dsh-web-fetch-playwright 0.2.15 → 0.2.18
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/README.md +30 -3
- package/README.zh-CN.md +30 -3
- package/lib/client.js +46 -21
- package/lib/index.js +1001 -84
- package/lib/types/autoscroll.d.ts +277 -0
- package/lib/types/challenge.d.ts +23 -9
- package/lib/types/client/controller.d.ts +7 -0
- package/lib/types/client/locales.d.ts +1 -1
- package/lib/types/config.d.ts +38 -3
- package/lib/types/index.d.ts +3 -1
- package/lib/types/markdown.d.ts +18 -7
- package/lib/types/provider.d.ts +25 -0
- package/lib/types/strip.d.ts +59 -0
- package/package.json +1 -1
- package/src/autoscroll.ts +653 -0
- package/src/challenge.ts +36 -11
- package/src/client/card.tsx +13 -0
- package/src/client/controller.ts +9 -0
- package/src/client/locales.ts +7 -0
- package/src/config.ts +46 -3
- package/src/index.ts +5 -0
- package/src/markdown.ts +371 -14
- package/src/provider.ts +57 -68
- package/src/strip.ts +72 -0
package/src/markdown.ts
CHANGED
|
@@ -5,8 +5,14 @@
|
|
|
5
5
|
* inline `data:` image payloads are elided to size placeholders (build
|
|
6
6
|
* tools inline images as base64, which would otherwise dominate the body);
|
|
7
7
|
* Mozilla Readability extracts the article (dropping nav bars, sidebars,
|
|
8
|
-
* footers, and ad chrome by scoring link density and text mass)
|
|
9
|
-
*
|
|
8
|
+
* footers, and ad chrome by scoring link density and text mass) — its result
|
|
9
|
+
* must then pass a coverage gate that distrusts tiny or sliver-thin
|
|
10
|
+
* extractions, which on list-shaped pages (portals, catalogs) is exactly
|
|
11
|
+
* what the link-density penalty leaves behind; an accepted extraction is
|
|
12
|
+
* then pruned of trailing link-dense sections — the post-article
|
|
13
|
+
* recommendation rails and share bars the scorer's vocabularies cannot name
|
|
14
|
+
* (they arrive as bare divs with no class to match); DOMPurify sanitizes
|
|
15
|
+
* whatever HTML remains and forbids the layout tags noise lives in;
|
|
10
16
|
* Turndown with the GFM plugin converts to markdown using the same style
|
|
11
17
|
* options and span-safe table rules as the shipped `dsh-tool-web` renderer,
|
|
12
18
|
* so output is consistent with what `web_fetch` produces elsewhere.
|
|
@@ -108,6 +114,105 @@ function humanSize(bytes: number): string {
|
|
|
108
114
|
return `${(bytes / 1_048_576).toFixed(1)}MB`
|
|
109
115
|
}
|
|
110
116
|
|
|
117
|
+
/**
|
|
118
|
+
* Tags whose subtrees are noise when measuring how much of a page an
|
|
119
|
+
* extraction kept: DOMPurify drops these outright (`KEEP_CONTENT: false`),
|
|
120
|
+
* and its default policy always removes `script`/`style` — so a text walk
|
|
121
|
+
* that skips them measures what the sanitized document fallback could
|
|
122
|
+
* actually return, not text that is already destined for deletion. Measuring
|
|
123
|
+
* the raw body instead would let a huge `<nav>`/`<footer>` mass shrink an
|
|
124
|
+
* article's honest coverage share and demote good extractions.
|
|
125
|
+
*/
|
|
126
|
+
const NOISE_SUBTREE_TAGS = new Set([...FORBID_TAGS, 'script', 'style'])
|
|
127
|
+
|
|
128
|
+
/**
|
|
129
|
+
* A page must have at least this much visible (post-noise) text before the
|
|
130
|
+
* article acceptance gate below has an opinion. Sparse pages — logins,
|
|
131
|
+
* stubs, `login-like` fixtures — have nowhere for content to hide, so
|
|
132
|
+
* Readability's pick, however short, is still the best reading of the page
|
|
133
|
+
* and keeps flowing through the article path exactly as before.
|
|
134
|
+
*/
|
|
135
|
+
const GATE_MIN_BODY_TEXT_CHARS = 1_000
|
|
136
|
+
|
|
137
|
+
/**
|
|
138
|
+
* Minimum non-whitespace characters a Readability extraction must carry
|
|
139
|
+
* before it can be trusted as a page's content. Measured on the 163.com
|
|
140
|
+
* homepage, the scorer's pick was a 96-character footer licensing block —
|
|
141
|
+
* real articles and even sparse pages clear two-to-three times this bar,
|
|
142
|
+
* while link-dense portal pages lose whole news columns to it.
|
|
143
|
+
*/
|
|
144
|
+
const GATE_MIN_ARTICLE_TEXT_CHARS = 250
|
|
145
|
+
|
|
146
|
+
/**
|
|
147
|
+
* Minimum share of a page's visible text an extraction must cover to be
|
|
148
|
+
* trusted. Measured separation on real pages is wide: portal homepages
|
|
149
|
+
* (163 0.01, sina 0.03, sohu 0.02) vs article pages (0.28–0.45), so 10%
|
|
150
|
+
* sits in an empty band with margin on both sides.
|
|
151
|
+
*/
|
|
152
|
+
const GATE_MIN_ARTICLE_COVERAGE = 0.1
|
|
153
|
+
|
|
154
|
+
/**
|
|
155
|
+
* Count non-whitespace characters under `root`, skipping noise subtrees
|
|
156
|
+
* ({@link NOISE_SUBTREE_TAGS}). Whitespace is ignored so CJK pages (no
|
|
157
|
+
* spaces) and latin pages (space-separated) measure comparable densities,
|
|
158
|
+
* mirroring how the extraction-quality ratios in the docs were computed.
|
|
159
|
+
*
|
|
160
|
+
* @param root - the subtree to measure; `null` counts as nothing.
