[richiemcilroy]

An idle AVAssetReaderDecoder throws away its native reader, decoded-frame cache and warm-ahead frames 250 ms after its last request. In the paused editor, almost any pause between hovers or scrub steps is longer than that. The next request then pays for a new reader (4 to 7 ms on the CI VM) and decodes again from a keyframe. In one traced example, a single hover frame after a teardown decoded 61 frames in 65 ms. Pressing Play after a pause also reopens the reader.

Measured on the GPUI editor playback profile (1080p30 fixture, Blacksmith macOS 26 VM): 6 to 6.5 teardowns per 10 s of paused hover, and 9 to 10 in the scrub hunt.

This PR keeps the 250 ms wake-up, but on each wake-up the reader is now released only when:

• it has been idle for 5 s, or
• every one of the 16 reader permits is taken, so another track may be waiting for one (the case the short timeout was added for: paused recordings with hundreds of segments).

Behaviour is otherwise unchanged. A request always resets the idle clock, and a released reader is recreated on the next request exactly as before.

Memory: an idle editor holds the frame cache for up to 5 s longer before releasing it, so steady idle memory is unchanged.

Measurements

A/B on ci/editor-playback-ab-idle (run 39955105504): base and head built in one job, 3 interleaved rounds, mean of 3. The fixture is a 40 s 1080p30 studio project, display only and display plus a 720p camera, on a Blacksmith macOS 26 VM.

The "scrub hunt" phase scrubs with 400 ms pauses, like someone looking for a moment:

fixture	metric	base	head
display	frames shown (seeks)	128 (183)	206 (204)
display	process CPU	37.7%	26.8%
display	frames decoded	1858	547
display	readers created (time spent)	13 (62 ms)	0
camera	frames shown (seeks)	122 (188)	207 (207)
camera	process CPU	53.2%	37.3%
camera	frames decoded	3371	1096
camera	readers created (time spent)	26 (132 ms)	0

The other phases:

• Paused hover: 16% more frames shown (display 365 to 424, camera 349 to 397), with process CPU within noise (24.3% to 25.3%, 34.0% to 33.6%). Reader creations fell from 4 and 10 to 0.
• Playback after a pause: no longer reopens the reader (5 ms and 12 ms saved).
• Forward and backward scrubs: lower process CPU (display forward 57.9% to 46.3%, backward 109.4% to 94.8%) and no reader creations. Forward scrubs decode a few more frames per request, because the kept reader decodes forward instead of reopening at a keyframe.

Memory was not measured by this harness. The frame cache is the same bounded cache as before; it is now released 5 s after the last request instead of 250 ms, so a paused editor's steady memory is unchanged.

Confidence Score: 5/5

The PR appears safe to merge.

Summary

Keeps idle AVAssetReaderDecoder readers and cached frames warm for roughly five seconds.

• Idle AVAssetReader readers stay warm for five seconds unless permits run out.

_{Reviews (2) · Last reviewed commit: "Merge main into perf/avassetreader-idle-..." · Reviewed by Greptile}

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
cap-status	Ready	Preview	Oct 23, 2026 5:18pm UTC

[richiemcilroy]

@greptileai review

[capy-ai]

Capy found no issues.

View 1 other finding in Capy.