v26.7.16

Architecture Upgrade

• Docker Sandbox Execution (run_terminal_command): Upgraded terminal execution tool to support an isolated environment (envType: "docker"). When enabled, commands run inside an ephemeral Docker container (default: python:3.11-slim), safely sandboxing code execution and file manipulation from the host OS.
• Trajectory Generation (TrajectoryLogger): Implemented automated dataset generation. The agent loop now hooks into TrajectoryLogger to save execution histories as JSONL files (~/.nyxora/trajectories.jsonl). Trajectories are formatted with <tool_call> and <tool_response> tags, ready for next-gen tool-calling model fine-tuning.
• Subagent Delegation (delegate_subagent): Empowered the core agent with the ability to recursively spawn isolated clone instances. Using the new delegate_subagent skill, the agent can dispatch long-running or complex tasks in parallel without polluting the main context window.
• Terminal User Interface (TUI): Developed a native TUI using blessed for non-GUI / VPS users. Features a multi-pane layout (Chat, Active Tools, Subagents) and utilizes SSE streaming to render real-time LLM responses identically to the Web Dashboard. Launch via nyxora tui.

Features & Platform Expansion

• Nyxora Desktop MVP: Successfully engineered and launched a native, standalone Desktop Application using Electron, Vite, and React.
  • The desktop client seamlessly mirrors the sleek aesthetic of the web dashboard while providing a native OS-level experience.
  • Introduced the new nyxora desktop CLI command. Running this command autonomously fetches dependencies, builds the renderer, and launches the Electron wrapper.
  • Zero-Touch Daemon Sync: Implemented an intelligent lifecycle manager within the Electron main process. The app autonomously bootstraps the Nyxora background daemon (nyxora start) upon launch and securely tears it down when closed.
  • Race Condition & Auth Security Fix: Engineered a robust asynchronous waiting mechanism and internal Vite proxy (vite.config.ts) to flawlessly synchronize the runtime.token injection between the backend daemon and the frontend UI, completely eradicating startup race conditions and CORS origin blocks.

Self-Learning & Continuous Improvement

• Proactive Tool-Iteration Trigger: Refactored the core agent loop (osAgent.ts) to proactively trigger the background review engine (ml-engine) after every 10 tool iterations, rather than waiting for an arbitrary session end or a 3-minute idle timeout.
• Aggressive Skill Creation: Upgraded the _COMBINED_REVIEW_PROMPT in ml-engine to explicitly command the AI to be highly active in creating and patching skills. Frustration and user corrections are now treated as "FIRST-CLASS skill signals".
• Deep Context Window: Expanded the history payload sent to the background review engine from 30 messages to 100 messages, ensuring the background reviewer has the complete context of long debugging struggles.
• UI Learning Notifications: The core gateway now awaits the background review's execution result. When the AI successfully creates or patches a skill, a 💾 Self-improvement review system message is automatically injected into the chat UI, providing immediate transparency that the agent is learning.
• Persistent System Corrections: Upgraded the Reflection Engine's schema (reflection.ts) to extract and permanently store system_correction memories. The AI now actively remembers explicit user rules regarding tool limitations or preferred workflows across all future sessions.
• UI Tool Status Sanitization: Fixed a visual bug where raw JSON tool arguments and execution states were leaking into the chat interface. Upgraded getToolLabel (osAgent.ts) to aggressively truncate tool outputs and provide clean status indicators.