End-to-end DJ-mix command. Loads the collection, resolves its natural
in-order track list, dedupes near-duplicate versions, samples up to
sample_size tracks under the per-album cap, replaces the queue, and
starts playback from index 0. Stamps queue_source_collection and bumps
the play counter (mirroring v2_enqueue_collection's replace mode), but
does not modify the collection's persistent play_mode.
Returns ShuffleTracksResult { requestedCount, actualCount } so the
frontend can show the trimmed-mix toast when the cap or dedup shrinks
the pool.
ThreadRng is not Send, so dedup + sample run inside a sync scope that
ends before the bridge awaits.
cc
Resolves the collection's natural in-order track list and returns the
count of distinct songs after similarity-based deduplication. Used by
the upcoming track-mix modal to size the slider of pickable queue
sizes. Auth gated; deterministic output.
cc
Round-robin sampling that picks up to `requested` tracks from a
deduped track list. Each album bucket is shuffled in place and
truncated to cap = max(2, ceil(requested * 0.30)); rounds permute
the album order so cross-album distribution is randomized too. May
return fewer than `requested` when the per-album cap and bucket
sizes do not add up — callers report this gap to the user.
7 new tests (37 total) including a 200-seed statistical test
confirming the cap is never exceeded.
cc
- unique_track_count returns the count of distinct songs after grouping
near-duplicate versions; deterministic, no RNG.
- dedup_by_similarity collapses each duplicate group to one survivor
picked at random via the supplied RNG. Output preserves input order
of survivors.
Both share a single bucketed Union-Find pass: tracks are first grouped
by exact normalized artist, then unioned within each artist bucket
when their normalized titles score >= 0.80 on token_set_ratio. Cross-
artist comparisons are skipped to avoid false positives like "Yesterday"
by different artists.
12 new tests (30 total in this module) including a 200-seed statistical
test confirming every duplicate version can be selected as the winner.
cc
Pure-function helpers used by the upcoming track-mix sampler:
- normalize_title strips parens/brackets, dash suffixes, feat patterns,
punctuation, and diacritics.
- normalize_artist keeps parens (band suffixes carry meaning) and just
lowercases + strips diacritics.
- token_set_ratio implements the RapidFuzz-style token-set similarity in
[0.0, 1.0] using strsim's normalized Levenshtein.
Promotes strsim from transitive to direct dep. 18 unit tests cover the
core normalization and similarity behaviors.
cc
Add audio_cache_l1_max_bytes to MemoryProfile and wire it into
AppState's AudioCache constructor. Normal hosts keep the historical
400 MB cap; LowMemory hosts drop to 50 MB.
Without this, a Pi 3B (1 GB total) starts with the L1 cache cap set
at ~40 % of physical RAM. Even with the Feature #3 prefetch caps
already in place, the runtime watchdog from the previous commit
ends up firing more often than necessary because nothing prevents
a single L1 fill cycle from claiming 400 MB. Capping at 50 MB caps
how much the cache can grow before either: (a) eviction picks up
naturally, or (b) the watchdog steps in if the rest of the system
is also under pressure.
Pure refactor in lib.rs — replaces a hard-coded 400 * 1024 * 1024
with a profile field read. The default-constructed AudioCache
fallback (when disk-spillover init fails) keeps its own crate
default; no need to thread the profile there since the failure
path is rare.
Two tests added: a presence-bound assertion in the existing Pi 3B
case, and an explicit comparison that LowMemory's cap is strictly
smaller than Normal's.
cc
Tauri setup() now spawns a tokio task that polls
qbz_core::system_capabilities::read_memory_pressure() every 10 s
when the resolved MemoryProfile classifies the host as LowMemory.
On Normal-class machines the task is never spawned, so polling cost
is zero where it isn't needed.
Behavior:
- pressure.is_critical (< 5 % MemAvailable) triggers
audio_cache.clear() once per critical episode and logs at warn.
Both L1 (in-memory tracks) and L2 (disk spillover) are flushed —
the disk hit on subsequent playback is far cheaper than letting
the OS swap-thrash through the audio thread.
- pressure.is_low (< 15 %) is logged once at info as an early
warning but does not evict.
- Recovery (back above 15 %) logs once and arms the next critical
episode.
The was_critical edge guard prevents log spam when pressure
oscillates near the threshold during heavy prefetch.
For codehd7's Pi 3B scenario (issue #331): when QBZ pushes
MemAvailable below ~47 MB (5 % of 938 MB), the watchdog flushes
the cache, frees memory back to the kernel, and the network stack
recovers from the swap thrash that was killing playback.
cc
StreamingConfig::from_speed_mbps inflates the initial buffer to 2 MB
on slow connections (< 1 MB/s) to absorb network jitter. On a healthy
desktop that's fine, but on a memory-constrained host the slow link
is itself a symptom: when QBZ on a Pi 3B starts allocating ~300 MB
of Hi-Res prefetch the kernel swaps, networking stalls, the speed
detector reports < 1 MB/s, and the buffer code reacts by allocating
ANOTHER 2 MB — exactly the wrong direction.
