Open Agent Harness — Development skill for Claude Code

open-agent-harness (`oah`)

`oah` is a small command-line tool that makes **Claude Code** and **Codex CLI** work as one disciplined team on your own repository. It is written for people who are starting to program with AI.

Two AI coding sessions on one repository will happily edit the same files. "Done" means whatever the agent says, and a review has no rule for when it is finished and no proof. `oah` gives both engines one rulebook:

• Claude implements first. Codex reviews, corrects and re-reviews until the review converges.
• The task's difficulty picks the model, the effort and the skills. Fixed rules decide first. An AI judge is optional, and every result says whether it was used.
• An impact map tells you what a change affects and which checks to run again.
• A project wiki remembers why things were done.
• "Done" is a ledger of evidence, and work lands only after the review converges and CI is green.

`oah` is Python 3.11+ with the standard library only. It runs natively on macOS and Windows (no WSL). It does not replace Claude Code or Codex: it drives them and keeps the records that make their work trustworthy.

[!WARNING] **Developer preview. Not ready for general use yet.** On `main` today: the project skeleton (`oah --version`, `oah doctor`, the privacy scan, CI on macOS and Windows), leases, the config loader, packets, intent and preflight. Routing, skills, gates, hooks, the impact ma