Kometa

Added

• Add semantic letterboxd_crew, letterboxd_studio, letterboxd_country, letterboxd_language, letterboxd_genre, letterboxd_theme, letterboxd_similar, and letterboxd_collection builders for Letterboxd discovery pages.
• Add value_filter overlay-file attribute to filter items at selection time based on a runtime-fetched numeric value; supports comparators gte, gt, lt, lte on any rating_sources variable using the normalised 0–10 scale.
• Add overlay_value_cache table (replaces overlay_special_text2) with a UNIQUE(rating_key, type) constraint and an expiration_date column; values refresh automatically after cache_expiration days.
• Add _overlay_state and _overlay_images per-library tables replacing the dual-use overlay TEXT column in _overlays; one row per overlay per item, written only on successful resolution.
• Add one-time migration that moves rows from overlay_special_text2 into overlay_value_cache (deduped), resets overlay application state for a full reprocess, and logs a visible "Overlay Cache Upgraded" separator.
• Rework defaults/overlays/ratings.yml to fetch ratings directly during the overlay run; Mass Rating Update operations are no longer required for fetched sources. Adds auto image-pick, Fresh/Rotten filtering via value_filter, a Direct image level for tomatoes/tomatoesaudience, and new image assets under defaults/overlays/images/rating/.
• Accept HTTP(S) URLs anywhere a text_file builder used to require a local file path. Kometa first tries to parse the response as JSON (matching today's behaviour) and falls back to a plain-text line-by-line parse on parse failure. Gzip-compressed responses are auto-decompressed. Mixed local/URL lists in a single text_file builder are supported.
• Add logo and square/square_art asset detection and setting via url_ and file_ methods for Collection builders, as well as Defaults via url_*_<<key>> template variables.
• Add Plex-compatible local asset names for posters, backgrounds, logos, and square art, including .tbn files and numbered variants such as poster-2.png; exact filenames such as poster.jpg take priority over numbered variants.
• Add trakt as a source for posters, backgrounds, and logos. Square art remains unsupported.
• Add tvdb as a source for posters, backgrounds, logos and square art.
• Add mass_metadata_update, a grouped Library Operation for genre, content rating, title, studio, date, rating, artwork, mapper, and backup updates. Existing mass metadata operations move under their matching sub-attributes..
• Add YamTrack connector support with yamtrack_list and yamtrack_list_details builders.
• Add yamtrack_tracked builder to collect YamTrack profile movie, TV, and anime items by tracked status.
• Add Plex show edition support wherever movie editions are supported, including edition filters, metadata matching/editing, overlays, dynamic collections, and item_edition.

Fixed

• Allow all semantic Letterboxd discovery builders to use collection_order: custom, preserving Letterboxd's returned order in Plex collections.
• Resolve TMDb episode ratings by Plex's episode-level tmdb:// GUID before falling back to season and episode numbers, allowing alternate Plex episode orderings such as split-cour anime to map correctly; persist the direct episode ID in the existing TMDb episode cache for zero-request warm lookups.
• modules/request.py/modules/plex.py: fix get_image() and delete_user_playlist() crashing a collection or playlist sync outright on a transient network failure (connection reset, timeout, plex.tv DNS resolution). Both now catch the underlying requests exception and raise Failed, matching how every other network-facing call in these modules already degrades.
• modules/tvdb.py: add a distinct Unavailable exception for TVDb requests that exhaust their retry budget without ever returning usable content (e.g. repeated HTTP 202/empty-body "still generating" responses), separate from NotFound's definitive 4xx. Previously both were re-raised with the identical "No Series/Movie found" message, so a confirmed-dead TVDb ID and a transient TVDb hiccup were indistinguishable in the log. All get_tvdb_obj() call sites (modules/builder.py, modules/operations.py) now log NotFound at debug (unchanged), Unavailable at warning (previously logged at error via the generic Failed handler), and any other Failed at error (unchanged).
• Fix custom rating<n>_file (and git/repo/url) overlay images being silently replaced by a built-in default/pmm asset.
