Model routing which takes into account task and current state of the session

We use the 2.5 pro model for all tasks and this might be overkill for some of the simpler tasks. Task specific routing will also save users pro quota for use in more complex tasks.

**Identify the token behavior of Gemini CLI
**Start with the flash model by default and switch over to the pro model when dealing with large context or cognitive loops.
**Thinking of implementing routing in a slightly different fashion.
1. Plan: We should start with pro to identify how we can break down the task into smaller steps. Each step can use flash or pro depending on the complexity of the task.
2. Execute: Use flash for execution. I would suspect this should suffice for most of the steps if they are small and well contained.