The mux skill

The mux skill is the one skill in the cyber-mux agent plugin. It passes what you type to the cyber-mux CLI, runs it, and reports the result. The skill adds nothing of its own: every command and flag it uses comes from the CLI’s --help.

Invoke the skill

In Claude Code, plugin skills are namespaced by plugin, so the skill is /cyber-mux:mux:

/cyber-mux:mux <command> [args] [options]

Cursor, Codex, and GitHub Copilot CLI load the same skill under the name mux. Call it the way that runtime calls skills, or ask in plain words. The agent loads the skill when your request is about opening, sending to, reading, waiting on, focusing, listing, or closing a pane, detecting the multiplexer, nudging a peer pane, or opening a git worktree in a new pane.

The skill needs Node.js 22 or later and a POSIX shell.

Argument forms

The skill reads its arguments as one of four forms:

You pass	The skill runs
nothing, --help, or -h	cyber-mux --help
<command> --help or <command> -h	cyber-mux <command> --help
<command> [args] [options]	cyber-mux <command> [args] [options], verbatim and in your quoting
a request in plain words	cyber-mux --help, then cyber-mux <command> --help, then the call built from the flags that help lists

With no arguments or --help, the CLI help is the only command the skill runs.

Commands

Each row lists the command’s key arguments and flags as its --help prints them. Most commands also accept --format. A <pane> argument takes a pane id or a label; see Pane.

Command	What it does	Key arguments and flags
doctor	Probe the multiplexer, self pane, and backend; print the fast-path pins	—
mode	Report the detected drivable backend	—
open	Open a new pane, tab, or workspace, optionally running a command in it	--at, --launch, --template, --cwd, --env, --label
send	Drive a pane without taking its turn: send text types text with no Enter, send keys presses named keys	text <pane> <text>, keys <pane> <keys...>
submit	Take a pane’s turn: type the text if given, then press Enter	<pane> [text]
read	Capture a pane’s output	<pane>, --lines or --full
wait	Block until a pane’s output matches or the timeout elapses	<pane>, --match or --regex, --timeout, --lines
focus	Move the attached client to a pane	<pane>
close	Close a pane	<pane>
list	List every live pane the backend can see	—
exists	Check whether a pane is still live (exit 0 live, 1 gone)	<pane>
worktree	Git worktree helpers: add, provision, open, list, remove, prune	add --branch [--at], open <path>, remove <path> [--force], prune [--force]
template	Manage named templates: list, show, validate, save, edit	show <name>, save <name> [--from], edit <name> [--set]
agent	Inspect and wait on a pane’s agent-lifecycle state: status, wait	status <pane>, wait <pane> [--until] [--timeout]

The CLI Reference documents every flag.

How the skill finds the CLI

You do not need to install the CLI first. The skill resolves the command once per invocation and uses it for every call, taking the first that works:

1. cyber-mux on your PATH.
2. The repository’s package manager, when its lockfile is present and the package is installed: pnpm exec cyber-mux, yarn exec cyber-mux, or bunx cyber-mux.
3. npx --yes cyber-mux@<version>, pinned to the plugin’s version.

Output

• Help is shown as the CLI prints it.
• Success is summarized in a line or two, naming any pane id the command returned. When the command’s help lists --format and you gave none, the skill adds --format agent, the compact AXI output for agents.
• Failure is quoted verbatim, with the exit code. The skill does not retry with different flags unless the error names the fix.

The skill only runs cyber-mux. It never calls tmux, herdr, or another multiplexer’s own CLI. Inside GNU Screen, commands that drive a pane fail with a named screen error: Screen is detected but cannot be driven, so the skill reports the error rather than trying to fix it. See Multiplexers.

Examples

Detect the multiplexer

/cyber-mux:mux mode

The skill runs cyber-mux mode --format agent and reports the backend, such as tmux or herdr.

Open a pane and read it

Open a pane to the right that runs the test watcher, then read its last 20 lines:

/cyber-mux:mux open --at pane:right --label tests --launch "pnpm test --watch"
/cyber-mux:mux read tests --lines 20

Send to a peer pane

Type a message into a pane and press Enter, then press C-c in another pane to stop what it runs:

/cyber-mux:mux submit reviewer "the tests pass, please review"
/cyber-mux:mux send keys tests C-c

submit takes the pane’s turn. send text and send keys drive the pane without pressing Enter for it.

Open a worktree in a new pane

Create a worktree on a new branch and open it in a workspace that runs your agent:

/cyber-mux:mux worktree add --branch feat/login --at workspace --launch claude

Ask in plain words

open a pane below and tail the server log

The agent loads the skill, reads cyber-mux --help and cyber-mux open --help, and runs cyber-mux open --at pane:down --launch "tail -f server.log" --format agent.