Allow to set reasoning_effort and capture reasoning content

[nforro]

Description

For models with thinking enabled, LiteLLM exposes the thoughts as reasoning_content. This PR captures it within AssistantMessage. To be able to enable/control model thinking, ChatModelOptions is extended with reasoning_effort.

This can be useful especially when not using/forcing the Think tool.

Which issue(s) does this pull-request address?

Closes: #1448

Checklist

General

• I have read the appropriate contributor guide: Python
  / TypeScript
• I have signed off on my commit: Python instructions / TypeScript instructions
• Commit messages and PR title follow conventional commits
• Appropriate label(s) added to PR: Python for Python changes, TypeScript for TypeScript changes

Code quality checks

• Code quality checks pass: mise check (mise fix to auto-fix)

Testing

• Unit tests pass: mise test:unit
• E2E tests pass: mise test:e2e
• Tests are included (for bug fixes or new features)

Documentation

• Documentation is updated
• Embedme embeds code examples in docs. To update after edits, run: Python mise docs:fix

[gemini-code-assist]

Code Review

This pull request introduces support for model reasoning content by adding a MessageReasoningContent type and updating AssistantMessage to include reasoning blocks. It also adds a reasoning_effort parameter to chat model options and enhances the LiteLLM adapter to handle reasoning and Anthropic-specific thinking_blocks. Review feedback identifies a logic error in the LiteLLM adapter where metadata could be lost if content parts are empty, suggests making text and tool calls non-exclusive, and recommends adding missing imports for improved type safety.

[nforro]

/gemini review

[gemini-code-assist]

Code Review

This pull request introduces support for model reasoning by adding MessageReasoningContent to the message framework and a reasoning_effort option to the chat model. It also updates the LiteLLM adapter to handle reasoning content and thinking_blocks for conversation history. A potential issue was identified in the LiteLLM adapter where the use of model_dump() might filter out necessary metadata like thinking_blocks if they are not explicitly defined as model fields; simplifying the condition to check for the existence of the update object is recommended.

[jpodivin]

This would partly address the #1448, but we would still lose the content if a tool_call was present in the delta. Could you integrate relevant parts of main...jpodivin:beeai-framework:fix/reasoning-content-drop into this?

[nforro]

This would partly address the #1448, but we would still lose the content if a tool_call was present in the delta.

I assume that issue has been resolved when addressing this feedback from gemini-code-assist:
#1447 (comment)

Could you integrate relevant parts of main...jpodivin:beeai-framework:fix/reasoning-content-drop into this?

Let me at least include the tests.

[nforro]

/gemini review

[gemini-code-assist]

Code Review

This pull request introduces support for model reasoning and 'thinking blocks' by adding MessageReasoningContent, updating AssistantMessage to handle reasoning parts, and including reasoning_effort parameters. The LiteLLM adapter is updated to extract reasoning content and manage signed thinking blocks for Anthropic models. Feedback suggests adding a type check for thinking_blocks in the LiteLLM adapter to prevent potential runtime errors during iteration.

[Tomas2D]

Great improvement. I am just worried whether LiteLLM handles the parameter for providers where thinkink_blocks attribute is not supported. Did you test it? If not I can do that.

[nforro]

I am just worried whether LiteLLM handles the parameter for providers where thinkink_blocks attribute is not supported. Did you test it? If not I can do that.

Good point, I didn't think of that. I didn't test it.

[jpodivin]

I am just worried whether LiteLLM handles the parameter for providers where thinkink_blocks attribute is not supported. Did you test it? If not I can do that.

Good point, I didn't think of that. I didn't test it.

Perhaps we could implement something in littellm.

[nforro]

Sorry, I'm not following, implement what in LiteLLM? I think I misunderstood the comment - are we talking about setting thinking_blocks attribute in assistant messages in _transform_input()? That should be stripped by exclude_none() if there are no thinking blocks in meta, and thinking blocks can only be there if we got them from an Anthropic chat model in the first place.