These are things Arclight does not do, or does only partly, at the time of writing. They come
from the package READMEs (packages/torrent-engine, packages/subtitles,
packages/download-clients), the specs in docs/spec/, the API reference (docs/API.md), and
what the UI really contains compared with docs/UI-INVENTORY.md.
Use this file while testing: if a check fails in a way described here, record it as Known
rather than as a new bug. If behaviour is worse than described, that is a real bug. When a
limitation is fixed, delete its row so the file stays honest.
“Where you notice it” names the checklist file or row that touches it.
The engine (@arclight/torrent-engine) is a from-scratch BitTorrent client written in
TypeScript. It is meant to be good enough to replace qBittorrent for ordinary public and
private-tracker use, not to match every feature.
Limitation
Effect
Where you notice it
No MSE/PE (message stream) encryption. The Torrent Engine setting for encryption accepts only “disabled”.
Peers that require encrypted connections cannot connect. Some swarms are slower than in qBittorrent. A tracker that forces encryption cannot be used.
05 (engine), 06
No uTP (BEP 29); TCP only.
Fewer reachable peers behind some NATs. Slower on congested links than clients with uTP.
05, 11
No IPv6 (no IPv6 peers, no peers6, no DHT over IPv6).
IPv6-only peers are never used.
05
No web seeds (BEP 17/19). url-list in a .torrent is read but ignored.
Torrents that depend on an HTTP seed with no peers stay at 0%.
05
No v2-only torrents (BEP 52). Hybrid torrents download through their v1 data; v2-only torrents go to the error state.
Rare; shows as an errored torrent.
05
No super-seeding, no secure DHT node ids (BEP 42), no read-only DHT or BEP 44, no UPnP or NAT-PMP port mapping, no IP-filter file formats.
Port forwarding must come from the VPN (Proton via Gluetun) or the router by hand.
05, 11
Fast extension is partial: incoming suggest and allowed fast messages are ignored and the engine never sends allowed-fast sets.
Slightly slower start on some swarms.
n/a
Rate limits are global and per connection. There are no per-torrent limits.
Cannot cap one torrent below the others.
05
Engine state lives in the config folder under torrents/ (fast-resume data). Losing the folder loses partial-download resume points.
Torrents recheck from scratch after restoring a backup that omits it.
rTorrent has no per-torrent seed limits in its API; whatever stops a finished torrent counts as the seeding policy.
A stopped finished torrent is treated as “seeding complete”.
05
A torrent stopped with no limit is removable only when it is known to be public (qBittorrent, Transmission).
Private-tracker torrents are left in the client after import when they have no seed limit.
05, 06
Completed-download handling, failed-download handling and auto re-search switches (enableCompletedDownloadHandling, autoRedownloadFailed) are stored in Media Management settings but have no screen.
Change them with PUT /api/v1/settings/mediamanagement.
05 (DLC-25), 06
The usenet clients (SABnzbd, NZBGet) need a real server and account to test end to end; automated tests use stubs only.
No CAPTCHA or Cloudflare solving. Captcha-walled providers report captcha and are throttled for 12 hours. Affected: Addic7ed (login reCAPTCHA), Subf2m (Cloudflare challenge), and any scraper site behind a challenge.
Provider shows throttled in Settings -> Subtitles until the throttle ends or you use “Reset throttles”.
08
Podnapisi is “best effort”: the upstream API shape could not be verified because the original provider was removed. Treat any result as a bonus.
May return nothing or error.
08
Providers marked implemented but not tested against live services in this build: gestdown, animetosho, jimaku, yifysubtitles, napisy24, embeddedsubtitles (needs ffprobe/ffmpeg), whisperai (needs your own Whisper-ASR web service). Only opensubtitles.com and subdl are expected to work with just an account.
Provider Test may fail for reasons outside Arclight.
08
Napisy24 requires your own credentials (no shared fallback account). Quotas for several scraper providers (Napisy24, Gestdown, Subf2m, YIFY) are unmeasured.
You may be rate-limited without a clear message.
08
rar and 7z subtitle archives are refused; Zip64 and password-protected zips are not supported. XZ support covers the LZMA2 filter only.
Some downloads fail with “unsupported archive”.
08
Only text subtitle tracks are extractable (ASS, SubRip, WebVTT). Bitmap subtitles (PGS, VobSub) are never offered and there is no OCR.
Embedded PGS tracks are listed as present but cannot be extracted or synced.
08
Sync needs ffsubsync or alass installed on the host; neither ships in the Docker image.
Sync reports “tool not found”; the toggle can still be saved.
08
Subtitle translation, mods catalogue beyond the first few, and mojibake repair are not implemented.
Wrongly encoded files stay wrong.
08
Provider throttle state is held in memory by the subtitle package; the API module persists and restores it.
