diff --git a/services/maintenance/node-ops/node-image-sweeper-daemonset.yaml b/services/maintenance/node-ops/node-image-sweeper-daemonset.yaml
index 6f029b92..db4b2a9e 100644
--- a/services/maintenance/node-ops/node-image-sweeper-daemonset.yaml
+++ b/services/maintenance/node-ops/node-image-sweeper-daemonset.yaml
@@ -16,7 +16,7 @@ spec:
   template:
     metadata:
       annotations:
-        atlas.bstein.dev/config-revision: "2026-10-04-preserve-active-pod-logs"
+        atlas.bstein.dev/config-revision: "2026-10-04-native-image-gc"
       labels:
         app: node-image-sweeper
     spec:
@@ -51,8 +51,6 @@ spec:
           env:
             - name: SWEEP_INTERVAL_SEC
               value: "7200"
-            - name: HIGH_USAGE_PERCENT
-              value: "70"
             - name: EMERGENCY_USAGE_PERCENT
               value: "80"
             - name: LOG_RETENTION_DAYS
diff --git a/services/maintenance/node-ops/scripts/node_image_sweeper.sh b/services/maintenance/node-ops/scripts/node_image_sweeper.sh
index 140560d9..52d389dd 100644
--- a/services/maintenance/node-ops/scripts/node_image_sweeper.sh
+++ b/services/maintenance/node-ops/scripts/node_image_sweeper.sh
@@ -3,7 +3,6 @@ set -eu
 
 ONE_SHOT=${ONE_SHOT:-false}
 SWEEP_INTERVAL_SEC=${SWEEP_INTERVAL_SEC:-21600}
-HIGH_USAGE_PERCENT=${HIGH_USAGE_PERCENT:-70}
 EMERGENCY_USAGE_PERCENT=${EMERGENCY_USAGE_PERCENT:-85}
 LOG_RETENTION_DAYS=${LOG_RETENTION_DAYS:-7}
 ORPHAN_POD_RETENTION_DAYS=${ORPHAN_POD_RETENTION_DAYS:-3}
@@ -12,11 +11,8 @@ JOURNAL_MAX_SIZE=${JOURNAL_MAX_SIZE:-200M}
 sweep_once() {
   usage=$(df -P /host | awk 'NR==2 {gsub(/%/,"",$5); print $5}') || usage=""
 
-  # crictl image metadata frequently omits createdAt on this cluster; prune by
-  # runtime reachability whenever rootfs crosses pressure thresholds.
-  if [ -n "${usage}" ] && [ "${usage}" -ge "${HIGH_USAGE_PERCENT}" ]; then
-    chroot /host /bin/sh -c "crictl rmi --prune >/dev/null 2>&1 || true"
-  fi
+  # Kubelet owns container images and runtime state, including garbage collection.
+  # Keep this legacy-named helper limited to host logs and package-cache cleanup.
 
   # Kubelet owns active logs; retain every active UID, including quiet pods.
   python3 /scripts/node_pod_log_cleanup.py --host-root /host \
@@ -26,12 +22,8 @@ sweep_once() {
     find /host/var/log.hdd/containers -xtype l -print -delete 2>/dev/null || true
   fi
 
-  find /host/var/lib/rancher/k3s/agent/images -type f -name "*.tar" -mtime +7 -print -delete 2>/dev/null || true
-  find /host/var/lib/rancher/k3s/agent/containerd -maxdepth 1 -type f -mtime +7 -print -delete 2>/dev/null || true
-
   if [ -n "${usage}" ] && [ "${usage}" -ge "${EMERGENCY_USAGE_PERCENT}" ]; then
     # Emergency pass for rootfs pressure on SD-backed nodes.
-    chroot /host /bin/sh -c "crictl rmi --prune >/dev/null 2>&1 || true"
     chroot /host /bin/sh -c "journalctl --vacuum-size='${JOURNAL_MAX_SIZE}' >/dev/null 2>&1 || true"
     # Pod logs are handled only by the UID-aware check above.
     find /host/var/log -path /host/var/log/pods -prune -o -type f -name "*.gz" -mtime +"${LOG_RETENTION_DAYS}" -print -exec rm -f {} \; 2>/dev/null || true