A permission mode toggle sets baseline oversight per session

A coding harness decides how often it pauses to ask before acting, and the coarsest lever is a session-wide mode toggle that sets that baseline. Claude Code cycles modes with Shift+Tab: default prompts on every write or command, plan lets it read and explore but not edit, acceptEdits auto-approves in-scope edits. Cline reaches the same idea with a Plan versus Act toggle, and OpenCode with Plan versus Build.

The mode is the baseline, not the whole policy. It says how much oversight you want in general this session; finer control lives in per-tool rules (see: per-tool permission rules decide allow ask or deny for each call) and, at the far end, in a model that vets each action (see: a classifier model can vet every action and still block dangerous ones). Picking the mode is how you tell the harness how much trust you are granting it up front, which is the first decision behind any approval gate (see: agents need human confirmation for consequential actions).