diff --git a/arcane/home/honeypot-elk/analysis/filebeat.yml b/arcane/home/honeypot-elk/analysis/filebeat.yml
index 8aea6240e..978d961be 100644
--- a/arcane/home/honeypot-elk/analysis/filebeat.yml
+++ b/arcane/home/honeypot-elk/analysis/filebeat.yml
@@ -202,6 +202,17 @@ filebeat.inputs:
     # content-hash window can.
     id: zeek-logs
     file_identity.native: ~
+    # #3284: this directory accumulates ~13,700 hourly rotation files (2026-09-03
+    # onward) and nothing prunes them, so every restart made filestream open a
+    # harvester per file until the Go runtime hit its 10000-thread limit. The
+    # cap is the fix, not GODEBUG: one harvester per log type is plenty, because
+    # Zeek writes one file per type per hour and only the newest is still being
+    # appended to. Older rotations are finished by definition.
+    harvester_limit: 64
+    # 72h is ~3 days of hourly rotations per log type, comfortably past any
+    # backlog while keeping the scanner off the 13,700 accumulated files. A
+    # rotation older than this is finished; nothing appends to it again.
+    ignore_older: 72h
     paths:
       - /logs/zeek/*.log
     parsers:
@@ -224,6 +235,11 @@ filebeat.inputs:
     # identity is immune to it because it never hashes file content.
     id: zeek-proxy-logs
     file_identity.native: ~
+    # #3284: same cap as zeek-logs, same reason -- this glob matches 834
+    # accumulated hourly rotations and one harvester per file is what exhausts
+    # the runtime's thread limit on restart.
+    harvester_limit: 64
+    ignore_older: 72h
     paths:
       - /logs/zeek-proxy/*.log
     parsers:
diff --git a/arcane/home/honeypot-utilities/analysis/log-maintenance.sh b/arcane/home/honeypot-utilities/analysis/log-maintenance.sh
index 5b764b1c2..a61d58d57 100644
--- a/arcane/home/honeypot-utilities/analysis/log-maintenance.sh
+++ b/arcane/home/honeypot-utilities/analysis/log-maintenance.sh
@@ -136,6 +136,15 @@ while true; do
   # its mtime current, so -mmin never fires on it (same shape vps/
   # suricata-log-maintenance.sh uses for eve.json).
   find /logs/zeek-proxy -maxdepth 1 -name '*.log' -mmin "+${json_retention_min}" -print -delete 2>/dev/null || true
+  # #3284: /logs/zeek here is an sshfs mount of the VPS, deliberately read-only
+  # (fuse.sshfs ro in /etc/fstab), so a delete attempted from this container
+  # fails EROFS on every file. Confirmed live: 12,390 paths printed by this
+  # script's find, file count unchanged, touch and find -delete both returning
+  # "Read-only file system". The 2>/dev/null above hides exactly that error.
+  #
+  # Retention for these files therefore lives on the VPS, in
+  # vps/zeek-log-maintenance.sh, which runs where they are actually written.
+  # Do not re-add a find here -- it cannot work, and it fails quietly.
   # #2323 part 2: extracted-file-importer.py copies carved bytes into ES
   # and tracks what it has seen in state/extracted-files.json, so the disk
   # copy only needs to outlive that importer's lag (IMPORT_INTERVAL=60s,
diff --git a/vps/zeek-log-maintenance.sh b/vps/zeek-log-maintenance.sh
new file mode 100755
index 000000000..503b54630
--- /dev/null
+++ b/vps/zeek-log-maintenance.sh
@@ -0,0 +1,43 @@
+#!/bin/sh
+set -eu
+
+# Prune rotated Zeek logs on the VPS, where they are actually written.
+#
+# #3284: this is the root cause of both open ops issues, and it was invisible
+# from the homeserver. honeypot-elk's filebeat reads /logs/zeek over an sshfs
+# mount that is deliberately read-only (root@10.8.0.1:/opt/stacks/apiary/logs/zeek
+# -> /var/dockge/stacks/apiary/logs/zeek, fuse.sshfs ro in /etc/fstab), so
+# log-maintenance.sh on the homeserver prints the paths it wants to delete and
+# then silently fails every one of them -- its find carries 2>/dev/null || true,
+# which is why 12,390 "deleted" zeek paths appeared in that container's log
+# with the file count unchanged.
+#
+# Zeek rotates hourly by rename-and-reopen and never truncates in place, so a
+# rotated generation is finished the moment it is closed: -mmin only ever fires
+# on files nothing will append to again. The bare-name match also covers the
+# live file of a dead sensor, whose mtime freezes (same shape
+# suricata-log-maintenance.sh uses for eve.json).
+#
+# Consequence of not having this: 13,717 accumulated hourly rotations (16 GB),
+# and one filestream harvester per file on every hp-filebeat restart, which is
+# what drove the Go runtime past its 10000-thread limit. It also fed #3283 --
+# one Elasticsearch index per day per log type, 836 of the cluster's 1091
+# shards, several holding a single document.
+#
+# #261: default derives from the shared HONEYPOT_RETENTION_DAYS knob, same
+# ratio and reasoning as analysis/log-maintenance.sh's json_retention_min and
+# suricata-log-maintenance.sh's retention_min. Elasticsearch holds the
+# searchable history; this is only the on-disk copy, which needs to outlive
+# Filebeat's ingest lag by a comfortable margin, not last forever.
+
+retention_min="${RETENTION_MINUTES:-$(( ${HONEYPOT_RETENTION_DAYS:-30} * 1440 / 10 ))}"
+interval="${CHECK_INTERVAL:-3600}"
+start_delay="${START_DELAY:-60}"
+log_dir="${LOG_DIR:-/opt/stacks/apiary/logs/zeek}"
+
+sleep "$start_delay"
+
+while true; do
+  find "$log_dir" -maxdepth 1 -name '*.log' -mmin "+${retention_min}" -print -delete 2>/dev/null || true
+  sleep "$interval"
+done