FunctionInvokingChatClient returns 400 Bad Request with OpenRouter Gemini 3 Preview models (reasoning_details not preserved)

[vbandi]

Description

When using FunctionInvokingChatClient with OpenRouter's Gemini 3 Preview models (google/gemini-3-flash-preview, google/gemini-3-pro-preview), tool call results fail with HTTP 400 Bad Request.

The root cause is that OpenRouter requires reasoning_details to be preserved and passed back when sending tool results to Gemini 3 reasoning models. The FunctionInvokingChatClient doesn't preserve this field when constructing the tool result message.

From OpenRouter's documentation:

Reasoning Details must be preserved when using multi-turn tool calling.

Reproduction Steps

Save and run this single-file .NET 10 script:

#!/usr/bin/env dotnet run
#:package Microsoft.Extensions.AI@10.1.1-preview.1.25612.2
#:package Microsoft.Extensions.AI.OpenAI@10.1.1-preview.1.25612.2
#:package OpenAI@2.8.0

using System.ClientModel;
using System.ComponentModel;
using Microsoft.Extensions.AI;
using OpenAI;

var apiKey = Environment.GetEnvironmentVariable("OPENROUTER_API_KEY") 
    ?? throw new InvalidOperationException("Set OPENROUTER_API_KEY environment variable");

var model = "google/gemini-3-flash-preview";  // FAILS with 400
// var model = "google/gemini-2.5-flash";     // WORKS

var options = new OpenAIClientOptions { Endpoint = new Uri("https://openrouter.ai/api/v1") };

var client = new OpenAI.Chat.ChatClient(model, new ApiKeyCredential(apiKey), options)
    .AsIChatClient()
    .AsBuilder()
    .UseFunctionInvocation()
    .Build();

var chatOptions = new ChatOptions { Tools = [AIFunctionFactory.Create(GetCurrentTime)] };
var messages = new List<ChatMessage> { new(ChatRole.User, "What time is it?") };

var response = await client.GetResponseAsync(messages, chatOptions);
Console.WriteLine(response.Text);

[Description("Gets the current time")]
static string GetCurrentTime() => DateTime.Now.ToString("HH:mm:ss");

Run with:

export OPENROUTER_API_KEY=your_key_here
dotnet run GeminiToolCallRepro.cs

Expected behavior

The tool call completes successfully and returns the time.

Actual behavior

ERROR: 400 - Service request failed.
Status: 400 (Bad Request)

The error occurs after:

1. Initial request sent with tool definition
2. Model responds with tool call request (including reasoning_details)
3. FunctionInvokingChatClient executes the tool and sends result back
4. 400 error because reasoning_details wasn't preserved in the follow-up request

Configuration

• .NET 10.0 Preview
• Microsoft.Extensions.AI 10.1.1
• Microsoft.Extensions.AI.OpenAI 10.1.1-preview.1.25612.2
• OpenAI 2.8.0
• OpenRouter API (OpenAI-compatible endpoint)
• Models affected: google/gemini-3-flash-preview, google/gemini-3-pro-preview

Related

• PR Add TextReasoningContent.ProtectedData #6784 added TextReasoningContent.ProtectedData for storing reasoning data, but FunctionInvokingChatClient doesn't automatically preserve it during tool call loops
• OpenRouter docs: https://openrouter.ai/docs/use-cases/reasoning-tokens

[vbandi]

Workaround: Custom PipelinePolicy to preserve reasoning_details

I found a workaround for this issue by using a custom PipelinePolicy that intercepts requests/responses to capture and re-inject the reasoning_details field.

The Problem

OpenRouter returns reasoning_details (containing encrypted thought signatures) in the assistant message when using Gemini 3 reasoning models. These must be preserved and sent back with subsequent tool result messages, but FunctionInvokingChatClient doesn't know about this provider-specific field.