|
|
161
|
+
* @returns the number of non-whitespace characters in kept subtrees.
|
|
162
|
+
*/
|
|
163
|
+
function visibleTextLength(root: Node | null): number {
|
|
164
|
+
if (root === null) return 0
|
|
165
|
+
let length = 0
|
|
166
|
+
const visit = (node: Node): void => {
|
|
167
|
+
// Text nodes contribute their non-whitespace characters (nodeType 3).
|
|
168
|
+
if (node.nodeType === 3) {
|
|
169
|
+
length += (node.nodeValue ?? '').replace(/\s+/g, '').length
|
|
170
|
+
return
|
|
171
|
+
}
|
|
172
|
+
// Only elements branch further (nodeType 1); comments and friends carry
|
|
173
|
+
// no text, and noise subtrees are skipped wholesale.
|
|
174
|
+
if (node.nodeType !== 1 || NOISE_SUBTREE_TAGS.has(node.nodeName.toLowerCase())) return
|
|
175
|
+
for (let child = node.firstChild; child !== null; child = child.nextSibling) visit(child)
|
|
176
|
+
}
|
|
177
|
+
visit(root)
|
|
178
|
+
return length
|
|
179
|
+
}
|
|
180
|
+
|
|
181
|
+
/**
|
|
182
|
+
* Whether a Readability extraction is trustworthy as *the* content of the
|
|
183
|
+
* page it came from — the acceptance gate that keeps list-shaped pages whole.
|
|
184
|
+
*
|
|
185
|
+
* Readability is an article extractor: it scores text mass and penalizes
|
|
186
|
+
* link density, so on a portal/catalog page — where the content *is* the
|
|
187
|
+
* links — every news column scores near zero, some sparse low-link block (a
|
|
188
|
+
* footer licensing notice, a province-switcher popup) wins by default, and
|
|
189
|
+
* Readability's own 500-char quality check cannot fail the parse because its
|
|
190
|
+
* retry loop eventually returns the longest of its bad attempts. The plugin
|
|
191
|
+
* used to accept any non-null result, so the 163.com homepage rendered as a
|
|
192
|
+
* 162-character licensing fragment while 320 headlines were dropped — a
|
|
193
|
+
* recall loss far worse than any precision the gate could cost.
|
|
194
|
+
*
|
|
195
|
+
* The gate therefore distrusts an extraction that is tiny in absolute terms
|
|
196
|
+
* or covers a sliver of the page's visible text; the pipeline then falls
|
|
197
|
+
* back to the sanitized whole document, which DOMPurify keeps clean and
|
|
198
|
+
* which on such pages retains effectively all list content (measured 0%
|
|
199
|
+
* text loss on the 163.com homepage). Only text-rich pages are gated
|
|
200
|
+
* ({@link GATE_MIN_BODY_TEXT_CHARS}); below that, behavior is unchanged.
|
|
201
|
+
*
|
|
202
|
+
* @param articleHtml - the `article.content` HTML Readability produced.
|
|
203
|
+
* @param document - the live document the article was extracted from.
|
|
204
|
+
* @returns whether the article should flow through the article path.
|
|
205
|
+
*/
|
|
206
|
+
function articleCoversPage(articleHtml: string, document: Document): boolean {
|
|
207
|
+
const bodyText = visibleTextLength(document.body)
|
|
208
|
+
if (bodyText < GATE_MIN_BODY_TEXT_CHARS) return true
|
|
209
|
+
const container = document.createElement('div')
|
|
210
|
+
container.innerHTML = articleHtml
|
|
211
|
+
const articleText = visibleTextLength(container)
|
|
212
|
+
return articleText >= GATE_MIN_ARTICLE_TEXT_CHARS
|
|
213
|
+
&& articleText / bodyText >= GATE_MIN_ARTICLE_COVERAGE
|
|
214
|
+
}
|
|
215
|
+
|
|
111
216
|
/** The shared converter: same style options as `dsh-tool-web`'s renderer. */
|
|
112
217
|
const turndown = new TurndownService({
|
|
113
218
|
headingStyle: 'atx',
|
|
@@ -170,17 +275,261 @@ export type DenoiseMode = 'article' | 'document'
|
|
|
170
275
|
export interface DenoiseResult {
|
|
171
276
|
/** The markdown body (title, when found, already prepended). */
|
|
172
277
|
markdown: string
|
|
173
|
-
/**
|
|
278
|
+
/**
|
|
279
|
+
* `article` = Readability extraction that passed the coverage gate;
|
|
280
|
+
* `document` = whole-document fallback (Readability returned nothing, or
|
|
281
|
+
* the gate distrusted what it returned — list-shaped pages).
|
|
282
|
+
*/
|
|
174
283
|
mode: DenoiseMode
|
|
175
284
|
}
|
|
176
285
|
|
|
286
|
+
/**
|
|
287
|
+
* Count non-whitespace characters under `root` that sit inside an anchor,
|
|
288
|
+
* skipping the same noise subtrees {@link visibleTextLength} skips so the
|
|
289
|
+
* numerator and denominator of a link-density ratio stay comparable.
|
|
290
|
+
*
|
|
291
|
+
* @param root - the subtree to measure; `null` counts as nothing.
|
|
292
|
+
* @returns the number of non-whitespace characters inside `<a>` descendants.
|
|
293
|
+
*/
|
|
294
|
+
function linkTextLength(root: Node | null): number {
|
|
295
|
+
if (root === null) return 0
|
|
296
|
+
let length = 0
|
|
297
|
+
const visit = (node: Node, insideAnchor: boolean): void => {
|
|
298
|
+
if (node.nodeType === 3) {
|
|
299
|
+
if (insideAnchor) length += (node.nodeValue ?? '').replace(/\s+/g, '').length
|
|
300
|
+
return
|
|
301
|
+
}
|
|
302
|
+
if (node.nodeType !== 1 || NOISE_SUBTREE_TAGS.has(node.nodeName.toLowerCase())) return
|
|
303
|
+
const anchor = insideAnchor || node.nodeName.toLowerCase() === 'a'
|
|
304
|
+
for (let child = node.firstChild; child !== null; child = child.nextSibling) visit(child, anchor)
|
|
305
|
+
}
|
|
306
|
+
visit(root, false)
|
|
307
|
+
return length
|
|
308
|
+
}
|
|
309
|
+
|
|
310
|
+
/**
|
|
311
|
+
* Minimum non-whitespace characters a block must carry before its link
|
|
312
|
+
* density is even evaluated. Blocks below this are spacers, empty wrappers,
|
|
313
|
+
* or one-line sign-offs — too small to be worth a removal decision, so the
|
|
314
|
+
* trailing walk skips past them instead of stopping at them.
