On the video record, per extracted file <subCid> (midhash256 of its bytes) in <lang3> (ISO 639-2/B — fre, ger, chi — normalised the same way meta-watch does; und when the stream declares no language):

Key	Value
`subtitles/<lang3>/<subCid>`	`"true"` — the leaf meta-watch lists
`extractedSubtitles/<lang3>/<subCid>`	`"true"` — provenance: came out of the container
`subtitleLanguages/<lang3>`	`"true"` (not for `und`)
`languages/<lang3>`	`"true"` (not for `und`)

On the subtitle file's own record: videos/<videoCid> = "true" and subtitleLanguage = <lang3> (not for und).

Writes are a single PATCH /meta/{cid} carrying only keys the record does not already hold; a failed write fails the task. Legacy comma-joined scalars written by the old version (extractedSubtitles, subtitleLanguages, subtitles) are deleted first — a scalar next to field/... keys breaks meta-sort's nested document.

Behaviour

Reads the ffmpeg plugin's stream table in any shape: meta-sort's nested stream array (JSON strings or objects), an index-keyed object, a JSON string, or meta-core's flat stream/{n}.
One ffmpeg pass extracts every text track (one -map/output per track). Subtitle packets are interleaved through the whole container, so that pass is still one full read of the video over WebDAV — but one, not one per track. If the shared pass fails, tracks are retried one by one.
Image codecs (PGS, VobSub, DVB, teletext, XSUB) are skipped explicitly; other unknown codecs are skipped and logged.
Skips a video that already carries the extractedSubtitles key-set (unless forceRecompute), and reuses an output file that already exists instead of re-reading the video.
Output: /files/plugin/subtitle-extractor/<Title> (<Year>)[<videoCid>]_subtitle.s<streamIndex>.<lang3>[.forced].<ext> — the stream index keeps two tracks in one language apart.
WebDAV endpoint resolved per request from the driving meta-core (/urls → webdavUrlInternal); WEBDAV_URL overrides.

Configuration

Option	Type	Default	Description
`forceRecompute`	boolean	`false`	Re-extract even if already processed
`outputFormat`	select	`native`	`native` keeps ASS as ASS (stream copy) and SubRip/WebVTT as-is, converts mov_text to SRT; `srt`, `vtt`, `ass` convert everything
`extractionTimeoutMs`	number	`270000`	Timeout of the single ffmpeg pass

Supported codecs

Text (extracted): subrip/srt, ass, ssa, webvtt, mov_text, text. Image (skipped): hdmv_pgs_subtitle, dvd_subtitle, dvb_subtitle, dvb_teletext, xsub.

Tests

pnpm install && pnpm test        # locally (ffmpeg tests need ffmpeg + ./test/fixtures/generate-test-fixtures.sh)
./test.sh                        # in Docker, fixtures generated in the image

Environment Variables

Variable	Description
`PORT`	HTTP server port (default: 8080)
`WEBDAV_URL`	Optional WebDAV override (otherwise resolved from the meta-core `/urls`)
`CACHE_PATH`	Temp root (default `/cache`; files go to `$CACHE_PATH/temp`)

License

MIT