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
588import { Router } from "express";
import type { Request, Response, NextFunction } from "express";
import multer from "multer";
import { writeFileSync, readFileSync, existsSync, rmSync } from "node:fs";
import { join } from "node:path";
import { ulid } from "ulid";
import type { IrisConfig } from "../config.ts";
import type { Store, SessionRecord } from "../store/db.ts";
import { encodeCursor, pageSessions, parseCursor } from "../store/db.ts";
import { Paths } from "../store/paths.ts";
import { runPipeline } from "../pipeline/orchestrator.ts";
import type { AuthedRequest } from "../auth/middleware.ts";
import { sendError } from "./errors.ts";
import { summarizeRun } from "../diagnostics.ts";
import { rasterizePdf, PdfTooLargeError, MAX_PDF_PAGES, type PageImage, type PdfLink } from "../util/pdf.ts";
import { outputBasenameFromUploads, convertedHtmlFilename, titledAs } from "../util/outputNames.ts";
import { captureFixtures } from "../pipeline/regression.ts";
import type { Fragment } from "../pipeline/fragment.ts";
import { RunQueue } from "../util/queue.ts";
import {
MAX_UPLOAD_FILES,
meterUploadBody,
requestSizeGate,
uploadGate,
uploadRateLimit,
} from "../util/requestLimits.ts";
import { RunLog } from "../store/runlog.ts";
import {
IMAGE_MEDIA_TYPES,
imageRejection,
rasterizedPageRejection,
resolveImageLimits,
} from "../providers/imageLimits.ts";
import { imageDimensions } from "../util/imageSize.ts";
// 50 MB is a memory bound, not the image limit. It stays well above what an image may
// be (see imageLimits.ts) because a PDF legitimately is: 25 pages of scans is a large
// file whose page images are small, and this ceiling has to admit it. Per-image size is
// checked in the handler instead, where the file's type is known.
//
// `files` bounds the OTHER half of that memory, which a per-file ceiling cannot: a
// multipart body may carry any number of parts, so 50 MB each with no count meant one
// request could buffer gigabytes before this handler ran. It cannot cost a real upload
// anything — a session is capped at 25 pages however they arrive, so 25 files is already
// more than an accepted request can use (see requestLimits.ts).
//
// Note what this pair is NOT: their product is 1.25 GB, so they are per-part backstops and
// never the total. What one request may buffer in all is enforced in two places that share
// one ceiling — `requestSizeGate`, which refuses a declared length before reading anything,
// and `meterUploadBody`, which counts an undeclared one as it arrives (requestLimits.ts).
// The numbers here only have to stay above any single legitimate part.
const upload = multer({
storage: multer.memoryStorage(),
limits: { fileSize: 50 * 1024 * 1024, files: MAX_UPLOAD_FILES },
});
// Wrap multer so its errors (e.g. file too large) become clean 400s.
function uploadImages(req: Request, res: Response, next: NextFunction): void {
// The total-bytes ceiling for a body that declared no length, which multer's per-part
// limits are not (see requestLimits.ts). Called here rather than mounted as its own
// middleware because it counts bytes off the request stream, and everything between it
// and multer's `req.pipe()` has to be synchronous — one statement is as close as that
// gets. It answers the request itself if the body overruns, and multer then never calls
// back at all.
meterUploadBody(req, res);
upload.array("images")(req, res, (err: unknown) => {
if (err) {
// meterUploadBody above can have already answered 413 and unpiped this body mid-parse.
// It does not expect multer to call back at all after that, but a second write to a
// sent response would throw from in here, where nothing is listening.
if (res.headersSent) return;
// Multer says "Too many files" for the part count, which does not say how many are
// too many. That cap is one Iris chose rather than a property of multipart, so the
// refusal owes the caller the number and what to do about it.
const message =
(err as { code?: string }).code === "LIMIT_FILE_COUNT"
? `Upload failed: at most ${MAX_UPLOAD_FILES} file parts per request, since one session ` +
`carries at most ${MAX_TOTAL_PAGES} pages however they arrive. Split a larger batch across sessions.`
: `Upload failed: ${(err as Error).message}`;
sendError(res, 400, "invalid_request", message);
return;
}
next();
});
}
// Built from the formats the vision model actually reads (providers/imageLimits.ts)
// rather than written out here, so this validator cannot go on accepting a format
// the model stopped supporting — or, as it did with TIFF, never supported.
