René's URL Explorer Experiment

Title: ci: run strict-no-cover in scripts/test to catch stale pragmas locally by maxisbey · Pull Request #2305 · modelcontextprotocol/python-sdk · GitHub

Open Graph Title: ci: run strict-no-cover in scripts/test to catch stale pragmas locally by maxisbey · Pull Request #2305 · modelcontextprotocol/python-sdk

X Title: ci: run strict-no-cover in scripts/test to catch stale pragmas locally by maxisbey · Pull Request #2305 · modelcontextprotocol/python-sdk

Description: Adds strict-no-cover as a final step in ./scripts/test so stale # pragma: no cover annotations are caught locally before pushing. Updates CLAUDE.md with accurate testing guidance and documents the three coverage-pragma types. Motivation and Context PR #2302 passed ./scripts/test (100% coverage, all green) but failed CI in two ways: strict-no-cover flagged lines 61–63 of src/mcp/server/stdio.py — the new test executed them, but they still carried # pragma: no cover. coverage report saw 100% because the pragma excludes the line; strict-no-cover saw that a "never covered" line was in fact executed. Python 3.14 Windows branch coverage — a nested async with in the test produced a missing ->exit arc, a known coverage.py quirk. CLAUDE.md claimed ./scripts/test "matches CI exactly." It didn't: CI runs strict-no-cover as a separate step after coverage report (Linux only), and CI runs the full 3.10–3.14 × {ubuntu, windows} matrix. The first gap is closed here; the second is documented. CI does not invoke scripts/test — it inlines the coverage commands in shared.yml — so this change is purely local and does not run strict-no-cover twice. How Has This Been Tested? Reproduced the #2302 failure mode: added # pragma: no cover to a line that tests/server/test_stdio.py executes, then ran both paths. Full path (./scripts/test): Before: coverage report → 100%, exit 0. Stale pragma undetected. After: coverage report → 100%, then strict-no-cover → ❎ 1 lines wrongly marked, exit 1. ~21s wall time (vs ~20s before). Fast targeted path (documented in CLAUDE.md): uv run --frozen coverage erase uv run --frozen coverage run -m pytest tests/server/test_stdio.py uv run --frozen coverage combine uv run --frozen coverage report --include='src/mcp/server/stdio.py' --fail-under=0 UV_FROZEN=1 uv run --frozen strict-no-cover ~4s, catches the stale pragma, no false positives on partial runs (strict-no-cover only reports excluded ∩ executed — fewer executed lines just means fewer chances to catch, never a wrong flag). uv.lock pollution fix: strict-no-cover internally spawns subprocess.run(['uv', 'run', 'coverage', 'json', ...]) — no --frozen. On machines with registry overrides this rewrites uv.lock. Setting UV_FROZEN=1 propagates to the child process. Verified uv.lock stays clean through a full run. Breaking Changes None. ./scripts/test takes ~1s longer and exits nonzero when a stale pragma is found. Types of changes Bug fix (non-breaking change which fixes an issue) 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 The # pragma: no branch quirk for nested async with is documented as what-it-is, not what-to-do-about-it; the existing ~100 annotations in the tree are the pattern reference.

Open Graph Description: Adds strict-no-cover as a final step in ./scripts/test so stale # pragma: no cover annotations are caught locally before pushing. Updates CLAUDE.md with accurate testing guidance and documents the ...

X Description: Adds strict-no-cover as a final step in ./scripts/test so stale # pragma: no cover annotations are caught locally before pushing. Updates CLAUDE.md with accurate testing guidance and documents the ...

Opengraph URL: https://github.com/modelcontextprotocol/python-sdk/pull/2305

X: @github

direct link

Domain: github.com

Links:

Viewport: width=device-width