use proper context for qwen3-coder

adds ast-grep tool for faster code exploration

.cargo/config.toml

@@ -0,0 +1,5 @@
+[target.aarch64-apple-darwin]
+rustflags = ["-C", "link-args=-Wl,-rpath,@executable_path"]
+
+[target.x86_64-apple-darwin]
+rustflags = ["-C", "link-args=-Wl,-rpath,@executable_path"]

.gitignore

@@ -2,10 +2,14 @@
 # will have compiled files and executables
 debug
 target
+.build
+appy/
 
 # These are backup files generated by rustfmt
 **/*.rs.bk
 
+**/.DS_Store
+
 # MSVC Windows builds of rustc generate these, which store debugging information
 *.pdb
 
@@ -19,3 +23,7 @@ target
 #  and can be added to the global gitignore or merged into this file.  For a more nuclear
 #  option (not recommended) you can uncomment the following to ignore the entire idea folder.
 #.idea/
+
+# Session logs directory
+logs/
+*.json

Cargo.toml

@@ -4,7 +4,8 @@ members = [
     "crates/g3-core", 
     "crates/g3-providers",
     "crates/g3-config",
-    "crates/g3-execution"
+    "crates/g3-execution",
+    "crates/g3-computer-control"
 ]
 resolver = "2"
 

DESIGN.md

@@ -1,157 +1,316 @@
-# G3 General Purpose AI Agent - Design Document
+# G3 - AI Coding Agent - Design Document
 
 ## Overview
-G3 is a **code-first AI agent** that helps you complete tasks by writing and executing code or scripts. Instead of just giving advice, G3 solves problems by generating executable code in the appropriate language.
+
+G3 is a **modular, composable AI coding agent** built in Rust that helps you complete tasks by writing and executing code. It provides a flexible architecture for interacting with various Large Language Model (LLM) providers while offering powerful code generation, file manipulation, and task automation capabilities.
+
+The agent follows a **tool-first philosophy**: instead of just providing advice, G3 actively uses tools to read files, write code, execute commands, and complete tasks autonomously.
 
 ## Core Principles
-1. **Code-First Philosophy**: Always try to solve problems with executable code
-2. **Multi-Language Support**: Generate scripts in Python, Bash, JavaScript, Rust, etc.
-3. **Unix Philosophy**: Small, focused tools that do one thing well
+
+1. **Tool-First Philosophy**: Solve problems by actively using tools rather than just providing advice
+2. **Modular Architecture**: Clear separation of concerns across multiple Rust crates
+3. **Provider Flexibility**: Support multiple LLM providers through a unified interface
 4. **Modularity**: Clear separation of concerns
 5. **Composability**: Components can be combined in different ways
-6. **Performance**: Blazing fast execution
+6. **Performance**: Built in Rust for speed and reliability
+7. **Context Intelligence**: Smart context window management with auto-summarization
+8. **Error Resilience**: Robust error handling with automatic retry logic
 
-## Architecture
+## Project Structure
 
-### High-Level Components
+G3 is organized as a Rust workspace with the following crates:
+
+```
+g3/
+├── src/main.rs                   # Main entry point (delegates to g3-cli)
+├── crates/
+│   ├── g3-cli/                   # Command-line interface, TUI, and retro mode
+│   ├── g3-core/                  # Core agent engine, tools, and streaming logic
+│   ├── g3-providers/             # LLM provider abstractions and implementations
+│   ├── g3-config/                # Configuration management
+│   ├── g3-execution/             # Code execution engine
+│   └── g3-computer-control/      # Computer control and automation
+├── logs/                         # Session logs (auto-created)
+├── README.md                     # Project documentation
+└── DESIGN.md                     # This design document
+```
+
+## Architecture Overview
+
+### High-Level Architecture
 
 ```
 ┌─────────────────┐    ┌─────────────────┐    ┌─────────────────┐
-│   CLI Module    │    │  Core Engine    │    │ LLM Providers   │
+│   g3-cli        │    │   g3-core       │    │ g3-providers    │
 │                 │    │                 │    │                 │
-│ - Task commands │◄──►│ - Task          │◄──►│ - OpenAI        │
-│ - Interactive   │    │   interpretation│    │ - Anthropic     │
-│   mode          │    │ - Code          │    │ - Embedded      │
-│ - Code exec     │    │   generation    │    │   (llama.cpp)   │
-│   approval      │    │ - Script        │    │ - Custom APIs   │
-│                 │    │   execution     │    │                 │
+│ • CLI parsing   │◄──►│ • Agent engine  │◄──►│ • Anthropic     │
+│ • Interactive   │    │ • Context mgmt  │    │ • Databricks    │
+│ • Retro TUI     │    │ • Tool system   │    │ • Embedded      │
+│ • Autonomous    │    │ • Streaming     │    │   (llama.cpp)   │
+│   mode          │    │ • Task exec     │    │ • OAuth flow    │
+│                 │    │ • TODO mgmt     │    │                 │
 └─────────────────┘    └─────────────────┘    └─────────────────┘
          │                       │                       │
          └───────────────────────┼───────────────────────┘
                                  │
-                    ┌─────────────────┐
-                    │   Execution     │
-                    │   Engine        │
-                    │                 │
-                    │ - Python        │
-                    │ - Bash/Shell    │
-                    │ - JavaScript    │
-                    │ - Rust          │
-                    │ - Sandboxing    │
+                    ┌─────────────────┐    ┌─────────────────┐
+                    │ g3-execution    │    │   g3-config     │
+                    │                 │    │                 │
+                    │ • Code exec     │    │ • TOML config   │
+                    │ • Shell cmds    │    │ • Env overrides │
+                    │ • Streaming     │    │ • Provider      │
+                    │ • Error hdlg    │    │   settings      │
+                    └─────────────────┘    │ • Computer      │
+                             │              │   control cfg   │
+                             │              └─────────────────┘
+                             │                       │
+                    ┌─────────────────┐             │
+                    │ g3-computer-    │◄────────────┘
+                    │   control       │
+                    │ • Mouse/kbd     │
+                    │ • Screenshots   │
+                    │ • OCR/Tesseract │
+                    │ • Windows/UI    │
                     └─────────────────┘
 ```
 
-### Module Breakdown
+## Core Components
 
-#### 1. CLI Module (`g3-cli`)
-- **Responsibility**: User interface and task interpretation
-- **New Features**:
-  - Progress indicators for script execution
+### 1. g3-core: Agent Engine
 
-#### 2. Core Engine (`g3-core`)
-- **Responsibility**: Task interpretation and code generation
-- **New Features**:
-  - Task analysis and decomposition
-  - Language selection based on task type
-  - Code generation with execution context
-  - Script template system
-  - Autonomous execution of generated code
+**Primary Responsibilities:**
+- Main orchestration logic for handling conversations and task execution
+- Context window management with intelligent token tracking
+- Built-in tool system for file operations and command execution
+- Streaming response parsing with real-time tool call detection
+- Error handling with automatic retry logic
 
-#### 3. LLM Providers (`g3-providers`)
-- **Responsibility**: LLM communication and model abstraction
-- **Supported Providers**:
-  - **OpenAI**: GPT-4, GPT-3.5-turbo via API
-  - **Anthropic**: Claude models via API  
-  - **Embedded**: Local open-weights models via llama.cpp
-- **Enhanced Prompts**:
-  - Code-first system prompts
-  - Language-specific generation instructions
+**Key Features:**
+- **Context Window Intelligence**: Automatic monitoring with percentage-based tracking (80% capacity triggers auto-summarization)
+- **Tool System**: Built-in tools for file operations (read, write, edit), shell commands, and structured output
+- **Streaming Parser**: Real-time parsing of LLM responses with tool call detection and execution
+- **Session Management**: Automatic session logging with detailed conversation history and token usage
+- **Error Recovery**: Sophisticated error classification and retry logic for recoverable errors
+- **TODO Management**: In-memory TODO list with read/write tools for task tracking
 
-#### 5. Embedded Provider (`g3-core/providers/embedded`) - NEW
-- **Responsibility**: Local model inference using llama.cpp
-- **Features**:
-  - GGUF model support (Llama, CodeLlama, Mistral, etc.)
-  - GPU acceleration via CUDA/Metal
-  - Configurable context length and generation parameters
-  - Async-compatible inference without blocking
-  - Thread-safe model access
-  - Stop sequence detection
+**Available Tools:**
+- `shell`: Execute shell commands with streaming output
+- `read_file`: Read file contents with optional character range support
+- `write_file`: Create or overwrite files with content
+- `str_replace`: Apply unified diffs to files with precise editing
+- `final_output`: Signal task completion with detailed summaries
+- `todo_read`: Read the entire TODO list content
+- `todo_write`: Write or overwrite the entire TODO list
+- `mouse_click`: Click the mouse at specific coordinates
+- `type_text`: Type text at the current cursor position
+- `find_element`: Find UI elements by text, role, or attributes
+- `take_screenshot`: Capture screenshots of screen, region, or window
+- `extract_text`: Extract text from images or screen regions using OCR
+- `find_text_on_screen`: Find text visually on screen and return coordinates
+- `list_windows`: List all open windows with IDs and titles
 
