Send periodic player states only while connected and playing, keeping debounced changes and reconnect state delivery. Skip remote playback publication for download-only updates. Keep transport ping/pong liveness while paused and avoid a server ping after recent client traffic.
Negative controls observed four eliminated paused phone states per minute, three redundant download-driven playback events reduced to zero, and server ping suppression with client pings. All 104 scoped Dart tests and 114 Rust tests pass, with strict analysis, clippy and format checks.
Bump Library to v21 and History to v16, add album artist indexes, and remove six Library and three History indexes whose seeks are served by existing queue indexes. Keep the named file-path index for released readers and synchronize native schema checks.
Persist the iOS container migration marker in the database transaction, exclude current paths, and bound Library and CUE reads to 256 rows. Repeat opens read one marker row. Test rollback, path keys, bookmarked sources and downgrade/reupgrade compatibility.
Four alternating 15,000-row host write pairs reduced median batch time from 3364.773 ms to 2782.200 ms (17.3 percent). Native staging now takes the shared History schema version.
Compare link candidates with the permissions of a normally created stage so a changed cached file cannot alter a new song cover's mode. Keep the copy fallback when they differ.
The new regression fails before this fix (0604 instead of 0644) and passes afterward. All three affected Rust crate tests, clippy, formatting, and diff checks pass.
json! turns every value into to_value(&value), so building the delta
with json!({"scanned": tracks, ...}) serialized every scanned row into
a second copy (wrapping the rows in Value::Array inside json! copies
them too). The object is now built as a Map that takes the rows and the
deleted paths by value, in the same key order, so the JSON is
byte-identical.
Measured with a standalone serde_json 1.0.154 program (release, 20,000
rows shaped like scan rows, 15.2 MB of JSON): 35 to 37 ms and 62 MB of
extra peak memory -> under 0.1 ms and 0.1 MB, same output hash.
Every NAS tag read that found a picture rewrote the content-addressed
network_cover_<sha256>.image through write_atomic (sync_all, which is
F_FULLFSYNC on Apple), then listed and trimmed the folder, even when
the same file was already there, as it is for every further song of an
album. A file that still holds exactly the same bytes is now kept and
its mtime renewed, as the rewrite did, so the 32-file trim keeps it
among the newest; nothing is added, so the trim is skipped. A damaged
file under the name, or one a concurrent trim removed, is still
rewritten.
Measured on this Mac (release, 350 KB cover, 32 cached files): a repeat
call 4.9 to 8.9 ms -> 0.23 to 0.39 ms, most of it the SHA-256 that
names the file.
Every song got its own thumbnail file even when an album's songs embed
the same picture, so a scan wrote and kept one 150 to 430 KB copy per
song. The in-memory thumbnail cache, which already reuses the resize by
content, now also records the file that holds the bytes. The next song
with the same picture hard-links that file into its own name (link a
stage name, then rename), but only while the file still holds exactly
those bytes, is current for the regenerate marker and has the expected
permissions; otherwise, or when linking fails, it writes the bytes as
before. Every name, its bytes and every stored coverPath stay the same.
Songs of one album hold the record until the first file is published,
so concurrent scan workers link instead of writing copies.
A new name keeps its inode's old mtime, so the Library cover prune now
counts a multi-link file from the newer of mtime and ctime (link(2)
updates ctime): the one-minute guard for background producers still
holds. Files with one name, which is every file on Android and Windows,
keep the mtime rule. Deleting a name never touches the others.
Android keeps one file per song (content-addressed names there are a
later unit) and Windows keeps copies too.
Measured (release, scan_timing_with_many_thumbnails, 1,000 songs over
100 pictures): 1,000 names on 113 to 120 files, 341 MB -> 38 to 40 MB
on disk; the scan time is unchanged (5.5 s). The extra files come from
the 8 MB in-memory cache evicting a picture before a lagging worker
reaches it.
Rust's sync_all is F_FULLFSYNC on Apple targets, a drive cache flush
that every Library thumbnail paid twice (file and folder), and the
device-wide flush does not overlap across scan workers. Scan workers
now publish thumbnails with a plain fsync(2), which still hands the
whole file to the drive before the rename gives it its final name, and
the scan flushes the thumbnail folder once with sync_all when its
workers finish (a drop guard covers failure and cancellation).
