feat(ai-writing): 大纲先行确认 + 新文章类型识别 + 参考文献 + 编辑器修复

- 写作配置 Q&A 改为「先流式生成大纲并多轮确认 → 再确认意图」，放宽轮数上限
- 意图识别出列表外的新文章类型时带上类型名+简介（仅本次写作临时使用）
- 成稿支持参考文献：正文 [n] 角标 + 文末按出现顺序的「参考文献」列表（表单开关，默认开）
- 流式接取未闭合 <think> 思考内容，首 token 前加「正在整理…」占位
- 修复右侧编辑器输入空格只出一次、光标跳字（避免自激回灌 setContent 重置光标）

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

frontend-web/src/components/workspace/messages/message-list.tsx

@@ -1,6 +1,6 @@
 import type { Message } from "@langchain/langgraph-sdk";
 import type { BaseStream } from "@langchain/langgraph-sdk/react";
-import { CheckIcon, SendIcon, XIcon } from "lucide-react";
+import { CheckIcon, PenLineIcon, SendIcon, XIcon } from "lucide-react";
 import { useCallback, useContext, useEffect, useMemo, useRef, useState } from "react";
 import { useStickToBottomContext } from "use-stick-to-bottom";
 
@@ -20,10 +20,10 @@ import {
   groupMessages,
   hasContent,
   hasPresentFiles,
-  hasReasoning,
-  hasToolCalls,
-  isHiddenFromUIMessage,
-} from "@/core/messages/utils";
+  hasReasoning,
+  hasToolCalls,
+  isHiddenFromUIMessage,
+} from "@/core/messages/utils";
 import { useRehypeSplitWordsIntoSpans } from "@/core/rehype";
 import { isThreadRagModeLocked, useLocalSettings } from "@/core/settings";
 import { useSkills } from "@/core/skills/hooks";
@@ -103,6 +103,7 @@ export function MessageList({
   onSetupAgentReject,
   onSetupWritingApprove,
   onSetupWritingReject,
+  onSetupWritingStart,
 }: {
   className?: string;
   contentClassName?: string;
@@ -120,6 +121,8 @@ export function MessageList({
   onSetupAgentReject?: (payload: SetupAgentApprovalPayload) => void | Promise<void>;
   onSetupWritingApprove?: (payload: SetupWritingApprovalPayload) => void | Promise<void>;
   onSetupWritingReject?: (payload: SetupWritingApprovalPayload) => void | Promise<void>;
+  /** 审批卡「直接开始写作」回调：应用配置（含已确认大纲）并立即启动写作流程（仅写作页提供，模板编辑等场景缺省隐藏该按钮）。 */
+  onSetupWritingStart?: (payload: SetupWritingApprovalPayload) => void | Promise<void>;
 }) {
   const baseRehypePlugins = useRehypeSplitWordsIntoSpans(thread.isLoading);
   const rehypePlugins = useMemo(
@@ -152,10 +155,10 @@ export function MessageList({
     });
   }, []);
 
-  const allMessages = useMemo(
-    () => thread.messages.filter((m) => !isHiddenFromUIMessage(m)),
-    [thread.messages],
-  );
+  const allMessages = useMemo(
+    () => thread.messages.filter((m) => !isHiddenFromUIMessage(m)),
+    [thread.messages],
+  );
   const messages = useMemo(
     () =>
       hiddenMessageIds.size === 0
@@ -430,6 +433,7 @@ export function MessageList({
                   message={message}
                   onApprove={onSetupWritingApprove}
                   onReject={onSetupWritingReject}
+                  onStart={onSetupWritingStart}
                 />
                 <MessageTokenUsageList
                   enabled={tokenUsageEnabled}
@@ -629,6 +633,8 @@ export type SetupAgentApprovalPayload = {
 export type SetupWritingApprovalPayload = {
   user_intent?: string;
   article_type?: string;
+  /** 当 article_type 是已有类型库里没有的新类型时，模型识别出的一句话特征说明（命中已有类型时为空）。 */
+  article_type_detail?: string;
   word_count_target?: number | null;
   target_audience?: string;
   writing_mode?: string;
@@ -820,14 +826,16 @@ function SetupWritingApprovalCard({
   message,
   onApprove,
   onReject,
+  onStart,
 }: {
   disabled: boolean;
   message: unknown;
   onApprove?: (payload: SetupWritingApprovalPayload) => void | Promise<void>;
   onReject?: (payload: SetupWritingApprovalPayload) => void | Promise<void>;
+  onStart?: (payload: SetupWritingApprovalPayload) => void | Promise<void>;
 }) {
   const payload = getSetupWritingApprovalPayload(message);
-  const [submitting, setSubmitting] = useState<"approve" | "reject" | null>(
+  const [submitting, setSubmitting] = useState<"approve" | "reject" | "start" | null>(
     null,
   );
 
@@ -838,7 +846,7 @@ function SetupWritingApprovalCard({
   const approvalPayload = payload;
   const busy = disabled || submitting !== null;
 
-  async function run(action: "approve" | "reject") {
+  async function run(action: "approve" | "reject" | "start") {
     if (busy) {
       return;
     }
@@ -846,6 +854,8 @@ function SetupWritingApprovalCard({
     try {
       if (action === "approve") {
         await onApprove?.(approvalPayload);
+      } else if (action === "start") {
+        await onStart?.(approvalPayload);
       } else {
         await onReject?.(approvalPayload);
       }
@@ -854,9 +864,14 @@ function SetupWritingApprovalCard({
     }
   }
 
+  const hasNewType = !!approvalPayload.article_type_detail?.trim();
   const rows: Array<{ label: string; value: string }> = [];
   if (approvalPayload.article_type) {
-    rows.push({ label: "文章类型", value: approvalPayload.article_type });
+    rows.push({
+      label: "文章类型",
+      // 新类型（列表外）标注「· 新」，让用户知道这是模型新识别的类型。
+      value: hasNewType ? `${approvalPayload.article_type} · 新` : approvalPayload.article_type,
+    });
   }
   if (typeof approvalPayload.word_count_target === "number") {
     rows.push({ label: "目标字数", value: `${approvalPayload.word_count_target} 字` });
@@ -903,6 +918,13 @@ function SetupWritingApprovalCard({
           </div>
         ) : null}
 
+        {hasNewType ? (
+          <div className="space-y-1">
+            <div className="text-muted-foreground text-xs">新类型说明</div>
+            <p className="text-sm leading-6">{approvalPayload.article_type_detail}</p>
+          </div>
+        ) : null}
+
         {approvalPayload.outline ? (
           <div className="space-y-1">
             <div className="text-muted-foreground text-xs">已有大纲</div>
@@ -926,6 +948,7 @@ function SetupWritingApprovalCard({
         </Button>
         <Button
           type="button"
+          variant={onStart ? "outline" : "default"}
           size="sm"
           disabled={busy || !onApprove}
           onClick={() => void run("approve")}
@@ -933,6 +956,17 @@ function SetupWritingApprovalCard({
           <CheckIcon className="h-4 w-4" />
           应用到表单
         </Button>
+        {onStart ? (
+          <Button
+            type="button"
+            size="sm"
+            disabled={busy}
+            onClick={() => void run("start")}
+          >
+            <PenLineIcon className="h-4 w-4" />
+            直接开始写作
+          </Button>
+        ) : null}
       </div>
     </div>
   );

frontend-web/src/core/messages/utils.ts

@@ -212,21 +212,33 @@ export function extractToolResultPayload(
 }
 
 const THINK_TAG_RE = /<think>\s*([\s\S]*?)\s*<\/think>/g;
+const THINK_OPEN = "<think>";
 
 function splitInlineReasoning(content: string) {
   const reasoningParts: string[] = [];
-  const cleaned = content
-    .replace(THINK_TAG_RE, (_, reasoning: string) => {
-      const normalized = reasoning.trim();
-      if (normalized) {
-        reasoningParts.push(normalized);
-      }
-      return "";
-    })
-    .trim();
+  // 1) 先抽出所有**已闭合**的 <think>...</think>。
+  let working = content.replace(THINK_TAG_RE, (_, reasoning: string) => {
+    const normalized = reasoning.trim();
+    if (normalized) {
+      reasoningParts.push(normalized);
+    }
+    return "";
+  });
+
+  // 2) 流式中常见**未闭合**的 <think>（还没等到 </think> 就先到达）：把开标签之后的内容
+  //    当作「实时思考」抽到 reasoning、从答案正文里移除。否则模型慢/先思考时，用户既会看到
+  //    字面 "<think>"，又会因答案区长时间空白而以为卡死。这样思考内容会实时流进思考区。
+  const openIdx = working.indexOf(THINK_OPEN);
+  if (openIdx !== -1 && !working.slice(openIdx).includes("</think>")) {
+    const partial = working.slice(openIdx + THINK_OPEN.length).trim();
+    if (partial) {
+      reasoningParts.push(partial);
+    }
+    working = working.slice(0, openIdx);
+  }
 
   return {
-    content: cleaned,
+    content: working.trim(),
     reasoning: reasoningParts.length > 0 ? reasoningParts.join("\n\n") : null,
   };
 }

frontend-web/src/open-canvas/components/ai-writing/AIWritingForm.tsx

@@ -65,6 +65,7 @@ export function AIWritingForm({
     audience, setAudience,
     wordCountInput, setWordCountInput,
     strictMode, setStrictMode,
+    addReferences, setAddReferences,
     materialSource, setMaterialSource,
     // 笔记来源（search_space_id + document_id 列表）已上提到 WritingFormContext，
     // 以便左栏「笔记本」点击可直接设为素材来源（需求 0616 #5）。
@@ -164,6 +165,7 @@ export function AIWritingForm({
       wordCountTarget: wordCount,
       keywordCount: DEFAULT_KEYWORD_COUNT,
       writingMode: hideAdvancedFields ? 'loose' : strictMode ? 'strict' : 'loose',
+      addReferences,
       materialSource: materialSource,
       modelName: selectedModel?.name || modelName || undefined,
       // 只有选了"我的空间"才传笔记参数
@@ -381,6 +383,23 @@ export function AIWritingForm({
           </Select>
         </div>
 
+        <div className="mt-4 flex items-start gap-2 rounded-xl border border-[#e8dedd] bg-[#f9f5f5] px-3 py-3 dark:border-[#1B477E] dark:bg-[#022248]/40">
+          <input
+            type="checkbox"
+            id="add-references"
+            checked={addReferences}
+            onChange={(e) => setAddReferences(e.target.checked)}
+            disabled={disabled}
+            className="mt-0.5 size-3.5 shrink-0 cursor-pointer accent-primary"
+          />
+          <label htmlFor="add-references" className="cursor-pointer select-none">
+            <span className="text-sm font-medium text-foreground/90">添加参考文献</span>
+            <p className="mt-0.5 text-xs leading-relaxed text-muted-foreground/80">
+              开启后，正文引用素材处自动加角标 [n]，文末按出现顺序生成「参考文献」列表（像论文一样）。
+            </p>
+          </label>
+        </div>
+
         <Button
           type="submit"
           disabled={disabled || !intent.trim()}

frontend-web/src/open-canvas/components/ai-writing/AIWritingPanel.tsx

@@ -41,6 +41,7 @@ import {
  * 把这个 flag 翻成 true 即可，UI / hook 全套代码都已就位。
  */
 const ENABLE_PAUSE_RESUME_UI = false;
+import { countArticleWords } from '../../utils/wordCount';
 import { AIWritingForm } from './AIWritingForm';
 import { AIWritingTimeline } from './AIWritingTimeline';
 import { InterventionCard } from './InterventionCard';
@@ -482,6 +483,8 @@ export function AIWritingPanel({
                   outline={state.outline}
                   reviewResult={state.reviewResult}
                   draftTitle={state.currentDraftTitle}
+                  draftWordCount={countArticleWords(state.currentDraftMarkdown)}
+                  wordCountTarget={state.request?.wordCountTarget}
                   revisionCount={state.revisionCount}
                   skillSearchSteps={state.skillSearchSteps}
                   skillSearchInProgress={state.status === 'researching' && state.skillSearchSteps.length > 0}

frontend-web/src/open-canvas/components/ai-writing/InterventionCard.tsx

@@ -24,6 +24,10 @@ interface InterventionCardProps {
   outline: AIOutline | null;
   reviewResult: ReviewResult | null;
   draftTitle: string;
+  /** 当前草稿实际字数（与目标对比展示，让用户在确认时即看到字数是否达标） */
+  draftWordCount?: number;
+  /** 用户设定的目标字数 */
+  wordCountTarget?: number;
   revisionCount: number;
   skillSearchSteps?: SkillSearchStep[];
   /** 是否正在检索中（researching/awaiting 切换之间，前端用于流式展示步骤进行中） */
@@ -34,7 +38,7 @@ interface InterventionCardProps {
 
 export function InterventionCard({
   pauseEvent, materials, outline, reviewResult,
-  draftTitle, revisionCount, skillSearchSteps, skillSearchInProgress, onSubmit, disabled,
+  draftTitle, draftWordCount, wordCountTarget, revisionCount, skillSearchSteps, skillSearchInProgress, onSubmit, disabled,
 }: InterventionCardProps) {
   const pausePoint = pauseEvent.pausePoint;
 
@@ -64,6 +68,8 @@ export function InterventionCard({
     return (
       <DraftConfirmCard
         draftTitle={draftTitle}
+        draftWordCount={draftWordCount}
+        wordCountTarget={wordCountTarget}
         revisionCount={revisionCount}
         reviewResult={reviewResult}
         onSubmit={onSubmit}
@@ -357,13 +363,15 @@ function OutlineConfirmCard({ outline: initialOutline, onSubmit, disabled }: Out
 
 interface DraftConfirmCardProps {
   draftTitle: string;
+  draftWordCount?: number;
+  wordCountTarget?: number;
   revisionCount: number;
   reviewResult: ReviewResult | null;
   onSubmit: (payload: UserInterventionPayload) => void;
   disabled?: boolean;
 }
 
-function DraftConfirmCard({ draftTitle, revisionCount, reviewResult, onSubmit, disabled }: DraftConfirmCardProps) {
+function DraftConfirmCard({ draftTitle, draftWordCount, wordCountTarget, revisionCount, reviewResult, onSubmit, disabled }: DraftConfirmCardProps) {
   const [notes, setNotes] = useState('');
   const [showNotes, setShowNotes] = useState(false);
   const [submitting, setSubmitting] = useState(false);
@@ -399,6 +407,11 @@ function DraftConfirmCard({ draftTitle, revisionCount, reviewResult, onSubmit, d
 
   const hasReview = reviewResult != null && reviewResult.revisionNotes.length > 0;
 
