Cast message to MessageType before creating StreamChunk in stream_broadcaster

[TheRealNeil]

Cast message to MessageType before creating StreamChunk in stream_broadcaster

Fixes #307

Summary

• Cast raw Hash messages to MessageType in stream_broadcaster so that chunk.message has the same typed interface as response.message

Root cause

During streaming, StreamChunk#message wraps the raw Hash from the provider's message stack — no type casting is applied. After generation completes, PromptResponse#messages are cast through Types::MessagesType into Common::Messages::* objects with method accessors (.content, .role).

This means chunk.message.content raises NoMethodError during streaming, while response.message.content works fine after generation. Users have to use chunk.message[:content] during streaming but response.message.content afterwards — an inconsistent API.

Fix

In stream_broadcaster, cast the message through Providers::Common::Messages::Types::MessageType when it's a Hash before passing it to StreamChunk.new. This gives chunk.message the same typed interface as response.message, so .content and .role work consistently in both streaming and non-streaming contexts.

Test plan

• Verify chunk.message.content works in on_stream callbacks
• Verify chunk.message.role works in on_stream callbacks
• Confirm non-Hash messages are passed through unchanged
• Confirm post-generation response.message.content is unaffected

[Copilot]

Pull request overview

This PR fixes an inconsistency in the streaming API by ensuring StreamChunk#message exposes the same typed interface as post-generation responses, so callers can use method accessors like .content/.role during streaming.

Changes:

• Cast streamed message values that are Hashes through Providers::Common::Messages::Types::MessageType before constructing StreamChunk.
• Preserve pass-through behavior for non-Hash messages.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

There are streaming callback tests, but none currently assert the new behavior that Hash messages are cast to a typed message with .content/.role during streaming. Add a test that calls stream_broadcaster with a Hash message and asserts chunk.message is cast (and that non-Hash messages still pass through unchanged) to prevent regressions.

[Copilot]

Providers::Common::Messages::Types::MessageType.new is instantiated for every streamed chunk. Since stream_broadcaster may be called many times per request, this creates avoidable allocations; consider memoizing the MessageType instance (e.g., build it once per stream_broadcaster call and close over it, or store it in an ivar) and reuse it for each cast.

[Copilot]

Casting Hash messages through Providers::Common::Messages::Types::MessageType drops provider-specific keys (e.g., OpenAI Chat streaming messages can include :tool_calls, and the common MessageType slices assistant hashes down to :role, :content, :name). This is an API change for streaming callbacks that currently can inspect the full provider message hash; consider preserving the original Hash alongside the typed wrapper (or otherwise exposing raw fields) so streaming users don’t lose access to metadata/tool-call details.

Suggested change

	cast_message = message.is_a?(Hash) ? Providers::Common::Messages::Types::MessageType.new.cast(message) : message
	self.stream_chunk = StreamChunk.new(cast_message, delta)
	self.stream_chunk = StreamChunk.new(message, delta)