Add a process-wide AtomicUsize cap (defaults to usize::MAX, i.e.
no-op) plus a small public set_max_initial_buffer_bytes setter.
Refactor from_speed_mbps to clamp through the cap; extract the
speed-to-bytes ladder into a pure raw_initial_buffer_for_speed
helper plus a from_speed_mbps_with_cap variant that exposes the
clamping for unit tests without touching global state.
src-tauri/lib.rs's run() wires the cap into qbz-player from
qbz-core's MemoryProfile right after logging the profile. Effective
on Pi 3B (issue #331): all StreamingConfig::from_speed_mbps results
pin to 256 KB regardless of detected speed.
Re-export max_initial_buffer_bytes / set_max_initial_buffer_bytes
from qbz_player::player so external callers don't need to reach
into the private streaming_source module.
Adds 5 unit tests: ladder behavior, pass-through when cap is loose,
clamp at the LowMemory cap (the issue #331 case), no-op at the
Normal cap, and that max_buffer_bytes stays unchanged.
cc
Wire qbz-core's detected MemoryProfile into the V2 prefetch loop so
memory-constrained hosts (Pi 3B and similar, < 2 GB RAM) cap their
in-memory footprint:
- V2_PREFETCH_COUNT becomes a function reading profile.prefetch_count
(5 on Normal, 1 on LowMemory). Reading it once per loop call is
cheap — OnceLock-cached behind the scenes.
- V2_MAX_CONCURRENT_PREFETCH likewise becomes a function; the
semaphore lazy_static initializes from it once at process start.
- Effective prefetch quality gets a low-memory cap before the
hardware-rate ladder runs. When profile.allow_hires_prefetch is
false we drop UltraHiRes/HiRes to Lossless so each cached track
stays under ~15 MB instead of ~60 MB. The hardware downgrade
still runs after, so a low-memory + 48 kHz-only DAC still gets
the right rate.
- lib.rs forces an early call to memory_profile() in run() so the
resolved profile is logged at boot rather than at first prefetch.
For codehd7's reported case (Pi 3B, 1 GB, HifiBerry Digi2 Pro,
issue #331): the active profile becomes prefetch_count=1,
max_concurrent_prefetch=1, allow_hires_prefetch=false. Cached
prefetch footprint drops from ~5 × 60 MB = 300 MB to a single
~15 MB Lossless track — well within the headroom that 1 GB minus
WebView (~150 MB) leaves.
cc
libasound's default error handler prints directly to stderr. During
device enumeration probes, PCMs that wrap dmix or surround routes emit
benign errors ("unable to open slave", "no matching channel map") that
cluttered the app log without indicating a real problem.
Install a process-wide snd_lib_error_set_handler shim from
AudioBackendFactory::create_backend (idempotent via std::sync::Once)
that forwards the file/line/function/err/fmt tuple to log::debug. The
messages stay available with RUST_LOG=debug and disappear during
normal runs.
Adds alsa-sys 0.3 as a direct Linux-only dependency on qbz-audio
(previously transitive via alsa/cpal). The handler is declared
non-variadic and transmuted into libasound's variadic function pointer;
ignoring the unused varargs is ABI-compatible on every platform we
support.
cc
Fixes GHSA-cq8v-f236-94qc (low: rand unsound with custom logger using
rand::rng()). Transitively bumped across both src-tauri and crates
workspaces via cargo update.
cc
Mixtapes & Collections release. Bumps all version references
(Cargo.toml, tauri.conf.json, package.json, snapcraft.yaml, both
Cargo.lock, package-lock.json), adds the 1.2.8 entry to the Flatpak
metainfo, and sets the About modal codename to "My QBZ".
cc
Offline-cached Qobuz tracks should appear in the Local Library on a
fresh install — the whole point of caching is to have them available
alongside local files. Users who set the preference explicitly (on or
off) are untouched: the INSERT OR IGNORE in DownloadSettingsStore
only writes the default row when no row exists yet.
Three coordinated changes inside download_settings.rs:
* DownloadSettings::default().show_in_library: false → true
* Schema DEFAULT on the column: 0 → 1 (covers rows added post-
migration from code paths that skip the explicit INSERT)
* The INSERT OR IGNORE seed row now interpolates
default_settings.show_in_library instead of hardcoding 0, so the
code default and the SQL default stay in sync.
cc
Plex thumbnail URLs carry X-Plex-Token (and sometimes
X-Plex-Client-Identifier) as a query parameter. Those rotate each
time a token refresh lands, so the *exact same* /library/metadata/…/
thumb resource hashes to a different cache key between sessions, and
the backend dutifully re-downloads and re-caches the same bytes under
the new hash. On a collection that's mostly Plex (the user's 67-item
case), that means every open re-pulls 67 images over the network —
which is indistinguishable from "caching isn't working".
