Add shutdown state checks in HTTP handlers

pkg/instance/proxy.go

@@ -154,16 +154,8 @@ func (p *proxy) build() (*httputil.ReverseProxy, error) {
 	return proxy, nil
 }
 
-// serveHTTP handles HTTP requests with inflight tracking and shutting down state checks
+// serveHTTP handles HTTP requests with inflight tracking
 func (p *proxy) serveHTTP(w http.ResponseWriter, r *http.Request) error {
-	// Check if instance is shutting down
-	status := p.instance.GetStatus()
-	if status == ShuttingDown {
-		w.WriteHeader(http.StatusServiceUnavailable)
-		w.Write([]byte("Instance is shutting down"))
-		return fmt.Errorf("instance is shutting down")
-	}
-
 	// Get the reverse proxy
 	reverseProxy, err := p.get()
 	if err != nil {

pkg/server/handlers_backends.go

@@ -109,6 +109,12 @@ func (h *Handler) LlamaCppProxy() http.HandlerFunc {
 			return
 		}
 
+		// Check if instance is shutting down before autostart logic
+		if inst.GetStatus() == instance.ShuttingDown {
+			writeError(w, http.StatusServiceUnavailable, "instance_shutting_down", "Instance is shutting down")
+			return
+		}
+
 		if !inst.IsRemote() && !inst.IsRunning() {
 			err := h.ensureInstanceRunning(inst)
 			if err != nil {

pkg/server/handlers_openai.go

@@ -4,6 +4,7 @@ import (
 	"bytes"
 	"encoding/json"
 	"io"
+	"llamactl/pkg/instance"
 	"llamactl/pkg/validation"
 	"net/http"
 )
@@ -106,6 +107,12 @@ func (h *Handler) OpenAIProxy() http.HandlerFunc {
 			return
 		}
 
+		// Check if instance is shutting down before autostart logic
+		if inst.GetStatus() == instance.ShuttingDown {
+			writeError(w, http.StatusServiceUnavailable, "instance_shutting_down", "Instance is shutting down")
+			return
+		}
+
 		if !inst.IsRemote() && !inst.IsRunning() {
 			err := h.ensureInstanceRunning(inst)
 			if err != nil {