dsh-voice-hotkey

Minimal push-to-talk voice input for DeepSeek Harness (dsh): press a key, talk, stop talking — the text lands in the composer and Enter sends it. No mouse on the way.

Recording, transcription and writing into the draft are still done by the official Voice Input bundle. This plugin only presses its buttons for you, and fills three gaps it leaves open.

中文

Three things

Voice hotkey	Default Alt+V, remappable. Tap to start, tap to stop; works anywhere on the page, no need to click the composer first.
Silence auto-stop	Once speech is detected, N seconds of silence ends the recording. Adjustable 0.5–10 s on a slider, or switch it off.
Focus the composer	Default Ctrl+Alt+I, one key puts the caret back at the end of the composer. DSH has no such command of its own.

What other tools use to focus their input

Tool	Shortcut	Why it can't be copied as-is
Cursor	Ctrl+L / Cmd+L	The browser address bar owns Ctrl+L, so it breaks on the web
VS Code · GitHub Copilot Chat	Ctrl+Alt+I	Free everywhere, safe on the web
VS Code · Claude Code	Ctrl+Esc / Cmd+Esc	On Windows this opens the Start menu
this plugin	Ctrl+Alt+I	Matches Copilot; same on web and desktop

Why focusing deserves its own key

The official voice plugin remembers the caret position at the moment recording starts and writes the transcript back there. So:

• pressing the hotkey inside the conversation puts the caret at the end of the composer first (otherwise the text has nowhere sensible to go);
• but if you clicked elsewhere in the meantime — sidebar, plugins page, another window — the caret is gone and you need the mouse again.

dsh ships no "focus the input" command, which is exactly where people get stuck. This plugin makes it a key: press it anytime and the caret is back in the composer, ready for Enter.

Turning a hotkey off

Both hotkeys can be disabled:

• click the key to start recording, then press Esc to clear it to Not set (Backspace / Delete work too).

When a shortcut shows "Not set" that action is fully disabled — no listener, no key swallowed. A shortcut is only saved once you press a regular key with Ctrl/Alt/Meta; modifiers alone just show a hint.

Requirements

• The official Voice Input bundle must be enabled. This plugin clicks its buttons; without it nothing happens.
• A browser with microphone access. Audio still travels the official path to the dsh server; this plugin adds no network requests.

Desktop app

The dsh desktop app is an Electron shell around the same web UI, so the plugin works there unchanged: hotkey, auto-stop and focus all live in the renderer process and need no extra permissions.

Two known differences, both handled:

• localStorage can be blocked under file:// or storage partitioning in Electron — settings then fall back to memory for the current session instead of throwing;
• microphone permission is granted by the desktop app; if the first recording is silent, allow the mic in the OS privacy settings.

Install

dsh plugin --profile web add dsh-voice-hotkey

Or from a local checkout:

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

Reload the page. The defaults work right away (Alt+V to talk, Ctrl+Alt+I to focus); change them under Settings → Voice hotkey.

Settings

Available both on the settings page and on the plugin's detail page, stored in localStorage (memory fallback on desktop), applied immediately:

• Voice input hotkey — Alt+V by default. Must include Ctrl / Alt / Meta so it never swallows normal typing.
• Focus the composer — Ctrl+Alt+I by default. The two hotkeys cannot be the same; a conflict is reported inline.
• Stop automatically when I stop talking — off means pure manual mode: press the hotkey again to finish.
• Silence before stopping — 0.5–10 s slider.

Limits

• Silence detection is energy-based, not semantic: pausing longer than the threshold ends the recording.
• The threshold adapts to ambient noise; in a loud room, hold-to-talk is steadier.
• No live/streaming captions — the official transcription returns the whole utterance at once.

License

MIT