feat(ai): langchain cached and reasoning tokens (#258)

* fix: reasoning and cached tokens

* test: new flows

* fix: missing field

* chore: bump

* fix: make sure we send write/read/reasoning tokens

.github/workflows/call-flags-project-board.yml

@@ -0,0 +1,17 @@
+# This workflow is used to call the flags-project-board workflow when a pull request is opened, ready for review, review requested, synchronized, converted to draft, or reopened.
+# It is used to update the feature flags project board with the pull request information.
+
+name: Call Feature Flags Project Workflow
+
+on:
+    pull_request:
+        types: [opened, ready_for_review, review_requested, synchronize, converted_to_draft, reopened]
+
+jobs:
+    call-flags-project:
+        uses: PostHog/.github/.github/workflows/flags-project-board.yml@main
+        with:
+            pr_number: ${{ github.event.pull_request.number }}
+            pr_node_id: ${{ github.event.pull_request.node_id }}
+            is_draft: ${{ github.event.pull_request.draft }}
+        secrets: inherit

CHANGELOG.md

@@ -1,3 +1,16 @@
+## 4.9.0 - 2025-06-13
+
+- feat(ai): track reasoning and cache tokens in the LangChain callback
+
+## 4.8.0 - 2025-06-10
+
+- fix: export scoped, rather than tracked, decorator
+- feat: allow use of contexts without error tracking
+
+## 4.7.0 - 2025-06-10
+
+- feat: add support for parse endpoint in responses API (no longer beta)
+
 ## 4.6.2 - 2025-06-09
 
 - fix: replace `import posthog` with direct method imports

posthog/__init__.py

@@ -15,7 +15,7 @@ new_context = new_context
 tag = tag
 get_tags = get_tags
 clear_tags = clear_tags
-tracked = scoped
+scoped = scoped
 
 """Settings."""
 api_key = None  # type: Optional[str]

posthog/ai/langchain/callbacks.py

@@ -14,7 +14,6 @@ from typing import (
     List,
     Optional,
     Sequence,
-    Tuple,
     Union,
     cast,
 )
@@ -569,9 +568,14 @@ class CallbackHandler(BaseCallbackHandler):
             event_properties["$ai_is_error"] = True
         else:
             # Add usage
-            input_tokens, output_tokens = _parse_usage(output)
-            event_properties["$ai_input_tokens"] = input_tokens
-            event_properties["$ai_output_tokens"] = output_tokens
+            usage = _parse_usage(output)
+            event_properties["$ai_input_tokens"] = usage.input_tokens
+            event_properties["$ai_output_tokens"] = usage.output_tokens
+            event_properties["$ai_cache_creation_input_tokens"] = (
+                usage.cache_write_tokens
+            )
+            event_properties["$ai_cache_read_input_tokens"] = usage.cache_read_tokens
+            event_properties["$ai_reasoning_tokens"] = usage.reasoning_tokens
 
             # Generation results
             generation_result = output.generations[-1]
@@ -647,9 +651,18 @@ def _convert_message_to_dict(message: BaseMessage) -> Dict[str, Any]:
     return message_dict
 
 
+@dataclass
+class ModelUsage:
+    input_tokens: Optional[int]
+    output_tokens: Optional[int]
+    cache_write_tokens: Optional[int]
+    cache_read_tokens: Optional[int]
+    reasoning_tokens: Optional[int]
+
+
 def _parse_usage_model(
-    usage: Union[BaseModel, Dict],
-) -> Tuple[Union[int, None], Union[int, None]]:
+    usage: Union[BaseModel, dict],
+) -> ModelUsage:
     if isinstance(usage, BaseModel):
         usage = usage.__dict__
 