Only scan workers use the batch. Single reads, CUE sheets, cover exports
and downloads keep the full sync, and other platforms keep the current
publication. A power loss inside the scan window can at worst leave an
empty thumbnail, which is now treated as missing and extracted again.
Measured on this Mac (release, ignored test
scan_timing_with_many_thumbnails, 348 KB thumbnails): 1,000 songs over
100 album pictures 9.3 s -> 5.5 s; 1,000 distinct pictures 42.4 s ->
41.0 s, where the serialized resizes dominate.
Bound queue waits and retain operation ownership through VM cleanup. Abandoned queued operations never execute. The short-deadline regression previously waited longer than one second; Rust workspace tests and strict clippy pass.
Recognize an advertised LAN origin only when the WebDAV response identifies the requested collection unambiguously. Keep browsing and playback on the saved connection so tunnel access and HTTPS trust continue to apply.
Validation: 76 Flutter tests and 7 Rust parser tests passed. Dart analyzer, Clippy with warnings denied, formatter checks, and git diff --check passed. Includes HTTP and HTTPS transport fixtures and rejected ambiguous origins; the reporter's NAS still needs a retest.
Read same-host WebDAV child paths advertised through another HTTP or HTTPS listener while retaining the configured transport origin. Accept successful property statuses without a reason phrase, and reject non-WebDAV documents instead of presenting empty folders.
Validation: 71 Flutter tests and 7 Rust parser tests passed. Dart analyzer, Clippy with warnings denied, formatter checks, and git diff --check passed. The reporter's NAS still needs a retest.
The daily cover check downloaded every stale Subsonic cover in full
and compared it with the cached file byte by byte, although the Link
desktop (and many servers) send an ETag and answer If-None-Match with
304. The first rescan of a day fetched all 500 album covers of a Link
library at original size.
The ETag of a downloaded cover is now kept beside it in
<cover>.etag together with the image length. A stale check sends
If-None-Match; a 304 keeps the path and touches the modification
time, like an unchanged download did, without reading a body. A 200
follows the existing rules (same bytes keep the path, changed bytes
get a new path behind .current) and stores the new ETag. A forced
refresh, a validator for other bytes, older caches and servers
without ETags behave as before.
The validator survives restarts. The Library cover prune (Dart and
Rust) keeps it with a referenced Subsonic cover, as it keeps the
.current pointer, and the song metadata cover cache still keeps 32
covers, each with its ETag, instead of counting ETags as covers.
Older builds ignore the files and their prune removes them.
A web cover copy in a catalog's coverPath outranked real pictures: an
album's cover is its first song with a cover file, and every id tries
cover files before embedded pictures, so an album mixing a song that
embeds its picture (or a later song with a cover file) with a web-only
song served the copy for every song.
Catalog lines now carry copies in their own webCoverPath field. The
server tries them only after the cover files and embedded pictures it
tried before, the song's own copy first and then its album's; an album
falls back to its first song with a copy only when no song has a cover
file or an embeddable format, and an artist to such an album only when
no other has a cover. Ids that had any cover serve exactly what they
served before, and a catalog without the field reads as before.
A failed read of an audio file is no longer remembered as a file
without a picture, for getCoverArt and linkEmbeddedPictures alike; only
a file that was read through counts.
The desktop should fetch a web cover only for a song whose file has no
picture of its own. linkEmbeddedPictures answers, for a list of paths,
whether each file embeds a picture getCoverArt would serve (null when it
is not a readable file). It reuses mornye_core's cover extraction and the
server's format table and remembered files without a picture, so a later
getCoverArt does not read them again. Works while the server is stopped.
The document scan slot test only compared the stem with the function that
computes it. A new test writes a thumbnail through the Rust scan's
cache_cover and expects library_cover_slot to find that file for the same
key, then checks that the newer of two current siblings wins.
The forced-scan test compared marker mtimes 20 ms apart, which can flake
on filesystems that store whole seconds; it now checks that each forced
scan rewrites the marker's contents. The padded ID3 cover test reuses the
shared backend and ID3 helpers.
The cover prune after a full scan treated a library_covers file as in use
only when library.cover_path named it. Playlists, favorites, the saved and
dismissed playback sessions, recent items, the download queue, listening
statistics and download history (including Link save offline copies) copy
those paths, so their covers broke after the next full scan.
The prune now also reads every such text column (read-only, in pages,
filtered with instr) and keeps the files they name, rebased onto today's
cover directory so file:// URIs and paths from an older iOS container
match. The names join the private native projection, so Rust still never
reads a live database and a failed read cancels the sweep.
