Voice input not isolated per session + captures external audio

[Ch1ronWang]

Problem

Voice input has two isolation issues:

1. No session isolation — voice is global across all sessions

activeVoice is a module-level variable in packages/opencode/src/cli/cmd/tui/component/prompt/index.tsx:91:

// Module-level voice state: survives component remounts and route changes
let activeVoice: {
  handle: Voice.StreamingHandle
  ...
}

When multiple sessions are open, they all share this single global voice handle. Speaking into the mic in one session writes text into all sessions' input boxes, because appendText/setText references get overwritten by whichever prompt component last mounted.

2. No audio source isolation — external sounds are captured

The recorder (packages/opencode/src/cli/cmd/tui/util/voice.ts:22-31) uses arecord or sox to capture from the default audio input device. There is no distinction between:

• The user's voice (intended input)
• System audio from playing videos, music, or other applications

The VAD (vad.ts) detects speech activity from whatever audio comes through, so any audio source with speech-like characteristics gets transcribed and written to the input box.

Expected behavior

1. Each session should have its own independent voice instance — enabling/disabling voice in one session should not affect others.
2. Voice should only capture from the microphone input, not from system/desktop audio output.

Environment

• MiMoCode CLI TUI
• Linux (WSL2) with PulseAudio
• Multiple sessions open simultaneously