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// A stand-in for the Iris API surface this harness uses, plus a PDF generator, so
// all three stages can be exercised without a token, a network, or a live
// deployment's queue. It is not a model of Iris — it is a fixture whose shapes match
// the endpoints and log events this harness reads, which is what the harness can
// actually get wrong.
//
// Its `max_pages` is deliberately tiny (3) so that splitting, the most intricate
// part of prepare, is exercised by a seven-page document rather than a 60-page one.
import { createServer } from "node:http";
// A minimal, valid, uncompressed multi-page PDF. Written by hand rather than shelled
// out to a generator so the test runs the same way on any machine — the point of the
// fixture is that `pdfinfo` reads a real page count and real page dimensions off it.
export function makePdf(pages, { widthPts = 612, heightPts = 792 } = {}) {
const objs = [];
const pageIds = [];
// 1: catalog, 2: page tree, 3: font, then two objects per page.
for (let i = 0; i < pages; i++) pageIds.push(4 + i * 2);
objs[1] = `<< /Type /Catalog /Pages 2 0 R >>`;
objs[2] = `<< /Type /Pages /Kids [${pageIds.map((id) => `${id} 0 R`).join(" ")}] /Count ${pages} >>`;
objs[3] = `<< /Type /Font /Subtype /Type1 /BaseFont /Helvetica >>`;
for (let i = 0; i < pages; i++) {
const pageId = pageIds[i];
const contentId = pageId + 1;
objs[pageId] =
`<< /Type /Page /Parent 2 0 R /MediaBox [0 0 ${widthPts} ${heightPts}] ` +
`/Resources << /Font << /F1 3 0 R >> >> /Contents ${contentId} 0 R >>`;
const stream = `BT /F1 24 Tf 72 700 Td (Page ${i + 1} of ${pages}) Tj ET`;
objs[contentId] = `<< /Length ${stream.length} >>\nstream\n${stream}\nendstream`;
}
let out = "%PDF-1.4\n";
const offsets = [];
for (let id = 1; id < objs.length; id++) {
offsets[id] = out.length;
out += `${id} 0 obj\n${objs[id]}\nendobj\n`;
}
const xref = out.length;
const count = objs.length; // object 0 is the free-list head
out += `xref\n0 ${count}\n0000000000 65535 f \n`;
for (let id = 1; id < count; id++) out += `${String(offsets[id]).padStart(10, "0")} 00000 n \n`;
out += `trailer\n<< /Size ${count} /Root 1 0 R >>\nstartxref\n${xref}\n%%EOF\n`;
return Buffer.from(out, "latin1");
}
// A delivered document with a defect of each kind the local re-lint is configured to
// catch: a duplicate id on inert elements, on active ones, and on one an ARIA/`for`
// reference points at (the case axe reports as `incomplete`, which lint.mjs promotes).
export const OUTPUT_HTML = `<!DOCTYPE html><html><head><title>Fixture</title></head><body>
<h1>Fixture</h1><p>Some delivered content, long enough not to look like an empty document.</p>
<ul><li id="x">a</li><li id="x">b</li></ul>
<label for="q1">Name</label><input id="q1"><input id="q1">
<a id="dup" href="#x">one</a><a id="dup" href="#x">two</a>
<img src="fig.png">
</body></html>`;
const limitsFor = (maxRequestBytes) => ({
pdf: { max_pages: 3 },
upload: { max_files: 25, max_request_bytes: maxRequestBytes },
image: { max_bytes: 3932160, max_long_edge_px: 1568, max_dimension_px: 8000 },
// High enough that the submit spacing does not dominate the test's runtime, while
// still exercising the code path that derives spacing from the published limit.
rate_limits: { general_per_minute: 240, auth_per_minute: 60, upload_per_minute: 600 },
});
const MODEL = "us.anthropic.claude-sonnet-4-6";
const call = (agent, capability, duration_ms, input_tokens, output_tokens) => ({
ts: "2026-01-01T00:00:02.000Z",
type: "model_call",
agent,
model: MODEL,
provider: "bedrock",
capability,
duration_ms,
ok: true,
input_tokens,
output_tokens,
});
const LOG_LINES = [
{ ts: "2026-01-01T00:00:00.000Z", type: "run_start", pages: 3 },
{ ts: "2026-01-01T00:00:01.000Z", type: "phase", phase: "extraction" },
call("page", "vision", 4000, 5200, 1400),
call("reader", "text", 2000, 3000, 500),
{ ts: "2026-01-01T00:00:04.000Z", type: "assembly", pages: 3, lint_ok: false, violations: 2 },
{ ts: "2026-01-01T00:00:05.000Z", type: "reader", iteration: 0, issues: 2 },
{ ts: "2026-01-01T00:00:06.000Z", type: "editor", iteration: 1 },
{ ts: "2026-01-01T00:00:07.000Z", type: "editor_links_dropped", iteration: 1, hrefs: ["#x"] },
{ ts: "2026-01-01T00:00:08.000Z", type: "reader", iteration: 1, issues: 0 },
{ ts: "2026-01-01T00:00:09.000Z", type: "run_complete", iterations: 1, unresolved: 0, mode: "initial" },
];
const FAIL_ERROR = "page 2: stream stalled (idle) after 60000ms";
const diagnostics = (id, failed) => ({
session_id: id,
status: failed ? "failed" : "ready_for_review",
phase: "done",
elapsed_ms: 9000,
in_flight: null,
in_flight_count: 0,
phase_durations_ms: { extraction: 5000, assembly: 1000, review: 3000 },
model_calls: { count: 2, failed: 0, total_ms: 6000, avg_ms: 3000, max_ms: 4000, concurrency_factor: 0.67 },
tokens: { input: 8200, output: 1900, cache_read: 0, cache_write: 0, calls_reported: 2 },
by_agent: {
page: { count: 1, total_ms: 4000, max_ms: 4000, input_tokens: 5200, output_tokens: 1400 },
reader: { count: 1, total_ms: 2000, max_ms: 2000, input_tokens: 3000, output_tokens: 500 },
},
slowest_calls: [],
errors: failed ? [{ ts: null, type: "run_failed", message: FAIL_ERROR }] : [],
});
// `failNth` makes one submission end in `status: "failed"`, so the test covers the
// capture path that matters most: a run whose only account of itself is its log.
