dsh-jev-skill-kit

dsh-jev-skill-kit

A DeepSeek Harness plugin that asks a local skill-selection model which skills a task needs, then injects that choice as a one-line preface at the start of the turn.

jev-decision: 已选定：docx、office-docx。不要读取技能，不要再挑选。

The agent then loads exactly those skills instead of browsing or picking again. Selection runs on the machine — no network call, no API key.

What it does

On every new turn (agent/pre-step), before the model speaks:

1. Reads the user's message text.
2. Asks the bundled helper process (a local HTTP service on 127.0.0.1) for a selection.
3. Prepends the resulting one-line preface to the first user message of the turn.

It also reports subagent start/stop events to the same helper, so the kit's local dashboard can show delegation history.

If the turn already starts with jev-decision: (or the legacy 【技能柜】), nothing is injected — a user typing the marker mid-sentence still gets a selection.

Requirement: the jev-decision-kit workspace

This plugin is the DeepSeek Harness adapter of jev-decision-kit. It does not ship the model or the scoring code; it calls the ones in that workspace:

• Python venv at <kit>/.venv (Windows Scripts/python.exe, macOS/Linux bin/python)
• <kit>/host/preface_worker.py and the trained weights in <kit>/models/jev/

Without the kit the plugin stays silent: it injects nothing and writes one kit-missing line to <kit>/runtime/harness-plugin.log. It will not prefix your turns with an error.

Install

1. Clone and set up the kit (README).
2. Install this plugin into a DSH profile:
   • GUI: 设置 → 插件 → 添加插件 → the absolute path to this directory, then restart DSH and open a new session.
   • CLI (non-desktop profiles): dsh plugin add <absolute path to this directory>
3. The profile must use a link: (or file:) dependency — that is how the plugin finds the kit after a reinstall. See the kit's INTEGRATION.md for the exact steps.

Where the kit lives, in order of precedence:

1. JEV_KIT_ROOT environment variable
2. <plugin dir>/kit-root.txt
3. %LOCALAPPDATA%/jev-skill-kit/root.txt (Windows) or ~/.local/share/jev-skill-kit/root.txt
4. the source path recorded in the profile's package.json (link:/file: dependency)
5. two directories above the plugin (a plain checkout)

Configuration

Privacy

Everything runs locally. The plugin starts a 127.0.0.1 helper process and posts the task text to it; nothing leaves the machine. The helper is spawned with windowsHide: true and no detached, so no console window appears.

License

MIT — same as the parent workspace.