Vercel sandbox backend¶

Bernstein's vercel sandbox backend talks to the Vercel Sandbox API (https://vercel.com/docs/sandbox) and conforms to the SandboxBackend protocol.

Module¶

src/bernstein/core/sandbox/backends/vercel.py - class VercelSandboxBackend, registered under the name vercel in bernstein.core.sandbox.registry.

Environment variables¶

Variable	Required	Purpose
`VERCEL_TOKEN`	yes	Personal/team API token (`https://vercel.com/account/tokens`).
`VERCEL_TEAM_ID`	conditional	Required for tokens scoped to a team. Forwarded as `?teamId=` on every request.
`VERCEL_API_URL`	no	API root override. Defaults to `https://api.vercel.com`.

Capabilities¶

FILE_RW, EXEC, NETWORK. SNAPSHOT is not declared because the Vercel Sandbox API does not currently expose a snapshot/restore primitive.

Selecting the backend¶

sandbox:
  backend: vercel
  options:
    runtime: node22
    region: iad1

Honest limitations¶

No exec streaming on the synchronous endpoint. The Vercel Sandbox HTTP API returns the buffered stdout/stderr after the command exits. For interactive workloads (tail-style log streaming), use the worktree or docker backends.
No stdin on the sync exec route. Passing stdin= raises NotImplementedError.
Snapshots not supported. Persist state via Vercel-managed storage (KV / Blob) rather than relying on session snapshotting.

Integration tests¶

Live integration tests are gated by CI_VERCEL_TEST=1 plus VERCEL_TOKEN. Without them the test in tests/integration/sandbox/test_vercel_backend.py skips cleanly.