2026-10-19, Jacks-iMac. Brandon said the other Mac creates sub chats, runs things in parallel, has better gates, and took skills off, and told me to find out. He was right on all four. Every one of those capabilities exists on this machine. Some are switched off, one is gated by policy, and three have simply never been used, which is a different and more awkward answer.
Two separate removals happened, by two different mechanisms.
LOCAL ~/.claude/skills 179 active skill directories
~/.claude/skills_disabled 26 switched off, with toggle.py to flip them back
24 of those auto-disabled 2026-07-17
reason recorded in MANIFEST.json: "unused+unreferenced"
ACCOUNT ~/.claude/skills/.trash 13 add-ons removed by the 2026-10-17 claude.ai sync
~/.claude/skills/synced 9 add-ons remain
The thirteen in the trash include deep-research and mcp-builder. That first one matters for his second claim, because deep-research was the parallel-subagent skill. Its own description says it coordinates research subagents and should be used only where a tool for spawning subagents exists, and its procedure spawns research subagents in parallel then a report writer. So the clean-up I verified as finished removed a parallel-research capability, and I reported the clean-up without noticing what came off with it.
The automatic fleet belongs to this machine, and it has been running all day.
41 com.drwu.* plists on disk, 40 loaded in launchctl
PIDs observed at measurement: n8n 914 reveal-server 897
claude-gpu-watch 905 arena-throttle-proxy 885
clip-sync 891
jobs include: claude-config-sync idea-miner conductor-rebuild
substack-hourly substack-daily etsy-delivery-audit
recents-rebuild transcript-vault usage-monitor-sweep
stem-map-refresher phone-chat-index disk-space-guardian
plus one crontab entry at 07:00 daily for idea_scanner.py
The n8n orchestrator held PID 914 at the moment of measurement. So the automatic fleet exists on this machine and was working through every turn of this conversation.
The MacBook authors gates and ships them back through iCloud, and several of the gates that refused me today arrived from it.
| Gate shipped back from the MacBook | Bundle it arrived in |
|---|---|
completion_stop_gate.py | skill-sync-2026-09-06-completion-controller |
no_idle_stop_gate.py, stop_gate.py, quit_gate.py | skill-sync-2026-06-05-no-quit-shift |
notation_evidence_gate.py, verdict_gate.py, gpif_share_gate.py | skill-sync-2026-09-06-notation-evidence-gate |
drum_reality_gate.py, impossible_gate.py | zappa-macbook-return-2026-08-31, kilgore-macbook-return-2026-08-13 |
The stop gate that refused four of my attempts to end a turn today is one of those imports. So the gate-authoring habit is a MacBook export, and this Mac is the consumer.
Nothing is missing here. One route is deliberately gated, and I wrongly reported that as the whole picture.
BLOCK 1 ~/CLAUDE.md, line 13 "Keep one active Claude session. Do not launch subagents, parallel chats, /loop, keepalive work, or browser automation unless Brandon authorizes that specific run." BLOCK 2 ~/.claude/hooks/opus_spawn_gate.py PreToolUse hook, matcher "Agent|Task" Hard-denies every sub-agent spawn until ~/.claude/.opus_ok is touched. One-shot marker, valid 15 minutes, consumed on use.
disableAgentView, worktree, disableWorkflows, enableWorkflows and teammateMode are all unset, the env block is empty so the concurrent-subagent cap stands at its default of 20, and permissions carries no Agent or Task rule.| Documented route | State on this Mac |
|---|---|
| Agent / Task tool subagents, up to 20 at once | gated by opus_spawn_gate.py and the CLAUDE.md clause |
/batch <instruction>, splits work across 5 to 30 subagents, each in its own worktree | nothing disables it, never tried here |
claude agents, Agent view, background sessions watched from one screen | disableAgentView unset, ~/.claude/agents absent, never used |
| Worktrees, separate CLI sessions in isolated checkouts | worktree unset, ~/.claude/worktrees absent, never used |
| Agent teams, shared task list and messaging | experimental and off by default upstream |
The absence of ~/.claude/agents and ~/.claude/worktrees is the evidence that those two routes have never run here at all. So "blocked" was the wrong word for three of the five: unused is the right one, and the difference between those two words is the thing Brandon has been pointing at.
CLAUDE.md is per-machine and is never written over, because it hardcodes the source Mac's home, drives and anchors. So the MacBook runs its own CLAUDE.md, which need not carry that one-session clause. The capability gap between the two Macs is most likely a single paragraph in a file that sync refuses to copy.| Skill | State | What it does | Still runnable |
|---|---|---|---|
fleet-dashboard | off 2026-07-17 | Shows what every part of the fleet is doing right now: the launchd jobs, the stall-scan cron, voice routing, n8n state, CLI session count. Renders one Markdown dashboard into Brandon-Vault. | Yes. Its script is intact at ~/Projects/fleet-dashboard/build_dashboard.py, 4,940 bytes. |
system-pulse | off 2026-07-17 | Ranks every skill by real usage from the session transcripts, flags skills older than the memory rules that govern them, and judges whether the connectors and newer systems are being used as much as they should. Powers a daily adoption nudge. | Partly. Its build scripts are intact, but deploy.sh pushes to Vercel through the CLI, which is forbidden from this Mac. |
Both carry the same recorded reason as the other twenty-two: unused+unreferenced. Neither was judged on what it was for.
Every one of these is reversible, and toggle.py moves directories rather than deleting them.
# see the fleet, without changing any policy
python3 ~/.claude/skills_disabled/toggle.py enable fleet-dashboard
python3 ~/Projects/fleet-dashboard/build_dashboard.py
# rank skills and connector adoption; do NOT run its deploy.sh on this Mac
python3 ~/.claude/skills_disabled/toggle.py enable system-pulse
python3 ~/.claude/skills/system-pulse/build_system_pulse.py
# restore the parallel-research skill from the trash
cp -R ~/.claude/skills/.trash/1792272791312-27828-YYTbyK/deep-research \
~/.claude/skills/deep-research
# allow one sub-agent spawn, 15 minutes, consumed on use
touch ~/.claude/.opus_ok
The one lever that is not a command is the first block: the one-session clause in ~/CLAUDE.md is Brandon's standing instruction, so sub chats stay off here until he changes that line. Nothing on this page was switched on while finding this out.