macOS Performance Monitor

Vitals

Read-only macOS performance inspection: a deterministic CLI gathers the numbers, Interpretation.md turns them into a diagnosis instead of a data dump.

Customization

Before executing, check for user customizations at: ~/.claude/LIFEOS/USER/CUSTOMIZATIONS/SKILLS/Vitals/

If this directory exists, load and apply any PREFERENCES.md, configurations, or resources found there. These override default behavior. If the directory does not exist, proceed with skill defaults.

Voice Notification

When executing a workflow, do BOTH:

1. Send voice notification:

   curl -s -X POST http://localhost:31337/notify \
     -H "Content-Type: application/json" \
     -d '{"message": "Running WORKFLOWNAME in Vitals"}' \
     > /dev/null 2>&1 &
   

2. Output text notification: Running **WorkflowName** in **Vitals**...

Workflow Routing

Workflow	Trigger	File
HealthCheck	"how's my system", "check my mac", "system health"	Workflows/HealthCheck.md
FindCulprit	"mac is slow", "what's eating CPU/GPU/memory", "fans loud"	Workflows/FindCulprit.md
DeepDiagnosis	"full diagnosis", "deep check", chronic/recurring slowness	Workflows/DeepDiagnosis.md

Interpretation reference (load with any workflow): Interpretation.md — thresholds, known-process table, diagnosis shape.

Quick Reference

bun ~/.claude/skills/Vitals/Tools/Vitals.ts check     # fast snapshot (<1s)
bun ~/.claude/skills/Vitals/Tools/Vitals.ts hogs      # live per-process CPU/energy (~3s)
bun ~/.claude/skills/Vitals/Tools/Vitals.ts full      # everything (~4s)
# also: gpu · memory · disk · thermal · startup · --json · --top N

The tool is read-only by contract: it never kills, renices, unloads, or writes system state. Remediation is recommended to the operator, never executed by the skill.

Examples

Example 1: Something feels slow

User: "My Mac is dragging, what's going on?"
→ FindCulprit: run hogs + check + gpu
→ Cross-reference Interpretation.md (is the hog a known background process?)
→ Report: named culprit with two evidence classes, recommended action for approval

Example 2: Routine check

User: "How's my system doing?"
→ HealthCheck: run check
→ One-screen verdict: 🟢/🟡/🔴 per subsystem, abnormalities flagged against thresholds

Example 3: Fans blasting

User: "Why are my fans so loud?"
→ FindCulprit: thermal + hogs + gpu
→ If kernel_task is the top consumer: that IS the cooling response — report the thermal cause, not the process

Gotchas

• top's first sample is garbage — CPU% is cumulative since boot. The tool always runs -l 2 and parses the SECOND sample; never "optimize" this away.
• ps -m does not reliably sort by memory on modern macOS (observed mis-ordering by GB-scale RSS). The tool sorts in code; keep it that way.
• kernel_task high CPU is the thermal system working, not a runaway process — it pins cores to force cooling. Never recommend killing it.
• macOS "used" memory is not a problem signal. Free RAM is deliberately spent on cache; kern.memorystatus_vm_pressure_level (1 normal / 2 warning / 4 critical) and live swapouts are the real signals.
• powermetrics requires sudo and runs forever without -n. It's the only accurate source of per-process energy and GPU power — keep it as the optional consent-gated leg in DeepDiagnosis.
• top -l truncates command names (~16 chars) regardless of COLUMNS in logging mode. Use the PID to identify the full process (ps -p <pid> -o comm=).
• GPU via ioreg -c IOAccelerator works without sudo on Apple Silicon (Device Utilization % in PerformanceStatistics) but is not guaranteed on every GPU — the tool degrades to an explicit unavailable-message, never a silent blank.
• pmset -g therm may not report CPU_Speed_Limit on desktops — the tool treats a missing key as not-throttled rather than erroring.