feat(voice): add voice mode proof-of-concept with Live API integration

[himanshu748]

Summary

Proof-of-concept implementation of voice mode for Gemini CLI, demonstrating the integration pattern for the Gemini Live API's bidirectional WebSocket streaming. This lays the groundwork for hands-free multimodal voice interaction (GSoC 2026 Project 11).

Related to #18067

What's included

Core module (packages/core/src/voice/)

• VoiceService class wrapping @google/genai Live API (GoogleGenAI.live.connect())
• Full client→server message support: sendText(), sendAudio(), sendAudioStreamEnd(), sendToolResponse(), sendInterrupt()
• Typed event emitter for all server→client messages: text responses, audio chunks, input/output transcriptions, tool calls, tool call cancellations, goAway, state changes
• State machine: IDLE → CONNECTING → CONNECTED → LISTENING ↔ RESPONDING
• VoiceConfig with sensible defaults (model, response modality, voice, VAD, sample rates)
• buildSpeechConfig() helper for SDK speech configuration
• 54 unit tests covering lifecycle, messaging, error handling, and event dispatch

CLI integration (packages/cli/)

• VoiceMode Ink component: state indicator, ASCII waveform visualization, transcription display, keyboard controls (ESC to exit, m to mute, Space to interrupt)
• /voice slash command using OpenCustomDialogActionReturn pattern (same pattern as /hooks)
• Wired into BuiltinCommandLoader

What's NOT included (future work)

• Audio I/O: Platform-specific microphone/speaker bindings (e.g., naudiodon/PortAudio) are not integrated. The VoiceService backend is fully functional for WebSocket streaming — it just needs audio bytes piped in.
• Tool execution bridge: The TOOL_CALL events are emitted but not yet routed to the existing ToolRegistry
• Session resumption: The Live API supports sessionResumption but it's not wired up yet
• Settings integration: No voice key in settings schema yet

Architecture decisions

1. Live API over REST+STT: Uses the native bidirectional WebSocket API (BidiGenerateContent) rather than bolting STT/TTS onto the existing HTTP streaming pipeline. This gives us server-side VAD, barge-in support, and lower latency.

2. Typed event emitter: Rather than extending EventEmitter (which loses type safety), VoiceService uses a composition pattern with typed on<E>() / off<E>() methods and a VoiceEventMap interface.

3. LiveServerMessage typed parameter: handleServerMessage() accepts the SDK's LiveServerMessage type directly (from onmessage callback), avoiding as casts and eslint-disable suppressions.

4. Custom dialog pattern: The /voice command follows the established OpenCustomDialogActionReturn pattern (same as /hooks), so it integrates cleanly with the existing command system.

Testing

npx vitest run packages/core/src/voice/voice-service.test.ts
# 54 tests passing

[github-actions]

You already have 7 pull requests open. Please work on getting existing PRs merged before opening more.

[anowardear062-svg]

Thanks for the update