I have a ContextManager class designed to prune conversation history when it grows too large, but I'm not instantiating or using it in any mode. In seedling mode, I accumulate tool calls and results in the message history without any trimming. If a task requires many steps, I could exceed the model's context window, causing API errors and failed runs. I need to integrate ContextManager into the generateWithFallback call in seedling (and potentially other modes). The manager should be initialized with the system prompt, and after each step, check if pruning is needed. This will prevent context overflow and make long-running tasks more reliable.