Normalize the cache key on both sides, keeping the original URL for
the actual authenticated fetch:
* src-tauri: cache_key_for() strips X-Plex-Token and
X-Plex-Client-Identifier before hashing. The `url` passed to
reqwest is still the original with token, so the GET authenticates
correctly; only the on-disk cache key is stable across sessions.
* frontend: same normalization in imageCacheService.cacheKeyFor()
covers the in-memory resolvedUrls map, pendingRequests dedup,
getResolvedIfCached sync hit, and preloadImages "skip if already
cached" short-circuit.
Non-Plex URLs are untouched — no '?' means no normalization work.
Qobuz's static.qobuz.com CDN URLs don't carry session params, so they
already cache correctly and this is a no-op for them.
cc
Final piece of the Plex-in-playlist plumbing. PlaylistDetailView now
loads the (ratingKey, position) pairs from playlist_plex_tracks and
hydrates metadata from plex_cache_tracks, then interleaves the results
with Qobuz and local tracks by playlist_position so the ordering the
user sees matches what got stored when they added each track.
Backend:
* plex_cache_get_tracks_by_keys(rating_keys) in plex/mod.rs returns
PlexTrack metadata for a set of ratingKey strings, skipping keys
the cache doesn't know about.
* v2_plex_cache_get_tracks_by_keys Tauri command exposing it,
registered in lib.rs.
Frontend:
* New PlaylistPlexTrack interface + plexTracks state alongside the
existing localTracks pipeline.
* loadPlexTracks() does the two-step hydrate (positions, then meta).
Called in parallel with loadLocalTracks and loadPlaylist.
* plexTrackToDisplay() converts to DisplayTrack. The frontend id is
derived from the ratingKey via a djb2 hash offset into a
PLEX_DISPLAY_ID_OFFSET-anchored negative range — stable and never
collides with local (small negative) or Qobuz (positive) ids.
* displayTracks interleaves local + plex + Qobuz tracks by position.
Fast path unchanged when neither source has content.
* hasLocalTracks, totalTrackCount, totalDuration, and the sidebar's
hasLocalContent indicator all pick up Plex rows automatically
because the backend's get_all_playlist_local_track_counts already
merged the plex count in an earlier commit.
Missing plex cache entries (purged, never hydrated) are silently
dropped instead of rendering placeholder rows — the playlist just
shows fewer Plex tracks until the next Plex cache refresh.
Fixes the bug: Plex tracks can now be added from LocalLibrary and
they show up in the playlist detail view, completely decoupled from
the Qobuz playlist API.
cc
Wires the qbz-library methods added in the previous commit to Tauri
invoke handlers:
* v2_playlist_add_plex_track(playlistId, ratingKey, position)
* v2_playlist_remove_plex_track(playlistId, ratingKey)
* v2_playlist_get_plex_tracks_with_position(playlistId)
All three lock LibraryState::db, mirror the existing local-track
commands, and go nowhere near the Qobuz API surface — no chance of a
Plex ratingKey being forwarded to /playlist/addTracks by mistake.
Registered in lib.rs invoke_handler next to v2_playlist_add_local_track.
cc
Plex has the same cast routing problem as local library: the Qobuz-only
play_track commands can't resolve Plex ratingKeys. Add dedicated commands
that fetch bytes through plex_resolve_track_media (existing), proxy them
through the cast media server, and load the URL on the renderer.
Proxying through the media server (rather than giving the renderer a
direct Plex URL) keeps the user's Plex token in-process. Renderers would
need the token to fetch from Plex directly and that's not a surface we
want to expose.
Content-type fallback to audio/mpeg when Plex omits it (transcode path),
with a plex_key_to_u64 namespaced hash (high bit set) so Plex ratingKeys
can't collide with Qobuz/library ids in the MediaServer keyspace.
cc
Companion to v2_dlna_play_local_track. Same bug, same fix applied to the
Chromecast protocol branch: reads library row + file, infers content_type,
registers bytes with the shared media server, loads the URL on the
Chromecast receiver. Reuses the local_track_content_type helper.
User flagged this on the #332 thread — same Qobuz-only assumption in the
Chromecast play command as in the DLNA one.
cc
Mirrors v2_dlna_play_track but for library row ids. Reads the file from
disk, infers content_type from AudioFormat (with extension fallback for
Unknown), registers bytes with the DLNA media server, sends AVTransport
URI + Play to the renderer.
