Ralph Harness: a tiny guarded loop scaffold for Codex CLI

[rxdt]

I built Ralph Harness, a small open-source Python scaffold for running Codex CLI and other terminal coding agents in short guarded loops.

The goal is not to replace Codex or build a large orchestrator. It is to make repeated agent runs safer and less sloppy:

• repo-local specs
• fresh context per iteration
• hard time/iteration caps
• git hook gate
• CI verification
• coverage gate
• no hosted backend
• no token service
• no agent lock-in

Try it:

uvx harness install my-project

Repo: https://github.com/rxdt/loopgate-harness
PyPI: https://pypi.org/project/ralph-harness/

I built it because the failure mode I kept hitting was not "the model is too weak." It was agents drifting, writing shallow tests to satisfy coverage, and making progress without enough repo-level proof.

I would like feedback from people already running Codex in loops: what guardrail is missing before you would let an agent work unattended on a real repo?