Add server-side AI subtitle translation backed by any OpenAI-compatible
chat endpoint (OpenAI, Groq, a local Ollama/llama.cpp server). A viewer
picks a source track and target language in the player; the server runs a
bounded, resumable job pipeline that translates SRT/VTT cues in batches and
streams them back over the realtime websocket so playback pauses, fills in
cues near the playhead, and resumes. The finished track is persisted as an
ordinary downloaded subtitle, so it reaches every client through the
existing subtitle pipeline with no client changes.
- Job lifecycle persisted in subtitle_ai_jobs (migration 168): enqueue with
idempotency, bounded concurrency, progress/heartbeat, cancellation, and
crash recovery.
- New realtime events (subtitle_ready + subtitle_translation_*) with a
per-session notifier; the player renders a synthetic "live" track fed by
websocket cues. Timestamps never leave the server, so timing can't drift.
- Admin settings card for endpoint / model / concurrency.
Player + lifecycle hardening (from the code review of this feature):
- Hand off from the live track to the persisted track on completion
(selected by downloaded-subtitle id) and on the subtitle_ready broadcast,
so the saved track survives a reload and a mid-stream socket drop.
- Never persist the synthetic live-track sentinel index as a subtitle
preference; restore the prior selection on failure; only auto-resume
playback if the viewer was actually playing.
- Resume promptly when the playhead is past the last cue; rebuild the live
track on a new job; O(batch) live-cue ingestion instead of O(n^2).
Reliability:
- Root translation jobs in the application context so shutdown cancels them.
- Heartbeat-based stale-job reaper (safe across multiple instances) replaces
the table-wide startup reset.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The shared WebSocket upgrader rejected handshakes unless the browser's
Origin host exactly matched r.Host. Behind a TLS-terminating CDN/proxy
that rewrites Host to the internal origin (carrying the public host in
X-Forwarded-Host), this comparison always failed and every realtime
socket 403'd at the handshake — playback control, events, watch-together
rooms, and admin log streaming all share the upgrader.
checkWebSocketOrigin now also accepts an Origin matching X-Forwarded-Host,
keeping the same-origin CSRF guard intact while supporting proxied
deployments. Extract a shared forwardedHost helper (first hop of a
multi-proxy list) and reuse it from requestBaseURL, replacing the
duplicated inline parse. Also reject opaque (empty-host) origins
explicitly.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
- Extract hardcoded trending snapshot source/window to named constants.
- Collapse the three identical personal-preset nil-checks into one case.
- Persist the selected preset through the shared storage util (try/catch
wrapped) instead of raw localStorage with manual SSR guards.
- Derive KNOWN_FILTERS from PRESET_OPTIONS so the lists can't drift.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Also drops the now-unused Filter/UserID/ProfileID fields from the
blendUpcomingIntoDiscoverRows CalendarFilter literal in recommendations.go,
which only wants an unrestricted windowed query.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The local-airtime change re-sorted calendar events in Go using air_at,
the absolute UTC instant, which is nil whenever air_timezone is unset.
Since air_timezone is only inferred for a few networks/countries, most
events fell through to the alphabetical title tiebreak while still
displaying their raw air_time, so each day appeared scrambled.
Sort each local day by the wall-clock time the viewer actually sees,
mirroring the client: zoned events convert air_at into the viewer
timezone, unzoned events use the raw air_time, and date-only entries
(no air_time) sort last. The timezone reasoning lives in the new
catalog.CalendarEventLocalTime helper.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
episodeCatalogSelectBody is the derived "mi" relation that episode catalog
hydration and preview read qualifiedListItemColumns("mi") from. The local
episode airtimes feature (26370616) added air_timezone to the shared column
lists but not to this hand-written subquery, so the outer projection
referenced mi.air_timezone, which the subquery never exposed.
Postgres returns SQLSTATE 42703 (undefined_column), which is not one of the
codes episodeCatalogEntriesUnavailable treats as "fast path unavailable" (it
only catches 42P01/42883), so episode catalog requests failed with HTTP 500
instead of degrading. movie and series scopes query media_items directly, so
the column is present there and only episode scope broke.
