Did they fix the streaming output in the API? Last check was buggy.

I've been evaluating CrewAI's API for a potential integration where low-latency, real-time feedback is critical for user experience. My last deep dive, about two months ago, revealed significant issues with their streaming response implementation. The stream would often hang or deliver incomplete JSON fragments, making it unusable for production.

My primary concern is the performance overhead and reliability. When I tested it, the implementation seemed to add substantial latency compared to a simple synchronous request, which defeats the purpose. I was using a straightforward setup with their `crew.kickoff()` method in streaming mode:

```python
# Simplified test from my previous evaluation
result = crew.kickoff(inputs={'topic': 'test'}, stream=True)
for chunk in result:
# Chunks would often be malformed or stall here
print(chunk)
```

Has anyone conducted recent tests? I'm particularly interested in:
* The stability of the HTTP stream connection—does it now close cleanly?
* The structure of the JSON chunks—are they now guaranteed to be parseable?
* Any measurable latency penalty versus non-streaming when considering time-to-first-token?

I'm hoping they've addressed the underlying issues, perhaps by moving away from a custom implementation to a more robust server-sent events (SSE) or similar standard. My stack is Go with PostgreSQL for state, and I need to know if I can reliably pipe this stream to our own WebSocket layer without constant timeouts and reconnection logic.

-- latency