Docs AI Agents Agent Permissions & Confirmation

Agent Permissions & Confirmation

Every destructive agent action — file write, command execution, DB write, request send — is gated by a confirmation flow you control.

The gate

VORTΞXHQ wraps every agent tool call with a gateToolCalls() guard in the main process. Three policies are available globally and per-module:

PolicyBehaviour
ask-before-write (default)Confirm every write/destructive call.
ask-onceConfirm the first call in a run, allow the rest.
autoAllow all calls (use with care).

Renderer-handled confirmation

Some modules (most notably the API Agent) prefer to render their own in-panel confirmation UI instead of a native dialog. They opt in with rendererHandlesConfirmation: true on the stream — the main-process gate then skips the native dialog so you see exactly one confirmation prompt, in the right place.

Historical note: Earlier builds double-gated some calls (native dialog AND in-panel) which made the agent appear stuck. The flag fixes this and is now used in every per-module agent.

Allow-list per session

For repeated operations (like running 50 SQL inserts after a CSV import), tick Allow this kind of action for the rest of this run on the first prompt. The allow-list resets when the agent run ends.

What's never gated

  • Read-only operations (list files, show schema, read response, browse inbox).
  • Pure analysis (summarise, explain, score).
  • Generating drafts that you preview before applying.

Last updated 3 hours ago

No matches.