+  // 实际字数 vs 目标字数对比：让用户在确认时直接看到是否达标，偏差超 100 字标橙提醒
+  const showWordCmp = wordCountTarget != null && wordCountTarget > 0 && draftWordCount != null && draftWordCount > 0;
+  const wordDiff = showWordCmp ? draftWordCount! - wordCountTarget! : 0;
+  const wordWithin = Math.abs(wordDiff) <= 100;
+
   return (
     <div className="rounded-md border bg-background space-y-3 p-3">
       <div className="flex items-start gap-2">
@@ -410,6 +423,12 @@ function DraftConfirmCard({ draftTitle, revisionCount, reviewResult, onSubmit, d
             {revisionCount > 0 ? `（第 ${revisionCount} 次修改稿）` : '（初稿）'}
             · 右侧可预览完整内容
           </p>
+          {showWordCmp && (
+            <p className={cn('text-xs mt-0.5', wordWithin ? 'text-muted-foreground' : 'text-orange-600 dark:text-orange-400')}>
+              实际 {draftWordCount} 字 / 目标 {wordCountTarget} 字
+              {wordDiff !== 0 && `（${wordDiff > 0 ? '超出' : '不足'} ${Math.abs(wordDiff)} 字${wordWithin ? '，在误差范围内' : ''}）`}
+            </p>
+          )}
         </div>
       </div>
 

frontend-web/src/open-canvas/components/ai-writing/OutlineCard.tsx

@@ -1,7 +1,21 @@
 import { useState } from 'react';
-import { ListTree, Plus } from 'lucide-react';
+import { ListTree, Plus, GripVertical } from 'lucide-react';
+import {
+  DndContext,
+  DragOverlay,
+  PointerSensor,
+  pointerWithin,
+  rectIntersection,
+  useSensor,
+  useSensors,
+  type CollisionDetection,
+  type DragEndEvent,
+  type DragStartEvent,
+} from '@dnd-kit/core';
 import { Button } from '@/components/ui/button';
-import type { AIOutline, OutlineSection as OutlineSectionType } from '../../types/ai-writing';
+import { cn } from '@/lib/utils';
+import type { AIOutline, KeyPoint, OutlineSection as OutlineSectionType } from '../../types/ai-writing';
+import { normalizeKeyPoint } from '../../types/ai-writing';
 import { OutlineSection } from './OutlineSection';
 import { CollapseTransition } from './CollapseTransition';
 
@@ -11,15 +25,177 @@ interface OutlineCardProps {
   readOnly?: boolean;
 }
 
+// 节点路径（章节/小节/子要点）
+export type DragPath =
+  | { type: 'section'; s: number }
+  | { type: 'keypoint'; s: number; k: number }
+  | { type: 'subpoint'; s: number; k: number; p: number };
+
+// 节点 id：s{S} / s{S}-k{K} / s{S}-k{K}-p{P}
+export function parseItemId(id: string): DragPath | null {
+  const m = id.match(/^s(\d+)(?:-k(\d+))?(?:-p(\d+))?$/);
+  if (!m) return null;
+  const s = parseInt(m[1], 10);
+  const k = m[2] !== undefined ? parseInt(m[2], 10) : undefined;
+  const p = m[3] !== undefined ? parseInt(m[3], 10) : undefined;
+  if (p !== undefined && k !== undefined) return { type: 'subpoint', s, k, p };
+  if (k !== undefined) return { type: 'keypoint', s, k };
+  return { type: 'section', s };
+}
+
+// 区域 id：before:<节点id> / after:<节点id> / into:<节点id>
+type Zone = 'before' | 'after' | 'into';
+function parseZoneId(zoneId: string): { zone: Zone; targetPath: DragPath; targetItemId: string } | null {
+  const m = zoneId.match(/^(before|after|into):(.+)$/);
+  if (!m) return null;
+  const targetItemId = m[2];
+  const targetPath = parseItemId(targetItemId);
+  if (!targetPath) return null;
+  return { zone: m[1] as Zone, targetPath, targetItemId };
+}
+
+function cloneOutline(outline: AIOutline): AIOutline {
+  return {
+    ...outline,
+    sections: outline.sections.map((s) => ({
+      ...s,
+      keyPoints: (s.keyPoints ?? []).map((kp) => {
+        const n = normalizeKeyPoint(kp);
+        return { ...n, subPoints: [...n.subPoints] } as KeyPoint;
+      }),
+      materialIds: [...(s.materialIds ?? [])],
+    })),
+  };
+}
+
+type SourceValue = OutlineSectionType | KeyPoint | string;
+
+// 拖到自身子树内非法（章节扔进自己的小节/子要点 / 小节扔进自己的子要点）
+function isDescendantTarget(from: DragPath, to: DragPath): boolean {
+  if (from.type === 'section') return to.s === from.s;
+  if (from.type === 'keypoint') return to.type !== 'section' && to.s === from.s && to.k === from.k;
+  return false;
+}
+
+function readAt(outline: AIOutline, p: DragPath): SourceValue {
+  if (p.type === 'section') return outline.sections[p.s];
+  const kps = outline.sections[p.s].keyPoints as KeyPoint[];
+  if (p.type === 'keypoint') return kps[p.k];
+  return kps[p.k].subPoints[p.p];
+}
+
+function removeAt(outline: AIOutline, p: DragPath): void {
+  if (p.type === 'section') { outline.sections.splice(p.s, 1); return; }
+  const kps = outline.sections[p.s].keyPoints as KeyPoint[];
+  if (p.type === 'keypoint') { kps.splice(p.k, 1); return; }
+  kps[p.k].subPoints.splice(p.p, 1);
+}
+
+function insertAt(outline: AIOutline, p: DragPath, value: SourceValue): void {
+  if (p.type === 'section') { outline.sections.splice(p.s, 0, value as OutlineSectionType); return; }
+  const kps = outline.sections[p.s].keyPoints as KeyPoint[];
+  if (p.type === 'keypoint') { kps.splice(p.k, 0, value as KeyPoint); return; }
+  kps[p.k].subPoints.splice(p.p, 0, value as string);
+}
+
+// 删源后，同一父列表里 target index 会前移一位
+function adjustTargetAfterRemove(to: DragPath, from: DragPath): DragPath {
+  if (from.type === 'section') {
+    if (to.s > from.s) return { ...to, s: to.s - 1 } as DragPath;
+    return to;
+  }
+  if (from.type === 'keypoint') {
+    if (to.type !== 'section' && to.s === from.s && to.k > from.k) {
+      return { ...to, k: to.k - 1 } as DragPath;
+    }
+    return to;
+  }
+  if (to.type === 'subpoint' && to.s === from.s && to.k === from.k && to.p > from.p) {
+    return { ...to, p: to.p - 1 };
+  }
+  return to;
+}
+
+type Level = DragPath['type'];
+
+function convertToLevel(value: SourceValue, fromLevel: Level, toLevel: Level): SourceValue {
+  if (fromLevel === toLevel) return value;
+  if (toLevel === 'section') {
+    if (fromLevel === 'keypoint') {
+      const kp = value as KeyPoint;
+      return {
+        sectionTitle: kp.pointTitle,
+        keyPoints: kp.subPoints.map((sp) => ({ pointTitle: sp, subPoints: [] })),
+        materialIds: [],
+      };
+    }
+    return { sectionTitle: value as string, keyPoints: [], materialIds: [] };
+  }
+  if (toLevel === 'keypoint') {
+    if (fromLevel === 'section') {
+      const sec = value as OutlineSectionType;
+      const sub = (sec.keyPoints ?? [])
+        .map((kp) => normalizeKeyPoint(kp).pointTitle)
+        .filter((s) => !!s);
+      return { pointTitle: sec.sectionTitle, subPoints: sub };
+    }
+    return { pointTitle: value as string, subPoints: [] };
+  }
+  if (fromLevel === 'section') return (value as OutlineSectionType).sectionTitle;
+  return (value as KeyPoint).pointTitle;
+}
+
+// into 区域：作为 target 节点的「末尾子项」插入
+function childPathOf(outline: AIOutline, target: DragPath): DragPath | null {
+  if (target.type === 'section') {
+    const kps = outline.sections[target.s]?.keyPoints ?? [];
+    return { type: 'keypoint', s: target.s, k: kps.length };
+  }
+  if (target.type === 'keypoint') {
+    const kp = normalizeKeyPoint((outline.sections[target.s]?.keyPoints ?? [])[target.k]);
+    return { type: 'subpoint', s: target.s, k: target.k, p: kp.subPoints.length };
+  }
+  return null; // 子要点没有子节点
+}
+
+// before/after：作为 target 节点的兄弟插入
+function siblingPathOf(target: DragPath, side: 'before' | 'after'): DragPath {
+  const delta = side === 'after' ? 1 : 0;
+  if (target.type === 'section') return { type: 'section', s: target.s + delta };
+  if (target.type === 'keypoint') return { type: 'keypoint', s: target.s, k: target.k + delta };
+  return { type: 'subpoint', s: target.s, k: target.k, p: target.p + delta };
+}
+
+// 自定义碰撞：pointerWithin 命中所有包含光标的 droppable，取 rect 最小的（最内层）
+// —— 这是嵌套区域 (section > keypoint > subpoint，每层又分 before/into/after) 能
+// 正确选中「最深的那个」的关键。
+function getRectArea(rect: { width: number; height: number } | null | undefined): number {
+  if (!rect) return Infinity;
+  return rect.width * rect.height;
+}
+
+const innermostCollision: CollisionDetection = (args) => {
+  const pointer = pointerWithin(args);
+  if (pointer.length === 0) return rectIntersection(args);
+  let best = pointer[0];
+  let bestArea = getRectArea(args.droppableRects.get(best.id));
+  for (const c of pointer.slice(1)) {
+    const a = getRectArea(args.droppableRects.get(c.id));
+    if (a < bestArea) { best = c; bestArea = a; }
+  }
+  return [best];
+};
+
 export function OutlineCard({ outline, onChange, readOnly = false }: OutlineCardProps) {
   const [collapsed, setCollapsed] = useState(false);
   const [editingTitle, setEditingTitle] = useState(false);
   const [titleDraft, setTitleDraft] = useState(outline.title);
-  const [dragIndex, setDragIndex] = useState<number | null>(null);
-  const [overIndex, setOverIndex] = useState<number | null>(null);
+  const [activeId, setActiveId] = useState<string | null>(null);
 
   const editable = !readOnly && !!onChange;
 
+  const sensors = useSensors(useSensor(PointerSensor, { activationConstraint: { distance: 4 } }));
+
   function commitTitle() {
     setEditingTitle(false);
     const trimmed = titleDraft.trim();
@@ -40,14 +216,6 @@ export function OutlineCard({ outline, onChange, readOnly = false }: OutlineCard
     onChange?.({ ...outline, sections: outline.sections.filter((_, idx) => idx !== i) });
   }
 
-  function reorderSections(from: number, to: number) {
-    if (from === to) return;
-    const sections = [...outline.sections];
-    const [moved] = sections.splice(from, 1);
-    sections.splice(to, 0, moved);
-    onChange?.({ ...outline, sections });
-  }
-
   function addSection() {
     onChange?.({
       ...outline,
@@ -55,9 +223,66 @@ export function OutlineCard({ outline, onChange, readOnly = false }: OutlineCard
     });
   }
 
+  function handleDragStart(event: DragStartEvent) {
+    setActiveId(String(event.active.id));
+  }
+  function handleDragCancel() {
+    setActiveId(null);
+  }
+
+  function handleDragEnd(event: DragEndEvent) {
+    setActiveId(null);
+    const { active, over } = event;
+    if (!over || !editable) return;
+
+    const fromId = String(active.id);
+    const overId = String(over.id);
+
+    const from = parseItemId(fromId);
+    const zoneInfo = parseZoneId(overId);
+    if (!from || !zoneInfo) return;
+    const { zone, targetPath, targetItemId } = zoneInfo;
+
+    // 落在源自身的任何区域 → 无操作
+    if (targetItemId === fromId) return;
+
+    // 计算插入路径：into = 末尾子项；before/after = 同级兄弟
+    const insertPath = zone === 'into'
+      ? childPathOf(outline, targetPath)
+      : siblingPathOf(targetPath, zone);
+    if (!insertPath) return;
+
+    // 拖入源自身子树 → 非法
+    if (isDescendantTarget(from, insertPath)) return;
+
+    const next = cloneOutline(outline);
+    const sourceValue = readAt(next, from);
+    removeAt(next, from);
+    const adjusted = adjustTargetAfterRemove(insertPath, from);
+    const converted = convertToLevel(sourceValue, from.type, adjusted.type);
+    insertAt(next, adjusted, converted);
+
+    onChange?.(next);
+  }
+
+  function renderDragOverlay() {
+    if (!activeId) return null;
+    const path = parseItemId(activeId);
+    if (!path) return null;
+    const section = outline.sections[path.s];
+    if (!section) return null;
+    if (path.type === 'section') {
+      return <OverlayCard level="section" label={section.sectionTitle} />;
+    }
+    const kp = normalizeKeyPoint((section.keyPoints ?? [])[path.k]);
+    if (path.type === 'keypoint') {
+      return <OverlayCard level="keypoint" label={kp.pointTitle} />;
+    }
+    return <OverlayCard level="subpoint" label={kp.subPoints?.[path.p] ?? ''} />;
+  }
+
   return (
     <div className="rounded-md border bg-muted/20 overflow-hidden">
-      {/* 头部 */}
       <div className="flex items-center gap-2 px-3 py-2 border-b">
         <ListTree className="size-3.5 text-purple-500 shrink-0" />
         <span className="text-xs font-medium flex-1">文章大纲</span>
@@ -97,34 +322,30 @@ export function OutlineCard({ outline, onChange, readOnly = false }: OutlineCard
             )}
           </div>
 
-          {/* 章节列表 */}
-          <div className="space-y-1.5">
-            {outline.sections.map((section, i) => (
-              <OutlineSection
-                key={i}
-                section={section}
-                index={i}
-                onChange={(updated) => updateSection(i, updated)}
-                onRemove={editable ? () => removeSection(i) : undefined}
-                readOnly={!editable}
-                isDragging={dragIndex === i}
-                isOver={overIndex === i && dragIndex !== null && dragIndex !== i}
-                onDragStart={() => setDragIndex(i)}
-                onDragEnter={() => setOverIndex(i)}
-                onDrop={() => {
-                  if (dragIndex !== null && dragIndex !== i) reorderSections(dragIndex, i);
-                  setDragIndex(null);
-                  setOverIndex(null);
-                }}
-                onDragEnd={() => {
-                  setDragIndex(null);
-                  setOverIndex(null);
-                }}
-              />
-            ))}
-          </div>
+          {/* 章节列表 —— 三区域拖放：顶/底边重排，中间嵌入为子项 */}
+          <DndContext
+            sensors={sensors}
+            collisionDetection={innermostCollision}
+            onDragStart={handleDragStart}
+            onDragEnd={handleDragEnd}
+            onDragCancel={handleDragCancel}
+          >
+            <div className="space-y-1.5">
+              {outline.sections.map((section, i) => (
+                <OutlineSection
+                  key={i}
+                  section={section}
+                  index={i}
+                  sectionIndex={i}
+                  onChange={(updated) => updateSection(i, updated)}
+                  onRemove={editable ? () => removeSection(i) : undefined}
+                  readOnly={!editable}
+                />
+              ))}
+            </div>
+            <DragOverlay dropAnimation={null}>{renderDragOverlay()}</DragOverlay>
+          </DndContext>
 
