Skip to main content
The mode picker at the bottom of the input box controls what the model can do on its own and which actions must ask you first. Claude Code provides all four modes. Codex and the other engines do not offer plan mode.

​
Four modes

Permission confirmation dialog
Auto mode skips file and command confirmation. Use it only when you accept the model changing any project file and running commands.

​
How to answer a permission dialog

Default mode, and some agent-mode actions, open a permission dialog: Requests from an external process are labeled separately. The dialog has a countdown. After timeout it closes and the model continues with the current policy. Set the timeout seconds in Settings → General → Behavior. When the model asks you a multiple-choice question, it uses an Ask user dialog. Turn on system notifications in Settings so you do not miss it after switching windows.

​
Plan mode

Plan mode only grants read-only tools. The model analyzes the code, proposes a plan, then waits for your approval. Typical flow:
  1. Switch to Plan mode
  2. Describe the change
  3. Review the plan and ask for a revision if needed
  4. Approve or reject it in the approval box
  5. After approval, let it execute
Plan approval Codex, Grok, Kimi, OpenCode, and Pi do not list plan mode. Use default / agent / auto on those runtimes.

​
Codex sandbox

Codex has a separate sandbox policy in Settings → Permissions. It is a second layer on top of chat mode:
  • Mode: whether to ask you
  • Sandbox: which directories and network the command can actually touch
You can combine them. For example, agent mode auto-accepts edits while a stricter sandbox limits commands.