diff --git a/.github/workflows/nightly-e2e.yaml b/.github/workflows/nightly-e2e.yaml
index f6195f4d0..3ec14f254 100644
--- a/.github/workflows/nightly-e2e.yaml
+++ b/.github/workflows/nightly-e2e.yaml
@@ -293,6 +293,35 @@ jobs:
           path: /tmp/nemoclaw-e2e-install.log
           if-no-files-found: ignore
 
+  # ── Sandbox operations (recovery + multi-sandbox isolation) ──
+  # Validates sandbox list, connect, status, logs, destroy, gateway
+  # auto-recovery after docker kill, registry rebuild, process recovery,
+  # multi-sandbox metadata, and cross-sandbox network isolation.
+  sandbox-operations-e2e:
+    if: github.repository == 'NVIDIA/NemoClaw'
+    runs-on: ubuntu-latest
+    timeout-minutes: 60
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v6
+
+      - name: Run sandbox operations E2E test
+        env:
+          NVIDIA_API_KEY: ${{ secrets.NVIDIA_API_KEY }}
+          NEMOCLAW_NON_INTERACTIVE: "1"
+          NEMOCLAW_ACCEPT_THIRD_PARTY_SOFTWARE: "1"
+          NEMOCLAW_POLICY_TIER: "open"
+          GITHUB_TOKEN: ${{ github.token }}
+        run: bash test/e2e/test-sandbox-operations.sh
+
+      - name: Upload test log on failure
+        if: failure()
+        uses: actions/upload-artifact@v4
+        with:
+          name: sandbox-operations-test-log
+          path: test-sandbox-operations-*.log
+          if-no-files-found: ignore
+
   # ── GPU E2E (Ollama local inference) ──────────────────────────
   # Enable by setting repository variable GPU_E2E_ENABLED=true
   # (Settings → Secrets and variables → Actions → Variables)
@@ -345,8 +374,8 @@ jobs:
 
   notify-on-failure:
     runs-on: ubuntu-latest
-    needs: [cloud-e2e, cloud-experimental-e2e, messaging-providers-e2e, sandbox-survival-e2e, hermes-e2e, skip-permissions-e2e, gpu-e2e]
-    if: ${{ always() && (needs.cloud-e2e.result == 'failure' || needs.cloud-experimental-e2e.result == 'failure' || needs.messaging-providers-e2e.result == 'failure' || needs.sandbox-survival-e2e.result == 'failure' || needs.hermes-e2e.result == 'failure' || needs.skip-permissions-e2e.result == 'failure' || needs.gpu-e2e.result == 'failure') }}
+    needs: [cloud-e2e, cloud-experimental-e2e, messaging-providers-e2e, sandbox-survival-e2e, hermes-e2e, skip-permissions-e2e, sandbox-operations-e2e, gpu-e2e]
+    if: ${{ always() && (needs.cloud-e2e.result == 'failure' || needs.cloud-experimental-e2e.result == 'failure' || needs.messaging-providers-e2e.result == 'failure' || needs.sandbox-survival-e2e.result == 'failure' || needs.hermes-e2e.result == 'failure' || needs.skip-permissions-e2e.result == 'failure' || needs.sandbox-operations-e2e.result == 'failure' || needs.gpu-e2e.result == 'failure') }}
     permissions:
       issues: write
     steps:
diff --git a/test/e2e/test-sandbox-operations.sh b/test/e2e/test-sandbox-operations.sh
new file mode 100755
index 000000000..01b64f5a5
--- /dev/null
+++ b/test/e2e/test-sandbox-operations.sh
@@ -0,0 +1,697 @@
+#!/usr/bin/env bash
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+#
+# =============================================================================
+# test-sandbox-operations.sh
+# NemoClaw Sandbox Operations E2E Test Suite
+#
+# Covers: TC-SBX-01 through TC-SBX-11
+# Assumes: NemoClaw is installed, no sandbox is currently onboarded
+#
+# Test ordering:
+#   Phase 1 — Basic operations (sandbox A alive)
+#   Phase 2 — Non-destructive recovery (sandbox A alive)
