Permissions
Permissions control how much autonomy the agent has.Modes (User View)
- Explore: discovery and planning mode with safety-first defaults.
- Ask to Edit: changes are proposed, then confirmed before execution.
- Execute: direct execution mode for implementation work.
How Permissions Fit the Product Flow
- Most work starts in Explore mode.
- Explore gathers context and prepares a plan.
- You approve transition into execution behavior.
- Task work proceeds with appropriate permission level.
Configuration
Permissions are defined at app/workspace/source levels and merged at runtime. Common rule categories include:- Allowed MCP tool patterns.
- Allowed API endpoints.
- Allowed shell command patterns.
- Optional blocked tools and allowed write paths.
Explore Mode Defaults
Explore mode is intentionally restrictive for safety:- Read/search operations are broadly allowed.
- Mutation-heavy operations are blocked by default.
- Plan creation is allowed so you can move from exploration to approved execution.
Source-Level Permission Design
Source permissions should enable meaningful read operations in Explore mode while preventing accidental writes. For MCP sources, source-level patterns are auto-scoped to reduce leakage across sources.UX Requirement: Plain-Language Prompts
When user confirmation is needed, prompts should explain the action in plain language rather than raw command syntax.Detailed Internal Guide
packages/shared/assets/docs/permissions.md