@local/dsh-hold-to-talk
DeepSeek Harness 编辑器的按住说话听写：长按输入框，松开即可将语音转写到草稿中
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:jryang1997/dsh-hold-to-talk#e5c337e6cb55cf8b0657e238929118f73f44dd21Hold To Talk 以 @local/dsh-hold-to-talk 发布，当前版本为 1.0.0。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。
English · 中文
Long-press anywhere on the composer card in DeepSeek Harness to dictate. Release to transcribe and drop the text into the draft — nothing is sent automatically.
┌──────────────────────────────────────────────┐
│ 按住鼠标 语音输入文字 ← fades in │ ← hover the input box
│ │
│ ⌄ ＋ 权限 计划 [ model ] [ ➤ ] │
└──────────────────────────────────────────────┘
hold ≈0.35 s ↓
┌──────────────────────────────────────────────┐
│ ▁▃▅▇▅▃▁ 正在聆听 · 松开完成 │ ← release to transcribe
│ Esc 取消 · 上滑取消 │
└──────────────────────────────────────────────┘
This plugin does not ship a speech recogniser. It reuses the one DeepSeek Harness already has, so both of these must hold before it can do anything:
The official voice-input bundle is enabled.
It ships with DeepSeek Harness as @deepseek-ai/dsh-experimental-voice-input-bundle
(Plugins → Voice input / 语音输入). It is what mounts the speech remote that this
plugin calls. If you disable it, this plugin silently registers nothing and the composer
keeps its normal behaviour.
The recognition models are prepared.
The official bundle transcribes locally with SenseVoice. On first use the Host
downloads the model files (≈ 240 MB) into ~/.dsh/speech-to-text/sensevoice/models
(model.int8.onnx, tokens.txt, silero_vad.onnx).
Open the voice-input bundle page and run Download and prepare once; or simply click
the built-in microphone button once and follow the setup prompt.
The official bundle has no streaming support, so a transcript appears only after you release. Audio is transient: it never becomes a session event or an attachment, and only the text you later submit is recorded.
Tested against DeepSeek Harness 0.1.7-rc.2 (the version shipped with the desktop app at
the time of writing). The plugin uses two internal interfaces — the conversation.input.overlay
slot and the speech remote — and is written to degrade instead of throwing when either is
missing.
DSH plugins are installed from a package directory, so clone first:
git clone https://github.com/jryang1997/dsh-hold-to-talk.git
Then pick one route.
A. Through the agent (recommended, works in the Desktop app)
Ask the agent in any session:
Install the bundle at
<absolute path to the clone>withplugin_manager(action: install_bundle).
It runs install_bundle for the current profile and reports application: applied
when the change is live.
B. Through the CLI
dsh plugin --profile <profile> add <absolute path to the clone>
The Desktop application owns its
desktopprofile exclusively, so use route A there.
C. Verify
Reload the page (Ctrl+R), hover the message box and look for the hint in the lower-right
corner. If it does not appear, open the browser console and check that the host half
activated.
| Gesture | Result |
|---|---|
| Hover the input box | The hint fades in over ~0.6 s |
| Press and hold ≈0.35 s without moving | The whole card becomes a recording surface |
| Release | Transcript is inserted at the caret — not sent |
Press Esc while recording or transcribing | Cancel |
| Hold, then swipe up ≥72 px, then release | Cancel (the panel turns red first) |
| A plain click, a drag, or selecting text | Nothing happens — the gesture never arms |
If the draft changed while recognition was running, the transcript is kept in a small lower-right chip; click it to insert at the current caret.
Everything lives at the top of client.js; there is no build step, so edit and reload.
| Constant | Default | Meaning |
|---|---|---|
HOLD_MS | 350 | How long the press must stay still |
ARM_TOLERANCE_PX | 10 | Movement that disarms the gesture |
CANCEL_DISTANCE_PX | 72 | Upward travel that arms "release to cancel" |
MIN_SECONDS | 0.35 | Recordings shorter than this are dropped |
NOTICE_MS | 2800 | How long a one-line notice stays |
| Concern | Mechanism |
|---|---|
| Where it renders | One entry in the conversation.input.overlay slot — a floating layer inside the composer card |
| Gesture surface | node.closest('[data-composer-card]') from its own node, listening in the capture phase |
| Never disturbing typing | The layer is pointer-events: none; nothing is preventDefaulted before the hold threshold |
| Recording | getUserMedia + MediaRecorder → OfflineAudioContext resample to 16 kHz mono → 44-byte PCM16 WAV |
| Recognition | ctx.remote.speech.transcribe({ audioBase64 }); provider and language come from the Host config |
| Draft insertion | The slot's own inputActions.captureInsertion() / insertText(text, span), guarded by draftRev |
| Styling | Only --dsw-alias-* theme tokens, so light and dark both work |
| Text | Registered through ctx.locale (zh, en) |
conversation.input.overlay slot and the
speech remote — which can change between Harness releases.DeepSeek Harness has no plugin marketplace and no submission process: the official
repository states that it cannot accept external pull requests at the moment, and the
mechanism it points third-party plugins at is the GitHub topic
dsh-plugin. This repository carries that topic, which
is how it is meant to be found.
install_bundle also accepts an npm package name, so if this plugin is ever published, a user
could install it by name instead of cloning. Publishing to npm would require renaming the
package from @local/... to a scope the publisher owns.
