97f5edb · 80d2e5b · b9e759e · 4ad7a6a · ef63cd6 · a9b30ef
@@ -181,6 +181,11 @@
export const SLOW: readonly Exclusion[] = [
    reason: "the cold home takes 20 timed requests at ~65 ms, each after a write, until its per-app reads are grouped",
    issue: 12310,
  },
  {
    file: "test/spawn-stranger-cell.test.ts",
    reason: "one test walks a whole forge in a real cell, from claim link to the first pull green; it is not a unit",
    issue: 12077,
      
        junipercrew
        1d ago
        
      
      The SLOW entry is acceptable for this walk, but the check must still run it unskipped. Prove that in the description.
    
  },
];
/** Everything wrong with one list of named files, `list` its name in the lines. */
@@ -0,0 +1,325 @@
/**
 * The stranger's path in a real cell (#12077), as `test/spawn-stranger-cell.test.ts`
 * drives it. Its own process, for the reason `real-spawn-e2e.ts` is: a check
 * sets `FORGE_CELLS=off` for the forges its tests boot.
 *
 * A parent with the real server is asked for a forge by name, as the home's
 * form does, with no account. The child comes up in a real cell at the plan's
 * size; the walk claims it through its claim link with the forge's name as
 * username and no email, asks "a page that says hello and shows today's date"
 * through the form on its first screen, and follows the ask to a new app, its
 * first pull landing on green, and the app's own address serving the page to
 * its owner as a browser reaches it. The model is `stranger-claude.sh`, a
 * fixed edit; the cell, the crew's clone and push, the pull, its check, the
 * merge and the deploy are the forge's own. The second ask is not walked.
 *
 * Each phase has its own budget and a message that names it. Printed: one JSON
 * line of what was seen, with `error` set to the phase that did not finish.
 */
import { mkdtempSync, rmSync } from "node:fs";
import { tmpdir } from "node:os";
import { join } from "node:path";
import { createApp } from "../../src/app.ts";
import { cellsAvailable } from "../../src/apps/contain.ts";
import { serve } from "../../src/http/server.ts";
import { useProbe } from "../../src/spawn/door.ts";
import { cellChildren, stopCellChildren } from "../../src/spawn/cell.ts";
import { lastStop } from "../../src/spawn/stopped.ts";
import "./roomy-disk.ts";
// One budget per phase, each counted from when it began: a slow phase never
// shortens the next, and a run that ran out says which one did. The test's
// timeout is their sum plus slack (see test/spawn-stranger-cell.test.ts).
const BUDGET_MS = { up: 120_000, ask: 60_000, pull: 120_000, serves: 60_000 } as const;
const NAME = "kid";
const ASK = "a page that says hello and shows today's date";
const PARENT = "forge.test";
const CHILD = `${NAME}--${PARENT}`;
process.env["FORGE_CLAUDE"] = join(import.meta.dir, "stranger-claude.sh");
let cleanup: () => Promise<void> = async () => {};
process.on("SIGTERM", () => void cleanup().finally(() => process.exit(143)));
const cells = await cellsAvailable();
if (!cells.ok) {
  console.log(JSON.stringify({ skipped: cells.reason }));
  process.exit(0);
}
const dataDir = mkdtempSync(join(tmpdir(), "real-stranger-"));
const parent = createApp({ dataDir, url: `https://${PARENT}`, port: 0 });
const servers = await serve(parent);
// The child's public address (https://kid--forge.test) is not a real host here: ask it
// the way every other request in this walk does, through the parent's own server with
// the child's Host, on /claim (the parent answers /healthz itself, whatever the host).
useProbe(parent, () => async () => {
  try {
    const res = await fetch(`${servers.url}/claim`, {
      headers: { host: CHILD },
      redirect: "manual",
      signal: AbortSignal.timeout(700),
    });
    await res.body?.cancel();
    return res.status === 200;
  } catch {
    return false;
  }
});
cleanup = async () => {
  servers.stop();
  await stopCellChildren(parent);
};
const seen: Record<string, unknown> = {};
const statuses = new Set<number>();
/** Cookies a browser would hold, by host. */
const jar = new Map<string, Map<string, string>>();
/** One request to `host` through the parent's real server; every status is noted, a 429 too. */
async function call(
  host: string,
  path: string,
  init: { method?: string; body?: URLSearchParams; accept?: string } = {},
): Promise<Response> {
  const cookies = [...(jar.get(host)?.entries() ?? [])].map(([k, v]) => `${k}=${v}`).join("; ");
  const res = await fetch(`${servers.url}${path}`, {
    method: init.method ?? "GET",
    redirect: "manual",
    signal: AbortSignal.timeout(20_000),
    ...(init.body === undefined ? {} : { body: init.body }),
    headers: {
      host,
      ...(init.method === "POST" ? { origin: `https://${host}` } : {}),
      ...(init.accept === undefined ? {} : { accept: init.accept }),
      ...(cookies === "" ? {} : { cookie: cookies }),
    },
  });
  statuses.add(res.status);
  for (const line of res.headers.getSetCookie()) {
    const [pair = ""] = line.split(";");
    const at = pair.indexOf("=");
    if (at < 1) continue;
    const held = jar.get(host) ?? new Map<string, string>();
    const value = pair.slice(at + 1);
    if (value === "" || /max-age=0/i.test(line)) held.delete(pair.slice(0, at));
    else held.set(pair.slice(0, at), value);
    jar.set(host, held);
  }
  return res;
}
/** Retries `read` each second until it answers, or says which phase ran out and what it saw. */
async function until<T>(
  phase: string,
  ms: number,
  read: () => Promise<T | undefined>,
  last: () => unknown = () => undefined,
): Promise<T> {
  const end = Date.now() + ms;
  for (;;) {
    const value = await read().catch((e: unknown) => {
      seen["lastError"] = String(e);
      return undefined;
    });
    if (value !== undefined) return value;
    if (Date.now() >= end) {
      const saw = JSON.stringify(last() ?? null);
      throw new Error(`the ${phase} phase did not finish in ${ms / 1000}s: ${saw}`);
    }
    await Bun.sleep(1000);
  }
}
function fail(phase: string, what: string): never {
  throw new Error(`the ${phase} phase failed: ${what}`);
}
/** The options of the ask form's target select, as a person sees them. */
const targets = (html: string): string[] =>
  [...html.matchAll(/<option[^>]*value="([^"]*)"/g)].map((m) => m[1]!);
/** Follows redirects as a browser does, from `start`, to the first answer that is not one. */
async function browse(
  start: string,
): Promise<{ status: number; body: string; hops: string[] }> {
  let url = new URL(start);
  const hops: string[] = [];
  for (let i = 0; i < 8; i++) {
    const res = await call(url.host, url.pathname + url.search, { accept: "text/html" });
    hops.push(`${res.status} ${url.host}${url.pathname}`);
    const where = res.headers.get("location");
    if (res.status >= 300 && res.status < 400 && where !== null) {
      await res.body?.cancel();
      url = new URL(where, url);
      continue;
    }
    return { status: res.status, body: await res.text(), hops };
  }
  return { status: 0, body: "", hops };
}
try {
  // Phase 1: a forge by name, from the home's own form; the child comes up in a cell.
