fix(canvas): 短剧分镜 - 支持多镜头并行生成并正确引用角色图声资产

web/src/services/api/generation-task.ts

@@ -124,6 +124,20 @@ export async function runBackendGenerationTask(
     return createAndWaitGenerationTask({ projectId, mode, prompt, config, referenceImages, referenceVideos, referenceAudios, textHistory, signal, metadata, onTaskUpdate }, prepared, dependencies);
 }
 
+// 分镜等后台生产流程只需要可靠提交任务；任务状态与产物由项目工作区轮询和
+// 后端自动回填负责，不能让页面 mutation 一直等待供应商完成。
+export async function submitBackendGenerationTask(
+    options: BackendGenerationTaskOptions,
+    dependencies: GenerationTaskDependencies = defaultDependencies,
+): Promise<GenerationTask> {
+    throwIfAborted(options.signal);
+    assertClientPromptLimit(options.mode, options.prompt, options.config, options.metadata);
+    if (usesLocalDreamina(options.config)) throw new Error("本机即梦任务暂不支持后台提交");
+    const prepared = await prepareGenerationReferences(options);
+    throwIfAborted(options.signal);
+    return createBackendGenerationTask(options, prepared, dependencies);
+}
+
 export async function runBackendToolGenerationTask(options: {
     prompt: string;
     config: AiConfig;
@@ -401,7 +415,14 @@ async function prepareGenerationReferences({
 }
 
 async function createAndWaitGenerationTask(options: BackendGenerationTaskOptions, prepared: PreparedGenerationReferences, dependencies: GenerationTaskDependencies) {
-    const { projectId, mode, prompt, config, signal, metadata, onTaskUpdate, onTextDelta } = options;
+    const task = await createBackendGenerationTask(options, prepared, dependencies);
+    const { signal, onTaskUpdate, onTextDelta } = options;
+    const completed = await dependencies.waitTask(task.id, { signal, initialTask: task, onTaskUpdate, onTextDelta });
+    return parseBackendGenerationResult(completed);
+}
+
+async function createBackendGenerationTask(options: BackendGenerationTaskOptions, prepared: PreparedGenerationReferences, dependencies: GenerationTaskDependencies) {
+    const { projectId, mode, prompt, config, metadata, onTaskUpdate } = options;
     const videoOperation = generationOperation(options);
     const workflow = resolveGenerationWorkflowExecution(config, mode);
     const logicalModelId = workflow ? "" : logicalModelIDForConfig(config);
@@ -428,8 +449,7 @@ async function createAndWaitGenerationTask(options: BackendGenerationTaskOptions
         },
     });
     onTaskUpdate?.(task);
-    const completed = await dependencies.waitTask(task.id, { signal, initialTask: task, onTaskUpdate, onTextDelta });
-    return parseBackendGenerationResult(completed);
+    return task;
 }
 
 async function prepareBackendMediaReference(media: ReferenceVideo | ReferenceAudio) {

web/src/services/api/video-contracts.ts

@@ -38,7 +38,12 @@ export type SeedanceTask = {
 };
 
 export type ApiEnvelope<T> = T | { code?: number; data?: T | null; msg?: string };
-export type RequestOptions = { signal?: AbortSignal };
+export type RequestOptions = {
+    signal?: AbortSignal;
+    videoEditOperation?: string;
+    videoStartFrameNodeId?: string;
+    videoEndFrameNodeId?: string;
+};
 
 export type VideoGenerationResult = { blob?: Blob; url?: string; mimeType?: string };
 export type VideoGenerationTask = { id: string; provider: "openai" | "agnes" | "seedance" | "video-generations" | "gemini-veo" | "novita" | "minimax"; model: string };

web/src/services/api/video-provider-minimax.ts

@@ -15,14 +15,16 @@ export async function createMiniMaxVideoTask(deps: VideoProviderDeps, config: Re
     const videoUrls = await Promise.all(videoReferences.slice(0, 3).map((video) => resolveMiniMaxMediaUrl(video, "参考视频")));
     const audioUrls = await Promise.all(audioReferences.slice(0, 3).map((audio) => resolveMiniMaxMediaUrl(audio, "参考音频")));
     const content: Array<Record<string, unknown>> = [{ type: "text", text: prompt.trim() }];
+    const explicitFrameMode = options?.videoEditOperation !== "reference_to_video" && Boolean(options?.videoStartFrameNodeId || options?.videoEndFrameNodeId);
     imageUrls.forEach((url, index) => {
-        // 首尾帧只允许成对出现；三张及以上图片统一作为多模态参考图，避免提交非法组合。
-        const role = videoUrls.length || audioUrls.length || imageUrls.length > 2 ? "reference_image" : index === 0 ? "first_frame" : "last_frame";
+        const reference = references[index];
+        const role = miniMaxImageRole(reference?.id, options);
+        if (explicitFrameMode && role === "reference_image") throw new Error("MiniMax 首尾帧模式不能混合未标记的参考图");
         content.push({ type: "image_url", image_url: { url }, role });
     });
     videoUrls.forEach((url) => content.push({ type: "video_url", video_url: { url }, role: "reference_video" }));
     audioUrls.forEach((url) => content.push({ type: "audio_url", audio_url: { url }, role: "reference_audio" }));
-    const frameMode = imageUrls.length > 0 && imageUrls.length <= 2 && videoUrls.length === 0 && audioUrls.length === 0;
+    const frameMode = explicitFrameMode;
     const payload = {
         model: modelOptionName(model),
         content,
@@ -41,6 +43,13 @@ export async function createMiniMaxVideoTask(deps: VideoProviderDeps, config: Re
     }
 }
 
+function miniMaxImageRole(imageId: string | undefined, options?: RequestOptions) {
+    if (options?.videoEditOperation === "reference_to_video") return "reference_image";
+    if (imageId && imageId === options?.videoStartFrameNodeId) return "first_frame";
+    if (imageId && imageId === options?.videoEndFrameNodeId) return "last_frame";
+    return "reference_image";
+}
+
 export async function pollMiniMaxVideoTask(deps: VideoProviderDeps, config: ResolvedAiConfig, task: VideoGenerationTask, options?: RequestOptions): Promise<VideoGenerationTaskState> {
     try {
         const response = await deps.transport.get<{ task?: MiniMaxVideoTask }>(miniMaxVideoUrl(config, `/query/video_generation/${encodeURIComponent(task.id)}`), options);