[celeroncoder]

Two thumbnail paths in the GPUI app did far more decoding than they painted.

Settings → Recordings / Screenshots. refresh_recordings and refresh_screenshots scan the full library, and set_recordings / set_screenshots then queued a thumbnail decode for every row without a cached image, through spawn_decode_pool. The lists only render visible_count rows (20, plus 20 per "Show more", reset on tab switch and search). On a library of a few thousand items that is thousands of decodes to show 20, repeated on every 2 s poll for rows that weren't cached yet.

Tray → Previous. load_thumbnail_png routed JPEG sources through library::thumbnail_rgba (the on-disk thumbnail cache, downsampled at decode) but decoded every PNG source at native resolution with image::load_from_memory and then resized it to a 32 px icon. Every screenshot is a PNG, and the menu is rebuilt after each capture and each recording stop.

Change

Commit 1, settings pages:

• visible_rows(matches, has_search, visible_count) is one pure helper that returns the indices of the rows that will be painted. Both renderers, both "Show more" handlers and the decode scheduling use it, so decodes and paints agree.
• pending_thumbnail_decodes queues decodes only for visible rows (a stale recording row with a path, or a screenshot row without an image).
• Each list owns a decode task separate from scan. A scan always starts a decode of the visible pending rows; "Show more", a tab click and search edits start one only if something visible is pending; a new decode replaces the running one, and the pool's workers stop after their current job once the receiver is dropped. The "install only if the row still exists and the path matches" check is unchanged.
• Flicker prevention is unchanged: rows still in the library keep their image, rows that vanish drop theirs. The recordings decode re-arms the 2 s poll when it finishes, as the scan task did before, including when nothing was pending.

Visible change: rows beyond the current page get their thumbnails when they are shown, so "Show more" briefly shows placeholders that fill in.

Commit 2, tray:

• source_may_have_alpha reads only the image header. JPEG is opaque; a PNG counts as alpha if its colour type has an alpha channel, which with image's EXPAND transformation also covers palette and RGB PNGs with a tRNS chunk.
• Opaque sources go through library::thumbnail_rgba (cached, downsample-at-decode, uniform scale so the cover crop is the same as before), with the full decode kept as the fallback. Sources that can carry alpha, such as window screenshots with transparent corners, still decode directly.

Tests: visible_rows across no search / search / show-more; a 500-row list queues 20 jobs, "Show more" queues rows 20-39, search keeps match order; a transparent PNG keeps its corner alpha and writes no cache file; an opaque PNG creates the cache file and a marker overwrite proves the warm path reads it.

Results

Debug build (dependencies at opt-level 2), 2 vCPU aarch64 Linux box.

	main @ 2a61427	This PR
Decode jobs queued for a 500-row uncached settings list	500	20
Tray icon from a 5120×2880 opaque PNG (24 MB), cold cache	766.7 ms	504.8 ms
Tray icon from the same PNG, warm cache	766.7 ms	7.3 ms

Medians of 5 runs. The live UI wasn't run; macOS wasn't built.

Overlap

• perf(desktop-gpui): decode settings library thumbnails through the decode pool #2496 (merged) routed the settings thumbnails through spawn_decode_pool; this bounds what gets queued into that pool.
• perf(desktop-gpui): build tray thumbnails from the Recents thumbnail cache #2485 (merged) built the tray thumbnails from the Recents cache for JPEG sources and noted that screenshot PNGs still decode the original; commit 2 closes that gap for opaque PNGs.
• perf(desktop-gpui): avoid decoding offscreen clip thumbnails #2707 (merged, closes Visible clip thumbnails wait behind offscreen decodes #2706) did the same visible-only decode for the editor clip list; this is the settings-page equivalent.
• perf(desktop-gpui): IDCT-scale JPEG library thumbnails #2721 (open, IDCT-scaled JPEG thumbnails) and feat: add {random} / {random:N} project name placeholder's #2431 (open) touch the same files and apply cleanly on top of this branch. feat: desktop design system, stabilization and dependency migration #2587 touches settings_window.rs and will need a rebase whichever lands second.

How this was made

This PR is agent-generated.

• The findings come from an audit of main by OpenCode scouts and Gemini 3.8 Flash, re-checked against the source by the orchestrating Claude session.
• Claude Opus 5.5 wrote both commits and their tests on current main and took the measurements above.
• Gemini 3.8 Flash (Antigravity) reviewed both commits independently: it checked every former filtered() call site, the decode-result install path, the pool's drop semantics, the 2 s poll re-arm, the tRNS handling in the png/image crates and the crop equivalence, re-ran cargo check, clippy and the settings_window::, tray:: and library:: tests, and approved both.

🤖 Generated with Claude Code

[vercel]

@celeroncoder is attempting to deploy a commit to the Cap Software Inc Team on Vercel.

A member of the Team first needs to authorize it.

[celeroncoder]

cc @richiemcilroy for review (GitHub doesn't let a fork request reviewers, so tagging instead).

[greptile-apps]

The comment repeats the test

This comment only restates the assignment and assertion below it. AGENTS.md prohibits comments that narrate what the code does. Remove it before merging to meet that requirement; the test already makes the case clear.

Suggested change

// A row without a thumbnail file has nothing to decode.

Context Used: AGENTS.md (source)

Prompt To Fix With AI

This is a comment left during a code review.
Path: apps/desktop-gpui/src/settings_window.rs
Line: 5090

Comment:
**The comment repeats the test**

This comment only restates the assignment and assertion below it. `AGENTS.md` prohibits comments that narrate what the code does. Remove it before merging to meet that requirement; the test already makes the case clear.

```suggestion

```

**Context Used:** AGENTS.md ([source](https://github.com/capsoftware/cap/blob/main/AGENTS.md))

---

For each issue above, determine whether it is valid and should be fixed. If so, fix it directly.

Note: If this suggestion doesn't match your team's coding style, reply to this and let me know. I'll remember it for next time!

[celeroncoder]

Removed in 8079013, together with one more narrating comment in the same test ("Load more" queues only the newly shown rows…) that a second pass flagged. The remaining added comments were checked one by one and each captures a non-obvious invariant (unfiltered indices, search order, task-drop cancellation, PNG header alpha check).

Reviewed by DeepSeek V4.1 Flash and Gemini 3.8 Flash.

[richiemcilroy]

Thanks @celeroncoder! Greptile hasn't scored this one yet, so I've asked it to review below. We only merge once Greptile reaches 5/5, so once its review is in, please address its findings (or reply to them if you think one is wrong) and comment @greptileai review to re-score.

@greptileai review