lsz-asd/dsh-plugin-device-info

dsh-device-info

Read-only Windows device info for DeepSeek Harness agents â€?12 win_* tools, one per Win32 device category.

@huanlin/dsh-plugin-device-info gives the model structured, read-only access to the host machine: time, system, CPU, memory, disk, GPU, network/WiFi, battery, processes, USB, audio, printers. No writes, no installs, no network.

Why this plugin

Ad-hoc PowerShell probing costs tokens every query â€?the model re-generates a command each time (~40â€?00+ tokens, more with retries after a typo or a permission prompt). A plugin call is ~6 tokens against pre-tested, versioned queries. Install once, call any tool; no command-writing, no trial-and-error. (Estimates: scripts/token-estimate.mjs; real benchmark planned, see Roadmap.)

Install

dsh plugin --profile <profile> add github:lsz-asd/dsh-plugin-device-info   # after publishing
dsh plugin --profile <profile> add link:C:/path/to/dsh-device-info           # local checkout

Restart the profile. lib/ ships pre-built (no prepare script).

Tools

Config

Example

User:  这台电脑还剩多少内存？电池呢�Model: [win_memory] �72.8% used, 2× Samsung 8 GB DDR5-4800
       [win_battery] �100%, AC online, scheme 381b4222�(平衡)

All tools return canonical JSON; nullable fields stay null (schema oneOf: [type, null]) for a stable shape:

{ "supported": true, "platform": "win32",
  "name": "13th Gen Intel(R) Core(TM) i7-13700HX",
  "cores": 16, "logical_processors": 24,
  "max_clock_mhz": 2100, "current_load_percent": 3 }

Architecture

model ──tools──�win_* tools ──�execute �canonical JSON value
                        �       render  �pure text projection
                        â–?              CollectContext {run, signal, platform, config}
                        �  Node os (no subprocess)         PowerShell child (WMI/CIM · netsh · powercfg)
  win_time, totals, addresses     killed on exec.signal · hard timeout

src/index.ts entry · src/tools.ts 12 tools + renders · src/collect.ts collectors · src/pwsh.ts runner + parsers · src/types.d.ts ambient peer types

Highlights

• Value/render split â€?execute returns plain bytes/ms/percents; render is a pure projection, replay-safe
• Locale-independent parsing â€?netsh wlan fields stay English on localized Windows; powercfg GUID regex language-independent; CIM /Date(ms)/ â†?ISO-8601
• Sentinel handling â€?battery runtime is junk on AC (0x04444444) â†?null; AdapterRAM uint32 wraparound documented
• Cancellation/timeouts â€?exec.signal kills the child; hard per-query timeout; no orphans, no hangs
• Domain vs infra failures â€?missing battery/WiFi/non-Windows â†?supported: false values; spawn/timeout/abort â†?thrown (isError)

Quality

59 unit tests (real-machine fixtures) · real powershell.exe smoke · real-composition smoke through production @deepseek-ai/dsh-tools (scripts/di-smoke.mjs) · strict TS, tsc + tsdown â†?lib/.

Development

pnpm install && pnpm run typecheck && pnpm test && pnpm run build

@deepseek-ai/dsh-tools is host-provided and not in devDependencies (its peer chain references an unpublished package); typecheck uses src/types.d.ts ambient declarations, tests mock the module, the real module runs in the smoke script/Loader.

Model Experience

The 12 fixed schemas join the generated tool catalog �no dynamic prompt injection. Token effect: fixed (constant schemas; one result message per call). KV cache: prefix-stable (schemas never change at runtime).

Roadmap

• Token-efficiency benchmark â€?real per-task usage: plugin call vs ad-hoc commands (incl. retries/prompts); results published here
• More platforms â€?Linux/macOS (/proc, sysctl, system_profiler)
• More categories â€?Bluetooth, displays, traffic stats, temperature (LibreHardwareMonitor), installed software

Known limitations

Sensors need a third-party daemon · Bluetooth/display/traffic stats deferred · process CPU is cumulative seconds, not live rate · only win_time on non-Windows.

License

AGPL-3.0