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// Stage 3: turn a directory of captured runs into the three numbers the campaign
// exists to produce — how often Iris succeeds, how long it takes, what it costs —
// plus the failure inventory that says what to fix first.
//
// Two things about accuracy are worth stating up front, because they decide what
// this report can and cannot claim.
//
// There is no ground truth. Nobody has hand-authored the correct accessible HTML
// for 2000 PDFs, so "accuracy" here is not agreement with a reference. It is the
// conjunction of signals Iris's own pipeline produces and one independent check:
//
// * The FINAL axe lint, re-run locally on the delivered HTML (see lint.mjs). This
// is the only accuracy number in the report that Iris did not compute — the
// deployment's own final lint goes to the aggregate `/v1/quality` tally and
// never to a per-session endpoint.
// * The review loop's own verdict: `iterations` used, and `unresolved` issues
// remaining when the iteration cap was hit. Hitting the cap is the pipeline
// saying, in its own words, that it could not finish fixing this document.
// * Signals the loop can only report and not fix: links the copy editor dropped,
// id collisions the assembler had to resolve, pages whose verify pass failed,
// specialists that declined.
// * Structure, from lint.mjs `shape()`. A clean lint on a nearly empty document
// is a silent failure no rule catches.
//
// And the denominator is not the corpus. Documents the corpus never ran — 404s,
// HTML login walls served as PDFs, encrypted files — are prepare's business and
// live in prepared.jsonl; a success rate computed here is over documents that were
// actually submitted. Both numbers are printed, because a campaign that ran 1400 of
// 2000 URLs and reports 95% has told you something misleading.
import { readdirSync, readFileSync, existsSync, statSync, writeFileSync } from "node:fs";
import { join } from "node:path";
import { runAxe, shape, AXE_VERSION } from "./lint.mjs";
import { costOf, loadRates, normalizeModelId, provenance } from "./pricing.mjs";
import { args, errorText, latestAttempts, log, num, pct, readJsonl } from "./util.mjs";
const readJson = (p) => {
try {
return JSON.parse(readFileSync(p, "utf8"));
} catch {
return null;
}
};
// --- one run ---------------------------------------------------------------
async function analyze(dir, rates, opts) {
const meta = readJson(join(dir, "meta.json"));
if (!meta) return null;
const status = readJson(join(dir, "status.json"));
const diag = readJson(join(dir, "diagnostics.json"));
const events = existsSync(join(dir, "log.jsonl")) ? readJsonl(join(dir, "log.jsonl")) : [];
const htmlPath = join(dir, "output.html");
const html = existsSync(htmlPath) ? readFileSync(htmlPath, "utf8") : null;
const ev = (type) => events.filter((e) => e.type === type);
const assembly = ev("assembly").at(-1) ?? null;
const readers = ev("reader");
const complete = ev("run_complete").at(-1) ?? null;
const failed = ev("run_failed").at(-1) ?? null;
const anchors = ev("assembly_anchors").at(-1) ?? null;
const tokens = diag?.tokens ?? null;
const models = new Set(events.filter((e) => e.type === "model_call" && e.model).map((e) => e.model));
// One model per run in practice (all three capabilities point at the same id), so
// the run's cost is priced against the one that did the most calls; a run that
// genuinely mixed models is flagged rather than averaged.
const modelCounts = {};
for (const e of events) if (e.type === "model_call" && e.model) modelCounts[e.model] = (modelCounts[e.model] ?? 0) + 1;
const primaryModel = Object.entries(modelCounts).sort((a, b) => b[1] - a[1])[0]?.[0] ?? null;
const cost = tokens ? costOf(tokens, primaryModel, rates) : { usd: null, priced: false, model: null };
// The independent check. Skippable (--no-lint) because axe over a few thousand
// documents is minutes of CPU, and the rest of the report does not depend on it.
let finalLint = null;
let structure = null;
if (html) {
structure = shape(html);
if (!opts.noLint) {
const r = await runAxe(html);
finalLint = {
ok: r.ok,
error: r.error ?? null,
violations: r.violations,
// Total offending elements, not rules: one document with 40 unlabelled
// inputs and one with a single missing lang are both "1 rule".