-#### 4. Execution Engine (`g3-execution`) - NEW
-- **Responsibility**: Safe code execution
-- **Features**:
-  - Multi-language script execution
-  - Sandboxing and security
-  - Resource limits
-  - Output capture and formatting
-  - Error handling and recovery
+### 2. g3-providers: LLM Provider Abstraction
 
-### Task Types and Language Selection
+**Primary Responsibilities:**
+- Unified interface for multiple LLM providers
+- Provider-specific optimizations and feature support
+- OAuth authentication flows
+- Streaming and non-streaming completion support
 
-| Task Type | Preferred Language | Use Cases |
-|-----------|-------------------|-----------|
-| Data Processing | Python | CSV/JSON analysis, data transformation |
-| File Operations | Bash/Shell | File manipulation, backups, organization |
-| System Admin | Bash/Shell | Process management, system monitoring |
-| Text Processing | Python/Bash | Log analysis, text transformation |
-| Database | Python/SQL | Data migration, queries, reporting |
-| Image/Media | Python | Image processing, format conversion |
-| Development | Rust | Code generation, project setup |
+**Supported Providers:**
+- **Anthropic**: Claude models via API with native tool calling support
+- **Databricks**: Foundation Model APIs with OAuth and token-based authentication (default provider)
+- **Embedded**: Local models via llama.cpp with GPU acceleration (Metal/CUDA)
+- **Provider Registry**: Dynamic provider management and hot-swapping
 
-## Implementation Plan
+**Key Features:**
+- **Native Tool Calling**: Full support for structured tool calls where available
+- **Fallback Parsing**: JSON tool call parsing for providers without native support
+- **OAuth Integration**: Built-in OAuth flow for secure provider authentication
+- **Context-Aware**: Provider-specific context length and token limit handling
+- **Streaming Support**: Real-time response streaming with tool call detection
 
-### Phase 1: Core Refactoring ✅
-1. ✅ Update CLI commands for task-oriented interface
-2. ✅ Enhance system prompts for code-first approach
-3. ✅ Add basic code execution capabilities
-4. ✅ Update interactive mode messaging
+### 3. g3-cli: Command-Line Interface
 
-### Phase 2: Enhanced Provider Support ✅
-1. ✅ Implement embedded model provider using llama.cpp
-2. ✅ Add GGUF model support for local inference
-3. ✅ Configure GPU acceleration and performance optimization
-4. ✅ Add comprehensive logging and debugging support
+**Primary Responsibilities:**
+- Command-line argument parsing and validation
+- Interactive terminal interface with history support
+- Retro-style terminal UI (80s sci-fi inspired)
+- Autonomous mode with coach-player feedback loops
+- Session management and workspace handling
 
-### Phase 3: Advanced Features (Future)
-1. Model quantization and optimization
-2. Multi-model ensemble support
-3. Advanced code execution sandboxing
-4. Plugin system for custom providers
-5. Web interface for remote access
+**Execution Modes:**
+- **Single-shot**: Execute one task and exit
+- **Interactive**: REPL-style conversation with the agent (default mode)
+- **Autonomous**: Coach-player feedback loop for complex projects
+- **Retro TUI**: Full-screen terminal interface with real-time updates
+
+**Key Features:**
+- **Multi-line Input**: Support for complex, multi-line prompts with backslash continuation
+- **Context Progress**: Real-time display of token usage and context window status
+- **Error Recovery**: Automatic retry logic for timeout and recoverable errors
+- **History Management**: Persistent command history across sessions
+- **Theme Support**: Customizable color themes for retro mode
+- **Cancellation**: Ctrl+C support for graceful operation cancellation
+
+### 4. g3-execution: Code Execution Engine
+
+**Primary Responsibilities:**
+- Safe execution of shell commands and scripts
+- Streaming output capture and display
+- Multi-language code execution support
+- Error handling and result formatting
+
+**Supported Execution:**
+- **Bash/Shell**: Direct command execution with streaming output (primary use case)
+- **Python**: Script execution via temporary files (legacy support)
+- **JavaScript**: Node.js-based execution (legacy support)
+
+**Key Features:**
+- **Streaming Output**: Real-time command output display
+- **Error Capture**: Comprehensive stderr and stdout handling
+- **Exit Code Tracking**: Proper success/failure detection
+- **Async Execution**: Non-blocking command execution
+- **Output Formatting**: Clean, user-friendly result presentation
+
+### 5. g3-config: Configuration Management
+
+**Primary Responsibilities:**
+- TOML-based configuration file management
+- Environment variable overrides
+- Provider-specific settings and credentials
+- CLI argument integration
+
+**Configuration Hierarchy:**
+1. Default configuration (Databricks provider with OAuth)
+2. Configuration files (`~/.config/g3/config.toml`, `./g3.toml`)
+3. Environment variables (`G3_*`)
+4. CLI arguments (highest priority)
+
+**Key Features:**
+- **Auto-generation**: Creates default configuration files if none exist
+- **Provider Overrides**: Runtime provider and model selection
+- **Validation**: Configuration validation with helpful error messages
+- **Flexible Paths**: Support for shell expansion (`~`, environment variables)
+
+### 6. g3-computer-control: Computer Control & Automation
+
+**Primary Responsibilities:**
+- Cross-platform computer control and automation
+- Mouse and keyboard input simulation
+- Window management and screenshot capture
+- OCR text extraction from images and screen regions
+
+**Platform Support:**
+- **macOS**: Core Graphics, Cocoa, screencapture integration
+- **Linux**: X11/Xtest for input, X11 for window management
+- **Windows**: Win32 APIs for input and window control
+
+**Key Features:**
+- **OCR Integration**: Tesseract-based text extraction from images
+- **Window Management**: List, identify, and capture specific application windows
+- **UI Automation**: Find elements, simulate clicks, type text
+- **Screenshot Capture**: Full screen, regions, or specific windows
+- **Accessibility**: Requires OS-level permissions for automation
+
+## Advanced Features
+
+### Context Window Management
+
+G3 implements sophisticated context window management:
+
+- **Automatic Monitoring**: Tracks token usage with percentage-based thresholds
+- **Smart Summarization**: Auto-triggers at 80% capacity to prevent context overflow
+- **Context Thinning**: Progressive thinning at 50%, 60%, 70%, 80% thresholds - replaces large tool results with file references
+- **Conversation Preservation**: Maintains conversation continuity through intelligent summaries
+- **Provider-Specific Limits**: Adapts to different model context windows (4k to 200k+ tokens)
+- **Cumulative Tracking**: Monitors total token usage across entire sessions
+
+### Error Handling & Resilience
+
+Comprehensive error handling system:
+
+- **Error Classification**: Distinguishes between recoverable and non-recoverable errors
+- **Automatic Retry**: Exponential backoff with jitter for rate limits, timeouts, and server errors
+- **Detailed Logging**: Comprehensive error context including stack traces and session data
+- **Error Persistence**: Saves detailed error logs to `logs/errors/` for analysis
+- **Graceful Degradation**: Continues operation when possible, fails gracefully when not
+
+### Session Management
+
+Automatic session tracking and logging:
+
+- **Session IDs**: Generated based on initial prompts for easy identification
+- **Complete Logs**: Full conversation history, token usage, and timing data
+- **JSON Format**: Structured logs for easy parsing and analysis
+- **Automatic Cleanup**: Organized in `logs/` directory with timestamps
+- **Status Tracking**: Records session completion status (completed, cancelled, error)
+
+### Autonomous Mode
+
+Advanced autonomous operation with coach-player feedback:
+
+- **Requirements-Driven**: Reads `requirements.md` for project specifications
+- **Dual-Agent System**: Separate player (implementation) and coach (review) agents
+- **Iterative Improvement**: Multiple rounds of implementation and feedback
+- **Progress Tracking**: Detailed reporting of turns, token usage, and final status
+- **Workspace Management**: Automatic workspace setup and file organization
 
 ## Provider Comparison
 
-| Feature | OpenAI | Anthropic | Embedded |
-|---------|--------|-----------|----------|
+| Feature | Anthropic | Databricks (Default) | Embedded |
+|---------|-----------|------------|----------|
 | **Cost** | Pay per token | Pay per token | Free after download |
 | **Privacy** | Data sent to API | Data sent to API | Completely local |
 | **Performance** | Very fast | Very fast | Depends on hardware |
 | **Model Quality** | Excellent | Excellent | Good (varies by model) |
 | **Offline Support** | No | No | Yes |
-| **Setup Complexity** | API key only | API key only | Model download required |
+| **Setup Complexity** | API key only | OAuth or token | Model download required |
+| **Context Window** | 200k tokens | Varies by model | 4k-32k tokens |
+| **Tool Calling** | Native support | Native support | JSON fallback |
 | **Hardware Requirements** | None | None | 4-16GB RAM, optional GPU |
 
