@gobing-ai/knowledge-kit 0.0.15 → 0.0.17

package/.claude-plugin/marketplace.json

@@ -6,7 +6,7 @@
     },
     "plugins": [
         {
-            "version": "0.0.15",
+            "version": "0.0.17",
             "name": "kk",
             "source": "./plugins/kk",
             "description": "knowledge-kit — a set of toolkit for knowledge management"

package/package.json

@@ -1,6 +1,6 @@
 {
     "name": "@gobing-ai/knowledge-kit",
-    "version": "0.0.15",
+    "version": "0.0.17",
     "description": "An ingest → create → publish content pipeline CLI (Bun).",
     "repository": {
         "type": "git",

package/plugins/generations/daily-article-gen/dist/index.js

@@ -21967,6 +21967,45 @@ function sourceLabelOf(doc2) {
     return "\u6765\u6E90";
   }
 }
+var SCAFFOLD_LABELS = ["\u53EF\u505A\u89D2\u5EA6", "\u4E3A\u4EC0\u4E48\u73B0\u5728\u503C\u5F97\u6CE8\u610F", "\u793E\u533A\u8BA8\u8BBA", "\u80CC\u666F", "\u6F5C\u5728\u5F71\u54CD", "\u5F71\u54CD"];
+var HEDGING_MARKERS = ["\u8BFB\u8005\u5B9C\u53E6\u884C\u6838\u5B9E", "\u8BFB\u8005\u6700\u597D", "\u4ECD\u5F85\u9A8C\u8BC1", "\u5C1A\u5F85\u9A8C\u8BC1", "\u9700\u53E6\u884C\u6838\u5B9E"];
+function scaffoldPrefixLength(text) {
+  const match = text.match(/^(?:#{1,6}\s*)?(?:\*\*)?\s*[\u300C\u300E]?\s*([^\u300D\u300F*\uFF1A:\n]{2,})\s*[\u300D\u300F]?\s*(?:\*\*)?\s*[\uFF1A:]?\s*/);
+  if (match === null || !SCAFFOLD_LABELS.includes((match[1] ?? "").trim()))
+    return 0;
+  return match[0].length;
+}
+function collapseScaffoldLabels(text) {
+  let rest = text;
+  for (;; ) {
+    const next = rest.slice(scaffoldPrefixLength(rest));
+    if (next === rest)
+      return rest;
+    rest = next;
+  }
+}
+function stripHedging(text) {
+  if (!HEDGING_MARKERS.some((marker) => text.includes(marker)))
+    return text;
+  return text.split(/(?<=\u3002)/).filter((sentence) => !HEDGING_MARKERS.some((marker) => sentence.includes(marker))).join("").trim();
+}
+function stripScaffold(body) {
+  const kept = [];
+  for (const line of body.split(`
+`)) {
+    const trimmed = line.trim();
+    if (trimmed === "") {
+      kept.push(line);
+      continue;
+    }
+    const collapsed = stripHedging(collapseScaffoldLabels(trimmed));
+    if (collapsed === "")
+      continue;
+    kept.push(collapsed === trimmed ? line : collapsed);
+  }
+  return kept.join(`
+`);
+}
 function renderDoc(doc2, index) {
   const heading = doc2.title?.trim() || `\u65B0\u95FB ${index + 1}`;
   const link = doc2.sourceUri ? `[\u6765\u6E90\uFF1A${sourceLabelOf(doc2)}](${doc2.sourceUri})` : "";
@@ -21977,7 +22016,7 @@ function renderDoc(doc2, index) {
 
 ${metaLine ? `${metaLine}
 
-` : ""}${(doc2.body ?? "").trim()}`;
+` : ""}${stripScaffold(doc2.body ?? "").trim()}`;
 }
 function renderDailyArticle(docs, date6) {
   const items = docs.map(renderDoc).join(`
@@ -22043,6 +22082,7 @@ if (import.meta.main) {
   process.exit(await main2());
 }
 export {
+  stripScaffold,
   renderDailyArticle,
   processGeneratorIO,
   main2 as main,

package/plugins/generations/daily-article-gen/src/index.ts

@@ -8,7 +8,8 @@ import { echoError } from '@gobing-ai/ts-utils';
  * daily-article-gen: render a merged `Doc[]` into one readable daily AI-news
  * article. Mechanical template fill — no LLM, no STORM. Every news item keeps
  * a human-clickable source link (`[来源](sourceUri)`) so readers can trace
- * claims to their origin.
+ * claims to their origin. Research scaffolding is stripped before render
+ * (task 0138 R4b — see `stripScaffold`).
  *
  * Run date: `ARTICLE_DATE` env var (YYYYMMDD or ISO), else today (UTC).
  * Language-neutral: the Doc bodies already carry the summary language.
@@ -26,6 +27,71 @@ function sourceLabelOf(doc: Doc): string {
     }
 }
 
