Keep OpenAI Agents traces connected across Temporal execution

CHANGELOG.md

@@ -56,6 +56,8 @@ to include examples, links to docs, or any other relevant information.
 
 ### Fixed
 
+- OpenAI Agents tracing preserves caller spans across workers and restores context
+  after activity, local-activity, and child-workflow completion callbacks.
 - `contrib.deepagents`: summarization middleware configured with a model name string now routes its LLM calls through Activities instead of running them in the Workflow.
 
 - **Experimental**: External storage metrics now report the wall-clock time storage was in flight.

temporalio/contrib/openai_agents/_otel_trace_interceptor.py

@@ -2,13 +2,14 @@
 
 from __future__ import annotations
 
+from collections.abc import Iterator
+from contextlib import contextmanager
 from typing import Any
 
 import opentelemetry.trace
 
 import temporalio.converter
 
-from ..opentelemetry._id_generator import TemporalIdGenerator
 from ._trace_interceptor import (
     OpenAIAgentsContextPropagationInterceptor,
     _InputWithHeaders,
@@ -22,19 +23,19 @@ class OTelOpenAIAgentsContextPropagationInterceptor(
 
     def __init__(
         self,
-        otel_id_generator: TemporalIdGenerator,
         payload_converter: temporalio.converter.PayloadConverter = temporalio.converter.default().payload_converter,
         add_temporal_spans: bool = True,
     ) -> None:
         """Initialize OTEL-aware context propagation interceptor.
 
         Args:
-            otel_id_generator: Generator for OTEL-compatible IDs.
             payload_converter: Converter for serializing trace context.
             add_temporal_spans: Whether to add Temporal-specific spans.
         """
-        super().__init__(payload_converter, add_temporal_spans, start_traces=True)
-        self._otel_id_generator = otel_id_generator
+        super().__init__(
+            payload_converter=payload_converter,
+            add_temporal_spans=add_temporal_spans,
+        )
 
     def header_contents(self) -> dict[str, Any]:
         """Get header contents enhanced with OpenTelemetry span context.
@@ -50,39 +51,46 @@ def header_contents(self) -> dict[str, Any]:
                 **super().header_contents(),
                 "otelSpanId": span_context.span_id,
                 "otelTraceId": span_context.trace_id,
+                "otelTraceFlags": int(span_context.trace_flags),
+                "otelTraceState": span_context.trace_state.to_header(),
             }
         else:
             return super().header_contents()
 
+    @contextmanager
     def context_from_header(
         self,
         input: _InputWithHeaders,
-    ):
-        """Extracts and initializes trace information the input header."""
+    ) -> Iterator[None]:
+        """Use propagated IDs as a remote parent without recording replicas."""
         span_info = self.get_header_contents(input)
-
-        if span_info is None:
-            return
-        otel_span_id = span_info.get("otelSpanId")
-        otel_trace_id = span_info.get("otelTraceId")
-
-        # Seed the trace id before the trace is reconstructed so the workflow's root
-        # OTEL span shares the caller's trace id rather than generating a new one.
-        if otel_trace_id and self._otel_id_generator:
-            self._otel_id_generator.seed_trace_id(otel_trace_id)
-
-        # If only a trace was propagated from the caller, we need to seed for trace context
-        if otel_span_id and self._otel_id_generator and span_info.get("spanId") is None:
-            self._otel_id_generator.seed_span_id(otel_span_id)
-
-        super().trace_context_from_header_contents(span_info)
-
-        # If a span was propagated from the caller, we need to seed for span context
-        if (
-            otel_span_id
-            and self._otel_id_generator
-            and span_info.get("spanId") is not None
-        ):
-            self._otel_id_generator.seed_span_id(otel_span_id)
-
-        super().span_context_from_header_contents(span_info)
+        with super().context_from_header(input=input):
+            if (
+                span_info is not None
+                and span_info.get("otelSpanId")
+                and span_info.get("otelTraceId")
+            ):
+                span_context: opentelemetry.trace.SpanContext = (
+                    opentelemetry.trace.SpanContext(
+                        trace_id=span_info["otelTraceId"],
+                        span_id=span_info["otelSpanId"],
+                        is_remote=True,
+                        trace_flags=opentelemetry.trace.TraceFlags(
+                            span_info.get(
+                                "otelTraceFlags", opentelemetry.trace.TraceFlags.SAMPLED
+                            )
+                        ),
+                        trace_state=opentelemetry.trace.TraceState.from_header(
+                            [span_info["otelTraceState"]]
+                            if span_info.get("otelTraceState")
+                            else []
+                        ),
+                    )
+                )
+                with opentelemetry.trace.use_span(
+                    opentelemetry.trace.NonRecordingSpan(span_context),
+                    end_on_exit=False,
+                ):
+                    yield
+            else:
+                yield

temporalio/contrib/openai_agents/_temporal_openai_agents.py

@@ -6,7 +6,9 @@
 import typing
 from collections.abc import AsyncIterator, Callable, Collection, Iterator, Sequence
 from contextlib import asynccontextmanager, contextmanager
+from contextvars import Token
 from datetime import timedelta
+from weakref import WeakKeyDictionary
 