  const home = await call(PARENT, "/");
  if (home.status !== 200 || !(await home.text()).includes('id="start-forge"')) {
    fail("up", `the home did not carry the form to start a forge (${home.status})`);
  }
  const started = await call(PARENT, "/forges/start", {
    method: "POST",
    body: new URLSearchParams({ name: NAME }),
  });
  const waiting = await started.text();
  const token = /href="[^"]*\/welcome#([A-Za-z0-9_-]{43})"/.exec(waiting)?.[1] ?? "";
  if (started.status !== 200 || token === "") {
    fail("up", `starting the forge answered ${started.status} with no claim link`);
  }
  let upSeen: unknown = null;
  await until(
    "cell up",
    BUDGET_MS.up,
    async () => {
      const res = await call(PARENT, `/forges/${NAME}/ready`, { accept: "application/json" });
      const body = (await res.json()) as { ready?: unknown };
      upSeen = { status: res.status, ...body, lastError: seen["lastError"] };
      return res.status === 200 && body.ready === true ? true : undefined;
    },
    () => upSeen,
  );
  const held = cellChildren(parent)
    .list()
    .find((c) => c.name === NAME);
  seen["placement"] = held === undefined ? "none" : "cell";
  seen["limits"] = held?.limits ?? null;
  if (held === undefined) fail("up", "the child is not in a cell");
  // Phase 2: the claim link, with the forge's own name as username, a password and no email.
  const form = await call(CHILD, "/claim");
  if (form.status !== 200) fail("claim", `the account form answered ${form.status}`);
  await form.body?.cancel();
  const claimed = await call(CHILD, "/claim", {
    method: "POST",
    body: new URLSearchParams({ token, username: NAME, password: "a-long-stranger-passphrase" }),
  });
  seen["claim"] = claimed.status;
  if (claimed.status !== 303 || jar.get(CHILD) === undefined) {
    fail("claim", `claiming answered ${claimed.status}`);
  }
  await claimed.body?.cancel();
  // Phase 3: the ask, through the form on the forge's first screen, and where it was filed.
  const landing = await call(CHILD, claimed.headers.get("location") ?? "/");
  const first = await landing.text();
  const tell = /<form[^>]*action="\/tell"[^>]*>([\s\S]*?)<\/form>/.exec(first)?.[1];
  if (landing.status !== 200 || tell === undefined) {
    fail("ask", `the forge's first screen answered ${landing.status} with no form to ask through`);
  }
  const fields = new URLSearchParams();
  for (const m of tell!.matchAll(/<input[^>]*type="hidden"[^>]*>/g)) {
    const name = /name="([^"]+)"/.exec(m[0])?.[1];
    if (name) fields.set(name, /value="([^"]*)"/.exec(m[0])?.[1] ?? "");
  }
  const select = /<select[^>]*name="([^"]+)"[^>]*>([\s\S]*?)<\/select>/.exec(tell!) ?? undefined;
  const chosen =
    select === undefined
      ? undefined
      : (/<option[^>]*value="([^"]*)"[^>]*\sselected/.exec(select[2]!)?.[1] ??
        targets(select[2]!)[0]);
  seen["firstTarget"] = chosen ?? fields.get("app") ?? null;
  if (select !== undefined && chosen !== undefined) fields.set(select[1]!, chosen);
  const textName = /<textarea[^>]*name="([^"]+)"/.exec(tell!)?.[1] ?? "";
  if (textName === "") fail("ask", "the form has no place to write the ask");
  fields.set(textName, ASK);
  const before = new Set(targets(first));
  const asked = await call(CHILD, "/tell", { method: "POST", body: fields });
  seen["tell"] = asked.status;
  await asked.body?.cancel();
  if (asked.status >= 400) fail("ask", `the ask answered ${asked.status}`);
  const filed = await until("ask filed on a new app", BUDGET_MS.ask, async () => {
    const res = await call(CHILD, "/");
    const now = targets(await res.text());
    // The forge's own copy may show among the targets once there is an app; never the
    // ask's home.
    const app = /^[a-z0-9][a-z0-9._-]*\/(?!forge$)[a-z0-9][a-z0-9._-]*$/;
    return now.find((t) => app.test(t) && !before.has(t));
  });
  seen["firstFiledOn"] = filed;
  const slug = filed.split("/")[1]!;
  const onForge = await call(CHILD, `/api/v1/repos/${NAME}/forge/issues?state=all`, {
    accept: "application/json",
  });
  const forgeBody = onForge.status === 200 ? ((await onForge.json()) as { items?: unknown[] }) : {};
  const forgeItems = forgeBody.items?.length ?? 0;
  await onForge.body?.cancel();
  seen["forgeIssues"] = onForge.status === 200 ? forgeItems : null;
  const onApp = await call(CHILD, `/api/v1/repos/${filed}/issues?state=all`, {
    accept: "application/json",
  });
  seen["appIssues"] =
    onApp.status === 200
      ? (((await onApp.json()) as { items?: unknown[] }).items?.length ?? 0)
      : onApp.status;
  // Phase 4: the first pull lands by itself, on green.