+#   Phase 3 — Destructive gateway kill (destroys sandbox A, then re-onboards)
+#   Phase 4 — Multi-sandbox (onboards sandbox B alongside re-onboarded A)
+#   Phase 5 — Cleanup verification
+# =============================================================================
+
+set -euo pipefail
+
+# ── Overall timeout (prevents hung CI jobs) ──────────────────────────────────
+if [ -z "${NEMOCLAW_E2E_NO_TIMEOUT:-}" ]; then
+  export NEMOCLAW_E2E_NO_TIMEOUT=1
+  TIMEOUT_SECONDS="${NEMOCLAW_E2E_TIMEOUT_SECONDS:-1800}"
+  if command -v timeout >/dev/null 2>&1; then
+    exec timeout -s TERM "$TIMEOUT_SECONDS" bash "$0" "$@"
+  elif command -v gtimeout >/dev/null 2>&1; then
+    exec gtimeout -s TERM "$TIMEOUT_SECONDS" bash "$0" "$@"
+  fi
+fi
+
+# ── Config ───────────────────────────────────────────────────────────────────
+SANDBOX_A="test-sbx-a"
+SANDBOX_B="test-sbx-b"
+LOG_FILE="test-sandbox-operations-$(date +%Y%m%d-%H%M%S).log"
+
+# macOS uses gtimeout (from coreutils); Linux uses timeout
+if command -v gtimeout &>/dev/null; then
+  TIMEOUT_CMD="gtimeout"
+elif command -v timeout &>/dev/null; then
+  TIMEOUT_CMD="timeout"
+else
+  echo "ERROR: Neither timeout nor gtimeout found. Install coreutils: brew install coreutils"
+  exit 1
+fi
+
+RED='\033[0;31m'
+GREEN='\033[0;32m'
+YELLOW='\033[1;33m'
+CYAN='\033[0;36m'
+NC='\033[0m'
+
+# ── Counters ─────────────────────────────────────────────────────────────────
+PASS=0
+FAIL=0
+SKIP=0
+TOTAL=0
+
+# ── Helpers ──────────────────────────────────────────────────────────────────
+log() { echo -e "${CYAN}[$(date +%H:%M:%S)]${NC} $*" | tee -a "$LOG_FILE"; }
+pass() {
+  ((PASS += 1))
+  ((TOTAL += 1))
+  echo -e "${GREEN}  PASS${NC} $1" | tee -a "$LOG_FILE"
+}
+fail() {
+  ((FAIL += 1))
+  ((TOTAL += 1))
+  echo -e "${RED}  FAIL${NC} $1 — $2" | tee -a "$LOG_FILE"
+}
+skip() {
+  ((SKIP += 1))
+  ((TOTAL += 1))
+  echo -e "${YELLOW}  SKIP${NC} $1 — $2" | tee -a "$LOG_FILE"
+}
+
+# Check that a sandbox is registered; skip the named test case if not.
+# Usage: require_sandbox "$SANDBOX_A" "TC-SBX-02" || return
+require_sandbox() {
+  if ! nemoclaw list 2>/dev/null | grep -q "$1"; then
+    skip "$2" "sandbox '$1' not available"
+    return 1
+  fi
+  return 0
+}
+
+# Run a command inside a named sandbox via SSH. Returns the command output.
+# Logs warnings on SSH config failure, empty config, timeout, or non-zero exit.
+sandbox_exec_for() {
+  local name="$1" cmd="$2"
+  local ssh_cfg
+  ssh_cfg="$(mktemp)"
+  if ! openshell sandbox ssh-config "$name" >"$ssh_cfg" 2>/dev/null; then
+    log "  [sandbox_exec] Failed to get SSH config for '$name'"
+    rm -f "$ssh_cfg"
+    echo ""
+    return 1
+  fi
+  if [[ ! -s "$ssh_cfg" ]]; then
+    log "  [sandbox_exec] SSH config for '$name' is empty"
+    rm -f "$ssh_cfg"
+    echo ""
+    return 1
+  fi
+  local result exit_code=0
+  result=$($TIMEOUT_CMD 60 ssh -F "$ssh_cfg" \
+    -o StrictHostKeyChecking=no -o UserKnownHostsFile=/dev/null \
+    -o ConnectTimeout=10 -o LogLevel=ERROR \
+    "openshell-${name}" "$cmd" 2>&1) || exit_code=$?
+  rm -f "$ssh_cfg"
+  if [[ $exit_code -eq 124 ]]; then
+    log "  [sandbox_exec] SSH command timed out after 60s for '$name'"
+  elif [[ $exit_code -ne 0 && -z "$result" ]]; then
+    log "  [sandbox_exec] SSH command failed (exit $exit_code) for '$name'"
+  fi
+  echo "$result"
+}
+
+# Shorthand: run a command inside sandbox A.
