Configure reusable roles
A role is a reusable bundle: prompt, model, thinking level, tool policy, approval mode, skills, MCP, and instruction files. Choose a role when creating a session; the orchestrator also starts workers from a role.
Create, copy, and edit
Section titled “Create, copy, and edit”Under Settings → Roles the left side lists roles with New / Import; the right side is the editor. Click a role to switch; unsaved changes ask before being discarded. Import reads a JSON file and renames roles whose ids collide.
The editor covers name, description, system prompt, model and thinking, tool policy, approval mode, skills, MCP servers, and instruction files. The system prompt is appended after the base prompt and may be empty. Saving applies to newly started sessions; a running session does not reread its role.
Built-in roles are editable. Reset defaults restores prompt, model, tool policy, approval, skills, and MCP; it applies only after you confirm and save. The menu also offers Copy as new role, Export JSON, and Delete. Built-in ids are fixed and cannot be deleted.
The two built-in roles
Section titled “The two built-in roles”| Built-in | Tool policy | Default approval |
|---|---|---|
| Orchestrator | Orchestration | Full |
| Worker | All tools | Auto-accept edits |
Tool policies and ids
Section titled “Tool policies and ids”There are three tool policies:
- All tools: read, write, and run commands. This is the Worker default.
- Read-only: no file edits and no arbitrary commands.
- Orchestration: read files, read-only git, and the dispatch, review, and merge tools. No writes and no arbitrary commands. This is why the Orchestrator still doesn’t edit code when its approval mode is Full.
A role id uses lowercase letters, digits, and hyphens, cannot start or end with a hyphen, and cannot change after save. Dispatch uses the Worker role when none is named.