Accepted ADR 0002 · 2026-10-16
#2. A consent gate, not a sandbox
#Context
AI agents need root for real work; copying sudo commands back and forth is slow and error-prone, and giving an agent passwordless sudo removes the human from the loop.
#Decision
anumati shows every command, with the agent's reason, in a window the user approves with their password. It defends against an agent running something unseen; it doesn't defend against malicious code already running as the user (which could fake any password prompt).
#Consequences
- The threat model is stated plainly in the README and SECURITY.md.
- Use it with agents you trust to ask honestly.