Fix MistralAiApi to handle thinking content blocks

[KyleKreuter]

Fixes #5019

Summary

Mistral's Magistral reasoning models return content as an array of content blocks
instead of a plain string.

Changes

• Update content() to extract the text block from thinking responses

Tested with

• Model: magistral-medium-2509
• Response format verified with direct API call

[nicolaskrier]

Hello @KyleKreuter, thanks for your contribution!

Here are a few suggestions to improve Magistral support:

• Add a MistralAiAssistantMessage class (like ZhiPuAiAssistantMessage) to handle both content and reasoning fields.
• Replace List/Map in MistralAiApi with dedicated classes: ThinkChunk, ReferenceChunk, and TextChunk (Mistral AI API documentation).
• Support the reasoning prompt mode as described in the API documentation (not related to the current issue).
• Add unit/integration tests to validate these features.

Happy to discuss or clarify if needed!

[KyleKreuter]

Hi, thanks for the feedback!

I've implemented the suggested improvements:

1. MistralAiAssistantMessage class

• Created MistralAiAssistantMessage extending AssistantMessage with a thinkingContent field
• Follows the same pattern as ZhiPuAiAssistantMessage and DeepSeekAssistantMessage
• Includes Builder pattern, proper equals()/hashCode()/toString() implementations

2. Dedicated chunk classes

• Added sealed interface ContentChunk with three implementations:
  • TextChunk - for text content blocks
  • ThinkChunk - for thinking/reasoning content blocks
  • ReferenceChunk - for citation references
• Updated ChatCompletionMessage with content(), thinkingContent(), and contentChunks() methods
• Note: The Mistral API returns thinking content in a nested format ({"type": "thinking", "thinking": [{"type": "text", "text": "..."}]}), which is now properly parsed

4. Unit/Integration tests

• MistralAiAssistantMessageTests - 23 test cases for the message class
• MistralAiContentParsingTests - 25 test cases for content parsing
• MistralAiMagistralIT - 6 integration tests with real Magistral API calls

Please let me know if you have any further suggestions.