Codex Review

/codex-review — Independent second opinion from Codex CLI

Argument: $ARGUMENTS — one of plan | diff | spec | last-response. Default last-response.

Never fall back to reviewing Claude's own output if Codex fails. If Codex can't run, say so and stop.

1. Preflight

Run in order. Stop cleanly on first failure:

command -v codex >/dev/null 2>&1 || { echo "codex CLI not found. Install: https://github.com/openai/codex"; exit 0; }
codex login status 2>&1 | grep -qi "logged in" || { echo "codex not logged in. Run: codex login"; exit 0; }

2. Resolve target

• last-response (default): identify your previous substantive assistant turn (the last turn with user-facing prose — skip turns that are pure tool calls or tool results). Write that prose verbatim to a temp file via the Write tool: $(mktemp -t codex-review).md. Record the path as $TARGET_FILE.
• plan: if there is an active in-conversation plan, write it to a temp file as above. Otherwise pick the newest file under ./.claude/plans/ (ls -t ./.claude/plans/*.md 2>/dev/null | head -1). If neither exists, report "no plan found" and stop.
• spec: pick the newest spec.md under ./specs/ (fd -t f 'spec\.md$' ./specs 2>/dev/null | xargs ls -t 2>/dev/null | head -1). If none, stop.
• diff: no file needed — Codex reads the repo directly via codex review --uncommitted.

3. Build rubric (skip for diff)

The diff target uses codex review's built-in review prompt — do not pass a custom rubric there (--uncommitted is mutually exclusive with [PROMPT] and will error).

For plan | spec | last-response, write this rubric to $(mktemp -t codex-rubric).txt:

You are an adversarial reviewer giving a second opinion. Be terse. Do not restate what is correct.

Check, in order:
1. Correctness — logic errors, off-by-one, wrong APIs, broken invariants.
2. Missing edge cases — nulls, empty inputs, concurrent writes, partial failures, auth gaps.
3. Simpler alternatives — is there a shorter, more conventional, or better-supported approach?
4. Risk — blast radius, reversibility, data loss, security implications.

Output format: 4 sections with those exact headings. Skip a section if you have nothing material. No preamble, no summary, no praise.

4. Invoke Codex

Use a single Bash call with 600000ms timeout. Capture stdout and exit code. Do not pass -c model=... — rely on the user's ~/.codex/config.toml defaults.

• diff:

  codex review --uncommitted
  

• plan | spec | last-response:

  { printf '=== CONTENT TO REVIEW ===\n'; cat "$TARGET_FILE"; printf '\n\n=== REVIEW INSTRUCTIONS ===\n'; cat "$RUBRIC_FILE"; } \
    | codex exec --sandbox read-only --skip-git-repo-check -
  

Pass all content via stdin. Never put file contents on argv.

5. Emit review

On exit code 0:

### Codex Review

<codex stdout verbatim — no edits, no paraphrase>

On non-zero exit or timeout:

### Codex Review (FAILED)

Codex exited with status <N>. Stderr tail:
<last 20 lines of stderr>

Stop after emitting failure. Do not substitute your own review.

6. Ask before reconciling

After a successful review, end your turn with one question to the user:

Want me to reconcile — decide which points to adopt, reject, or defer?

Only if the user says yes, produce a ### Reconciliation section with three bullet lists: Adopt, Reject (each with a one-line reason), Defer (each with a one-line reason to revisit later). Do not rubber-stamp Codex; reject points that are wrong or off-scope.