All AI model are not 100% deterministic. So you can't exactly define why and what. Thinking model and other are a way to understand this non-deterministic nature.

For your tools, try to provide custom wrapper or build your own tool on top of existing tool like GoogleSearch. With this you can have control to know what, what and how a Tools is used.

If you use Google ADK (build on top of Google GenAI SDK), you have https://google.github.io/adk-docs/callbacks/types-of-callbacks/#tool-execution-callbacks to control and log these responses