• Prevent Kometa from creating duplicate collections when Plex search misses an existing same-named collection by falling back to the full collection inventory before creating.
• Refine the run summary output: group repeated overlay, Letterboxd/TMDb, missing TVDb season/episode and IMDb episode, Plex resolution-regex, Trakt TVDb-miss, and missing local artwork/theme path messages; update the supported-artwork summary wording; normalize asset-directory and logo warning summaries; print overlay and convert summaries in the same Count | Message format as warning/error summaries; add a visual divider before the summary intro text; rename the overlay section to Overlay Summary; and keep the run-status table hidden when there is no status data to show.
• Group missing TMDb collection errors into a single end-of-run summary row while preserving each collection ID in the main log.
• Fix overlay cache poisoning where application state was written for unresolved overlays (e.g. no IMDb rating), causing the item to be permanently skipped even after a rating became available.
• Fix duplicate rows accumulating in overlay_special_text2 on every run due to a missing UNIQUE(rating_key, type) constraint.
• Ignore episode logo and square-art files during asset-directory discovery because Plex episodes only support poster and background artwork.
• modules/textfile.py: fix tmdb: entries in playlist/mixed mode (where is_movie=None) being silently dropped. The old is_movie is not False check treated None as True, tagging TV show TMDb IDs as "tmdb" type which only checks movie_map. Now returns both (id, "tmdb") and (id, "tmdb_show") so the builder resolves against both movie and show maps. Also updates mdblist and simkl three-state contract comments and docs.
• modules/poster.py: fix 18 Optional-member-access errors flagged by pyright. The ImageData.__init__ triple-nested-ternary is unwound into an explicit if/elif/else (also skips a redundant os.stat() call when compare= is provided explicitly). apply_vars and adjust_text_width early-return on self.text is None (matches the callers' existing if component.text: guard). get_generated_layer raises Failed with a descriptive message instead of crashing with 'NoneType' object is not subscriptable when Image.load() or text-dimension computation returns unexpectedly. Baseline drops 1223 → 1205.
• modules/letterboxd.py: fix 9 Optional-call/iterable errors flagged by pyright by removing dead defensive code. letterboxdpy is a pinned hard dependency in requirements.txt, so the speculative try/except ImportError (and the runtime _require_library() guard it required) was never reachable in any supported install. Imports are now unconditional, matching every other integration module (plex.py, tmdb.py, etc.). The two util.get_list(...) for-loops in validate_letterboxd_lists / validate_letterboxd_user_pages now pass return_none=False and fall back to [], so they never try to iterate None. Net diff: -19 lines, no behavioural change for any real install. Baseline drops 1205 → 1196 (and modules/letterboxd.py falls out of the per-file table entirely).
• modules/anidb.py: switch from lxml import etree to import lxml.etree as etree so pyright can resolve the symbol. lxml.etree is a C extension (no Python wrapper module), and the from X import Y form trips pyright's reportAttributeAccessIssue while the equivalent import X.Y as Y form does not. Identical runtime behaviour. Baseline drops 1196 → 1195.
• modules/convert.py: fix 7 Optional-cluster errors flagged by pyright and harden two real latent bugs in the process. Extracted the thrice-repeated re.search("-(.*)", check_id).group(1) pattern into a _hama_suffix() static helper that raises MappingConvertError("Malformed Hama ID ...") if the regex doesn't match, instead of letting an AttributeError escape from .group(1) on None. Restructured the int(self.tmdb_to_tvdb(tmdb_id, fail=True)) call to extract the result first and check for None (the surrounding except Failed: pass only catches the documented failure mode; int(None) would have been an AttributeError slipping past it). The two util.get_list(...) for-loops now pass return_none=False and fall back to [] so they never try to iterate None (matches the fix already used in modules/letterboxd.py). Added 2 new tests covering the helper's happy path + malformed-input behaviour. Baseline drops 1140 → 1067.
