mcp: fix a race condition in logging

Reading the buffer outside of a critical section can lead to data corruption.

mcp/logging.go

@@ -10,6 +10,7 @@ import (
 	"context"
 	"encoding/json"
 	"log/slog"
+	"slices"
 	"sync"
 	"time"
 )
@@ -165,8 +166,7 @@ func (h *LoggingHandler) Handle(ctx context.Context, r slog.Record) error {
 
 func (h *LoggingHandler) handle(ctx context.Context, r slog.Record) error {
 	// Observe the rate limit.
-	// TODO(jba): use golang.org/x/time/rate. (We can't here because it would require adding
-	// golang.org/x/time to the go.mod file.)
+	// TODO(jba): use golang.org/x/time/rate.
 	h.mu.Lock()
 	skip := time.Since(h.lastMessageSent) < h.opts.MinInterval
 	h.mu.Unlock()
@@ -175,6 +175,7 @@ func (h *LoggingHandler) handle(ctx context.Context, r slog.Record) error {
 	}
 
 	var err error
+	var data json.RawMessage
 	// Make the buffer reset atomic with the record write.
 	// We are careful here in the unlikely event that the handler panics.
 	// We don't want to hold the lock for the entire function, because Notify is
@@ -185,6 +186,8 @@ func (h *LoggingHandler) handle(ctx context.Context, r slog.Record) error {
 		defer h.mu.Unlock()
 		h.buf.Reset()
 		err = h.handler.Handle(ctx, r)
+		// Clone the buffer as Bytes() references the internal buffer.
+		data = json.RawMessage(slices.Clone(h.buf.Bytes()))
 	}()
 	if err != nil {
 		return err
@@ -197,7 +200,7 @@ func (h *LoggingHandler) handle(ctx context.Context, r slog.Record) error {
 	params := &LoggingMessageParams{
 		Logger: h.opts.LoggerName,
 		Level:  slogLevelToMCP(r.Level),
-		Data:   json.RawMessage(h.buf.Bytes()),
+		Data:   data,
 	}
 	// We pass the argument context to Notify, even though slog.Handler.Handle's
 	// documentation says not to.