Show HN: Pu.sh – a full coding-agent harness in 400 lines of shell

TL;DR Highlight

ShellAgent runs LLM-powered coding tasks with just curl and awk, ditching npm, pip, and Docker.

Who Should Read

Developers wanting to quickly experiment with AI coding agents without heavy frameworks, or backend/infrastructure developers eager to learn by dissecting agent harness architectures.

Core Mechanics

pu.sh implements an agent harness—the execution framework—in 400 lines of pure Shell script, functioning with only curl, awk, and an LLM API key, bypassing npm, pip, and Docker.
Installation completes with a single command: `curl -sL pu.dev/pu.sh -o pu.sh && chmod +x pu.sh`, followed by direct execution with `./pu.sh`.
Released under the MIT license, pu.sh provides both source code and documentation on GitHub, aiming for rapid code generation—self-described as a 'slop cannon'.
Community criticism centers on the code's low readability due to minification, a constraint imposed to meet the 400-line limit.
Its complete lack of dependencies makes it ideal for minimal environments like busybox or container-based development setups.
The harness prioritizes 'tool call recording/replay and failure mode handling'—the most complex area of agent debugging—as a core evaluation metric.

Evidence

"The code's minification to meet the 400-line limit sparked debate, with critics arguing it created security vulnerabilities and resulted in '100% vibe coding.' Requests for the uncompressed 6KB source were made."

How to Apply

If you want to quickly test an AI coding agent without setting up Node.js or Python locally, install with `curl -sL pu.dev/pu.sh -o pu.sh && chmod +x pu.sh` and run it after setting your LLM API key.
When deploying AI coding agents in minimal container environments based on busybox or Alpine Linux, leverage pu.sh's curl and awk dependency to avoid adding further requirements.
If you're learning to implement agent harnesses, use pu.sh's structure as a reference for creating a similar single-file agent in Shell or Node.js; the community-shared Node.js version (willhanlen.com) provides a helpful example.
If you need an agent runtime independent of specific vendors like Claude, consider integrating pu.sh or aloop (github.com/zackham/aloop) – vendor-agnostic open-source harnesses – into your projects.

Code Example

snippet

# Installation and execution
curl -sL pu.dev/pu.sh -o pu.sh && chmod +x pu.sh
./pu.sh

# Requirements: curl, awk, LLM API key
# npm, pip, docker not required

Terminology

agent harnessThe execution framework that enables AI agents to use tools like code writing, file reading, and command execution. An agent cannot function without it.

tool callAn LLM's ability to go beyond text generation and invoke external functions or system commands. For example, if asked to 'read a file,' the LLM directly calls a file reading function.

vibe codingA development practice where code generated by AI is used without understanding its underlying logic. It's fast but prone to debugging issues and security risks.

minifyReducing code size by removing whitespace, comments, and long variable names. Commonly used for browser JavaScript, but can make Shell scripts unreadable.

vendor-agnosticAn architecture that isn't tied to a specific company's (OpenAI, Anthropic, etc.) APIs or SDKs, allowing for interchangeable LLM services.

record/replayThe ability to store and re-execute an agent's tool call history. Essential for bug reproduction and analysis.