-          {/* 添加章节 */}
           {editable && (
             <Button
               type="button"
@@ -142,3 +363,23 @@ export function OutlineCard({ outline, onChange, readOnly = false }: OutlineCard
     </div>
   );
 }
+
+// 拖动时跟随光标的浮动预览。DragOverlay 把这个节点 portal 到 body，
+// 不受任何祖先 overflow:hidden 影响，所以拖出卡片也始终可见。
+function OverlayCard({ level, label }: { level: Level; label: string }) {
+  const styleByLevel: Record<Level, string> = {
+    section: 'rounded border-2 border-blue-400 bg-background text-xs font-medium px-2.5 py-2 shadow-lg',
+    keypoint: 'rounded border-2 border-blue-400 bg-muted/40 text-xs px-2 py-1.5 shadow-lg',
+    subpoint: 'rounded border-2 border-blue-400 bg-background text-xs px-2 py-1 shadow-lg',
+  };
+  return (
+    <div className={cn('flex items-center gap-1.5 cursor-grabbing min-w-[180px] max-w-[420px]', styleByLevel[level])}>
+      <GripVertical className="size-3 text-muted-foreground/50 shrink-0" />
+      {level === 'keypoint' && <span className="text-purple-500/70 shrink-0">###</span>}
+      {level === 'subpoint' && <span className="text-pink-500/70 shrink-0 text-[10px] font-mono">####</span>}
+      <span className="truncate">
+        {label || <span className="italic text-muted-foreground">（未命名）</span>}
+      </span>
+    </div>
+  );
+}

frontend-web/src/open-canvas/components/ai-writing/OutlineSection.tsx

@@ -1,5 +1,6 @@
 import { useState, useMemo } from 'react';
 import { ChevronRight, Plus, X, GripVertical } from 'lucide-react';
+import { useDraggable, useDroppable } from '@dnd-kit/core';
 import { cn } from '@/lib/utils';
 import type { KeyPoint, OutlineSection as OutlineSectionType } from '../../types/ai-writing';
 import { normalizeKeyPoint } from '../../types/ai-writing';
@@ -8,36 +9,34 @@ import { CollapseTransition } from './CollapseTransition';
 interface OutlineSectionProps {
   section: OutlineSectionType;
   index: number;
+  /** 章节在 outline.sections 中的下标，用于生成稳定 dnd id */
+  sectionIndex: number;
   onChange: (updated: OutlineSectionType) => void;
   onRemove?: () => void;
   readOnly?: boolean;
-  // Drag-and-drop (章节级别)
-  isDragging?: boolean;
-  isOver?: boolean;
-  onDragStart?: () => void;
-  onDragEnter?: () => void;
-  onDrop?: () => void;
-  onDragEnd?: () => void;
 }
 
-/**
- * 章节卡片(Level 1)。内部渲染 keyPoints[] 为「小节子卡片」(Level 2),小节里再渲染
- * subPoints[] 为 bullet 列表(Level 3)。
- *
- * 兼容旧 2 级大纲:section.keyPoints 可能是 (string | KeyPoint)[];进入组件前通过
- * normalizeKeyPoint 统一成 KeyPoint[],对外输出也一律是 KeyPoint[](自然升级,用户
- * 编辑过的旧 session 大纲会被持久化为新格式)。
- */
+// 各级 grip 标题提示文案
+const DRAG_HINT = '按住拖动：项的顶/底边 = 同级重排；中间区域 = 嵌入为子项';
+const DRAG_HINT_LEAF = '按住拖动重排（顶半区 = 上插入，底半区 = 下插入）';
+
+// ─── 章节 (Level 1) ───────────────────────────────────────────────────────────
 export function OutlineSection({
-  section, index, onChange, onRemove, readOnly,
-  isDragging, isOver, onDragStart, onDragEnter, onDrop, onDragEnd,
+  section, index, sectionIndex, onChange, onRemove, readOnly,
 }: OutlineSectionProps) {
-  const [expanded, setExpanded] = useState(false);
+  const [expanded, setExpanded] = useState(true);
   const [editingTitle, setEditingTitle] = useState(false);
   const [titleDraft, setTitleDraft] = useState(section.sectionTitle);
 
-  // 把 keyPoints 统一成 KeyPoint[];旧 string 元素会被 lift。useMemo 避免每次 render
-  // 都生成新引用导致子组件无谓重渲染。
+  const id = `s${sectionIndex}`;
+  const { attributes, listeners, setNodeRef: setDragRef, isDragging } = useDraggable({
+    id, disabled: readOnly,
+  });
+  // 三个 drop 区：before / into / after
+  const beforeDrop = useDroppable({ id: `before:${id}`, disabled: readOnly });
+  const intoDrop = useDroppable({ id: `into:${id}`, disabled: readOnly });
+  const afterDrop = useDroppable({ id: `after:${id}`, disabled: readOnly });
+
   const keyPoints: KeyPoint[] = useMemo(
     () => (section.keyPoints ?? []).map(normalizeKeyPoint),
     [section.keyPoints],
@@ -52,128 +51,153 @@ export function OutlineSection({
       setTitleDraft(section.sectionTitle);
     }
   }
-
   function updateKeyPoint(i: number, updated: KeyPoint) {
     const next = [...keyPoints];
     next[i] = updated;
     onChange({ ...section, keyPoints: next });
   }
-
   function removeKeyPoint(i: number) {
     onChange({ ...section, keyPoints: keyPoints.filter((_, idx) => idx !== i) });
   }
-
   function addKeyPoint() {
     onChange({ ...section, keyPoints: [...keyPoints, { pointTitle: '', subPoints: [] }] });
   }
 
-  // 子要点总数,显示在章节头给个粗略概览
   const totalSubPoints = keyPoints.reduce((sum, kp) => sum + (kp.subPoints?.length || 0), 0);
 
   return (
     <div
-      draggable={!readOnly}
-      onDragStart={(e) => { e.stopPropagation(); onDragStart?.(); }}
-      onDragOver={(e) => { e.preventDefault(); e.stopPropagation(); onDragEnter?.(); }}
-      onDrop={(e) => { e.preventDefault(); e.stopPropagation(); onDrop?.(); }}
-      onDragEnd={(e) => { e.stopPropagation(); onDragEnd?.(); }}
+      ref={setDragRef}
+      {...attributes}
       className={cn(
-        'rounded border bg-background text-xs transition-all duration-150 select-none',
-        isDragging ? 'opacity-40 border-border/60' : isOver ? 'border-blue-400 ring-1 ring-blue-400/40' : 'border-border/60',
+        'relative rounded border text-xs transition-colors duration-150',
+        isDragging ? 'opacity-40 border-blue-300' : 'border-border/60',
+        intoDrop.isOver ? 'ring-2 ring-blue-400 ring-inset bg-blue-50/40 dark:bg-blue-950/30' : 'bg-background',
       )}
     >
-      {/* 章节头 */}
-      <div className="flex items-center gap-1.5 px-2.5 py-2">
-        {!readOnly && (
-          <GripVertical className="size-3 text-muted-foreground/40 shrink-0 cursor-grab active:cursor-grabbing" />
-        )}
-        <button
-          type="button"
-          onClick={() => setExpanded((v) => !v)}
-          className="shrink-0"
-        >
-          <ChevronRight
-            className={cn(
-              'size-3.5 text-muted-foreground transition-transform duration-200',
-              expanded && 'rotate-90',
-            )}
-          />
-        </button>
+      {/* before：顶部 12px 命中区，命中时顶边显示蓝色线 */}
+      <div ref={beforeDrop.setNodeRef} className="absolute top-0 inset-x-0 h-3 z-20 pointer-events-none" />
+      {beforeDrop.isOver && <DropLine position="top" />}
 
-        <span className="text-muted-foreground/60 shrink-0 w-4 text-center">{index + 1}.</span>
-
-        {editingTitle && !readOnly ? (
-          <input
-            autoFocus
-            value={titleDraft}
-            onChange={(e) => setTitleDraft(e.target.value)}
-            onBlur={commitTitle}
-            onKeyDown={(e) => {
-              if (e.key === 'Enter') commitTitle();
-              if (e.key === 'Escape') { setTitleDraft(section.sectionTitle); setEditingTitle(false); }
-            }}
-            className="flex-1 bg-transparent border-b border-primary outline-none font-medium"
-          />
-        ) : (
-          <span
-            className={cn('flex-1 font-medium', !readOnly && 'cursor-text hover:text-primary')}
-            onClick={() => !readOnly && setEditingTitle(true)}
-          >
-            {section.sectionTitle || <span className="text-muted-foreground italic">（未命名）</span>}
-          </span>
-        )}
-
-        <span className="text-muted-foreground/50 shrink-0">
-          {keyPoints.length} 小节{totalSubPoints > 0 ? ` · ${totalSubPoints} 子要点` : ''}
-        </span>
-        {!readOnly && onRemove && (
-          <button type="button" onClick={onRemove} className="shrink-0 text-muted-foreground/40 hover:text-destructive">
-            <X className="size-3" />
-          </button>
-        )}
-      </div>
-
-      {/* 小节列表 (Level 2) */}
-      <CollapseTransition open={expanded}>
-        <div className="px-6 pb-2.5 pt-1 space-y-1.5 border-t">
-          {keyPoints.map((kp, i) => (
-            <KeyPointRow
-              key={i}
-              keyPoint={kp}
-              onChange={(updated) => updateKeyPoint(i, updated)}
-              onRemove={readOnly ? undefined : () => removeKeyPoint(i)}
-              readOnly={readOnly}
-            />
-          ))}
+      {/* into：包住头部 + 折叠体；本身就是 drop 容器 */}
+      <div ref={intoDrop.setNodeRef}>
+        {/* 章节头 */}
+        <div className="flex items-center gap-1.5 px-2.5 py-2">
           {!readOnly && (
             <button
               type="button"
-              onClick={addKeyPoint}
-              className="flex items-center gap-1 text-muted-foreground/60 hover:text-primary pt-0.5"
+              {...listeners}
+              className="shrink-0 cursor-grab active:cursor-grabbing touch-none p-0.5 -m-0.5 text-muted-foreground/40 hover:text-muted-foreground"
+              aria-label="拖动排序章节"
+              title={DRAG_HINT}
             >
-              <Plus className="size-3" />
-              添加小节
+              <GripVertical className="size-3" />
+            </button>
+          )}
+          <button
+            type="button"
+            onClick={() => setExpanded((v) => !v)}
+            className="shrink-0"
+          >
+            <ChevronRight
+              className={cn(
+                'size-3.5 text-muted-foreground transition-transform duration-200',
+                expanded && 'rotate-90',
+              )}
+            />
+          </button>
+
+          <span className="text-muted-foreground/60 shrink-0 w-4 text-center">{index + 1}.</span>
+
+          {editingTitle && !readOnly ? (
+            <input
+              autoFocus
+              value={titleDraft}
+              onChange={(e) => setTitleDraft(e.target.value)}
+              onBlur={commitTitle}
+              onKeyDown={(e) => {
+                if (e.key === 'Enter') commitTitle();
+                if (e.key === 'Escape') { setTitleDraft(section.sectionTitle); setEditingTitle(false); }
+              }}
+              className="flex-1 bg-transparent border-b border-primary outline-none font-medium"
+            />
+          ) : (
+            <span
+              className={cn('flex-1 font-medium', !readOnly && 'cursor-text hover:text-primary')}
+              onClick={() => !readOnly && setEditingTitle(true)}
+            >
+              {section.sectionTitle || <span className="text-muted-foreground italic">（未命名）</span>}
+            </span>
+          )}
+
+          <span className="text-muted-foreground/50 shrink-0">
+            {keyPoints.length} 小节{totalSubPoints > 0 ? ` · ${totalSubPoints} 子要点` : ''}
+          </span>
+          {!readOnly && onRemove && (
+            <button type="button" onClick={onRemove} className="shrink-0 text-muted-foreground/40 hover:text-destructive">
+              <X className="size-3" />
             </button>
           )}
         </div>
-      </CollapseTransition>
+
+        {/* 小节列表 (Level 2) */}
+        <CollapseTransition open={expanded}>
+          <div className="px-6 pb-2.5 pt-1 space-y-1.5 border-t">
+            {keyPoints.map((kp, i) => (
+              <KeyPointRow
+                key={i}
+                keyPoint={kp}
+                sectionIndex={sectionIndex}
+                keyPointIndex={i}
+                onChange={(updated) => updateKeyPoint(i, updated)}
+                onRemove={readOnly ? undefined : () => removeKeyPoint(i)}
+                readOnly={readOnly}
+              />
+            ))}
+            {!readOnly && (
+              <button
+                type="button"
+                onClick={addKeyPoint}
+                className="flex items-center gap-1 text-muted-foreground/60 hover:text-primary pt-0.5"
+              >
+                <Plus className="size-3" />
+                添加小节
+              </button>
+            )}
+          </div>
+        </CollapseTransition>
+      </div>
+
+      {/* after：底部 12px 命中区 */}
+      <div ref={afterDrop.setNodeRef} className="absolute bottom-0 inset-x-0 h-3 z-20 pointer-events-none" />
+      {afterDrop.isOver && <DropLine position="bottom" />}
     </div>
   );
 }
 
-// ─── 小节行(Level 2,内嵌 Level 3 子要点) ───────────────────────────────────────
+// ─── 小节 (Level 2) ───────────────────────────────────────────────────────────
 interface KeyPointRowProps {
   keyPoint: KeyPoint;
+  sectionIndex: number;
+  keyPointIndex: number;
   onChange: (updated: KeyPoint) => void;
   onRemove?: () => void;
   readOnly?: boolean;
 }
 
-function KeyPointRow({ keyPoint, onChange, onRemove, readOnly }: KeyPointRowProps) {
-  const [expanded, setExpanded] = useState(false);
+function KeyPointRow({ keyPoint, sectionIndex, keyPointIndex, onChange, onRemove, readOnly }: KeyPointRowProps) {
+  const [expanded, setExpanded] = useState(true);
   const [editingTitle, setEditingTitle] = useState(false);
   const [titleDraft, setTitleDraft] = useState(keyPoint.pointTitle);
 
