dsh-qwen-voice

简体中文 | English

Experimental voice-control plugin for DeepSeek Harness Web, built on top of Qwen Audio Agent.

Version 1.0.0 connects only to DeepSeek Harness Web at http://127.0.0.1:3080. It can be installed independently from GitHub, but first use still requires a DashScope API key and a real DSH session selected as the Coordinator in the Web UI.

This project would not exist without Qwen Audio Agent. It reuses Qwen Audio Agent as the realtime voice engine and adds the DeepSeek Harness plugin UI, ACP bridge, and multi-session routing integration around it. Our sincere thanks to the Qwen Audio Agent maintainers and every upstream contributor for releasing their excellent work as open source.

It adds a floating voice orb that remains connected while switching DSH conversations. One browser tab can dispatch multiple named DSH sessions, show live task state, interrupt playback, and announce results as tasks finish.

Compatibility

• DeepSeek Harness Web 0.1.0-rc.6 (the current development environment also verifies 0.1.0-rc.7)
• Qwen Audio Agent 1.10.0
• Node.js supported by both upstream projects
• Windows launcher included; the plugin and bridge are cross-platform Node.js

This is a community integration, not an official DeepSeek or Qwen release. Both upstream projects are evolving quickly, so exact versions matter.

Current support scope

The realtime voice frontend and the backend task Agent are separate layers. This release supports:

• DashScope Qwen-Audio-Realtime: qwen-audio-3.0-realtime-plus (default), qwen-audio-3.0-realtime-flash, qwen3.5-omni-flash-realtime, and qwen3.5-omni-plus-realtime;
• the optional local Hugging Face speech-to-speech frontend provided by upstream Qwen Audio Agent;
• DeepSeek Harness Web as the backend task Agent through this project's ACP bridge. The complete chain is pinned to 127.0.0.1:3080; it never probes or connects to a simultaneously running Desktop instance, which prevents split session ownership.

This release does not directly support OpenAI Realtime, Gemini Live, or other cloud realtime providers not registered by upstream Qwen Audio Agent. Backend model-provider support does not imply realtime voice-provider support.

Install

One-command Windows installer (recommended)

Start DSH Web and confirm http://127.0.0.1:3080 opens, then run:

git clone https://github.com/leaveimagination/dsh-qwen-voice.git
cd dsh-qwen-voice
.\scripts\install.cmd

The installer validates Node.js, installs pnpm 11 when missing, installs all dependencies, applies the Qwen compatibility patch, builds and registers both DSH plugins, asks for the DashScope API key with hidden input, and starts the voice runtime. The real key is written only to %USERPROFILE%\.config\qwaudio\config.env and never to the repository.

To install without starting immediately:

.\scripts\install.cmd -SkipStart

Prerequisites

Manual install

# 1. Clone and install (includes the pinned Qwen Audio Agent 1.10.0)
git clone https://github.com/leaveimagination/dsh-qwen-voice.git
cd dsh-qwen-voice
pnpm install

# 2. Apply compatibility patches + build + register into a DSH profile
pnpm setup

pnpm setup does three things automatically:

1. patches the project-pinned Qwen Audio Agent 1.10.0 (never touches a global install);
2. installs the ACP Bridge dependencies and runs typecheck + build;
3. registers the plugin into --profile web via the DSH CLI (override with the DSH_PROFILE env var, e.g. $env:DSH_PROFILE = 'voice-test').

Start

pnpm start

Starts the voice runtime, Gateway, and ACP Bridge (foreground, stop with Ctrl+C). Then refresh the DSH Web page (http://127.0.0.1:3080) — a floating voice orb appears in the bottom-right corner.

By default the current working directory is the ACP workspace. To override:

$env:ACP_WORKSPACE = 'C:\path\to\workspace'
pnpm start

Configure voice credentials (required)

The default frontend uses DashScope realtime voice and needs your own API key:

The Windows installer guides this step automatically. For a manual install, edit %USERPROFILE%\.config\qwaudio\config.env and set the following values (see config.env.example):

DASHSCOPE_API_KEY=sk-your-key
QWEN_AUDIO_REALTIME_PROVIDER=dashscope
QWEN_AUDIO_REALTIME_MODEL=qwen-audio-3.0-realtime-plus

Restart pnpm start for the config to take effect.

Credentials stay in the local config; no API key is included in this repository or in the browser plugin bundle.

First use: bind the coordinator session

Open the voice panel and click Set as coordinator session in the target DSH session. Voice tasks are dispatched by that Coordinator session to multiple child sessions in parallel; switching Harness conversations keeps the voice connection alive.

Troubleshooting

• No floating orb on the page: confirm pnpm setup succeeded, DSH Web was restarted, and the plugin shows up in dsh plugin --profile web list.
• Gateway fails to start: check the startup output and confirm QWEN_AUDIO_REALTIME_PROVIDER=dashscope in config.env (no other value is supported).
• DSH Web is reported unreachable: start DSH Web and confirm http://127.0.0.1:3080 opens. This release does not fall back to Desktop.
• Cancelling a task hangs: the cancellation deadlock is fixed in v1.0.0+; if it still misbehaves, open an issue with the Gateway log.

The GitHub Actions workflow re-runs install → patch → typecheck → build → bridge tests → CLI check on a fresh Windows runner to prove that "clone and run" works.

What the compatibility patch changes

pnpm setup:qwen applies version-checked integration changes only to the Qwen Audio Agent 1.10.0 runtime bundled inside this project:

• separate coordinator lanes for simultaneous voice turns;
• keep one persistent ACP Coordinator Session per authenticated owner;
• allow the fixed local DSH Web origin to access the Gateway;
• install the authenticated local DSH Session API and Task Manager lifecycle;
• expose target Session identity in task snapshots;
• disable the obsolete first-session continuation fallback when present.

Task cards use the Gateway's authoritative task status. Cancelling a task now shows an in-progress state, confirms the Gateway response, and reports failures instead of silently sending a request.

The script refuses unknown Qwen Audio Agent versions and is safe to rerun. Reinstalling the upstream package removes the patch; run pnpm setup:qwen again afterward to restore the integration.

Repository layout

pnpm setup builds and registers both the src/ main plugin and the tools/ coordinator tools plugin into the same DSH profile.

Development

pnpm typecheck
pnpm build
pnpm --dir bridge test

The ACP bridge accepts loopback DSH URLs only. The floating client defaults to the local gateway at 127.0.0.1:3101.

Known limitations

• Developer-preview compatibility is verified with DSH Web 0.1.0-rc.6 and 0.1.0-rc.7.
• DSH Desktop and automatic Web/Desktop multi-instance switching are not supported in this release.
• The only verified cloud realtime voice frontend is DashScope Qwen Realtime.
• The Qwen compatibility patch is temporary and should eventually be replaced by upstream extension points.
• Session titles are generated from the spoken task objective and may still require manual renaming for ambiguous requests.

Acknowledgements

Special thanks to the Qwen Audio Agent team and contributors. They built and open-sourced the realtime voice frontend, gateway, audio transport, provider integration, and agent coordination foundation on which this DSH integration is based. This repository focuses on connecting that foundation to DeepSeek Harness; it is not a replacement for or independent reimplementation of Qwen Audio Agent.

Thanks also to the DeepSeek Harness team for the plugin platform and community.

Community discussions

• Qwen Audio Agent community: DSH Qwen Voice
• DeepSeek Harness community: DSH Qwen Voice

License

MIT. Adapted audio transport logic is noted in NOTICE; Qwen Audio Agent is Apache-2.0 licensed. Please retain the upstream attribution when redistributing this integration.