Augmenting CLIs and APIs for LLM Agents

The author encountered limitations in existing command-line tools and APIs when using Large Language Model (LLM) agents for reverse engineering automation, especially with the small context windows of local models. APIs need to balance providing enough information to reduce tool calls while avoiding context window overflow. Solutions explored include improved docstrings, helper functions, and pre-commit hooks. Further improvements suggested involve wrappers that cache output, structure it, and report remaining lines, as well as shell hooks providing directory information. The author concludes that existing CLIs need LLM enhancements; perhaps even a whole set of LLM-enhanced CLIs or a custom LLM shell is needed to improve the user experience for LLM agents.