Claude Code가 짠 코드, 검수도 맡겨도 될까? Codex로 검수자 나누는 법

0:00Here's how to review code Claude Code wrote, using Codex. Ever handed off code because the AI said it was done, then found gaps later? With four commands inside Claude Code, you can add Codex as a reviewer. Say Claude Code just fixed a login function. Ask in the same chat if it's okay, and the fixer ends up grading itself.

0:19Instead, type /codex:review and Codex reads pre-commit changes and returns only findings. We'll install the plugin, run reviews, and roll back if things go wrong. Without a check that returns pass or fail, "looks done" is the only signal. So we have a different model challenge the result in a fresh context. Reviewing in a fresh context means less bias toward the code just written.

0:44Just ask it to find gaps, and it'll report something even on solid work. So we narrow the instruction to gaps affecting correctness and requirements. Split the builder and the checker, but ask narrowly. That's today's takeaway. There are 2 review commands, and one lets you specify in words what to focus on. The rewind menu has 6 options, so you can restore just the code, not everything.

1:07Checkpoints keep up to the latest 100 per session, but some changes aren't covered. We'll unpack what those are in the rollback group.

1:18The map has 6 slots, split into a review group of 3 and a rollback group of 3. It covers standard review, adversarial review, and the review gate. The rollback group is where Claude Code protects your code. We'll look at the rewind menu, changes you can't undo, and git's role.

1:36From here, we follow the order in the official repository exactly. First, you need a ChatGPT subscription or an OpenAI API key. ChatGPT subscriptions include the free plan. Node.js must be 18.18 or higher. Also note that reviews count against your Codex usage limits. Installing means entering four commands in order into the Claude Code prompt.

1:58The first adds the marketplace, targeting openai/codex-plugin-cc. The second, /plugin install, installs codex@openai-codex. The third, /reload-plugins, reloads the plugins. The fourth, /codex:setup, tells you whether Codex is ready to use. If Codex isn't installed at all, setup may offer to install it via npm. To install it yourself, use npm install -g @openai/codex. Once installed, commands starting with /codex should appear in the list.

2:32The codex:codex-rescue subagent also shows up in the /agents list. If /codex:setup says you're not logged in, enter !codex login in the prompt. For a first run, the example is /codex:review --background to review in the background. Then check progress and results with /codex:status and /codex:result. We didn't run the install in this video; we only summarized the official steps.

3:01To change the review model or reasoning effort, set it in config.toml. Config files can live in two places: user level and project level. Project config is only read if you've trusted that project. This plugin doesn't run separately; it works through the local Codex CLI and app server. So it uses the same install, same login, and same repository working tree.

3:25The code the reviewer reads is the exact working tree Claude Code just edited. So what should you ask your reviewer, and how?

3:37The top slot is /codex:review, the most basic read-only review. Bottom line: it doesn't touch the code, it only returns findings. It's described as the same quality as running /review inside Codex. It looks at 2 things: uncommitted changes and your branch versus a base branch. To review the whole branch, add a base, like /codex:review --base main.

4:02If you changed many files, it takes time, so running in the background is recommended. But this command doesn't take text on where to focus. Standard review slot: "Reads pre-commit changes and branches only." /codex:status shows Codex jobs running or recently finished in this repo. When done, /codex:result pulls up the saved final result. If a session ID comes with it, codex resume reopens that run in Codex.

4:30To stop a background Codex job midway, use /codex:cancel. The command from earlier that lets you specify focus in words is /codex:adversarial-review. It challenges your chosen direction itself, more than code details. It picks review targets the same way as standard review, and also doesn't edit code. Use it to question design choices, tradeoffs, hidden assumptions, and alternatives.

4:55The example asks whether the caching and retry design was right. You can also point it at risky areas like auth, data loss, rollback, race conditions, reliability. Per the earlier principle, narrow the text to gaps touching correctness and requirements. For your code, instead of "check everything," say "check if this payment rollback loses data." Adversarial review slot: "Challenges direction and hidden assumptions; I pick the focus."

5:21The review gate is an optional setting managed in /codex:setup. When on, a Codex review runs via a Stop hook when Claude finishes responding. If it finds issues, it blocks the stop so Claude fixes them first. But there's a warning: the back-and-forth can drain usage limits quickly. So it says to enable it only when you plan to watch the session.

5:41To toggle it, add --enable-review-gate or --disable-review-gate after /codex:setup. There's also /codex:rescue to hand fixes to Codex, but today we only use review. Review gate slot: "Automatic review, only while watching." But what if applying the feedback actually breaks the code?