+sandbox_exec() {
+  sandbox_exec_for "$SANDBOX_A" "$1"
+}
+
+# Onboard a sandbox by name. Removes stale locks, runs nemoclaw onboard in
+# non-interactive mode, and returns 0 if the sandbox appears in nemoclaw list.
+onboard_sandbox() {
+  local name="$1"
+  log "  Onboarding sandbox '$name'..."
+
+  # Remove stale lock from previous crashed runs
+  rm -f "$HOME/.nemoclaw/onboard.lock" 2>/dev/null || true
+
+  local onboard_exit=0
+  NEMOCLAW_SANDBOX_NAME="$name" \
+    NEMOCLAW_NON_INTERACTIVE=1 \
+    NEMOCLAW_ACCEPT_THIRD_PARTY_SOFTWARE=1 \
+    NEMOCLAW_RECREATE_SANDBOX=1 \
+    nemoclaw onboard --non-interactive --yes-i-accept-third-party-software \
+    2>&1 | tee -a "$LOG_FILE" || onboard_exit=$?
+
+  if [[ $onboard_exit -ne 0 ]]; then
+    log "  [onboard_sandbox] nemoclaw onboard exited with code $onboard_exit"
+    return 1
+  fi
+
+  if ! nemoclaw list 2>/dev/null | grep -q "$name"; then
+    log "  [onboard_sandbox] Sandbox '$name' not found in nemoclaw list after onboard"
+    return 1
+  fi
+  return 0
+}
+
+# ── Resolve repo root ────────────────────────────────────────────────────────
+SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]:-$0}")" && pwd)"
+if [ -f "$SCRIPT_DIR/../../install.sh" ]; then
+  REPO_ROOT="$(cd "$SCRIPT_DIR/../.." && pwd)"
+elif [ -f "./install.sh" ]; then
+  REPO_ROOT="$(pwd)"
+else
+  echo "ERROR: Cannot find install.sh — run from the repo root or test/e2e/"
+  exit 1
+fi
+
+# ── Install NemoClaw if not present ──────────────────────────────────────────
+# Matches the pattern from test-sandbox-survival.sh and test-full-e2e.sh:
+# each E2E test installs NemoClaw from source so it runs on a fresh CI runner.
+install_nemoclaw() {
+  if command -v nemoclaw &>/dev/null; then
+    log "nemoclaw already installed: $(nemoclaw --version 2>/dev/null || echo 'unknown')"
+    return 0
+  fi
+
+  log "=== Installing NemoClaw via install.sh ==="
+
+  local install_exit=0
+  bash "$REPO_ROOT/install.sh" --non-interactive --yes-i-accept-third-party-software \
+    2>&1 | tee -a "$LOG_FILE" || install_exit=$?
+
+  # Source shell profile to pick up PATH changes from install.sh
+  if [ -f "$HOME/.bashrc" ]; then
+    # shellcheck source=/dev/null
+    source "$HOME/.bashrc" 2>/dev/null || true
+  fi
+  export NVM_DIR="${NVM_DIR:-$HOME/.nvm}"
+  if [ -s "$NVM_DIR/nvm.sh" ]; then
+    # shellcheck source=/dev/null
+    . "$NVM_DIR/nvm.sh"
+  fi
+  if [ -d "$HOME/.local/bin" ] && [[ ":$PATH:" != *":$HOME/.local/bin:"* ]]; then
+    export PATH="$HOME/.local/bin:$PATH"
+  fi
+
+  if [[ $install_exit -ne 0 ]]; then
+    echo -e "${RED}FATAL: install.sh failed (exit $install_exit)${NC}"
+    exit 1
+  fi
+
+  if ! command -v nemoclaw &>/dev/null; then
+    echo -e "${RED}FATAL: nemoclaw not found on PATH after install${NC}"
+    exit 1
+  fi
+
+  log "nemoclaw installed: $(nemoclaw --version 2>/dev/null || echo 'unknown')"
+
+  # Destroy the sandbox that install.sh created (we create our own)
+  local install_sandbox
+  install_sandbox="${NEMOCLAW_SANDBOX_NAME:-my-assistant}"
+  if nemoclaw list 2>/dev/null | grep -q "$install_sandbox"; then
+    log "Destroying install sandbox '$install_sandbox'..."
