[azooz2003-bit]

Summary

The Feed's rows are now UIKit: a UICollectionView with a diffable data source keyed by event id, and cells made of plain UIKit views (no SwiftUI hosting in rows). Each row's layout (every frame and attributed string) is computed off the main thread and kept until its event or row state changes. Markdown bodies are native attributed text, code cards and tables using the web renderer's metrics; only math still uses the pooled web view, once scrolling settles. Replaces #24467, whose SwiftUI trimming showed no gain.

Why: each SwiftUI row that scrolled in cost 14–20 ms of main-thread work (SwiftUI building and measuring the whole row tree), about one per row at fling speed. A store update also re-walked all ~300 rows.

What stays the same: the reply box (including terminal-agent replies typed into the terminal), question cards with drafts, Allow/Deny with Face ID, Answer/Open in Terminal, Decline, the … and long-press menus, the rounded swipe actions with the drawer handoff, quote bubbles, Show more, Copy/Share, links (web opens Safari, a file path opens the event), pull to refresh, the "N new" pill with the reading position kept, routing to the source Mac, and the demo driver.

Details, numbers and evidence: cmux-next-ios/verify/feed-perf-20261023.md.

Results (iPhone 16 Plus simulator, 320 events, 2400 pt/s, base vs fix in the same hour)

Run	Base hitch ms/s · p99 · main CPU ms/s	Fix hitch ms/s · p99 · main CPU ms/s
no live traffic (4 runs)	38–56 · 35–37 ms · ~183	0.7–1.7 · 16.67 ms · ~58
4 live changes/s (4 runs)	49–62 · 35–38 ms · ~190	1.1–3.3 · 16.67 ms · ~80

The < 5 ms/s target is met in every run. Without live traffic the latest runs have no frame over 33.3 ms. With live traffic a few single frames reach 34–51 ms per run, from a new cell's first commit or an insert landing in that frame.

Real use: two rig Macs, real Codex and Claude sessions on one, permission asks on the other, the phone linked to both. See the verify doc.

Changelog

• Changed: the Feed scrolls smoothly with many events: rows are native views laid out ahead of time.

Testing

• swift test --filter "Feed|TerminalFeed|MultiMac|FrameInterval|Markdown" on mini-cmux7s: 83 tests in 21 suites passed.
• New FeedListPlanTests committed red first (8ef597f55f73): one layout per event, state and width; a change reconfigures only its row; snapshot edits; only never-shown events count as new.
• Demo simulator builds on the fleet; screenshots of base vs new for every item kind, light and dark, under verify/evidence/feedui/shots/ (not committed; numbers in the doc).

Observed / Not run

• Observed: frame-log A/B runs and sample profiles as listed; side-by-side screenshots.
• Not run: Instruments traces (xctrace does not record over SSH on the minis; hq CronCreate durable: true silently downgraded to session-only inside cmux #3395 documents the fix), any device run (simulator only, as asked).

Review

Two correctness subagent reviews (at 26bc5b4 and 5144b51); findings and fixes are in a PR comment.

No capture: the Feed is iOS; light and dark simulator stills of every item kind are listed above and described in the verify doc.

🤖 Generated with Claude Code

---

^{Need help on this PR? Tag @codesmith-bot with what you need.}

---

Summary by cubic

Replaces the Feed's SwiftUI rows with UIKit UICollectionView cells laid out off the main thread, cutting main-thread hitches from 38–62 ms/s to 0.7–3.3 ms/s on a 320-event iPhone simulator feed (p99 frame time 16.67 ms, within the 33.3 ms budget). The prior SwiftUI trimming attempt showed no gain.

Refactors

• Each row's frames and attributed strings are computed off the main thread and cached until its event, row state, or text size changes; a store update re-keys only the rows whose event changed, and a reconfigure re-sums only the frames below the first changed row.
• Markdown bodies, code cards, and tables render as native attributed text using the web renderer's metrics; only math uses the pooled web view, after scrolling settles.
• Fixes a crash when a row height changed right after an insert, before the layout pass ran: the quick re-sum path now only runs on current frames.
• All Feed interactions are preserved: reply box (including terminal-agent replies), question cards with drafts, Allow/Deny with Face ID, Open in Terminal, swipe actions, quote bubbles, Show more, copy/share, links (web opens Safari, a file path opens the event), and the "N new" pill with the reading position kept.

