Claude Code, Codex, or Grok — one gate discipline
The implement loop is provider-blind by construction: whichever agent writes the diff, the same independent capture, the same cross-check, and the same human gate stand between it and your repo.
How it works
Claude Code, the OpenAI Codex CLI, and the xAI Grok CLI all sit behind the same agent-agnostic adapter — the loop neither knows nor cares who did the typing.
A registry-driven model picker routes each implement task by complexity and cost, so a one-line fix and a cross-cutting refactor are allowed to land on different providers.
Every provider is held to the same standard: the diff is captured independently and cross-checked against the agent's claims. If capture can't verify, the proposal blocks — a new provider gets no trust discount.
Output lands at the same proposal gate with the same test evidence and reviews, whoever wrote it — switching providers never changes what a human gets to decide.
The agents behind it
Human control
The proposal gate is provider-blind — every coding agent's diff is independently captured, cross-checked, and human-decided.