f52a7b6 · 4cb357e · 9316a1c
@@ -0,0 +1,6 @@
# Real-cell proof: the build half, in process (#11999)
- `test/helpers/build-standin.ts` (its work in `build-standin-builder.ts`) is a deterministic stand-in for the model: it adds a `/reading` route and its test to the template's app (`READING-V1` in the prompt), or changes what it says (`READING-V2`), runs the app's real `bun test`, and pushes only when that passes.
- `test/build-journey.test.ts` walks a first ask to production in one process: `newApp` copies the template, the crew's stand-in builds, the pull's preview run says its tests passed, a person lands it, production serves `/reading`, and a second change lands the same way. About 8 s.
Not proved here, and tracked in #12001: the same walk inside a 2 GiB, 1-core cell, and a real memory breach restarting a child.
@@ -0,0 +1,144 @@
/**
 * The build half of a child's first hour (#11999), in this process: a first
 * ask builds an app from the template, the app's own tests pass in the
 * check, a person lands the pull, production serves the change, and a
 * second change lands the same way. The model is `build-standin.ts`, which
 * is deterministic and runs the app's real `bun test` before it pushes; the
 * template, the crew, the check, the merge and the runtime are the forge's
 * own. It runs in this process; the same walk in a Small cell is its own issue.
 */
import { afterAll, beforeAll, expect, setDefaultTimeout, test } from "bun:test";
import { mkdtemp, rm } from "node:fs/promises";
import { tmpdir } from "node:os";
import { join } from "node:path";
import { newApp } from "../src/actions/apps.ts";
import { createAgent } from "../src/actions/crew.ts";
import { assignIssue, openIssue } from "../src/actions/issues.ts";
import { mergePull } from "../src/actions/pulls.ts";
import { createApp, type App } from "../src/app.ts";
import { AppsRuntime } from "../src/apps/runtime.ts";
import { startChecks, type ChecksHandle } from "../src/checks/service.ts";
import { getRepoById, type Repo } from "../src/core/repos.ts";
import { listOpenPulls, type Pull } from "../src/core/pulls.ts";
import { listRuns } from "../src/core/runs.ts";
import { createUser, type User } from "../src/core/users.ts";
import { startCrew, type CrewHandle } from "../src/crew/dispatch.ts";
import { serve, type Servers } from "../src/http/server.ts";
import { READING_V1, READING_V2, writeBuildStandIn } from "./helpers/build-standin.ts";
import { quietLogs } from "./helpers/globals.ts";
import "./helpers/roomy-disk.ts";
quietLogs();
// The walk takes about 8 s; the rest is room for a loaded check machine.
setDefaultTimeout(60_000);
const DOMAIN = "apps.test";
let dir = "";
let app: App;
let servers: Servers;
let runtime: AppsRuntime;
let checks: ChecksHandle;
let crew: CrewHandle;
let alice: User;
let bot: User;
let repo: Repo;
let slug = "";
const viaHost = (path: string) => fetch(`${servers.url}${path}`, { headers: { host: `${slug}.${DOMAIN}` }, redirect: "manual" });
async function until<T>(what: string, read: () => T | undefined | Promise<T | undefined>, ms = 30_000): Promise<T> {
  const deadline = Date.now() + ms;
  while (Date.now() < deadline) {
    const value = await read();
    if (value !== undefined) return value;
    await Bun.sleep(100); // real-timer: polling a real forge, crew and runtime that run on their own clocks
  }
  throw new Error(`${what} did not happen in ${ms / 1000}s`);
}
beforeAll(async () => {
  dir = await mkdtemp(join(tmpdir(), "forge-build-journey-"));
  app = createApp({ dataDir: dir, port: 0 });
  servers = await serve(app);
  runtime = new AppsRuntime(app, { domain: DOMAIN, scheme: "http", publicPort: servers.port, forgeUrl: servers.url });
  await runtime.start();
