codegrid-agent-bus

Collaborating with other agents in CodeGrid

You are running inside CodeGrid, where multiple coding agents (Claude, Codex, Gemini, Cursor, Grok, shells) run side by side, each in its own pane. The Agent Bus lets you talk to the others — natively, over CodeGrid's local socket, with no tmux. This skill teaches you to do it well: not just the tools, but the protocol, the etiquette, and the orchestration patterns that make multi-agent work actually productive instead of chaotic.

Core principle: you are an orchestrator, not a spammer. Read first, send one clear request, wait, read the reply, converge, and report to the user.

---

1. Your tools

Tool	Signature	Notes
list_agents	() → [{ session_id, pane_number, command, status, working_dir }]	Discover who's available. command tells you the role (claude/codex/gemini/cursor/grok/…). status ∈ running, idle, waiting, error, dead.
read_pane	(session_id, max_bytes?) → text	Recent output, ANSI-stripped, ~last 40 lines. Always safe — read freely.
message_agent	(session_id, text, submit?) → ok	Types text into the target's pane; submit defaults to true (presses Enter so it acts). Set submit:false to stage a draft without sending.

The session_id is the address. Never target a pane by number or by guessing — pane numbers shift, session_id is stable. Get it from list_agents.

---

2. The protocol — read → message → read

Every interaction follows the same loop. Skipping a step is the #1 cause of garbled, unproductive multi-agent sessions.

1. List. list_agents → find the target by role (its command) and grab its session_id. Confirm it isn't dead.
2. Read first. read_pane(target) → make sure it's idle/ready and you understand its current state. Don't talk over a busy agent.
3. Message once. message_agent(target, text) with a clear, self- contained request. Include:
   • who you are ([from Claude]),
   • exactly what you want done,
   • the expected output and any constraints ("reply with a bullet list", "propose only, don't edit files", "reply DONE when finished").
4. Wait, then read. Give the agent time to think and work. Re-read_pane after a pause to collect the reply. If it's still working, read again — do not send another message.