Un playlist EVENT cuyas entradas empiezan en 0 mientras los fragmentos
llevan tfdt desplazado (-ss + -output_ts_offset) es exactamente la forma
que el parser HLS nativo de iOS no traga: resume a 368s → error del player
y bucle de re-bootstrap de sesión en iPhone (observado 2026-06-10).
Copy produce siempre desde 0 con PTS absolutos reales: adelanta a la
reproducción a velocidad de I/O, así que el punto de resume aparece en la
timeline creciente en segundos y el seek de startPosition del player
aterriza con normalidad. Test de resume actualizado: el playlist debe
cubrir la timeline completa.
Nuevo modo VideoCopy en el engine HLS: ffmpeg -c:v copy (el vídeo jamás se
re-encodea — I/O puro, funciona en un NAS sin GPU), audio copy si ya es AAC
o AAC 192k si no, muxeado a segmentos fMP4 con ffmpeg escribiendo SU PROPIO
playlist (EVENT mientras corre, ENDLIST al acabar, EXTINF exactos en los
keyframes del source). Sustituye al remux growing-fMP4 servido por HTTP
Range artesanal, cuya fragilidad estructural produjo tres incidentes en un
día (init malformado/delay_moov, loop de re-seek por total inventado, iOS
rechazando total desconocido).
Diferencias deliberadas respecto al modo encode:
- playlist de ffmpeg servido desde disco (los cortes van a keyframe del
source → duraciones imposibles de pre-renderizar; medido: probar
keyframes antes cuesta 8-24s, inviable para TTFF)
- sin seek-restart ni auto-restart (la copia va a velocidad de disco y
adelanta a cualquier viewer; el -ss de segmentos uniformes corrompería
la timeline de cortes variables)
- sin caché HLS (regenerar no cuesta encode; cachear solo quema disco)
- resume vía -ss (snap a keyframe) + -output_ts_offset
- master playlist sin CODECS (un string hardcodeado equivocado hace que
iOS rechace la variante; omitirlo es legal y universal)
Validación: TTFB seg-0 510ms sobre el MKV real del incidente (HEVC Main10
+ EAC3, 6.7GB). Suite de integración con ffmpeg real (tag smoke): h264+aac
(copy total), h264+ac3 (re-encode de audio con priming dts — la clase
delay_moov), hevc10+eac3 (la forma exacta del incidente, tag hvc1), resume
con StartSec, y serving del playlist; asserts de codecs vía ffprobe sobre
el playlist servido, suma EXTINF ≈ duración, segmentos completos en disco
(+temp_file = rename atómico).
El wiring web (plan remux→hls+videoCopy con gate de versión ≥1.0.10) va en
el repo web. Plan: docs/plans/hls-copy-remux-replacement.md (web).
iOS/WebKit abre todo <video src> con una sonda "bytes=0-1" y se niega a
reproducir si el 206 no trae una longitud concreta en Content-Range —
"/*" (total desconocido, el fix anterior del loop de re-seek) le hacía
abortar y re-bootstrapear la sesión sin parar.
Vuelve a anunciar siempre un total numérico (exacto si ffmpeg terminó, el
estimado mientras crece). El loop de re-seek real no era el total
anunciado sino el init segment malformado, ya arreglado con +delay_moov
en buildFFmpegArgs. Test nuevo: la sonda 0-1 debe llevar total concreto.
Keep an NVENC downscale of an SDR source entirely on the GPU
(decode -> scale_cuda -> h264_nvenc) instead of copying every frame to the
CPU for `scale=` and back. That GPU->CPU->GPU round-trip is the wall on
modest GPUs; even a strong box gains ~37% (scale_cuda 14.9x vs CPU 10.9x
on a 4K SDR HEVC -> 1080p encode).
Strictly gated so every case that needs CPU frames is unchanged:
- HDR (libplacebo Vulkan / zscale CPU tonemap can't consume a CUDA surface),
- burn-in (the scale2ref+overlay composite runs on CPU frames),
- non-NVENC encoders, and no-op when not actually downscaling.
- hwscale.go: FFmpegSupportsScaleCuda — a functional 1-frame probe mirroring
the libplacebo probe (presence in -filters lies; needs a real CUDA device).