|
|
315
|
+
*/
|
|
316
|
+
const TRAILING_MIN_TEXT_CHARS = 40
|
|
317
|
+
|
|
318
|
+
/**
|
|
319
|
+
* Link-density share above which a trailing block reads as a link farm:
|
|
320
|
+
* recommendation feeds, related-article rails, tag clouds, share bars.
|
|
321
|
+
* Measured on the 163.com dy-article badcase the section's pick carried 84
|
|
322
|
+
* linked characters inside 183 total (0.46) — recommendation cards dilute
|
|
323
|
+
* their density with author/date lines, so a 0.5 bar would miss them by
|
|
324
|
+
* construction; an honest closing paragraph with a citation or two measures
|
|
325
|
+
* 0.1–0.3, leaving 0.4 in the empty band between.
|
|
326
|
+
*/
|
|
327
|
+
const TRAILING_LINK_DENSITY = 0.4
|
|
328
|
+
|
|
329
|
+
/**
|
|
330
|
+
* A trailing block qualifies as a link farm only when it is link-heavy in
|
|
331
|
+
* aggregate too: at least {@link TRAILING_LINK_COUNT} anchors or
|
|
332
|
+
* {@link TRAILING_LINK_TEXT_CHARS} linked characters. Density alone would
|
|
333
|
+
* let a single long linked sentence (an inline citation) drag a legitimate
|
|
334
|
+
* paragraph under the knife.
|
|
335
|
+
*/
|
|
336
|
+
const TRAILING_LINK_COUNT = 3
|
|
337
|
+
const TRAILING_LINK_TEXT_CHARS = 60
|
|
338
|
+
|
|
339
|
+
/**
|
|
340
|
+
* A trailing block may never hold more than this share of the whole
|
|
341
|
+
* article's text. The prune targets *appended* widgets; a block that IS most
|
|
342
|
+
* of the extraction says the scorer picked the widget, which is the coverage
|
|
343
|
+
* gate's failure mode to handle, not this walk's.
|
|
344
|
+
*/
|
|
345
|
+
const TRAILING_MAX_ARTICLE_SHARE = 0.6
|
|
346
|
+
|
|
347
|
+
/**
|
|
348
|
+
* A trailing run may never hold more than this share of its container's
|
|
349
|
+
* text — the aggregate form of the guard above, and the one that keeps
|
|
350
|
+
* list-shaped extractions whole. A recommendation rail is a *minority*
|
|
351
|
+
* appended to real content (the 163 rail is 44% of its container); a
|
|
352
|
+
* portal or feed whose extraction IS a list of link rows (measured on a
|
|
353
|
+
* Hacker News homepage capture: ~91 trailing rows, every one individually
|
|
354
|
+
* link-dense at 0.69–0.94) has a tail run holding ~97% of the container —
|
|
355
|
+
* there the tail IS the content, the walk declines wholesale, and only a
|
|
356
|
+
* run that is a minority of its container is removed.
|
|
357
|
+
*/
|
|
358
|
+
const TRAILING_MAX_RUN_SHARE = 0.6
|
|
359
|
+
|
|
360
|
+
/**
|
|
361
|
+
* Link-density band within which a run and the block the walk stopped at
|
|
362
|
+
* read as the same material. When the content immediately before the run
|
|
363
|
+
* is itself as link-dense as the run — measured on a Hacker News homepage
|
|
364
|
+
* capture, the rows after the run's cut measure 0.94 against the run's
|
|
365
|
+
* 0.91 — the container is a *uniform list* (feed rows, portal headlines)
|
|
366
|
+
* and the tail is that list's own content, not a widget appended after it;
|
|
367
|
+
* a widget hangs after genuinely different material (the 163 rail's 0.44
|
|
368
|
+
* follows body text at 0.00). Only a run denser than what precedes it is
|
|
369
|
+
* removed.
|
|
370
|
+
*/
|
|
371
|
+
const TRAILING_UNIFORM_DENSITY_BAND = 0.25
|
|
372
|
+
|
|
373
|
+
/**
|
|
374
|
+
* Prune trailing link-dense sections from an accepted Readability
|
|
375
|
+
* extraction — the post-article widget class the scorer itself cannot be
|
|
376
|
+
* taught to drop.
|
|
377
|
+
*
|
|
378
|
+
* Why this exists (badcase: `https://www.163.com/dy/article/…`): Chinese
|
|
379
|
+
* portal article pages append a recommendation feed inside the very wrapper
|
|
380
|
+
* Readability scores as the article (`#content` holds the body, the
|
|
381
|
+
* platform disclaimer, then a rail of image/title/author/date cards). Every
|
|
382
|
+
* existing noise layer misses it: Readability's `unlikelyCandidates` /
|
|
383
|
+
* `negative` vocabularies know `related` but not `recommend` (nor the
|
|
384
|
+
* pinyin families `tuijian`/`xiangguan`), and its own cleaning pass draws
|
|
385
|
+
* the "mostly links" line at per-block density 0.5 — these cards measure
|
|
386
|
+
* 0.42–0.48 (the title links over title+author+date text), so every card
|
|
387
|
+
* clears it; in the extracted HTML the rail is a *bare* div — class and id
|
|
388
|
+
* are gone — so no class-based rule can ever see it again; DOMPurify's
|
|
389
|
+
* FORBID_TAGS works on tags, and the rail is plain div/p/h3/a; and the
|
|
390
|
+
* coverage gate cannot help because the noise rides *inside* the
|
|
391
|
+
* extraction it measures (the badcase passed at 0.103 — the rail's text
|
|
392
|
+
* was in both numerator and denominator). Modern sites with CSS-module
|
|
393
|
+
* class names (`css-1b3x7r`) present the same shape with no vocabulary to
|
|
394
|
+
* match at all, so the only signal that survives all of this is
|
|
395
|
+
* structural: a *trailing* run of blocks whose text is mostly links.
