build_rss_rule joined the comma-separated filter string with '|' via
str.join, which splits it into characters ('720,480' -> '7|2|0|,|4|8|0')
instead of alternating whole terms ('720|480') the way every other filter
site does with .replace(',', '|'). Latent since the rule dict was first
written; low real-world impact (the download path uses add_torrent, not
qB-native RSS rules) but the extracted helper is now correct.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014w1Z6Nxy6XTRgkFXqPr9Zh
No production code imports the sync `engine` after the async flip. Remove it
and its foreign_keys connect listener; the migration tests build their own
throwaway sync engine via create_engine and the sync create_tables/
run_migrations helpers in migrations.py, which are kept for that purpose.
DATA_PATH stays only to derive ASYNC_DATA_PATH.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014w1Z6Nxy6XTRgkFXqPr9Zh
Convert the whole DB stack from a sync Session to AsyncSession in one atomic
change. The async boundary cannot be split into individually-green commits:
async repositories and sync consumers cannot coexist, and the API service-mock
tests break the moment the consumers flip, so the planned repos/consumers/test
commits land together here.
- Repositories (bangumi/rss/torrent/user) take an AsyncSession; every method is
async. user.py drops sqlmodel session.exec for session.execute().scalars().
- Database is an async context manager over async_session_factory (with an
optional engine= DI seam for tests); add stays sync, commit/rollback/refresh/
close and create_table/run_migrations/migrate are async. get_db is async.
- migrations.py gains create_tables_async / run_migrations_async wrappers that
drive the existing sync-Connection bodies via async_engine.begin() + run_sync.
- Delete the module-level bangumi TTL cache entirely. Its real job — matching
every torrent in an RSS refresh without re-querying per call — is now done by
loading the active bangumi list ONCE per refresh cycle in RSSEngine.refresh_rss
and matching in memory via match_bangumi_in_list().
- Flip every consumer to `async with Database()` + awaits: rss engine/analyser,
managers (torrent/renamer/collector), core loops/context/offset_scanner,
notification manager (dropping the asyncio.to_thread poster shim for a direct
async query), security/api auth helpers, all api routes, mcp tools/resources,
and the update/* startup chain.
- Fix two pre-existing latent bugs: update/rss.py::update_main_rss and
update/data_migration.py::data_migration called the async engine.add_rss(...)
WITHOUT await, so the calls were silently never executed. Both are now awaited
and their enclosing functions made async up to AppContext.startup.
- RSSEngine/TorrentManager lose their transitional __enter__/__exit__; routes
construct them plainly. Test fixtures go async (in-memory StaticPool engine;
bare Database() rebound to a per-test temp-file DB; get_db overridden).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014w1Z6Nxy6XTRgkFXqPr9Zh
Concurrent aiosqlite connections replace the single serialized sync Session,
so add PRAGMA journal_mode=WAL and PRAGMA busy_timeout=5000 to the async
engine's connect listener (keeping foreign_keys=ON). Extract the listener as
_set_sqlite_pragmas_async so tests can attach it to a throwaway engine.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014w1Z6Nxy6XTRgkFXqPr9Zh
AppContext is built in create_app and owned by the lifespan: startup() is
awaited before serving (errors abort boot), and start_tasks() runs as a
supervised background task. A generic PeriodicTask/Scheduler replaces the four
Thread mixins, with per-tick bodies moved to plain functions in loops.py. The
MCP app now takes the context instead of importing the api.program global.
Config updates go through reload_settings(), restoring the RSS re-application
that was previously commented out. SessionStore adds session expiry and
constant-time token comparison, and downloader shutdown() is wired into app
shutdown.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014w1Z6Nxy6XTRgkFXqPr9Zh
Cache one concrete client per connection-settings key at module level and reuse
it across `async with DownloadClient()` blocks instead of logging in on enter and
out on exit every operation (~1/min from the rename loop). QbDownloader.auth is
now idempotent — a live, authenticated client short-circuits without re-posting
login — and data operations go through a _request wrapper that re-authenticates
once and retries on a 403 (expired session). __aexit__ no longer logs out; a new
shutdown() module function tears the cached client down (wired up by the Phase 5
composition root). MockDownloader/Aria2Downloader gain the same auth idempotency.
When connection settings change, the stale client is retired and closed lazily on
the next enter. A failed fresh login still closes the pool and drops the cache
entry before raising (leak fix, #1043).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014w1Z6Nxy6XTRgkFXqPr9Zh
Introduce a DownloaderClient Protocol with per-backend DownloaderCapabilities;
the DownloadClient facade now consults capabilities and skips-and-logs
unsupported operations instead of relying on NotImplementedError stubs. Aria2
advertises only its core surface (auth/logout/add_torrents) and its ~20
NotImplementedError bodies are deleted; qB and Mock advertise the full surface
(Mock gains get_torrents_by_tag).
