fix(cmux-next): Cmd-T keeps the pane's content until the agent page paints; new terminal tabs titled with their folder

[teamleaderleo]

Problem

Two pieces of known chrome changed after they first appeared (#17485, rows 4 and 5):

• Cmd-T showed an empty pane. The new tab's strip entry appeared right away, but the pane went to bare wallpaper until the agent chat page drew. On a warm app that lasted about 230 ms (frames 21.983 to 22.183 s below). op-next-layout measured about 700 ms on a cold app. An agent page is transparent until its document paints, and PaneContentView.show removed the outgoing view as soon as the new one was installed.
• Cmd-D titled the new terminal "Terminal" for one frame, then "~". An untitled terminal fell back to the generic name until the shell's first title arrived.

Before, Cmd-T from a terminal pane (fleet build of 4328967, the right pane only, 30 fps capture):

Fix

• Paint hold. The agent page now reports pane.painted once per document, on the second animation frame after its first post-handshake render. That is the first frame that shows what the page is, with the new chat's hero.
  • When a pane switches to an agent page that hasn't reported, PaneContentView keeps the outgoing view on screen. The new page sits above it at alpha 0 and keeps rendering.
  • The hold ends when the page reports, or after 500 ms. Both views then swap in one transaction.
  • Another switch or a detach ends a hold early.
  • Content that has already painted, and every non-agent content type, still swaps at once.
• Terminal title. An untitled terminal shows its abbreviated cwd ("~") instead of "Terminal". That is what the shell's first title usually is, so the label doesn't change a frame after the tab opens. With no known cwd it keeps "Terminal".

The hold reports after the handshake, not after mount, because of an intermediate capture (tag v3). Reporting after mount swapped in the pre-handshake frame: "Connecting" with no hero. The hero then appeared 83 ms later.

After, Cmd-T from a terminal pane (fleet build of the head, tag next-flicker-v4). The terminal stays until the agent page replaces it in one frame:

Tooling

• scripts/cmux-next/flicker-check.py takes a recording of one scripted interaction and finds two patterns. It exits 1 when it finds any.
  • Blinks: something changes and returns within 100 ms.
  • Staggered settles: three or more separate steps within 300 ms.
• scripts/cmux-next/flicker-capture.py records scripted scenarios on a capture Mac and runs the check on each. A scenario is real pointer moves, clicks and hotkeys, recorded by Cua Driver's ScreenCaptureKit recorder.
  • screencapture over SSH has no Screen Recording grant on the capture minis.
  • The recorder delivers up to about 30 fps, so a single 60 Hz frame can fall between captures. Anything lasting two frames or more shows.

Verification

• Swift tests (not run locally, the build host can't compile Swift):
  • PanePaintHoldTests (CmuxNextAppTests): the outgoing view stays until paint, painted content swaps at once, another switch ends the hold, and the 500 ms limit.
  • AgentPanePaintTests (CmuxNextAgentPaneTests): pane.painted parses and is a registered op, waiters run once, and painting doesn't count as a user touch.
  • UntitledTerminalTabTitleTests (CmuxNextBridgeTests).
• Each fix's test is committed before it, for the red/green record. CI runs them.
• bun test src/agent-session/acpmux: 982 pass. tsc --noEmit is clean, and build-agent-pane-web.sh --check says the bundle is current.
• Fleet builds (--backend-mode local): 1ff7c80e (baseline) and a32f99ed (head before the handshake change).
  • The after captures use a capture-only branch that also merges fix(cmux-next): keep next-turn pacer alive during tab replacement #17515 (flicker-capture-v3; v3 80da76e0, v4 56f8124f).
• Captures on a capture Mac, a capture Mac and a capture Mac: hover a workspace row, Home and +, open a workspace, Cmd-D, Cmd-T, and back to Home.
  • On the head, flicker-check finds nothing except the deliberate tab-insert slide.
  • The Cmd-D title change wasn't caught at 30 fps in my baseline, so the before frames are in cmux-next: layout shifts and click alignment, frame-by-frame survey #17485.
• Not covered: the cold New Tab (first agent page after launch). The 500 ms limit caps the hold there, so a cold page longer than that still shows the page before it paints.