6:06Each time you send a prompt that starts a turn, the prior code state is saved automatically. Bottom line: Claude's file edits can be undone prompt by prompt. Type /rewind, or press Esc twice with an empty prompt, to open the menu. If there's text in the prompt, pressing Esc twice clears it instead. Pick a point in the menu, and you can handle code and conversation separately.

6:30Restore both, conversation only, code only, two summarize options, and cancel: 6 in total. The "restore just the code" from earlier is this code restore. We couldn't verify the bottom row's name in the source. You keep the conversation and return only to the code before the feedback. Summarizing doesn't change files on disk. Rewinding the conversation puts the original prompt from that point back in the input.

6:54Checkpoints are saved with the conversation, so you can rewind even after reopening a session. After rewinding, run /codex:review again to see if the same findings remain. File snapshots are kept for up to the latest 100 checkpoints per session. Snapshots may be deleted after about 30 days, by default, since the session was last saved.

7:14Rewinding after that may fail with No files were restored. Rewind slot: "Restore code and conversation selectively, per prompt." Here are the changes that checkpoints don't cover, as mentioned earlier. Bottom line: only changes made by Claude's file editing tools are tracked. Files changed by Bash commands aren't tracked, so rewind can't recover them. Examples include deleting with rm, moving with mv, or copying with cp.

7:41Except for a foreground forked skill, subagents' edits to other files aren't undone and stay. So it advises rolling those back with git. Your own edits outside Claude Code, or edits from other sessions, usually aren't captured. Files linked by symbolic or hard links are skipped on restore. If Bash commands were part of applying feedback, don't rely on rewind alone.

8:04Can't-undo slot: "Bash, subagents, outside changes." The last slot, git, is what checkpoints don't replace. Checkpoints are designed for quick recovery within a session. The docs say to keep commits, branches, and long history in version control like git. So it makes sense to commit only changes that pass review. Fill the git slot with "Permanent record outside the session," and the map is done.

8:29So under what conditions do these tools fit?

8:35Claude Code implements, then runs pass/fail checks like tests or builds. If it passes, review with Codex, apply the feedback, then review again. If applying it goes wrong, go back to the prior checkpoint. Only commit and merge after it passes. Instead of just asking for an email validator, include pass and fail examples and running tests.

8:56Then Claude runs the tests, reads the results, and fixes until they pass. Instead of "it's done," ask for evidence like test output or commands run. If it's heading the wrong way, stop it with Esc. The context stays intact. If you've corrected the same issue more than 2 times in a session, the context is cluttered with failures.

9:15Then the docs say to /clear and start fresh with a more specific prompt that includes what you learned. If you can't use Codex, you can still set up a separate reviewer inside Claude Code. Subagents run separately with their own context and allowed tools. The security review example checks injection, auth and permission flaws, secrets in code, and unsafe data handling.

9:37But it's a reviewer within the same Claude, so keep it distinct from cross-review with another tool. To document review rules, there's CLAUDE.md. It's a special file Claude reads at the start of every conversation. But CLAUDE.md instructions are advisory; to enforce them, use hooks. Another condition: if the file is too long, instructions may get ignored.

9:58So keep the review steps short, and leave must-run checks to hooks. Let me also be clear about what this video left out. We didn't cover codex exec and AGENTS.md, since we couldn't get official sources this time. We also didn't include numbers on plugin pricing or review effectiveness. We also didn't actually run the review or rewind in this video.

10:18All that's left is picking something to try today.

10:23Installation ends with checking readiness via /codex:setup. For review, have standard review read changes, and if direction worries you, add a narrow prompt to adversarial review. If things break, use /rewind to restore just the code, and handle out-of-checkpoint changes with git. One key number again: checkpoints keep up to the latest 100 per session. Rather than doing it all, try adding /codex:review to your next change first.

10:48Only when the design feels shaky, write the risky areas into adversarial review. If applying feedback breaks things, keep the conversation and just rewind the code. The condition for the review gate: only turn it on when you can watch the session.

11:03The AI that built it stops when it looks done and grades its own answers, so it misses things. If you build code with Claude Code, use a fresh-context reviewer and narrow it to correctness and requirements. What gaps have you found later in code an AI said was done? If you got stuck on installation or review prompts, let us know in the comments too.

11:22If this helped, please like and subscribe. Next time, we'll cover running read-only AI review in CI without a chat.