AI-triage should have separate steps for analyzing and executing code

[renanrodrigo]

The AI assistant implemented in #3 has on_permission_request=PermissionHandler.approve_all set; this may be dangerous when running on a non-isolated environment.

The safe way would be separating the triage task in two main steps, as described in #3 (review)