dsh-composer-dictation

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.

┌──────────────────────────────────────────────────────┐
│                                                      │
│  ⌄  ＋  权限  计划    🎤 按住鼠标语音输入文字  [◉] [➤] │
│                       ↑ the hint tucks into the tool  │
│                         row — never over your draft   │
└──────────────────────────────────────────────────────┘

        press and hold ≈0.3 s ↓  (a ring draws at the pointer)

                  ╭──────────────────────────────╮
                  │  ●  ▁▂▅▇▅▃▂▁▂▄▆▄▂▁   松开完成  │   ← the bubble floats
                  ╰──────────────────────────────╯     above the card
┌──────────────────────────────────────────────────────┐
│  your draft is still here — readable, typeable       │
│  ⌄  ＋  权限  计划                          [◉] [➤] │
└──────────────────────────────────────────────────────┘

Recording does not take the composer over. The bubble is a floating capsule above the card, and the card is left completely alone, so you can keep reading and editing what you were writing while you dictate.

---

Requirements — read this first

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:

1. 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.

2. 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.

Install

A. Install straight from GitHub — no clone needed

The plugin manager accepts a git URL as an install target, so the whole install is one command. From the CLI:

dsh plugin --profile <profile> add github:jryang1997/dsh-composer-dictation

or ask the agent in any session:

Install the bundle github:jryang1997/dsh-composer-dictation with plugin_manager (action: install_bundle).

A git install is pinned to the commit it was installed from. To pin it to a release instead — so you know exactly what you are running, and can move deliberately — name the tag:

dsh plugin --profile <profile> add github:jryang1997/dsh-composer-dictation#v1.1.0

See Releases for what changed in each one, and CHANGELOG.md for the detail.

B. Install from a clone

Cloning first also works, and is the route to take when you want to modify the plugin locally:

git clone https://github.com/jryang1997/dsh-composer-dictation.git

Then install the directory: ask the agent to run plugin_manager (action: install_bundle) with the clone's absolute path as target, or run

dsh plugin --profile <profile> add <absolute path to the clone>

The Desktop application owns its desktop profile exclusively, so there ask the agent rather than using the CLI.

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.

Usage

Gesture	Result
Hover the input box	A hint appears in the tool row, between the mode chips and the model selector
Press and hold ≈0.3 s without moving	A ring draws at the pointer, then a capsule floats up above the card: a live dot, the level waveform, and one short word
Release	Transcript is inserted at the caret — not sent
Press Esc while recording or transcribing	Cancel
Hold, then drag up ≥48 px or move the pointer off the box	The capsule and the card's outline wash red and the word changes to "release to discard"; release there to drop it, come back to keep it
A plain click, a drag, or selecting text	Nothing happens — the arc retracts and 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.

Tuning

Everything lives at the top of client.js; there is no build step, so edit and reload.

Constant	Default	Meaning
HOLD_MS	300	How long the press must stay still
ARM_TOLERANCE_PX	10	Movement that disarms the gesture
CANCEL_ARM_PX	48	Upward travel that arms "release to discard" (leaving the box arms it too)
CANCEL_RELEASE_PX	38	Where it disarms again — the 10 px band is what stops the capsule flickering
CANCEL_FLICK_PX_PER_S	150	Release speed that overrides position when deciding discard vs. keep
MIN_SECONDS	0.35	Recordings shorter than this are dropped
NOTICE_MS	2800	How long a one-line notice stays
EXIT_MS	180	How long any surface takes to leave
METER_SETTLE_MS	240	How long the level trace keeps draining after you release

Easing, durations and materials are not tuned by hand here — they are read from the host app's own tokens. See Motion.

Motion

The plugin deliberately owns no motion vocabulary of its own; it borrows DeepSeek Harness's, so it reads as part of the app rather than as a guest.

