fix: enable mTLS, inference routing, and SSH session for Kind

[pdettori]

Summary

• Add TLS volume mounts and env vars for mTLS between sandbox pod and gateway
• Set infrastructure env vars for inference routing (ANTHROPIC_BASE_URL=https://inference.local, OPENAI_BASE_URL=https://inference.local/v1)
• Add SSH socket path and handshake secret for supervisor SSH server
• Add kagenti.io/inject=disabled label to prevent authbridge webhook sidecar injection
• Set CLAUDE_CODE_DISABLE_EXPERIMENTAL_BETAS=1 for LiteLLM proxy compatibility

Context

The compute driver creates Sandbox CR pod specs. For the full OpenShell flow to work on Kind (gateway login → sandbox create → Claude Code), the pod needs:

1. mTLS certs mounted as volumes so the supervisor can authenticate to the gateway
2. SSH socket path so the supervisor starts its SSH server (required for ConnectSupervisor relay)
3. Inference routing env vars pointing to inference.local — the sandbox proxy intercepts these and routes to the upstream LLM
4. Inject-disabled label to prevent the kagenti authbridge webhook from injecting sidecars into sandbox pods

Test plan

• Deploy Kind cluster with kagenti + OpenShell
• openshell sandbox create --provider claude --no-auto-providers -- claude --print hello returns LLM response
• Sandbox pod has TLS volume mounts (/tls/ca, /tls/client)
• SSH session env contains ANTHROPIC_BASE_URL=https://inference.local

Generated with Claude Code