Repository navigation
Description
Summary
Client.new() is a shallow copy, so a clone shares the original's MCPConfigTransport instance. That transport keeps the backend transports of its most recent connection in a transport-wide _transports list which _build_composite() overwrites on every connect. Calling Client.close() on either client — even one that never connected — therefore closes whatever list was last written, tearing down the other client's live backends. The victim's next call fails with ToolError: Error calling tool '<name>': Connection closed.
This is the symmetric case of #5618/#5596: just as a clone connecting must not close the original's backends, an unconnected client closing must not either.
Reproduction (offline, current main adf90b6)
Two real local stdio servers (tests/client/minimal_stdio_server.py), no network:
original = Client(config) # two servers
clone = original.new()
async with clone: # clone's backends are live
await clone.call_tool("a_pid") # ok
await original.close() # original never connected
await clone.call_tool("a_pid")
Observed:
clone pid before: <pid>
CLONE BROKEN: ToolError Error calling tool 'pid': Connection closed
The same happens in the other direction (an unconnected clone close kills a live original), and when the original was connected before the clone connected last. Expected: each client's close() only tears down the backends owned by its own connection; the other client keeps serving.
Suggested fix
Do not keep connection-scoped transports in a transport-wide mutable list. Give Client.new() an independent MCPConfigTransport over the same MCPConfig object (shared configuration, separate connection state), and/or have close() only act on transports owned by the calling client's connection. A connected client's close() must still tear down its own backends, and connect_session's exit stack remains the owner for normal teardown.
Tests
A fix with regression tests for both close directions plus connected-close teardown (in-memory close-recording transports and real stdio subprocesses) is ready. I'll open the linked PR with the red/green evidence.
Activity