// Get an independent second opinion from OpenAI Codex CLI on a plan, diff, spec, or Claude's last response. Use when the user asks to "get a second opinion", "have codex review", "cross-check with codex", or wants adversarial review of Claude's output. Argument is one of plan|diff|spec|last-response (default last-response).
Run mutant, read mutation reports, fix alive mutations, and verify coverage. Use when running mutation testing, responding to alive mutations, or improving test quality. Triggers: "mutation testing", "mutant", "alive mutation", "mutation coverage". WHEN NOT: Writing tests from scratch (use rspec-agent), fixing failing tests, or general code review.
Multi-axis adversarial review using friction engineering. Routes a spec, plan, ADR, service design, schema, or any design artifact through 5 specialized reviewers with explicit prohibitions. Each reviewer tags positions as [sound], [contestable], [blind_spot], or [refuted]. Returns a consolidated friction report for human arbitration before implementation begins. Use when reviewing feature specs, architecture decisions, service designs, database schemas, or any artifact where hidden assumptions must be surfaced. WHEN NOT: code already written (use code-review), routine CRUD with no design decisions, quick one-off questions, or post-implementation reviews.
Reviews Dependabot gem upgrade PRs for breaking changes, codebase impact, and merge readiness. Use when user pastes a Dependabot PR URL, asks about a gem version bump, or wants to audit open dependency PRs ("which dep PRs are safe to merge", "audit our deps", "check dependabot"). WHEN NOT: Non-Dependabot PRs, npm/yarn upgrades, or general code review.
PostgreSQL database patterns for query optimization, schema design, indexing, and security.
Behavioral guidelines to reduce common LLM coding mistakes. Use when writing, reviewing, or refactoring code to avoid overcomplication, make surgical changes, surface assumptions, and define verifiable success criteria.
Audits Rails application accessibility against WCAG 2.2 Level AA, detects violations with axe-core / Lighthouse / Pa11y, and reports remediation guidance for ERB views, ViewComponents, Stimulus controllers, and Turbo-powered interactions. Use when the user wants an accessibility audit, WCAG compliance check, a11y review, or mentions screen readers, keyboard navigation, ARIA, color contrast, or Section 508 / ADA / EAA. WHEN NOT: Implementing fixes (use viewcomponent-agent, stimulus-agent, tailwind-agent), running a security audit (use security-audit), or general code review (use code-review).
| name | codex-review |
| description | Get an independent second opinion from OpenAI Codex CLI on a plan, diff, spec, or Claude's last response. Use when the user asks to "get a second opinion", "have codex review", "cross-check with codex", or wants adversarial review of Claude's output. Argument is one of plan|diff|spec|last-response (default last-response). |
| argument-hint | [plan|diff|spec|last-response] |
| user-invocable | true |
| allowed-tools | ["Read","Write","Bash(codex *)","Bash(command -v *)","Bash(ls *)","Bash(fd *)","Bash(mktemp *)","Bash(cat *)","Bash(printf *)"] |
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.
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; }
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.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 $(mktemp -t codex-rubric).txt:
You are an adversarial reviewer giving a second opinion. Be terse. Do not restate what is correct.
Check, in order:
1. Correctness — logic errors, off-by-one, wrong APIs, broken invariants.
2. Missing edge cases — nulls, empty inputs, concurrent writes, partial failures, auth gaps.
3. Simpler alternatives — is there a shorter, more conventional, or better-supported approach?
4. Risk — blast radius, reversibility, data loss, security implications.
Output format: 4 sections with those exact headings. Skip a section if you have nothing material. No preamble, no summary, no praise.
Use a single Bash call with 600000ms timeout. Capture stdout and exit code. Do not pass -c model=... — rely on the user's ~/.codex/config.toml defaults.
diff:
codex review --uncommitted
plan | spec | last-response:
{ printf '=== CONTENT TO REVIEW ===\n'; cat "$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.
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.
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.