  let pulls: unknown = null;
  const landed = await until(
    "first pull",
    BUDGET_MS.pull,
    async () => {
      const res = await call(CHILD, `/api/v1/repos/${filed}/pulls?state=all`, {
        accept: "application/json",
      });
      if (res.status !== 200) return undefined;
      type Pull = { number: number; state: string; mergedAt?: unknown; check?: string };
      const items = ((await res.json()) as { items?: Pull[] }).items ?? [];
      pulls = items.map((p) => ({
        number: p.number,
        state: p.state,
        merged: p.mergedAt != null,
        check: p.check,
      }));
      return items.find((p) => p.mergedAt != null);
    },
    () => pulls,
  );
  seen["firstPull"] = {
    number: landed.number,
    state: landed.state,
    merged: true,
    check: landed.check,
  };
  // Phase 5: the app's own address serves the page to its owner, as a browser reaches it.
  const host = `${slug}--${CHILD}`;
  const day = () => new Date().toISOString().slice(0, 10);
  const from = day();
  let page: { status: number; body: string; hops: string[] } | undefined;
  await until(
    "app serving",
    BUDGET_MS.serves,
    async () => {
      page = await browse(`https://${host}/hello`);
      const hello = /^Hello! Today is \d{4}-\d{2}-\d{2}\.$/;
      return page.status === 200 && hello.test(page.body) ? page : undefined;
    },
    () => ({ host, page }),
  );
  seen["firstServes"] = {
    host,
    status: page!.status,
    body: page!.body.slice(0, 80),
    hops: page!.hops,
  };
  const shown = /\d{4}-\d{2}-\d{2}/.exec(page!.body)?.[0] ?? "";
  seen["dateShown"] = shown === from || shown === day();
  seen["stop"] = lastStop(parent, held!.bodyId)?.why ?? null;
} catch (e) {
  seen["error"] = e instanceof Error ? e.message : String(e);
} finally {
  seen["statuses"] = [...statuses].sort((a, b) => a - b);
  console.log(JSON.stringify(seen));
  await cleanup();
  parent.close();
  rmSync(dataDir, { recursive: true, force: true });
}
@@ -0,0 +1,78 @@
/**
 * What the stranger's walk (`real-stranger-e2e.ts`) puts in the model's place
 * (#12077), run by `bun` in the clone the crew gives it: one fixed edit, the
 * app's own tests, and only then a commit and a push. Given a prompt that asks
 * for a page that says hello and shows today's date, it adds a `/hello` route
 * to the app the template made, and a test of it. Any other prompt, or an app
 * that already has the route (a reviewer's clone, a second run), changes
 * nothing and says so. Argument: none; the prompt is read from stdin.
 */
import { readFileSync, writeFileSync } from "node:fs";
import { ANCHOR } from "./build-standin.ts";
/** What `/hello` says, before the date. */
export const HELLO = "Hello! Today is ";
const today = 'new Date().toISOString().slice(0, 10)';
const body = `${JSON.stringify(HELLO)} + ${today} + "."`;
const route = `  { method: "GET", path: "/hello", handle: () => text(${body}) },\n\n`;
const test = `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(), "hello-"));
const server = serve({ port: 0, dataDir, identityKey: "a-key-only-the-forge-holds" });
afterAll(() => {
  server.stop(true);
  rmSync(dataDir, { recursive: true, force: true });
});
test("/hello says hello and today's date", async () => {
  const res = await fetch(\`http://\${server.hostname}:\${server.port}/hello\`);
  expect(res.status).toBe(200);
  expect(await res.text()).toMatch(/^Hello! Today is \\d{4}-\\d{2}-\\d{2}\\.$/);
});
`;
const finish = (ok: boolean, result: string): void =>
  console.log(
    JSON.stringify({
      type: "result",
      subtype: ok ? "success" : "error_during_execution",
      is_error: !ok,
      result,
      total_cost_usd: 0.01,
      num_turns: 1,
      session_id: "stranger-1",
      terminal_reason: ok ? "completed" : "tool_error",
    }),
  );
const git = (...args: string[]): void => {
  const r = Bun.spawnSync(["git", ...args], { stdout: "pipe", stderr: "pipe" });
  if (r.exitCode !== 0) throw new Error(`${args.join(" ")}: ${r.stderr.toString()}`);
};
const prompt = (await Bun.stdin.text()).toLowerCase();
const routes = readFileSync("src/routes.ts", "utf8");
if (!prompt.includes("hello") || !prompt.includes("date") || routes.includes('path: "/hello"')) {
  finish(true, "DECLINED: nothing here I am asked to build.");
  process.exit(0);
}
if (!routes.includes(ANCHOR)) {
  finish(false, "the template's routes have changed shape");
  process.exit(0);
}
writeFileSync("src/routes.ts", routes.replace(ANCHOR, route + ANCHOR));
writeFileSync("src/hello.test.ts", test);
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", "Add a page that says hello and today's date");
git("push", "-q", "-u", "origin", "HEAD");
finish(true, "Added /hello, tests pass.");
@@ -0,0 +1,9 @@
#!/bin/sh
# The model's place in test/spawn-stranger-cell.test.ts (#12077): Claude Code's
# stream-json, with a fixed edit made by stranger-builder.ts. It lives in the
# repository because a cell hides the temp directory a stand-in would be
# written to; the parent hands its path to the child as FORGE_CLAUDE.
if [ "$1" = "--version" ]; then echo "2.1.0 (stranger stand-in)"; exit 0; fi
IFS= read -r first
echo '{"type":"system","subtype":"init","session_id":"stranger-1","model":"stand-in"}'
printf '%s\n' "$first" | bun "$(dirname "$0")/stranger-builder.ts"
@@ -0,0 +1,96 @@
/**
 * The stranger's path in a real cell (#12077, v1 item 1 of #11541): a forge
 * named from the home's form comes up at size small, is claimed through its
 * claim link with the forge's name as username and no email, and is asked for
 * "a page that says hello and shows today's date" through the form on its
 * first screen. The model's call is a fixed edit (`helpers/stranger-claude.sh`);
 * everything else is the forge's own. The ask is filed on a new app and never
 * `<name>/forge`, the first pull lands on green by itself, the app's address
 * serves the page to its owner, and nothing was answered 429 or stopped for
 * memory. The second ask is not walked here.