A Subsonic cover kept by a reference also keeps its
subsonic_<hash>.image.current pointer, in Rust and in the Dart fallback;
before, the pointer was always deleted and the next scan fetched and
renamed the cover again.
Scan thumbnails were named by a 32-bit djb2 hash of "path|size|mtime" (or
the document key) and an existing file was reused without checking its
source, so two songs whose keys collided showed one cover. The Android
document scan had its own copy of the same hash.
Thumbnails are now named cover_<first 128 bits of SHA-256 of the key>, in
Rust for every scan and through library_cover_slot for the Kotlin document
scan, so reuse by name is safe. Rows that name older files keep working;
those files are not looked up again and are pruned once unreferenced.
A forced rescan can touch a .regenerate marker in the cover folder:
thumbnails older than it are extracted again under the same name instead
of being reused, and the newer of a .jpg and .png sibling wins.
The ETag and the resized copy of an embedded picture came from its
extracted copy in the temp cache: path, mtime and length. Once the cache
trimmed that copy (4000 files) or the OS purged the temp folder, the
picture was extracted again with a new mtime, so the same image got a new
ETag (clients downloaded it again) and a second resized copy that
stayed in the cache beside the first.
The stamp of an embedded picture is now the audio file's path, mtime and
length, which already named the extracted copy. Every id that resolves
to one image (album, artist or song id) still shares one ETag and one
resized copy. Cover files keep their stamp and their ETags.
getCoverArt resolved an id to a single source: a song's own cover file,
else its album's, else embedded pictures. When the listed file was gone
(a pruned or moved Library thumbnail, a stale export) it answered 404,
although the album's cover file or the picture embedded in the song
could still serve it. The phone then showed a placeholder.
The catalog now returns every fallback in order (the song's own file,
its album's file, then the songs that can embed a picture) and the
server serves the first one on disk. 404 remains for ids with none.
Responses, content types and ETags for covers that resolve today are
unchanged.
Albums, singles and playlists, online and in the Library, end with
"Suggested songs": up to ten songs like the ones above that the collection
does not hold yet, each with a preview, an add button and a click or tap
that downloads it. On a user playlist the add button adds the song there
and it leaves the suggestions; elsewhere it opens the playlist picker.
Phones list the songs as rows, desktop in the page's song table.
Suggestions come from a new optional extension method, getRelatedTracks,
which extensions declare with the relatedTracks capability: it takes up to
three of the collection's songs and returns songs in the shape of search
results. The extension most seeds came from is preferred so it can use
their IDs. Without such an extension, the first metadata provider is
searched by the collection's artists. A refresh button starts from other
songs, and loaded suggestions stay for ten minutes.
Publish all split tracks and their visibility rule atomically while preserving the original files and playlist rows. Keep the rule across rescans, restore originals when outputs leave the visible Library, and stabilize Android SAF identities and artwork caching.
Validation: 84 Flutter tests, 19 native Library tests, scoped analyzer, formatting checks, and Android Kotlin compilation passed.
A microphone button in the empty search field listens to the room and
names the song, on phones and desktop.
- The audio fingerprint and the Shazam request run in Rust behind a new
recognize_song data job, so every platform shares one path.
- Listening sends the audio so far at 4, 8 and 12 seconds and stops at
the first match; quiet or short clips are reported instead of sent.
- A match can be played from the library (by ISRC, then title and
artist), searched in Mornye, or opened in Shazam. Recent matches are
kept in a local history.
- Android asks for the microphone only when listening starts; iOS and
macOS gain the microphone usage description and entitlement. The Linux
package recommends pulseaudio-utils for capture.
Add an audio_features native data job that reads one or two decoded
mono windows and returns a compact, versioned vector of 42 numbers:
tempo and its confidence, loudness and its spread, brightness,
rolloff, flatness, onset rate and flux, low/mid/high band shares,
key, mode and key strength, and 13 MFCC means and spreads. Silent or
too short audio yields no features. The job reuses the hi-res FFT and
the AutoMix tempo detector and checks for cancellation as it goes.
Prevent stale transport and lyric updates, cancel abandoned requests, bound Link and Subsonic buffering, and preserve native audio lifecycle ownership. Add regression coverage for startup, search and transfer cleanup.
