Streamable HTTP client stays closed after an interrupted response and raises ExceptionGroup on cleanup

從 mcp.client.streamable_http 中的 streamable_http_client 入口點開始，並針對 mcp-failure-lab 執行提供的 repro.py。追蹤中斷的斷線回應如何影響用戶端和工作群組清理；當斷線失敗得到回報、同一個用戶端上的後續 ping 成功，且離開內容脈絡時不會引發 ExceptionGroup，即表示完成。

Description

With the current 2.x release, one interrupted Streamable HTTP tool response permanently closes the Python client session. The failing request is reported, but a subsequent normal request on the same client fails with MCPError: Connection closed. Exiting streamable_http_client(...) then raises an unhandled ExceptionGroup from its background task group.

The same MCP server remains healthy. Under the identical sequence, @modelcontextprotocol/sdk@1.30.0 reports the interrupted request and successfully completes the next ping on the same client.

Expected behavior

The interrupted request should fail, while a subsequent request should be able to use a fresh HTTP connection and succeed. Transport cleanup should not raise an additional ExceptionGroup.

Actual behavior

1. disconnect interrupts its active HTTP response.
2. Python reports the request failure.
3. A following ping fails immediately with MCPError: Connection closed.
4. Context cleanup raises ExceptionGroup: unhandled errors in a TaskGroup whose nested cause is httpx2.RemoteProtocolError: Server disconnected without sending a response.

Comparison

The official TypeScript client 1.30.0 receives TypeError: fetch failed for the deliberately interrupted request, then successfully sends ping through the same client. This suggests the server remains available and the problem is in Python client recovery/cleanup.

I also searched the existing issues. #3494 concerns a server-side incomplete SSE body during initialize, and #3441 concerns an unresolved 1.x request; neither appears to cover this v2 same-client recovery failure.

Minimal reproduction

# Terminal 1
npx mcp-failure-lab@0.9.0 serve --transport http --host 127.0.0.1 --port 43118 --path /mcp

# Terminal 2: save as repro.py, then run: python repro.py
import anyio
from mcp import ClientSession
from mcp.client.streamable_http import streamable_http_client

URL = "http://127.0.0.1:43118/mcp"

async def main():
    try:
        async with streamable_http_client(URL) as streams:
            async with ClientSession(
                streams[0], streams[1], read_timeout_seconds=1.0
            ) as session:
                await session.initialize()
                try:
                    await session.call_tool(
                        "disconnect", {}, read_timeout_seconds=0.6
                    )
                except BaseException as error:
                    print("disconnect:", type(error).__name__, error)

                # Expected: succeeds. Actual: MCPError: Connection closed
                print("ping:", await session.call_tool("ping", {}))
    except BaseException as error:
        # Actual: ExceptionGroup from streamable_http_client cleanup
        print("cleanup:", type(error).__name__, error)

anyio.run(main)

Environment

Python 3.12.14; mcp 2.2.0; macOS arm64. Server: mcp-failure-lab 0.9.0; protocol 2026-07-28.