results comes from asyncio.gather over to_search, so they're always equal length — strict=True asserts the invariant and satisfies ruff B905. Carried in with #896.
#3 priority mode is quality-agnostic again: search_with_fallback (the
priority/hybrid path — best_quality has its own search_all_sources) returned
the first source whose results met a target, so an mp3 with bitrate=None (slskd
omits it often) was deemed unsatisfied and deprioritised, changing which source
wins for users who never opted in. Restore "first source with tracks wins",
byte-for-byte; cross-source quality pooling stays in best_quality mode.
#4 metadata-less FLAC no longer over-claims a hi-res target: matches_target let
a FLAC with no sample_rate/bit_depth satisfy a 24-bit/192k target while a real
16/44 FLAC failed it, so unknown-spec files outranked and discarded genuine CD
FLAC under audiophile/hi-res profiles. An unconfirmable spec now fails the strict
tier and falls to the plain-flac bucket.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Three fixes from on-device testing of best-quality mode:
1. clear_completed_local no longer prunes terminal tasks that belong to a
STILL-ACTIVE batch (one with non-terminal work remaining). The 5-min
"Clean Completed Downloads" automation was yanking completed/failed/
unverified rows out of download_tasks mid-run — and failed/cancelled
aren't in library_history — so they only reappeared after the batch
ended. Now the whole active batch stays intact until it finishes.
2. search_all_sources runs every source CONCURRENTLY (asyncio.gather)
instead of sequentially, so the pool waits only for the slowest source
(e.g. usenet/Prowlarr) in parallel rather than summing all latencies.
3. The pool log now reports per-source contribution counts
(e.g. "usenet=0, hifi=11, soulseek=1") instead of just echoing the
chain, so a release-level source that returns nothing for a track-title
query is visible rather than appearing to have been searched.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Adds an opt-in search strategy toggle in the Quality Profile:
- priority (default): unchanged — first source in the hybrid chain that
meets a quality target wins.
- best_quality: pool candidates from EVERY source per query and download
them best→worst by actual audio quality; source order only breaks ties.
Implementation reuses existing plumbing so the retry system is untouched:
- engine.search_all_sources pools raw tracks across all configured,
non-exhausted sources (no first-source short-circuit).
- candidates.order_candidates: new quality_first sort path — profile
quality rank dominates, confidence/peer signals break ties. Priority
path is byte-for-byte unchanged (regression-locked by tests).
- task_worker passes quality_first + targets through; skips the redundant
hybrid-fallback block in best-quality mode (pool already covered it).
- Per-source retry budgets unchanged: a source that spends its budget is
added to exhausted_download_sources and thus dropped from the whole
pool. Independent of post_processing.retry_exhaustive.
- Query generator NOT touched.
Also clarifies the "Allow fallback" setting wording: it accepts OFF-LIST
quality as a last resort (not "walk down my list"), and notes that
lossy_copy.downsample_hires also bypasses the quality gate — the cause of
16-bit/MP3 files slipping through a 24-bit-only profile.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Add core/quality/selection.py: rank_with_targets() returns (ranked,
satisfied) where satisfied = a candidate meets a real target (strict).
load_profile_targets()/rank_for_profile() are the DB-backed wrappers.
search_with_fallback now skips a source that can deliver no target-meeting
quality and escalates to the next (source priority still wins among
satisfying sources; first source's results kept as fallback unless the
profile disables it). Returns RAW tracks — the satisfied check is a coarse
source gate; match-filtering + final ranking stay in the orchestrator so
the correct track is never pruned. Ranking is fail-open: a ranking error
never drops a source's real results.
Tested: rank_with_targets satisfied/fallback matrix + engine escalation,
stop-on-first, raw-not-pruned, fallback on/off. Amazon field test updated
for the corrected format token.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
These files had silent `except Exception: pass` blocks but no module
logger. Added `import logging` + `logger = logging.getLogger(__name__)`
at the top of each, then replaced the silent excepts with
`logger.debug(...)`.
- core/replaygain.py — 4 sites (id3 txxx + vorbis + mp4 atom reads)
- core/wishlist/presence.py — 3 sites (wishlist row parsing + queries)
- core/runtime_state.py — 1 site (activity toast emit)
- core/automation/signals.py — 1 site (collect known signals)
- core/download_engine/rate_limit.py — 1 site (plugin rate_limit_policy)
- api/system.py — 1 site (hydrabase status probe)
- api/search.py — 1 site (hydrabase search)
Refs #369
Two findings from JohnBaumb on the engine refactor.
(1) Every download client returned None when self._engine was None,
just logging an error. The orchestrator's download_with_fallback
treated None as "source declined", so the user got no feedback —
download silently disappeared. Now each client raises a RuntimeError
on the engine-not-wired path. download_with_fallback already catches
plugin exceptions, logs a warning, and tries the next source — so
the visible behavior is "real error in logs + fallback to next
source" instead of "silent drop". Six clients touched (deezer, hifi,
qobuz, soundcloud, tidal, youtube). Pinning tests updated to expect
raise.