The Solution

Add a pipeline policy to OpenAIClientOptions that:

1. Captures reasoning_details from responses (keyed by tool_call_id)
2. Injects them back into outgoing requests containing tool results

using System.ClientModel;
using System.ClientModel.Primitives;
using System.Text.Json.Nodes;

/// <summary>
/// A pipeline policy that preserves OpenRouter's reasoning_details field for
/// Gemini 3 Preview and other reasoning models that require thought signatures.
/// </summary>
public class ReasoningDetailsPolicy : PipelinePolicy
{
    private readonly Dictionary<string, JsonNode> _reasoningDetailsByCallId = new();
    
    public override void Process(PipelineMessage message, IReadOnlyList<PipelinePolicy> pipeline, int currentIndex)
    {
        ProcessRequest(message);
        ProcessNext(message, pipeline, currentIndex);
        ProcessResponse(message);
    }
    
    public override async ValueTask ProcessAsync(PipelineMessage message, IReadOnlyList<PipelinePolicy> pipeline, int currentIndex)
    {
        ProcessRequest(message);
        await ProcessNextAsync(message, pipeline, currentIndex);
        ProcessResponse(message);
    }
    
    private void ProcessRequest(PipelineMessage message)
    {
        if (message.Request?.Content == null) return;
        
        try
        {
            using var stream = new MemoryStream();
            message.Request.Content.WriteTo(stream, default);
            stream.Position = 0;
            
            var requestJson = JsonNode.Parse(stream);
            if (requestJson == null) return;
            
            var messagesArray = requestJson["messages"]?.AsArray();
            if (messagesArray == null) return;
            
            bool modified = false;
            
            foreach (var msg in messagesArray)
            {
                if (msg?["role"]?.GetValue<string>() != "assistant") continue;
                
                var toolCalls = msg["tool_calls"]?.AsArray();
                if (toolCalls == null || toolCalls.Count == 0) continue;
                
                foreach (var toolCall in toolCalls)
                {
                    var callId = toolCall?["id"]?.GetValue<string>();
                    if (callId != null && _reasoningDetailsByCallId.TryGetValue(callId, out var reasoningDetails))
                    {
                        if (msg["reasoning_details"] == null)
                        {
                            msg["reasoning_details"] = reasoningDetails.DeepClone();
                            modified = true;
                        }
                        break;
                    }
                }
            }
            
            if (modified)
            {
                var newContent = requestJson.ToJsonString();
                message.Request.Content = BinaryContent.Create(BinaryData.FromString(newContent));
            }
        }
        catch { }
    }
    
    private void ProcessResponse(PipelineMessage message)
    {
        if (message.Response?.Content == null) return;
        
        try
        {
            var responseJson = JsonNode.Parse(message.Response.Content.ToString());
            if (responseJson == null) return;
            
            var choices = responseJson["choices"]?.AsArray();
            if (choices == null) return;
            
            foreach (var choice in choices)
            {
                var msgNode = choice?["message"];
                if (msgNode == null) continue;
                
                var reasoningDetails = msgNode["reasoning_details"];
                if (reasoningDetails == null) continue;
                
                var toolCalls = msgNode["tool_calls"]?.AsArray();
                if (toolCalls != null)
                {
                    foreach (var toolCall in toolCalls)
                    {
                        var callId = toolCall?["id"]?.GetValue<string>();
                        if (callId != null)
                        {
                            _reasoningDetailsByCallId[callId] = reasoningDetails.DeepClone();
                        }
                    }
                }
            }
        }
        catch { }
    }
}

Usage

var options = new OpenAIClientOptions
{
    Endpoint = new Uri("https://openrouter.ai/api/v1"),
};

// Add the policy for reasoning models
options.AddPolicy(new ReasoningDetailsPolicy(), PipelinePosition.PerCall);

var client = new OpenAI.Chat.ChatClient("google/gemini-3-flash-preview", new ApiKeyCredential(apiKey), options)
    .AsIChatClient()
    .AsBuilder()
    .UseFunctionInvocation()  // Works now!
    .Build();

This workaround works transparently with FunctionInvokingChatClient and doesn't require any changes to the calling code.

