1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116#!/usr/bin/env node
// Crawl the seed lists and extract every functional image candidate found.
//
// This replaces an agent that searched the web and typed records by hand. The
// work it does is mechanical, so a script does it: fetch the page, find images
// inside interactive elements, compute what each control already announces, slice
// the markup out of the bytes we fetched, and write a row. Nothing here makes a
// judgment. Deciding whether an image is really functional, and whether the alt
// text the site shipped is any good, is left to the review pass, which is the
// only part that needs a model.
//
// The one rule enforced here, from directives/00-corpus-goals.md: an image whose
// control announces nothing is skipped entirely and never written. This corpus
// pairs images with descriptions that were really shipped. A control with no
// name has no description to pair with.
//
// Usage:
// node tools/harvest.mjs crawl every seed list
// node tools/harvest.mjs --sector news crawl one seed list
// node tools/harvest.mjs --pages 40 cap pages fetched per host
// node tools/harvest.mjs --follow 0 seeds only, follow no links
// node tools/harvest.mjs --dry-run report, write nothing
// node tools/harvest.mjs --selftest offline, no network
// --seeds DIR, --pool FILE work somewhere else
//
// Politeness is not optional: robots.txt is honoured, one request at a time, one
// second between requests to the same host, and a cap on pages per host.
//
// Exit codes: 0 candidates were found, 1 nothing was found, 2 refused because
// the pool file or seeds are unusable, 3 bad usage or self-test failure.
import { readFileSync, writeFileSync, appendFileSync, existsSync, readdirSync,
mkdirSync, mkdtempSync, rmSync } from 'node:fs'
import { tmpdir } from 'node:os'
import { dirname, join, resolve, basename } from 'node:path'
import { fileURLToPath } from 'node:url'
import { parse, outerHtml, ancestors, visibleText, imageName, accessibleName,
collapse } from './html.mjs'
const HERE = dirname(fileURLToPath(import.meta.url))
const PROJECT = resolve(HERE, '..')
const USER_AGENT = 'benchmarks-ai-alt corpus harvester ' +
'(+https://github.com/EqualifyEverything/benchmarks-ai-alt)'
const FETCH_TIMEOUT_MS = 20000
const MAX_REDIRECTS = 5
const MAX_PAGE_BYTES = 4 * 1024 * 1024
const HOST_DELAY_MS = 1000
const SURROUNDING_LIMIT = 240
// SVG elements that actually put marks on the canvas. An inline SVG with none of
// them draws nothing on its own, which in practice means its shape comes from a
// `<use>` reference to a sprite symbol defined elsewhere.
const SVG_PAINTS = /<(path|circle|rect|polygon|polyline|line|ellipse|text|image|foreignObject)\b/i
const SVG_NS = 'http://www.w3.org/2000/svg'
const XLINK_NS = 'http://www.w3.org/1999/xlink'
// Roles that make an element interactive for our purposes, mapped to the role we
// record. A custom control with role="button" is as functional as a `<button>`.
const ROLE_MAP = {
link: 'link', button: 'button', checkbox: 'button', switch: 'button',
tab: 'button', menuitem: 'link', menuitemcheckbox: 'button',
radio: 'button', option: 'button',
}
// --- extraction ------------------------------------------------------------
const attr = (node, name) => {
const v = node.attrs[name]
return typeof v === 'string' ? v : null
}
function absolute(url, base) {
if (url === null || url.trim() === '') return null
try {
return new URL(url.trim(), base).href
} catch {
return null
}
}
// The interactive element an image sits inside, or null when there is none.
// `<input type="image">` and `<area>` are their own controls.
function interactiveAncestor(image) {
if (image.tag === 'input' || image.tag === 'area') return image
for (const node of ancestors(image)) {
if (node.tag === 'a') {
const href = attr(node, 'href')
if (href !== null && href.trim() !== '') return node
}
if (node.tag === 'button' || node.tag === 'summary') return node
const role = (attr(node, 'role') ?? '').toLowerCase().trim()
if (role in ROLE_MAP) return node
// Stop at the next block of content. An image ten divs away from a link is
// not inside that link in any meaningful sense.
if (node.tag === 'main' || node.tag === 'body' || node.tag === '#root') break
}
return null
}
function roleOf(control) {
if (control.tag === 'a') return 'link'
if (control.tag === 'button' || control.tag === 'summary') return 'button'
if (control.tag === 'input') return 'input-image'
if (control.tag === 'area') return 'area'
const role = (attr(control, 'role') ?? '').toLowerCase().trim()
return ROLE_MAP[role] ?? 'custom'
}
// A first guess at the taxonomy, from the markup alone. The review pass corrects
// it and assigns the two sub-types that need judgment, which are functional
// emojis and structural or reader controls.
function classify(control, image, pageUrl) {
if (control.tag === 'input') {
return { category: 2, subtype: 'form-control-or-image-button' }
}
if (control.tag === 'area') {
return { category: 4, subtype: 'linked-complex-graphic-or-image-map' }
}
if (control.tag === 'a') {
const target = absolute(attr(control, 'href'), pageUrl)
if (target !== null) {
try {
const to = new URL(target)
const from = new URL(pageUrl)
const atRoot = to.pathname === '/' || to.pathname === ''
// A link to the root of its own or another site, carrying an image, is
// the logo pattern. Anything else is navigation to a document.