^{Written for commit 659e639. Summary will update on new commits.}

[github-actions]

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

Review 659e639: 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=pr24683 -f ref=feat-cmux-next-ios-feed-uikit -f platform=macos -f backend_mode=local

[azooz2003-bit]

Correctness reviews (subagent) and resolutions

Review 1, at 26bc5b4: 4 blocking findings, all fixed in 1314ffb

Finding	Fix
The unread dot and swipe actions were not redrawn on a read-only change (empty plan returned early)	redraw computed before the early return
A width or Dynamic Type change dropped rows, lost row state and counted a false "N new"	relayoutAll keeps order and state, waits for all layouts, then reconfigures every row; nothing counts as new
Question card Back, Next and Answer were offset by the card's x	Frames used as given (card coordinates)
A saved question draft was not shown, but was acted on	Drafts are seeded before layout; state kept until the question stops pending; Other takes the keyboard only on .beginOtherEditing
Non-blocking: stale accessibility ids, header not re-measured on text size change, false new counts, Other refocusing, reply observation chains piling up, submit dropping picks	All fixed
Risks: duplicate ids, snapshots applied mid-layout, minute timer not invalidated	Ids deduplicated; width change deferred; timer invalidated in deinit

Review 2, at 5144b51: nothing blocking; fixed in 6da6b1c

Finding	Fix
A submitted question did not redraw at once	mutateState
VoiceOver kept reading "Unread" after a read on the Mac	accessibilityElementsHidden updated in redraw
A prefetched cell missed a read flag or minute change	Catch-up in willDisplay
Header measured during a trait-change layout pass	Deferred, using the collection view's size category
A web body's view was created mid-scroll	Created only once the row settles
Nits: Set lookups, reply height kept, Other resigned on reuse	Done

Not changed: the tab bar badge rebuild cost (SwiftUI TabView, in RootView, also in the base) is left for a follow-up.

[blacksmith-sh]

Found 2 test failures on Blacksmith runners:

Failures

Test	View Logs
test_base_passes (main.NoticesPairTests.test_base_passes)/ test_base_passes (main.NoticesPairTests.test_base_passes)	View Logs
test_the_repository_highlights_validate (main.DigestTests.test_the_repository_highl ights_validate)/ test_the_repository_highlights_validate (main.DigestTests.test_the_repository_highl ights_validate)	View Logs

^{Need help on this PR? Tag @codesmith-bot with what you need.}

[teamleaderleo]

Bead cx-frdwq created for this PR by the cmux coordinator.

[azooz2003-bit]

Review 3, the merge carry-over (at ee9e53c): nothing blocking; fixed in 3687fdf

• Links in short bodies opened the event instead of Safari. The hit test is now shared and also checks the inline label.
• A tap on blank space near a link opened Safari. The hit is now bounded by the glyph's rect.
• The hit test's text container now has unbounded height.
• "Open in Terminal" is an explicit action. Before, a "terminal" option id could have been sent by an agent and taken as that action.
• File paths in code spans are no longer underlined, since the web Feed never linked them.

Review 4 (at 3687fdf): one likely crash, fixed in 659e639

• FeedListLayout.setHeights quick path: after a full reset, a height change that arrived before prepare() re-summed frames of the old length, writing past the end of the array.
  • Red: FeedListLayoutTests, iOS simulator at 9e2ed53, "Crash: xctest at FeedListLayout.resum(from:)".
  • Green: at 659e639, 12 passed with FeedListPlanTests. The quick path now runs only on current frames.
• A link tap in the hidden overflow under Show more now opens the event.
• The test moved into CmuxMobileLiveTests (a direct Feed dependency).

Measured at 3687fdf and 659e639

• Synthetic, no live traffic: 0.2 and 0.9 ms/s.
• Synthetic, 4 live changes a second: 1.7 and 4.0 ms/s.
• Real use (two rig Macs, real Codex and Claude, a loaded real Feed): 0.0–0.1 ms/s against base's 23–35 ms/s. Details are in verify/feed-perf-20261023.md.

CI

• CmuxMobileKit tests (macOS swift test, iOS simulator): success.
• defaults-snapshot and nightly-guards / release-notary are base findings: both are red on cmux next new tab: cmd-t spam and fast closes no longer flash black, draw the chip row twice or show the chat screen #24692, cmux-next iOS: TUI scroll tail probes recover a flick's lost last packet #24691, fix(cmux-next): PageScriptPipeline compiles again (no covariant Self in a default argument) #24687, next: preserve notification producer identity (#3419) #24668, cmux next: the diff tab picks split or unified from its width, like the codex app #24664 and fix(cmux-next tests): CmuxNextAppTests compile again (ThemeStore.shared removed; let SidebarGroup) #24655.
• cargo check (x86_64-pc-windows-gnu) failed on acpmux/hub/titles.rs. That came in with the base merge (fix(next): title chats from the first user message #24643), and base fixed it in f9d7e1ab3cb1, which is merged here.