@@ -657,15 +670,23 @@ def _parse_usage_model(
         # https://pypi.org/project/langchain-anthropic/ (works also for Bedrock-Anthropic)
         ("input_tokens", "input"),
         ("output_tokens", "output"),
+        ("cache_creation_input_tokens", "cache_write"),
+        ("cache_read_input_tokens", "cache_read"),
         # https://cloud.google.com/vertex-ai/generative-ai/docs/multimodal/get-token-count
         ("prompt_token_count", "input"),
         ("candidates_token_count", "output"),
+        ("cached_content_token_count", "cache_read"),
+        ("thoughts_token_count", "reasoning"),
         # Bedrock: https://docs.aws.amazon.com/bedrock/latest/userguide/monitoring-cw.html#runtime-cloudwatch-metrics
         ("inputTokenCount", "input"),
         ("outputTokenCount", "output"),
+        ("cacheCreationInputTokenCount", "cache_write"),
+        ("cacheReadInputTokenCount", "cache_read"),
         # Bedrock Anthropic
         ("prompt_tokens", "input"),
         ("completion_tokens", "output"),
+        ("cache_creation_input_tokens", "cache_write"),
+        ("cache_read_input_tokens", "cache_read"),
         # langchain-ibm https://pypi.org/project/langchain-ibm/
         ("input_token_count", "input"),
         ("generated_token_count", "output"),
@@ -683,13 +704,45 @@ def _parse_usage_model(
 
             parsed_usage[type_key] = final_count
 
-    return parsed_usage.get("input"), parsed_usage.get("output")
+    # Caching (OpenAI & langchain 0.3.9+)
+    if "input_token_details" in usage and isinstance(
+        usage["input_token_details"], dict
+    ):
+        parsed_usage["cache_write"] = usage["input_token_details"].get("cache_creation")
+        parsed_usage["cache_read"] = usage["input_token_details"].get("cache_read")
+
+    # Reasoning (OpenAI & langchain 0.3.9+)
+    if "output_token_details" in usage and isinstance(
+        usage["output_token_details"], dict
+    ):
+        parsed_usage["reasoning"] = usage["output_token_details"].get("reasoning")
+
+    field_mapping = {
+        "input": "input_tokens",
+        "output": "output_tokens",
+        "cache_write": "cache_write_tokens",
+        "cache_read": "cache_read_tokens",
+        "reasoning": "reasoning_tokens",
+    }
+    return ModelUsage(
+        **{
+            dataclass_key: parsed_usage.get(mapped_key) or 0
+            for mapped_key, dataclass_key in field_mapping.items()
+        },
+    )
 
 
-def _parse_usage(response: LLMResult):
+def _parse_usage(response: LLMResult) -> ModelUsage:
     # langchain-anthropic uses the usage field
     llm_usage_keys = ["token_usage", "usage"]
-    llm_usage: Tuple[Union[int, None], Union[int, None]] = (None, None)
+    llm_usage: ModelUsage = ModelUsage(
+        input_tokens=None,
+        output_tokens=None,
+        cache_write_tokens=None,
+        cache_read_tokens=None,
+        reasoning_tokens=None,
+    )
+
     if response.llm_output is not None:
         for key in llm_usage_keys:
             if response.llm_output.get(key):

posthog/ai/openai/openai.py

@@ -230,6 +230,42 @@ class WrappedResponses:
                 groups=posthog_groups,
             )
 
+    def parse(
+        self,
+        posthog_distinct_id: Optional[str] = None,
+        posthog_trace_id: Optional[str] = None,
+        posthog_properties: Optional[Dict[str, Any]] = None,
+        posthog_privacy_mode: bool = False,
+        posthog_groups: Optional[Dict[str, Any]] = None,
+        **kwargs: Any,
+    ):
+        """
+        Parse structured output using OpenAI's 'responses.parse' method, but also track usage in PostHog.
+
+        Args:
+            posthog_distinct_id: Optional ID to associate with the usage event.
+            posthog_trace_id: Optional trace UUID for linking events.
+            posthog_properties: Optional dictionary of extra properties to include in the event.
+            posthog_privacy_mode: Whether to anonymize the input and output.
+            posthog_groups: Optional dictionary of groups to associate with the event.
+            **kwargs: Any additional parameters for the OpenAI Responses Parse API.
+
+        Returns:
+            The response from OpenAI's responses.parse call.
+        """
+        return call_llm_and_track_usage(
+            posthog_distinct_id,
+            self._client._ph_client,
+            "openai",
+            posthog_trace_id,
+            posthog_properties,
+            posthog_privacy_mode,
+            posthog_groups,
+            self._client.base_url,
+            self._original.parse,
+            **kwargs,
+        )
+
 
 class WrappedChat:
     """Wrapper for OpenAI chat that tracks usage in PostHog."""

posthog/ai/openai/openai_async.py

@@ -230,6 +230,42 @@ class WrappedResponses:
                 groups=posthog_groups,
             )
 
