fix(ai): report OpenAI Agents cached and reasoning tokens

.sampo/changesets/openai-agents-cache-reasoning-tokens.md

@@ -0,0 +1,5 @@
+---
+pypi/posthog: patch
+---
+
+Report cached input and reasoning tokens from OpenAI Agents generation and response spans, which nest them under the usage details fields.

posthog/ai/openai_agents/processor.py

@@ -43,6 +43,27 @@ def _parse_iso_timestamp(iso_str: Optional[str]) -> Optional[float]:
         return None
 
 
+def _usage_detail(usage: Any, details_keys: tuple, key: str) -> Optional[int]:
+    """Read a token count nested under a usage details field.
+
+    Agents span usage is a dict (generation spans) or an OpenAI usage object
+    (response spans), with details under ``input_tokens_details``/
+    ``output_tokens_details`` (Responses) or ``prompt_tokens_details``/
+    ``completion_tokens_details`` (Chat Completions).
+    """
+
+    def get(container: Any, name: str) -> Any:
+        if isinstance(container, dict):
+            return container.get(name)
+        return getattr(container, name, None)
+
+    for details_key in details_keys:
+        value = get(get(usage, details_key), key)
+        if isinstance(value, int):
+            return value
+    return None
+
+
 class PostHogTracingProcessor(TracingProcessor):
     """
     A tracing processor that sends OpenAI Agents SDK traces to PostHog.
@@ -528,11 +549,30 @@ def _handle_generation_span(
             ),
         }
 
-        # Add optional token fields if present
-        if usage.get("reasoning_tokens"):
-            properties["$ai_reasoning_tokens"] = usage["reasoning_tokens"]
-        if usage.get("cache_read_input_tokens"):
-            properties["$ai_cache_read_input_tokens"] = usage["cache_read_input_tokens"]
+        # Add optional token fields if present. The Agents SDK nests these under
+        # the usage details fields; flat keys are kept for custom models.
+        reasoning_tokens = usage.get("reasoning_tokens") or _usage_detail(
+            usage,
+            ("output_tokens_details", "completion_tokens_details"),
+            "reasoning_tokens",
+        )
+        if reasoning_tokens:
+            properties["$ai_reasoning_tokens"] = reasoning_tokens
+        cache_read_tokens = usage.get("cache_read_input_tokens")
+        if not cache_read_tokens:
+            cache_read_tokens = _usage_detail(
+                usage,
+                ("input_tokens_details", "prompt_tokens_details"),
+                "cached_tokens",
+            )
+            if cache_read_tokens:
+                # OpenAI-shaped usage counts cached tokens inside input_tokens.
+                # Say so explicitly: the cost backend otherwise infers exclusive
+                # reporting from Claude model names served through compatible
+                # endpoints. A caller-supplied value still takes precedence.
+                properties["$ai_cache_reporting_exclusive"] = False
+        if cache_read_tokens:
+            properties["$ai_cache_read_input_tokens"] = cache_read_tokens
         if usage.get("cache_creation_input_tokens"):
             properties["$ai_cache_creation_input_tokens"] = usage[
                 "cache_creation_input_tokens"
@@ -708,6 +748,20 @@ def _handle_response_span(
         if total_cost_usd is not None:
             properties["$ai_total_cost_usd"] = total_cost_usd
 
+        if usage:
+            reasoning_tokens = _usage_detail(
+                usage, ("output_tokens_details",), "reasoning_tokens"
+            )
+            if reasoning_tokens:
+                properties["$ai_reasoning_tokens"] = reasoning_tokens
+            cache_read_tokens = _usage_detail(
+                usage, ("input_tokens_details",), "cached_tokens"
+            )
+            if cache_read_tokens:
+                properties["$ai_cache_read_input_tokens"] = cache_read_tokens
+                # Responses input_tokens include cached tokens.
+                properties["$ai_cache_reporting_exclusive"] = False
+
         # Extract output content from response
         if response:
             output_items = getattr(response, "output", None)

posthog/test/ai/openai_agents/test_processor.py

@@ -288,6 +288,163 @@ def test_generation_span_with_reasoning_tokens(
         call_kwargs = mock_client.capture.call_args[1]
         assert call_kwargs["properties"]["$ai_reasoning_tokens"] == 400
 