Probes the worst-case real input (10-bit p010 -> 8-bit yuv420p) so a host
whose scale_cuda can't do the 10->8-bit conversion fails closed to CPU.
- hls.go: useCudaScale gate + `-hwaccel_output_format cuda` + a
`scale_cuda=-2:H:format=yuv420p` filter branch. Output is 8-bit
(format=yuv420p + `-profile:v main`), browser-safe.
- transcode_quality.go / player_session_registry.go / daemon.go: HasScaleCuda
flag, populated + warmed at startup like the other ffmpeg capability probes.
Fail-closed: probe absent/fails -> keep the CPU scale path, no regression.
Verified live (real 4K SDR HEVC Main10 session emitted scale_cuda, 5.54x
realtime, nvenc at 100%) + 8 arg-builder unit tests for the gate.
El remux reencodea el audio no-AAC (eac3→aac); la pista AAC arranca con un
dts negativo (priming/encoder-delay). Con empty_moov el moov se escribía
ANTES de conocer ese delay, así que el primer fragmento quedaba mal formado
y un demuxer estricto (Safari / la forma en que Apple decodifica HEVC) nunca
inicializaba la reproducción: el <video> cargaba bytes (se veía en Network)
pero no arrancaba, y el player re-bootstrapeaba la sesión cada pocos segundos.
Añade +delay_moov: retiene el moov hasta el primer paquete y maneja el dts
de priming. ffmpeg deja de emitir el warning "nonzero dts ... moov already
written" y el fMP4 reproduce. Reproducido con Hoppers (HEVC Main 10 + EAC3).
serveGrowing anunciaba en Content-Range total = EstimatedSize() = el tamaño
del MKV fuente mientras ffmpeg aún corría. Pero el fMP4 resultante no mide
eso (el audio re-encodea a AAC y la fragmentación cambian el byte count), así
que el <video> nativo mapeaba su timeline sobre una longitud falsa, pedía
offsets que no cuadraban, re-seekeaba y reabría la conexión cientos de veces
por segundo (el loop de reproducción remux).
Mientras crece (!Final) la longitud real es DESCONOCIDA: ahora se sirve
Content-Range "bytes start-end/*" (RFC 7233 §4.2) sin Content-Length, y el
cliente lee secuencial en vez de re-seekear. Cuando ffmpeg termina, el tamaño
real se conoce y se anuncia como antes. El 416 y el Content-Length del HEAD
solo cuando el total es real (final).
- failSession usa un contexto fresco (no el del daemon): los fallos se
concentran justo cuando el daemon se apaga (la cancelación mata arranques
en vuelo) y un report derivado de ese contexto moría antes de llegar a la
web; el cap de 10s sigue acotándolo
- consts sessErrFfmpegMissing/sessErrStartFailed sustituyen los 7 literales
inline (un typo habría producido un code que el z.enum de la web rechaza
con 400 — exactamente el fallo mudo que este canal elimina)
- markReady() unifica los tres goroutines idénticos de MarkSessionReady de
los caminos sin transcode (direct-play, remux, debrid direct)
- nuevo agentClient.ReportSessionError → POST /agent/session-error;
failSession() en todos los abortos del handler de sesiones (path muerto,
ffmpeg ausente, remux, provider debrid, StartHLSSession). Antes eran
returns mudos y el player quedaba en "Preparando sesión" hasta agotar el
deadline de probes
- resolvePlayableFile() unifica la resolución de paths del /stream raw y de
las sesiones HLS/remux/direct (remap de base path + stat con retries NFS +
directorio→vídeo, antes duplicada y divergente) y distingue file_missing
(la web self-heala filas stale) de path_rejected (el fichero existe fuera
de los roots = config; la web no debe podar nada)
- library.SyncBatches: el batching del sync de biblioteca vive en un solo
sitio; el scan manual y el auto-scan sincronizan todos los roots en UNA
sesión con scanRoots/fullCycle, en vez de una sesión por root que dejaba
al server podar filas de roots que la sesión nunca visitó
El web empuja SubtitleFetchRequest por el sync (URL del proxy, ya
charset-fixed a WebVTT); el daemon lo descarga y lo escribe como sidecar
<base>.<lang>.vtt junto al medio (contención en scan paths con
EvalSymlinks, cap 10 MiB) y reporta done/failed en el siguiente sync
para que el web marque el job. Config nueva library.subtitles
(auto_fetch + languages) para el auto-fetch en scan, off por defecto.