if (atRoot && to.search === '' && to.host !== '') {
return {
category: 1,
subtype: to.host === from.host
? 'linked-standalone-logo'
: 'linked-standalone-logo',
}
}
} catch { /* fall through to navigation */ }
}
return { category: 1, subtype: 'standalone-navigational-link' }
}
return { category: 2, subtype: 'action-or-toggle-icon' }
}
// An `<area>` describes a region of the image that declares `usemap`, so the
// picture being described lives on a different element entirely.
function mapOwner(tree, image) {
const map = ancestors(image).find((n) => n.tag === 'map')
if (map === undefined) return null
const name = attr(map, 'name') ?? attr(map, 'id')
if (name === null) return null
return findNode(tree, (n) =>
(n.tag === 'img' || n.tag === 'object') &&
(attr(n, 'usemap') ?? '').replace(/^#/, '') === name)
}
// The pixel space an `<area>`'s `coords` are measured in. Per the HTML
// specification they are CSS pixels of the map image as rendered, not of the
// file, so a page that ships `width="370"` on a 512-pixel-wide image puts the
// region 1.38 times too far right for anyone reading the coordinates against
// the archived bytes. weather.gov/forecastmaps does exactly that, so the
// validation tool needs the rendered size recorded alongside the URL. Only the
// `width` and `height` attributes are readable from markup; a map scaled by a
// stylesheet stays undetectable here, which is why this is `null` rather than a
// guess when the attributes are absent.
function coordSpaceFor(tree, image) {
if (image.tag !== 'area') return null
const owner = mapOwner(tree, image)
if (owner === null) return null
const w = (attr(owner, 'width') ?? '').trim()
const h = (attr(owner, 'height') ?? '').trim()
if (!/^\d+$/.test(w) || !/^\d+$/.test(h)) return null
if (Number(w) === 0 || Number(h) === 0) return null
return `${w}x${h}`
}
// The image URL, resolved against the page. Inline SVG has none: its bytes are
// the markup itself, which tools/fetch-images.mjs writes out.
function imageUrlFor(tree, image, pageUrl) {
if (image.tag === 'svg') return null
if (image.tag === 'area') {
const owner = mapOwner(tree, image)
return owner === null ? null : absolute(attr(owner, 'src'), pageUrl)
}
const src = attr(image, 'src')
if (src !== null && src.trim() !== '') return absolute(src, pageUrl)
// Some responsive images ship only a srcset. Take its first candidate.
const srcset = attr(image, 'srcset')
if (srcset !== null && srcset.trim() !== '') {
const first = srcset.split(',')[0].trim().split(/\s+/)[0]
return absolute(first, pageUrl)
}
return null
}
// The alt text the site actually shipped, which is the thing under review. An
// empty string means `alt=""` was present and deliberate; null means no
// description of any kind was attached to the image element.
function observedAlt(image) {
if (image.tag === 'svg') {
const label = attr(image, 'aria-label')
if (label !== null && collapse(label) !== '') return collapse(label)
const { name, source } = imageName(image)
return source === 'svg-title' ? name : null
}
const alt = attr(image, 'alt')
return alt === null ? null : collapse(alt)
}
function surroundingText(control, image) {
const inside = control === image ? '' : visibleText(control, { skip: image })
if (inside !== '') return inside.slice(0, SURROUNDING_LIMIT)
const parent = control.parent
if (parent === null || parent === undefined) return ''
return visibleText(parent, { skip: control }).slice(0, SURROUNDING_LIMIT)
}
function descendants(node) {
const out = []
const walk = (n) => {
for (const child of n.children) {
if (child.text !== undefined) continue
out.push(child)
walk(child)
}
}
walk(node)
return out
}
// A slice of a page is not a document. Inside HTML an `<svg>` inherits its
// namespaces from the parser and its size from CSS; written to a file on its own
// it has to declare both or it renders as nothing.
function asSvgDocument(svg, viewBox) {
const openEnd = svg.indexOf('>')
const open = svg.slice(0, openEnd)
const has = (name) => new RegExp(`\\s${name}\\s*=`, 'i').test(open)
let add = ''
if (!has('xmlns')) add += ` xmlns="${SVG_NS}"`
if (/xlink:/i.test(svg) && !has('xmlns:xlink')) {
add += ` xmlns:xlink="${XLINK_NS}"`
}
if (!has('viewBox') && typeof viewBox === 'string' && viewBox.trim() !== '') {
add += ` viewBox="${viewBox.trim()}"`
}
return open + add + svg.slice(openEnd)
}
// A standalone SVG document for an inline SVG: its own markup, plus the sprite
// symbols its `<use>` elements point at, copied into a `<defs>` block. This is
// what makes the archived file a picture instead of a blank rectangle, and it is
// why 28 percent of named controls are collectable at all.
