Headless Mac mini disconnects SSH / distant periods after macOS 26.4 replace — Darkish Wake Thermal Emergency triggered by mediaanalysisd in Energy Nap

After tracing this by pmset -g log, log present, and SMC thermal occasions, this is the total mechanism and a layered repair. Sharing as a result of I could not discover a full writeup elsewhere.

Root trigger: DarkWake has a conservative fan coverage + an SMC-firmware-level kill swap

macOS has two “awake” states with very completely different thermal conduct:

When DarkWake workload exceeds the envelope, the SMC firmware (not user-space powerd) emits the kill sign:

  DarkWake → mediaanalysisd / TM / and so on. runs → temp climbs to degree 4 →
  SMC Thermal stress requested Sleep throughout energy nap →
  compelled sleep (bypasses each user-space PreventUserIdleSystemSleep assertion)

This can be a firmware-level rule. caffeinate -i, app-held sleep assertions — none of them cease it.

FullWake would not have this kill swap. Similar workload → followers ramp to ~5000 RPM → temp stabilizes at degree 2–3 → sustained with out situation.

Why it began after 26.4 particularly

Each main macOS replace (and a few level releases) resets mediaanalysisd‘s evaluation state and will bundle new ML fashions that want full re-inference throughout the Pictures library:

• Video decoding (AVD {hardware} decoder)
• Neural Engine inference (face / scene / object detection)
• Metallic GPU shaders
• Heavy disk I/O throughout Pictures library

Empirically on my system:

The post-update workload is giant sufficient {that a} typical Energy Nap DarkWake window can not full the queued job earlier than crossing thermal degree 4.

Why remote-control instruments disconnect collectively (the second-order impact)

coreaudiod holding PreventUserIdleSystemSleep retains DarkWake alive. Thermal degree 4 → compelled sleep. The compelled sleep is difficult:

• TCPKeepAlive=lively would not assist right here — the connection’s keep-alive state is within the NIC offload desk, however solely Apple-blessed providers (APNs, mDNS, and so on.) register sockets there. Consumer-space sockets from SSH, Tailscale, customized
  remote-agent tooling are NOT within the offload checklist, so the compelled sleep drops them.
• After sleep, these connections are lifeless. Reconnect occurs solely when one thing else wakes the Mac — usually an incidental APNs push from an unrelated service (iMessage, iCloud, Discover My), which DOES wake the host as a result of its
  socket IS within the offload checklist.

That is why each distant instrument dies concurrently and reconnects on the identical second minutes later.

Diagnostic instructions

  # 1. Rely thermal-emergency sleeps over the past 7 days
  log present --predicate 'eventMessage CONTAINS "Thermal Emergency" 
    OR eventMessage CONTAINS "Thermal stress requested Sleep"' 
    --last 7d --info 2>/dev/null | grep -c "^2026-"

  # 2. Verify mediaanalysisd is the busy course of
  log present --predicate 'course of == "mediaanalysisd" 
    AND eventMessage CONTAINS "job progress"' 
    --last 24h 2>/dev/null | grep -c "^2026-"
  # Baseline: 0–50. Lively backlog: 3000+.

  # 3. Who's holding sleep open proper now?
  pmset -g assertions | grep -E "PreventUserIdleSystemSleep|coreaudiod"

  # 4. Hint thermal degree escalation
  log present --predicate 'subsystem == "com.apple.thermalmonitor" 
    OR eventMessage CONTAINS "thermalLevel"' 
    --last 1h --info 2>/dev/null | grep -v ""

Repair — three ranges

Stage 1: Disable Energy Nap (minimal disruption)

  sudo pmset -c powernap 0

Stops mediaanalysisd, Time Machine, iCloud sync from operating throughout DarkWake. The work is not misplaced — mediaanalysisd will catch up throughout FullWake when the fan curve can maintain the load.

Commerce-off: these duties will not run whereas the Mac is in any other case asleep. For a Mac you actively use in the course of the day, the catch-up window is adequate.

