[azooz2003-bit]

Summary

Follow-ups left by the reviews of #27664 (landed as 5e4c938), in the cmux-next iOS chat transcript and agent tab cards.

1. Very tall images re-decoded in a loop. SentAttachments.fill scaled only the short side, so a 1290x20000 screenshot kept a 600x9300 thumbnail (about 22 MB). A card whose last turns carried two of them passed the 32 MB bound, each landing evicted the other, resident() never held for both, and AgentCardRasterizer.frame re-queued and decoded again at 60 Hz.
   • A thumbnail now also cuts its long side to the middle 2 x side that its fill frame shows (SentAttachments.thumbnail): 600x1200 px, 2.9 MB.
   • The rasterizer holds the thumbnails a waiting card already has (AgentChatPreview.PictureHold) and lends them to the cache while it draws, so an eviction cannot race the draw.
   • The bubble's own .task, which ImageRenderer starts during the draw, finds the lent thumbnail and does not decode it again.
   • A decode lands directly in the holds waiting for it, so with many cards queued a fresh decode is not lost before the card's turn (from this PR's review).
2. Native-cell measurements were dropped. ChatNativeRowCell.preferredLayoutAttributesFitting returned attributes without ChatRowAttributes.shown, and the layout keeps only measurements that carry it. A native row that is not exact went back to its estimate on the next change above it. The native cell now carries the state it was configured with, and keeps a measurement only at the width its layout was made for.
3. aRowsHeightForAStateIsTheOneItSettledAt now requires a kept height (knownHeight != nil) for both the open and closed states, so it fails if measurements are dropped.
4. measuresSettle is per row. While a disclosure change waits for the list's reload, only rows whose disclosures (ChatRowLookup.fold) it moves are held. Another row's growth, such as a streaming answer, is kept.

Fixes, tests, red/green

Command for every red/green below (aws-m4pro-6, simulator cmux-listperf-1, Xcode 26.6):
xcodebuild test -scheme CmuxMobileKit-Package -destination "id=<cmux-listperf-1>,arch=arm64" -enableCodeCoverage NO CODE_SIGNING_ALLOWED=NO -only-testing:CmuxMobileAgentChatTests/SentAttachmentThumbnailsTests -only-testing:CmuxMobileAgentChatTests/NativeRowFoldTests -only-testing:CmuxMobileWorkspacesTests/AgentCardPictureTests

Red commit 0139cca holds the tests only (plus test hooks: SentAttachmentThumbnails.evict(attachmentID:), and AgentCardRasterizer.frame made internal).

Fix	Test	Red (0139cca)	Green
1, long-side cap	SentAttachmentThumbnailsTests.twoVeryTallImagesSettleUnderTheBound (two 600x9000 images, 32 MB bound)	not resident; decodes 62 -> 82 while settled; 21.6 MB kept; thumbnail not in memory	passes
1, rasterizer holds	AgentCardPictureTests.anEvictedThumbnailIsDecodedOnceThenTheCardIsDrawn	2 decodes (the held image was decoded again)	1 decode; card drawn with both images
1, hold filled on landing	SentAttachmentThumbnailsTests.aWaitingPictureHoldsWhatIsDecodedForIt	red at d11d586: not resident, 3 decodes	green at ee1dc9c: 2 decodes
2	NativeRowFoldTests.aNativeRowKeepsTheHeightItsCellMeasured	nothing kept; row back to its estimate (30 pt off)	passes
3	NativeRowFoldTests.aRowsHeightForAStateIsTheOneItSettledAt (strengthened)	passes (measurements of SwiftUI rows were already kept)	passes
4	NativeRowFoldTests.anotherRowsGrowthIsKeptWhileADisclosureIsPending	the other row's growth was dropped	passes

The review subagent's P3s are fixed in d11d586:

• A card whose key no longer needs drawing drops its hold.
• When the scaled copy cannot be made, the fallback draws the whole image scaled into the crop.
• A native measurement is kept only at its layout's width.
• A reused native cell clears shown before a configure that may return early.