+/**
+ * Research-scaffold labels that must never reach reader copy (task 0138 R4b). The last30days
+ * research style emits them as `**「label」**` prefixes; the plan-translate prompt forbids them,
+ * but this deterministic pass catches what the agent misses. Grows with sources.
+ */
+const SCAFFOLD_LABELS = ['可做角度', '为什么现在值得注意', '社区讨论', '背景', '潜在影响', '影响'];
+
+/** Self-referential hedging markers — research voice, not news copy (task 0138 R4b). */
+const HEDGING_MARKERS = ['读者宜另行核实', '读者最好', '仍待验证', '尚待验证', '需另行核实'];
+
+/**
+ * Length of a leading scaffold label in any shape the research bodies use — `**「背景」**`,
+ * `「背景」`, `**背景**`, `### 背景`, `背景：` — or 0 when the line does not open with a known label.
+ * The capture is greedy and unbounded on purpose: it must reach the whole label (so `可做` cannot
+ * satisfy it where `可做角度` is the real one) and must not impose a length the label list outgrows.
+ * The `SCAFFOLD_LABELS` membership check, not the regex, decides.
+ */
+function scaffoldPrefixLength(text: string): number {
+    const match = text.match(
+        /^(?:#{1,6}\s*)?(?:\*\*)?\s*[「『]?\s*([^」』*：:\n]{2,})\s*[」』]?\s*(?:\*\*)?\s*[：:]?\s*/,
+    );
+    if (match === null || !SCAFFOLD_LABELS.includes((match[1] ?? '').trim())) return 0;
+    return match[0].length;
+}
+
+/** Drop every leading scaffold label, keeping the prose after it (`**「可做角度」** 可做角度：X` -> `X`). */
+function collapseScaffoldLabels(text: string): string {
+    let rest = text;
+    for (;;) {
+        const next = rest.slice(scaffoldPrefixLength(rest));
+        if (next === rest) return rest;
+        rest = next;
+    }
+}
+
+/** Drop the sentences carrying a hedging marker; a line that is only hedging collapses to ''. */
+function stripHedging(text: string): string {
+    if (!HEDGING_MARKERS.some((marker) => text.includes(marker))) return text;
+    return text
+        .split(/(?<=。)/)
+        .filter((sentence) => !HEDGING_MARKERS.some((marker) => sentence.includes(marker)))
+        .join('')
+        .trim();
+}
+
+/**
+ * Strip research scaffolding from a Doc body before it reaches reader copy (task 0138 R4b):
+ * standalone label lines are dropped, inline label prefixes collapse to their prose, and hedging
+ * sentences are removed. A clean body passes through byte-identical.
+ */
+export function stripScaffold(body: string): string {
+    const kept: string[] = [];
+    for (const line of body.split('\n')) {
+        const trimmed = line.trim();
+        if (trimmed === '') {
+            kept.push(line); // paragraph breaks survive untouched
+            continue;
+        }
+        const collapsed = stripHedging(collapseScaffoldLabels(trimmed));
+        if (collapsed === '') continue;
+        kept.push(collapsed === trimmed ? line : collapsed);
+    }
+    return kept.join('\n');
+}
+
 /** Render one news item section: heading, source link, and summary body. */
 function renderDoc(doc: Doc, index: number): string {
     const heading = doc.title?.trim() || `新闻 ${index + 1}`;
@@ -33,7 +99,7 @@ function renderDoc(doc: Doc, index: number): string {
     const meta = doc.metadata as Record<string, unknown> | undefined;
     const publishedAt = typeof meta?.publishedAt === 'string' ? meta.publishedAt.slice(0, 10) : '';
     const metaLine = [link, publishedAt].filter(Boolean).join(' · ');
-    return `## ${index + 1}. ${heading}\n\n${metaLine ? `${metaLine}\n\n` : ''}${(doc.body ?? '').trim()}`;
+    return `## ${index + 1}. ${heading}\n\n${metaLine ? `${metaLine}\n\n` : ''}${stripScaffold(doc.body ?? '').trim()}`;
 }
 
 /** Render the article markdown for a merged Doc[] (pure — exported for tests). */

