A Comma and a Question Mark

Summary

I’ve been using a terminal for more than two decades now, building muscle memory for find flags and git commands. I’m not sure how it happened, but reaching for --help has become less and less natural. Instead of typing the command I would start typing the sentence: “find the 5 largest files”, “show me the last 3 commits”. command not found kept reminding me: computers can’t talk. Or can’t they? In the glorious age of slop, any(one)(thing) can talk. And sure it did; all it took was about a hundred lines of code to stitch together zsh, llama-server, Qwen3.6, Pi and three commands: A comma for “give me a command” A question mark for “answer my question” It was fun, easy, and only cost $7k for a M5 Max MBP with 128GB of unified memory. The comma Now I can type a comma followed by plain English, a description of what I want to do. A few seconds later I get a short list of commands, each with a one-line explanation. I pick one, it drops onto my prompt line. I read it, maybe edit it, and press Enter myself. Of course, being the CEO of the Structured Outputs Company™, I had to use a few tricks: JSON Schema to get a list of {command, note}, and some grammar fun to force the command prefixes (ls, git, etc.). The thing answering my commas is a 27B parameter model running locally through llama.cpp. It is not a frontier model and it doesn’t have to be. I’m not asking it to be brilliant; I’m asking it to propose four ways to list large files. Pinning the shape and a local model is more than enough, so is my laptop. The question mark The question mark does a different job, so it gets a different tool. A question might be answered from what the model already knows, or it might need to read a file on my disk, or check something current on the web. So the question mark hands the prompt to a small local agent instead (the Amazing Pi). I’ve given it a deliberately narrow toolset: it can read files and it can search the web, and nothing else. No writing, no editing, no running shell co...