@xiaohhhh1/canvas-agent 0.4.39 → 0.4.40

package/dist/video-intelligence/local-analysis.d.ts

@@ -2,6 +2,8 @@ import type { AgentAttachment } from "../agent/types.js";
 import { type FastMossIntegration } from "../integrations/fastmoss.js";
 export type LocalVideoLearningSource = {
     id?: string;
+    source?: string;
+    platform?: string;
     sourceUrl?: string;
     source_url?: string;
     platformVideoId?: string;
@@ -29,6 +31,10 @@ type VideoArchiveUpload = {
     headers?: Record<string, string>;
     maxBytes?: number;
 };
+export declare function shouldUploadVideoArchive(input: {
+    kind: "uploaded-mp4" | "tiktok";
+    url: string;
+}, publicUrl: string): boolean;
 export declare class LocalVideoIntelligence {
     private readonly fastmoss;
     constructor(fastmoss: FastMossIntegration);
@@ -56,6 +62,10 @@ export declare class LocalVideoIntelligence {
         };
     }>;
 }
+export declare function verifiedVideoLearningSource(source: LocalVideoLearningSource): {
+    kind: "tiktok" | "uploaded-mp4";
+    url: string;
+};
 export declare function localVideoAnalysisPrompt(source: LocalVideoLearningSource, durationMs: number, transcript: string, attachments: AgentAttachment[]): string;
 export declare function assertJointVisualAndSpokenEvidence(analysis: Record<string, unknown>): void;
 export declare function timedTranscriptFromVtt(value: string): string;

