feat: context-window limits handling

[pkarw]

The context window sizes should be defined by the supported model.
When we get to 80-85% (?) of the context window limit (we can estimate the usage with tiktoken port) there should be a way to handle it, ideas:

• splitting per-agent context rfc: Per-agent context splitting #6
• summarization of previous context chunks
• context-windowing (last X chunks)
• RAG (context windowing + vector store, relevant information retrieval)