Introduce KGAudioTrack and KGAudioRegion as new track/region types
alongside existing MIDI tracks. Users can import WAV, MP3, OGG, FLAC,
and AAC files into audio tracks with full playback, looping, and
solo/mute/volume support.
New core models:
- KGAudioTrack (extends KGTrack with TrackType.Wave)
- KGAudioRegion (extends KGRegion with audio file reference)
- KGAudioPlayerBus (Tone.Gain + ToneBufferSource-based playback)
- KGAudioFileStorage (OPFS media/ directory for binary audio files)
New commands:
- AddAudioTrackCommand (create audio track with player bus)
- ImportAudioCommand (import audio file, create region at playhead,
auto-expand maxBars)
Playback features:
- Audio regions scheduled via Tone.Transport alongside MIDI events
- Loop support with duration capping at loop boundaries
- Resume safety offset to prevent ToneBufferSource timing misses
- Cross-track buffer copy when moving audio regions between tracks
- Proper buffer cleanup on track deletion and project switch
UI changes:
- Separate "+ MIDI" and "+ Audio" buttons in top-left spacer
- Audio track shows upload button instead of instrument selector
- FileImportModal reused for audio file drag-and-drop upload
- Real waveform rendering on audio region canvas
- Green color scheme for audio regions (vs blue for MIDI)
- Pencil icon and resize handles hidden for audio regions
- Cross-type region moves blocked (MIDI ↔ Audio)
- Piano roll blocked for audio regions
- Audio region deletion support
Infrastructure:
- Project structure version bumped to 4 (no-op migration)
- class-transformer discriminators updated for new subtypes
- RemoveTrackCommand updated to clean up audio player buses
- Replace custom XML-based tool parsing (XMLToolExecutor) with OpenAI SDK's
native tool_calls via `openai` npm package (dangerouslyAllowBrowser)
- Consolidate 4 LLM providers (OpenAI, Claude, Gemini, ClaudeOpenRouter)
into a single OpenAI SDK-based LLMProvider compatible with any
OpenAI-style API (OpenAI, OpenRouter, Ollama, vLLM)
- Move agentic tool execution loop from ChatBox into AgentCore
- Update Message type to support tool roles, tool_calls, and tool_call_id
- Update system prompt to remove XML formatting instructions (~45% smaller)
- Remove AttemptCompletionTool (replaced by stop_reason detection),
ThinkTool, and ThinkingTool
- Polish tool descriptions for OpenAI function calling schema compliance
- Normalize base URLs by stripping /chat/completions suffix
Xiaohan-Tian