Extract the qB RSS rule dict into rules.build_rss_rule and route both the
facade set_rule and manager.update_rule through a new DownloadClient.set_rss_rule
guarded by can_rss_rules. Convert TorrentPath methods into module-level
functions in downloader.path and update all callers.
Recompose Renamer, SeasonCollector, and SearchTorrent to take/hold their
collaborators instead of inheriting DownloadClient / RequestContent / RSSAnalyser,
and update their call sites (sub_thread, api.rss, api.search, mcp.tools).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014w1Z6Nxy6XTRgkFXqPr9Zh
Break the DB-side inheritance chains so no service IS-A Session:
- Database no longer subclasses sqlmodel.Session; it now owns a Session
(self.session) and constructs the repos from it. It keeps its
context-manager role plus thin add/commit/rollback/refresh/close delegates
for the call sites that relied on Database being a Session, and retains
create_table/run_migrations/drop_table/migrate.
- RSSEngine and TorrentManager are plain services that take a Database in
__init__ and reach the repos through self.db. They expose rss/torrent/bangumi
as read-only properties over self.db and keep a trivial __enter__/__exit__ so
the existing `with Service(db) as x:` call sites and their black-box tests
stay valid.
- RSSAnalyser no longer subclasses TitleParser; it calls TitleParser.raw_parser /
mikan_parser / tmdb_parser statically. SearchTorrent keeps composing
RSSAnalyser and works unchanged because those methods are now static calls.
- API routes obtain a request-scoped Database through a new get_db FastAPI
dependency (module/database/__init__.py) and build the service from it.
- Lifecycle-only call sites that previously went through an RSSEngine (create
table / run migrations / migrate / default user) now use Database directly.
No behavior change; the REST API contract is frozen. Added minimal black-box
tests for the previously-uncovered routes (/rss/analysis, /rss/collect,
/rss/subscribe, PATCH /bangumi/{id}/weekday) before refactoring. Full suite:
741 passed / 71 skipped / 2 xfailed.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014w1Z6Nxy6XTRgkFXqPr9Zh
Move the migration machinery out of the Database session class into
module/database/migrations.py. Each Migration entry now carries a
declarative already_applied guard (column_exists / table_exists),
replacing the hardcoded per-version if-ladder in run_migrations().
Version numbers 1-9 and their SQL are unchanged — existing databases
upgrade identically. CURRENT_SCHEMA_VERSION is derived from the list, so
adding a migration is one tuple entry with its guard. The runner works on
a single sync Connection (each migration in a SAVEPOINT, stop on first
failure), which the async flip can later drive via conn.run_sync().
Database.create_table()/run_migrations() stay as thin delegates for
existing callers; they will move to the composition root in a later phase.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014w1Z6Nxy6XTRgkFXqPr9Zh
Post-review fixes for the batch-3.2.8 branch:
- auth() and the setup wizard again require a positive success marker on
200 responses ('Ok.' body) so a proxy or non-qB service answering
200 + HTML is not mistaken for a login; 204 stays the qB >= 5.2 success.
- delete_rule no longer wraps a failed torrent deletion into a 200 reply,
and the WebUI torrents-delete endpoint reports the failure instead of
answering 'Torrents deleted' unconditionally.
- The close-on-failed-auth teardown now lives once in
DownloadClient.__aenter__ (client.logout() before raising) instead of
being copy-pasted into each concrete client.
- search_all() records a cache generation before querying and skips the
cache write if an invalidation landed in between, so a stale snapshot
can no longer overwrite a newer invalidation.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014w1Z6Nxy6XTRgkFXqPr9Zh
Delete the legacy notification stack (notification.py + plugin/, five
providers duplicated by the active providers/ registry, including a broken
copy-pasted slack plugin), the unused SIGALRM timeout decorator, the unused
ConflictError, the unused bgm_parser, the 0-byte transmission client stub,
unused Lock/Event state in ProgramStatus, the fully-skipped legacy RSS
engine test module, and the dev_server/test_passkey_server scratch entry
points that duplicated app assembly.
The zh notification message template now exists in exactly one place
(NotificationProvider._format_message) — the other seven copies lived in
the deleted legacy plugins.
No behavior change; full suite stays green (726 passed).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014w1Z6Nxy6XTRgkFXqPr9Zh
OpenAIParser now uses AsyncOpenAI — the previous sync SDK call ran directly
on the event loop (submitting to a ThreadPoolExecutor and immediately
blocking on .result() serialized it anyway). TitleParser.raw_parser and its
callers become async accordingly.
