mcp: a canceled write on IOTransport/StdioTransport cannot leave the write queue #1498

Checklist

• I searched the existing issues, and this has not been reported before.
• This is a real problem that I ran into myself, using the latest release of the SDK.
• I ran into it in a production system, not in a toy example, a benchmark, or an automated or AI-driven scan of the code.
• I (a human) examined the relevant SDK code, and wrote or reviewed every line of this report. It is not unreviewed AI output.
• This report is brief and concise, and sticks to the facts needed to understand the problem.

What happened?

ioConn.Write checks the context only before taking its write mutex. If a write is blocked because the peer stopped reading, another Write whose context is canceled keeps waiting for the mutex instead of returning, and the blocked write itself cannot be interrupted by canceling its context. Callers that bound a request or notification with a deadline therefore hang until the peer reads again.

Expected: a write that has not started returns ctx.Err() without sending anything, and canceling the write in progress stops it (closing the connection, since a partial frame cannot be withdrawn).

AI assistance: the investigation was done with Claude; I reviewed the report and the SDK code.

How to reproduce

package main

import (
	"context"
	"fmt"
	"net"
	"time"

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

func main() {
	local, peer := net.Pipe() // the peer never reads, as with a stalled server
	defer peer.Close()
	conn, err := (&mcp.IOTransport{Reader: local, Writer: local}).Connect(context.Background())
	if err != nil {
		panic(err)
	}

	write := func(name string, timeout time.Duration) {
		ctx, cancel := context.WithTimeout(context.Background(), timeout)
		defer cancel()
		done := make(chan error, 1)
		go func() { done <- conn.Write(ctx, &jsonrpc.Request{Method: name}) }()
		select {
		case err := <-done:
			fmt.Printf("%s: returned %v\n", name, err)
		case <-time.After(timeout + 2*time.Second):
			fmt.Printf("%s: still blocked 2s after its %v deadline\n", name, timeout)
		}
	}
	go write("active", 200*time.Millisecond)
	time.Sleep(50 * time.Millisecond)
	write("queued", 100*time.Millisecond)
	time.Sleep(time.Second)
}

Output with v1.8.0:

queued: still blocked 2s after its 100ms deadline
active: still blocked 2s after its 200ms deadline

Where did you hit this?

Frame hosts MCP plugin processes over stdio; when a plugin stopped reading, shutdown hung on pending writes, so Frame had to wrap SDK writes in its own cancelable layer.

Go MCP SDK version

v1.8.0 (also on main at f91620f)

Go version

go version go1.27.3 darwin/arm64

Logs

No response