fix(tracing): keep list-shaped eval spans searchable and drop stale eval attrs

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>

Author: mohammadatallah-scale

src/agentex/lib/core/temporal/workers/worker.py

@@ -275,7 +275,7 @@ async def run(
             build_id=str(uuid.uuid4()),
             debug_mode=debug_enabled,  # Disable deadlock detection in debug mode
             # Temporal inherits client tracing before these business interceptors.
-            interceptors=[*self.interceptors, SGPEvalsInterceptor()],
+            interceptors=[SGPEvalsInterceptor(), *self.interceptors],
         )
 
         logger.info(f"Starting workers for task queue: {self.task_queue}")

src/agentex/lib/core/tracing/processors/sgp_tracing_processor.py

@@ -83,16 +83,19 @@ def _sgp_metadata(span: Span) -> Any:
     leak like that today. Left as-is: changing five long-shipped fields is not
     this change's business.)
     """
-    extra: dict[str, Any] = dict(sgp_evals.attrs_for_span(span))
+    eval_attrs = sgp_evals.attrs_for_span(span)
+    extra: dict[str, Any] = dict(eval_attrs)
     commit_sha = code_revision.commit_sha()
     if commit_sha is not None:
         extra[code_revision.COMMIT_SHA_KEY] = commit_sha
     if not extra:
         return span.data
     if isinstance(span.data, dict):
         return {**span.data, **extra}
-    # List-shaped data is an accepted `data` shape and has nowhere to put a
-    # metadata key; leave it untouched rather than dropping the caller's data.
+    # List-shaped data has nowhere to put a key. Eval spans must stay searchable by run and row,
+    # so their list moves under "data". Otherwise it is left untouched rather than reshaped.
+    if eval_attrs and isinstance(span.data, list):
+        return {**extra, "data": span.data}
     return span.data
 
 

src/agentex/lib/core/tracing/sgp_evals.py

@@ -52,6 +52,9 @@ def register_task_metadata(task_id: str, task_metadata: Any) -> dict[str, Any] |
     attrs = span_attrs_from_task_metadata(task_metadata)
     if attrs is not None:
         register_task(task_id, attrs)
+    elif _attrs_by_task:
+        with _lock:
+            _attrs_by_task.pop(task_id, None)
     return attrs
 
 

tests/lib/core/tracing/test_sgp_evals.py

@@ -98,6 +98,15 @@ def test_other_tasks_and_non_eval_tasks_are_untouched(self) -> None:
         assert _sgp_metadata(_span(trace_id="task-2", data={"k": 1})) == {"k": 1}
         assert _sgp_metadata(_span(trace_id="task-3", data={"k": 1})) == {"k": 1}
 
+    def test_eval_span_with_list_data_keeps_list_and_gains_the_ids(self) -> None:
+        sgp_evals.register_task_metadata("task-1", EVAL_METADATA)
+        assert _sgp_metadata(_span(data=[{"a": 1}])) == {**EXPECTED_ATTRS, "data": [{"a": 1}]}
+
+    def test_task_that_stops_being_an_eval_task_stops_being_stamped(self) -> None:
+        sgp_evals.register_task_metadata("task-1", EVAL_METADATA)
+        sgp_evals.register_task_metadata("task-1", {"team": "x"})
+        assert _sgp_metadata(_span(data={"k": 1})) == {"k": 1}
+
     def test_registry_is_bounded(self) -> None:
         with patch.object(sgp_evals, "_MAX_TASKS", 2):
             for i in range(3):