package/plugins/generations/dailynews-gen/dist/index.js

@@ -21999,34 +21999,85 @@ function normalizeBroadcastText(raw, isZh) {
   }
   return text;
 }
-function synthesizePersonalUnderstanding(title, body, isZh) {
+function synthesizePersonalUnderstanding(title, body, isZh, used) {
   const combined = `${title} ${body}`.toLowerCase();
-  if (isZh) {
-    if (/jakub|szymon|\u9996\u5E2D\u79D1\u5B66\u5BB6|\u521B\u59CB\u56E2\u961F|\u6280\u672F\u9886\u8896|\u9EC4\u91D1\u642D\u6863/i.test(combined)) {
-      return "\u5728\u5404\u79CD\u6280\u672F\u5149\u73AF\u80CC\u540E\uFF0C\u6B63\u662F\u8FD9\u6837\u4E13\u6CE8\u800C\u4E0D\u77E5\u75B2\u5026\u7684\u5E95\u5C42\u653B\u575A\u8005\uFF0C\u4E3A\u4ECA\u5929\u6574\u4E2A\u5927\u6A21\u578B\u751F\u6001\u5960\u5B9A\u4E86\u6700\u575A\u5B9E\u7684\u6280\u672F\u57FA\u77F3\u3002";
-    }
-    if (/\u7B97\u529B|\u5409\u74E6|\u5DE5\u5382|\u7535\u529B|\u82AF\u7247|\u57FA\u7840\u8BBE\u65BD|\u57FA\u5EFA|\u786C\u4EF6|\u80FD\u6E90/i.test(combined)) {
-      return "\u8FD9\u5176\u5B9E\u91CA\u653E\u4E86\u4E00\u4E2A\u975E\u5E38\u660E\u786E\u7684\u4FE1\u53F7\uFF1AAI\u7ADE\u4E89\u7684\u4E0B\u534A\u573A\uFF0C\u5DF2\u7ECF\u4E0D\u518D\u53EA\u662F\u5355\u7EAF\u62FC\u6A21\u578B\u53C2\u6570\uFF0C\u66F4\u662F\u7535\u529B\u3001\u80FD\u6E90\u548C\u5168\u6808\u5DE5\u4E1A\u57FA\u7840\u8BBE\u65BD\u7684\u786C\u6838\u8F83\u91CF\u3002";
-    }
-    if (/sora|\u89C6\u9891|\u751F\u6210|cameo|chatgpt/i.test(combined)) {
-      return "\u6211\u4E2A\u4EBA\u89C9\u5F97\uFF0C\u8FD9\u6B21\u6700\u5927\u7684\u770B\u70B9\u662F\u5185\u5BB9\u521B\u4F5C\u95E8\u69DB\u7684\u518D\u6B21\u964D\u4F4E\u3002\u4E0D\u8FC7\u6B63\u5982\u56E2\u961F\u63D0\u5230\u7684\uFF0C\u5982\u4F55\u5E73\u8861\u521B\u4F5C\u81EA\u7531\u4E0E\u9632\u6B62\u4F4E\u8D28\u5185\u5BB9\u6CDB\u6EE5\uFF0C\u4F1A\u662F\u6574\u4E2A\u884C\u4E1A\u63A5\u4E0B\u6765\u6700\u503C\u5F97\u89C2\u5BDF\u7684\u5173\u952E\u8003\u9898\u3002";
-    }
-    if (/\u878D\u8D44|\u6295\u8D44|\u5E02\u573A|\u5546\u4E1A|\u4F30\u503C|\u5408\u4F5C/i.test(combined)) {
-      return "\u8FD9\u9879\u8FDB\u5C55\u53CD\u6620\u51FA\u8D44\u672C\u4E0E\u4EA7\u4E1A\u754C\u6B63\u5728\u52A0\u901F\u5F62\u6210\u5408\u529B\uFF0C\u540E\u7EED\u5B9E\u9645\u4E1A\u52A1\u573A\u666F\u7684\u843D\u5730\u6548\u679C\u975E\u5E38\u503C\u5F97\u6211\u4EEC\u6301\u7EED\u8DDF\u8FDB\u3002";
-    }
-    return "\u8FD9\u9879\u65B0\u52A8\u6001\u53CD\u6620\u51FA\u6280\u672F\u6F14\u8FDB\u6B63\u5728\u8FDB\u4E00\u6B65\u63D0\u901F\uFF0C\u5BF9\u76F8\u5173\u9886\u57DF\u7684\u5B9E\u9645\u5F71\u54CD\u4E5F\u975E\u5E38\u503C\u5F97\u6211\u4EEC\u6301\u7EED\u5173\u6CE8\u3002";
+  const line = isZh ? zhInsightFor(combined) : enInsightFor(combined);
+  if (used.has(line))
+    return null;
+  used.add(line);
+  return line;
+}
+function zhInsightFor(combined) {
+  if (/jakub|szymon|\u9996\u5E2D\u79D1\u5B66\u5BB6|\u521B\u59CB\u56E2\u961F|\u6280\u672F\u9886\u8896|\u9EC4\u91D1\u642D\u6863/i.test(combined)) {
+    return "\u5728\u5404\u79CD\u6280\u672F\u5149\u73AF\u80CC\u540E\uFF0C\u6B63\u662F\u8FD9\u6837\u4E13\u6CE8\u800C\u4E0D\u77E5\u75B2\u5026\u7684\u5E95\u5C42\u653B\u575A\u8005\uFF0C\u4E3A\u4ECA\u5929\u6574\u4E2A\u5927\u6A21\u578B\u751F\u6001\u5960\u5B9A\u4E86\u6700\u575A\u5B9E\u7684\u6280\u672F\u57FA\u77F3\u3002";
+  }
+  if (/\u7B97\u529B|\u5409\u74E6|\u7535\u529B|\u80FD\u6E90|\u6570\u636E\u4E2D\u5FC3|\u57FA\u7840\u8BBE\u65BD|\u57FA\u5EFA/.test(combined)) {
+    return "\u8FD9\u5176\u5B9E\u91CA\u653E\u4E86\u4E00\u4E2A\u975E\u5E38\u660E\u786E\u7684\u4FE1\u53F7\uFF1AAI\u7ADE\u4E89\u7684\u4E0B\u534A\u573A\uFF0C\u5DF2\u7ECF\u4E0D\u518D\u53EA\u662F\u5355\u7EAF\u62FC\u6A21\u578B\u53C2\u6570\uFF0C\u66F4\u662F\u7535\u529B\u3001\u80FD\u6E90\u548C\u5168\u6808\u5DE5\u4E1A\u57FA\u7840\u8BBE\u65BD\u7684\u786C\u6838\u8F83\u91CF\u3002";
+  }
+  if (/\u82AF\u7247|\u786C\u4EF6|\u5904\u7406\u5668|enclave/i.test(combined)) {