  checks = startChecks(app, { forgeUrl: servers.url, concurrency: 1 });
  crew = await startCrew(app, {
    forgeUrl: servers.url,
    localUrl: servers.url,
    binary: writeBuildStandIn(dir).fake,
    concurrency: 1,
    budgetUsd: 1,
    runTimeoutMs: 30_000,
    networkRetryScale: 0,
    tickMs: 100,
  });
  const made = await createUser(app.db, { origin: "seed", trusted: true, username: "alice", email: "[email protected]", password: "a-good-passphrase", adminIfFirst: true });
  if (!made.ok) throw new Error(made.error);
  alice = made.value;
  const agent = await createAgent(app, alice, { username: "bot", brief: "Build what is asked." });
  if (!agent.ok) throw new Error(String(agent.error));
  bot = agent.value.user;
});
afterAll(async () => {
  await crew?.stop();
  await checks?.stop();
  await runtime?.stop();
  servers?.stop();
  app?.close();
  await rm(dir, { recursive: true, force: true });
});
const fresh = () => getRepoById(app.db, repo.id)!;
/** The pull the crew's run opened for `issueId`'s work, once there is one that is not `known`. */
const newPull = (known: ReadonlySet<number>) =>
  until<Pull>("the crew's pull", () => listOpenPulls(app.db, repo.id).find((p) => !known.has(p.number)));
/**
 * Land `pull` as the person, once its preview has been built and its tests
 * have passed: the pull's own deploy run says "its tests passed".
 */
async function land(pull: Pull): Promise<void> {
  const built = await until("the pull's preview and its tests", () =>
    listRuns(app.db, repo.id, { kind: "deploy", env: `pull-${pull.number}` }).find((r) => r.state === "ok" || r.state === "failed"),
  );
  expect(built.state, built.summary ?? "").toBe("ok");
  expect(built.summary ?? "").toContain("its tests passed");
  const merged = await mergePull(app, alice, fresh(), pull, "squash");
  if (!merged.ok) throw new Error(JSON.stringify(merged.error));
}
test("a first ask builds an app from the template; its tests pass; production serves it", async () => {
  const made = await newApp(app, alice, { name: "reading", description: "A page that lists my reading. READING-V1", agent: "bot", visibility: "public" });
  if (!made.ok) throw new Error(JSON.stringify(made.error));
  repo = made.value.repo;
  slug = made.value.app.slug;
  // The template's files are the app's, whole: its own tests came with it.
  expect(made.value.issue).toBeDefined();
  const pull = await newPull(new Set());
  await land(pull);
  const served = await until("production to serve the change", async () => {
    const res = await viaHost("/reading").catch(() => undefined);
    const body = res === undefined ? "" : await res.text();
    return res?.status === 200 && body === READING_V1 ? body : undefined;
  });
  expect(served).toBe(READING_V1);
});
test("a second change lands the same way and production serves that", async () => {
  const known = new Set(listOpenPulls(app.db, repo.id).map((p) => p.number));
  const asked = openIssue(app, alice, fresh(), { title: "Two books", body: "Say two books. READING-V2" });
  if (!asked.ok) throw new Error(String(asked.error));
  const assigned = assignIssue(app, alice, fresh(), asked.value, bot);
  if (!assigned.ok) throw new Error(String(assigned.error));
  const pull = await newPull(known);
  await land(pull);
  const served = await until("production to serve the second change", async () => {
    const res = await viaHost("/reading").catch(() => undefined);
    const body = res === undefined ? "" : await res.text();
    return body === READING_V2 ? body : undefined;
  });
  expect(served).toBe(READING_V2);
});
@@ -0,0 +1,47 @@
/**
 * What `build-standin.ts`'s stand-in does, run by `bun` in the clone it is
 * given: edit, run the app's own tests, and only then commit and push.
 * Argument: the file holding the prompt.