[stephentoub]

This isn't specific to Microsoft.Extensions.AI, is it? Can repro just using OpenAI's ChatClient directly if you do the function invocation manually and don't do anything special for that reasoning property?

[vbandi]

This isn't specific to Microsoft.Extensions.AI, is it? Can repro just using OpenAI's ChatClient directly if you do the function invocation manually and don't do anything special for that reasoning property?

Correct. Here is the OpenAI only repro:

#!/usr/bin/env dotnet run
#:package OpenAI@2.8.0

using System.ClientModel;
using OpenAI;
using OpenAI.Chat;

// Direct OpenAI SDK repro - no Microsoft.Extensions.AI
// Manually handle tool invocation WITHOUT preserving reasoning_details

var apiKey = Environment.GetEnvironmentVariable("OPENROUTER_API_KEY") 
    ?? throw new InvalidOperationException("Set OPENROUTER_API_KEY environment variable");

var model = "google/gemini-3-flash-preview";  // FAILS with 400
// var model = "google/gemini-2.5-flash";     // WORKS

Console.WriteLine($"Testing model: {model}");
Console.WriteLine(new string('-', 50));

var options = new OpenAIClientOptions
{
    Endpoint = new Uri("https://openrouter.ai/api/v1")
};

var client = new ChatClient(model, new ApiKeyCredential(apiKey), options);

// Define the tool
var tool = ChatTool.CreateFunctionTool(
    functionName: "GetCurrentTime",
    functionDescription: "Gets the current time",
    functionParameters: BinaryData.FromString("{\"type\":\"object\",\"properties\":{},\"required\":[]}")
);

var messages = new List<ChatMessage>
{
    new UserChatMessage("What time is it right now?")
};

var chatOptions = new ChatCompletionOptions();
chatOptions.Tools.Add(tool);

try
{
    Console.WriteLine("Step 1: Sending initial request with tool...");
    var response = await client.CompleteChatAsync(messages, chatOptions);
    
    Console.WriteLine($"  Finish reason: {response.Value.FinishReason}");
    
    if (response.Value.FinishReason == ChatFinishReason.ToolCalls)
    {
        Console.WriteLine($"  Tool calls: {response.Value.ToolCalls.Count}");
        
        // Add assistant message (with tool calls)
        messages.Add(new AssistantChatMessage(response.Value));
        
        // Execute each tool call and add results
        foreach (var toolCall in response.Value.ToolCalls)
        {
            Console.WriteLine($"  Executing tool: {toolCall.FunctionName}");
            var result = DateTime.Now.ToString("HH:mm:ss");
            Console.WriteLine($"  Result: {result}");
            
            // Add tool result - NOT preserving any reasoning_details
            messages.Add(new ToolChatMessage(toolCall.Id, result));
        }
        
        Console.WriteLine("\nStep 2: Sending tool results back (without reasoning_details)...");
        var finalResponse = await client.CompleteChatAsync(messages, chatOptions);
        
        Console.WriteLine($"Success! Response: {finalResponse.Value.Content[0].Text}");
    }
    else
    {
        Console.WriteLine($"Response: {response.Value.Content[0].Text}");
    }
}
catch (ClientResultException ex)
{
    Console.WriteLine($"\nERROR: {ex.Status} - {ex.Message}");
    Console.WriteLine("\nThis confirms the issue: without preserving reasoning_details,");
    Console.WriteLine("Gemini 3 Preview models fail on the second request.");
}

[stephentoub]

@vbandi, are you still hitting failures here? I'm trying to repro but struggling to get it to 400. Did OpenRouter maybe improve since you filed this?

[vbandi]

I just checked and yes, it seems like this got fixed in the meantime, likely on OpenRouter's side.