+    nemoclaw "$install_sandbox" destroy --yes 2>/dev/null || true
+  fi
+}
+
+# ── Pre-flight ───────────────────────────────────────────────────────────────
+# Verify prerequisites (Docker, API key), install NemoClaw if needed, and
+# clean up leftover sandboxes and stale locks from previous crashed runs.
+preflight() {
+  log "=== Pre-flight checks ==="
+
+  if ! docker info &>/dev/null; then
+    echo -e "${RED}ERROR: Docker is not running.${NC}"
+    exit 1
+  fi
+  log "Docker is running"
+
+  if [[ -z "${NVIDIA_API_KEY:-}" && -z "${OPENAI_API_KEY:-}" && -z "${ANTHROPIC_API_KEY:-}" ]]; then
+    echo -e "${YELLOW}WARNING: No API key detected.${NC}"
+  fi
+
+  install_nemoclaw
+
+  log "nemoclaw: $(nemoclaw --version 2>/dev/null || echo 'unknown')"
+  log "openshell: $(openshell --version 2>&1 | head -1 || echo 'unknown')"
+  log "timeout: $TIMEOUT_CMD"
+
+  # Remove stale onboard lock from previous crashed runs
+  if [[ -f "$HOME/.nemoclaw/onboard.lock" ]]; then
+    log "Removing stale onboard lock"
+    rm -f "$HOME/.nemoclaw/onboard.lock"
+  fi
+
+  for sb in "$SANDBOX_A" "$SANDBOX_B"; do
+    if nemoclaw list 2>/dev/null | grep -q "$sb"; then
+      log "Cleaning up leftover sandbox: $sb"
+      nemoclaw "$sb" destroy --yes 2>/dev/null || true
+    fi
+  done
+
+  log "Pre-flight complete"
+  echo ""
+}
+
+# ── Setup: Onboard sandbox A ────────────────────────────────────────────────
+# Create the primary test sandbox. Exits the script on failure since all
+# subsequent test cases depend on sandbox A being available.
+setup_sandbox_a() {
+  log "=== Setup: Onboarding sandbox '$SANDBOX_A' ==="
+  log "This may take a few minutes..."
+
+  if ! onboard_sandbox "$SANDBOX_A"; then
+    echo -e "${RED}FATAL: Onboard failed — sandbox '$SANDBOX_A' not found.${NC}"
+    exit 1
+  fi
+
+  log "Sandbox '$SANDBOX_A' onboarded successfully"
+  echo ""
+}
+
+# =============================================================================
+# Phase 1: Basic operations (sandbox A alive)
+# =============================================================================
+
+# ── TC-SBX-01: List Sandboxes ───────────────────────────────────────────────
+test_sbx_01_list_sandboxes() {
+  log "=== TC-SBX-01: List Sandboxes ==="
+
+  local output
+  output=$(nemoclaw list 2>&1)
+
+  if echo "$output" | grep -q "$SANDBOX_A"; then
+    pass "TC-SBX-01: nemoclaw list shows '$SANDBOX_A'"
+  else
+    fail "TC-SBX-01: List Sandboxes" "'$SANDBOX_A' not found in nemoclaw list output"
+  fi
+}
+
+# ── TC-SBX-02: Connect & Chat ───────────────────────────────────────────────
+test_sbx_02_connect_chat() {
+  log "=== TC-SBX-02: Connect & Chat ==="
+  require_sandbox "$SANDBOX_A" "TC-SBX-02" || return
+
+  log "  Sending one-shot message to agent via SSH..."
+  local reply
+  reply=$(sandbox_exec "openclaw agent --agent main --local -m 'Say exactly: HELLO_E2E' --session-id e2e-test" 2>&1) || true
+
+  if echo "$reply" | grep -qi "HELLO_E2E"; then
+    pass "TC-SBX-02: Agent replied with expected token"
+  else
+    fail "TC-SBX-02: Connect & Chat" "Got: $(echo "$reply" | head -3)"
+  fi
+}
+
+# ── TC-SBX-03: Status Fields ────────────────────────────────────────────────
+test_sbx_03_status_fields() {
+  log "=== TC-SBX-03: Status Fields ==="
+  require_sandbox "$SANDBOX_A" "TC-SBX-03" || return
+
+  local output
+  output=$(nemoclaw "$SANDBOX_A" status 2>&1)
+
+  local all_good=true
+  for field in "Sandbox" "Model" "Provider" "GPU"; do
+    if echo "$output" | grep -qi "$field"; then
+      log "  Found field: $field"
+    else
+      log "  MISSING field: $field"
+      all_good=false
+    fi
+  done
+
+  if $all_good; then
+    pass "TC-SBX-03: Status output contains all expected fields"
+  else
+    fail "TC-SBX-03: Status Fields" "Missing expected fields. Output: $(echo "$output" | head -10)"
+  fi
+}
+
+# ── TC-SBX-04: Log Streaming ────────────────────────────────────────────────
+test_sbx_04_log_streaming() {
+  log "=== TC-SBX-04: Log Streaming ==="
+  require_sandbox "$SANDBOX_A" "TC-SBX-04" || return
+
+  local output logs_exit=0
+  output=$($TIMEOUT_CMD 10 nemoclaw "$SANDBOX_A" logs 2>&1) || logs_exit=$?
