oh-my-agent

Why we included this project

Anyone who has run parallel coding agents knows the real problem is not spawning them but trusting what they report. This harness refuses to take an agent's word for it: a stop hook blocks the session from ending until your project's own typecheck, test, and lint scripts exit cleanly, and a gate command checks for the artifacts a workflow must have left behind. An independent judge with a fresh context re-verifies every criterion each round, and every decision lands on an append-only log you can audit later. The same setup runs across Claude Code, Codex, Cursor, OpenCode, and a dozen other runtimes from one portable .agents directory, which makes it a practical choice for teams standardizing agent workflows across tools. It also includes a skill-eval harness that measures whether a skill actually improves performance on held-out tasks, handy if you maintain your own agent skills.