Closing a legacy Client during initialize leaves an unhandled rejection in workerd

[zackbart]

Closing a legacy Client during initialize leaves an unhandled rejection under workerd

@modelcontextprotocol/client@2.0.0 can reject a Promise without a handler when its transport closes during the pre-2026 initialize handshake. The application's client.connect(transport) Promise already has a rejection handler. Under workerd, Vitest reports SdkError: Not connected from Client._notificationViaCodec via _legacyHandshake; Node does not report an unhandled rejection for the same sequence.

Reproduction in a Vitest project using workerd:

import { Client, InMemoryTransport } from "@modelcontextprotocol/client";
import { McpServer } from "@modelcontextprotocol/server";
import { it } from "vitest";

it("closes during legacy initialize", async () => {
  for (let ticks = 0; ticks < 30; ticks++) {
    const [clientTransport, serverTransport] = InMemoryTransport.createLinkedPair();
    const server = new McpServer({ name: "downstream", version: "1.0.0" });
    await server.connect(serverTransport);
    const client = new Client(
      { name: "probe", version: "1.0.0" },
      { versionNegotiation: { mode: "legacy" } },
    );
    const connecting = client.connect(clientTransport).catch(() => {});
    for (let tick = 0; tick < ticks; tick++) await Promise.resolve();
    await clientTransport.close();
    await connecting;
    await server.close();
  }
});

Expected: all failures caused by close belong to connecting; no unhandled rejection.

Actual: workerd reports four unhandled NOT_CONNECTED rejections at _legacyHandshake's notifications/initialized send. The first such report was reproduced with workerd 1.20260722.1 and @modelcontextprotocol/client@2.0.0. The issue can occur when a host cancels a connection after initialize replies but before notifications/initialized is sent.

Connecta temporarily skips that notification only when its own connection close has already aborted the handshake. A fix in the SDK would let us remove this workaround.

[DorianChn]

I’d like to take this with a regression test that exercises the close-during-legacy-initialize sequence in workerd, then keep the fix scoped to the rejection path. I’m checking the reproduction against current main before proposing the implementation.