Streaming response
Server-sent event-style delivery of LLM output token-by-token. Lets the UI render as the model writes, collapsing time-to-first-token.
SumTube's side panel consumes an SSE stream with 15-second heartbeats, rendering in `meta → oneLiner → bullets → chapters` order with sub-second TTFT. Users don't wait the full 30 seconds for the complete summary — they can decide whether to watch the video within the first few rendered events.