//
// Null means the graphic cannot be rebuilt from this page: the sprite lives in
// another file, or the element paints nothing and references nothing. Unlike
// `element_html`, which is always a character-exact slice, the archived image is
// allowed to be assembled. It has to be: an SVG needs its namespace declared.
export function standaloneSvg(tree, node) {
const html = outerHtml(tree, node)
const openEnd = html.indexOf('>')
if (!/^<svg[\s/>]/i.test(html) || openEnd === -1) return null
if (html.lastIndexOf('</svg>') < openEnd) return null
const defs = []
let viewBox = null
for (const el of descendants(node)) {
if (el.tag !== 'use') continue
const ref = (attr(el, 'href') ?? attr(el, 'xlink:href') ?? '').trim()
// An external reference, `href="/sprite.svg#icon"`, needs a request we are
// not making and bytes this page never carried.
if (!ref.startsWith('#') || ref.length < 2) return null
const target = tree.byId.get(ref.slice(1))
if (target === undefined) return null
const symbol = outerHtml(tree, target)
if (!SVG_PAINTS.test(symbol)) return null
// The symbol carries the coordinate space the icon was drawn in. Without it
// the shape lands in a corner of a default-sized canvas.
if (viewBox === null) viewBox = attr(target, 'viewbox')
if (!defs.includes(symbol)) defs.push(symbol)
}
if (defs.length === 0) {
return SVG_PAINTS.test(html) ? asSvgDocument(html, null) : null
}
const withDefs = html.slice(0, openEnd + 1) +
`<defs>${defs.join('')}</defs>` + html.slice(openEnd + 1)
return asSvgDocument(withDefs, viewBox)
}
function findNode(tree, pred) {
let hit = null
const walk = (n) => {
for (const child of n.children) {
if (child.text !== undefined) continue
if (hit === null && pred(child)) hit = child
if (hit === null) walk(child)
}
}
walk(tree.root)
return hit
}
function eachElement(tree, visit) {
const walk = (n) => {
for (const child of n.children) {
if (child.text !== undefined) continue
visit(child)
walk(child)
}
}
walk(tree.root)
}
const IMPLEMENTATION = {
img: 'img', svg: 'inline-svg', input: 'input-image', area: 'area',
}
function isImage(node) {
if (node.tag === 'img' || node.tag === 'area') return true
if (node.tag === 'svg') return true
if (node.tag === 'input') {
return (attr(node, 'type') ?? '').toLowerCase().trim() === 'image'
}
return false
}
// Pull every candidate out of one page. Pure: no network, no filesystem, which
// is what lets the self-test run the real code path against fixtures.
export function harvestDoc(doc, pageUrl, sector, retrieved) {
const tree = parse(doc)
const found = []
const skipped = { noControl: 0, noName: 0, noImage: 0, spriteMissing: 0,
duplicateOnPage: 0 }
const seen = new Set()
let domain
try {
domain = new URL(pageUrl).host.toLowerCase().replace(/^www\./, '')
} catch {
return { items: [], skipped, error: `unusable page URL ${pageUrl}` }
}
eachElement(tree, (node) => {
if (!isImage(node)) return
// An inline SVG nested in another SVG is one graphic, not two.
if (node.tag === 'svg' && ancestors(node).some((a) => a.tag === 'svg')) return
const control = interactiveAncestor(node)
if (control === null) { skipped.noControl++; return }
const { name, source } = accessibleName(tree, control, node)
if (name === '') { skipped.noName++; return }
const imageUrl = imageUrlFor(tree, node, pageUrl)
if (imageUrl === null && node.tag !== 'svg') { skipped.noImage++; return }
// An inline SVG is archived from the page's own bytes, sprite symbols and
// all. When that cannot be done the graphic is a blank rectangle, and a
// blank rectangle cannot be judged against its alt text, so it is not an
// item. Same reason directive 00 leaves icon fonts out.
let svg = null
if (node.tag === 'svg') {
svg = standaloneSvg(tree, node)
if (svg === null) { skipped.spriteMissing++; return }
}
const role = roleOf(control)
const key = `${imageUrl ?? outerHtml(tree, node)}|${role}|${name}`
if (seen.has(key)) { skipped.duplicateOnPage++; return }
seen.add(key)
const { category, subtype } = classify(control, node, pageUrl)
found.push({
id: null,
status: 'unreviewed',
page_url: pageUrl,
domain,
sector,
image_url: imageUrl,
image_coord_space: coordSpaceFor(tree, node),
image_svg: svg,
image_file: null,
image_sha256: null,
implementation: IMPLEMENTATION[node.tag],
element_role: role,
element_html: outerHtml(tree, control),
surrounding_text: surroundingText(control, node),
observed_alt: observedAlt(node),
accessible_name: name,
accessible_name_source: source,
category,
subtype,
retrieved,
review_verdict: null,
review_reason: null,
alt_quality: null,
})
})
return { items: found, skipped }
}
// Same-host links worth following from a page, in document order.
export function outboundLinks(doc, pageUrl) {
const tree = parse(doc)
let here
try { here = new URL(pageUrl) } catch { return [] }
const out = []
const seen = new Set()
eachElement(tree, (node) => {
if (node.tag !== 'a') return
const href = absolute(attr(node, 'href'), pageUrl)
if (href === null) return
let url
try { url = new URL(href) } catch { return }
if (url.host !== here.host) return
if (url.protocol !== 'http:' && url.protocol !== 'https:') return
if (/\.(pdf|zip|jpg|jpeg|png|gif|svg|mp4|mp3|css|js|xml|json)$/i
.test(url.pathname)) return
url.hash = ''
if (url.href === here.href || seen.has(url.href)) return
seen.add(url.href)
out.push(url.href)
})
return out
}
// --- robots.txt ------------------------------------------------------------
// Enough of the exclusion standard to be a good citizen: the groups that apply
// to us, longest matching rule wins, Allow beats Disallow at equal length. A
// robots.txt we cannot fetch is treated as permissive, which is the convention.
