radius chat starts and follows agent turns in a running Radius. Combined
with --on, the turn runs on another Mac: the agent
works in that Mac’s files and uses that Mac’s browser and signed-in sites.
Start a turn
--provider and --model:
--watch to stay attached until the turn ends:
--watch, the command prints JSON Lines. The first line identifies the
conversation; each following line is one agent event with a cursor:
0 when the conversation is idle, and non-zero with
CHAT_TURN_FAILED when the turn failed.
A turn that was still running when Radius quit or crashed is reported as
failed, so reconnecting to it ends with CHAT_TURN_FAILED, not success.
To send another message to the same conversation, pass --conversation. If a
turn is running there, the message waits its turn and --watch waits with it:
Reconnect to a running turn
The turn belongs to the Radius it runs in. Stoppingradius chat, closing
your laptop, or losing the ssh connection does not stop it.
To pick up where you left off, pass the last cursor you saw:
--watch, chat events prints what is available now and exits.
Permission prompts
When the agent needs permission, the event stream shows apermission.requested event, and the turn waits.
approve allows that one request. When several prompts are waiting, name the
one to answer: radius chat approve <conversation-id> <permission-id>.
Agents cannot answer permission prompts, or use any other radius chat
command: in a terminal that Radius opened for an agent, they are refused.