 */
import { readFileSync, writeFileSync } from "node:fs";
import { ANCHOR, READING_V1, READING_V2, routeFor, testFor } from "./build-standin.ts";
const prompt = readFileSync(process.argv[2]!, "utf8");
const say = prompt.includes("READING-V2") ? READING_V2 : prompt.includes("READING-V1") ? READING_V1 : undefined;
const finish = (ok: boolean, result: string): void =>
  console.log(
    JSON.stringify({ type: "result", subtype: ok ? "success" : "error_during_execution", is_error: !ok, result, total_cost_usd: 0.01, num_turns: 1, session_id: "build-1", terminal_reason: ok ? "completed" : "tool_error" }),
  );
const git = (...args: string[]): void => {
  const r = Bun.spawnSync(["git", ...args], { stdout: "pipe", stderr: "pipe" });
  if (r.exitCode !== 0) throw new Error(`${args.join(" ")}: ${r.stderr.toString()}`);
};
if (say === undefined) {
  finish(true, "DECLINED: nothing here I know how to build.");
  process.exit(0);
}
const routes = readFileSync("src/routes.ts", "utf8");
if (say === READING_V1) {
  if (!routes.includes(ANCHOR)) {
    finish(false, "the template's routes have changed shape");
    process.exit(0);
  }
  writeFileSync("src/routes.ts", routes.replace(ANCHOR, routeFor(READING_V1) + ANCHOR));
} else {
  if (!routes.includes(JSON.stringify(READING_V1))) {
    finish(false, "there is no first change to change");
    process.exit(0);
  }
  writeFileSync("src/routes.ts", routes.replace(JSON.stringify(READING_V1), JSON.stringify(READING_V2)));
}
writeFileSync("src/reading.test.ts", testFor(say));
const tested = Bun.spawnSync(["bun", "test"], { stdout: "pipe", stderr: "pipe" });
if (tested.exitCode !== 0) {
  finish(false, `its tests failed: ${tested.stderr.toString().slice(-300)}`);
  process.exit(0);
}
git("add", "-A");
git("commit", "-qm", say === READING_V1 ? "Add the reading list" : "Say two books");
git("push", "-q", "-u", "origin", "HEAD");
finish(true, `Added ${say}, tests pass.`);
@@ -0,0 +1,68 @@
/**
 * A deterministic stand-in for the model that builds on the template
 * (#11999). It speaks stream-json as Claude Code does, but what it does is
 * fixed: given a prompt carrying `READING-V1` it adds a `/reading` route and
 * a test of it to the app the template made; given `READING-V2` it changes
 * what that route says and its test. Then it runs the app's own `bun test`
 * in the clone, and only commits and pushes when that passes; a red test
 * ends the run as a failure, as a model that cannot make its tests pass
 * would. No network, no login, no cost, the same bytes every time.
 *
 * It is the model's place in `build-journey.test.ts`.
