🚀 SoloEngine v0.2.1 Release — LLM Provider Architecture Refactoring, Stop Mechanism & Streaming Output Optimization

[v0.2.1] - 2026-06-03

🚀 Added

• SubAgent Token Tracking — SubAgent completion events now carry token usage data. The frontend Agent group view displays real-time token consumption per Agent.
• Project Isolation — All file-operation endpoints accept an agentic_flow_id parameter, isolating sandbox file systems per flow instance.
• MiMo Model Support — Added Xiaomi MiMo as a new LLM provider (LLMProvider.MIMO), implemented via an OpenAI-compatible interface.

🔧 Improved

• LLM Stop Mechanism — Removed blocking await response.aclose() calls from individual model layers. Cancellation now raises CancelledError immediately, delegating cleanup to the upper layer and reducing stop-response latency.
• LLM Streaming Callbacks — stream_callback no longer buffers chunks until the loop finishes. It now pushes incremental deltas in real time, making intermediate output visible on the frontend as it arrives.
• Standalone Provider Configuration — LLM provider settings have been extracted from hardcoded constants into data/config/llm_providers.json. The new get_provider_config() interface supports custom providers, default models, and model lists.
• Unified Model Creation Interface — All providers now pass connection parameters through client_kwargs. Added support for frequency_penalty and presence_penalty.
• Centralized LLM Configuration — The canvas introduces a configMap for centralized LLM management. Components read config from the store. The /providers API endpoint now returns a color field alongside provider data.
• Streaming Data Updates — Streaming output now batches UI updates via requestAnimationFrame instead of direct setState calls. Added duplicate-block detection. useRunWebSocket respects active stop signals.
• Message & File-Change Management — Undo operations follow a unified path and clean up associated file_changes and blob records simultaneously. Removed CAS counting — cleanup now follows reference relationships directly.
• Canvas Interactions — Canvas connectors now render arrow markers, making flow direction explicit.
• Run Panel — SubAgent groups display token counts with improved nested layout. The file browser resets its state automatically when switching projects.
• Project Management — get_recent_projects filters directly by agentic_flow_id, correcting data-field mapping issues.

🗑️ Removed

• CAS reference-counting methods (ref_count, decrement_ref_count, cleanup_orphan_blobs, etc.)
• SoloAgentConfig.memory field
• frontend/src/config/providerDefaults.ts — its functionality is now handled by canvasStore.configMap
• frontend/src/store/runProjectStore.ts — merged into runPanelStore.ts
• Hardcoded model lists, default-model mappings, and redundant string constants from LLMFactory

---

🤝 Join Us

We're looking for like-minded contributors who share our passion for SoloEngine and Agentic AI. Every contribution — from a typo fix to a full feature — makes SoloEngine better.

📝 Contributing Guide · 💬 Discussions · 📧 Contact Us