fix: populate AgentGraphResult.evaluations and clean up remaining review items

- Fix AgentGraphResult.evaluations type from Optional[List[Any]] to Optional[List[JudgeResult]]
- Populate evaluations in both LangGraph and OpenAI runners with all judge results
- Remove stray `if tracker:` guard in OpenAI _handle_handoff (tracker is always set)
- Add comment documenting why output_text is empty at handoff time in OpenAI runner
- flush() now returns List[JudgeResult] instead of None

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Author: jsonbailey

packages/ai-providers/server-ai-langchain/src/ldai_langchain/langgraph_agent_graph_runner.py

@@ -330,7 +330,7 @@ async def run(self, input: Any) -> AgentGraphResult:
             output = extract_last_message_content(messages)
 
             # Flush per-node metrics to LD trackers
-            await handler.flush(self._graph, pending_eval_tasks)
+            all_eval_results = await handler.flush(self._graph, pending_eval_tasks)
 
             tracker.track_path(handler.path)
             tracker.track_duration(duration)
@@ -341,6 +341,7 @@ async def run(self, input: Any) -> AgentGraphResult:
                 output=output,
                 raw=result,
                 metrics=LDAIMetrics(success=True),
+                evaluations=all_eval_results,
             )
 
         except Exception as exc:

packages/ai-providers/server-ai-langchain/src/ldai_langchain/langgraph_callback_handler.py

@@ -5,6 +5,7 @@
 from langchain_core.callbacks import BaseCallbackHandler
 from langchain_core.outputs import ChatGeneration, LLMResult
 from ldai.agent_graph import AgentGraphDefinition
+from ldai.providers.types import JudgeResult
 from ldai.tracker import TokenUsage
 
 from ldai_langchain.langchain_helper import get_ai_usage_from_response
@@ -188,7 +189,9 @@ def on_tool_end(
     # Flush
     # ------------------------------------------------------------------
 
-    async def flush(self, graph: AgentGraphDefinition, eval_tasks=None) -> None:
+    async def flush(
+        self, graph: AgentGraphDefinition, eval_tasks=None
+    ) -> List[JudgeResult]:
         """
         Emit all collected per-node metrics to the LaunchDarkly trackers.
 
@@ -198,8 +201,10 @@ async def flush(self, graph: AgentGraphDefinition, eval_tasks=None) -> None:
         :param eval_tasks: Optional dict mapping node key to a list of awaitables that
             return judge evaluation results. Multiple tasks arise when a node is visited
             more than once (e.g. in a graph with cycles).
+        :return: All judge results collected across all nodes.
         """
         node_trackers: Dict[str, Any] = {}
+        all_eval_results: List[JudgeResult] = []
         for node_key in self._path:
             if node_key in node_trackers:
                 continue
@@ -229,6 +234,9 @@ async def flush(self, graph: AgentGraphDefinition, eval_tasks=None) -> None:
 
             for eval_task in eval_tasks.get(node_key, []):
                 results = await eval_task
+                all_eval_results.extend(results)
                 for r in results:
                     if r.success:
                         config_tracker.track_judge_result(r)
+
+        return all_eval_results

packages/ai-providers/server-ai-openai/src/ldai_openai/openai_agent_graph_runner.py

@@ -91,8 +91,10 @@ async def run(self, input: Any) -> AgentGraphResult:
             root_agent = self._build_agents(path, state, tracker)
             result = await Runner.run(root_agent, input_str)
             self._flush_final_segment(state, result, input_str)
+            all_eval_results = []
             for node_tracker, eval_task in state.pending_eval_tasks:
                 eval_results = await eval_task
+                all_eval_results.extend(eval_results)
                 for r in eval_results:
                     if r.success:
                         node_tracker.track_judge_result(r)
@@ -111,6 +113,7 @@ async def run(self, input: Any) -> AgentGraphResult:
                 output=str(result.final_output),
                 raw=result,
                 metrics=LDAIMetrics(success=True, usage=token_usage),
+                evaluations=all_eval_results,
             )
         except Exception as exc:
             if isinstance(exc, ImportError):
@@ -251,8 +254,7 @@ def _handle_handoff(
     ) -> None:
         path.append(tgt)
         state.last_node_key = tgt
-        if tracker:
-            tracker.track_handoff_success(src, tgt)
+        tracker.track_handoff_success(src, tgt)
 
         now_ns = time.perf_counter_ns()
         duration_ms = (now_ns - state.last_handoff_ns) // 1_000_000
@@ -275,6 +277,8 @@ def _handle_handoff(
 
             src_node = self._graph.get_node(src)
             if src_node is not None:
+                # The OpenAI Agents SDK does not expose the agent's text output at
+                # handoff time via RunContextWrapper, so output_text is empty here.
                 eval_task = src_node.get_config().evaluator.evaluate(input_str, '')
                 state.pending_eval_tasks.append((config_tracker, eval_task))
 

packages/sdk/server-ai/src/ldai/providers/types.py

@@ -110,4 +110,4 @@ class AgentGraphResult:
     output: str
     raw: Any
     metrics: LDAIMetrics
-    evaluations: Optional[List[Any]] = None
+    evaluations: Optional[List[JudgeResult]] = None