+    async def parse(
+        self,
+        posthog_distinct_id: Optional[str] = None,
+        posthog_trace_id: Optional[str] = None,
+        posthog_properties: Optional[Dict[str, Any]] = None,
+        posthog_privacy_mode: bool = False,
+        posthog_groups: Optional[Dict[str, Any]] = None,
+        **kwargs: Any,
+    ):
+        """
+        Parse structured output using OpenAI's 'responses.parse' method, but also track usage in PostHog.
+
+        Args:
+            posthog_distinct_id: Optional ID to associate with the usage event.
+            posthog_trace_id: Optional trace UUID for linking events.
+            posthog_properties: Optional dictionary of extra properties to include in the event.
+            posthog_privacy_mode: Whether to anonymize the input and output.
+            posthog_groups: Optional dictionary of groups to associate with the event.
+            **kwargs: Any additional parameters for the OpenAI Responses Parse API.
+
+        Returns:
+            The response from OpenAI's responses.parse call.
+        """
+        return await call_llm_and_track_usage_async(
+            posthog_distinct_id,
+            self._client._ph_client,
+            "openai",
+            posthog_trace_id,
+            posthog_properties,
+            posthog_privacy_mode,
+            posthog_groups,
+            self._client.base_url,
+            self._original.parse,
+            **kwargs,
+        )
+
 
 class WrappedChat:
     """Async wrapper for OpenAI chat that tracks usage in PostHog."""

posthog/scopes.py

@@ -12,31 +12,34 @@ def _get_current_context() -> Dict[str, Any]:
 
 
 @contextmanager
-def new_context(fresh=False):
+def new_context(fresh=False, capture_exceptions=True):
     """
-     Create a new context scope that will be active for the duration of the with block.
-     Any tags set within this scope will be isolated to this context. Any exceptions raised
+    Create a new context scope that will be active for the duration of the with block.
+    Any tags set within this scope will be isolated to this context. Any exceptions raised
     or events captured within the context will be tagged with the context tags.
 
-     Args:
-         fresh: Whether to start with a fresh context (default: False).
-                If False, inherits tags from parent context.
-                If True, starts with no tags.
+    Args:
+        fresh: Whether to start with a fresh context (default: False).
+               If False, inherits tags from parent context.
+               If True, starts with no tags.
+        capture_exceptions: Whether to capture exceptions raised within the context (default: True).
+               If True, captures exceptions and tags them with the context tags before propagating them.
+               If False, exceptions will propagate without being tagged or captured.
 
-     Examples:
-         # Inherit parent context tags
-         with posthog.new_context():
-             posthog.tag("request_id", "123")
-             # Both this event and the exception will be tagged with the context tags
-             posthog.capture("event_name", {"property": "value"})
-             raise ValueError("Something went wrong")
+    Examples:
+        # Inherit parent context tags
+        with posthog.new_context():
+            posthog.tag("request_id", "123")
+            # Both this event and the exception will be tagged with the context tags
+            posthog.capture("event_name", {"property": "value"})
+            raise ValueError("Something went wrong")
 
-         # Start with fresh context (no inherited tags)
-         with posthog.new_context(fresh=True):
-             posthog.tag("request_id", "123")
-             # Both this event and the exception will be tagged with the context tags
-             posthog.capture("event_name", {"property": "value"})
-             raise ValueError("Something went wrong")
+        # Start with fresh context (no inherited tags)
+        with posthog.new_context(fresh=True):
+            posthog.tag("request_id", "123")
+            # Both this event and the exception will be tagged with the context tags
+            posthog.capture("event_name", {"property": "value"})
+            raise ValueError("Something went wrong")
 
     """
     from posthog import capture_exception
@@ -49,7 +52,8 @@ def new_context(fresh=False):
     try:
         yield
     except Exception as e:
-        capture_exception(e)
+        if capture_exceptions:
+            capture_exception(e)
         raise
     finally:
         _context_stack.reset(token)
@@ -88,13 +92,14 @@ def clear_tags() -> None:
 F = TypeVar("F", bound=Callable[..., Any])
 
 
-def scoped(fresh=False):
+def scoped(fresh=False, capture_exceptions=True):
     """
     Decorator that creates a new context for the function. Simply wraps
     the function in a with posthog.new_context(): block.
 
