fix(xiaohongshu): extract direct comment reply target

cli-manifest.json

@@ -30631,7 +30631,7 @@
         "type": "boolean",
         "default": false,
         "required": false,
-        "help": "Include nested replies (楼中楼)"
+        "help": "Include nested replies; reply_to is the direct target shown by the page"
       }
     ],
     "columns": [
@@ -42583,7 +42583,7 @@
         "type": "boolean",
         "default": false,
         "required": false,
-        "help": "Include nested replies (楼中楼)"
+        "help": "Include nested replies; reply_to is the direct target shown by the page"
       }
     ],
     "columns": [

clis/rednote/comments.js

@@ -31,7 +31,7 @@ cli({
     args: [
         { name: 'note-id', required: true, positional: true, help: 'Full rednote note URL with xsec_token' },
         { name: 'limit', type: 'int', default: 20, help: 'Number of top-level comments (max 50)' },
-        { name: 'with-replies', type: 'boolean', default: false, help: 'Include nested replies (楼中楼)' },
+        { name: 'with-replies', type: 'boolean', default: false, help: 'Include nested replies; reply_to is the direct target shown by the page' },
     ],
     columns: ['rank', 'author', 'text', 'likes', 'time', 'is_reply', 'reply_to', 'images'],
     func: async (page, kwargs) => {

clis/xiaohongshu/comments.js

@@ -262,12 +262,17 @@ export function buildCommentsExtractJs(withReplies, limit = 20) {
             p.querySelectorAll('.reply-container .comment-item-sub, .sub-comment-list .comment-item').forEach(sub => {
               const sAuthor = clean(sub.querySelector('.name, .user-name'))
               const sAuthorHrefRaw = extractAuthorHref(sub)
-              const sText = clean(sub.querySelector('.content, .note-text'))
+              const sContent = sub.querySelector('.content, .note-text')
+              const sText = clean(sContent)
+              // Xiaohongshu renders the direct target only for reply-to-reply
+              // rows (回复 <span class="nickname">Bob</span> : ...). A nested
+              // row without that marker is a direct reply to the thread root.
+              const sReplyTo = clean(sContent?.querySelector(':scope > .nickname')) || author
               const sLikes = parseLikes(sub.querySelector('.count'))
               const sTime = clean(sub.querySelector('.date, .time'))
               const sImages = extractImages(sub)
               if (!sText) return
-              results.push({ author: sAuthor, authorHrefRaw: sAuthorHrefRaw, text: sText, likes: sLikes, time: sTime, is_reply: true, reply_to: author, images: sImages })
+              results.push({ author: sAuthor, authorHrefRaw: sAuthorHrefRaw, text: sText, likes: sLikes, time: sTime, is_reply: true, reply_to: sReplyTo, images: sImages })
             })
           }
         }
@@ -287,7 +292,7 @@ export const command = cli({
     args: [
         { name: 'note-id', required: true, positional: true, help: 'Full Xiaohongshu note URL with xsec_token' },
         { name: 'limit', type: 'int', default: 20, help: 'Number of top-level comments (max 50)' },
-        { name: 'with-replies', type: 'boolean', default: false, help: 'Include nested replies (楼中楼)' },
+        { name: 'with-replies', type: 'boolean', default: false, help: 'Include nested replies; reply_to is the direct target shown by the page' },
     ],
     columns: ['rank', 'author', 'userId', 'profileUrl', 'text', 'likes', 'time', 'is_reply', 'reply_to', 'images'],
     func: async (page, kwargs) => {

clis/xiaohongshu/comments.test.js

@@ -27,20 +27,35 @@ function createPageMock(evaluateResult) {
     };
 }
 
-async function runCommentsExtract(html) {
+async function runCommentsExtract(html, withReplies = false) {
     const dom = new JSDOM(html, { url: 'https://www.xiaohongshu.com/search_result/abc123?xsec_token=tok' });
+    const hadDocument = Object.hasOwn(globalThis, 'document');
+    const hadLocation = Object.hasOwn(globalThis, 'location');
+    const hadWindow = Object.hasOwn(globalThis, 'window');
+    const hadHTMLElement = Object.hasOwn(globalThis, 'HTMLElement');
     const previousDocument = globalThis.document;
     const previousLocation = globalThis.location;
+    const previousWindow = globalThis.window;
+    const previousHTMLElement = globalThis.HTMLElement;
     globalThis.document = dom.window.document;
     globalThis.location = dom.window.location;
+    globalThis.window = dom.window;
+    globalThis.HTMLElement = dom.window.HTMLElement;
     try {
         // limit=1 so the scroll-loading loop's initial "already have enough"
         // check short-circuits instead of burning through stall retries — the
         // JSDOM fixtures below are fully static, there's nothing more to load.
-        return await eval(buildCommentsExtractJs(false, 1));
+        return await eval(buildCommentsExtractJs(withReplies, 1));
     } finally {
-        globalThis.document = previousDocument;
-        globalThis.location = previousLocation;
+        if (hadDocument) globalThis.document = previousDocument;
+        else delete globalThis.document;
+        if (hadLocation) globalThis.location = previousLocation;
+        else delete globalThis.location;
+        if (hadWindow) globalThis.window = previousWindow;
+        else delete globalThis.window;
+        if (hadHTMLElement) globalThis.HTMLElement = previousHTMLElement;
+        else delete globalThis.HTMLElement;
+        dom.window.close();
     }
 }
 
