feat(digital-human): 前端加脚本模型选择器(llm_model)

Co-Authored-By: Claude <noreply@anthropic.com>

src/web/pages/DigitalHuman.tsx

@@ -5,6 +5,7 @@ import { getSessionToken, type SessionAudience } from '../authSession.js';
 type WorkflowItem = { name: string; display_name?: string; source?: string; key?: string };
 type TemplateItem = { name: string; display_name?: string; orientation?: string; size?: string; width?: number; height?: number; key?: string };
 type BgmItem = { name: string; path: string; source?: string };
+type MarketplaceModel = { name: string };
 type GenerationMode = 'ai' | 'fixed';
 
 type TaskStatus = 'pending' | 'running' | 'success' | 'failed' | string;
@@ -33,6 +34,7 @@ type DigitalHumanDraft = {
   previewTtsText?: string;
   imageWorkflow?: string;
   apiVideoWorkflow?: string;
+  llmModel?: string;
   imageWidth?: number;
   imageHeight?: number;
   promptPrefix?: string;
@@ -68,6 +70,13 @@ function basename(value: string): string {
   return value.split('/').pop() || value;
 }
 
+const NON_TEXT_MODEL_PATTERN =
+  /(image|video|dall|flux|\bsd\b|sdxl|stable-?diffusion|wan|seedance|kling|banana|tts|whisper|audio|speech|voice|embedding|rerank|moderation|midjourney|\bmj\b|cogview|kolors|hunyuan-?video|vidu|pika|luma|sora|recraft|ideogram)/i;
+
+function isTextModelName(name: string): boolean {
+  return Boolean(name) && !NON_TEXT_MODEL_PATTERN.test(name);
+}
+
 function readDigitalHumanDraft(): DigitalHumanDraft {
   if (typeof window === 'undefined') return {};
   try {
@@ -167,17 +176,32 @@ function firstString(...values: unknown[]): string | undefined {
 
 function firstOutputUrl(value: any): string | undefined {
   if (!value) return undefined;
-  const direct = firstString(value.video_url, value.videoUrl, value.url, value.file_url, value.fileUrl);
-  if (direct) return direct;
   if (Array.isArray(value)) {
+    const archived = value
+      .map((item) => firstOutputUrl(item))
+      .find((url) => url && (url.startsWith('/api/') || !/^https?:/i.test(url)));
+    if (archived) return archived;
     for (const item of value) {
       const found = firstOutputUrl(item);
       if (found) return found;
     }
+    return undefined;
   }
+  const direct = firstString(
+    value.video_url,
+    value.videoUrl,
+    value.url,
+    value.file_url,
+    value.fileUrl,
+    value.remote_url,
+    value.remoteUrl,
+    value.download_url,
+    value.downloadUrl,
+  );
+  if (direct) return direct;
   if (typeof value === 'object') {
-    return firstOutputUrl(value.outputs)
+    return firstOutputUrl(value.assets)
-      || firstOutputUrl(value.assets)
+      || firstOutputUrl(value.outputs)
       || firstOutputUrl(value.result)
       || firstOutputUrl(value.task)
       || firstOutputUrl(value.data);
@@ -187,8 +211,16 @@ function firstOutputUrl(value: any): string | undefined {
 
 function resolvePixelleMediaUrl(path?: string): string {
   if (!path) return '';
-  if (/^(https?:|blob:|data:)/i.test(path) || path.startsWith('/api/')) return path;
+  const normalized = String(path).trim();
-  return api.pixelle.fileUrl(path);
+  if (/^(blob:|data:)/i.test(normalized) || normalized.startsWith('/api/')) return normalized;
+  try {
+    const url = new URL(normalized, typeof window !== 'undefined' ? window.location.origin : 'http://127.0.0.1');
+    if ((url.hostname === '127.0.0.1' || url.hostname === 'localhost') && url.pathname.startsWith('/api/files/')) {
+      return api.pixelle.fileUrl(url.pathname.replace(/^\/api\/files\/?/, ''));
+    }
+  } catch {}
+  if (/^https?:/i.test(normalized)) return normalized;
+  return api.pixelle.fileUrl(normalized);
 }
 