 */
import { chmodSync, writeFileSync } from "node:fs";
import { join } from "node:path";
/** What `/reading` says after the first change and after the second. */
export const READING_V1 = "reading list: one book";
export const READING_V2 = "reading list: two books";
/** The route the first change inserts, before the template's health route. */
export const ANCHOR = "  // Up or not, for the forge and for anyone watching";
export const routeFor = (say: string): string =>
  `  { method: "GET", path: "/reading", handle: () => text(${JSON.stringify(say)}) },\n\n`;
export const testFor = (say: string): string => `import { afterAll, expect, test } from "bun:test";
import { mkdtempSync, rmSync } from "node:fs";
import { tmpdir } from "node:os";
import { join } from "node:path";
import { serve } from "./server.ts";
const dataDir = mkdtempSync(join(tmpdir(), "reading-"));
const server = serve({ port: 0, dataDir, identityKey: "a-key-only-the-forge-holds" });
afterAll(() => {
  server.stop(true);
  rmSync(dataDir, { recursive: true, force: true });
});
test("/reading says the reading list", async () => {
  const res = await fetch(\`http://\${server.hostname}:\${server.port}/reading\`);
  expect(res.status).toBe(200);
  expect(await res.text()).toBe(${JSON.stringify(say)});
});
`;
export interface BuildStandIn {
  /** The binary to hand `startCrew` (or `FORGE_CLAUDE`). */
  readonly fake: string;
}
/** Write the stand-in into `dir`: a shell script that reads the prompt and runs the builder. */
export function writeBuildStandIn(dir: string): BuildStandIn {
  const builder = join(import.meta.dir, "build-standin-builder.ts");
  const promptFile = join(dir, "build-standin.prompt");
  const fake = join(dir, "build-claude");
  writeFileSync(
    fake,
    `#!/bin/sh
if [ "$1" = "--version" ]; then echo "2.1.0 (build stand-in)"; exit 0; fi
IFS= read -r first; printf '%s\\n' "$first" > "${promptFile}"
echo '{"type":"system","subtype":"init","session_id":"build-1","model":"stand-in"}'
exec ${JSON.stringify(process.execPath)} "${builder}" "${promptFile}"
`,
  );
  chmodSync(fake, 0o755);
  return { fake };
}
		@@ -0,0 +1,6 @@
	1	# Real-cell proof: the build half, in process (#11999)
	2
	3	- `test/helpers/build-standin.ts` (its work in `build-standin-builder.ts`) is a deterministic stand-in for the model: it adds a `/reading` route and its test to the template's app (`READING-V1` in the prompt), or changes what it says (`READING-V2`), runs the app's real `bun test`, and pushes only when that passes.
	4	- `test/build-journey.test.ts` walks a first ask to production in one process: `newApp` copies the template, the crew's stand-in builds, the pull's preview run says its tests passed, a person lands it, production serves `/reading`, and a second change lands the same way. About 8 s.
	5
	6	Not proved here, and tracked in #12001: the same walk inside a 2 GiB, 1-core cell, and a real memory breach restarting a child.
		@@ -0,0 +1,144 @@
	1	/**
	2	* The build half of a child's first hour (#11999), in this process: a first
	3	* ask builds an app from the template, the app's own tests pass in the
	4	* check, a person lands the pull, production serves the change, and a
	5	* second change lands the same way. The model is `build-standin.ts`, which
	6	* is deterministic and runs the app's real `bun test` before it pushes; the
	7	* template, the crew, the check, the merge and the runtime are the forge's
	8	* own. It runs in this process; the same walk in a Small cell is its own issue.
	9	*/
	10	import { afterAll, beforeAll, expect, setDefaultTimeout, test } from "bun:test";
	11	import { mkdtemp, rm } from "node:fs/promises";
	12	import { tmpdir } from "node:os";
	13	import { join } from "node:path";
	14	import { newApp } from "../src/actions/apps.ts";
	15	import { createAgent } from "../src/actions/crew.ts";
	16	import { assignIssue, openIssue } from "../src/actions/issues.ts";
	17	import { mergePull } from "../src/actions/pulls.ts";
	18	import { createApp, type App } from "../src/app.ts";
	19	import { AppsRuntime } from "../src/apps/runtime.ts";
	20	import { startChecks, type ChecksHandle } from "../src/checks/service.ts";
	21	import { getRepoById, type Repo } from "../src/core/repos.ts";
	22	import { listOpenPulls, type Pull } from "../src/core/pulls.ts";
	23	import { listRuns } from "../src/core/runs.ts";
	24	import { createUser, type User } from "../src/core/users.ts";
	25	import { startCrew, type CrewHandle } from "../src/crew/dispatch.ts";
	26	import { serve, type Servers } from "../src/http/server.ts";
	27	import { READING_V1, READING_V2, writeBuildStandIn } from "./helpers/build-standin.ts";
	28	import { quietLogs } from "./helpers/globals.ts";
	29	import "./helpers/roomy-disk.ts";
	30
	31	quietLogs();
	32	// The walk takes about 8 s; the rest is room for a loaded check machine.
