Mac mini audit

1. fleet-medic (local-bin, every 5 min). Its settings revive any session it has ever seen and poke every idle one. A reboot looks to it like 83 sessions dying, so it brought them back 3 at a time for two hours. Each revival is a paid turn. It also loops on one dead window, about 14 times today.
2. openclaw (~/openclaw, a March clone of someone else’s repo). Its voice API answers anyone on the home network or Tailscale with no password, and a paid ElevenLabs key is written in its code. Its dashboard accepts the built-in default password. “voicebox” has crash-looped every 30 s since its folder was deleted, and its error log is 168 MB. One piece is used by Agent Smith: a local Claude proxy. That moves into local-bin; it doesn’t get deleted.
3. Check-in pokes (watcher-doctor + nudge). Every 3 minutes, watcher-doctor re-arms a watcher for each of 19 Driver lanes. The watchers poke after 10 idle minutes and never expire. Two sessions got 82 of the 142 pokes. Each check-in asks for a model ruling. The August off-switch for this was lifted on 08-31.
4. ccswitch respawns sessions that hit the usage limit onto other accounts, both after each turn and every 30 seconds. fleet-medic competes with it for the same accounts, with 202 refused moves.
5. Codex is registered for every session in ~/.claude.json. It was last used Sept 10.
6. Agent Smith keeps its Temporal Cloud jobs, which is fine. But its task watcher revives Drivers without checking pause/stop switches, and it reconnects to Temporal about every 2 seconds, which is what fills its log.
7. WhatsApp runs from a third-party clone inside ~/code/PropFlow. Only the personal bridge is paired. The other two print QR codes in a loop (605 MB log).
8. Security. Also on: old-style screen sharing, Remote Apple Events and the AirPlay receiver. FileVault is off and auto-login is on, which is what makes unattended restarts work. Tailscale SSH is off, and nothing is published to the internet except the 4 intended PropFlow webhook addresses.
9. Disk has 333 GB free, so space isn’t the problem; clutter is.
   • 22 copies of the docs site (32 GB) and Python cache (19.6 GB).
   • Claude desktop VM image (11.4 GB), old build copies (8.3 GB), Trash (5.3 GB) and iOS simulators (5.3 GB).
   • 85 worktrees pointing at missing folders, plus 133 orphaned ones parked since September.
   • Logs that nothing ever trims.
10. Also not PropFlow: Blender and a Python AI stack (torch, mlx), Codex (about 9 GB of state), ChatGPT and Figma, a cookie-export browser extension, Google updaters and a Tesla display project.

Two read-only scripts are written:

• a snapshot to take before the restart;
• a PASS/FAIL check to run 3 minutes after.

The order matters:

1. Switch fleet-medic off first, or it re-creates the storm.
2. Take the snapshot.
3. Restart.
4. Run the check.

Passing means: every service above is up, no more than 5 sessions came back, there are 0 revivals, and the openclaw crash loop is gone.

/Users/miniclaw
├─ code/                          368 GB
│  ├─ PropFlow/                   350 GB  [PropFlow] 758 worktrees across 14 repos
│  │  ├─ propflowai/              108 GB  app checkout; 74 GB of it is nested worktrees, .git 18.8 GB
│  │  ├─ .smith-review-docs/       32 GB  22 copies of the docs site, keep 2
│  │  ├─ propflow-docs/            15 GB  docs checkout + 77 worktrees
│  │  ├─ appfolio-browser-agent/  9.5 GB  8.7 GB of it is nested worktrees
│  │  ├─ next-old-1e*/            8.3 GB  [STALE] old builds, Sep 30
│  │  ├─ agent-smith/ 4.0 GB · propflowai-smith/ 3.0 GB · brand-assets/ 3.4 GB
│  │  ├─ ~35 docs worktrees ≈ 35 GB (Oct 3–6) · ~130 ingest worktrees (live — leave)
│  │  ├─ whatsapp-mcp/            [NOT PropFlow] third-party clone, runs the bridges
│  │  └─ 1,200 loose *.log files   [STALE] command output, 71 MB
│  ├─ TheGarzaWedding/           7.9 GB  [NOT PropFlow]
│  ├─ tmp/ 5.9 GB · Mofu/ 1.2 GB [PropFlow] · datasets/ 1.1 GB
│  └─ tesla-second-display/, graphify/   [NOT PropFlow]
├─ .cache/                        52 GB  uv 19.6 [CACHE] · shared node_modules 18.4 [KEEP] · gh 5.0 · whisper 4.2
├─ Library/                       51 GB
│  ├─ Claude desktop VM image    11.4 GB  [NOT PropFlow] Jul 6
│  ├─ Chrome 6.9 GB · Codex updater cache 6.0 GB [CACHE] · iOS simulators 5.3 GB [STALE]
│  └─ Logs/ 2.2 GB                never rotated
├─ .claude/                       45 GB  [PropFlow] dotclaude
│  ├─ projects/                  25.3 GB  session transcripts (needed for resume)
│  ├─ jobs/ 8.3 GB · scripts/ 8.1 GB (venvs and outputs, should be ~0)
│  └─ ccswitch/ 0.9 GB            config for 398 sessions, ~50 alive
├─ .Trash/                       5.3 GB  installers — empty it
├─ resources/ 5.4 GB · openclaw/ 1.1 GB · .codex/ 1.2 GB   [NOT PropFlow]
├─ .wt-trash-parked/              133 orphaned worktrees, Sep 16–17
└─ 17,052 .claude.json.tmp.* files  1.2 GB  [STALE] failed saves, Jul–Oct 7