dsh-plugin-judge
DeepSeek Harness 插件价值审计器：安装前判断插件是否值得安装，安装后审计是否仍应保留——启发式扫描 + LLM 评判，并在模型切换时提醒重新审计。
插件会安装到这里；不确定时保持 web。
npx -y @deepseek-ai/dsh plugin --profile web add dsh-plugin-judge@0.1.1Plugin Judge 以 dsh-plugin-judge 发布，当前版本为 0.1.1。Plugin Hub 会校验它的 manifest，并保存精确安装来源，便于复现安装结果。
Judge whether a DeepSeek Harness plugin is worth using: pre-install review, post-install audit, and re-audit reminders when the model changes.
Plugins fall into two natures:
agent/pre-step / tools/pre-execute, restrict / guard on tools, or
complete: true wholesale prompt replacement) — like putting the model
"in a cage".Constraint plugins carry a hidden cost: the model will follow the rules faithfully (compliance looks great), but it performs worse on the task you actually care about — rules crowd the context and suppress the model's own judgment, putting a low ceiling over a capable model. Worse, as models get upgraded, many "patch rules for the old model" turn from patches into ceilings, and the plugin becomes obsolete or even harmful.
So a plugin's value is not a property of the plugin alone — it is a binary relation of "plugin × current model". That is exactly what this plugin evaluates.
| Feature | When | How |
|---|---|---|
| Pre-install review | Before installing | /plugin-audit <github:owner/repo or npm:pkg> command, the judge_plugin tool, or the settings-page input. Fetches the plugin source → static scan + LLM judge → a "worth installing?" verdict |
| Post-install audit | Already installed | Enumerates installed bundles from the profile → scans and classifies each (capability / constraint / cosmetic / hybrid) → constraint-risk score → settings-page report panel |
| Model-switch reminders | When the model changes | Watches the default model → flags plugins whose verdict depended on the old model → overlay popup reminder to re-audit |
# Install (requires an existing web profile)
dsh plugin --profile web add dsh-plugin-judge
# Or install from source
dsh plugin --profile web add <path-to-this-repo>
After installing and restarting the harness:
/plugin-audit github:some/repo in chat for a pre-install review;judge_plugin tool directly;Installed plugins: profile manifest (~/.dsh/profiles/<profile>/package.json → dsh.profile.bundles)
+ each package's package.json / lib source (read locally via node:fs)
Candidate plugins: package.json + entry source fetched from npm registry / jsDelivr / GitHub raw (Node fetch)
Heuristic scan: capability signals (+) vs constraint signals (−), weighted → class + risk score
LLM judge: llm.stream (generative verdict, carrying the current model identity) → verdict + reasons
Model watch: agentDefaultModel.currentSelection() polling (fallback: agent/request waterfall observation)
Persistence: ~/.dsh/plugin-judge/audits.json (audit history + model fingerprint)
Bridge: Host ctx.connection.rpc.handle('/plugin-judge', …) ← Client rpc.call polling
Reminders: Host produces pending-reminders → Client polls → shell.overlay popup
Reports: Client settings.section "Plugin Judge" page + judge_plugin tool card
Capability signals: tools.register / defineTool / skills / MCP /
webServer.register / registerFetchProvider / commands.register /
provide( …
Constraint signals: systemPrompt.section / ctx.systemPrompt / persona /
writing AGENTS.md / agent/pre-step / tools/pre-execute / tools/execute
waterfall listeners / restrict( / guard( / complete: true wholesale
replacement / strong-directive density (must, never, always, MUST, NEVER,
必须、不得、禁止) / fixed-template instructions (strictly follow, exact
format, 严格按照、按以下格式) …
judgePlugin/
├── package.json # dsh.bundle.patch + dsh.client.inject + exports (Node >= 18)
├── cordis.patch.yml # insert: plugin-judge
├── lib/
│ └── index.js # Host half: inventory / source fetching / heuristic scan / LLM judge /
│ # model watch & reminders / tool & command / connection RPC bridge / persistence
└── client/
└── client.js # Client half (__ModuleLoader__ format, no build step):
# settings.section report panel + shell.overlay model-switch reminder popup
Verdicts are advisory opinions produced by heuristics plus a model judge — not a security audit. Installing any plugin runs third-party code on your machine with your own permissions; read the source before you install.