 *
 * It runs where bwrap and socat are, and says it skipped where they are not
 * (under FORGE_TEST_SOCKET, as `spawn-cell-e2e.test.ts` does). Its budget, as
 * that file's: each phase has its own (cell up 120 s, ask filed 60 s, first
 * pull 120 s, app serving 60 s, set in the helper and named in its failure),
 * and the helper's timeout and this test's are their sum plus slack: 380 s and
 * 400 s. The first real run took 47 s whole. It is a whole forge made and
 * built on, not a unit.
 */
import { test } from "bun:test";
import { join } from "node:path";
import { portInsteadOfSocket } from "../src/git/hook-address.ts";
import { missing } from "./helpers/tools.ts";
import { failedWith } from "./helpers/machine-note.ts";
import { resultLine } from "./helpers/result-line.ts";
const cellTools =
  process.platform === "linux" && !portInsteadOfSocket() && !missing("bwrap", "socat");
const name =
  "a stranger claims a small child, asks for a page, and the first change lands live, " +
  "with no 429 and no memory stop";
test.skipIf(!cellTools)(name, async () => {
  const env: Record<string, string> = {};
  for (const [k, v] of Object.entries(process.env)) {
    if (v !== undefined && k !== "FORGE_CELLS") env[k] = v;
  }
  const walk = join(import.meta.dir, "helpers", "real-stranger-e2e.ts");
  const proc = Bun.spawn([process.execPath, walk], {
    env,
    stdout: "pipe",
    stderr: "pipe",
    timeout: 380_000,
  });
  const [out, errText] = await Promise.all([
    new Response(proc.stdout).text(),
    new Response(proc.stderr).text(),
    proc.exited,
  ]);
  let seen: Record<string, unknown>;
  try {
    seen = JSON.parse(resultLine(out)) as Record<string, unknown>;
  } catch {
    throw new Error(failedWith("the stranger's walk printed no result", out.slice(-800), errText));
  }
  if (typeof seen["skipped"] === "string") {
    const why = seen["skipped"];
    console.log(`SKIPPING: this host refuses a cell inside the test (${why}); no forge started`);
    return;
  }
      
        junipercrew
        1d ago
        outdated
      
      700 s of timeout for a test that has never been seen to pass. Land the first real run in the description before trusting these budgets.
    
  // The walk's own line, in the check's log, whether it passed or not.
  console.log(`stranger walk: ${JSON.stringify(seen)}`);
  const bad = (what: string) => new Error(failedWith(what, seen, errText));
  if (typeof seen["error"] === "string") throw bad(`the stranger's path broke: ${seen["error"]}`);
  const limits = seen["limits"] as { memoryMiB?: number; cpus?: number } | null;
  if (seen["placement"] !== "cell" || limits?.memoryMiB !== 2048 || limits.cpus !== 1) {
    throw bad("the forge did not come up in a small cell");
  }
  if (seen["claim"] !== 303) throw bad("the claim link did not sign the stranger in");
  if (seen["firstTarget"] !== "new-app") {
    throw bad("the first screen's form did not default to a new app");
  }
  const filed = String(seen["firstFiledOn"]);
  if (!/^kid\/(?!forge$)[a-z0-9][a-z0-9._-]*$/.test(filed)) {
    throw bad("the ask was not filed on a new app");
  }
  if (seen["forgeIssues"] !== null && seen["forgeIssues"] !== 0) {
    throw bad("the ask was filed on the forge's own copy");
  }
  if (typeof seen["appIssues"] !== "number" || seen["appIssues"] < 1) {
    throw bad("the new app holds no ask");
  }
  const pull = seen["firstPull"] as { merged?: boolean; check?: string } | undefined;
  if (pull?.merged !== true || pull.check !== "green") {
    throw bad("the first pull did not land on green");
  }
  const serves = seen["firstServes"] as { status?: number } | undefined;
  if (serves?.status !== 200 || seen["dateShown"] !== true) {
    throw bad("the app's address did not serve the page with today's date");
  }
  if ((seen["statuses"] as number[]).includes(429)) {
    throw bad("a request was answered 429, the sentence for a full host");
  }
  if (seen["stop"] !== null) throw bad("the child was stopped for memory");
}, 400_000);
		@@ -181,6 +181,11 @@ export const SLOW: readonly Exclusion[] = [
181	181	reason: "the cold home takes 20 timed requests at ~65 ms, each after a write, until its per-app reads are grouped",
182	182	issue: 12310,
183	183	},
	184	{
	185	file: "test/spawn-stranger-cell.test.ts",
	186	reason: "one test walks a whole forge in a real cell, from claim link to the first pull green; it is not a unit",
	187	issue: 12077,
		junipercrew 1d ago The SLOW entry is acceptable for this walk, but the check must still run it unskipped. Prove that in the description.
	188	},
184	189	];
185	190
186	191	/** Everything wrong with one list of named files, `list` its name in the lines. */
		@@ -0,0 +1,325 @@
	1	/**
	2	* The stranger's path in a real cell (#12077), as `test/spawn-stranger-cell.test.ts`
	3	* drives it. Its own process, for the reason `real-spawn-e2e.ts` is: a check
	4	* sets `FORGE_CELLS=off` for the forges its tests boot.
	5	*
	6	* A parent with the real server is asked for a forge by name, as the home's
	7	* form does, with no account. The child comes up in a real cell at the plan's
	8	* size; the walk claims it through its claim link with the forge's name as
	9	* username and no email, asks "a page that says hello and shows today's date"
	10	* through the form on its first screen, and follows the ask to a new app, its
	11	* first pull landing on green, and the app's own address serving the page to
	12	* its owner as a browser reaches it. The model is `stranger-claude.sh`, a
	13	* fixed edit; the cell, the crew's clone and push, the pull, its check, the
	14	* merge and the deploy are the forge's own. The second ask is not walked.
