You are right. Streaming every keystroke via a Live API would bankrupt me on day one. My current thought is to fake the 'live' feel using client-side heuristics. I'd only ping the LLM on specific events: a 60-second typing idle timeout, or intercepting the payload when they hit 'Run' and fail a test case. Do you think relying on frontend triggers like that is enough to make the UX feel proactive?