[austinywang]

Summary

This makes cmux NEXT's blue unread pane ring behave the way Austin asked on 2026-10-20, matching cmux classic's ring.

• Arrival: when an agent finishes, a steady blue ring appears. notifications.attention.style now defaults to steady; blink and pulse stay available.
• Dismiss: the ring stays until you go into that pane, by clicking or typing in it or opening its notification. notifications.dismissal now defaults to click; focus stays a choice. Then the ring blinks blinkCount times and goes away, as classic's dismiss flash does. A ring cleared any other way (Mark Read, a timeout, a mute, Dismiss Highlight, another device) fades without the blink.
• Clicks read only the clicked pane: a click reads the pane under the pointer, found with a layout hit test. A click on the sidebar, the titlebar or another pane no longer clears the focused pane's ring.
• Finishes in a pane you can see: a turn or an OSC 7501 done that finishes in a visible pane keeps its ring under click. It posts no banner (cx-kxa2, cx-slk1h), the feed is told it didn't alert, and the timeout dismissal still gets its deadline.
• GUI agent chats: these now ring when their turn ends. acpmux reads a turn that ends in a shown chat at once (cx-n8ou), so nothing marked the pane. AgentTurnStateStore.onTurnEnded feeds NotificationCenterService.agentFinished. Mark Workspace Read, Mark All Read and Clear All clear those rings too.

Takes back 78680d5 (pushed directly on 2026-10-19, no PR), which set the focus dismissal default. Why: Austin dogfooded it and rings cleared before he clicked the pane. Lawrence's u2414 words were also "you click and it gets dismissed".

Changelog

• Changed: In cmux NEXT, a pane's blue unread ring appears steady when an agent finishes, including agent chats, and stays until you click or type in that pane; it then blinks out.

Testing

Command: cmux-ci run --class light --script scripts/ci/package-test-lane.sh --arg=suite --arg=Packages/macOS/CmuxNext --arg=<suites>

• Red 22646d3 (step 7ae3e9d2): UnreadRingClickReadTests fails. A click in the other pane read the focused pane (ack surface:1), and a click outside the panes read one too. The default tests fail on focus.
• Red fa5f7fd (step f7e502f3): no dismiss flash, and the style default was still blink.
• Head 65aae85 (step 454e5af5): 12 suites pass: AgentTurnRingTests, AttentionDismissFlashTests, AttentionDefaultTests, PaneAttentionDismissFlashTests, AttentionSteadyTests, StaticStatusIndicatorTests, BlinkRateCapTests, UnreadRingClickReadTests, UnreadDismissDefaultsTests, UnreadRingDismissTests, WatchedDoneReadTests and AgentChatDoneDotTests.
• scripts/cmux-next/prepush-gate.py: no new violations.

Observed / Not run

• Observed: Austin dogfooded tags 22678-unread-ring-click-only and 22678-unread-ring-steady-flash, which led to these changes. Tag 22678-unread-ring-gui-agents (1549a2a) is with him now.
• Not run: the generated docs/mdm/* still list the old defaults until the autofix regenerates them. The agentTurnEnded tab lookup and the observeAttention entry-read collection have no direct test. Nothing was run on the full CmuxNext suite.

🤖 Generated with Claude Code

[austinywang]

Correctness review (subagent), head 8c42df8

Round 1 on the first version: FIX

1. The shown set read state that isn't observed (focus topology, window occlusion). A tab switch in a pane without focus ran no pass, so that tab could be read later at an unrelated pass. Fixed: shownTabs(in:) reads TabSelectionMemory and LayoutModel.visiblePanes (both observed) and no longer filters on occlusion. Classic counted a rendered pane whatever covered the window.
2. Columns scrolled off screen in a strip or canvas counted as on screen. Fixed: only visiblePanes count.
3. An on-screen read retires the pane's approval banner. Kept on purpose: classic dismissed a visible pane's notification the same way.
4. Test gaps (the trivial log.count assertion, no unfocused tab-switch case). Fixed: the test counts ack lines and covers a background tab selected in the pane without focus.

Round 2: PASS. Workspace switches re-run the pass through viewedRevision (SettleKey includes the topology's workspace). visiblePanes is reported in a window that is never ordered in. PaneModel.id is the TabSelectionMemory key. Optional note applied: when the stored selection names a tab that just closed, fall back to the pane's default tab.

My own check: the branch diff is 4 files (two notification service files, ProgramStatusSeen.swift and the new test). Fleet: red 6905d76 fails the two new tests (step a567469d); the fix head passes UnreadRingClassicReadTests, UnreadRingDismissTests and WatchedDoneReadTests (step 641089e8). No tagged-build dogfood was run.

[github-actions]

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

Review 1549a2a: 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=pr22678 -f ref=feat-cmux-next-unread-ring-classic -f platform=macos -f backend_mode=local

Not verified: macOS jobs did not run: compile and app tests were skipped on 1549a2ad2a; ci-status is green without them.

CI passes on 1549a2ad2a (run 39550917171 attempt 1).
macOS jobs did not run: compile and app tests were skipped: Fast static checks skipped (the macos job needs both to succeed). Re-run CI once they pass, before merging.

Written by scripts/ci/classify_failures.py (ci-failure-attribution.yml); signatures are its SIGNATURES table. A machine verdict is the runner's fault, not this PR's; yours means the failing file is one this PR changes, also red on main that main's latest full suite fails the same way, seen on other PRs that it failed on another pull request's run lately.

web and size checks: green

[teamleaderleo]

Bead cx-zo71h created for this PR by the cmux coordinator.

[teamleaderleo]

cmux-next guards / defaults-snapshot is red on 23 of 23 fresh feat-cmux-next PR heads in the last 2:00:00, so it is most likely a base red, not this PR's. The base-reds lane has it; a base merge picks up the fix once it lands.

[austinywang]

Correctness reviews (subagents) since the first review comment. Head 65aae85.

1. Click default (4dc3b8f), FIX:
   • Visible finishes were read at arrival → kept unread under click, with no banner (fc67117).
   • Child windows fell back to the focused tab → hit-test in the host window (fc67117).
   • Overlapping hits were picked arbitrarily → no read unless one is focused (fc67117).
   • Feed note and timeout skipped on the early return → keepSeenUnread (d031e3b).
   • Re-review: PASS.
2. Steady ring and dismiss flash (f47f16e), FIX:
   • The flash keyed on focus, so Mark Read, a timeout or a mute on the focused pane blinked → only entry reads blink (entryReads → LayoutModel.attentionReadByEntry) (9dac6eb).
   • Static indicators → one fade (9dac6eb).
   • The flash is now in BlinkRateCapTests (9dac6eb).
3. Entry reads and GUI agent ring (9dac6eb, 1549a2a), FIX:
   • A stale entry read could blink a later Mark Read → taken on every change (65aae85).
   • Workspace, all-read and clear-all paths left chat rings → readAgentTurns (65aae85).
   • A chat ring now needs a local or mirrored host, and focus mode skips a viewed chat (65aae85).
   • Not re-reviewed after 65aae85.

Fleet run on 65aae85: 12 suites green (step 454e5af5). Not labeled ready-to-land: waiting on Austin's dogfood and the docs/mdm autofix.

[github-actions]

Shipped in:

• cmux NEXT nightly 1.0.0-nightly.3955987471201 (03960f5a64)
• cmux NEXT nightly 1.0.0-nightly.3956132659001 (233cc81754)