@@ -66,6 +81,32 @@ describe('parseXhsLikeCountText', () => {
 
 describe('xiaohongshu comments', () => {
     const command = getRegistry().get('xiaohongshu/comments');
+    it('restores JSDOM globals after DOM extraction', async () => {
+        const keys = ['document', 'location', 'window', 'HTMLElement'];
+        const before = keys.map(key => ({
+            key,
+            hadOwnProperty: Object.hasOwn(globalThis, key),
+            value: Reflect.get(globalThis, key),
+        }));
+
+        await runCommentsExtract(`
+          <main>
+            <section class="parent-comment">
+              <div class="comment-item">
+                <span class="name">Alice</span>
+                <div class="content">Root comment</div>
+              </div>
+            </section>
+          </main>
+        `);
+
+        for (const entry of before) {
+            expect(Object.hasOwn(globalThis, entry.key)).toBe(entry.hadOwnProperty);
+            if (entry.hadOwnProperty) {
+                expect(Reflect.get(globalThis, entry.key)).toBe(entry.value);
+            }
+        }
+    });
     it('returns ranked comment rows for signed full URLs', async () => {
         const page = createPageMock({
             loginWall: false,
@@ -404,6 +445,41 @@ describe('xiaohongshu comments', () => {
         expect(result[0]).toMatchObject({ rank: 1, author: 'Alice' });
     });
     describe('--with-replies', () => {
+        it('extracts the direct reply target from nested reply DOM', async () => {
+            const data = await runCommentsExtract(`
+              <main>
+                <section class="parent-comment">
+                  <div id="comment-root" class="comment-item">
+                    <div class="author-wrapper"><span class="name">Alice</span></div>
+                    <div class="content">Root comment</div>
+                  </div>
+                  <div class="reply-container">
+                    <div id="comment-direct" class="comment-item-sub">
+                      <div class="author-wrapper"><span class="name">Bob</span></div>
+                      <div class="content"><span class="note-text">Direct reply</span></div>
+                    </div>
+                    <div id="comment-nested" class="comment-item-sub">
+                      <div class="author-wrapper"><span class="name">Carol</span></div>
+                      <div class="content">
+                        <span>回复 </span><span class="nickname">Bob</span> :
+                        <span class="note-text">Nested reply</span>
+                      </div>
+                    </div>
+                  </div>
+                </section>
+              </main>
+            `, true);
+
+            expect(data.results).toHaveLength(3);
+            expect(data.results[0]).toMatchObject({ author: 'Alice', is_reply: false, reply_to: '' });
+            expect(data.results[1]).toMatchObject({ author: 'Bob', is_reply: true, reply_to: 'Alice' });
+            expect(data.results[2]).toMatchObject({
+                author: 'Carol',
+                text: '回复 Bob : Nested reply',
+                is_reply: true,
+                reply_to: 'Bob',
+            });
+        });
         it('includes reply rows with is_reply=true and reply_to set', async () => {
             const page = createPageMock({
                 loginWall: false,

docs/adapters/browser/rednote.md

@@ -39,6 +39,7 @@ opencli rednote download "https://www.rednote.com/search_result/<id>?xsec_token=
 ```
 
 > Note: `note`, `comments`, and `download` require a full signed rednote.com note URL with `xsec_token`. Bare note IDs and xhslink.com short links are not accepted because they cannot prove the rednote host/cookie identity before navigation.
+> With `comments --with-replies`, `reply_to` is the direct reply target displayed by the page. Replies without an explicit `回复 <name>` marker target the enclosing top-level comment.
 
 ## Prerequisites
 

docs/adapters/browser/xiaohongshu.md

@@ -72,6 +72,7 @@ opencli xiaohongshu delete-note 6a08ba0b000000000702a893 --execute
 ```
 
 > Note: `note` and `comments` now require a full signed note URL with `xsec_token`. `download` accepts either a signed note URL or an `xhslink` short link. Bare note IDs are no longer reliable on xiaohongshu.
+> With `comments --with-replies`, `reply_to` is the direct reply target displayed by the page. Replies without an explicit `回复 <name>` marker target the enclosing top-level comment.
 > `ask` is separate from ordinary `search`: it submits the question to 点点, returns `answer`, `source_count`, and `sources[]`, and keeps `xsec_token` in JSON when Xiaohongshu returns one. The current 点点 source API may return bare note IDs without `xsec_token`; in that case `url` falls back to `/explore/<note_id>` and `xsec_token` is an empty string. Each source also carries the engagement and identity metadata 点点 returns: `like_count`, `note_type` (`normal`/`video`), `user_id`, and `published_at` (each omitted when 点点 does not provide it), so citation analysis can read likes and note format without a follow-up `search`/`note` round-trip.
 > `delete-note` operates in creator center and accepts a 24-character note ID or exact Xiaohongshu note URL; it defaults to dry-run verification and only deletes with `--execute`.
 > `follow` and `unfollow` are write commands on the public profile page. They verify the browser stayed on the requested `/user/profile/<id>` target before clicking, and verify the visible follow-state button after the action.