fix: preserve additional_kwargs in tool-call messages for reasoning models

[octo-patch]

Fixes #6108

Problem

When using an OpenAI reasoning model (gpt-5, o-series) with a tool assigned to an Agent node, the API returns a 400 error:

Error in Agent node: 400 Item 'msg_xxxxx' of type 'message' was provided without its required 'reasoning' item: 'rs_xxxx'

This happens because when the agent receives an assistant response with tool calls, it pushes a reconstructed message object to the messages array — but omits additional_kwargs. For reasoning models, additional_kwargs contains the reasoning content (and structured reasoning items with IDs like rs_xxxx) that the OpenAI API requires to be present in subsequent API calls alongside their paired assistant messages.

Solution

Include additional_kwargs: response.additional_kwargs when pushing the assistant message in both handleToolCalls and handleResumedToolCalls. This preserves the reasoning items in the conversation history so the next LLM invocation satisfies the OpenAI API constraint.

Testing

• Reproduce with an AgentFlow V2 using a gpt-5 (or o-series) model with reasoning enabled and at least one custom tool
• Before fix: 400 error on the first tool-call round trip
• After fix: agent completes the tool call and returns a response

[gemini-code-assist]

Code Review

This pull request updates the Agent_Agentflow class to include additional_kwargs when pushing LLM responses to the message history. This change ensures that reasoning items from models such as OpenAI's o-series are preserved, preventing 400 errors in subsequent API calls that require this context. I have no feedback to provide as there were no review comments.