Without this command the cast path had no way to send local audio to a
DLNA device — v2_dlna_play_track resolves the id as a Qobuz track and
fails (404 in offline sessions, wrong-song match in online sessions where
the library row id happens to collide with a real Qobuz track id). Issue
#332 reported the symptom as "DLNA not operational"; logs showed the
renderer connected successfully but never received a SOAP SetAVTransportURI.
Frontend wiring in the cast dispatch + playbackService guard removal lands
in follow-up commits.
cc
Backend:
- New v2_enqueue_collection_item(collectionId, position, mode) routes a
single collection item through the existing ProdItemResolver — same
path v2_enqueue_collection uses, scoped to one position. Supports
replace / play_next / append modes. Unlocks per-album Play for local
and Plex items from the Collection detail view (previously only Qobuz
had a working fast-path there).
- enqueue.rs: resolve_local_album now detects plex: prefixes and routes
to plex_cache_get_album_tracks instead of querying local_tracks, so
Collections with Plex albums actually enqueue end-to-end.
- repo.rs: list_collections now hydrates items for every collection so
the item count in CollectionsView / MixtapesView is real, not 0.
- mixtapeCollectionsStore.ts: bindings updated for new + changed
commands (add_item with allow_duplicate, item_exists, custom cover
upload/remove).
cc
Lets users pick any image (png/jpg/jpeg/webp) as a Mixtape's or
Collection's cover. Mirrors v2_library_set_custom_album_cover:
source → artwork cache dir → resize to 1000x1000 via image crate →
store path in mixtape_collections.custom_artwork_path.
- v2_mixtape_upload_custom_cover(id, source_path): validates extension,
copies + resizes, stamps the resulting path, deletes the previous
cover file after the DB write succeeds.
- v2_mixtape_remove_custom_cover(id): nulls the column + deletes the
on-disk file.
- Both registered in lib.rs.
- Frontend store: uploadCustomCover / removeCustomCover wrappers that
also patch collectionsStore optimistically so UIs update without a
full reload.
Uses mixtape::repo::get_custom_artwork (added in the previous commit)
to find the previous path before overwriting.
cc
QueueManager::clear previously always preserved the current track;
plus bridge.play_index only mutates queue state, never triggers
audio. Result: when the user hit Clear Queue with nothing playing,
a stale track stayed in NOW PLAYING across app restarts.
- QueueManager::clear(keep_current: bool) — existing callers pin to
true (legacy behavior). New test covers the false path.
- QbzCore::clear_queue / CoreBridge::clear_queue take the flag through.
- v2_clear_queue accepts an optional includeCurrent flag from the
frontend; when true, calls bridge.stop() first so the audio pipeline
actually stops, then wipes the current slot too.
- qconnect sync path + qbzd daemon pin keep_current=true (no behavior
change).
cc
Plex exposes album release year (parentYear / parentOriginallyAvailableAt)
and genre (inline attr or nested <Genre tag>). The parser dropped both;
LocalLibrary and DiscographyBuilder could never show a Plex album's year
even though the metadata was available upstream.
- Extend PlexTrack + PlexCachedAlbum structs with year/genre.
- Parse both in parse_track_block, including the ISO-date fallback.
- ALTER TABLE plex_cache_tracks ADD COLUMN year INTEGER / genre TEXT
(idempotent).
- INSERT + SELECT paths + the album aggregation carry the new columns.
cc
list_collections only read the mixtape_collections row and left
items=[], so every consumer (sidebar, Collections/Mixtapes grid,
AddToMixtapeModal) saw zero-item counts for every collection. Load
items per collection at the end of the listing, matching what
get_collection already does.
cc
In-memory Option<String> tracks which Mixtape/Collection the current
queue was built from. Set by v2_enqueue_collection, cleared by any
non-Mixtape queue replacement (set_queue / clear_queue). Append-style
ops (add/add_next/bulk) preserve the context.
Persistence across restart is deferred — the session_queue_state
table does not exist yet; when it arrives, the ALTER in
mixtape::schema will add the source_collection_id column automatically,
and this state just needs a read+write on the appropriate bootstrap
and commit points.
cc
Two commands that read the current queue + index + elapsed time from
the runtime and call next_item_index / previous_item_index helpers
(already tested in mixtape::enqueue) to compute the target seek
position. Boundary detection uses source_item_id_hint with album_id
fallback, so these also work for queues built via "play album"
outside the Mixtape flow.
Commands are registered and invocable from Tauri but not wired to
any frontend control — per spec §4.5, NowPlayingBar changes are
deferred to a future spec.
cc
Polymorphic resolver handles album / track / playlist x qobuz / local
variants; each maps to the canonical QueueTrack the existing "play
album" paths use. Album shuffle applied at enqueue time, tracks
within each item stay in their native order. source_item_id_hint
stamped centrally so skip-to-item boundary detection (next task)
works consistently.
