Silence annoying OpenAI client shutdown error (#4157)

skyvern/forge/forge_app.py

@@ -9,6 +9,7 @@ from openai import AsyncAzureOpenAI, AsyncOpenAI
 from skyvern.config import Settings
 from skyvern.forge.agent import ForgeAgent
 from skyvern.forge.agent_functions import AgentFunction
+from skyvern.forge.forge_openai_client import ForgeAsyncHttpxClientWrapper
 from skyvern.forge.sdk.api.azure import AzureClientFactory
 from skyvern.forge.sdk.api.llm.api_handler_factory import LLMAPIHandlerFactory
 from skyvern.forge.sdk.api.llm.models import LLMAPIHandler
@@ -95,7 +96,10 @@ def create_forge_app() -> ForgeApp:
     app.EXPERIMENTATION_PROVIDER = NoOpExperimentationProvider()
 
     app.LLM_API_HANDLER = LLMAPIHandlerFactory.get_llm_api_handler(settings.LLM_KEY)
-    app.OPENAI_CLIENT = AsyncOpenAI(api_key=settings.OPENAI_API_KEY or "")
+    app.OPENAI_CLIENT = AsyncOpenAI(
+        api_key=settings.OPENAI_API_KEY or "",
+        http_client=ForgeAsyncHttpxClientWrapper(),
+    )
     if settings.ENABLE_AZURE_CUA:
         app.OPENAI_CLIENT = AsyncAzureOpenAI(
             api_key=settings.AZURE_CUA_API_KEY,
@@ -113,6 +117,7 @@ def create_forge_app() -> ForgeApp:
         app.UI_TARS_CLIENT = AsyncOpenAI(
             api_key=settings.VOLCENGINE_API_KEY,
             base_url=settings.VOLCENGINE_API_BASE,
+            http_client=ForgeAsyncHttpxClientWrapper(),
         )
 
     app.SECONDARY_LLM_API_HANDLER = LLMAPIHandlerFactory.get_llm_api_handler(

skyvern/forge/forge_openai_client.py

@@ -0,0 +1,24 @@
+import asyncio
+
+from openai import DefaultAsyncHttpxClient
+
+
+class ForgeAsyncHttpxClientWrapper(DefaultAsyncHttpxClient):
+    """
+    Wrapper around OpenAI's AsyncHttpxClientWrapper to mask teardown races.
+
+    The upstream `__del__` checks `self.is_closed`, but during interpreter
+    shutdown httpx internals may already be None, which raises:
+
+    AttributeError: 'NoneType' object has no attribute 'CLOSED'
+
+    We defensively swallow that destructor error so shutdown logs stay clean.
+    """
+
+    def __del__(self) -> None:
+        try:
+            if self.is_closed:
+                return
+            asyncio.get_running_loop().create_task(self.aclose())
+        except Exception:
+            pass

skyvern/forge/sdk/api/llm/api_handler_factory.py

@@ -18,6 +18,7 @@ from pydantic import BaseModel
 from skyvern.config import settings
 from skyvern.exceptions import SkyvernContextWindowExceededError
 from skyvern.forge import app
+from skyvern.forge.forge_openai_client import ForgeAsyncHttpxClientWrapper
 from skyvern.forge.sdk.api.llm.config_registry import LLMConfigRegistry
 from skyvern.forge.sdk.api.llm.exceptions import (
     DuplicateCustomLLMProviderError,
@@ -1100,7 +1101,11 @@ class LLMCaller:
         self.openai_client = None
         if self.llm_key.startswith("openrouter/"):
             self.llm_key = self.llm_key.replace("openrouter/", "")
-            self.openai_client = AsyncOpenAI(api_key=settings.OPENROUTER_API_KEY, base_url=settings.OPENROUTER_API_BASE)
+            self.openai_client = AsyncOpenAI(
+                api_key=settings.OPENROUTER_API_KEY,
+                base_url=settings.OPENROUTER_API_BASE,
+                http_client=ForgeAsyncHttpxClientWrapper(),
+            )
 
     def add_tool_result(self, tool_result: dict[str, Any]) -> None:
         self.current_tool_results.append(tool_result)
@@ -1513,7 +1518,8 @@ class LLMCaller:
         if isinstance(response, UITarsResponse):
             ui_tars_usage = response.usage
             return LLMCallStats(
-                llm_cost=0,  # TODO: calculate the cost according to the price: https://www.volcengine.com/docs/82379/1544106
+                llm_cost=0,
+                # TODO: calculate the cost according to the price: https://www.volcengine.com/docs/82379/1544106
                 input_tokens=ui_tars_usage.get("prompt_tokens", 0),
                 output_tokens=ui_tars_usage.get("completion_tokens", 0),
                 cached_tokens=0,  # only part of model support cached tokens