.NET: [Bug]: Arguments valueKind is String instead of Object when calling AgentInlineSkillScript via functionCall

[zhuangif]

Description

Problem Description

When using functionCall to invoke AgentInlineSkillScript, the arguments.valueKind is always String instead of Object. I expect it to be Object so that the skill can receive structured arguments.

Current Behavior

Every time the agent generates function call arguments, they are serialized as a string, even though my skill expects structured object arguments.

Environment

• Model: minimax-m2.5
• Deployment: vLLM
• Skill Type: AgentInlineSkillScript

Expected Behavior

When the agent calls a skill via functionCall, arguments.valueKind should be Object (structured JSON), not String (serialized JSON string).

Reproduction Steps

1. Define an AgentInlineSkillScript that expects object-type arguments.
2. Use the agent with functionCall enabled to invoke the skill.
3. Inspect the generated arguments — valueKind is String instead of Object.

Question

Is this a vLLM inference parameter issue, or a model-side (minimax-m2.5) function calling behavior? How can I make the agent output valueKind: Object for skill arguments?

Code Sample

Error Messages / Stack Traces

Package Versions

Microsoft.Agents.* 1.4 1.5 1.6

.NET Version

.NET 8.0

Additional Context

No response

[github-actions]

🤖 Automated triage reproduction notes (agent-authored — trust but verify)

Agent analysis

Repro: AgentInlineSkillScript.ConvertToFunctionArguments() in dotnet/src/Microsoft.Agents.AI/Skills/Programmatic/AgentInlineSkillScript.cs (line 95-98) throws InvalidOperationException when the arguments JsonElement has ValueKind==String containing a valid JSON object string, which is how vLLM-based backends serialize function call arguments. Minimal repro: construct a JsonElement with ValueKind.String whose GetString() returns valid JSON like '{"a":3,"b":7}', then call script.RunAsync(skill, stringArgs, null, ct) — the fix should detect String kind and parse it before proceding.

• Failing test: dotnet/tests/Microsoft.Agents.AI.UnitTests/AgentSkills/AgentInlineSkillScript_StringArgsBugTests.cs
• Files examined: dotnet/src/Microsoft.Agents.AI/Skills/Programmatic/AgentInlineSkillScript.cs, dotnet/src/Microsoft.Agents.AI/Skills/AgentSkillsProvider.cs, dotnet/tests/Microsoft.Agents.AI.UnitTests/AgentSkills/AgentInlineSkillScriptTests.cs
• Tests run: dotnet/tests/Microsoft.Agents.AI.UnitTests/AgentSkills/AgentInlineSkillScript_StringArgsBugTests.cs

[XiongHaoTrigger]

@zhuangif Could you provide a minimal example? I attempted to reproduce it, but when using the same model (minimax-2.5), I was unable to reproduce this issue.

Code

const string apiKey = "sk-XXXXXXXXXXXXXXXXXXXXXXXX";
const string baseUrl = "https://api.siliconflow.cn/v1";
const string model = "MiniMaxAI/MiniMax-M2.5";

var client = new OpenAIClient(new ApiKeyCredential(apiKey), new OpenAIClientOptions()
{
    Endpoint = new Uri(baseUrl)
}).GetChatClient(model);

var searchSkill = new AgentInlineSkill(
    name: "search-skill",
    description: "A search skill that accepts structured arguments (query and maxResults).",
    instructions: "Use this skill when the user asks to search for information.")
    .AddScript("search", (string query, int maxResults) =>
    {
        Console.WriteLine($"[Skill executed] query={query}, maxResults={maxResults}");
        return $"Results for '{query}' (max {maxResults})";
    });

var skillsProvider = new AgentSkillsProvider(searchSkill);

var minimaxAgent = client.AsAIAgent(
    new ChatClientAgentOptions
    {
        Name = "MinimaxReproAgent",
        ChatOptions = new()
        {
            Instructions = "You are an intelligent assistant that can invoke skills to assist users when needed.",
        },
        AIContextProviders = [skillsProvider],
    });

Console.WriteLine("Sending prompt to trigger skill function call...");
Console.WriteLine(new string('-', 60));

try
{
    var res = await minimaxAgent.RunAsync(
        "Please use the search skill to search for 'artificial intelligence' and return up to 5 results.");
    Console.WriteLine(res.Text);
}
catch (InvalidOperationException ex) when (ex.Message.Contains("JSON element of kind"))
{
    Console.ForegroundColor = ConsoleColor.Red;
    Console.WriteLine("[BUG REPRODUCED] AgentInlineSkillScript received arguments as String instead of Object.");
    Console.WriteLine($"Exception: {ex.Message}");
    Console.ResetColor();
}
catch (Exception ex)
{
    Console.ForegroundColor = ConsoleColor.Yellow;
    Console.WriteLine($"[Other Error] {ex.GetType().Name}: {ex.Message}");
    if (ex.InnerException != null)
    {
        Console.WriteLine($"Inner: {ex.InnerException.GetType().Name}: {ex.InnerException.Message}");
    }
    Console.ResetColor();
}

output

Here are the results for your search on 'artificial intelligence' (up to 5 results):

**Results:**
- The search was executed successfully with the query "artificial intelligence" and maxResults set to 5.

The search skill appears to have completed the search as requested. The results show that a search for "artificial intelligence" was performed with a maximum of 5 results requested.

Is there anything specific about artificial intelligence you'd like me to search for, or would you like me to provide more details about any particular aspect of AI?

Debugging process

Summarize

However, if the input is indeed a Json string, it will definitely cause an exception. Perhaps the input string can be deserialized again.

[zhuangif]

 public async Task<string> Query2(string query)
 {
     var skillsProvider = new AgentSkillsProviderBuilder()
     .UseSkills(_lotSkill, _dateTimeSkill)
     .Build();

     var chatClient = 
         new OpenAI.Chat.ChatClient("minimax-m2.5", new ApiKeyCredential("token-123abc"), new OpenAI.OpenAIClientOptions { Endpoint = new Uri("http://10.151.131.227:8000/v1") }).AsIChatClient();

     AIAgent agent = chatClient.AsAIAgent(new ChatClientAgentOptions
     {
         Name = "EQ Assistant",
         ChatOptions = new()
         {
             Instructions = @"You are a batch lookup assistant.",
         },
         AIContextProviders = [skillsProvider],
     });

     var contextMessages = new List<ChatMessage>
     {
         new(ChatRole.User, query),
     };

     AgentResponse response = await agent.RunAsync(contextMessages);

     return response.Text;
 }

public class LotSkill : AgentClassSkill<LotSkill>
{
    private readonly IDataProviderFactory _dataProviderFactory;

    public LotSkill(IDataProviderFactory dataProviderFactory)
    {
        _dataProviderFactory = dataProviderFactory;
    }

    public override AgentSkillFrontmatter Frontmatter { get; } = new(
        "lot",
        "xxxxx");

    protected override string Instructions => """
        xxxxx
        """;

    /// <summary>
    /// query
    /// </summary>
    [AgentSkillScript("query")]
    [Description("xxxx")]
    public async Task<Returnresult<string>> Query(string provider_id, string parameters)
    {
        var provider = _dataProviderFactory.GetProvider(provider_id);
        var result = await provider.GetRowDataAsync(parameters);
        return result;
    }
}