Description

_is_openai_client_closed in run_agent.py always returns True for real OpenAI client instances due to a type mismatch with the openai SDK.

Root cause

run_agent.py, AIAgent._is_openai_client_closed() (around line 3488):

@staticmethod
def _is_openai_client_closed(client: Any) -> bool:
    from unittest.mock import Mock
    if isinstance(client, Mock):
        return False
    if bool(getattr(client, "is_closed", False)):   # <-- BUG
        return True
    http_client = getattr(client, "_client", None)
    return bool(getattr(http_client, "is_closed", False))

In the openai SDK (tested on v2.26.0, constraint is >=2.21.0,<3), OpenAI.is_closed is a method, not a property:

# openai/_base_client.py
class SyncAPIClient:
    def is_closed(self) -> bool:
        return self._client.is_closed

So getattr(client, "is_closed", False) returns the bound method object, and bool(<bound method>) is always True. The check short-circuits before reaching the _client.is_closed fallback (which works correctly since httpx.Client.is_closed is a plain bool property).

Impact

Every API call:

1. Falsely detects the shared client as closed
2. Creates a new shared client (TCP + TLS handshake to provider)
3. Closes the old healthy client
4. Creates a per-request client (another TCP + TLS handshake)
5. Closes the per-request client after the stream

Steps 1-3 are entirely unnecessary — the shared client is alive. This adds ~100-200ms of connection overhead per API call.

Why tests don't catch it

test_openai_client_lifecycle.py uses FakeRequestClient which doesn't define is_closed as an attribute. So getattr(fake_client, "is_closed", False) returns the default False, falls through to _client.is_closed (a SimpleNamespace attribute), and works correctly. The bug only manifests with real OpenAI instances.

Reproduction

from openai import OpenAI

client = OpenAI(api_key="test", base_url="https://openrouter.ai/api/v1")
val = getattr(client, "is_closed", False)
print(repr(val))   # <bound method SyncAPIClient.is_closed of ...>
print(bool(val))    # True  <-- always truthy, even though client is open

Suggested fix

In run_agent.py, the _is_openai_client_closed static method should call is_closed() when it's callable instead of checking the truthiness of the method object:

@staticmethod
def _is_openai_client_closed(client: Any) -> bool:
    from unittest.mock import Mock
    if isinstance(client, Mock):
        return False
    is_closed = getattr(client, "is_closed", None)
    if callable(is_closed):
        if is_closed():
            return True
    elif bool(is_closed):
        return True
    http_client = getattr(client, "_client", None)
    if http_client is not None:
        return bool(getattr(http_client, "is_closed", False))
    return False

The test in test_openai_client_lifecycle.py should also add a case using a real OpenAI client (or a fake with is_closed as a method) to prevent regression.

Environment

• openai SDK: 2.26.0 (bug applies to any version where is_closed is a method — all of >=2.21.0)
• hermes-agent: v2026.3.23
• Observed via gateway adapter (custom AgentRunner wrapping AIAgent)