fix: handle None choices in ChatCompletion response (provider error payloads)

When a provider (e.g. OpenRouter) returns an error payload, the ChatCompletion
response object has choices=None. Previously this caused a silent no-op or a
downstream TypeError. Now a ModelBehaviorError is raised before choices[0] is
accessed, including the provider's error field in the message when present.

Author: Spectual

src/agents/models/openai_chatcompletions.py

@@ -20,7 +20,7 @@
 
 from .. import _debug
 from ..agent_output import AgentOutputSchemaBase
-from ..exceptions import UserError
+from ..exceptions import ModelBehaviorError, UserError
 from ..handoffs import Handoff
 from ..items import ModelResponse, TResponseInputItem, TResponseStreamEvent
 from ..logger import logger
@@ -131,6 +131,16 @@ async def get_response(
                 prompt=prompt,
             )
 
+            if not response.choices:
+                provider_error = getattr(response, "error", None)
+                error_details = (
+                    f": {provider_error}" if provider_error is not None else ""
+                )
+                raise ModelBehaviorError(
+                    f"ChatCompletion response has no choices (possible provider error payload)"
+                    f"{error_details}"
+                )
+
             message: ChatCompletionMessage | None = None
             first_choice: Choice | None = None
             if response.choices and len(response.choices) > 0: