[v2] Dispatcher/ServerRunner receive-path swap — replaces BaseSession (#2710)

tests/conftest.py

@@ -1,6 +1,44 @@
+import os
+from collections.abc import Iterator
+
 import pytest
 
+# OpenTelemetry's `set_tracer_provider` is set-once per process, so the suite
+# uses a single span-capture mechanism: logfire's `capfire` fixture (its
+# `configure()` swaps span processors on repeat calls rather than re-setting
+# the provider). Logfire's default `distributed_tracing=None` emits a
+# RuntimeWarning + diagnostic span when incoming W3C trace context is
+# extracted; several tests exercise that propagation deliberately, so opt in
+# suite-wide. Set before logfire is imported anywhere.
+os.environ.setdefault("LOGFIRE_DISTRIBUTED_TRACING", "true")
+
+import opentelemetry.trace  # noqa: E402  (env var must be set before logfire import below)
+from logfire.testing import CaptureLogfire  # noqa: E402
+
+import mcp.shared._otel  # noqa: E402
+
 
 @pytest.fixture
 def anyio_backend():
     return "asyncio"
+
+
+@pytest.fixture(name="capfire")
+def _capfire_isolated(capfire: CaptureLogfire) -> Iterator[CaptureLogfire]:
+    """Override of logfire's `capfire` that scopes the MCP tracer to the test.
+
+    `capfire` installs a real tracer provider, and logfire's proxy machinery
+    mutates the cached `mcp.shared._otel._tracer` to delegate to it for the
+    rest of the process. Without isolation, every subsequent test in the same
+    worker would emit real spans, and `send_raw_request` would inject a real
+    `traceparent` into outbound `_meta`, breaking the interaction-suite
+    snapshots that pin `_meta={}` under a no-op tracer.
+
+    Setup points `_tracer` at the now-live provider so MCP spans record;
+    teardown replaces it with a `NoOpTracer`.
+    """
+    mcp.shared._otel._tracer = opentelemetry.trace.get_tracer_provider().get_tracer("mcp-python-sdk")
+    try:
+        yield capfire
+    finally:
+        mcp.shared._otel._tracer = opentelemetry.trace.NoOpTracer()