Add si.air_timezone to the subquery, and add a regression test asserting that
episodeCatalogSelectBody exposes every column qualifiedListItemColumns reads
off mi, so future additions to the shared column lists cannot silently drift
from the episode read model again.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
- Interleave Trakt movies/shows by rank so the mixed row shows both types
instead of burying all series past the display limit.
- Treat any Trakt sub-fetch failure as fatal (errors.Join) so a partial
result never overwrites the last-good snapshot with a media type missing.
- Skip non-title entries (TMDB trending/all returns media_type "person") in
both ID batching and ordering so they can't match an unrelated library title.
- Guard the refresh task against a nil refresher.
- Tests: person skip, Trakt interleave, Trakt partial-failure preserves
last-good, snapshot read error propagation.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Extract newTrendingEntry, reuse orderMediaItems, and collapse concurrent
cache-miss loads with singleflight. Baseline for the persistent snapshot work.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
A library-agnostic home section that surfaces external global trending
(TMDB or Trakt, admin-selectable) mixing movies + series, matched to
titles in the viewer's enabled libraries. TMDB uses /trending/all/{window}
(natively mixed); Trakt merges trending movies + shows. Fetched live with
a 1h in-process cache, so no background job or stored collection — and no
per-library duplication.
Appears in the admin section gallery via its recipe presets (TMDB Trending
Today/This Week, Trakt Trending); featured -> hero via the existing flag.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Login identifiers were compared case-sensitively, so "John" and "john"
were distinct accounts and a user could not log in unless they matched the
exact casing used at registration.
Convert users.username and users.email to the citext type (migration 165).
citext compares case-insensitively while preserving the originally stored
casing for display, so the existing unique constraints become
case-insensitive and `WHERE username = $1` / `email = $1` lookups match
regardless of case with no change to the query code itself.
Also add auth.NormalizeUsername/NormalizeEmail (trim-only; case preserved),
applied at the repository chokepoints (Create, Update, GetByUsername,
GetByEmail) and before validation in the create paths, so surrounding
whitespace no longer defeats matching or creates lookalike accounts.
Verified non-destructively against the dev DB: mixed-case lookups resolve
to the same row, case-variant inserts are rejected by the unique
constraint, and the down migration cleanly reverts to text.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
scanItemsWithTotal was not updated for the new air_timezone column, yet the
shared column lists it reads (itemColumns, qualifiedListItemColumns) include
it. Search and BrowseFavorites build their SELECTs from those lists with
COUNT(*) OVER (), so each row carried one more column than the scan had
destinations and every call failed at scan time with a pgx mismatch. Add the
missing &item.AirTimezone target between AirTime and ShowStatus.
Found during PR #21 review (critical: Search/Favorites runtime regression).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Clearing a previously-set air timezone sent JSON null, which decodes to a
nil *string that UpdateMetadata treats as "skip this column", so the old
value remained. The dialog now sends "" (accepted by ValidateAirTimezone),
and UpdateMetadata maps air_timezone through NULLIF so an empty value
persists as SQL NULL (matching the nullable column) rather than "".
Addresses PR #21 review (P2).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Orphan detection moved outside the media_items delete in the batched
library-delete rewrite, opening a TOCTOU race: a concurrent scan/import
could attach one of the collected content IDs to another library between
collectOrphanBatch and the delete, after which the unconditional
`DELETE FROM media_items WHERE content_id = ANY($1)` would still remove the
shared row and cascade away the newly-added membership — dropping the item
from the other library. Re-check the orphan invariant inside the delete
(NOT EXISTS a membership in another folder) and count rows actually deleted.
Addresses PR #21 review (P1).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Episode stills used as backdrops only exist at w500/w300 in the cache, so
requesting a w1280/w1920 backdrop width 404s. Add catalog.BackdropVariantPath
+ imageTypeFromCachedPath and route featured (w1920) and Continue Watching /
Next Up (w1280) backdrops through it; still/poster/logo paths clamp to their
type's largest cached variant while real backdrops keep the requested width.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Replaces the single multi-minute delete transaction with phased, batched
autocommit deletes (orphan items, media files, memberships, folder row),
each retried on deadlock. Holds only short locks, survives concurrent
writers, and is resumable on failure.
Add rowQuerier interface satisfied by both *pgxpool.Pool and pgx.Tx.