• modules/imdb.py: fix 7 pyright errors spanning four patterns and harden two latent bugs. (1) validate_imdb iterates util.get_list(...) with the now-familiar return_none=False or [] belt-and-suspenders. (2) main_data = imdb_dict[X].strip() becomes str(imdb_dict[X]).strip() so a non-string YAML value (e.g. a bare numeric id) doesn't crash with AttributeError. (3) In _graphql_json.check_constraint, the if range_data: block reads range_name[i] — but range_name defaults to None, and pyright noticed the invariant 'range_data only populates when range_name is not None' isn't visible across loop iterations. Made the invariant explicit: if range_data and range_name is not None:. (4) re.search(r"(\d+) of (\d+).*", relevant).group(...) in _imdb_keywords now guards on if result is not None, falling back to (0, 0) (matches the else branch already used for inputs without "of"). Also added @overload signatures to IMDb._interface(interface) so "ratings"/"basics" return dict and "episode" returns list[list[str]], fixing the self.ratings[imdb_id] type-mismatch at the cached-property layer. Baseline drops 1188 → 1181 (and modules/imdb.py falls out of the per-file table entirely).
• Bundle: fix 9 pyright errors across four small files — all the same util.get_list(...) iterable + .strip()-on-possibly-non-str pattern that's been fixed in letterboxd.py, convert.py, imdb.py, anilist.py. Touched files: modules/mdblist.py (2 errors — validate_mdblist_lists), modules/icheckmovies.py (3 errors — validate_icheckmovies_lists), modules/textfile.py (3 errors — validate_file + get_ids; the or [] belt-and-suspenders was the missing piece since return_none=False alone isn't visible to pyright through the untyped helper), modules/config.py (1 of 2 errors — notify; the other 'Code too complex to analyze' error needs a refactor). All three small files fall out of the per-file error table. Baseline drops 1181 → 1172.
• modules/validator.py: switch except ryaml.error.YAMLError to except ryaml.YAMLError at both call sites (lines 144 and 291). Same C-extension import-form trap as modules/anidb.py and modules/request.py: pyright resolves the top-level ruamel.yaml.YAMLError alias but not the .error.YAMLError chain. Identical runtime behaviour — ruamel.yaml re-exports YAMLError from ruamel.yaml.error. Baseline drops 1181 → 1179 (and modules/validator.py falls out of the per-file table entirely).
• modules/anilist.py: fix 7 pyright errors across three patterns and harden one latent bug. (1) int(util.validate_date(value, return_as="%Y%m%d")) becomes int(str(...)) to satisfy pyright's int(str | datetime) complaint — validate_date returns str when return_as is provided but the helper is not typed to reflect that. (2) The mod == "gte"/"lte" blocks did value -= 1 / value += 1 where pyright saw value as int | LiteralString | Unknown. The invariant 'gte/lte modifiers only apply to numeric attributes (start/end/episodes/duration/score/popularity)' is true at runtime but not visible across the dispatch table. Cast with int(value) to make the contract explicit. (3) validate(name, data) iterated util.get_list(data) (possibly None) and called d.lower() on items pyright saw as dict | int | str. Added return_none=False or [] + wrapped d in str(...). Added 1 test (numeric input doesn't AttributeError). Baseline drops 1181 → 1174.
• modules/request.py: fix 7 pyright errors across three patterns and harden three latent bugs. (1) get_stream did dl / total_length * 100 inside the per-chunk log line, but total_length is None when the server doesn't send a Content-Length header — division would TypeError. Added an if total_length: branch that falls back to a bytes-only progress line. (2) get_header had an implicit return None path that the get_cloudscrape_html caller would crash on with cloud_headers.pop("User-Agent"). Made the fallback return {} so callers always get a dict (semantically identical to None for requests.get(headers=...)). (3) The YAML class let self.path=None silently flow into os.path.exists() / open() when the caller forgot to pass either path or input_data. Added an explicit raise Failed("Either path or input_data must be provided") guard at the start of the no-input-data branch. Also rewrote the except ruamel.yaml.error.YAMLError clause to use ruamel.yaml.YAMLError (which pyright can resolve) and replaced e.problem direct access with getattr(e, "problem", None) since problem only exists on MarkedYAMLError (a bare YAMLError would have crashed the error handler with AttributeError). Added 4 tests (get_header matrix + YAML() failure path). Baseline drops 1181 → 1174 (and modules/request.py falls out of the per-file table entirely).
