dsh-guarded-live-voice
受防护的 DeepSeek Harness 语音基础，具备精确会话披露和仅提案权限
插件会安装到这里；不确定时保持 web。
npx -y @deepseek-ai/dsh plugin --profile web add github:Jstn-1g/dsh-guarded-live-voice#7807f175d6f3e47e235e5852aefddb20990d1140Guarded Live Voice 以 dsh-guarded-live-voice 发布，当前版本为 0.2.0。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。
A guarded DeepSeek Harness voice foundation with an exact-session Host boundary and a lazy browser disclosure UI. Live microphone and audio remain disabled.
Milestone two adds the browser-side consent surface to the milestone-one Host boundary:
Milestone three groundwork adds an internal, configuration-only Qwen session transport with deterministic fake-provider coverage. It can send only one fixed text-only, manual-turn configuration event and exposes no audio, transcript, instruction, tool, DSH context, raw socket, or send capability. The provider must confirm the exact requested model, session identity, modalities, and turn detection before the transport reports readiness.
The registered plugin and installed package do not request microphone access,
transmit audio or text, open a Qwen provider connection, perform transcription
or playback, or insert a proposal into the composer. The internal transport is
not exported from the package root and is not called by apply. Binary
WebSocket frames remain rejected. Version 0.2.0 is a development milestone, not
a marketplace-ready voice product.
The disclosure flow is user-visible, but it is not cryptographic proof that a
human accepted it. The one-shot challenge proves control of that local client
connection only. Loopback, Host, Origin, and Sec-Fetch-Site checks mitigate
remote access, DNS rebinding, and cross-site requests; they do not authenticate
a human or resist a malicious same-user local process.
The registered plugin sends no data to Qwen because it opens no provider connection. A future audio milestone is designed to exclude DSH history, files, workspace instructions, memory, and project context. Its provider output may create only bounded proposal data for the exact bound session. Composer integration is not implemented yet; when added, it must fill the ordinary draft without submitting a message, calling a tool, writing a custom session event, or executing work.
Requirements: Node.js 22.19+ and pnpm 11.7.
pnpm install
pnpm check
pnpm check runs strict Host and browser TypeScript checking, deterministic
tests, Host and browser builds, package linting, browser-bundle materialization,
and a dry-run package-content check.
The bundle inserts a guarded-live-voice Cordis row. A profile override will
need an Alibaba Cloud Model Studio workspace id before provider authorization:
- id: guarded-live-voice
name: dsh-guarded-live-voice
config:
credentialRef: DASHSCOPE_API_KEY
dashscopeWorkspaceId: your-workspace-id
model: qwen-audio-3.0-realtime-plus
route: /guarded-voice
trustedHosts: localhost,127.0.0.1,[::1]
maxConnections: 8
After accepted disclosure, the registered plugin validates the allowlisted endpoint and credential availability only; it does not connect to Qwen. The internal configuration-only transport will remain outside the public package API until it is composed behind the same exact authority and consumed-consent boundary.
Configure the credential through DSH's credential provider. Never put a secret
value in cordis.patch.yml, browser storage, an issue, or a log.
MIT licensed. This repository is a clean-room implementation with new public history. It does not contain private product source, assets, identifiers, or repository history.