (fails when
the semaphore ignores the bound).
- The NULL-aware backward album predicate is not sargable, so a backward seek
from a far cursor scans the whole album partition. Fall back to the offset
query when the anchor is more than twice the offset away; near-cursor
backward reads still use keyset.
- Cover the playlist keyset path: a DAO test for `playlistSongsBefore` and a
repository test scanning a playlist forward and backward. Both fail if the
playlist order key stops driving the seek.
`SongListItem` carried a nullable `position` that only playlist rows have,
so the album/song queries projected a fake `NULL AS position` just to map
the shared type — a leak of the playlist concept into unrelated reads.
Split it into `AlbumSongItem` (song, coverArt) and `PlaylistSongItem` (song,
coverArt, position), plus a `SongItem` interface so the queue stays generic
over kinds. The queue's `QueueWindowItem` now carries the optional
`playlistPosition` for the keyset cursor, so no projection fabricates it.
Playlist detail and its paging use `PlaylistSongItem`.
- `albumSongsBefore` used the null-intolerant row-value `<`, so an album
with NULL `disc`/`track` (untagged files) dropped exactly the rows a
backward window needs, returning the wrong track at a position. Use a
NULL-aware predicate (NULL sorts first, so it belongs to any `before`
set); `albumSongsFrom` is unaffected.
- The new `(sourceId, albumId, disc, track, id)` index has the old
`(sourceId, albumId)` index as a prefix, so drop the latter in the
migration and from the entity.
- Tests: mixed NULL disc/track album for the before-window and a full
backward scan; both fail on the previous commit.
- ArtworkStore's fetcher still did `response.body ?: throw` from before the body
size limit landed; OkHttp's `body` is not nullable, so the elvis was dead.
- DownloadRepository's `downloadStates` uses `flatMapLatest` without the opt-in
its callers annotate.
- HeroHeader's subtitle tap invoked `onSubtitleClick?.invoke()` inside the branch
where it is already known to be non-null.
- MediaItemTest used a safe call on `clippingConfiguration`, which never returns
null.
No behaviour change: a clean full rebuild now compiles without warnings.
#71 converted only two reads; the rest of `QueueViewModelTest` still read
`viewModel.rows` — a snapshot list mutated on the main dispatcher — from the
test thread, so `removeAndUndoKeepTheBoundedWindowConsistent` could map a torn
list (the removed position missing) under load. Route every test-thread read
through `rowsSnapshot()` and run the synchronous `reorder` mutations on the
dispatcher too.
- SourceRepository.coverArt and networkCoverArt read the source id and the
source object separately, so a source switch between the two could pair one
source's cache key with another source's URL; for the store that meant writing
the wrong server's artwork into the per-source file for good. Both now resolve
from a single active snapshot.
- ArtworkStore wrote through a fixed `<name>.part`, so two downloads of songs
from one album could interleave: one publishing the file while the other still
had the shared inode open, which leaves a torn image for Coil to decode. Each
write gets its own partial now.
- OkHttpArtworkFetcher read the whole body into memory whatever its size. It
refuses anything past 8 MB, checked by declared length and again while reading.
- Artwork whose library rows could not all be resolved is no longer swept:
pruning an album while its songs remain used to delete art a download still
held. A source with no download rows still sweeps everything.
- fetchArtworkSafely and storeArtwork no longer swallow CancellationException,
and the poll loop fetches a resumed row's artwork off the tick instead of
holding up the next one.
- Reconcile reads the download rows once rather than twice, and download() no
longer repeats the check enqueueIfAbsent makes under the lock.
- The artwork step no longer holds up the media: the song goes to the platform
first and the art is fetched afterwards, in the same foreground window. A
stalled artwork server can no longer delay a download, and a process death
during the fetch can no longer lose the download entirely.
- Resuming a queued download after a restart fetches its artwork now too; only a
fresh download did before.
- The artwork budget is a minute rather than ten seconds, which costs nothing but
a live coroutine now that nothing waits on it for the media.
- OkHttpArtworkFetcher derives a client with a per-call timeout. A blocking
OkHttp call is not interrupted by a coroutine timeout, so without it the outer
budget bounded our wait but not the socket.
Second block of #65.
- ArtworkStore keeps the album's and the artist's cover art, thumbnail and full,
in downloads/<sourceId>/art, keyed by the cache key the rest of the app already
uses for Coil and for palettes. Keys are base64url encoded because cover art ids
come from the server and must not reach outside the art directory.
