Files

T

Antigravity Agent 93ec7e9056 [G1-Sync] Manual knowledge update

2026-05-09 21:08:02 +09:00

5.0 KiB

Raw Blame History

id, title, category, status, source_trust_level, verification_status, created_at, updated_at, tags, tech_stack, applied_in, aliases

title

Code Interpreter

LLM 이 작성한 코드를 안전 실행. 격리된 sandbox 가 핵심. E2B / Daytona / Modal / Cloudflare Containers / 자체 K8s. OpenAI Code Interpreter, Anthropic Bash tool 도 같은 패턴.

📖 핵심 개념

Sandbox: 격리 + 자원 제한 + 시간 제한.
Stateful: 한 세션 안 변수 유지 (Jupyter kernel).
File I/O: upload + 결과 download.
Streaming output.

💻 코드 패턴

E2B Code Interpreter

import { Sandbox } from '@e2b/code-interpreter';

const sb = await Sandbox.create({ apiKey: process.env.E2B_API_KEY });

const exec = await sb.runCode(`
import pandas as pd
df = pd.read_csv('/data/sales.csv')
df.groupby('region').sum()
`);

console.log(exec.text);   // stdout
console.log(exec.results); // 차트, table 등 rich result
console.log(exec.error);

await sb.kill();

File upload / download

await sb.files.write('/data/sales.csv', csvBuffer);

const out = await sb.runCode(`
import matplotlib.pyplot as plt
df.plot()
plt.savefig('/out/chart.png')
`);

const png = await sb.files.read('/out/chart.png', 'binary');

Streaming

const exec = await sb.runCode(longTask, {
  onStdout: (line) => stream.write(line),
  onStderr: (line) => stream.write(line),
});

LLM tool 로 wrap

const tools = [{
  name: 'execute_python',
  description: 'Execute Python code in a sandbox. Files in /data/. Save outputs to /out/. State persists between calls in the same conversation.',
  input_schema: {
    type: 'object',
    properties: { code: { type: 'string' } },
    required: ['code'],
  },
}];

async function executeTool(name: string, input: { code: string }) {
  if (name === 'execute_python') {
    const r = await sb.runCode(input.code);
    return JSON.stringify({
      stdout: r.text.slice(-2000),
      error: r.error?.value,
      results: r.results.map(x => ({ type: x.type, ...x })),
    });
  }
}

자원 제한

const sb = await Sandbox.create({
  template: 'code-interpreter-v1',
  timeoutMs: 60_000,        // 세션 최대 1분
  cpuCount: 1,
  memoryMB: 512,
});

const exec = await sb.runCode(code, { timeoutMs: 30_000 }); // 1 cell 30초

Persistent state (Jupyter)

await sb.runCode('x = 42');
const r = await sb.runCode('print(x * 2)'); // 84

자체 sandbox (Docker)

FROM python:3.12-slim
RUN useradd -m sandbox
USER sandbox
WORKDIR /home/sandbox
RUN pip install pandas numpy matplotlib

import { spawn } from 'node:child_process';

async function runInDocker(code: string): Promise<string> {
  return new Promise((resolve, reject) => {
    const p = spawn('docker', [
      'run', '--rm',
      '--network', 'none',
      '--memory', '512m',
      '--cpus', '0.5',
      '-i', 'python:3.12-slim',
      'python', '-c', code,
    ]);
    let out = ''; let err = '';
    p.stdout.on('data', (d) => { out += d.toString(); });
    p.stderr.on('data', (d) => { err += d.toString(); });
    p.on('close', (c) => c === 0 ? resolve(out) : reject(new Error(err)));
    setTimeout(() => p.kill(), 30_000);
  });
}

⚠️ Docker 자체로 sandbox 부족 — gVisor / firecracker / E2B 권장.

출력 안전 처리

function sanitizeOutput(s: string): string {
  return s
    .replace(/[\x00-\x08\x0b-\x1f]/g, '') // control chars
    .slice(0, 10_000);                     // 길이 제한
}

결과 caching (같은 코드)

const key = sha256(code + ':' + dataHash);
const cached = await cache.get(key);
if (cached) return cached;
const r = await sb.runCode(code);
await cache.set(key, r);

🤔 의사결정 기준

사용	추천
ChatGPT-style data analysis	E2B Code Interpreter
Long-running compute	Modal / Daytona
Cloudflare 환경	Cloudflare Containers
Self-hosted K8s	Pod per session + gVisor
단순 calc	Python eval 위험 — math.js / sandbox js
Untrusted user code (CTF, 강의)	gVisor / Firecracker / E2B

❌ 안티패턴

eval() 직접: RCE. sandbox 필수.
Docker --privileged: escape 가능.
Network 허용 + 무제한: SSRF / 외부 호출.
Memory / CPU 무제한: fork bomb / OOM.
Output 무제한 → LLM 으로: 다음 turn 비싸짐.
Sandbox 재사용 cross-user: state leak.
Filesystem 무제한 write: 디스크 채움.
Timeout 없음: 영원 hang.

🤖 LLM 활용 힌트

E2B / Daytona 가 가장 modern.
Sandbox = network none + memory limit + cpu limit + timeout.
Output truncate 후 LLM 으로.

5.0 KiB Raw Blame History