If persistence fails, a restart clears throttles early.
08
Blacklisting is possible only from the per-file subtitle menu; the Blacklist page lists and removes entries but cannot add one (add via POST /api/v1/subtitle/blacklist).
Bundled public-tracker definitions (1337x, eztv, limetorrents, nyaa, thepiratebay, torlock, torrentgalaxy, yts) are untested against the live sites in this build. Public sites move domains and add challenges often.
Test may fail or return no rows; try a mirror via “site link”, add a FlareSolverr proxy, or use a Torznab URL.
04
No remote definition catalogue: only the bundled set plus your own YAML files in <config>/definitions/. There is no “outdated definition” health check.
New trackers need a hand-written definition.
04
The definition engine does not paginate: one page per search path. Template functions beyond the documented set (now, pipelines, arithmetic) are not supported. Only urlencoded POST is supported through FlareSolverr.
Deep interactive searches return only the site’s first page.
04
FlareSolverr is stateless: every solve is a new session, so first requests to a protected site are slow.
04
Custom YAML definitions have no UI editor; upload and delete exist in the API only (POST /api/v1/indexer/definition, DELETE /api/v1/indexer/definition/:id, POST .../reload).
04
Arclight does not expose a Torznab server for other apps and cannot push indexers to a separate Lidarr or Readarr. Other tools (Prowlarr) push into Arclight through the compat API instead.
Node’s fetch cannot bind to a local address. The native public-IP check (publicIpCheckUrl, default https://api.ipify.org?format=json) follows the operating-system routing table. With a full-tunnel WireGuard config (AllowedIPs 0.0.0.0/0) it goes through the tunnel; with a split tunnel it may not.
The VPN card can show your ISP address while the engine is correctly bound. Set “home public IP” so a check that leaves via your ISP is flagged as a leak instead of passing silently.
11
The Windows WireGuard provider needs the official WireGuard for Windows client and an elevated (Administrator) process.
Without elevation, Connect fails with a permission message.
11
With the Docker VPN stacks Arclight shares Gluetun’s network. Stopping Gluetun also cuts the UI.
The kill-switch drill (VPN-30) is done from the Docker host with docker stop gluetun, then watching the Docker host side, not through the browser.
11
Newer Gluetun releases require authentication on the control server (API key or password) and the Proton port-forward port can take about a minute to appear after connect.
Forwarded port shows empty until Gluetun reports it.
11
Only Gluetun, WireGuard for Windows and interface binding are supported. No OpenVPN management, no other VPN client integrations.
11
Port forwarding is learnt from Gluetun only; the engine has no UPnP or NAT-PMP of its own.
Clients that wait for push updates poll or show stale data until refreshed.
14
calendar and wanted/missing on /api/v3 are episode-only (Sonarr shape). Radarr’s movie calendar is not served on the shared path.
Radarr-only apps do not show upcoming movies from Arclight.
14
Lookups are TMDB-backed: tvdbId is 0 unless TMDB knows it; TV Rage and TV Maze ids are always 0; airTime is UTC, not network time.
Some mobile apps show odd times or cannot search by TVDB id.
14
Read-only on /api/v3: customformat, delayprofile, releaseprofile, downloadclient, notification, importlist, config/*, and system/backup (always empty).
Recyclarr can sync quality profiles but not every setting.
14
POST /api/v3/indexer/test validates the body only; it does not contact the indexer.
Prowlarr’s “Test” always passes.
14
Stored indexer base URLs are normalised (trailing slash removed), so Prowlarr may report a difference and resend the indexer on its next sync.
Harmless repeated updates.
14
Queue records expand multi-episode downloads into one record per episode. languageprofile is a single fixed “Any” profile. Custom-format specification values are best-effort translations.
Approved requests add the title directly to the Arclight library; there are no separate Radarr/Sonarr “server” targets as in Seerr. 4K is a second profile on the same item, not a second server.
Differs from docs/spec/requests.md section 9.
09
The Approve button has no override options; set profile, root folder and tags by editing the pending request first. The API accepts overrides on approve (POST /api/v1/request/:id/approve with root folder, profile, tags, monitor option, minimum availability, seasons).
09
Out of scope in this build: a Plex library scanner for availability, voting, and the “Recently added” discover slider. By default availability comes from the Arclight library only; the optional “use media server availability” setting also asks Jellyfin or Emby connections (not Plex).
A title that exists on Plex but not in Arclight shows as not available.
09
Sign-in through Jellyfin/Emby/Plex accounts is not built; users are local accounts created in Settings -> Security.
01, 09
Seerr-style partial-request enforcement on the server for partialRequestsEnabled = false is unverified.
09
A failed subtitle or language request sends no notification; created, approved, declined and completed do.
The requester sees “Failed” in My requests only.
