@xiaohhhh1/canvas-agent 0.4.60 → 0.4.62

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

@@ -8,6 +8,7 @@ import { videoEvidenceTimestamps } from "../integrations/fastmoss.js";
 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";
+const VIDEO_INTELLIGENCE_SCHEMA_VERSION = "commerce-video-intelligence-v5";
 let localAsrPipeline;
 export function shouldUploadVideoArchive(input, publicUrl) {
     return input.kind !== "uploaded-mp4" || input.url !== publicUrl;
@@ -153,11 +154,11 @@ export function localVideoAnalysisPrompt(source, durationMs, transcript, attachm
             durationMs,
         })}\n\n` +
         `带时间码的口播字幕：\n${transcript}\n\n` +
-        `只返回一个 JSON 对象，不要 Markdown。字段必须是：schemaVersion="commerce-video-intelligence-v3"；durationMs；language；summary；` +
+        `只返回一个 JSON 对象，不要 Markdown。字段必须是：schemaVersion="${VIDEO_INTELLIGENCE_SCHEMA_VERSION}"；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）；` +
-        `creativeMutationProfile{coreVisualPremise,mutationAxes[至少2项],culturalSignals[{market,signal,narrativeRole,aestheticRole,transferConditions,avoidStereotype,evidence[{startMs,endMs,observation}]}],firstFrameComposition,spectacleMechanism,productIntegration,transferableRule,culturalLimitations,doNotCopy,evidence[{startMs,endMs,observation}],confidence}；` +
+        `creativeMutationProfile{coreVisualPremise,mutationAxes[至少2项],culturalSignals[{market,signal,narrativeRole,aestheticRole,transferConditions,avoidStereotype,evidence[{startMs,endMs,observation}]}],firstFrameComposition,spectacleMechanism,productIntegration,captureProfile{subjectDominance,depthAndDensity,colorLightContrast,materialTexture,cameraEnergy,captureImperfections},transferableRule,culturalLimitations,doNotCopy,evidence[{startMs,endMs,observation}],confidence}；` +
         `mechanisms[{type,label,mechanism,whyItMayWork,evidence[{startMs,endMs,observation}],confidence,replicationRisk}]（type 仅 hook/conflict/proof/pacing/trust/product-reveal/offer-framing/cta/audio/visual-grammar/environment/visual-mutation/culture-code/spectacle）；` +
         `viralHypotheses[{hypothesis,supportingMetrics,supportingEvidence[{startMs,endMs,observation}],confounders,confidence}]；` +
         `nonReplicableFactors；complianceRisks；originalityGuidance{preserveMechanisms,mustRewrite,forbiddenCopying}。`;

package/dist/workflow/creative-candidates.d.ts

@@ -1,5 +1,5 @@
 type JsonSchema = Record<string, unknown>;
-export declare const FLOW_C_CREATIVE_CANDIDATE_CONTRACT_VERSION = "flow-c-creative-candidates-v3";
+export declare const FLOW_C_CREATIVE_CANDIDATE_CONTRACT_VERSION = "flow-c-creative-candidates-v4";
 export declare function flowCCreativeCandidateOutputSchema(count: number): JsonSchema;
 export declare function parseFlowCCreativeCandidateOutput(value: string, expectedOrdinals: number[]): Record<string, unknown>[];
 export {};

package/dist/workflow/creative-candidates.js

@@ -1,11 +1,13 @@
 import { assertStrictResponseSchema } from './script-output.js';
-export const FLOW_C_CREATIVE_CANDIDATE_CONTRACT_VERSION = 'flow-c-creative-candidates-v3';
+export const FLOW_C_CREATIVE_CANDIDATE_CONTRACT_VERSION = 'flow-c-creative-candidates-v4';
 const text = { type: 'string', minLength: 1 };
+const nullableText = { anyOf: [text, { type: 'null' }] };
 function object(properties) {
     return { type: 'object', properties, required: Object.keys(properties), additionalProperties: false };
 }
 function candidateSchema() {
     return object({
+        learnedTemplateId: nullableText,
         visualPremise: text,
         mutationAxes: { type: 'array', minItems: 2, maxItems: 4, items: text },
         culturalAnchors: {

