OptionalatHow full the context window has to get before the conversation is compacted, as a fraction between 0 and 1. Defaults to 0.8.
Measured against the model's window, which has to be known: a model whose window neither the provider nor the catalog reports is never auto-compacted, because there is nothing to take a fraction of.
OptionalkeepHow many trailing turns to keep intact when compacting. Defaults to 2. The first turn is always kept as well — it is where the task was set.
When to shorten a conversation that is filling the context window, and how much to keep.
Example: ```javascript B.ai.call({ message: B.net.request.parameter("q"), autoCompact: { atRatio: 0.75, keepLastTurns: 4 }, }); ```