	33	setDefaultTimeout(60_000);
	34
	35	const DOMAIN = "apps.test";
	36	let dir = "";
	37	let app: App;
	38	let servers: Servers;
	39	let runtime: AppsRuntime;
	40	let checks: ChecksHandle;
	41	let crew: CrewHandle;
	42	let alice: User;
	43	let bot: User;
	44	let repo: Repo;
	45	let slug = "";
	46
	47	const viaHost = (path: string) => fetch(`${servers.url}${path}`, { headers: { host: `${slug}.${DOMAIN}` }, redirect: "manual" });
	48
	49	async function until<T>(what: string, read: () => T \| undefined \| Promise<T \| undefined>, ms = 30_000): Promise<T> {
	50	const deadline = Date.now() + ms;
	51	while (Date.now() < deadline) {
	52	const value = await read();
	53	if (value !== undefined) return value;
	54	await Bun.sleep(100); // real-timer: polling a real forge, crew and runtime that run on their own clocks
	55	}
	56	throw new Error(`${what} did not happen in ${ms / 1000}s`);
	57	}
	58
	59	beforeAll(async () => {
	60	dir = await mkdtemp(join(tmpdir(), "forge-build-journey-"));
	61	app = createApp({ dataDir: dir, port: 0 });
	62	servers = await serve(app);
	63	runtime = new AppsRuntime(app, { domain: DOMAIN, scheme: "http", publicPort: servers.port, forgeUrl: servers.url });
	64	await runtime.start();
	65	checks = startChecks(app, { forgeUrl: servers.url, concurrency: 1 });
	66	crew = await startCrew(app, {
	67	forgeUrl: servers.url,
	68	localUrl: servers.url,
	69	binary: writeBuildStandIn(dir).fake,
	70	concurrency: 1,
	71	budgetUsd: 1,
	72	runTimeoutMs: 30_000,
	73	networkRetryScale: 0,
	74	tickMs: 100,
	75	});
	76	const made = await createUser(app.db, { origin: "seed", trusted: true, username: "alice", email: "[email protected]", password: "a-good-passphrase", adminIfFirst: true });
	77	if (!made.ok) throw new Error(made.error);
	78	alice = made.value;
	79	const agent = await createAgent(app, alice, { username: "bot", brief: "Build what is asked." });
	80	if (!agent.ok) throw new Error(String(agent.error));
	81	bot = agent.value.user;
	82	});
	83
	84	afterAll(async () => {
	85	await crew?.stop();
	86	await checks?.stop();
	87	await runtime?.stop();
	88	servers?.stop();
	89	app?.close();
	90	await rm(dir, { recursive: true, force: true });
	91	});
	92
	93	const fresh = () => getRepoById(app.db, repo.id)!;
	94
	95	/** The pull the crew's run opened for `issueId`'s work, once there is one that is not `known`. */
	96	const newPull = (known: ReadonlySet<number>) =>
	97	until<Pull>("the crew's pull", () => listOpenPulls(app.db, repo.id).find((p) => !known.has(p.number)));
	98
	99	/**
	100	* Land `pull` as the person, once its preview has been built and its tests
	101	* have passed: the pull's own deploy run says "its tests passed".
	102	*/
	103	async function land(pull: Pull): Promise<void> {
	104	const built = await until("the pull's preview and its tests", () =>
	105	listRuns(app.db, repo.id, { kind: "deploy", env: `pull-${pull.number}` }).find((r) => r.state === "ok" \|\| r.state === "failed"),
	106	);
	107	expect(built.state, built.summary ?? "").toBe("ok");
	108	expect(built.summary ?? "").toContain("its tests passed");
	109	const merged = await mergePull(app, alice, fresh(), pull, "squash");
	110	if (!merged.ok) throw new Error(JSON.stringify(merged.error));
	111	}
	112
	113	test("a first ask builds an app from the template; its tests pass; production serves it", async () => {
	114	const made = await newApp(app, alice, { name: "reading", description: "A page that lists my reading. READING-V1", agent: "bot", visibility: "public" });
	115	if (!made.ok) throw new Error(JSON.stringify(made.error));
	116	repo = made.value.repo;
	117	slug = made.value.app.slug;
	118	// The template's files are the app's, whole: its own tests came with it.
