ralph

About

Ralph operationalizes a simple claim about AI coding: an agent with fresh context every iteration outperforms one long session that degrades under context rot. The loop reads prd.json, spawns a clean instance of Amp or Claude Code for each iteration, and that instance picks the highest-priority story marked failing, implements it, runs typecheck and tests, and commits only when they pass — then marks the story done and appends what it learned to progress.txt and an evolving AGENTS.md. No state lives in the model's window between iterations; git history and flat files carry all memory, which is why the README insists stories be small enough to fit one context window. A completion phrase or iteration cap ends the run, and typecheck/test gates prevent broken code from compounding across iterations. Teams use it for long unattended pushes through a feature backlog, accepting the project's own warning to run it in isolated environments.