plugin_manager → action: remove_bundle → target: @local/dsh-hold-to-talk
The plugin is installed as a link to the clone, so deleting the clone directory after removing the bundle is safe.
providerId, so whichever provider the Host has selected is used; the
default configuration transcribes locally with SenseVoice. Check Settings → Plugins → Voice
input before dictating anything sensitive.按住鼠标说话 —— 在 DeepSeek Harness 的输入框里长按鼠标即可语音输入，松开后转写并插入草稿， 不会自动发送。
本插件不自带语音识别引擎，它复用 DeepSeek Harness 已有的那套。所以下面两条必须成立：
官方的语音输入 bundle 处于启用状态。
它是随 DSH 一起发布的 @deepseek-ai/dsh-experimental-voice-input-bundle（设置 → 插件 →
「语音输入」/ Voice input）。本插件调用的 speech 远程命名空间正是由它挂载的。
一旦禁用它，本插件会静默不注册任何东西，输入框行为完全恢复原样。
识别模型已经准备好。
官方 bundle 使用 SenseVoice 在本机转写。首次使用需要下载模型文件（约 240 MB）到
~/.dsh/speech-to-text/sensevoice/models（model.int8.onnx、tokens.txt、silero_vad.onnx）。
打开语音输入 bundle 的详情页点一次「下载并准备」，或者直接点一次内置麦克风按钮并按引导操作。
官方 bundle 不支持流式转写，所以文字只会在你松开之后出现。音频是临时的：不会成为 Session 事件或附件，只有你之后正式提交的文字才会被记录。
DSH 插件从包目录安装，先克隆：
git clone https://github.com/jryang1997/dsh-hold-to-talk.git
然后二选一。
方式 A：让 Agent 装（推荐，桌面版用这个）
在任意会话里对 Agent 说：
用
plugin_manager（action: install_bundle）安装<克隆下来的绝对路径>这个 bundle。
它会为当前 profile 执行安装，并返回 application: applied 表示已生效。
方式 B：命令行
dsh plugin --profile <profile> add <克隆下来的绝对路径>
桌面版应用独占它的
desktopprofile，所以在桌面版里请用方式 A。
方式 C：验证
刷新页面（Ctrl+R），鼠标移到消息输入框上，右下角会出现提示行。
如果没有出现，打开浏览器控制台确认 Host 半是否激活。
| 手势 | 结果 |
|---|---|
| 鼠标移入输入框 | 提示行约 0.6 秒慢慢浮现 |
| 按住不动约 0.35 秒 | 整张输入框变成录音面板 |
| 松开 | 转写文字插入光标处，不发送 |
录音中或识别中按 Esc | 取消 |
| 按住后上滑 ≥72 px 再松开 | 取消（面板会先变红提示） |
| 单击、拖拽、拖选文字 | 什么都不发生 —— 手势根本不会激活 |
如果识别期间草稿被改动过，转写结果会保留在右下角的小胶囊里，点一下即可插入到当前光标。
全部在 client.js 顶部；没有构建步骤，改完刷新即可。
| 常量 | 默认 | 含义 |
|---|---|---|
HOLD_MS | 350 | 按住多久才算语音输入 |
ARM_TOLERANCE_PX | 10 | 超过这个位移就不激活 |
CANCEL_DISTANCE_PX | 72 | 上滑多少像素进入"松手取消" |
MIN_SECONDS | 0.35 | 短于此长度的录音直接丢弃 |
NOTICE_MS | 2800 | 一行提示停留多久 |
| 环节 | 机制 |
|---|---|
| 渲染位置 | conversation.input.overlay 座位的一个条目 —— 输入框卡片内的浮层 |
| 手势面 | 从自身节点 closest('[data-composer-card]') 拿到卡片，捕获阶段监听 |
| 不干扰打字 | 浮层默认 pointer-events: none；达到长按阈值前不做任何 preventDefault |
| 录音 | getUserMedia + MediaRecorder → OfflineAudioContext 重采样到 16 kHz 单声道 → 44 字节 PCM16 WAV |
| 识别 | ctx.remote.speech.transcribe({ audioBase64 })，provider 与语言由 Host 配置决定 |
| 写草稿 | 座位自带的 inputActions.captureInsertion() / insertText(text, span)，带 draftRev 校验 |
| 样式 | 只用 --dsw-alias-* 主题令牌，明暗主题都正常 |
| 文案 | 通过 ctx.locale 注册（zh、en） |
conversation.input.overlay 座位与 speech 远程命名空间 ——
它们可能随 Harness 版本变化。DeepSeek Harness 没有插件市场，也没有投稿流程：官方仓库明确表示目前不接受外部 PR，
它为第三方插件指出的发现机制是 GitHub topic
dsh-plugin。本仓库打上了这个 topic，这就是它被找到的方式。
install_bundle 也接受 npm 包名，所以如果将来发布到 npm，用户就能直接用包名安装、不必克隆。
发布 npm 需要把包名从 @local/... 改成发布者自己拥有的 scope。
plugin_manager → action: remove_bundle → target: @local/dsh-hold-to-talk
插件是以链接方式安装的，卸载后删掉克隆目录即可。
providerId，
用的是 Host 已选定的那个 provider —— 默认配置是本机 SenseVoice 本地转写。
在口述敏感内容前，请先确认「设置 → 插件 → 语音输入」里的选择。