Fix/truncation strategy

[philipph-askui]

Refactors truncation strategies to align with prompt caching

• makes truncation strategy configurable as parameter of the agent
• removes obsolete factory logic for truncation strategies
• replaces the old SimpleTruncationStrategy (which deleted messages to stay under limits) with a new summarization-based strategy SummarizingTruncationStrategy that use the VLM to compress history instead of discarding it
• adds a new experimental SlidingImageWindowSummarizingTruncationStrategy, that additionally strips old base64 images beyond a configurable window and uses dual cache breakpoints (image-removal boundary + last user message).

For reviewing: I think you don't need to review the experimental SlidingImageWindowSummarizingTruncationStrategy in detail at this point.

[programminx-askui]

This are my first comments.

I need to continue the review.

[programminx-askui]

I assume the description of the image is partly in the messages history or? So the LLM can assume what the App has done.

e.g. I see that an Error Message popped up. ...

[philipph-askui]

the rest of the content in the message history, e.g. the interpretation of the LLM remains unchanged. we just remove the image here. Generating a dedicated description for each image would create additional traffic and cost, hence I decided against that option

[programminx-askui]

Is this Token Consumption tracked in our Report? Please say yes. :-D

[philipph-askui]

https://tenor.com/de/view/homer-simpson-the-simpsons-hiding-gif-13035563

[programminx-askui]

How did you came up this the message?

Can me make the message configurable?

I would add something like. You are an CUA agent how should continue the conversation.. But

[philipph-askui]

addressed in b23dd88

[philipph-askui]

for the exact content: lets keep it simple in the beginning. This simple prompt worked quite well in my test. If we see that the agent loses track after summarizing, we can improve it