mcp: make the streamable client transport less strict (#524)

In practice, the streamable client transport was having trouble
connecting to various backends, because they are nonconformant in
various ways.

While it would be nice if all servers conformed to the spec, in practice
there are certain spec violations that are recoverable, and we can and
should recover them. Specifically:

- tolerate 404 instead of 405 for the hanging GET (#393)
- tolerate (=ignore) spurious response body for notifications and
responses, since we know none are expected

Additionally, fix a bug that we weren't parsing Content-Type correctly.

Fixes #521

mcp/streamable.go

@@ -12,6 +12,7 @@ import (
 	"fmt"
 	"io"
 	"iter"
+	"log/slog"
 	"math"
 	"math/rand/v2"
 	"net/http"
@@ -981,6 +982,14 @@ type StreamableClientTransport struct {
 	// MaxRetries is the maximum number of times to attempt a reconnect before giving up.
 	// It defaults to 5. To disable retries, use a negative number.
 	MaxRetries int
+
+	// TODO(rfindley): propose exporting these.
+	// If strict is set, the transport is in 'strict mode', where any violation
+	// of the MCP spec causes a failure.
+	strict bool
+	// If logger is set, it is used to log aspects of the transport, such as spec
+	// violations that were ignored.
+	logger *slog.Logger
 }
 
 // These settings are not (yet) exposed to the user in
@@ -1025,6 +1034,8 @@ func (t *StreamableClientTransport) Connect(ctx context.Context) (Connection, er
 		incoming:   make(chan jsonrpc.Message, 10),
 		done:       make(chan struct{}),
 		maxRetries: maxRetries,
+		strict:     t.strict,
+		logger:     t.logger,
 		ctx:        connCtx,
 		cancel:     cancel,
 		failed:     make(chan struct{}),
@@ -1039,6 +1050,8 @@ type streamableClientConn struct {
 	cancel     context.CancelFunc
 	incoming   chan jsonrpc.Message
 	maxRetries int
+	strict     bool         // from [StreamableClientTransport.strict]
+	logger     *slog.Logger // from [StreamableClientTransport.logger]
 
 	// Guard calls to Close, as it may be called multiple times.
 	closeOnce sync.Once
@@ -1152,9 +1165,11 @@ func (c *streamableClientConn) Write(ctx context.Context, msg jsonrpc.Message) e
 	}
 
 	var requestSummary string
+	var isCall bool
 	switch msg := msg.(type) {
 	case *jsonrpc.Request:
 		requestSummary = fmt.Sprintf("sending %q", msg.Method)
+		isCall = msg.IsCall()
 	case *jsonrpc.Response:
 		requestSummary = fmt.Sprintf("sending jsonrpc response #%d", msg.ID)
 	default:
@@ -1209,11 +1224,24 @@ func (c *streamableClientConn) Write(ctx context.Context, msg jsonrpc.Message) e
 		}
 	}
 	if resp.StatusCode == http.StatusNoContent || resp.StatusCode == http.StatusAccepted {
+		// [§2.1.4]: "If the input is a JSON-RPC response or notification:
+		// If the server accepts the input, the server MUST return HTTP status code 202 Accepted with no body."
+		//
+		// [§2.1.4]: https://modelcontextprotocol.io/specification/2025-06-18/basic/transports#listening-for-messages-from-the-server
+		resp.Body.Close()
+		return nil
+	} else if !isCall && !c.strict {
+		// Some servers return 200, even with an empty json body.
+		// Ignore this response in non-strict mode.
+		if c.logger != nil {
+			c.logger.Warn(fmt.Sprintf("unexpected status code %d from non-call", resp.StatusCode))
+		}
 		resp.Body.Close()
 		return nil
 	}
 
-	switch ct := resp.Header.Get("Content-Type"); ct {
+	contentType := strings.TrimSpace(strings.SplitN(resp.Header.Get("Content-Type"), ";", 2)[0])
+	switch contentType {
 	case "application/json":
 		go c.handleJSON(requestSummary, resp)
 
@@ -1223,7 +1251,7 @@ func (c *streamableClientConn) Write(ctx context.Context, msg jsonrpc.Message) e
 
 	default:
 		resp.Body.Close()
-		return fmt.Errorf("%s: unsupported content type %q", requestSummary, ct)
+		return fmt.Errorf("%s: unsupported content type %q", requestSummary, contentType)
 	}
 	return nil
 }
@@ -1294,18 +1322,36 @@ func (c *streamableClientConn) handleSSE(requestSummary string, initialResp *htt
 		newResp, err := c.reconnect(lastEventID)
 		if err != nil {
 			// All reconnection attempts failed: fail the connection.
-			c.fail(fmt.Errorf("%s: failed to reconnect: %v", requestSummary, err))
+			c.fail(fmt.Errorf("%s: failed to reconnect (session ID: %v): %v", requestSummary, c.sessionID, err))
 			return
 		}
 		resp = newResp
 		if resp.StatusCode == http.StatusMethodNotAllowed && persistent {
+			// [§2.2.3]: "The server MUST either return Content-Type:
+			// text/event-stream in response to this HTTP GET, or else return HTTP
+			// 405 Method Not Allowed, indicating that the server does not offer an
+			// SSE stream at this endpoint."
+			//
+			// [§2.2.3]: https://modelcontextprotocol.io/specification/2025-06-18/basic/transports#listening-for-messages-from-the-server
+
 			// The server doesn't support the hanging GET.
 			resp.Body.Close()
 			return
 		}
+		if resp.StatusCode == http.StatusNotFound && persistent && !c.strict {
+			// modelcontextprotocol/gosdk#393: some servers return NotFound instead
+			// of MethodNotAllowed for the persistent GET.
+			//
+			// Treat this like MethodNotAllowed in non-strict mode.
+			if c.logger != nil {
+				c.logger.Warn("got 404 instead of 405 for hanging GET")
+			}
+			resp.Body.Close()
+			return
+		}
 		// (see equivalent handling in [streamableClientConn.Write]).
 		if resp.StatusCode == http.StatusNotFound {
-			c.fail(fmt.Errorf("%s: failed to reconnect: %w", requestSummary, errSessionMissing))
+			c.fail(fmt.Errorf("%s: failed to reconnect (session ID: %v): %w", requestSummary, c.sessionID, errSessionMissing))
 			return
 		}
 		if resp.StatusCode < 200 || resp.StatusCode >= 300 {