tests/test_routing_golden.py

@@ -1,8 +1,9 @@
 """Golden-set routing test for the structured router.
 
-20 labelled tasks → expected agent. Skipped if provider keys are missing
-(ANTHROPIC for the router LLM, COMPOSIO because importing the graph eagerly
-loads tools). Run with:
+20 labelled tasks → expected agent. Live tests require ANTHROPIC_API_KEY +
+COMPOSIO_API_KEY; without them the parametrized cases skip at runtime but
+the module still collects (so `pytest --collect-only` in CI works without
+secrets). Run with:
 
     pytest tests/test_routing_golden.py -v
 """
@@ -14,17 +15,12 @@ from dotenv import load_dotenv
 
 load_dotenv(os.path.join(os.path.dirname(__file__), "..", ".env"))
 
-if not os.getenv("ANTHROPIC_API_KEY"):
-    pytest.skip(
-        "ANTHROPIC_API_KEY not set; skipping live router tests.",
-        allow_module_level=True,
-    )
-if not os.getenv("COMPOSIO_API_KEY"):
-    pytest.skip(
-        "COMPOSIO_API_KEY not set; skipping live router tests.", allow_module_level=True
-    )
-
-from graph import router_node  # noqa: E402
+_REQUIRED_KEYS = ("ANTHROPIC_API_KEY", "COMPOSIO_API_KEY")
+_missing = [k for k in _REQUIRED_KEYS if not os.getenv(k)]
+requires_keys = pytest.mark.skipif(
+    bool(_missing),
+    reason=f"Requires {', '.join(_REQUIRED_KEYS)}; missing: {', '.join(_missing)}",
+)
 
 
 GOLDEN_SET: list[tuple[str, str]] = [
@@ -107,8 +103,13 @@ def test_golden_set_size():
     assert len(GOLDEN_SET) == 20
 
 
+@requires_keys
 @pytest.mark.parametrize("task,expected", GOLDEN_SET)
 def test_router_picks_expected_agent(task: str, expected: str):
+    # Lazy import — graph.py loads composio_tools at module import and needs
+    # COMPOSIO_API_KEY. Keep the import inside the skipif-guarded test body.
+    from graph import router_node
+
     out = router_node({"task": task, "messages": []})
     assert out["route"] == expected, (
         f"task={task!r} got={out['route']} expected={expected}"