Context overflow protection

[ahenkes1]

It would be nice to have context overflow protection to prevent the IDE from crashing or the workflow from getting stuck.

Sometimes the model tries to (accidentally) read a very long file, e.g., an extensive log. This can easily result in >3M tokens. In order to recover from this and stop Roo from trying to condense the context (which might not be possible) and protect from exploding costs, the following feature would be helpful:

1. Detect too large context.

   This is already there in the auto-condense feature.

2. Automatically set a recovery point in order to be able to roll back JUST before the problematic function call.

This shouldn't be too big of a change, as the functionallity for all of this is there. I am curious about your opinion!