- hint_extractor: add _normalize_vertical_tracklist() to handle bare-number/
title/duration format (Tufaranka-style tracklists)
- hint_extractor: fix level-1 tracklist match — allow disc_num=None (single-CD)
by assuming disc=1; previously no tracklist title was ever applied to single-
CD tracks because the guard required disc_num to be set
- music_enricher: register module in sys.modules before exec_module() so
@dataclass definitions in jellyfin_playlist_generator work correctly
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
The previous regex lookbehind (?<![A-Za-z0-9_-]) excluded _ as valid preceding
character, so IDs after underscores were never matched. New approach: split stem
by _ and check if the last token is an 11-char YouTube ID (mixed case + digit).
Also strips the ID token from the stem before _parse_filename() to prevent it
from leaking into the track title or being misread as an artist-title separator.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- Extract 11-char YouTube video IDs from audio filenames
- Fetch title, uploader, chapters via yt-dlp (--dump-json)
- Use chapters as tracklist when no .txt tracklist is available
- Store yt_title / yt_uploader in AlbumHints for LLM prompt context
- Fall back to YouTube video title as track title for single-file albums
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
hint_extractor: _ocr_back_cover() sends back/inlay images to Ollama Vision
when no tracklist .txt/.htm/.nfo is present. Model priority:
qwen3-vl:latest → minicpm-v:latest → deepseek-ocr:latest (configurable
via OLLAMA_OCR_MODEL env var). Timeout 180s. OCR text is fed into the
same _parse_tracklist() pipeline as regular text files.
music_enricher: extract_hints(use_ocr=not args.no_api) — OCR is skipped
with --no-api to allow fully offline/fast runs.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
executor: disc=1 now generates '1-01' prefix (same as disc=2 → '2-01'),
so multi-disc albums have consistent D-TT scheme throughout.
Single-disc tracks without disc tag stay as plain 'TT'.
hint_extractor: tracklist pattern 2 now requires '.' ')' or ':' as separator
(not bare whitespace) — prevents false-positive matches like
'2 x CD, Compilation, Remastered' being parsed as track 2.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
hint_extractor:
- _norm_for_match(): strips all non-alnum for punctuation-agnostic comparison
- _catalog_key(): extracts BWV/Op./K./HWV/... catalog number for matching
(fixes abbreviated filenames like "Fantasia_Cm_BWV_562" vs "Fantasia In C Minor, BWV 562")
- Matching priority: exact number+disc → exact title → fuzzy title → catalog number
- Tracklist disc+track OVERRIDE M3U position when a match is found
(M3U is only used as last fallback; fixes wrong alphabetical ordering)
metadata_resolver:
- LLM prompt now defines artist/albumartist roles explicitly
(artist = composer for classical; albumartist = performer/interpreter)
- LLM albumartist can override dir_artist when confidence < 0.4
- _build_track_proposals: when track artist == albumartist (performer from filename),
composer (album-level artist) is used as track artist instead
- Tracklist header (first lines before tracks) included in LLM prompt
for label/year/album-title discovery
- import re added (was missing)
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
scanner: nicht in Unterordner wenn Root Audio-Dateien enthält (verhindert
Doppel-Scan bei versehentlichen Unterordner-Kopien); nur Disc-Ordner
(CD1, Disc 2…) werden bei Multi-CD-Alben rekursiert.
hint_extractor: M3U/Playlist-Dateien als Track-Reihenfolge-Quelle; BOM-
Bereinigung; Tracklist-Matching auch per Titel (nicht nur per Nummer);
tracknumber=0 wird als 'keine Nummer' gewertet.
metadata_resolver: sequenzielle Fallback-Nummerierung (1,2,3…) für Tracks
ohne Tracknummer — verhindert '00'-Präfix beim --rename; dir_artist hat
Vorrang vor 'Various Artists'-Heuristik; LLM darf bei Konfidenz <0.3
auch bestehende Werte korrigieren (Tippfehler im Verzeichnisnamen).
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
AI-powered per-album pipeline: scan → local hints → MusicBrainz/Discogs/Claude
resolve → cover art → interactive or auto review → tag write + rename + report.
All external dependencies optional; 17/17 unit tests passing.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
