During stress-testing of the latest build, I found that the AI could be “convinced” to perform file writes even while in Plan Mode. To ensure strict adherence to mode boundaries, I propose implementing directory-level read/write permissions tied directly to the active mode:
-
Plan Mode: Write Access: Restricted exclusively to
.cursor/*.plan.md.- Read Access: Global (or project-wide), but with all other file paths mounted as read-only.
-
Act Mode: Full Read/Write access as currently implemented.
Stripping write permissions at the system/plugin level (rather than relying on model instructions) would prevent accidental or “persuaded” edits during the planning phase.