Can Tool-augmented Large Language Models be Aware of Incomplete Conditions?