pu.sh

pu.sh is a minimal, portable coding-agent harness written entirely in shell script — no Node, no pip, no Docker, no runtime. It implements a full agentic loop (send prompt → parse response → execute tool → append to history → repeat) in roughly 400 lines, weighing less than 40 KB. It supports both Anthropic and OpenAI APIs, offers an interactive REPL with multi-turn memory, and can be piped together with other agents for composable workflows.

• Zero-dependency install: Download a single shell file with curl, chmod +x, and run — no package manager required.
• Dual provider support: Works with Anthropic Messages API and OpenAI Responses API; auto-detects provider from API key or model name.
• 7 built-in tools: bash, read, write, edit, grep, find, ls — a minimal but complete surface area for coding tasks.
• Interactive REPL: Multi-turn sessions with memory, REPL commands like /model, /effort, /compact, /export, /skill:name, and inline shell via !cmd.
• Reasoning effort control: Set AGENT_EFFORT from none to max; gated by model support for both providers.
• Auto-compaction: Automatically summarizes older turns when the context budget is exceeded; manual /compact [focus] also available.
• Pipe mode: --pipe flag produces clean stdout for composing agents together in shell pipelines.
• Checkpoint and resume: Writes .pu-history.json for automatic session resume; every step logged to .pu-events.jsonl as structured JSONL.
• Skills and prompt templates: /skill:name loads SKILL.md from local or user directories; /template expands .pi/prompts/template.md.
• Context files: Auto-loads AGENTS.md / CLAUDE.md from the current directory upward.
• Confirmation mode: AGENT_CONFIRM=1 prompts before every tool execution for safer operation.
• Regression test suite: 90 no-API behavioral tests via bash eval/test_real.sh covering JSON parsing, tool behavior, compaction, and more.
• MIT licensed: Fully open source — read every line, fork it, modify it, deploy it anywhere.