# Modes and permissions Source: https://docs.ideharmony.com/tui/modes Mode is the authority setting for a session: how much an agent may do before it has to stop and ask. ```text /mode safe /mode ask /mode yolo /permissions ``` ## The three modes | Mode | Consequential actions | Use when | |---|---|---| | `safe` | Gated | You are working in a repository you care about | | `ask` | Prompts before acting | Default working posture for most people | | `yolo` | Minimal friction | Throwaway sandboxes, or work you will discard | > [!WARNING] > `yolo` is not "faster mode". It is "fewer chances to stop it" mode. The speed > you gain is the speed of not being asked, which is the same thing as not > getting to say no. ## Mode is per session Changing mode affects the current session. Two sessions in the same workspace can run at different authority levels, which is useful: a read-only exploration session and a gated implementation session can coexist. Check what you are actually in before assuming: ```text /permissions ``` ## What an approval prompt is for The prompt is the last cheap moment to stop something. Everything reaching an agent is untrusted input — model output, repository contents, browser pages, provider responses, messages from peers. > [!SECURITY] > A repository can contain text written specifically to steer an agent. Reading > a file is not neutral. This is the reason gates exist at the action boundary > rather than the input boundary. ## Reading a prompt properly 1. **What is the exact action?** A path, a command, a network destination. 2. **Does it match what I asked for?** Scope creep appears here first. 3. **Is it reversible?** If you find yourself clearing prompts without reading, that is a signal to change mode, not to keep approving. ## Desktop YOLO defaults on The desktop's YOLO setting defaults to **on** and injects provider full-access flags at launch. It is separate from the per-session mode here, and separate again from Crew Mode ([Experimental](/docs/reference/feature-status)), which is about delegation scale rather than authority. Three different switches. Turning one on does not turn the others on. See [security permissions](/docs/security/permissions). ## Extension permission ceilings An extension cannot exceed its permission ceiling, which is how a read-only skill does not quietly acquire shell access. > [!IMPORTANT] > A ceiling is not a sandbox. Extension code is not sandboxed by default — it > runs with the privileges of the process that loaded it. ## Related - [Security permissions](/docs/security/permissions) - [Permissions and approvals](/docs/concepts/permissions-and-approvals) - [The composer](/docs/tui/composer)