Project
Claude Code operating framework
A version-controlled operating layer for Claude Code. Every project I run goes through it. It sets which model does which job and what each agent is allowed to touch.
I started it on 14 July 2026, and by November it had over 1,600 commits. I set the rules and the structure. Agents do the work inside it. Most of the commits to the framework itself are theirs.
A lead agent plans and reviews. It hands the rest out by task to cheaper models: Haiku explores and reads, Sonnet builds. Opus reviews any change to the rules themselves, from a fresh context that hasn't seen the reasoning behind it. Four agent tiers in all.
Each project has its own named agent, with its own git identity and its own permissions. The history shows which agent made which commit, and an agent working outside its home proposes on a branch for the owning agent to merge.
63 deny rules keep credentials out of agent context. 21 hook scripts check commands and commits before they run. A check I haven't seen fail isn't one I trust, so they get tested against known-bad input.
No rule changes without an independent review and my sign-off. All the agents share one GitHub account, so no model gets to approve its own rules.
The voice AI pipelines were built this way.
The repository is private, so there's nothing to link. I'm happy to walk through it.