CI test runner

Run a suite in a disposable sandbox, watch it fail, apply a fix, watch it pass — then turn the output into a structured report.

A CI job is just a sandbox that clones a repo, runs the tests, and reports back. This workflow does the smallest honest version of that: a three-assertion suite against an implementation with a planted bug, a fix applied mid-run, and the raw node --test output parsed into the pass/fail summary a real job would post to a PR.

Not connected
createSandboxwriteFileexecdeleteSandbox~40s
Connect to a sandbox server above to run this.

What just happened

  • The project (package.json, sum.js, sum.test.js) was written into /proj with sb.writeFile. sum.js subtracts instead of adds.
  • sb.exec("node --test") ran the suite. Its stdout was streamed live and also captured so the step could parse # pass / # fail counts out of it.
  • The fix from the editor overwrote sum.js — same writeFile call, no new sandbox.
  • The suite ran again, green this time, and the final step emitted a JSON report comparing before and after.

Everything ran in one container. A real runner would fork per shard — that's the Parallel Test Shards cookbook, and the fork workflow here.

Run untrusted code

Execute an arbitrary Python snippet in a disposable, resource-capped container — and prove it's isolated.

Checkpoint a pipeline

Run an extract → transform → load pipeline that snapshots the container after every stage.

On this page

What just happened