`exo-eventlog-rotate-guard.sh`

32 lines
#!/bin/bash
# exo-eventlog-rotate-guard.sh — bound the unbounded exo active event log (TK-10236 ENOSPC class).
#
# WHY: exo's DiskEventLog._rotate() compresses events.bin -> events.<stamp>.bin.zst (keeps <=5
#      archives) ONLY on startup/close — never in-session. So a long-running exo (kept alive for
#      weeks) grows ~/.exo/event_log/api/events.bin without limit (3.2G over ~5 days, Aug 2026).
#      A 3.2G network-event log compresses to ~100M .zst, so a restart reclaims ~97%.
#
# WHAT: if events.bin > THRESHOLD_GB, kickstart exo-keepalive. exo dies + restarts; on restart it
#       sees the stale events.bin and rotates+compresses it, then starts a fresh empty log.
#       NO manual file deletion — exo's own rotation does it safely. Reversible: it's a restart.
#
# COST of firing: one brief exo-cluster restart (in-flight inference drops, tunnel blips ~seconds).
#       The exo-cluster-watchdog already restarts exo routinely when down, so this is well-tolerated.
#
# Tune THRESHOLD_GB up to restart less often (default 2.0G ≈ every ~3 days at current growth).
set -uo pipefail
LOG="$HOME/.exo/event_log/api/events.bin"
THRESHOLD_GB="${THRESHOLD_GB:-8.0}"  # raised 2.0->8.0 2026-08-26 TK-10707: let the wedged follower finish converging (idx 873K->6M) without the hourly restart resetting it. REVERT to 2.0 once converged.
STAMP="$(date -u +%FT%TZ)"
[ -f "$LOG" ] || { echo "$STAMP no events.bin — nothing to do"; exit 0; }
sz=$(stat -f '%z' "$LOG" 2>/dev/null || echo 0)
human=$(awk -v b="$sz" 'BEGIN{printf "%.2fG", b/1073741824}')
thr=$(awk -v g="$THRESHOLD_GB" 'BEGIN{printf "%d", g*1073741824}')
if [ "$sz" -gt "$thr" ]; then
  echo "$STAMP events.bin ${human} > ${THRESHOLD_GB}G — restarting exo to trigger native rotation"
  launchctl kickstart -k "gui/$(id -u)/com.steve.exo-keepalive"
  echo "$STAMP kickstarted com.steve.exo-keepalive (exo will rotate events.bin -> .zst on restart)"
else
  echo "$STAMP events.bin ${human} <= ${THRESHOLD_GB}G — ok"
fi