|
|
396
|
+
*
|
|
397
|
+
* For every container in the extraction the walk gathers the maximal
|
|
398
|
+
* trailing run of qualifying children — link-farm blocks per the constants
|
|
399
|
+
* above, plus sub-threshold spacers (empty wrapper divs) — and removes the
|
|
400
|
+
* run only when BOTH structural guards say "appended widget, not content":
|
|
401
|
+
* the run is a minority of the container's text
|
|
402
|
+
* ({@link TRAILING_MAX_RUN_SHARE}), and the run is clearly denser than the
|
|
403
|
+
* block the walk stopped at ({@link TRAILING_UNIFORM_DENSITY_BAND}). The
|
|
404
|
+
* guards split the two worlds the per-block view cannot: a rail is
|
|
405
|
+
* one-or-few appended blocks after genuinely different material (the 163
|
|
406
|
+
* rail's 0.44 density follows body text at 0.00), while a link-list
|
|
407
|
+
* extraction — feed rows, portal headlines — is *uniform*: its rows before
|
|
408
|
+
* any cut measure the same density as the rows after it (0.91 vs 0.94 on
|
|
409
|
+
* the Hacker News capture), so no tail of it can be mistaken for a widget.
|
|
410
|
+
*
|
|
411
|
+
* Deliberately NOT applied to the whole-document fallback: that path
|
|
412
|
+
* exists for list-shaped pages where the links *are* the content (the
|
|
413
|
+
* 163.com homepage fix), and pruning link-dense regions there would
|
|
414
|
+
* re-introduce exactly the recall loss the coverage gate exists to
|
|
415
|
+
* prevent.
|
|
416
|
+
*
|
|
417
|
+
* @param container - parsed `article.content`; mutated in place.
|
|
418
|
+
*/
|
|
419
|
+
function pruneTrailingLinkDenseSections(container: Element): void {
|
|
420
|
+
const articleText = visibleTextLength(container)
|
|
421
|
+
if (articleText === 0) return
|
|
422
|
+
// The container itself leads the walk — Readability usually wraps its
|
|
423
|
+
// content in div#readability-page-1, but an unwrapped extraction's
|
|
424
|
+
// top-level blocks are the container's direct children and must be
|
|
425
|
+
// examined too. querySelectorAll snapshots before any removal; entries
|
|
426
|
+
// inside an already-removed subtree are skipped by the contains() check,
|
|
427
|
+
// and parents are visited before children, so a deeper run inside a block
|
|
428
|
+
// an outer run is taking never competes with it.
|
|
429
|
+
for (const element of [container, ...Array.from(container.querySelectorAll('*'))]) {
|
|
430
|
+
if (!container.contains(element) && element !== container) continue
|
|
431
|
+
removeTrailingLinkDenseRun(element, articleText)
|
|
432
|
+
}
|
|
433
|
+
}
|
|
434
|
+
|
|
435
|
+
/**
|
|
436
|
+
* Remove one container's maximal trailing run of link-dense children when
|
|
437
|
+
* that run is a minority of the container's text AND clearly different
|
|
438
|
+
* material from the content that precedes it.
|
|
439
|
+
*
|
|
440
|
+
* @param parent - the container whose element children are examined from
|
|
441
|
+
* the end.
|
|
442
|
+
* @param articleText - the whole article's text, for the per-block share
|
|
443
|
+
* guard.
|
|
444
|
+
*/
|
|
445
|
+
function removeTrailingLinkDenseRun(parent: Element, articleText: number): void {
|
|
446
|
+
const children = Array.from(parent.children)
|
|
447
|
+
if (children.length === 0) return
|
|
448
|
+
const containerText = visibleTextLength(parent)
|
|
449
|
+
if (containerText === 0) return
|
|
450
|
+
const run: Element[] = []
|
|
451
|
+
let runText = 0
|
|
452
|
+
let runLinked = 0
|
|
453
|
+
let stopIndex = -1
|
|
454
|
+
for (let index = children.length - 1; index >= 0; index--) {
|
|
455
|
+
const child = children[index]
|
|
456
|
+
if (child === undefined) break
|
|
457
|
+
const text = visibleTextLength(child)
|
|
458
|
+
// Too small to judge (spacer, empty wrapper): it belongs to the run
|
|
459
|
+
// but cannot extend it — keep walking toward earlier siblings.
|
|
460
|
+
if (text < TRAILING_MIN_TEXT_CHARS) {
|
|
461
|
+
run.push(child)
|
|
462
|
+
continue
|
|
463
|
+
}
|
|
464
|
+
const linked = linkTextLength(child)
|
|
465
|
+
const linkCount = child.querySelectorAll('a').length
|
|
466
|
+
if (
|
|
467
|
+
linked / text >= TRAILING_LINK_DENSITY
|
|
468
|
+
&& (linkCount >= TRAILING_LINK_COUNT || linked >= TRAILING_LINK_TEXT_CHARS)
|
|
469
|
+
&& text <= TRAILING_MAX_ARTICLE_SHARE * articleText
|
|
470
|
+
) {
|
|
471
|
+
run.push(child)
|
|
472
|
+
runText += text
|
|
473
|
+
runLinked += linked
|
|
474
|
+
continue
|
|
475
|
+
}
|
|
476
|
+
// Real content that is not a link farm: everything before it is
|
|
477
|
+
// anchored by actual article flow — stop and compare against it.
