Chats and agents

Each chat keeps its own settings in Session Setup, the panel on the right: its title, language, mode, main model and subagents.

Modes

Mode	For
General	One model answers.
Code	The main model plans and writes; subagents can help. Files and commands go through approvals.
Hypothesis	A debate: a supporter and an attacker argue, a judge weighs the evidence.

Local and cloud models can be mixed in one chat, for example a local supporter with a cloud judge.

Subagents

Add subagents in Session Setup, each with its own model and access. Call one with @Name in a message, or ask the chat to spawn a subagent. Up to four work at once; their results appear as expandable cards, and the main model writes the final answer.

Access and approvals

The access icon beside the model in the composer sets what the chat’s agents may do. It applies to the chat and every agent in it.

Access	Reads in the workspace	Writes in the workspace	Delete, commands, outside paths
Ask for approval	allowed	asks	asks
Approve for me	allowed	allowed	asks
Full access	allowed	allowed	allowed

An approval card shows the exact paths and contents, or the command and its arguments. Approve runs that proposal once; Cancel skips it. Text in a prompt or an attachment can never grant access, and writes through plugins always ask, even with full access.

Note

Access modes are the app’s policy, not a sandbox: approved commands run with your user’s permissions.

Projects

A project is a folder on your disk with its own chats. Add one under Projects in the sidebar; the app never moves or deletes the folder. Agents in a project chat read, search, edit and run checks inside it.

Reviewing edited files

When an agent edits a file, Review opens it in the right panel. Select text, add a comment and send it back to the chat; the agent changes only that part.

Images

Attach images to a chat with a vision model (one with a vision adapter). Models without one say so instead of guessing.