+    return "\u771F\u6B63\u88AB\u91CD\u65B0\u5B9A\u4E49\u7684\uFF0C\u662F\u786C\u4EF6\u5C42\u627F\u62C5\u7684\u5B89\u5168\u4E0E\u7B97\u529B\u8D23\u4EFB\uFF1A\u628A\u80FD\u529B\u505A\u8FDB\u7845\u7247\u5C42\u66F4\u53EF\u9760\u3001\u4E5F\u66F4\u96BE\u7ED5\u8FC7\uFF0C\u4F46\u751F\u6001\u9501\u5B9A\u7684\u4EE3\u4EF7\u9700\u8981\u4E00\u8D77\u7B97\u8FDB\u53BB\u3002";
   }
+  if (/sora|\u89C6\u9891|\u751F\u6210\u89C6\u9891|\u89C6\u9891\u751F\u6210|\u6587\u751F\u89C6\u9891|\u5185\u5BB9\u521B\u4F5C|cameo/i.test(combined)) {
+    return "\u6211\u4E2A\u4EBA\u89C9\u5F97\uFF0C\u8FD9\u6B21\u6700\u5927\u7684\u770B\u70B9\u662F\u5185\u5BB9\u521B\u4F5C\u95E8\u69DB\u7684\u518D\u6B21\u964D\u4F4E\u3002\u4E0D\u8FC7\u6B63\u5982\u56E2\u961F\u63D0\u5230\u7684\uFF0C\u5982\u4F55\u5E73\u8861\u521B\u4F5C\u81EA\u7531\u4E0E\u9632\u6B62\u4F4E\u8D28\u5185\u5BB9\u6CDB\u6EE5\uFF0C\u4F1A\u662F\u6574\u4E2A\u884C\u4E1A\u63A5\u4E0B\u6765\u6700\u503C\u5F97\u89C2\u5BDF\u7684\u5173\u952E\u8003\u9898\u3002";
+  }
+  if (/\u878D\u8D44|\u6295\u8D44|\u5E02\u573A|\u5546\u4E1A|\u4F30\u503C|\u5408\u4F5C/.test(combined)) {
+    return "\u8FD9\u9879\u8FDB\u5C55\u53CD\u6620\u51FA\u8D44\u672C\u4E0E\u4EA7\u4E1A\u754C\u6B63\u5728\u52A0\u901F\u5F62\u6210\u5408\u529B\uFF0C\u540E\u7EED\u5B9E\u9645\u4E1A\u52A1\u573A\u666F\u7684\u843D\u5730\u6548\u679C\u975E\u5E38\u503C\u5F97\u6211\u4EEC\u6301\u7EED\u8DDF\u8FDB\u3002";
+  }
+  return "\u8FD9\u9879\u65B0\u52A8\u6001\u53CD\u6620\u51FA\u6280\u672F\u6F14\u8FDB\u6B63\u5728\u8FDB\u4E00\u6B65\u63D0\u901F\uFF0C\u5BF9\u76F8\u5173\u9886\u57DF\u7684\u5B9E\u9645\u5F71\u54CD\u4E5F\u975E\u5E38\u503C\u5F97\u6211\u4EEC\u6301\u7EED\u5173\u6CE8\u3002";
+}
+function enInsightFor(combined) {
   if (/jakub|szymon|chief scientist|founding team|technical leader|partnership/i.test(combined)) {
     return "Behind the headlines, it is relentless technical leadership and deep engineering rigor that continue to drive these foundational milestones.";
   }
-  if (/compute|gigawatt|factory|power|chip|infrastructure|datacenter/i.test(combined)) {
+  if (/compute|gigawatt|power|energy|datacenter|infrastructure/.test(combined)) {
     return "This signals clearly that the AI frontier is no longer just about algorithms, but a massive contest of power, energy, and full-stack industrial engineering.";
   }
-  if (/sora|video|generation|cameo|chatgpt/i.test(combined)) {
+  if (/chip|hardware|processor|enclave/i.test(combined)) {
+    return "What stands out is the hardware layer quietly taking on the security and compute burden: putting the guarantee in silicon is harder to bypass, and it carries its own lock-in cost.";
+  }
+  if (/sora|video|video generation|content creation|cameo/i.test(combined)) {
     return "In my view, the biggest takeaway here is the dramatic lowering of creation barriers, while content quality governance will be the key test ahead.";
   }
   return "This development highlights how rapidly the ecosystem is moving from experimental proofs of concept into large-scale execution.";
 }
+var MAX_FACT_SEGMENT_CHARS = 300;
+var TITLE_ECHO_NOISE_RE = /[\uFF0C\u3002\u3001\uFF1A:\uFF1B;!\uFF01?\uFF1F\u2026\s\u300C\u300D\u300E\u300F\u3010\u3011]/g;
+function stripLeadingTitleEcho(body, title) {
+  const wanted = title.replace(TITLE_ECHO_NOISE_RE, "");
+  if (wanted.length < 4)
+    return body;
+  let matched = "";
+  let end = -1;
+  for (let i = 0;i < body.length && matched.length < wanted.length; i++) {
+    const kept = (body[i] ?? "").replace(TITLE_ECHO_NOISE_RE, "");
+    if (kept === "")
+      continue;
+    matched += kept;
+    if (matched.length >= wanted.length)
+      end = i + 1;
+  }
+  if (end === -1 || matched !== wanted)
+    return body;
+  return body.slice(end).replace(/^[\uFF0C\u3002\u3001\uFF1A:\uFF1B;!\uFF01?\uFF1F\u2026\s\u300C\u300D\u300E\u300F\u3010\u3011]+/, "");
+}
+function splitFactText(text, maxChars = MAX_FACT_SEGMENT_CHARS) {
+  if (text.length <= maxChars)
+    return [text];
+  const clauses = text.match(/[^\u3002\uFF01\uFF1F\u2026\uFF1B]+[\u3002\uFF01\uFF1F\u2026\uFF1B]*|[\s\S]/g) ?? [text];
