SDK 0.2.x rejects ReasoningEffort for BYOK gpt-5.4-mini model

[mit2nil]

Description

When using the .NET Copilot SDK with a BYOK (Bring Your Own Key) Azure OpenAI provider, setting ReasoningEffort on SessionConfig for gpt-5.4-mini throws:

System.IO.IOException: Communication error with Copilot CLI: Request session.create failed with message: Model 'gpt-5.4-mini' does not support reasoning effort configuration. Use models.list to check which models support reasoning effort.

Expected Behavior

gpt-5.4-mini is a reasoning model and supports reasoning effort configuration via the Azure OpenAI Responses API. The SDK should either:

1. Pass ReasoningEffort through to the provider API without validating against its own model registry, OR
2. Include gpt-5.4-mini (and other GPT-5.x models) in its reasoning model allowlist

When using BYOK providers, the SDK should not gate ReasoningEffort based on its own internal model catalog — the provider endpoint knows what its deployed models support.

Reproduction

var sessionConfig = new SessionConfig
{
    Model = "gpt-5.4-mini",
    ReasoningEffort = "high",
    Provider = new ProviderConfig
    {
        Type = "openai",
        BaseUrl = "https://my-aoai.cognitiveservices.azure.com/openai/v1/",
        BearerToken = "...",
    },
};

// Throws: "Model 'gpt-5.4-mini' does not support reasoning effort configuration"
await client.CreateSessionAsync(sessionConfig);

Environment

• SDK versions tested: 0.1.32 and 0.2.1 (NuGet: GitHub.Copilot.SDK) — fails on both
• Runtime: .NET 8.0
• Provider: Azure OpenAI (BYOK) with WireApi = "responses"

Workaround

Catching the exception and retrying without ReasoningEffort, but this degrades model behavior since reasoning effort cannot be configured for BYOK deployments.

[github-actions]

Investigation findings

Thanks for the detailed report. I investigated this and can confirm the behavior, but it turns out the SDK already provides a sanctioned override mechanism for exactly this scenario — it's just completely undocumented. I've labeled this as documentation.

What I found

The error originates in the Copilot CLI, not the SDK. The .NET SDK itself does no client-side validation of ReasoningEffort; it sends the value verbatim via the session.create RPC (see Client.cs:489). The CLI then validates whether the requested model has capabilities.supports.reasoningEffort = true in its internal registry, which doesn't know about BYOK custom models like gpt-5.4-mini.

A ModelCapabilities override exists for this purpose. The SDK exposes ModelCapabilitiesOverride (Rpc.cs:393), which is deep-merged over the CLI's runtime defaults before the validation runs. Setting Supports.ReasoningEffort = true here tells the CLI "trust the caller, this model supports it." The field is wired all the way through:

• SessionConfig.ModelCapabilities → CreateSessionRequest.ModelCapabilities → sent in the RPC body

Immediate workaround

var sessionConfig = new SessionConfig
{
    Model = "gpt-5.4-mini",
    ReasoningEffort = "high",
    Provider = new ProviderConfig
    {
        Type = "openai",
        BaseUrl = "(myaoai.cognitiveservices.azure.com/redacted),
        BearerToken = "...",
    },
    ModelCapabilities = new ModelCapabilitiesOverride
    {
        Supports = new ModelCapabilitiesOverrideSupports
        {
            ReasoningEffort = true,
        },
    },
};
await client.CreateSessionAsync(sessionConfig);

Documentation gaps

The docs need to be fixed in two places in dotnet/README.md:

1. ReasoningEffort entry (currently says "Use ListModelsAsync() to check which models support this option") should also note that for BYOK providers with models not in the Copilot model registry, ModelCapabilities.Supports.ReasoningEffort = true must be set to bypass the registry check.

2. ModelCapabilities is not documented at all in any SDK README. It should be added to the CreateSessionAsync config table with a description of when to use it (e.g., BYOK providers, new model IDs not yet in the registry).

Regarding the regression from 0.1.32: the CLI's model validation may have been tightened in a recent release. Even if that change is rolled back, documenting the override property remains important for forward compatibility.

Generated by Bug Handler for issue #1041 · ◷

[mit2nil]

Thanks for the investigation. Unfortunately ModelCapabilitiesOverride and ModelCapabilitiesOverrideSupports types don't exist in the published SDK 0.2.1 NuGet package. SessionConfig only has a Model string property — no ModelCapabilities property is exposed.

\
// SDK 0.2.1 — SessionConfig has no ModelCapabilities property
var type = typeof(GitHub.Copilot.SDK.SessionConfig);
type.GetProperties().Where(p => p.Name.Contains("Model"))
// Returns only: Model (string)
\\

Are these types available in an unreleased version or a different package? We're using:

• NuGet package: GitHub.Copilot.SDK version 0.2.1
• Runtime: .NET 8.0

For now we're using a retry fallback (catch the exception, set ReasoningEffort = null, retry CreateSessionAsync), but this means BYOK deployments can't configure reasoning effort at all.

Would appreciate either:

1. Publishing a new SDK version with ModelCapabilities exposed, or
2. Removing the CLI-side model validation for ReasoningEffort when a BYOK provider is configured (the provider endpoint knows what its models support)

[SteveSandersonMS]

@mit2nil ModelCapabilitiesOverride is a new API only added this week. It's currently available in packages versioned 0.2.2-preview.0. It will be included in the next stable release as well, and we'll update docs then too.

I'll close this as it seems the scenario is covered.