nodes: r.violations.reduce((s, v) => s + v.nodes, 0),
};
}
}
const pages = assembly?.pages ?? meta.pages ?? null;
const wall = meta.round_trip_ms ?? null;
const serverMs = diag?.elapsed_ms ?? null;
return {
id: meta.id,
url: meta.url,
session_id: meta.session_id ?? null,
outcome: meta.outcome,
// Was this a whole document or a slice of an oversize one? A chunk's review
// score is not comparable — the reviewer saw a document with no beginning — so
// the summary reports both populations separately.
is_chunk: Boolean(meta.parent_sha),
page_from: meta.page_from ?? null,
page_to: meta.page_to ?? null,
risks: meta.risks ?? [],
bytes: meta.bytes ?? null,
pages,
// --- time ---
// Three clocks, because they answer different questions: wall is what a caller
// experiences, server elapsed is what Iris spent once it had a slot, and queue
// wait is the difference concurrency causes.
wall_ms: wall,
server_ms: serverMs,
queue_wait_ms: meta.queue_wait_ms ?? null,
ms_per_page: pages && wall ? Math.round(wall / pages) : null,
phase_ms: diag?.phase_durations_ms ?? null,
model_calls: diag?.model_calls?.count ?? null,
model_calls_failed: diag?.model_calls?.failed ?? null,
concurrency_factor: diag?.model_calls?.concurrency_factor ?? null,
// --- cost ---
tokens,
// Lower than model_calls means these sums cover only part of the run, so the
// cost below is a floor. Surfaced per document, not just in aggregate.
tokens_partial: tokens ? tokens.calls_reported < (diag?.model_calls?.count ?? 0) : null,
model: primaryModel,
models: [...models],
mixed_models: models.size > 1,
usd: cost.usd,
usd_per_page: cost.usd != null && pages ? cost.usd / pages : null,
tokens_per_page: tokens && pages ? Math.round((tokens.input + tokens.output) / pages) : null,
by_agent: diag?.by_agent ?? null,
// --- accuracy ---
final_lint: finalLint,
// The pre-review lint, from the assembly event. The pair is the interesting
// thing: pre-review violations that survive to final are what the review loop
// failed to fix.
assembly_lint_ok: assembly?.lint_ok ?? null,
assembly_violations: assembly?.violations ?? null,
assembly_lint_error: assembly?.lint_error ?? null,
iterations: complete?.iterations ?? status?.iterations_completed ?? null,
unresolved: complete?.unresolved ?? null,
// The pipeline's own admission of defeat: the cap was reached with issues left.
hit_iteration_cap: complete ? complete.unresolved > 0 : null,
reader_issues: readers.map((r) => r.issues),
links_dropped: ev("editor_links_dropped").reduce((s, e) => s + (e.hrefs?.length ?? 0), 0),
id_collisions: anchors?.collisions?.length ?? 0,
ambiguous_refs: anchors?.ambiguous?.length ?? 0,
skipped_pages: anchors?.skipped_pages?.length ?? 0,
pages_verify_failed: ev("page_verify_failed").length,
links_unrecovered: ev("page_links_unrecovered").length,
specialists_dispatched: ev("specialist_dispatched").length,
specialists_declined: ev("specialist_declined").length,
specialists_unresolved: ev("specialist_unresolved").length,
reextracts: ev("reextract_start").length,
structure,
// --- what went wrong ---
error: errorText(failed?.error ?? meta.error ?? status?.error) ?? null,
errors: (diag?.errors ?? []).map((e) => e.message),
// Every run that files one of these opens a GitHub issue on the upstream repo.
// Counted so the campaign can see what it generated.
agent_issues: ev("agent_issue").length,
};
}
// --- aggregation -----------------------------------------------------------
// Errors are grouped by shape, not text: a provider message carrying a request id,
// a page number or a byte count is one class of failure, not two hundred.
function errorClass(msg) {
if (!msg) return "unknown";
return String(msg)
.replace(/\b[0-9a-f]{8,}\b/gi, "<id>")
// Every digit run, not just whole words: the timeouts and byte counts that make
// two identical failures look different are written `60000ms`, `3932160b`.
