Sphinx uses a reviewer model isolated from the coding conversation to inspect committed changes and ask about behavior, failure conditions and invariants. Developers can consult the code while answering, then confirm before continuing a Git push. The aim is to practice explaining AI-written changes. Its configuration documentation prevents repositories from changing the model provider or raising usage limits. Model calls may incur costs and transmit code, and the check can be bypassed, so passing it is not a substitute for security review or tests.