Interview tool: epics, which agent runs it, and the 0-questions bug

[mhenke]

Findings: Two threads about the interview tool. One proposes organizing interview questions into epics (auth, deployment, and so on) with per-epic progress and the ability to drill into one epic. The other reports a bug (the interview ended after "0 questions" and left the UI waiting on browser input) and debates which agent should run the interview (orchestrator versus librarian for web access and speed versus the smartest model), plus whether it should run in a fresh session.

What was said:

• @alvinunreal (issue Interview tool; part 2 #271): proposed grouping interview questions into epics like authentication and production deployment, showing how many questions each epic has and letting a user pick one epic to deepen its spec. No further discussion yet.
• @alvinunreal (issue Interview no more questions #297): the interview feels odd in Orchestrator mode and should probably run in a fresh session; wondered if the librarian (better web access, fast) should run it.
• @rustanacexd (issue Interview no more questions #297): the smartest model should own the interview as the entry point of an idea, and suggested the orchestrator use the librarian for research (ExaSearch) during the interview.

Open questions:

1. Should the interview group questions by epic instead of a flat list, and what does "clarified per epic" look like?
2. Which agent should own the interview: orchestrator, librarian, or the smartest available model, and should it run in a fresh session?
3. How should the interview signal "done" so the UI doesn't gray out waiting for input that never comes (the 0-questions bug)?

Original issues: #271 and #297