Full Opus-max adversarial audit found 2 CRIT, 7 HIGH, 9 MED, 9 LOW.
This commit lands all CRIT, all 7 HIGH, the highest-impact MEDs, and
the addon-manifest LOW. The remaining MED/LOW items are cosmetic or
defensive polish — captured in the audit report but not blocking.
CRIT-1: _attach_sponsorblock had a duplicate block stacked on top of
itself. After a video ended the second monitor instantiated, blocked
30s waiting for a player that wasn't there, logged 'timed out' and
returned. Net effect on the family-room TV: Kodi froze for 30s after
every single video. Deleted the duplicate (lines 512–530).
CRIT-2: MPD HTTP server bound to 0.0.0.0 + emitted
Access-Control-Allow-Origin: *. The MPD embeds signed googlevideo
segment URLs — anything on the LAN (guest phones, IoT, malicious sites
loaded in any browser) could scrape the manifest and grab those URLs.
Bind to the LAN IP only with a 127.0.0.1 fallback if that fails; drop
the gratuitous CORS header. Setting is off by default so the live
exposure was small, but locked down before M7 ever flips it on.
HIGH-1: _extract_id accepted arbitrary 'v=' query values without
validating the 11-char [A-Za-z0-9_-] shape. New _validate_id helper +
strict netloc-suffix matching (was: 'youtube.com' in netloc, which let
'myyoutube.com.evil.example' through). Every branch now pipes through
_validate_id. Same shape enforced in the sidecar via
validate_youtube_id() at every op entry point — defense in depth.
HIGH-2: _call_sidecar parser now picks the LAST non-empty stdout line
(robust against future stray println!/log lines), with a dedicated
JSONDecodeError catch that returns a clear 'sidecar stdout was not
JSON: <repr>' rather than a confusing 'sidecar exited 0' message.
HIGH-3: _pv_youtube_installed() now logs the probe exception before
returning False, so silent degradation to 360p is observable in
kodi.log instead of mysterious.
HIGH-4: SponsorBlockMonitor.run()'s getTime() catch widened from
(RuntimeError, OSError) to bare Exception — historically Kodi has
thrown other types when the player goes stale mid-poll, and we don't
want one of those to escape past _play's finally block and leak the
MPD HTTP server.
HIGH-5: Wrapped _attach_sponsorblock in try/except at both call sites
(pv.youtube delegate path + progressive fallback) so a SB monitor bug
can't pop a 'Plugin error' dialog on the TV after successful playback.
HIGH-6: Sidecar Search/ChannelVideos/Playlist now clamp 'limit' at
MAX_LIMIT=200. Prevents an unbounded limit=u32::MAX from OOMing Kodi
on a malicious or buggy addon request.
HIGH-7: Sidecar Rip op now requires dest_dir to be a prefix of an
allowlist (/storage/.kodi/temp/ or
/storage/.kodi/userdata/addon_data/plugin.video.torttube/). Op is
dormant today (no Python caller), but the protocol was a wide-open
arbitrary-write primitive.
MED-1: Bumped search / channel / playlist sidecar timeouts from 15s
to 25s — first-search-after-Kodi-boot on a slow LAN routinely hit 8s+
of rustypipe TLS-handshake + Innertube initial parse.
MED-2: _resolved_listitem now checks urlparse(url).path.endswith('.mpd'
/'.m3u8') instead of substring scan of the whole URL — yt-dlp URLs
have base64 blobs in the query string that can accidentally contain
those substrings.
MED-7: Error classifiers (classify_yt_dlp_error, classify_rustypipe_error)
now match on word-level patterns ('private video', 'age-restrict',
'region-restrict' etc.) instead of bare substrings — fixes
'private network' in a TLS error misclassifying as PrivateVideo,
'package' triggering AgeRestricted, etc.
LOW-1: addon.xml's plugin.video.youtube dep marked optional='true' —
our _pv_youtube_installed() check and fallback paths assumed
optionality; the manifest now matches.
Addon v0.0.13. Verified live via two browse-only smokes:
- Sidecar bad-id rejection ('../../etc/passwd'):
{ok:false, error:'invalid youtube id (length 16 != 11)', kind:'bad_request'}
- Files.GetDirectory ?q=cat: 12 results, formatted labels intact.
NOT VERIFIED VIA PLAYBACK (Leia was watching the TV).