+  const id = `s${sectionIndex}-k${keyPointIndex}`;
+  const { attributes, listeners, setNodeRef: setDragRef, isDragging } = useDraggable({
+    id, disabled: readOnly,
+  });
+  const beforeDrop = useDroppable({ id: `before:${id}`, disabled: readOnly });
+  const intoDrop = useDroppable({ id: `into:${id}`, disabled: readOnly });
+  const afterDrop = useDroppable({ id: `after:${id}`, disabled: readOnly });
+
   function commitTitle() {
     setEditingTitle(false);
     const trimmed = titleDraft.trim();
@@ -181,120 +205,222 @@ function KeyPointRow({ keyPoint, onChange, onRemove, readOnly }: KeyPointRowProp
       onChange({ ...keyPoint, pointTitle: trimmed });
     }
   }
-
   function updateSubPoint(i: number, value: string) {
     const next = [...keyPoint.subPoints];
     next[i] = value;
     onChange({ ...keyPoint, subPoints: next });
   }
-
   function removeSubPoint(i: number) {
     onChange({ ...keyPoint, subPoints: keyPoint.subPoints.filter((_, idx) => idx !== i) });
   }
-
   function addSubPoint() {
     onChange({ ...keyPoint, subPoints: [...keyPoint.subPoints, ''] });
-    setExpanded(true);   // 添加时自动展开,免得用户看不到刚加的输入框
+    setExpanded(true);
   }
 
   return (
-    <div className="rounded border border-border/40 bg-muted/10">
-      {/* 小节头 */}
-      <div className="flex items-center gap-1.5 px-2 py-1.5">
-        <button
-          type="button"
-          onClick={() => setExpanded((v) => !v)}
-          className="shrink-0"
-          aria-label="展开/收起子要点"
-        >
-          <ChevronRight
-            className={cn(
-              'size-3 text-muted-foreground/60 transition-transform duration-200',
-              expanded && 'rotate-90',
-            )}
-          />
-        </button>
+    <div
+      ref={setDragRef}
+      {...attributes}
+      className={cn(
+        'relative rounded border border-border/40 transition-colors duration-150',
+        isDragging && 'opacity-40 border-blue-300',
+        intoDrop.isOver ? 'ring-2 ring-blue-400 ring-inset bg-blue-50/40 dark:bg-blue-950/30' : 'bg-muted/10',
+      )}
+    >
+      <div ref={beforeDrop.setNodeRef} className="absolute top-0 inset-x-0 h-2 z-20 pointer-events-none" />
+      {beforeDrop.isOver && <DropLine position="top" />}
 
-        <span className="text-purple-500/70 shrink-0">###</span>
-
-        {editingTitle && !readOnly ? (
-          <input
-            autoFocus
-            value={titleDraft}
-            onChange={(e) => setTitleDraft(e.target.value)}
-            onBlur={commitTitle}
-            onKeyDown={(e) => {
-              if (e.key === 'Enter') commitTitle();
-              if (e.key === 'Escape') { setTitleDraft(keyPoint.pointTitle); setEditingTitle(false); }
-            }}
-            className="flex-1 bg-transparent border-b border-primary outline-none"
-            placeholder="小节标题"
-          />
-        ) : (
-          <span
-            className={cn('flex-1', !readOnly && 'cursor-text hover:text-primary')}
-            onClick={() => !readOnly && setEditingTitle(true)}
-          >
-            {keyPoint.pointTitle || <span className="text-muted-foreground italic">（未命名小节）</span>}
-          </span>
-        )}
-
-        <span className="text-muted-foreground/40 shrink-0 text-[10px]">
-          {keyPoint.subPoints.length > 0 ? `${keyPoint.subPoints.length} 子要点` : '无子要点'}
-        </span>
-        {!readOnly && onRemove && (
-          <button
-            type="button"
-            onClick={onRemove}
-            className="shrink-0 text-muted-foreground/40 hover:text-destructive"
-            aria-label="删除小节"
-          >
-            <X className="size-3" />
-          </button>
-        )}
-      </div>
-
-      {/* 子要点 (Level 3) — 每个子要点会变成正文里的 #### 标题 + 一段独立论述 */}
-      <CollapseTransition open={expanded}>
-        <div className="px-6 pb-1.5 space-y-0.5 border-t border-border/30 bg-background/40">
-          {keyPoint.subPoints.map((sp, i) => (
-            <div key={i} className="flex items-center gap-1.5 pt-1">
-              {/* 用 #### 标记替代 bullet,提示这是「四级标题」而非简单列表项 */}
-              <span className="text-pink-500/70 shrink-0 text-[10px] font-mono">####</span>
-              {readOnly ? (
-                <span className="flex-1 text-muted-foreground">{sp}</span>
-              ) : (
-                <input
-                  value={sp}
-                  onChange={(e) => updateSubPoint(i, e.target.value)}
-                  placeholder="子要点标题(会展开成一段正文)"
-                  className="flex-1 bg-transparent border-b border-transparent hover:border-border focus:border-primary outline-none text-muted-foreground"
-                />
-              )}
-              {!readOnly && (
-                <button
-                  type="button"
-                  onClick={() => removeSubPoint(i)}
-                  className="shrink-0 text-muted-foreground/30 hover:text-destructive"
-                  aria-label="删除子要点"
-                >
-                  <X className="size-3" />
-                </button>
-              )}
-            </div>
-          ))}
+      <div ref={intoDrop.setNodeRef}>
+        {/* 小节头 */}
+        <div className="flex items-center gap-1.5 px-2 py-1.5">
           {!readOnly && (
             <button
               type="button"
-              onClick={addSubPoint}
-              className="flex items-center gap-1 text-muted-foreground/50 hover:text-primary pt-0.5 text-[11px]"
+              {...listeners}
+              className="shrink-0 cursor-grab active:cursor-grabbing touch-none p-0.5 -m-0.5 text-muted-foreground/30 hover:text-muted-foreground"
+              aria-label="拖动排序小节"
+              title={DRAG_HINT}
             >
-              <Plus className="size-3" />
-              添加子要点
+              <GripVertical className="size-3" />
+            </button>
+          )}
+          <button
+            type="button"
+            onClick={() => setExpanded((v) => !v)}
+            className="shrink-0"
+            aria-label="展开/收起子要点"
+          >
+            <ChevronRight
+              className={cn(
+                'size-3 text-muted-foreground/60 transition-transform duration-200',
+                expanded && 'rotate-90',
+              )}
+            />
+          </button>
+
+          <span className="text-purple-500/70 shrink-0">###</span>
+
+          {editingTitle && !readOnly ? (
+            <input
+              autoFocus
+              value={titleDraft}
+              onChange={(e) => setTitleDraft(e.target.value)}
+              onBlur={commitTitle}
+              onKeyDown={(e) => {
+                if (e.key === 'Enter') commitTitle();
+                if (e.key === 'Escape') { setTitleDraft(keyPoint.pointTitle); setEditingTitle(false); }
+              }}
+              className="flex-1 bg-transparent border-b border-primary outline-none"
+              placeholder="小节标题"
+            />
+          ) : (
+            <span
+              className={cn('flex-1', !readOnly && 'cursor-text hover:text-primary')}
+              onClick={() => !readOnly && setEditingTitle(true)}
+            >
+              {keyPoint.pointTitle || <span className="text-muted-foreground italic">（未命名小节）</span>}
+            </span>
+          )}
+
+          <span className="text-muted-foreground/40 shrink-0 text-[10px]">
+            {keyPoint.subPoints.length > 0 ? `${keyPoint.subPoints.length} 子要点` : '无子要点'}
+          </span>
+          {!readOnly && onRemove && (
+            <button
+              type="button"
+              onClick={onRemove}
+              className="shrink-0 text-muted-foreground/40 hover:text-destructive"
+              aria-label="删除小节"
+            >
+              <X className="size-3" />
             </button>
           )}
         </div>
-      </CollapseTransition>
+
+        {/* 子要点 (Level 3) */}
+        <CollapseTransition open={expanded}>
+          <div className="px-6 pb-1.5 space-y-0.5 border-t border-border/30 bg-background/40">
+            {keyPoint.subPoints.map((sp, i) => (
+              <SubPointRow
+                key={i}
+                value={sp}
+                sectionIndex={sectionIndex}
+                keyPointIndex={keyPointIndex}
+                subPointIndex={i}
+                onChange={(value) => updateSubPoint(i, value)}
+                onRemove={readOnly ? undefined : () => removeSubPoint(i)}
+                readOnly={readOnly}
+              />
+            ))}
+            {!readOnly && (
+              <button
+                type="button"
+                onClick={addSubPoint}
+                className="flex items-center gap-1 text-muted-foreground/50 hover:text-primary pt-0.5 text-[11px]"
+              >
+                <Plus className="size-3" />
+                添加子要点
+              </button>
+            )}
+          </div>
+        </CollapseTransition>
+      </div>
+
+      <div ref={afterDrop.setNodeRef} className="absolute bottom-0 inset-x-0 h-2 z-20 pointer-events-none" />
+      {afterDrop.isOver && <DropLine position="bottom" />}
     </div>
   );
 }
+
+// ─── 子要点 (Level 3，叶节点，无 into) ────────────────────────────────────────
+interface SubPointRowProps {
+  value: string;
+  sectionIndex: number;
+  keyPointIndex: number;
+  subPointIndex: number;
+  onChange: (value: string) => void;
+  onRemove?: () => void;
+  readOnly?: boolean;
+}
+
+function SubPointRow({
+  value, sectionIndex, keyPointIndex, subPointIndex, onChange, onRemove, readOnly,
+}: SubPointRowProps) {
+  const id = `s${sectionIndex}-k${keyPointIndex}-p${subPointIndex}`;
+  const { attributes, listeners, setNodeRef: setDragRef, isDragging } = useDraggable({
+    id, disabled: readOnly,
+  });
+  // 叶节点：没有 into；顶半区 = before、底半区 = after，把整行高度铺满
+  const beforeDrop = useDroppable({ id: `before:${id}`, disabled: readOnly });
+  const afterDrop = useDroppable({ id: `after:${id}`, disabled: readOnly });
+
+  return (
+    <div
+      ref={setDragRef}
+      {...attributes}
+      className={cn(
+        'relative flex items-center gap-1.5 pt-1 rounded transition-colors duration-150',
+        isDragging && 'opacity-40',
+      )}
+    >
+      <div
+        ref={beforeDrop.setNodeRef}
+        className="absolute left-0 right-0 top-0 bottom-1/2 z-20 pointer-events-none"
+      />
+      {beforeDrop.isOver && <DropLine position="top" thin />}
+
+      {!readOnly && (
+        <button
+          type="button"
+          {...listeners}
+          className="shrink-0 cursor-grab active:cursor-grabbing touch-none p-0.5 -m-0.5 text-muted-foreground/30 hover:text-muted-foreground"
+          aria-label="拖动排序子要点"
+          title={DRAG_HINT_LEAF}
+        >
+          <GripVertical className="size-2.5" />
+        </button>
+      )}
+      <span className="text-pink-500/70 shrink-0 text-[10px] font-mono">####</span>
+      {readOnly ? (
+        <span className="flex-1 text-muted-foreground">{value}</span>
+      ) : (
+        <input
+          value={value}
+          onChange={(e) => onChange(e.target.value)}
+          placeholder="子要点标题(会展开成一段正文)"
+          className="flex-1 bg-transparent border-b border-transparent hover:border-border focus:border-primary outline-none text-muted-foreground"
+        />
+      )}
+      {!readOnly && onRemove && (
+        <button
+          type="button"
+          onClick={onRemove}
+          className="shrink-0 text-muted-foreground/30 hover:text-destructive"
+          aria-label="删除子要点"
+        >
+          <X className="size-3" />
+        </button>
+      )}
+
+      <div
+        ref={afterDrop.setNodeRef}
+        className="absolute left-0 right-0 top-1/2 bottom-0 z-20 pointer-events-none"
+      />
+      {afterDrop.isOver && <DropLine position="bottom" thin />}
+    </div>
+  );
+}
+
+// ─── 蓝色插入指示线 ───────────────────────────────────────────────────────────
+function DropLine({ position, thin }: { position: 'top' | 'bottom'; thin?: boolean }) {
+  return (
+    <div
+      className={cn(
+        'absolute inset-x-1 bg-blue-500 rounded-full z-30 pointer-events-none',
+        thin ? 'h-[2px]' : 'h-[3px]',
+        position === 'top' ? '-top-px' : '-bottom-px',
+      )}
+    />
+  );
+}

frontend-web/src/open-canvas/components/ai-writing/WritingSetupChat.tsx

@@ -21,6 +21,7 @@ import {
   type SetupWritingApprovalPayload,
 } from '@/components/workspace/messages';
 import { ThreadContext } from '@/components/workspace/messages/context';
+import { extractContentFromMessage, extractReasoningContentFromMessage, hasToolCalls } from '@/core/messages/utils';
 import { getAPIClient } from '@/core/api';
 import { useModels } from '@/core/models/hooks';
 import { useThreadStream } from '@/core/threads/hooks';
@@ -29,6 +30,8 @@ import { cn } from '@/lib/utils';
 
 import { useArticleTypes } from '../../hooks/useArticleTypes';
 import { useWritingFormOptional } from '../../contexts/WritingFormContext';
+import { useAIWritingOptional } from '../../contexts/AIWritingContext';
+import type { AIWritingRequest } from '../../types/ai-writing';
 import { WritingWelcome } from './WritingWelcome';
 
 /**
@@ -61,6 +64,9 @@ export function WritingSetupChat({
   // 缺省回填写作表单；onApprove 传入时脱离 WritingFormProvider（如模板创建页）。
   const formCtx = useWritingFormOptional();
   const applyProposal = onApproveProp ?? formCtx?.applyProposal;
+  // 「直接开始写作」：仅在真正的写作上下文（AIWritingProvider 内、未被复用为模板编辑）提供。
+  const aiWriting = useAIWritingOptional();
+  const canStartWriting = !onApproveProp && !!aiWriting;
 
   // 首个模型可用时默认选中
   useEffect(() => {
@@ -116,9 +122,68 @@ export function WritingSetupChat({
     [applyProposal],
   );
 
+  // 「直接开始写作」：意图确认卡里已带上**用户已确认的大纲**（setup agent 在确认大纲后才提交），
+  // 这里把审批建议（含该大纲）直接组装成写作请求并启动 ai_writing 流程，同时回填表单保持一致。
+  const handleStart = useCallback(
+    async (payload: SetupWritingApprovalPayload) => {
+      if (!aiWriting) return;
+      applyProposal?.(payload);
+
+      const wordCount =
+        typeof payload.word_count_target === 'number' && payload.word_count_target > 0
+          ? payload.word_count_target
+          : Math.max(100, parseInt(formCtx?.wordCountInput ?? '', 10) || 800);
+      const materialSource = formCtx?.materialSource ?? 'general';
+
+      const request: AIWritingRequest = {
+        userIntent: payload.user_intent?.trim() || formCtx?.intent?.trim() || '',
+        // 已与用户确认的大纲：作为用户大纲传入，后端据此规划（留空则后端自动规划）
+        userOutline: payload.outline?.trim() || formCtx?.outlineText?.trim() || undefined,
+        articleType: payload.article_type?.trim() || formCtx?.articleType || 'analysis',
+        // 新（列表外）文章类型的一句话特征，仅本次写作临时使用（命中已有类型时为空）。
+        articleTypeDetail: payload.article_type_detail?.trim() || undefined,
+        targetAudience: payload.target_audience?.trim() || formCtx?.audience?.trim() || '普通读者',
+        wordCountTarget: wordCount,
+        keywordCount: 1,
+        writingMode: payload.writing_mode === 'strict' ? 'strict' : 'loose',
+        // 跟随表单的「添加参考文献」开关（默认开）。
+        addReferences: formCtx?.addReferences ?? true,
+        materialSource,
+        modelName: chatModel || formCtx?.modelName || undefined,
+        noteSearchSpaceId:
+          materialSource === 'notebook' ? (formCtx?.selectedNotebookId ?? undefined) : undefined,
+        noteSourceIds:
+          materialSource === 'notebook' && formCtx?.selectedNotebookId != null
+            ? formCtx.selectedDocumentIds
+            : undefined,
+      };
+
+      if (!request.userIntent) {
+        toast.error('写作意图为空，请先补充意图');
+        return;
+      }
+      await aiWriting.startWriting(request);
+    },
+    [aiWriting, applyProposal, formCtx, chatModel],
+  );
+
   // 对话还没有任何消息时显示欢迎/空状态屏（与原型图一致）
   const showWelcome = thread.messages.length === 0 && !thread.isLoading;
 