Decision	Where it comes from
Entrance curve cubic-bezier(.16,1,.3,1)	the curve DSH's own menu and preset entrances use
Exit curve cubic-bezier(.4,0,.2,1)	--ds-ease-in-out
140–200 ms durations	inside the app's --ds-transition-duration band
The card's outline in the discard state	--dsw-radius-panel, the composer card's own 28 px — which the app also renders as a squircle via corner-shape: superellipse(1.5)
The capsule's surface	--dsw-specific-menu over --dsw-menu-backdrop-filter: the recipe MenuSurface.module.css uses for every floating layer
The capsule's shadow	--dsw-elevation-prominent, the host's shadow for a surface that floats above content
Level meter	the shipped voice input's waveform: 28 RMS samples in a shift register, redrawn at 20 fps

Four rules hold everywhere, and tests/render.test.mjs enforces them:

1. Recording runs alongside you, it does not take over. Apple's rule is dim to focus, separate to keep flow: a panel that is parallel to what you are doing uses translucency and offset without a scrim. The capsule floats above the card; the card is untouched, so your draft stays readable and typeable for the whole recording.
2. Entry and exit are transitions, never keyframes. Entry rides @starting-style, exit a data-leaving attribute. Anything the user reverses mid-flight — a hold that turns into a drag, a bubble reopening during its own exit — retargets from where it actually is instead of restarting.
3. Only transform, opacity, translate and scale are animated. The meter used to animate height, which cost a layout per bar per frame.
4. The gesture is a continuous quantity, not a boolean. How far you have dragged toward "discard" is one number (--dsh-htt-cancel) that drives the red wash, the ✕ glyph, the card's hairline and the meter's retreat together, so the state is steerable rather than switched.

Text is treated as a last resort: the capsule shows a word only while a release would actually discard the recording. While everything is fine, the live dot and the waveform say it without any words at all.

How it works

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
Hint placement	Measured from the tool row: the hint is aimed at the gap in front of the row's trailing group, so it can never cover the draft or a control
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 plus the host's own surface/material/radius tokens, so light, dark and prefers-reduced-transparency all work
Text	Registered through ctx.locale (zh, en)

Interface-by-interface notes, verified against the shipped packages, live in docs/design.md.

Known limitations

• Pointer-only. The long press is the only entry point; there is no keyboard equivalent yet, so the feature is not reachable by keyboard alone. A focusable trigger is the obvious next addition.
• Mouse-first. Touch input is untested: on a touch screen a long press also drives text selection, so the thresholds would probably need tuning there.
• The hint needs room. It lives in the tool row's empty middle, so on a very narrow composer — where the mode chips and the model selector leave no gap — it is dropped rather than allowed to overlap a control.
• It leans on two internal DSH interfaces — the conversation.input.overlay slot and the speech remote — which can change between Harness releases.

Development

No dependencies and no build step. npm run check parses both halves, validates the bundle manifest, checks the locale dictionaries for key and placeholder parity, and renders the client component across every gesture state — including the motion rules from Motion, which would otherwise regress silently. The same checks run in CI.

Distribution and discovery

DeepSeek Harness has no official 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.

Community directories index that topic automatically — notably dsh-market (dsh.market), whose daily crawler collects repositories tagged dsh-plugin, so no submission is needed there either.

install_bundle also accepts an npm package name, but this plugin is not published to npm. Installing from the git URL above is the supported route, needs no registry account, and is the one this repository is tested against.

Updating

A git install is pinned to the commit it was installed from, so a new release needs one reinstall — there is no auto-update:

plugin_manager → action: remove_bundle → target: @jryang1997/dsh-composer-dictation
plugin_manager → action: install_bundle → target: github:jryang1997/dsh-composer-dictation#v1.1.0

Removing first matters: re-installing over an existing row can report ambiguous-install, because the dependency spec itself has not changed.

Reload the page afterwards. The client half is a browser module, and it keeps the copy it already loaded until the page is refreshed.

Changelog

Every release is documented in CHANGELOG.md, in Keep a Changelog form, and tagged so an install can be pinned to it. The short version:

Version	What it was about
1.1.0	Recording stopped taking the composer over: a floating capsule, an acknowledged press, real exits, and a motion vocabulary borrowed from the host
1.0.0	The hold-to-talk gesture itself, and everything it needs to be safe to use

Uninstall

plugin_manager → action: remove_bundle → target: @jryang1997/dsh-composer-dictation

A directory install (route B) links to the clone, so deleting the clone after removing the bundle is safe. A git install (route A) lives inside the profile and needs no cleanup.

