Shell-only agent

/** * Shell-only agent: an agent run in its own session with no browser, only a shell. It generates a CSV by a rule, computes * statistics from it with Python and writes a report; then the example reads the files back (`files.readText`) and * recomputes every number itself, so the result does not rest on the agent's word. * *   npx tsx node/index.ts            (BOXLINE_API_KEY; a plan with shell sessions and agent runs) * * The data is made up by a rule (no site is visited). Writes output/sales.csv, output/stats.json, output/report.md and * output/result.json. */import { mkdirSync, writeFileSync } from "node:fs";import { join } from "node:path";import { Boxline } from "@boxline/sdk";const out = process.env.OUTPUT_DIR ?? "output";const bx = new Boxline();// The rule the agent follows, and the example checks against.const ROWS = 240;const REGIONS = ["north", "south", "east", "west"];const rule = (i: number) => ({ id: i, region: REGIONS[i % 4]!, units: ((i * 37) % 23) + 1, price: 4.5 + (i % 5) * 1.25 });const round2 = (x: number) => Math.round(x * 100) / 100;/** The numbers the agent should find, from the rule alone. */function expectedStats() {  const rows = Array.from({ length: ROWS }, (_, k) => rule(k + 1));  const byRegion = Object.fromEntries(REGIONS.map((r) => [r, round2(rows.filter((x) => x.region === r).reduce((s, x) => s + x.units * x.price, 0))]));  const units = rows.map((x) => x.units).sort((a, b) => a - b);  const mean = units.reduce((s, u) => s + u, 0) / units.length;  const mid = units.length / 2;  return {    total_units: units.reduce((s, u) => s + u, 0),    total_revenue: round2(rows.reduce((s, x) => s + x.units * x.price, 0)),    revenue_by_region: byRegion,    top_region: Object.entries(byRegion).sort((a, b) => b[1] - a[1])[0]![0],    units_mean: round2(mean),    units_median: units.length % 2 ? units[Math.floor(mid)]! : (units[mid - 1]! + units[mid]!) / 2,    units_stdev: round2(Math.sqrt(units.reduce((s, u) => s + (u - mean) ** 2, 0) / units.length)),  };}const task =  `Work in /workspace with Python 3 (standard library only). Do these three things, in this order:\n` +  `1. Write sales.csv with the header "id,region,units,unit_price" and one row for each i from 1 to ${ROWS}: ` +  `id = i, region = ["north", "south", "east", "west"][i % 4], units = (i * 37) % 23 + 1, unit_price = 4.50 + (i % 5) * 1.25 (two decimals).\n` +  `2. Read sales.csv back with a Python script and write stats.json with these keys: total_units (integer), total_revenue ` +  `(the sum of units * unit_price), revenue_by_region (an object with the four regions), top_region (the region with the highest revenue), ` +  `units_mean, units_median and units_stdev (the population standard deviation, statistics.pstdev). Round every decimal to 2 places.\n` +  `3. Write report.md: a heading, a table of revenue by region, and one sentence that names the top region and the total revenue ` +  `with two decimals. Take every number from stats.json.\n` +  `When all three files exist, reply with one sentence.`;const started = await bx.agent.run({  task,  // The run's own session: no browser, only a shell. keepSession leaves it running so the files can be read.  session: { browser: false, shell: true, timeout: 900, idleTimeout: 300, userMetadata: { example: "shell-only-agent" } },  keepSession: true,  maxSteps: 20,});console.log(`Session: ${started.sessionId}`);console.log(`Agent run ${started.id} (${started.model}) is working in a shell-only session…`);try {  const run = await bx.agent.wait(started.id);  if (run.status !== "completed") throw new Error(`the run ${run.status}: ${run.error ?? "no result"}`);  const toolSteps = run.steps.filter((s) => s.type === "tool");  console.log(`The agent finished after ${toolSteps.length} tool steps: ${String(run.result).slice(0, 200)}`);  // The files the agent wrote, read through the API.  const files = bx.sessions.files;  const csv = await files.readText(started.sessionId, "sales.csv");  const reported = JSON.parse(await files.readText(started.sessionId, "stats.json"));  const report = await files.readText(started.sessionId, "report.md");  const lines = csv.trim().split(/\r?