Skip to content

README.md

Latest commit

 

History

History
68 lines (51 loc) · 4.21 KB

README.md

File metadata and controls

68 lines (51 loc) · 4.21 KB
Raw

Codex on Tenki (headless, on a real repo)

Run OpenAI Codex non-interactively against a real Git checkout inside a disposable Tenki Sandbox microVM, then read back the diff the agent produced and throw the machine away.

The code (run.mjs)

import { TenkiSandbox, stdoutText } from "@tenkicloud/sandbox";

const REPO = "https://github.com/sindresorhus/yocto-queue";
const TASK = "Add a toArray() method to the Queue class in index.js that returns the queued values as an array, oldest first. Declare it in index.d.ts and add a test for it in test.js.";

const tenki = new TenkiSandbox({ authToken: process.env.TENKI_AUTH_TOKEN });

// cloneRepoUrl checks the repo out to ./repo before createAndWait resolves.
await using sandbox = await tenki.createAndWait({
  cpuCores: 2,
  memoryMb: 4096,
  cloneRepoUrl: REPO,
  workspaceId: process.env.TENKI_WORKSPACE_ID,
});

// Outbound is on by default, so no allowOutbound is needed to reach the npm registry.
await sandbox.exec("npm", { args: ["i", "-g", "@openai/codex"] });

const decoder = new TextDecoder();
await sandbox.exec("codex", {
  // Codex sandboxes itself with bubblewrap. Inside a microVM that layer is redundant, and
  // bypassing it drops both the nesting and the "could not find bubblewrap" warning.
  args: ["exec", "--dangerously-bypass-approvals-and-sandbox", TASK],
  cwd: "repo", // relative paths resolve under the workdir, /home/tenki
  timeoutMs: 10 * 60_000,
  env: { OPENAI_API_KEY: process.env.OPENAI_API_KEY }, // scoped to this one process
  onOutput: ({ data }) => process.stdout.write(decoder.decode(data)), // data is a Uint8Array
});

// sandbox.git.* runs at the workdir, and the checkout is one level down — so use `git -C`.
console.log(stdoutText(await sandbox.exec("git", { args: ["-C", "repo", "diff"] })));

The last line is the point: a unified diff against a real upstream checkout, produced by an agent that had free run of a machine you are about to delete.

Run it

npm install
export TENKI_AUTH_TOKEN=...      # from `tenki login` (~/.config/tenki/config.yaml)
export TENKI_WORKSPACE_ID=...
export OPENAI_API_KEY=sk-...     # the agent turn
node run.mjs                     # streams the agent's turn, then prints the diff

run.mjs uses top-level await using, which needs Node 24+.

Verify the Tenki half without a model key — this is what CI runs:

node verify.mjs   # create + clone → install the CLI → check its version and flags → edit → git diff

Notes

  • Codex brings its own sandbox, and nesting it inside a microVM buys you nothing. Under the default --sandbox workspace-write it looks for bubblewrap, warns could not find bubblewrap on PATH, and falls back to a bundled copy. --dangerously-bypass-approvals-and-sandbox reports sandbox: danger-full-access and skips the layer — the right call here precisely because the blast radius is one disposable VM.
  • sandbox.git.* runs at the sandbox workdir (/home/tenki), but cloneRepoUrl checks out one level down into ./repo — so sandbox.git.diff() fails with "not a git repository". Read the checkout with sandbox.exec("git", { args: ["-C", "repo", "diff"] }), or clone at the workdir root if you want the helpers.
  • codex exec runs fine outside a Git repo, so --skip-git-repo-check is not needed here; the checkout just gives it something to diff. Pass -C, --cd <DIR> if you would rather point Codex at a directory than set cwd.
  • The default image already ships Node 24, npm 11, and git 2.43, and outbound network is on for a bare create. npm i -g @openai/codex finishes in about seven seconds — no allowOutbound, no custom image, no baked-in CLI.
  • exec(command, { args }) runs a bare binary and its args with no shell splitting, so the whole task prompt goes through as a single argument. ExecOptions.env is scoped to that one process — the model key never lands in the repo or the image. A key whose value is undefined arrives as an empty string rather than unset, so filter before adding more vars.
  • The same shape with Anthropic's CLI is in claude-code-sandbox; for the OpenAI Agents SDK as a code-interpreter backend, see openai-agents-sdk.