Do you think this might fix #6467?

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 78.75%. Comparing base (6427c07) to head (ec8358b).
Report is 1 commits behind head on main.

@@            Coverage Diff             @@
##             main    #6497      +/-   ##
==========================================
- Coverage   78.75%   78.75%   -0.01%     
==========================================
  Files         225      225              
  Lines       16581    16589       +8     
==========================================
+ Hits        13058    13064       +6     
- Misses       3523     3525       +2     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

While attempting to improve code coverage, I discovered that the previous fallback logic in del (used to stop the actor if McpWorkbench is GC’ed) did not behave correctly under certain loop conditions.

As a result, the shutdown logic has been rewritten.

This change simplifies the behavior, avoids misleading fallback attempts, and ensures the shutdown path is verifiable via testing (e.g., pytest.warns()).

@jackgerrits
Regarding #6467
This change might help in some scenarios, especially when McpWorkbench is deleted unintentionally (e.g., via GC or teardown).

That said, I’m not entirely confident about the lifecycle behavior on Windows — I’m not a specialist in how del interacts with the asyncio loop on that platform.

Ideally, McpWorkbench should be used via async with, in which case aexit guarantees proper shutdown of the MCP session.

The fact that it doesn’t close cleanly in #6467 is quite weird and deserves separate investigation.