package/dist/video-intelligence/local-analysis.js

@@ -9,28 +9,35 @@ const ANALYSIS_TIMEOUT_MS = 10 * 60_000;
 const TRANSCRIPT_LANGUAGES = "en.*,es.*,zh.*,pt.*,fr.*,de.*,vi.*,th.*,id.*,ms.*,ja.*,ko.*";
 const LOCAL_ASR_MODEL = "onnx-community/whisper-base";
 let localAsrPipeline;
+export function shouldUploadVideoArchive(input, publicUrl) {
+    return input.kind !== "uploaded-mp4" || input.url !== publicUrl;
+}
 export class LocalVideoIntelligence {
     fastmoss;
     constructor(fastmoss) {
         this.fastmoss = fastmoss;
     }
     async archive(source, upload) {
-        const sourceUrl = verifiedTikTokSourceUrl(source);
+        const input = verifiedVideoLearningSource(source);
         const uploadUrl = verifiedHttpsUrl(upload?.uploadUrl, "站内上传地址无效");
         const publicUrl = verifiedHttpsUrl(upload?.publicUrl, "站内视频地址无效");
         const maxBytes = Math.min(500 * 1024 * 1024, Math.max(1, Number(upload?.maxBytes || 100 * 1024 * 1024)));
         const workDir = await mkdtemp(path.join(os.tmpdir(), "canvas-video-archive-"));
         try {
-            const downloaded = await downloadVideoArchive(sourceUrl, workDir, maxBytes);
+            const downloaded = input.kind === "uploaded-mp4"
+                ? await downloadUploadedMp4(input.url, workDir, maxBytes)
+                : await downloadVideoArchive(input.url, workDir, maxBytes);
             const bytes = await readFile(downloaded.file);
-            const response = await fetch(uploadUrl, {
-                method: "PUT",
-                headers: { "Content-Type": "video/mp4" },
-                body: bytes,
-                signal: AbortSignal.timeout(5 * 60_000),
-            });
-            if (!response.ok)
-                throw new Error(`站内视频上传失败（HTTP ${response.status}）`);
+            if (shouldUploadVideoArchive(input, publicUrl)) {
+                const response = await fetch(uploadUrl, {
+                    method: "PUT",
+                    headers: { "Content-Type": "video/mp4" },
+                    body: bytes,
+                    signal: AbortSignal.timeout(5 * 60_000),
+                });
+                if (!response.ok)
+                    throw new Error(`站内视频上传失败（HTTP ${response.status}）`);
+            }
             return {
                 archive: {
                     publicUrl,
@@ -50,11 +57,10 @@ export class LocalVideoIntelligence {
         }
     }
     async analyze(source) {
-        const sourceUrl = verifiedTikTokSourceUrl(source);
         const archivedVideoUrl = String(source.archivedVideoUrl || source.archived_video_url || "").trim();
         const archivedTranscript = String(source.archivedTranscript || source.archived_transcript || "").trim();
         const hasArchivedMedia = Boolean(archivedVideoUrl);
-        const mediaUrl = archivedVideoUrl ? verifiedHttpsUrl(archivedVideoUrl, "站内视频副本地址无效") : sourceUrl;
+        const mediaUrl = archivedVideoUrl ? verifiedHttpsUrl(archivedVideoUrl, "站内视频副本地址无效") : verifiedTikTokSourceUrl(source);
         const platformVideoId = String(source.platformVideoId || source.platform_video_id || "").trim();
         const workDir = await mkdtemp(path.join(os.tmpdir(), "canvas-video-intelligence-"));
         try {
@@ -68,7 +74,7 @@ export class LocalVideoIntelligence {
                 transcriptSource = transcript ? `local-asr:${LOCAL_ASR_MODEL}` : "local-asr:no-speech-detected";
             }
             else if (!transcript) {
-                transcript = await extractTimedTranscript(sourceUrl, workDir).catch(() => "");
+                transcript = await extractTimedTranscript(mediaUrl, workDir).catch(() => "");
                 transcriptSource = transcript ? "TikTok native/automatic timed captions via local yt-dlp" : "caption-unavailable";
             }
             if (!visualEvidence.frames.length)
@@ -100,6 +106,16 @@ function verifiedTikTokSourceUrl(source) {
     }
     return sourceUrl;
 }
+export function verifiedVideoLearningSource(source) {
+    if (String(source.source || "").trim() === "manual-upload") {
+        const sourceUrl = verifiedHttpsUrl(String(source.sourceUrl || source.source_url || "").trim(), "上传视频地址无效");
+        const pathname = new URL(sourceUrl).pathname;
