Menu
Orchestrate multiple Pi agents via tmux windows. Use when spawning background agents, sending them tasks, and capturing their output.
Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.
Based on SOC occupation classification
Bosun configuration — models, sandbox, daemon, Pi settings. Use when changing models, editing config, or understanding how bosun is set up.
Bootstrap a new project using bosun as a foundation — via bun dependency (recommended) or git submodule. Creates a downstream project that inherits bosun's multi-agent infrastructure while adding custom agents, skills, and extensions. Works standalone — fetchable by any pi agent via raw GitHub URL.
Search curated markdown memory like sessions, plans, docs, and skills. Use when recalling prior context or looking for relevant historical/project knowledge.
Analyze Pi session JSONL files using jq patterns. Use when extracting metrics, tool usage, costs, or reviewing session history. Load for session export, summarization, or workflow analysis.
Session-based browser plan review primitives for Bosun. Provides a local-first state model, markdown-aware anchor extraction, and review-session persistence for reviewing existing plan files in dedicated CDP browser windows. Triggers: "plan-review", "review this plan", "review markdown plan".
Session-based browser diff review primitives for Bosun. Provides transport- agnostic v1 schemas plus local state and git snapshot helpers for immutable review rounds. Triggers: "diff-review", "reround", "review round", "snapshot-backed review".
| name | tmux-orchestration |
| description | Orchestrate multiple Pi agents via tmux windows. Use when spawning background agents, sending them tasks, and capturing their output. |
| license | MIT |
| compatibility | pi |
| metadata | {"audience":"developers","category":"orchestration"} |
Spawn, control, and monitor multiple Pi agents running in parallel tmux windows.
Use this skill when:
Do NOT use for:
Must be running inside tmux. Start with just start or tmux.
| Tool | Purpose |
|---|---|
spawn_agent | Start agent in background window |
send_keys | Send text/commands to a window |
capture_pane | Read output from a window (non-mesh agents only) |
list_windows | List all windows |
Most agents have mesh tools and will report back automatically. Do NOT poll them. The mesh is for coordination, not conversation — ask for concise, substantive reports rather than chatty progress pings.
spawn_agent({
agent: "verify",
task: "Run tests and send one concise mesh_send report to bosun with pass/fail summary and any blockers"
})
// Spawns in background, returns immediately
"I've spawned verify to run the tests. I'll process their report when it arrives."
The agent's mesh_send message arrives automatically as a follow-up. NEVER use capture_pane to check on mesh agents. If the user has other work, handle that instead. Do not send acknowledgment-only replies back over mesh unless you need the agent to change course.
When the mesh message arrives, summarize findings and take action.
Don't kill agent windows when they finish. The user may want to inspect output or interact further.
Only use capture_pane for agents that do NOT have mesh tools (e.g., Q) or interactive programs.
spawn_agent({ agent: "q", task: "Plan the refactor" })
list_windows({})
// Shows: 1: bosun-1 (active), 2: q
capture_pane({
window: "q",
lines: 50
})
send_keys({
window: "q",
text: "What tasks are blocked?"
})
Tasks with special characters (quotes, apostrophes) are automatically escaped:
spawn_agent({
agent: "lite",
task: "What's the user's name?"
})
-d flag keeps focus on current windowmesh_send results when done — just waitcapture_pane to check on agents that have mesh toolsgot it, thanks, or emoji-only mesh replieslist_windows clearer