.NET: [Bug]: "Empty" AIContextProvider breaks skill loading

[VaclavK]

Description

we have agent decorated with skill loading provider context provider - no issue

we wanted to add another augmenting context provider but even if the provider literally returns

new AIContext()

in ProvideAIContextAsync and nothing else, we see skill loading breaking with

Task failed: HTTP 400 (invalid_request_error: )
Parameter: input

Item 'fc_007f6e0e4e84d6ca006a5e107039ec8196aea3835af4b14dce' of type 'function_call' was provided without its required 'reasoning' item: 'rs_007f6e0e4e84d6ca006a5e106fd43881969b6860989376dab2'.

that's when the agent is invoked w/o session

Code Sample

protected override async ValueTask<AIContext> ProvideAIContextAsync(InvokingContext context, CancellationToken cancellationToken = default)
   {
      
       return new AIContext();

   }

Error Messages / Stack Traces

Task failed: HTTP 400 (invalid_request_error: )
Parameter: input

Item 'fc_007f6e0e4e84d6ca006a5e107039ec8196aea3835af4b14dce' of type 'function_call' was provided without its required 'reasoning' item: 'rs_007f6e0e4e84d6ca006a5e106fd43881969b6860989376dab2'.

Package Versions

1.13

.NET Version

.net 10

Additional Context

No response

[github-actions]

👋 Hi @VaclavK — thanks for the report!

Your error about 'function_call provided without its required reasoning item' is an OpenAI Responses API constraint for reasoning models (o3, o4-mini, etc.). We wrote unit tests matching your scenario — an empty AIContextProvider returning new AIContext() combined with tool calls that include TextReasoningContent — and reasoning is preserved correctly through the provider pipeline in both PerServiceCall and end-of-run persistence modes. To help reproduce on our side, could you share: (1) which OpenAI model you're using, (2) whether you're using AsIChatClientWithStoredOutputDisabled or the standard AsIChatClient, (3) whether RequirePerServiceCallChatHistoryPersistence is enabled, and (4) the full agent configuration/pipeline setup? The framework code itself does not strip reasoning items when an empty provider is added.

Once you can share the details above, we'll take another look. 🙏

[VaclavK]

it seems the underlying differentiator is whether the agent is invoked with or without session

var result = await agent.RunAsync<OrchestrationEnvelope>(BuildMessage(request));

or

var result = await agent.RunAsync<OrchestrationEnvelope>(BuildMessage(request), session);

completes with session, throws the error without session

we are using gpt-5.5

we use standard AI client set up as follows

 builder.AddAIAgent(Consts.RouterValidatorAgentName, (sp, agentKey) =>
 {
     var chatClient = sp.GetRequiredService<IChatClient>();
     var options = new ChatClientAgentOptions
     {
         Name = Consts.RouterValidatorAgentName,
         AIContextProviders = [sp.GetRequiredService<TestAIContextProvider>(), routerValidatorSkillsProvider],
         //ChatHistoryProvider = sp.GetRequiredService<InMemoryChatHistoryProvider>(),
         ChatOptions = new ChatOptions
         {
             Seed = 123456789,
             Instructions = Router.RouterAgentInstructions
         }
     };
     return new ChatClientAgent(chatClient, options)
         .AsBuilder()
         .UseToolApproval(new ToolApprovalAgentOptions { AutoApprovalRules = [autoApprove] })
         .Build();
 });

the inner client is set up as

 .GetResponsesClient()
 .AsIChatClient(deploymentName)
 .AsBuilder()
 .Build();

[VaclavK]

👋 Hi @VaclavK — thanks for the report!

Your error about 'function_call provided without its required reasoning item' is an OpenAI Responses API constraint for reasoning models (o3, o4-mini, etc.). We wrote unit tests matching your scenario — an empty AIContextProvider returning new AIContext() combined with tool calls that include TextReasoningContent — and reasoning is preserved correctly through the provider pipeline in both PerServiceCall and end-of-run persistence modes. To help reproduce on our side, could you share: (1) which OpenAI model you're using, (2) whether you're using AsIChatClientWithStoredOutputDisabled or the standard AsIChatClient, (3) whether RequirePerServiceCallChatHistoryPersistence is enabled, and (4) the full agent configuration/pipeline setup? The framework code itself does not strip reasoning items when an empty provider is added.