NVENC (ffmpeg 6.1 + drivers actuales) emite los keyframes forzados por
-force_key_frames como I-frames NO-IDR; el muxer HLS solo corta en IDR,
así que cada segmento se estiraba en silencio al GOP por defecto
(250 frames ≈ 10.4 s @24fps) mientras la playlist server-side seguía
prometiendo 2 s por segmento. Con los PTS reales ~5× fuera del mapa de
la playlist, los seeks aterrizaban donde podían y los subtítulos se
desincronizaban en cuanto se mezclaban segmentos de runs distintos
(seek-restart) en el mismo dir.
Medido: 3 segmentos por 30 s de encode en vez de 15; con -forced-idr 1
exactamente 15, y post-fix seg-150/151/158 arrancan en 300.0/302.0/316.0
clavados. Afecta a TODO el HLS por NVENC histórico (no era del rate
control nuevo: la config de bitrate fijo producía lo mismo). QSV recibe
su grafía -forced_idr. Las entradas de caché viejas nunca llegaron a
sellarse (el conteo de segmentos no cuadraba), así que no hay migración:
solo sesiones vivas estaban afectadas.
- StreamSession.Prewarm → HLSSessionConfig.Prewarm: el daemon difiere el
encode de un prewarm hasta que no haya encode vivo (poll 10s, tope
30min) y lo registra vía RegisterKeep (side-by-side, sin desalojar).
Antes todo pasaba por Register(), que cierra las demás sesiones — un
prewarm de next-episode reclamado en mitad de la reproducción mataba
el stream del usuario ("closed (cache discarded)" → master 404,
verificado 2026-06-10). Una sesión REAL nueva primero reapea los
prewarms en vuelo (CloseWhere(IsPrewarm)) para liberar el writer-lock
de la caché — un prewarm SELLADO sobrevive como cache HIT — y luego
desaloja normal vía Register.
- Trickplay: -skip_frame nokey + fps=...:eof_action=pass — solo
decodifica keyframes (12x menos CPU medido: 233s→19s en un episodio
de 24min 1080p; importa porque corre junto al streaming en vivo).
Los ticks siguen siendo uniformes (fps repite el último keyframe),
así que manifest y clientes cacheados no cambian. eof_action=pass
cubre clips con un único keyframe (el filtro fps no emite nada de un
stream de 1 frame con el eof por defecto).
- HLSSessionConfig.StartSec (sync StreamSession.startSec): el primer
ffmpeg arranca ya seekeado en el punto de resume (-ss +
-output_ts_offset + -start_number, misma maquinaria que el
seek-restart) en vez de encodear desde seg-0 para morir en el
seek-restart inmediato del player (doble spawn, resume lento).
readyMax se pre-siembra al índice de arranque; el ready-watcher
compara ReadyCount() > WriterStartIdx() para no marcar "ready" antes
del primer segmento real. startSec >= duración → arranque desde 0
(resume obsoleto de un fichero reemplazado).
- Rate control: capped constant-quality donde el encoder lo hace bien —
libx264 -crf 23, NVENC -cq 23 -b:v 0 — con el mismo -maxrate de
siempre y -bufsize 2x (antes 1x estrangulaba picos). Escenas fáciles
emiten muchos menos bits (menos stalls vía funnel/LTE); el peor caso
no cambia. QSV/VideoToolbox/VAAPI conservan el triple de bitrate fijo
probado (sus knobs de calidad tienen gotchas de vendor).
- Limpieza: wrapper buildHLSFFmpegArgs y guard startIdx<0 muertos.
Close the recurring "video has subtitles but the web player shows none" gap
with a source-agnostic pipeline:
- Discover EXTERNAL sidecar subs in the scan (Video.es.ass siblings + a Subs/
bundle), parse lang/forced/SDH from the filename, skip VobSub (.sub+.idx).
ffprobe-only scanning ignored these (ToonsHub/anime "MSubs" releases).
