feat(provider/amazon-bedrock): Add reasoning support

[Und3rf10w]

This adds reasoning support for Claude models in Amazon Bedrock, as noted in issue #4977.

I have tested these changes in my environment, and confirmed that we receive reasoning in both streaming and non-streaming configurations.

In streaming and non-streaming, both the signature and reasoning text are properly extracted. You may want to convert this to use providerOptions instead of relying on additionalModelRequestFields in BedrockChatSettings, but I don't know the best practice for that.

[shaper]

Thank you! I will take a look at this soon.

[Und3rf10w]

After some testing (from e7aeec0), it looks like something goes wrong when you have tool use involved with reasoning enabled.

I can at least get this error back when I'm doing a generation with a tool enabled, but not sure how to resolve this:

The model returned the following errors: messages.1.content.0.type: Expected thinking or redacted_thinking, but found text. When thinking is enabled, a final assistant message must start with a thinking block (preceeding the lastmost set of tool_use and tool_result blocks). We recommend you include thinking blocks from previous turns. To avoid this requirement, disable thinking.

EDIT: Resolved in 31bef7c

The issue was I wasn't adding the reasoningContent part in convertToBedrockChatMessages. I can confirm this now works in all the cases I have tested.

[david-arteaga]

For what it's worth I took this patch and manually applied it locally and it works with sonnet 3.7 with both generateText and streamText

[shaper]

This is on my radar and I aim to get to it soon, thank you for your continued work here and contributions. We plan to work with you to get this landed as soon as we can.

[Und3rf10w]

This has literally been my experience with trying to write this, I swear:

I feel like every time I think I have it done right, I do some testing, then experience some random edge case and find I did something wrong.

---

@shaper, do you want to populate reasoningDetails[] on a response from doGenerate that includes a tool invocation?

For example, right now, as of 468a750, if we call generateText without any tools, then the reasoningDetails comes back populated:

However, if we call generateText with tools, then it will be empty, but can be extracted from the steps or response messages:

And later in the response.messages:

I genuinely have no idea how that's supposed to return, but something important to note.