RequestResponder.cancel sends JSON-RPC response, violating cancellation spec

[q-thomasdickson]

Description

RequestResponder.cancel in mcp/shared/session.py sends an ErrorData(code=0, message="Request cancelled") JSON-RPC response back to the sender after a notifications/cancelled is received. The cancellation spec explicitly says receivers SHOULD NOT do this:

Receivers of cancellation notifications SHOULD:

• Stop processing the cancelled request
• Free associated resources
• Not send a response for the cancelled request

— https://github.com/modelcontextprotocol/modelcontextprotocol/blob/main/docs/specification/draft/basic/utilities/cancellation.mdx#L33-L36

Reproduction

mcp 1.27.0, Python 3.11. Driving the server over stdio:

1. Send initialize + notifications/initialized.
2. Send any tools/call whose handler enters an await (e.g. await asyncio.sleep).
3. Send notifications/cancelled with the request id.
4. Observe stdout: an unsolicited response is sent for the cancelled request.

{"jsonrpc": "2.0", "id": 99, "error": {"code": 0, "message": "Request cancelled"}}

Source:
https://github.com/modelcontextprotocol/python-sdk/blob/v1.27.0/src/mcp/shared/session.py#L137-L150

async def cancel(self) -> None:
    ...
    self._cancel_scope.cancel()
    self._completed = True
    # Send an error response to indicate cancellation
    await self._session._send_response(
        request_id=self.request_id,
        response=ErrorData(code=0, message="Request cancelled", data=None),
    )

Impact

Strict clients treat the unexpected response as an unknown message id and drop the transport. Claude Code, for example, logs Received a response for an unknown message ID, closes the stdio transport, and reconnects. With long-running tools (custom polling tools, long DB queries, etc.) every cancellation costs a reconnect cycle.

Suggested fix

Drop the _send_response call from RequestResponder.cancel. The cancel scope is already cancelled and _completed is already set, so the request is correctly cleaned up server-side without violating the spec.

Workaround

Until this lands, downstream servers can monkey-patch RequestResponder.cancel:

from mcp.shared.session import RequestResponder

async def _cancel_without_response(self) -> None:
    if not self._entered:
        raise RuntimeError("RequestResponder must be used as a context manager")
    if not self._cancel_scope:
        raise RuntimeError("No active cancel scope")
    self._cancel_scope.cancel()
    self._completed = True

RequestResponder.cancel = _cancel_without_response

Environment

• mcp 1.27.0
• Python 3.11
• stdio transport