.replace(/\d+/g, "<n>")
.slice(0, 140);
}
const sum = (rows, f) => rows.reduce((s, r) => s + (f(r) ?? 0), 0);
const defined = (rows, f) => rows.map(f).filter((v) => v != null);
function summarize(rows, prepared, rates, opts) {
const delivered = rows.filter((r) => r.outcome === "ready_for_review" || r.outcome === "closed");
const walls = defined(delivered, (r) => r.wall_ms);
const perPage = defined(delivered, (r) => r.ms_per_page);
const usd = defined(delivered, (r) => r.usd);
const linted = delivered.filter((r) => r.final_lint);
const lintable = linted.filter((r) => !r.final_lint.error);
const outcomes = {};
for (const r of rows) outcomes[r.outcome] = (outcomes[r.outcome] ?? 0) + 1;
// Which axe rules actually fire, ranked by how many documents they affect rather
// than by node count — a rule that breaks 300 documents once matters more to fix
// than one that fires 900 times inside a single table.
const rules = {};
for (const r of lintable) {
for (const v of r.final_lint.violations) {
const cur = (rules[v.id] ??= { rule: v.id, impact: v.impact, documents: 0, nodes: 0 });
cur.documents += 1;
cur.nodes += v.nodes;
}
}
const errors = {};
for (const r of rows) {
if (r.outcome === "ready_for_review" || r.outcome === "closed") continue;
const k = errorClass(r.error ?? r.errors[0]);
const cur = (errors[k] ??= { error: k, count: 0, examples: [] });
cur.count += 1;
if (cur.examples.length < 3) cur.examples.push(r.id);
}
// Which agent spends the money, and which one spends the time. Iris reports these
// per run; the campaign's question is which one to optimize first.
const agents = {};
for (const r of delivered) {
for (const [name, a] of Object.entries(r.by_agent ?? {})) {
const cur = (agents[name] ??= { agent: name, calls: 0, total_ms: 0, input_tokens: 0, output_tokens: 0 });
cur.calls += a.count;
cur.total_ms += a.total_ms;
cur.input_tokens += a.input_tokens;
cur.output_tokens += a.output_tokens;
}
}
// URL-level accounting, which is not the same as document-level. One oversize PDF
// becomes several runnable chunks, so counting prepared.jsonl rows would both
// inflate the number of URLs and count a single URL's `ok` several times. Rows
// with a `parent_sha` are chunk children and are excluded from the URL tallies.
const preparedCounts = {};
const urls = new Set();
for (const p of prepared) {
if (p.url) urls.add(p.url);
if (p.parent_sha) continue;
preparedCounts[p.klass ?? "unknown"] = (preparedCounts[p.klass ?? "unknown"] ?? 0) + 1;
}
// A URL counts as covered only if EVERY item it produced was delivered. A 40-page
// PDF whose second chunk failed is not a document anyone received.
const byUrl = new Map();
for (const r of rows) {
const cur = byUrl.get(r.url) ?? { total: 0, delivered: 0 };
cur.total += 1;
if (r.outcome === "ready_for_review" || r.outcome === "closed") cur.delivered += 1;
byUrl.set(r.url, cur);
}
const urlsCovered = [...byUrl.values()].filter((u) => u.total > 0 && u.delivered === u.total).length;
const pagesDelivered = sum(delivered, (r) => r.pages);
const partial = delivered.filter((r) => r.tokens_partial).length;
return {
generated_at: new Date().toISOString(),
axe_version: AXE_VERSION,
lint_skipped: Boolean(opts.noLint),
corpus: {
// The honest denominator: URLs the CSV offered, what prepare made of them,
// and how many of those were actually submitted.
urls_prepared: urls.size || null,
prepared_classes: Object.keys(preparedCounts).length ? preparedCounts : null,
submitted: rows.length,
urls_submitted: byUrl.size,
urls_covered: urlsCovered,
chunks: rows.filter((r) => r.is_chunk).length,
pages_delivered: pagesDelivered,
},
outcomes,
success_rate: rows.length ? delivered.length / rows.length : null,
// The same question asked of every URL the CSV contained — the number a "can
// Iris handle real-world PDFs" claim has to survive, since it charges the
// harness for the 404s, sign-in walls and encrypted files too.
