Fix worker status for unusable terminal backend responses

src/guidellm/backends/openai/http.py

@@ -483,6 +483,7 @@ async def _resolve_non_streaming(
         response.raise_for_status()
         data = response.json()
         gen_response = request_handler.compile_non_streaming(request, arguments, data)
+        request_handler.post_validation(gen_response)
         yield gen_response, request_info
         self._check_tool_call_expectations(request, gen_response)
 
@@ -553,6 +554,7 @@ async def _resolve_streaming(
 
             request_info.timings.request_end = time.time()
             gen_response = request_handler.compile_streaming(request, arguments)
+            request_handler.post_validation(gen_response)
             self._check_tool_call_expectations(request, gen_response)
             yield gen_response, request_info
         except asyncio.CancelledError as err:

src/guidellm/backends/openai/request_handlers.py

@@ -126,6 +126,18 @@ def compile_streaming(
         """
         ...
 
+    def post_validation(self, response: GenerationResponse) -> None:
+        """Validate a compiled response before returning it.
+
+        Default implementation is permissive (no-op). Handlers override
+        this to reject responses that lack usable output for their
+        endpoint type.
+
+        :param response: The compiled generation response to validate.
+        :raises ValueError: If the response is unusable.
+        """
+        ...
+
 
 class OpenAIRequestHandlerFactory(RegistryMixin[type[OpenAIRequestHandler]]):
     """
@@ -272,6 +284,18 @@ def compile_streaming(
         """
         ...
 
+    def post_validation(self, response: GenerationResponse) -> None:
+        """Validate a compiled response before returning it.
+
+        Default implementation is permissive (no-op). Handlers override
+        this to reject responses that lack usable output for their
+        endpoint type.
+
+        :param response: The compiled generation response to validate.
+        :raises ValueError: If the response is unusable.
+        """
+        ...
+
 
 class OpenAIWSRequestHandlerFactory(RegistryMixin["type[OpenAIWSRequestHandler]"]):
     """Factory for registering and creating WebSocket request handlers by path."""
@@ -317,6 +341,24 @@ def _apply_tool_call_metrics(
         output_metrics.mixed_content_tool_tokens = output_metrics.text_tokens
 
 
+def _validate_text_response(response: GenerationResponse) -> None:
+    """Reject a compiled response that has no usable output.
+
+    A response is considered usable if it has non-empty text, tool calls,
+    or output tokens. Used by text/chat completions and responses handlers.
+
+    :param response: The compiled generation response to validate.
+    :raises ValueError: If the response contains no usable output.
+    """
+    has_text = bool(response.text and response.text.strip())
+    has_tool_calls = bool(response.tool_calls)
+    output_tokens = response.output_metrics.total_tokens or 0
+    if not has_text and not has_tool_calls and output_tokens <= 0:
+        raise ValueError(
+            "[UNUSABLE_BACKEND_RESPONSE] backend resolved with empty response payload"
+        )
+
+
 _DEFAULT_REASONING_TEMPLATE = "<think>{reasoning}</think>"
 
 
@@ -571,6 +613,10 @@ def compile_streaming(
             output_metrics=output_metrics,
         )
 
+    def post_validation(self, response: GenerationResponse) -> None:
+        """Reject responses with no text, tool calls, or output tokens."""
+        _validate_text_response(response)
+
     def extract_line_data(self, line: str) -> dict[str, Any] | None:
         """
         Extract JSON data from a streaming response line.
@@ -1848,6 +1894,10 @@ def compile_streaming(
             streaming_reasoning_texts=self.streaming_reasoning_texts,
         )
 
+    def post_validation(self, response: GenerationResponse) -> None:
+        """Reject responses with no text, tool calls, or output tokens."""
+        _validate_text_response(response)
+
     def extract_line_data(self, line: str) -> dict[str, Any] | None:
         """Parse a Responses API SSE line.
 
@@ -2147,6 +2197,9 @@ class PoolingRequestHandler(ChatCompletionsRequestHandler):
     pooling-specific request structure with nested data fields.
     """
 
+    def post_validation(self, response: GenerationResponse) -> None:  # noqa: ARG002
+        """Pooling responses produce non-text output; skip validation."""
+
     def format(
         self,
         data: GenerationRequest,

src/guidellm/backends/openai/websocket.py

@@ -506,7 +506,9 @@ async def resolve(  # type: ignore[override, misc]  # noqa: C901, PLR0912, PLR09
                                 f"(last type={event.get('type')!r})."
                             )
 
-                yield handler.compile_streaming(request, arguments), request_info
+                compiled = handler.compile_streaming(request, arguments)
+                handler.post_validation(compiled)
+                yield compiled, request_info
 
         except asyncio.CancelledError as err:
             yield handler.compile_streaming(request, arguments), request_info

tests/unit/backends/openai/test_http.py

@@ -449,7 +449,11 @@ def mock_fail(*args, **kwargs):
     @async_timeout(10.0)
     async def test_resolve_with_history(self, httpx_mock: HTTPXMock):
         """Test resolve method handles conversation history."""
-        backend = _make_backend(target="http://test", request_format="/v1/completions")
+        backend = _make_backend(
+            target="http://test",
+            request_format="/v1/completions",
+            stream=False,
+        )
 
         # Mock the models endpoint
         httpx_mock.add_response(