Execution modes
Decide what the agent can do, and where it should stop for your input.
Four levels of control
| Mode | File edits | Shell and delegation | Best for |
|---|---|---|---|
plan | Read-only | Restricted by read-only policy | Exploration and proposing changes. |
manual | Approval required | Approval required | Reviewing each change and command. |
auto-edit | Allowed automatically | Approval required | Moving quickly through code changes. |
auto | Allowed automatically | Allowed automatically | Tasks with clear scope and trusted execution. |
Switch modes during a session
The UI and API can change the mode while a session is active. The agent also has a set_mode tool. The mode is a policy layer over the tool-approval gate, rather than a different agent implementation.
To set it in a streaming API request, include a mode field alongside the messages.
{
"messages": [{ "role": "user", "content": "Explain this repository" }],
"mode": "plan"
}Planning and clarification
A plan can be offered for review before implementation. Vibes can also call ask_user with a structured questionnaire. That halts the run until your response arrives, keeping decisions in the conversation instead of guessing.
Understand the execution boundary
Structured file operations are contained within the workspace by the sandbox. Shell commands run real host binaries. A working directory is not an operating-system jail, so choose an execution mode appropriate to the repository and task.