Remaining MED/LOW items (mostly cosmetic): MED-3 endOfDirectory
cacheToDisc, MED-4 0-is-falsy in _format_duration/_format_views, MED-5
codec regex single-quote, MED-6 Response::Ok ok-clobber, MED-8 SB
response-size cap, MED-9 thumbnail dict type check, LOW-2..LOW-9
defensive polish. Tracked in audit report; can hit in a follow-up
sprint.
|
||
|---|---|---|
| addon/plugin.video.torttube | ||
| docs | ||
| scripts | ||
| sidecar | ||
| .gitignore | ||
| LICENSE | ||
| MILESTONES.md | ||
| README.md | ||
torttube
Kodi addon for YouTube via RustyPipe extraction + SponsorBlock segment skipping.
Replaces the dead plugin.video.youtube on LibreELEC RPi TVs after Google
required account-linking for the upstream addon.
Architecture
Kodi (LibreELEC, RPi)
└── plugin.video.torttube [Python addon — UI, browse, SponsorBlock]
├── torttube-sidecar [Rust binary — JSON-over-stdio]
│ ├── rustypipe [Native Rust Innertube for browse]
│ ├── yt-dlp subprocess [Fallback resolve]
│ └── sponsorblock [REST client, SHA-256 prefix lookup]
└── plugin.video.youtube [DEPENDENCY — handles HD playback]
└── inputstream.adaptive [DASH demux + decode]
plugin.video.youtube is declared as a Kodi addon dependency in
addon.xml. When a user installs
torttube, Kodi auto-fetches pv.youtube from the official Kodi addon
repository — user only manages torttube; the dep is transparent.
torttube does what it's faster at: rustypipe-backed search/channel/playlist
browse, SponsorBlock auto-skip via a tight xbmc.Player() monitor loop,
JSON-RPC remote-control for share-to-TV. Playback hands off to
pv.youtube via plugin://plugin.video.youtube/play/?video_id=<id> —
they've spent years getting the DASH-MPD + multi-client Innertube
fallback right. Our SponsorBlock monitor runs in parallel because
xbmc.Player() is a global accessor that works regardless of which
addon initiated playback.
Kodi addons are Python — the engine layer (n-param sig decoding, Innertube, SponsorBlock hashing) lives in a Rust sidecar so we get a single maintained extraction surface and clean aarch64/armv7 cross-compiles.
Three-tier resolve because YouTube actively fights every extractor:
- rustypipe (Rust) — preferred. Fast, in-process, no Python dep on the RPi.
- yt-dlp subprocess — fallback when rustypipe sig-decoding falls behind YouTube's deobfuscator changes. yt-dlp updates weekly; we shell out, parse
-jJSON. - Rip-to-temp — last resort when stream URLs 403 mid-playback (poToken expiry, cookie session mismatch). yt-dlp downloads to
/storage/.kodi/temp/torttube/<id>.<ext>, Kodi plays the local file. Temp dir has size cap + age cleanup.
Status
M0 scaffold. Nothing playable yet — see MILESTONES.md.
Upstream — we fight with the FOSS extractor ecosystem, not next to it
YouTube's anti-scraping changes hit every extractor: NewPipe, yt-dlp, Invidious, rustypipe. Every fix we make in our sidecar gets evaluated for "is this upstreamable?" — if yes, the fix lands at the upstream project, not just here.
Active lanes:
- rustypipe (Rust, codeberg.org/ThetaDev/rustypipe) — maintenance has slowed.
Open PR #77 "Some fixes" is unmerged as of 2026-05-23. We will either help land
it (review + ping maintainer) or fork to
Sulkta-Coop/rustypipeif upstream stays quiet. Forking is the worst case, not the first move. - NewPipeExtractor (Java, github.com/TeamNewPipe/NewPipeExtractor) — actively maintained, 177 open issues. We use it as the reference implementation for Innertube behaviour. PRs to NPE land in Rust here via rustypipe, and vice versa.
- yt-dlp (Python, github.com/yt-dlp/yt-dlp) — the gold standard. We're more consumers than contributors here, but if our rip-to-temp tier surfaces a specific extractor bug we file it.
Issues we're watching:
- NPE #1339 — n-parameter deobfuscation
- NPE #1444 — distinguish unavailable vs unextractable
- NPE #1360 — refactor link handlers (help wanted)
- NPE #1357 — JDoc checks in PR pipeline (good first issue)
- rustypipe PR #77 — open as of 2026-05-23, unmerged
Contribution log lives at docs/upstream.md — every PR we file lands there with its outcome.
License
GPL-3.0-or-later. Matches RustyPipe and NewPipeExtractor.