|
|
478
|
+
stopIndex = index
|
|
479
|
+
break
|
|
480
|
+
}
|
|
481
|
+
if (runText === 0 || stopIndex === -1) {
|
|
482
|
+
// Nothing to remove — or the run reached the top of the container (no
|
|
483
|
+
// preceding content to differ from): a tail that IS the whole
|
|
484
|
+
// container is list content, never a widget.
|
|
485
|
+
return
|
|
486
|
+
}
|
|
487
|
+
// The tail is the same material as what precedes it (a uniform list's
|
|
488
|
+
// rows): decline wholesale.
|
|
489
|
+
const stopChild = children[stopIndex]
|
|
490
|
+
if (stopChild === undefined) return
|
|
491
|
+
const stopDensity = linkTextLength(stopChild) / Math.max(1, visibleTextLength(stopChild))
|
|
492
|
+
if (Math.abs(runLinked / runText - stopDensity) <= TRAILING_UNIFORM_DENSITY_BAND) return
|
|
493
|
+
// The tail holds most of the container (a link-list page's rows): decline
|
|
494
|
+
// wholesale.
|
|
495
|
+
if (runText > TRAILING_MAX_RUN_SHARE * containerText) return
|
|
496
|
+
for (const child of run) child.remove()
|
|
497
|
+
}
|
|
498
|
+
|
|
499
|
+
/**
|
|
500
|
+
* Run {@link pruneTrailingLinkDenseSections} over one extraction with a
|
|
501
|
+
* misfire guard: a prune that takes a healthy article below the gate's
|
|
502
|
+
* absolute floor ({@link GATE_MIN_ARTICLE_TEXT_CHARS}) or empties it
|
|
503
|
+
* entirely says the walk misread the page, and Readability's original pick
|
|
504
|
+
* is returned untouched. A *sparse* article (under the floor before the
|
|
505
|
+
* prune, auto-trusted by the gate) keeps its pruned form — its trailing
|
|
506
|
+
* links were noise all the same.
|
|
507
|
+
*
|
|
508
|
+
* @param articleHtml - the `article.content` HTML Readability produced.
|
|
509
|
+
* @param document - the live document, for parsing the fragment.
|
|
510
|
+
* @returns the pruned article HTML, or the original on a misfire.
|
|
511
|
+
*/
|
|
512
|
+
function prunedArticleContent(articleHtml: string, document: Document): string {
|
|
513
|
+
const container = document.createElement('div')
|
|
514
|
+
container.innerHTML = articleHtml
|
|
515
|
+
const before = visibleTextLength(container)
|
|
516
|
+
if (before === 0) return articleHtml
|
|
517
|
+
pruneTrailingLinkDenseSections(container)
|
|
518
|
+
const after = visibleTextLength(container)
|
|
519
|
+
if (after === 0 || (before >= GATE_MIN_ARTICLE_TEXT_CHARS && after < GATE_MIN_ARTICLE_TEXT_CHARS)) {
|
|
520
|
+
return articleHtml
|
|
521
|
+
}
|
|
522
|
+
return container.innerHTML
|
|
523
|
+
}
|
|
524
|
+
|
|
177
525
|
/**
|
|
178
526
|
* Convert one rendered HTML document to denoised markdown.
|
|
179
527
|
*
|
|
180
|
-
* Readability failure (non-article pages
|
|
181
|
-
*
|
|
182
|
-
*
|
|
183
|
-
*
|
|
528
|
+
* Readability failure (non-article pages, or extractions the coverage gate
|
|
529
|
+
* {@link articleCoversPage} distrusts) degrades to converting the sanitized
|
|
530
|
+
* whole document — layout tags are still forbidden, so the fallback stays
|
|
531
|
+
* cleaner than raw turndown, and a degraded page beats an error for a body
|
|
532
|
+
* the browser already rendered.
|
|
184
533
|
*
|
|
185
534
|
* @param html - the rendered page HTML (`page.content()`).
|
|
186
535
|
* @param url - the page URL, used to resolve relative links during parsing.
|
|
@@ -204,15 +553,23 @@ export function htmlToMarkdown(html: string, url: string): DenoiseResult {
|
|
|
204
553
|
let source: string | null = null
|
|
205
554
|
let title: string | undefined
|
|
206
555
|
try {
|
|
207
|
-
// Ungated: isProbablyReaderable is a conservative hint that
|
|
208
|
-
// sparse-but-real articles (measured on a browser-rendered fixture),
|
|
209
|
-
// the extraction is simply attempted; a null
|
|
210
|
-
// to the sanitized whole document below. parse() mutates,
|
|
211
|
-
// (typed as Document: DOM lib types cloneNode's return as
|
|
556
|
+
// Ungated readability: isProbablyReaderable is a conservative hint that
|
|
557
|
+
// rejects sparse-but-real articles (measured on a browser-rendered fixture),
|
|
558
|
+
// so the extraction is simply attempted; a null, empty, or gate-rejected
|
|
559
|
+
// result falls back to the sanitized whole document below. parse() mutates,
|
|
560
|
+
// hence the clone (typed as Document: DOM lib types cloneNode's return as
|
|
561
|
+
// Node).
|
|
212
562
|
const cloned = document.cloneNode(true) as typeof document
|
|
213
563
|
const article = new Readability(cloned).parse()
|
|
214
|
-
if (
|
|
215
|
-
|
|
564
|
+
if (
|
|
565
|
+
article !== null && typeof article.content === 'string' && article.content !== ''
|
|
566
|
+
&& articleCoversPage(article.content, document)
|
|
567
|
+
) {
|
|
568
|
+
// Gate first, prune second, and in that order on purpose: the gate
|
|
569
|
+
// must judge Readability's pick as delivered (the widget rides inside
|
|
570
|
+
// it), and the prune then strips the post-article link farms the
|
|
571
|
+
// scorer could not be taught — see pruneTrailingLinkDenseSections.
|
|
572
|
+
source = prunedArticleContent(article.content, document)
|
|
216
573
|
title = article.title ?? undefined
|
|
217
574
|
}
|
|
218
575
|
} catch {
|
package/src/provider.ts
CHANGED
|
@@ -50,6 +50,31 @@
|
|
|
50
50
|
* fails with {@link WEB_FETCH_CHALLENGE_CODE} instead of returning the
|
|
51
51
|
* interstitial as content.