(2) Monitor's engine.get_all_downloads() walked every plugin
including soulseek, but the same monitor loop already pulled slskd
transfers via the transfers/downloads endpoint a few lines earlier —
soulseek's records were being fetched twice per tick. Same issue in
web_server.py's get_cached_transfer_data path. Added an exclude
parameter to engine.get_all_downloads(); both call sites now pass
('soulseek',). New test pins the exclude semantic.
Also fixed a stray 8-space over-indent on the for-loop body in
get_cached_transfer_data (cosmetic, JohnBaumb flagged the same
pattern in monitor.py earlier).
Per JohnBaumb: the single state_lock serialized progress callbacks
across every source. Pre-refactor each client owned its own download
lock, so Deezer / YouTube / Tidal workers never blocked each other.
Multi-source concurrent downloads under the unified lock fought for
the same RLock on every progress update.
Replaced the engine-wide state_lock with per-source RLocks. Each
source gets its own lock, lazily created via _source_lock() on first
use (meta-lock guards the create-race). All record mutations
(add/update/update_unless_state/remove/get/iter) take only that
source's lock — Deezer progress updates no longer block Tidal writes.
Cancelled-preserve semantics still hold because cancel + worker
terminal write target the same source, so they share that source's
lock. New test pins lock independence: holding source-A's lock from
one thread does not block a write on source-B from another.
Per JohnBaumb's review: iter_records_for_source() walked every
(source, id) tuple across the entire engine state to filter one
source — O(total_records) instead of O(source_records). Fine in
practice because total active downloads is usually small, but the
shape was wrong.
Switched the engine's _records storage from a single composite-key
dict (Dict[Tuple[str, str], DownloadRecord]) to a nested dict
(Dict[str, Dict[str, DownloadRecord]]). Per-source iteration now
only touches that source's bucket. add/get/update/remove all
adjusted to the nested layout. remove_record drops the empty source
bucket so future iterations don't see stale source keys.
Public surface unchanged. New test pins the empty-bucket-cleanup
behavior.
Three correctness fixes from kettui's PR review plus the web_server
migration to generic accessors.
- Engine alias map: register_plugin accepts aliases tuple; get_plugin
+ cancel_download resolve through it. Fixes deezer_dl cancels
silently routing to soulseek.
- Orchestrator hybrid_order normalization: _resolve_source_chain
routes raw config names through registry.get_spec() so legacy
deezer_dl entries don't drop deezer from hybrid mode.
- Atomic update_record_unless_state on the engine: holds state_lock
across the check + write. Both _mark_terminal AND the success path
use it now so a Cancelled state set mid-impl can't be clobbered.
- web_server.py: 30 soulseek_client.<source> reaches migrated to
client("<source>"); shutdown-check setup migrated to generic
registry iteration; 4 hifi reload sites use reload_instances('hifi').
- 18 new tests pin every fix.
Three findings from a final review pass:
1. **Worker clobbered Cancelled with Errored when impl returned
None / raised mid-cancel.** The legacy per-client thread workers
each had a guard (``if state != 'Cancelled': state = 'Errored'``);
the shared worker dropped it. Fix: new ``_mark_terminal`` helper
in BackgroundDownloadWorker reads current state before writing
the terminal one and leaves Cancelled alone. SoundCloud test
updated back to the strict Cancelled-only assertion (had been
loosened to accept Errored as a workaround). Two new pinning
tests catch the regression.
2. **Dead code in engine.py.** ``find_record`` and
``iter_all_records`` had no production callers — only tests.
Removed them. Concurrent-add stress test rewritten to use the
per-source iterator that's actually in use.
3. **Silent ``except Exception: pass`` in cross-source query
methods.** Faithful to legacy behavior (one source failing
shouldn't take down aggregation) but Cin's standard is "log
even when you swallow." Each silent-swallow site now logs at
debug level so the source name + exception are inspectable
without adding warning-level noise.
Suite still green (2049 passed).
`engine.search_with_fallback(query, source_chain, ...)` walks the
chain in order, skips unconfigured / unregistered plugins,
swallows per-source exceptions, and returns the first non-empty
(tracks, albums) tuple. Replaces orchestrator's hand-rolled
hybrid search loop.
`engine.download_with_fallback(username, filename, file_size,
source_chain)` falls through the chain when a source returns
None / raises. Username hint promotes a matching source-chain
entry to head of order. NOT yet wired into orchestrator.download
— today's username comes from a search result and represents
the user's explicit source pick, so silently falling through
would override their choice. Engine method is available for
future callers that want fallback semantics
(search_and_download_best, automation).