+  // 首 token 到达前的「正在整理…」占位：模型在跑（isLoading），但当前助手回合还没有任何
+  // 可见输出（无答案文本、无 <think> 思考、无工具调用）—— 这段纯网络/排队延迟下答案区空白，
+  // 加个轻量提示避免用户以为卡死（思考一旦开始流式，此提示自动消失，改由思考区接力）。
+  const showThinkingPlaceholder = useMemo(() => {
+    if (!thread.isLoading) return false;
+    const last = thread.messages[thread.messages.length - 1];
+    if (!last || last.type !== 'ai') return true; // 还停在用户消息，AI 尚未开口
+    const hasOutput =
+      !!extractContentFromMessage(last).trim() ||
+      !!extractReasoningContentFromMessage(last) ||
+      hasToolCalls(last);
+    return !hasOutput;
+  }, [thread.isLoading, thread.messages]);
+
   const handleReject = useCallback(async () => {
     await handleChatSubmit('这份配置先不应用，请继续帮我调整。');
   }, [handleChatSubmit]);
@@ -147,10 +212,21 @@ export function WritingSetupChat({
                 onClarificationSubmit={handleChatSubmit}
                 onSetupWritingApprove={handleApprove}
                 onSetupWritingReject={handleReject}
+                onSetupWritingStart={canStartWriting ? handleStart : undefined}
               />
             )}
           </div>
 
+          {/* 首 token 前的「正在整理…」占位：避免模型慢/排队时答案区空白像卡死 */}
+          {showThinkingPlaceholder && (
+            <div className="shrink-0 border-t bg-background px-4 py-2 dark:bg-transparent">
+              <div className="flex items-center gap-2 text-xs text-muted-foreground">
+                <span className="inline-block size-1.5 animate-pulse rounded-full bg-blue-500" />
+                正在整理…
+              </div>
+            </div>
+          )}
+
           <div className="bg-background dark:bg-transparent shrink-0 border-t px-4 py-4">
             {/* 与原型图一致：白色圆角卡片包裹「模型选择行 + 输入框」 */}
             <div className="space-y-3 rounded-2xl border border-[#ece5e4] bg-card p-3 shadow-sm dark:border-[#1B477E] dark:bg-[#022248]/40">

frontend-web/src/open-canvas/contexts/AIWritingContext.tsx

@@ -23,7 +23,7 @@ import {
 
 import { useCanvasGraphContext } from '../contexts/CanvasGraphContext';
 import { saveAIWritingTranscript } from '../api/ai-writing-sessions';
-import { countWords } from '../utils/wordCount';
+import { countArticleWords } from '../utils/wordCount';
 import {
   INITIAL_STATE,
   customToProgressEvent,
@@ -280,10 +280,14 @@ export function AIWritingProvider({ children }: { children: ReactNode }) {
           // 用户自带大纲原文（留空则后端走自动规划）
           user_outline_raw: request.userOutline?.trim() || '',
           article_type: request.articleType,
+          // 新（列表外）文章类型的一句话特征，仅本次写作临时使用；命中已有类型时为空。
+          article_type_detail: request.articleTypeDetail?.trim() || '',
           target_audience: request.targetAudience,
           word_count_target: request.wordCountTarget,
           keyword_count: Math.max(1, Math.min(10, request.keywordCount)),
           writing_mode: request.writingMode === 'strict' ? 'strict' : 'loose',
+          // 是否加参考文献（正文 [n] + 文末「参考文献」列表）；缺省 false 不改其它入口行为。
+          add_references: request.addReferences ?? false,
           material_source: request.materialSource,
           notebook_search_space_id: request.noteSearchSpaceId ?? null,
           notebook_source_ids: request.noteSourceIds ?? null,
@@ -328,7 +332,7 @@ export function AIWritingProvider({ children }: { children: ReactNode }) {
     else if (pausePoint === 'outline_confirm') record.outline = state.outline ?? undefined;
     else if (pausePoint === 'draft_confirm') {
       record.draftTitle = state.currentDraftTitle;
-      record.draftWordCount = countWords(state.currentDraftMarkdown);
+      record.draftWordCount = countArticleWords(state.currentDraftMarkdown);
       record.reviewResult = state.reviewResult ?? undefined;
     } else if (pausePoint === 'review_confirm') {
       record.reviewResult = state.reviewResult ?? undefined;
@@ -342,7 +346,7 @@ export function AIWritingProvider({ children }: { children: ReactNode }) {
       pendingRevisionContext = {
         submittedIssues: allIssues.filter((_, i) => selectedSet.has(i)),
         skippedIssues: allIssues.filter((_, i) => !selectedSet.has(i)),
-        oldWordCount: countWords(state.currentDraftMarkdown),
+        oldWordCount: countArticleWords(state.currentDraftMarkdown),
       };
     }
 

frontend-web/src/open-canvas/contexts/WritingFormContext.tsx

@@ -34,6 +34,9 @@ export interface WritingFormValue {
   setAudience: (v: string) => void;
   strictMode: boolean;
   setStrictMode: (v: boolean) => void;
+  /** 是否在成稿里加参考文献（正文 [n] 角标 + 文末「参考文献」列表，像论文）。默认开。 */
+  addReferences: boolean;
+  setAddReferences: (v: boolean) => void;
   materialSource: 'general' | 'notebook';
   setMaterialSource: (v: 'general' | 'notebook') => void;
   /** 我的空间（笔记本）来源：选中的 search_space_id；上提到 Context 以便左栏「笔记本」点击直接设为素材来源。 */
@@ -81,6 +84,7 @@ export function WritingFormProvider({
   const [wordCountInput, setWordCountInput] = useState('800');
   const [audience, setAudience] = useState('普通读者');
   const [strictMode, setStrictMode] = useState(false);
+  const [addReferences, setAddReferences] = useState(true);
   const [materialSource, setMaterialSource] = useState<'general' | 'notebook'>(initialMaterialSource ?? 'general');
   const [selectedNotebookId, setSelectedNotebookId] = useState<number | null>(initialNotebookId ?? null);
   const [selectedDocumentIds, setSelectedDocumentIds] = useState<number[]>(initialDocumentIds ?? []);
@@ -117,13 +121,14 @@ export function WritingFormProvider({
       wordCountInput, setWordCountInput,
       audience, setAudience,
       strictMode, setStrictMode,
+      addReferences, setAddReferences,
       materialSource, setMaterialSource,
       selectedNotebookId, setSelectedNotebookId,
       selectedDocumentIds, setSelectedDocumentIds,
       modelName, setModelName,
       applyProposal,
     }),
-    [intent, formExpanded, showOutline, outlineText, articleType, wordCountInput, audience, strictMode, materialSource, selectedNotebookId, selectedDocumentIds, modelName, applyProposal],
+    [intent, formExpanded, showOutline, outlineText, articleType, wordCountInput, audience, strictMode, addReferences, materialSource, selectedNotebookId, selectedDocumentIds, modelName, applyProposal],
   );
 
   return <WritingFormContext.Provider value={value}>{children}</WritingFormContext.Provider>;

frontend-web/src/open-canvas/hooks/useAIWritingStream.ts

@@ -21,7 +21,7 @@
  * 这样 reducer 逻辑（INITIAL_STATE + switch 各分支）与老 hook 共享同一份代码，
  * 只换"事件源"。阶段 7 删冗余时把老 hook 整个删掉就行。
  */
-import { countWords } from '../utils/wordCount';
+import { countArticleWords } from '../utils/wordCount';
 import type {
   AIWritingRequest,
   AIWritingState,
@@ -339,7 +339,7 @@ function handleEvent(prev: AIWritingState, event: ProgressEvent): AIWritingState
 
     case 'draft_ready': {
       const newMarkdown = event.draftMarkdown ?? '';
-      const newWordCount = countWords(newMarkdown);
+      const newWordCount = countArticleWords(newMarkdown);
       const summary = prev.pendingRevisionContext
         ? {
             revisedIssues: prev.pendingRevisionContext.submittedIssues,
@@ -701,6 +701,10 @@ export type AIWritingValues = {
   user_intent?: string;
   user_outline_raw?: string;
   article_type?: string;
+  /** 新（列表外）文章类型的一句话特征，仅本次写作临时使用。 */
+  article_type_detail?: string;
+  /** 是否加参考文献（正文 [n] 角标 + 文末「参考文献」列表）。 */
+  add_references?: boolean;
   status?: string;
   material_package?: unknown;
   current_outline?: unknown;

frontend-web/src/open-canvas/hooks/useWritingEditor.ts

@@ -25,25 +25,33 @@ export function useWritingEditor({
   const onChangeRef = useRef(onChange);
   useEffect(() => { onChangeRef.current = onChange; }, [onChange]);
 
+  // 编辑器自身最近一次 emit 出去的 markdown。父级把它存进 state 后会作为新的 `markdown` prop
+  // 回灌进来，下面的同步 effect 据此识别「这是我自己刚产生的变更」并跳过 setContent。
+  const lastEmittedRef = useRef<string | null>(null);
+
   const editor = useEditor({
     extensions,
     editable,
     content: markdown,
     onUpdate({ editor }) {
-      const fn = onChangeRef.current;
-      if (!fn) return;
       // eslint-disable-next-line @typescript-eslint/no-explicit-any
       const md = (editor as any).storage?.markdown?.getMarkdown?.() ?? editor.getText();
-      fn(md);
+      lastEmittedRef.current = md;
+      onChangeRef.current?.(md);
     },
   }, []);
 
-  // Sync external markdown changes (e.g. after streaming completes)
+  // Sync external markdown changes (e.g. after streaming completes / AI 改写)
   useEffect(() => {
     if (!editor) return;
     if (markdown === prevMarkdownRef.current) return;
     prevMarkdownRef.current = markdown;
 
+    // 跳过「编辑器自身刚 emit、又被父级回灌」的变更：此时编辑器内容已经是最新的，再 setContent
+    // 会重新解析整篇文档、重置光标。markdown 序列化对空格/换行是有损的（如行尾空格被裁剪），
+    // 自激回灌一旦触发 setContent，就会把刚输入的空格吞掉并让光标跳到下一个字（本次修的 bug）。
+    if (markdown === lastEmittedRef.current) return;
+
     const { from, to } = editor.state.selection;
     editor.commands.setContent(markdown, { emitUpdate: false });
     const docSize = editor.state.doc.content.size;

frontend-web/src/open-canvas/types/ai-writing.ts

@@ -254,6 +254,8 @@ export interface AIWritingRequest {
   /** 用户粘贴的「已有大纲」原文（可选）。非空时后端按此大纲生成，不再自动规划 */
   userOutline?: string;
   articleType: string;
+  /** 当 articleType 是已有类型库里没有的新类型时，配置助手识别的一句话特征说明（仅本次写作临时使用，不入库）。 */
+  articleTypeDetail?: string;
   targetAudience: string;
   wordCountTarget: number;
   modelName?: string;
@@ -261,6 +263,8 @@ export interface AIWritingRequest {
   keywordCount: number;
   /** 写作模式：strict=严格依据素材 / loose=可结合通用知识（默认 loose） */
   writingMode: "strict" | "loose";
+  /** 是否在成稿里加参考文献（正文 [n] 角标 + 文末「参考文献」列表，像论文）。默认 true。 */
+  addReferences?: boolean;
   /** 检索类型：general=通用检索（素材收集专家按配置技能检索，默认知识库检索技能）/ notebook=我的空间 */
   materialSource: "general" | "notebook";
   /** 笔记来源（可选）：空间 search_space_id */

frontend-web/src/open-canvas/utils/wordCount.ts

@@ -10,3 +10,15 @@ export function countWords(text: string): number {
   const alnumRuns = (text.match(/[A-Za-z0-9]+/g) ?? []).length;
   return cjk + cjkPunct + alnumRuns;
 }
+
+/**
+ * 正文字数：与后端 word_count 同口径——**不计文末「参考文献」列表**。
+ * 后端在字数收敛后才追加 `## 参考文献` 段，其 word_count 只统计正文；前端展示「实际 vs 目标」
+ * 也必须排除该段，否则会因文献列表虚高几十~几百字，让用户误以为严重超标。
+ */
+export function countArticleWords(markdown: string): number {
+  if (!markdown) return 0;
+  const idx = markdown.lastIndexOf('## 参考文献');
+  const body = idx >= 0 ? markdown.slice(0, idx) : markdown;
+  return countWords(body);
+}

offline-backend-20260512/backend/packages/harness/deerflow/agents/ai_writing/graph.py

@@ -209,11 +209,17 @@ async def pause_draft_confirm(state: AIWritingState, config: RunnableConfig) ->
             "progress_events": [{"type": "resumed", "agent_name": "系统", "message": f"按用户意见重写（{notes[:30]}）"}],
         }
 
-    # to_editor
+    # to_editor：用户可能在「您的修改意见」里填了关注点再点「让编辑继续审核」，
+    # 把它带给编辑节点，让三维度审查重点核查用户在意的方面（不再被丢弃）。
+    editor_notes = (user_input.get("userRevisionNotes") or "").strip()
+    msg = (
+        f"提交编辑审核（已附带你的修改意见：{editor_notes[:30]}）"
+        if editor_notes else "提交编辑审核"
+    )
     return {
-        "last_user_intervention": user_input,
+        "last_user_intervention": {**user_input, "_editor_notes": editor_notes},
         "status": "reviewing",
-        "progress_events": [{"type": "resumed", "agent_name": "系统", "message": "提交编辑审核"}],
+        "progress_events": [{"type": "resumed", "agent_name": "系统", "message": msg}],
     }
 
 

offline-backend-20260512/backend/packages/harness/deerflow/agents/ai_writing/nodes/editor.py

@@ -10,6 +10,7 @@ from langchain_core.runnables import RunnableConfig
 from ..state import AIWritingState, ReviewResult, ReviewIssue
 from ..prompts.editor_prompts import (
     EDITOR_SYSTEM, FACT_REVIEW_USER, LOGIC_REVIEW_USER, LANGUAGE_REVIEW_USER,
+    USER_FOCUS_BLOCK,
 )
 from ._agent_runner import (
     EDITOR_AGENT_ID,
@@ -34,6 +35,19 @@ logger = logging.getLogger(__name__)
 _DEFAULT_PASS_THRESHOLD = 75
 