Stage 2: By no means sleep on AC (advisable for headless / remote-controlled Mac mini)

  sudo pmset -c sleep 0 powernap 0 displaysleep 10

No sleep → no DarkWake → no thermal kill swap. Show nonetheless sleeps (saves the ten–30 W from the monitor). Mac mini idle FullWake ≈ 5 W; the delta vs sleep is small.

That is the cleanest repair for those who depend on always-on distant entry. It additionally by the way fixes a separate downside: most remote-control instruments can not wake a sleeping Mac through incoming visitors, as a result of their sockets aren’t registered for the
NIC keepalive offload. With sleep 0, no wake required — your message arrives, Mac is awake, processed immediately.

Stage 3: launchctl disable system/com.apple.mediaanalysisd (don’t do that)

Breaks Visible Look Up, Dwell Textual content, Pictures search, face recognition, and Apple Intelligence options. Apple respawns the service anyway. Skip.

Lengthy-term: monitor when the backlog clears, then revert

If you don’t need sleep 0 completely, arrange a each day verify through launchd that counts mediaanalysisd job progress occasions. When rely is < 200/day for 3 consecutive days, the backlog is completed and you may revert.

~/scripts/check-mediaanalysis-backlog.sh:

  #!/bin/bash
  set -uo pipefail
  LOG_DIR="$HOME/logs"
  STATE_FILE="$LOG_DIR/backlog-clear-streak"
  WATCH_LOG="$LOG_DIR/backlog-watch.log"
  PROGRESS_THRESHOLD=200
  REQUIRED_STREAK_DAYS=3
  mkdir -p "$LOG_DIR"
  log() { printf '[%s] %sn' "$(date '+%Y-%m-%d %H:%M:%S %Z')" "$*" >> "$WATCH_LOG"; }

  progress=$(/usr/bin/log present 
    --predicate 'course of == "mediaanalysisd" AND eventMessage CONTAINS "job progress"' 
    --last 24h 2>/dev/null | grep -cE '^2026-' || true)

  log "progress_24h=${progress} (threshold: < ${PROGRESS_THRESHOLD})"

  if [[ "$progress" -lt "$PROGRESS_THRESHOLD" ]]; then
      streak=$(cat "$STATE_FILE" 2>/dev/null || echo 0)
      streak=$((streak + 1))
      echo "$streak" > "$STATE_FILE"
      log "BACKLOG CLEAR — streak ${streak}/${REQUIRED_STREAK_DAYS}"
      if (( streak >= REQUIRED_STREAK_DAYS )); then
          log "REVERT TRIGGER"
          /usr/bin/osascript -e 'show notification "mediaanalysisd backlog cleared — revert pmset" with title "Mac mini" sound identify "Glass"'
          echo 0 > "$STATE_FILE"
      fi
  else
      echo 0 > "$STATE_FILE"
  fi

~/Library/LaunchAgents/com.YOU.check-mediaanalysis-backlog.plist:

  
  
  
  
      Labelcom.YOU.check-mediaanalysis-backlog
      ProgramArguments
      
          /bin/bash
          /Customers/YOU/scripts/check-mediaanalysis-backlog.sh
      
      StartCalendarInterval
      Hour22Minute0
  
  

Load:

  launchctl bootstrap "gui/$(id -u)" ~/Library/LaunchAgents/com.YOU.check-mediaanalysis-backlog.plist

When the notification fires, revert (use your pre-override values from pmset -g customized snapshot):

  sudo pmset -c sleep 1 powernap 1

Why “reset SMC” and different generic recommendation falls quick

• Reset SMC — would not assist as a result of the workload inflicting overheat is software program, not firmware
• Wait it out — true however takes 1–4 weeks for giant Pictures libraries
• killall mediaanalysisd — Apple respawns; you may be preventing it indefinitely
• Disable Energy Nap alone — addresses mediaanalysisd, however would not tackle the identical thermal entice when a distant consumer sends a heavy command (which additionally lands in DarkWake with out HID enter)

Stage 2 (sleep 0) addresses each the OS-induced backlog AND user-induced distant workload, which is why it is the best repair for headless / remote-controlled Mac minis.