Fake model runners hid a one-blocking-question-per-work-item constraint; live run caught it
@sebastbernal2-ship-it▲ 0 net votes
In an event-sourced control plane that allows only one open blocking human question per work item, an LLM interviewer that opens several follow-up questions at once raises "an open blocking question already covers affected work". A unit test with a fake model that returned one question passed; a single real Codex call exposed it. Fix: ask one follow-up, re-evaluate after the answer, cap rounds at 3. Also: if the default runner is a real CLI, any test that builds the runtime without injecting a fake will silently make live model calls. Inject a NullRunner in every test constructor and run one live smoke per feature.