The desktop now answers OpenSubsonic getLyricsBySongId with the lyrics
a song's file embeds, else its .lrc sidecar, as line-level structured
lyrics, and advertises songLyrics. The entry also carries the stored
text as mornyeRaw, so the phone can keep word timing, voices and
backing vocals that the line-level shape drops.
mDNS answers on Wi-Fi carried every link-local IPv6 address of the
machine, including the loopback's fe80::1, which on most networks is
the router. Android before 14 keeps a single resolved address and could
pick it, leaving the phone with an unresolvable .local name, so finding
the computer failed and pairing needed a typed address. Loopback is not
advertised either; nobody on the network can use it.
Play on desktop no longer waits for the whole file. The phone posts
handoff/play once the bytes up to the resume point and a 512 KiB margin
are sent (all of them for an MP4 whose index follows the audio), naming
the upload in `arriving`; the rest uploads in the background, and later
songs wait for it.
The desktop serves an arriving file over loopback HTTP with Range
support, reading the part file as it grows, and the player opens that
URL through MusicPlayerRuntime.streamWhileArriving until the file is
whole. The queue entry keeps the final path, and the embedded cover is
read once the file lands. On Windows the Rust rename into place is
retried briefly, since an open reader blocks it.
Play on desktop used to skip songs only the phone has. The phone now
asks the desktop which songs it lacks, uploads the current one first
(PUT /mornye/v1/files/<id>.<ext>, streamed to <app support>/link/received
and pruned to 4 GiB) and starts the desktop there. Later songs follow in
the background and handoff/append adds them in order.
getCoverArt also serves covers embedded in the audio file, so downloads
whose cover is only in the file no longer show up on the phone without
one.
The silent test client answers queued pings on the closed connection,
and under load the reset can beat the close frame. The test still checks
the 1001 code when the frame is read and that the client is removed.
The app reported favorite artists as ids it computed with a looser
normalization than the catalog, so some never matched. getStarred now
sends artist names and the server maps them with its own rule; unstar
matches favorites by name as well.
Index the NDJSON catalog export into albums, artists, genres and a
case- and diacritic-insensitive search, with contract ids (al-/ar- +
16 hex of SHA-256). Answer the catalog endpoints in OpenSubsonic JSON
or Subsonic XML, stream and download files with single-range support,
transcode through ffmpeg when asked (killed when the client goes
away), resize and cache cover art, and forward playlists, stars,
scrobbles and the play queue to the app with song ids expanded both
ways.
Implements the Mornye Link server core (unit L1a) in mornye-link:
- tls: self-signed ECDSA P-256 certificate for mornye.local, generated
once into cert.pem/key.pem, pinned by its SHA-256 fingerprint
- store: server.json and devices.json, atomic private writes
- auth: Subsonic apiKey, salted token and password schemes, Bearer keys
- pairing: 6-digit PIN window with HMAC proof, attempt limit, expiry
and a per-address rate limit
- server/api: own Tokio runtime, port policy with persistence, dual-stack
TLS listener, /rest routing to the Subsonic layer, /mornye/v1 info,
pair, ws and forwarding with a request timeout
- ws: hello/welcome sessions, client list events, pushes, pings and
idle drop, newest connection per device wins
- mdns: _mornye._tcp advertisement, best effort
mornye-desktop gains the link* methods and bridges forwarded requests
to Dart as linkRequest inbound calls answered by linkReply.
Mornye Link lets the phone app use the desktop app over the local
network. This adds the mornye-link crate with the types its server and
Subsonic layers share, and the Dart protocol models, desktop request
router and phone client interface the feature units build on.
mornye.exe is a GUI-subsystem app, so spawning the ffmpeg.exe console
program without creation flags makes Windows open a visible console
window for every conversion. Pass CREATE_NO_WINDOW, matching what
dart:io's Process.start already does for the Dart-side FFmpeg calls.
Rust 1.99.1, Tokio 1.54.1, zip 9.0.5, libc 0.2.191 and
rustls-native-certs 0.8.5. scraper 0.29 moves to html5ever 0.40, so
the parser pin follows it. The Unicode 15 table pins and the vendored
rustix, cap-std and libusb1-sys sources stay as they are.
Extensions that require the new deferredPreview runtime feature return
mornye-preview:<extension>/<track> links instead of a sample URL. The
preview player resolves such a link through the extension's
resolvePreview function when the preview starts, so a search no longer
waits for one sample lookup per track.