Found, not fixed here

• Crash on Cmd-T, reported on fix(cmux-next): keep next-turn pacer alive during tab replacement #17515: an intermittent EXC_BREAKPOINT in AgentPaneNextTurnPacer deinit during AgentPaneView.installTransport. It hit v2 and v4, including with fix(cmux-next): keep next-turn pacer alive during tab replacement #17515 merged.
• First launch of a downloaded build: after Gatekeeper's "Verifying cmux…", the window drew black for about 100 ms, then empty, then "Conversations need a newer cmux daemon", then the UI. This is the daemon handoff and is left to the launch lane.
• Agent page chrome that mounts as data arrives, from a code search and not yet recorded. These are for op-next-agent-polish and op-next-layout:
  • The composer chips (mode, model, effort, context ring) each appear when their data does.
  • The header's trailing buttons do the same.
  • The composer is hidden while handoff state loads.
• Sidebar rows (sent to op-next-sidebar-merge): hover swaps the unread badge for the close button and re-truncates the title, and a group header's count appears on hover.

Changelog

Fixed: Cmd-T no longer shows an empty pane before the agent chat page appears, and a new terminal tab is titled with its folder from the first frame.

cmux-b2 (unregistered)

🤖 Generated with Claude Code

---

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

[coderabbitai]

Important

Review skipped

Auto reviews are disabled on base/target branches other than the default branch.

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

• Configuration used: Repository: manaflow-ai/cmux/.coderabbit.yaml
• Review profile: ASSERTIVE
• Plan: Advanced
• Run ID: f3e32088-6de3-45a6-980b-c4a9386880ac

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

• Autopilot · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[teamleaderleo]

Pre-merge review (subagent, correctness first): ship with nits, nothing blocking. Every hold ends: a page that paints, the 500 ms deadline, another switch, a detach, a moved tab or teardown. Late waiters and deadlines are tied to their own hold and a released pane. Parked spares and LaunchReveal are unaffected.

Fixed in 53144ff:

• The untitled-terminal test used "kind":"terminal", which decodes as .other, so it never covered a real terminal tab. It now uses pty.
• An outgoing browser whose header band the strip was pinned to would have shown its header jump during the hold. A band-pinned browser no longer holds.
• Outgoing internal pages (Bookmarks, History…) hide themselves when withdrawn, so holding them kept an empty pane anyway. Hidden content no longer holds; there's a new test for it.
• The deadline test now has 1 s of slack, up from 200 ms, for loaded runners.

Left as is:

• During a hold (500 ms at most), keys and clicks go to the new page, which is still invisible. That's the same target as before, when it was visible but blank, and type-ahead into the new chat is wanted.
• whenPainted waiters from holds that hit the deadline stay until the page paints. They're small closures holding a weak reference to the pane.
• An untitled terminal's strip title now repeats its cwd subtitle, and the sidebar still shows the empty title.

cmux-b2 (unregistered)

[github-actions]

Not verified: macOS jobs did not run: compile and app tests were skipped on beea666c0b; ci-status is green without them.

CI passes on beea666c0b (run 37491052595 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.

[teamleaderleo]

Batch queue: green. Stack next-batch/20261006 at 954248bfe072 (feat-cmux-next 7cf558be8de1 + #17527 #17537 #17542 #17553 #17554 #17559, generated files regenerated) built on the fleet, production: job c8f563681211aa65c6e94453, tag next-batch-f16f40ea (cmux-ci artifact c8f563681211aa65c6e94453). Validated this PR at 19fea9e56aef.

Own check still red: web / react-apps-check, formatting. Run bun run check:fix in webviews/ and push.

Next: land it yourself with gh-merge-green on this receipt.