Second-opinion review from another model

# AGENTS.md: second-opinion review from another model

Before you call a non-trivial change done, get an independent read-only review from a different agent or model.
A different model misses different things; a reviewer that cannot edit cannot "fix" its way past a problem.

## When to ask
- The change touches three or more files, a public interface, auth, payments, data migrations or concurrency.
- You fixed a bug you could not reproduce with a test.
- You are about to tell the human the work is finished.

Skip it for typo fixes, formatting, dependency bumps with green CI, and changes the human will review line by line anyway.

## Rules
- Use the `cross-model-review` skill. The reviewer runs read-only (Codex: default `codex exec` sandbox; Claude Code: plan or
  dontAsk mode with read tools only).
- Give the reviewer the goal, the diff and the test command. Do not give it your reasoning; it should judge the code, not your story.
- Treat findings as claims. Verify each against the code before changing anything. Reject wrong findings with a one-line reason.
- Maximum two review rounds per change. If round two still finds blocking issues, stop and report to the human.
- Never paste secrets, `.env` contents or customer data into a review prompt.
- In the final report, list what the reviewer found and what you did about each finding.

# CLAUDE.md

@AGENTS.md

## Claude Code notes
- Skills from this loadout live in `.claude/skills/`: `cross-model-review`. Use them when their description matches the task.
- Follow AGENTS.md above; where this file and AGENTS.md disagree, AGENTS.md wins.

<!-- From the Agent Loadout "Second-opinion review from another model" v1.0.0 — https://loadout.jnw-it.com/loadouts/second-opinion-review/ -->

# GEMINI.md

@./AGENTS.md

## Gemini CLI notes
- Skills from this loadout live in `.agents/skills/`: `cross-model-review`. Use them when their description matches the task.
- Follow AGENTS.md above; where this file and AGENTS.md disagree, AGENTS.md wins.

<!-- From the Agent Loadout "Second-opinion review from another model" v1.0.0 — https://loadout.jnw-it.com/loadouts/second-opinion-review/ -->

---
name: cross-model-review
description: Ask a second agent (Codex from Claude Code, or Claude Code from Codex) for a read-only review of the current diff, then verify and act on its findings. Use before reporting a non-trivial change as done.
---

# Cross-model review

1. **Prepare.** Commit or stage the work. Pick the base (usually `origin/main`). Write the diff to a temporary file:
   `git diff origin/main...HEAD > "${TMPDIR:-/tmp}/review.diff"` (add `git diff` of uncommitted work if any).
   Check its size: over ~3,000 lines, review per directory instead.
2. **Write the prompt** from `references/review-prompt.md`: goal, test command, focus areas. No reasoning, no secrets.
3. **Run the reviewer**, read-only, from the repository root:
   - From Claude Code, ask Codex:
     `codex exec --ephemeral -o "${TMPDIR:-/tmp}/review.md" "$(cat "${TMPDIR:-/tmp}/review-prompt.txt")" < "${TMPDIR:-/tmp}/review.diff"`
     (`codex exec` is read-only unless you pass `--sandbox`; do not pass it.) `codex review` is an alternative for the current branch.
   - From Codex or another agent, ask Claude Code:
     `claude -p --permission-mode dontAsk --allowedTools "Read,Grep,Glob,Bash(git diff *),Bash(git log *)" "$(cat "${TMPDIR:-/tmp}/review-prompt.txt")" < "${TMPDIR:-/tmp}/review.diff" > "${TMPDIR:-/tmp}/review.md"`
   If the reviewer CLI is not installed or not logged in, stop and tell the human; do not substitute a different tool silently.
4. **Triage.** For each finding: open the cited code, decide valid / invalid / unsure. Fix valid ones. Write a reason for invalid ones.
5. **Round two** only if you changed code because of blocking findings. Same prompt plus "Previously reported: …". Never a third round.
6. **Report** a table: finding, verdict, action. Delete the temporary files.

# Review prompt template

Save as review-prompt.txt, fill the angle brackets, keep it under 40 lines.

```
You are reviewing a diff you did not write. You cannot edit files. The diff is on stdin.

Goal of the change: <one or two sentences>.
How it is tested: <command>, which <passes / was not run because …>.
Focus on: <for example: error handling in the API client, the migration's down step, race conditions in the cache>.

Report only problems that would cause wrong behaviour, security issues, data loss, broken callers or missing tests
for changed behaviour. Ignore style and naming unless they hide a bug.

Format, most severe first, at most 10 items:
<blocking|should-fix|nit> — <path>:<line> — <problem> — <suggested fix>

If you find nothing blocking, say "No blocking issues" and stop.
```