01 / collectorRead pressure and headroom locally
The Rust core gathers memory, swap, disk, and process state on the Mac it is protecting.
Rust · sysinfo · macOSPerf Pulse00
Perf Pulse
Crash Guard for developer Macs.
Project film · Out of memory
Mid-screen-share, macOS runs out of application memory and the call freezes. Rewind nine minutes: inside the 18 GB, one app never stops growing. Crash Guard forecasts the freeze, warns first, and Stop safely re-checks the process before it acts. 14:39:07, same second, still on the call. Score: “Tues” by Hope Atina.
Perf Pulse came from a failure I kept hitting. Agents, builds, and desktop apps would eat memory or disk until the Mac stopped responding, while Activity Monitor sat there unopened. Crash Guard turns that late diagnosis into an always-on warning, a local dashboard, and a deliberate next move.
CoreRust
ProtectionMemory + disk
RuntimePer-user launchdno root daemon · no system extension
Installbrew install perf-pulsepublic tap · pinned SHA-256 · real binaries
brew install hopeatina/perf-pulse/perf-pulsepublic tap · Apple Silicon + Intel · protection remains opt-in01
The consequential problem
The warning arrived after the damage
Activity Monitor can show a problem. It does not watch for the moment before the machine goes down.
The failure rarely begins with one obvious process. Agent sessions accumulate, a build spikes, swap grows, or generated artifacts quietly consume the remaining writable disk. By the time the Mac becomes visibly slow, opening another inspection tool is already late.
A useful guard had to run continuously, explain which resource was approaching danger, and lead straight to a safe response. No root access, and no new heavyweight service.
- Warn while there is still time to act.
- Open the local incident view. Not an editor, not an opaque script.
- Never delete files automatically; keep destructive action explicit.
02
What I saw
A warning is a doorway
Detection only becomes useful when the next move is attached.
Crash Guard connects three moments that system utilities usually separate: the trend before failure, a concise native alert, and the exact live dashboard for that incident. Developers can tune the guard to their machine, inspect bounded storage hotspots, and decide whether to stop a process with its identity revalidated at action time.
03
The decision that changed the system
Always on, still local-first
Use a per-user agent for protection and a local dashboard for judgment.
A headless launchd agent samples memory pressure and writable-disk headroom even when the dashboard is closed. When a threshold is crossed, a native alert can reopen the dedicated Crash Guard tab on the live localhost server. The person at the keyboard decides what to stop, which thresholds to change, and when to turn protection off. The monitor doesn't.
01Watch headroom
A per-user agent samples memory pressure and writable disk locally.
02Detect trajectory
Thresholds and disk forecasting surface risk before the machine becomes unresponsive.
03Open the incident
The alert leads to the dedicated live Crash Guard dashboard with the relevant context.
04Intervene deliberately
The user tunes protection, inspects storage, or requests an identity-checked process stop.
04
System anatomy / rationale / surfaces
One guard, one operating picture
Continuous detection and human-controlled action share a small local-first architecture.
The guard can keep watch without an open browser, while the dashboard appears only when the developer needs context or control. CLI, TUI, and overview surfaces read from the same underlying system state.
Decision anatomy / select a knot
Pressure did not decorate the architecture. It determined it.
Before / tension
Manual inspection begins only after the developer notices slowdown, low disk, or a frozen machine.
The knot / decision
Run Crash Guard as an opt-in per-user launchd agent with tunable memory and disk thresholds.
After / consequence
Protection persists across dashboard sessions without a root daemon or macOS system extension.
02 / guardModel risk before failure
Tunable thresholds, trend checks, and disk forecasting distinguish current load from a deteriorating trajectory.
Tokio · local rules03 / incidentBind the alert to the live context
Native notifications can open the dedicated Crash Guard tab on the exact localhost dashboard serving that Mac.
launchd · Axum · native alerts04 / authorityKeep intervention in human hands
Configuration is explicit, storage diagnostics stay read-only, and stop requests revalidate process identity before signaling.
PID + name + start timeArchitecture rationale / pressure → choice → consequence
PressureDecisionWhy it mattered
A snapshot can miss the path toward a crashTrack pressure, growth, and remaining disk headroom
The product can warn while the developer still has room to respond.
A generic notification creates another huntOpen the dedicated live incident surface
Risk, controls, and current process context arrive together.
A stale PID can point at a different processValidate PID, name, and start time before a stop request
Convenient action does not erase the safety boundary.
Surface atlas
The system becomes tangible through the places people encounter and use it.
Crash Guard
Persistent memory and disk protection with tunable thresholds and native alerts.
Local dashboard
The exact live operating picture for protection status, recent incidents, and intervention.
CLI + TUI
Fast checks and sustained inspection stay where developers already work.
Homebrew
Installation becomes the first coherent product interaction, not a README obstacle.
Core
Rust · Tokio · sysinfo
Runtime
launchd · Axum · local API
Delivery
Homebrew · native alerts · optional webhooks
Technology atlas / select a tool
The tools behind the decisions.
Select a tool to see the role it plays in this system.
- Core / Perf Pulse
Rust
A memory-safe, compact core keeps continuous collection predictable on the machine it is protecting.
05
Authentic proof
Crash Guard / one local operating picture
The notification opens the exact live dashboard developers need: current risk, tunable thresholds, recent incidents, and identity-checked stop controls in one place.
Early warning / Crash Guard
Protection state, threshold controls, incident history, and the next safe action stay in one focused local surface.
Immediate / CLI
The shortest path from machine state to an actionable explanation stays in the terminal and can feed another tool when needed.
Sustained / TUI
The same core signal earns a persistent terminal surface when the operator needs to watch change over time.
Reversible action / Meeting Mode
The interface makes the intervention, resources freed, paused processes, and restore path visible in one state.
06
What changed in my operating model
The interface begins at the interruption
A warning earns trust when it carries the user to a clear choice.
Crash Guard changed the product from something a developer remembers to open into protection that can notice first. The hard part was not adding another alert; it was joining early detection, useful context, configurable sensitivity, and carefully bounded intervention without overstating what a signal request can guarantee.
- Design the alert and destination as one interaction.
- Expose sensitivity without making safe defaults feel unfinished.
- Keep automation ahead of failure and behind user authority.