	119	expect(made.value.issue).toBeDefined();
	120	const pull = await newPull(new Set());
	121	await land(pull);
	122	const served = await until("production to serve the change", async () => {
	123	const res = await viaHost("/reading").catch(() => undefined);
	124	const body = res === undefined ? "" : await res.text();
	125	return res?.status === 200 && body === READING_V1 ? body : undefined;
	126	});
	127	expect(served).toBe(READING_V1);
	128	});
	129
	130	test("a second change lands the same way and production serves that", async () => {
	131	const known = new Set(listOpenPulls(app.db, repo.id).map((p) => p.number));
	132	const asked = openIssue(app, alice, fresh(), { title: "Two books", body: "Say two books. READING-V2" });
	133	if (!asked.ok) throw new Error(String(asked.error));
	134	const assigned = assignIssue(app, alice, fresh(), asked.value, bot);
	135	if (!assigned.ok) throw new Error(String(assigned.error));
	136	const pull = await newPull(known);
	137	await land(pull);
	138	const served = await until("production to serve the second change", async () => {
	139	const res = await viaHost("/reading").catch(() => undefined);
	140	const body = res === undefined ? "" : await res.text();
	141	return body === READING_V2 ? body : undefined;
	142	});
	143	expect(served).toBe(READING_V2);
	144	});
		@@ -0,0 +1,47 @@
	1	/**
	2	* What `build-standin.ts`'s stand-in does, run by `bun` in the clone it is
	3	* given: edit, run the app's own tests, and only then commit and push.
	4	* Argument: the file holding the prompt.
	5	*/
	6	import { readFileSync, writeFileSync } from "node:fs";
	7	import { ANCHOR, READING_V1, READING_V2, routeFor, testFor } from "./build-standin.ts";
	8
	9	const prompt = readFileSync(process.argv[2]!, "utf8");
	10	const say = prompt.includes("READING-V2") ? READING_V2 : prompt.includes("READING-V1") ? READING_V1 : undefined;
	11	const finish = (ok: boolean, result: string): void =>
	12	console.log(
	13	JSON.stringify({ type: "result", subtype: ok ? "success" : "error_during_execution", is_error: !ok, result, total_cost_usd: 0.01, num_turns: 1, session_id: "build-1", terminal_reason: ok ? "completed" : "tool_error" }),
	14	);
	15	const git = (...args: string[]): void => {
	16	const r = Bun.spawnSync(["git", ...args], { stdout: "pipe", stderr: "pipe" });
	17	if (r.exitCode !== 0) throw new Error(`${args.join(" ")}: ${r.stderr.toString()}`);
	18	};
	19
	20	if (say === undefined) {
	21	finish(true, "DECLINED: nothing here I know how to build.");
	22	process.exit(0);
	23	}
	24	const routes = readFileSync("src/routes.ts", "utf8");
	25	if (say === READING_V1) {
	26	if (!routes.includes(ANCHOR)) {
	27	finish(false, "the template's routes have changed shape");
	28	process.exit(0);
	29	}
	30	writeFileSync("src/routes.ts", routes.replace(ANCHOR, routeFor(READING_V1) + ANCHOR));
	31	} else {
	32	if (!routes.includes(JSON.stringify(READING_V1))) {
	33	finish(false, "there is no first change to change");
	34	process.exit(0);
	35	}
	36	writeFileSync("src/routes.ts", routes.replace(JSON.stringify(READING_V1), JSON.stringify(READING_V2)));
	37	}
	38	writeFileSync("src/reading.test.ts", testFor(say));
	39	const tested = Bun.spawnSync(["bun", "test"], { stdout: "pipe", stderr: "pipe" });
	40	if (tested.exitCode !== 0) {
