Entering & exiting

Agent conversations are designed to be cheap to enter and cheap to leave. You can drop into one mid-command, pop back to the shell, and pick the thread up later — nothing is lost in between.

Starting a conversation

There are two front doors:

• ctrl+enter (cmd+enter on macOS) from the terminal prompt. Whatever you've already typed comes along as the first message — so you can start writing a shell command, realize it's really a question, and press ctrl+enter instead of enter.
• /agent from the slash menu: type / at the prompt, pick /agent, and the conversation opens on its landing view.

Either way the pane switches into the conversation view, and the tab title becomes the conversation title once one is set — so a sidebar full of tabs reads like a list of what each agent is working on.

The slash menu also offers /orchestrate-agent and /orchestrate-cli for multi-agent runs (Orchestration) and /plan to ask for research and a plan before any code is written.

Leaving with esc

Press esc (when the conversation input isn't focused — press esc once to unfocus it first if needed) and the conversation collapses out of the way:

• The pane returns to the normal terminal.
• The conversation shrinks to a mini block in your scrollback — a compact row with a checkmark, the conversation title, and a chevron — sitting just above the prompt.
• The conversation is saved and appears in Recent Activity on the agent landing page and in the Tools panel's conversations list.

Clicking the mini block reopens the conversation with its full transcript and model selection intact. You can accumulate several mini blocks in one pane and hop between conversations and shell work freely.

Resuming

• ctrl+shift+Y (cmd+shift+Y on macOS) reopens the last conversation in the pane — the prompt even hints this after you collapse one.
• /conversations opens a searchable picker over your whole conversation history. See Conversation management.
• The agent landing page lists Recent Activity — click any entry to resume it.

Resuming restores both the visible transcript and the underlying context, so the agent remembers what it was doing.

Split panes are independent

Each pane hosts its own agent. Split with ctrl+D or ctrl+shift+D (Split panes) and you can run two conversations side by side; esc and ctrl+shift+Y act only on the active pane.