Testing

• iOS simulator, full CmuxMobileKit-Package suite at ee1dc9c, aws-m4pro-6: TEST SUCCEEDED. 1,904 Swift Testing tests in 13 test runs, all passed.
• macOS swift test in CmuxMobileKit at ee1dc9c, aws-m4pro-5: 1,598 tests, 4 failures. All 4 are the known host-only PreviewInsufficientSpace failures (LinkFilePreviewStreamTests x3, TerminalFloodTests x1); nothing else failed.
• python3 scripts/cmux-next/prepush-gate.py: no new violations.
• Correctness review (subagent) of ac20cd3: SHIP. The P2 and P3s above are fixed; one P3 was not taken, see below.

Observed / Not run

• Observed: the red and green runs above, both full suites, and the gate.
• Not taken from the review: P3 Disambiguate Sentry environment between production and dev #2 said the lent check in SentAttachmentThumbnails.start is dead because the bubble's .task runs after the draw. A diagnostic run showed otherwise: the decode started inside ImageRenderer's synchronous draw (2 decodes at draw time). The check is what turns anEvictedThumbnailIsDecodedOnceThenTheCardIsDrawn green, so it stays.
• Not run: no on-device or tagged-build dogfood, and no perf trace of the tab grid with real tall screenshots. The loop is covered by the unit tests only.
• AgentCardPictureTests uses SentAttachmentThumbnails.shared. Its decode-count assertion could be disturbed by another suite decoding into the shared cache at the same moment; its suite is .serialized, and it passed in every run here.

No capture: no visible change apart from tall images in agent cards no longer re-decoding. The thumbnails draw the same pixels: the crop keeps everything the bubble's fill frame shows.

Changelog

Fixed: an agent tab card showing two very tall screenshots no longer re-decodes its images in a loop.

🤖 Generated with Claude Code

[github-actions]

cmux status _{(one comment, edited in place; each line is updated by the job that owns it)}

Review 6758f50: HQ feed card · gallery before/after runs · checks

App of this head (app.zip, kept 1 day)

gh workflow run reload-build.yml -R manaflow-ai/cmux --ref feat-cmux-next -f tag=pr30681 -f ref=ios-lists-followups -f platform=macos -f backend_mode=local

Shipped in:

• cmux NEXT nightly 1.0.0-nightly.4153956844201 (48fc806e8b)
• cmux NEXT nightly 1.0.0-nightly.4154067381701 (ecce1e0d8f)
• cmux NEXT nightly 1.0.0-nightly.4154301631801 (11e5aeaa08)
• cmux NEXT nightly 1.0.0-nightly.4154452830701 (f9f6f341a5)

[cursor]

Bugbot needs on-demand usage enabled

Bugbot uses usage-based billing for this team and requires on-demand usage to be enabled.

A team admin can enable on-demand usage in the Cursor dashboard.

[azooz2003-bit]

Review (Claude subagents, read-only, correctness first): SHIP at 6758f50.

• At 0953544 (the four fixes plus a clean merge of feat-cmux-next): SHIP. Tall thumbnails keep the centred crop the bubble shows (checked tall, wide and 288 px); hold and lent keyed like the cache; native heights kept only at the layout's width; per-row measuresSettle covers every disclosure key. One P2: held images were not bounded and outlived the grid.
• That P2 fixed in 6758f50: one waiting card holds at a time (it stays first until its decodes land), released when the grid moves, on a memory warning, on drop or an undrawable key. No starvation, no re-decode loop, the two-tall-images anti-loop guarantee kept. The decode-count test counts only its own attachments with per-run ids.

iOS 26.5 simulator (aws-m4pro-6): full CmuxMobileKit-Package suite green; AgentCardPictureTests and SentAttachmentThumbnailsTests 5/5 iterations.
P3s left (order only, memory warning does not purge the cache itself, | in ids for test counts).

