One-Line Pitch

Add a 🔊 button next to every AI response in DeepSeek Harness. Click it to have the local Omi engine call Doubao TTS 1.0 to read out the final response text. The Doubao API Key stays in your own Mac's Keychain (BYOK).

Core Features

• Inject a three-state 🔊 button next to each assistant response: click once to read, click again while playing to pause, click again to resume from pause position
• Only read the final response body; tool call logs, thinking processes, command executions, pure code fences, pure tables, and box drawings/ASCII graphics are filtered by the engine before the request. Button shows as disabled when there's nothing to read
• Engine actively polls /v1/status to sync play/pause/failed/idle states. Show specific failure reason toast on read failure. Button auto-resets after engine disconnection
• All text and playback control goes through local 127.0.0.1:8765 HTTP protocol. Plugin side does not store or touch the Doubao Key
• Engine implements cost control: 3-second deduplication for identical text, 300ms throttle for any read, in-process LRU cache of 3 items/≤5MB, pure memory cleared on exit

Technical Implementation

• Language: JavaScript (plugin side, runs in browser) + Swift 6 (Omi engine, separate source repo directory)
• Key Dependencies: No runtime dependencies (dependencies in package.json is empty); browser side uses DSH built-in React; engine uses Foundation / AVFoundation / Network / Swift Concurrency
• Architecture Pattern: DSH cordis client plugin, mounted via cordis.patch.yml with minimal Node placeholder entry, exports["./client"] + dsh.client.platform: web triggers client bundle auto-load; client registers read-aloud button component to conversation.chat.assistant-actions slot (client/lib/client.js:247-258), does not modify host Node side logic
• Entry Files: Plugin entry host/lib/index.js (placeholder) + client entry client/lib/client.js; engine entries engine/Sources/ReadAloudConfig/main.swift and engine/Sources/ReadAloudService/main.swift, HTTP service in engine/Sources/ReadAloudService/LocalTTSService.swift

Use Cases

When you want to "listen" instead of "read" long DSH responses, replay AI answers while doing housework/commuting, or want visually impaired/low-vision users to hear AI answers on desktop. The cost is keeping an unsigned Omi engine running on your Mac and paying Doubao's per-character billing.

Prerequisites and Compatibility

Installation

dsh plugin --profile web add github:PolinniZhong/dsh-omi-voice

The engine (engine/, macOS menu bar App) needs to be separately built per engine/README.md and ditto to ~/Applications/Omi DSH.app. This repo's plugin install command won't install the engine for you.

Configuration

Doubao API Key, current voice, speech rate, etc. are managed by Omi engine independently and not in this plugin's config items; this plugin does not implement DSH cordis style ctx.config Schema.

FAQ

Q: Clicking 🔊 prompts "Omi DSH engine not detected" - what to do?

A: Engine is not running or closed. Open "Omi DSH" (Applications folder or ⌘+space search "Omi"), recommend checking "App Preferences > Launch at Login" in settings.

Q: Prompt says "Please configure Doubao API Key in Omi settings page first" - how to handle?

A: Engine is running but Key not saved. Open Omi settings page "API Key" item, follow README "Get Doubao API Key" three steps to enable "Speech Synthesis 1.0" in Volcano Engine console, create associated Access Key and fill it in; plugin itself has no Key input entry.

Q: Button is gray, clicking does nothing?

A: This response has no readable content after engine filtering (pure code, pure table, pure box drawing/ASCII graphics). Starting from v0.1.1, button shows as disabled and hover says "This response has no readable content", no Doubao request will be made.

Q: Can I switch Doubao voice? Is there voice selection UI in plugin?

A: No voice UI on plugin side, v1 protocol only returns current voice field; switching needs to be done in Omi engine settings page "Voice ID". Voice list endpoint /v1/voices is marked as v1.1 reserved in protocol.

Q: Is Windows or Linux supported?

A: Not supported. Omi engine only targets macOS 13+ Apple Silicon, needs to be built from source; although plugin JS part can load in any dsh web, without macOS engine there's no sound output.

Q: What's the difference from zero-config read-aloud plugins (like dsh-voice-chat)?

A: Zero-config types usually use system TTS, no Key needed, robotic voice, read entire paragraphs; this plugin uses Doubao TTS 1.0 natural voice, only reads final response and filters noise content, cost is BYOK per-character billing, needs Omi engine running locally.

Q: Where is data stored? Does it write chat history?

A: Engine does not save read text to disk, does not write history, does not callback any remote endpoints; plugin only communicates via 127.0.0.1. PRIVACY.md and docs/API.md §5 clearly state text and Key are not sent externally.

Q: How to uninstall?

A: Use dsh plugin --profile web remove github:PolinniZhong/dsh-omi-voice to remove plugin, then rm -rf "$HOME/Applications/Omi DSH.app" to delete engine; plugin uninstall also cleans up old version residual localStorage items (dsh-omi-voice/settings, dsh-omi-voice/autoSeq/*) at client.js:46-55 startup.

Getting Started Difficulty

Beginner — the plugin itself only exposes one button and optional engine address override. But to make it actually produce sound requires completing three extra things: build unsigned engine on Mac, apply for Doubao Access Key, fill into Omi settings page.

