wj2514939573-ui / dsh-security-gate

🛡️ dsh-security-gate · 插件安检门 / SecurityGate

装插件，先过安检门。 A DSH (DeepSeek Harness / Cordis) plugin that inspects other plugins before you install them — static scan + AI review + runtime watch, community ratings and bug bounty.

• Plugin ID: gate-2 · Package: pkg-18 (v1.0.0 / v10)
• Platform: DeepSeek Harness dynamic Cordis plugin (Host + Client)

---

Why SecurityGate

dsh-plugin-marketplace validates plugins and dsh-suite tests compatibility, but those are after-the-fact checks. SecurityGate moves security to proactive defense before install:

                   ┌──────────────────────────────────────────┐
                   │   🛡️ SecurityGate · 插件安检门            │
   new plugin ───▶ │ ① Static scan (malicious score + caps)    │
                   │ ② 🤖 LLM AI review (false-positive check) │
                   │ ③ Runtime registry watch (live arrivals)  │
                   │ ④ Community ratings + bug bounty          │
                   │  → verdict & advice → install with trust  │
                   └──────────────────────────────────────────┘

Features

#	Feature	Description
1	Dual-dimension scan	恶意风险分 (malicious-risk score 0-100: dynamic code / shell / raw fs / obfuscation / policy bypass / persistence / exfil) separated from 权限能力面 (capability surface: network / file write / env read / services / tools — functional permissions, not malice)
2	🤖 LLM AI review	One click: the current model re-judges static findings to filter false positives (gate_scan ai_review param)
3	Runtime registry watch	Baselines the visible tool set at startup, subscribes tools/change, tracks added/removed tools and recent arrivals
4	New-tool alert	A floating "🚨 new plugin tools detected" toast appears when any plugin registers tools during the session
5	Quick access	A 🛡️ button in the session header opens a floating quick-scan panel — no need to open Settings
6	Scan history	Last 20 scans persisted with the community store; per-entry verdict copy
7	Share report	One-click export + copy of a shareable inspection report with the "✅ passed SecurityGate" stamp
8	Community ratings	1-5 star security ratings + comments, leaderboard (local store; cloud sync on the roadmap)
9	Bug bounty flow	Submit reports (critical 1000 / high 500 / medium 200 / low 50 pts), audit flow: pending → under review → confirmed (points awarded) / dismissed / fixed / blacklisted
10	i18n	zh-CN / en-US UI via the locale service (auto-falls back to Chinese)
11	Onboarding	A "SecurityGate is on duty" card in the Run card + in-app usage guide

Installation

This is a dynamic Cordis plugin for DSH. Two ways:

A. From the define payload (recommended, reproducible)

The file plugin/source.json contains the exact cordis_define payload:

plugin:  { kind: 'new', idPrefix: 'gate' }
name:    security-gate 插件安检门
purpose: 安装插件前先过安检门：双维静态安全扫描（恶意风险分 + 权限能力面）、
         LLM AI 复核、运行时注册观察、社区安全评分与漏洞赏金。
code.host:    plugin/host.js (or host.raw.js)
code.client:  plugin/client.js (or client.raw.js)

Ask your DSH agent to define it with code.host = plugin/host.raw.js content and code.client = plugin/client.raw.js content, then run it. Approve the client half when prompted.

B. Manual copy

plugin/host.raw.js and plugin/client.raw.js are the bare function bodies exactly as recorded by the Host — paste them into cordis_define.

How to use

Settings UI: Sidebar → Settings → 🛡️ 插件安检门

• 🧪 Preview Sandbox: paste source or scan a local file/directory → risk score, capability surface, findings (with line numbers), dry-run trace, advice; 🤖 AI Review button
• ⭐ Community Ratings: rate plugins 1-5 stars, leaderboard
• 🐛 Bug Bounty: submit reports, drive the audit status flow

Conversation: just say "用安检门扫一下这个插件" — the model calls gate_scan automatically. Four model tools:

Tool	Purpose
gate_scan	static scan + dry-run preview; source or path; optional ai_review
gate_registry	runtime registry watch: tools added/removed since gate start, recent arrivals
gate_rate	submit a 1-5 star community rating
gate_report	submit a bug bounty report

Architecture

Host (Node process):
  static scan engine (7 malicious rules × weights → 0-100 score; 4 capability markers)
  AI review (llm service + agentDefaultModel, JSON verdict parsing)
  runtime watcher (tools baseline + tools/change subscription, arrivals log)
  community store (ratings + reports + history, persisted to <workspace>/.security-gate-store.json)
  scan history + share-text exporter
  4 model tools + 11 package-private RPCs
Client (browser):
  settings page (3 tabs) + usage guide (i18n)
  session-header 🛡️ quick button + floating quick-scan panel
  floating new-arrival alert (8s poll)
  Run card "on duty" status card

Data & privacy

• Community data lives locally: <workspace>/.security-gate-store.json (ratings, reports, scan history)
• SecurityGate itself makes no network calls (the plugin's own runtime sends nothing anywhere)
• Scanning is read-only: inspected code is never executed
• AI review sends only scan findings/snippets to the configured model

Roadmap

• v2.0: cloud community backend (shared ratings / blacklist / bounty across users — minimal JSON API or serverless)
• Cloud preview sandbox backend (e2b / container) behind the backend seam
• Marketplace integration (install button → inspection gate)
• Behavior signature library (known-malicious fingerprints + auto-blacklist)

License

MIT — see LICENSE.