[lawrencecchen]

Merge service note: this PR has ready-to-land, but the merge round leaves it out with "no comment whose first line is "Review (): SHIP at 6758f50"". The receipt from 20:14Z has markdown bold and a final period on its first line (**SHIP at …**.), so it does not match. The reviewer can post the same receipt again with a plain first line, exactly:

Review (Claude subagents, read-only, correctness first): SHIP at 6758f5003ca2bd8477873809b97db9996c8eb481

The red release-notary / Guard status checks are a base finding (fixed on feat by #32548), so they do not hold this PR.

[lawrencecchen]

Review (claude-lead-review-subagent): SHIP at 6758f50

Read-only review of the diff from merge-base d9d62ba to 6758f50 (9 files). I did not build or run anything. Correctness came first, then the CLAUDE.md rules.

Checked and found correct:

• SentAttachments.thumbnail and crop: the visible part of a 200x150 (4:3) or 96x96 scaledToFill frame along the long side is at most 4/3 of the short side, so a centered crop to 2 x side keeps every pixel the bubble shows. This holds for portrait and landscape images and also when the short side is below side (no upscale). The full-screen preview uses the file bytes, not the thumbnail, so the crop does not lose content there. The fallback (scaled ?? image) draws into the same centered frame. UIGraphicsImageRenderer off the main thread is safe.
• resident(_:holding:) and landed: a decode that lands goes into each live hold that waits for it. The hold keeps only the keys that turns draws. Undecodable keys (settled == false) are skipped, so the waiting card cannot stall on them. Weak waiters are cleared on landing.
• AgentCardRasterizer.frame: the waiting card stays first until its decodes land. Only one hold exists at a time. The hold is released on want with an undrawable key, on drop, on setIdle(false), on a memory warning and before the draw (the local hold keeps the images alive through drawing). The ahead path inserts a card that is not resident at the front. I found no ping-pong and no starvation path, because every decode completes.
• drawing lends the images only during the synchronous ImageRenderer pass and restores lent with defer. start and image both look in lent.
• ChatNativeRowCell.preferredLayoutAttributesFitting: shown comes from the layout item as configured (the same pattern as ChatRowCell). It is cleared before the early return in configureNative. It is set only when layoutWidth (environment width + 40) matches the attributes width. Because ChatRowAttributes.isEqual compares shown, a measurement at the same height still reaches shouldInvalidateLayout, which records it and returns false. Exact native rows still ignore measured.
• measureSettles(item): ChatRowLookup.fold covers every disclosure key prefix the rows read (tool-, group-, thought-, diff-, code-). So a row whose fold the pending change does not move draws the state of its item, and its measurement can be kept.
• The tests assert the changed behavior, and the red evidence in the description matches what the code paths do. No try!, no blocking call on the main thread, no new font literals, no animations and no new user-facing strings.

Non-blocking notes (P3):

1. SentAttachmentThumbnails.swift:39 decodesByAttachment is a counter for tests only. It lives in production, grows with every distinct attachment id and is never trimmed, and it splits the key string on each decode. It is small, but you could bound it or put it behind DEBUG.
2. AgentChatPreview.picturesResident(_:) (the overload with no hold) and SentAttachmentThumbnails.resident(_:) have no production callers now (REVIEW.md, "remove the API your change made unreachable"). Also, each call of resident(_:) registers a throwaway hold in waiting while a decode runs. That is harmless, but it is extra work.
3. New test hooks in production source: evict(attachmentID:), holdsThumbnails, and frame() made internal. This follows the existing decodes/renders pattern, but REVIEW.md section 8 discourages it.
4. While a card waits, other queued cards wait behind it for one decode duration. This is intended (it bounds the hold to one card), but a tall image can delay later cards by the decode time.

Checks: the only reds are nightly-guards / Guard status and release-notary, the known base finding (fixed on feat by #32548). The PR causes no red of its own.