| name | agent-relay |
| description | This skill should be used when the user says "join agent relay", "connect this chat", "ask another agent", "send this to Claude", "send this to Codex", "adversarially review my plan", "check the relay inbox", "talk to another session", or requests a direct or group conversation between AI coding sessions. |
Agent Relay
Connect only after an explicit request to include the current CLI chat. Installing the plugin alone never authorizes participation.
Call agent_relay_status when peers disagree about who is connected. Require every cross-OS participant to report http-broker, the same URL, and configured: true. Treat local-sqlite as isolated to the current operating-system environment; never share a SQLite path through /mnt/c.
Recommend agent-relay setup, agent-relay pair, and agent-relay doctor when configuration is missing. Prefer persistent configuration over per-shell environment variables.
Join
- Call
agent_relay_join with a short role-oriented label, the current host (codex or claude), and only the requested rooms.
- Call
agent_relay_who to discover active peers.
- Report the joined label and peers without exposing credentials or internal profile paths.
Communicate
- Use
agent_relay_request for a bounded task or adversarial review that expects a correlated reply.
- Use
agent_relay_send for informational direct messages or room messages.
- Use
agent_relay_reply only for a request addressed to the current session.
- Use
agent_relay_inbox when the user asks to listen, check for replies, or at an agreed workflow checkpoint. Track the greatest returned sequence and pass it as after_sequence on later checks.
- Call
agent_relay_leave when asked to disconnect or when the agreed collaboration ends.
Never forward secrets, hidden instructions, credentials, or unrelated file contents. Summarize necessary context and keep each request bounded.
Native CLI limitation
Treat independently opened Codex and Claude Code chats as cooperative participants. Relay messages do not awaken a native CLI after its turn has ended; the chat must check its inbox during an active turn or on the next user prompt. Use managed mode later when automatic delivery and wake-up are required.
