My take is that when an LLM calls a CLI, a lot can go wrong that has nothing to do with the model. It's just that the CLI itself was not designed for LLM use, ultimately creating issues, sudden stops, or token over-consumption.

I'd be interested in collecting your opinion on this tool: https://github.com/Camil-H/cli-agent-lint

For the record, this is not commercial software, just an open-source hobbyist project.

Thanks in advance!