dsh-voice-ptt
DeepSeek Harness 的按键听写：在编辑框中按住一个按键，松开后，转录文本会插入草稿。
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:eicon-xyz/dsh-voice-ptt#97e1d3602a37a3165d0b3edd87fa62a6013d2793Voice Ptt 以 dsh-voice-ptt 发布，当前版本为 0.1.0。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。
点击下方 Badge 复制 Markdown，粘贴到 README 即可。
验证 package.json 声明的 GitHub 仓库，即可管理这个公开页面。认领后，Hub 会优先安排当前版本的安全扫描，并在通过后公开展示结果。
认领这个 Plugin →English | 中文
Push-to-talk dictation for DeepSeek Harness: hold a single key in the composer, speak, release — the transcript is inserted into your draft.
No chord to memorise. Hold F2, release, done.
Hold F2 to dictate.Recording — release F2 to finish.While dictating:
Esc, or press any other key (say, the letter you meant to type), to cancel and discard the recording.The hint is a button. Click it to expand the candidate keys, click one, and it applies immediately and is remembered (in localStorage).
| Key | Notes |
|---|---|
| F2 (default) | Function key: types no character, minimal conflicts |
| F4 | Function key, further right |
| Insert | Right hand on full-size keyboards; absent on many laptops |
| Right Ctrl | Comfortable for the right pinky |
| Right Shift | Right pinky; on Windows, five taps can trigger Sticky Keys |
Modifier keys work properly: the plugin knows the hold key is a modifier, so it does not mistake it for a chord and pass the event through.
The hint, the status line, and these key names follow your DSH language (en / zh).
To offer a different key, edit KEY_OPTIONS at the top of lib/client.js (a {code, labelKey} array, code spelled as KeyboardEvent.code, labelKey naming an entry in the zh / en dictionaries at the top of the same file).
dsh plugin --profile web add dsh-voice-ptt
The package declares dsh.bundle.patch, so the command appends it to the profile's bundle stack and mounts it — no hand-editing of cordis.patch.yml. From a Git checkout, pin the revision:
dsh plugin --profile web add 'github:eicon-xyz/dsh-voice-ptt#<commit>'
Requires the speech recognizer. Install and enable @deepseek-ai/dsh-experimental-voice-input-bundle and prepare a model once (roughly 230 MB). If the recognizer is not ready, holding the key reports a visible reason in the status line instead of failing silently.
# mount the checkout directly instead of installing
ln -sfn "$PWD" "$DSH_HOME/profiles/web/node_modules/dsh-voice-ptt"
# then add to the profile's own cordis.patch.yml:
# - insert:
# - id: dsh-voice-ptt
# name: 'dsh-voice-ptt'
Remove that manual mount line before switching to the bundle channel, or the row is mounted twice.
The official shortcut service (@deepseek-ai/dsh-client-shortcuts) dispatches keydown only — there is no keyup event to end a hold. It also rejects modifier-free bindings (bindingIssue returns modifier-required) and, on the web, restricts chords to a platform allow-list.
So "hold one key and talk" cannot be expressed as a shortcut command. This plugin owns the keyboard instead: document-level keydown starts, keyup finishes. The trade-off is that it cannot appear in Settings → Shortcuts, which is why the key picker lives in the composer.
| Situation | Behaviour |
|---|---|
| Focus is not in a composer | Does not start recording |
| Any other key pressed while recording | Cancels the take; that key works normally (hold Right Ctrl, press A → Ctrl+A still applies) |
| A dialog or menu is open | Yields the key |
Modified press (Ctrl+F2, Alt+F2, …) | Not intercepted at all |
| Focus is in a terminal | Not intercepted (xterm needs function keys) |
inputActions (captureInsertion + insertText). The plugin registers a resident entry in conversation.input.left to record each session's actions; that same entry renders the status line, the focus hint, and the key picker.MediaRecorder and resamples through OfflineAudioContext to 16 kHz mono PCM16 WAV — the same format the official voice-input plugin sends — then calls the host's speech.transcribe.dsh.client.| Path | Purpose |
|---|---|
lib/client.js | Browser half: keys, recording, transcription, status line, key picker |
lib/index.js | Host half (intentionally empty, so the bundle row is mountable) |
cordis.patch.yml | dsh.bundle.patch layer the installer merges |
test/ | Headless-Chrome end-to-end tests |
They drive a real browser against a running DSH web profile, with Chrome's fake microphone fed the bundled fixture:
npm install # provides ws
export PTT_TOKEN=<token from the dsh web URL>
node test/e2e.cjs # hold → release → transcript inserted
node test/picker.cjs # picker opens, a new key applies and persists
node test/modifier-key.cjs # Right Ctrl records; Ctrl+A is not hijacked
node test/nonempty-draft.cjs # works with a non-empty draft, text is appended
node test/modal-yield.cjs # yields while a dialog is open, works again after
test/make-fixture.cjs regenerates test/fixtures/speech-16k.wav from any audio Chrome can decode:
node test/make-fixture.cjs input.mp3 # needs Chrome; no ffmpeg required
KEY_OPTIONS.MIT