|
|
52
52
|
*
|
|
53
|
+
* Deep capture (viewport-virtualized feeds and app-shell documents): feed-
|
|
54
|
+
* shaped pages — Discourse topic streams, comment threads, timelines —
|
|
55
|
+
* render only the items near the scroll position; everything else is an
|
|
56
|
+
* empty, height-only placeholder that keeps the item's id but none of its
|
|
57
|
+
* content, so a single-instant capture hands back the skeleton while the
|
|
58
|
+
* bodies silently vanish (and the denoise pipeline's coverage gate cannot
|
|
59
|
+
* notice: the page's own visible text is the same windowed fragment as the
|
|
60
|
+
* extraction). App-shell document editors (Feishu docs) push the same
|
|
61
|
+
* pathology one level down: the body never scrolls, blocks are keyed by
|
|
62
|
+
* data-block-id with no id, and off-window blocks are REMOVED behind an
|
|
63
|
+
* anonymous height spacer — the fetch would return just the statically
|
|
64
|
+
* rendered first screen. Before serialization the fetch therefore runs one
|
|
65
|
+
* bounded in-page scroll walk — over the window or, when the document does
|
|
66
|
+
* not scroll, the page's dominant inner scroll container — stashing what
|
|
67
|
+
* each rendered window shows, terminating only AT the bottom with the
|
|
68
|
+
* height stable (that is what pages load-more sentinels; height stability
|
|
69
|
+
* elsewhere proves nothing), then rehydrating degraded keyed elements from
|
|
70
|
+
* their stashed snapshots and splicing removed keys back at their document
|
|
71
|
+
* position so the final DOM holds the union of every window in document
|
|
72
|
+
* order. The existing pipeline then runs unchanged. Bounded by
|
|
73
|
+
* `scrollBudgetMs` (default 20s; 0 restores the legacy single-instant
|
|
74
|
+
* behavior) and by the remaining fetch deadline; any walk failure silently
|
|
75
|
+
* degrades to the legacy capture (see `src/autoscroll.ts` for the measured
|
|
76
|
+
* linux.do and Feishu badcases).
|
|
77
|
+
*
|
|
53
78
|
* @module dsh-web-fetch-playwright/provider
|
|
54
79
|
*/
|
|
55
80
|
|
|
@@ -57,10 +82,12 @@ import { WebError } from '@deepseek-ai/dsh-web'
|
|
|
57
82
|
import type { WebFetchProvider, WebFetchRequest, WebFetchResult } from '@deepseek-ai/dsh-web'
|
|
58
83
|
import { CHALLENGE_DOM_PROBE, CHALLENGE_FINISH_RESERVE_MS, CHALLENGE_POLL_INTERVAL_MS, classifyChallengeHtml, classifyChallengeResponse, isChallengeCompatibleResponse } from './challenge.ts'
|
|
59
84
|
import type { ChallengeVerdict } from './challenge.ts'
|
|
60
|
-
import {
|
|
85
|
+
import { deepCaptureWalk, SCROLL_FINISH_RESERVE_MS } from './autoscroll.ts'
|
|
86
|
+
import { DEFAULT_MAX_CONCURRENCY_CDP, DEFAULT_MAX_CONCURRENCY_LOCAL, effectiveChallengeRetries, effectiveChallengeWaitMs, effectiveContextMode, effectiveMaxConcurrency, effectiveScrollBudgetMs, normalizeCdpEndpoint } from './config.ts'
|
|
61
87
|
import type { ResolvedConfig } from './config.ts'
|
|
62
88
|
import { CdpConnectionPool } from './cdp-pool.ts'
|
|
63
89
|
import { htmlToMarkdown } from './markdown.ts'
|
|
90
|
+
import { STRIP_SCRIPTING_ELEMENTS } from './strip.ts'
|
|
64
91
|
import { resolveCdpBackend, resolvePlaywrightBackend } from './playwright-resolve.ts'
|
|
65
92
|
import type { PlaywrightBrowser, PlaywrightContext, PlaywrightDownload, PlaywrightPage, PlaywrightResponse, PlaywrightRoute } from './types.ts'
|
|
66
93
|
|
|
@@ -535,14 +562,31 @@ export class PlaywrightFetchProvider implements WebFetchProvider {
|
|
|
535
562
|
// what domcontentloaded already produced.
|
|
536
563
|
await page.waitForLoadState('networkidle', { timeout: Math.min(SETTLE_MS, deadline.remainingMs()) }).catch(() => {})
|
|
537
564
|
|
|
538
|
-
//
|
|
539
|
-
//
|
|
540
|
-
//
|
|
541
|
-
//
|
|
542
|
-
//
|
|
543
|
-
|
|
565
|
+
// Deep capture (viewport-virtualized feeds and app-shell documents lose
|
|
566
|
+
// their off-window content at rest — see `src/autoscroll.ts`): a bounded
|
|
567
|
+
// walk stashes what each rendered window shows and, when it restored
|
|
568
|
+
// anything, hands back its own merged, already-stripped serialization.
|
|
569
|
+
// HTML documents only, inside the deadline with a decode reserve,
|
|
570
|
+
// silently degrading to the legacy capture; `scrollBudgetMs: 0` skips it.
|
|
571
|
+
const scrollBudgetMs = Math.min(
|
|
572
|
+
effectiveScrollBudgetMs(config),
|
|
573
|
+
deadline.remainingMs() - SCROLL_FINISH_RESERVE_MS,
|
|
574
|
+
)
|
|
575
|
+
const captured = scrollBudgetMs > 0 ? await deepCaptureWalk(page, scrollBudgetMs) : undefined
|
|
544
576
|
|
|
545
|
-
|
|
577
|
+
// Serialization safety for the live read: the document is about to be
|
|
578
|
+
// flattened into an HTML string and re-parsed by jsdom, and scripting
|
|
579
|
+
// elements are the one place that round-trip can break — see
|
|
580
|
+
// `src/strip.ts` for exactly what is removed, what is kept (JSON-LD,
|
|
581
|
+
// noscript), and why each removal cannot change what the pipeline
|
|
582
|
+
// extracts. (The walk's merged capture went through the same strip.)
