an approval should expire when the thing it approved changes. how are you binding approvals to exact actions?

scrolling this sub today, half the threads are the same problem from different angles: sub-agents spawning sub-agents (whose authority?), retries that look like the same op but aren't, one-click approvals in slack, agents with creds nobody remembers granting. the pattern i keep landing on: most "approval" in agent systems is approval of an intent, not an action. someone says "yes, send the invoice" and the agent treats that as a license for whatever it does next. then the amount changes, or the recipient, or the draft gets regenerated, and the old yes still counts. what i think an approval needs to carry: the exact thing approved (diff hash, recipient, amount, file id), not a description who approved it, and whether that person is allowed to approve this kind of action an expiry a rule that any change to the bound thing voids it and asks again scope that can only narrow when passed to a child agent, never widen the annoying part is the ux. if every regenerated draft needs a fresh approval, people start rubber-stamping, which is worse than no gate. questions for people running this in prod: do you bind approvals to a hash of the action, or to a task/ticket id? how do you handle "small" changes after approval (typo fix vs changed amount)? do you have a materiality rule? when a child agent acts, do you log the original human approval id on the write? has anyone found a way to keep approval fatigue down without loosening the binding?