- Transcode sidecar charset -> UTF-8 before WebVTT (BOM/UTF-16/code-page by
language). Chinese SCRIPT matters: chs/sc -> GBK, cht/tc/big5 -> Big5
(decoding one as the other is mojibake).
- /sub now serves a standalone sidecar file (i=-1, p=file, &l=lang hint) and a
remote debrid URL (ffmpeg reads http, no local stat) — not just embedded
streams of a local file.
- probe.json emits a tokened vttUrl per TEXT track so torrent/debrid HLS streams
(never library-scanned) get subtitles too. Embedded index is counted among
embedded streams only, so -map 0:s:N stays aligned when sidecars are appended.
Tested against a real 347-file gallery: 26/26 sidecars and embedded ass/srt/
mov_text all extract to valid WebVTT; bitmap (pgs/dvd_subtitle) correctly stays
burn-in. Manual harness gated behind GALLERY_DIR.
The runtime stage's `ARG TARGETARCH=amd64` default shadowed buildx's
per-leg value, so even the published arm64 image bundled x86-64
cloudflared and ffmpeg alongside a native arm64 unarr binary. The daemon
spawning cloudflared hit "exec format error", the funnel never came up,
and TV/Stremio connect failed with "Failed to get add-on manifest".
Read the real arch from `dpkg --print-architecture` (the emulated base
image's arch) for both the ffmpeg and cloudflared RUN steps. Correct
under buildx cross-builds and plain `docker build` alike. Drop the
poisoned TARGETARCH default.
Reported-by: Serge <s@bongiozzo.ru>
- IsRevoked no longer matches a bare 401. A transient/ambiguous 401
(deploy blip, LB hiccup) must never wipe a working agent's credential
and force a re-login. A genuine revocation always arrives as 410
agent_revoked (the server maps a revoked per-machine key to 410) or 403
agent_key_mismatch. Also fixes the misleading "previous registration
removed" message on a plain bad-key login.
- Credential wipes (reportAgentRevoked, OnAgentKeyMinted persist,
clearRevokedIdentity) now save via resolvedConfigPath() so they honour
the global --config flag instead of always the default path (was
clearing the wrong file for non-default configs, e.g. unarr-dev).
--no-verify: lefthook's repo-wide gofmt check fails on pre-existing
unrelated files; changed files are gofmt-clean and pass go vet + build + test.
Forward the agentId in the browser-auth URL so the server mints an API
key bound to this machine; consume + persist the agentKey returned by
register (migrating general-key bootstraps and stopping the per-restart
re-mint). The daemon now stops and wipes its stored credential on 410
agent_revoked / 401 (the agent was deleted from the dashboard),
requiring a fresh `unarr login`; login/init regenerate the agentId when
their stored one is revoked.
Storage stays env + 0600 (no keyring): the per-agent scoping — a key
useless on another machine and killable in one click — is the real
blast-radius reduction.
--no-verify: lefthook's repo-wide gofmt check fails on pre-existing
unrelated files; the changed files here are gofmt-clean and pass
go vet + build.
Parse ffmpeg's -stats progress line (speed=Yx, fps=) from the HLS encoder's
stderr into a per-session EWMA, and report a health snapshot to the web side a
few seconds after seg-0. Lets the player name a too-slow transcode from a
direct measurement (~5-7s) instead of inferring it from stall shape (~15-30s).
- hls.go: add -stats; rewrite hlsStderrCapture.Write to frame on \r and \n,
parse speed=/fps= (telemetry only, never logged), flag input-bound on source
read errors. EWMA on HLSSession + GetTranscodeStats(); warmup-skip the first
cold-start frames so a healthy encoder isn't reported as struggling.
- client.go: MarkSessionReady takes an optional *SessionHealth.
- daemon.go: watcher reports one health snapshot once >=4 post-warmup samples
settle; classifyAgentHealth maps the speed ratio to ok/marginal/struggling.
Additive: old web replicas ignore the extra field; cache-hit/direct-play
sessions and short encodes report nil (the web keeps its stall heuristic).