The SPA catch-all route listed dist/ on every request; snapshot it once at
startup. The module-level bangumi TTL cache is now lock-guarded — it is
written from asyncio.to_thread workers in notification paths while the event
loop reads it.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014w1Z6Nxy6XTRgkFXqPr9Zh
refresh_rss fired every feed at once, so sites with many subscribed feeds
(e.g. nyaa.si) rate-limited the whole batch. Feeds are now grouped by host:
serial with a 2s delay within a host, still parallel across hosts, with the
existing global concurrency cap kept.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014w1Z6Nxy6XTRgkFXqPr9Zh
/setup/test-downloader now validates the URL scheme (http/https only) like
test-rss already did, and accepts qBittorrent 5.2's 204 login response.
Raw exception and response detail is no longer echoed back from the
pre-authentication setup endpoints; it goes to the server log only.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014w1Z6Nxy6XTRgkFXqPr9Zh
torrents_delete received a Python list, which httpx form-encodes as repeated
'hashes' fields that qBittorrent ignores — pipe-join it as the API expects.
Check the response status instead of logging success unconditionally, and
propagate the failure up through delete_torrent to the API ResponseModel.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014w1Z6Nxy6XTRgkFXqPr9Zh
qBittorrent >= 5.2 answers a successful login with HTTP 204 and an empty
body instead of 200 + 'Ok.', so every login was treated as bad credentials.
Accept both variants and keep 200 + 'Fails.' as a failure.
Also close the httpx.AsyncClient when auth() fails: DownloadClient.__aenter__
raises before __aexit__ can run, so each failed connect leaked a client and
its connection pool (qb and aria2).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014w1Z6Nxy6XTRgkFXqPr9Zh
Closes#984
Apply httpx.Limits(keepalive_expiry=30, max_connections=10, max_keepalive_connections=5) to the qBittorrent AsyncClient — same recipe as #1018 for the RSS side. Prevents the "Server disconnected" cascade when a proxy / NAS silently reaps idle sockets between renamer cycles.
The release artifact only zipped backend/src, omitting pyproject.toml and
uv.lock, so local deployments couldn't install dependencies. #994 also
reported the missing requirements.txt after the uv migration in 3.2.0.
Install uv in the release job and generate a production-only
requirements.txt via `uv export --no-dev`, then bundle all three alongside
src in the app zip.
Closes#994Closes#1015
prefix_process called re.sub(f".{group}.", "", raw) unconditionally. When
group was empty (torrents that don't start with [group]), the pattern
degenerated to ".." and every pair of characters in the title was deleted,
leaving a stub the downstream splitter couldn't turn into title_en/zh/jp.
Guarded the substitution with `if group:`. Titles like
"冰之城墙「氷の城壁」The Ramparts of Ice S01E02 1080p 日英双语-多国字幕"
and "Girls Band Cry S01E05 ..." now parse correctly. Updated #764's test
which had been pinned to the broken behavior.
Closes#1025
Mikanime's mirror (and some CDN-fronted sources) respond with 302 to the
canonical host. httpx AsyncClient defaults to follow_redirects=False, so
raise_for_status surfaced the 302 as an HTTPStatusError and the RSS pull
failed in a retry loop.
Enable follow_redirects=True for every construction of the shared client
(proxy, socks5, and direct branches) via a shared kwargs dict.
Closes#983
When AB runs in a Linux container and qBittorrent runs on a Windows host,
qB returns save_path strings with backslash separators. PurePosixPath
treats those as a single unsplittable segment, so _path_to_bangumi never
matched "Season N" and every non-S1 bangumi collapsed to Season 1.
PureWindowsPath accepts both "\\" and "/" as separators, so using it for
the parsing side handles both layouts. Path generation (_gen_save_path /
_join_path) still uses the platform-conditional Path alias, so AB keeps
emitting native paths for the host it runs on.
Diagnosis and fix credited to the reporter in #1016.
Closes#1016
RSS loop runs every 900s but reuses the same httpx.AsyncClient connection pool, well past server keep-alive expiry (60-120s). Reusing dead sockets produced ConnectTimeout errors over time.
- Configure httpx.Limits with keepalive_expiry=60s so idle connections are dropped proactively
- reset_shared_client() called on retry paths to force a fresh pool after ConnectError
- asyncio.Semaphore(5) in refresh_rss() caps concurrent RSS fetches to avoid tripping site rate limits
Closes#1008
Related #1010, #742, #701
Deleting an RSSItem failed with sqlite3.IntegrityError: FOREIGN KEY constraint failed when child torrents still referenced it, leaving the sidebar entry stuck in the UI.
