Gemini no longer does a per-response cleanup that could race when multiple responses were in flight at once. Removing it clears a source of intermittent failures under concurrent usage, so agents hitting Gemini from many requests in parallel behave reliably instead of tripping over shared state.
View the Gemini docs for more.