This is basically the setup I ended up using after trying to check the following boxes:

- using GPT‑5.6‑sol via OAuth instead of an API key, with Claude Code as the harness

- sandboxing the reviewer agent completely (GPT‑5.6‑sol seems to be a model that believes "the end justifies the means", I read tons of horror stories on X and Reddit)

- using herdr to view and control the current process of the review agent (can jump in at any time)

- the main agent can refute certain points made by the reviewing agent