export function parseRobots(text) {
const groups = []
let current = null
for (const raw of text.split('\n')) {
const line = raw.replace(/#.*$/, '').trim()
if (line === '') continue
const colon = line.indexOf(':')
if (colon === -1) continue
const field = line.slice(0, colon).trim().toLowerCase()
const value = line.slice(colon + 1).trim()
if (field === 'user-agent') {
if (current === null || current.rules.length > 0) {
current = { agents: [], rules: [] }
groups.push(current)
}
current.agents.push(value.toLowerCase())
} else if ((field === 'allow' || field === 'disallow') && current !== null) {
current.rules.push({ allow: field === 'allow', path: value })
}
}
const mine = groups.filter((g) => g.agents.some((a) =>
a === '*' || USER_AGENT.toLowerCase().includes(a)))
const specific = groups.filter((g) => g.agents.some((a) =>
a !== '*' && USER_AGENT.toLowerCase().includes(a)))
const rules = (specific.length > 0 ? specific : mine).flatMap((g) => g.rules)
return {
allows(path) {
let best = null
for (const rule of rules) {
if (rule.path === '') continue
if (!matchesRobotsPath(rule.path, path)) continue
if (best === null || rule.path.length > best.path.length ||
(rule.path.length === best.path.length && rule.allow)) {
best = rule
}
}
return best === null ? true : best.allow
},
}
}
function matchesRobotsPath(pattern, path) {
const anchored = pattern.endsWith('$')
const body = anchored ? pattern.slice(0, -1) : pattern
const parts = body.split('*')
let at = 0
for (let i = 0; i < parts.length; i++) {
const part = parts[i]
if (part === '') continue
const found = i === 0 ? (path.startsWith(part) ? 0 : -1) : path.indexOf(part, at)
if (found === -1) return false
at = found + part.length
}
return anchored ? at === path.length : true
}
// --- seeds and pool --------------------------------------------------------
export function loadSeeds(seedsDir, only) {
if (!existsSync(seedsDir)) return { seeds: [], errors: [`no seeds at ${seedsDir}`] }
const errors = []
const seeds = []
const files = readdirSync(seedsDir).filter((n) => n.endsWith('.txt')).sort()
for (const file of files) {
const sector = basename(file, '.txt')
if (only !== null && sector !== only) continue
const text = readFileSync(join(seedsDir, file), 'utf8')
text.split('\n').forEach((raw, i) => {
const line = raw.replace(/#.*$/, '').trim()
if (line === '') return
try {
const url = new URL(line)
if (url.protocol !== 'http:' && url.protocol !== 'https:') throw new Error()
seeds.push({ url: url.href, sector })
} catch {
errors.push(`${file}:${i + 1}: not an http URL, "${line}"`)
}
})
}
if (only !== null && seeds.length === 0 && files.length > 0) {
errors.push(`no seed list named ${only}.txt`)
}
return { seeds, errors }
}
function readPool(poolPath) {
if (!existsSync(poolPath)) return { rows: [], errors: [] }
const rows = []
const errors = []
readFileSync(poolPath, 'utf8').split('\n').forEach((raw, i) => {
const line = raw.trim()
if (line === '') return
try { rows.push(JSON.parse(line)) } catch (e) {
errors.push(`${poolPath}:${i + 1}: invalid JSON, ${e.message}`)
}
})
return { rows, errors }
}
const nextIdFrom = (rows) => {
let highest = 0
for (const row of rows) {
const m = /^fi-(\d{4,})$/.exec(typeof row.id === 'string' ? row.id : '')
if (m) highest = Math.max(highest, Number(m[1]))
}
return highest + 1
}
const formatId = (n) => `fi-${String(n).padStart(4, '0')}`
// A candidate already in the pool. Keyed on the image and the control's name, so
// re-crawling a page adds what changed and nothing else.
const poolKey = (row) =>
`${row.page_url}|${row.image_url ?? row.element_html}|${row.element_role}|` +
`${row.accessible_name}`
// --- network ---------------------------------------------------------------
const sleep = (ms) => new Promise((done) => setTimeout(done, ms))
async function getText(url, { manual = false } = {}) {
let res
try {
res = await fetch(url, {
redirect: manual ? 'manual' : 'follow',
headers: { 'user-agent': USER_AGENT, accept: 'text/html,*/*' },
signal: AbortSignal.timeout(FETCH_TIMEOUT_MS),
})
} catch (e) {
return { error: `request failed, ${e.message}` }
}
// A redirect is reported rather than followed, so the caller can consult the
// destination's robots.txt before asking for it. Following redirects blindly is
// how a crawler ends up on a host that never gave it permission: a link on one
// retailer's site redirecting to a sibling brand's, for one real example.
if (manual && res.status >= 300 && res.status < 400) {
const to = absolute(res.headers.get('location'), url)
if (to === null) return { error: `HTTP ${res.status} with no usable location` }
return { redirect: to }
}
if (!res.ok) return { error: `HTTP ${res.status}` }
const type = (res.headers.get('content-type') ?? '').toLowerCase()
if (type !== '' && !type.includes('html') && !type.includes('xml')) {
return { error: `not HTML, content type ${type.split(';')[0]}` }
}
const bytes = Buffer.from(await res.arrayBuffer())
if (bytes.length > MAX_PAGE_BYTES) {
return { error: `${bytes.length} bytes, over the ${MAX_PAGE_BYTES} limit` }
}
// res.url is the URL after redirects, which is the one the markup is relative
// to. Recording the pre-redirect URL would break every relative image path.
return { doc: bytes.toString('utf8'), finalUrl: res.url || url }
}
// Walk a redirect chain one hop at a time, asking permission before each hop.
