v0.2.0

Highlights

This release focuses on agent intelligence improvements and prompt architecture refactoring for better LLM performance and maintainability.

New Features

• Unified Tool Registry: Dynamic tool registration with ToolDescriptor interface for consistent tool documentation and execution
• Shared Streaming Execution: Common streaming logic across CommandTool and tfcmd via CommandExecutor
• Enhanced Error Feedback: Tool failures now include command output in LLM feedback, enabling better error diagnosis

Improvements

Agent & LLM

• Stricter JSON Response Format: Enforced JSON-only responses with explicit rules against markdown wrapping
• Reorganized System Prompt: Condensed from ~190 lines to ~90 lines with clear markdown sections
• Tool-Specific Instructions: Moved CLI rules (flag groups, validation, SSH keys) to respective tool definitions

GUI

• Preserve Profiles on Logout: User profiles now persist after logout (only credentials cleared)
• Error Handling in Workflows: Fixed bug where workflow steps disappeared on error; now shows error as additional step
• Answer/Explanation Separation: Added empty line between answer and explanation in responses

Bug Fixes

• Fixed: Workflow steps disappearing when error occurs mid-workflow
• Fixed: LLM not receiving command output on tool failure (only got "exit status 1")
• Fixed: Answer and explanation concatenated without separation