 
+def _build_user_focus(state: AIWritingState) -> str:
+    """从干预数据取用户在「编辑审核」时附带的修改意见，拼成注入审查提示词的关注块。
+
+    仅当用户点「让编辑继续审核」且填了意见（graph.py `to_editor` 分支 stash 的
+    ``_editor_notes``）时返回非空块；否则返回空串，三维度审查保持原样。
+    """
+    intervention = state.get("last_user_intervention") or {}
+    notes = (intervention.get("_editor_notes") or "").strip()
+    if not notes:
+        return ""
+    return USER_FOCUS_BLOCK.format(notes=notes)
+
+
 def _materials_summary(pkg: dict) -> str:
     lines = []
     for m in (pkg.get("materials") or [])[:6]:
@@ -129,10 +143,13 @@ async def editor_node(state: AIWritingState, config: RunnableConfig) -> dict:
     ai_writing_config = get_ai_writing_config()
     pass_threshold = int(ai_writing_config.get("pass_threshold", _DEFAULT_PASS_THRESHOLD))
 
+    # 用户在草稿确认时附带的修改意见 → 注入三维度审查，让编辑重点核查（空则不影响原审查）
+    user_focus = _build_user_focus(state)
+
     REVIEW_TASKS = [
-        ("fact",     FACT_REVIEW_USER.format(materials_summary=mats, article=article)),
-        ("logic",    LOGIC_REVIEW_USER.format(article=article)),
-        ("language", LANGUAGE_REVIEW_USER.format(article=article)),
+        ("fact",     FACT_REVIEW_USER.format(materials_summary=mats, article=article, user_focus=user_focus)),
+        ("logic",    LOGIC_REVIEW_USER.format(article=article, user_focus=user_focus)),
+        ("language", LANGUAGE_REVIEW_USER.format(article=article, user_focus=user_focus)),
     ]
 
     # 内置 agent 模式（config.yaml ai_writing.use_builtin_agents.editor）：

offline-backend-20260512/backend/packages/harness/deerflow/agents/ai_writing/nodes/intent_parser.py

@@ -151,7 +151,9 @@ def _resolve_intent_fields(
     llm_type = raw_type if raw_type in valid_keys else fallback_type
 
     preconfigured_type = (state.get("article_type") or "").strip()
-    if preconfigured_type in valid_keys:
+    if preconfigured_type:
+        # 用户/配置助手已明确指定文章类型（含**列表外的自定义/新类型**）→ 原样尊重，
+        # 不再因不在默认表里被丢弃回退成默认（否则自定义/新类型在写作时会被还原成默认）。
         final_article_type = preconfigured_type
     else:
         final_article_type = llm_type

offline-backend-20260512/backend/packages/harness/deerflow/agents/ai_writing/nodes/writer_draft.py

@@ -1,6 +1,7 @@
 """作家节点 Phase 2：按大纲逐段写正文，支持流式 chunk 输出。"""
 from __future__ import annotations
 
+import json
 import logging
 import re
 
@@ -9,9 +10,16 @@ from langchain_core.runnables import RunnableConfig
 
 from ..state import AIWritingState, DraftArticle
 from ..prompts.writer_prompts import (
-    DRAFT_SECTION_SYSTEM, DRAFT_SECTION_USER, DRAFT_EXPAND_USER,
+    DRAFT_SECTION_SYSTEM, DRAFT_SECTION_USER, DRAFT_EXPAND_USER, DRAFT_COMPRESS_USER,
+    OUTLINE_REVISE_SYSTEM, OUTLINE_REVISE_USER,
     WRITER_MODE_LOOSE, WRITER_MODE_STRICT, SECTION_BLOCKED_MARKER,
 )
+# 复用大纲节点的解析/素材摘要：修订模式下先据意见更新大纲（含标题），再重写正文。
+# writer_outline 不反向 import 本模块，无循环。
+from .writer_outline import (
+    _parse_outline as _revise_parse_outline,
+    _materials_summary as _revise_materials_summary,
+)
 from ._agent_runner import (
     WRITER_AGENT_ID,
     WritingAgentUnavailable,
@@ -27,13 +35,116 @@ from ._model_fallback import (
     emit_model_switch_event,
     fallback_chain,
 )
-from ._utils import resolve_writing_model_name, _content_to_text, count_words, reasoning_delta, ThinkTagParser, safe_get_stream_writer
+from ._utils import resolve_writing_model_name, _content_to_text, count_words, llm_json, reasoning_delta, ThinkTagParser, safe_get_stream_writer
 
 logger = logging.getLogger(__name__)
 
 # 章节正文里的 Markdown 标题行(允许前导空白) — 分组捕获缩进 + 井号串
 _HEADING_PREFIX_RE = re.compile(r"^([ \t]*)(#{1,6})[ \t]+", re.MULTILINE)
 
+# ── 字数控制 ────────────────────────────────────────────────────────────────
+# 全文允许偏离目标字数的误差带（用户要求「误差不超过 100 字」）：初稿装配后双向收敛
+# 到 [target-tol, target+tol] 内，太短就扩写、太长就压缩。
+_WORD_COUNT_TOLERANCE = 100
+# 单节字数下限：均分/压缩时给章节兜个底，避免某节被压成一两句话。
+_SECTION_MIN_WORDS = 80
+# 收敛回合上限：每回合调一节（按相对应得字数偏差最大的节优先）。放宽到 8，让大幅超量
+# （需逐节压多节）也能压到位；命中容差或所有节都压不动即提前停。
+_MAX_REBALANCE_ROUNDS = 8
+
+
+def _allocate_section_words(word_count_target: int, n_sections: int) -> int:
+    """把总目标字数均分到各章节。
+
+    历史实现按 1.1 过量系数放大以补偿模型「写不够」，但实测模型更常「写太多」
+    （800 字写成 2000）。这里去掉放大、按目标如实均分，只留一个保底下限防止单节退化；
+    初稿超量/欠量统一交给装配后的双向收敛阶段纠正，确保全文落在误差带内。
+    """
+    n = max(n_sections, 1)
+    return max(_SECTION_MIN_WORDS, round(word_count_target / n))
+
+
+def _section_word_range(per_section: int) -> tuple[int, int]:
+    """单节目标对应的「硬上下限」，喂给模型卡字数。
+
+    关键是给出**上限**——旧提示词只说「不少于 N 字」，模型逐节超额累加才把 800 写成 2000。
+    """
+    return int(per_section * 0.85), int(per_section * 1.2)
+
+
+def _pick_rebalance(
+    counts: list[int], fair: int, target: int, tol: int, skip: set[int] | tuple = ()
+) -> tuple[int, int] | None:
+    """决定下一步调整哪一节、调到多少字；全文已在容差内或无可调项则返回 None。
+
+    返回 ``(idx, sec_target)``：
+    - 全文偏长 → 选「超应得份额最多」且未跳过的节，**压到它的应得份额**（可达目标，不强求
+      一节吃下全部缺口；各超额节轮流压到份额，累计即收敛到目标）。没有超份额的节时，退而
+      把最长的一节按缺口压一压。
+    - 全文偏短 → 选「最欠份额」的节，把整体缺口加到它头上扩写（扩写对模型更容易）。
+    ``skip`` 是本轮已尝试但「压不动 / 没改善」的章节集合，跳过它们换下一节，避免一节压不动
+    就放弃整个收敛。配合调用方「确有改善才采用」的护栏，保证单调收敛、自然终止。
+    """
+    if not counts:
+        return None
+    total = sum(counts)
+    delta = target - total  # >0 需补，<0 需删
+    if abs(delta) <= tol:
+        return None
+    pool = [k for k in range(len(counts)) if k not in skip]
+    if not pool:
+        return None
+    if delta < 0:
+        over = [k for k in pool if counts[k] > fair]
+        if over:
+            idx = max(over, key=lambda k: counts[k] - fair)
+            return idx, max(_SECTION_MIN_WORDS, fair)
+        # 各节都不高于均分却仍超目标（节数偏多 / 下限抬高）→ 压最长的一节
+        idx = max(pool, key=lambda k: counts[k])
+        return idx, max(_SECTION_MIN_WORDS, counts[idx] + delta)
+    under = [k for k in pool if counts[k] < fair]
+    target_pool = under or pool
+    idx = min(target_pool, key=lambda k: counts[k] - fair)
+    return idx, counts[idx] + delta
+
+
+# 触发「先改大纲再重写正文」的修订动作（直接按用户意见改 / 采纳编辑意见 / 自定义修订）
+_REVISE_ACTIONS = {"user_revise", "accept_review", "custom_revise"}
+
+
+async def _revise_outline_with_notes(
+    state: AIWritingState, config: RunnableConfig, outline: dict, feedback: str
+) -> dict:
+    """修订模式下，先按用户/编辑意见更新大纲（标题/结构），再据此重写正文。
+
+    章节标题、文章标题都由大纲注入，正文生成本身不产出标题——所以「把第一章标题改成…」
+    「加一节 / 删一节 / 调整顺序」这类指令必须先落到大纲，成稿里才看得到变化（否则只是
+    正文被重写、标题仍沿用旧大纲）。复用 OUTLINE_REVISE 提示词：它只改用户明确提到的
+    部分，纯正文类意见（如「第二段数据来源不清晰」）不会误动结构。任何异常都保留原大纲。
+    """
+    pkg = state.get("material_package") or {}
+    model_name = (state.get("model_name") or "").strip() or None
+    try:
+        result = await llm_json(
+            OUTLINE_REVISE_SYSTEM,
+            OUTLINE_REVISE_USER.format(
+                user_intent=state.get("user_intent", ""),
+                article_type=state.get("article_type", "analysis"),
+                target_audience=state.get("target_audience", "普通读者"),
+                word_count_target=state.get("word_count_target", 800),
+                current_outline=json.dumps(outline, ensure_ascii=False, indent=2),
+                feedback=feedback,
+                materials_summary=_revise_materials_summary(pkg),
+            ),
+            outline,  # 解析失败 → 回退原大纲
+            config,
+            model_name=model_name,
+        )
+        return _revise_parse_outline(result)
+    except Exception as e:
+        logger.warning("draft revise: outline update failed, keep original outline: %s", e)
+        return outline
+
 
 def _normalize_section_headings(text: str) -> str:
     """把章节正文里的 Markdown 标题层级归一到允许的两级:### / ####。
@@ -117,19 +228,112 @@ def _looks_like_refusal(text: str) -> bool:
     return any(marker in text for marker in _REFUSAL_MARKERS)
 
 
-def _get_section_materials(material_ids: list[str], pkg: dict) -> str:
+def _build_citation_map(pkg: dict) -> dict[str, int]:
+    """material_id → 1-based 序号（按 package 顺序），写作时作为角标 [n] 的稳定编号。
+
+    成稿装配后再按「正文出现顺序」重排成连续 [1][2][3]（见 ``_append_references``）。
+    """
+    return {m["id"]: i + 1 for i, m in enumerate(pkg.get("materials", [])) if m.get("id")}
+
+
+def _get_section_materials(material_ids: list[str], pkg: dict, cite_map: dict[str, int] | None = None) -> str:
+    """渲染本节可用素材喂给作家。
+
+    ``cite_map`` 非空（开启参考文献）时，每条素材前缀用其**全局引用编号** ``[n]``，作家据此在
+    正文里打角标；为空时沿用旧行为（前缀素材 id，仅作辨识、不要求引用）。
+    """
     mat_map = {m["id"]: m for m in pkg.get("materials", [])}
+
+    def _prefix(mid: str) -> str:
+        if cite_map and mid in cite_map:
+            return f"[{cite_map[mid]}]"
+        return f"[{mid}]"
+
     lines = []
     for mid in material_ids:
         if mid in mat_map:
             m = mat_map[mid]
-            lines.append(f"[{mid}] {m['title']}: {m['content'][:300]}")
+            lines.append(f"{_prefix(mid)} {m['title']}: {m['content'][:300]}")
     if not lines:
         for m in pkg.get("materials", [])[:2]:
-            lines.append(f"[{m['id']}] {m['title']}: {m['content'][:300]}")
+            lines.append(f"{_prefix(m['id'])} {m['title']}: {m['content'][:300]}")
     return "\n".join(lines) if lines else "（无相关素材）"
 
 
+# 开启参考文献时，追加到每节写作指令末尾：要求作家在引用素材处打角标。
+_CITATION_DIRECTIVE = (
+    "\n\n【引用要求】上面每条素材前的 [n] 是它的编号。"
+    "当你在正文里用到某条素材的事实/数据/观点时，在该句句末加上对应角标，如 [1] 或 [1][3]；"
+    "只能使用确实给出的素材编号，不要编造编号，不要给通用常识加角标。"
+    "不要在本节末尾自行罗列参考文献——文末的「参考文献」列表由系统统一生成。"
+)
+
+
+_CITATION_TOKEN_RE = re.compile(r"\[(\d{1,3})\]")
+
+
+def _format_reference_entry(m: dict) -> str:
+    """单条参考文献的展示文本：标题. 来源. 日期. url（缺项自动省略）。"""
+    title = (m.get("title") or "").strip() or "（无标题）"
+    parts = [title]
+    source = (m.get("source") or "").strip()
+    if source:
+        parts.append(source)
+    date = (m.get("published_at") or "").strip()
+    if date:
+        parts.append(date)
+    entry = ". ".join(parts)
+    url = (m.get("url") or "").strip()
+    if url:
+        entry = f"{entry}. {url}"
+    return entry
+
+
+def _append_references(markdown: str, pkg: dict, cite_map: dict[str, int]) -> str:
+    """把正文里的引用角标按「出现顺序」重排为连续 [1][2][3]，并在文末追加「参考文献」列表。
+
+    只收录正文真正引用到（且编号有效）的素材，列表顺序与正文出现顺序一致——像论文一样。
+    没有任何有效引用时原样返回（不加空列表）。
+    """
+    num_to_mat: dict[int, dict] = {}
+    for m in pkg.get("materials", []):
+        n = cite_map.get(m.get("id"))
+        if n:
+            num_to_mat[n] = m
+    if not num_to_mat:
+        return markdown
+
+    # 1) 按正文出现顺序收集被引用且有效的原始编号
+    order: list[int] = []
+    seen: set[int] = set()
+    for match in _CITATION_TOKEN_RE.finditer(markdown):
+        k = int(match.group(1))
+        if k in num_to_mat and k not in seen:
+            seen.add(k)
+            order.append(k)
+    if not order:
+        return markdown
+
+    # 2) 原始编号 → 连续新编号（出现先后）；逐 match 独立映射，避免顺序替换串号
+    renumber = {old: new for new, old in enumerate(order, start=1)}
+
+    def _sub(match: re.Match) -> str:
+        k = int(match.group(1))
+        return f"[{renumber[k]}]" if k in renumber else match.group(0)
+
+    body = _CITATION_TOKEN_RE.sub(_sub, markdown)
+
+    # 3) 文末参考文献列表（按新编号顺序）。
+    #    每条之间用**空行**分隔，使其各成独立段落——否则 Markdown 会把单换行的多行并成一段，
+    #    所有文献会挤成一坨（见排版反馈）。与正文一致沿用 [n] 角标，不再叠加列表序号。
+    entries = [
+        f"[{renumber[old]}] {_format_reference_entry(num_to_mat[old])}"
+        for old in order
+    ]
+    references_md = "## 参考文献\n\n" + "\n\n".join(entries)
+    return f"{body}\n\n{references_md}"
+
+
 async def _save_draft_to_repo(config: RunnableConfig, draft: DraftArticle) -> None:
     """把 current_draft 异步落到业务表（best-effort，写失败不影响流程）。
 
