provision

Create one ephemeral SandboxClaim from a pre-installed SandboxTemplate and return its endpoint.

Skills to load

• /agent-sandboxing:k8s-agent-sandbox
• /agent-sandboxing:claude-code-on-sandbox
• /core:nushell
• /core:anti-fabrication

Steps

1. Determine session id

If --session-id is not provided, generate one:

SESSION_ID=$(uuidgen | tr 'A-Z' 'a-z' | tr -d '-' | head -c 12)

2. Determine template name

If --template is not provided, list installed templates and ask the user which to use:

kubectl get sandboxtemplate -o name

Typical names: claude-code-kata, claude-code-gvisor, claude-code-kina.

3. Render the claim

nu <CLAUDE_PLUGIN_ROOT>/scripts/render-claim.nu \
  --session-id $SESSION_ID \
  --template $TEMPLATE_NAME \
  ${SHUTDOWN_TIME:+--shutdown-time $SHUTDOWN_TIME} \
  | kubectl apply -f -

The script reads templates/SandboxClaim.session.yaml, substitutes placeholders, and emits the final YAML to stdout. If the operator wants to inspect before applying, drop the pipe to kubectl apply.

4. Wait for Bound

nu <CLAUDE_PLUGIN_ROOT>/scripts/wait-bound.nu --session-id $SESSION_ID --timeout 120

The script polls kubectl get sandboxclaim every 2 seconds until status.phase == Bound or the timeout elapses. Reports the bound pod name on success; reports controller log excerpts on timeout.

If the claim sticks in Pending, common causes:

• Missing RuntimeClass — kubectl describe sandboxclaim session-$SESSION_ID will say so.
• Missing image — kubectl describe pod ... -n <claim-namespace> will show ImagePullBackOff.
• Missing anthropic secret — pod will be Running but claude --version will fail at exec time.

5. Report the endpoint

kubectl get sandboxclaim session-$SESSION_ID -o jsonpath='{.status.endpoint}'
echo

Tell the user the session id, the bound pod name, and the endpoint. Suggest the next command is /agent-sandboxing:exec $SESSION_ID to run something, or /agent-sandboxing:reap $SESSION_ID when done.