Includes next_item_index / previous_item_index helpers with unit
tests for both the exact-hint and album_id-fallback cases.
local_track_to_queue_track converts LocalTrack sample_rate from Hz
to kHz and wraps artwork_path as a file:// URL.
cc
Twelve v2_* commands wrap mixtape::repo with the standard app state
pattern used by commands_v2/playlists.rs. Enum values travel as
strings across the Tauri boundary; camelCase for frontend, snake_case
for Rust.
Also adds LibraryDatabase::with_connection_mut for the reorder command,
which needs a transaction (&mut Connection).
cc
Full surface for create/get/list/rename/description/play_mode/kind/
delete/touch_play on collections and add/list/remove/reorder on items.
Dedup on exact (source, source_item_id) within a collection. Artist
collection kind is locked (cannot be converted to/from). Cascade
delete covers items. Position compaction on remove; atomic transaction
on reorder.
Covered by 10 unit tests using in-memory SQLite.
cc
Two new tables (mixtape_collections, mixtape_collection_items) with
CASCADE on delete and indexes for the two common query patterns. One
additive column on session_queue_state for queue-source tracking.
Idempotent (CREATE TABLE IF NOT EXISTS, ALTER only when missing),
safe on every app start. The ALTER is skipped gracefully if
session_queue_state hasn't been created yet.
Adds LibraryDatabase::with_connection() helper (qbz-library) so
src-tauri can run external DDL on the shared Connection without
exposing it as a field. Migration is wired in session_lifecycle
immediately after library.init_at() in both the online and offline
activation paths.
cc
Reserved for upcoming Mixtape/Collection enqueue paths to mark item
boundaries. Existing enqueue paths populate it with the track's
album_id as a safe fallback so skip-to-item commands behave
correctly even for queues built via "play album" today.
cc
Linux Secret Service (GNOME Keyring / KWallet) can block indefinitely on
an unlock dialog when the user's login keyring is out of sync with their
session password. Each credential touch (load at bootstrap, save after
login, save-credentials on password path) used to produce its own dialog,
so every login triggered 3-4 prompts the user had to dismiss before the
app could continue (issue #329).
Make credentials/mod.rs defensive:
- Every keyring operation runs on a worker thread with a 2.5 s wall-clock
timeout (run_with_keyring_timeout). If the user is staring at an
unlock dialog they can't satisfy, control flow continues after the
timeout while the worker stays blocked in the background.
- A process-wide atomic (KEYRING_STATE) latches to BROKEN on the first
failure or timeout. Subsequent keyring_get / keyring_set /
keyring_delete calls short-circuit and the app falls back to the
encrypted file path for the rest of the session. One prompt max per
process. NoEntry does not count as a failure.
- The encrypted file in the config directory is the single source of
truth for every credential. save_qobuz_credentials already worked this
way; save_oauth_token was inverted from keyring-first to file-first so
the keyring is no longer on the success path of OAuth login.
- Load paths try the keyring first (cheap when it works) and fall back
to the file; the obsolete "migrate file to keyring" branches are gone.
Public API is unchanged. All existing callers compile unmodified. Existing
encrypted files and keyring entries remain readable; no file-format
changes.
qbz-secrets (offline-cache master key) is deliberately untouched — adding
a KDF fallback there without a migration plan would make existing
BACKEND_MARKER_KEYRING blobs undecipherable. Filed as Phase 2 in the
analysis doc at qbz-nix-docs/bugfix/2026-04-18-issue-329-keyring-unlock-dialog.md.
Fixes#329.
cc
User feedback from live testing: even with gapless working correctly,
the cold first-play of an offline CMAF bundle takes ~500ms to decrypt
on AES-NI hardware, 5-6s on the software fallback. No visible
indicator of what was happening — UX felt like the app was stalling.
Solution: when a track is being decrypted, replace its play glyph in
the track row with an animated padlock (lucide-svelte Lock icon +
CSS keyframe shake/scale). Tells the user "we're unlocking your
content", which is literally what's happening.
Implementation:
- Backend: new helper `load_cmaf_bundle_with_ui_events` wraps the
spawn_blocking decrypt with two Tauri events:
- `offline:unlock_start { trackId }`
- `offline:unlock_end { trackId, success }`
The helper takes two ids: `display_track_id` (what the frontend
knows this track as — Qobuz id or library row id depending on
flow) and `cmaf_track_id` (the Qobuz id used to key the bundle
inside load_cmaf_bundle itself). Same track shows the animation
whether clicked from Qobuz view or Local Library.
