Skip to content

Self-evolving Context Database for AI Agents. Unify Agent Memory, Knowledge RAG and Skills.

Fix Codex VLM response stream parsing (#2355)

Zayn Jarvisfd101200f58a
diff --git a/openviking/models/vlm/backends/codex_responses_adapter.py b/openviking/models/vlm/backends/codex_responses_adapter.py
index 5388be29e..43e94ac19 100644
--- a/openviking/models/vlm/backends/codex_responses_adapter.py
+++ b/openviking/models/vlm/backends/codex_responses_adapter.py
@@ -185,6 +185,35 @@ def _build_chat_completion_like_response(final_response: Any, model: str) -> Any
)
+def _build_final_response_from_stream_events(
+ completed_response: Any,
+ collected_output_items: List[Any],
+ collected_text_deltas: List[str],
+ has_function_calls: bool,
+) -> Any:
+ output = getattr(completed_response, "output", None) if completed_response is not None else None
+ if output:
+ return completed_response
+
+ fallback_output: List[Any] = []
+ if collected_output_items:
+ fallback_output = list(collected_output_items)
+ elif collected_text_deltas and not has_function_calls:
+ fallback_output = [
+ SimpleNamespace(
+ type="message",
+ role="assistant",
+ status="completed",
+ content=[SimpleNamespace(type="output_text", text="".join(collected_text_deltas))],
+ )
+ ]
+
+ return SimpleNamespace(
+ output=fallback_output,
+ usage=getattr(completed_response, "usage", None) if completed_response is not None else None,
+ )
+
+
class CodexCompletionsAdapter:
def __init__(self, client_factory: Callable[[], Any], model: str):
self._client_factory = client_factory
@@ -218,7 +247,9 @@ class CodexCompletionsAdapter:
collected_output_items: List[Any] = []
collected_text_deltas: List[str] = []
has_function_calls = False
- with client.responses.stream(**response_kwargs) as stream:
+ completed_response = None
+ stream = client.responses.create(**response_kwargs, stream=True)
+ try:
for event in stream:
event_type = getattr(event, "type", "")
if event_type == "response.output_item.done":
@@ -233,22 +264,19 @@ class CodexCompletionsAdapter:
continue
if "function_call" in event_type:
has_function_calls = True
- final_response = stream.get_final_response()
- output = getattr(final_response, "output", None)
- if not output:
- if collected_output_items:
- final_response.output = list(collected_output_items)
- elif collected_text_deltas and not has_function_calls:
- final_response.output = [
- SimpleNamespace(
- type="message",
- role="assistant",
- status="completed",
- content=[
- SimpleNamespace(type="output_text", text="".join(collected_text_deltas))
- ],
- )
- ]
+ continue
+ if event_type == "response.completed":
+ completed_response = getattr(event, "response", None)
+ finally:
+ close = getattr(stream, "close", None)
+ if callable(close):
+ close()
+ final_response = _build_final_response_from_stream_events(
+ completed_response,
+ collected_output_items,
+ collected_text_deltas,
+ has_function_calls,
+ )
return _build_chat_completion_like_response(final_response, model)
def create(self, **kwargs) -> Any:
diff --git a/openviking/models/vlm/backends/codex_vlm.py b/openviking/models/vlm/backends/codex_vlm.py
index 60bcc5129..b7059fb02 100644
--- a/openviking/models/vlm/backends/codex_vlm.py
+++ b/openviking/models/vlm/backends/codex_vlm.py
@@ -13,8 +13,8 @@ The complexity in this file arises from the need to shim/adapt standard Chat Com
requests (used by OpenViking) into Responses API requests. This involves:
1. Converting `text` and `image_url` parts into `input_text` and `input_image`.
2. Adapting tool calls and schemas.
-3. Translating the `client.responses.stream` event stream back into a format
- compatible with standard Chat Completion responses.
+3. Translating raw `client.responses.create(stream=True)` events back into a
+ format compatible with standard Chat Completion responses.