Orchestrator gains _resolve_source_chain helper that builds
the ordered list (hybrid_order config, falling back to legacy
primary/secondary pair). Orchestrator.search hands chain off
to engine.search_with_fallback for hybrid mode.
8 new tests pin the fallback semantics: chain ordering,
unconfigured-skip, exception-continue, empty-when-exhausted,
username-hint promotion. Suite still green (2050 passed).
`core/download_engine/rate_limit.py` introduces a per-source
policy declaration: download_concurrency + download_delay_seconds.
Plugins declare via `RATE_LIMIT_POLICY` class attribute or a
`rate_limit_policy()` method.
Engine applies the declared policy to engine.worker at
register_plugin time — set_concurrency + set_delay get pushed
in automatically. Plugins without a declaration get the
conservative default (1 / 0). The set_engine callback fires
AFTER policy registration so config-driven sources (YouTube
reads user-tunable youtube.download_delay) can override.
Plan doc updated to reflect Phase D skip (search code is 90%
source-specific, not 60% — lifting it would be lossy or
bloated).
Pure additive — no plugin migrated yet. 8 tests pin the
resolution priority + engine wire-up + override semantics.
Suite still green (327 download tests).
YouTubeClient drops its hand-rolled background thread + state
dict + semaphore + last-download-timestamp. download() now
delegates to engine.worker.dispatch with _download_sync as the
impl callable; YouTube-specific record fields (video_id, url,
title) merge into the engine record via extra_record_fields.
Engine wires itself in via plugin.set_engine(engine) callback
on register_plugin. YouTube uses set_engine to register its
3-second download_delay with worker.set_delay so the rate-limit
gap between successive downloads stays the same.
Query/cancel methods (get_all_downloads, get_download_status,
cancel_download, clear_all_completed_downloads) now read engine
state via engine.iter_records_for_source / get_record /
update_record / remove_record. Net: ~120 LOC of thread+state
boilerplate removed from youtube_client.py.
Phase A pinning tests updated to assert engine state instead of
client.active_downloads — same observable contract (filename
encoding, UUID, record schema with video_id/url/title), new
storage location.
Suite still green (2025 passed). Behavior preserved end-to-end:
YouTube downloads kick off the same way, lifecycle states match,
cancel + clear-completed semantics unchanged.
`BackgroundDownloadWorker` lives on the engine and owns the
boilerplate every streaming download client currently
hand-rolls: thread spawn, per-source semaphore, rate-limit
delay, state lifecycle (Initializing → InProgress → Completed
or Errored), exception capture.
Plugins provide only the atomic download op (`impl_callable`).
Per-source rate-limit policy (concurrency, delay) is configured
on the worker via `set_concurrency` / `set_delay`. Source-
specific record fields merge in via `extra_record_fields` so
existing consumer code that reads `video_id`, `track_id`,
`permalink_url`, etc. keeps working post-migration. Username
slot supports override (Deezer's legacy `'deezer_dl'`).
Phase C1 scope: worker exists. No client migrated yet — C2-C7
migrate sources one at a time, each gated by the Phase A
pinning tests so per-source contract drift fails fast.
10 new tests pin the worker contract: UUID id format, initial
record shape, extra-fields merge, username override, state
transitions on success / impl-returns-None / impl-raises,
semaphore serialization (default + parallel), rate-limit
delay between successive downloads.
Suite still green (308 download tests). Pure additive.
`DownloadEngine` grows async query methods that wrap plugin
iteration: `get_all_downloads` (concatenates every plugin's
active downloads), `get_download_status` (first plugin to
recognize the id wins), `cancel_download` (with source-hint
routing — streaming sources go direct, unknown hints route to
Soulseek as peer username), and `clear_all_completed_downloads`
(skips unconfigured plugins).
Code moved from the orchestrator's hand-iterated loops into the
engine. Orchestrator delegation comes in B3 — for B2 the engine
methods exist but nothing calls them yet.
Per-plugin behavior preserved verbatim (defensive `try ... except`
swallows per-iteration, unconfigured-skip on clear, source-hint
routing semantics). Phase A pinning tests + 8 new engine query
tests catch any drift.
Pure additive — zero behavior change for users.
`core/download_engine/` package with the engine class that will own
cross-source state, threading, search retry, rate-limits, and
fallback chains. Orchestrator constructs an engine and registers
each plugin with it.
Phase B1 scope: skeleton only. Engine stores active_downloads
records keyed by (source, download_id), provides thread-safe
add/update/remove/iterate primitives, and holds plugin references
for later phases. NOT on any code path yet — pure additive
scaffolding so subsequent commits can introduce engine-driven
behavior one piece at a time without a big-bang switchover.
15 new tests pin the engine's state-storage contract: shallow-copy
reads, partial-patch updates, no-op-on-missing semantics,
per-source iteration, id-only find, concurrent-add safety.
Suite still 290 (download subset) green. Zero behavior change.
