Show HN: Keep large tool output out of LLM context: 3x accuracy 95% fewer tokensgithub.com·10 pts·loumaciel·1