OP here - happy to answer any questions.

This was one of the more unexpectedly tricky layers of building real-time LLM suggestions, and I’d love to hear how others have approached timing, cancellation, and speculative prediction in their editors or agents.