const IMAGE_EXT = new RegExp(`(?:${Object.keys(IMAGE_MEDIA_TYPES).map((e) => `\\${e}`).join("|")})$`, "i");
const PDF_EXT = /\.pdf$/i;
const MAX_TOTAL_PAGES = MAX_PDF_PAGES; // overall cap across all uploaded files
// How the allowed types read in an error message: ".png,.jpg" -> "PNG, JPG".
const ALLOWED_TYPES = [...Object.keys(IMAGE_MEDIA_TYPES).map((e) => e.slice(1).toUpperCase()), "PDF"].join(", ");
// A rasterized PDF page the vision model cannot accept. Thrown from the expansion
// loop rather than returned, because the check sits inside an `await` over several
// files and the alternative is threading a rejection out through both branches; it
// lands in the same catch as the other PDF failures. It carries the finished message
// so the handler decides the status code and nothing else re-words the diagnosis.
class PageTooLargeError extends Error {}
// One-line axe-core summary for the PR description (from sessions/<id>/lint.json).
function sessionSummary(s: SessionRecord) {
return {
session_id: s.session_id,
status: s.status,
image_count: s.image_count,
created_at: s.created_at,
updated_at: s.updated_at,
};
}
// Owned-by-caller lookup. Returns undefined (caller sends 404) when missing or
// owned by another user, so a token cannot probe others' sessions.
function ownedSession(store: Store, id: string, userId: number): SessionRecord | undefined {
const s = store.getSession(id);
if (!s || s.github_user_id !== userId) return undefined;
return s;
}
export function sessionsRouter(cfg: IrisConfig, store: Store): Router {
const r = Router();
const paths = new Paths(cfg);
// What one page image may be, for the model this deployment runs. Resolved once:
// config does not change without a restart, and the same numbers are published by
// GET /v1/limits so a client can check before uploading.
const imageLimits = resolveImageLimits(cfg);
// One queue for the whole process, bounding how many pipelines run at once
// across all sessions (see util/queue.ts for why global rather than per-user,
// and why waiting rather than rejecting).
const queue = new RunQueue(cfg.defaults.max_concurrent_runs);
// What the queue cannot bound, because it is spent before a handler runs: upload bytes
// held in memory at once, and uploads accepted from one user per minute. Built once per
// router, since both hold the counters they enforce (util/requestLimits.ts).
//
// The in-flight gate is checked before the per-minute budget so a request refused for
// congestion does not also spend a minute's allowance — it was told to retry in
// seconds, and that advice has to still be true when it does.
const inFlightUploads = uploadGate(cfg);
const uploadBudget = uploadRateLimit(cfg);
// Submit a run to the queue instead of starting it immediately.
//
// The queue wait is recorded in the session's own run log, at both ends: a
// `run_queued` event when it is admitted and a `run_dequeued` event carrying
// `waited_ms` when it actually starts. Without those, a session sitting in
// `queued` behind someone else's 25-page PDF is indistinguishable from a hung
// one — the operator sees a status that never changes and a log with nothing
// in it. Logging is best-effort: a run must never fail to start because its
// log could not be appended to.
// Returns whether the run got a slot immediately, so the caller can report the
// session's real state rather than assuming it started.
const enqueueRun = (args: Parameters<typeof runPipeline>[0]): boolean => {
const submittedAt = Date.now();
const log = new RunLog(paths.sessionLog(args.sessionId));
const stats = queue.stats();
try {
log.event("run_queued", { running: stats.running, waiting: stats.waiting, limit: stats.limit });
} catch {
// ignore — observability must not block the run
}
let startedImmediately = false;
// `void` is safe: RunQueue.submit() always resolves, even when the task
// throws (see its docblock), so this cannot become an unhandled rejection.
// submit() pumps synchronously, so onStart has already fired by the time
// this returns if a slot was free.
void queue.submit(
() => runPipeline(args),
() => {
startedImmediately = true;
try {
log.event("run_dequeued", { waited_ms: Date.now() - submittedAt });
} catch {
// ignore
}
},
);
return startedImmediately;
};
// GET /v1/sessions — list this user's sessions, newest first.