     Args:
         fresh: Whether to start with a fresh context (default: False)
+        capture_exceptions: Whether to capture and track exceptions with posthog error tracking (default: True)
 
     Example:
         @posthog.scoped()
@@ -114,7 +119,7 @@ def scoped(fresh=False):
 
         @wraps(func)
         def wrapper(*args, **kwargs):
-            with new_context(fresh=fresh):
+            with new_context(fresh=fresh, capture_exceptions=capture_exceptions):
                 return func(*args, **kwargs)
 
         return cast(F, wrapper)

posthog/test/ai/langchain/test_callbacks.py

@@ -1378,11 +1378,11 @@ def test_langgraph_agent(mock_client):
     )
     graph.invoke(inputs, config={"callbacks": [cb]})
     calls = [call[1] for call in mock_client.capture.call_args_list]
-    assert len(calls) == 21
+    assert len(calls) == 15
     for call in calls:
         assert call["properties"]["$ai_trace_id"] == "test-trace-id"
     assert len([call for call in calls if call["event"] == "$ai_generation"]) == 2
-    assert len([call for call in calls if call["event"] == "$ai_span"]) == 18
+    assert len([call for call in calls if call["event"] == "$ai_span"]) == 12
     assert len([call for call in calls if call["event"] == "$ai_trace"]) == 1
 
 
@@ -1435,11 +1435,13 @@ def test_span_set_parent_ids_for_third_level_run(mock_client, trace_id):
 
     assert mock_client.capture.call_count == 3
 
-    span2, span1, trace = [
-        call[1]["properties"] for call in mock_client.capture.call_args_list
-    ]
-    assert span2["$ai_parent_id"] == span1["$ai_span_id"]
-    assert span1["$ai_parent_id"] == trace["$ai_trace_id"]
+    calls = mock_client.capture.call_args_list
+    span_props_2 = calls[0][1]["properties"]
+    span_props_1 = calls[1][1]["properties"]
+    trace_props = calls[2][1]["properties"]
+
+    assert span_props_2["$ai_parent_id"] == span_props_1["$ai_span_id"]
+    assert span_props_1["$ai_parent_id"] == trace_props["$ai_trace_id"]
 
