The safety model: why every mutating action stops for a human

[MSKazemi]

KubeIntellect can act — scale, restart, delete — not just diagnose. The thing that makes that safe is a hard invariant: every mutating operation pauses for explicit human approval, with a server-side dry-run diff and RBAC, before anything executes. Read-only queries run immediately.

This is deliberately different from both read-only assistants (which can't act) and raw automation (which acts without judgment). We'd love your input on the edges:

• What should the approval UX show you to make a safe decision fast?
• Where should RBAC roles (admin / operator / readonly) draw the line?
• How do we best defend against indirect prompt injection — cluster logs/events that try to talk the agent into an action?

If you've operated clusters at scale, your scar tissue is exactly what we want here.

(We'll add a dedicated Architecture & RFCs category soon; posting here in General for now.)