Split collectImageDirs into collectRawImageDirs (raw collection) and a
thin wrapper that filters via filterUnreferencedImageDirs. Both helpers
now accept rowQuerier so a later task can call them from pool-level
batch deletes without an open transaction.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
When inferProviderTagRe was broadened to handle unsubstituted Sonarr token
placeholders ({TvdbId}/{imdb-}), the regex grew long enough that gofmt prefers
single-space rather than column-aligned spacing across the var block.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Folder years routinely differ from provider release years by a year or two
(festival vs wide release, regional dates), zeroing the year bonus and leaving
a lone exact-title candidate at 63-68 — just under the single-candidate >=70
gate (e.g. Dead Reckoning 1947 vs 1946, 17 Blocks 2021 vs 2019, Stasi FC). Add
title corroboration to the existing lone-result rule: a sole distinct show whose
normalized title exactly matches and whose year is within +/-2 is accepted. The
55 floor still rejects low-similarity titles (e.g. Hotel Transylvania Puppy! vs
Puppy!). No 55/70/15 threshold change.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Folders tagged with a bare IMDb id in brackets (Plex/Kodi style, e.g.
'17 Blocks (2021) [tt10011226]') had the id silently dropped — folderIDPattern
needs an 'imdb-' prefix and trailingImdbIDPattern needs an un-bracketed trailing
tt-id. Recognize bracketed bare tt-ids so these items get the trusted-ID match
path instead of falling to title+year scoring.
Adds DEBUG logs in the ModeInitialMatch search path: each provider's result
count for the query, and the assembled raw/candidate totals. Lets us see when a
provider search returns zero ('no metadata found') vs a scoring/tie issue.
Zero behavior change.
When the top tie-group is one distinct show returned by 2+ distinct providers
(candidatesAreSingleDistinctShow already verifies matching title+year), accept it
even if the hint has no parsed year (year-less folders like '100 Deeds for Eddie
McDowd'). Multi-source agreement substitutes for the year guard; lone single-source
no-year results stay subject to the single-candidate >=70 gate. Thresholds unchanged.
Accept a year-corroborated single distinct show when the TOP tie-group (within
15 pts of best) is one show across providers, ignoring low-score noise below it,
and pick the winner by the library's metadata-provider chain order (providerPriority,
highest-first; falls back to top-scored). Recovers items like '100 Days Wild'
that are returned identically by TVDB and TMDB. Thresholds (55/70/15) unchanged.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Adds a DEBUG-gated log in selectInitialMatchCandidate printing each scored
candidate (title/year/type/sources/provider_ids/score) against the hint, so
operators can see why an item did or didn't auto-match. Zero-cost when debug
logging is off; no change to matching behavior.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
A scan drainer and the background MatchWorker can process the same folder
concurrently. When a provider-ID merge moves a series' episodes to the survivor
and deletes the source, an in-flight ensureSeriesEpisodeLinks(sourceID) hits
catalog.ErrItemNotFound and was failing the whole scan. The episodes are already
reattached, so this is benign: log and continue (matching the lenient call sites)
instead of failing. Genuine errors still abort.
A search that resolves to one distinct show (one candidate, or the same
title+year returned once per source as unmerged TVDB/TMDB rows) whose year
matches the parsed year is now auto-accepted via the existing top-ranked
candidate, even when the fuzzy title score is in the 55-69 band. The 55/70/15
thresholds are unchanged; this only adds a year-gated acceptance for
effectively-unique results (recovers lone-correct-result items like 1201 (1993)).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
'86' / '22 7' were parsed as trailing tvdb ids, tripping the trusted-ID gate
so the correct title match was rejected. Require a letter in the name before
treating a trailing number as a bare id.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
These tokens survived the provider-tag regex ([\w]+ doesn't match braces),
polluting parsed titles (e.g. 'A Girl & Her Guard Dog [tvdb-{TvdbId}]') so
they could not score-match. Broaden the regex to drop {...} and empty tokens.
The unmatched-items search filtered only the current page's rows client-side.
Push the query server-side: HandleListUnmatchedItems takes an optional 'q' param
and filters title/library/type/status with parameterized ILIKE across all rows,
paginating the filtered set. Frontend hook takes a debounced search, resets to
page 1 on change, keeps the section mounted while searching. Also fixes stale
test mocks that returned the pre-pagination array shape instead of {items,total}.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