 
 def test_captures_error_with_details_in_span(mock_client):
@@ -1478,3 +1480,250 @@ def test_captures_error_without_details_in_span(mock_client):
         == "ValueError"
     )
     assert mock_client.capture.call_args_list[1][1]["properties"]["$ai_is_error"]
+
+
+def test_openai_reasoning_tokens(mock_client):
+    """Test that OpenAI reasoning tokens are captured correctly."""
+    prompt = ChatPromptTemplate.from_messages(
+        [("user", "Think step by step about this problem")]
+    )
+
+    # Mock response with reasoning tokens in output_token_details
+    model = FakeMessagesListChatModel(
+        responses=[
+            AIMessage(
+                content="Let me think through this step by step...",
+                usage_metadata={
+                    "input_tokens": 10,
+                    "output_tokens": 25,
+                    "total_tokens": 35,
+                    "output_token_details": {"reasoning": 15},  # 15 reasoning tokens
+                },
+            )
+        ]
+    )
+
+    callbacks = [CallbackHandler(mock_client)]
+    chain = prompt | model
+    result = chain.invoke({}, config={"callbacks": callbacks})
+
+    assert result.content == "Let me think through this step by step..."
+    assert mock_client.capture.call_count == 3
+
+    generation_args = mock_client.capture.call_args_list[1][1]
+    generation_props = generation_args["properties"]
+
+    assert generation_args["event"] == "$ai_generation"
+    assert generation_props["$ai_input_tokens"] == 10
+    assert generation_props["$ai_output_tokens"] == 25
+    assert generation_props["$ai_reasoning_tokens"] == 15
+
+
+def test_anthropic_cache_write_and_read_tokens(mock_client):
+    """Test that Anthropic cache creation and read tokens are captured correctly."""
+    prompt = ChatPromptTemplate.from_messages([("user", "Analyze this large document")])
+
+    # First call with cache creation
+    model_write = FakeMessagesListChatModel(
+        responses=[
+            AIMessage(
+                content="I've analyzed the document and cached the context.",
+                usage_metadata={
+                    "total_tokens": 1050,
+                    "input_tokens": 1000,
+                    "output_tokens": 50,
+                    "cache_creation_input_tokens": 800,  # Anthropic cache write
+                },
+            )
+        ]
+    )
+
+    callbacks = [CallbackHandler(mock_client)]
+    chain = prompt | model_write
+    result = chain.invoke({}, config={"callbacks": callbacks})
+
+    assert result.content == "I've analyzed the document and cached the context."
+    assert mock_client.capture.call_count == 3
+
+    generation_args = mock_client.capture.call_args_list[1][1]
+    generation_props = generation_args["properties"]
+
+    assert generation_args["event"] == "$ai_generation"
+    assert generation_props["$ai_input_tokens"] == 1000
+    assert generation_props["$ai_output_tokens"] == 50
+    assert generation_props["$ai_cache_creation_input_tokens"] == 800
+    assert generation_props["$ai_cache_read_input_tokens"] == 0
+    assert generation_props["$ai_reasoning_tokens"] == 0
+
+    # Reset mock for second call
+    mock_client.reset_mock()
+
+    # Second call with cache read
+    model_read = FakeMessagesListChatModel(
+        responses=[
+            AIMessage(
+                content="Using cached analysis to provide quick response.",
+                usage_metadata={
+                    "input_tokens": 200,
+                    "output_tokens": 30,
+                    "total_tokens": 1030,
+                    "cache_read_input_tokens": 800,  # Anthropic cache read
+                },
+            )
+        ]
+    )
+
+    chain = prompt | model_read
+    result = chain.invoke({}, config={"callbacks": callbacks})
+
+    assert result.content == "Using cached analysis to provide quick response."
+    assert mock_client.capture.call_count == 3
+
+    generation_args = mock_client.capture.call_args_list[1][1]
+    generation_props = generation_args["properties"]
+
+    assert generation_args["event"] == "$ai_generation"
+    assert generation_props["$ai_input_tokens"] == 200
+    assert generation_props["$ai_output_tokens"] == 30
+    assert generation_props["$ai_cache_creation_input_tokens"] == 0
+    assert generation_props["$ai_cache_read_input_tokens"] == 800
+    assert generation_props["$ai_reasoning_tokens"] == 0
+
+
+def test_openai_cache_read_tokens(mock_client):
+    """Test that OpenAI cache read tokens are captured correctly."""
+    prompt = ChatPromptTemplate.from_messages(
+        [("user", "Use the cached prompt for this request")]
+    )
+
+    # Mock response with cache read tokens in input_token_details
+    model = FakeMessagesListChatModel(
+        responses=[
+            AIMessage(
+                content="Response using cached prompt context.",
+                usage_metadata={
+                    "input_tokens": 150,
+                    "output_tokens": 40,
