Sessions

A session is one unit of work with an agent. Conversations, one-shot tasks, scheduled runs and orchestrations all render in the same session view.

Starting a session

From New Session in the sidebar (or any agent’s “New Session” button):

1. Pick the target: an agent, or a project directory.
2. Type the goal. Use @ to reference agents/projects, / to invoke skills.
3. Optionally adjust runtime, model and thinking level (off / low / medium / high / xhigh) from the composer bar.
4. Attach files if needed, then send.

The draft you’re typing is saved per workspace, so a stray reload doesn’t lose it.

Reading a session

• Tool cards: every tool call renders as a card (posted tweet, search results, DB rows…) with its input and output expandable; routine Write/Edit calls fold into the collapsed group instead of taking a row each.
• Status badge (next to the title): Queued, Running, Waiting for you (needs an approval), Succeeded, Failed, Timed out. Failure badges carry the concrete reason and a Continue action.
• Metadata: turn count, duration, and cost are in the header’s details toggle.
• Summary: finished autonomous runs expose a Summary modal (achievements / blockers / insights).

Context usage

The ring around the composer shows how much of the model’s context window the session is using: green below 70%, amber below 90%, red at 90%+. During compaction it pulses: older turns are being summarized to free space. You can also start a fresh turn with a bigger-context model mid-session; the choice persists for the session.

Steering mid-run

• Model switching: change the model from the composer; it applies to the next turn.
• Interrupt: stop the current turn; the conversation stays intact for follow-ups.
• Ask a question: queued messages are delivered to the agent between turns.

Failure & recovery

Failed sessions don’t lose their history. The failure card shows the reason (for example “Runtime request timed out after 600s”) with a Continue button that resumes the work in a new turn.

Session types at a glance

Copilot conversations, direct tasks and autonomous runs all live in the same list (sidebar → agent). Autonomous and scheduled runs additionally show up under Scheduled tasks with their run history.