+
+  if [[ $logs_exit -ne 0 ]]; then
+    fail "TC-SBX-04: Log Streaming" "nemoclaw logs exited with code $logs_exit"
+  elif [[ -n "$output" ]]; then
+    pass "TC-SBX-04: Log streaming produced output ($(echo "$output" | wc -l | tr -d ' ') lines)"
+  else
+    fail "TC-SBX-04: Log Streaming" "nemoclaw logs succeeded but produced no output"
+  fi
+
+  $TIMEOUT_CMD 5 nemoclaw "$SANDBOX_A" logs --follow &>/dev/null &
+  local pid=$!
+  sleep 3
+
+  if ! ps -p "$pid" &>/dev/null; then
+    fail "TC-SBX-04: Log --follow" "Process exited before kill (was not streaming)"
+  else
+    kill "$pid" 2>/dev/null || true
+    wait "$pid" 2>/dev/null || true
+    if ps -p "$pid" &>/dev/null; then
+      fail "TC-SBX-04: Log --follow cleanup" "Orphaned log process still running"
+    else
+      pass "TC-SBX-04: Log --follow exited cleanly after kill"
+    fi
+  fi
+}
+
+# =============================================================================
+# Phase 2: Non-destructive recovery (sandbox A stays alive)
+# =============================================================================
+
+# ── TC-SBX-07: Registry Rebuild ─────────────────────────────────────────────
+test_sbx_07_registry_rebuild() {
+  log "=== TC-SBX-07: Registry Rebuild ==="
+  require_sandbox "$SANDBOX_A" "TC-SBX-07" || return
+
+  local registry="$HOME/.nemoclaw/sandboxes.json"
+  if [[ ! -f "$registry" ]]; then
+    skip "TC-SBX-07" "sandboxes.json not found"
+    return
+  fi
+
+  cp "$registry" "${registry}.bak"
+  log "  Backed up and deleted sandboxes.json"
+  rm -f "$registry"
+
+  local output
+  output=$($TIMEOUT_CMD 60 nemoclaw list 2>&1) || true
+
+  if echo "$output" | grep -q "$SANDBOX_A"; then
+    pass "TC-SBX-07: Registry rebuilt — '$SANDBOX_A' found after deletion"
+    rm -f "${registry}.bak"
+  else
+    fail "TC-SBX-07: Registry Rebuild" "Not found after rebuild. Restoring backup."
+    mv "${registry}.bak" "$registry"
+  fi
+}
+
+# ── TC-SBX-08: Process Recovery ─────────────────────────────────────────────
+test_sbx_08_process_recovery() {
+  log "=== TC-SBX-08: Process Recovery ==="
+  require_sandbox "$SANDBOX_A" "TC-SBX-08" || return
+
+  log "  Killing OpenClaw gateway process inside sandbox..."
+  local kill_output
+  kill_output=$(sandbox_exec "pkill -9 -f 'openclaw gateway' 2>/dev/null || kill -9 \$(pgrep -f 'openclaw gateway') 2>/dev/null || kill -9 \$(ps aux | grep 'openclaw.*gateway' | grep -v grep | awk '{print \$2}') 2>/dev/null; echo EXIT_\$?" 2>&1) || true
+
+  if echo "$kill_output" | grep -q "EXIT_0"; then
+    log "  Process kill confirmed"
+  else
+    log "  WARNING: Could not confirm process was killed (output: $kill_output)"
+  fi
+  sleep 5
+
+  log "  Running nemoclaw status (expect process recovery)..."
+  local status_output status_exit=0
+  status_output=$($TIMEOUT_CMD 120 nemoclaw "$SANDBOX_A" status 2>&1) || status_exit=$?
