fix: normalize root run_eval fixture and add agent-specific stream rationale

Apply STREAM constant to the root conftest.py for a uniform interface,
and document why streaming is disabled in autogen/mcp_agent and agentic_rag.

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

Author: andrewdonheiser

agents/autogen/mcp_agent/tests/behavioral/conftest.py

@@ -58,6 +58,7 @@ def eval_config() -> dict[str, Any]:
 
 
 FIXTURES_DIR = Path(__file__).parent / "fixtures"
+# AutoGen MCP exposes tool_invocations only in non-streaming JSON responses.
 STREAM = False
 
 

agents/langgraph/agentic_rag/tests/behavioral/conftest.py

@@ -50,6 +50,7 @@ def _find_repo_root() -> Path:
 
 
 FIXTURES_DIR = Path(__file__).parent / "fixtures"
+# Agentic RAG does not expose tool_calls in streaming; MLflow traces are the only source.
 STREAM = False
 
 

tests/behavioral/conftest.py

@@ -115,6 +115,9 @@ async def http_client() -> AsyncGenerator[httpx.AsyncClient, None]:
         yield client
 
 
+STREAM = False
+
+
 @pytest.fixture
 def run_eval(
     agent_url: str, http_client: httpx.AsyncClient
@@ -133,7 +136,6 @@ async def _run(
         timeout_seconds: float = 30.0,
         max_tokens_budget: int | None = None,
         model: str | None = None,
-        stream: bool = False,
     ) -> TaskResult:
         config = TaskConfig(
             agent_url=agent_url,
@@ -142,7 +144,7 @@ async def _run(
             timeout_seconds=timeout_seconds,
             max_tokens_budget=max_tokens_budget,
             model=model,
-            stream=stream,
+            stream=STREAM,
         )
         return await run_task(config, client=http_client)