…rror: Event loop is closed"

Issue

- #1009

Summary:

Fix RuntimeError: Event loop is closed in OpenAI client initialization by separating the HTTP/2 probe from production client creation. The probe thread previously returned the AsyncOpenAI client itself, leaving its SSL connections bound to a closed event loop. The production
client is now created after the thread exits, in the caller's event loop.

Details:

The HTTP/2 probe ran in a background thread with its own event loop,
created an AsyncOpenAI client, and returned it directly. Closing the
thread's event loop while the client still held open keep-alive SSL
connections caused `RuntimeError: Event loop is closed` whenever the
client was later used in the caller's (different) event loop.

Bug Fix: Decouple probe from production client lifetime

- Rename probe_and_initialize() → probe_http2(), which now returns a
  bool instead of an AsyncOpenAI client
- Explicitly close the probe client in a finally block so all
  connections are drained before the thread's event loop is closed
- Create the production AsyncOpenAI client after the thread exits,
  in the caller's event loop context, so its connections bind to the
  correct running loop

HTTP/2 Fallback Logic

- Preserve HTTP/2 vs HTTP/1.1 selection: use_http2=True creates a
  default AsyncOpenAI(); use_http2=False creates one with an explicit
  httpx.AsyncClient(http2=False) transport
- No change to probe timeout (5s) or overall initialization timeout (15s)

…rror: Event loop is closed"

Issue

- #1009

Summary:

Fix two CI-only failures caused by environment variable bleed-through in the CI test pipeline: Docker containers receiving localhost as the OpenSearch hostname (instead of the opensearch service name), and the backend crashing with invalid literal for int() with base 10: '' due
to empty Langflow retry env vars being injected into the container.

Detials:

The GitHub Actions workflow sets OPENSEARCH_HOST=localhost as a step-level
env var so the in-process pytest backend can reach OpenSearch via Docker
port mapping. This bled into docker-compose, overriding the opensearch
service-name default and breaking all OpenSearch calls from inside the
backend and Langflow containers. Separately, LANGFLOW_KEY_RETRIES and
LANGFLOW_KEY_RETRY_DELAY had no :-default fallbacks in docker-compose.yml,
so when unset in CI they were injected as empty strings, causing
int("") to raise ValueError inside get_langflow_api_key().

CI Configuration: Docker container startup

- Pin OPENSEARCH_HOST=opensearch when invoking docker-compose up in the
  test-ci and test-ci-local Makefile targets, isolating it from the
  OPENSEARCH_HOST=localhost value the surrounding CI shell sets for pytest

docker-compose.yml: Environment variable defaults

- Add :-15 default for LANGFLOW_KEY_RETRIES so unset env produces "15"
  rather than an empty string that crashes int() in settings.py
- Add :-2.0 default for LANGFLOW_KEY_RETRY_DELAY for the same reason;
  defaults match the hardcoded fallbacks already in settings.py:227-228