 ## Configuration Examples
 
-### Cloud-First Setup
+### Cloud-First Setup (Anthropic)
 ```toml
 [providers]
-default_provider = "openai"
+default_provider = "anthropic"
 
-[providers.openai]
-api_key = "sk-..."
-model = "gpt-4"
+[providers.anthropic]
+api_key = "sk-ant-..."
+model = "claude-3-5-sonnet-20241022"
+max_tokens = 8192
+temperature = 0.1
 ```
 
-### Privacy-First Setup  
+### Enterprise Setup (Databricks - Default)
+```toml
+[providers]
+default_provider = "databricks"
+
+[providers.databricks]
+host = "https://your-workspace.cloud.databricks.com"
+model = "databricks-claude-sonnet-4"
+max_tokens = 32000
+temperature = 0.1
+use_oauth = true
+```
+
+### Privacy-First Setup (Local Models)
 ```toml
 [providers]
 default_provider = "embedded"
 
 [providers.embedded]
-model_path = "~/.cache/g3/models/codellama-7b-instruct.Q4_K_M.gguf"
-model_type = "codellama"
+model_path = "~/.cache/g3/models/qwen2.5-7b-instruct-q3_k_m.gguf"
+model_type = "qwen"
+context_length = 32768
+max_tokens = 2048
+temperature = 0.1
 gpu_layers = 32
+threads = 8
 ```
 
 ### Hybrid Setup
@@ -159,14 +318,109 @@ gpu_layers = 32
 [providers]
 default_provider = "embedded"
 
-# Use embedded for most tasks
+# Local model for most tasks
 [providers.embedded]
 model_path = "~/.cache/g3/models/codellama-7b-instruct.Q4_K_M.gguf"
 model_type = "codellama"
+context_length = 16384
 gpu_layers = 32
 
