agent group drives that same agent from your terminal: open a conversation, send a message, then read or stream the reply.
The agent already knows your workspace, so you don’t hand it any context. You talk to it in a conversation, the thread that holds a title and its messages. Each message kicks off a turn, one run of the agent for that message. A turn does real work, reading your runs, issues, and signals before it answers, so it runs in the background: you send the message, get a turn_id back right away, then read or stream that turn as it finishes. Each entry in the thread is a message, yours or the agent’s reply. Sometimes a turn proposes a confirmation, an action the agent wants to take that needs your approved or rejected sign-off.
Commands
Talk to the agent
Start a conversation.title is optional; leave it off and the agent names it:
message is required; pass the conversation_id to keep the thread going. Skip the first step entirely by calling enqueue-chat with only a message, and the agent starts a fresh conversation and returns both ids. The response carries the turn_id:
--after-seq:
Review the thread
List your conversations, newest activity first:Confirm an action
When the agent proposes an action that needs sign-off, answer with a decision.conversation_id and action are required; decision is approved or rejected:
See also
Issues
The tracked problems the agent investigates.
Signals
The detectors behind those problems.
Overview
What the CLI is and how the groups fit together.