Privacy and disclaimer

• Audio is captured locally, handed to the Host's speech service over the local API, and discarded. It never becomes a session event or an attachment.
• If the Host is configured with a cloud speech provider, that audio leaves the machine. This plugin sends no 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.
• The plugin makes no network request of its own and stores nothing.
• Unofficial community plugin. Not affiliated with, endorsed by, or supported by DeepSeek; "DeepSeek Harness" is named only to describe compatibility. The code here is original and carries no code from the Harness packages.

License

MIT

---

---

中文

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。

前置依赖（先读这一节）

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：

1. 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。

2. 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。

安装

方式 A：直接从 GitHub 安装，无需克隆

插件管理器接受 git 地址作为安装目标，整条命令就能装完。命令行：

dsh plugin --profile <profile> add github:jryang1997/dsh-composer-dictation

或在任意会话里对 Agent 说：

用 plugin_manager（action: install_bundle）安装 github:jryang1997/dsh-composer-dictation 这个 bundle。

方式 B：从克隆安装

先克隆再装目录也可以，想改本地代码时用这条路：

git clone https://github.com/jryang1997/dsh-composer-dictation.git

然后安装目录：让 Agent 执行 plugin_manager（action: install_bundle，target 填克隆目录的 绝对路径），或运行：

dsh plugin --profile <profile> add <克隆下来的绝对路径>

桌面版应用独占它的 desktop profile，所以在桌面版里请对 Agent 说，而不是用命令行。

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。

使用

手势	结果
鼠标移入输入框	工具行中间（模式按钮与模型选择器之间）出现一行提示
按住不动约 0.3 秒	指针处先画出一个进度环，随后输入框上方浮起一颗胶囊：一个实心状态点、实时波形、四个字
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 ≥48 px，或把鼠标移出输入框	胶囊与输入框外框一起泛红、文字变成「松开丢弃」；在那里松手即丢弃，移回来则保留
单击、拖拽、拖选文字	什么都不发生 —— 进度环自己退回，手势不会激活

录音不会接管输入框。 胶囊浮在卡片上方，卡片本身完全不动，所以你说的整段时间里草稿一直是 可读、可编辑的。

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。

调参

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	300	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_ARM_PX	48	上滑多少像素进入「松手丢弃」（移出输入框同样会进入）
CANCEL_RELEASE_PX	38	退回多少才解除 —— 这 10 px 的滞回带就是胶囊不再闪红闪蓝的原因
CANCEL_FLICK_PX_PER_S	150	松手速度，优先于位置决定「丢弃还是保留」
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
EXIT_MS	180	任何一个临时表面退场所需时间
METER_SETTLE_MS	240	松手后电平轨迹继续消退多久

缓动、时长与材质都不在这里手调 —— 它们直接取自宿主应用自己的令牌，见下节。

动效

插件刻意不发明自己的动效语汇，而是借用 DeepSeek Harness 的，这样它读起来像应用的一部分， 而不是一个外来户。

决定	出处
入场曲线 cubic-bezier(.16,1,.3,1)	DSH 自己的菜单与预设座位入场用的曲线
退场曲线 cubic-bezier(.4,0,.2,1)	--ds-ease-in-out
140–200 ms 时长	落在应用的 --ds-transition-duration 区间内
丢弃态下输入框的那道外框	--dsw-radius-panel，也就是输入框卡片自己的 28 px —— 而应用还通过 corner-shape: superellipse(1.5) 把它渲染成超椭圆
胶囊材质	--dsw-specific-menu 叠 --dsw-menu-backdrop-filter，即 MenuSurface.module.css 给所有浮动层用的那套配方
胶囊投影	--dsw-elevation-prominent，宿主给「浮在内容之上」的表面的阴影档
电平表	官方语音输入的波形：28 个 RMS 采样移位寄存器，20 fps 重绘

