cmux-next: run ! commands in place instead of jumping to a terminal

[teamleaderleo]

Typing ! in the agent composer or on the New Tab screen used to replace the tab with a terminal. Now ! as the first character switches the prompt to shell mode in place, and Enter runs the command in the chat's folder with its output in the transcript. You stay in the chat.

Behavior

• ! first enters shell mode: a ! glyph in the theme's magenta (ANSI 5), a monospace field, a 1px edge on the box. Same padding and line height as the prompt. No popups.
• Backspace on an empty command and Escape leave it. What was typed moves to the prompt.
• A file or image pasted in shell mode stays attached for the next prompt, and the mode stays.
• Enter runs the command in the chat's folder (the location row). The block shows where it ran: command, Stop (Ctrl-C) while it runs, then exit status and time, the last 8 lines of output (Show all output for the rest), and Open in terminal. Open in terminal types the command into a new terminal tab in that folder without running it. Nothing opens a terminal on its own.
• The command stays in the composer as a removable $ cmd chip. When the next prompt goes, the chip carries the command, folder, exit status and the last 16 KB of output to the agent.
• New Tab screen (layout b): !cmd then Enter turns the page into a chat in the page's folder that shows the block. Suggestion rows hide in shell mode. Recent chat cards stay. The default New Tab page that 63ff230 restored keeps its ! Terminal kind prefix unchanged.
• Cloud chats have no shell mode for now. ! there is plain text.

Host

shell.run, shell.read, shell.stop on the agent pane bridge (AgentPaneShell):

• $SHELL -l -c cmd in the chat's folder, in its own process group, stdin /dev/null, stdout and stderr on one pipe, TERM=dumb, PAGER=cat.
• shell.run needs a real key press or click in the pane (AgentPaneUserGestures.consume()), so page script cannot run a command. One gesture runs one command.
• shell.read long-polls: it answers when there is output or the command ended, else after 10 s. Reads never split a UTF-8 character.
• Stop sends SIGINT to the group, then SIGTERM and SIGKILL 2 s apart. Closing the pane hangs up running commands.
• Deadlines (a read's wait, Stop's escalation, the 250 ms drain after exit) are one-shot DemandTimers. The event stream is bounded and ends only once the run has both its exit and its EOF.
• Bounds: 4 running commands per pane, 1 MB kept per command, 16 runs kept. No timers while idle.
• Shell requests are left out of the bridge's request log, since a command can carry secrets.

The old ! path (tab.typeAhead, the R81 terminal placeholder) is no longer used by the page. Its host handler and the bench_bang debug bench stay for now and will be removed separately.

Validation

• Red, at 752666e (tests only): 9 failing in bun test src/agent-session/acpmux/newtab src/agent-session/acpmux/quickComposer.test.tsx.
• Green at the head: bun test src/agent-session/acpmux: 1038 pass, 0 fail. vp check (format, lint, types): clean.
• scripts/cmux-next/build-agent-pane-web.sh --check: bundle current.
• verify-local.py --only swift-syntax --swift-changed origin/feat-cmux-next: 8 files pass. The localization parity check reports 27380 errors with or without this change, all inherited.
• Swift tests added in AgentPaneShellTests (run, exit code, Stop on a process group, missing folder, 4-command cap, UTF-8 boundary, request decode, gesture gate). Not run locally: they run on CI.
• check-concurrency.sh and check-no-godfiles.sh --only swift: pass.
• Fleet build shell-3 at 29be512, before/after stills on a capture mini: see the comment.

Localization: 10 new pane strings in all 21 pane locales, 5 host statuses in all 21 module locales.

Part of the cmux-next parity work listed in #16174 (comment).

Changelog

Changed: ! in the agent composer and on the New Tab screen runs a shell command in place, with its output in the chat, instead of opening a terminal.

🤖 Generated with Claude Code

[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: 72d63a7e-04c6-470d-8d11-41fc39c37425

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.}

[github-actions]

Not verified: macOS jobs did not run: compile and app tests were skipped on f7ecbbac0d; ci-status is green without them.

CI passes on f7ecbbac0d (run 37516213292 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: dropped from batch next-batch/20261006 (feat-cmux-next 7cf558be8de1), conflict with feat-cmux-next.