	15	*
	16	* Each phase has its own budget and a message that names it. Printed: one JSON
	17	* line of what was seen, with `error` set to the phase that did not finish.
	18	*/
	19	import { mkdtempSync, rmSync } from "node:fs";
	20	import { tmpdir } from "node:os";
	21	import { join } from "node:path";
	22	import { createApp } from "../../src/app.ts";
	23	import { cellsAvailable } from "../../src/apps/contain.ts";
	24	import { serve } from "../../src/http/server.ts";
	25	import { useProbe } from "../../src/spawn/door.ts";
	26	import { cellChildren, stopCellChildren } from "../../src/spawn/cell.ts";
	27	import { lastStop } from "../../src/spawn/stopped.ts";
	28	import "./roomy-disk.ts";
	29
	30	// One budget per phase, each counted from when it began: a slow phase never
	31	// shortens the next, and a run that ran out says which one did. The test's
	32	// timeout is their sum plus slack (see test/spawn-stranger-cell.test.ts).
	33	const BUDGET_MS = { up: 120_000, ask: 60_000, pull: 120_000, serves: 60_000 } as const;
	34	const NAME = "kid";
	35	const ASK = "a page that says hello and shows today's date";
	36	const PARENT = "forge.test";
	37	const CHILD = `${NAME}--${PARENT}`;
	38
	39	process.env["FORGE_CLAUDE"] = join(import.meta.dir, "stranger-claude.sh");
	40
	41	let cleanup: () => Promise<void> = async () => {};
	42	process.on("SIGTERM", () => void cleanup().finally(() => process.exit(143)));
	43
	44	const cells = await cellsAvailable();
	45	if (!cells.ok) {
	46	console.log(JSON.stringify({ skipped: cells.reason }));
	47	process.exit(0);
	48	}
	49
	50	const dataDir = mkdtempSync(join(tmpdir(), "real-stranger-"));
	51	const parent = createApp({ dataDir, url: `https://${PARENT}`, port: 0 });
	52	const servers = await serve(parent);
	53	// The child's public address (https://kid--forge.test) is not a real host here: ask it
	54	// the way every other request in this walk does, through the parent's own server with
	55	// the child's Host, on /claim (the parent answers /healthz itself, whatever the host).
	56	useProbe(parent, () => async () => {
	57	try {
	58	const res = await fetch(`${servers.url}/claim`, {
	59	headers: { host: CHILD },
	60	redirect: "manual",
	61	signal: AbortSignal.timeout(700),
	62	});
	63	await res.body?.cancel();
	64	return res.status === 200;
	65	} catch {
	66	return false;
	67	}
	68	});
	69	cleanup = async () => {
	70	servers.stop();
	71	await stopCellChildren(parent);
	72	};
	73
	74	const seen: Record<string, unknown> = {};
	75	const statuses = new Set<number>();
	76	/** Cookies a browser would hold, by host. */
	77	const jar = new Map<string, Map<string, string>>();
	78
	79	/** One request to `host` through the parent's real server; every status is noted, a 429 too. */
	80	async function call(
	81	host: string,
	82	path: string,
	83	init: { method?: string; body?: URLSearchParams; accept?: string } = {},
	84	): Promise<Response> {
	85	const cookies = [...(jar.get(host)?.entries() ?? [])].map(([k, v]) => `${k}=${v}`).join("; ");
	86	const res = await fetch(`${servers.url}${path}`, {
	87	method: init.method ?? "GET",
	88	redirect: "manual",
	89	signal: AbortSignal.timeout(20_000),
	90	...(init.body === undefined ? {} : { body: init.body }),
	91	headers: {
	92	host,
	93	...(init.method === "POST" ? { origin: `https://${host}` } : {}),
	94	...(init.accept === undefined ? {} : { accept: init.accept }),
	95	...(cookies === "" ? {} : { cookie: cookies }),
	96	},
	97	});
	98	statuses.add(res.status);
	99	for (const line of res.headers.getSetCookie()) {
	100	const [pair = ""] = line.split(";");
	101	const at = pair.indexOf("=");
	102	if (at < 1) continue;
	103	const held = jar.get(host) ?? new Map<string, string>();
	104	const value = pair.slice(at + 1);
	105	if (value === "" \|\| /max-age=0/i.test(line)) held.delete(pair.slice(0, at));
	106	else held.set(pair.slice(0, at), value);
	107	jar.set(host, held);
	108	}
	109	return res;
	110	}
	111
	112	/** Retries `read` each second until it answers, or says which phase ran out and what it saw. */
	113	async function until<T>(
	114	phase: string,
	115	ms: number,
	116	read: () => Promise<T \| undefined>,
	117	last: () => unknown = () => undefined,
	118	): Promise<T> {
	119	const end = Date.now() + ms;
	120	for (;;) {
	121	const value = await read().catch((e: unknown) => {
	122	seen["lastError"] = String(e);
	123	return undefined;
	124	});
	125	if (value !== undefined) return value;
	126	if (Date.now() >= end) {
	127	const saw = JSON.stringify(last() ?? null);
	128	throw new Error(`the ${phase} phase did not finish in ${ms / 1000}s: ${saw}`);
	129	}
	130	await Bun.sleep(1000);
	131	}
	132	}
	133
	134	function fail(phase: string, what: string): never {
	135	throw new Error(`the ${phase} phase failed: ${what}`);
	136	}
	137
	138	/** The options of the ask form's target select, as a person sees them. */
	139	const targets = (html: string): string[] =>
	140	[...html.matchAll(/<option[^>]value="([^"])"/g)].map((m) => m[1]!);
	141
	142	/** Follows redirects as a browser does, from `start`, to the first answer that is not one. */
	143	async function browse(
	144	start: string,
	145	): Promise<{ status: number; body: string; hops: string[] }> {
	146	let url = new URL(start);
	147	const hops: string[] = [];
	148	for (let i = 0; i < 8; i++) {
	149	const res = await call(url.host, url.pathname + url.search, { accept: "text/html" });
	150	hops.push(`${res.status} ${url.host}${url.pathname}`);
	151	const where = res.headers.get("location");
	152	if (res.status >= 300 && res.status < 400 && where !== null) {
	153	await res.body?.cancel();
	154	url = new URL(where, url);
	155	continue;
	156	}
	157	return { status: res.status, body: await res.text(), hops };
	158	}
	159	return { status: 0, body: "", hops };
	160	}
	161
	162	try {
	163	// Phase 1: a forge by name, from the home's own form; the child comes up in a cell.
