architect-loop, an open-source pair of Claude Code skills from Dan McInerney, enforces a strict division of labour: Claude Fable specs and reviews the work, GPT-5.5 Codex writes it, and the two never swap roles.
Per the project's README, Fable commits acceptance "gates" to a docs/gates/ folder before any builder runs, and those gate files are read-only. If a Codex builder edits one, the slice fails automatically. Each builder runs in its own git worktree, may touch only its declared files, and has no commit access; Fable then runs the gate commands itself, on the logic that "builder claims are hearsay," and merges only the lanes that pass. Nothing carries between sessions except the repo: the gates, per-lane notes, and a pruned docs/HANDOFF.md.
It is a bet that the expensive failure in agentic coding is not bad code but an agent grading its own homework, and that freezing the rubric before work starts is cheaper than trusting the marker afterwards.