dsh-client-ui-voice-input

Composer voice-input control for DeepSeek Harness: a mic button in the composer tool row that transcribes browser Web Speech recognition into the draft via the official inputActions.setDraft write path.

• Zero backend, zero key — recognition runs entirely in the browser (Chrome/Edge → Google, Safari → Apple Web Speech).
• Accuracy: interim segments replace (no duplication), final segments commit, the pre-existing draft is preserved.
• Convenience: click to start / click to stop, listening state (aria-pressed + title), unsupported browsers degrade to a disabled button with a hint.
• Configurable: language (default zh-CN), continuous, interimResults.

Source distribution (dsh-plugin topic). The browser bundle is produced by the DSH web build's tsdown client preset; the package ships source + tests for a DSH checkout.

How it works

MicButton (conversation.input.left)
  → window.SpeechRecognition / webkitSpeechRecognition
  → TranscriptAccumulator (final + interim, no duplication)
  → inputActions.setDraft(base + transcript)

Install in a DSH checkout

Mount in the web-app browser roster (packages/bundle/web-app/cordis.patch.yml):

- id: ui-voice-input
  name: '@zhangbo-cn/dsh-client-ui-voice-input'

Build the client bundle with the repo's tsdown preset:

pnpm --filter @zhangbo-cn/dsh-client-ui-voice-input run bundle

Tests

npx vitest run   # 9 tests: transcript accuracy, toggling, unsupported state, apply wiring

License

MIT