Once you can share the details above, we'll take another look. 🙏

it seems the underlying differentiator is whether the agent is invoked with or without session

var result = await agent.RunAsync<OrchestrationEnvelope>(BuildMessage(request));

or

var result = await agent.RunAsync<OrchestrationEnvelope>(BuildMessage(request), session);

completes with session, throws the error without session

we are using gpt-5.5

we use standard AI client set up as follows

 builder.AddAIAgent(Consts.RouterValidatorAgentName, (sp, agentKey) =>
 {
     var chatClient = sp.GetRequiredService<IChatClient>();
     var options = new ChatClientAgentOptions
     {
         Name = Consts.RouterValidatorAgentName,
         AIContextProviders = [sp.GetRequiredService<TestAIContextProvider>(), routerValidatorSkillsProvider],
         //ChatHistoryProvider = sp.GetRequiredService<InMemoryChatHistoryProvider>(),
         ChatOptions = new ChatOptions
         {
             Seed = 123456789,
             Instructions = Router.RouterAgentInstructions
         }
     };
     return new ChatClientAgent(chatClient, options)
         .AsBuilder()
         .UseToolApproval(new ToolApprovalAgentOptions { AutoApprovalRules = [autoApprove] })
         .Build();
 });

the inner client is set up as

 .GetResponsesClient()
 .AsIChatClient(deploymentName)
 .AsBuilder()
 .Build();

[VaclavK]

so I had AI generate a kind of demo console app to demo what we have been seeing

/*
 * Router Agent Session Demo
 * =========================
 * This console app registers the RouterAgent exactly as the FluxBackend does
 * (the USE_MAF_AGENTS path) and then demonstrates two scenarios:
 *
 *   1. WITHOUT a session  → fails with InvalidOperationException because
 *      ExternalMemoryAIContextProvider requires a "sessionId" in run options.
 *
 *   2. WITH a session + sessionId  → succeeds and prints the OrchestrationEnvelope.
 *
 * Prerequisites (user-secrets or environment variables):
 *   flux-backend:azure-openai-endpoint          (Azure OpenAI endpoint URL)
 *   flux-backend:azure-openai-deployment-name   (model deployment name)
 */

using Flux.Backend;
using Flux.Backend.AgentInstructions;
using Flux.Backend.Contracts;
using Flux.Backend.Services;
using Microsoft.Agents.AI;
using Microsoft.Agents.AI.Hosting;
using Microsoft.Extensions.AI;
using Microsoft.Extensions.Configuration;
using Microsoft.Extensions.DependencyInjection;
using Microsoft.Extensions.Hosting;
using Microsoft.Extensions.Logging;
using System.Text.Json;

// ──────────────────────────────────────────────
// 1. Build configuration (user secrets + env)
// ──────────────────────────────────────────────
var config = new ConfigurationBuilder()
    .AddJsonFile("appsettings.json", optional: true)
    .AddUserSecrets<Program>()
    .AddEnvironmentVariables()
    .Build();

string endpoint = config["flux-backend:azure-openai-endpoint"]
    ?? throw new InvalidOperationException("flux-backend:azure-openai-endpoint is not set.");

string deployment = config["flux-backend:azure-openai-deployment-name"]
    ?? throw new InvalidOperationException("flux-backend:azure-openai-deployment-name is not set.");

// ──────────────────────────────────────────────
// 2. Build the host / DI container
// ──────────────────────────────────────────────
#pragma warning disable MAAI001

var host = Host.CreateDefaultBuilder()
    .ConfigureLogging(logging =>
    {
        logging.ClearProviders();
        logging.AddConsole();
        logging.SetMinimumLevel(LogLevel.Warning); // keep output tidy
    })
    .ConfigureServices((_, services) =>
    {
        // Chat client (same factory as the backend)
        services.AddSingleton<IChatClient>(_ =>
            OpenAIChatClientFactory.Create(endpoint, deployment)
                .AsBuilder()
                .Build());

        // Empty dummy AI context provider for demo purposes
        services.AddSingleton<EmptyDummyAIContextProvider>();

        // Router agent skills provider (points at the Artefacts output directory)
        var skillsBasePath = Path.Combine(AppContext.BaseDirectory, "skills");

        var routerValidatorSkillsProvider = new AgentSkillsProvider(
            skillPath: Path.Join(skillsBasePath, "routervalidator"));

        // Auto-approve all tool calls (same as backend)
        Func<FunctionCallContent, ValueTask<bool>> autoApprove =
            _ => ValueTask.FromResult(true);

        // Session store (in-memory, same as backend)
        services.AddSingleton<AgentSessionStore, InMemoryAgentSessionStore>();

        // RouterValidator agent  ← registered exactly as in Program.cs (USE_MAF_AGENTS)
        services.AddAIAgent(Consts.RouterValidatorAgentName, (sp, _) =>
        {
            var chatClient = sp.GetRequiredService<IChatClient>();
            var options = new ChatClientAgentOptions
            {
                Name = Consts.RouterValidatorAgentName,
                AIContextProviders =
                [
                    sp.GetRequiredService<EmptyDummyAIContextProvider>(),
                    routerValidatorSkillsProvider
                ],
                ChatOptions = new ChatOptions
                {
                    Seed = 123456789,
                    Instructions = Router.RouterAgentInstructions
                }
            };
            return new ChatClientAgent(chatClient, options)
                .AsBuilder()
                .UseToolApproval(new ToolApprovalAgentOptions { AutoApprovalRules = [autoApprove] })
                .Build();
        });

