docs: add compaction concept
This commit is contained in:
@@ -48,7 +48,7 @@ Pruning uses an estimated context window (chars ≈ tokens × 4). The window siz
|
||||
|
||||
## Interaction with other limits
|
||||
- Built-in tools already truncate their own output; session pruning is an extra layer that prevents long-running chats from accumulating too much tool output in the model context.
|
||||
- Compaction is separate: compaction summarizes and persists, pruning is transient per request.
|
||||
- Compaction is separate: compaction summarizes and persists, pruning is transient per request. See [/concepts/compaction](/concepts/compaction).
|
||||
|
||||
## Defaults (when enabled)
|
||||
- `keepLastAssistants`: `3`
|
||||
|
||||
Reference in New Issue
Block a user