Work / Open source

Ralph Harness: coding-agent loops that only ship verified work

Open-source CLI that runs Claude Code in a loop for days; a commit ships only if your own checks pass. 95 of 106 iterations kept.

Organisation
Open source
My role
Author and maintainer
Period
Sep 2026 - present
Ralph Harness: every iteration is a fresh claude -p, and git decides what shipped; a log shows one iteration shipped and the next reverted by the verify gate

Context

Long-running coding-agent loops are useful for backlog work, but they drift, and a model saying "done" is not evidence that anything works.

Problem

Run an agent unattended for days without letting broken or unverified changes reach the main branch.

Approach

  • Every iteration starts a fresh claude -p, so context from a bad attempt does not leak into the next one.
  • A commit is kept only if the checks you configure pass: tests, type checks, lint, or anything with an exit code.
  • Zero runtime dependencies; built for Bun and published on npm and Homebrew.

Key decisions

The checks decide, not the model
An agent can claim success on anything. An exit code from your own test suite cannot be argued with.
Fresh process per iteration
Restarting clears a polluted context and makes each iteration independent, so one failure does not cascade.
No dependencies
A tool that runs for days on your machine should be easy to audit and hard to break with an upstream update.

Outcomes

  • Kept 95 of 106 iterations across 8 real loops; the other 11 were discarded instead of merged.Source: Project README
  • Passed 2,000 npm downloads in its first two months (September to November 2026).Source: npm

Stack

  • TypeScript
  • Bun
  • Claude Code
  • CLI
  • npm
  • Homebrew