Nothing particularly insightful other than avoiding messing with previous messages so as not to mess with the cache.

Summary by Claude:

    Codex works by repeatedly sending a growing prompt to the model, executing any tool calls it requests, appending the results, and repeating until the model returns a text response