pi-coding-agent

What problem does it solve?

This Skill provides a framework and guidance for building minimal, self-contained AI coding agents that are reproducible and easy to customize, inspired by the Pi project's design principles.

Core Features & Use Cases

• Deterministic Agent Loop: Operates on a strict REPL loop (session -> LLM -> thought -> tool call -> result -> append -> next LLM loop).
• Structured Session Modeling: Uses append-only JSON logs for full reproducibility and "time travel" debugging.
• Minimalist, Deterministic Tools: Employs a fixed set of tools (read, write, edit, bash, search, plan) for predictable interactions.
• Context Engineering: Carefully crafts prompts using only necessary session elements, avoiding prompt bloat.
• Multi-Model Support: Facilitates testing and integration of different LLMs within the same session.
• Use Case: Develop a new AI coding assistant that can reliably generate boilerplate code, refactor existing files, and execute shell commands within a project, with every step auditable and replayable.

Quick Start

Use the pi-coding-agent skill to guide the creation of a new LLM coding agent that follows the Pi design principles.