Prewarm research context while the user types

Goal

Reduce time to first token when large research context is reused.

Steps

1. When the user focuses the input or starts typing, issue a background one-token cache-warming request using the same context that will be submitted. [source13]
2. After submission, send the real request and inspect cache_read_input_tokens; keep the warm-up and final context identical and scoped to the same session. [source13]
3. Compare TTFT and cost with ordinary caching; stop warming after an idle timeout and never share private context across users. [source13]

How to verify

The real request records a cache read for the same large context, reduces TTFT against the cold baseline, and shows no cross-user data reuse.

Troubleshooting

If there is no cache hit, compare the warm-up and final context byte-for-byte and check TTL; small contexts may not benefit, so avoid warming abandoned queries.