Fleet Manager

Manage a fleet of AI coding agents running in tmux panes using the control MCP tools.

The Loop

digraph fleet {
  "list_panes" -> "Filter to claude-code only";
  "Filter to claude-code only" -> "get_pane_conversation for each";
  "get_pane_conversation for each" -> "Classify each pane";
  "Classify each pane" -> "Present digest to human";
  "Present digest to human" -> "Human picks action";
  "Human picks action" -> "Execute (send_to_pane, stop_pane, or skip)";
  "Execute (send_to_pane, stop_pane, or skip)" -> "list_panes" [label="next cycle"];
}

Step 1: Gather

Call list_panes with no filters. Then for each claude-code pane, call get_pane_conversation (count: 10) to get recent context.

Step 2: Classify

Read each conversation and classify:

Status	Signal	What it means
Working	Recent tool calls, thinking, actively producing output	Leave it alone
Needs input	Asked a question, waiting for user response, permission prompt	Human needs to respond
Stuck	Repeated errors, retrying same thing, no progress in last few messages	May need redirection
Idle	At the prompt, no recent activity, conversation ended	Available for new work
Done	Completed its task, committed, reported back	Can be stopped or given new work

Use your judgment. Read the conversation content — don't just look at the status line.

Step 3: Present Digest

Format as a brief table:

Fleet Status: 5 claude-code panes

  %4  nl          Working    editing graph layout code
  %19 tracker     Needs input  asking about test strategy
  %20 barnstormer Idle       at prompt, finished last task
  %50 hex         Stuck      retrying failed SSH connection
  %79 dot-files   Done       committed linting fixes

1 needs your input, 1 looks stuck. Want me to handle any of these?

Keep descriptions to ~5 words. The human wants a glance, not a novel.

Step 4: Act

When the human says to act on a pane:

• Respond to a question: send_to_pane with the answer
• Redirect a stuck agent: send_to_pane with new instructions ("try a different approach — skip the SSH and test locally instead")
• Give idle agent work: send_to_pane with a new task
• Stop a done agent: stop_pane
• Read more context: get_pane_conversation with higher count (30-50)

When sending to a pane, be direct. The text you send is typed into the pane and Enter is pressed. Write it like you're talking to the agent.

Autonomous Mode

If the human says "handle it" or "you decide":

• Needs input: Only respond if you can confidently answer from context. If unsure, flag it for the human.
• Stuck: Suggest an alternative approach via send_to_pane.
• Idle: Ask the human what to assign. Don't invent work.
• Done: Leave it. Don't stop unless asked.

Never stop a working agent autonomously. Never send instructions you're unsure about.

Quick Commands

User says	What to do
"check on my agents"	Full loop: list, classify, digest
"what needs attention"	Same, but only show needs-input and stuck
"brief me"	Full digest with one-line summaries
"handle the stuck ones"	Read stuck conversations, send redirections
"send X to the tracker agent"	send_to_pane using project name fuzzy match
"stop the idle ones"	stop_pane for each idle pane
"what's %4 doing"	get_pane_conversation for that specific pane

Prerequisites

The control daemon must be running and the MCP server registered. If tools aren't available, tell the user:

The control MCP tools aren't available. Make sure:
1. The daemon is running: control daemon start --background
2. MCP is registered in ~/.claude/settings.json