The anacrolix/torrent client maps the listen port on the router via
UPnP/NAT-PMP for inbound peers. Many home routers reject AddPortMapping
with '500 Internal Server Error' and the lib retries every lease cycle,
spamming the daemon log. The rejection is harmless (downloads work over
DHT + outbound peers), so wrap the log handlers and drop just that line
while keeping the mapping attempts for routers that do support it.
The libplacebo HDR->SDR tonemap needs a Vulkan device, but the nvidia
Vulkan ICD (libGLX_nvidia.so.0) pulls in libnvidia-glcore, which
references glibc malloc hooks removed in glibc 2.34 (__malloc_hook etc.)
and the Xorg symbol ErrorF. On any headless modern-glibc container these
go unresolved so vkCreateInstance returns VK_ERROR_INCOMPATIBLE_DRIVER
and the agent correctly falls back to the CPU zscale tonemap chain.
Document why we deliberately do NOT chase it (graphics cap + X11 libs +
1.4 loader + desktop glibc/Xorg, fragile + distro/driver coupled).
nvenc/nvdec (CUDA, not Vulkan) work regardless.
The agent obtains a valid wildcard cert for *.<hash>.agent.unarr.app from
the web broker (ACME DNS-01) so the https web player reaches it directly
over HTTPS instead of the CloudFlare funnel.
- internal/acme: generate EC P-256 key + CSR locally (private key never
leaves the machine), fetch the signed chain from the broker, persist it
atomically, NeedsIssue renewal check
- daemon: generate + persist a stable agent_hash in config.toml; register
before requesting the cert (broker ownership check needs the row); arm
the HTTPS listener with the cert; 6h renewal poll hot-swaps it (no restart)
- report httpsStreamPort + agentHash on register/sync
- stream_server: emit Access-Control-Allow-Private-Network on PNA preflight
so an https page can reach the agent on loopback / LAN
On a docker agent the web DB holds host paths (e.g. /mnt/nas/peliculas/…)
while the container mounts that media at /downloads, so the runtime allowed
root (cfg.Download.Dir=/downloads) rejects the host path. The raw /stream
handler already self-heals via relocateUnreachable, but the HLS/remux session
handler did not — it logged "path outside allowed dirs" and returned, so the
web silently fell back to the raw /stream path (no transcode, slow funnel
start) and HLS/remux never ran. The path-scoped sidecar handlers
(/thumbnail, /trickplay, /sub) had the same skew → 404 for every scrubber
frame, trickplay sprite and external subtitle.
- HLS handler (OnStreamSession): apply the same relocateUnreachable remap as
the raw handler before the dir-resolve.
- StreamServer: add SetPathResolver/healMediaPath, applied in /thumbnail,
/trickplay, /sub AFTER token verification (the token still binds the
original web path; the resolver is a pure function of that path and
re-validates containment, so it can't be abused to serve a different file).
- Hoist the allowed-roots list into streamAllowedRoots(cfg) so the raw, HLS
and sidecar handlers can't drift apart.
Note: relocateUnreachable needs a ≥3-segment path tail, so flat media layouts
are not self-healed (same limitation as /stream; a re-scan rewrites the DB
path). The HLS handler replicates only the lexical remap, not the raw
handler's transient-NFS os.Stat retry.
Supersedes the previous "disable signature verification" stop-gap. The two
checksum tests now run verifyChecksum with signature verification ENABLED using a
per-test ed25519 keypair (withReleasePubKey) and a matching checksums.txt.sig
served over the exact body — so they cover the real production path end to end
instead of skipping it. Adds verifyChecksum-level coverage for the cases that
actually protect a self-update: a checksums file signed by the wrong key is
rejected, a missing .sig is rejected, and verifyChecksumOnly (--allow-unsigned)
still passes on the checksum alone. No production code change.
TestVerifyChecksumWithHTTPTest and TestVerifyChecksumCaseInsensitive predate
release signing (commit 1757bda baked the release pubkey). With the pubkey set,
verifyChecksum now requires a valid checksums.txt.sig and fails at signature
decode before reaching the SHA256 comparison these tests assert. They exercise
the checksum-matching path only, so clear releasePubKeyBase64 for their duration
(t.Cleanup restore) — mirroring the existing pattern in signature_test.go. The
signature path itself keeps its dedicated coverage there. No production change.
