Convert CoreMessage to UIMessage where it makes sense

[ianmacartney]

When interacting with AI SDK frontend code, there's a preference for UIMessage. There may also be benefits to passing that into the AI SDK (Mastra mentioned the AI SDK choking on "role: tool" messages).

It's mostly a matter of rearranging messages into a friendlier format, with a few differences:

• The id field which is assigned by AI SDK (currently populated for response messages but not for the initial user message).
• The sources field which is in our steps model but not on each message.
• The annotations field which I'm not sure the origin of (it mentions StreamData so maybe it's an internal signal for marking the start/end/segment of streaming text/objects)

Some references:
vercel/ai#3161 (comment)
https://github.com/mastra-ai/mastra/pull/3230/files

[ianmacartney]

If we used UIMessage instead of CoreMessage, the only things we'd miss that I can see are extra metadata like providerOptions, experimental_providerMetadata, experimental_content, and tool result error status.

I suspect we could always go from a UIMessage to a CoreMessage, so it might be worth having that be the primitive value, then steps would only differentiate by having:

• finishReason
• usage
• warnings
• request & response http metadata (headers, ids, timestamp, modelId)

[ianmacartney]

Strategy for now:

• keep saving core message
• save other fields on the high-level message
• have a function to convert them to UIMessage
• consider in the future passing UIMessage to the AI SDK internally instead of core messages