A benchmark for LLMs on complicated tasks in the terminal