        // Session agent (dummy, used only to hold session state)
        services.AddAIAgent(Consts.SessionAgent, (sp, _) =>
        {
            var chatClient = sp.GetRequiredService<IChatClient>();
            var options = new ChatClientAgentOptions
            {
                Name = Consts.SessionAgent,
                ChatOptions = new ChatOptions
                {
                    Seed = 123456789,
                    Instructions = "you are fake agent and only used internally to store and load sessions",
                }
            };
            return new ChatClientAgent(chatClient, options)
                .AsBuilder()
                .UseToolApproval(new ToolApprovalAgentOptions { AutoApprovalRules = [autoApprove] })
                .Build();
        });
    })
    .Build();

#pragma warning restore MAAI001

// ──────────────────────────────────────────────
// 3. Resolve agents
// ──────────────────────────────────────────────
var sp = host.Services;

var routerAgent   = sp.GetRequiredKeyedService<AIAgent>(Consts.RouterValidatorAgentName);
var sessionAgent  = sp.GetRequiredKeyedService<AIAgent>(Consts.SessionAgent);
var sessionStore  = sp.GetRequiredService<AgentSessionStore>();

const string demoPrompt = "onboard project Helios with owner jane@example.com";

// ──────────────────────────────────────────────
// 4. DEMO A — without session (expected failure)
// ──────────────────────────────────────────────
Console.ForegroundColor = ConsoleColor.Yellow;
Console.WriteLine("════════════════════════════════════════════════");
Console.WriteLine(" DEMO A — Running WITHOUT a session");
Console.WriteLine("════════════════════════════════════════════════");
Console.ResetColor();
Console.WriteLine($"Prompt: {demoPrompt}");
Console.WriteLine();

try
{
    var sessionId = Guid.NewGuid().ToString();

    var session = await SessionService.GetOrCreateSessionAsync(sessionStore, sessionAgent, sessionId);

    // No session, no sessionId in run options → ExternalMemoryAIContextProvider will throw
    var resultA = await StructuredAgentRunner.RunAsync<OrchestrationEnvelope>(
         routerAgent,
        demoPrompt,
        session: null,
        sessionId: sessionId,
        storeContext: true,
        cancellationToken: CancellationToken.None);

    Console.ForegroundColor = ConsoleColor.Red;
    Console.WriteLine("⚠  Expected a failure but got a result — this is unexpected:");
    Console.ResetColor();
    Console.WriteLine(JsonSerializer.Serialize(resultA, new JsonSerializerOptions { WriteIndented = true }));
}
catch (InvalidOperationException ex) when (ex.Message.Contains("Session ID is missing"))
{
    Console.ForegroundColor = ConsoleColor.Green;
    Console.WriteLine($"✓  Got the expected failure: {ex.Message}");
    Console.ResetColor();
}
catch (Exception ex)
{
    Console.ForegroundColor = ConsoleColor.Red;
    Console.WriteLine($"✗  Got an unexpected exception: {ex.GetType().Name}: {ex.Message}");
    Console.ResetColor();
}

Console.WriteLine();

// ──────────────────────────────────────────────
// 5. DEMO B — with session (expected success)
// ──────────────────────────────────────────────
Console.ForegroundColor = ConsoleColor.Cyan;
Console.WriteLine("════════════════════════════════════════════════");
Console.WriteLine(" DEMO B — Running WITH a session");
Console.WriteLine("════════════════════════════════════════════════");
Console.ResetColor();
Console.WriteLine($"Prompt: {demoPrompt}");
Console.WriteLine();

try
{
    var sessionId = Guid.NewGuid().ToString();

    // GetOrCreateSession gives us a proper AgentSession
    var session = await SessionService.GetOrCreateSessionAsync(sessionStore, sessionAgent, sessionId);

    // Pass the session AND sessionId in run options, exactly as RouterAgentActivity does
    var resultB = await StructuredAgentRunner.RunAsync<OrchestrationEnvelope>(
        routerAgent,
        demoPrompt,
        session,
        sessionId: sessionId,
        storeContext: true,
        cancellationToken: CancellationToken.None);