+
+  if [[ $status_exit -ne 0 ]]; then
+    fail "TC-SBX-08: Process Recovery (status)" "nemoclaw status exited with code $status_exit"
+  elif echo "$status_output" | grep -qiE "recover|running|healthy|OpenClaw"; then
+    pass "TC-SBX-08: Status detected and recovered dead OpenClaw process"
+  else
+    fail "TC-SBX-08: Process Recovery (status)" "Output: $(echo "$status_output" | head -5)"
+  fi
+
+  log "  Verifying SSH still works..."
+  local check
+  check=$(sandbox_exec "echo process-recovery-ok" 2>&1) || true
+  if echo "$check" | grep -q "process-recovery-ok"; then
+    pass "TC-SBX-08: SSH works after process recovery"
+  else
+    fail "TC-SBX-08: Process Recovery (SSH)" "Cannot SSH after recovery"
+  fi
+}
+
+# ── TC-SBX-05: Destroy Cleanup ──────────────────────────────────────────────
+test_sbx_05_destroy_cleanup() {
+  log "=== TC-SBX-05: Destroy Cleanup ==="
+  local target="$1"
+
+  if ! nemoclaw list 2>/dev/null | grep -q "$target"; then
+    skip "TC-SBX-05" "Sandbox '$target' not present"
+    return
+  fi
+
+  log "  Destroying sandbox '$target'..."
+  local destroy_exit=0
+  nemoclaw "$target" destroy --yes 2>&1 | tee -a "$LOG_FILE" || destroy_exit=$?
+
+  if [[ $destroy_exit -ne 0 ]]; then
+    fail "TC-SBX-05: Destroy ($target)" "nemoclaw destroy exited with code $destroy_exit"
+  fi
+
+  if nemoclaw list 2>/dev/null | grep -q "$target"; then
+    fail "TC-SBX-05: Destroy ($target)" "Still in nemoclaw list after destroy (exit $destroy_exit)"
+  else
+    pass "TC-SBX-05: '$target' removed from nemoclaw list"
+  fi
+
+  if openshell sandbox list 2>/dev/null | grep -q "$target"; then
+    fail "TC-SBX-05: Destroy ($target)" "Still in openshell sandbox list after destroy"
+  else
+    pass "TC-SBX-05: '$target' removed from openshell sandbox list"
+  fi
+}
+
+# =============================================================================
+# Phase 3: Destructive gateway kill (destroys sandbox A, then re-onboards)
+# =============================================================================
+
+test_sbx_06_gateway_recovery() {
+  log "=== TC-SBX-06: Gateway Auto-Recovery (destructive) ==="
+  require_sandbox "$SANDBOX_A" "TC-SBX-06" || return
+
+  local container="openshell-cluster-nemoclaw"
+  if ! docker ps -q --filter "name=$container" | grep -q .; then
+    skip "TC-SBX-06" "Gateway container '$container' not running"
+    return
+  fi
+
+  log "  Killing gateway container (simulates Docker crash)..."
+  docker kill "$container" 2>/dev/null || true
+  sleep 5
+
+  # Verify the container is actually stopped
+  local container_state
+  container_state=$(docker inspect -f '{{.State.Running}}' "$container" 2>/dev/null || echo "removed")
+  log "  Gateway container state after kill: $container_state"
+  if [[ "$container_state" == "true" ]]; then
+    skip "TC-SBX-06" "Container still running after docker kill — cannot test recovery"
+    return
+  fi
+
+  log "  Running nemoclaw status (expect recovery attempt — may take several minutes)..."
+  local status_output status_exit=0
+  status_output=$($TIMEOUT_CMD 600 nemoclaw "$SANDBOX_A" status 2>&1) || status_exit=$?
+  log "  Status output (exit $status_exit): $(echo "$status_output" | head -5)"
+
+  if echo "$status_output" | grep -qiE "recover|healthy|Ready|running|gateway"; then
+    pass "TC-SBX-06: Status handled gateway kill gracefully"
+  else
+    fail "TC-SBX-06: Gateway Recovery (status)" "No recovery indication (exit $status_exit). Output: $(echo "$status_output" | head -5)"
+  fi
+
+  # Re-onboard to restore sandbox A for subsequent tests
+  log "  Re-onboarding sandbox A after gateway destruction..."
