[lawrencecchen]

Summary

SWITCH FRAMES SF-5 (part B): a hidden terminal that prints made the daemon send scroll-changed to the Mac app's event subscription for every frame of its output. With 8 hidden busy terminals the app's subscription got 604 events in 5 s; with 24, 1910 (about 15 per terminal per second; 2x display, cmux24, round sw-sf5-events). The app's store applies nothing for this event: it reads a terminal's scroll position on that terminal's attach stream. Each event still woke the store's event pump and asked for a main-thread drain frame: the wakeup ledger showed DaemonStore.drain rising from about 0 to 57.8 frames/s, and 4106 display links made, with 24 busy hidden terminals.

Fix (an additive capability, both sides in this PR):

• cmux-tui: a connection that echoes subscribe-scroll-on-attach-v1 in set-client-info gets no scroll-changed on its subscribe stream. Every other client keeps it. identify lists the capability. Spec: cmux-tui/spec/commands.md, subscribe.
• App: DaemonCapability.subscribeScrollOnAttach (optional tier) is advertised, so the app's connection says it; plans/cmux-next/daemon-capabilities.json is updated.
• Compatibility: an older daemon ignores the capability (no change); an older app does not send it (no change). Attach streams still carry scroll-changed.

This is the daemon event feed, not key routing or output bytes, so it does not overlap typing-fix-2's area.

Testing

• cmux-tui-core server::tests::subscribe_scroll_on_attach (2 tests): red at 9d222a2 on Linux (step 153c8b00): the app connection got 2 scroll-changed events, and identify did not list the capability. Green at 1e68b01 (step 96deee45): 2 of 2 passed. (The other connection gets at least one scroll event, not 20: the subscriber mailbox keeps the newest scroll per surface.)
• After the merge with feat, 8c485a1: Swift compile (CmuxNext, app, cmux-tui) passed. The head 839b72a changes comments and docs only (review follow-up).
• App-level gate: the busy hidden terminal A/B on one mini (cmux24, same feat base, bench builds), with part A (terminal: a hidden terminal's output costs the launch-frame keeper no wakes (SWITCH FRAMES SF-5 A) #32384): base vs fix, CPU, wakeups and daemon events at 0 / 8 / 24 busy hidden terminals. Result: see the A/B table below.

Busy-terminal A/B (lead's gate), one mini

cmux24, 2x virtual display at 120 Hz, bench builds of the same feat base 507438d: base test/sf5-ab-base-bench (build bcd1b411) vs fix = #32384 (at f2fa334) + #32380 (build 9b6ddda3). Hidden terminals each print 20 lines/s; a quiet workspace is shown; 20 s windows.

busy hidden terminals	app CPU base -> fix	app main-thread wakes (debug.wakeups ledger, /s) base -> fix	DaemonStore.drain frames/s	launch-frame keeper wakes/s	process wakeups/s (proc_pid_rusage) base -> fix
0	2.70 -> 2.34 %	13.8 -> 17.5 (CEF message pump)	0 -> 0	0.1 -> 0	4.0 -> 4.2
8	5.78 -> 3.54 %	75.5 -> 15.9	23.4 -> 1.6	14.5 -> 0	221.1 -> 134.8
24	3.81 -> 2.01 %	72.7 -> 7.9	55.1 -> 4.8	14.9 -> 0	311.3 -> 114.4

• App CPU with 24 busy hidden terminals is within 1 point of the 0-terminal baseline (target met).
• The app's main-thread wakes are back near baseline (the CEF message pump is the baseline's own).
• Process-level wakeups stay at about 115/s with 8 or 24 busy hidden terminals: they come from the 8 warm hidden terminals' output being parsed off the main thread (the model update itself, about 14 wakeups/s per warm terminal). That is the next SF-5 cause; it is not in this PR.
• The daemon-event column of the probe (not shown) still counts scroll-changed, because the probe's own connection does not opt in; the app's DaemonStore.drain column shows its own subscription no longer gets them.

Observed / Not run

• Observed: the wakeup ledger and daemon event counts above (rounds sw-sf5-busy-ledger, sw-sf5-events on build 0ccecd0c); the Rust red and green; the prepush gate.
• Not run: the full cmux-tui suite (hosted --full); the A/B is below.

Known, not fixed here

• The daemon's own event thread still wakes for each scroll event before it drops it for this client (a daemon cost, not the app's).

Changelog

• Fixed: terminals that print while their workspace is hidden no longer wake cmux many times a second.

Proof

No capture: no visible change.

[github-actions]

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

Review 839b72a: 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=pr32380 -f ref=fix/switch-subscribe-scroll-on-attach -f platform=macos -f backend_mode=local

Shipped in:

• cmux NEXT nightly 1.0.0-nightly.4147460604401 (87f859ba07)
• cmux NEXT nightly 1.0.0-nightly.4147660705901 (b27c0b3daf)

[lawrencecchen]

Review (subagent, correctness first): SHIP at 839b72a

An independent review agent read the full diff (read only; it built and ran nothing), in two rounds.

• Round 1 at 8c485a1: SHIP, no must-fix. Checked: no app connection applies scroll-changed from subscribe (store, stage, coalesce key, mobile compat); attach streams keep the event; the phone lane forwards the phone's own set-client-info; set-client-info runs before subscribe on one connection, in order; the Rust continue sits with the other filters and skips no overflow or incognito bookkeeping; capabilities only add, so a late set-client-info is harmless; the capability registration is complete (no CI file missing); the Rust tests compile and assert what the PR says.
• Round 2 at 839b72a: comments and docs only (the stated reason was wrong: the app applies no scroll-changed; the identify capability paragraph now lists the name). The earlier review stands.
• Known: the daemon's event thread still wakes per scroll event before it drops it.

Proofs: Rust red 153c8b00 (both tests fail on asserts), Rust green 96deee45 (2/2), Swift compile at 8c485a1 passed. ready-to-land waits for the busy-terminal A/B on one mini (lead's gate).