package/dist/workflow/manager.d.ts

@@ -21,6 +21,10 @@ type ScriptRecord = {
 };
 type ReferenceStyleCard = {
     id?: unknown;
+    templateKind?: unknown;
+    templateProfile?: unknown;
+    modelFamily?: unknown;
+    sourceDurationSeconds?: unknown;
     visualPremisePattern?: unknown;
     mutationAxes?: unknown;
     culturePattern?: unknown;
@@ -31,6 +35,10 @@ type ReferenceStyleCard = {
     rhythmPattern?: unknown;
     capturePattern?: unknown;
     voiceTone?: unknown;
+    targetDurationPolicy?: unknown;
+    retimingRule?: unknown;
+    segmentRule?: unknown;
+    omniAdaptation?: unknown;
     sourceEvidenceTier?: unknown;
     truthBoundaries?: unknown;
     applicability?: {
@@ -51,6 +59,7 @@ type ProductInput = {
 type SelectedCandidate = {
     ordinal: number;
     productIndex: number;
+    learnedTemplateId?: string | null;
     visualPremise: string;
     mutationAxes: string[];
     culturalAnchors: Array<Record<string, unknown>>;

package/dist/workflow/manager.js

@@ -645,22 +645,24 @@ ${compactJson([...selected.values()], 16_000)}
 ${durationRules}
 写作要求：
 1. visualHook、conflict、escalation/turn、productIntervention、visibleProof、purchaseReason、callbackMotivation 与所有 shots 必须执行选中卡的 visualPremise、firstFrame、spectacleEscalation、productProofAction、purchaseReason、truthBoundary 以及七项视觉执行字段；不得扩大功效、稀释构图或换成普通模板。productProofAction 必须同时写清一个真实动作及其镜头内可见结果；purchaseReason 只解释该结果为何值得目标用户购买，不能变成泛泛 CTA。
+1a. executionBindings 必须逐项指出 opening、proof、purchaseReason 和七项 visualExecution 实际落在哪个 segment/shot；executedAs 必须同时忠实对应 selected candidate 原要求和该 shot 的真实内容。opening 必须包含 segment 1 shot 1。它是可审计索引，不是新增剧情，也不能把未执行的要求写成已执行。
 2. 每条生成一份稳定 voiceProfile；每段只写简短 voiceCue。口播必须是目标市场 ${task.market} 的自然原生语言、偏快但清晰；导演说明统一用简洁制作英文。
 3. 每段永远是独立 0–10 秒，含 1–8 个按剧情需要决定的 shots、准确 voiceover、onScreenText、evidence、soundBgm、emotionalNote 与 endingState；模型不要输出 continuity 或 continuityMode。
 4. shot.visual 要直接写出主体位置、景别、光线、材质、动作峰值和当地自然环境，并保留真实皮肤、细发、布料/商品微纹理和实景混合光；禁止塑料蜡感、静态举产品、品牌广告片、电视购物或伪造工厂/销量/来源。
 5. 20/30 秒的后一段剧情承接前段 endingState，但每段仍只使用自己的局部 0–10 秒执行内容。
-6. creativePlan 只写 format、visualHook、conflict、escalation/turn、productIntervention、visibleProof、purchaseReason、callbackMotivation 与 differentiationKey。callbackMotivation 必须自然回扣开场与已经展示的 purchaseReason；禁止编造折扣、库存、销量、价格或未提供的量化功效。Agent 会把选中卡的构图、材质、光线、文化、真实性边界和 fingerprint 确定性并回，中心会程序化计算质量门；模型不要重复这些字段。
+6. creativePlan 只写 format、visualHook、conflict、escalation/turn、productIntervention、visibleProof、purchaseReason、callbackMotivation 与 differentiationKey。callbackMotivation 必须自然回扣开场与已经展示的 purchaseReason；禁止编造折扣、库存、销量、价格或未提供的量化功效。Agent 会把选中卡的构图、材质、光线、文化、真实性边界和 fingerprint 确定性并回，中心会程序化计算质量门；模型不要重复这些字段。若 selected candidate 的 selectionMode=user-framework，用户原框架的开头、事件顺序、核心剧情和结尾是最高权威；只补真实证明、局部时序和执行细节，不得改成另一个创意。
 固定使用 GPT-5.6 Terra 高推理。storyboardLayoutVersion=${layoutVersion} 仅由中心管理。只返回当前结构化 jobs，不调用工具、不创建任何图片或视频任务。`;
 }
 export function creativeCandidatePrompt(id, task, ordinals) {
     const products = relevantProductInputs(task, ordinals);
-    const styleCards = relevantStyleCards(task.reference_style_cards || [], products).map(compactReferenceStyleCard).filter(Boolean).slice(0, 6);
+    const styleCards = relevantStyleCards(task.reference_style_cards || [], products, task.market).map(compactReferenceStyleCard).filter(Boolean).slice(0, 6);
     const ledger = (task.creative_fingerprint_ledger || []).slice(-60);
     return `Flow C creative-candidate stage for handoff ${id}. Set contractVersion exactly to ${FLOW_C_CREATIVE_CANDIDATE_CONTRACT_VERSION}. Return exactly one group for each ordinal: ${ordinals.join(", ")}; mapping: ${scriptProductAssignments(task.product_quantities, ordinals)}.
-Market: ${task.market}. Product facts: ${compactJson(products, 12_000)}
+Market: ${task.market}. Target output duration: ${Number(task.duration_seconds || 10)} seconds. Product facts: ${compactJson(products, 12_000)}
 Optional learned style abstractions: ${compactJson(styleCards, 5_000)}
 Already accepted fingerprint ledger: ${compactJson(ledger, 8_000)}
 For each ordinal return exactly 6 short, auditable candidate cards. These are visible structured alternatives, not hidden reasoning. Vary the overall creative skeleton through open dimensions such as relationship/identity, scene mismatch, scale/quantity, group reaction, viewpoint, sensory texture and proof action; do not rotate a fixed template list.
+Set learnedTemplateId to the id of one supplied learned-video-structure card when its product/market fit is credible; otherwise set it to null. A fitting learned template has priority over free ideation. Preserve its hook-to-proof-to-purchase-reason causal structure, but re-time it to the current target duration instead of stretching, truncating or copying the source timeline. The template is model-neutral evidence: never output Seedance wording or source-specific people, dialogue, wardrobe, set or exact shots. The same learnedTemplateId may be used for multiple ordinals when useful, but each adapted candidate must still have truthful product-specific content, native VO intent and a meaningfully changed opening/proof execution. If the product has a non-empty user-authored creativeBrief, this candidate stage is bypassed entirely and no learned template may influence the script.