end_to_end_rate: urls.size ? urlsCovered / urls.size : null,
time: {
// Percentiles, not means: the tail decides how long a campaign takes and it is
// where the timeouts live.
wall_ms: { p50: pct(walls, 50), p90: pct(walls, 90), p99: pct(walls, 99), max: pct(walls, 100) },
ms_per_page: { p50: pct(perPage, 50), p90: pct(perPage, 90), max: pct(perPage, 100) },
queue_wait_ms: { p50: pct(defined(rows, (r) => r.queue_wait_ms), 50), p90: pct(defined(rows, (r) => r.queue_wait_ms), 90) },
total_hours: sum(rows, (r) => r.wall_ms) / 3_600_000,
// >1 means model calls overlapped (extraction runs pages in parallel). Near 1
// on multi-page documents means the concurrency is not being used.
concurrency_factor_p50: pct(defined(delivered, (r) => r.concurrency_factor), 50),
},
cost: {
tokens: {
input: sum(delivered, (r) => r.tokens?.input),
output: sum(delivered, (r) => r.tokens?.output),
cache_read: sum(delivered, (r) => r.tokens?.cache_read),
cache_write: sum(delivered, (r) => r.tokens?.cache_write),
},
tokens_per_page_p50: pct(defined(delivered, (r) => r.tokens_per_page), 50),
usd_total: usd.length ? usd.reduce((a, b) => a + b, 0) : null,
usd_per_document: { p50: pct(usd, 50), p90: pct(usd, 90), max: pct(usd, 100) },
usd_per_page_p50: pct(defined(delivered, (r) => r.usd_per_page), 50),
// Projected from the per-page median, which is the number to quote for "what
// would 100,000 pages cost" — and the number to distrust if unpriced > 0.
usd_per_1000_pages: pct(defined(delivered, (r) => r.usd_per_page), 50) * 1000 || null,
unpriced_documents: delivered.filter((r) => r.usd == null).length,
// Documents whose token sums cover only part of the run: their cost is a floor.
partial_token_documents: partial,
rates: provenance(new Set(delivered.flatMap((r) => r.models)), rates),
caveat:
"The deployment runs on Amazon Bedrock, which is partner-operated and billed by AWS at " +
"AWS's own rates. Dollar figures here are estimates for comparing documents, not an invoice.",
},
accuracy: {
// Documents delivered with zero final axe violations, by the deployment's own
// rule configuration, re-checked locally.
lint_clean: lintable.filter((r) => r.final_lint.ok).length,
lint_checked: lintable.length,
lint_clean_rate: lintable.length ? lintable.filter((r) => r.final_lint.ok).length / lintable.length : null,
// A document axe could not examine is not a clean document.
lint_errored: linted.filter((r) => r.final_lint.error).length,
violation_nodes_p50: pct(defined(lintable, (r) => r.final_lint.nodes), 50),
top_rules: Object.values(rules).sort((a, b) => b.documents - a.documents).slice(0, 20),
// The review loop's self-assessment.
iterations: {
p50: pct(defined(delivered, (r) => r.iterations), 50),
max: pct(defined(delivered, (r) => r.iterations), 100),
// Cap reached with issues remaining: the pipeline could not finish the job.
hit_cap: delivered.filter((r) => r.hit_iteration_cap).length,
unresolved_total: sum(delivered, (r) => r.unresolved),
},
// Losses the loop reports but cannot repair.
links_dropped: sum(rows, (r) => r.links_dropped),
documents_losing_links: rows.filter((r) => r.links_dropped > 0).length,
id_collisions: sum(rows, (r) => r.id_collisions),
ambiguous_refs: sum(rows, (r) => r.ambiguous_refs),
skipped_pages: sum(rows, (r) => r.skipped_pages),
pages_verify_failed: sum(rows, (r) => r.pages_verify_failed),
specialists: {
dispatched: sum(rows, (r) => r.specialists_dispatched),
declined: sum(rows, (r) => r.specialists_declined),
unresolved: sum(rows, (r) => r.specialists_unresolved),
},
reextracts: sum(rows, (r) => r.reextracts),
// Structural floors — a clean lint on an empty document.