+  if onboard_sandbox "$SANDBOX_A"; then
+    pass "TC-SBX-06: Re-onboard after gateway kill succeeded"
+  else
+    fail "TC-SBX-06: Re-onboard" "Could not re-onboard after gateway kill"
+  fi
+}
+
+# =============================================================================
+# Phase 4: Multi-sandbox (onboards sandbox B alongside A)
+# =============================================================================
+
+test_sbx_10_multi_sandbox_metadata() {
+  log "=== TC-SBX-10: Multi-Sandbox Metadata ==="
+  require_sandbox "$SANDBOX_A" "TC-SBX-10" || return
+
+  log "  Onboarding second sandbox '$SANDBOX_B'..."
+  if ! onboard_sandbox "$SANDBOX_B"; then
+    fail "TC-SBX-10: Multi-Sandbox" "Sandbox '$SANDBOX_B' failed to onboard"
+    return
+  fi
+
+  local output
+  output=$(nemoclaw list 2>&1)
+
+  local found_a=false found_b=false
+  echo "$output" | grep -q "$SANDBOX_A" && found_a=true
+  echo "$output" | grep -q "$SANDBOX_B" && found_b=true
+
+  if $found_a && $found_b; then
+    pass "TC-SBX-10: Both sandboxes visible in nemoclaw list"
+  else
+    fail "TC-SBX-10: Multi-Sandbox" "Missing sandbox (A=$found_a, B=$found_b)"
+    return
+  fi
+
+  local meta_ok=true
+  for sb in "$SANDBOX_A" "$SANDBOX_B"; do
+    local sb_meta
+    sb_meta=$(echo "$output" | grep -A1 "$sb" | tail -1)
+    if [[ -z "$sb_meta" ]] || ! echo "$sb_meta" | grep -q "model:"; then
+      log "  $sb: metadata line missing or no model field"
+      meta_ok=false
+    elif echo "$sb_meta" | grep -q "model: unknown"; then
+      log "  $sb: model is unknown"
+      meta_ok=false
+    fi
+    if [[ -z "$sb_meta" ]] || ! echo "$sb_meta" | grep -q "provider:"; then
+      log "  $sb: metadata line missing or no provider field"
+      meta_ok=false
+    elif echo "$sb_meta" | grep -q "provider: unknown"; then
+      log "  $sb: provider is unknown"
+      meta_ok=false
+    fi
+  done
+
+  if $meta_ok; then
+    pass "TC-SBX-10: Both sandboxes have non-empty metadata"
+  else
+    fail "TC-SBX-10: Multi-Sandbox Metadata" "One or more sandboxes have unknown model/provider"
+  fi
+}
+
+test_sbx_11_network_isolation() {
+  log "=== TC-SBX-11: Sandbox Network Isolation ==="
+  require_sandbox "$SANDBOX_A" "TC-SBX-11" || return
+  require_sandbox "$SANDBOX_B" "TC-SBX-11" || return
+
+  # Use node (always available) instead of curl (removed by hardening).
+  # Isolation is enforced by the OpenShell proxy — blocked requests return
+  # HTTP 403. Connection errors (ENOTFOUND, ECONNREFUSED, TIMEOUT) also
+  # count as isolation. Only HTTP 200 would indicate a breach.
+  log "  Testing: sandbox A cannot reach sandbox B by hostname..."
+  local probe_a
+  probe_a=$(sandbox_exec_for "$SANDBOX_A" "node -e \"
+const http = require('http');
+const req = http.get('http://${SANDBOX_B}:18789/', (res) => {
+  console.log('STATUS_' + res.statusCode);
+  res.resume();
+});
+req.on('error', (e) => console.log('ERROR: ' + e.message));
+req.setTimeout(5000, () => { req.destroy(); console.log('TIMEOUT'); });
+\"" 2>&1) || true
+
+  if [[ -z "$probe_a" ]]; then
+    fail "TC-SBX-11: Isolation (A→B)" "Empty response — SSH or infrastructure failure"
+  elif echo "$probe_a" | grep -qiE "STATUS_403|ERROR|TIMEOUT"; then
+    pass "TC-SBX-11: Sandbox A cannot reach sandbox B ($(echo "$probe_a" | grep -oE 'STATUS_[0-9]+|ERROR|TIMEOUT' | head -1))"
+  elif echo "$probe_a" | grep -qE "STATUS_[0-9]+"; then
+    fail "TC-SBX-11: Isolation (A→B)" "Sandbox A reached sandbox B ($(echo "$probe_a" | grep -oE 'STATUS_[0-9]+' | head -1))"
+  else
+    fail "TC-SBX-11: Isolation (A→B)" "Unexpected probe output: $(echo "$probe_a" | head -3)"
+  fi
+
+  log "  Testing reverse: sandbox B cannot reach sandbox A..."
