Pu.sh: 400-Line Shell Script Coding-Agent Harness from HN

Pu.sh is a full coding-agent harness written in ~400 lines of shell script, designed for maximum portability. It requires only sh, curl, and awk — no npm, pip, Docker, or other dependencies. The project is inspired by Pi (pi.dev) and uses the same 7-tool surface and exact-text edit model.
Key Features
- Supported models: Anthropic (Claude) and OpenAI
- 7 built-in tools: bash, read, write, edit, grep, find, ls
- REPL mode for interactive use
- Auto-compaction of conversation context
- Checkpoint/resume capabilities
- Pipe mode for scripting
- 90 no-API tests to verify functionality without external calls
Notably missing: TUI, streaming, image support, OAuth, Windows compatibility — and, as the author puts it, "dignity."
Code and Philosophy
The author admits they cannot read most of the awk code, which was written by Pi/Claude/Codex. The reasoning loop handles JSON parsing and tool calls across turns using awk. The agent loop itself is tiny — most of the complexity in real agent CLIs comes from DX and hardening. The project is released under a self-imposed rule: no new dependencies and under 500 LOC.
Setup
curl -sL pu.dev/pu.sh -o pu.sh && chmod +x pu.sh
./pu.sh
All you need is curl, awk, and an API key. The author encourages readers to build their own harness adapted to their needs.
📖 Read the full source: HN AI Agents
👀 See Also

SkyClaw: An Open Agent Runtime Written in Rust
SkyClaw is an open-source agent runtime written in Rust with 34 new features across 7 development phases. It includes task checkpointing, SQLite-backed persistent queues, parallel tool execution, and multi-tenancy support.

Structured Claude Skill for B2B SaaS Growth Workflows
A developer has open-sourced a Claude Skill that structures B2B SaaS growth knowledge into playbooks and case studies to improve Claude's output quality. The repository includes 5 SaaS case studies, a 4-stage growth flywheel, and 6 structured playbooks.

LLMs Leak Reasoning into Structured Output Despite Explicit Instructions
A developer building a tool that makes parallel API calls to Claude and parses structured output found that validation models intermittently output reasoning text before corrected content, despite explicit instructions to return only corrected text. The fix involved prompt tightening plus a defensive strip function that runs before parsing.

Hollow AgentOS reduces Claude Code token usage by 68.5% with JSON-native OS for AI agents
Hollow AgentOS is a JSON-native operating system for AI agents that cuts Claude Code's token usage by 68.5% by eliminating wasteful shell command overhead. It plugs into Claude Code via MCP, runs local inference through Ollama, and is MIT licensed.