Responses API: AssistantReasoningDeltaEvent and AssistantReasoningEvent triggered with empty content via BYOK providers

[lawrence6140791]

When using GitHub Copilot SDK in a BYOK (Bring Your Own Key) setup—whether the provider is OpenAI or vLLM—the Responses API triggers both AssistantReasoningDeltaEvent and AssistantReasoningEvent, but the content (reasoning or explanation) for these events is always empty.

Steps to reproduce:

1. Set up Copilot SDK to use BYOK with OpenAI or vLLM as the provider (Windows or WSL). Example .NET code configuration:

await using var session = await client.CreateSessionAsync(new SessionConfig
{
    Provider = new ProviderConfig
    {
        Type = "openai",
        BaseUrl = "https://api.openai.com/v1",
        ApiKey = "your_api_key",
        WireApi = "responses",
    },
    ReasoningEffort = "high",
    Model = "gpt-5.4-2026-03-05",
    OnPermissionRequest = PermissionHandler.ApproveAll,
    Streaming = true
});

2. Trigger chat completions or requests that should provide reasoning content through the Responses API.
3. Observe that AssistantReasoningDeltaEvent and AssistantReasoningEvent are triggered, but the content is always empty.

Expected behavior:
These events should include the reasoning/explanation content returned by the provider's API. Content should not be empty if the provider returns a reasoning field.

Context:

• Environment: Windows, WSL
• Copilot SDK version: 0.3.0
• BYOK provider: OpenAI or vLLM

[github-actions]

Investigation findings

Thanks for the report! I investigated this by examining the relevant source code and documentation.

What I found

For OpenAI BYOK (type: "openai", wireApi: "responses"): This behavior is by design. OpenAI's API does not expose plaintext reasoning chain content to API consumers — instead, the Copilot runtime captures an encrypted form of it. Looking at AssistantMessageData in the generated types:

/**
 * Encrypted reasoning content from OpenAI models. Session-bound and stripped on resume.
 */
encryptedContent?: string;

The assistant.reasoning and assistant.reasoning_delta events do fire, but their content/deltaContent fields are empty because OpenAI does not return readable reasoning text via the API. The encrypted reasoning is available on the assistant.message event's encryptedContent field (usable only within the same session; it is stripped on resume).

For vLLM BYOK: Open-source models served via vLLM (e.g., DeepSeek-R1) can expose plaintext reasoning content. If these events are also empty for vLLM, that may indicate the runtime isn't extracting reasoning text from vLLM's response format. However, since the runtime (CLI) is not part of this SDK repository, that warrants separate investigation.

Documentation gap

The current documentation in docs/features/streaming-events.md for assistant.reasoning and assistant.reasoning_delta states that content/deltaContent is required and describes it as "The complete extended thinking text" — with no caveat that, for OpenAI models, the text will always be empty because the content is encrypted. Likewise, docs/auth/byok.md does not mention any reasoning content limitations per provider.

What needs to be added/fixed

1. docs/features/streaming-events.md — The assistant.reasoning and assistant.reasoning_delta event tables should note that for OpenAI models, content/deltaContent will be an empty string (reasoning is encrypted); direct users to the encryptedContent field on the assistant.message event.
2. docs/auth/byok.md — Add a section explaining how reasoning content differs by provider: OpenAI encrypts it (use encryptedContent), Anthropic exposes it as readable text (via reasoningOpaque/reasoningText on the message event or assistant.reasoning events), and open-source providers like vLLM may expose plaintext reasoning.

This issue has been relabelled as documentation.

Generated by Bug Handler for issue #1230 · ● 1.2M · ◷