-# Fallback to cloud for complex tasks
-[providers.openai]
-api_key = "sk-..."
-model = "gpt-4"
+# Cloud fallback for complex tasks
+[providers.anthropic]
+api_key = "sk-ant-..."
+model = "claude-3-5-sonnet-20241022"
 ```
+
+## Usage Examples
+
+### Single-Shot Mode
+```bash
+g3 "implement a fibonacci function in Rust"
+```
+
+### Interactive Mode
+```bash
+g3
+g3> read the README and suggest improvements
+g3> implement the suggestions you made
+```
+
+### Autonomous Mode
+```bash
+g3 --autonomous --max-turns 10
+# Reads requirements.md and implements iteratively
+```
+
+### Retro TUI Mode
+```bash
+g3 --retro --theme dracula
+# Full-screen terminal interface
+```
+
+## Implementation Details
+
+### Planned Features
+- **Plugin System**: Custom tool and provider plugins
+- **Web Interface**: Browser-based UI for remote access
+- **Model Quantization**: Optimized local model deployment
+- **Multi-Model Ensemble**: Combine multiple models for better results
+- **Advanced Sandboxing**: Enhanced security for code execution
+- **Collaborative Mode**: Multi-user sessions and shared workspaces
+
+### Technical Improvements
+- **Performance Optimization**: Faster streaming and tool execution
+- **Memory Management**: Better handling of large contexts and files
+- **Caching System**: Intelligent caching of model responses and computations
+- **Monitoring**: Built-in metrics and performance monitoring
+- **Testing**: Comprehensive test suite and CI/CD integration
+
+## Development Guidelines
+
+### Code Organization
+- **Modular Design**: Each crate has a single, well-defined responsibility
+- **Trait-Based**: Use traits for abstraction and testability
+- **Error Handling**: Comprehensive error types with context
+- **Documentation**: Inline docs and examples for all public APIs
+- **Testing**: Unit tests, integration tests, and property-based testing
+
+### Performance Considerations
+- **Async-First**: All I/O operations are asynchronous (Tokio runtime)
+- **Streaming**: Real-time response processing where possible
+- **Memory Efficiency**: Careful memory management for large contexts
+- **Caching**: Strategic caching of expensive operations
+- **Profiling**: Regular performance profiling and optimization
+
+This design document reflects the current state of G3 as a mature, production-ready AI coding agent with sophisticated architecture and comprehensive feature set.
+
+## Current Implementation Status
+
+### Fully Implemented
+- ✅ **Core Agent Engine**: Complete with streaming, tool execution, and context management
+- ✅ **Provider System**: Anthropic, Databricks, and Embedded providers with OAuth support
+- ✅ **Tool System**: 13 tools including file ops, shell, TODO management, and computer control
+- ✅ **CLI Interface**: Interactive mode, single-shot mode, retro TUI
+- ✅ **Autonomous Mode**: Coach-player feedback loop with requirements.md processing
+- ✅ **Configuration**: TOML-based config with environment overrides
+- ✅ **Error Handling**: Comprehensive retry logic and error classification
+- ✅ **Session Logging**: Automatic session tracking and JSON logs
+- ✅ **Context Management**: Context thinning (50-80%) and auto-summarization at 80% capacity
+- ✅ **Computer Control**: Cross-platform automation with OCR support
+- ✅ **TODO Management**: In-memory TODO list with read/write tools
+
+### Architecture Highlights
+- **Workspace**: 6 crates with clear separation of concerns
+- **Dependencies**: Modern Rust ecosystem (Tokio, Clap, Serde, etc.)
+- **Streaming**: Real-time response processing with tool call detection
+- **Cross-Platform**: Works on macOS, Linux, and Windows
+- **GPU Support**: Metal acceleration for local models on macOS, CUDA on Linux
+- **OCR Support**: Tesseract integration for text extraction from images
+
+### Key Files
+- `src/main.rs`: main entry point delegating to g3-cli
+- `crates/g3-core/src/lib.rs`: main agent implementation
+- `crates/g3-cli/src/lib.rs`: CLI and interaction modes
+- `crates/g3-providers/src/lib.rs`: provider trait and registry
+- `crates/g3-config/src/lib.rs`: configuration management
+- `crates/g3-execution/src/lib.rs`: code execution engine
+- `crates/g3-computer-control/src/lib.rs`: computer control and automation
+- `crates/g3-computer-control/src/platform/`: platform-specific implementations

OLLAMA_CONFIG.md

@@ -0,0 +1,456 @@
+# Configuring Ollama Provider in G3
+
+This guide shows you how to configure G3 to use Ollama as your LLM provider.
+
+## Quick Start
+
+### 1. Install Ollama
+
+```bash
+# Visit https://ollama.ai to download and install
+# Or use curl:
+curl https://ollama.ai/install.sh | sh
+```
+
+### 2. Pull a Model
+
+```bash
+ollama pull llama3.2
+# or any other model you prefer
+```
+
+### 3. Create Configuration File
+
+Copy the example configuration:
+
+```bash
+cp config.ollama.example.toml ~/.config/g3/config.toml
+```
+
+Or create it manually:
+
+```toml
+[providers]
+default_provider = "ollama"
+
+[providers.ollama]
+model = "llama3.2"
+```
+
+### 4. Run G3
+
+```bash
+g3
+# G3 will now use Ollama with llama3.2!
+```
+
+## Configuration Options
+
+### Basic Configuration
+
+```toml
+[providers]
+default_provider = "ollama"
+
+[providers.ollama]
+model = "llama3.2"
+```
+
+This is the minimal configuration needed. It uses all defaults:
+- Base URL: `http://localhost:11434`
+- Temperature: `0.7`
+- Max tokens: Not limited (uses model default)
+
+### Full Configuration
+
+```toml
+[providers]
+default_provider = "ollama"
+
+[providers.ollama]
+model = "llama3.2"
+base_url = "http://localhost:11434"
+max_tokens = 2048
+temperature = 0.7
+```
+
+### Custom Ollama Host
+
+If you're running Ollama on a different machine or port:
+
+```toml
+[providers.ollama]
+model = "llama3.2"
+base_url = "http://192.168.1.100:11434"
+```
+
+### Different Models
+
+You can use any Ollama model:
+
+```toml
+[providers.ollama]
+model = "qwen2.5:7b"  # Alibaba's Qwen model
+```
+
+```toml
+[providers.ollama]
+model = "mistral"  # Mistral AI
+```
+
+```toml
+[providers.ollama]
+model = "llama3.1:70b"  # Larger Llama model
+```
+
+## Multiple Provider Configuration
+
+You can configure multiple providers and switch between them:
+
+```toml
+[providers]
+default_provider = "ollama"  # Default for most operations
+
+# Ollama for local, fast responses
+[providers.ollama]
+model = "llama3.2:3b"
+temperature = 0.7
+
+# Databricks for more complex tasks
+[providers.databricks]
+host = "https://your-workspace.cloud.databricks.com"
+model = "databricks-claude-sonnet-4"
+max_tokens = 4096
+temperature = 0.1
+use_oauth = true
+```
+
+Then switch providers with:
+
+```bash
+g3 --provider databricks
+```
+
+## Autonomous Mode (Coach-Player)
+
+Use different providers for code review (coach) and implementation (player):
+
+```toml
+[providers]
+default_provider = "ollama"
+coach = "databricks"  # Use powerful cloud model for review
+player = "ollama"     # Use local model for implementation
+
+[providers.ollama]
+model = "qwen2.5:14b"  # Larger local model for coding
+
+[providers.databricks]
+host = "https://your-workspace.cloud.databricks.com"
+model = "databricks-claude-sonnet-4"
+use_oauth = true
+```
+
+This gives you the best of both worlds:
+- Fast local execution for coding tasks
+- Powerful cloud review for quality assurance
+
+## Recommended Models
+
+### For Coding Tasks
+
+| Model | Size | Speed | Quality | Notes |
+|-------|------|-------|---------|-------|
+| **qwen2.5:7b** | 7B | Fast | Excellent | Best balance for coding |
+| **llama3.2:3b** | 3B | Very Fast | Good | Great for quick tasks |
+| **llama3.1:8b** | 8B | Medium | Very Good | Solid all-rounder |
+| **mistral** | 7B | Fast | Good | Good for general use |
+
+### For Complex Tasks
+
+| Model | Size | Speed | Quality | Notes |
+|-------|------|-------|---------|-------|
+| **qwen2.5:14b** | 14B | Medium | Excellent | Best local model for coding |
+| **qwen2.5:32b** | 32B | Slow | Outstanding | If you have the resources |
+| **llama3.1:70b** | 70B | Very Slow | Outstanding | Requires significant RAM/GPU |
+
+## Temperature Settings
+
+Temperature controls randomness in responses:
+
+- **0.1-0.3**: Deterministic, good for code generation
+- **0.5-0.7**: Balanced, good for most tasks
+- **0.8-1.0**: Creative, good for brainstorming
+
+```toml
+[providers.ollama]
+model = "qwen2.5:7b"
+temperature = 0.2  # Focused code generation
+```
+
+## Max Tokens
+
+Control response length:
+
+```toml
+[providers.ollama]
+model = "llama3.2"
+max_tokens = 1024  # Shorter responses
+```
+
+```toml
+[providers.ollama]
+model = "qwen2.5:7b"
+max_tokens = 4096  # Longer, detailed responses
+```
+
+Leave it unset for model defaults (recommended).
+
+## Performance Tuning
+
+### GPU Acceleration
+
+Ollama automatically uses GPU if available. To check:
+
+```bash
+ollama ps
+```
+
+### Quantized Models
+
+For faster responses with less RAM:
+
+```toml
+[providers.ollama]
+model = "llama3.2:3b-q4_0"  # 4-bit quantization
+```
+
+Quantization options:
+- `q4_0`: 4-bit, fastest, lowest quality
+- `q5_0`: 5-bit, balanced
+- `q8_0`: 8-bit, slower, better quality
+
+### Multiple Models
+
+You can pull multiple models and switch easily:
+
+```bash
+ollama pull llama3.2:3b    # Fast for chat
+ollama pull qwen2.5:7b     # Better for code
+ollama pull mistral        # General purpose
+```
+
+Then change your config:
+
+```toml
+[providers.ollama]
+model = "qwen2.5:7b"  # Just change this line
+```
+
+## Troubleshooting
+
+### Ollama Not Running
+
+```bash
+# Check if Ollama is running
+curl http://localhost:11434/api/version
+
+# Start Ollama (macOS/Linux)
+ollama serve
+
+# Or just run a model (auto-starts)
