is there a necessary difference between a harness and a shell?
what if your shell could distinguish between prompts and commands, and allow the LLM to run commands in the foreground?
a PTY, a command-prompt classifier, a wrapper around the user's shell (only tested with fish)
go run .
vibe-coded. glhf