	164	const home = await call(PARENT, "/");
	165	if (home.status !== 200 \|\| !(await home.text()).includes('id="start-forge"')) {
	166	fail("up", `the home did not carry the form to start a forge (${home.status})`);
	167	}
	168	const started = await call(PARENT, "/forges/start", {
	169	method: "POST",
	170	body: new URLSearchParams({ name: NAME }),
	171	});
	172	const waiting = await started.text();
	173	const token = /href="[^"]*\/welcome#([A-Za-z0-9_-]{43})"/.exec(waiting)?.[1] ?? "";
	174	if (started.status !== 200 \|\| token === "") {
	175	fail("up", `starting the forge answered ${started.status} with no claim link`);
	176	}
	177	let upSeen: unknown = null;
	178	await until(
	179	"cell up",
	180	BUDGET_MS.up,
	181	async () => {
	182	const res = await call(PARENT, `/forges/${NAME}/ready`, { accept: "application/json" });
	183	const body = (await res.json()) as { ready?: unknown };
	184	upSeen = { status: res.status, ...body, lastError: seen["lastError"] };
	185	return res.status === 200 && body.ready === true ? true : undefined;
	186	},
	187	() => upSeen,
	188	);
	189	const held = cellChildren(parent)
	190	.list()
	191	.find((c) => c.name === NAME);
	192	seen["placement"] = held === undefined ? "none" : "cell";
	193	seen["limits"] = held?.limits ?? null;
	194	if (held === undefined) fail("up", "the child is not in a cell");
	195
	196	// Phase 2: the claim link, with the forge's own name as username, a password and no email.
	197	const form = await call(CHILD, "/claim");
	198	if (form.status !== 200) fail("claim", `the account form answered ${form.status}`);
	199	await form.body?.cancel();
	200	const claimed = await call(CHILD, "/claim", {
	201	method: "POST",
	202	body: new URLSearchParams({ token, username: NAME, password: "a-long-stranger-passphrase" }),
	203	});
	204	seen["claim"] = claimed.status;
	205	if (claimed.status !== 303 \|\| jar.get(CHILD) === undefined) {
	206	fail("claim", `claiming answered ${claimed.status}`);
	207	}
	208	await claimed.body?.cancel();
	209
	210	// Phase 3: the ask, through the form on the forge's first screen, and where it was filed.
	211	const landing = await call(CHILD, claimed.headers.get("location") ?? "/");
	212	const first = await landing.text();
	213	const tell = /<form[^>]action="\/tell"[^>]>([\s\S]*?)<\/form>/.exec(first)?.[1];
	214	if (landing.status !== 200 \|\| tell === undefined) {
	215	fail("ask", `the forge's first screen answered ${landing.status} with no form to ask through`);
	216	}
	217	const fields = new URLSearchParams();
	218	for (const m of tell!.matchAll(/<input[^>]type="hidden"[^>]>/g)) {
	219	const name = /name="([^"]+)"/.exec(m[0])?.[1];
	220	if (name) fields.set(name, /value="([^"]*)"/.exec(m[0])?.[1] ?? "");
	221	}
	222	const select = /<select[^>]name="([^"]+)"[^>]>([\s\S]*?)<\/select>/.exec(tell!) ?? undefined;
	223	const chosen =
	224	select === undefined
	225	? undefined
	226	: (/<option[^>]value="([^"])"[^>]*\sselected/.exec(select[2]!)?.[1] ??
	227	targets(select[2]!)[0]);
	228	seen["firstTarget"] = chosen ?? fields.get("app") ?? null;
	229	if (select !== undefined && chosen !== undefined) fields.set(select[1]!, chosen);
	230	const textName = /<textarea[^>]*name="([^"]+)"/.exec(tell!)?.[1] ?? "";
	231	if (textName === "") fail("ask", "the form has no place to write the ask");
	232	fields.set(textName, ASK);
	233	const before = new Set(targets(first));
	234	const asked = await call(CHILD, "/tell", { method: "POST", body: fields });
	235	seen["tell"] = asked.status;
	236	await asked.body?.cancel();
	237	if (asked.status >= 400) fail("ask", `the ask answered ${asked.status}`);
	238	const filed = await until("ask filed on a new app", BUDGET_MS.ask, async () => {
	239	const res = await call(CHILD, "/");
	240	const now = targets(await res.text());
	241	// The forge's own copy may show among the targets once there is an app; never the
	242	// ask's home.
	243	const app = /^[a-z0-9][a-z0-9._-]\/(?!forge$)[a-z0-9][a-z0-9._-]$/;
	244	return now.find((t) => app.test(t) && !before.has(t));
	245	});
	246	seen["firstFiledOn"] = filed;
	247	const slug = filed.split("/")[1]!;
	248	const onForge = await call(CHILD, `/api/v1/repos/${NAME}/forge/issues?state=all`, {
	249	accept: "application/json",
	250	});
	251	const forgeBody = onForge.status === 200 ? ((await onForge.json()) as { items?: unknown[] }) : {};
	252	const forgeItems = forgeBody.items?.length ?? 0;
	253	await onForge.body?.cancel();
	254	seen["forgeIssues"] = onForge.status === 200 ? forgeItems : null;
	255	const onApp = await call(CHILD, `/api/v1/repos/${filed}/issues?state=all`, {
	256	accept: "application/json",
	257	});
	258	seen["appIssues"] =
	259	onApp.status === 200
	260	? (((await onApp.json()) as { items?: unknown[] }).items?.length ?? 0)
	261	: onApp.status;
	262
	263	// Phase 4: the first pull lands by itself, on green.
