Show HN: Pu.sh – a full coding-agent harness in 400 lines of shell
400 lines of shell. No dependencies. A fully functional coding agent that actually works—and you can read (most of) it.
Why it matters
A developer shipped a minimalist, portable AI coding agent that strips away framework bloat and demonstrates that agent loops don't require heavy tooling. It's a contrarian take on agent architecture and a working proof that control and simplicity can coexist with capability.
The key facts
17 to know~400 lines of shell script (sub-500 LOC constraint)
7 tools: bash, read, write, edit, grep, find, ls
Supports both Anthropic and OpenAI APIs
Features: REPL, auto-compaction, checkpoint/resume, pipe mode
90 no-API tests included
Zero external dependencies—sh, curl, awk only
Inspired by Pi (pi.dev) agent architecture
Published on Hacker News with 63 points
Pu.sh: 400-line shell script coding agent harness
Supports both Anthropic Claude and OpenAI models
7 integrated tools: bash, read, write, edit, grep, find, ls
Zero external dependencies—uses only sh, curl, awk, system primitives
Features: auto-compaction, checkpoint/resume, pipe mode, 90 no-API tests
Inspired by pi.dev architecture; same 7-tool surface and exact-text edit model
Implements full OpenAI tool-use loop with reasoning items across turns
Originally started as 6 KB proof-of-concept, iterated to sub-500 LOC constraint
Creator used Claude and Codex to write portions of the awk logic
Go to the source
Hacker Newspu.dev
Publisher excerpt: I 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…