+  local probe_b
+  probe_b=$(sandbox_exec_for "$SANDBOX_B" "node -e \"
+const http = require('http');
+const req = http.get('http://${SANDBOX_A}:18789/', (res) => {
+  console.log('STATUS_' + res.statusCode);
+  res.resume();
+});
+req.on('error', (e) => console.log('ERROR: ' + e.message));
+req.setTimeout(5000, () => { req.destroy(); console.log('TIMEOUT'); });
+\"" 2>&1) || true
+
+  if [[ -z "$probe_b" ]]; then
+    fail "TC-SBX-11: Isolation (B→A)" "Empty response — SSH or infrastructure failure"
+  elif echo "$probe_b" | grep -qiE "STATUS_403|ERROR|TIMEOUT"; then
+    pass "TC-SBX-11: Sandbox B cannot reach sandbox A ($(echo "$probe_b" | grep -oE 'STATUS_[0-9]+|ERROR|TIMEOUT' | head -1))"
+  elif echo "$probe_b" | grep -qE "STATUS_[0-9]+"; then
+    fail "TC-SBX-11: Isolation (B→A)" "Sandbox B reached sandbox A ($(echo "$probe_b" | grep -oE 'STATUS_[0-9]+' | head -1))"
+  else
+    fail "TC-SBX-11: Isolation (B→A)" "Unexpected probe output: $(echo "$probe_b" | head -3)"
+  fi
+}
+
+# ── Teardown ─────────────────────────────────────────────────────────────────
+teardown() {
+  # Disable errexit during teardown — cleanup must be best-effort
+  set +e
+  log ""
+  log "=== Teardown ==="
+  for sb in "$SANDBOX_B" "$SANDBOX_A"; do
+    if nemoclaw list 2>/dev/null | grep -q "$sb"; then
+      log "Destroying sandbox '$sb'..."
+      nemoclaw "$sb" destroy --yes 2>/dev/null || true
+    fi
+  done
+  # Clean up gateway if no sandboxes remain
+  openshell gateway destroy -g nemoclaw 2>/dev/null || true
+  rm -f "$HOME/.nemoclaw/onboard.lock" 2>/dev/null || true
+  log "Teardown complete"
+  set -e
+}
+
+# ── Summary ──────────────────────────────────────────────────────────────────
+summary() {
+  echo ""
+  echo "============================================================"
+  echo "  TEST SUMMARY"
+  echo "============================================================"
+  echo -e "  ${GREEN}PASS: $PASS${NC}"
+  echo -e "  ${RED}FAIL: $FAIL${NC}"
+  echo -e "  ${YELLOW}SKIP: $SKIP${NC}"
+  echo "  TOTAL: $TOTAL"
+  echo "============================================================"
+  echo "  Log: $LOG_FILE"
+  echo "============================================================"
+  echo ""
+
+  if [[ $FAIL -gt 0 ]]; then
+    exit 1
+  fi
+  exit 0
+}
+
+# ── Main ─────────────────────────────────────────────────────────────────────
+main() {
+  echo ""
+  echo "============================================================"
+  echo "  NemoClaw Sandbox Operations E2E Test Suite"
+  echo "  $(date)"
+  echo "============================================================"
+  echo ""
+
+  preflight
+  setup_sandbox_a
+
+  # Phase 1: Basic operations (sandbox A alive)
+  test_sbx_01_list_sandboxes
+  test_sbx_02_connect_chat
+  test_sbx_03_status_fields
+  test_sbx_04_log_streaming
+
+  # Phase 2: Non-destructive recovery (sandbox A stays alive)
+  test_sbx_07_registry_rebuild
+  test_sbx_08_process_recovery
+
+  # Phase 3: Destructive gateway kill (destroys + re-onboards sandbox A)
+  test_sbx_06_gateway_recovery
+
+  # Phase 4: Multi-sandbox (onboards sandbox B)
+  test_sbx_10_multi_sandbox_metadata
+  test_sbx_11_network_isolation
+
+  # Phase 5: Cleanup verification (destroys sandbox B)
+  test_sbx_05_destroy_cleanup "$SANDBOX_B"
+
+  # Report — teardown runs via EXIT trap, no need to call explicitly
+  trap - EXIT
+  teardown
+  summary
+}
+
+trap teardown EXIT
+main "$@"