personal-codex-review

Verify codex exec supports -s read-only, -c, -o, and stdin prompts, then write a brief file specifying uncommitted mode with git status --porcelain, git diff --cached, and git diff. Launch codex exec from a herdr pane and read the result file after the nonce-marked done file confirms exit=0.

It only runs cross-review when a trusted deterministic preflight verifies author=claude and reviewer=codex. Codex-authored changes are routed back to Claude, mixed or unknown authors go to human adjudication, and explicit second opinions are labeled second-opinion only, never independent approval.

Yes, via base mode: the caller supplies expected base and head OIDs, and the skill verifies local HEAD, the base ref via check-ref-format and rev-parse with --end-of-options, and a clean worktree before reviewing. It never checks out, fetches, resets, or mutates the worktree itself.

The skill returns Status: BLOCKED with Blocked at: capability-preflight instead of guessing flags or substituting pane output for the result file. It hands off to a verified reviewer route or a human, and never modifies user configuration.

Inside a Claude Code or Codex sandbox, nested sandboxes are impossible, so codex exec cannot run commands. Without herdr and without verifiable unsandboxed settings (including managed settings), the skill blocks at launch-path and hands a manual run script to a human.

It performs no GitHub reads, comments, approvals, or merges, and no repository writes; the PR lifecycle belongs to the caller. Its verdict covers only the code review process, not CI, required checks, or overall merge readiness.