- delete() now removes referencing torrents in the same transaction before removing the RSSItem
- Adds missing session.rollback() in the exception path
- rss_loop wraps each iteration to skip stale RSSItems deleted mid-loop via API
- Tests enable PRAGMA foreign_keys=ON to match production and gain 5 cascade-delete regression tests
Closes#1010Closes#1017
Startup probe hardcoded http:// ignoring settings.downloader.ssl, causing 5-minute startup delay when qBittorrent is HTTPS-only or redirects http→https.
Now derives the scheme from settings.downloader.ssl.
Reduces container startup time on low-spec devices (TrueNAS etc.) by limiting recursive chown to /app/data, /app/config, /home/ab instead of /app.
Closes#969 (related performance).
- #992: title_parser returns None for non-episodic resources instead of crashing
- #1005: BangumiDatabase.search_official_title method exists and works correctly
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
When raw_parser fails to parse non-episodic resources (movies, collections),
it returns None. Add guard clause in title_parser to skip these gracefully
instead of crashing on attribute access. Downgrade log level from error to
info since this is expected behavior.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
When an MCP client disconnects, Starlette tries to call the return value
of the handle_sse endpoint as an ASGI response. Since the function had
no return statement it returned None, causing:
TypeError: 'NoneType' object is not callable
The mcp library's own SseServerTransport docstring explicitly documents
this requirement. Fix by adding 'return Response()' and importing
starlette.responses.Response.
The per-bangumi default in Bangumi model already uses \d+-\d+ to filter
batch/collection torrents (e.g. '01-13'). The global RSSParser default
was inconsistently using \d+-\d (one fewer +), which has subtly different
matching behaviour.
Align both defaults to \d+-\d+ for consistency.
BangumiDatabase.search_official_title() was called in
notification/manager.py (_get_poster) but never implemented,
causing a RenameThread crash:
'BangumiDatabase' object has no attribute 'search_official_title'
Add the method as a simple exact-match lookup on Bangumi.official_title,
consistent with the existing search_id / search_rss pattern.
When GET /config/get returns config to the frontend, sensitive fields
(password, token, api_key, secret) are masked as '********'. If the
user changes any non-sensitive setting (e.g. ssl: true → false) and
saves, the frontend sends back the masked placeholder verbatim. The
backend was saving it directly, overwriting the real credential with
the literal string '********', breaking authentication silently.
Add _restore_sensitive() to substitute any '********' placeholder with
the current in-memory value before writing to disk. Non-sensitive
fields and genuinely new values are unaffected.
Fixes#995
- Decouple HTTPS scheme selection from TLS certificate verification:
`verify=False` always, since self-signed certs are the norm for
home-server/NAS/Docker qBittorrent setups
- Bump connect timeout from 3.1s to 5.0s for slow TLS handshakes
- Add actionable error messages when HTTPS connection fails
- Fix `continue` → `break` bug in torrents_rename_file verification loop
- Consolidate json imports to top-level
- Add 31 unit tests for QbDownloader constructor, auth, and error handling
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Prevent memory leaks by ensuring the search EventSource connection is
closed when the modal unmounts and setTimeout handles are cleared in
copy-to-clipboard flows across modal components.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Add _fallback_parse() tried when TITLE_RE.match() returns None, using two
regex patterns to extract episode numbers from formats the main regex misses:
- digits before [ bracket (issues #876, #910)
- compound [02(57)] format (issue #773)
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
10 tests covering the full bug chain:
- raw_parser misparses leading number as episode
- TitleParser.raw_parser returns None for unparseable titles
- add_title_alias rejects None and empty string
- _get_aliases_list filters null values from JSON
- get_all_title_patterns skips None title_raw
- match_torrent and match_list handle corrupted data
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Titles like "29 岁单身冒险家的日常" cause the regex to match the leading
number as episode, leaving title_raw as None. This cascades into storing
null aliases and crashing match_torrent with TypeError.
- Fall back to title_jp when title_en and title_zh are both None
- Return None from raw_parser when no title can be extracted
- Reject None/empty aliases in add_title_alias
- Filter null values from parsed title_aliases JSON
- Skip None title_raw in get_all_title_patterns
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Allow users to drag bangumi cards from the "Unknown" section into weekday
columns in the calendar view. Manual assignments are locked so calendar
refresh from Bangumi.tv doesn't overwrite them. A reset button lets users
unlock and send cards back to Unknown.
Backend:
- Add weekday_locked field to Bangumi model (migration v9)
- Add PATCH /api/v1/bangumi/{id}/weekday endpoint
- Skip locked items in refresh_calendar()
Frontend:
- Add vuedraggable for smooth drag-and-drop
- Pin indicator and unpin button on manually-assigned cards
- Drop zone highlighting during drag
- i18n strings for drag/pin/unpin
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>