\n/); // Python's csv module ends lines with \r\n  const matchesRule =    lines[0] === "id,region,units,unit_price" &&    lines.length === ROWS + 1 &&    lines.slice(1).every((line, k) => {      const [id, region, units, price] = line.split(",");      const want = rule(k + 1);      return Number(id) === want.id && region === want.region && Number(units) === want.units && Math.abs(Number(price) - want.price) < 1e-9;    });  const expected = expectedStats();  console.log(`sales.csv: ${lines.length - 1} rows${matchesRule ? ", as the rule says" : ", NOT as the rule says"}`);  console.log(`The agent's total revenue ${reported.total_revenue}, expected ${expected.total_revenue}; top region ${reported.top_region}, expected ${expected.top_region}`);  // What kind of session the run made.  const session = (await bx.sessions.get(started.sessionId)).data;  console.log(`The run's session: browser ${session.browser}, shell ${session.shell}, liveUrl ${session.liveUrl}`);  mkdirSync(out, { recursive: true });  writeFileSync(join(out, "sales.csv"), csv);  writeFileSync(join(out, "stats.json"), JSON.stringify(reported, null, 2));  writeFileSync(join(out, "report.md"), report);  writeFileSync(    join(out, "result.json"),    JSON.stringify(      {        csv: { rows: lines.length - 1, matchesRule },        expected,        reported,        report,        session: { id: session.id, browser: session.browser, shell: session.shell, liveUrl: session.liveUrl },        toolSteps: toolSteps.map((s) => ({ name: s.name, input: s.input })),        agentRuns: [{ id: run.id }],      },      null,      2,    ),  );} finally {  await bx.sessions.stop(started.sessionId).catch(() => undefined);}

"""Shell-only agent: an agent run in its own session with no browser, only a shell. It generates a CSV by a rule, computesstatistics from it with Python and writes a report; then the example reads the files back (``files.read_text``) andrecomputes every number itself, so the result does not rest on the agent's word.    python python/main.py            (BOXLINE_API_KEY; a plan with shell sessions and agent runs)The data is made up by a rule (no site is visited). Writes output/sales.csv, output/stats.json, output/report.md andoutput/result.json."""import jsonimport osimport statisticsfrom pathlib import Pathfrom boxline import Boxlineout = Path(os.environ.get("OUTPUT_DIR", "output"))bx = Boxline()# The rule the agent follows, and the example checks against.ROWS = 240REGIONS = ["north", "south", "east", "west"]def rule(i: int) -> dict:    return {"id": i, "region": REGIONS[i % 4], "units": (i * 37) % 23 + 1, "price": 4.5 + (i % 5) * 1.25}def expected_stats() -> dict:    """The numbers the agent should find, from the rule alone."""    rows = [rule(i) for i in range(1, ROWS + 1)]    by_region = {r: round(sum(x["units"] * x["price"] for x in rows if x["region"] == r), 2) for r in REGIONS}    units = [x["units"] for x in rows]    return {        "total_units": sum(units),        "total_revenue": round(sum(x["units"] * x["price"] for x in rows), 2),        "revenue_by_region": by_region,        "top_region": max(by_region, key=by_region.get),        "units_mean": round(statistics.mean(units), 2),        "units_median": statistics.median(units),        "units_stdev": round(statistics.pstdev(units), 2),    }TASK = (    "Work in /workspace with Python 3 (standard library only). Do these three things, in this order:\n"    f'1. Write sales.csv with the header "id,region,units,unit_price" and one row for each i from 1 to {ROWS}: '    'id = i, region = ["north", "south", "east", "west"][i % 4], units = (i * 37) % 23 + 1, unit_price = 4.50 + (i % 5) * 1.25 (two decimals).