+  const chunks = [];
+  let current = "";
+  for (const clause of clauses) {
+    if (current !== "" && current.length + clause.length > maxChars) {
+      chunks.push(current);
+      current = clause;
+      continue;
+    }
+    current += clause;
+  }
+  if (current !== "")
+    chunks.push(current);
+  return chunks.length > 0 ? chunks : [text];
+}
 function buildNewsVoiceScript(docs, options) {
   const language = exports_external.enum(["zh", "en"]).parse((options?.language || "zh").toLowerCase());
   const isZh = language === "zh";
@@ -22086,26 +22137,31 @@ function buildNewsVoiceScript(docs, options) {
     gap_ms: 700,
     emotion: "confident, warm"
   });
+  const usedInsights = new Set;
   docs.forEach((doc2, idx) => {
     const isLast = idx === docs.length - 1 && docs.length > 1;
     const cleanTitle = (doc2.title ?? "").trim();
-    const cleanBody = (doc2.body ?? "").trim();
+    const cleanBody = stripLeadingTitleEcho((doc2.body ?? "").trim(), cleanTitle);
     const lead = isZh ? idx === 0 ? `\u9996\u5148\u6765\u804A\u804A\u5927\u5BB6\u975E\u5E38\u5173\u6CE8\u7684\u3010${cleanTitle}\u3011\u3002` : isLast ? `\u6700\u540E\uFF0C\u6765\u770B\u770B\u4ECA\u5929\u7684\u6700\u540E\u4E00\u6761\u8D44\u8BAF\uFF0C\u3010${cleanTitle}\u3011\u3002` : `\u63A5\u7740\u6211\u4EEC\u628A\u76EE\u5149\u8F6C\u5411\u53E6\u4E00\u6761\u91CD\u8981\u8FDB\u5C55\uFF0C\u3010${cleanTitle}\u3011\u3002` : idx === 0 ? `First up today, let's look at ${cleanTitle}.` : isLast ? `And finally today, let's wrap up with ${cleanTitle}.` : `Next, turning our attention to ${cleanTitle}.`;
     const normalizedLead = normalizeBroadcastText(lead, isZh);
-    const normalizedBody = normalizeBroadcastText(cleanBody, isZh);
-    const insight = synthesizePersonalUnderstanding(cleanTitle, cleanBody, isZh);
-    const normalizedInsight = normalizeBroadcastText(insight, isZh);
+    const normalizedBody = cleanBody === "" ? "" : normalizeBroadcastText(cleanBody, isZh);
+    const insight = synthesizePersonalUnderstanding(cleanTitle, cleanBody, isZh, usedInsights);
+    const normalizedInsight = insight === null ? null : normalizeBroadcastText(insight, isZh);
     const factText = isZh ? `${normalizedLead}${normalizedBody}` : `${normalizedLead} ${normalizedBody}`;
-    segments.push({
-      text: factText,
-      gap_ms: 500,
-      emotion: "professional, clear"
-    });
-    segments.push({
-      text: normalizedInsight,
-      gap_ms: 800,
-      emotion: "thoughtful, personal"
-    });
+    for (const chunk of splitFactText(factText)) {
+      segments.push({
+        text: chunk,
+        gap_ms: 500,
+        emotion: "professional, clear"
+      });
+    }
+    if (normalizedInsight !== null) {
+      segments.push({
+        text: normalizedInsight,
+        gap_ms: 800,
+        emotion: "thoughtful, personal"
+      });
+    }
   });
   const outroText = isZh ? `\u597D\u4E86\uFF0C\u4EE5\u4E0A\u5C31\u662F\u4ECA\u5929AI\u79D1\u6280\u6F2B\u8C08\u7684\u5168\u90E8\u5185\u5BB9\u3002\u4E0D\u77E5\u9053\u4ECA\u5929\u8FD9\u51E0\u6761\u6D88\u606F\u6709\u6CA1\u6709\u7ED9\u60A8\u5E26\u6765\u4E00\u4E9B\u65B0\u7684\u542F\u53D1\uFF1F\u611F\u8C22\u60A8\u7684\u6536\u542C\uFF0C\u6211\u662F${hostName}\uFF0C\u795D\u60A8\u62E5\u6709\u5145\u5B9E\u9AD8\u6548\u7684\u4E00\u5929\uFF0C\u6211\u4EEC\u4E0B\u671F\u8282\u76EE\u4E0D\u89C1\u4E0D\u6563\uFF01` : `That wraps up today's AI Daily Briefing. I hope today's highlights gave you some useful takeaways. Thank you for listening. This is ${hostName}, wishing you a productive and wonderful day. Until next time!`;
   segments.push({
@@ -22330,6 +22386,8 @@ if (import.meta.main) {
 }
 export {
   synthesizePersonalUnderstanding,
+  stripLeadingTitleEcho,
+  splitFactText,
   processGeneratorIO,
   normalizeBroadcastText,
   main2 as main,
@@ -22340,5 +22398,6 @@ export {
   VoiceSpeakerSchema,
   VoiceScriptSegmentSchema,
   VoiceScriptSchema,
+  MAX_FACT_SEGMENT_CHARS,
   EffectConfigSchema
 };