@@ -171,6 +375,10 @@ async def writer_draft_node(state: AIWritingState, config: RunnableConfig) -> di
     intervention = state.get("last_user_intervention") or {}
     revision_count = state.get("revision_count", 0)
 
+    # 参考文献：开启且有素材时，写作给素材编号 + 要求作家打角标 [n]，成稿装配后统一生成文献列表。
+    add_references = bool(state.get("add_references", False)) and bool(pkg.get("materials"))
+    cite_map = _build_citation_map(pkg) if add_references else None
+
     # 写作模式；用户在「素材不足求助」里选「按通用知识续写」的章节强制走宽松
     writing_mode = state.get("writing_mode", "loose")
     loose_titles: set[str] = set(intervention.get("_loose_sections") or [])
@@ -198,11 +406,23 @@ async def writer_draft_node(state: AIWritingState, config: RunnableConfig) -> di
         if parts:
             revision_notes = "\n修改要求：\n" + "\n".join(parts)
 
+    # 修订模式：先把上面这条意见落实到大纲（标题/结构），再据更新后的大纲逐节重写。
+    # 章节标题、文章标题都由大纲注入、正文生成本身不产出标题，所以「改第一章标题」「加一节/
+    # 删一节」这类指令必须先改大纲才能在成稿里生效（否则只是正文被重写、标题仍是旧的）。
+    outline_revised = False
+    if intervention.get("action") in _REVISE_ACTIONS and revision_notes.strip():
+        revised_outline = await _revise_outline_with_notes(state, config, outline, revision_notes.strip())
+        if revised_outline and revised_outline != outline:
+            outline = revised_outline
+            outline_revised = True
+
     word_count_target = state.get("word_count_target", 800)
     sections = outline.get("sections", [])
     n_sections = max(len(sections), 1)
-    # 单节硬下限：总目标 / 节数，乘 1.1 过量系数补偿模型对短目标的天然欠量；单节至少 150 字
-    words_per_section = max(150, int(word_count_target * 1.1 / n_sections))
+    # 按目标如实均分到各节（不再放大），并给模型一个硬上下限区间卡字数；
+    # 装配后再双向收敛把全文夹逼进误差带。
+    words_per_section = _allocate_section_words(word_count_target, n_sections)
+    sec_min_words, sec_max_words = _section_word_range(words_per_section)
     article_title = outline.get("title", "文章")
 
     # 兼容后端 snake_case 与前端编辑后回传的 camelCase
@@ -265,7 +485,7 @@ async def writer_draft_node(state: AIWritingState, config: RunnableConfig) -> di
         key_points = section.get("key_points") or section.get("keyPoints") or []
         mat_ids    = section.get("material_ids") or section.get("materialIds") or []
 
-        section_materials = _get_section_materials(mat_ids, pkg)
+        section_materials = _get_section_materials(mat_ids, pkg, cite_map)
         # 三级结构:把每个 KeyPoint 展开成「小节标题 + 缩进子要点」喂给作家。
         # 兼容旧 session 里 key_points 为 List[str] 的情况(走相同 helper)。
         key_points_str = _render_key_points_for_prompt(key_points)
@@ -289,7 +509,12 @@ async def writer_draft_node(state: AIWritingState, config: RunnableConfig) -> di
             revision_section=revision_notes,
             mode_section=mode_section,
             target_words=words_per_section,
+            min_words=sec_min_words,
+            max_words=sec_max_words,
         )
+        # 开启参考文献时，追加角标引用要求（两条路径——内置 agent 与旧路径——共用 user_msg）。
+        if add_references:
+            user_msg += _CITATION_DIRECTIVE
 
         section_buf = ""   # 只累积正文（非思考）用于拼接 Markdown
         section_succeeded = False
@@ -440,61 +665,95 @@ async def writer_draft_node(state: AIWritingState, config: RunnableConfig) -> di
             ],
         }
 
-    def _assemble() -> str:
+    def _assemble(bufs: list[str]) -> str:
         # 章节正文里的标题统一归一到 ###（要点小标题），章节标题用 ##，文章标题用 #
         body = "\n\n".join(
             f"## {t}\n\n{_normalize_section_headings(b)}"
-            for t, b in zip(section_titles, section_bufs)
+            for t, b in zip(section_titles, bufs)
         )
         return f"# {article_title}\n\n{body}"
 
-    full_markdown = _assemble()
+    async def _rewrite_section(idx: int, sec_target: int, *, shrink: bool) -> str:
+        """重写第 idx 节使其趋近 sec_target 字：shrink=True 走压缩、否则扩写（单轮，非流式）。
+
+        兜底优化，不再做模型链容错：拿当前已成功的模型试一次即可，失败由调用方保留原文。
+        """
+        sec = sections[idx]
+        tmpl = DRAFT_COMPRESS_USER if shrink else DRAFT_EXPAND_USER
+        msg = tmpl.format(
+            article_title=article_title,
+            section_title=section_titles[idx],
+            key_points=_render_key_points_for_prompt(
+                sec.get("key_points") or sec.get("keyPoints") or []
+            ),
+            section_materials=_get_section_materials(
+                sec.get("material_ids") or sec.get("materialIds") or [], pkg, cite_map
+            ),
+            existing_text=section_bufs[idx],
+            target_words=sec_target,
+        )
+        # 扩写/压缩时保留正文里已有的引用角标 [n]，避免重排后文献缺失。
+        if add_references:
+            msg += "\n\n注意：保留原文中已有的引用角标（形如 [1]、[2]），不要删除或改动它们。"
+        if use_agent:
+            return (await stream_writing_agent(
+                agent_id=WRITER_AGENT_ID,
+                user_payload=msg,
+                state=state,
+                config=config,
+                on_text=None,
+            )).strip()
+        rewrite_model = create_chat_model(name=current_model_name) if current_model_name else create_chat_model()
+        resp = await rewrite_model.ainvoke([
+            SystemMessage(content=DRAFT_SECTION_SYSTEM),
+            HumanMessage(content=msg),
+        ])
+        parser = ThinkTagParser()
+        return "".join(
+            t for is_thinking, t in parser.feed(_content_to_text(resp.content)) if not is_thinking
+        ).strip()
+
+    full_markdown = _assemble(section_bufs)
     word_count = count_words(full_markdown)
 
-    # 扩写兜底：总字数不足目标 90% 时，重写并扩充最短的章节（单轮，非流式）
-    if section_bufs and word_count < word_count_target * 0.9:
-        idx = min(range(len(section_bufs)), key=lambda k: count_words(section_bufs[k]))
-        target_sec = sections[idx]
-        expand_target = count_words(section_bufs[idx]) + (word_count_target - word_count)
+    # 双向收敛兜底：把全文字数夹逼进 [target±tol]。太短就扩写最欠的节、太长就压缩最超的节，
+    # 每回合只动一节，且仅在「确实更接近目标」时才采用。某节压不动 / 没改善 → 记入 skip 换
+    # 下一节继续（不再一节失败就放弃整个收敛，否则弱模型一次压缩失败就留下超量稿）。
+    rounds = 0
+    skip: set[int] = set()
+    while section_bufs and rounds < _MAX_REBALANCE_ROUNDS:
+        counts = [count_words(b) for b in section_bufs]
+        pick = _pick_rebalance(counts, words_per_section, word_count_target, _WORD_COUNT_TOLERANCE, skip)
+        if pick is None:
+            break
+        rounds += 1
+        idx, sec_target = pick
+        shrink = sec_target < counts[idx]
         try:
-            expand_msg = DRAFT_EXPAND_USER.format(
-                article_title=article_title,
-                section_title=section_titles[idx],
-                key_points=_render_key_points_for_prompt(
-                    target_sec.get("key_points") or target_sec.get("keyPoints") or []
-                ),
-                section_materials=_get_section_materials(
-                    target_sec.get("material_ids") or target_sec.get("materialIds") or [], pkg
-                ),
-                existing_text=section_bufs[idx],
-                target_words=expand_target,
-            )
-            if use_agent:
-                expanded = (await stream_writing_agent(
-                    agent_id=WRITER_AGENT_ID,
-                    user_payload=expand_msg,
-                    state=state,
-                    config=config,
-                    on_text=None,
-                )).strip()
-            else:
-                # 扩写是兜底优化，不再做模型链兜底：拿当前已成功的模型试一次即可，失败就保留原文。
-                expand_model = create_chat_model(name=current_model_name) if current_model_name else create_chat_model()
-                resp = await expand_model.ainvoke([
-                    SystemMessage(content=DRAFT_SECTION_SYSTEM),
-                    HumanMessage(content=expand_msg),
-                ])
-                parser = ThinkTagParser()
-                expanded = "".join(
-                    t for is_thinking, t in parser.feed(_content_to_text(resp.content)) if not is_thinking
-                ).strip()
-            # 仅在扩写确实更长时才采用，避免越改越短
-            if count_words(expanded) > count_words(section_bufs[idx]):
-                section_bufs[idx] = expanded
-                full_markdown = _assemble()
-                word_count = count_words(full_markdown)
+            new_text = await _rewrite_section(idx, sec_target, shrink=shrink)
         except Exception as e:
-            logger.warning("draft expand failed for section [%s]: %s", section_titles[idx], e)
+            logger.warning("draft rebalance failed for section [%s]: %s", section_titles[idx], e)
+            skip.add(idx)
+            continue
+        if not new_text:
+            skip.add(idx)
+            continue
+        candidate = list(section_bufs)
+        candidate[idx] = new_text
+        new_md = _assemble(candidate)
+        new_wc = count_words(new_md)
+        if abs(new_wc - word_count_target) < abs(word_count - word_count_target):
+            section_bufs = candidate
+            full_markdown = new_md
+            word_count = new_wc
+            skip.discard(idx)  # 改善了，后续若仍最超可再压
+        else:
+            skip.add(idx)      # 这一节压不动 / 反而更差，跳过它，换下一节
+
+    # 参考文献：在全文（含字数收敛）定稿后统一处理——把正文角标按出现顺序重排为连续 [1][2][3]，
+    # 并在文末追加「参考文献」列表（只收录被引用到的素材）。word_count 保持正文字数、不计文献列表。
+    if add_references and cite_map:
+        full_markdown = _append_references(full_markdown, pkg, cite_map)
 
     draft = DraftArticle(
         title=outline.get("title", "文章"),
@@ -513,6 +772,13 @@ async def writer_draft_node(state: AIWritingState, config: RunnableConfig) -> di
     notices: list[dict] = []
     if model_default_notice:
         notices.append(model_default_notice)
+    if outline_revised:
+        notices.append({
+            "type": "notice",
+            "agent_name": agent_label,
+            "title": "已按意见调整大纲",
+            "detail": "已根据你的修改意见更新大纲（如标题/结构），并据此重写正文。",
+        })
     if section_errors:
         notices.append({
             "type": "notice",
@@ -521,7 +787,7 @@ async def writer_draft_node(state: AIWritingState, config: RunnableConfig) -> di
             "detail": "部分章节生成失败，对应段落已占位。原因：" + "；".join(section_errors[:3]),
         })
 
-    return {
+    updates: dict = {
         "current_draft": draft,
         "drafts": [draft],
         "status": "awaiting_draft_confirm",
@@ -539,3 +805,7 @@ async def writer_draft_node(state: AIWritingState, config: RunnableConfig) -> di
             },
         ],
     }
+    # 大纲被意见改过 → 同步回 state，使后续修订/确认基于更新后的大纲（标题不再回退）
+    if outline_revised:
+        updates["current_outline"] = outline
+    return updates

offline-backend-20260512/backend/packages/harness/deerflow/agents/ai_writing/nodes/writer_outline.py

@@ -271,9 +271,14 @@ async def writer_outline_node(state: AIWritingState, config: RunnableConfig) ->
             logger.warning("outliner builtin agent unavailable, falling back to legacy prompts: %s", e)
             use_agent = False
 
+    # 新（列表外）文章类型时附上配置助手识别的一句话特征，让大纲生成体现该类型侧重；
+    # 命中已有类型时 detail 为空，article_type 维持原样。
+    _article_type = state.get("article_type", "analysis")
+    _article_type_detail = (state.get("article_type_detail") or "").strip()
+    article_type_text = f"{_article_type}（{_article_type_detail}）" if _article_type_detail else _article_type
     fmt_args = dict(
         user_intent=state["user_intent"],
-        article_type=state.get("article_type", "analysis"),
+        article_type=article_type_text,
         target_audience=state.get("target_audience", "普通读者"),
         word_count_target=state.get("word_count_target", 800),
         materials_summary=_materials_summary(pkg),

offline-backend-20260512/backend/packages/harness/deerflow/agents/ai_writing/prompts/editor_prompts.py

@@ -24,8 +24,17 @@ EDITOR_SYSTEM = """你是一名资深主编（编辑），标准极高，要求
   ]
 }"""
 
+# 用户在提交编辑审核时附带的修改意见 / 关注点（无则为空串，不注入）。
+# 放在「判断标准」之后、待审文章之前，提醒编辑在本维度重点核查用户在意的方面。
+USER_FOCUS_BLOCK = """
+
+【用户特别关注（高优先级）】
+用户在提交审核时提出了以下修改意见 / 关注点，请在本维度审查时**重点核查**这些方面；
+若发现文章未妥善处理，必须在 issues 中明确列出对应问题并给出针对性的 suggestion：
+{notes}"""
+
 FACT_REVIEW_USER = """请审查以下文章的【事实准确性】维度。
-判断标准：引用的数据/事实是否有素材支撑，是否有捏造或夸大。
+判断标准：引用的数据/事实是否有素材支撑，是否有捏造或夸大。{user_focus}
 
 参考素材：
 {materials_summary}
@@ -36,7 +45,7 @@ FACT_REVIEW_USER = """请审查以下文章的【事实准确性】维度。
 只输出 JSON（score: 0-100 整数，issues 列表，每条含 severity）："""
 
 LOGIC_REVIEW_USER = """请审查以下文章的【逻辑结构】维度。
-判断标准：论点是否清晰，论据是否充分，段落是否连贯，结构是否合理。
+判断标准：论点是否清晰，论据是否充分，段落是否连贯，结构是否合理。{user_focus}
 
 待审文章：
 {article}
@@ -44,7 +53,7 @@ LOGIC_REVIEW_USER = """请审查以下文章的【逻辑结构】维度。
 只输出 JSON（score: 0-100 整数，issues 列表，每条含 severity）："""
 
 LANGUAGE_REVIEW_USER = """请审查以下文章的【语言文字】维度。
-判断标准：语言是否流畅，用词是否准确，有无错别字或病句。
+判断标准：语言是否流畅，用词是否准确，有无错别字或病句。{user_focus}
 
