How to run read-only code review via external LLMs in Claude Code
This describes a way to run a read-only using an external model (such as GPT-5.6) through Claude Code's Codex plugin, by directly invoking the plugin's companion script from the CLI. The key practice is explicitly forcing which model and level get used, rather than leaving it to defaults. Prompts should include both cases that must pass cleanly and cases designed to fail, as a way to check the review logic actually works.
Execution should be verified by checking directly rather than trusting the model's own claims about what it did. The approach also tackles a subtler problem: telling apart a review that found nothing because it never actually read the code from one that read the code and genuinely found no issues.
Key points
- Invokes the Codex plugin's companion script directly from the CLI for read-only reviews with an external model
- Explicitly forces which model and level the review uses
- Prompts include both cases that should pass and cases that should fail, to test the review logic
- Verifies actual execution through instead of the model's self-reported claims
- Distinguishes a review that read nothing from one that read the code and found no issues