VoxAI_IDE_v2.2.0

Windows installer release for VoxAI IDE.

Included asset:

• VoxAI-Setup.exe

Highlights:

• Packaged Windows executable build
• Full Inno Setup installer
• Startup and shutdown reliability fixes
• Improved packaged resource/path handling

VoxAI IDE v2.1.0 — Terminal Overhaul, RAG Engine Rebuild & UI Polish

What's New

Claude Code-Style Terminal Mode

The terminal has been completely expanded into a full development CLI. Minimize the GUI to tray and work entirely from the console with 25+ slash commands:

• Model management — /model, /model <name>, /models to switch and list models on the fly
• Full Git suite — /status, /branch, /branches, /log, /diff, /commit, /push, /pull, /fetch, /stash, /checkout
• Project tools — /files, /search, /run, /index, /tokens
• Fuzzy model matching by name or list number

UI Overhaul

• Chat rendering rewritten for a Claude Code-like appearance — clean role labels, collapsible tool results, dimmed thinking blocks
• Input bar redesigned Cursor-style with inline model selector, mode selector, attachment button, and send button
• Icon bar replaces the old toolbar — slim 34px strip with toggle icons and centered project title
• Global theme polished: darker base (#111113), slimmer scrollbars, refined splitter handles, consistent widget styling
• Tool result folding — tool outputs now appear as one-line collapsible summaries instead of flooding the chat

RAG System Rebuild

• Unified Go engine — merged separate server and CLI binaries into a single main.go entry point
• Pre-compiled binary — vox-vector-engine.exe is now built ahead of time, eliminating the slow go run fallback
• Server failure caching — if the HTTP server can't start, it won't retry on every request
• Incremental indexing — hash-based manifest (index_manifest.json) ensures only changed files are re-indexed
• Fixed CLI bugs — correct TopKCandidates (was using token count), consistent doc_id format between HTTP and CLI paths
• Silenced log spam — urllib3, requests, PIL, matplotlib, and embedding loader logs suppressed

Token Optimization

• Automatic truncation of tool outputs, file attachments, and command results
• Old conversation history compressed into recaps
• System prompts condensed
• ~30-50% fewer tokens per request

New IDE Features

• Quick File Switcher (Ctrl+P) — fuzzy search overlay to jump to any file
• Code Outline (Ctrl+Shift+L) — classes, functions, methods sidebar (AST for Python, regex for others)
• Live Change Highlighting — AI-modified lines highlight green (added) and red (removed) in open editor tabs
• Context Window Visualization — color-coded progress bar in the status bar

Build & Release

• GitHub Actions now auto-builds a Windows .exe on every release tag
• Includes Go RAG binary, all Python packages, UI resources, and IronGate web client
• Download the zip below — no Python install required

Bug Fixes

• Chat scrolling: replaced timer-based auto-scroll with signal-based (rangeChanged/valueChanged)
• Background tool execution no longer makes messages appear silent
• Phase/Siege mode now ends with detailed summaries instead of generic "phase completed"
• Settings dialog provider section no longer squashed (now uses tabbed layout)
• Fixed Co-authored-by injection in git commits
• Fixed /dev/null → os.devnull for Windows compatibility
• Fixed content_type NameError in IronGate web client
• Fixed shell injection and path traversal vulnerabilities in terminal mode
• Fixed platform.system() replacing os.uname() for cross-platform file reveal
• Fixed shortcut collision (Ctrl+Shift+O for both Open Project and Code Outline)

Install

From release (recommended):

1. Download VoxAI_IDE-v2.1.0-agentic-windows.zip below
2. Extract and run VoxAI_IDE.exe
3. (Optional) Drop .gguf models into models/llm/ for offline use

VoxAI_IDE

Release Notes: VoxAI IDE v1.7.0 "Agentic"
"The Local-First Autonomy Update"

This release marks a major shift towards fully autonomous, local-first development. VoxAI now runs standard GGUF models natively, allowing for privacy-focused, offline coding agents that can read your codebase, plan architectures, and execute complex refactors without sending a single byte to the cloud.

🌟 New Features
⚡ Native Local LLM Support (GGUF)
No Ollama Required: Added a built-in inference engine powered by llama-cpp-python.
Drag & Drop: Simply place .gguf models (Llama-3, Mistral, Gemma) into models/llm/ and they appear instantly in the provider list.
Privacy: Run 100% offline with zero data leakage.
🛡️ Command & Control Modes
Phased Strategy (Mode 1): The agent drafts a plan and pauses for your approval before executing each phase. Best for architectural changes.
Siege Mode (Mode 2): "Go Limitless." The agent enters a continuous loop of Code -> Error -> Debug -> Patch until the test passes. Best for stubborn bugs.
🧠 Deep RAG Integration
Semantic Memory: Now defaults to a high-speed local vector search.
Auto-Indexing: The IDE automatically indexes your project in the background for retrieval.
Tooltips: Added explainers for advanced RAG settings (Top-K, Min Score).
🎨 UI & UX Polish
Live Appearance: Change chat colors (User/AI) instantly without restarting.
Tool Visualization: XML tool calls are now parsed and displayed as interactive "Tool Used" widgets in the chat, keeping the interface clean.
Stop Button: Improved state management ensures the "Stop" button actually stops the agent and resets correctly.
🐛 Bug Fixes
Local Model Hallucinations: Fixed an issue where small models (3B/7B) would aggressively hallucinate tool calls. Implemented a CODING_AGENT_LITE system prompt specifically for them.
Settings Crash: Fixed a regression where opening Settings would crash due to legacy network RAG fields.
Layout: Moved "Appearance" settings to the main dialog for better accessibility.
⚠️ Breaking Changes
Legacy RAG Removed: Support for external vector engine URLs has been removed in favor of the optimized internal engine.
Get Started:

Run ./setup_integration.bat to install dependencies.
Drop a model in models/llm/.
Run ./start_IDE.bat.