// Pulled out and given its collaborators as arguments so the self-test can drive
// the real sequencing with no network.
export async function followRedirects(start, { allowed, get, max = MAX_REDIRECTS }) {
let url = start
const chain = [start]
for (let hop = 0; hop <= max; hop++) {
const permission = await allowed(url)
if (permission !== true) return { error: permission, chain }
const got = await get(url)
if (got.redirect === undefined) return { ...got, chain }
if (chain.includes(got.redirect)) return { error: 'redirect loop', chain }
url = got.redirect
chain.push(url)
}
return { error: `more than ${max} redirects`, chain }
}
function makeFetcher() {
const lastAt = new Map()
const robots = new Map()
const fetcher = {
// True, or the reason this URL is off limits. Consulted once per hop, so a
// redirect to another host is checked against that host's rules.
async allowed(url) {
let at
try { at = new URL(url) } catch { return 'unusable URL' }
const origin = at.origin
if (!robots.has(origin)) {
await fetcher.wait(origin)
const got = await getText(`${origin}/robots.txt`)
robots.set(origin, got.error !== undefined
? parseRobots('')
: parseRobots(got.doc))
}
return robots.get(origin).allows(at.pathname + at.search)
? true
: 'disallowed by robots.txt'
},
async page(url) {
return followRedirects(url, {
allowed: (at) => fetcher.allowed(at),
get: async (at) => {
await fetcher.wait(new URL(at).origin)
return getText(at, { manual: true })
},
})
},
async wait(host) {
const last = lastAt.get(host)
if (last !== undefined) {
const since = Date.now() - last
if (since < HOST_DELAY_MS) await sleep(HOST_DELAY_MS - since)
}
lastAt.set(host, Date.now())
},
}
return fetcher
}
// --- crawl -----------------------------------------------------------------
export async function crawl(seeds, fetcher, opts) {
const { pagesPerHost, follow, retrieved, onPage } = opts
const queues = new Map()
for (const seed of seeds) {
let host
try { host = new URL(seed.url).host } catch { continue }
if (!queues.has(host)) queues.set(host, { pending: [], sector: seed.sector })
queues.get(host).pending.push(seed.url)
}
const items = []
const totals = { pages: 0, failed: 0, disallowed: 0 }
const skipped = { noControl: 0, noName: 0, noImage: 0, spriteMissing: 0,
duplicateOnPage: 0 }
for (const [host, queue] of queues) {
const done = new Set()
let followed = 0
while (queue.pending.length > 0 && done.size < pagesPerHost) {
const url = queue.pending.shift()
if (done.has(url)) continue
done.add(url)
const got = await fetcher.page(url)
if (got.error !== undefined) {
if (got.error.includes('robots')) totals.disallowed++
else totals.failed++
if (onPage) onPage({ url, error: got.error })
continue
}
totals.pages++
const pageUrl = got.finalUrl ?? url
// A redirect means we asked for one URL and read another. Marking the
// destination done as well stops us fetching the same page twice when a
// link points at where the redirect landed, which is common: /docs/stable/
// redirecting to /docs/6.1/ and the page then linking to /docs/6.1/.
done.add(pageUrl)
const result = harvestDoc(got.doc, pageUrl, queue.sector, retrieved)
for (const key of Object.keys(skipped)) skipped[key] += result.skipped[key]
items.push(...result.items)
if (onPage) onPage({ url: pageUrl, found: result.items.length })
if (followed < follow) {
for (const link of outboundLinks(got.doc, pageUrl)) {
if (done.has(link) || queue.pending.includes(link)) continue
queue.pending.push(link)
followed++
if (followed >= follow) break
}
}
}
void host
}
return { items, totals, skipped }
}
// --- self-test -------------------------------------------------------------
const FIXTURE_PAGES = {
'shop.html': `<!doctype html><html><body>
<svg hidden><symbol id="i-search" viewBox="0 0 16 16"><path d="M1 1 8 8"/></symbol></svg>
<header>
<a href="/"><img src="/img/logo.png" alt="Northwind"></a>
<a href="/cart"><img src="/img/cart.svg" alt="Cart, 3 items"></a>
<a href="/help"><img src="/img/q.svg" alt="">Help centre</a>
<a href="/x"><img src="/img/unnamed.svg"></a>
<button type="button"><svg viewBox="0 0 2 2"><title>Close</title><path d="M0 0 2 2"/></svg></button>
<button type="button"><svg class="i" aria-label="Search site"><use xlink:href="#i-search"></use></svg></button>
<button type="button"><svg class="i" aria-label="Filter"><use href="/sprite.svg#i-filter"></use></svg></button>
<form action="/s"><input type="image" src="/img/go.gif" alt="Search"></form>
</header>
<p>Body text with <a href="/deep/page">a text link</a> and no image.</p>
<img src="/img/decorative.png" alt="A field of wheat">
<map name="regions">
<area shape="rect" coords="0,0,9,9" href="/north" alt="North region">
<area shape="rect" coords="9,0,18,9" href="/south" alt="South region">
</map>
<img src="/img/map.png" usemap="#regions" alt="Sales regions" width="180" height="90">
<map name="zones">
<area shape="rect" coords="0,0,9,9" href="/east" alt="East region">
</map>
<img src="/img/map2.png" usemap="#zones" alt="Zones" width="100%">
</body></html>`,
'dupes.html': `<!doctype html><html><body>
<a href="/pay"><img src="/b.gif" alt="Buy now"></a>
<a href="/pay"><img src="/b.gif" alt="Buy now"></a>
<a href="/pay"><img src="/b.gif" alt="Buy now"></a>
</body></html>`,
}
async function selftest() {
let failures = 0
const check = (name, cond, detail) => {
if (cond) process.stdout.write(`PASS ${name}\n`)
else { process.stdout.write(`FAIL ${name}: ${detail ?? ''}\n`); failures++ }
}
const dir = mkdtempSync(join(tmpdir(), 'alt-harvest-'))
const page = harvestDoc(FIXTURE_PAGES['shop.html'],
'https://shop.example.com/', 'commerce', '2026-08-27')
const by = (id) => page.items.find((i) => i.accessible_name === id)
check('every named functional image is found, and nothing else',
page.items.length === 9,
`found ${page.items.length}: ${page.items.map((i) => i.accessible_name).join(' | ')}`)
check('an unnamed control is skipped, not recorded',
page.skipped.noName === 1 && by('') === undefined,
`noName ${page.skipped.noName}`)
// The sprite pattern is 28 percent of the named controls a real harvest finds.
