You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Sign-in without a password opened the API from anywhere — With UI
authentication turned off, any client could ask for a session without a
password, and that session opened every API route and handed out the API
key, although a key was set. Over the plugin installer this reached code
execution. A password-less session is now only given to clients on the local
network (loopback, private ranges, Tailscale), never through a reverse proxy
and never before the first setup, and it ends as soon as UI authentication
is turned on. Behind a reverse proxy, turn UI authentication on.
The first setup could be claimed by anyone — Until UI authentication is
set up, the setup now needs the local network or a one-time token that
Sublarr writes to the container log at start.
The API key alone no longer switches sign-in off — Turning UI
authentication on or off with the API key now also needs the UI password.
Other apps on the same host could act with your session — A page of
another app on the same host (another port) could make the browser send
changes to Sublarr with its sign-in cookie. Changes that rely on the cookie
are now refused when they come from another origin.
Sessions outlived logout and password changes — Logging out now ends
that session on the server too, so a copied cookie stops working; changing
the password or turning UI authentication on or off ends every other
session. Existing sessions end once with this update; sign in again.
Webhooks no longer need the API key — Sonarr, Radarr and Jellyfin held
the API key for their webhooks, and that key could install plugins: whoever
read their settings could run code in Sublarr. Webhooks now have their own
secret, generated on first start and shown under Settings → System → Hooks,
which opens the webhooks and nothing else. The API key keeps working on
webhooks until you turn "Keep accepting the API key on webhooks" off there;
do that once every connection uses the webhook secret.
Plugins install only from the plugin index — The installer took the
download URL and hash from the request, so any HTTPS file could be
installed. URL and hash now come from the plugin index, installing or
removing a plugin needs a signed-in session (the API key alone is refused),
plugin names are checked (the name "." removed every installed plugin), and
each download redirect is checked like the first URL.
Sync engine installs are pinned — "Install alass" downloaded whatever
the latest GitHub release was and made it executable without checking it,
and "Install ffsubsync" installed any version. alass is now version 2.0.0
and only installed when its SHA-256 matches; ffsubsync is pinned to 0.5.1.
alass 2.0.0 ships a Linux x86-64 binary only; on other systems install it by
hand.
Fonts came from Google — The interface loaded Inter and JetBrains Mono
from Google's servers, which sent every visitor's IP address to Google. Both
fonts now ship with Sublarr, and the Content Security Policy no longer
allows Google's font hosts.
Fixed
Wanted scans on large libraries never completed (#248) — A full scan
that needed longer than the job's two-hour ceiling was stopped, pruned
nothing and kept its watermark, so every following scan was full again and
stopped again; rows for moved or deleted files stayed forever. A stopped
full scan now resumes where it stopped on the next run, and the pass
finished in parts prunes against everything its parts covered. A stopped
scan also still removes rows whose file is gone or whose subtitle is
already there, stops between series instead of fetching and probing the
rest only to discard it, and releases the scan lock before its final
report, so a stalled report no longer answers every refresh with 409.
Thanks @Svnow132 for the detailed report.
A Sonarr hiccup could delete a series' wanted items — When Sonarr did
not answer the episode list of one series during a full scan, the scan saw
that series as empty and its cleanup removed all its wanted items (one
series never trips the cleanup fuse). Its items are now kept.
Dubtitle sweep only ever looked at the first 1,600 files — Every run read
the same alphabetically first paths, so a larger library was never covered.
The sweep now continues where the last run stopped.
Standalone: a profile read error deleted pending items — The scan fell
back to the global language and "no forced subtitles" and dropped the
pending items of the title's real languages. A title whose profile cannot be
read is now skipped and logged.
Standalone: deleted or moved films stayed in the library — Films whose
file is gone are now removed with their wanted items (only under enabled,
reachable watched folders); a moved film no longer shows up twice.
"Already running" until a restart — The dedup scan, the subtitle repair,
the wrong-episode requeue and the MT approval batch now release their
running flag when their background worker cannot be started.
Foreign-track sweep never finished on slow storage — Every run walked
the library from the start; a walk longer than the job's two hours (a
Raspberry Pi, a slow network share, a very large library) was stopped every
time, so the sweep never got past it. A stopped walk now continues where it
stopped. Files in a folder the walk cannot read keep their cached result
instead of being treated as deleted.
A stopped wanted scan waited for every probe of the current series —
Probes not yet started are now dropped when the scan is asked to stop.
MT: a profile read error deleted the wanted item — With "keep seeking
the original" on, the item was deleted instead of kept provisional when the
profile could not be read. It is now left alone until the profile reads.
Scheduler: a stalled history write blocked the job — The run-history
row is now written after the job lock is released.
An offline share deleted its library's items — Cleanups treated every
file they could not see as deleted: a NAS that dropped, an I/O error or a
share that was not mounted removed wanted items (up to the fuse), standalone
series, movies and episode ids, download history and probe cache entries. A
file now counts as deleted only when the disk vouches for it; unreadable or
unmounted ones are kept. A full wanted scan also keeps the items of episodes
it could not read, and standalone cleanups only judge enabled watched
folders. A disk whose mount folder disappears on unmount (udisks under
/media, autofs) is treated as off, not deleted.
Standalone: a large first folder starved the others — Folders are now
scanned least recently scanned first, so a pass the scheduler stops after a
slow first folder continues with the next one.
MT re-seek starved behind pinned items — Pinned items and items whose
profile stopped seeking no longer hold the front of every batch.
Several Sonarr or Radarr instances — Profile assignments are keyed by the
title id, which two instances number independently. While more than one
instance is configured, the scan no longer drops items for languages "not in
the profile", since that profile may belong to another instance's title.
Windows path mappings ignored case — D:\TV and d:\tv now map the same.
Lighter on small hardware — The upgrade scan reads four columns instead
of whole download rows and commits in batches; a large series' episode ids
are written in one transaction; the wrong-episode audit loads only the
newest download per file; the probe-cache cleanup deletes in chunks below
SQLite's parameter limit.
"Keep seeking the original" never sought it for translated ASS — The
scan filed an item served by Sublarr's own .ass translation as provisional,
and the cleanup right after it removed the item because an .ass was there.
Every full scan added and removed the same items (about 1,600 on a test
library) and the re-seek never saw them. The cleanup now keeps them. An
embedded track next to our own .ass no longer sends the item into an
extraction that never writes beside an .ass.
Provenance lookups scanned the whole download history — The index on subtitle_downloads.file_path is back (migration), so scans and cleanups
no longer read the full table once per file.
Wanted scans slowed down on libraries with many subtitles — The check
for signs-only sidecars keeps a cache of each file's dialogue count. It held
8,192 files and was emptied whenever it filled, so a library with more
sidecars than that parsed every file again on every scan (a cleanup pass on
a test library: 172 s instead of 59 s). The cache now holds 65,536 files and
drops only its oldest entries when full.
Errors no longer vanish silently — 44 places that caught an error and
carried on without a trace now write it to the debug log.