ui : cache the local backend's model list

Assisted-by: pi:llama.cpp/DeepSeek-V4.1-Flash

tools/ui/src/lib/services/backends.service.ts

@@ -6,8 +6,10 @@
  * consumed by the backends settings UI and the per-backend model cache.
  */
 
+import { API_MODELS, LOCAL_BACKEND_ID } from '$lib/constants';
 import type { Backend, ModelOption } from '$lib/types';
 import { isAbortError } from '$lib/utils/abort';
+import { apiUrl } from '$lib/utils/api-base';
 import { getAuthHeadersForBackend } from '$lib/utils/api-headers';
 import { backendModelsUrl } from '$lib/utils/backend';
 
@@ -35,12 +37,17 @@ export class BackendsService {
 	 * @param signal - Optional abort signal for a cancelled request.
 	 */
 	static async listModels(backend: Backend, signal?: AbortSignal): Promise<BackendModelsResult> {
-		if (!backend.baseUrl.trim()) {
+		// the local backend has no base URL; its models endpoint is base relative
+		const url = backend.baseUrl.trim()
+			? backendModelsUrl(backend)
+			: apiUrl(API_MODELS.LIST, LOCAL_BACKEND_ID);
+
+		if (!backend.baseUrl.trim() && backend.id !== LOCAL_BACKEND_ID) {
 			return { error: 'Backend URL is required', models: [], ok: false, status: null };
 		}
 
 		try {
-			const response = await fetch(backendModelsUrl(backend), {
+			const response = await fetch(url, {
 				headers: getAuthHeadersForBackend(backend),
 				signal
 			});

tools/ui/src/lib/stores/backendsModels.svelte.ts

@@ -1,10 +1,10 @@
 /**
  * backendsModelsStore - Per-backend model catalog cache.
  *
- * The models selector prefetches every enabled backend's model list the first
- * time it opens, so switching backends is instant and the switcher can show
- * load state. The active backend's list still lives in modelsStore, which owns
- * selection and chat wiring; this cache is the prefetch layer for the others.
+ * Prefetches every enabled backend's model list, so switching backends is
+ * instant and the switcher can show load state. The active backend's list
+ * still lives in modelsStore, which owns selection and chat wiring; this
+ * cache is the prefetch layer the switches start from.
  */
 
 import { BackendsService } from '$lib/services/backends.service';
@@ -28,13 +28,12 @@ class BackendsModelsStore {
 	}
 
 	/**
-	 * Load a backend's models once. Local backends are skipped: their list is
-	 * owned by modelsStore.
+	 * Load a backend's models once.
 	 */
 	async ensureLoaded(backendId: string): Promise<void> {
 		const backend = backendsStore.enabled.find((candidate) => candidate.id === backendId);
 
-		if (!backend || !backend.baseUrl.trim()) return;
+		if (!backend) return;
 
 		const state = this.states[backendId];
 

tools/ui/src/lib/stores/models/index.svelte.ts

@@ -410,17 +410,28 @@ class ModelsStore implements ModelPropsHost, ModelStatusHost {
 		this.clearSelection();
 		this.routerModels = [];
 		this.error = null;
-		serverStore.clear();
+
+		const backend = backendsStore.active;
+
+		// server props describe the local server; drop them only when the next
+		// backend is not the one they describe, refresh in place otherwise
+		if (backend.protocol !== 'llama.cpp') {
+			serverStore.clear();
+		}
 
 		// prefer the prefetched list so switching does not refetch
-		const backend = backendsStore.active;
-		const cached = backend.baseUrl.trim() ? backendsModelsStore.get(backend.id) : null;
+		const cached = backendsModelsStore.get(backend.id);
 
-		if (cached?.loaded) {
+		if (cached.loaded) {
 			this.activeModels = cached.models;
 			this.loading = false;
 
-			await serverStore.fetch();
+			await serverStore.fetch({ background: true });
+
+			// the cache carries names only; reload the router load status in place
+			if (serverStore.isRouterMode) {
+				await this.fetchRouterModels();
+			}
 
 			if (this.activeModels.length > 0) {
 				await this.ensureFirstModelSelected();