// An earlier version skipped all of it, and the version before that archived
// every one of them as a blank rectangle.
const sprite = by('Search site')
check('a sprite defined on the page is resolved into a standalone graphic',
sprite !== undefined && sprite.image_svg !== null &&
sprite.image_svg.includes('xmlns="http://www.w3.org/2000/svg"') &&
sprite.image_svg.includes('xmlns:xlink=') &&
sprite.image_svg.includes('viewBox="0 0 16 16"') &&
sprite.image_svg.includes('<defs><symbol id="i-search"') &&
SVG_PAINTS.test(sprite.image_svg),
JSON.stringify(sprite?.image_svg))
check('element_html stays a slice even when the graphic is assembled',
sprite !== undefined &&
FIXTURE_PAGES['shop.html'].includes(sprite.element_html) &&
!FIXTURE_PAGES['shop.html'].includes(sprite.image_svg),
JSON.stringify(sprite?.element_html))
check('a sprite in another file is not collected',
page.skipped.spriteMissing === 1 && by('Filter') === undefined,
`spriteMissing ${page.skipped.spriteMissing}`)
check('an image outside any control is skipped',
page.skipped.noControl >= 1 &&
page.items.every((i) => i.accessible_name !== 'A field of wheat'),
`noControl ${page.skipped.noControl}`)
check('element_html is an exact slice of the page source',
page.items.every((i) => FIXTURE_PAGES['shop.html'].includes(i.element_html)),
page.items.map((i) => i.element_html).join(' ~ '))
const logo = by('Northwind')
check('a link to the site root is classified as a logo',
logo !== undefined && logo.category === 1 &&
logo.subtype === 'linked-standalone-logo' &&
logo.image_url === 'https://shop.example.com/img/logo.png' &&
logo.observed_alt === 'Northwind' && logo.accessible_name_source === 'alt',
JSON.stringify(logo))
const cart = by('Cart, 3 items')
check('a link to a document is classified as navigation',
cart !== undefined && cart.subtype === 'standalone-navigational-link',
JSON.stringify(cart?.subtype))
const help = by('Help centre')
check('an empty alt named by the link text is kept, with the text as source',
help !== undefined && help.observed_alt === '' &&
help.accessible_name_source === 'control-text' &&
help.surrounding_text === 'Help centre',
JSON.stringify(help))
const close = by('Close')
check('an inline SVG in a button has no image URL and keeps its markup',
close !== undefined && close.implementation === 'inline-svg' &&
close.image_url === null && close.subtype === 'action-or-toggle-icon' &&
close.observed_alt === 'Close' && close.element_html.includes('<svg'),
JSON.stringify(close))
const search = by('Search')
check('an image button is its own control',
search !== undefined && search.implementation === 'input-image' &&
search.element_role === 'input-image' &&
search.subtype === 'form-control-or-image-button' &&
search.image_url === 'https://shop.example.com/img/go.gif',
JSON.stringify(search))
const north = by('North region')
check('an area takes its image from the element declaring usemap',
north !== undefined && north.implementation === 'area' &&
north.category === 4 &&
north.image_url === 'https://shop.example.com/img/map.png',
JSON.stringify(north))
check('an area records the rendered size its coords are measured in',
north !== undefined && north.image_coord_space === '180x90',
JSON.stringify(north?.image_coord_space))
check('a map sized by percentage records no coord space rather than a guess',
by('East region')?.image_coord_space === null,
JSON.stringify(by('East region')?.image_coord_space))
check('coord space is an image-map concern only',
by('Cart, 3 items')?.image_coord_space === null,
JSON.stringify(by('Cart, 3 items')?.image_coord_space))
const dupes = harvestDoc(FIXTURE_PAGES['dupes.html'],
'https://pay.example.com/', 'commerce', '2026-08-27')
check('identical controls on one page collapse to a single candidate',
dupes.items.length === 1 && dupes.skipped.duplicateOnPage === 2,
`${dupes.items.length} item(s), ${dupes.skipped.duplicateOnPage} folded`)
// robots.txt, including the longest-match rule that decides real cases.