• modules/logs.py: fix 5 pyright errors across two clusters. (1) remove_main_handler and remove_playlists_handler called self._logger.removeHandler(self.main_handler) directly, but main_handler / playlists_handler are initialised to None in __init__ and only assigned in the corresponding add_* method. Calling remove_* before add_* would crash with a type error from logging.removeHandler(None). Added explicit if self.main_handler is not None: guards. (2) MyLogger.findCaller used while hasattr(f, "f_code"): to walk the frame chain — semantically correct (every non-None frame has f_code) but pyright doesn't narrow Optional[FrameType] through hasattr. Switched to the equivalent while f is not None: which is both clearer and pyright-friendly. Baseline drops 1156 → 1151 (and modules/logs.py falls out of the per-file table entirely).
• Singles bundle: fix 4 pyright errors across three small files (one error each in modules/ntfy.py and modules/tvdb.py, two in modules/ergast.py). (a) modules/ntfy.py: change title: str = None to title: str | None = None — classic PEP 484 Optional-default annotation. (b) modules/tvdb.py: wrap released in str(...) before datetime.strptime — the local parse_page helper returns list | str | None and the existing if released else released ternary handles None, but pyright can't see the str-only branch. (c) modules/ergast.py: Race.session_info subtracted timedelta from self.date unconditionally, but self.date is set to None when the API returns a malformed/missing date string (existing except (ValueError, TypeError): self.date = None). Added an explicit if self.date is None: video_date = None early branch — real latent bug that would crash with TypeError rather than gracefully degrade. Added 1 regression test in tests/test_ergast.py. Three files drop off the per-file error table. Baseline drops 1151 → 1147.
• modules/library.py: fix 7 pyright errors across three patterns. (1) Added missing @abstractmethod def get_ids(self, item) -> tuple: ... to the Library(ABC) base — the method is concrete on Plex but wasn't declared abstract, so pyright saw self.get_ids(item) as an attribute access on Library* (the unbound subclass type) and flagged it. (2) Added -> list return annotations to the existing @abstractmethod declarations for get_all and item_labels and replaced their pass bodies with ... — the implicit None return type was being propagated to call-sites like for la in self.item_labels(item) and for item in items which then crashed reportOptionalIterable. (3) Matched the existing poster.compare if poster else "" defensive ternary for the parallel background, logo, and square_art cache-update calls in image_update — the X_uploaded flag guarantees X is non-None at runtime but pyright can't follow that invariant across the dispatch chain. Baseline drops 1147 → 1140 (and modules/library.py falls out of the per-file table entirely).
• modules/util.py: high-leverage fix — chip 26 pyright errors out of the file itself AND cascade-fix 56 more errors across other files for a total −82 baseline reduction in one PR. Three patterns: (1) Added @overload declarations to util.get_list so pyright knows return_none=False guarantees list (never None). This single change fixed errors in 8 caller modules (kometa.py, anilist, builder, icheckmovies, imdb, mdblist, meta.py dropped 59 → 25, textfile.py dropped off). (2) The module-level logger: MyLogger = None was lying — the runtime default is None but the annotation said MyLogger. Switched to a TYPE_CHECKING split (logger: MyLogger for pyright, logger = None for runtime) so consumers can still call logger.info(...) without an Optional guard. (3) Hardened two real latent bugs in the process: HTTPError.response is Response | None per requests typing, but the retry_if_http_429_error predicate and wait_for_retry_after_header.__call__ both used it unconditionally — added explicit is not None checks. And parse()'s error-message branch did for o in options when options could be None (the elif condition allowed entry via the translation clause without options being set) — now falls back to translation keys. Added 4 regression tests. Baseline drops 1181 → 1100 across 9 files; util.py and textfile.py drop off the per-file table entirely.