//
// Paged by keyset on (created_at, session_id), not by created_at alone: that is
// not unique — a burst of uploads shares a millisecond — and paging on it drops
// and repeats rows at tie boundaries (see store.listSessions). The cursor is
// therefore compound, and an unparseable one is a 400 rather than being
// compared as a string, which used to silently hand back page one forever.
r.get("/", (req: AuthedRequest, res) => {
// One rule for every unusable value: fall back to the default. Written as
// `Math.max(parseInt(x) || 20, 1)` it was two rules — `?limit=0` is falsy so
// it became 20, while `?limit=-1` clamped to 1 — so two equally invalid
// inputs got different page sizes, neither of them documented.
//
// Clamping the low end at all matters because a negative limit is destructive
// twice over: SQLite reads `LIMIT -4` as NO limit, so the whole table comes
// back, and `slice(0, -5)` then trims rows off the end of the page while
// still leaving an extra row held back — a short page carrying a next_cursor.
// (`listSessions` and `pageSessions` clamp too; that symptom is invisible in
// the response, so it should not depend on one caller getting this right.)
const asked = parseInt(String(req.query.limit ?? ""), 10);
const limit = Number.isInteger(asked) && asked >= 1 ? Math.min(asked, 100) : 20;
const status = req.query.status ? String(req.query.status) : undefined;
const raw = req.query.cursor ? String(req.query.cursor) : undefined;
const cursor = raw ? parseCursor(raw) : undefined;
if (raw && !cursor) {
// Echo the offending value — a 400 that doesn't say what it rejected is
// undiagnosable from a client log — but truncated. The reflected input is
// whatever the client sent, and a query string can be kilobytes; there is no
// legible cursor longer than this, so a longer one is already the answer.
const shown = raw.length > 80 ? `${raw.slice(0, 80)}… (${raw.length} chars)` : raw;
sendError(
res,
400,
"invalid_request",
`Invalid cursor: ${shown}. Pass a next_cursor value from a previous response verbatim ` +
`(<created_at>|<session_id>, where created_at is UTC ISO-8601 with milliseconds), ` +
`percent-encoded in the query string.`,
);
return;
}
const rows = store.listSessions(req.user!.github_user_id, {
limit: limit + 1,
status,
cursor: cursor ?? undefined,
});
// The slice-and-cursor decision lives with the query (pageSessions), not
// here: the over-fetch above, the page boundary, and the cursor are one
// contract, and a route that re-derives it can drift from the query silently.
const { page, next } = pageSessions(rows, limit);
res.json({
sessions: page.map(sessionSummary),
next_cursor: next ? encodeCursor(next) : null,
});
});
// POST /v1/sessions — create a session. Accepts images and/or PDFs; PDFs are
// rasterized to one image per page, expanded in submitted order.
//
// Three gates run BEFORE `uploadImages`, and the order is the point: multer buffers the
// whole body into memory as it parses, so anything that runs after it has already paid
// for the request it was going to refuse. So the size the caller declares is checked
// first (free), then how many bytes of upload are already arriving, then the caller's
// upload budget for the minute — and only then is a byte of this body read.
r.post("/", requestSizeGate(), inFlightUploads, uploadBudget, uploadImages, async (req: AuthedRequest, res) => {
const files = (req.files as Express.Multer.File[] | undefined) ?? [];
if (files.length === 0) {
sendError(res, 400, "invalid_request", "At least one file part named 'images' is required (image or PDF)");
return;
}
for (const f of files) {
if (!IMAGE_EXT.test(f.originalname) && !PDF_EXT.test(f.originalname)) {
sendError(res, 400, "invalid_request", `Unsupported file type: ${f.originalname} (allowed: ${ALLOWED_TYPES})`);
return;
}
// Too heavy for the vision model to accept, checked before anything is written
// to disk or queued. Without this the upload succeeded, the run started, and
// the first vision call ate the whole payload and then failed — reaching the
// user two to four minutes later as a message about a stream that produced no
// output, naming neither the file nor anything to do about it.