- Downloading a song fetches that artwork first, in our process and while the app
is in the foreground, under a timeout, and hands the media to the platform
afterwards whether or not the artwork arrived. A song whose album or artist has
no art, or whose art fetch fails, still downloads.
- SourceRepository.coverArt resolves a stored file before the server, at the one
point every CoverArt, the notification bitmap loader and the palette prefetch go
through, so downloaded artwork works without the network.
- The sweep prunes art that no download references, so removing the last download
of an album drops its art while art another download still needs stays.
`QueueViewModelTest` read `viewModel.rows` (a `mutableStateListOf` mutated
on the main dispatcher) from the test thread, racing the view model: CI saw
`ConcurrentModificationException` in `removingWhileShuffledReconcilesTheRows`
and a torn read in `removingATrackUnderTheBlockReloadsRows`. Copy the rows
on the dispatcher via a `rowsSnapshot()` helper before asserting.
It measured pixel bounds to assert which line the indicator sits on and which
side of the artist name it lands. That is a presentational choice with exactly
one place in the code that makes it, so the test could only fail when someone
deliberately changed that line, and would break on unrelated layout changes.
The state-to-indicator tests and the accessibility descriptions stay.
The glyph and the progress ring sat in the title row after the star, spaced 8dp;
they now lead the subtitle row, 4dp from the artist name, which is where the
artist line has room for them. `theDownloadedIndicatorSitsOnTheArtistLineToItsLeft`
measures the unmerged bounds and fails if the indicator moves back up to the
title line or to the artist's right.
`single { SystemDownloadEngine(...) }` registers the concrete class, but
DownloadRepository asks for the DownloadEngine interface, so resolving
PlaybackController threw NoDefinitionFoundException and the app died on launch
with a Koin InstanceCreationException. It is now `single<DownloadEngine>`, the
way the other interface singletons in the module are declared.
Nothing in the suite ever started Koin — the repository tests build the
repository directly with a fake — which is why a launch-crashing binding got
through. AppModuleTest now resolves the playback graph from the real module, and
it failed with the same NoDefinitionFoundException before this fix.
Second review round on the downloads block:
- `loaded` was completed upstream of the state flow that publishes the map, so
`awaitLoaded()` could return while the map was still the initial empty one and
the first window after a cold start would stream a track that is already on
disk. The collector now assigns the map before completing, and the map is a
plain state flow instead of a lazy eager `stateIn` with a bare touch to start
it.
- A failed download now deletes its partial file, in both the engine-failure
and the recovery paths. A failed row used to protect its own partial bytes
from the sweep, and the UI offers no delete for it.
- A queued row is left queued when the download URL cannot be resolved yet,
instead of being failed. At launch the repository's reconcile can run before
the source repository has built its client, and failing there turned a
resumable download into a dead one.
- The poller only considers the active source's rows, so a queued row belonging
to another source no longer keeps it polling the database every second.
- `SourceRepository.downloadUri` takes the expected source id and returns null
unless it is the active one, so a download cannot be enqueued against the
previous server's URL during a source switch.
- Tests: a stream with no content length still gets its declared length through
the production data source factory (the http half of the chain, which the
first round left uncovered), and an engine failure is asserted to leave no
partial file. The engine's status mapping now runs as a plain unit test.
Review findings on the first downloads block:
- ExoPlayer could not open a downloaded file at all: the service's data source
factory was OkHttp only, so a file:// URI went to the HTTP stack and failed
with a malformed URL. Wrap it in DefaultDataSource so the file, asset and
content schemes are resolved before the stream chain, and cover it with a
test that reads a local file through the production factory.
- Reconcile no longer restarts a failed download, or a running one whose engine
entry vanished: only a queued row is started again, and the rest become failed
so the explicit Retry in the UI means something.
- song_downloads also references songs with CASCADE now, so a song pruned from
the mirror takes its download row with it and the sweep can reclaim the file.
- The mutating methods run on the injected dispatcher rather than the caller's,
which was the main thread for every call site, and the repository exposes one
shared state flow instead of a second map plus five subscriptions.
- Playback stops when the active source disappears, including deleting the last
source while a downloaded track is playing.
- Dropped the unused request mime type, the queued-at column nothing read, the
metered constructor parameter nobody varied, and a dead index guard.
- The running indicator is asserted positively, the engine's status and reason
mapping has a test, and the failure paths use the fake's failure and forget
helpers.
austinried