Optimisarr 0.2.17 — more reliable sidecars and a Linux worker preview

This update improves sidecar RAM storage, GPU decoding and recovery when work is cancelled. A Linux sidecar preview adds a dedicated page for pairing and monitoring a worker.

What's new

• Run a Linux worker. Pair the separate sidecar container from its web page and see its active media, CPU/GPU activity and available encoders. Capacity remains controlled through Compose.
• Collect VMAF v1 research separately. Optional server measurements compare short samples under the current and new models. The existing quality model still decides whether a file can replace its original.

Smoother and safer

• RAM jobs start and clean up more reliably on Mac. Concurrent jobs share a bounded RAM budget, and downloads stream directly into working storage.
• GPU jobs handle more of the real pipeline. Capable workers receive the correct hardware-decode commands, Intel decoding has room for frames still held by the encoder, and VAAPI keeps 10-bit precision where required.
• Cancelled work stays cancelled. Late claims and upload retries no longer keep obsolete remote work alive.
• Affected VC-1 files avoid false timestamp failures. Source verification reconstructs missing presentation timestamps consistently with encoding; candidate checks remain strict.

Before you update

Update the main container and sidecars together to receive the full set of fixes. Preserve existing pairing/configuration volumes. Deploying the Linux worker requires its separate sidecar Compose service; the main container remains the control plane.

Good to know

• The Linux sidecar is an amd64 preview; Intel acceleration and RAM working storage require explicit Compose configuration. Follow the Linux sidecar setup guide. The dashboard has no built-in authentication: keep it private or use an authenticated proxy.
• The Windows MSI is an unsigned Windows preview. See the code signing policy. macOS downloads are Developer ID signed, notarised and stapled.
• VMAF v1 research is disabled by default. Enabling OPTIMISARR_VMAF_SHADOW_SERVER=1 adds bounded CPU scoring on the main server, including for sidecar-verified jobs. It does not change quality thresholds or replacement decisions.

See the full changelog for the complete technical record.

Optimisarr 0.2.16 — good encodes stop failing the quality check

Optimisarr 0.2.16 fixes the last known cause of good encodes failing the quality check, catches broken sources before anything is encoded, and tells you when a sidecar needs updating. Update your sidecars too: the quality fix also runs inside them.

What's new

• Sidecars tell you when they're out of date. When a Mac or Windows sidecar is older than your server, its menu or tray shows Update available with a link to the release page, and its card under Settings → Workers says the same. Sidecars never update themselves; installing stays your choice.
• A clearer Dashboard. Total space saved leads, next to what's in progress and what needs you. A Saved per day chart covers the last 30 days, with a daily average and a readout for any day you point at, and Recent results lists each finished file's size before and after, its quality score and where it ran.
• A status strip on every page shows what the queue is doing (encoding here, On workers, paused, holding, waiting or idle) and why, with free working space and a Pause/Resume button, so you can pause from anywhere.

Smoother and safer

• No more false quality failures from drifting timestamps. Some encodes keep every frame of the original but stamp stretches of them a frame early. The quality check then compared each picture with its neighbour and failed good encodes with scores around 15–40. When an encode has exactly as many frames as the original, each sample is now compared frame for frame. Encodes that were failing now score in the 90s, and normal encodes score exactly as before.
• Quality samples line up exactly. On some titles, samples were cut a few milliseconds apart from the original, which could send a quality search back to the library's default quality.
• Broken sources are caught before encoding. A file whose picture stops well before its audio now fails straight away, naming Source video timeline, instead of after a full encode (and, on a worker, a full download and upload). The original is untouched.
• The sidebar shows what's really happening. It no longer says a sidecar is encoding after the sidecar has handed its work back. It names the stage (encoding, verifying, returning a result and so on) and where it's happening.
• Settings now uses the full page width like every other page, and the interface widens to fill large screens.

Before you update

• Update your Mac and Windows sidecars to 0.2.16. The frame-by-frame quality check also runs on sidecars. An older sidecar keeps measuring the old way, so some good encodes can still fail until it's updated. Your originals stay safe either way.
• One small database update runs automatically on start-up. It records each finished job's original size for the Dashboard, and fills it in once for jobs verified earlier. No action needed.

Good to know

• The Mac sidecar now bundles libvmaf 3.2.1, the same VMAF library as the server and the Windows sidecar. Quality scores don't change.
• The Windows installer is still an unsigned preview. Follow the code signing policy, and don't weaken Windows or company security settings to install it. The Mac app is signed and notarised by Apple.

See the full changelog for the complete technical record.

Optimisarr 0.2.15 — no more false quality failures

Optimisarr 0.2.15 stops good encodes from failing the quality check, and skips full transcodes that were never going to save space. Update your sidecars too: half of the quality fix runs inside them.

