One agent writes. Another reviews in a loop.

Review changes you're already working on#

If Claude Code has changes ready for review, send this prompt to the Claude Code session that wrote them:

Use Facade to start Codex with gpt-6-sol and high effort to review the changes in this project folder without editing files.
Fix valid issues and have it recheck in a loop until clean.

Claude starts Codex in the same project folder. Codex sends its findings to Claude. Claude fixes valid issues and asks for another review.

Start a new task with an automatic review loop#

To start a new task with review, use this example setup:

Step 1. Launch and configure Claude Code

In VS Code, open the Facade Code panel in the sidebar, click New Agent, and select Claude Code. Confirm that the working directory is your project folder.

Set the model with /model opus and effort with /effort high.

Step 2. Send your task with review instructions

Send this prompt to Claude Code, replacing the first line with your task:

[Describe your task and the expected result.]

Implement it, then use Facade to start Codex with gpt-6-sol and high effort to review the changes in this project folder without editing files.
Fix valid issues and have it recheck in a loop until clean.

What happens next#

1. Claude implements
2. Codex reviews
3. Claude fixes
4. Codex rechecks (loops until clean)

Claude uses Facade to start Codex in the same project folder. Codex checks the changes against your requirements without editing files. It sends its findings directly to Claude.

Claude fixes valid issues and asks Codex to recheck until no valid issues remain. Claude then summarizes the changes, review findings, checks run, and anything still unverified.

You can open the Codex session at any time to clarify the requirements or discuss its review.

Reuse a long-running reviewer across features

You can also send work to an existing Codex reviewer. Keeping the same reviewer lets you build on earlier discussions about the project and review criteria.

Name the existing reviewer in your prompt, using its name (such as @Kite) or its #aN ID:

[Describe your task and the expected result.]

Implement it, then use Facade to message @Kite to review the changes in this project folder without editing files.
Fix valid issues and have it recheck in a loop until clean.

Pin the reviewer in the Facade panel to return to it across features. When you reopen the project, Facade restores its saved CLI conversation where supported.

Why separate implementation and review#

• Catch overlooked issues. A second agent checks the changes against your requirements and may find issues the writer missed.
• Fixes get another review. Claude sends the fixes back to Codex, so the review covers the updated code as well as the original changes.
• Agents exchange feedback directly. Facade delivers the messages between writer and reviewer, saving you from relaying their feedback.
• Different models can offer another perspective. Models from different providers may spot different issues. Check their findings against the code and tests.

Review adds time and model usage. A single agent is enough for small changes; use this pattern when the work warrants a separate review.

Use it in your project#

• To change the writer, select another installed CLI in New Agent and set its model and effort with that CLI’s commands.
• To change the reviewer, replace Codex, gpt-6-sol, and high effort in the prompt with a combination you can use.
• Both roles can use the same CLI or different CLIs. You can also use the same model in separate sessions.

For saved launch settings and other providers, see Agent settings, Claude Code with another provider, and Codex with another provider.