fix: suppress ClosedResourceError tracebacks in streamable HTTP transport

[elijahgjacob]

Summary

• Suppresses noisy ClosedResourceError tracebacks from the MCP SDK's streamable HTTP session manager
• These fire when a client disconnects mid-request (timeout, network drop, Cursor closing connection)
• Harmless in stateless_http=True mode — each session is isolated, the server self-heals — but the full ExceptionGroup traceback clutters deploy logs
• Sets mcp.server.streamable_http_manager logger to CRITICAL as a targeted suppression

Root cause

Known bug in mcp Python SDK v1.26.0. The error handler tries to send a log message back to an already-closed write stream, causing an unhandled anyio.ClosedResourceError. Upstream fix: python-sdk PR #2072.

What this does NOT fix

• The underlying race condition (that's an SDK fix)
• Uvicorn graceful shutdown timeout messages during Railway redeploys (benign, 0 tasks affected)

Test plan

• 61 unit tests pass
• 22 smoke tests pass against local Docker
• Verified server starts and responds to MCP requests normally
• Change is a single-line logger level override — no functional behavior change

Made with Cursor