Fullsend

Appearance

Purpose ​

Run the host-side API server experiment to validate that agents inside an OpenShell sandbox can call API servers running on the host through the L7 proxy.

Requirements ​

RequirementLink
Go 1.23+https://go.dev/dl/
Python 3.11+https://www.python.org/downloads/
OpenShell CLI v0.0.43+https://github.com/NVIDIA/OpenShell
openshell-gateway v0.0.43+https://github.com/NVIDIA/OpenShell/releases
Podman (rootless)https://podman.io/docs/installation
curl(pre-installed on most systems)
githttps://git-scm.com/downloads

OpenShell setup ​

This experiment uses the standalone openshell-gateway binary with the Podman driver — not the older K3s-in-Docker approach (openshell gateway start, removed in v0.0.37).

  1. Install Podman and start the socket:

    bash
    systemctl --user start podman.socket

  2. Download the openshell-gateway binary from the OpenShell releases page and place it in your $PATH.

  3. Pull the required images:

    bash
    podman pull ghcr.io/nvidia/openshell/supervisor:latest
podman pull ghcr.io/nvidia/openshell-community/sandboxes/base:latest

  4. Start the gateway:

    bash
    OPENSHELL_SSH_HANDSHAKE_SECRET="$(python3 -c 'import secrets; print(secrets.token_hex(16))')" \
OPENSHELL_SUPERVISOR_IMAGE="ghcr.io/nvidia/openshell/supervisor:latest" \
OPENSHELL_SANDBOX_IMAGE="ghcr.io/nvidia/openshell-community/sandboxes/base:latest" \
OPENSHELL_SANDBOX_IMAGE_PULL_POLICY="missing" \
openshell-gateway \
  --bind-address 0.0.0.0 \
  --port 18080 \
  --health-port 18081 \
  --drivers podman \
  --disable-tls \
  --db-url "sqlite:/tmp/openshell-gateway.db?mode=rwc" \
  --log-level info &

  5. Register the gateway with the CLI:

    bash
    CONFIG_DIR="${XDG_CONFIG_HOME:-$HOME/.config}/openshell"
mkdir -p "${CONFIG_DIR}/gateways/podman-local"
cat > "${CONFIG_DIR}/gateways/podman-local/metadata.json" <<'EOF'
{
  "name": "podman-local",
  "gateway_endpoint": "http://127.0.0.1:18080",
  "is_remote": false,
  "gateway_port": 18080,
  "auth_mode": "plaintext"
}
EOF
printf 'podman-local' > "${CONFIG_DIR}/active_gateway"

  6. Verify the gateway is healthy:

    bash
    curl -sf http://127.0.0.1:18081/healthz

Environment variables ​

VariableDescription
FULLSEND_GCP_PROJECT_IDGCP project with Vertex AI API enabled
CLOUD_ML_REGIONVertex AI region (default: global)
GOOGLE_APPLICATION_CREDENTIALSPath to a service account key JSON with roles/aiplatform.user in the project

Create a service account and key if you don't have one:

bash
gcloud iam service-accounts create fullsend-runner \
  --display-name="Fullsend Runner" \
  --project=<your-gcp-project>

gcloud projects add-iam-policy-binding <your-gcp-project> \
  --member="serviceAccount:fullsend-runner@<your-gcp-project>.iam.gserviceaccount.com" \
  --role="roles/aiplatform.user" \
  --condition=None

gcloud iam service-accounts keys create /tmp/sa-key.json \
  --iam-account=fullsend-runner@<your-gcp-project>.iam.gserviceaccount.com

export GOOGLE_APPLICATION_CREDENTIALS=/tmp/sa-key.json
export FULLSEND_GCP_PROJECT_ID=<your-gcp-project>

The fullsend run CLI must also be installed and in your $PATH.

Steps ​

Automated (fullsend run) ​

  1. Navigate to the experiment directory:

    bash
    cd experiments/host-side-api-server

  2. Run the setup script to build Go binaries and verify prerequisites:

    bash
    ./setup.sh

  3. Run a harness:

    bash
    ./run.sh baked-instructions-full

    Available harnesses (3 discovery methods × 2 policies):

    • baked-instructions-full / baked-instructions-restricted
    • openapi-discovery-full / openapi-discovery-restricted
    • tooluse-discovery-full / tooluse-discovery-restricted

  4. Results are saved to results/.

Expected Output ​

  • Both API servers start and pass health checks
  • Sandbox is created with L7 policy applied
  • From inside the sandbox, curl to API server endpoints succeeds for allowed endpoints and returns 403 for restricted ones
  • Container build via host API completes successfully
  • Repo provisioning clones, scans, and reports results