fix: keep stdin open in stdio test to survive mcp>=1.27.0 transport-close cancel

[Copilot]

MCP 1.27.0 added finally: tg.cancel_scope.cancel() to Server.run() (PR modelcontextprotocol/python-sdk#2334). When stdin closes, all in-flight handlers are immediately cancelled — including tools/list — before they can respond. The previous test used subprocess.run() which closes stdin immediately after writing all messages, creating a race where the handler was cancelled before it could flush its response.

Changes

• tests/test_mcp_server.py: Replace subprocess.run() with subprocess.Popen() in test_stdio_initialize_and_tools_list
  • Write messages to stdin one at a time, keeping the pipe open
  • Collect responses in a daemon thread; block until id=2 (tools/list) arrives or 5 s elapses
  • Close stdin only after the response is confirmed received, then wait() for clean exit
  • Capture stderr before assertions; add explicit timeout assertion for clearer failure messages

# Before — stdin closed immediately after writes, races against handler cancel
proc = subprocess.run([...], input=input_text, ...)

# After — stdin stays open until tools/list response is in hand
proc = subprocess.Popen([...], stdin=PIPE, stdout=PIPE, ...)
for msg in (init, initialized, tools_list):
    proc.stdin.write(json.dumps(msg) + "\n")
    proc.stdin.flush()

t = threading.Thread(target=_collect, daemon=True)  # stops when id==2 seen
t.start()
t.join(timeout=5)

proc.stdin.close()   # only now signal EOF to the server
proc.wait(timeout=5)

[Copilot]

Pull request overview

Updates the stdio integration test to keep the subprocess stdin open long enough for tools/list to respond, avoiding handler cancellation introduced by MCP >=1.27.0 when the transport closes.

Changes:

• Switch test_stdio_initialize_and_tools_list from subprocess.run() to subprocess.Popen() to avoid closing stdin immediately.
• Send JSON-RPC messages incrementally and keep stdin open until the tools/list response is observed (or a timeout occurs).
• Collect stdout responses asynchronously and improve failure diagnostics by asserting on explicit timeouts and capturing stderr.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[teh-hippo]

Addressed the reviewer's race-condition concern in 7e669fa: the reader thread now signals a threading.Event the instant the tools/list response (id=2) arrives, the main thread waits on that event instead of t.join(timeout=…), and the reader is explicitly joined after the subprocess exits so no further mutations to responses can race the assertions. Stress-tested 50/50 locally; full suite still passes.