+ollama run llama3.2
+```
+
+### Model Not Found
+
+```bash
+# List available models
+ollama list
+
+# Pull the model
+ollama pull llama3.2
+```
+
+### Slow Responses
+
+1. Use a smaller model:
+   ```toml
+   model = "llama3.2:1b"  # Smallest, fastest
+   ```
+
+2. Use quantized version:
+   ```toml
+   model = "llama3.2:3b-q4_0"
+   ```
+
+3. Reduce max_tokens:
+   ```toml
+   max_tokens = 512
+   ```
+
+### Out of Memory
+
+1. Switch to smaller model
+2. Use quantized version
+3. Close other applications
+4. Check GPU memory: `ollama ps`
+
+### Connection Refused
+
+Check base_url is correct:
+
+```toml
+[providers.ollama]
+model = "llama3.2"
+base_url = "http://localhost:11434"  # Default
+```
+
+For remote Ollama:
+
+```toml
+base_url = "http://your-server:11434"
+```
+
+## Complete Example Configs
+
+### Minimal Local Setup
+
+```toml
+[providers]
+default_provider = "ollama"
+
+[providers.ollama]
+model = "llama3.2"
+
+[agent]
+max_context_length = 8192
+enable_streaming = true
+timeout_seconds = 60
+```
+
+### Optimized for Coding
+
+```toml
+[providers]
+default_provider = "ollama"
+
+[providers.ollama]
+model = "qwen2.5:7b"
+temperature = 0.2
+max_tokens = 2048
+
+[agent]
+max_context_length = 16384
+enable_streaming = true
+timeout_seconds = 120
+```
+
+### Fast Responses
+
+```toml
+[providers]
+default_provider = "ollama"
+
+[providers.ollama]
+model = "llama3.2:3b-q4_0"
+temperature = 0.7
+max_tokens = 1024
+
+[agent]
+max_context_length = 4096
+enable_streaming = true
+timeout_seconds = 30
+```
+
+### High Quality (Requires Good Hardware)
+
+```toml
+[providers]
+default_provider = "ollama"
+
+[providers.ollama]
+model = "qwen2.5:32b"
+temperature = 0.3
+max_tokens = 4096
+
+[agent]
+max_context_length = 32768
+enable_streaming = true
+timeout_seconds = 300
+```
+
+### Hybrid (Local + Cloud)
+
+```toml
+[providers]
+default_provider = "ollama"
+coach = "databricks"
+player = "ollama"
+
+[providers.ollama]
+model = "qwen2.5:14b"
+temperature = 0.2
+
+[providers.databricks]
+host = "https://your-workspace.cloud.databricks.com"
+model = "databricks-claude-sonnet-4"
+use_oauth = true
+
+[agent]
+max_context_length = 16384
+enable_streaming = true
+timeout_seconds = 120
+```
+
+## Environment Variables
+
+You can override config with environment variables:
+
+```bash
+# Override model
+G3_PROVIDERS_OLLAMA_MODEL=qwen2.5:7b g3
+
+# Override base URL
+G3_PROVIDERS_OLLAMA_BASE_URL=http://192.168.1.100:11434 g3
+
+# Override default provider
+G3_PROVIDERS_DEFAULT_PROVIDER=ollama g3
+```
+
+## Best Practices
+
+1. **Start Small**: Begin with llama3.2:3b, scale up if needed
+2. **Use Quantization**: q4_0 or q5_0 for best speed/quality balance
+3. **Match Task to Model**: 
+   - Quick edits: 1B-3B models
+   - Code generation: 7B-14B models
+   - Complex refactoring: 14B-32B models
+4. **Temperature for Code**: Use 0.1-0.3 for deterministic output
+5. **Enable Streaming**: Always enable for better UX
+6. **Local First**: Use Ollama by default, cloud for special cases
+
+## Comparison with Other Providers
+
+| Feature | Ollama | Databricks | OpenAI | Anthropic |
+|---------|--------|------------|--------|-----------|
+| Cost | Free | Paid | Paid | Paid |
+| Privacy | Full | Medium | Low | Low |
+| Speed (small models) | Fast | Fast | Medium | Medium |
+| Speed (large models) | Slow | Fast | Fast | Fast |
+| Setup Complexity | Low | Medium | Low | Low |
+| Authentication | None | OAuth/Token | API Key | API Key |
+| Offline Support | Yes | No | No | No |
+| Tool Calling | Yes | Yes | Yes | Yes |
+
+## Next Steps
+
+1. Try different models: `ollama pull mistral`, `ollama pull qwen2.5`
+2. Experiment with temperature settings
+3. Set up hybrid config with cloud provider for complex tasks
+4. Share your config in the community!
+
+## Getting Help
+
+- Ollama docs: https://ollama.ai/docs
+- G3 issues: https://github.com/your-repo/issues
+- Test your config: `g3 --help`

OLLAMA_EXAMPLE.md

@@ -0,0 +1,315 @@
+# Ollama Provider for g3
+
+A simple, local LLM provider implementation for g3 that connects to Ollama.
+
+## Features
+
+- ✅ **Simple Setup**: No API keys or authentication required
+- ✅ **Local Execution**: Runs entirely on your machine
+- ✅ **Tool Calling Support**: Native tool calling for compatible models
+- ✅ **Streaming**: Full streaming support with real-time responses
+- ✅ **Flexible Configuration**: Custom base URL, temperature, and max tokens
+- ✅ **Model Discovery**: Automatic detection of available models
+
+## Quick Start
+
+### Prerequisites
+
+1. Install and start Ollama: https://ollama.ai
+2. Pull a model: `ollama pull llama3.2`
+
+### Basic Usage
+
+```rust
+use g3_providers::{OllamaProvider, LLMProvider, CompletionRequest, Message, MessageRole};
+
+#[tokio::main]
+async fn main() -> anyhow::Result<()> {
+    // Create provider with default settings (localhost:11434)
+    let provider = OllamaProvider::new(
+        "llama3.2".to_string(),
+        None,  // base_url: defaults to http://localhost:11434
+        None,  // max_tokens: optional
+        None,  // temperature: defaults to 0.7
+    )?;
+
+    // Create a simple request
+    let request = CompletionRequest {
+        messages: vec![
+            Message {
+                role: MessageRole::User,
+                content: "What is the capital of France?".to_string(),
+            },
+        ],
+        max_tokens: Some(1000),
+        temperature: Some(0.7),
+        stream: false,
+        tools: None,
+    };
+
+    // Get completion
+    let response = provider.complete(request).await?;
+    println!("Response: {}", response.content);
+    println!("Tokens: {}", response.usage.total_tokens);
+
+    Ok(())
+}
+```
+
+### Streaming Example
+
+```rust
+use futures_util::StreamExt;
+
+let request = CompletionRequest {
+    messages: vec![
+        Message {
+            role: MessageRole::User,
+            content: "Write a short poem about coding".to_string(),
+        },
+    ],
+    max_tokens: Some(500),
+    temperature: Some(0.8),
+    stream: true,
+    tools: None,
+};
+
+let mut stream = provider.stream(request).await?;
+
+while let Some(chunk_result) = stream.next().await {
+    match chunk_result {
+        Ok(chunk) => {
+            print!("{}", chunk.content);
+            if chunk.finished {
+                println!("\n\nDone!");
+                if let Some(usage) = chunk.usage {
+                    println!("Total tokens: {}", usage.total_tokens);
+                }
+            }
+        }
+        Err(e) => eprintln!("Error: {}", e),
+    }
+}
+```
+
+### Tool Calling Example
+
+```rust
+use serde_json::json;
+
+let tools = vec![Tool {
+    name: "get_weather".to_string(),
+    description: "Get current weather for a location".to_string(),
+    input_schema: json!({
+        "type": "object",
+        "properties": {
+            "location": {
+                "type": "string",
+                "description": "City name"
+            },
+            "unit": {
+                "type": "string",
+                "enum": ["celsius", "fahrenheit"],
+                "description": "Temperature unit"
+            }
+        },
+        "required": ["location"]
+    }),
+}];
+
+let request = CompletionRequest {
+    messages: vec![
+        Message {
+            role: MessageRole::User,
+            content: "What's the weather in Paris?".to_string(),
+        },
+    ],
+    max_tokens: Some(500),
+    temperature: Some(0.5),
+    stream: false,
+    tools: Some(tools),
+};
+
+let response = provider.complete(request).await?;
+println!("Response: {}", response.content);
+```
+
+### Custom Ollama Host
+
+```rust
+// Connect to remote Ollama instance
+let provider = OllamaProvider::new(
+    "llama3.2".to_string(),
+    Some("http://192.168.1.100:11434".to_string()),
+    None,
+    None,
+)?;
+```
+
+### Fetch Available Models
+
+```rust
+// Discover what models are available
+let models = provider.fetch_available_models().await?;
+println!("Available models:");
+for model in models {
+    println!("  - {}", model);
+}
+```
+
+## Supported Models
+
+The provider works with any Ollama model, including:
+
+- **llama3.2** (1B, 3B) - Meta's latest Llama models
+- **llama3.1** (8B, 70B, 405B) - Previous generation
+- **qwen2.5** (7B, 14B, 32B) - Alibaba's Qwen models  
+- **mistral** - Mistral AI models
+- **mixtral** - Mixture of experts model
+- **phi3** - Microsoft's Phi-3
+- **gemma2** - Google's Gemma 2
+
+## Configuration
+
+### Constructor Parameters
+
+```rust
+OllamaProvider::new(
+    model: String,           // Model name (e.g., "llama3.2")
+    base_url: Option<String>, // Ollama API URL (default: http://localhost:11434)
+    max_tokens: Option<u32>,  // Maximum tokens to generate (optional)
+    temperature: Option<f32>, // Sampling temperature (default: 0.7)
+)
+```
+
+### Request Options
+
+```rust
+CompletionRequest {
+    messages: Vec<Message>,      // Conversation history
+    max_tokens: Option<u32>,     // Override provider's max_tokens
+    temperature: Option<f32>,    // Override provider's temperature
+    stream: bool,                // Enable streaming responses
+    tools: Option<Vec<Tool>>,    // Tools for function calling
+}
+```
+
+## Comparison with Other Providers
+
+| Feature | Ollama | OpenAI | Anthropic | Databricks |
+|---------|--------|--------|-----------|------------|
+| Local Execution | ✅ | ❌ | ❌ | ❌ |
+| Authentication | None | API Key | API Key | OAuth/Token |
+| Tool Calling | ✅ | ✅ | ✅ | ✅ |
+| Streaming | ✅ | ✅ | ✅ | ✅ |
+| Cost | Free | Paid | Paid | Paid |
+| Privacy | High | Low | Low | Medium |
+
+## Implementation Details
+
+### API Endpoints
+
