multi-model-review
Multi-Model Review
Two independent review lanes run in parallel, then every merged finding is
adversarially verified against the actual code. Replaces the retired
multi-model-review workflow: per MODELS.md there is no Claude shim - the
main loop (Root) runs the codex lane itself via direct background Bash.
Inputs: repo (absolute path, required), base (default main), focus
(default: correctness, security, edge cases, API contracts, maintainability).
Shared review contract (put in BOTH lane prompts):
Repo:
<repo>. Review the diffgit diff <base>...HEAD; if that diff is empty, review staged+unstaged changes instead. Focus:<focus>. You are READ-ONLY: do not edit any file. Read the changed files fully - not just the diff hunks. Report findings with file:line, why each matters, and the precise fix. If the change is clean, use verdict "ship" with zero findings - do not invent issues.