- All three v2 decrypt call sites now route through the helper:
v2_play_track offline branch, v2_play_next_gapless (both the
Qobuz-id and Library-qobuz_download branches), and
v2_library_play_track.
- Frontend: new store `unlockingStore` maintains a `Set<trackId>`
updated from the events, with subscribe/isUnlocking API. Started
from +page.svelte alongside the existing player/queue pollers.
- TrackRow.svelte: new `isUnlocking` reactive state derived from
the store, a new branch in the play-number template that renders
the Lock icon with `lock-shake` animation, scoped CSS keyframes
(qbz-unlocking — slight rotate + scale + opacity pulse at 1.2s
loop, accent-primary colour, cursor: progress).
The animation fires on cold decrypt only; warm L1/L2 hits bypass the
decrypt entirely and don't emit the events, so the user never sees
the padlock for a cached track — correct behavior.
cc
Qobuz-cached tracks (both v1 legacy and new v2 bundles) weren't
showing album art in Local Library. Root cause was a pair of bugs:
1. Legacy v1 path DID download cover.jpg to the organized album
directory (via save_album_artwork) but passed None as artwork_path
when inserting into the library index. So the file existed on
disk but the DB row didn't point at it — the UI had no way to
find it. Pre-existing bug, not introduced by the v2 work.
2. v2 CMAF bundle path didn't save artwork at all. Worked fine for
encrypted playback but the tracks-cmaf/<id>/ directory had no
cover.jpg for the UI to fall back to.
Fix: in both paths, call save_album_artwork next to the cached
content (organized FLAC parent for v1, tracks-cmaf/<id>/ for v2),
then pass the cover.jpg path to insert_qobuz_cached_track_with_
grouping's artwork_path parameter. Library rows now carry the
on-disk path to the artwork and the UI picks it up just like for
user-owned local FLACs.
New downloads will have artwork immediately. Tracks already in the
cache from before this commit will remain artwork-less until
re-downloaded — acceptable since the user can just re-trigger
download per track if they want the art back, and the 30-day
subscription wipe refreshes everything eventually anyway.
cc
Library auto-advance had a ~6s gap between two Qobuz-cached v2 tracks.
Tracks within Qobuz flow had gapless working because those pass the
Qobuz track id to v2_play_next_gapless and the offline cache is keyed
by that id. From Local Library the frontend passes the LIBRARY row id
(small autoincrement, e.g. 542), so:
- L1 keyed by library row id → miss (prefetches key by Qobuz id)
- L2 keyed by library row id → miss
- offline cache get_cmaf_bundle(542) → miss (cache is keyed by Qobuz id)
- library fallback: file_path for qobuz_download v2 tracks is a
directory, std::fs::read blows up → gapless returns Ok(false)
Result: gapless declined, track ended, frontend called play_track as
normal transition, which went through v2_library_play_track's full
decrypt path from cold → 5-6s silence.
Fix: in the library-fallback branch of v2_play_next_gapless, detect
source='qobuz_download' and re-query the offline cache by the row's
qobuz_track_id. If a v2 bundle exists, decrypt it on the blocking
pool and feed to play_next using the LIBRARY row id (so the player
keys match what the frontend is waiting for — same pattern as the
earlier library track-id fix). Warm L1 keyed by library row id too,
so next transition between tracks in this library view hits memory
instead of re-decrypting.
This leaves the v1 path working (plain FLAC at segments_path) and
the regular user-owned local FLAC path unchanged.
cc
Live testing from Local Library with a v2 CMAF-cached Qobuz track
showed a cascade of UI breakage: seekbar never appeared, track never
auto-advanced, queue-manager fired next_track every 1s for 30+
iterations in a loop, currently-playing song card stayed on the first
track. Audio was playing fine; the UI just couldn't see it.
Root cause: my v2_library_play_track CMAF path was passing the Qobuz
track id (from the library row's qobuz_track_id field) to
player.play_data. Every other piece of the UI state machine — queue
sync, seekbar progress events, auto-advance trigger, now-playing card
— keys off whatever id the frontend invoked with, which is the
library row id (e.g. 542), not the Qobuz id (e.g. 95787326). So the
player was emitting updates about "playing 95787326" while the
frontend was waiting for updates about 542 that never came.
Fix: play_data now takes track_id (library row id) on both the v2
CMAF decrypt branch and the v1 plain-FLAC branch. The offline-cache
DB lookup still uses qobuz_track_id (that DB is keyed by Qobuz id),
the player gets the library id. AudioCache insert also keyed by
library id so Library-side replay hits L1 correctly.
cc
Live testing showed gapless between two offline-cached v2 tracks still
had a 6s gap even with the L1 warming + spawn_blocking + track_id fixes.