//
// A PDF is not measured here — the file the caller sent is not what reaches the
// model. A 20 MB PDF of 25 light pages is a perfectly convertible document, so
// the size that matters is the size of each page AFTER rasterizing, and that is
// checked where the pages exist (below).
if (!PDF_EXT.test(f.originalname)) {
const size = imageDimensions(f.buffer);
const why = imageRejection(
{
name: f.originalname,
bytes: f.buffer.length,
width: size?.width,
height: size?.height,
},
imageLimits,
);
if (why) {
sendError(res, 400, "invalid_request", why);
return;
}
}
}
// The review cap is the caller's account default (seeded from the
// deployment's `defaults.max_review_iterations`). A per-request `config`
// JSON part used to override it; it was removed because it asked the caller
// for a number they are not positioned to choose — how many rounds a
// document needs is a property of the document, discovered by the loop, and
// the only thing the knob could usefully express was "spend fewer rounds
// than this deployment budgeted", i.e. ask for a worse document. It was also
// unvalidated: any number was accepted, including `0` (one reader pass, no
// fix ever applied) and negatives (review skipped outright). A `config` part
// on the request is now ignored rather than rejected, so an older client
// still sending one keeps working, at the deployment's cap.
const maxIter = req.user!.max_review_iterations;
// Expand uploads into ordered page images (PDF -> one PNG per page), each
// carrying the link annotations on that page — the one part of a PDF that
// rasterizing destroys, so it travels alongside the image (see pipeline/links.ts).
const pages: PageImage[] = [];
try {
for (const f of files) {
if (PDF_EXT.test(f.originalname)) {
const rendered = await rasterizePdf(f.buffer, f.originalname);
// The uploaded PDF was exempt from the per-image limits; its pages are not.
// Rasterizing at a fixed DPI means the page image's size follows the
// PHYSICAL page, so a large-format page (a drawing, a fold-out) renders past
// what the model accepts — and the run would then die inside the first
// vision call, minutes in, which is the failure this route exists to catch.
for (const [i, p] of rendered.entries()) {
const size = imageDimensions(p.buffer);
const why = rasterizedPageRejection(
f.originalname,
i + 1,
{ bytes: p.buffer.length, width: size?.width, height: size?.height },
imageLimits,
);
if (why) throw new PageTooLargeError(why);
}
pages.push(...rendered);
} else {
pages.push({ name: f.originalname, buffer: f.buffer, links: [] });
}
}
} catch (e) {
if (e instanceof PageTooLargeError || e instanceof PdfTooLargeError) {
sendError(res, 400, "invalid_request", e.message);
} else {
sendError(res, 422, "pdf_conversion_failed", `Could not process a PDF: ${(e as Error).message}`);
}
return;
}
if (pages.length === 0) {
sendError(res, 400, "invalid_request", "No pages found in the uploaded files");
return;
}
if (pages.length > MAX_TOTAL_PAGES) {
sendError(res, 400, "invalid_request", `Too many pages (${pages.length}); the maximum is ${MAX_TOTAL_PAGES}.`);
return;
}
const sessionId = `ses_${ulid()}`;
paths.initSession(sessionId);
// Remember the source basename so the output title + download filename
// mirror the upload (e.g. report.pdf -> report_converted.html).
writeFileSync(paths.sessionSourceName(sessionId), outputBasenameFromUploads(files));
// Persist page images with an order prefix so submitted order survives.
pages.forEach((p, i) => {
const order = String(i + 1).padStart(4, "0");
writeFileSync(join(paths.sessionInput(sessionId), `${order}__${p.name}`), p.buffer);
});
// …and the link annotations beside them, keyed by that same order. Only when
// there are some, so a session of plain images gains no empty file.
const linksByOrder: Record<string, PdfLink[]> = {};
pages.forEach((p, i) => {
if (p.links.length) linksByOrder[String(i + 1)] = p.links;
});
if (Object.keys(linksByOrder).length) {
writeFileSync(paths.sessionLinks(sessionId), JSON.stringify(linksByOrder, null, 2));
}
const record = store.createSession({
session_id: sessionId,
github_user_id: req.user!.github_user_id,
image_count: pages.length,
iterations_max: maxIter,
});
// Queue the pipeline; clients poll GET /v1/sessions/{id}. The session stays
// in the `queued` status it was created with until a slot frees up, so a
// client polling sees queued -> running -> ready_for_review.
enqueueRun({ cfg, store, sessionId, maxReviewIterations: maxIter, githubToken: req.token });
res.status(201).json({
session_id: record.session_id,
status: record.status,
image_count: record.image_count,
created_at: record.created_at,
});
});
// GET /v1/sessions/{id} — status.