"""
from __future__ import annotations
@@ -91,6 +91,18 @@ class CodexVLM(OpenAIVLM):
self.api_base = explicit_api_base or credentials["base_url"]
return self.api_key, self.api_base
+ def _build_text_kwargs(self, *args, **kwargs):
+ request_kwargs = super()._build_text_kwargs(*args, **kwargs)
+ if self.stream:
+ request_kwargs["stream"] = True
+ return request_kwargs
+
+ def _build_vision_kwargs(self, *args, **kwargs):
+ request_kwargs = super()._build_vision_kwargs(*args, **kwargs)
+ if self.stream:
+ request_kwargs["stream"] = True
+ return request_kwargs
+
def get_client(self):
if openai is None:
raise ImportError("Please install openai: pip install openai")
diff --git a/tests/models/vlm/test_timeout_config.py b/tests/models/vlm/test_timeout_config.py
index 15613792b..e6bc71ca3 100644
--- a/tests/models/vlm/test_timeout_config.py
+++ b/tests/models/vlm/test_timeout_config.py
@@ -131,7 +131,7 @@ def test_codex_vlm_propagates_config_timeout():
)
with mock.patch("openviking.models.vlm.backends.codex_vlm.openai.OpenAI") as fake:
- fake.return_value.responses.stream.return_value = _MockResponsesStream(
+ fake.return_value.responses.create.return_value = _MockResponsesStream(
_build_final_response("timeout ok")
)
assert vlm.get_completion("hello") == "timeout ok"
diff --git a/tests/unit/test_codex_vlm.py b/tests/unit/test_codex_vlm.py
index b96a714ee..35c12c187 100644
--- a/tests/unit/test_codex_vlm.py
+++ b/tests/unit/test_codex_vlm.py
@@ -17,6 +17,7 @@ from openviking_cli.utils.config.vlm_config import VLMConfig
class _MockResponsesStream:
def __init__(self, final_response):
self._final_response = final_response
+ self.closed = False
def __enter__(self):
return self
@@ -25,12 +26,32 @@ class _MockResponsesStream:
return False
def __iter__(self):
- return iter(())
+ events = [
+ SimpleNamespace(type="response.output_item.done", item=item)
+ for item in self._final_response.output or []
+ ]
+ events.append(SimpleNamespace(type="response.completed", response=self._final_response))
+ return iter(events)
+
+ def close(self):
+ self.closed = True
def get_final_response(self):
return self._final_response
+class _MockResponsesEventStream:
+ def __init__(self, events):
+ self._events = events
+ self.closed = False
+
+ def __iter__(self):
+ return iter(self._events)
+
+ def close(self):
+ self.closed = True
+
+
def _build_final_response(text: str):
return SimpleNamespace(
output=[
@@ -45,7 +66,7 @@ def _build_final_response(text: str):
def _build_fake_openai_client(text: str):
client = MagicMock()
- client.responses.stream.return_value = _MockResponsesStream(_build_final_response(text))
+ client.responses.create.return_value = _MockResponsesStream(_build_final_response(text))
return client
@@ -57,7 +78,7 @@ def test_codex_text_completion_uses_responses_api(mock_resolve, mock_openai_clas
"base_url": "https://chatgpt.com/backend-api/codex",
}
mock_real_client = MagicMock()
- mock_real_client.responses.stream.return_value = _MockResponsesStream(
+ mock_real_client.responses.create.return_value = _MockResponsesStream(
_build_final_response("hello from codex")
)
mock_openai_class.return_value = mock_real_client
@@ -66,10 +87,12 @@ def test_codex_text_completion_uses_responses_api(mock_resolve, mock_openai_clas
result = vlm.get_completion("hello")
assert result == "hello from codex"
- call_kwargs = mock_real_client.responses.stream.call_args.kwargs
+ call_kwargs = mock_real_client.responses.create.call_args.kwargs
assert call_kwargs["model"] == "gpt-5.3-codex"
assert call_kwargs["input"] == [{"role": "user", "content": "hello"}]
+ assert call_kwargs["stream"] is True
assert "messages" not in call_kwargs
+ mock_real_client.responses.stream.assert_not_called()
@patch("openviking.models.vlm.backends.codex_vlm.openai.OpenAI")
@@ -80,7 +103,7 @@ def test_codex_vision_completion_converts_images(mock_resolve, mock_openai_class
"base_url": "https://chatgpt.com/backend-api/codex",
}
mock_real_client = MagicMock()
