etanhey/cmux

cmux Terminal Control

Teach agents to drive cmux: split panes, notify, open browser, coordinate with other agents.

For agent workflows, prefer /cmux-agents and the agent-based MCP tools (spawn_agent, send_to_agent, wait_for, list_agents, my_agents, get_agent_state, stop_agent). This skill stays primitive-first on purpose: panes, tabs, browser surfaces, raw terminal delivery, and low-level inspection.

Detect cmux

Always check first — skip gracefully if not in cmux:

if ! command -v cmux &>/dev/null || ! cmux identify --json &>/dev/null; then
  echo "Not in cmux — skipping cmux operations"
  exit 0
fi

Self-Identification

# Get your own location
cmux identify --json
# → { "surface_ref": "surface:4", "pane_ref": "pane:5", "workspace_ref": "workspace:2" }

Identify & Navigate

# Get your current surface/pane/workspace refs
cmux identify

# List all workspaces
cmux list-workspaces

# List panes in current workspace
cmux list-panes

# List surfaces (tabs) in a pane
cmux list-pane-surfaces --pane pane:1

Splits (Most Common)

# Split current surface left/right/up/down
cmux new-split right    # adds a right split in current workspace
cmux new-split down

# Get the new surface ref after splitting
cmux list-pane-surfaces --pane pane:1

# Close a surface
cmux close-surface --surface surface:5

Workspaces (New Tabs)

# New workspace (new sidebar tab)
cmux new-workspace

# Focus a workspace
cmux focus-pane --pane pane:2

# Rename the tab label
cmux rename-tab --surface surface:3 "🤖 Agent Name"

Send Commands to Panes

# Send text (as if typed) to a surface
cmux send --surface surface:3 "cd ~/Gits/golems && claude 'prompt here'\n"

# Send a single key
cmux send-key --surface surface:3 Return

# Note: \n at end of send = auto-press Enter to run

Compound commands can split across PTY buffer writes. Separate source and launch:

# SAFE: separate commands
cmux send --surface surface:N "source ~/.zshrc\n"
sleep 3
cmux send --surface surface:N "cd ~/Gits/repo && claude -s\n"

# UNSAFE: long compound — may corrupt shell parsing
# cmux send --surface surface:N "source ~/.zshrc && cd ~/Gits/repo && claude -s\n"

Notifications

# Notification ring (tab lights up in sidebar)
cmux notify "Title" "Body text"

# Workspace-action for custom sidebar label
cmux workspace-action --action set-title --title "🤖 orcClaude"

In-App Browser

# Open browser pane (splits alongside terminal)
cmux new-pane --type browser --direction right --url "http://localhost:3000"

# Navigate existing browser surface
cmux browser navigate "http://localhost:3000/new-page"

# Browser tab management
cmux browser tab new "https://docs.example.com"
cmux browser tab list

Agent-to-Agent Messaging

# Send text to another agent's surface
cmux send --surface surface:6 "STATUS: done with auth module, ready for review"

# Envelope format for deterministic messaging:
# [FROM=surface:A TO=surface:B TYPE=TYPE] key=val
cmux send --surface surface:6 \
  "[FROM=surface:1 TO=surface:6 TYPE=TASK] repo=golems task=fix-tests"

Parallel Agent Fan-out Pattern

# 1. Identify self
MY_SURFACE=$(cmux identify | jq -r '.caller.surface_ref')
MY_WS=$(cmux identify | jq -r '.caller.workspace_ref')

# 2. Spawn 3 splits, each running claude on different repos
for repo in brainlayer golems voicelayer; do
  NEW_SURFACE=$(cmux new-split right --workspace "$MY_WS" | awk '{print $2}')
  cmux rename-tab --surface "$NEW_SURFACE" "🤖 $repo"
  sleep 0.5
  cmux send --surface "$NEW_SURFACE" "cd ~/Gits/$repo && claude 'your task here'\n"
done

golem-terminal Integration Note

Future reference only — golem-terminal is not yet built.

These cmux patterns are the reference implementation for golem-terminal's UDS API. The golem-terminal equivalents:

| cmux | golem-terminal | |------|---------------| | cmux split | orchestrate.py split <slot> | | cmux notify | HTTP POST localhost:3847/notify | | cmux sidebar set | UDS status command | | cmux send | UDS send_input command | | cmux open-browser | Built into sidebar pane |

Verify Delivery (CRITICAL — from session mining)

send_input returns ok:true even on frozen terminals. Never trust it blindly.

Boot verification (before sending task prompt)

# Claude takes 8-15s to boot + MCP init
sleep 15
cmux read-screen --surface surface:N --lines 5
# "❯" with "0 tokens" → Claude ready, send prompt
# "zsh%" or bare shell prompt → Claude didn't start, retry launch
# Blank screen → wait longer, re-read in 5s

Post-send verification (after sending task prompt)

sleep 8
cmux read-screen --surface surface:N --lines 5
# Check: did token count jump? Is there new output?
# If token count SAME after 2 checks → terminal frozen → kill → new_split → resend

Stuck-state detection cheat sheet:

• "Press up to edit queued messages" → STUCK — send Enter key to unblock
• "Twisting/Channelling + timer" → THINKING (verify with token count delta)
• Token count same between two reads → STUCK or DONE (investigate)
• "❯" prompt with 0 tokens → boot not complete, wait longer
• "zsh%" or bare shell prompt → agent didn't boot, retry launch command
• "tools not available" or "MCP connection failed" → MCP server down, don't proceed
• High token count but no tool calls in last 50 lines → agent may be rambling, not executing

read_screen Depth

Default for agent monitoring: lines: 50, scrollback: true — NOT 15.

Bottom 15 lines = status bars and thinking indicators. Actual work (file edits, tool calls, decisions) is ABOVE the fold. Use 50+ lines to see what's really happening.

Exception: Post-send verification (checking token count jumped) → 5 lines is fine.

Rules

1. Always detect cmux first — don't assume you're in cmux
2. Use envelope format for agent messages — prevents cross-pane confusion
3. Set sidebar status at task start + end — gives user visibility
4. Notify on completion — ring is better than polling
5. Don't spawn too many panes — 4-6 max, cmux gets crowded
6. Verify delivery after every send — ok:true is optimistic, not verified
7. read_screen 50+ lines for monitoring — 15 lines sees only status bars
8. Respawn > absorb — frozen agent → read_screen 50 lines (salvage what it did) → brain_store accomplishments → kill → new_split → resend SAME task with "already done: X" context. NEVER pull agent work into your own context.
9. Don't navigate interactive menus via send_key — menu state is invisible in read_screen. Instead send the INSTRUCTION to the agent: "Run /mcp, reconnect brainlayer, verify with brain_search('test')."
10. Report evidence, not vibes — cite WHAT the agent produced (file edits, tool calls, test counts), not just THAT it appears active. "Token count +5K, edited 3 files" > "Making progress."

See Also

• /cmux-agents — visible worker spawning, spawn_agent / send_to_agent / wait_for lifecycle, and recovery protocols
• /orc — orchestration decisions, design iteration gates, collab protocols