Conversations & composer controls
Choose a model, attach context, read tool activity, and continue or fork a conversation.
On this page
Create or reopen a chat
Choose a repository or conversation space and use its + button to create a new AI chat. Click an existing chat in the sidebar to load its saved conversation. You can read a teammate’s chat when you have access; only the current driver can continue that same chat.
The centre panel distinguishes human prompts from AI responses. Tool activity and message totals come from the conversation data. Expand a response’s steps to inspect available commands, tools, and output. A recorded tool call is evidence of an attempted action, not proof that the final result is correct.
Use the composer controls
| Control | What it does |
|---|---|
| Provider | Chooses the installed AI tool that will execute your request |
| Model | Searches available models reported by that tool |
| Reasoning effort | Selects one of the selected model’s supported effort levels |
| Fast | Selects low reasoning effort when supported; unavailable otherwise |
| Plan | Adds an instruction asking for a plan before changes; it is not a separate security sandbox |
| Attach files | Reads supported text/code files into the draft prompt |
| Send | Sends the current draft using the selected setup |
The pickers open above the composer. Use search where available, arrow keys to choose an option, Enter to select, and Escape to close. Model names and effort choices can differ by account, tool version, and connected computer.
Use Enter to send and Shift+Enter for a newline. Starter prompts only fill or guide the draft; choosing a starter does not run an agent.
Attach files and reference conversations
The attachment control accepts UTF-8 text/code files up to 100 KB per file. Their contents are inserted into the prompt draft. It is not an image upload, repository import, or a way to attach unlimited binary files. Review the draft before sending.
Type @ to select an available room/chat reference from the picker. Pick an actual result so the app can resolve its context; simply typing another person’s name is not the same as selecting a conversation reference. Where shown, Import context from… starts a new chat using another accessible conversation’s context.
Context is limited by access and prompt budgets. The system cannot give your AI unrestricted access to every private room.
Fork instead of taking control
When watching a teammate’s chat, fork a new chat creates your own conversation with imported context. It leaves the original chat and driver intact.
A fork is not automatically a new repository branch: chats in the same room repository share that room’s worktree. For independent code work, use your own room checkout of a shared repository. Use Handoff when you want to continue the original conversation itself.
Rename, archive, or reference a chat
Hover a chat row to reveal its available actions. Rename changes its title. Archive this chat moves an eligible chat out of the active list; expand Archived and use Restore this chat to bring it back. Management actions depend on who drives the chat and its current state.
Copy chat path copies a room/chat reference for use in another conversation. It is a context reference, not a guest invitation and not a way to grant access. Use Share chat when someone outside the room needs to read it.
While a task is running
Follow the streamed response and expand available activity. Queue message lets the driver add instructions while work is active. Detach, when shown, detaches the view; do not assume that closing the view stops the underlying task.
Read queued instructions for provider-specific timing, and history and checkpoints for recovery after a disconnect.