test(crewai): keep the tests off the network

instrumentation-loongsuite/loongsuite-instrumentation-crewai/CHANGELOG.md

@@ -12,6 +12,11 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 - Keep incidental provider requests out of the CrewAI VCR cassettes so a
   first-use model-metadata fetch cannot fail the recorded tests.
   ([#270](https://github.com/alibaba/loongsuite-python/pull/270))
+- Keep the CrewAI tests off the network: turn off crewAI's telemetry and
+  version check, stop VCR from passing requests through to real hosts, and
+  block real connections. A telemetry export in flight let concurrent LLM
+  calls skip their cassettes, so tests failed at random.
+  ([#287](https://github.com/alibaba/loongsuite-python/pull/287))
 
 ## Version 0.9.0 (2026-09-07)
 

instrumentation-loongsuite/loongsuite-instrumentation-crewai/tests/conftest.py

@@ -13,6 +13,7 @@
 # limitations under the License.
 
 import os
+from pathlib import Path
 
 import pytest
 
@@ -24,6 +25,35 @@
 # ``litellm.litellm_core_utils.get_model_cost_map``.
 os.environ.setdefault("LITELLM_LOCAL_MODEL_COST_MAP", "True")
 
+# crewAI exports anonymous telemetry to telemetry.crewai.com from a background
+# thread, and checks PyPI for a newer release when a verbose crew starts.
+# Neither belongs in these offline tests. Passing the telemetry requests
+# through VCR (``ignore_hosts``) made them flaky: while vcrpy sends such a
+# request to the real host, ``vcr.patch.force_reset`` undoes its patches for
+# every HTTP client in the process, so an LLM call crewAI made meanwhile on its
+# worker thread skipped the cassette and reached the real provider. When that
+# window spanned the end of a test, the patches went back out of order and the
+# next test was replayed from the previous test's cassette.
+os.environ["CREWAI_DISABLE_TELEMETRY"] = "true"
+os.environ["CREWAI_DISABLE_VERSION_CHECK"] = "true"
+
+_TESTS_DIR = Path(__file__).parent
+
+
+def pytest_collection_modifyitems(items):
+    """Block real network access in these tests.
+
+    Every response they need is recorded in the cassettes, so a connection
+    attempt means a request got past VCR: pytest-recording's
+    ``block_network`` mark makes it fail with "Network is disabled" instead
+    of reaching a real endpoint. As with ``--block-network``, recording new
+    cassettes (``--record-mode`` other than ``none``) still has network
+    access.
+    """
+    for item in items:
+        if item.path.is_relative_to(_TESTS_DIR):
+            item.add_marker(pytest.mark.block_network)
+
 
 def scrub_response_headers(response):
     """Clean up common sensitive response headers before recording."""
@@ -58,15 +88,7 @@ def vcr_config():
         ],
         "before_record_response": scrub_response_headers,
         "decode_compressed_response": True,
-        # Requests incidental to the code under test must not reach the
-        # cassettes, or they fail playback under ``record_mode=none``. The
-        # litellm package's own conftest ignores its telemetry hosts for the
-        # same reason.
-        "ignore_hosts": [
-            "raw.githubusercontent.com",
-            "us.i.posthog.com",
-            "app.posthog.com",
-            "api.litellm.ai",
-            "telemetry.crewai.com",
-        ],
+        # No ``ignore_hosts``: vcrpy sends requests for ignored hosts to the
+        # real network, and while it does, concurrent LLM calls skip the
+        # cassette (see the telemetry switches above).
     }