Known Issues and Limitations

• Engine only supports macOS 13+ Apple Silicon (engine/Package.swift:6, engine/CHANGELOG.md:101), Windows / Linux have no corresponding build artifacts
• Engine is currently "macOS Source Build Developer Preview", not completed Developer ID signing, notarization, installer distribution (engine/SECURITY.md:24, multiple declarations in engine/CHANGELOG.md), needs evaluation in isolated development environment
• Protocol v1 does not implement long text auto-segmentation and prefetch, needs to wait for v1.1 (engine/CHANGELOG.md:103, docs/API.md:23); current text plays in ≤900 UTF-8 byte natural segments sequentially
• Protocol v1 has no authentication token, any local process can call engine to trigger read; current impact is only "local speaker makes sound", both docs/API.md:172 and engine/CHANGELOG.md:104 note that if local file reading is supported later, token must be added
• Starting from v0.1.1, auto-read and 📢 toggle removed (CHANGELOG.md:13), deliberately only click-to-read remains; product decision, please do not re-introduce
• Current Keychain access strategy targets unsigned local development packages (engine/CHANGELOG.md:105),正式签名版本需要迁移到 Data Protection Keychain 与正式 Access Group

dsh-omi-voice

沉浸式听朗读 · 豆包音质

点一下 🔊，把 AI 回复用豆包的自然音色念给你听——无需复制，不做自动朗读

DeepSeek Harness 插件 · BYOK · MIT

让 DeepSeek Harness（DSH）桌面端里的 AI 回复，用豆包自然音色读给你听。语音由你本机常驻的 Omi 引擎合成，豆包 API Key 只留在你自己的钥匙串里（BYOK）。

快速概览

三步开始朗读

1. 安装插件 + 构建并打开 Omi 引擎（见下方「获取豆包 API Key」与「安装」）。
2. 在 Omi 引擎设置页保存一次豆包 API Key。
3. 在 DSH 对话里点 AI 回复旁的 🔊，即可朗读。

flowchart LR
    A[点 🔊] --> B[插件取回复的最终回答文本]
    B --> C[POST 127.0.0.1:8765/v1/speak]
    C --> D[Omi 引擎清洗 + 分段]
    D --> E[豆包 TTS 流式合成]
    E --> F[本机扬声器播放]

获取豆包 API Key（新手必读）

本插件是 BYOK（自带 Key）：豆包语音由你自己的火山引擎账户按字符计费。三步拿到 Key：

第 1 步 · 找到「豆包语音」：登录火山引擎控制台，进入「豆包语音」（语音技术）服务。

第 2 步 · 开通「语音合成 1.0」：在产品列表里开通「语音合成大模型 / 语音合成 1.0」。

第 3 步 · 创建 API Key：创建 Access Key 时，关联第 2 步开通的「语音合成 1.0」服务；把得到的 API Key 填进 Omi 引擎「设置 > API Key」并保存。

API Key 只保存进 Omi 引擎的 macOS Keychain，插件侧零 Key、不出本机。

安装

dsh plugin --profile web add "github:PolinniZhong/dsh-omi-voice#v0.1.2&path:/"

本地开发可直接装目录：

dsh plugin --profile web add /path/to/dsh-omi-voice

引擎（Omi DSH）构建见 engine/README.md：./engine/build/build-service.sh 后 ditto 到 ~/Applications/Omi DSH.app。

使用

• 🔊 点读；播放中再点 = 暂停，再点 = 从暂停处继续；点其它消息的 🔊 = 打断并读新的。
• 只读最终回答：工具执行日志、思考过程不会读；代码围栏、表格、纯图形（盒绘/ASCII）在请求前过滤，回复若只有这些内容，按钮呈禁用态并提示"没有可朗读的内容"。
• 引擎未启动 / 未配置 Key 时，点击会给出明确提示（含去哪打开 Omi）。

成本透明

豆包 TTS（seed-tts-1.0）按字符数计费，由你在火山引擎账户自行承担。为此：

• 只有手动点 🔊 才合成，不自动朗读；
• 引擎对相同文本 3 秒内去重，且进程内 LRU 缓存最近 3 条（≤5MB、纯内存、退出即清），重复朗读不重复计费；
• 纯表格/纯代码等无有效内容不发请求（invalid_text）。

架构与协议

插件只是"遥控器"：播放、变速、暂停/继续、缓存、文本清洗全部在 Omi 引擎内完成。协议见 docs/API.md（/v1/status、/v1/speak、/v1/pause、/v1/resume、/v1/stop）。

FAQ

相关

• 引擎：本仓库 engine/（Omi DSH 本地引擎，Swift 源码 + 构建脚本）
• 生态：awesome-dsh-plugin

项目文档

• AGENTS.md — 给 AI 编码代理的项目说明（Codex 约定）
• docs/DESIGN.md — 架构与设计
• docs/DECISIONS.md — 设计决策记录
• docs/MEMORY.md — 长期知识库（坑/结论）
• docs/HANDOFF.md — 交接与续作
• docs/API.md — 本地协议

License

MIT