the token split is wild - 93% input vs 4% output. makes sense to optimize output but forcing short responses can hurt coherence in longer agentic sessions