mcp: closing a session right after a cancelled call drops notifications/cancelled, so the server handler keeps running

未关闭
  • 哪里不对？
#1,505 1 条评论 0 个 reaction 已指派 0 人 在 GitHub 查看

维护者通常 1 天内回复

@vshalpnjabi 已经在做这个了。

开始于 2026年11月6日。

  • #1 来自 @vshalpnjabi —— 已关闭，未合并
  • #1508 来自 @vshalpnjabi —— 已关闭，未合并

评估

难度
4/5
预计耗时
3-5 天
新手友好度
48/100
Issue 类型
缺陷
描述清晰度
描述清楚
活跃度
活跃
技术栈
go
领域
backend-api-design

调研方向

Start by tracing call() and ClientSession.Close through the JSON-RPC connection shutdown path, then inspect how outgoing notifications are tracked and drained. The issue’s repro can confirm the failure and provide a starting test. Done means closing immediately after a cancelled call still delivers notifications/cancelled, cancels the server handler, and avoids waiting for the 5-second timeout.

由索引模型根据 Issue 内容生成。

描述

Summary

Since v1.8.0 (#1151), call() retires a cancelled call and sends notifications/cancelled from a goroutine. If the client closes the session right after the call returns (a graceful shutdown that cancels in-flight tool calls and then closes, for example), the notification is dropped. On a stateful streamable HTTP server the tool handler is then never cancelled, and ClientSession.Close blocks for closeDeleteTimeout (5s).

MCPGODEBUG=blockingcancelnotify=1 restores the v1.7.0 ordering, but it is scheduled for removal in v1.9.0.

Repro (v1.8.0, also matches main)
package repro

import (
	"context"
	"errors"
	"net/http"
	"net/http/httptest"
	"testing"
	"time"

	"github.com/modelcontextprotocol/go-sdk/mcp"
)

func TestCloseAfterCanceledCallCancelsTheServerHandler(t *testing.T) {
	entered := make(chan struct{}, 1)
	canceled := make(chan struct{}, 1)
	release := make(chan struct{})

	server := mcp.NewServer(&mcp.Implementation{Name: "server", Version: "1"}, nil)
	server.AddTool(&mcp.Tool{Name: "wait", InputSchema: map[string]any{"type": "object"}},
		func(ctx context.Context, _ *mcp.CallToolRequest) (*mcp.CallToolResult, error) {
			entered <- struct{}{}
			select {
			case <-ctx.Done():
				canceled <- struct{}{}
				return nil, ctx.Err()
			case <-release:
				return &mcp.CallToolResult{}, nil
			}
		})
	httpServer := httptest.NewServer(mcp.NewStreamableHTTPHandler(func(*http.Request) *mcp.Server { return server }, nil))
	t.Cleanup(httpServer.Close)
	t.Cleanup(func() { close(release) })

	client := mcp.NewClient(&mcp.Implementation{Name: "client", Version: "1"}, nil)
	session, err := client.Connect(t.Context(), &mcp.StreamableClientTransport{Endpoint: httpServer.URL}, nil)
	if err != nil {
		t.Fatal(err)
	}

	ctx, cancel := context.WithCancel(t.Context())
	go func() {
		<-entered
		cancel()
	}()
	if _, err := session.CallTool(ctx, &mcp.CallToolParams{Name: "wait"}); !errors.Is(err, context.Canceled) {
		t.Fatalf("CallTool: got %v, want context.Canceled", err)
	}

	start := time.Now()
	_ = session.Close()
	closed := time.Since(start)

	select {
	case <-canceled:
		t.Logf("handler canceled; Close took %v", closed.Round(time.Millisecond))
	case <-time.After(2 * time.Second):
		t.Fatalf("server handler never saw the cancellation; Close took %v", closed.Round(time.Millisecond))
	}
}

Go 1.26.10, linux/amd64:

  • Default: fails 20/20, server handler never saw the cancellation; Close took 5s.
  • MCPGODEBUG=blockingcancelnotify=1: passes 20/20, handler canceled; Close took 1ms.
Why
  1. call() runs conn.Retire(call, ...), starts the notify goroutine and returns.
  2. ClientSession.Close → jsonrpc2.Connection.Close marks the connection closing.
  3. The goroutine's Notify finds no outgoing calls and no incoming requests, so shuttingDown(ErrClientClosing) rejects it. The notification is never written.
  4. The streamable client sends DELETE. serveStatefulDELETE → ServerSession.Close cancels only subscriptions/listen handlers, then conn.Close waits for in-flight requests, including the tool handler that is still blocked on its context.
  5. The DELETE hangs until the client gives up after closeDeleteTimeout and cancels its connection. The handler keeps running.

So closing the connection does not signal the cancellation to a stateful streamable server, which is the assumption in #1241.

Suggestion

Keep the caller's return immediate (the point of #1150), but register the pending notification synchronously before call() returns, e.g. bump outgoingNotifications (or an equivalent "pending cancel" counter) before spawning the goroutine. Then Close allows the write and its drain waits for it, bounded by notifyCancellationTimeout.

We hit this in a graceful-shutdown path that cancels tool calls, waits for them to return, then closes the session; we're running with blockingcancelnotify=1 until there is a fix.

主要语言
Go
星标
5.3k
派生
607
平均合并
1 天 15 分钟
30 天内合并 PR
76

环境准备

在 Codespaces 中打开

在浏览器里用你自己的 GitHub 账号启动这个项目的开发容器。

从这里开始

  1. 先读完整个 Issue，再读项目的贡献指南。
  2. 在 Issue 下留言说明你要接手 —— 这能避免两个人做同样的事。
  3. Fork 仓库，在一个分支上完成修改。
  4. 提交 Pull Request，并在描述里引用这个 Issue 编号。

modelcontextprotocol/go-sdk 的其他 Issue

查看 modelcontextprotocol/go-sdk 的全部 Issue

相似的 Issue

更多 Go Issue

把新 issue 发到你的邮箱

精选适合新手参与的 GitHub issue 摘要。