- mock_real_client.responses.stream.return_value = _MockResponsesStream(
+ mock_real_client.responses.create.return_value = _MockResponsesStream(
_build_final_response("image result")
)
mock_openai_class.return_value = mock_real_client
@@ -89,7 +112,7 @@ def test_codex_vision_completion_converts_images(mock_resolve, mock_openai_class
result = vlm.get_vision_completion("describe", [b"\x89PNG\r\n\x1a\n0000"])
assert result == "image result"
- call_kwargs = mock_real_client.responses.stream.call_args.kwargs
+ call_kwargs = mock_real_client.responses.create.call_args.kwargs
content = call_kwargs["input"][0]["content"]
assert content[0]["type"] == "input_image"
assert content[0]["image_url"].startswith("data:image/png;base64,")
@@ -108,7 +131,7 @@ async def test_codex_async_client_defers_runtime_credential_resolution(
"base_url": "https://chatgpt.com/backend-api/codex",
}
mock_real_client = MagicMock()
- mock_real_client.responses.stream.return_value = _MockResponsesStream(
+ mock_real_client.responses.create.return_value = _MockResponsesStream(
_build_final_response("async hello")
)
mock_openai_class.return_value = mock_real_client
@@ -126,6 +149,42 @@ async def test_codex_async_client_defers_runtime_credential_resolution(
mock_resolve.assert_called_once()
+@patch("openviking.models.vlm.backends.codex_vlm.openai.OpenAI")
+@patch("openviking.models.vlm.backends.codex_vlm.resolve_codex_runtime_credentials")
+async def test_codex_async_stream_uses_text_deltas_when_completed_output_is_none(
+ mock_resolve,
+ mock_openai_class,
+):
+ mock_resolve.return_value = {
+ "api_key": "oauth-token",
+ "base_url": "https://chatgpt.com/backend-api/codex",
+ }
+ final_response = SimpleNamespace(
+ output=None,
+ usage=SimpleNamespace(input_tokens=11, output_tokens=7, total_tokens=18),
+ )
+ mock_stream = _MockResponsesEventStream(
+ [
+ SimpleNamespace(type="response.output_text.delta", delta="overview "),
+ SimpleNamespace(type="response.output_text.delta", delta="ready"),
+ SimpleNamespace(type="response.completed", response=final_response),
+ ]
+ )
+ mock_real_client = MagicMock()
+ mock_real_client.responses.create.return_value = mock_stream
+ mock_openai_class.return_value = mock_real_client
+
+ vlm = CodexVLM({"provider": "openai-codex", "model": "gpt-5.3-codex"})
+ response = await vlm.get_completion_async("summarize")
+
+ assert response == "overview ready"
+ assert mock_stream.closed is True
+ call_kwargs = mock_real_client.responses.create.call_args.kwargs
+ assert call_kwargs["stream"] is True
+ mock_real_client.responses.stream.assert_not_called()
+
+
@patch("openviking.models.vlm.backends.codex_auth.has_codex_auth_available", return_value=True)
def test_vlm_config_accepts_codex_without_api_key(_mock_auth_available):
config = VLMConfig(provider="openai-codex", model="gpt-5.3-codex")
@@ -384,7 +443,7 @@ def test_codex_streaming_is_rejected(mock_resolve, mock_openai_class):
with pytest.raises(NotImplementedError, match="Streaming is not supported"):
vlm.get_completion("hello")
- mock_real_client.responses.stream.assert_not_called()
+ mock_real_client.responses.create.assert_not_called()
@patch("openviking.models.vlm.backends.codex_vlm.openai.OpenAI")
@@ -431,7 +490,7 @@ def test_codex_translates_tool_history_into_responses_input(mock_resolve, mock_o
"base_url": "https://chatgpt.com/backend-api/codex",
}
mock_real_client = MagicMock()
- mock_real_client.responses.stream.return_value = _MockResponsesStream(
+ mock_real_client.responses.create.return_value = _MockResponsesStream(
_build_final_response("final answer")
)
mock_openai_class.return_value = mock_real_client
@@ -476,7 +535,7 @@ def test_codex_translates_tool_history_into_responses_input(mock_resolve, mock_o
)
assert response.content == "final answer"
- input_items = mock_real_client.responses.stream.call_args.kwargs["input"]
+ input_items = mock_real_client.responses.create.call_args.kwargs["input"]
assert input_items == [
{"role": "user", "content": "What is the weather?"},
{