09
Language requests have no entry in universal search.
Find them on the Requests page under “Subtitles & languages”.
09
The Wanted page shows no unreleased warning for season or series interactive searches.
The warning banner appears only when searching a single item.
07
Movies on Wanted show the date that counts for their minimum availability rather than a single release-date column.
The cinema, digital and physical dates can differ from the one used for “Upcoming”.
The inventory in docs/UI-INVENTORY.md lists these as planned or partly built. They are not
bugs unless the checklist for that area says otherwise.
Area
What is missing
Library
No poster-size slider, no Overview view, no bulk Refresh/Search/Tags actions (API: PUT /api/v1/series/bulk and /movie/bulk take tags, refresh, search), no tag-management page, no import-exclusions screen, no “import existing library from folder”. The series “Refresh” button is greyed out (API: POST /api/v1/command/refresh). No series History tab; the movie History tab shows “History is coming soon”. Rename preview, rescan and library statistics are API only.
Calendar
Month and Agenda views only. No week or day view, no tag/type/date-kind filters, no popovers. No “regenerate token” button for the iCal feed (API: POST /api/v1/calendar/feed/regenerate; GET /calendar also takes mediaType).
Wanted
No filters and no monitored toggle on Missing. Cutoff Unmet shows “Cutoff unmet is not available yet”; the native API now has GET /api/v1/wanted/cutoff/episodes and /movies but the page is not wired to them yet.
Activity
History has no Redownload action (API: POST /api/v1/history/:id/redownload); the Imported, Failed and Ignored filter chips use event names the API does not store, so they may show nothing. The Queue has no status filter. Blocklist cannot be cleared or added to from the UI (API only). “Mark as failed” cannot set skip-redownload.
Settings
Notification flags enabled and includeHealthWarnings, compat settings, and the log level have no screen. Changing your own password has no screen (API only: PUT /api/v1/user/me/password). Settings -> General still shows a stale “Backups are coming soon” card; backups are under System -> Backups.
Shell
The newer UI kit pieces (Backdrop, Rail, CommandPalette, KeyboardSheet, MobileTabBar) are not mounted in the live app. The only keyboard shortcut is / (focus search). No keyboard-shortcut help sheet, no command palette.
Mobile / PWA
No web app manifest and no service worker, so “Install app” is not offered. The layout collapses to a drawer below 900 px; there is no bottom tab bar and no push notification.
System
No Updates page (current/available, changelog) and no audit/events log.
One API key per install, and it always grants admin. There are no per-user API keys.
Give the key only to trusted tools.
Roles are admin and user; finer control is through the permission bitfield managed in Settings -> Users -> Manage.
No two-factor sign-in, no external identity provider.
With method “None on the local network” only loopback, private-range and link-local addresses skip sign-in. If Arclight sits behind a reverse proxy on the same LAN that forwards public addresses without X-Forwarded-For, every visitor looks local. Put the proxy’s authentication first.
Failed logins are limited to 10 per address per 15 minutes (HTTP 429). The counter is in memory and resets on restart.
Wave-2 features (cross-seed, trackers, transcode, Discord, scene numbering, policies)
Reported by the modules’ authors and docs/API.md; there are no QA checklist files for these
areas yet.
Limitation
Effect
Cross-seed handles single-file torrents only. Season packs, multi-season releases, multi-file torrents and magnet-only releases are skipped.
Only individual episode or movie files are cross-seeded; a library built from season packs gets few or no matches.
Cross-seed adds only to the Built-in engine or qBittorrent, and verifies only the first and last piece before seeding with checking skipped.
Other clients are not used. A file that differs in the middle is only caught by the client later, as a failed piece.
Cross-seed links a differently named torrent with a hard link and falls back to a symlink when the link folder is on another filesystem.
Deleting or upgrading the library file breaks a symlinked cross-seed.
The hit-and-run guard is enforced only by the Built-in download client.
qBittorrent and other external clients can still remove a torrent before its seeding rule is met.
Tracker account statistics (ratio, buffer, H&R count, bonus) are scraped only for definitions that declare an account block; torrents are attributed to trackers by grab history and announce hosts.
Other trackers show engine totals only; a torrent added by hand with an unknown announce host is not counted.
Freeleech-only mode trusts the indexer’s freeleech flags and download volume factor (rejection notFreeleech).
An indexer that does not report freeleech rejects every release while the mode is on.
Request approval by size (autoApproveIfSizeBelowGb, budget autoApproveUnder.sizeGb) works only for titles already in the library; the size comes from a quick interactive search (20 s).
Requests for new titles are never size-approved. In movieGb budgets, unknown sizes count as 0.
Approval windows use server local time.
Users in other time zones see holds and releases at shifted hours.