+- **Chat Completion**: `POST /api/chat`
+- **Model List**: `GET /api/tags`
+
+### Response Format
+
+Ollama uses a simple JSON-per-line streaming format:
+
+```json
+{"message":{"role":"assistant","content":"Hello"},"done":false}
+{"message":{"role":"assistant","content":" there"},"done":false}
+{"done":true,"prompt_eval_count":10,"eval_count":20}
+```
+
+### Tool Call Format
+
+Tool calls are returned in the message structure:
+
+```json
+{
+  "message": {
+    "role": "assistant",
+    "content": "",
+    "tool_calls": [
+      {
+        "function": {
+          "name": "get_weather",
+          "arguments": {"location": "Paris", "unit": "celsius"}
+        }
+      }
+    ]
+  },
+  "done": true
+}
+```
+
+## Troubleshooting
+
+### Connection Errors
+
+If you see connection errors, ensure Ollama is running:
+
+```bash
+# Check if Ollama is running
+curl http://localhost:11434/api/version
+
+# Start Ollama (if needed)
+ollama serve
+```
+
+### Model Not Found
+
+Pull the model first:
+
+```bash
+ollama pull llama3.2
+ollama list  # Check available models
+```
+
+### Performance Issues
+
+- Use smaller models (1B, 3B) for faster responses
+- Reduce `max_tokens` to limit generation length
+- Enable GPU acceleration if available
+- Consider quantized models (e.g., `llama3.2:3b-q4_0`)
+
+## Testing
+
+Run the included tests:
+
+```bash
+cargo test --package g3-providers ollama
+```
+
+All tests should pass:
+```
+running 4 tests
+test ollama::tests::test_custom_base_url ... ok
+test ollama::tests::test_message_conversion ... ok
+test ollama::tests::test_provider_creation ... ok
+test ollama::tests::test_tool_conversion ... ok
+```
+
+## Architecture
+
+The provider follows the same architecture as other g3 providers:
+
+1. **OllamaProvider**: Main struct implementing `LLMProvider` trait
+2. **Request/Response Structures**: Internal types for Ollama API
+3. **Streaming Parser**: Handles line-by-line JSON parsing
+4. **Tool Call Handling**: Accumulates and converts tool calls
+5. **Error Handling**: Robust error handling with retries
+
+## Contributing
+
+The provider is part of the g3-providers crate. To contribute:
+
+1. Add features to `ollama.rs`
+2. Update tests
+3. Run `cargo test --package g3-providers`
+4. Update this documentation
+
+## License
+
+Same as the g3 project.

README.md

@@ -1,3 +1,247 @@
-# G3
+# G3 - AI Coding Agent
 
-An experiment in a code-first AI agent that helps you complete tasks by writing and executing code.
+G3 is a coding AI agent designed to help you complete tasks by writing code and executing commands. Built in Rust, it provides a flexible architecture for interacting with various Large Language Model (LLM) providers while offering powerful code generation and task automation capabilities.
+
+## Architecture Overview
+
+G3 follows a modular architecture organized as a Rust workspace with multiple crates, each responsible for specific functionality:
+
+### Core Components
+
+#### **g3-core**
+The heart of the agent system, containing:
+- **Agent Engine**: Main orchestration logic for handling conversations, tool execution, and task management
+- **Context Window Management**: Intelligent tracking of token usage with context thinning (50-80%) and auto-summarization at 80% capacity
+- **Tool System**: Built-in tools for file operations, shell commands, computer control, TODO management, and structured output
+- **Streaming Response Parser**: Real-time parsing of LLM responses with tool call detection and execution
+- **Task Execution**: Support for single and iterative task execution with automatic retry logic
+
+#### **g3-providers**
+Abstraction layer for LLM providers:
+- **Provider Interface**: Common trait-based API for different LLM backends
+- **Multiple Provider Support**: 
+  - Anthropic (Claude models)
+  - Databricks (DBRX and other models)
+  - Local/embedded models via llama.cpp with Metal acceleration on macOS
+- **OAuth Authentication**: Built-in OAuth flow support for secure provider authentication
+- **Provider Registry**: Dynamic provider management and selection
+
+#### **g3-config**
+Configuration management system:
+- Environment-based configuration
+- Provider credentials and settings
+- Model selection and parameters
+- Runtime configuration options
+
+#### **g3-execution**
+Task execution framework:
+- Task planning and decomposition
+- Execution strategies (sequential, parallel)
+- Error handling and retry mechanisms
+- Progress tracking and reporting
+
+#### **g3-computer-control**
+Computer control capabilities:
+- Mouse and keyboard automation
+- UI element inspection and interaction
+- Screenshot capture and window management
+- OCR text extraction via Tesseract
+
+#### **g3-cli**
+Command-line interface:
+- Interactive terminal interface
+- Task submission and monitoring
+- Configuration management commands
+- Session management
+
+### Error Handling & Resilience
+
+G3 includes robust error handling with automatic retry logic:
+- **Recoverable Error Detection**: Automatically identifies recoverable errors (rate limits, network issues, server errors, timeouts)
+- **Exponential Backoff with Jitter**: Implements intelligent retry delays to avoid overwhelming services
+- **Detailed Error Logging**: Captures comprehensive error context including stack traces, request/response data, and session information
+- **Error Persistence**: Saves detailed error logs to `logs/errors/` for post-mortem analysis
+- **Graceful Degradation**: Non-recoverable errors are logged with full context before terminating
+
+## Key Features
+
+### Intelligent Context Management
+- Automatic context window monitoring with percentage-based tracking
+- Smart auto-summarization when approaching token limits
+- **Context thinning** at 50%, 60%, 70%, 80% thresholds - automatically replaces large tool results with file references
+- Conversation history preservation through summaries
+- Dynamic token allocation for different providers (4k to 200k+ tokens)
+
+### Interactive Control Commands
+G3's interactive CLI includes control commands for manual context management:
+- **`/compact`**: Manually trigger summarization to compact conversation history
+- **`/thinnify`**: Manually trigger context thinning to replace large tool results with file references
+- **`/readme`**: Reload README.md and AGENTS.md from disk without restarting
+- **`/stats`**: Show detailed context and performance statistics
+- **`/help`**: Display all available control commands
+
+These commands give you fine-grained control over context management, allowing you to proactively optimize token usage and refresh project documentation. See [Control Commands Documentation](docs/CONTROL_COMMANDS.md) for detailed usage.
+
+### Tool Ecosystem
+- **File Operations**: Read, write, and edit files with line-range precision
+- **Shell Integration**: Execute system commands with output capture
+- **Code Generation**: Structured code generation with syntax awareness
+- **TODO Management**: Read and write TODO lists with markdown checkbox format
+- **Computer Control** (Experimental): Automate desktop applications
+  - Mouse and keyboard control
+  - macOS Accessibility API for native app automation (via `--macax` flag)
+  - UI element inspection
+  - Screenshot capture and window management
+  - OCR text extraction from images and screen regions
+  - Window listing and identification
+- **Final Output**: Formatted result presentation
+
+### Provider Flexibility
+- Support for multiple LLM providers through a unified interface
+- Hot-swappable providers without code changes
+- Provider-specific optimizations and feature support
+- Local model support for offline operation
+
+### Task Automation
+- Single-shot task execution for quick operations
+- Iterative task mode for complex, multi-step workflows
+- Automatic error recovery and retry logic
+- Progress tracking and intermediate result handling
+
+## Language & Technology Stack
+
+- **Language**: Rust (2021 edition)
+- **Async Runtime**: Tokio for concurrent operations
+- **HTTP Client**: Reqwest for API communications
+- **Serialization**: Serde for JSON handling
+- **CLI Framework**: Clap for command-line parsing
+- **Logging**: Tracing for structured logging
+- **Local Models**: llama.cpp with Metal acceleration support
+
+## Use Cases
+
+G3 is designed for:
+- Automated code generation and refactoring
+- File manipulation and project scaffolding
+- System administration tasks
+- Data processing and transformation  
+- API integration and testing
+- Documentation generation
+- Complex multi-step workflows
+- Desktop application automation and testing
+
+## Getting Started
+
+### Default Mode: Accumulative Autonomous
+
+The default interactive mode now uses **accumulative autonomous mode**, which combines the best of interactive and autonomous workflows:
+
+```bash
+# Simply run g3 in any directory
+g3
+
+# You'll be prompted to describe what you want to build
+# Each input you provide:
+# 1. Gets added to accumulated requirements
+# 2. Automatically triggers autonomous mode (coach-player loop)
+# 3. Implements your requirements iteratively
+
+# Example session:
+requirement> create a simple web server in Python with Flask
+# ... autonomous mode runs and implements it ...
+requirement> add a /health endpoint that returns JSON
+# ... autonomous mode runs again with both requirements ...
+```
+
+### Other Modes
+
+```bash
+# Single-shot mode (one task, then exit)
+g3 "implement a function to calculate fibonacci numbers"
+
+# Traditional autonomous mode (reads requirements.md)
+g3 --autonomous
+
+# Traditional chat mode (simple interactive chat without autonomous runs)
