René's URL Explorer Experiment

Title: [v1.x] Backport transport test deflakes from main (tests-only) by maxisbey · Pull Request #2837 · modelcontextprotocol/python-sdk · GitHub

Open Graph Title: [v1.x] Backport transport test deflakes from main (tests-only) by maxisbey · Pull Request #2837 · modelcontextprotocol/python-sdk

X Title: [v1.x] Backport transport test deflakes from main (tests-only) by maxisbey · Pull Request #2837 · modelcontextprotocol/python-sdk

Description: Backports the transport test deflaking work from main (#2764, #2765, #2767, #2788) to v1.x. Tests-only: the diff touches nothing under src/. Stacked on #2717 (maxisbey/v1x-interaction-backport), which supplies the in-process streaming bridge these tests are built on. Motivation and Context The v1.x transport tests still run over real sockets with uvicorn servers in threads or subprocesses, and CI pays for it: ephemeral port races, server startup polling, and a real-clock session-timeout bet that loses under scheduler load. main fixed all of this by moving the tests in process; this PR ports those rewrites to v1.x so the v1.x matrix stops flaking on them. Two of the rewritten files have direct v1.x CI evidence (test_streamable_http_multiple_reconnections and test_json_response failures on Windows legs); the rest share the same socket-based harness and mechanism. The session-timeout deflake matters for #2717 itself: the suite that PR backports still carries the pre-#2788 real-clock test, which fails roughly 1 in 3 full-suite runs under coverage load, once per transport leg. Folding the first three commits here into #2717 instead would work equally well. Adaptations made for v1.x, each noted in the commit that makes it: The 4-line src/ delta from #2767 (a pragma removal and a stale docstring note) is dropped; v1.x has no strict pragma checking, so the retained pragma is harmless. 9 wildcard-Accept-header test params are omitted because v1.x predates the RFC 7231 wildcard parsing from #2152. The v1.x-only deprecation tests (streamablehttp_client alias, deprecated transport params) are re-added in in-process form, so no existing v1.x test is lost. In-process serving exposes two v1.x-specific teardown behaviors that the subprocess harness hid: sse-starlette < 3.0 binds a module-global exit event to the first event loop that serves a response (guarded with a reset fixture, exercised by the lowest-direct legs), and the server transport leaks anyio memory streams whose GC-time warnings would otherwise escape the per-test filters (handled with scoped filters plus a per-test gc.collect() so isolated runs stay clean). The streaming bridge now ignores response messages after the response has completed, matching what a client of a real ASGI server observes, with contract tests pinning the behavior. One ported test verifies inside its connection loop because Python 3.11 loses line tracing after an async-with teardown throws (python/cpython#106749), which otherwise misreports coverage on 3.11 legs. How Has This Been Tested? Full v1.x gate (pytest plus 100% line and branch coverage) green throughout the series. The rewritten files were run repeatedly in a scratch CI environment replicating the v1.x matrix (Windows and Ubuntu, Python 3.10 to 3.14, locked and lowest-direct resolutions): roughly 900 repetitions of the affected files with zero failures, including the lowest-direct legs that install sse-starlette 1.6.1. Contrast evidence for the timeout deflake: on the pre-deflake base, 5 of 8 full-suite runs under coverage load failed on test_session_level_timeout_applies_to_every_request; with the virtual-clock version the failure has never reproduced (28 full-suite runs plus 150 isolated repetitions). The full-suite failures that did appear during matrix stress were all in files this PR does not touch (the known Windows stdio child-cleanup and experimental-tasks flakes), which are out of scope here because fixing them requires src/ changes. Breaking Changes None. Tests-only; no public interface is touched. Types of changes Bug fix (non-breaking change which fixes an issue) New feature (non-breaking change which adds functionality) Breaking change (fix or feature that would cause existing functionality to change) Documentation update Checklist I have read the MCP Documentation My code follows the repository's style guidelines New and existing tests pass locally I have added appropriate error handling I have added or updated documentation as needed Additional context Draft while #2717 is in review. If #2717 is amended, the only coupling points are tests/interaction/transports/__init__.py (the re-export added here) and the small _bridge.py and conftest.py changes; everything else is independent of its internals. Once #2717 merges this can be retargeted to v1.x. AI Disclaimer

Open Graph Description: Backports the transport test deflaking work from main (#2764, #2765, #2767, #2788) to v1.x. Tests-only: the diff touches nothing under src/. Stacked on #2717 (maxisbey/v1x-interaction-backport), wh...

X Description: Backports the transport test deflaking work from main (#2764, #2765, #2767, #2788) to v1.x. Tests-only: the diff touches nothing under src/. Stacked on #2717 (maxisbey/v1x-interaction-backport), wh...

Opengraph URL: https://github.com/modelcontextprotocol/python-sdk/pull/2837

X: @github

direct link

Domain: github.com

Links:

Viewport: width=device-width