 import pydantic
 from agents import ModelProvider, Trace, set_trace_provider
@@ -52,6 +54,10 @@
 from temporalio.worker.workflow_sandbox import SandboxedWorkflowRunner
 
 if typing.TYPE_CHECKING:
+    from openinference.instrumentation.openai_agents._processor import (
+        OpenInferenceTracingProcessor,
+    )
+
     from temporalio.contrib.openai_agents import (
         SandboxClientProvider,
         StatefulMCPServerProvider,
@@ -62,32 +68,51 @@
 _otel_trace_start_patch_lock = threading.RLock()
 _otel_trace_start_patch_ref_count = 0
 _otel_trace_start_original: Callable[..., typing.Any] | None = None
+_otel_trace_end_original: (
+    Callable[["OpenInferenceTracingProcessor", Trace], None] | None
+) = None
 _otel_trace_start_instrumentor: typing.Any | None = None
 
 
 def _install_otel_instrumentation(tracer_provider: typing.Any) -> None:
     """Configure OpenInference while at least one tracing context is active."""
     global _otel_trace_start_instrumentor
     global _otel_trace_start_original
+    global _otel_trace_end_original
     global _otel_trace_start_patch_ref_count
 
     from openinference.instrumentation.openai_agents import OpenAIAgentsInstrumentor
     from openinference.instrumentation.openai_agents._processor import (
         OpenInferenceTracingProcessor,
     )
-    from opentelemetry.context import attach
+    from opentelemetry.context import Context, attach, detach
     from opentelemetry.trace import set_span_in_context
 
     with _otel_trace_start_patch_lock:
         if _otel_trace_start_patch_ref_count == 0:
             original_on_trace_start = OpenInferenceTracingProcessor.on_trace_start
+            original_on_trace_end = OpenInferenceTracingProcessor.on_trace_end
             _otel_trace_start_original = original_on_trace_start
+            _otel_trace_end_original = original_on_trace_end
+            trace_tokens: WeakKeyDictionary[Trace, Token[Context]] = WeakKeyDictionary()
+
+            def on_trace_start(
+                self: OpenInferenceTracingProcessor, trace: Trace
+            ) -> None:
+                original_on_trace_start(self=self, trace=trace)
+                trace_tokens[trace] = attach(
+                    set_span_in_context(self._root_spans[trace.trace_id])
+                )
 
-            def on_trace_start(self: typing.Any, trace: Trace) -> None:  # type: ignore[reportUnusedFunction]
-                original_on_trace_start(self, trace)
-                attach(set_span_in_context(self._root_spans[trace.trace_id]))
+            def on_trace_end(self: OpenInferenceTracingProcessor, trace: Trace) -> None:
+                try:
+                    original_on_trace_end(self=self, trace=trace)
+                finally:
+                    if token := trace_tokens.pop(trace, None):
+                        detach(token)
 
             setattr(OpenInferenceTracingProcessor, "on_trace_start", on_trace_start)
+            setattr(OpenInferenceTracingProcessor, "on_trace_end", on_trace_end)
             try:
                 _otel_trace_start_instrumentor = OpenAIAgentsInstrumentor()
                 _otel_trace_start_instrumentor.instrument(
@@ -99,7 +124,13 @@ def on_trace_start(self: typing.Any, trace: Trace) -> None:  # type: ignore[repo
                     "on_trace_start",
                     _otel_trace_start_original,
                 )
+                setattr(
+                    OpenInferenceTracingProcessor,
+                    "on_trace_end",
+                    _otel_trace_end_original,
+                )
                 _otel_trace_start_original = None
+                _otel_trace_end_original = None
                 _otel_trace_start_instrumentor = None
                 raise
         _otel_trace_start_patch_ref_count += 1
@@ -109,6 +140,7 @@ def _uninstall_otel_instrumentation() -> None:
     """Tear down OpenInference after the final tracing context exits."""
     global _otel_trace_start_instrumentor
     global _otel_trace_start_original
+    global _otel_trace_end_original
     global _otel_trace_start_patch_ref_count
 
     from openinference.instrumentation.openai_agents._processor import (
@@ -130,7 +162,14 @@ def _uninstall_otel_instrumentation() -> None:
                         "on_trace_start",
                         _otel_trace_start_original,
                     )
+                if _otel_trace_end_original is not None:
+                    setattr(
+                        OpenInferenceTracingProcessor,
+                        "on_trace_end",
+                        _otel_trace_end_original,
+                    )
                 _otel_trace_start_original = None
+                _otel_trace_end_original = None
                 _otel_trace_start_instrumentor = None
 
 
@@ -426,7 +465,6 @@ def workflow_runner(runner: WorkflowRunner | None) -> WorkflowRunner:
 
             interceptor = OTelOpenAIAgentsContextPropagationInterceptor(
                 add_temporal_spans=add_temporal_spans,
-                otel_id_generator=provider.id_generator(),
             )
 
         @asynccontextmanager