package/plugins/generations/dailynews-gen/src/script-builder.ts

@@ -144,39 +144,131 @@ export function normalizeBroadcastText(raw: string, isZh: boolean): string {
 /**
  * Synthesizes a natural, concise spoken perspective / takeaway based on the story facts.
  * Gives the monologue a demonstrative tech podcast talk feel rather than mechanical news reading.
+ *
+ * `used` is the per-episode set of lines already spoken (task 0139 R2). A colliding item gets
+ * `null` — no commentary segment at all — because a repeated line is worse than none and a
+ * second-choice line would be a different topic misfire (Q&A item 5).
  */
-export function synthesizePersonalUnderstanding(title: string, body: string, isZh: boolean): string {
+export function synthesizePersonalUnderstanding(
+    title: string,
+    body: string,
+    isZh: boolean,
+    used: Set<string>,
+): string | null {
     const combined = `${title} ${body}`.toLowerCase();
+    const line = isZh ? zhInsightFor(combined) : enInsightFor(combined);
+    if (used.has(line)) return null;
+    used.add(line);
+    return line;
+}
 
-    if (isZh) {
-        if (/jakub|szymon|首席科学家|创始团队|技术领袖|黄金搭档/i.test(combined)) {
-            return '在各种技术光环背后，正是这样专注而不知疲倦的底层攻坚者，为今天整个大模型生态奠定了最坚实的技术基石。';
-        }
-        if (/算力|吉瓦|工厂|电力|芯片|基础设施|基建|硬件|能源/i.test(combined)) {
-            return '这其实释放了一个非常明确的信号：AI竞争的下半场，已经不再只是单纯拼模型参数，更是电力、能源和全栈工业基础设施的硬核较量。';
-        }
-        if (/sora|视频|生成|cameo|chatgpt/i.test(combined)) {
-            return '我个人觉得，这次最大的看点是内容创作门槛的再次降低。不过正如团队提到的，如何平衡创作自由与防止低质内容泛滥，会是整个行业接下来最值得观察的关键考题。';
-        }
-        if (/融资|投资|市场|商业|估值|合作/i.test(combined)) {
-            return '这项进展反映出资本与产业界正在加速形成合力，后续实际业务场景的落地效果非常值得我们持续跟进。';
-        }
-        return '这项新动态反映出技术演进正在进一步提速，对相关领域的实际影响也非常值得我们持续关注。';
+/**
+ * Bucket order is first-match-wins, and the buckets are deliberately disjoint (task 0139 R2).
+ * The energy line requires energy-side keywords: the pre-fix bucket matched bare
+ * `芯片|硬件|工厂`, so it fired after the Nvidia-regulation and Apple Secure-Enclave items on the
+ * 2026-09-16 run. Hardware/chip content now has its own line instead.
+ */
+function zhInsightFor(combined: string): string {
+    if (/jakub|szymon|首席科学家|创始团队|技术领袖|黄金搭档/i.test(combined)) {
+        return '在各种技术光环背后，正是这样专注而不知疲倦的底层攻坚者，为今天整个大模型生态奠定了最坚实的技术基石。';
+    }
+    if (/算力|吉瓦|电力|能源|数据中心|基础设施|基建/.test(combined)) {
+        return '这其实释放了一个非常明确的信号：AI竞争的下半场，已经不再只是单纯拼模型参数，更是电力、能源和全栈工业基础设施的硬核较量。';
     }
+    if (/芯片|硬件|处理器|enclave/i.test(combined)) {
+        return '真正被重新定义的，是硬件层承担的安全与算力责任：把能力做进硅片层更可靠、也更难绕过，但生态锁定的代价需要一起算进去。';
+    }
+    // Video/creation only — bare `生成` matched every generative-AI body (Meta and both Gemini
+    // items on the 2026-09-16 run) and bare `chatgpt` matched the Meta item, which is what made
+    // this line a topic misfire. A product name is not a creation-barrier story.
+    if (/sora|视频|生成视频|视频生成|文生视频|内容创作|cameo/i.test(combined)) {
+        return '我个人觉得，这次最大的看点是内容创作门槛的再次降低。不过正如团队提到的，如何平衡创作自由与防止低质内容泛滥，会是整个行业接下来最值得观察的关键考题。';
+    }
+    if (/融资|投资|市场|商业|估值|合作/.test(combined)) {
+        return '这项进展反映出资本与产业界正在加速形成合力，后续实际业务场景的落地效果非常值得我们持续跟进。';
+    }
+    return '这项新动态反映出技术演进正在进一步提速，对相关领域的实际影响也非常值得我们持续关注。';
+}
 