+    def test_generation_span_reads_nested_cache_and_reasoning_tokens(
+        self, processor, mock_client, mock_span
+    ):
+        """The Agents SDK nests cached and reasoning tokens under the usage
+        details fields of generation span usage."""
+        from agents.usage import Usage, model_usage_to_span_usage
+        from openai.types.responses.response_usage import (
+            InputTokensDetails,
+            OutputTokensDetails,
+        )
+
+        usage = Usage(
+            requests=1,
+            input_tokens=100,
+            output_tokens=50,
+            total_tokens=150,
+            input_tokens_details=InputTokensDetails(
+                cached_tokens=60, cache_write_tokens=0
+            ),
+            output_tokens_details=OutputTokensDetails(reasoning_tokens=40),
+        )
+        mock_span.span_data = GenerationSpanData(
+            model="gpt-5-mini", usage=model_usage_to_span_usage(usage)
+        )
+
+        processor.on_span_start(mock_span)
+        processor.on_span_end(mock_span)
+
+        props = mock_client.capture.call_args[1]["properties"]
+        assert props["$ai_input_tokens"] == 100
+        assert props["$ai_cache_read_input_tokens"] == 60
+        assert props["$ai_output_tokens"] == 50
+        assert props["$ai_reasoning_tokens"] == 40
+        assert props["$ai_cache_reporting_exclusive"] is False
+
+    def test_response_span_reads_cache_and_reasoning_tokens(
+        self, processor, mock_client, mock_span
+    ):
+        """Response spans carry the OpenAI Responses usage object, with cached
+        and reasoning tokens in its details fields."""
+        from openai.types.responses import Response
+
+        response = Response.model_validate(
+            {
+                "id": "resp_123",
+                "object": "response",
+                "created_at": 0,
+                "model": "gpt-5-mini",
+                "status": "completed",
+                "parallel_tool_calls": True,
+                "tool_choice": "auto",
+                "tools": [],
+                "output": [],
+                "usage": {
+                    "input_tokens": 100,
+                    "output_tokens": 50,
+                    "total_tokens": 150,
+                    "input_tokens_details": {
+                        "cached_tokens": 60,
+                        "cache_write_tokens": 0,
+                    },
+                    "output_tokens_details": {"reasoning_tokens": 40},
+                },
+            }
+        )
+        mock_span.span_data = ResponseSpanData(response=response, input="Hi")
+
+        processor.on_span_start(mock_span)
+        processor.on_span_end(mock_span)
+
+        props = mock_client.capture.call_args[1]["properties"]
+        assert props["$ai_input_tokens"] == 100
+        assert props["$ai_cache_read_input_tokens"] == 60
+        assert props["$ai_output_tokens"] == 50
+        assert props["$ai_reasoning_tokens"] == 40
+        assert props["$ai_cache_reporting_exclusive"] is False
+
+    def test_cached_compatible_claude_usage_is_declared_inclusive(self):
+        """A Claude model behind an OpenAI-compatible endpoint reports inclusive
+        cached tokens. Without an explicit convention, the cost backend infers
+        exclusive reporting from the model name and overcounts input cost."""
+        from agents import set_trace_processors
+        from agents.tracing import generation_span, trace
+
+        client = MagicMock(privacy_mode=False, enable_full_ai_capture=False)
+        set_trace_processors([])
+        processor = instrument(client, distinct_id="review-user")
+        try:
+            with trace("compatible-model"):
+                with generation_span(
+                    model="anthropic/claude-sonnet-4.6",
+                    usage={
+                        "input_tokens": 10339,
+                        "output_tokens": 60,
+                        "input_tokens_details": {"cached_tokens": 10318},
+                    },
+                ):
+                    pass
+            properties = next(
+                call.kwargs["properties"]
+                for call in client.capture.call_args_list
+                if call.kwargs["event"] == "$ai_generation"
+            )
+            assert properties["$ai_cache_read_input_tokens"] == 10318
+            assert properties.get("$ai_cache_reporting_exclusive") is False
+        finally:
+            set_trace_processors([])
+            processor.shutdown()
+
+    def test_flat_cache_keys_keep_their_own_convention(
+        self, processor, mock_client, mock_span
+    ):
+        """Flat cache keys come from custom models with their own convention, so
+        no inclusive/exclusive flag is declared for them."""
+        mock_span.span_data = GenerationSpanData(
+            model="custom-model",
+            usage={
+                "input_tokens": 100,
+                "output_tokens": 50,
+                "cache_read_input_tokens": 60,
+            },
+        )
+
+        processor.on_span_start(mock_span)
+        processor.on_span_end(mock_span)
+
+        props = mock_client.capture.call_args[1]["properties"]
+        assert props["$ai_cache_read_input_tokens"] == 60
+        assert "$ai_cache_reporting_exclusive" not in props
+
+    def test_caller_cache_reporting_convention_takes_precedence(
+        self, mock_client, mock_span
+    ):
+        """A processor-wide $ai_cache_reporting_exclusive overrides the default
+        declared for nested OpenAI-shaped usage."""
+        processor = PostHogTracingProcessor(
+            client=mock_client,
+            distinct_id="test-user",
+            privacy_mode=False,
+            properties={"$ai_cache_reporting_exclusive": True},
+        )
+        mock_span.span_data = GenerationSpanData(
+            model="gpt-5-mini",
+            usage={
+                "input_tokens": 100,
+                "output_tokens": 50,
+                "input_tokens_details": {"cached_tokens": 60},
+            },
+        )
+
+        processor.on_span_start(mock_span)
+        processor.on_span_end(mock_span)
+
+        props = mock_client.capture.call_args[1]["properties"]
+        assert props["$ai_cache_read_input_tokens"] == 60
+        assert props["$ai_cache_reporting_exclusive"] is True
+
     def test_function_span_mapping(self, processor, mock_client, mock_span):
         """Test FunctionSpanData maps to $ai_span event with type=tool."""
         span_data = FunctionSpanData(