• modules/radarr.py + modules/sonarr.py + stubs/arrapi/: eliminate 108 pyright errors across radarr.py and sonarr.py (54 each) with a local stub package for arrapi. Root cause and fix are identical to the tmdbapis stubs in the same PR: BaseObj._parse() in the arrapi library (also Kometa-maintained, Kometa-Team/ArrAPI) has no return-type annotation, producing the same ~N-member union explosion on every attribute of every Movie, Series, and Tag object. Stub files created: stubs/arrapi/{__init__,exceptions}.pyi, stubs/arrapi/apis/{__init__,base,radarr,sonarr}.pyi, stubs/arrapi/objs/{__init__,base,reload,simple}.pyi. RadarrAPI.add_multiple_movies and SonarrAPI.add_multiple_series are fully annotated with their actual positional parameter lists (they had 8 and 11 positional params respectively); _raw: Any is declared as an instance attribute on SonarrAPI (it's an internal raw-API object with .v3/.v4/.new_codebase attributes, not a method). Four remaining Optional.id/name errors in each file (qp: QualityProfile | None loop-assigned variable used without a None guard) are pre-existing real bugs and left for a separate PR. Baseline drops 397 → 297 (radarr.py and sonarr.py each drop 54 → 4).
• modules/tmdb.py + stubs/tmdbapis/: eliminate all 670 pyright errors in modules/tmdb.py with a local stub package for tmdbapis. The root cause: TMDbObj._parse() in the tmdbapis library has no return-type annotation, so pyright infers a ~49-member union from all its elif branches (every TMDb object type — Movie, TVShow, Credit, Episode, int, float, str, datetime, etc.). Every attribute assigned via self._parse(...) then carries that union, and any subscript or iteration on such an attribute emits one error per union member — hence 49 errors on a single results.tv_results[0] subscript and 40 errors on each for i in person.movie_cast comprehension. Fix: stubs/tmdbapis/ directory (new) with pyproject.toml: stubPath = "stubs". The stub declares _parse() -> Any on TMDbObj, which collapses all attribute union explosions. Critical attributes that Kometa actually uses are typed precisely — FindResults.{movie_results,tv_results,tv_episode_results}: list[Movie|TVShow|Episode], Person.{movie_cast,tv_cast,movie_crew,tv_crew}: list[Credit], Credit.{movie,tv_show}: Movie|TVShow, TMDbPagination.total_results: int + __iter__ — so code that reads those attributes gets the right narrow type. __getattr__ -> Any on every class serves as the safe fallback for the hundreds of other attributes Kometa doesn't use. One small code fix in modules/tmdb.py: added self.tvdb_id: None = None to TMDbMovie.__init__ (the attribute is only meaningful on TMDbShow, but convert_from() accesses item.tvdb_id on the union type — pyright correctly flagged the missing attribute on the movie branch). Baseline drops 1067 → 397 (modules/tmdb.py falls off the per-file table entirely). tmdbapis is maintained by the Kometa team; a follow-up PR to add py.typed + inline annotations upstream would eliminate the need for these stubs.
• Add a 3-strike retry around Plex saveMultiEdits() read/connect timeouts: retry after 2 seconds on attempt 1, retry after 5 seconds on attempt 2, and fail gracefully on attempt 3 while preserving the traceback in meta.log.
• Update the Sight & Sound Letterboxd default collection to use the correct sightsoundmag owner for the list.
• Fix mixed-library playlist plex_search by building the Plex search per library type while logging the criteria once.
• Fix asset-directory logos being ignored during metadata image updates.
• Skip episode rating operations for movie libraries.
• Show the configured assets directory in missing-asset warnings instead of 'None' when no matching flat asset file is found.
• Suppress full stack traces for a small set of known, non-critical logger patterns so expected Plex not-found noise prints as a warning instead of a stack trace.
• Preserve Plex batch multi-edit state across timeout retries so a transient saveMultiEdits() timeout no longer raises Batch multi-editing mode not enabled on retry.