Trickplay sprite generation (one full-decode ffmpeg pass per file) could pin a
machine: multiple agents on the same library decoded the same 4K file at once, no
CPU throttling, and crashed/restarted agents orphaned ffmpeg to init (it ran the
full 45-min decode to completion). Stacked orphans spiked a box to load ~140.
- Single-flight lock: O_CREATE|O_EXCL .lock in the shared sidecar dir so two
agents watching the same library never decode the same file twice (stale locks
reclaimed after a TTL). Returns ErrTrickplayInProgress → prewarm skips, not fail.
- Load gate: defer the heavy decode until 1-min load ≤ max(ratio×NumCPU, 1.5),
capped at 15 min so it throttles without ever becoming a permanent off-switch on
busy / small hosts. New knob library.prewarm_max_load_ratio (default 0.7).
- Concurrency: trickSem caps trickplay to ONE decode at a time per agent.
- CPU priority: setLowCPUPriority (nice 19) alongside the existing idle ionice.
- No orphans: hardenCmd sets Setpgid + Pdeathsig=SIGKILL, with runtime.LockOSThread
around the child so the kernel kills ffmpeg exactly when the agent dies (and not
spuriously — golang/go#27505).
Tests: single-flight/stale-reclaim, load-gate immediate/cancel, and an e2e
Pdeathsig orphan-kill check.
handleStreamTask now serves a mode=stream task FROM a resolved debrid HTTPS link
(when the web set preferredMethod=debrid + the torrent is cached) instead of
joining the P2P swarm — served over the SAME /stream endpoint so VLC and other
external players consume it identically (and far faster). No HLS transcode:
external players handle any container. Falls through to the P2P StreamEngine
when there is no direct URL. Uses the mutex-safe SetStreamURL setter.
Also widen the debrid HEAD size-probe timeout 10s -> 15s to match the transport's
TLS handshake budget, so a slow CDN no longer trips it and falls back to a
guessed size.
Bump 1.0.2-beta.
Pre-generate ONE trickplay sprite (montage JPEG of frames sampled every
library.trickplay.interval, default 10s) + a JSON manifest per file during the
scan/auto-scan prewarm, cached in .unarr next to the media. The web scrubber
shows tiles from it instead of extracting frames live — removing the ffmpeg
contention with the active stream that broke seekbar previews (the original
'no thumbnail' report was the auto-scan prewarm decoding the same file the HLS
transcode was reading, not a seek-index fault).
- config: [library.trickplay] enabled/interval/width (default on, 10s, 240px),
editable + a toggle; IntervalSeconds() with a 10s fallback.
- mediainfo: GenerateTrickplay (one ffmpeg fps=1/interval,scale,tile pass; idle
I/O priority; ceil() frame count so no black trailing tile; a 16.7M-px cap
coarsens the interval for long media so a single sprite stays decodable on
iOS/Safari) + sprite/manifest sidecar cache helpers.
- engine: /trickplay endpoint (manifest JSON, ?kind=sprite JPEG); the agent owns
the tile width so the web requests by path only; thumb:<sha256> token reused.
- prewarm: a trickplay job per item, gated; scan.go + daemon.go wire the config.
Tests: parseDims; synthetic 3x2 / exact-multiple / 1x1; real-file e2e smoke
(S02E08 → 143 tiles, 662KB sprite). Non-breaking: the existing 5-frame panel
prewarm + on-demand /thumbnail stay until the web migrates to the sprite.
The smoke check extracted the docker image version with `grep -oE 'v[0-9.]+'`,
which truncates "v1.0.1-beta" to "v1.0.1" and false-warns on a correct image.
Match the trailing suffix too.
Releases were shipping UNSIGNED: ship.sh never invoked sign-checksums, the
goreleaser pubkey ldflag defaulted to empty, and publish-cli-release.sh did not
upload a .sig — so the self-updater's signature check was silently skipped
(1.0.0-beta had no checksums.txt.sig). Make signing unconditional:
- internal/upgrade/signature.go: bake the canonical release public key as the
compiled-in default (public, safe to commit; removes the empty-env footgun).
