Support creating sandbox CRs in a configurable target namespace (multi-tenant isolation)

Problem Statement

Problem

The OpenShell gateway only manages sandbox CRs in its own namespace. It tracks sandboxes internally by UUID and requires the CRD to exist in the same namespace as the gateway pod. There is no --namespace flag or multi-namespace watch capability.

This means all agent sandbox pods from all tenants run in a single shared namespace, regardless of organizational boundaries.

Current Behavior

• Gateway deployed in namespace openshell
• All sandbox CRs must be created in openshell
• All agent pods run in openshell
• No way to isolate tenants at the Kubernetes namespace level

Proposed Design

Desired Behavior

Option A: Gateway watches multiple namespaces

# Gateway configuration
args:
  - --watch-namespaces=project-dev,project-prod,openshell

Option B: Sandbox CRD accepts target namespace

apiVersion: agents.x-k8s.io/v1alpha1
kind: Sandbox
metadata:
  name: my-agent
  namespace: project-dev   # Sandbox CR AND pod created HERE

Option C: Per-tenant gateway instances (documented pattern)

Deploy lightweight gateway replicas scoped to specific namespaces.

Option B is preferred.

Alternatives Considered

Why This Matters

Requirement	Current (single NS)	With Multi-NS
Kubernetes RBAC per tenant	Not possible	Tenants manage their own NS
NetworkPolicy isolation	All agents share NS	Per-namespace NetworkPolicy
Resource quotas per team	Single quota for all	Per-namespace ResourceQuota
Istio PeerAuthentication	Shared identity	Per-namespace mTLS policies
Audit attribution	All in one NS	Clear namespace ownership
Compliance (data residency)	Cannot segregate	Namespace-level controls

Use Case

Enterprise platform with multiple teams/tenants deploying agents. Each team has their own Kubernetes namespace with RBAC, NetworkPolicy, ResourceQuota, and audit requirements. Currently all their agents are forced into a single shared namespace, breaking existing organizational security controls.

Workaround

• All sandboxes run in the gateway namespace (openshell)
• Logical tenant isolation via custom AgentClient CRD (controls which client IDs can call which agents)
• Auth enforcement via Istio AuthorizationPolicy with label selectors on agent pods
• No true namespace-level isolation between tenants' workloads

Agent Investigation

No response

Checklist

• I've reviewed existing issues and the architecture docs
• This is a design proposal, not a "please build this" request

No implementation files or tests are named. Start by locating the gateway's namespace handling and Sandbox CRD processing, then review the architecture docs and compare the proposed multi-namespace, target-namespace, and per-tenant gateway approaches. Done requires an agreed design and defined behavior for namespace-scoped sandbox creation and isolation.