|
|
583
|
+
let html: string
|
|
584
|
+
if (captured?.html !== undefined) {
|
|
585
|
+
html = captured.html
|
|
586
|
+
} else {
|
|
587
|
+
await stripScriptingElements(page)
|
|
588
|
+
html = await page.content()
|
|
589
|
+
}
|
|
546
590
|
if (!config.denoise) {
|
|
547
591
|
// The tool layer's own turndown renders raw HTML; the checkbox only
|
|
548
592
|
// governs the Readability/DOMPurify stage this provider owns.
|
|
@@ -907,68 +951,13 @@ async function probeStillOnChallenge(page: PlaywrightPage): Promise<boolean> {
|
|
|
907
951
|
}
|
|
908
952
|
}
|
|
909
953
|
|
|
910
|
-
/**
|
|
911
|
-
* In-page script that removes non-content scripting subtrees from the live
|
|
912
|
-
* document before `page.content()` reads it as an HTML string.
|
|
913
|
-
*
|
|
914
|
-
* WHAT goes, and why each removal is output-neutral versus the old pipeline:
|
|
915
|
-
*
|
|
916
|
-
* - `script` — three consumers ever saw script text before, and none loses
|
|
917
|
-
* it here: Readability's `parse()` deletes script/noscript nodes BEFORE
|
|
918
|
-
* extraction and scoring (so script text never reached the article or
|
|
919
|
-
* influenced the scorer); DOMPurify's default allowlist drops script; the
|
|
920
|
-
* tool layer's renderer removes SCRIPT wholesale on the raw path. What the
|
|
921
|
-
* removal buys is the reason this exists: a page's runtime JS may rewrite
|
|
922
|
-
* a data script's text with a literal `</script>` embedded — measured on
|
|
923
|
-
* techcommunity.microsoft.com (Khoros/Aurora), whose client writes its
|
|
924
|
-
* Apollo cache back into `script#__NEXT_DATA__` via `JSON.stringify`,
|
|
925
|
-
* turning the wire's 712 `\u003c` escapes back into raw `<` — and
|
|
926
|
-
* browsers serialize script text verbatim, so the flattened string ends
|
|
927
|
-
* that script at the FIRST embedded end tag and re-parses the remaining
|
|
928
|
-
* ~270KB of JSON as markup. The phantom elements that survive (unknown
|
|
929
|
-
* tags like `<id>`, anchors full of JSON text) are exactly what
|
|
930
|
-
* Readability's scorer reads as article mass, so the extracted "article"
|
|
931
|
-
* is CSS and i18n strings while the real body — still present in the DOM —
|
|
932
|
-
* loses. No string-level patch can undo this after serialization (the
|
|
933
|
-
* string is genuinely ambiguous); only the browser's own parse holds the
|
|
934
|
-
* document's true structure.
|
|
935
|
-
* - `script[type="application/ld+json"]` — KEPT, but defused: Readability
|
|
936
|
-
* extracts metadata (title/byline/excerpt — e.g. the site-suffix-free
|
|
937
|
-
* "Remote Desktop Protocol" instead of "… - Wikipedia") from JSON-LD
|
|
938
|
-
* scripts before removing them, so deleting them would regress the title
|
|
939
|
-
* on every JSON-LD-carrying page. Rewriting `</` as `<\/` keeps the text
|
|
940
|
-
* valid JSON (`\/` is a legal escape, so the re-parsed text parses to the
|
|
941
|
-
* same object) while making the serialization harmless: the tokenizer
|
|
942
|
-
* only terminates on a literal `</script`.
|
|
943
|
-
* - `noscript` — KEPT: Readability's `_unwrapNoscriptImages` salvages the
|
|
944
|
-
* real image out of lazy-load `<noscript><img …>` wrappers before
|
|
945
|
-
* `_removeScripts` deletes the noscript; pre-stripping it would silently
|
|
946
|
-
* drop those images from the markdown.
|
|
947
|
-
* - `template` — removed: its children parse into the inert `.content`
|
|
948
|
-
* fragment (invisible to Readability's traversal), DOMPurify's
|
|
949
|
-
* FORBID_TAGS drops it, and the tool layer's renderer removes it too.
|
|
950
|
-
* - `style` stays — its text cannot terminate a re-parse except at its own
|
|
951
|
-
* end tag, and the pipeline already drops it.
|
|
952
|
-
*/
|
|
953
|
-
const STRIP_SCRIPTING_ELEMENTS = `(() => {
|
|
954
|
-
for (const el of Array.from(document.querySelectorAll('script, template'))) {
|
|
955
|
-
if (el.tagName === 'SCRIPT' && (el.getAttribute('type') || '').trim().toLowerCase() === 'application/ld+json') {
|
|
956
|
-
// '<\\/' is still valid JSON, so the re-parsed textContent parses to
|
|
957
|
-
// the same object — while the serialized form can no longer close the
|
|
958
|
-
// script element early.
|
|
959
|
-
el.textContent = el.textContent.replace(/<\\//g, '<\\\\/')
|
|
960
|
-
continue
|
|
961
|
-
}
|
|
962
|
-
el.remove()
|
|
963
|
-
}
|
|
964
|
-
})()`
|
|
965
|
-
|
|
966
954
|
/**
|
|
967
955
|
* Remove non-content scripting subtrees from the live document before
|
|
968
|
-
* serialization
|
|
969
|
-
*
|
|
970
|
-
*
|
|
971
|
-
*
|
|
956
|
+
* serialization (see `stripScriptingElementsIn` in `src/strip.ts` for exactly
|
|
957
|
+
* what goes and why). Best-effort, mirroring {@link probeStillOnChallenge}: a backend
|
|
958
|
+
* without `evaluate` (fakes, minimal engines) simply keeps the document
|
|
959
|
+
* as-is — the fetch never fails here, and the pipeline degrades to whatever
|
|
960
|
+
* the serialized string itself says.