-    // English spoken insights
+/** English mirror of {@link zhInsightFor} — same bucket order, same disjointness rule. */
+function enInsightFor(combined: string): string {
     if (/jakub|szymon|chief scientist|founding team|technical leader|partnership/i.test(combined)) {
         return 'Behind the headlines, it is relentless technical leadership and deep engineering rigor that continue to drive these foundational milestones.';
     }
-    if (/compute|gigawatt|factory|power|chip|infrastructure|datacenter/i.test(combined)) {
+    if (/compute|gigawatt|power|energy|datacenter|infrastructure/.test(combined)) {
         return 'This signals clearly that the AI frontier is no longer just about algorithms, but a massive contest of power, energy, and full-stack industrial engineering.';
     }
-    if (/sora|video|generation|cameo|chatgpt/i.test(combined)) {
+    if (/chip|hardware|processor|enclave/i.test(combined)) {
+        return 'What stands out is the hardware layer quietly taking on the security and compute burden: putting the guarantee in silicon is harder to bypass, and it carries its own lock-in cost.';
+    }
+    if (/sora|video|video generation|content creation|cameo/i.test(combined)) {
         return 'In my view, the biggest takeaway here is the dramatic lowering of creation barriers, while content quality governance will be the key test ahead.';
     }
     return 'This development highlights how rapidly the ecosystem is moving from experimental proofs of concept into large-scale execution.';
 }
 
+/**
+ * Longest fact segment the script emits (task 0139 R4).
+ * ponytail: observed zh speech rates 3.4–6.1 chars/s put 300 chars at 49–88s per chunk; raise only
+ * with measured evidence. 400 would reach ~117s at the slowest observed rate, where whisper
+ * transcription fidelity collapsed to 0.25–0.41 on the 2026-09-16 run.
+ */
+export const MAX_FACT_SEGMENT_CHARS = 300;
+
+/** Punctuation and space the translator drifts between a title and its body restatement. */
+const TITLE_ECHO_NOISE_RE = /[，。、：:；;!！?？…\s「」『』【】]/g;
+
+/**
+ * Drop a leading restatement of the item title from its body (task 0139 R1).
+ *
+ * `plan-translate` derives each title from the body's first sentence, so the echo is always a
+ * prefix with at most punctuation drift (`：` → `，`) — the rule is a deterministic significant-
+ * character prefix compare, never a fuzzy/LLM similarity match (anti-pattern list). The body is
+ * returned untouched whenever the normalized prefix does not equal the normalized title, and a
+ * body that is *only* the restatement collapses to the empty string (the lead alone carries it).
+ */
+export function stripLeadingTitleEcho(body: string, title: string): string {
+    const wanted = title.replace(TITLE_ECHO_NOISE_RE, '');
+    if (wanted.length < 4) return body; // a short title would match too much of the body
+    let matched = '';
+    let end = -1;
+    for (let i = 0; i < body.length && matched.length < wanted.length; i++) {
+        const kept = (body[i] ?? '').replace(TITLE_ECHO_NOISE_RE, '');
+        if (kept === '') continue; // drift punctuation between title and echo
+        matched += kept;
+        if (matched.length >= wanted.length) end = i + 1;
+    }
+    if (end === -1 || matched !== wanted) return body;
+    return body.slice(end).replace(/^[，。、：:；;!！?？…\s「」『』【】]+/, '');
+}
+
+/**
+ * Split a fact text into chunks of at most `maxChars` (task 0139 R4).
+ *
+ * Greedy packing at clause boundaries: a chunk closes when the next clause would push it past the
+ * cap, so `chunks.join('') === text` always holds — the `[\s\S]` alternative is what keeps leading
+ * or stray punctuation from being dropped by the match. `；` is a boundary too: a clause-joined
+ * sentence would otherwise stay whole and blow past the ~90s budget the cap exists to protect. A
+ * single clause longer than the cap is emitted whole — splitting mid-thought is worse than a long
+ * segment — and is left to the QC duration audit.
+ */
+export function splitFactText(text: string, maxChars: number = MAX_FACT_SEGMENT_CHARS): string[] {
+    if (text.length <= maxChars) return [text];
+    const clauses = text.match(/[^。！？…；]+[。！？…；]*|[\s\S]/g) ?? [text];
+    const chunks: string[] = [];
+    let current = '';
+    for (const clause of clauses) {
+        if (current !== '' && current.length + clause.length > maxChars) {
+            chunks.push(current);
+            current = clause;
+            continue;
+        }
+        current += clause;
+    }
+    if (current !== '') chunks.push(current);
+    return chunks.length > 0 ? chunks : [text];
+}
+
 export function buildNewsVoiceScript(docs: Doc[], options?: ScriptBuilderOptions): VoiceScript {
     const language = z.enum(['zh', 'en']).parse((options?.language || 'zh').toLowerCase());
     const isZh = language === 'zh';
@@ -259,10 +351,13 @@ export function buildNewsVoiceScript(docs: Doc[], options?: ScriptBuilderOptions
     });
 
     // 2. Story segments: demonstrative talk with facts + conversational insight
+    const usedInsights = new Set<string>(); // one per episode (task 0139 R2)
     docs.forEach((doc, idx) => {
         const isLast = idx === docs.length - 1 && docs.length > 1;
         const cleanTitle = (doc.title ?? '').trim();
-        const cleanBody = (doc.body ?? '').trim();
+        // R1: drop the title restatement plan-translate prepends to the body, so the lead's
+        // 【title】 is the only place the title is spoken for this item.
+        const cleanBody = stripLeadingTitleEcho((doc.body ?? '').trim(), cleanTitle);
 
