Plan Mode and approvals
What the assistant asks before acting, and how to change it.
What the assistant asks before acting
| Action | By default | Can you change it? |
|---|---|---|
| Reading and searching files in the workspace | Allowed | — |
| Editing file contents | Applied directly, with the diff shown in the chat | Undo any task with /undo |
| Deleting a whole file | Asks you, showing what will be removed | Per task: remember your answer for the rest of the task |
| Running a command, or typing into a running program | Asks you, showing the exact command | Per workspace: allow commands without asking |
| Calling a tool from an MCP server | Asks you, showing the server, tool and arguments | Per tool, in mcp.json |
The assistant can only read and write files inside the open workspace folder. Commands are different: they run with your own user permissions, like anything you type in a terminal.
Plan Mode
Click Plan Mode next to the message box before sending. With it on, the assistant investigates, then proposes a plan and waits for your approval before it writes files, runs commands, or changes anything else. Approve it, or reply with changes.
- Use it for anything large, unfamiliar, or hard to undo.
- It stays on for later messages until you click it again, or restart Lamied.
- Approving a plan doesn't approve its commands. Each command still asks, unless you've allowed commands for the workspace.
Approving commands
When the assistant wants to run something, a card in the chat shows the exact command and why. You can:
- allow it, this once;
- decline it. The assistant is told, and continues another way or asks you;
- or tick Don't ask again for commands in this workspace to stop being asked in this project.
Command output streams live into the chat. Long-running servers the assistant starts keep running after its answer, and stop when you close Lamied or ask the assistant to stop them.
Allow commands without asking
Open AI settings (⚙) → Command permissions → Allow commands without asking in this workspace.
- It applies to the open workspace only, and is remembered on this computer until you turn it off.
- It includes typing into programs the assistant runs, such as answering a prompt.
- It doesn't change file-deletion approvals, and never approves MCP tools.
- No AI tool can turn it on. Only you can, in settings.
Project files can't grant permissions
Instructions in a repository's LAMIED.md or AGENTS.md are guidance only. They can't switch off approvals, because Lamied enforces approvals itself, not by asking the model. See Project rules.