四条规则贯穿始终，并由 tests/render.test.mjs 守着：

1. 录音是并行的，不是接管。 苹果的原则是「聚焦用压暗，并行用分区」：跟手头的事并行的 面板靠半透明与位移来分层，不压暗背景。所以胶囊浮在卡片上方、卡片分毫不动，你说的整段 时间里草稿都可读可写。
2. 入场与退场都是 transition，绝不用 keyframes。 入场走 @starting-style，退场走 data-leaving 属性。凡是用户可能中途反转的动作 —— 按住后改成拖拽、胶囊正在退场时又 开始新录音 —— 都从它当前实际所在的位置接着走，而不是重播。
3. 只动 transform / opacity / translate / scale。 电平条原先动的是 height， 等于每帧为每根条付一次 layout。
4. 手势是一个连续量，不是布尔值。 你朝「丢弃」拖了多远是一个数 （--dsh-htt-cancel），红晕、✕ 图标、输入框外框、电平条的退让由它一起驱动，所以状态是被你 操纵的，而不是被切换的。

文字在这里是最后的选项：胶囊只在「松手会真的丢东西」时才显示一个词。一切正常的时候， 状态点和波形不用任何文字就把话说完了。

实现原理

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
提示行位置	从工具行量出来：提示行对齐到工具行尾部控件组前面的空档，因此永远不会压住草稿或某个控件
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，外加宿主自己的表面 / 材质 / 圆角令牌，因此明暗主题与 prefers-reduced-transparency 都正常
文案	通过 ctx.locale 注册（zh、en）

逐接口的源码笔记（对照发行包核实过）在 docs/design.md。

已知限制

• 只能用指针。 长按是唯一入口，目前没有键盘等价操作，纯键盘用户无法触达。加一个可聚焦的 触发按钮是下一步最该做的事。
• 以鼠标为主。 触摸屏未验证：触摸长按同时会驱动文本选择，阈值大概需要另调。
• 提示行需要空间。 它待在工具行中间的空档里，所以输入框特别窄时（模式按钮与模型选择器 之间挤不出空档）它会被直接省掉，而不是允许它压住控件。
• 依赖 DSH 的两个内部接口 —— conversation.input.overlay 座位与 speech 远程命名空间 —— 它们可能随 Harness 版本变化。

开发

零依赖、无构建步骤。npm run check 会解析两个半边、校验 bundle manifest、检查两份文案的 键与占位符是否一致，并把客户端组件在每一种手势状态下渲染一遍 —— 其中包含动效 那四条规则，否则它们会悄无声息地退化。CI 里跑的就是这几项。

分发与发现

DeepSeek Harness 没有官方插件市场，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。

社区目录会自动索引这个 topic —— 比如 dsh-market（dsh.market）， 它的每日爬虫会收集打了 dsh-plugin 的仓库，在那里同样无需投稿。

install_bundle 也接受 npm 包名，所以发布到 npm 之后，用户可以直接用包名安装、不必走 git：

plugin_manager → action: install_bundle → target: @jryang1997/dsh-composer-dictation

更新

git 安装是钉在安装时那个提交上的，没有自动更新，出新版本要重装一次：

plugin_manager → action: remove_bundle → target: @jryang1997/dsh-composer-dictation
plugin_manager → action: install_bundle → target: github:jryang1997/dsh-composer-dictation#v1.1.0

必须先 remove：直接在原行上重装会报 ambiguous-install，因为依赖声明本身没有变化。

装完记得刷新页面。客户端半是浏览器模块，不刷新会继续用已经加载的那份。

变更日志

每个版本都记在 CHANGELOG.md 里（Keep a Changelog 格式）， 并且打了 tag，所以安装可以钉在某个版本上。一句话版：

版本	这一版在解决什么
1.1.0	录音不再接管输入框：悬浮胶囊、按下即有反馈、真正的退场，以及一套借自宿主的动效语汇
1.0.0	长按说话这个手势本身，以及让它安全可用所需的全部东西

卸载

plugin_manager → action: remove_bundle → target: @jryang1997/dsh-composer-dictation

方式 B（目录安装）是以链接方式装的，卸载后删掉克隆目录即可；方式 A（git 安装）装在 profile 内部，无需清理。

隐私与声明

• 音频在本机采集，经本地 API 交给 Host 的语音服务，用完即弃；不会成为 Session 事件或附件。
• 如果 Host 配置的是云端语音服务，音频会离开这台机器。 本插件不指定 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。
• 插件自身不发起任何网络请求，也不存储任何内容。
• 非官方社区插件，与 DeepSeek 无隶属、背书或支持关系；提及「DeepSeek Harness」仅为说明兼容性。 本仓库代码为原创，不含 Harness 各包中的代码。

许可

MIT