Root cause: v2_play_next_gapless checked the offline cache FIRST, which
meant for tracks that were BOTH on disk AND already in L1 (prefetched
from the CDN into memory), we took the slow offline decrypt path even
though instant bytes were sitting in RAM.
Log evidence:
16:52:09 Prefetch put track 47683552 in L1 (61MB)
16:53:00 play_next_gapless for 47683552
16:53:06 [V2/GAPLESS] Track 47683552 from OFFLINE cache (CMAF v2)
(decrypted 58MB — took ~5s while L1 had the bytes)
16:53:06 Gapless: no engine, ignoring PlayNext
(engine dropped at 16:53:05 when old track ended)
The very next transition (47683552 → 47683553), where 47683553 happened
to be ONLY in L1 and not in the offline cache, worked perfectly:
16:55:58 [V2/GAPLESS] Track 47683553 from MEMORY cache (72MB) ← 60ms
16:56:02 Gapless transition: track 47683552 -> 47683553 (clean)
Fix: reorder tiers to L1 → L2 → offline → library. The offline v2
decrypt path is expensive precisely because we already warm L1 after
every decrypt — so the second time around (or whenever prefetch beat
us to it), L1 hit is instant. The offline tier is correctly reserved
for the "truly cold" case: track on disk but not in memory.
Also added a "warm L1 from L2" step so a disk-cache hit also populates
memory, matching the warm-L1-on-decrypt pattern the offline path
already did.
cc
Live testing showed v2 offline tracks had a 3-6s gap on gapless
transitions even after the track_id + spawn_blocking fixes, because
the decrypt itself takes 5-7 seconds for a typical 60-200MB FLAC
(disk read of encrypted segments.bin + AES-CTR per-frame decrypt is
slower than the track's natural tail). By the time load_cmaf_bundle
returns, the previous track has ended and the player's gapless engine
has been dropped.
Workable only second time: once a track's decrypted bytes are in L1
memory cache, subsequent accesses are ~60ms. That's what happened on
the second transition we observed — prefetched tracks hit L1 for
free, so gapless landed in time.
Fix: after successfully decrypting a v2 bundle, insert the resulting
plain-FLAC bytes into the AudioCache (L1) before handing them to the
player. Applies to three call sites:
- v2_play_track offline cache branch → warm on first play
- v2_play_next_gapless CMAF v2 branch → warm on gapless success
- v2_library_play_track CMAF v2 branch → warm on Library click
Next access to the same track (replay, scrub, gapless re-queue,
upcoming-track peek) is now a <100ms L1 hit regardless of the v2
cache_format the row has. The decrypt cost is paid once per track
per session, not once per access.
This still leaves the "first play of an offline track" taking 5-7s
before audio starts. That's a separate optimization — candidates are
(a) offline-track prefetch that pre-decrypts upcoming tracks in the
background, (b) storing decrypted FLAC in L2 disk cache so cold
restarts don't have to decrypt, (c) speeding up the decrypt itself.
cc
Two bugs the live testing surfaced.
1. v2_library_play_track was passing the library row id (autoincrement
small integer) to the offline-cache DB lookup. The offline cache is
keyed by the Qobuz track id (the big id like 95787327). So the
lookup always missed and my earlier "orphan" error fired for EVERY
v2 track clicked from Local Library. Fix: read qobuz_track_id from
the library row and use THAT for the offline-cache lookup.
2. load_cmaf_bundle is synchronous (std::fs::read + AES-CTR decrypt +
FLAC assemble). When called directly from an async fn on the tokio
executor, a 200MB offline track took ~22 seconds end-to-end during
concurrent prefetch activity — because the sync work starved the
executor thread, delaying the queue's auto-advance event, which
fired play_track as a fallback AFTER the old track stopped (which
killed the gapless engine). Net: 6-8 second audible gap between
two offline-cached tracks even though the bytes were on disk.
Fix: wrap load_cmaf_bundle in tokio::task::spawn_blocking in all
three call sites (v2_play_track, v2_play_next_gapless,
v2_library_play_track). The executor stays responsive, gapless
finishes before the engine drops, transition is clean.
Together these close the two loudest symptoms from the live session:
"Is a directory" when playing v2 from Library, and the seconds-long
gap between consecutive v2 tracks even with gapless enabled.
cc
Two symptoms with the same root cause — code paths that touched the
offline cache read file_path as "a plain FLAC file" and handed it
straight to std::fs::read + Symphonia. For v2 bundles the file_path is
the encrypted segments.bin, so both paths exploded.
1. v2_library_play_track: error "Is a directory (os error 21)" when
clicking a v2-cached track from Local Library.
