Mac resource triage

Mac Vibe Triage

Mac first aid for AI coding leftovers. Triage resource bloat left behind by vibe coding, local agents, MCP servers, dev tools, and leaked browser helpers. Find the evidence first, then kill only the vetted (user-greenlit) PIDs.

This skill is user-invocable only. Never auto-run.

Safety baseline (universal, never relax)

1. Never kill a root-owned process.
2. Never kill anything owned by another user.
3. Never kill anything under /System/Library/, /usr/libexec/, or matching com.apple.*.
4. Never kill kernel_task, WindowServer, launchd, loginwindow, coreaudiod.
5. Never kill anything in the current agent PID chain (walk via ps -o ppid= -p $$).
6. Never rewrite JSON, delete LaunchAgents, uninstall apps/packages, or run launchctl bootout without a separate explicit user confirmation for that exact action.

The safety baseline lives in this skill and is enforced again by scripts/kill-pid-list.sh. The persistent profile can add user preferences; it can never weaken these rules.

Workflow

1. Read persistent profile at $HOME/.config/mac-vibe-triage/profile.md. Use references/profile-template.md as the schema. If the file is malformed or missing, use an empty profile and require explicit confirmation for every candidate.
2. Run diagnostics — see references/diagnostics.md. Each step has a "Done iff" criterion; satisfy it before moving on.
3. Map playbooks — see references/playbooks.md. If evidence matches a known incident shape, name the playbook and its evidence; if no playbook fits, say "unknown / general resource bloat" rather than guessing.
4. Vet candidates. Every gathered candidate gets one of:
   • 🟢 matches an Always-green rule → surface as a low-risk candidate, still requiring this-run confirmation
   • 🔴 matches an Always-keep rule → drop from report entirely
   • 🟡 matches a Sit-down rule, or no rule → ask before placing in any kill batch
5. Run a vetted kill. Once the user greenlights a PID (or kill all green after seeing the exact PID list), exec scripts/kill-pid-list.sh <PID …>. Each PID is vetted individually before being added; SIGTERM → verify → SIGKILL is the runtime half.
6. Emit Triage Report — see references/triage-report.md. Include symptom, evidence, suspects, actions taken, decisions needed, and what remains unknown.
7. Emit post-run recommendations — see references/post-run-recommendations.md. Pick 3–5 items grounded in what was just vetted.
8. Close. Ask the user which (if any) of this run's decisions to memorize (k keep / s sit-down / t threshold / n no). Append to profile on confirmation, bump updated:.

Vetted-kill protocol (the leading concept)

A vetted kill = a SIGTERM/SIGKILL pair sent to a PID the user has explicitly greenlit in this session. Vetting is per-PID, per-run; a PID vetted last run is not automatically vetted this run.

A vet cycle is: candidate identified → surfaced to user → user says kill <PID> (or batched kill all green) → only then SIGTERM. If the user says no or stays silent for >1 turn, drop the candidate.

scripts/kill-pid-list.sh enforces the runtime half (refuses non-numeric, other-user/root, system paths, com.apple.*, protected process names, the current agent PID chain, claude-agent-sdk-darwin-arm64/bin/claude, and ssh-agent). The protocol above enforces the user-greenlit half.

Helper scripts

Script	Behavior
scripts/kill-pid-list.sh	Runtime half of a vetted kill — SIGTERM → verify → SIGKILL with per-PID verify loop
scripts/inspect-pid-tree.sh	Diagnostic — walks ppid chain upward for a PID, read-only
scripts/find-high-freq-configs.sh	Diagnostic — scans settings JSON for sub-threshold polling intervals, read-only

Only the first writes (sends signals). The rest are read-only — even if the rest are run from the same shell as a vetted kill.

Persistent profile

Schema for $HOME/.config/mac-vibe-triage/profile.md:

• Always-green — patterns the user has confirmed as low-risk candidates; still require this-run confirmation.
• Always-keep — patterns never to surface (e.g. cmux, current Claude Code PID chain).
• Sit-down-and-ask — surface but require per-run user input (e.g. Token Monitor).
• Custom thresholds, Output preferences, Anti-rules (user-pinned) — user preferences and learned negations.

Failure mode: profile unreadable / malformed → use an empty profile and ask for every candidate. Never infer Always-green from an unknown or broken profile.

Anti-patterns (skill-side, not safety)

• ❌ Concluding "Electron == 耗电大户" without TIME-based evidence.
• ❌ Acting like a generic Mac cleanup tool. Stay in the vibe-coding/local-agent incident frame unless evidence proves otherwise.
• ❌ Recommending reboot as a first reaction for WindowServer drift.
• ❌ Trusting ps snapshot alone for "is X installed?" — cross-check /Applications + LaunchAgents per Step 5.
• ❌ Auto-rewriting settings.json — tell the user what to change; they edit unless they explicitly ask you to edit that exact file.
• ❌ Bundling this with /neat-freak-style cleanup (that skill is for skills + memory + docs, not processes).

The profile is for preferences, not universal safety. A profile entry that conflicts with the safety baseline is ignored and reported as invalid.

Edge cases

• ps returns zero user rows → "你名下当前没有任何用户进程，跑空。" Done.
• User pastes a 50-PID list → truncate to first 20 in the report table, offer to show the rest.
• Process respawns within 3 s with a new PID → flag warn: PID <X> respawned as <Y> and ask before re-killing.
• User asks to kill current Claude Code → refuse; suggest /exit + claude restart.
• User says "nuke everything" → never comply silently; enumerate every category and require exact kill <PID> commands or kill all green after showing the exact green PID list.