 待审文章：
 {article}

offline-backend-20260512/backend/packages/harness/deerflow/agents/ai_writing/prompts/writer_prompts.py

@@ -208,7 +208,8 @@ DRAFT_SECTION_SYSTEM = """你是一名专业内容创作者（作家）。
 - 直接返回这一段的内容（含 `### 小节标题` + 概述段 + `#### 子要点标题` + 各自正文），不要 JSON，不要解释。
 - 不要写"以下是正文""本段内容如下"之类的开场白。
 - 紧扣本章节主题，不要写「文章整体结构」中其他章节负责的内容，避免重复。
-- 充分展开每一层：小节级 2-3 句概述，子要点级 3-5 句具体论述，尽量结合素材中的事实与数据说理，写满要求的篇幅。"""
+- 充分展开每一层：小节级 2-3 句概述，子要点级 3-5 句具体论述，尽量结合素材中的事实与数据说理。
+- 【字数纪律，硬性要求】严格遵守用户给定的字数区间：既不要明显写不够，更不要超出上限。靠控制论述详略来卡字数，不要为凑字数注水，也不要为收尾而草草带过。"""
 
 # ── 写作模式说明：注入 DRAFT_SECTION_USER 的 {mode_section} ───────────────────
 
@@ -240,14 +241,14 @@ DRAFT_SECTION_USER = """文章标题：{article_title}
 相关素材：
 {section_materials}
 {revision_section}{mode_section}
-请写出「{section_title}」这一章节的内容，不少于 {target_words} 字。
+请写出「{section_title}」这一章节的内容，目标约 {target_words} 字，并**严格控制在 {min_words}~{max_words} 字之间**——不要明显少于下限，更绝不要超过上限。
 
 按下面的结构两层展开写：
 - 每个主要 bullet（顶格）→ `### 小节标题`，标题下先写 2-3 句概述/引入正文。
 - 该小节的每个缩进 bullet（子要点）→ `#### 子要点标题`，标题下展开 3-5 句具体论述。
 - 没有缩进 bullet 的小节，直接在 ### 标题下连贯成段论述，不出现 ####。
 
-论述充分、层次清晰、内容充实。"""
+论述层次清晰、内容充实，但务必在上述字数上限内完成：宁可精炼也不要为凑字数注水。"""
 
 DRAFT_EXPAND_USER = """文章标题：{article_title}
 当前章节：{section_title}
@@ -266,3 +267,21 @@ DRAFT_EXPAND_USER = """文章标题：{article_title}
 - 保留原有的 ### 小节标题与 #### 子要点标题；不要新增或删除小节/子要点，也不要把 #### 改成 bullet。
 - 不要输出 # 或 ##（文章标题、章节标题由系统添加）。
 - 直接返回扩充后的完整章节内容（含 ### 与 #### 两层标题及各自正文段），不要解释。"""
+
+DRAFT_COMPRESS_USER = """文章标题：{article_title}
+当前章节：{section_title}
+章节内的小节与子要点（层级 bullet 列表）：
+{key_points}
+
+相关素材：
+{section_materials}
+
+这一章节已有内容如下：
+{existing_text}
+
+以上内容篇幅明显偏长。请在保留原意、关键事实和「### 小节 / #### 子要点」两级标题结构的基础上**大幅精简**改写这一章节，使内容**压到 {target_words} 字以内**——这是硬性上限，宁可删减细节也绝不能超过。
+要求：
+- 这是「压缩」任务，不是「重写扩充」：必须比原文明显更短。大胆删减冗余表述、合并重复论点、砍掉次要细节与铺陈，只保留关键事实、数据与核心结论。
+- 保留原有的 ### 小节标题与 #### 子要点标题；不要新增或删除小节/子要点，也不要把 #### 改成 bullet。
+- 不要输出 # 或 ##（文章标题、章节标题由系统添加）。
+- 直接返回精简后的完整章节内容（含 ### 与 #### 两层标题及各自正文段），不要解释。"""

offline-backend-20260512/backend/packages/harness/deerflow/agents/ai_writing/state.py

@@ -129,12 +129,17 @@ class AIWritingState(TypedDict):
     # 直接解析这份大纲并按用户结构生成，做到「完全按我的大纲写」。留空走正常自动规划。
     user_outline_raw: str
     article_type: str
+    # 当 article_type 是「已有类型库里没有」的新类型时，配置助手识别出的一句话特征说明。
+    # 仅本次写作临时使用（不入库）：在大纲生成时作为 hint 附在类型后，影响全文结构与侧重。
+    article_type_detail: str
     target_audience: str
     word_count_target: int
     # 素材检索关键词数量（用户指定，1~10，默认 6）
     keyword_count: int
     # 写作模式：strict=严格依据素材 / loose=可结合通用知识（默认 loose）
     writing_mode: str
+    # 是否在成稿里加参考文献：正文引用素材处加角标 [n]，文末按出现顺序生成「参考文献」列表（像论文）。
+    add_references: bool
     # 检索类型：general=通用检索（按素材收集专家 agent 配置的技能编排，默认）/
     # notebook=我的空间（仅笔记检索）。旧值 knowledge_base / skill 由 researcher
     # 节点按 general 兼容处理。

offline-backend-20260512/backend/packages/harness/deerflow/agents/lead_agent/agent.py

@@ -27,7 +27,9 @@ from deerflow.models import create_chat_model
 logger = logging.getLogger(__name__)
 
 # 写作配置 Q&A：意图澄清最多几轮（一条用户消息 = 一轮），到顶强制提交配置卡。
-WRITING_SETUP_MAX_ROUNDS = 3
+# 写作配置 Q&A 现在「先生成大纲并与用户多轮确认 → 再提交配置」，需要更多来回，故上限放宽。
+# 仍是兜底安全阀：到顶才用 tool_choice 强制提交，正常流程由用户确认大纲后自然提交。
+WRITING_SETUP_MAX_ROUNDS = 8
 
 
 def _get_runtime_config(config: RunnableConfig) -> dict:

offline-backend-20260512/backend/packages/harness/deerflow/agents/lead_agent/prompt.py

@@ -1128,29 +1128,40 @@ def build_writing_setup_prompt(article_types: list[dict] | None = None) -> str:
         type_lines = "\n".join(
             f"  - `{t.get('key', '')}`：{t.get('label', '')}" for t in article_types if t.get("key")
         )
-        type_block = f"【可选文章类型】article_type 只能填下列 key 之一：\n{type_lines}"
+        type_block = (
+            "【文章类型】优先从下列已有类型里选，命中就把它的 key 填进 `article_type`：\n"
+            f"{type_lines}\n"
+            "若上述类型**都不贴切**，可以**自拟一个新的文章类型**：`article_type` 填一个简短的中文类型名"
+            "（如「政策解读」「人物特写」「复盘总结」），并在 `article_type_detail` 里用一句话说明该类型的"
+            "写作侧重/特征。能用已有类型就不要新造，确实不贴切才新建。"
+        )
     else:
-        type_block = "【文章类型】若不确定 article_type，可留空，交给用户在表单里选。"
+        type_block = (
+            "【文章类型】根据意图判断文章类型：`article_type` 填一个简短的中文类型名，并在 "
+            "`article_type_detail` 里用一句话说明其写作侧重；不确定可留空，交给用户在表单里选。"
+        )
 
-    return f"""你是「AI 写作」功能的**配置助手**。你的唯一职责：通过简短对话把用户的写作需求聊清楚，然后调用 `setup_writing` 工具把写作配置提交给用户确认。提交后由用户在表单里应用、再点「开始写作」。
+    return f"""你是「AI 写作」功能的**配置助手**。你的职责分两段：① 先与用户把写作需求聊清楚，**先生成并与用户确认一份大纲**；② 待用户对大纲满意后，再调用 `setup_writing` 工具提交写作配置（含已确认的大纲），交用户开始写作。
 
 【铁律——最高优先级】
-- 你**绝对不撰写**文章正文、分析内容、大纲内容或任何成稿文字。即使用户的输入看起来就是一个写作任务（例如「新能源汽车市场分析」「帮我写一篇……」），你也**只把它当作"要配置的写作目标"**，绝不替用户动笔、绝不直接完成这篇写作。
+- 你**绝对不撰写**文章正文、分析内容或任何成稿文字。即使用户的输入看起来就是一个写作任务（例如「新能源汽车市场分析」「帮我写一篇……」），你也**只产出「大纲」（仅结构与标题）与「写作配置」**，绝不替用户动笔、绝不直接完成这篇写作。
 - 你只有两个工具：① `ask_clarification`（向用户提问澄清，需要时用）；② `setup_writing`（提交写作配置）。**没有任何其它工具**，不要做检索、不要做研究、不要假装调用别的工具。
-- 不要在回复里输出文章段落、要点清单或大纲（用户主动粘贴的大纲除外——原样放进 `outline` 参数）。
+- `setup_writing` **只能在用户明确确认大纲之后**才调用；在那之前绝不调用它。
 
-【工作流程】（与用户交互确认，但最多 3 轮内必须提交）
-1. 读用户输入。当关键项（文章类型、目标字数、目标读者）缺失或模糊时，调用 `ask_clarification` 向用户提问——**优先给出 `options` 选项**让用户点选（例如目标字数给「800 / 1500 / 3000」，读者给「普通读者 / 行业专家 / 政策决策者」）。一次可就 1-2 个独立的关键项各发一个 ask_clarification。
-2. 与用户的澄清来回**最多 2 次**。一旦关键信息明确、或已经问过 2 次，就**立即进入第 3 步提交**，缺失字段用合理默认值，绝不无休止追问。
-3. 调用 `setup_writing` 提交配置：
+【工作流程】
+1. **聊清需求**：读用户输入，当关键项（文章类型、目标字数、目标读者）缺失或模糊时，调用 `ask_clarification` 提问——**优先给出 `options` 选项**让用户点选（如目标字数「800 / 1500 / 3000」、读者「普通读者 / 行业专家 / 政策决策者」）。澄清**最多 2 次**，信息够了就进入下一步，不要无休止追问。
+2. **生成大纲（先于确认意图）**：信息足够后，直接以**普通回复文本**输出一份**多级层级**的中文大纲——用 Markdown 标题层级（`#` 一级、`##` 二级，必要时 `###` 三级），每节可用**一句话**点明这一节要写什么，但**绝不写正文**；大纲规模与目标字数匹配。**在回复的最后一句**明确请用户确认，例如：「以上是建议大纲，你看是否需要调整？确认后我就整理写作配置、开始写作。」这一步**不要调用任何工具**，输出完大纲就结束本轮、等待用户回应。
+3. **按反馈灵活改大纲**：用户要求调整（增删章节、改层级、换角度等）就**重新输出修订后的完整大纲**，并再次在最后一句请其确认；如此反复，直到用户表示满意/确认。用户若只是补充信息也照常更新大纲再确认。
+4. **提交配置**：用户确认大纲后，调用 `setup_writing` 提交：
    - `user_intent`：精炼成一句话的写作目标（必填）。
-   - `article_type`：见下方可选类型；不确定就留空。
+   - `article_type`：见下方【文章类型】——优先选已有类型的 key；都不贴切才自拟一个新的中文类型名。
+   - `article_type_detail`：**仅当 article_type 是新类型时**，用一句话说明该类型的写作侧重/特征；命中已有类型则留空。
    - `word_count_target`：正整数；用户没明说就给合理默认（如 800）。
    - `target_audience`：如「普通读者」「行业专家」；没提就填「普通读者」。
    - `writing_mode`：用户强调「严格依据素材/不要编造」→ `strict`，否则 `loose`。
    - `material_source`：**默认 `knowledge_base`（知识库）**；只有当用户明确要求用「技能/技能检索」时才填 `skill`，否则一律 `knowledge_base`。
-   - `outline`：用户提供的大纲原文（保留其层级与标题），否则留空。
-4. 调用 `setup_writing` 前，用**一句话**说明你将提交的配置要点；调用之后不要再输出任何文章内容。
+   - `outline`：**填入刚刚与用户确认的那份大纲全文**（保留其层级与标题）。
+   调用前用**一句话**说明将提交的配置要点；调用之后不再输出任何内容。
 
 {type_block}
 """

offline-backend-20260512/backend/packages/harness/deerflow/agents/middlewares/setup_writing_approval_middleware.py

@@ -74,6 +74,8 @@ class SetupWritingApprovalMiddleware(AgentMiddleware[SetupWritingApprovalMiddlew
         return {
             "user_intent": str(args.get("user_intent") or "").strip(),
             "article_type": str(args.get("article_type") or "").strip(),
+            # 新（列表外）文章类型的一句话特征说明；命中已有类型时为空。
+            "article_type_detail": str(args.get("article_type_detail") or "").strip(),
             "word_count_target": self._coerce_word_count(args.get("word_count_target")),
             "target_audience": str(args.get("target_audience") or "").strip(),
             "writing_mode": str(args.get("writing_mode") or "").strip(),

offline-backend-20260512/backend/packages/harness/deerflow/tools/builtins/setup_writing_tool.py

@@ -28,24 +28,30 @@ def setup_writing(
     writing_mode: str | None = None,
     material_source: str | None = None,
     outline: str | None = None,
+    article_type_detail: str | None = None,
 ) -> Command:
     """把与用户聊清的写作配置提交给前端表单（不直接开始写作）。
 
     Args:
         user_intent: 用户的写作意图（精炼成一句话的写作目标）。
-        article_type: 文章类型的 key（须是系统给定的合法类型之一；不确定可留空）。
+        article_type: 文章类型。优先填系统给定的合法类型 key；若已有类型都不贴切，可**自拟一个新的
+            文章类型**——此时填一个简短的中文类型名（如「政策解读」「人物特写」），并在
+            ``article_type_detail`` 里说明该类型特征。不确定可留空。
         word_count_target: 目标字数（正整数）。
         target_audience: 目标读者（如「行业专家」「普通读者」）。
         writing_mode: 写作模式，'strict'（严格依据素材）或 'loose'（可结合通用知识）。
         material_source: 素材来源，'knowledge_base'（知识库）或 'skill'（技能检索）。
-        outline: 若用户已提供大纲，原样填入其大纲文本（保留其层级与标题），否则留空。
+        outline: **已与用户确认的大纲全文**（保留其层级与标题）。本工具在用户确认大纲后调用，
+            故通常应带上这份大纲；用户未提供/未确认大纲时才留空。
+        article_type_detail: 仅当 ``article_type`` 是**新类型（不在已有列表里）**时填写：用一句话说明
+            这个类型的写作侧重/特征（如「以政策原文为锚、逐条拆解影响并给出执行建议」）。命中已有类型时留空。
     """
     logger.info("[setup_writing] proposed writing config submitted (intercepted by approval middleware)")
     return Command(
         update={
             "messages": [
                 ToolMessage(
-                    content="写作配置已生成，请在右侧确认后应用到表单。",
+                    content="写作配置已生成，请在右侧确认后应用到表单，或直接开始写作。",
                     tool_call_id=runtime.tool_call_id,
                 )
             ]