[Model Routing] claude-sonnet-5 swallows streaming assistant.reasoning_delta; flushes all thinkingContent at completion

[JaretRussell]

Environment Details:

Working Model: claude-sonnet-4.5 (streams perfectly)

Failing Model: claude-sonnet-5

Integration: GitHub Copilot SDK (subscribe('assistant.reasoning_delta'))

Observed Behavior:
When routing requests to claude-sonnet-5, the SDK emits zero assistant.reasoning_delta events during the generation lifecycle. Instead, it accumulates them in the background, resulting in a ~25,000ms Time-to-First-Token (TTFT) and blocking client-side UI rendering. A complete, non-empty thinkingContent block is delivered atomically only when the final text stream buffer flushes or the file completes.

Expected Behavior:
The SDK should map Anthropic's streaming reasoning frames into real-time assistant.reasoning_delta events immediately upon receiving them from the upstream model router, preserving low operational TTFT just as it successfully does for claude-sonnet-4.5.

[github-actions]

Investigation Findings

Thanks for the detailed report. I investigated the relevant SDK code and want to clarify how the SDK relates to the behavior you're describing.

What I looked at

• nodejs/src/session.ts — _dispatchEvent and the event subscription system
• nodejs/src/generated/session-events.ts — assistant.reasoning_delta type definitions
• All SDK source files for any model-specific logic, buffering, or filtering

What I found

The SDK is a thin JSON-RPC pass-through between your application and the Copilot CLI server. The _dispatchEvent method dispatches events directly to subscribers as they arrive, with no buffering, batching, or model-specific filtering:

_dispatchEvent(event: SessionEvent): void {
    // Dispatches immediately to typed handlers and wildcard handlers
    const typedHandlers = this.typedEventHandlers.get(event.type);
    if (typedHandlers) {
        for (const handler of typedHandlers) {
            handler(event);  // immediate, no delay
        }
    }
    // ...
}

There is no code in the SDK that:

• Buffers assistant.reasoning_delta events
• Differentiates behavior based on model name (e.g., claude-sonnet-5 vs claude-sonnet-4.5)
• Accumulates reasoning content before emitting

The string "claude-sonnet-5" does not appear anywhere in the repository.

Conclusion

If assistant.reasoning_delta events are not being emitted in real-time for a specific model, the root cause would be in the Copilot CLI server or the upstream model router (CAPI), not the SDK. The SDK emits exactly what it receives from the server.

This has been relabeled as a question since the SDK is working as designed. If you are seeing this behavior, I'd suggest:

1. Confirming which version of the Copilot CLI you're using
2. Checking whether the CLI server itself is emitting assistant.reasoning_delta events for that model (you can subscribe to all events to verify)
3. Filing the issue against the Copilot service/CLI if the events are missing from the server stream

Happy to help clarify how the event subscription system works if needed.

Generated by Bug Handler for #2346 · sonnet46 35.2 AIC · ⌖ 5.22 AIC · ⊞ 5.9K · ◷