test: diagnose OpenTelemetry replay span duplication

.github/workflows/ci.yml

@@ -26,10 +26,7 @@ jobs:
           - os: ubuntu-latest
             python: "3.14"
             docsTarget: true
-            openaiTestTarget: true
             clippyLinter: true
-          - python: "3.10"
-            pytestExtraArgs: '--reruns 3 --only-rerun "RuntimeError: Failed validating workflow"'
           - os: ubuntu-arm
             runsOn: ubuntu-24.04-arm64-2-core
           - os: macos-arm
@@ -62,17 +59,13 @@ jobs:
       - run: poe build-develop
       - run: poe lint
       - run: mkdir junit-xml
-      - run: poe test ${{matrix.pytestExtraArgs}} -s --junit-xml=junit-xml/${{ matrix.python }}--${{ matrix.os }}.xml
+      - name: Repeat OpenTelemetry tracing tests
+        shell: bash
+        run: |
+          for attempt in {1..20}; do
+            poe test -s tests/contrib/opentelemetry/test_opentelemetry.py tests/contrib/opentelemetry/test_opentelemetry_plugin.py
+          done
         timeout-minutes: 15
-      # Time skipping doesn't yet support ARM
-      - if: ${{ !endsWith(matrix.os, '-arm') }}
-        run: poe test ${{matrix.pytestExtraArgs}} -s --workflow-environment time-skipping --junit-xml=junit-xml/${{ matrix.python }}--${{ matrix.os }}--time-skipping.xml
-        timeout-minutes: 10
-      - if: ${{ matrix.openaiTestTarget && (github.event.pull_request.head.repo.full_name == '' || github.event.pull_request.head.repo.full_name == 'temporalio/sdk-python') }}
-        run: poe test tests/contrib/openai_agents/test_openai.py ${{matrix.pytestExtraArgs}} -s  --junit-xml=junit-xml/${{ matrix.python }}--${{ matrix.os }}--openai.xml
-        timeout-minutes: 10
-        env:
-          OPENAI_API_KEY: ${{ secrets.OPENAI_API_KEY }}
       - name: "Upload junit-xml artifacts"
         uses: actions/upload-artifact@ea165f8d65b6e75b540449e92b4886f43607fa02 # v4
         if: always()

tests/contrib/google_adk_agents/test_google_adk_agents.py

@@ -56,7 +56,7 @@
 from temporalio.contrib.opentelemetry import OpenTelemetryPlugin, create_tracer_provider
 from temporalio.worker import Worker
 from temporalio.workflow import ActivityConfig
-from tests.contrib.opentelemetry.test_opentelemetry import dump_spans
+from tests.contrib.opentelemetry.test_opentelemetry import assert_span_hierarchy
 
 logger = logging.getLogger(__name__)
 
@@ -578,22 +578,24 @@ async def test_single_agent_telemetry(
         assert result.content.parts is not None
         assert result.content.parts[0].text == "warm and sunny"
 
-    print("\n".join(dump_spans(exporter.get_finished_spans(), with_attributes=False)))
-    assert dump_spans(exporter.get_finished_spans(), with_attributes=False) == [
-        "StartWorkflow:WeatherAgent",
-        "  RunWorkflow:WeatherAgent",
-        "    invocation [test_app]",
-        "      agent_run [test_agent]",
-        "        call_llm",
-        "          StartActivity:invoke_model",
-        "            RunActivity:invoke_model",
-        "          execute_tool get_weather",
-        "            StartActivity:get_weather",
-        "              RunActivity:get_weather",
-        "        call_llm",
-        "          StartActivity:invoke_model",
-        "            RunActivity:invoke_model",
-    ]
+    assert_span_hierarchy(
+        exporter.get_finished_spans(),
+        [
+            "StartWorkflow:WeatherAgent",
+            "  RunWorkflow:WeatherAgent",
+            "    invocation [test_app]",
+            "      agent_run [test_agent]",
+            "        call_llm",
+            "          StartActivity:invoke_model",
+            "            RunActivity:invoke_model",
+            "          execute_tool get_weather",
+            "            StartActivity:get_weather",
+            "              RunActivity:get_weather",
+            "        call_llm",
+            "          StartActivity:invoke_model",
+            "            RunActivity:invoke_model",
+        ],
+    )
 
 
 async def test_unsetting_timeout():