What's new

• Size check before the full transcode. While choosing a quality, Optimisarr now compares each 40-second sample with how much space the original used for the same scenes. If the finished file looks likely to be bigger than allowed, or to miss your minimum saving, the job waits under Queue → Needs review instead of spending hours on a full encode. Choose Encode anyway to run it regardless; the final size and quality checks still apply. The original is never touched while a job waits.

Smoother and safer

• No more false quality failures. Some encodes that looked fine were failed with a VMAF score near zero. The encodes were good, but the check was sometimes comparing each frame with the frame next to it. Both causes are fixed: one in the check that runs after a full encode, and one in how remote workers score their quality samples. Encodes that were failing at 10–20 now score in the 90s, and quality searches on workers pick a real quality instead of giving up.
• Encode anyway works straight away on workers. Approving a job that a worker's samples paused no longer leaves it waiting ten minutes before a worker picks it up again.

Before you update

• Update your Mac and Windows sidecars to 0.2.15. Part of the quality fix runs on the sidecar. An older sidecar keeps failing good encodes until it's updated. Your originals stay safe either way.
• One small database update runs automatically on start-up. No action needed.

Good to know

• The optional Stellar icon (Settings → System → Appearance) has been redrawn. The default icon hasn't changed.
• The Windows installer is still an unsigned preview. Follow the code signing policy, and don't weaken Windows or company security settings to install it. The Mac app is signed and notarised by Apple.

See the full changelog for the complete technical record.

Optimisarr 0.2.14

Optimisarr 0.2.14 brings the Precession cube back as the default identity across the web app and both native sidecars.

What changed

• One icon everywhere. Browser sessions, favicons, the macOS app and menu bar, the Windows executable, tray, Start shortcut, and installer now use the Precession cube.
• Visible worker activity. The macOS and Windows tray icons rotate while their worker is processing a job, return to the exact static mark when idle, and respect reduced-motion settings.
• Matching project header. The README uses the same production Precession motion, with dark, light, and reduced-motion variants.
• Published signing policy. The repository now documents the Windows signing scope, controls, verification, and incident response while the SignPath Foundation application is reviewed.

Installation note

The Windows MSI is currently an unsigned preview. Follow the project’s code-signing policy and do not weaken Windows or organisational security controls to install it.

See the full changelog.

Optimisarr 0.2.13 — clearer controls and compact workers

Optimisarr 0.2.13 makes workers easier to monitor and library settings easier to navigate. It brings the refreshed application interface, a matching Mac menu-bar client and the first Windows tray installer preview into one versioned release.

What’s new

• Compact worker monitors. Mac and Windows show current work, resource readings, processing details and pause controls in a small native panel. Panels stay anchored when their contents resize, and the Stellar icon appears consistently in the application and native launchers.
• Verification can stay on the worker. Enable strict worker verification for supported remote video jobs to run probes, decode, timing, audio and quality measurements on the sidecar. Missing or invalid evidence stops the job instead of starting a container verification fallback.
• Library controls follow the workflow. Choose files, Encode, Verify, and Schedule & replace have focused views, breadcrumbs, advanced sections and explanatory tooltips.
• Media details stay in view. Inventory and Queue use artwork-led dialogs. Quarantine opens a full review page with side-by-side playback and verification evidence.
• A consistent interface and clearer guides. Settings, light and dark themes, cards, hover shadows and the animated Stellar mark share the refreshed appearance. Updated guides and current screenshots explain worker setup, placement and which responsibilities remain on the server.

Smoother and safer

• Queue artwork remains visible during live updates, and older responses no longer overwrite newer progress.
• Worker verification is bound to the assigned contract and both file hashes. Existing pairings renegotiate capabilities after an upgrade without needing a new pairing code.
• Software-decode retries retain a valid output directory. The Mac media bundle includes AV1 decoding and 10/12-bit x265 support.
• The container release passes real-media acceptance tests before image publication. Native packaging checks include installer validation and rendering from a relocated Mac app.

Downloads

• Container: ghcr.io/jellman86/optimisarr:0.2.13 (also :latest).
• macOS, Apple Silicon: DMG or ZIP, Developer ID signed, notarised and stapled. Requires macOS 14 or later.
• Windows x64 — unsigned preview: MSI installer, with the tray app, background worker, media tools and private .NET runtimes. It is not code-signed and may be blocked by Windows or organisational policy.

SHA-256 files accompany the downloads. Exact corresponding media sources, manifests and build instructions are provided beside the installers. Windows sources include split archives and a checksum-checked reassembly script; follow the source download guide. Native version numbers now align with the container version, so the Mac version advances from 0.2.0 to 0.2.13.