 Every card must be surprising but physically filmable and product-relevant. productProofAction must state one truthful intended-use action and the observable result that remains visible in the same scene. purchaseReason must state one concrete reason the supported result matters to a real target buyer; it is not a generic CTA and must not invent price, discount, scarcity, sales, inventory or unsupported quantified performance. productProofAction, purchaseReason and truthBoundary must follow supplied facts. For food, baby, medical/efficacy and personal-hygiene products, reject unhygienic display, fear marketing, dangerous use and unsupported promises.
 Every card must also contain a concrete visual execution card, not mood adjectives: firstFrameFocus names one dominant subject and its approximate 35%-85% frame share; visualContrast defines one readable scale/color/light/relationship contrast; sensoryTexture names truthful category-specific material evidence; motionPeak names the single peak physical instant; compositionLighting specifies foreground/midground/background, practical light direction and one plausible exposure/optical imperfection; localNativeDetail names observable local behavior/environment/aesthetic density with the same evidence discipline as culturalAnchors; antiFlatness names the exact generic catalogue treatment this concept must avoid. The firstFrame must pass a screenshot test: even without dialogue, a still image must show the unusual relationship, product relevance and depth hierarchy.
 Cultural anchors are optional (0–3). source must be learned-style-card, product-market-evidence, or conservative-everyday-detail. A learned-style-card anchor must cite the supplied card id in evidence. A product-market-evidence anchor must cite a concrete supplied product/market fact; if none exists, use conservative-everyday-detail and explicitly say it is ordinary context, not a custom claim. Never use flags, national costumes, royalty/public figures, religion, stereotypes or cultural mockery as shortcut hooks. Locality belongs in consumption context, natural interaction, language rhythm and aesthetic density.
@@ -669,7 +671,7 @@ creativeFingerprint must use five concise semantic labels: characterRelation, sc
 function compactReferenceStyleCard(value) {
     if (!value || typeof value !== "object")
         return "";
-    const fields = ["id", "visualPremisePattern", "mutationAxes", "culturePattern", "spectaclePattern", "productIntegrationPattern", "openingPattern", "proofPattern", "rhythmPattern", "capturePattern", "voiceTone", "sourceEvidenceTier"];
+    const fields = ["id", "templateKind", "templateProfile", "modelFamily", "sourceDurationSeconds", "visualPremisePattern", "mutationAxes", "culturePattern", "spectaclePattern", "productIntegrationPattern", "openingPattern", "proofPattern", "rhythmPattern", "capturePattern", "voiceTone", "targetDurationPolicy", "retimingRule", "segmentRule", "omniAdaptation", "sourceEvidenceTier"];
     const compact = Object.fromEntries(fields.flatMap((field) => {
         const text = String(value[field] || "").trim().replace(/\s+/g, " ").slice(0, 400);
         return text ? [[field, text]] : [];
@@ -680,13 +682,15 @@ function compactReferenceStyleCard(value) {
         Object.assign(compact, { applicability: { categories, markets } });
     return Object.keys(compact).length ? compact : null;
 }
-function relevantStyleCards(cards, products) {
+function relevantStyleCards(cards, products, market = "") {
     const categories = new Set(products.map((product) => String(product.category || "").trim().toLowerCase()).filter(Boolean));
-    if (!categories.size)
-        return cards.slice(0, 3);
     const matched = cards.filter((card) => {
         const allowed = Array.isArray(card.applicability?.categories) ? card.applicability.categories.map((item) => String(item || "").trim().toLowerCase()).filter(Boolean) : [];
-        return !allowed.length || allowed.includes("all") || allowed.includes("global") || allowed.some((item) => categories.has(item));
+        const markets = Array.isArray(card.applicability?.markets) ? card.applicability.markets.map((item) => String(item || "").trim().toLowerCase()).filter(Boolean) : [];
+        const categoryFit = !categories.size || !allowed.length || allowed.includes("all") || allowed.includes("global") || allowed.some((item) => categories.has(item));
+        const targetMarket = String(market || "").trim().toLowerCase();
+        const marketFit = !markets.length || markets.includes("all") || markets.includes("global") || markets.includes(targetMarket);
+        return categoryFit && marketFit;
     });
     return matched.length ? matched : [];
 }
@@ -709,6 +713,7 @@ function selectedCandidatePlan(value) {
         throw new Error("中心缺少选中的创意候选");
     return {
         visualPremise: value.visualPremise,
+        learnedTemplateId: value.learnedTemplateId || null,
         mutationAxes: value.mutationAxes,
         culturalAnchors: value.culturalAnchors,
         spectacleEscalation: value.spectacleEscalation,