• webviews/src/agent-session/acpmux/conversation/conversation.css: both sides changed it
• webviews/src/agent-session/acpmux/model.ts: both sides changed it
• webviews/src/agent-session/acpmux/newtab/NewTabScreen.tsx: both sides changed it
• webviews/src/agent-session/acpmux/newtab/screen.css: both sides changed it
• webviews/src/agent-session/acpmux/newtab/screenActions.test.ts: both sides changed it
• webviews/src/agent-session/acpmux/newtab/screenActions.ts: both sides changed it

Next: merge feat-cmux-next, resolve, push.

[cursor]

Bugbot is paused — on-demand spend limit reached

Bugbot uses usage-based billing for this team and has hit its on-demand spend limit.

A team admin can raise the spend limit in the Cursor dashboard, or wait for the next billing cycle to continue.

[teamleaderleo]

Fleet before/after on a capture mini. Before: shell-base (feat-cmux-next at f3c86006). After: shell-3 (this PR at 29be512). Both use the showcase's seeded Claude Code chat.

Before. ! is plain text in the agent composer. !git --version; ls /Applications | head -5 goes to the agent as a prompt.

! typed	Return

After. ! turns the prompt into shell mode in place: magenta ! glyph, monospace field, 1px edge. Same box, nothing moves. Return runs the command in the chat's folder. The block shows the command, exit status and time, output and Open in terminal. The next prompt carries the run as a removable $ chip. No terminal opens.

! typed	command	Return

Stills only: CuaDriver's recording captured no frames (known finalize bug, REPAIR.md).

[teamleaderleo]

Batch queue: pushed bun run check:fix output (web formatting and lint autofix) to this branch; landing once its checks rerun.

[teamleaderleo]

Batch queue: green. Batch stack 87bf7dcfd4d7 on feat-cmux-next 7d209e8ddbed with #17520 #17534 #17541 #17551 #17552 #17558 #17582. Fleet production build: job 3b88261eeebcf0f42557a5e7 (cmux-ci artifact 3b88261eeebcf0f42557a5e7 (job 3b88261eeebcf0f42557a5e7, tag nb-local-20261006-153816-1)). heavy tier did not run on this branch (run); your PR's own checks are the tests. Validated this PR at 29be512b2eb2.

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

[teamleaderleo]

Batch queue: green. Batch stack c03ab5ca52f8 on feat-cmux-next 3892c6ce7b55 with #17520 #17527 #17534 #17537 #17541 #17542 #17551 #17552 #17553. Fleet production build: job d1e25b8a9144e8ef3ef2287e (cmux-ci artifact d1e25b8a9144e8ef3ef2287e (job d1e25b8a9144e8ef3ef2287e, tag nb-local-20261006-160146-1)). heavy tier did not run on this branch (run); your PR's own checks are the tests. Validated this PR at 988359b22471.

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

[cursor]

Bugbot is paused — on-demand spend limit reached

Bugbot uses usage-based billing for this team and has hit its on-demand spend limit.

A team admin can raise the spend limit in the Cursor dashboard, or wait for the next billing cycle to continue.

[teamleaderleo]

Batch queue: pushed bun run check:fix output (web formatting and lint autofix) to this branch; landing once its checks rerun.

[teamleaderleo]

Batch queue: green. Batch stack 23ce15a389b7 on feat-cmux-next 96f7421369d0 with #17541 #17600 #17602 #17605 #17611 #17620. Fleet production build: job 7c149b4a67df53c3864a6572 (cmux-ci artifact 7c149b4a67df53c3864a6572 (job 7c149b4a67df53c3864a6572, tag nb-local-20261006-182815-1)). heavy tier did not run on this branch (run); your PR's own checks are the tests. Validated this PR at 2862af7f1cfa.

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

[teamleaderleo]

Batch queue: green. Batch stack ed71230a0e11 on feat-cmux-next ff267c4b41d1 with #17541 #17611 #17624 #17625. Fleet production build: job 362a7927afa7714ace3f1b33 (cmux-ci artifact 362a7927afa7714ace3f1b33 (job 362a7927afa7714ace3f1b33, tag nb-local-20261006-190840-1)). heavy tier did not run on this branch (run); your PR's own checks are the tests. Validated this PR at f7ecbbac0dc4.

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

[teamleaderleo]

merge-override: only cmux-next swift test is red, on base-owned DebugMouseButtonTests committed red on feat-cmux-next; shell tests pass