Plan mode

plan mode is for “figure out what to do, show me, then do it” - the model explores read-only, presents a plan, and only after you approve it can it write or run anything.

Enter it with permissions.mode: "plan", or press Shift+Tab mid-session to cycle into it.

What runs in plan mode

• Allowed automatically: read_file, grep, glob, web_fetch, web_search - reads and network research, no local side effects.
• Hard-denied: write_file, edit_file, bash, task. The model is told these will fail, so it doesn’t waste a turn trying.

The approval step

When the model is done exploring it calls exit_plan_mode with a plan (in markdown). You get a dedicated approve / reject prompt:

• Approve → the session drops into manual mode for the rest of the conversation, and write/execute tools become available.
• Reject → the turn stops. Reply with what to change.
• Comment → stop and send free-text redirection as your next message.

Every proposed plan is saved to ~/.polyglot/plans/ regardless of the outcome - see Data handling.

The model cannot get approval by asking in prose (“let me know if this looks good”) - the only path is the exit_plan_mode call, and the system prompt tells it so.