Before you update

• Back up the container’s /config directory and let active worker leases finish before replacing a sidecar. Keep existing pairing data and work directories.
• Quit the Mac sidecar, replace the app in Applications, then reopen it. Existing signed pairings are retained. Upgrade an existing MSI installation with the new MSI; it retains pairing data. After installation, open Optimisarr Sidecar from Start and choose Preferences → Start worker (or reboot), because the installer leaves the service stopped. Manually registered Windows services need the migration steps in the installer guide, because the MSI deliberately refuses to overwrite them.
• Remote workers and strict verification remain opt-in. Preserve your chosen verification policy during upgrades; jobs can wait until a compatible worker is available. See worker setup and placement.

Good to know

Worker-only placement and strict verification are separate controls. Strict verification does not make the server idle: scanning, initial probes, scheduling, transfers, hashing, policy evaluation and safe replacement remain server responsibilities. Remote placement covers supported video re-encodes; audio, images and remux work remain local.

The Windows preview uses CPU VMAF; NVIDIA hardware encoding and decoding are supported separately. Hardware smoke tests do not certify every driver, source format or full-length title. Originals remain protected by the configured verification and quarantine/rollback checks.

See the full changelog for the complete technical record.

macOS sidecar 0.2.0

Four ways a job could be lost, and the measurement that was failing good encodes.

Update if you are running 0.1.x. Everything below was found on real hardware over one evening, and every one of them cost finished work.

A job could stop forever with nothing running

A Mac sat in "Measuring" for fifty-two minutes with no FFmpeg running at all — lease renewing perfectly, menu bar animating, check-in answering every ten seconds, and the job never moving.

The runner treated end-of-file on FFmpeg's output as FFmpeg ending. A pipe is only at its end when the last handle on its write side closes, and a pipe macOS creates is inherited by anything the app spawns while it is open — a load probe, a preview frame, another job. Nothing could recover, because every part that might have noticed was working correctly.

The pipes are now kept out of other children, and the process's exit decides when a job is over.

A finished encode could be thrown away

A candidate that had been encoded, measured and accepted was lost to Delivering the candidate failed (HTTP 502) — the server had restarted while the bytes were going up, and the job went back on the queue to be done again from nothing.

Delivery has always been resumable. Only a narrow set of errors was being retried, and a restarting server was not among them.

A moment's silence could abandon the work

One refused lease renewal ended the renewal loop and took the job with it, so restarting the server discarded every encode running at that moment, minutes in.

A lease is now given up when the server says it is gone, or when no renewal has landed for as long as the server granted the lease for — not on the first failure.

Good encodes were being failed on quality

Whole seasons failed the perceptual-quality gate with harmonic means in single figures while the encodes themselves were sound.

Two causes. FFmpeg's default frame-rate handling was dropping frames on sources ffprobe calls constant — about fifty an episode on one VC-1 source — which both damaged the file and made every comparison meaningless, because frame N of the candidate stopped being frame N of the original. And the correction meant to line the two up was derived from the containers' headers, which say the same thing for a file that needs a correction and one that is ruined by it; it computed as zero for every file this sidecar had ever measured.

The candidate is now tried a frame either way against a short window and whichever matched is what the measurement uses. On the pair that exposed it, the same encode went from a harmonic mean of 0.21 to 92.32.

Nothing unsafe ever happened. Misalignment makes scores worse, never better, so the failure mode was rejecting good encodes — never accepting a bad one. No file was replaced that should not have been.

The panel is now a readout

It was a stack of rounded cards that happened to carry figures. The machine's name and state now sit opposite each other in a bar across the top, everything it knows runs down one column of figures you can compare vertically, and the controls sit along the bottom where they cannot be mistaken for readings.

Progress is lit segments rather than a bar — and during an encode there is no bar at all, because the worker knows how many seconds it has done and not how many it owes. A bar that guessed would be a lie in the one place this panel exists to be honest.

The colours are Optimisarr's own now, taken from the web interface, in both light and dark. A fault on this panel is the same colour as the same fault on your dashboard.

Also

• The menu-bar mark no longer turns on an idle Mac.
• Abandoned scratch directories are swept, so a sidecar that was killed mid-job stops holding gigabytes for ever.
• Each measured window reports its own score, so a measurement that comes back wrong can be diagnosed after the fact.

About the version number

0.1.5 was the last public build. There is no 0.1.6 — this is a bigger step than a patch, and the jump says so.

Installing

Download the .dmg, drag it to Applications, and open it. It is signed and notarised. Your existing pairing survives the upgrade.

sidecar-v0.1.5

macOS sidecar 0.1.5

Stable menu that names the file and shows transfer speed, and an Options panel for where work happens.

sidecar-v0.1.4

macOS sidecar 0.1.4

AV1 encoding via SVT-AV1, AV1 hardware decode, audio encoder reporting.

sidecar-v0.1.3

macOS sidecar 0.1.3

Reports its audio encoders, and refreshes what it can do on every check-in.

sidecar-v0.1.2

macOS sidecar 0.1.2

Fixes a bundled FFmpeg that could not start on any Mac but the build machine.