	264	let pulls: unknown = null;
	265	const landed = await until(
	266	"first pull",
	267	BUDGET_MS.pull,
	268	async () => {
	269	const res = await call(CHILD, `/api/v1/repos/${filed}/pulls?state=all`, {
	270	accept: "application/json",
	271	});
	272	if (res.status !== 200) return undefined;
	273	type Pull = { number: number; state: string; mergedAt?: unknown; check?: string };
	274	const items = ((await res.json()) as { items?: Pull[] }).items ?? [];
	275	pulls = items.map((p) => ({
	276	number: p.number,
	277	state: p.state,
	278	merged: p.mergedAt != null,
	279	check: p.check,
	280	}));
	281	return items.find((p) => p.mergedAt != null);
	282	},
	283	() => pulls,
	284	);
	285	seen["firstPull"] = {
	286	number: landed.number,
	287	state: landed.state,
	288	merged: true,
	289	check: landed.check,
	290	};
	291
	292	// Phase 5: the app's own address serves the page to its owner, as a browser reaches it.
	293	const host = `${slug}--${CHILD}`;
	294	const day = () => new Date().toISOString().slice(0, 10);
	295	const from = day();
	296	let page: { status: number; body: string; hops: string[] } \| undefined;
	297	await until(
	298	"app serving",
	299	BUDGET_MS.serves,
	300	async () => {
	301	page = await browse(`https://${host}/hello`);
	302	const hello = /^Hello! Today is \d{4}-\d{2}-\d{2}\.$/;
	303	return page.status === 200 && hello.test(page.body) ? page : undefined;
	304	},
	305	() => ({ host, page }),
	306	);
	307	seen["firstServes"] = {
	308	host,
	309	status: page!.status,
	310	body: page!.body.slice(0, 80),
	311	hops: page!.hops,
	312	};
	313	const shown = /\d{4}-\d{2}-\d{2}/.exec(page!.body)?.[0] ?? "";
	314	seen["dateShown"] = shown === from \|\| shown === day();
	315
	316	seen["stop"] = lastStop(parent, held!.bodyId)?.why ?? null;
	317	} catch (e) {
	318	seen["error"] = e instanceof Error ? e.message : String(e);
	319	} finally {
	320	seen["statuses"] = [...statuses].sort((a, b) => a - b);
	321	console.log(JSON.stringify(seen));
	322	await cleanup();
	323	parent.close();
	324	rmSync(dataDir, { recursive: true, force: true });
	325	}
		@@ -0,0 +1,78 @@
	1	/**
	2	* What the stranger's walk (`real-stranger-e2e.ts`) puts in the model's place
	3	* (#12077), run by `bun` in the clone the crew gives it: one fixed edit, the
	4	* app's own tests, and only then a commit and a push. Given a prompt that asks
	5	* for a page that says hello and shows today's date, it adds a `/hello` route
	6	* to the app the template made, and a test of it. Any other prompt, or an app
	7	* that already has the route (a reviewer's clone, a second run), changes
	8	* nothing and says so. Argument: none; the prompt is read from stdin.
	9	*/
	10	import { readFileSync, writeFileSync } from "node:fs";
	11	import { ANCHOR } from "./build-standin.ts";
	12
	13	/** What `/hello` says, before the date. */
	14	export const HELLO = "Hello! Today is ";
	15
	16	const today = 'new Date().toISOString().slice(0, 10)';
	17	const body = `${JSON.stringify(HELLO)} + ${today} + "."`;
	18	const route = ` { method: "GET", path: "/hello", handle: () => text(${body}) },\n\n`;
	19
	20	const test = `import { afterAll, expect, test } from "bun:test";
	21	import { mkdtempSync, rmSync } from "node:fs";
	22	import { tmpdir } from "node:os";
	23	import { join } from "node:path";
	24	import { serve } from "./server.ts";
	25
	26	const dataDir = mkdtempSync(join(tmpdir(), "hello-"));
	27	const server = serve({ port: 0, dataDir, identityKey: "a-key-only-the-forge-holds" });
	28	afterAll(() => {
	29	server.stop(true);
	30	rmSync(dataDir, { recursive: true, force: true });
	31	});
	32
	33	test("/hello says hello and today's date", async () => {
	34	const res = await fetch(\`http://\${server.hostname}:\${server.port}/hello\`);
	35	expect(res.status).toBe(200);
	36	expect(await res.text()).toMatch(/^Hello! Today is \\d{4}-\\d{2}-\\d{2}\\.$/);
	37	});
	38	`;
	39
	40	const finish = (ok: boolean, result: string): void =>
	41	console.log(
	42	JSON.stringify({
	43	type: "result",
	44	subtype: ok ? "success" : "error_during_execution",
	45	is_error: !ok,
	46	result,
	47	total_cost_usd: 0.01,
	48	num_turns: 1,
	49	session_id: "stranger-1",
	50	terminal_reason: ok ? "completed" : "tool_error",
	51	}),
	52	);
	53	const git = (...args: string[]): void => {
	54	const r = Bun.spawnSync(["git", ...args], { stdout: "pipe", stderr: "pipe" });
	55	if (r.exitCode !== 0) throw new Error(`${args.join(" ")}: ${r.stderr.toString()}`);
	56	};
	57
	58	const prompt = (await Bun.stdin.text()).toLowerCase();
	59	const routes = readFileSync("src/routes.ts", "utf8");
	60	if (!prompt.includes("hello") \|\| !prompt.includes("date") \|\| routes.includes('path: "/hello"')) {
	61	finish(true, "DECLINED: nothing here I am asked to build.");
	62	process.exit(0);
	63	}
	64	if (!routes.includes(ANCHOR)) {
	65	finish(false, "the template's routes have changed shape");
	66	process.exit(0);
	67	}
	68	writeFileSync("src/routes.ts", routes.replace(ANCHOR, route + ANCHOR));
	69	writeFileSync("src/hello.test.ts", test);
	70	const tested = Bun.spawnSync(["bun", "test"], { stdout: "pipe", stderr: "pipe" });
	71	if (tested.exitCode !== 0) {
	72	finish(false, `its tests failed: ${tested.stderr.toString().slice(-300)}`);
	73	process.exit(0);
	74	}
	75	git("add", "-A");
	76	git("commit", "-qm", "Add a page that says hello and today's date");
	77	git("push", "-q", "-u", "origin", "HEAD");
	78	finish(true, "Added /hello, tests pass.");
		@@ -0,0 +1,9 @@
	1	#!/bin/sh
	2	# The model's place in test/spawn-stranger-cell.test.ts (#12077): Claude Code's
	3	# stream-json, with a fixed edit made by stranger-builder.ts. It lives in the
	4	# repository because a cell hides the temp directory a stand-in would be
	5	# written to; the parent hands its path to the child as FORGE_CLAUDE.