|
|
972
961
|
*/
|
|
973
962
|
async function stripScriptingElements(page: PlaywrightPage): Promise<void> {
|
|
974
963
|
if (typeof page.evaluate !== 'function') return
|
package/src/strip.ts
ADDED
|
@@ -0,0 +1,72 @@
|
|
|
1
|
+
/**
|
|
2
|
+
* Serialization safety for documents about to be flattened into an HTML
|
|
3
|
+
* string and re-parsed by jsdom: the one in-page cleanup both serialization
|
|
4
|
+
* paths share — the provider's live-document read and the deep-capture
|
|
5
|
+
* walk's merged clone.
|
|
6
|
+
*
|
|
7
|
+
* @module dsh-web-fetch-playwright/strip
|
|
8
|
+
*/
|
|
9
|
+
|
|
10
|
+
/**
|
|
11
|
+
* In-page function that removes non-content scripting subtrees from a
|
|
12
|
+
* document before it is flattened into an HTML string — the live document
|
|
13
|
+
* before `page.content()`, or the deep-capture walk's merged clone before it
|
|
14
|
+
* serializes it itself (`src/autoscroll.ts`).
|
|
15
|
+
*
|
|
16
|
+
* It is evaluated BY SOURCE (`fn.toString()`), so it must stay
|
|
17
|
+
* self-contained: no imports, no module-level references.
|
|
18
|
+
*
|
|
19
|
+
* WHAT goes, and why each removal is output-neutral versus the old pipeline:
|
|
20
|
+
*
|
|
21
|
+
* - `script` — three consumers ever saw script text before, and none loses
|
|
22
|
+
* it here: Readability's `parse()` deletes script/noscript nodes BEFORE
|
|
23
|
+
* extraction and scoring (so script text never reached the article or
|
|
24
|
+
* influenced the scorer); DOMPurify's default allowlist drops script; the
|
|
25
|
+
* tool layer's renderer removes SCRIPT wholesale on the raw path. What the
|
|
26
|
+
* removal buys is the reason this exists: a page's runtime JS may rewrite
|
|
27
|
+
* a data script's text with a literal `</script>` embedded — measured on
|
|
28
|
+
* techcommunity.microsoft.com (Khoros/Aurora), whose client writes its
|
|
29
|
+
* Apollo cache back into `script#__NEXT_DATA__` via `JSON.stringify`,
|
|
30
|
+
* turning the wire's 712 `\u003c` escapes back into raw `<` — and
|
|
31
|
+
* browsers serialize script text verbatim, so the flattened string ends
|
|
32
|
+
* that script at the FIRST embedded end tag and re-parses the remaining
|
|
33
|
+
* ~270KB of JSON as markup. The phantom elements that survive (unknown
|
|
34
|
+
* tags like `<id>`, anchors full of JSON text) are exactly what
|
|
35
|
+
* Readability's scorer reads as article mass, so the extracted "article"
|
|
36
|
+
* is CSS and i18n strings while the real body — still present in the DOM —
|
|
37
|
+
* loses. No string-level patch can undo this after serialization (the
|
|
38
|
+
* string is genuinely ambiguous); only the browser's own parse holds the
|
|
39
|
+
* document's true structure.
|
|
40
|
+
* - `script[type="application/ld+json"]` — KEPT, but defused: Readability
|
|
41
|
+
* extracts metadata (title/byline/excerpt — e.g. the site-suffix-free
|
|
42
|
+
* "Remote Desktop Protocol" instead of "… - Wikipedia") from JSON-LD
|
|
43
|
+
* scripts before removing them, so deleting them would regress the title
|
|
44
|
+
* on every JSON-LD-carrying page. Rewriting `</` as `<\/` keeps the text
|
|
45
|
+
* valid JSON (`\/` is a legal escape, so the re-parsed text parses to the
|
|
46
|
+
* same object) while making the serialization harmless: the tokenizer
|
|
47
|
+
* only terminates on a literal `</script`.
|
|
48
|
+
* - `noscript` — KEPT: Readability's `_unwrapNoscriptImages` salvages the
|
|
49
|
+
* real image out of lazy-load `<noscript><img …>` wrappers before
|
|
50
|
+
* `_removeScripts` deletes the noscript; pre-stripping it would silently
|
|
51
|
+
* drop those images from the markdown.
|
|
52
|
+
* - `template` — removed: its children parse into the inert `.content`
|
|
53
|
+
* fragment (invisible to Readability's traversal), DOMPurify's
|
|
54
|
+
* FORBID_TAGS drops it, and the tool layer's renderer removes it too.
|
|
55
|
+
* - `style` stays — its text cannot terminate a re-parse except at its own
|
|
56
|
+
* end tag, and the pipeline already drops it.
|
|
57
|
+
*/
|
|
58
|
+
export function stripScriptingElementsIn(doc: Document): void {
|
|
59
|
+
for (const el of Array.from(doc.querySelectorAll('script, template'))) {
|
|
60
|
+
if (el.tagName === 'SCRIPT' && (el.getAttribute('type') ?? '').trim().toLowerCase() === 'application/ld+json') {
|
|
61
|
+
// '<\/' is still valid JSON, so the re-parsed textContent parses to the
|
|
62
|
+
// same object — while the serialized form can no longer close the
|
|
63
|
+
// script element early.
|
|
64
|
+
el.textContent = (el.textContent ?? '').replace(/<\//g, '<\\/')
|
|
65
|
+
continue
|
|
66
|
+
}
|
|
67
|
+
el.remove()
|
|
68
|
+
}
|
|
69
|
+
}
|
|
70
|
+
|
|
71
|
+
/** {@link stripScriptingElementsIn} over the live document, as an evaluate expression. */
|
|
72
|
+
export const STRIP_SCRIPTING_ELEMENTS = `(${stripScriptingElementsIn.toString()})(document)`
|