\n'    "2. Read sales.csv back with a Python script and write stats.json with these keys: total_units (integer), total_revenue "    "(the sum of units * unit_price), revenue_by_region (an object with the four regions), top_region (the region with the highest revenue), "    "units_mean, units_median and units_stdev (the population standard deviation, statistics.pstdev). Round every decimal to 2 places.\n"    "3. Write report.md: a heading, a table of revenue by region, and one sentence that names the top region and the total revenue "    "with two decimals. Take every number from stats.json.\n"    "When all three files exist, reply with one sentence.")started = bx.agent.run(    TASK,    # The run's own session: no browser, only a shell. keep_session leaves it running so the files can be read.    session={"browser": False, "shell": True, "timeout": 900, "idleTimeout": 300, "userMetadata": {"example": "shell-only-agent"}},    keep_session=True,    max_steps=20,)session_id = started["sessionId"]print(f"Session: {session_id}", flush=True)print(f"Agent run {started['id']} ({started['model']}) is working in a shell-only session…", flush=True)try:    run = bx.agent.wait(started["id"])    if run["status"] != "completed":        raise SystemExit(f"the run {run['status']}: {run.get('error') or 'no result'}")    tool_steps = [s for s in run["steps"] if s["type"] == "tool"]    print(f"The agent finished after {len(tool_steps)} tool steps: {str(run['result'])[:200]}", flush=True)    # The files the agent wrote, read through the API.    csv = bx.sessions.files.read_text(session_id, "sales.csv")    reported = json.loads(bx.sessions.files.read_text(session_id, "stats.json"))    report = bx.sessions.files.read_text(session_id, "report.md")    lines = csv.strip().splitlines()  # also \r\n, which Python's csv module writes    matches_rule = lines[0] == "id,region,units,unit_price" and len(lines) == ROWS + 1    if matches_rule:        for k, line in enumerate(lines[1:]):            id_, region, units, price = line.split(",")            want = rule(k + 1)            if not (int(id_) == want["id"] and region == want["region"] and int(units) == want["units"] and abs(float(price) - want["price"]) < 1e-9):                matches_rule = False                break    expected = expected_stats()    print(f"sales.csv: {len(lines) - 1} rows" + (", as the rule says" if matches_rule else ", NOT as the rule says"))    print(f"The agent's total revenue {reported.get('total_revenue')}, expected {expected['total_revenue']}; top region {reported.get('top_region')}, expected {expected['top_region']}")    # What kind of session the run made.    session = bx.sessions.get(session_id).data    print(f"The run's session: browser {session['browser']}, shell {session['shell']}, liveUrl {session['liveUrl']}")    out.mkdir(parents=True, exist_ok=True)    (out / "sales.csv").write_text(csv)    (out / "stats.json").write_text(json.dumps(reported, indent=2))    (out / "report.md").write_text(report)    result = {        "csv": {"rows": len(lines) - 1, "matchesRule": matches_rule},        "expected": expected,        "reported": reported,        "report": report,        "session": {"id": session["id"], "browser": session["browser"], "shell": session["shell"], "liveUrl": session["liveUrl"]},        "toolSteps": [{"name": s.get("name"), "input": s.get("input")} for s in tool_steps],        "agentRuns": [{"id": run["id"]}],    }    (out / "result.json").write_text(json.dumps(result, indent=2, ensure_ascii=False))finally:    try:        bx.sessions.stop(session_id)    except Exception:        pass

# Node 18 or newer, in the example's folderexport BOXLINE_API_KEY=…   # your API keynpm install @boxline/sdk tsxnpx tsx node/index.ts

# Python 3.9 or newer, in the example's folderexport BOXLINE_API_KEY=…   # your API keypip install boxline-sdkpython python/main.py