pi-hashline-edit

A Pi extension, to substitute for the built-in edit tool, which uses a hashline-based file editing strategy to improve the accuracy and efficiency of agents' file-editing abilities, reducing the likelihood of failed edits that require extra turns and tokens by the agent.

This editing strategy has improved the edit success rate of Gemini models by 8%, Claude Sonnet 4.5 by 14.4%, etc. and reduced overall token usage, without requiring any additional training to these models.

Heavily inspired by/borrowed from Can Bölük's oh-my-pi fork of Pi. This is a pure JavaScript implementation of the hashline editing feature, extracted into a standalone Pi extension with adjustments for Node and Bun compatibility. See Can's blog post, I Improved 15 LLMs at Coding in One Afternoon. Only the Harness Changed to understand how hashline editing works and how it improves coding agent accuracy.

Note: oh-my-pi's hashline edit tool has been rewritten as a Rust crate, which significantly changed the underlying implementation. This extension originally provided parity with oh-my-pi's JS-based solution, but cannot guarantee it as omp's solution continues to evolve.

Install

pi install npm:@the-agency/pi-hashline-edit

This will enable the hashline_read and hashline_edit tools, and disable the built-in edit tool, to ensure models are not confused about which editing tool to use.