+                    "total_tokens": 190,
+                    "input_token_details": {
+                        "cache_read": 100,  # 100 tokens read from cache
+                        "cache_creation": 0,
+                    },
+                },
+            )
+        ]
+    )
+
+    callbacks = [CallbackHandler(mock_client)]
+    chain = prompt | model
+    result = chain.invoke({}, config={"callbacks": callbacks})
+
+    assert result.content == "Response using cached prompt context."
+    assert mock_client.capture.call_count == 3
+
+    generation_args = mock_client.capture.call_args_list[1][1]
+    generation_props = generation_args["properties"]
+
+    assert generation_args["event"] == "$ai_generation"
+    assert generation_props["$ai_input_tokens"] == 150
+    assert generation_props["$ai_output_tokens"] == 40
+    assert generation_props["$ai_cache_read_input_tokens"] == 100
+    assert generation_props["$ai_cache_creation_input_tokens"] == 0
+    assert generation_props["$ai_reasoning_tokens"] == 0
+
+
+def test_openai_cache_creation_tokens(mock_client):
+    """Test that OpenAI cache creation tokens are captured correctly."""
+    prompt = ChatPromptTemplate.from_messages(
+        [("user", "Create a cache for this large prompt context")]
+    )
+
+    # Mock response with cache creation tokens in input_token_details
+    model = FakeMessagesListChatModel(
+        responses=[
+            AIMessage(
+                content="Created cache for the prompt context.",
+                usage_metadata={
+                    "input_tokens": 2000,
+                    "output_tokens": 25,
+                    "total_tokens": 2025,
+                    "input_token_details": {
+                        "cache_creation": 1500,  # 1500 tokens written to cache
+                        "cache_read": 0,
+                    },
+                },
+            )
+        ]
+    )
+
+    callbacks = [CallbackHandler(mock_client)]
+    chain = prompt | model
+    result = chain.invoke({}, config={"callbacks": callbacks})
+
+    assert result.content == "Created cache for the prompt context."
+    assert mock_client.capture.call_count == 3
+
+    generation_args = mock_client.capture.call_args_list[1][1]
+    generation_props = generation_args["properties"]
+
+    assert generation_args["event"] == "$ai_generation"
+    assert generation_props["$ai_input_tokens"] == 2000
+    assert generation_props["$ai_output_tokens"] == 25
+    assert generation_props["$ai_cache_creation_input_tokens"] == 1500
+    assert generation_props["$ai_cache_read_input_tokens"] == 0
+    assert generation_props["$ai_reasoning_tokens"] == 0
+
+
+def test_combined_reasoning_and_cache_tokens(mock_client):
+    """Test that both reasoning tokens and cache tokens can be captured together."""
+    prompt = ChatPromptTemplate.from_messages(
+        [("user", "Think through this cached problem")]
+    )
+
+    # Mock response with both reasoning and cache tokens
+    model = FakeMessagesListChatModel(
+        responses=[
+            AIMessage(
+                content="Let me reason through this using cached context...",
+                usage_metadata={
+                    "input_tokens": 500,
+                    "output_tokens": 100,
+                    "total_tokens": 600,
+                    "input_token_details": {"cache_read": 300, "cache_creation": 0},
+                    "output_token_details": {"reasoning": 60},  # 60 reasoning tokens
+                },
+            )
+        ]
+    )
+
+    callbacks = [CallbackHandler(mock_client)]
+    chain = prompt | model
+    result = chain.invoke({}, config={"callbacks": callbacks})
+
+    assert result.content == "Let me reason through this using cached context..."
+    assert mock_client.capture.call_count == 3
+
+    generation_args = mock_client.capture.call_args_list[1][1]
+    generation_props = generation_args["properties"]
+
+    assert generation_args["event"] == "$ai_generation"
+    assert generation_props["$ai_input_tokens"] == 500
+    assert generation_props["$ai_output_tokens"] == 100
+    assert generation_props["$ai_cache_read_input_tokens"] == 300
+    assert generation_props["$ai_cache_creation_input_tokens"] == 0
+    assert generation_props["$ai_reasoning_tokens"] == 60
+
+
+@pytest.mark.skipif(not OPENAI_API_KEY, reason="OPENAI_API_KEY is not set")
+def test_openai_reasoning_tokens(mock_client):
+    model = ChatOpenAI(
+        api_key=OPENAI_API_KEY, model="o4-mini", max_completion_tokens=10
+    )
+    cb = CallbackHandler(
+        mock_client, trace_id="test-trace-id", distinct_id="test-distinct-id"
+    )
+    model.invoke("what is the weather in sf", config={"callbacks": [cb]})
+    call = mock_client.capture.call_args_list[0][1]
+    assert call["properties"]["$ai_reasoning_tokens"] is not None
+    assert call["properties"]["$ai_input_tokens"] is not None
+    assert call["properties"]["$ai_output_tokens"] is not None