+g3 --chat
+```
+
+```bash
+# Build the project
+cargo build --release
+
+# Run from the build directory
+./target/release/g3
+
+# Or copy both files to somewhere in your PATH (macOS only needs both files)
+cp target/release/g3 ~/.local/bin/
+cp target/release/libVisionBridge.dylib ~/.local/bin/  # macOS only
+
+# Execute a task
+g3 "implement a function to calculate fibonacci numbers"
+```
+
+## WebDriver Browser Automation
+
+G3 includes WebDriver support for browser automation tasks using Safari.
+
+**One-Time Setup** (macOS only):
+
+Safari Remote Automation must be enabled before using WebDriver tools. Run this once:
+
+```bash
+# Option 1: Use the provided script
+./scripts/enable-safari-automation.sh
+
+# Option 2: Enable manually
+safaridriver --enable  # Requires password
+
+# Option 3: Enable via Safari UI
+# Safari → Preferences → Advanced → Show Develop menu
+# Then: Develop → Allow Remote Automation
+```
+
+**For detailed setup instructions and troubleshooting**, see [WebDriver Setup Guide](docs/webdriver-setup.md).
+
+**Usage**: Run G3 with the `--webdriver` flag to enable browser automation tools.
+
+## macOS Accessibility API Tools
+
+G3 includes support for controlling macOS applications via the Accessibility API, allowing you to automate native macOS apps.
+
+**Available Tools**: `macax_list_apps`, `macax_get_frontmost_app`, `macax_activate_app`, `macax_get_ui_tree`, `macax_find_elements`, `macax_click`, `macax_set_value`, `macax_get_value`, `macax_press_key`
+
+**Setup**: Enable with the `--macax` flag or in config with `macax.enabled = true`. Grant accessibility permissions:
+- **macOS**: System Preferences → Security & Privacy → Privacy → Accessibility → Add your terminal app
+
+**For detailed documentation**, see [macOS Accessibility Tools Guide](docs/macax-tools.md).
+
+**Note**: This is particularly useful for testing and automating apps you're building with G3, as you can add accessibility identifiers to your UI elements.
+
+## Computer Control (Experimental)
+
+G3 can interact with your computer's GUI for automation tasks:
+
+**Available Tools**: `mouse_click`, `type_text`, `find_element`, `take_screenshot`, `extract_text`, `find_text_on_screen`, `list_windows`
+
+**Setup**: Enable in config with `computer_control.enabled = true` and grant OS accessibility permissions:
+- **macOS**: System Preferences → Security & Privacy → Accessibility  
+- **Linux**: Ensure X11 or Wayland access
+- **Windows**: Run as administrator (first time only)
+
+## Session Logs
+
+G3 automatically saves session logs for each interaction in the `logs/` directory. These logs contain:
+- Complete conversation history
+- Token usage statistics
+- Timestamps and session status
+
+The `logs/` directory is created automatically on first use and is excluded from version control.
+
+## License
+
+MIT License - see LICENSE file for details
+
+## Contributing
+
+G3 is an open-source project. Contributions are welcome! Please see CONTRIBUTING.md for guidelines.

TODO

@@ -0,0 +1,19 @@
+next tasks
+
+x get something working with autonomous mode
+- g3d
+- bug where it prints everything in a conversation turn all over again before final_output
+x ui abstraction from core
+- context token counting bug
+- embedded model
+  - prompt rewriting
+  - generates status messages "ruffling feathers..."
+  - project description?
+- treesitter + friends
+x error where it just gives up turn
+- "project" behaviors (read readme first)
+- advance project mgmt
+  - git for reverting
+  - swarm
+  - ui tests / computer controller
+

config.coach-player.example.toml

@@ -0,0 +1,24 @@
+[providers]
+default_provider = "databricks"
+# Specify different providers for coach and player in autonomous mode
+coach = "databricks"    # Provider for coach (code reviewer) - can be more powerful/expensive
+player = "anthropic"    # Provider for player (code implementer) - can be faster/cheaper
+
+[providers.databricks]
+host = "https://your-workspace.cloud.databricks.com"
+# token = "your-databricks-token"  # Optional - will use OAuth if not provided
+model = "databricks-claude-sonnet-4"
+max_tokens = 4096
+temperature = 0.1
+use_oauth = true
+
+[providers.anthropic]
+api_key = "your-anthropic-api-key"
+model = "claude-3-haiku-20240307"  # Using a faster model for player
+max_tokens = 4096
+temperature = 0.3  # Slightly higher temperature for more creative implementations
+
+[agent]
+max_context_length = 8192
+enable_streaming = true
+timeout_seconds = 60

config.example.toml

@@ -1,38 +1,25 @@
-# Example configuration file for G3
-# Copy to ~/.config/g3/config.toml and customize
-
 [providers]
-default_provider = "embedded"
+default_provider = "databricks"
+# Optional: Specify different providers for coach and player in autonomous mode
+# If not specified, will use default_provider for both
+# coach = "databricks"    # Provider for coach (code reviewer)
+# player = "anthropic"    # Provider for player (code implementer)
+# Note: Make sure the specified providers are configured below
 
-[providers.openai]
-# Get your API key from https://platform.openai.com/api-keys
-api_key = "sk-your-openai-api-key-here"
-model = "gpt-4"
-# Optional: custom base URL for OpenAI-compatible APIs
-# base_url = "https://api.openai.com/v1"
-max_tokens = 2048
-temperature = 0.1
-
-[providers.anthropic]
-# Get your API key from https://console.anthropic.com/
-api_key = "your-anthropic-api-key-here"
-model = "claude-3-5-sonnet-20241022"
+[providers.databricks]
+host = "https://your-workspace.cloud.databricks.com"
+# token = "your-databricks-token"  # Optional - will use OAuth if not provided
+model = "databricks-claude-sonnet-4"
 max_tokens = 4096
 temperature = 0.1
-
-[providers.embedded]
-# Path to your GGUF model file
-model_path = "~/.cache/g3/models/codellama-7b-instruct.Q4_K_M.gguf"
-model_type = "codellama"
-context_length = 16384  # Use CodeLlama's full context capability
-max_tokens = 2048  # Default fallback, but will be calculated dynamically
-temperature = 0.1
-# Number of layers to offload to GPU (0 for CPU only)
-gpu_layers = 32
-# Number of CPU threads to use
-threads = 8
+use_oauth = true
 
 [agent]
 max_context_length = 8192
 enable_streaming = true
 timeout_seconds = 60
+
+[computer_control]
+enabled = false  # Set to true to enable computer control (requires OS permissions)
+require_confirmation = true
+max_actions_per_second = 5

config.ollama.example.toml

@@ -0,0 +1,26 @@
+# Example G3 configuration using Ollama provider
+# Copy this to ~/.config/g3/config.toml or ./g3.toml to use it
+
+[providers]
+default_provider = "ollama"
+
+# Ollama configuration (local LLM)
+[providers.ollama]
+model = "llama3.2"  # or qwen2.5, mistral, etc.
+# base_url = "http://localhost:11434"  # Optional, defaults to localhost
+# max_tokens = 2048  # Optional
+# temperature = 0.7  # Optional
+
+# Optional: Specify different providers for coach and player in autonomous mode
+# coach = "ollama"    # Provider for coach (code reviewer)
+# player = "ollama"   # Provider for player (code implementer)
+
+[agent]
+max_context_length = 8192
+enable_streaming = true
+timeout_seconds = 60
+
+[computer_control]
+enabled = false  # Set to true to enable computer control (requires OS permissions)
+require_confirmation = true
+max_actions_per_second = 5

crates/g3-cli/Cargo.toml

@@ -12,9 +12,13 @@ tokio = { workspace = true }
 anyhow = { workspace = true }
 tracing = { workspace = true }
 tracing-subscriber = { workspace = true, features = ["env-filter"] }
-serde = { workspace = true }
+serde = { workspace = true, features = ["derive"] }
 serde_json = { workspace = true }
 rustyline = "17.0.1"
 dirs = "5.0"
 tokio-util = "0.7"
 indicatif = "0.17"
+chrono = { version = "0.4", features = ["serde"] }
+crossterm = "0.29.0"
+ratatui = "0.29"
+termimad = "0.34.0"

crates/g3-cli/src/machine_ui_writer.rs

@@ -0,0 +1,94 @@
+use g3_core::ui_writer::UiWriter;
+use std::io::{self, Write};
+
+/// Machine-mode implementation of UiWriter that prints plain, unformatted output
+/// This is designed for programmatic consumption and outputs everything verbatim
+pub struct MachineUiWriter;
+
+impl MachineUiWriter {
+    pub fn new() -> Self {
+        Self
+    }
+}
+
+impl UiWriter for MachineUiWriter {
+    fn print(&self, message: &str) {
+        print!("{}", message);
+    }
+
+    fn println(&self, message: &str) {
+        println!("{}", message);
+    }
+
+    fn print_inline(&self, message: &str) {
+        print!("{}", message);
+        let _ = io::stdout().flush();
+    }
+
+    fn print_system_prompt(&self, prompt: &str) {
+        println!("SYSTEM_PROMPT:");
+        println!("{}", prompt);
+        println!("END_SYSTEM_PROMPT");
+        println!();
+    }
+
+    fn print_context_status(&self, message: &str) {
+        println!("CONTEXT_STATUS: {}", message);
+    }
+
+    fn print_context_thinning(&self, message: &str) {
+        println!("CONTEXT_THINNING: {}", message);
+    }
+
+    fn print_tool_header(&self, tool_name: &str) {
+        println!("TOOL_CALL: {}", tool_name);
+    }
+
+    fn print_tool_arg(&self, key: &str, value: &str) {
+        println!("TOOL_ARG: {} = {}", key, value);
+    }
+
+    fn print_tool_output_header(&self) {
+        println!("TOOL_OUTPUT:");
+    }
+
+    fn update_tool_output_line(&self, line: &str) {
+        println!("{}", line);
+    }
+
+    fn print_tool_output_line(&self, line: &str) {
+        println!("{}", line);
+    }
+
+    fn print_tool_output_summary(&self, count: usize) {
+        println!("TOOL_OUTPUT_LINES: {}", count);
+    }
+
+    fn print_tool_timing(&self, duration_str: &str) {