documents_without_headings: delivered.filter((r) => r.structure && r.structure.headings === 0).length,
documents_under_200_chars: delivered.filter((r) => r.structure && r.structure.text_chars < 200).length,
chars_per_page_p50: pct(
defined(delivered, (r) => (r.structure && r.pages ? Math.round(r.structure.text_chars / r.pages) : null)),
50,
),
},
failures: Object.values(errors).sort((a, b) => b.count - a.count).slice(0, 25),
by_agent: Object.values(agents).sort((a, b) => b.total_ms - a.total_ms),
agent_issues_filed: sum(rows, (r) => r.agent_issues),
};
}
// --- output ----------------------------------------------------------------
const money = (v) => (v == null ? "n/a" : `$${v >= 1 ? v.toFixed(2) : v < 0.01 ? v.toFixed(5) : v.toFixed(4)}`);
const secs = (v) => (v == null ? "n/a" : `${(v / 1000).toFixed(1)}s`);
const rate = (v) => (v == null ? "n/a" : `${(v * 100).toFixed(1)}%`);
function print(s) {
const l = log;
l("=== equalify-iris-bench ===");
l(`corpus: ${s.corpus.submitted} item(s) submitted from ${s.corpus.urls_submitted} of`,
`${s.corpus.urls_prepared ?? "?"} prepared URL(s) — ${s.corpus.pages_delivered} page(s) delivered,`,
`${s.corpus.chunks} item(s) were chunks of oversize PDFs`);
l("");
l("--- outcomes ---");
for (const [k, v] of Object.entries(s.outcomes).sort((a, b) => b[1] - a[1])) l(` ${k}: ${v}`);
l(` success rate (of items submitted): ${rate(s.success_rate)}`);
l(` end-to-end (URLs fully delivered, of every URL in the CSV):`,
`${rate(s.end_to_end_rate)} — ${s.corpus.urls_covered}/${s.corpus.urls_prepared ?? "?"}`);
l("");
l("--- time ---");
l(` per document: p50 ${secs(s.time.wall_ms.p50)} p90 ${secs(s.time.wall_ms.p90)} max ${secs(s.time.wall_ms.max)}`);
l(` per page: p50 ${secs(s.time.ms_per_page.p50)} p90 ${secs(s.time.ms_per_page.p90)}`);
l(` queue wait: p50 ${secs(s.time.queue_wait_ms.p50)} p90 ${secs(s.time.queue_wait_ms.p90)}`);
l(` campaign wall clock: ${s.time.total_hours.toFixed(1)}h; call concurrency p50 ${s.time.concurrency_factor_p50 ?? "n/a"}x`);
l("");
l("--- cost (estimate; see caveat) ---");
l(` tokens: ${s.cost.tokens.input.toLocaleString()} in / ${s.cost.tokens.output.toLocaleString()} out`,
`(cache ${s.cost.tokens.cache_read.toLocaleString()} read / ${s.cost.tokens.cache_write.toLocaleString()} write)`);
l(` per document: p50 ${money(s.cost.usd_per_document.p50)} p90 ${money(s.cost.usd_per_document.p90)}`);
l(` per page: p50 ${money(s.cost.usd_per_page_p50)} → ${money(s.cost.usd_per_1000_pages)} per 1000 pages`);
l(` total: ${money(s.cost.usd_total)} (median ${s.cost.tokens_per_page_p50 ?? "n/a"} tokens/page)`);
if (s.cost.unpriced_documents) l(` WARNING: ${s.cost.unpriced_documents} document(s) ran on an unpriced model — cost excludes them`);
if (s.cost.partial_token_documents) l(` WARNING: ${s.cost.partial_token_documents} document(s) reported tokens for only part of the run — cost is a floor`);
l("");
l("--- accuracy ---");
if (s.lint_skipped) l(" (local re-lint skipped: --no-lint)");
l(` axe-clean: ${s.accuracy.lint_clean}/${s.accuracy.lint_checked} (${rate(s.accuracy.lint_clean_rate)}) [axe ${s.axe_version}]`);
if (s.accuracy.lint_errored) l(` axe could not examine ${s.accuracy.lint_errored} document(s) — not counted as clean`);
l(` review iterations: p50 ${s.accuracy.iterations.p50 ?? "n/a"}, max ${s.accuracy.iterations.max ?? "n/a"};`,
