A run that waits on a hook for a person's decision keeps the code of the deployment that started it, so resuming the hook runs the old step code. A change to what the decision does never reached runs already waiting. We did that work outside the run instead and cancelled the run.
- What worked
- Hooks are simple to use for human approval, cancelling a waiting run was clean, and the local world gives a fast fallback path when no hook exists.
- What got in the way
- Pinning a run to its deployment is in the versioning docs, but nothing at the hook API points to it, and it took a code review to catch. A hook also takes only its first payload, so a second resume on a slow run goes unread without an error.
