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.
Provider credentials are bootstrap config and never reach the database
(DESIGN.md §10), so which providers exist is settled once at startup;
which of them a search runs is the `providers_enabled` row the API reads
per request. OpenSubtitles.com cannot be called without a registered API
key, so without one it is not offered at all.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Issue #199, DESIGN.md §15 and §9.1. Lists what exists per media file and
per title, runs the enabled providers for one language and returns every
candidate with §9.3's verdict vocabulary — including the rejected ones
naming the rule that killed each — then grabs, translates and deletes.
Inline rather than 202-and-poll like the release deck: a subtitle search
is one or two HTTP calls and nothing persists its candidates, so there
is nothing to come back for. The cost is that a grab repeats the
`forced` and `sdh` facts the search reported, since the server does not
remember them.
Every write ends by marking the language satisfied, whether or not it is
in the wanted set. That is §15's "manual actions bypass the wanted-set
logic": the operator asking for Spanish gets Spanish, and the loop does
not then read it as a gap. A forced track is the exception §15 names — it
covers signs only — so it is recorded and satisfies nothing.
`alass` (#194) does not run yet: a fetched sidecar is recorded unsynced.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
DESIGN.md §15 reads satisfaction off the files, so a `satisfied` attempt
row whose sidecar was just deleted by hand is a stale claim that hides
the gap from the reconcile loop's work list. `unsatisfy` withdraws only
that claim: the attempt count and timestamp stay, because the backoff is
a fact about what providers were already asked and a delete does not
un-ask them.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The Podnapisi zip fixtures never reached the branch: a global gitignore
excludes *.zip, so git add skipped them silently and the worker's CI passed
against untracked files. Recovered, with a fixtures .gitignore that keeps
the next binary fixture from vanishing the same way.
Two providers landing in parallel also disagreed on names and on which
crate dependencies each integration-test target uses. PodnapisiProvider is
now Podnapisi, matching OpenSubtitles, and both test targets declare the
dependencies they do not use so the per-target lint stays quiet.
Searches by moviehash computed from the media file and by TMDB id
with season/episode for TV, ranks candidates through
arr_core::subs::rank, and downloads under a lazily-fetched user
token. The daily download cap (429/406) surfaces as Error::RateLimited
so the loop can show a queue state. Credentials come from config or
environment only; tests run against wiremock fixtures.
ffmpeg, spawned and left to die like ffprobe, maps one subtitle stream
and converts it to SRT under §15's sidecar name. Text formats become
legal translation sources; bitmap tracks never extract.
SubtitleTrack now carries the codec, split text formats from bitmap
ones per DESIGN.md §15, plus the forced and SDH dispositions ffprobe
reports. Without the forced flag a file carrying only a forced track
read as satisfied for that language.
The Backend trait plus everything the backends (#191-#193) share, so no
backend can skip it: chunking into character-budgeted batches, rejection
of replies whose cue count or numbering drifted, and reassembly onto the
original timings. Timing data never leaves arr; pt-PT and pt-BR are
distinct targets a backend must refuse rather than conflate.
Translation reassembles translated text onto original timings (§15), so
cues need a structured form. Parsing is tolerant of real files (CRLF,
BOM, missing indices, dot milliseconds); rendering is strict and
renumbers from 1.
#184 and #185 were built in parallel and their candidate types did not
meet. Ranking returned no identity for a candidate, so the winner of a
rank() could not be handed back to Provider::download -- rank() reorders,
so the caller could not recover it by position either.
RankedSubtitle now carries the index of the candidate in the slice it was
given, and Candidate::to_core is the one place the two shapes are mapped:
hash_match against a compared hash, sdh against hearing_impaired, group
against release_group, and the two optional tiebreakers defaulted to sort
last rather than block a candidate.
Recording a subtitle is idempotent on both keys the schema carries: the
sidecar path, and the language an embedded track satisfies, so a second
probe or a re-import converges instead of duplicating. Attempts are
upserted per (media file, language); the work list is every language that
is not satisfied, newest import first, which is the order §15 wants the
daily allowance spent in.
DESIGN.md §15 needs two shapes: what subtitles exist for a media file,
and what arr has tried per wanted language. An embedded track carries no
path — it is recorded because it satisfies a language, not because there
is a file — and the CHECK constraints tie provider, engine and path to
the origin so an impossible row cannot be written.
The skeleton the subtitles milestone hangs off: provider domain types, an
object-safe Provider trait, the crate's own error type, and the cargo
features the translation backends will sit behind. No provider, no
translation, no ranking — DESIGN.md §15 keeps ranking pure in arr-core.
Records the subtitle decisions settled during Subtitles milestone
planning: global wanted set (pt + en), embedded-track extraction,
providers, ranking, translation backends, sidecar naming, alass sync,
budgets, and stated non-goals. Points §13 item 9 at §15 and adds
arr-subs to the crate lists.
Closes-Issue: #183
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Size bands describe one episode, so a pack's size is divided by the
episodes it covers before the floor and target are applied — a 2160p
pack no longer loses to a 1080p one on a penalty that scaled with
season length.
A failed pack grab now quiets the pack lane on the shared §6.2 backoff
curve instead of disabling it forever, and a manual season search
waives the window. The season deck names which state it is actually in
rather than blaming backoff for a sweep that was never going to run.
Closes#179, #180, #181, #182
Miguel Palhas