OpenCode Interrupt Plugin

Streaming TTS + voice interruption for OpenCode. Speaks responses as they arrive and detects when you talk over it. Walkie-talkie mode lets you hold the spacebar to redirect the model on the fly.

Installation

Add both the server plugin and the TUI plugin to your ~/.config/opencode/opencode.json:

{
  "plugin": [
    "opencode-interrupt-plugin",
    "opencode-interrupt-plugin/tui"
  ]
}

The server plugin handles TTS streaming, voice overlap detection, and interrupt injection. The TUI plugin adds the walkie-talkie keybinding (space).

Usage

Walkie-Talkie Mode

Hold down the spacebar while speaking, then release it to redirect the model:

1. The model is generating a response — TTS is playing
2. Press and hold spacebar — generation aborts, TTS stops, mic starts recording
3. Speak your correction while holding spacebar
4. Release spacebar — audio is transcribed via Whisper and sent as your next message
5. The model responds to your correction

[Model speaking]  →  hold space → speak correction → release space → model redirects

Requires whisper.cpp with the base model installed. Run the one-command install script:

bash scripts/install-whisper.sh

Or set OPENAI_API_KEY in your environment as a fallback.

Voice Interruption (server plugin)

When TTS is playing, just speak — the plugin detects your voice, stops TTS, and marks the session for correction injection.

Configuration

Environment Variables

• OPENAI_API_KEY — optional fallback for Whisper transcription (used if whisper CLI is not found)
• WHISPER_MODEL — path to whisper.cpp model file (default: ~/.local/bin/ggml-base.bin)
• EDGE_TTS_VOICE — TTS voice (default: en-US-AriaNeural)

How It Works

• Server plugin: Monitors mic via sox (continuous PCM pipe), detects voice during TTS playback, injects correction context into the next LLM request
• TUI plugin: Registers a space keybinding in OpenCode's TUI keymap; key repeats while held keep the recording active; 300ms of silence after release triggers transcription and redirect