r.get("/:id", (req: AuthedRequest, res) => {
const s = ownedSession(store, req.params.id, req.user!.github_user_id);
if (!s) {
sendError(res, 404, "session_not_found", "No such session");
return;
}
const body: Record<string, unknown> = {
session_id: s.session_id,
status: s.status,
phase: s.phase,
iterations_completed: s.iterations_completed,
iterations_max: s.iterations_max,
image_count: s.image_count,
created_at: s.created_at,
updated_at: s.updated_at,
};
if (s.status === "failed" && s.error) body.error = s.error;
res.json(body);
});
// GET /v1/sessions/{id}/output — the HTML document.
r.get("/:id/output", (req: AuthedRequest, res) => {
const s = ownedSession(store, req.params.id, req.user!.github_user_id);
if (!s) {
sendError(res, 404, "session_not_found", "No such session");
return;
}
if (s.status !== "ready_for_review" && s.status !== "closed") {
sendError(res, 409, "invalid_state", "Output not available until session is ready_for_review");
return;
}
const outPath = paths.sessionOutput(s.session_id);
if (!existsSync(outPath)) {
sendError(res, 409, "invalid_state", "Output not available");
return;
}
// Title + download filename mirror the uploaded file's name.
const base = existsSync(paths.sessionSourceName(s.session_id))
? readFileSync(paths.sessionSourceName(s.session_id), "utf8").trim() || "document"
: "document";
const html = titledAs(readFileSync(outPath, "utf8"), base);
res.setHeader("Content-Disposition", `inline; filename="${convertedHtmlFilename(base)}"`);
res.type("text/html").send(html);
});
// GET /v1/sessions/{id}/logs — the run log as ndjson.
r.get("/:id/logs", (req: AuthedRequest, res) => {
const s = ownedSession(store, req.params.id, req.user!.github_user_id);
if (!s) {
sendError(res, 404, "session_not_found", "No such session");
return;
}
const logPath = paths.sessionLog(s.session_id);
res.type("application/x-ndjson").send(existsSync(logPath) ? readFileSync(logPath, "utf8") : "");
});
// GET /v1/sessions/{id}/diagnostics — machine-readable timing/health summary
// for maintainers (human or AI): phase + per-call durations, the slowest
// calls, and any in-flight call (the likely culprit when a run seems hung).
r.get("/:id/diagnostics", (req: AuthedRequest, res) => {
const s = ownedSession(store, req.params.id, req.user!.github_user_id);
if (!s) {
sendError(res, 404, "session_not_found", "No such session");
return;
}
const logPath = paths.sessionLog(s.session_id);
const text = existsSync(logPath) ? readFileSync(logPath, "utf8") : "";
res.json(summarizeRun(text, { sessionId: s.session_id, status: s.status, phase: s.phase, now: Date.now() }));
});
// POST /v1/sessions/{id}/feedback — re-run within the same session.
r.post("/:id/feedback", (req: AuthedRequest, res) => {
const s = ownedSession(store, req.params.id, req.user!.github_user_id);
if (!s) {
sendError(res, 404, "session_not_found", "No such session");
return;
}
if (s.status !== "ready_for_review") {
sendError(res, 409, "invalid_state", "Feedback can only be submitted when ready_for_review");
return;
}
const feedback = (req.body as { feedback?: string } | undefined)?.feedback;
if (!feedback || typeof feedback !== "string") {
sendError(res, 400, "invalid_request", "feedback (string) is required");
return;
}
// Mark it non-terminal before enqueueing, both so a second feedback POST
// hits the ready_for_review guard and so a client polling a waiting re-run
// sees `queued` rather than a stale `ready_for_review` alongside a 202 that
// promised a new run.
//
// The claim is what makes that guard a lock rather than a check: the status
// test and the write are one statement, so of two concurrent callers exactly
// one proceeds to enqueue. The validation above stays ahead of it so a request
// with no feedback body still gets its 400 without disturbing the session.
