Stop AI coding agents from quietly wrecking your repo.
AI coding agents are fast. They also leave dirty worktrees, out-of-scope files, generated residue, vague "done" claims, and tiny cleanup tasks that slowly bloat a project.
Loop Governor is a local post-iteration gate for Codex, Claude Code, Cline, Aider, OpenHands, Goose, SWE-agent, and other AI coding workflows.
It does not run the agent. It does not write code. It does not clean, repair, commit, or merge.
It reads:
- a narrow
project-os-loop.yamlcontract - real
git status - allowed paths and lanes
- required evidence
- claim boundaries
Then it returns a governance decision:
continuestop_for_userupdate_railisolate_workcommit
Use it when you want AI agents to move fast without turning your repository into a pile of mystery diffs.
Public alpha goal: collect failure reports from real AI coding workflows before claiming production readiness.
measured worktree
-> temporary project-os-loop.yaml
-> validate-config-only
-> real git-status gate
-> classify residue
-> claimStatus narrowing
-> commit target change only
docs/quickstart.mddocs/contract-authoring.mddocs/external-measured-project-recipe.mddocs/cannot-claim.mddocs/decisions.md
Loop Governor does not control Codex or any agent.
It is not a runner, hook, sandbox, registry, dashboard, plugin, Guardian Agent, or autonomous supervisor.
continue does not mean final claim complete.
A valid config does not mean an iteration is safe.
npm install
npm test
npm run typecheck