dual-reviewlisted

# Dual-model review gate Two reviewers from different vendors have decorrelated blind spots. The independence is the point: **neither reviewer may see the other's findings before producing its own.** Agreement is strong evidence a finding is real; a single-model finding is a hypothesis that needs verification before acting. ## Procedure 1. **Scope the diff.** Uncommitted changes by default. If a base ref argument was given, review the WORKING TREE against the merge-base - `BASE=$(git merge-base <base> HEAD) && git diff $BASE` - so uncommitted implementation is reviewed too (committed-only `<base>...HEAD` silently skips everything not yet committed), **plus** untracked files: list them with `git ls-files --others --exclude-standard` and read each one directly. If `merge-base` fails (unborn HEAD, unknown base), fall back to `git diff HEAD` plus the untracked files. 2. **Run both reviews independently - in parallel, same input:** - **Claude side:** launch the `code-reviewer` subagent on the diff. - **Codex side:** call the `codex` MCP tool with the diff and instructions equivalent to the code-reviewer's: report EVERY defect found - including ones it is uncertain about, marked as such - each with file:line, severity, and a concrete failure scenario. The bar is concrete: anything that could cause incorrect behavior, a test failure, a security hole, or a misleading result; omit pure style/naming nits. Reviewers must not self-filter for confidence - the triage step below d