• Create the per-library _backgrounds and _logos image-map tables unconditionally so caches created before those tables existed self-heal on the next run, instead of raising no such table: image_map_<n>_logos and failing every collection that sets a logo.
• Report transient TMDb network failures as a warning and a timeout-style error instead of dumping the raw connection traceback into the run summary.
• Asset folder would be reported as None when asset_folders was set to false
• Skip Recommendation Hub sorting on a targeted --run-collections run instead of resorting every pinned collection; a partial run only knows the hub_priority values of the collections it rebuilt, so treating the rest as unprioritised pushed them out of position. Also log the configured hub_priority value during attribute validation, matching other methods like collection_order.
• Extend the --run-collections Recommendation Hub sort skip to also cover --run-files, which has the same partial-run limitation.
• Warn and skip item_edition edits when Plex Pass is unavailable instead of attempting the edit and surfacing a Plex 403 Forbidden error.
• Fix the default AU content-rating overlay so the MA15+ badge matches both Plex rating spellings: canonical au/MA 15+ from explicit TMDb Australian certifications and Plex-derived au/MA15+ for titles without an AU certification.
• Fix "Movies Added" report entries missing release year; titles now match the "Movies Missing"/"Movies Removed" format e.g. The Grapes of Wrath (1939).

Changed

• Update requirements
• Playlist libraries is now optional; when omitted, playlists use every library processed as part of the run. Defining libraries on a playlist still overrides that default.
• modules/request.py: every outbound HTTP request now sends a 30-second per-socket timeout (DEFAULT_TIMEOUT), so a stalled external server can no longer hang a run indefinitely. Retries on Requests.get/post switch from a fixed 10-second wait (up to 50s of sleeping per failing URL) to exponential backoff capped at 10 seconds (~25s worst case, much less for transient blips).
• modules/request.py: get_stream throttles its download-progress log updates to ~4 per second (plus a final 100% line) instead of logging once per 8 KB chunk.
• modules/cache.py: the cache now holds one shared SQLite connection (WAL journal mode) for the life of the run instead of opening a new connection for every query — previously each of the ~60 cache methods opened a connection per call and never closed it, which added measurable overhead on large overlay runs. Transaction-per-block commit behaviour is unchanged.
• Updated assets to accept all filenames and filetype extensions that Plex allows as per https://support.plex.tv/articles/200220677-local-media-assets-movies/
• Internal: replace mypy (which was running with continue-on-error: true and producing output nobody read) with pyright using a ratcheting baseline. .pyright-baseline.json pins the current per-file error counts; the new pyright CI job (powered by scripts/pyright_baseline.py --check) fails any PR that introduces new errors in a file but lets maintainers chip away at existing errors at their own pace. Today's baseline: 1223 errors across modules/ + kometa.py. See scripts/README.md for the --update workflow.

Security

• modules/cache.py: dynamic table/column names interpolated into SQL are now validated by a sql_identifier() guard at every method that accepts them. All current callers pass internal identifiers, so this hardens against future misuse rather than fixing an exploitable path (and addresses bandit B608).
• modules/request.py: a per-library verify_ssl: false (e.g. on a Plex connection) no longer disables InsecureRequestWarning process-wide; only the config-level global SSL opt-out does.
• modules/letterboxd.py: short boxd.it URLs must use http/https; other schemes now fail with a clear error before any network call.
• modules/logs.py: registered secrets are also redacted in their URL-encoded (quote/quote_plus) forms, so tokens embedded in logged query strings no longer slip past redaction.
• Internal: replace mypy (which was running with continue-on-error: true and producing output nobody read) with pyright using a ratcheting baseline. .pyright-baseline.json pins the current per-file error counts; the new pyright CI job (powered by scripts/pyright_baseline.py --check) fails any PR that introduces new errors in a file but lets maintainers chip away at existing errors at their own pace. Today's baseline: 1223 errors across modules/ + kometa.py. See scripts/README.md for the --update workflow.