Events not triggered: reasoning content from vLLM BYOK not surfaced in Copilot SDK

[lawrence6140791]

When using GitHub Copilot SDK with BYOK (Bring Your Own Key) and vLLM as the provider, reasoning content returned by vLLM in the reasoning field is not processed. As a result, key events such as AssistantMessageDeltaEvent and AssistantReasoningEvent are not triggered as expected, and the reasoning data is not received by the client.

Steps to reproduce:

1. Configure Copilot SDK to use BYOK with vLLM as provider (Windows or WSL).
2. Send a request that should generate reasoning content.
3. Observe that AssistantMessageDeltaEvent and AssistantReasoningEvent do not fire or do not surface reasoning content.
4. Inspect API response: vLLM returns reasoning in reasoning, but SDK only checks reasoning_content.

Expected behavior:
Copilot SDK should handle both reasoning and reasoning_content fields, so events work and reasoning data is surfaced, regardless of provider.

Context:

• Environment: Windows, WSL
• Copilot SDK version: 0.3.0
• Copilot Cli version: v1.0.43
• BYOK provider: vLLM

[github-actions]

Bug Investigation Findings

Thanks for the detailed report. I investigated this issue and confirmed it is a genuine bug. Here's what I found:

What I investigated

• nodejs/src/session.ts, python/copilot/session.py, go/session.go, dotnet/src/Session.cs — the SDK session event-processing layer
• nodejs/src/generated/session-events.ts and equivalents — SDK event type definitions
• docs/auth/byok.md, docs/troubleshooting/compatibility.md — BYOK documentation and feature support tables
• All SDK source files for any reference to reasoning_content

What I found

1. The SDK itself contains no reasoning_content field handling.
A search across all SDK source files (*.ts, *.py, *.go, *.cs) returns zero matches for reasoning_content. The issue description's framing is slightly off — the SDK does not directly inspect raw LLM responses at all.

2. The SDK is a thin JSON-RPC client over the Copilot CLI.
The architecture is: Your App → SDK → JSON-RPC → Copilot CLI → LLM Provider (vLLM). The Copilot CLI is the component that actually calls vLLM, parses its response (including reasoning fields), and emits events like assistant.reasoning and assistant.reasoning_delta to the SDK. The SDK then passes these events through to your application. If the CLI doesn't emit them, the SDK has no data to surface.

3. The real issue is in the Copilot CLI's response parsing.
The CLI apparently only handles reasoning_content (the field name used by DeepSeek-style models) and not reasoning (used by some vLLM model configurations). This means the CLI silently drops vLLM reasoning content without emitting the corresponding SDK events.

4. vLLM is an officially documented, supported BYOK provider.
From docs/auth/byok.md:

| Other OpenAI-compatible | "openai" | vLLM, LiteLLM, etc. |

There is no documented limitation around reasoning event support for vLLM. Users reasonably expect AssistantReasoningEvent / AssistantReasoningDeltaEvent to work with vLLM just as with other providers.

Why this is a bug

vLLM is an explicitly supported BYOK provider, and reasoning events are an advertised SDK capability. A provider-specific field name difference (reasoning vs reasoning_content) silently suppressing reasoning content — without any documentation of this limitation — is a functional regression for users of a documented feature combination.

The fix belongs in the Copilot CLI (to also handle the reasoning field name used by vLLM), but this repository is the right place to track the issue since it affects the SDK's documented behavior.

Generated by Bug Handler for issue #1229 · ● 421.4K · ◷