it-ram-doctor

The snapshot only reads; step 4 kills, and only on the owner's word.

When memory is short

"13 of 16 GB used" and the swap size do not show load: macOS keeps memory busy with cache and compression, and swap does not shrink until a reboot. Memory is short when any of these holds:

• pressure: sysctl kern.memorystatus_vm_pressure_level is 4 or stays at 2 for long (threshold: on Apple silicon 2 appears under ordinary compression); memory_pressure | tail -1 shows less than 25–30% free;
• Swapouts in vm_stat grow between two readings a minute apart;
• swap grows three readings in a row.

No signs: clean only on a direct request. Measure a process by top -o mem or footprint; ps understates the compressed part.

Procedure

1. Snapshot. From the skill root:

   bash scripts/tails-macos.sh                                                     # local
   ssh -o BatchMode=yes -o ConnectTimeout=8 <alias> 'bash -s' < scripts/tails-<os>.sh   # remote
   

   Keep the script's summary as the "before" and take the top of top -o mem: the table omits launchd services (on Linux, .service units), system processes, Claude.app and children of active sessions, and one live job (a local whisper run, 9 GB) can outweigh all the tails. Done when you hold the table, the summary and the top of top.

2. Decide. The "Action" column sets what to do with a class (the script prints the Russian labels shown):

   Class	What it is	Action
   claude-сессия (claude session)	Claude Code CLI, bundled with Desktop or standalone	kill if a tail
   хвост закрытого приложения (closed-app tail)	processes of X.app with PPID 1 while X is not running, and everything under them	kill
   сирота (orphan)	a worker with PPID 1: omp/bun, Chrome for Testing, npx/MCP, grok/codex; duplicates marked	kill
   dev-сервер (dev server)	http.server, vite, npm run dev without a parent, older than a day	ask
   зомби (zombie)	STAT Z older than a minute	its parent reaps it: name the parent

   The script does not know session activity: match the id from the row against list_sessions. A tail is a session with lastActivityAt older than 12 h, not isRunning, with no work going on. Id "не найдена" (not found) and sessions on a remote machine (list_sessions does not know them): ask. Done when every row has an action and every claude row has a verdict: tail, live or ask.

3. Show. The table to the owner: row, action, what dies with it. Before killing a claude process, name its dev servers and ports from the "dies with it" column. "Kill" and "clean up" go straight to step 4; "slow" and "swap" wait for the word. Done when the owner sees the table and the word is received or not needed.

4. Kill. Only rows that passed step 2, in bash with an array, with a check and a follow-up KILL:

   # bash -c: in zsh `kill $pids` with a string silently kills nothing
   bash -c 'pids=(111 222 333); kill -TERM "${pids[@]}"; sleep 4
     for p in "${pids[@]}"; do kill -0 "$p" 2>/dev/null && kill -KILL "$p"; done'
   

   • Idle claude sessions ignore SIGTERM: the KILL pass is required.
   • A dev server started with & inside a session outlives it as an orphan: check the port with lsof -iTCP:<port> -sTCP:LISTEN.

   Done when no pid from the array answers kill -0 and the ports of killed dev servers are free.

5. Verify. Run the script again. Report to the owner: pressure and the candidates' RSS before and after; swap does not shrink until a reboot, so it is not the measure. A killed claude session stays in the app's list and comes back on the next message. Killing the CLI does not free the memory of Desktop's renderer and main processes: that needs an app restart, on the owner's word, after checking it will not interrupt needed work, with another reading. Done when the run shows no killed candidates and "before" and "after" are named.

Off limits

• the current session, active and isRunning claude sessions;
• launchd services;
• Claude.app.