- .goreleaser.yml: drop the pubkey ldflag (committed default is authoritative)
+ add a signs: block that runs scripts/sign-checksums over checksums.txt.
sign-checksums requires -key, so an unset RELEASE_SIGNING_KEY fails the build
instead of shipping unsigned.
- scripts/ship.sh: source RELEASE_SIGNING_KEY from ~/.config/unarr-release/signing.key
(or the env), die if absent, and assert checksums.txt.sig was produced.
Private key lives outside the repo (gitignored keyfile + operator's vault);
public key verified to match (priv[32:] == baked pubkey).
Fast input seek (-ss before -i) fails on files whose seek index is imprecise
or mildly corrupt: the demuxer lands mid-EBML element ("invalid as first byte
of an EBML number") and decodes no frame, so the web scrubber showed a broken
image (2026-06-03, anime MKVs: 15/15 prewarm thumbnails failed). When the fast
path yields no frame, retry once with output seek (-ss after -i, decode from
the start) + -err_detect ignore_err. Applied in both the on-demand handler
(buildThumbnailArgsAccurate) and the prewarm extractor (ExtractThumbnailJPEG).
Cost is paid only when the fast path fails, so healthy files keep the cheap path.
Regression test: TestBuildThumbnailArgsAccurate.
The web persists the chosen audioIndex globally, so a value from a
multi-track file can arrive for a file with fewer tracks. buildHLSFFmpegArgsAt
mapped `-map 0🅰️N?` verbatim; the optional `?` then matched nothing and the
HLS output had NO audio stream (video-only — 2026-06-03, Wistoria S02E08 had
one audio track but the session carried audioIndex=2). Clamp an out-of-range
index to the first track so audio is never silently dropped.
Regression test: TestBuildHLSFFmpegArgsAudioClamp.
The binary self-update hard-stops inside a container (internal/upgrade
refuses when /.dockerenv exists), so the web's in-app 'force update' button
is futile for Docker agents. Report whether we run in a container via a new
RunningInDocker() helper (UNARR_DOCKER env baked into the image, /.dockerenv
fallback for podman/containerd) on every register + sync, so the web can
swap the button for a copy-paste 'docker compose pull' command.
Stop treating the absolute path as a file's identity so a base-path change
(host binary→docker remap, moved media folder, remount) no longer makes the
server duplicate and orphan library rows.
- fingerprint.go: ComputeFingerprint = sha256(size ‖ first 1MiB ‖ last 1MiB),
a stable content identity that survives rename/move/base-path change. Cached
in LibraryItem and reused on incremental scans when size+mtime are unchanged.
- sync: send fingerprint + rel_path (relative to the scan root) + agent_id in
the library-sync request, so the server can move a row in place and scope
stale-cleanup per agent.
- daemon: force a FULL re-scan (with a user-facing WARNING) when the scan root
changed since the last cache, so the server re-maps by fingerprint instead of
duplicating. basePathChanged compares filepath.Clean'd roots.
- daemon: relocateUnreachable self-heals a stream request whose path is under an
old root but whose file still exists under a current allowed root, so playback
works immediately without waiting for the re-scan. Conservative: requires a
3-segment tail and re-checks containment after resolving symlinks so it can
neither serve the wrong file nor escape the allowed dirs.
See docs/plans/unarr-path-resilience.md in the web repo.
Second critico pass on the functional probe.
- The probe does real Vulkan device init, which can transiently fail when the
box is busy (notably the startup warm racing the encode benchmark). Caching
that timeout as a permanent 'no' would pin HDR to the zscale CPU chain until
daemon restart. Now a deadline is NOT cached — only a clean non-zero exit
(filter absent / no ICD), which is a stable result. zscale stays cached as
before (cheap deterministic grep, can't flake).
- Surface the exec error when ffmpeg never produced stderr (timeout / ENOENT):
the fallback log now shows err.Error() instead of a blank tail, so 'no
Vulkan' is distinguishable from 'ffmpeg never ran'.
- Dockerfile comment: clarify the Vulkan ICD (not GLX) is the load-bearing
mount that 'graphics' adds; 'compute' alone doesn't mount it.
Probe still returns true on a Vulkan host (verified); engine tests green.
Make libplacebo actually reachable in the shipped agent image, and refuse it
where it would be a regression.
