Cross-process against the real arr binary, real ffmpeg and real alass:
a provider fetch synced and named under DESIGN.md §15, never
re-searched once satisfied; an embedded English track extracted and
translated into a .mt sidecar; an implausible sync kept unsynced and
surfaced in the missing-subtitles queue. Podnapisi and the translation
backend are stubbed at the HTTP/process boundary, never a live tracker
or a live translation API.
Daemon::spawn_with_env for extra child environment, database_path()
and media_root() accessors so a test can seed a media file directly,
the daemon built with translate-command, and the plausible/farfetched
Podnapisi zip fixtures the subtitle tests download from.
ARR_PODNAPISI_URL, same shape as the existing tmdb_url seam: a test
harness can point the provider at a wiremock fake without touching
DESIGN.md §10's config surface.
Syncer::run passed [subtitle, video] where alass expects
<reference-file> <incorrect-sub-file>; every real invocation failed
before comparing timings, silently degrading to SyncState::NotRun.
An attempt row survives after a language leaves wanted_languages —
missing_for (#201) already bounds by the current wanted set, so the
queue reads the same way instead of showing a stale gap forever.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
GET /api/queues/subtitles, grouped by title with why each language
is a gap (#186's attempt states, plus a sync alass rejected). Series
episodes collapse into one season row when the gap is uniform, the
same restraint §9.5 gives the TV attention queues.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
#193 moved them out of openai.rs so the command backend could share them,
but left them ungated: with neither LLM backend compiled in they are dead
code. just ci missed it because it lints the workspace with --all-features;
building arr-api alone does not enable arr-subs' features.
Providers are charged one unit per download, claimed atomically
right before the call; a provider at its cap is skipped in favour of
the next-ranked candidate rather than failing the whole gap.
Translators are charged the source character count before
translating. Either cap is a queue state, same as a provider's own
429.
One row per provider/translator per day; try_spend is a single
atomic upsert so concurrent reconcile closes can't both slip a spend
past the daily cap. No allowance configured reads as unlimited.
Backend::characters_billed() polled a cumulative counter that races
under concurrent closes and can't attribute cost to one call. Drop it
in favour of the caller counting source characters before it sends
anything, which #197's budget needs anyway.
The manual subtitle view shows the facts that decided a row (§9.3), and
release-name match is §15's second ranking tier. The file's own release
name is not otherwise on the wire, so the UI cannot derive it.
remove_library_files only resolved video paths from media_files, so a
season/episode-scoped delete dropped subtitle_files rows via cascade
but left the .srt sidecars on disk (#218).
#196 built SubtitleAction with an empty backend list, correctly: no
translation backend existed when it was written. #216 then built
translation_backends() and wired it into AppState, so the API can translate.
Merged, the reconcile loop still got Vec::new() and its translate step
reported "no engine" in every real deployment.
Neither branch was wrong alone; the gap only exists once both are in.
Exposes per-media-file subtitles and missing wanted languages, with the
attempt reason, so #201's UI has one call per title (movies, episodes)
and one bulk call per series instead of one per episode.
Closes each unsatisfied wanted language per DESIGN.md §8/§15: embedded
tracks satisfy for free (recording them as #189 left to this issue),
then provider search + ranked fetch + alass sync + sidecar write, then
immediate machine translation — extracting a text-format embedded track
when that is the only source — and otherwise the reason lands on the
attempt row for the missing-subtitles queue.
Closes run as detached tasks because alass and translation outlive the
25 s reconcile action budget; every outcome is recorded in domain rows
first, so a crash converges on the next tick. Failures back off on the
same §6.2 curve as movie searches; a rate-limited provider is the
'capped' queue state; unreachable providers and translators fold into
the existing §9.5 broken notification, edge-triggered.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Forward each translate-* feature from arr-daemon to arr-subs, add the
missing OpenAI model bootstrap key, and construct the compiled and
credentialed backends at startup so the translate endpoint stops
answering 503 unconditionally.
Both #195 and #214 added a builder call to the AppState chain, which pushed
run past the too-many-lines limit. The chain grows a line per upstream the
API learns to talk to, so it gets its own function.
DESIGN.md §7.5's watcher gap applies to a sidecar dropped next to a
file Jellyfin already knows about, same as an imported file. A grab
or translation now makes the same refresh call import does; a
refresh failure logs and never fails the write that already landed.
arr-daemon depends on arr-api, so a handler in arr-api can never
reach the daemon's private JellyfinClient. Move it into arr-api and
attach an instance to AppState, so a manual subtitle write can ask
for the same refresh import already does (#195).
The grab handler runs alass before recording the row, replacing the
sidecar with the synced text on acceptance and flagging it as
rejected otherwise (§15). Wires a Syncer into AppState, defaulting
to alass on PATH; the daemon binary points it at config.alass_path.
Folds an implausible result and an unusable alass binary into one
SyncState both the grab handler (#199) and reconcile loop (#196)
can record without re-deriving the same match arms.
Turning on --all-features compiled every translation backend, so the test
asserting that a known-but-uncompiled engine is refused had nothing left to
refuse and failed. It hardcoded "deepl" and a comment that no feature was
on, which stopped being true in the same commit that made CI see it.
It now picks whichever engine this build did not compile, and when all of
them are compiled asserts the complementary truth instead: a compiled
engine is accepted. Meaningful under either feature set.
Every translation backend sits behind a default-off cargo feature, and the
gate ran with the default set, so clippy and the test run never saw a line
of arr-subs' openai, deepl or google modules. #191 and #192 each reported
it after verifying their own work by hand. lint and test now pass
--all-features.
Also drops a duplicated wiremock suppression the #188 merge left behind and
puts the module list back in order.
Behind the existing translate-openai feature. One JSON-in/JSON-out
chat/completions request per batch; refusal, content-filter and
truncated replies fail loudly instead of validating as a short batch.
#199 shipped the manual grab before #213 existed, so srt_text answered 422
for any provider not serving SRT. Fetched::to_srt already decodes and
converts, so this is that call. A format with no parser still fails rather
than reaching the disk.
Leaves the alass half of #214 open; that waits on #194.
Miguel Palhas