{
const robots = parseRobots([
'User-agent: *', 'Disallow: /private', 'Allow: /private/public',
'Disallow: /*.json$',
].join('\n'))
check('robots.txt allows, disallows and prefers the longest match',
robots.allows('/open') && !robots.allows('/private/x') &&
robots.allows('/private/public/y') && !robots.allows('/a/b.json') &&
robots.allows('/a/b.json.html'),
'see rules')
}
{
const robots = parseRobots('User-agent: *\nDisallow: /\n')
check('a blanket disallow blocks everything', !robots.allows('/anything'))
}
{
check('an unreachable robots.txt is treated as permissive',
parseRobots('').allows('/anything'))
}
// Seed loading, including a bad line being reported rather than crawled.
{
const seedsDir = join(dir, 'seeds')
mkdirSync(seedsDir, { recursive: true })
writeFileSync(join(seedsDir, 'news.txt'),
'# a comment\nhttps://a.example.com/\n\nnot a url\nhttps://b.example.com/x\n')
writeFileSync(join(seedsDir, 'docs.txt'), 'https://c.example.com/\n')
const all = loadSeeds(seedsDir, null)
const one = loadSeeds(seedsDir, 'news')
check('seed lists load, carry their sector, and report bad lines',
all.seeds.length === 3 && all.errors.length === 1 &&
one.seeds.length === 2 && one.seeds[0].sector === 'news',
`${all.seeds.length} seeds, ${all.errors.join('; ')}`)
}
// The crawl: robots respected, per-host cap obeyed, links followed once.
{
const fetched = []
const fetcher = {
async page(url) {
fetched.push(url)
if (url.includes('/blocked')) return { error: 'disallowed by robots.txt' }
if (url.includes('/gone')) return { error: 'HTTP 404' }
return {
doc: '<a href="/next"><img src="/i.png" alt="Next page"></a>' +
'<a href="/other">text</a>',
finalUrl: url,
}
},
}
const r = await crawl([
{ url: 'https://h.example.com/', sector: 'docs' },
{ url: 'https://h.example.com/blocked', sector: 'docs' },
{ url: 'https://h.example.com/gone', sector: 'docs' },
], fetcher, { pagesPerHost: 4, follow: 1, retrieved: '2026-08-27' })
check('the crawl respects robots, counts failures, and follows one link',
r.totals.disallowed === 1 && r.totals.failed === 1 && r.totals.pages === 2 &&
r.items.length === 2 && fetched.length === 4,
JSON.stringify(r.totals) + ` fetched ${fetched.length}`)
}
{
const fetcher = {
async page(url) {
return {
doc: '<a href="/a"><img src="/i.png" alt="A"></a>' +
'<a href="/b">b</a><a href="/c">c</a><a href="/d">d</a>',
finalUrl: url,
}
},
}
const r = await crawl([{ url: 'https://cap.example.com/', sector: 'docs' }],
fetcher, { pagesPerHost: 2, follow: 10, retrieved: '2026-08-27' })
check('the per-host page cap stops the crawl',
r.totals.pages === 2, `${r.totals.pages} pages`)
}
{
// Redirects are walked one hop at a time so permission is asked before each
// one. A crawler that follows redirects blindly ends up on hosts that never
// gave it permission.
const asked = []
const allowed = async (url) => {
asked.push(url)
return url.includes('//off.example.com') ? 'disallowed by robots.txt' : true
}
const get = async (url) => url.endsWith('/here')
? { doc: '<p>arrived</p>', finalUrl: url }
: { redirect: 'https://a.example.com/here' }
const r = await followRedirects('https://a.example.com/start',
{ allowed, get })
check('a redirect is followed and permission asked at every hop',
r.doc === '<p>arrived</p>' && asked.length === 2 &&
r.chain.length === 2 && r.finalUrl === 'https://a.example.com/here',
JSON.stringify({ asked, chain: r.chain }))
const off = await followRedirects('https://a.example.com/start', {
allowed,
get: async () => ({ redirect: 'https://off.example.com/here' }),
})
check('a redirect to a host that refuses crawlers is not fetched',
off.error === 'disallowed by robots.txt' && off.doc === undefined,
JSON.stringify(off))
const loop = await followRedirects('https://a.example.com/one', {
allowed,
get: async (url) => ({
redirect: url.endsWith('/one')
? 'https://a.example.com/two'
: 'https://a.example.com/one',
}),
})
check('a redirect loop stops rather than spinning',
loop.error === 'redirect loop', JSON.stringify(loop))
const long = await followRedirects('https://a.example.com/0', {
allowed,
get: async (url) => ({
redirect: `https://a.example.com/${Number(url.split('/').pop()) + 1}`,
}),
})
check('a redirect chain is bounded',
long.error === `more than ${MAX_REDIRECTS} redirects`, JSON.stringify(long))
}
{
// A seed that redirects, to a page that then links to where the redirect
// landed. Fetching that twice is rude and yields nothing but duplicates.
