.NET: OpenAI-compatible extra body field thinking is not forwarded when using Microsoft.Agents.AI.OpenAI

[ssccinng]

Summary

When using Microsoft.Agents.AI.OpenAI with an OpenAI-compatible backend (ARK), provider-specific body field thinking is not forwarded to the final outgoing payload.

Expected outgoing request should be:

curl https://ark.cn-beijing.volces.com/api/v3/responses \
  -H "Authorization: Bearer $ARK_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
      "model": "doubao-seed-2-0-lite-260215",
      "input": "常见的十字花科植物有哪些？",
      "thinking":{"type": "enabled"},
      "stream": true
  }'

Minimal repro code

using System;
using System.ClientModel;
using System.Text.Json;
using Microsoft.Agents.AI;
using Microsoft.Extensions.AI;
using OpenAI;

var client = new OpenAIClient(
    new ApiKeyCredential(Environment.GetEnvironmentVariable("ARK_API_KEY")!),
    new OpenAIClientOptions { Endpoint = new Uri("http://127.0.0.1:24424") }); // local proxy

var chatClient = client.GetChatClient("doubao-seed-2-0-lite-260215");
var thinkingConfig = JsonDocument.Parse("{\"type\":\"enabled\"}").RootElement.Clone();

var chatOptions = new ChatOptions
{
    Instructions = "You are a helpful assistant",
    AdditionalProperties = new AdditionalPropertiesDictionary { ["thinking"] = thinkingConfig }
};

var agent = chatClient.AsAIAgent(new ChatClientAgentOptions { ChatOptions = chatOptions });

var runOptions = new AgentRunOptions
{
    AdditionalProperties = new AdditionalPropertiesDictionary { ["thinking"] = thinkingConfig }
};

var session = await agent.CreateSessionAsync();
await agent.RunAsync("常见的十字花科植物有哪些？", options: runOptions, session: session);

Actual

Intercepted outgoing JSON (via local proxy) contains standard fields (messages/model/temperature/reasoning_effort), but does not include thinking.

Expected

A supported way to pass through provider-specific extra body fields for OpenAI-compatible endpoints (for example an ExtraBody map or documented passthrough mechanism).

Question

Is this behavior by design, or a bug/feature gap?
If by design, what is the recommended way to send provider-specific request fields such as thinking?

[strikene]

The same issue applies: any value specified in ChatClientAgentOptions.ChatOptions.AdditionalProperties will ultimately not become part of the request payload.
In version 1.0, if you try to modify ChatOptions.Reasoning.Effort while using an OpenAI-compatible backend, you can still see this parameter included in the request. In some cases, it even causes request errors — for example, LM Studio throws an error saying XXX model does not support ExtraHigh. This indicates that the Reasoning parameters are still being sent in the request.
Starting from version 1.5, any modifications to ChatClientAgentOptions.ChatOptions.Reasoning.Effort or ChatClientAgentOptions.ChatOptions.Reasoning.Output are no longer reflected in the actual request payload. I built a mock OpenAI server specifically for inspecting this kind of request traffic to verify this behavior.

Microsoft.Agents.AI="1.5.0"
Microsoft.Agents.AI.DevUI="1.5.0-preview.260507.1"
Microsoft.Agents.AI.Hosting.A2A="1.5.0-preview.260507.1"
Microsoft.Agents.AI.Hosting.A2A.AspNetCore="1.5.0-preview.260507.1"
Microsoft.Agents.AI.Hosting.OpenAI="1.5.0-alpha.260507.1"
Microsoft.Agents.AI.OpenAI="1.5.0"
Microsoft.Extensions.AI="10.5.2"
Microsoft.Extensions.AI.OpenAI="10.5.2"