One "fix" is for the caller to correctly classify those fundamentally impossible tasks and pass them to a subprocess.

Some future "AI" could be a billion benchmark-hacks and a way to tell which one is needed.