	6	if [ "$1" = "--version" ]; then echo "2.1.0 (stranger stand-in)"; exit 0; fi
	7	IFS= read -r first
	8	echo '{"type":"system","subtype":"init","session_id":"stranger-1","model":"stand-in"}'
	9	printf '%s\n' "$first" \| bun "$(dirname "$0")/stranger-builder.ts"
		@@ -0,0 +1,96 @@
	1	/**
	2	* The stranger's path in a real cell (#12077, v1 item 1 of #11541): a forge
	3	* named from the home's form comes up at size small, is claimed through its
	4	* claim link with the forge's name as username and no email, and is asked for
	5	* "a page that says hello and shows today's date" through the form on its
	6	* first screen. The model's call is a fixed edit (`helpers/stranger-claude.sh`);
	7	* everything else is the forge's own. The ask is filed on a new app and never
	8	* `<name>/forge`, the first pull lands on green by itself, the app's address
	9	* serves the page to its owner, and nothing was answered 429 or stopped for
	10	* memory. The second ask is not walked here.
	11	*
	12	* It runs where bwrap and socat are, and says it skipped where they are not
	13	* (under FORGE_TEST_SOCKET, as `spawn-cell-e2e.test.ts` does). Its budget, as
	14	* that file's: each phase has its own (cell up 120 s, ask filed 60 s, first
	15	* pull 120 s, app serving 60 s, set in the helper and named in its failure),
	16	* and the helper's timeout and this test's are their sum plus slack: 380 s and
	17	* 400 s. The first real run took 47 s whole. It is a whole forge made and
	18	* built on, not a unit.
	19	*/
	20	import { test } from "bun:test";
	21	import { join } from "node:path";
	22	import { portInsteadOfSocket } from "../src/git/hook-address.ts";
	23	import { missing } from "./helpers/tools.ts";
	24	import { failedWith } from "./helpers/machine-note.ts";
	25	import { resultLine } from "./helpers/result-line.ts";
	26
	27	const cellTools =
	28	process.platform === "linux" && !portInsteadOfSocket() && !missing("bwrap", "socat");
	29
	30	const name =
	31	"a stranger claims a small child, asks for a page, and the first change lands live, " +
	32	"with no 429 and no memory stop";
	33
	34	test.skipIf(!cellTools)(name, async () => {
	35	const env: Record<string, string> = {};
	36	for (const [k, v] of Object.entries(process.env)) {
	37	if (v !== undefined && k !== "FORGE_CELLS") env[k] = v;
	38	}
	39	const walk = join(import.meta.dir, "helpers", "real-stranger-e2e.ts");
	40	const proc = Bun.spawn([process.execPath, walk], {
	41	env,
	42	stdout: "pipe",
	43	stderr: "pipe",
	44	timeout: 380_000,
	45	});
	46	const [out, errText] = await Promise.all([
	47	new Response(proc.stdout).text(),
	48	new Response(proc.stderr).text(),
	49	proc.exited,
	50	]);
	51	let seen: Record<string, unknown>;
	52	try {
	53	seen = JSON.parse(resultLine(out)) as Record<string, unknown>;
	54	} catch {
	55	throw new Error(failedWith("the stranger's walk printed no result", out.slice(-800), errText));
	56	}
	57	if (typeof seen["skipped"] === "string") {
	58	const why = seen["skipped"];
	59	console.log(`SKIPPING: this host refuses a cell inside the test (${why}); no forge started`);
	60	return;
	61	}
		junipercrew 1d ago outdated 700 s of timeout for a test that has never been seen to pass. Land the first real run in the description before trusting these budgets.
	62	// The walk's own line, in the check's log, whether it passed or not.
	63	console.log(`stranger walk: ${JSON.stringify(seen)}`);
	64	const bad = (what: string) => new Error(failedWith(what, seen, errText));
	65	if (typeof seen["error"] === "string") throw bad(`the stranger's path broke: ${seen["error"]}`);
	66	const limits = seen["limits"] as { memoryMiB?: number; cpus?: number } \| null;
	67	if (seen["placement"] !== "cell" \|\| limits?.memoryMiB !== 2048 \|\| limits.cpus !== 1) {
	68	throw bad("the forge did not come up in a small cell");
	69	}
	70	if (seen["claim"] !== 303) throw bad("the claim link did not sign the stranger in");
	71	if (seen["firstTarget"] !== "new-app") {
	72	throw bad("the first screen's form did not default to a new app");
	73	}
	74	const filed = String(seen["firstFiledOn"]);
	75	if (!/^kid\/(?!forge$)[a-z0-9][a-z0-9._-]*$/.test(filed)) {
	76	throw bad("the ask was not filed on a new app");
	77	}
	78	if (seen["forgeIssues"] !== null && seen["forgeIssues"] !== 0) {
	79	throw bad("the ask was filed on the forge's own copy");
	80	}
	81	if (typeof seen["appIssues"] !== "number" \|\| seen["appIssues"] < 1) {
	82	throw bad("the new app holds no ask");
	83	}
	84	const pull = seen["firstPull"] as { merged?: boolean; check?: string } \| undefined;
	85	if (pull?.merged !== true \|\| pull.check !== "green") {
	86	throw bad("the first pull did not land on green");
	87	}
	88	const serves = seen["firstServes"] as { status?: number } \| undefined;
	89	if (serves?.status !== 200 \|\| seen["dateShown"] !== true) {
	90	throw bad("the app's address did not serve the page with today's date");
	91	}
	92	if ((seen["statuses"] as number[]).includes(429)) {
	93	throw bad("a request was answered 429, the sentence for a full host");
	94	}
	95	if (seen["stop"] !== null) throw bad("the child was stopped for memory");
	96	}, 400_000);