	41	finish(false, `its tests failed: ${tested.stderr.toString().slice(-300)}`);
	42	process.exit(0);
	43	}
	44	git("add", "-A");
	45	git("commit", "-qm", say === READING_V1 ? "Add the reading list" : "Say two books");
	46	git("push", "-q", "-u", "origin", "HEAD");
	47	finish(true, `Added ${say}, tests pass.`);
		@@ -0,0 +1,68 @@
	1	/**
	2	* A deterministic stand-in for the model that builds on the template
	3	* (#11999). It speaks stream-json as Claude Code does, but what it does is
	4	* fixed: given a prompt carrying `READING-V1` it adds a `/reading` route and
	5	* a test of it to the app the template made; given `READING-V2` it changes
	6	* what that route says and its test. Then it runs the app's own `bun test`
	7	* in the clone, and only commits and pushes when that passes; a red test
	8	* ends the run as a failure, as a model that cannot make its tests pass
	9	* would. No network, no login, no cost, the same bytes every time.
	10	*
	11	* It is the model's place in `build-journey.test.ts`.
	12	*/
	13
	14	import { chmodSync, writeFileSync } from "node:fs";
	15	import { join } from "node:path";
	16
	17	/** What `/reading` says after the first change and after the second. */
	18	export const READING_V1 = "reading list: one book";
	19	export const READING_V2 = "reading list: two books";
	20
	21	/** The route the first change inserts, before the template's health route. */
	22	export const ANCHOR = " // Up or not, for the forge and for anyone watching";
	23
	24	export const routeFor = (say: string): string =>
	25	` { method: "GET", path: "/reading", handle: () => text(${JSON.stringify(say)}) },\n\n`;
	26
	27	export const testFor = (say: string): string => `import { afterAll, expect, test } from "bun:test";
	28	import { mkdtempSync, rmSync } from "node:fs";
	29	import { tmpdir } from "node:os";
	30	import { join } from "node:path";
	31	import { serve } from "./server.ts";
	32
	33	const dataDir = mkdtempSync(join(tmpdir(), "reading-"));
	34	const server = serve({ port: 0, dataDir, identityKey: "a-key-only-the-forge-holds" });
	35	afterAll(() => {
	36	server.stop(true);
	37	rmSync(dataDir, { recursive: true, force: true });
	38	});
	39
	40	test("/reading says the reading list", async () => {
	41	const res = await fetch(\`http://\${server.hostname}:\${server.port}/reading\`);
	42	expect(res.status).toBe(200);
	43	expect(await res.text()).toBe(${JSON.stringify(say)});
	44	});
	45	`;
	46
	47	export interface BuildStandIn {
	48	/** The binary to hand `startCrew` (or `FORGE_CLAUDE`). */
	49	readonly fake: string;
	50	}
	51
	52	/** Write the stand-in into `dir`: a shell script that reads the prompt and runs the builder. */
	53	export function writeBuildStandIn(dir: string): BuildStandIn {
	54	const builder = join(import.meta.dir, "build-standin-builder.ts");
	55	const promptFile = join(dir, "build-standin.prompt");
	56	const fake = join(dir, "build-claude");
	57	writeFileSync(
	58	fake,
	59	`#!/bin/sh
	60	if [ "$1" = "--version" ]; then echo "2.1.0 (build stand-in)"; exit 0; fi
	61	IFS= read -r first; printf '%s\\n' "$first" > "${promptFile}"
	62	echo '{"type":"system","subtype":"init","session_id":"build-1","model":"stand-in"}'
	63	exec ${JSON.stringify(process.execPath)} "${builder}" "${promptFile}"
	64	`,
	65	);
	66	chmodSync(fake, 0o755);
	67	return { fake };
	68	}