dsh-sysload

Live CPU, GPU and memory load for the DeepSeek Harness Web UI — three small cells in the Session header, where the session utilities already live, instead of another floating window.

Each cell is a label, a percentage and a 2 px bar. The bar and the number take the warning colour at 75 % and the error colour at 90 %. Hovering a cell shows the detail: core count, memory in GiB, VRAM, the adapter name, and which strategy produced the GPU reading.

CPU 37%   GPU 23%   MEM 77%
▬▬▬▬▬▬    ▬▬▬▬     ▬▬▬▬▬▬▬▬▬▬▬▬

How it works

A plugin is two halves. The Host half samples the machine and serves one JSON document; the Client half polls it and renders.

Reading	Source
CPU	os.cpus() — idle/total delta between two polls, so the number is load since the last poll; the very first sample of a process falls back to the average since boot
Memory	os.totalmem() and os.freemem()
GPU	nvidia-smi --query-gpu=utilization.gpu,memory.used,memory.total,name; with several adapters the busiest one drives the bar and VRAM is summed. Without NVIDIA, Windows falls back to the GPU Engine(*)\Utilization Percentage performance counter set (slow: ~7 s per reading, so it is bounded at 25 s and parked for five minutes after a failure). Neither path available → the cell shows —

A GPU reading younger than 1.8 s is reused, so overlapping polls never spawn two probes. The route takes no input: it reads no query string, no body and no headers, so nothing a caller sends can reach the commands above.

Polling runs at 2 s in a visible tab, drops to 15 s when the tab is hidden, and backs off exponentially to a 20 s ceiling while requests fail — a closed tab stops polling entirely. The Client requests the route with a path relative to the page, so the widget keeps working behind a reverse-proxy prefix and inside the desktop shell.

Requirements

• DeepSeek Harness with the webServer service (the Web UI). Without that service the plugin registers nothing and stays inert.
• CPU and memory: any platform Node runs on.
• GPU: nvidia-smi on PATH (Linux and Windows), or Windows performance counters for Intel/AMD adapters. macOS reports —.

Install

dsh plugin add github:k623615746/dsh-sysload

From a local checkout instead:

dsh plugin add file:/absolute/path/to/dsh-sysload

The widget appears in the Session header on the next page load; no restart.

Endpoint

curl http://127.0.0.1:56142/dsh-sysload/stats

{
  "at": 1763000000000,
  "cpu": { "pct": 35.5, "cores": 20 },
  "mem": { "pct": 77.1, "used": 12947177472, "total": 16962281472 },
  "gpu": {
    "pct": 23,
    "memUsed": 1509949440,
    "memTotal": 8589934592,
    "name": "NVIDIA GeForce RTX 3070 Laptop GPU",
    "source": "nvidia-smi"
  }
}

source is one of nvidia-smi, perf-counter or unavailable; any field the machine cannot supply is null. Only GET and HEAD are answered (other methods get 405 with allow: GET, HEAD), and every response is cache-control: no-store.

Privacy

Reads local kernel counters and, when a GPU exists, spawns a local CLI. No network egress, no telemetry, no files written.

License

MIT — see LICENSE.

---

中文

在 DeepSeek Harness 会话顶栏显示本机的 CPU、GPU 与内存占用率：三个小格，每格一行标签、一个百分比和一条 2 px 细条，75 % 变告警色、90 % 变错误色；悬停可看核数、内存 GiB、显存、显卡型号，以及 GPU 读数来自哪条采集路径。

• CPU：os.cpus() 两次轮询之间的 idle/total 差值，即「距上次轮询以来的负载」；进程内第一次采样退回开机以来的平均值。
• 内存：os.totalmem() 与 os.freemem()。
• GPU：优先 nvidia-smi（多张卡时取最忙的一张驱动进度条、显存求和）；没有 NVIDIA 时，Windows 退回 GPU Engine(*)\Utilization Percentage 性能计数器（一次约 7 秒，因此限时 25 秒、失败后停用五分钟）；两条路都取不到就显示 —。

轮询在前台 2 秒一次，标签页切到后台降为 15 秒，请求失败时指数退避到 20 秒上限；关闭标签页即停止。前端用相对路径请求，放在反向代理前缀下或桌面壳里都不会打歪。

安装：

dsh plugin add github:k623615746/dsh-sysload

数据接口 GET /dsh-sysload/stats 返回一份 JSON，只接受 GET/HEAD，cache-control: no-store；插件只读本机计数器并调用本机命令行工具，不上报任何数据。

MIT 许可。