Fix: for source='qobuz_download' rows, the offline-cache DB is the
authoritative source of file paths, not the library index (library
carries display metadata only). Query get_cmaf_bundle:
- cache_format=2 → load_cmaf_bundle (decrypt + play)
- cache_format=1 → read row.segments_path (which for v1 IS the FLAC)
- no row → clean error instead of reading the library directory
2. v2_play_next_gapless: "Symphonia probe failed: end of stream" with
~6 failed retries per track transition → ~6s gap between two v2
offline-cached tracks.
Fix: same cache_format branching as v2_play_track already does.
Reuse load_cmaf_bundle so the probe sees plain FLAC bytes, not
encrypted mp4 boxes.
The prefetch AES-CBC unpad error during concurrent downloads is the
other piece of today's gap story — session race in ensure_cmaf_session.
That's a separate fix in the next commit.
cc
The UI's offline-download progress ring listens to
'offline:caching_progress' events. The legacy StreamFetcher::fetch_to_file
emits those per HTTP chunk via app_handle. The new CMAF path went
silent because cmaf_download_raw lives in the qbz-qobuz crate with no
Tauri access — the user saw the completion toast but never saw the
ring fill.
Fix:
- Add an optional CmafProgressCallback to the qbz_qobuz::cmaf public
API. Typed as Arc<dyn Fn(CmafProgressUpdate) + Send + Sync>.
- New download_full_with_progress and download_raw_with_progress
variants; the callback-less versions delegate to them with None.
- fetch_all_segments fires the callback once per completed segment
with cumulative (k, n, bytes_this_segment). Callback is invoked on
the worker thread after the bytes land, before the rate-limit
cooldown.
- try_cmaf_offline_download passes a callback that emits
'offline:caching_progress' events in the exact shape the UI already
expects (trackId, progressPercent 0..100, bytesDownloaded,
totalBytes: null, status: 'downloading'). Also emits one
zero-progress event up front so the ring doesn't sit at "queued"
while the first segment fetches.
Percentage is computed as (segments_completed / n_segments) * 100,
rounded and clamped. It's segment-granular rather than byte-granular
so the ring advances in N discrete steps (typically 8-20 for a 4-min
track) rather than smoothly — matching what the CMAF pipeline can
actually observe, which is segment-complete events.
cc
Two loose ends after v2 landed:
1. Local Library integration. Qobuz-cached offline tracks already show
up in the Local Library view (via insert_qobuz_cached_track_with_
grouping with source='qobuz_download'). v2_library_play_track was
reading track.file_path and doing std::fs::read() — that works for
plain-FLAC v1 entries, but for v2 the path is the bundle directory,
not a playable file. Fixed:
- Extract the CMAF-load-and-decrypt logic into a shared module
(offline_cache::playback::load_cmaf_bundle) used by both
v2_play_track and v2_library_play_track.
- v2_library_play_track checks source='qobuz_download' +
cache_format=2 on the row and routes through the shared helper.
Legacy v1 entries and regular user local files unchanged.
2. purge_all_cached_files. Previously deleted individual file_path
entries only. For v2 the file_path points inside a track dir
(segments.bin) — removing just that file orphans init.mp4 and
manifest.json. Fixed:
- When the file_path looks like tracks-cmaf/<id>/segments.bin
(matched by the path shape, not a schema flag), remove the whole
enclosing <id>/ directory.
- Belt-and-suspenders: also delete the top-level tracks-cmaf/ dir
after the per-row pass, so orphaned bundles from corrupted rows
are swept up too.
Net: Local Library view preserves the "cached Qobuz tracks appear
here" UX, AND the 30-day grace wipe still removes every encrypted
segment + every library row in one atomic operation. No contradiction
between the two behaviors.
Also extends the compliance whitepaper with a section explaining the
Local Library integration so Franck's team can see that the wipe
contract isn't worked around by the UX shortcut.
cc
Matches Qobuz's own mobile app posture. The previous 3-day window was
both user-hostile (punishes anyone on a flaky network) and a compliance
red flag (shorter windows read as 'we're hiding something', longer
windows read as 'we're lenient for a reason'). 30 days is the exact
parity answer.
The useful protection against bulk-download-and-ghost attacks comes
from the CMAF-at-rest offline cache (v2 format), not from aggressive
revalidation. This grace window is a belt over the suspenders.
Changes:
- GRACE_PERIOD_SECS: 3 → 30 days in config/subscription_state.rs, with
a comment explaining the reasoning for future readers.
- Log message in session_lifecycle no longer says "3 days" (it quoted
a constant, now says "grace window").
- Disclaimer string in all 5 locales (en/es/de/fr/pt) updated from
"3 days" to "30 days" so the Offline Playback settings screen
reflects the actual behavior.
No behavior changes for users with an active subscription. Users who
were within the old 3-day window don't lose cache — the window
extended, not contracted.
cc
vicrodh