+        if (!/\.mp4$/i.test(pathname))
+            throw new Error("手动上传学习只接受 MP4 文件");
+        return { kind: "uploaded-mp4", url: sourceUrl };
+    }
+    return { kind: "tiktok", url: verifiedTikTokSourceUrl(source) };
+}
 function verifiedHttpsUrl(value, message) {
     try {
         const url = new URL(value);
@@ -119,8 +135,9 @@ export function localVideoAnalysisPrompt(source, durationMs, transcript, attachm
         `1. 每张图片是实际视频在指定毫秒的画面，文件名和时间映射如下。不能把封面、榜单或网页文字当视频内容。\n${frameTimeline}\n` +
         `2. 下面口播证据来自原生字幕、自动字幕或本机音轨语音识别。只概述，不输出连续逐字稿；若为空，表示没有识别到人声，不代表视频没有音轨，也不得凭画面猜口播。\n` +
         `3. segments 每段必须写 visual；若该段有人声，spokenText 必须概述口播；可见字幕写 onScreenText。画面、口播或字幕缺失就写 null。没有听觉音轨输入，audio 必须写 null，不得猜音乐、语气或音效。\n` +
-        `4. 0-3 秒至少按 500ms 证据判断；后续按镜头变化。所有机制和“为什么可能爆”的假设都必须引用真实 startMs/endMs。榜单指标只是相关性，不是因果。\n` +
-        `5. 目标是学习、融合、进化：只提炼可迁移机制，不复制原文案、人物、视觉资产或连续镜头顺序。\n\n` +
+        `4. 必须先判断整条视频的主带货形式 primary，只能选 factory-demo/comedy/mini-drama/ugc-testimonial/review-demo/tutorial/problem-solution/price-shock/unboxing/expert-explainer/lifestyle/live-cut/comparison/other。secondary 可选其他辅助形式。不能把 Hook 或实验假设冒充带货形式。\n` +
+        `5. 0-3 秒至少按 500ms 证据判断；后续按镜头变化。所有机制、带货形式和“为什么可能爆”的假设都必须引用真实 startMs/endMs。榜单指标只是相关性，不是因果。\n` +
+        `6. 目标是学习、融合、进化：只提炼可迁移机制，不复制原文案、人物、视觉资产或连续镜头顺序。\n\n` +
         `样本：${JSON.stringify({
             sourceUrl: source.sourceUrl || source.source_url,
             platformVideoId: source.platformVideoId || source.platform_video_id,
@@ -134,7 +151,8 @@ export function localVideoAnalysisPrompt(source, durationMs, transcript, attachm
             durationMs,
         })}\n\n` +
         `带时间码的口播字幕：\n${transcript}\n\n` +
-        `只返回一个 JSON 对象，不要 Markdown。字段必须是：schemaVersion="commerce-video-intelligence-v1"；durationMs；language；summary；` +
+        `只返回一个 JSON 对象，不要 Markdown。字段必须是：schemaVersion="commerce-video-intelligence-v2"；durationMs；language；summary；` +
+        `sellingFormat{primary,label,secondary,rationale,evidence[{startMs,endMs,observation}],confidence}；` +
         `hook{startMs,endMs,visual,spokenText,onScreenText,patternInterrupt,openLoop}；productFirstSeenMs；` +
         `segments[{startMs,endMs,role,visual,spokenText,onScreenText,audio,editing,confidence}]（至少2段）；` +
         `mechanisms[{type,label,mechanism,whyItMayWork,evidence[{startMs,endMs,observation}],confidence,replicationRisk}]（type 仅 hook/conflict/proof/pacing/trust/product-reveal/offer-framing/cta/audio）；` +
@@ -292,6 +310,25 @@ async function downloadVideoArchive(sourceUrl, workDir, maxBytes) {
     const transcript = await transcriptFromDirectory(workDir, lastError).catch(() => "");
     return { file: video.file, transcript, media };
 }
+async function downloadUploadedMp4(sourceUrl, workDir, maxBytes) {
+    const response = await fetch(sourceUrl, { signal: AbortSignal.timeout(5 * 60_000) });
+    if (!response.ok)
+        throw new Error(`读取上传视频失败（HTTP ${response.status}）`);
+    const declaredLength = Number(response.headers.get("content-length") || 0);
+    if (declaredLength > maxBytes)
+        throw new Error(`视频超过站内存储上限（${Math.ceil(maxBytes / 1024 / 1024)} MiB）`);
+    const bytes = Buffer.from(await response.arrayBuffer());
+    if (!bytes.length)
+        throw new Error("上传视频为空");
+    if (bytes.length > maxBytes)
+        throw new Error(`视频超过站内存储上限（${Math.ceil(maxBytes / 1024 / 1024)} MiB）`);
+    if (!bytes.subarray(0, 64).includes(Buffer.from("ftyp")))
+        throw new Error("上传文件不是可识别的 MP4");
+    const file = path.join(workDir, "manual-upload.mp4");
+    await writeFile(file, bytes);
+    const media = await inspectVideoFile(file, workDir);
+    return { file, transcript: "", media };
+}
 export async function inspectVideoFile(videoFile, workDir = path.dirname(videoFile)) {
     const probe = await runProcess(process.platform === "win32" ? "ffprobe.exe" : "ffprobe", [
         "-v", "error", "-show_entries", "stream=codec_type,codec_name,width,height,duration:format=duration,format_name", "-of", "json", videoFile,

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@xiaohhhh1/canvas-agent",
-  "version": "0.4.39",
+  "version": "0.4.40",
   "type": "module",
   "main": "./dist/index.js",
   "types": "./dist/index.d.ts",