@xiaohhhh1/canvas-agent 0.4.66 → 0.4.67

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

@@ -171,11 +171,22 @@ export function localVideoAnalysisPrompt(source, durationMs, transcript, attachm
 }
 export function assertJointVisualAndSpokenEvidence(analysis) {
     const segments = Array.isArray(analysis.segments) ? analysis.segments : [];
-    if (segments.length < 2)
-        throw new Error("本机 Codex 返回的镜头时间轴不足 2 段，未计为已理解");
-    if (segments.some((segment) => !String(segment.visual || "").trim())) {
-        throw new Error("视频分析存在缺少画面证据的镜头段，未计为已理解");
-    }
+    const evidencedSegments = segments.filter((segment) => String(segment.visual || "").trim());
+    if (evidencedSegments.length < 2)
+        throw new Error("本机 Codex 返回的有效画面时间轴不足 2 段，未计为已理解");
+    if (evidencedSegments.length === segments.length)
+        return;
+    // Never invent a visual description for a speech-only/null segment. Keep the
+    // independently evidenced beats and record the omission so the server can
+    // build an auditable template without throwing away the whole Omni result.
+    analysis.segments = evidencedSegments;
+    const warnings = Array.isArray(analysis.analysisWarnings)
+        ? analysis.analysisWarnings.filter((item) => typeof item === "string")
+        : [];
+    analysis.analysisWarnings = [
+        ...warnings,
+        `已忽略 ${segments.length - evidencedSegments.length} 个缺少独立画面描述的时间段；未臆造画面事实`,
+    ];
 }
 export function timedTranscriptFromVtt(value) {
     const rows = String(value || "").replace(/^\uFEFF/, "").split(/\r?\n/);

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@xiaohhhh1/canvas-agent",
-  "version": "0.4.66",
+  "version": "0.4.67",
   "type": "module",
   "main": "./dist/index.js",
   "types": "./dist/index.d.ts",