4. Run in OpenShell (Optional)
Run a Hermes code-review agent inside an OpenShell sandbox, controlled by the NVIDIA NeMo Fabric Python SDK on the host. OpenShell supplies isolation and policy; the adapter and harness execute the task inside the sandbox.
This optional continuation uses the repository’s Hermes code-review example. It requires a source checkout and a Linux host. It packages a sample repository and skill in a task image rather than automatically moving the scripts or MCP server from the preceding guides into a sandbox.
Prepare a Linux Host
Use a NeMo Fabric source checkout with its Python SDK and native extension installed, following Development Setup. The shared launcher requires Docker with Linux containers, Rust/Cargo, uv, Python 3, curl, openssl, and tar. The Docker daemon must be running. The launcher currently supports Linux x86_64 and arm64 hosts; this procedure does not run directly on macOS or Windows.
From the repository root, select your installed host SDK environment and provide the NVIDIA API key used for model access:
The image contains its own Hermes installation and adapter. You do not need to reuse the local Hermes interpreter from quickstart two. Refer to the shared launcher requirements if your host environment is elsewhere.
Run the Sandboxed Code Review
Run from the repository root:
The launcher builds the provider and task images, downloads its pinned OpenShell release, starts a local gateway, and configures model access. The host consumer prepares a sandbox, runs two review turns on one runtime, then stops the runtime and releases the sandbox. NeMo Relay is off by default.
The example uses its own configured NVIDIA model. To select another model available to your key, set OPENSHELL_CODE_REVIEW_MODEL before launching. Alternate endpoints also need a matching OpenShell provider policy, as described in the example’s model-access instructions.
Inspect the Result
Open .tmp/code-review-hermes/fabric_owned-report.json. A completed run has top-level status: "succeeded" and two entries in reviews, each with status succeeded. Inspect each review’s output.response for the finding and follow-up explanation. The second review should recall the example’s cedar-4827 review tag. Collected runtime artifacts appear under .tmp/code-review-hermes/artifacts/.
The consumer attempts runtime shutdown and environment release even after a failure. The launcher checks sandbox removal in this ownership mode and cleans up its gateway resources. If the command fails, inspect the report’s error field and launcher output; the OpenShell example guide explains setup, networking, and ownership modes.
Next Steps
Use Environments and Sandboxes to integrate sandbox preparation and release into your application. The custom LangGraph example demonstrates the same deployment path with an application-owned agent and adapter.