Would that be useful in other harnesses/models like Codex? Wouldn't that lead to increased token usage?