cmux-delegate — Delegate Task to Another Agent

Launch an AI agent in a separate cmux pane or workspace, send it a task, monitor completion, and collect results.

Language Rules

1. Auto-detect input language → output in the same language
2. Japanese input → Japanese output, use references/delegate-guide.ja.md
3. English input → English output, use references/delegate-guide.md
4. Explicit override takes priority

Prerequisites

BEFORE delegating, verify cmux environment:

echo $CMUX_SOCKET_PATH

• If empty or unset → stop with error:
  • EN: "Error: Not running inside a cmux session. Please start Claude Code from within cmux."
  • JA: "エラー: cmux セッション内で実行されていません。cmux 内で Claude Code を起動してください。"
• If set → proceed

Verify target agent is available (if not default):

which codex    # or whichever agent was specified

Agent Profiles

Select the agent based on user instruction. Default is Claude Code with auto-approve.

For agent details and custom agent definitions, see references/agent-profiles.md.

Execution Steps

Step 1: Parse User Request

Extract from the user's instruction:

• Task: What to do (the prompt to send to the agent)
• Agent: Which agent to use (default: Claude Code)
• Mode: Auto-approve (default) or interactive. Use interactive only if user explicitly requests it (e.g., "interactive", "対話モード", "承認あり")
• Direction: Where to place (default: new workspace)
• Directory: Target directory if cross-directory (optional)

Step 2: Create Surface

Default is pane split right. cmux new-workspace may create a non-terminal surface that cannot receive commands — use cmux new-split instead.

# Default: split right (recommended)
cmux new-split right
# Output: OK surface:{N} workspace:{N}

# Alternative: split down
cmux new-split down

Extract the surface handle from the output.

Step 3: Change Directory (if cross-directory)

If the user specified a different directory:

cmux send --surface surface:{N} "cd /path/to/target/repo\n"
sleep 1

Step 4: Launch Agent

Select the launch command from the Agent Profiles table based on the agent and mode determined in Step 1. Default is auto-approve mode.

Auto-approve (default):

# Claude Code
cmux send --surface surface:{N} "claude --dangerously-skip-permissions\n"

# Codex
cmux send --surface surface:{N} "codex --dangerously-bypass-approvals-and-sandbox\n"

# Gemini CLI (no auto-approve available — always interactive)
cmux send --surface surface:{N} "gemini\n"

Interactive mode (if user explicitly requested):

cmux send --surface surface:{N} "claude\n"
cmux send --surface surface:{N} "codex\n"

Step 5: Detect Agent Prompt (Polling)

Poll until the agent shows its prompt (ready for input):

sleep 3
cmux read-screen --surface surface:{N}

Check output for the agent's prompt pattern. If not detected, retry up to 5 times with 3-second intervals (15 seconds total timeout).

If prompt not detected after retries → report error and stop.

Step 6: Send Task

Send the task prompt to the agent:

cmux send --surface surface:{N} "{task_prompt}"
cmux send-key --surface surface:{N} return

For multi-line prompts, send line by line:

cmux send --surface surface:{N} "line 1"
cmux send-key --surface surface:{N} return
cmux send --surface surface:{N} "line 2"
cmux send-key --surface surface:{N} return

After sending, verify the agent accepted the task by reading the screen and confirming the prompt disappeared.

Step 7: Monitor Completion (Polling)

Poll the surface to detect task completion. Use graduated intervals:

Completion signals:

• Agent prompt reappears → task completed
• Error message pattern detected → task failed

Timeout: No hard timeout by default. Report progress to user periodically (every 60 seconds).

cmux read-screen --surface surface:{N}

Step 8: Collect Results

Once completion is detected, collect the full output:

cmux read-screen --surface surface:{N} --scrollback 500

Analyze the output:

• Extract the conclusion/summary
• Determine success or failure
• Report results to the user

Step 9: Cleanup (Optional)

Ask the user or auto-cleanup based on context:

# Close the workspace
cmux close-workspace --workspace workspace:{N}

# Or close just the surface
cmux close-surface --surface surface:{N}

Error Handling

Notes

• This skill launches new agent sessions, not forked conversations. For conversation forking, use cmux-fork.
• The parent agent (running this skill) monitors and collects results. The user can also observe the child agent in real-time via the cmux UI.
• For detailed agent profiles, custom agent definitions, and advanced patterns, see the reference guides.