agent-bridge

Agent Bridge

Coordinate independent chats while keeping each chat's current turn open. The sender uses ask and resumes when the receiver calls reply; do not replace this with polling or a new user turn.

Shared identifiers

Before using the bridge, determine:

• project_key: a stable shared scope agreed by both chats, such as the repository name or a user-provided collaboration name.
• agent_name: a unique short identity for this chat, such as architect or reviewer.

Names use letters, digits, _, and -. Keep the same identifiers throughout the collaboration.

Register

Call register_agent once at the beginning of the chat's active bridge turn. Include a short description so peers can distinguish roles. If the other peer is not registered yet, report that fact and wait for the user to start it; do not invent another identity.

Receive and remain available

When the user asks this chat to receive work:

1. Register the identity.
2. Call listen with a long timeout, normally 3600 seconds.
3. When it returns a request, preserve request_id exactly.
4. Complete the requested analysis or work using the chat's normal Codex capabilities.
5. Call reply with the result and exact request_id.
6. If the user requested persistent availability, call listen again in the same turn.

An idle result means no request arrived before the timeout. Explain it briefly; do not claim that the peer answered.

Ask and wait synchronously

When the user asks this chat to consult another chat:

1. Register the identity if needed.
2. Optionally call bridge_status to verify the target name.
3. Call ask with the full, self-contained request and a suitable timeout.
4. Remain in the same turn while the tool is pending. Do not poll, schedule a task, or ask the user to copy messages.
5. When ask returns replied, treat response as the peer's answer and continue the task.

If ask returns expired or cancelled, state that no usable peer response arrived. Do not fabricate one.

Safety and lifecycle

• Limit autonomous dialogue to the number of rounds requested by the user; default to one request/reply round when no limit is specified.
• Never send secrets unless the user explicitly placed them in scope for both chats.
• Use distinct project keys for unrelated work.
• Stop listening when the user interrupts the turn.
• Use cancel_request only for a request created by this chat.

Example setup

In the receiver chat:

Use Agent Bridge. Register as reviewer in project checkout-redesign. Wait for architect,
review its proposal, reply with concrete findings, then wait again. Maximum 3 requests.

In the sender chat:

Use Agent Bridge. Register as architect in project checkout-redesign. Ask reviewer to inspect
my plan for correctness and wait for its response before continuing.