// Losing the claim means someone else moved the session out of
// ready_for_review in between, which is the same condition the guard reports.
if (!store.claimSession(s.session_id, "ready_for_review", { status: "queued", phase: "extraction", error: null })) {
sendError(res, 409, "invalid_state", "Feedback can only be submitted when ready_for_review");
return;
}
const started = enqueueRun({
cfg,
store,
sessionId: s.session_id,
maxReviewIterations: s.iterations_max,
feedback,
githubToken: req.token,
});
// Report what actually happened. Under the run cap a re-run waits, and
// saying "running" then would make a queued session look hung. Both report
// `extraction` — the phase the run starts in; the queued branch used to say
// `triage`, a phase the pipeline never enters, and this 202 was the one place
// a client could actually observe it.
res.status(202).json(
started
? { session_id: s.session_id, status: "running", phase: "extraction" }
: { session_id: s.session_id, status: "queued", phase: "extraction" },
);
});
// POST /v1/sessions/{id}/close — finalize the session and clean tmp.
// Agent contributions are auto-filed as GitHub issues during the run (see
// pipeline/contribute.ts), so close no longer opens PRs.
r.post("/:id/close", (req: AuthedRequest, res) => {
const s = ownedSession(store, req.params.id, req.user!.github_user_id);
if (!s) {
sendError(res, 404, "session_not_found", "No such session");
return;
}
if (s.status !== "ready_for_review") {
sendError(res, 409, "invalid_state", "Session is not ready_for_review");
return;
}
// Claim the session BEFORE doing any of the work below, unlike the feedback
// endpoint where the claim sits last. Everything after this point mutates
// state outside the database — it captures fixtures into the shared agent
// library and deletes the session's tmp tree — and a loser that discovers it
// lost only afterwards has already done both. Two concurrent closes would
// file the same fixtures twice, so the claim goes first and the work happens
// only on the winning path.
if (!store.claimSession(s.session_id, "ready_for_review", { status: "closed" })) {
sendError(res, 409, "invalid_state", "Session is not ready_for_review");
return;
}
// Auto-capture regression fixtures from the accepted output: the
// page agent's per-page output + its source image, used to gate future agent
// updates so a change can't break a use it already handled. Best-effort —
// never block closing a session on fixture capture.
try {
const finalPath = paths.sessionFinalFragments(s.session_id);
if (existsSync(finalPath)) {
const saved = JSON.parse(readFileSync(finalPath, "utf8")) as {
fragments?: Fragment[];
failedPages?: number[];
uncorrectedPages?: number[];
};
// A page whose extraction failed is excluded. Its fragment is a `@page-failed`
// comment, and a fixture is an assertion that this HTML is the RIGHT output for
// this image — capturing one would gate every future page-agent update on
// reproducing a failure, and the first page of a document is exactly the
// representative image the fixture is keyed to. Read from the recorded set
// rather than sniffed out of the HTML: the run already knows which pages those
// are (pipeline/orchestrator.ts).
//
// And a page the fidelity check REJECTED, on the same argument carried one step
// further (#328). That page's markup is not a marker, it is content — content Iris's
// own verifier named problems in and the correction pass did not replace — so it
// would be captured as though it were right, and `captureFixtures` keys the fixture
// to the first fragment it is given. Every future page-agent update would then be
// gated on reproducing markup this run had already declared wrong. Accepting the
// session is a human saying the DOCUMENT is good enough to close, which is not the
// same claim as this page being the correct output for its image; before the set was
// recorded there was no way to tell them apart here, since the verdict is nowhere in
// the markup.
const excluded = new Set([...(saved.failedPages ?? []), ...(saved.uncorrectedPages ?? [])]);
captureFixtures(paths, s.session_id, (saved.fragments ?? []).filter((f) => !excluded.has(f.order)));
}
} catch {
// ignore — fixture capture must not prevent accepting a session
}
// Best-effort for the same reason as the capture above, and newly REQUIRED by
// the claim-first ordering: the row is already `closed`, so a throw here is
// unrecoverable rather than transient. `force: true` suppresses ENOENT only —
// an unwritable subdirectory still gives ENOTEMPTY. Previously the status write
// came last, so a throw left the session at ready_for_review and the client's
// retry re-ran the cleanup; now the retry would get a 409 while
// `data_dir/tmp/<id>` is orphaned anyway (nothing else touches tmpDir, and
// failStaleSessions only rewrites statuses). Swallowing keeps the leak a leak
// instead of also stranding the session.
try {
const tmp = paths.tmpDir(s.session_id);
if (existsSync(tmp)) rmSync(tmp, { recursive: true, force: true });
} catch {
// ignore — the session is closed either way
}
res.json({ session_id: s.session_id, status: "closed" });
});
return r;
}