Support follow-ups on previous tasks

The reference agent does not allow for following-up on previous tasks, because the reference agent only sends the last message to the LLM. We could tweak things as follows:

• To decide if it's a task for MOSAICO agents, we can send over only the human messages.
• To actually produce output, we send over the full chat history.

It should be possible, for example, to have this interaction:

• Generate a metamodel for a pet shop.
• (... have the reference agent product a solution ...)
• Add a concept for the Shop.
• (... reference agent gets the previous attempt and produces an updated version ...)