feat: Added ping command

server/internal/grpc/pb/vantage.pb.go

@@ -158,8 +158,14 @@ type ServerCommand struct {
 	RunStep          *RunStepCmd          `json:"run_step,omitempty"`
 	CleanupWorkspace *CleanupWorkspaceCmd `json:"cleanup_workspace,omitempty"`
 	OpenProxy        *OpenProxyCmd        `json:"open_proxy,omitempty"`
+	Ping             *PingCmd             `json:"ping,omitempty"`
 }
 
+// PingCmd is a server-originated liveness beat. It carries nothing and expects
+// no reply: its arrival is the entire message. See the .proto for why gRPC
+// keepalive is not sufficient on its own.
+type PingCmd struct{}
+
 type CleanupWorkspaceCmd struct {
 	WorkspaceId string `json:"workspace_id"`
 }

server/internal/grpc/server.go

@@ -212,11 +212,27 @@ func (s *vantageServer) CommandStream(stream pb.Vantage_CommandStreamServer) err
 		}
 	}()
 
+	// The heartbeat is what lets the agent tell a live stream from an orphaned
+	// one. gRPC keepalive cannot: behind an L7 proxy the agent's connection
+	// terminates at the proxy, which answers pings on its own behalf, so a dead
+	// pod leaves the agent blocked in Recv forever with commands vanishing into
+	// a stream nobody is serving. A message that originates here is the only
+	// thing that proves this process is still on the other end.
+	ping := time.NewTicker(pingInterval)
+	defer ping.Stop()
+
 	ctx := stream.Context()
 	for {
 		select {
 		case <-ctx.Done():
 			return nil
+		case <-ping.C:
+			// A failed send is the point: it is how this side learns the stream
+			// is gone, which runs the deferred release and frees the agent's
+			// presence claim for whichever pod it reconnects to.
+			if err := stream.Send(&pb.ServerCommand{Ping: &pb.PingCmd{}}); err != nil {
+				return err
+			}
 		case cmd, ok := <-ch:
 			if !ok {
 				return nil
@@ -228,6 +244,11 @@ func (s *vantageServer) CommandStream(stream pb.Vantage_CommandStreamServer) err
 	}
 }
 
+// How often the server beats on an idle command stream. Comfortably under the
+// agent's staleness threshold, so a single dropped beat does not cost a
+// reconnect.
+const pingInterval = 20 * time.Second
+
 // StartGRPC serves the agent API until stop is called.
 //
 // It returns a stop function rather than serving forever because an abrupt exit