---
title: ShellBench task runtime
sdk: docker
app_port: 7860
---

# ShellBench HF Sandbox runtime

This is the source of the public dependency-only image used by all 115 private
ShellBench task packages through HF Sandbox:

```text
hf.co/spaces/osolmaz/shellbench-runtime-b12d083f9044
```

The image contains no task data. Harbor uploads each private task environment
after creating its Sandbox. Node 22 and the benchmark's exact OpenClaw version
are baked into the image because files installed during one HF Sandbox command
are not reliably executable in a later command.

Update the existing Space rather than creating another runtime Space:

```sh
hf upload osolmaz/shellbench-runtime-b12d083f9044 \
  runtime/hf-sandbox . \
  --repo-type space \
  --commit-message "fix: bake shellbench agent runtime"
```

After the Space build succeeds, create a temporary CPU Sandbox from the Space
image. Run these as separate Sandbox commands and require every command to
succeed:

```sh
. ~/.nvm/nvm.sh && nvm install 22
. ~/.nvm/nvm.sh && nvm use 22 && node --version && npm --version
. ~/.nvm/nvm.sh && nvm use 22 && npm install -g openclaw@2026.7.1
. ~/.nvm/nvm.sh && nvm use 22 && openclaw --version
```

Kill the temporary Sandbox after the gate. Record the resulting Space revision
in the active launch catalog before submitting profiles. The Space should stay
paused when it is not being built or inspected.