[irvingouj@Devolutions (irvingoujAtDevolution)]

Fixed-rate recordings (web sessions at 1 fps) keep growing at the bitrate target even when nothing on screen changes. After a WebM recording is remuxed, Gateway now re-encodes it as VP8 at a fixed quantizer of 30, keeping only the frames whose picture changed, and swaps the file in only when the result is strictly smaller.

On 8 lab recordings (RDP and web, busy and idle, VP8 and VP9 sources) Gateway stores 68.0 MB of 116.0 MB, with no visible change lost. Idle recordings drop to 3–8 % of their size and busy web ones to about 70 %. A busy RDP recording either shrinks a little or comes out larger, and then the original stays.

The encoding costs about 3 CPU seconds per recorded minute on one thread. VP9 at a matching fixed quantizer saved about 6.6 MB more for three times the CPU, which a 2-vCPU Gateway can't spare beside its live sessions.

The rules live in crates/video-streamer/src/compact.intent.md and the new Compaction section of devolutions-gateway/src/recording.intent.md.

[github-actions]

Let maintainers know that an action is required on their side

• Add the label release-required Please cut a new release (Devolutions Gateway, Devolutions Agent, Jetsocat, PowerShell module) when you request a maintainer to cut a new release (Devolutions Gateway, Devolutions Agent, Jetsocat, PowerShell module)

• Add the label release-blocker Follow-up is required before cutting a new release if a follow-up is required before cutting a new release

• Add the label publish-required Please publish libraries (`Devolutions.Gateway.Utils`, OpenAPI clients, etc) when you request a maintainer to publish libraries (Devolutions.Gateway.Utils, OpenAPI clients, etc.)

• Add the label publish-blocker Follow-up is required before publishing libraries if a follow-up is required before publishing libraries

[irvingouj@Devolutions (irvingoujAtDevolution)]

Implementation notes:

• video_streamer::compact_webm (compact.rs) decodes every frame, diffs full-res luma in 8x8 blocks against the last kept picture (any block with mean |dY| > 12 counts), and re-encodes kept frames at their original timestamps in the input codec with max_quantizer(30). First and last frames are always kept, so duration doesn't change.
• No RDP/web gate. It bails as soon as the output is bigger than the input file, or bigger than the input consumed so far after 60 s of media. Gateway remuxes the result and only replaces the recording when it's strictly smaller. Resolution change inside a file, multiple tracks, or a non-1 ms timescale: skipped, original kept.
• Runs inside RemuxJob right after the existing remux, so no new job type. Failures just log a warning.
• Lab numbers (16 recordings): web idle 0.08x, web active 0.50x, every RDP file skipped after 0.1-0.4 CPU-s. Outputs decode clean in ffmpeg and duration matches to the ms.
• examples/compact.rs is the driver I used for those numbers.

Draft because it needs a cadeau release: cadeau#87 (quantizer range) is merged but not published, so this patches cadeau to git e56c47c for now. To land: merge cadeau#88, cut a native release, then bump the crates and ci/download-cadeau.ps1 (still on v2026.8.21.0). XMF is loaded with dlopen and every symbol is resolved at init, so the crate and the native lib have to move together.

Note

Human-tuned, LLM-assisted content.

[Copilot]

🟡 Changes recommended

An output path that aliases the input can truncate the source recording before it is read, creating a critical data-loss risk.

5 open issues

• Prevent output paths from truncating the input · New
• Handle hidden final frames before dropping the last block · New
• Preserve source duration in the remuxed header · New
• Preserve or convert source colour metadata · New
• Limit concurrent compaction to protect live sessions · New

What changed in this PR

This PR adds post-recording WebM compaction to Gateway, using video-streamer to reduce storage for fixed-rate recordings with unchanged pictures.

Changes:

• Re-encodes retained frames as VP8 at a fixed quantizer of 30.
• Runs compaction after remuxing and replaces the recording only when the final output is smaller.
• Adds compaction tests, an example tool, XMF CI coverage, and updated native dependencies.

File	Description
testsuite/​tests/​video_streamer/​mod.rs	Registers compaction integration tests.
testsuite/​tests/​video_streamer/​compact.rs	Tests frame retention, timestamps, size savings, and empty inputs.
testsuite/​tests/​main.rs	Registers video-streamer tests.
testsuite/​Cargo.toml	Adds dependencies for compaction tests.
devolutions-gateway/​src/​recording.rs	Integrates compaction, size-checked replacement, and temporary-file cleanup.
devolutions-gateway/​src/​recording.intent.md	Documents remuxing and replacement guarantees.
crates/​video-streamer/​src/​lib.rs	Exports the compaction API.
crates/​video-streamer/​src/​compact.rs	Implements frame comparison, VP8 re-encoding, and output generation.
crates/​video-streamer/​src/​compact.intent.md	Defines frame-retention and encoding decisions.
crates/​video-streamer/​examples/​compact.rs	Adds a command-line compaction example.
ci/​download-cadeau.ps1	Updates the native Cadeau release.
Cargo.toml	Pins Cadeau patches supporting quantizer controls.
Cargo.lock	Records patched dependencies and test additions.
.github/​workflows/​ci.yml	Runs XMF-backed tests on Windows and Linux.

_{Review effort: Balanced}

[irvingouj@Devolutions (irvingoujAtDevolution)]

Before/after numbers from the lab experiment, per scenario. Compaction only replaces a recording when the result is strictly smaller; otherwise the (remuxed) original stays untouched, as in rdp-heavy-vp8 below.

Scenario	Source	Before	Compacted (VP8, Q30)	Frames kept	Replaced?	Stored after	CPU (1 thread)	VMAF vs source
RDP idle, VP8	1280x800, 3.1 min	5.97 MB	0.23 MB (4 %)	32/1471	yes	0.23 MB	2.2 s	94.3
RDP idle, VP9	1536x961, 5.0 min	20.33 MB	1.08 MB (5 %)	320/2410	yes	1.08 MB	9.0 s	95.9
RDP busy, VP8	1280x800, 3.7 min	27.04 MB	31.15 MB (115 %)	1769/1781	no, original kept	27.04 MB	31.3 s	99.1
RDP busy, VP9	1536x960, 4.0 min	19.50 MB	16.76 MB (86 %)	1691/1824	yes	16.76 MB	23.0 s	97.8
Web idle, VP8	1024x768, 4.0 min	7.36 MB	0.65 MB (9 %)	25/1920	yes	0.65 MB	2.9 s	88.9
Web idle, VP9	1024x768, 4.0 min	5.29 MB	0.23 MB (4 %)	6/1919	yes	0.23 MB	2.0 s	89.7
Web busy, VP8	1024x768, 3.0 min	15.03 MB	11.12 MB (74 %)	1396/1440	yes	11.12 MB	11.8 s	100.0
Web busy, VP9	1024x768, 3.0 min	15.47 MB	10.85 MB (70 %)	1391/1440	yes	10.85 MB	13.7 s	100.0
Total		116.0 MB				68.0 MB

• CPU is the median of 3 runs. VMAF compares every source frame with what the compacted file shows at that time, so on idle recordings it also counts the moments between kept frames.
• End to end through Gateway (push, remux, compaction, pull): web idle VP8 7.36 MB → 0.65 MB and RDP idle VP9 20.33 MB → 1.08 MB, same duration (239.877 s and 301.388 s), fully decodable and seekable. Played back after replacement in the web player and the RDM player (seek to middle, seek to end, scripted seek): 8/8 passed over two runs on a VM.

Note

LLM-assisted content (no human feedback).