`${s.accuracy.iterations.hit_cap} document(s) hit the cap with ${s.accuracy.iterations.unresolved_total} issue(s) unresolved`);
l(` links dropped by the editor: ${s.accuracy.links_dropped} across ${s.accuracy.documents_losing_links} document(s)`);
l(` id collisions ${s.accuracy.id_collisions}, ambiguous refs ${s.accuracy.ambiguous_refs}, pages left as written ${s.accuracy.skipped_pages}`);
l(` page verify failures ${s.accuracy.pages_verify_failed}, re-extractions ${s.accuracy.reextracts}`);
l(` specialists: ${s.accuracy.specialists.dispatched} dispatched, ${s.accuracy.specialists.declined} declined, ${s.accuracy.specialists.unresolved} unresolved`);
l(` suspiciously thin output: ${s.accuracy.documents_without_headings} with no headings, ${s.accuracy.documents_under_200_chars} under 200 chars`,
`(median ${s.accuracy.chars_per_page_p50 ?? "n/a"} chars/page)`);
if (s.accuracy.top_rules.length) {
l(" top axe rules, by documents affected:");
for (const r of s.accuracy.top_rules.slice(0, 10)) l(` ${r.documents}× ${r.rule} [${r.impact ?? "?"}] (${r.nodes} nodes)`);
}
if (s.failures.length) {
l("");
l("--- failures ---");
for (const f of s.failures.slice(0, 10)) l(` ${f.count}× ${f.error} e.g. ${f.examples[0]}`);
}
if (s.by_agent.length) {
l("");
l("--- where the time and money go ---");
for (const a of s.by_agent) {
l(` ${a.agent}: ${a.calls} call(s), ${(a.total_ms / 1000).toFixed(0)}s,`,
`${a.input_tokens.toLocaleString()} in / ${a.output_tokens.toLocaleString()} out`);
}
}
if (s.agent_issues_filed) {
l("");
l(`NOTE: this campaign filed ${s.agent_issues_filed} agent-suggestion issue(s) upstream.`);
}
}
// --- main ------------------------------------------------------------------
async function main() {
const a = args();
const runsDir = a.runs ?? "runs";
const outDir = a.out ?? runsDir;
const rates = loadRates(a.rates);
const opts = { noLint: Boolean(a["no-lint"]) };
if (!existsSync(runsDir)) {
console.error(`no such directory: ${runsDir}`);
process.exit(2);
}
const dirs = readdirSync(runsDir)
.map((d) => join(runsDir, d))
.filter((p) => statSync(p).isDirectory());
if (!dirs.length) {
console.error(`${runsDir} has no run directories — run src/run.mjs first`);
process.exit(2);
}
const prepared = latestAttempts(readJsonl(a.prepared ?? "prepared.jsonl"));
log(`analyzing ${dirs.length} run(s)${opts.noLint ? "" : ` with a local axe re-lint (axe-core ${AXE_VERSION})`}`);
// Serial on purpose. axe in jsdom is CPU-bound, and the point of this stage is a
// report, not throughput; a progress line every 100 is enough.
const rows = [];
for (const [i, dir] of dirs.entries()) {
const row = await analyze(dir, rates, opts);
if (row) rows.push(row);
if ((i + 1) % 100 === 0) log(` analyzed ${i + 1}/${dirs.length}`);
}
const resultsPath = join(outDir, "results.jsonl");
writeFileSync(resultsPath, rows.map((r) => JSON.stringify(r)).join("\n") + (rows.length ? "\n" : ""));
const s = summarize(rows, prepared, rates, opts);
const summaryPath = join(outDir, "summary.json");
writeFileSync(summaryPath, JSON.stringify(s, null, 2));
print(s);
log("");
log(`wrote ${resultsPath} (one row per document) and ${summaryPath}`);
// stdout is the summary and nothing else, so this can be piped into jq while the
// narration above goes to stderr.
process.stdout.write(`${JSON.stringify(s, null, 2)}\n`);
}
await main();