Show HN: Pu.sh – a full coding-agent harness in 400 lines of shell
pu.devI originally was just messing with pi-autoresearch. Gave it a sample task to build the most portable coding agent.
First cut was 6 KB of shell. Great for one-shots, unusable interactively. I was shocked it actually worked.
Started building up -- adding features — but with a self-imposed rule: no new dependencies, and sub 500 LOC. This thing had to be truly portable. Just sh, curl, awk. System primitives only.
Which means I did some genuinely disgusting things in awk, including JSON parsing and the OpenAI Responses tool loop with reasoning items carried across turns.
It's now ~400 lines. In the box: Anthropic + OpenAI, 7 tools (bash, read, write, edit, grep, find, ls), REPL, auto-compaction, checkpoint/resume, pipe mode, 90 no-API tests. Not in the box: TUI, streaming, images, OAuth, Windows, dignity.
Two honest things:
1. I stole/modified the system prompt and the architecture. Pi/Claude/Codex wrote the awk. I cannot read most of this code. This wasn't possible for me a year ago.
2. Heavily inspired by Pi (pi.dev) — same 7-tool surface, same exact-text edit model. Credit where it's due. Pi is awesome -- you should probably use them.
The agent loop itself is tiny. Almost everything else in a "real" agent CLI is DX and hardening. You can probably build your own harness exactly how you like it. Mario Zechner's AI Engineer talk on taking back control of your tools nudged me here.
The name is because it's a .sh file. The other thing it sounds like is, regrettably, also accurate. Really like the looks of it, but minifying the code to achieve the “400 lines” marketing gimmick is a huge turn-off. It’s also a security nightmare, and ensures it remains 100% vibe coded. Would rather have a readable source with an honest line count. Agreed. It would be one thing if it was a regular bash project that got minified via a script or something, since at least you could tell what it's doing. This is just spaghetti. Maybe it's spaghetti that runs, but code of this quality does not engender confidence. I think it's fine that it's minified, code-golf style, but yes I agree that this would be much more convincing and useful with an unminified source. Isn’t unwrapping and commenting just one prompt away? Assuming the agent doesn't forget or misinterpret anything. It would be much better to just have a deterministic minification script. … no, you shouldn’t use AI for simple reformatting of code without the newlines. We have syntax parsers and reformatters for this. Use them, then put that into the LLM for commenting, if you really want to. Caught by surprise by the people who want to read the code -- and thank you for the great ideas. Slapped together an 'unminified version' that has been ran against the test suite. Hopefully has more educational value https://github.com/NahimNasser/pu/blob/main/pu-unminified.sh I'm not even sure why it stopped at 500 line limit. If it's just a shell script then why not make it one line? Could you post the initial 6KB version? I'd love to have a version that I can actually read and understand. I don't care if it's 2K LOC. > Could you post the initial 6KB version? I'd love to have a version that I can actually read and understand. I don't care if it's 2K LOC. I don't think that even with APL you could get an average of three characters per line of code to get a 6 Kb file in 2 K LOC. Maybe in Wényán lang? heh... i'm afraid i didn't commit it -- as it was in the original pi-autoresearch yolo'ing phase. I love this. This is the grown-up brother of my one-liner bash+python at https://github.com/kkovacs/kkrc/blob/ecff1f65d5ccfa24d5f6695... :) Does it work in just-bash? https://github.com/vercel-labs/just-bash/blob/main/packages/... your 500 LOC requirement has made this thing completely and unnecessarily inscrutable and unreadable. Pass. Interested in something like this if the code can be more readable, it would be interesting to see it in languages like lua etc. too in future. I like the small-surface-area approach. The question I’d use to evaluate this is how well the harness records/replays tool calls and failure modes, since that is where debugging agent behavior usually gets messy. I love it. I think I'm going to have fun with this and possibly learn a bit, too. I'm pondering a container based dev environment at the moment and might throw it in with busybox and see how far I get :) Kudos. I set on this exact journey a couple of days back and Pi is what I started reading for inspiration as well. I really can't stand the text boxes and the animations of the mainstream harnesses. I'd love to learn how to build something like this, but unfortunately the code is unreadable. If author could write a heavily commented version that would be amazing. I built full agentic-loop from scratch with full MCP support and session save and resume capabilities just this week. The intent was to learn by building full system from scratch, and share the learnings with others. Working demo: https://asciinema.org/a/983581 Code: https://github.com/freakynit/agentic-loop Tutorial: https://building-an-agent.pagey.site/ this inspired me to build something similar. i like the idea but i wanted slightly different constraints: - one file - zero dependencies (other than node) - <1k loc - ascii only - suckless configuration - works with ollama models - more readble sourcecode - pretty markdown output - toggle-able read-only mode https://willhanlen.com/~~/outbox/26/5/ai.js/?code took me only a few hours. amazing times we live in! if you are looking for a customizable agent harness that provides everything except a tui (cli, acp, embeddable python package) i built and use this for a handful of projects where i needed a vendor-agnostic replacement for claude agent sdk: https://github.com/zackham/aloop In only a time someone will do it in shell ---- oh wow you did it!!! why the LOC limit? just because? Originally a metric to drive prioritization and constraints on the project (amongst others, like zero deps, etc) -- but clearly ended up getting abused