package/dist/workflow/script-output.js

@@ -37,6 +37,24 @@ function creativePlanSchema() {
     };
     return object(properties);
 }
+function executionBindingSchema() {
+    return object({
+        segmentIndex: { type: "integer", minimum: 1, maximum: 3 },
+        shotIndexes: { type: "array", minItems: 1, maxItems: 8, items: { type: "integer", minimum: 1, maximum: 8 } },
+        executedAs: text,
+    });
+}
+function executionBindingsSchema() {
+    return object({
+        opening: executionBindingSchema(),
+        proof: executionBindingSchema(),
+        purchaseReason: executionBindingSchema(),
+        visualExecution: object(Object.fromEntries([
+            "firstFrameFocus", "visualContrast", "sensoryTexture", "motionPeak",
+            "compositionLighting", "localNativeDetail", "antiFlatness",
+        ].map((field) => [field, executionBindingSchema()]))),
+    });
+}
 function segmentSchema() {
     return object({
         voiceCue: text,
@@ -65,6 +83,7 @@ export function flowCScriptOutputSchema(durationSeconds, count) {
         productIndex: { type: "integer", minimum: 0 },
         sellingFormId: text,
         creativePlan: creativePlanSchema(),
+        executionBindings: executionBindingsSchema(),
         voiceProfile: voiceProfileSchema(),
         openingState: openingStateSchema(),
     };

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@xiaohhhh1/canvas-agent",
-  "version": "0.4.60",
+  "version": "0.4.62",
   "type": "module",
   "main": "./dist/index.js",
   "types": "./dist/index.d.ts",