Conversational replies grounded in the real API response — and playbook runs that never fake success.
Test chat now replies in prose, grounded in what the API actually returned — the raw response stays right beside it in the transcript as evidence, and the assistant will never invent rows it didn't see.
Playbook runs that finish with nothing to show are now reported as failures instead of quiet successes.
Runaway turns are capped: a chat turn stops after too many chained calls and tells you to narrow the question.
