I built a deterministic policy engine that stops AI coding agents before they run a bad command (opens in new tab)

Output filters check what the model says. I wanted something that checks what it does, before the tool call executes.