// `unauthorizedAfter` starts refusing submissions partway through, which is what an
// expiring GitHub user token looks like to a campaign that outlives it.
export function startStub({
pages = 7,
failNth = 3,
unauthorizedAfter = Infinity,
// Tiny in the one test that needs a chunk to come out unsendable — the real cap is
// 128 MB, which no fixture is going to reach.
maxRequestBytes = 134217728,
} = {}) {
const LIMITS = limitsFor(maxRequestBytes);
const pdf = makePdf(pages);
// Distinct byte counts so these three are not each other's duplicates.
const dribbled = makePdf(2);
const flakyPdf = makePdf(1);
const sessions = new Map();
let submissions = 0;
let flakyRequests = 0;
const server = createServer(async (req, res) => {
const path = new URL(req.url, "http://stub").pathname;
const send = (code, body, type = "application/json") => {
res.writeHead(code, { "content-type": type });
res.end(typeof body === "string" || Buffer.isBuffer(body) ? body : JSON.stringify(body));
};
if (path === "/v1/limits") return send(200, LIMITS);
// Two URLs, same bytes, so duplicate collapsing is exercised.
if (path === "/a.pdf" || path === "/copy.pdf") return send(200, pdf, "application/pdf");
// An HTML sign-in page served as application/pdf — why prepare checks magic bytes.
if (path === "/signin") return send(200, "<html>please sign in</html>", "application/pdf");
if (path === "/missing") return send(404, "not found", "text/plain");
// The three shapes a real document server takes, which one flat deadline cannot
// tell apart. Headers and a few bytes, then silence forever — a hung transfer:
if (path === "/stall.pdf" || (path === "/flaky.pdf" && flakyRequests++ === 0)) {
res.writeHead(200, { "content-type": "application/pdf" });
res.write(pdf.subarray(0, 64));
return; // deliberately never ended
}
// ...the same host on a second attempt, working:
if (path === "/flaky.pdf") return send(200, flakyPdf, "application/pdf");
// ...and slow but always progressing, which must succeed however long it takes.
if (path === "/dribble.pdf") {
res.writeHead(200, { "content-type": "application/pdf" });
const step = Math.ceil(dribbled.length / 5);
for (let at = 0; at < dribbled.length; at += step) {
res.write(dribbled.subarray(at, at + step));
await new Promise((r) => setTimeout(r, 250));
}
return res.end();
}
if (path === "/v1/sessions" && req.method === "POST") {
for await (const chunk of req) void chunk; // drain the multipart body
submissions += 1;
if (submissions > unauthorizedAfter) {
return send(401, { error: "unauthorized", message: "Bad credentials" });
}
const id = `sess-${submissions}`;
sessions.set(id, { polls: 0, failed: submissions === failNth });
return send(200, { session_id: id, status: "queued" });
}
const m = path.match(/^\/v1\/sessions\/([^/]+)(\/.*)?$/);
if (m) {
const s = sessions.get(m[1]);
if (!s) return send(404, { error: "no such session" });
switch (m[2] ?? "") {
case "":
// queued -> running -> terminal, so the queue-wait measurement has
// something to measure.
s.polls += 1;
if (s.polls === 1) return send(200, { session_id: m[1], status: "queued", phase: "queued" });
if (s.polls === 2) return send(200, { session_id: m[1], status: "running", phase: "extraction" });
return send(200, {
session_id: m[1],
status: s.failed ? "failed" : "ready_for_review",
phase: s.failed ? "extraction" : "done",
iterations_completed: 1,
...(s.failed ? { error: FAIL_ERROR } : {}),
});
case "/logs":
return send(200, LOG_LINES.map((e) => JSON.stringify(e)).join("\n"), "application/x-ndjson");
case "/diagnostics":
return send(200, diagnostics(m[1], s.failed));
case "/output":
return send(200, OUTPUT_HTML, "text/html");
case "/close":
return send(200, { ok: true });
}
}
send(404, { error: `stub has no route for ${path}` });
});
return new Promise((resolve) => {
server.listen(0, "127.0.0.1", () => {
const { port } = server.address();
resolve({ base: `http://127.0.0.1:${port}/v1`, origin: `http://127.0.0.1:${port}`, server });
});
});
}