[celeroncoder]

Every capped screenshot view scanned the whole screenshots library before truncating. list_screenshots_in parsed recording-meta.json for every .cap bundle, then read_dir'd the bundle to find its PNG, then stat'ed it, and only afterwards did the callers keep 9 (RECENT_MEDIA_LIMIT, Recents and the tray's Previous menu via recent_media_in) or 20 (LIBRARY_PANEL_LIMIT, the main window's Screenshots tab). Each screenshot capture triggers this three times (tray refresh, settings refresh, library refresh). #2532 fixed the same shape for recordings and left screenshots as is.

Change

screenshot_items_in mirrors recording_items_in from #2532:

• Phase 1 is stat-only: for each .cap directory it stats original.png (the file screenshot.rs and import.rs write) and falls back to the old first-*.png read_dir search for legacy bundles. Standard bundles now cost one stat instead of read_dir plus stat.
• Stable sort by the PNG's timestamp, newest first, so equal timestamps keep read_dir order exactly as before.
• Phase 2 is lazy: recording-meta.json is parsed only for the items a caller consumes, and bundles whose meta fails to parse are skipped as before.

list_screenshots_in / list_screenshots are now collect() wrappers, so the settings page keeps its full listing. New list_screenshots_limited(limit); scan_screenshots takes RECENT_MEDIA_LIMIT and the Screenshots tab takes LIBRARY_PANEL_LIMIT.

-let mut items = library::list_screenshots();
-items.truncate(library::LIBRARY_PANEL_LIMIT);
+let items = library::list_screenshots_limited(library::LIBRARY_PANEL_LIMIT);

One new test builds a temp library (30 bundles 6 ms apart, a legacy shot.png bundle, a bundle without a PNG, a non-.cap directory, a loose.cap file, a newest bundle with corrupt meta, and two bundles whose PNGs are hard links so they share a timestamp) and asserts: the full listing equals a verbatim copy of the old code; the tied pair keeps read_dir order; take(K) equals the first K of the full listing for K in {0, 1, 9, 20, 100}; and after taking 9, deleting the remaining metas makes the same iterator yield nothing, which proves the parse is deferred.

Results

Real crate functions on a temp library, debug build, warm page cache, 2 vCPU aarch64 Linux box. Medians of 21 samples, p95 in brackets.

Bundles	Limit	main @ 2a61427 (scan + truncate)	This PR (take(limit))	This PR, full listing
400	9	12.6 ms [15.5]	2.7 ms [2.9]	7.7 ms [7.9]
400	20	11.9 ms [12.3]	2.9 ms [3.6]	8.2 ms [10.6]
1,200	9	38.6 ms [41.4]	9.2 ms [9.5]	27.7 ms [29.6]
1,200	20	37.7 ms [48.8]	9.1 ms [9.6]	26.9 ms [31.7]

The capped scan is still O(N) stats plus a sort; only the meta parsing is bounded. Cold-cache and macOS numbers weren't measured.

Two small behaviour notes: a bundle with several top-level PNGs now prefers original.png instead of whichever read_dir returned first (the screenshot editor already assumes original.png), and the sort happens before the meta parse, the same window recording_items_in already has.

Overlap

• perf(gpui): cut capped recording scans by 80% in large libraries #2532 (merged) is the recording-side version of this change and explicitly left screenshots as they were; this is its counterpart.
• Open PRs that touch library.rs or main_window.rs (perf(desktop-gpui): IDCT-scale JPEG library thumbnails #2721, feat(desktop-gpui): recover recordings that crashed instead of leaving them failed #2630, fix: stop waiting forever on a hung picker thumbnail capture #2460, fix(desktop): hide terminal window when opening explorer on Windows (#2309) #2316) all apply cleanly on top of this branch. fix(desktop-gpui): request the Instant link while the capture is built on macOS #2592 doesn't apply on current main either, so that is not caused by this PR.
• feat: desktop design system, stabilization and dependency migration #2587, feat: bring desktop Studio UI and local playback to web #2312 and feat: add imported media tracks and editor waveforms #2308 are large branches that touch most GPUI files; whichever lands second needs a rebase.

How this was made

This PR is agent-generated.

• The finding comes from an audit of main by OpenCode scouts and Gemini 3.8 Flash, re-checked against the source by the orchestrating Claude session; an earlier attempt at this change was rejected in review for a compile error and a test that read the real library, and was not reused.
• Claude Opus 5.5 wrote this change and the test on current main and took the measurements above with the real crate functions.
• Gemini 3.8 Flash (Antigravity) reviewed the diff independently: it checked the ordering and tie-break equivalence, the original.png fallback, the laziness test, that no test touches the real library, re-ran cargo check and the library:: and main_window:: tests, and approved.

🤖 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]

Timestamp ties fail the test

If the temporary filesystem has timestamp precision coarser than 6 ms, consecutive PNGs can share a timestamp despite the sleeps. The strict > assertion then fails even when the listing is correct. The assertions that the hard-linked pair occupies the first two positions have the same dependency.

Use controlled sort times, or allow other ties and locate the hard-linked pair by path.

Prompt To Fix With AI

This is a comment left during a code review.
Path: apps/desktop-gpui/src/library.rs
Line: 1589-1593

Comment:
**Timestamp ties fail the test**

If the temporary filesystem has timestamp precision coarser than 6 ms, consecutive PNGs can share a timestamp despite the sleeps. The strict `>` assertion then fails even when the listing is correct. The assertions that the hard-linked pair occupies the first two positions have the same dependency.

Use controlled sort times, or allow other ties and locate the hard-linked pair by path.

---

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

[celeroncoder]

Fixed in 34c4f1c. set_modified is not an option here because the sort key prefers created(), so the test now tolerates ties instead: the hard-linked pair is located by bundle path and must share a timestamp, and every adjacent pair in the listing must be strictly newer-to-older or, on equal timestamps, keep read_dir order (the stable-sort property the production code relies on). The assertions that assumed the sleeps produce distinct times (legacy last, the pair at positions 0-1, the strict > chain) are gone; the equivalence with the old listing, the take(K) prefix loop and the laziness check are unchanged. The pair is deliberately not asserted adjacent, since on a coarse filesystem a third bundle could share the timestamp and sit between them in directory order.

Fix written by Claude Opus 5.5, 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