Cmux

cmux

Purpose

You are driving cmux, a CLI + socket for controlling terminal surfaces (and the agents running inside them). Every window, workspace, pane, and surface is a real, addressable object you can spawn, prompt, read, and close from the command line.

Instructions

Discover commands first (--help)

Before doing anything, run:

cmux --help

Then drill into any subcommand you intend to use:

cmux <command> --help     # e.g. cmux workspace --help, cmux send --help

cmux evolves; trust --help over memory. Never guess flags — confirm them.

Understand the hierarchy

Everything nests in one tree. Learn the boxes and the verbs fall out:

• Window → a top-level OS window.
• Workspace → a sidebar entry ("tab") inside a window.
• Pane → a split region within a workspace.
• Surface → a tab within a pane (a terminal or a browser).

Use cmux tree --all (or cmux workspace list / cmux list-pane-surfaces) to see the current state before you act.

Create a workspace and inject credentials (--env-file)

Create a workspace and capture BOTH refs in one call. --json returns the workspace_ref and the initial surface_ref — grab and thread them; never guess positional refs.

cmux workspace create --name <name> --cwd <dir> --env-file .env --json

• cmux workspace create supports --env-file, which loads that file's environment variables into every surface in the workspace — so an agent launched in a pane (claude, pi, codex, gemini) comes up already authenticated, no manual export needed.
• Default --env-file to .env (the repo's .env) unless told otherwise: --env-file .env. That is the canonical source for OPENROUTER_API_KEY, ANTHROPIC_API_KEY, etc.
• Pair it with --layout <compact-json> to boot a whole multi-pane team declaratively in one call (each pane's command auto-launches its agent).
• Don't inject over a working login. If an agent is already authenticated (e.g. Claude Code), don't push a placeholder key over it via --env-file; scope credential injection to the agents that actually need it.
• Assume the keys are already set up — and never read their values. By default, just point --env-file at .env and proceed; do not cat .env, echo $OPENROUTER_API_KEY, or read-screen a surface to capture a key. Only if an agent actually fails to authenticate should you validate, and do it safely: cmux workspace env --workspace <ref> --mask shows that a var is present without revealing it, and [ -n "$VAR" ] confirms it is non-empty. Report the masked/presence result, never the secret itself.

The control loop

You operate surfaces the way a person would, but over the CLI:

• cmux send --surface <ref> "<text>" — type text into a surface.
• cmux send-key --surface <ref> enter — submit it (press a key). send types; send-key submits — they are separate steps.
• cmux read-screen --surface <ref> — read what's on screen (add --scrollback for history). This is your eyes.
• cmux close-surface --surface <ref> — end a surface cleanly.

Wait for agents via notification events (don't busy-poll)

Instead of looping on read-screen, subscribe to cmux's push channel and block until an agent finishes its turn. This is verified working for pi, codex, and Claude Code.

cmux events is the wait channel — not cmux wait-for. cmux wait-for <name> is an unrelated named-token rendezvous (a manual semaphore you signal yourself); it does not know when an agent finishes. The agent-completion signal is the notification event category.

Prerequisite — install the notification hooks once:

cmux hooks setup            # wires pi, codex, opencode, gemini, … to emit on turn-stop
# or per agent:  cmux hooks pi install   /   cmux hooks codex install

Claude Code emits notifications out of the box when launched inside cmux (no cmux hooks entry needed). Without hooks, an agent stays silent and you're back to polling — so install them before relying on the wait.

What an agent emits when its turn ends — one event per completed turn:

{ "name": "notification.requested", "category": "notification",
  "workspace_id": "120FC732-…", "surface_id": null, "seq": 1512, … }

Match on workspace_id — for hook-emitted notifications surface_id is usually null, but workspace_id is always set. The title/body are redacted in the event (you get the signal, not the text), so once it fires, read-screen that workspace's surface for the actual reply. Filter to --name notification.requested; a sibling notification.clear_requested fires when a surface gains focus and is just noise.

