This 400-line shell script runs AI coding agents. Nobody can audit it.

Pu.sh is a coding-agent framework packed into 400 lines of POSIX shell. Created by software engineer Deluan Quintão, the tool runs AI coding agents with just curl, awk, and an API key. You don't need npm, pip, or Docker. It supports OpenAI and Anthropic models, handling state management and JSON parsing through custom AWK functions, all within its self-imposed constraint. Quintão describes it as a "slop cannon small enough to fit your pocket."

But the 400-line limit comes at a real cost. The code is minified to hit that number, and developers on Hacker News aren't happy. Commenter ricardobeat called the minification a "marketing gimmick." Others labeled it a "security nightmare." When a script handles your API keys and can modify your file system, readability isn't optional. It's the whole point. The lack of comments and clear formatting makes the code nearly impossible to audit.

Quintão has a track record with tools like Inertia and CushyStudio. He deliberately chose density over clarity here. Some commenters asked for a readable version or a port to Lua. That's what would actually make Pu.sh useful beyond a technical stunt. The project exposes a real tension in agent tooling: minimalism is appealing, but not when it means you can't trust what you're running.