         // Conversational lead
         const lead = isZh
@@ -278,23 +373,28 @@ export function buildNewsVoiceScript(docs: Doc[], options?: ScriptBuilderOptions
                 : `Next, turning our attention to ${cleanTitle}.`;
 
         const normalizedLead = normalizeBroadcastText(lead, isZh);
-        const normalizedBody = normalizeBroadcastText(cleanBody, isZh);
-        const insight = synthesizePersonalUnderstanding(cleanTitle, cleanBody, isZh);
-        const normalizedInsight = normalizeBroadcastText(insight, isZh);
+        const normalizedBody = cleanBody === '' ? '' : normalizeBroadcastText(cleanBody, isZh);
+        const insight = synthesizePersonalUnderstanding(cleanTitle, cleanBody, isZh, usedInsights);
+        const normalizedInsight = insight === null ? null : normalizeBroadcastText(insight, isZh);
 
         const factText = isZh ? `${normalizedLead}${normalizedBody}` : `${normalizedLead} ${normalizedBody}`;
 
-        // Emit facts followed by personal insight commentary
-        segments.push({
-            text: factText,
-            gap_ms: 500, // Natural breathing pause before giving personal perspective
-            emotion: 'professional, clear',
-        });
-        segments.push({
-            text: normalizedInsight,
-            gap_ms: 800, // Distinct 800ms pause before the next story
-            emotion: 'thoughtful, personal',
-        });
+        // R4: facts first, chunked so no single segment runs past ~90s of speech; then the insight
+        // only when this episode has not already spoken that line (R2 collision ⇒ no segment).
+        for (const chunk of splitFactText(factText)) {
+            segments.push({
+                text: chunk,
+                gap_ms: 500, // Natural breathing pause before giving personal perspective
+                emotion: 'professional, clear',
+            });
+        }
+        if (normalizedInsight !== null) {
+            segments.push({
+                text: normalizedInsight,
+                gap_ms: 800, // Distinct 800ms pause before the next story
+                emotion: 'thoughtful, personal',
+            });
+        }
     });
 
     // 3. Outro segment: Warm, conversational wrap-up

package/plugins/generations/episode-plan-gen/dist/index.js

@@ -21968,6 +21968,7 @@ var SOURCE_BASE = {
 };
 var DEFAULT_SOURCE_BASE = 5;
 var TITLE_SIMILARITY_THRESHOLD = 0.6;
+var STUB_BODY_FLOOR = 300;
 var FRESHNESS_DAYS = 7;
 var PER_SOURCE_CAP = 5;
 var NON_NEWS_TAGS = new Set(["geopolitics", "military-conflict", "war"]);
@@ -22166,8 +22167,9 @@ function classifyCategory(doc2) {
 }
 function scoreDimensions(doc2, planDate) {
   const meta3 = asRecord(doc2.metadata);
-  let quality = 1 + Math.min(4, Math.round((doc2.body ?? "").length / 1500));
-  if (doc2.sourceUri)
+  const bodyLen = (doc2.body ?? "").length;
+  let quality = bodyLen < STUB_BODY_FLOOR ? 1 : 1 + Math.min(4, Math.round(bodyLen / 1500));
+  if (doc2.sourceUri && bodyLen >= STUB_BODY_FLOOR)
     quality = Math.min(5, quality + 1);
   const source = typeof meta3.source === "string" ? meta3.source : "";
   let importance = IMPORTANCE_BASE[source] ?? DEFAULT_IMPORTANCE;

package/plugins/generations/episode-plan-gen/src/index.ts

@@ -35,6 +35,11 @@ const SOURCE_BASE: Record<string, number> = {
 };
 const DEFAULT_SOURCE_BASE = 5;
 const TITLE_SIMILARITY_THRESHOLD = 0.6; // character-bigram Jaccard
+// Minimum body length for any substance credit (task 0138 D2). Scoring runs on blend-time bodies:
+// the 2026-09-16 aihot stubs are 234-255 EN chars, substantive items are >=348 — a headline-echo
+// stub must not reach `qc_min_quality=2` on its sourceUri bonus alone.
+// ponytail: re-calibrate with evidence if a source ships legitimate 256-347-char items.
+const STUB_BODY_FLOOR = 300;
 const FRESHNESS_DAYS = 7; // a daily episode drops items published before this window (dogfood 20260906: aihot served a 2020-2025 blog archive as "selected")
 const PER_SOURCE_CAP = 5; // one feed can't flood the episode even when all its items score high
 const NON_NEWS_TAGS = new Set(['geopolitics', 'military-conflict', 'war']); // horizon finance-news tags (dogfood 20260906: Iran oil tankers in an AI episode)
@@ -308,8 +313,10 @@ export function classifyCategory(doc: Doc): Category {
  */
 export function scoreDimensions(doc: Doc, planDate: string): DimensionScores {
     const meta = asRecord(doc.metadata);
-    let quality = 1 + Math.min(4, Math.round((doc.body ?? '').length / 1500));
-    if (doc.sourceUri) quality = Math.min(5, quality + 1);
+    // The floor short-circuits the sourcedness bonus: "quality 1 regardless of sourceUri" (R2).
+    const bodyLen = (doc.body ?? '').length;
+    let quality = bodyLen < STUB_BODY_FLOOR ? 1 : 1 + Math.min(4, Math.round(bodyLen / 1500));
+    if (doc.sourceUri && bodyLen >= STUB_BODY_FLOOR) quality = Math.min(5, quality + 1);
 
     const source = typeof meta.source === 'string' ? meta.source : '';
     let importance = IMPORTANCE_BASE[source] ?? DEFAULT_IMPORTANCE;