Block until a specific agent finishes (capture its workspace_id first via cmux list-workspaces --json --id-format both):

WS=<agent-workspace-uuid>
# Start the listener to a file BEFORE sending the prompt, then poll the file.
cmux events --name notification.requested --no-heartbeat --no-ack > /tmp/cmux.ev &
EV=$!
cmux send --surface <ref> "<task>"; cmux send-key --surface <ref> enter
# wait (bounded) for this workspace's turn-done event
until grep -q "\"workspace_id\":\"$WS\"" /tmp/cmux.ev; do sleep 1; done
kill $EV
cmux read-screen --surface <ref> --scrollback --lines 40   # now read the reply

Pitfall: a cmux events | jq … & pipeline in a one-liner can stall on stdout buffering — stream to a file and poll the file (above), or pass jq --unbuffered. For a durable cursor across reconnects use cmux events --cursor-file <path> --reconnect.

Launching the pi agent

pi is an interactive TUI agent — launch it as pi --model … "<task>".

• Launch it inside a pane (via cmux send + send-key enter), not from your own non-interactive/batch shell.

Launching Codex — run it in yolo / auto mode

When launching Codex in a pane, start it unattended so it doesn't stall on approval prompts (it's running inside cmux, driven by an orchestrator). Pass the flag at launch — do not edit Codex's global config:

• Yolo (full, no sandbox): codex --dangerously-bypass-approvals-and-sandbox "<task>" — skips every approval prompt and the sandbox. Use only because the run is orchestrated/observed.
• Auto (sandboxed): codex --full-auto "<task>" — automatic execution inside a workspace-write sandbox; safer when full access isn't needed.

Default to yolo for hands-off fleet runs; reach for --full-auto when you want a sandbox. These are per-launch flags, so they never change the user's global Codex setup.

Always launch Codex with the gpt-5.5 model unless a prompt specifies otherwise — pass -m gpt-5.5 at launch, e.g. codex -m gpt-5.5 --dangerously-bypass-approvals-and-sandbox "<task>". If a prompt names a different Codex model/effort, use that instead; gpt-5.5 is just the default.

Launching Claude Code — use cc bypass mode

Plain claude launches in ask-for-permission mode: it will decline to run Bash/edits and instead print instructions, then end its turn. For a hands-off fleet agent, launch it the same way you yolo Codex — bypass permissions at launch:

• cc bypass: claude --dangerously-skip-permissions "<task>" — Claude's equivalent of Codex yolo. The composer then shows ⏵⏵ bypass permissions on and it executes shell/edits without prompting. (--dangerously-skip-permissions is a per-launch flag; it doesn't change global Claude settings.)

Caveat verified in testing: a notification still fires on turn-completion even when Claude refused to do the work — so if you only watch events, you can mistake a "declined, nothing happened" turn for success. Always read-screen (or check the artifacts) after the event, don't trust the event alone. Claude Code emits cmux notifications out of the box (no cmux hooks entry needed); see Wait for agents via notification events above.

Best practices

1. --help before every unfamiliar verb. Confirm the subcommand and flags exist.
2. Look before you leap. Inspect with tree / list / read-screen before sending or closing anything.
3. Refs are positional and renumber. surface:N / workspace:N shift as things open and close. Re-read the tree right before you act; for anything long-lived, anchor to a stable window UUID, not a positional ref.
4. Type then submit. A prompt isn't sent until you send-key enter. Give agents a beat before you read-screen their reply.
5. Read back to verify. After sending a command, read-screen to confirm it actually ran and got the result you expected — don't assume.
6. Close scoped, never broad. Close only surfaces you just created or explicitly identified. Never loop a close over the whole tree — you'll kill things you didn't mean to. close/close-window may no-op while a live agent occupies a pane; use close-surface per pane.
7. One window per team. Keep a unit of work to a single window so it stays monitorable and tearable as a unit.
8. Never print secrets. If a surface has credentials/keys loaded, read results back without echoing the secret values.
9. Prefer push over poll. Use cmux events --category notification (see Wait for agents via notification events above) to know the instant an agent finishes instead of polling read-screen in a tight loop. Install hooks first (cmux hooks setup); match events on workspace_id. cmux wait-for is a manual named-token semaphore, not an agent-finished signal.

Workflows

Drive a surface end-to-end

The default loop for any single-surface task: discover, inspect, act, verify, report.

1. cmux --help (and per-subcommand --help) to confirm the verbs.
2. Inspect current state (tree --all / workspace list).
3. Take the action (create / send + send-key / read).
4. Read back to verify the result.
5. Report concisely what happened, citing the surfaces/refs involved.

Report Format

Report concisely in plain English: what you did, the surfaces/refs involved, and what read-screen confirmed. Cite refs (e.g. workspace:2 / surface:3) and never echo secret values.