Running a Session
Start one
Go to Assistant and click New session.
Give it a title, pick the repositories you want the agent to look at and the branch for each, then choose the agent and model it should think with. Claude Code on Opus 4.8 is a good default; Codex is there if you prefer it.
Click Start session.
Your first question can take a few minutes to come back while the agent gets set up. After that, replies are quick.
Ask it things
The conversation sits in the middle, your other sessions on the left, and the spec on the right once there is one.
Replies stream in as they're written. When the agent goes off to look through your code, the steps it takes collapse into a summary you can expand if you're curious — the answer itself is always shown in full.
You can only send one message at a time: while the agent is working, sending is paused until it's finished. You can keep typing meanwhile.
Close the tab whenever you like. Come back and the conversation is still there.
Sessions end on their own
A session closes after 30 minutes without activity, or 4 hours in total — whichever comes first. Turning the spec into a Task also closes it.
When that happens the conversation becomes read-only. Everything stays readable, and any spec it produced can still be turned into a Task.
Sessions can't be reopened, so start a new one if you need to keep going.
Sessions and Tasks draw on the same LLM account, and a session holds its place for as long as it's open. Convert it or leave it once you have what you need.
If something goes wrong
The agent stops mid-conversation. You'll see a message explaining why — usually credits, a rate limit, or an authentication problem with your LLM key. The session ends and you'll need to start a new one once it's sorted.
A message won't send. Your text is kept, so just try again.
Replies stop appearing. The connection recovers on its own and catches up — no need to reload.