feat: implement decision-based routing with plugin architecture

Signed-off-by: bitliu <[email protected]>

Author: Xunzhuo

src/semantic-router/cmd/main.go

@@ -54,6 +54,10 @@ func main() {
 		logging.Fatalf("Failed to load config: %v", err)
 	}
 
+	// Set the initial configuration in the global config
+	// This is important for Kubernetes mode where the controller will update it
+	config.Replace(cfg)
+
 	// Initialize distributed tracing if enabled
 	ctx := context.Background()
 	if cfg.Observability.Tracing.Enabled {

src/semantic-router/pkg/extproc/server.go

@@ -18,6 +18,7 @@ import (
 	"google.golang.org/grpc"
 	"google.golang.org/grpc/credentials"
 
+	"github.com/vllm-project/semantic-router/src/semantic-router/pkg/config"
 	"github.com/vllm-project/semantic-router/src/semantic-router/pkg/observability/logging"
 	tlsutil "github.com/vllm-project/semantic-router/src/semantic-router/pkg/utils/tls"
 )
@@ -158,6 +159,15 @@ func (rs *RouterService) Process(stream ext_proc.ExternalProcessor_ProcessServer
 
 // watchConfigAndReload watches the config file and reloads router on changes.
 func (s *Server) watchConfigAndReload(ctx context.Context) {
+	// Check if we're using Kubernetes config source
+	// If so, skip file watching as config is managed by the Kubernetes controller
+	cfg := config.Get()
+	if cfg != nil && cfg.ConfigSource == config.ConfigSourceKubernetes {
+		logging.Infof("ConfigSource is kubernetes, skipping file watcher")
+		<-ctx.Done() // Just wait for context cancellation
+		return
+	}
+
 	watcher, err := fsnotify.NewWatcher()
 	if err != nil {
 		logging.LogEvent("config_watcher_error", map[string]interface{}{