Leverage context caching provided by models to minimize token usage and increase composition speed.

Use the same conversation ID from the original auto-composition on subsequent recomposes. This may mean that we don't need to send diffs again.