 function currentAudience(): SessionAudience {
@@ -370,6 +402,8 @@ export default function DigitalHuman() {
   const [previewTtsText, setPreviewTtsText] = useState(draftRef.current.previewTtsText || '你好，欢迎使用 BoosAPI 数字人。');
   const [imageWorkflow, setImageWorkflow] = useState(draftRef.current.imageWorkflow || '');
   const [apiVideoWorkflow, setApiVideoWorkflow] = useState(draftRef.current.apiVideoWorkflow || '');
+  const [llmModel, setLlmModel] = useState(draftRef.current.llmModel || '');
+  const [llmModels, setLlmModels] = useState<MarketplaceModel[]>([]);
   const [imageWidth, setImageWidth] = useState(() => normalizeNumber(draftRef.current.imageWidth, 1024, 64));
   const [imageHeight, setImageHeight] = useState(() => normalizeNumber(draftRef.current.imageHeight, 1024, 64));
   const [promptPrefix, setPromptPrefix] = useState(draftRef.current.promptPrefix || DEFAULT_PROMPT_PREFIX);
@@ -414,6 +448,16 @@ export default function DigitalHuman() {
       setMediaWorkflows(Array.isArray(media?.workflows) ? media.workflows : []);
       setTemplates(Array.isArray(tpl?.templates) ? tpl.templates : []);
       setBgmList(Array.isArray(bgm?.bgm_files) ? bgm.bgm_files : []);
+      void (async () => {
+        try {
+          const marketplace = (await api.getModelsMarketplace()) as { models?: MarketplaceModel[] };
+          const textModels = (Array.isArray(marketplace?.models) ? marketplace.models : [])
+            .filter((m) => m && typeof m.name === 'string' && isTextModelName(m.name));
+          setLlmModels(textModels);
+        } catch {
+          setLlmModels([]);
+        }
+      })();
     } catch (error) {
       setResourceError(getErrorMessage(error, '加载 Pixelle 资源失败'));
     } finally {
@@ -463,6 +507,7 @@ export default function DigitalHuman() {
       previewTtsText,
       imageWorkflow,
       apiVideoWorkflow,
+      llmModel,
       imageWidth,
       imageHeight,
       promptPrefix,
@@ -482,6 +527,7 @@ export default function DigitalHuman() {
     previewTtsText,
     imageWorkflow,
     apiVideoWorkflow,
+    llmModel,
     imageWidth,
     imageHeight,
     promptPrefix,
@@ -579,6 +625,7 @@ export default function DigitalHuman() {
         n_scenes: nScenes,
         media_workflow: imageWorkflow || undefined,
         api_video_workflow: apiVideoWorkflow || undefined,
+        llm_model: llmModel || undefined,
         tts_workflow: ttsWorkflow || undefined,
         frame_template: frameTemplate,
         ratio: selectedTemplateSize.ratio,
@@ -614,6 +661,7 @@ export default function DigitalHuman() {
     title,
     nScenes,
     imageWorkflow,
+    llmModel,
     ttsWorkflow,
     promptPrefix,
     videoFps,
@@ -1026,6 +1074,18 @@ export default function DigitalHuman() {
                 ))}
             </select>
           </label>
+          <label className="digital-human-field">
+            <span>脚本模型（可选）</span>
+            <select value={llmModel} onChange={(event) => setLlmModel(event.target.value)}>
+              <option value="">默认（deepseek-v4-flash）</option>
+              {llmModel && !llmModels.some((m) => m.name === llmModel) && (
+                <option value={llmModel}>{llmModel}</option>
+              )}
+              {llmModels.map((m) => (
+                <option key={m.name} value={m.name}>{m.name}</option>
+              ))}
+            </select>
+          </label>
           <div className="digital-human-form-grid">
             <label className="digital-human-field">
               <span>宽度</span>