# Sys Health

> >-

- **Type:** Skill
- **Install:** `agentstack add skill-phanghonghao-thu-awesome-skills-sys-health`
- **Verified:** Yes — security-reviewed for prompt injection and unsafe behavior
- **Seller:** [phanghonghao](https://agentstack.voostack.com/s/phanghonghao)
- **Installs:** 0
- **Category:** [Agent Skills](https://agentstack.voostack.com/c/agent-skills)
- **Latest version:** 0.1.0
- **License:** MIT
- **Upstream author:** [phanghonghao](https://github.com/phanghonghao)
- **Source:** https://github.com/phanghonghao/THU-Awesome-Skills/tree/main/skills/sys-health

## Install

```sh
agentstack add skill-phanghonghao-thu-awesome-skills-sys-health
```

Requires the [AgentStack CLI](https://agentstack.voostack.com/docs/cli). Works with Claude Code, Cursor, and any MCP-compatible agent.

## About

# sys-health

One-shot system health snapshot for this UniLab dev laptop. RL training here is
**CPU-only and runs inside WSL2**. A long run stresses three things, in order
of risk for a laptop:

1. **Heat** — sustained 100%-on-all-cores cooks a mobile CPU.
2. **SSD write endurance** — every `model_*.pt` checkpoint and tfevents line is
   a write; small `save_interval` × many iters = real wear.
3. **C: drive space** — the WSL2 ext4 filesystem is a `.vhdx` that lives on C:;
   if C: fills, the vhdx can't grow and training crashes mid-run.

This skill measures all three in one command (heat needs one-time setup, below).

## How to invoke

```bash
wsl -d Ubuntu -u u20174 -- bash -c \
  "tr -d '\r' /dev/null
```

- `tr -d '\r'` makes the script immune to CRLF a Windows editor may introduce.
- `2>/dev/null` mutes the WSL localhost-proxy Chinese-garbage warning.
- The script is **read-only and idempotent** — safe to run anytime, including
  mid-training.

After it prints, give the user a **one-paragraph health verdict + the single
most important action** (see Interpretation). Do not just dump the raw output.

## Sections & how to read them

| Section | Healthy | Watch / Act |
|---|---|---|
| Load (1/5/15 min) | below 16 sustained | 15-min load pinned near/over 16 = box fully committed; nothing else runs well |
| MEMORY & SWAP | Swap `0B` used, available > 2GB | Swap in use = training spilling to SSD (slow + wear). Low available RAM → reduce `algo.num_envs` |
| DISK USAGE | Use%  85% → clean up; this ext4 is the WSL vhdx |
| TOP CPU PROCESSES | see what's running | a `python` at ~`cores×100%` for a long ELAPSED = a UniLab run is active (note the run path under TRAINING WRITE ACTIVITY) |
| TRAINING WRITE ACTIVITY | few files in last 10 min | every listed file = an SSD write; checkpoint `size × save_interval × iters` = cumulative wear |
| NVMe SMART | Data Units Written low, temp  60GB | C: < 40GB risky — vhdx + pagefile live here and grow with training data |
| WINDOWS CPU TEMP | < 85°C (laptop) | needs LibreHardwareMonitor; without it, temp is the one signal this skill can't see |

## Known gaps (and the one-time fixes)

- **CPU temperature** is invisible by default. WSL2's virtual kernel does not
  expose host CPU temp, and stock Windows WMI hides it. This is the #1 laptop
  risk signal, so enable it once:
  1. Install **LibreHardwareMonitor** (free), launch it, enable its **WMI**
     provider (and ideally run at startup).
  2. Re-run this skill — the WINDOWS CPU TEMP section auto-detects it via the
     `root/LibreHardwareMonitor` namespace. No script edit needed.
- **SSD wear numbers** (bytes written, % life left) need `nvme-cli` + sudo.
  Until then the script reports *proxies*: checkpoint count/size and recent
  write activity. For real numbers:
  `sudo apt install nvme-cli`, then re-run.

## Verdict guidance tuned to THIS box

- **It's a laptop** (8745HS mobile APU, no discrete GPU). Heat is the real risk
  under long CPU-only runs. Until LibreHardwareMonitor is on, recommend:
  elevate the laptop, keep vents clear, run in a cool room, and consider not
  pegging all 16 threads 100% (smaller `num_envs`).
- RAM ~14GB in WSL / ~28GB on host — memory is not the bottleneck; do **not**
  let training swap.
- The Samsung 1TB NVMe is roomy but an OEM low-power part with modest write
  endurance — checkpoint spam (very small `save_interval`) is the main wear
  source. Move old checkpoints off the SSD (see `checkpoints/` convention in
  the project CLAUDE.md).
- If C: drops under ~40GB free, move/clean **before** starting a long run,
  not after it crashes.

## Files

- `check_sys.sh` — the probe script (read-only, idempotent, safe anytime).
  Edit the probe list here if new hardware is added or a section misbehaves.

## Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

- **Author:** [phanghonghao](https://github.com/phanghonghao)
- **Source:** [phanghonghao/THU-Awesome-Skills](https://github.com/phanghonghao/THU-Awesome-Skills)
- **License:** MIT

Install and usage instructions live in the source repository linked above.

## Pricing

- **Free** — Free

## Security capabilities

Automated source analysis of v0.1.0 — what this tool can access:

- **Network access:** no
- **Filesystem access:** no
- **Shell / process execution:** no
- **Environment & secrets:** no
- **Dynamic code execution:** no

*"Yes" means the capability is present in the source — more access means more to trust, not that it is unsafe.*


## Versions

- **0.1.0** — security scan: passed — Imported from the upstream source.

## Links

- Listing page: https://agentstack.voostack.com/l/skill-phanghonghao-thu-awesome-skills-sys-health
- Seller: https://agentstack.voostack.com/s/phanghonghao
- Browse the marketplace: https://agentstack.voostack.com/browse

---
Listed on AgentStack — the marketplace for AI agent skills and MCP servers. Every listing is security-reviewed. Creators keep 70%.