fix: address remaining review issues (3/5)

- Drop cosmetic reranking 'running' event — reranking completes inside
tool execution, so only emit 'done' with results (issue #4)
- Replace for/else with explicit response.tool_calls check, matching
the pattern used by run() (issue #5)
- Fix test_stream_events_ordered to filter legacy events, accommodating
new stage events in the stream (issue #3)

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

agent_bench/agents/orchestrator.py

@@ -266,11 +266,8 @@ class Orchestrator:
                         "chunks_pre_rerank": pre_rerank,
                     })
 
-                    # --- Reranking stage (if reranking happened) ---
+                    # --- Reranking stage (already completed inside tool execution) ---
                     if pre_rerank > 0:
-                        yield StreamEvent(type="stage", metadata={
-                            "stage": "reranking", "status": "running", "iteration": iteration,
-                        })
                         yield StreamEvent(type="stage", metadata={
                             "stage": "reranking", "status": "done", "iteration": iteration,
                             "chunks": result.metadata.get("chunks", []),
@@ -278,8 +275,10 @@ class Orchestrator:
 
                 if "sources" in result.metadata:
                     all_sources.extend(result.metadata["sources"])
-        else:
-            # Max iterations hit — force text answer without tools
+
+        # Max iterations hit — force text answer without tools
+        # (same pattern as run(): explicit call after loop)
+        if response.tool_calls:
             yield StreamEvent(type="stage", metadata={
                 "stage": "llm", "status": "running", "iteration": iteration,
             })
@@ -293,7 +292,6 @@ class Orchestrator:
                 "stage": "llm", "status": "done", "iteration": iteration,
             })
 
-        # Handle max_iterations=0
         if self.max_iterations == 0:
             response = await self.provider.complete(
                 messages, tools=None, temperature=self.temperature

tests/test_serving.py

@@ -451,7 +451,7 @@ class TestStreaming:
 
     @pytest.mark.asyncio
     async def test_stream_events_ordered(self, test_app):
-        """Event sequence: sources → chunk* → done."""
+        """Legacy event sequence preserved: sources → chunk* → done."""
         import json as json_mod
 
         async with AsyncClient(
@@ -461,15 +461,17 @@ class TestStreaming:
                 "/ask/stream", json={"question": "How do path parameters work?"}
             )
 
-        events = []
+        all_events = []
         for line in response.text.strip().split("\n"):
             if line.startswith("data: "):
-                events.append(json_mod.loads(line[6:]))
-
-        assert len(events) >= 3  # at least sources + 1 chunk + done
-        assert events[0]["type"] == "sources"
-        assert events[-1]["type"] == "done"
-        assert all(e["type"] == "chunk" for e in events[1:-1])
+                all_events.append(json_mod.loads(line[6:]))
+
+        # Filter to legacy event types only (stage events are additive)
+        legacy = [e for e in all_events if e["type"] in ("sources", "chunk", "done", "_orchestrator_done")]
+        assert len(legacy) >= 3  # at least sources + 1 chunk + done
+        assert legacy[0]["type"] == "sources"
+        assert legacy[-1]["type"] in ("done", "_orchestrator_done")
+        assert all(e["type"] == "chunk" for e in legacy[1:-1])
 
     @pytest.mark.asyncio
     async def test_stream_chunks_assemble(self, test_app):

tests/test_stream_stages.py

@@ -47,7 +47,9 @@ class TestOrchestratorStageEvents:
 
         stage_events = [e for e in events if e.type == "stage"]
         reranking_events = [e for e in stage_events if e.metadata.get("stage") == "reranking"]
-        assert len(reranking_events) >= 1  # at least done (running may be instant)
+        assert len(reranking_events) >= 1  # done event with chunk details
+        # Reranking completes inside tool execution, so only a done event is emitted
+        assert all(e.metadata.get("status") == "done" for e in reranking_events)
 
     @pytest.mark.asyncio
     async def test_stream_emits_llm_stage(self, orchestrator):