GitHub radar

backpass: An agent that learns from past experience

A command line interface (CLI) that analyzes transcripts from seven different coding agents running locally on your computer, suggesting data-driven changes to your AGENTS.md file (or CLAUDE.md). It’s like gradient descent for your AI’s instructions.

01kunchenguid/backpass 328JavaScript

backpass is a command line interface (CLI) that helps you iterate on your AGENTS.md file (or CLAUDE.md) based on your agent usage. It parses local agent transcript files from seven different agent harnesses (Claude, Codex, Pi, OpenCode, Grok, Cursor CLI, Hermes) without any API keys or uploading. Each suggested change includes direct quotes from actual sessions to support it, and each new rule must be supported by at least two unique sessions (no single session can support a change). Each execution will suggest up to five changes (called a “gradient step”), which are small, noisy, and generally safe. All secrets are redacted before being passed into any model calls. Changes are only written via the backpass apply command, which displays all proposed changes along with their supporting evidence and allows you to manually approve or deny each change. backpass requires Node.js version 22.5 or higher and the acpx utility on your PATH.

Why a vibe-coder should care

If you’re using Claude Code, Codex, or another agent frequently, then your AGENTS.md likely becomes outdated over time as you only update it after manually recalling a failure that occurred. backpass helps automate this process by identifying commonalities in your actual sessions and suggesting specific changes to improve your agent over time, without requiring you to manually read through the transcripts.

How to install

Copy this and send it to your agent — Claude Code, Codex, any of them:

Install backpass from https://github.com/kunchenguid/backpass, analyse the agent transcripts in this project, show me the proposed edits to AGENTS.md and wait for my decision on each one. Ask if you need any permissions.

Runs on a regular laptop — you'll need Node.js 22.5+ and the acpx utility (instructions in the README).

Open on GitHub