const fetched = []
const fetcher = {
async page(url) {
fetched.push(url)
const landed = 'https://r.example.com/v2/'
return {
doc: `<a href="/v2/"><img src="/i.png" alt="Version 2"></a>`,
finalUrl: url === 'https://r.example.com/latest/' ? landed : url,
}
},
}
const r = await crawl([{ url: 'https://r.example.com/latest/', sector: 'docs' }],
fetcher, { pagesPerHost: 4, follow: 4, retrieved: '2026-08-27' })
check('a page reached by redirect is not fetched again by its own link',
fetched.length === 1 && r.totals.pages === 1,
`fetched ${JSON.stringify(fetched)}`)
}
rmSync(dir, { recursive: true, force: true })
process.stdout.write(failures === 0
? '\nharvest self-test passed\n'
: `\nharvest self-test failed, ${failures} case(s)\n`)
return failures === 0 ? 0 : 3
}
// --- entry point -----------------------------------------------------------
const today = () => new Date().toISOString().slice(0, 10)
async function main(argv) {
let seedsDir = join(PROJECT, 'seeds')
let poolPath = join(PROJECT, 'pool', 'candidates.jsonl')
let only = null
let pagesPerHost = 12
let follow = 8
let dryRun = false
for (let i = 0; i < argv.length; i++) {
const arg = argv[i]
if (arg === '--selftest') return selftest()
else if (arg === '--dry-run') dryRun = true
else if (arg === '--sector') only = argv[++i] ?? null
else if (arg === '--pages') pagesPerHost = Number(argv[++i])
else if (arg === '--follow') follow = Number(argv[++i])
else if (arg === '--seeds') seedsDir = resolve(argv[++i] ?? '')
else if (arg === '--pool') poolPath = resolve(argv[++i] ?? '')
else {
process.stderr.write(`harvest.mjs: unknown argument "${arg}"\n` +
'usage: harvest.mjs [--sector NAME] [--pages N] [--follow N] ' +
'[--dry-run] [--seeds DIR] [--pool FILE] [--selftest]\n')
return 3
}
}
if (!Number.isFinite(pagesPerHost) || pagesPerHost < 1 ||
!Number.isFinite(follow) || follow < 0) {
process.stderr.write('harvest.mjs: --pages must be 1 or more and --follow 0 ' +
'or more\n')
return 3
}
const { seeds, errors: seedErrors } = loadSeeds(seedsDir, only)
for (const e of seedErrors) process.stdout.write(` ${e}\n`)
if (seeds.length === 0) {
process.stdout.write('no seeds to crawl. Add URLs to seeds/SECTOR.txt, one ' +
'per line.\n')
return 2
}
const { rows: existing, errors: poolErrors } = readPool(poolPath)
if (poolErrors.length > 0) {
process.stdout.write('refused to harvest, the pool file is unreadable:\n')
for (const e of poolErrors) process.stdout.write(` ${e}\n`)
return 2
}
const hosts = new Set(seeds.map((s) => {
try { return new URL(s.url).host } catch { return '?' }
}))
process.stdout.write(`harvest: ${seeds.length} seed(s) across ${hosts.size} ` +
`host(s), up to ${pagesPerHost} page(s) each, following up to ${follow} ` +
'link(s) per host\n\n')
const { items, totals, skipped } = await crawl(seeds, makeFetcher(), {
pagesPerHost,
follow,
retrieved: today(),
onPage: ({ url, error, found }) => {
process.stdout.write(error !== undefined
? ` skip ${url} ${error}\n`
: ` ${String(found).padStart(4)} ${url}\n`)
},
})
const known = new Set(existing.map(poolKey))
const fresh = []
let already = 0
for (const item of items) {
if (known.has(poolKey(item))) { already++; continue }
known.add(poolKey(item))
fresh.push(item)
}
let next = nextIdFrom(existing)
for (const item of fresh) item.id = formatId(next++)
process.stdout.write(`\npages fetched: ${totals.pages}\n`)
process.stdout.write(`pages failed: ${totals.failed}\n`)
process.stdout.write(`pages disallowed: ${totals.disallowed}\n`)
process.stdout.write(`candidates found: ${items.length}\n`)
process.stdout.write(`already in pool: ${already}\n`)
process.stdout.write(`new candidates: ${fresh.length}\n`)
process.stdout.write('\nskipped while reading the pages:\n')
process.stdout.write(` no interactive ancestor: ${skipped.noControl}\n`)
process.stdout.write(` no accessible name: ${skipped.noName}\n`)
process.stdout.write(` no image URL: ${skipped.noImage}\n`)
process.stdout.write(` sprite not on the page: ${skipped.spriteMissing}\n`)
process.stdout.write(` duplicate on the page: ${skipped.duplicateOnPage}\n`)
if (fresh.length === 0) {
process.stdout.write('\nnothing new. Add seeds, or raise --pages.\n')
return 1
}
if (dryRun) {
process.stdout.write('\nnothing written, --dry-run\n')
return 0
}
mkdirSync(dirname(poolPath), { recursive: true })
const lines = fresh.map((r) => JSON.stringify(r)).join('\n') + '\n'
if (existsSync(poolPath)) appendFileSync(poolPath, lines)
else writeFileSync(poolPath, lines)
process.stdout.write(`\nappended ${fresh.length} candidate(s) to ` +
`${poolPath.replace(PROJECT + '/', '')}\n`)
process.stdout.write('Next: node tools/fetch-images.mjs, then ' +
'node tools/select.mjs\n')
return 0
}
process.exit(await main(process.argv.slice(2)))