Dockerfile (so a Vulkan-capable host can use the GPU tonemap path):
- install libvulkan1 (the Vulkan loader libplacebo links at runtime; ~150 KB)
- add 'graphics' to NVIDIA_DRIVER_CAPABILITIES so the nvidia container runtime
mounts the Vulkan ICD (nvidia_icd.json + GLX libs) under --gpus all
Both are inert without a working Vulkan GPU — the functional probe gates use.
hls.go: gate libplacebo on a real HW encoder (HWAccel != none). A software-only
host with mesa would expose lavapipe (CPU Vulkan); the functional probe accepts
it but its tonemap is SLOWER than the zscale CPU chain, so libplacebo there is a
regression. No HW encoder -> stay on zscale.
Verified on the GPU dev box: nvenc session still picks libplacebo (-c:v
h264_nvenc -vf ...,libplacebo=...:tonemapping=bt.2390); new unit test locks the
software-encoder path onto zscale.
Review (critico) caught a regression: the prod agent image ships a BtbN GPL
ffmpeg with libplacebo COMPILED IN but no Vulkan runtime (debian-slim, no
libvulkan1/mesa-vulkan-drivers/nvidia ICD). The presence probe (ffmpeg
-filters) would flip HasLibplacebo on, the filter's Vulkan device creation
would fail at runtime, and HDR sources that previously tonemapped via zscale
would break.
- FFmpegSupportsLibplacebo now RUNS the filter on one synthetic frame and
requires a clean exit (forces Vulkan device init + filtergraph negotiation),
so it is honest about THIS host: works on Vulkan-capable hosts, falls back to
zscale where Vulkan is absent. Logs the real ffmpeg error on failure.
- Warm the libplacebo (Vulkan init ~1.7s) + zscale caches in a background
goroutine at startup so the first stream session doesn't pay the probe and
risk its setup timeout.
- Benchmark: margin 1.5x -> 2.0x (the probe measures encode only; real decode
of HEVC/10-bit + busier content needs more headroom), per-probe timeout
12s -> 6s + overall 45s -> 20s (it blocks registration on software hosts),
and a 'no rung measured' case (missing lavfi/wedged ffmpeg) now keeps the
1080 default instead of flooring at 480 — an infra failure isn't a slow host.
Verified e2e on the fixed binary: LOTR Two Towers (HEVC 3840x1608 10-bit
HDR10/PQ, 12GB) on desktop-Chrome caps -> hls, ffmpeg runs h264_nvenc with
-vf ...,libplacebo=...:format=yuv420p:tonemapping=bt.2390 (zscale chain
replaced), 45 fMP4 segments, ffprobe confirms output h264 yuv420p bt709
(tonemapped from bt2020/smpte2084), no ffmpeg errors.
Replace the guessed transcode ceiling (CPU->1080, GPU->2160) with a measured
one. HW encoders still return 2160 instantly. A software-only host runs a
bounded encode benchmark — 3s testsrc2 through the real libx264 superfast
settings at 1080/720/480, top-down — and reports the rung it sustains at
>=1.5x realtime (margin for real decode + busier content).
Fixes risk 2: a weak NAS/old CPU that is ffmpeg-capable but can't keep up
with a 1080p software encode no longer advertises a 1080 ceiling, so
decideStreamPlan routes oversized sources to an external player instead of a
stuttering transcode. Floors at 480; each probe is timeout-bounded so a
wedged ffmpeg can't stall daemon startup.
Prefer the single-pass Vulkan libplacebo filter over the CPU zscale chain
for HDR->SDR tonemapping when the agent ffmpeg has it. One GPU pass does
tonemap + BT.709 primaries/transfer/matrix + 8-bit yuv420p, replacing the
four-stage zscale chain and its trailing format=/setparams. Higher quality,
far cheaper than the CPU path, and present on builds that lack zscale.
- FFmpegSupportsLibplacebo probe (cached, mirrors FFmpegSupportsZscale)
- HasLibplacebo on TranscodeRuntime, wired from buildTranscodeRuntime
- hls.go: videoTail picks libplacebo when present (not h264_vaapi), else
keeps the zscale tonemap + format chain
- test: libplacebo replaces the zscale chain, never runs alongside it