    Console.ForegroundColor = ConsoleColor.Green;
    Console.WriteLine("✓  Agent returned a result:");
    Console.ResetColor();
    Console.WriteLine(JsonSerializer.Serialize(resultB, new JsonSerializerOptions { WriteIndented = true }));
}
catch (Exception ex)
{
    Console.ForegroundColor = ConsoleColor.Red;
    Console.WriteLine($"✗  Unexpected exception: {ex.GetType().Name}: {ex.Message}");
    if (ex.InnerException is not null)
        Console.WriteLine($"   Inner: {ex.InnerException.Message}");
    Console.ResetColor();
}

internal sealed class EmptyDummyAIContextProvider : AIContextProvider
{
    public override IReadOnlyList<string> StateKeys => [];

    protected override ValueTask<AIContext> ProvideAIContextAsync(InvokingContext context, CancellationToken cancellationToken = default)
    {
        return ValueTask.FromResult(new AIContext());
    }
}

you will see the difference being session null for Demo A and that produces skill load tool call exception as opposed to session being provided for Demo B and that completes as expected

which produces

 DEMO A - Running WITHOUT a session
════════════════════════════════════════════════
Prompt: onboard project Helios with owner jane@example.com

?  Got an unexpected exception: ClientResultException: HTTP 400 (invalid_request_error: )
Parameter: input

Item 'fc_08513de623800d8e006a5f9ef17e7c81939dbab52b49fa67d3' of type 'function_call' was provided without its required 'reasoning' item: 'rs_08513de623800d8e006a5f9ef0ab4c8193a52f5167e57774f7'.

════════════════════════════════════════════════
 DEMO B - Running WITH a session
════════════════════════════════════════════════
Prompt: onboard project Helios with owner jane@example.com

√  Agent returned a result:
{
  "SubPrompt": "onboard project Helios with owner jane@example.com",
  "WorkItems": [
    {
      "Domain": "ado",
      "Prompt": "create or ensure Azure DevOps project with the name \u0027Helios\u0027 with project administrator \u0027jane@example.com\u0027",
      "RouteAfterCompletion": "return_to_router"
    },
    {
      "Domain": "teams",
      "Prompt": "create or ensure Team with the name \u0027Helios\u0027 with owner of \u0027jane@example.com\u0027 and ensure the private \u0027Configuration\u0027 channel exists in the team and ensure \u0027jane@example.com\u0027 is a member of the channel and team",
      "RouteAfterCompletion": "return_to_router"
    },
    {
      "Domain": "github",
      "Prompt": "create github organization team \u0027Helios\u0027",
      "RouteAfterCompletion": "return_to_router"
    }
  ],
  "Context": [
    "Project onboarding requested for project \u0027Helios\u0027 with owner \u0027jane@example.com\u0027.",
    "After GitHub project-level onboarding is completed, schedule GitHub user onboarding for owner \u0027jane@example.com\u0027 to GitHub organization team \u0027Helios\u0027; this requires the target user\u0027s GitHub handle before the GitHub user onboarding work can be emitted."
  ],
  "EngagementQueueItems": [],
  "RequiresEngagement": false,
  "EngagementType": 0,
  "Succeeded": true,
  "Error": null,
  "ResultSummary": "Routed project onboarding for \u0027Helios\u0027 into Azure DevOps, Teams, and GitHub core work items. GitHub user onboarding for owner \u0027jane@example.com\u0027 is deferred until after core project onboarding and requires the owner\u2019s GitHub handle."
}

C:\projects\Flux\src\RouterAgentDemo\bin\Debug\net10.0\RouterAgentDemo.exe (process 56004) exited with code 0 (0x0).
To automatically close the console when debugging stops, enable Tools->Options->Debugging->Automatically close the console when debugging stops.
Press any key to close this window . . .

[westey-m]

@VaclavK, which responses service are you using? E.g. OpenAI directly, or are you using Foundry or some other responses compatible service?

[VaclavK]

@VaclavK, which responses service are you using? E.g. OpenAI directly, or are you using Foundry or some other responses compatible service?

Foundry hosted chatgpt-5.5 model deployment

[westey-m]

@VaclavK, I haven't been able to repro your exact failure, but I believe it's most likely related to the tool auto approval feature, which does multiple sub agent invocations without a session, and this is a bug. I've created a PR to fix.

Some general tips:

If you just want to approve everything (yolo mode), there is a built in approval rule already you can use: ToolApprovalAgent.AllToolsAutoApprovalRule, so you don't need to write your own.

Also, the SkillsProvider allows you to pass options that can switch off tool approval entirely for loading skills/resources and executing skill scripts, which means you may not need .UseToolApproval(new ToolApprovalAgentOptions { AutoApprovalRules = [autoApprove] }).
See https://github.com/microsoft/agent-framework/blob/main/dotnet/src/Microsoft.Agents.AI/Skills/AgentSkillsProviderOptions.cs#L44.