posthog/test/ai/openai/test_openai.py

@@ -26,6 +26,11 @@ try:
         ResponseOutputMessage,
         ResponseOutputText,
         ResponseUsage,
+        ParsedResponse,
+    )
+    from openai.types.responses.parsed_response import (
+        ParsedResponseOutputMessage,
+        ParsedResponseOutputText,
     )
 
     from posthog.ai.openai import OpenAI
@@ -115,6 +120,59 @@ def mock_openai_response_with_responses_api():
     )
 
 
+@pytest.fixture
+def mock_parsed_response():
+    return ParsedResponse(
+        id="test",
+        model="gpt-4o-2024-08-06",
+        object="response",
+        created_at=1741476542,
+        status="completed",
+        error=None,
+        incomplete_details=None,
+        instructions=None,
+        max_output_tokens=None,
+        tools=[],
+        tool_choice="auto",
+        output=[
+            ParsedResponseOutputMessage(
+                id="msg_123",
+                type="message",
+                role="assistant",
+                status="completed",
+                content=[
+                    ParsedResponseOutputText(
+                        type="output_text",
+                        text='{"name": "Science Fair", "date": "Friday", "participants": ["Alice", "Bob"]}',
+                        annotations=[],
+                        parsed={
+                            "name": "Science Fair",
+                            "date": "Friday",
+                            "participants": ["Alice", "Bob"],
+                        },
+                    )
+                ],
+            )
+        ],
+        output_parsed={
+            "name": "Science Fair",
+            "date": "Friday",
+            "participants": ["Alice", "Bob"],
+        },
+        parallel_tool_calls=True,
+        previous_response_id=None,
+        usage=ResponseUsage(
+            input_tokens=15,
+            output_tokens=20,
+            input_tokens_details={"prompt_tokens": 15, "cached_tokens": 0},
+            output_tokens_details={"reasoning_tokens": 5},
+            total_tokens=35,
+        ),
+        user=None,
+        metadata={},
+    )
+
+
 @pytest.fixture
 def mock_embedding_response():
     return CreateEmbeddingResponse(
@@ -646,3 +704,73 @@ def test_responses_api(mock_client, mock_openai_response_with_responses_api):
         assert props["$ai_http_status"] == 200
         assert props["foo"] == "bar"
         assert isinstance(props["$ai_latency"], float)
+
+
+def test_responses_parse(mock_client, mock_parsed_response):
+    with patch(
+        "openai.resources.responses.Responses.parse",
+        return_value=mock_parsed_response,
+    ):
+        client = OpenAI(api_key="test-key", posthog_client=mock_client)
+        response = client.responses.parse(
+            model="gpt-4o-2024-08-06",
+            input=[
+                {"role": "system", "content": "Extract the event information."},
+                {
+                    "role": "user",
+                    "content": "Alice and Bob are going to a science fair on Friday.",
+                },
+            ],
+            text={
+                "format": {
+                    "type": "json_schema",
+                    "json_schema": {
+                        "name": "event",
+                        "schema": {
+                            "type": "object",
+                            "properties": {
+                                "name": {"type": "string"},
+                                "date": {"type": "string"},
+                                "participants": {
+                                    "type": "array",
+                                    "items": {"type": "string"},
+                                },
+                            },
+                            "required": ["name", "date", "participants"],
+                        },
+                    },
+                }
+            },
+            posthog_distinct_id="test-id",
+            posthog_properties={"foo": "bar"},
+        )
+
+        assert response == mock_parsed_response
+        assert mock_client.capture.call_count == 1
+
+        call_args = mock_client.capture.call_args[1]
+        props = call_args["properties"]
+
+        assert call_args["distinct_id"] == "test-id"
+        assert call_args["event"] == "$ai_generation"
+        assert props["$ai_provider"] == "openai"
+        assert props["$ai_model"] == "gpt-4o-2024-08-06"
+        assert props["$ai_input"] == [
+            {"role": "system", "content": "Extract the event information."},
+            {
+                "role": "user",
+                "content": "Alice and Bob are going to a science fair on Friday.",
+            },
+        ]
+        assert props["$ai_output_choices"] == [
+            {
+                "role": "assistant",
+                "content": '{"name": "Science Fair", "date": "Friday", "participants": ["Alice", "Bob"]}',
+            }
+        ]
+        assert props["$ai_input_tokens"] == 15
+        assert props["$ai_output_tokens"] == 20
+        assert props["$ai_reasoning_tokens"] == 5
+        assert props["$ai_http_status"] == 200
+        assert props["foo"] == "bar"
+        assert isinstance(props["$ai_latency"], float)

posthog/version.py

@@ -1,4 +1,4 @@
-VERSION = "4.6.2"
+VERSION = "4.9.0"
 
 if __name__ == "__main__":
     print(VERSION, end="")  # noqa: T201

pyproject.toml

@@ -68,10 +68,11 @@ test = [
     "django",
     "openai",
     "anthropic",
-    "langgraph",
-    "langchain-community>=0.2.0",
-    "langchain-openai>=0.2.0",
-    "langchain-anthropic>=0.2.0",
+    "langgraph>=0.4.8",
+    "langchain-core>=0.3.65",
+    "langchain-community>=0.3.25",
+    "langchain-openai>=0.3.22",
+    "langchain-anthropic>=0.3.15",
     "google-genai",
     "pydantic",
     "parameterized>=0.8.1",