The real GPU paths (VA-API, Quick Sync, NVENC) are not exercised in CI, which has no GPU. The code and error handling are tested with fake ffmpeg output only.
Report GPU-specific failures with the Reason text from Settings -> Transcode -> Hardware acceleration.
The arm64 image has no VA-API or Quick Sync (Jellyfin’s ffmpeg does not include them there). Only NVENC can work.
On a Raspberry Pi or other ARM host, VAAPI and Quick Sync always show Unavailable.
Hardware test results are cached in memory until Arclight restarts or you press Test again.
A GPU added or removed after start is not noticed until you test again.
NAS models whose CPU has no integrated GPU have no /dev/dri.
Hardware encoding is unavailable and jobs run on the CPU (or fail, if “Fall back to software” is off).
Transcoding runs one file at a time. There is no HDR tone mapping, mp4 output drops image subtitles, and hardware encoding needs the device passed into the container (/dev/dri for VAAPI/QSV, the NVIDIA runtime for NVENC).
An HDR source encoded to h264 keeps PQ/HLG tags on 8-bit video. PGS/VobSub tracks are lost when remuxing to mp4.
runAfterImport transcodes wait in memory (up to 6 h) for media analysis. With keepOriginal and a container change, the original stays on disk but is no longer tracked by the library.
A restart while a file waits drops that automatic transcode (run it by hand). The untracked original is not cleaned up.
The community scene-numbering feed is not published yet (/scenemapping/feed reports it as not published). Applying rules re-registers remapped files with new file ids.
Only manual rules and the anime absolute fallback apply. Records keyed by the old file id (subtitle files, cross-seeds, transcode jobs) are not carried over to the new id.
The Discord bot posts only request created, approved and completed events; /request for a series requests all seasons; /pending shows at most 5.
Declines and failures are not posted. Request single seasons from the web UI.
The engine agent accepts only the API key (not a browser session) and always answers 409 to a remote listen-port change. It serves nothing while the built-in engine is stopped.
The driving instance’s “follow VPN port” has no effect on the agent; set the port on the agent instance.
Multi-version libraries allow at most 8 version slots per series or movie, one file per slot.
Whisper subtitles need an external Whisper ASR service (ARCLIGHT_WHISPER_URL, compose profile whisper); it is not bundled in the image.
Without it, /subtitle/whisper/status reports it as not configured and the provider is never used.
Container update items carry the container’s Docker labels (labels), but the Updates page does not show them yet.
Multi-season packs are rejected (“Multi-season releases are not supported” by the decision engine; “Multi-season download, unable to import automatically” on import).
Grab season packs one season at a time.
Partial season packs and season extras are not imported automatically (“Partial season packs are not supported”, “Extras are not supported”).
Use Activity -> Manual Import.
Archive files (.rar, .zip) in a completed download are not extracted by Arclight (“Found archive file, might need to be extracted”).
Configure the download client to unpack, or extract by hand then Manual Import.
Language profiles for audio language live on the quality profile; subtitle language profiles are separate (Settings -> Subtitles).
Metadata comes from TMDB only (no TVDB fallback, no image caching on disk).
Series without TMDB episode data show gaps; posters load from TMDB each time.
Docker image includes Jellyfin’s ffmpeg and ffprobe but not ffsubsync or alass.
Subtitle sync with the tool method is unavailable unless you install them in a derived image; the builtin sync check works.
Docker permission and read-only-mount health checks assume a Linux container user mapping (PUID/PGID).
docs/ARCHITECTURE.md, docs/ROADMAP.md, docs/API.md (module list and the “not built yet”
statements) and the Status paragraph in README.md still describe an earlier phase. The code is
the authority. docs/UI-INVENTORY.md marks pages as unbuilt that now exist (Discover, Requests,
Issues, Watchlist, Subtitles, Indexer Proxies, Indexer Stats, Import Lists, Backups, Import,
VPN, Torrent Engine) and lists features that do not (week view, bulk tags, command palette).
Open questions from the specs (undecided behaviour)
Podnapisi replacement; which Whisper service shape; legal and terms-of-service review for scraper providers; CAPTCHA strategy (build, require a solver, or decline).
Subtitles (docs/spec/subtitles.md, section 16)
Backoff reset when a file is replaced; tie-breaking of equal scores; embedded extraction precedence over downloads; post-processing timeout (default 60 s); history retention after item removal.
Remote catalogue; licensing of bundled site definitions; default timezone for dates without a zone; cookie-jar fidelity; seeding-rule precedence on private trackers; whether failed queries count against query limits.
Health checks (docs/spec/health-checks.md, section 8)
Update-available logic; exact meaning of “folder empty” and “folder writable”.
Requests (docs/spec/requests.md)
Whether quota windows without days mean unlimited; server-side enforcement of partial requests.
