---
name: codex-review
description: Independent gpt-5.6 diff review via the Codex CLI, normal or steerable adversarial with JSON findings. Use before shipping nontrivial changes.
---

# Codex Review

Independent gpt-5.6 review of local changes through the Codex CLI. Read-only - never
lets Codex modify the repo. Works identically from the main loop, subagents, and
workflow stages (no plugin runtime required).

Reviews run on `gpt-5.6-luna` at `"max"` (MODELS.md: Luna max is the Codex
default for almost everything, review included). Escalate to `gpt-5.6-terra` at
`"max"` when the review is the load-bearing check - a one-way door, a
security or money path, or an adversarial pass whose verdict you will act on
directly. Drop to `gpt-5.6-luna` at `"high"` only for a fast look at a small
diff where latency matters more than depth.

Pin BOTH the model and the effort on every command below rather than inheriting
CLI defaults. For `codex review`, `-m` is a GLOBAL flag and must come before the
subcommand. Paths written as `<skill-dir>` mean this skill's base directory
(provided when the skill is invoked).

## Mode selection

- **Normal review** (default): overall code-quality pass on a diff, same quality as
  running `/review` inside Codex.
- **Adversarial review**: when the user (or you) wants a specific decision, risk area,
  or design pressure-tested. Steerable with focus text; returns structured JSON.

## Normal review

```bash
codex -m gpt-5.6-luna -c model_reasoning_effort="max" review --uncommitted   # staged + unstaged + untracked
codex -m gpt-5.6-luna -c model_reasoning_effort="max" review --base main     # branch vs base
codex -m gpt-5.6-luna -c model_reasoning_effort="max" review --commit <sha>  # a single commit
```

- Multi-file/large diffs take minutes: run via Bash `run_in_background: true`; the
  harness notifies on completion (no polling needed).
- Run as ONE bare command - no pipes, no `cd &&` chains (keeps the RTK hook inert).

## Adversarial review

1. Build the prompt from `references/adversarial-prompt.md`: fill `{{TARGET_LABEL}}`
   (e.g. "diff vs main"), `{{USER_FOCUS}}` (focus text or "general adversarial pass"),
   `{{REVIEW_COLLECTION_GUIDANCE}}` (usually: "Collect the diff yourself with git; read
   changed files fully."), and `{{REVIEW_INPUT}}` (either the diff inline, or an
   instruction to run `git diff <base>...HEAD`). Write it to the scratchpad.
2. Run as a single bare command, feeding the prompt file via stdin (`-` reads the
   prompt from stdin; a plain `< file` redirect keeps the command pipe-free):

```bash
codex exec -m gpt-5.6-luna -c model_reasoning_effort="max" -s read-only --cd <repo> --output-schema <skill-dir>/references/review-output.schema.json --output-last-message <scratchpad>/codex-adversarial-<ts>.json - < <scratchpad>/prompt.md
```

3. Read the output JSON: `verdict` (`approve` | `needs-attention`), `findings[]`
   (severity, title, body, file, line_start/line_end, confidence 0–1, recommendation),
   `next_steps[]`.

## Presenting results

- Present findings ranked by severity, verbatim in substance - do not soften.
- **Never auto-apply fixes from a review.** Present first; fix only when the user asks
  or when you already have an approved implementation task that covers it.
- Note the session can be reopened in Codex with `codex resume` if deeper follow-up
  is wanted.
- If Codex disagrees with a deliberate design choice, report the disagreement and your
  own position - don't silently adopt either side.

## Effort

Always pin effort explicitly - never inherit a vendor default. Per MODELS.md,
`gpt-5.6-luna` at `"max"` is the review tier; `"high"` only for a fast look at a
small diff where latency matters more than depth. The Sol worker tiers are
retired. If a Luna max review still leaves material doubt, escalate per
MODELS.md: one independent `gpt-5.6-terra` `"max"` adversarial pass, then Root
resolves the hard part inline.

References adapted from openai/codex-plugin-cc (Apache-2.0; see references/NOTICE
and the full license text in references/LICENSE).