+        println!("TOOL_DURATION: {}", duration_str);
+        println!("END_TOOL_OUTPUT");
+        println!();
+    }
+
+    fn print_agent_prompt(&self) {
+        println!("AGENT_RESPONSE:");
+        let _ = io::stdout().flush();
+    }
+
+    fn print_agent_response(&self, content: &str) {
+        print!("{}", content);
+        let _ = io::stdout().flush();
+    }
+
+    fn notify_sse_received(&self) {
+        // No-op for machine mode
+    }
+
+    fn flush(&self) {
+        let _ = io::stdout().flush();
+    }
+    
+    fn wants_full_output(&self) -> bool {
+        true  // Machine mode wants complete, untruncated output
+    }
+}

crates/g3-cli/src/simple_output.rs

@@ -0,0 +1,32 @@
+/// Simple output helper for printing messages
+pub struct SimpleOutput {
+    machine_mode: bool,
+}
+
+impl SimpleOutput {
+    pub fn new() -> Self {
+        SimpleOutput { machine_mode: false }
+    }
+
+    pub fn new_with_mode(machine_mode: bool) -> Self {
+        SimpleOutput { machine_mode }
+    }
+
+    pub fn print(&self, message: &str) {
+        if !self.machine_mode {
+            println!("{}", message);
+        }
+    }
+
+    pub fn print_smart(&self, message: &str) {
+        if !self.machine_mode {
+            println!("{}", message);
+        }
+    }
+}
+
+impl Default for SimpleOutput {
+    fn default() -> Self {
+        Self::new()
+    }
+}

crates/g3-cli/src/theme.rs

@@ -0,0 +1,147 @@
+use ratatui::style::Color;
+use serde::{Deserialize, Serialize};
+use std::fs;
+use std::path::Path;
+use anyhow::Result;
+
+/// Color theme configuration for the retro TUI
+#[derive(Debug, Clone, Serialize, Deserialize)]
+pub struct ColorTheme {
+    /// Name of the theme
+    pub name: String,
+    
+    /// Main terminal text color (for general output)
+    pub terminal_green: ColorValue,
+    
+    /// Warning/system messages color
+    pub terminal_amber: ColorValue,
+    
+    /// Border and dim text color
+    pub terminal_dim_green: ColorValue,
+    
+    /// Background color
+    pub terminal_bg: ColorValue,
+    
+    /// Highlight/emphasis color
+    pub terminal_cyan: ColorValue,
+    
+    /// Error/negative diff color
+    pub terminal_red: ColorValue,
+    
+    /// READY status color
+    pub terminal_pale_blue: ColorValue,
+    
+    /// PROCESSING status color
+    pub terminal_dark_amber: ColorValue,
+    
+    /// Bright/punchy text color
+    pub terminal_white: ColorValue,
+    
+    /// Success status color (for tool completions)
+    pub terminal_success: ColorValue,
+}
+
+/// Represents a color value that can be serialized/deserialized
+#[derive(Debug, Clone, Serialize, Deserialize)]
+#[serde(untagged)]
+pub enum ColorValue {
+    /// RGB color with r, g, b components
+    Rgb { r: u8, g: u8, b: u8 },
+    /// Named color
+    Named(String),
+}
+
+impl ColorValue {
+    /// Convert to ratatui Color
+    pub fn to_color(&self) -> Color {
+        match self {
+            ColorValue::Rgb { r, g, b } => Color::Rgb(*r, *g, *b),
+            ColorValue::Named(name) => match name.to_lowercase().as_str() {
+                "black" => Color::Black,
+                "red" => Color::Red,
+                "green" => Color::Green,
+                "yellow" => Color::Yellow,
+                "blue" => Color::Blue,
+                "magenta" => Color::Magenta,
+                "cyan" => Color::Cyan,
+                "gray" | "grey" => Color::Gray,
+                "darkgray" | "darkgrey" => Color::DarkGray,
+                "lightred" => Color::LightRed,
+                "lightgreen" => Color::LightGreen,
+                "lightyellow" => Color::LightYellow,
+                "lightblue" => Color::LightBlue,
+                "lightmagenta" => Color::LightMagenta,
+                "lightcyan" => Color::LightCyan,
+                "white" => Color::White,
+                _ => Color::White, // Default fallback
+            },
+        }
+    }
+}
+
+impl ColorTheme {
+    /// Load a theme from a JSON file
+    pub fn from_file<P: AsRef<Path>>(path: P) -> Result<Self> {
+        let content = fs::read_to_string(path)?;
+        let theme: ColorTheme = serde_json::from_str(&content)?;
+        Ok(theme)
+    }
+    
+    /// Get the default retro sci-fi theme (inspired by Alien terminals)
+    pub fn default() -> Self {
+        ColorTheme {
+            name: "Retro Sci-Fi".to_string(),
+            terminal_green: ColorValue::Rgb { r: 136, g: 244, b: 152 },
+            terminal_amber: ColorValue::Rgb { r: 242, g: 204, b: 148 },
+            terminal_dim_green: ColorValue::Rgb { r: 154, g: 174, b: 135 },
+            terminal_bg: ColorValue::Rgb { r: 0, g: 10, b: 0 },
+            terminal_cyan: ColorValue::Rgb { r: 0, g: 255, b: 255 },
+            terminal_red: ColorValue::Rgb { r: 239, g: 119, b: 109 },
+            terminal_pale_blue: ColorValue::Rgb { r: 173, g: 234, b: 251 },
+            terminal_dark_amber: ColorValue::Rgb { r: 204, g: 119, b: 34 },
+            terminal_white: ColorValue::Rgb { r: 218, g: 218, b: 219 },
+            terminal_success: ColorValue::Rgb { r: 136, g: 244, b: 152 }, // Same as terminal_green for retro theme
+        }
+    }
+    
+    /// Get the Dracula theme
+    pub fn dracula() -> Self {
+        ColorTheme {
+            name: "Dracula".to_string(),
+            terminal_green: ColorValue::Rgb { r: 248, g: 248, b: 242 }, // Use Dracula foreground (white) for main text
+            terminal_amber: ColorValue::Rgb { r: 255, g: 184, b: 108 }, // Dracula orange
+            terminal_dim_green: ColorValue::Rgb { r: 98, g: 114, b: 164 }, // Dracula comment
+            terminal_bg: ColorValue::Rgb { r: 40, g: 42, b: 54 },      // Dracula background
+            terminal_cyan: ColorValue::Rgb { r: 139, g: 233, b: 253 },  // Dracula cyan
+            terminal_red: ColorValue::Rgb { r: 255, g: 85, b: 85 },     // Dracula red
+            terminal_pale_blue: ColorValue::Rgb { r: 189, g: 147, b: 249 }, // Dracula purple
+            terminal_dark_amber: ColorValue::Rgb { r: 255, g: 121, b: 198 }, // Dracula pink
+            terminal_white: ColorValue::Rgb { r: 248, g: 248, b: 242 }, // Dracula foreground
+            terminal_success: ColorValue::Rgb { r: 80, g: 250, b: 123 }, // Dracula green for success
+        }
+    }
+    
+    /// Get a theme by name or from file
+    pub fn load(theme_name: Option<&str>) -> Result<Self> {
+        match theme_name {
+            None => Ok(Self::default()),
+            Some("default") | Some("retro") => Ok(Self::default()),
+            Some("dracula") => Ok(Self::dracula()),
+            Some(path) => {
+                // Try to load from file
+                if Path::new(path).exists() {
+                    Self::from_file(path)
+                } else {
+                    // Try to find in standard locations
+                    let home = dirs::home_dir().ok_or_else(|| anyhow::anyhow!("Could not find home directory"))?;
+                    let theme_file = home.join(".config").join("g3").join("themes").join(format!("{}.json", path));
+                    if theme_file.exists() {
+                        Self::from_file(theme_file)
+                    } else {
+                        Err(anyhow::anyhow!("Theme '{}' not found", path))
+                    }
+                }
+            }
+        }
+    }
+}

crates/g3-cli/src/tui.rs

@@ -0,0 +1,160 @@
+use crossterm::style::Color;
+use crossterm::style::{SetForegroundColor, ResetColor};
+use std::io::{self, Write};
+use termimad::MadSkin;
+
+/// Simple output handler with markdown support
+pub struct SimpleOutput {
+    mad_skin: MadSkin,
+}
+
+impl SimpleOutput {
+    pub fn new() -> Self {
+        let mut mad_skin = MadSkin::default();
+        // Dracula color scheme
+        // Background: #282a36, Foreground: #f8f8f2
+        // Colors: Cyan #8be9fd, Green #50fa7b, Orange #ffb86c, Pink #ff79c6, Purple #bd93f9, Red #ff5555, Yellow #f1fa8c
+        
+        mad_skin.set_headers_fg(Color::Rgb { r: 189, g: 147, b: 249 }); // Purple for headers
+        mad_skin.bold.set_fg(Color::Rgb { r: 255, g: 121, b: 198 });    // Pink for bold
+        mad_skin.italic.set_fg(Color::Rgb { r: 139, g: 233, b: 253 });  // Cyan for italic
+        mad_skin.code_block.set_bg(Color::Rgb { r: 68, g: 71, b: 90 }); // Dracula background variant
+        mad_skin.code_block.set_fg(Color::Rgb { r: 80, g: 250, b: 123 }); // Green for code text
+        mad_skin.inline_code.set_bg(Color::Rgb { r: 68, g: 71, b: 90 }); // Same background for inline code
+        mad_skin.inline_code.set_fg(Color::Rgb { r: 241, g: 250, b: 140 }); // Yellow for inline code
+        mad_skin.quote_mark.set_fg(Color::Rgb { r: 98, g: 114, b: 164 }); // Comment purple for quote marks
+        mad_skin.strikeout.set_fg(Color::Rgb { r: 255, g: 85, b: 85 });  // Red for strikethrough
+        
+        Self { mad_skin }
+    }
+
+    /// Detect if text contains markdown formatting
+    fn has_markdown(&self, text: &str) -> bool {
+        // Check for common markdown patterns
+        text.contains("**") ||
+        text.contains("```") ||
+        text.contains("`") ||
+        text.lines().any(|line| {
+            let trimmed = line.trim();
+            trimmed.starts_with('#') ||
+            trimmed.starts_with("- ") ||
+            trimmed.starts_with("* ") ||
+            trimmed.starts_with("+ ") ||
+            (trimmed.len() > 2 && 
+             trimmed.chars().next().is_some_and(|c| c.is_ascii_digit()) &&
+             trimmed.chars().nth(1) == Some('.') &&
+             trimmed.chars().nth(2) == Some(' ')) ||
+            (trimmed.contains('[') && trimmed.contains("]("))
+        }) ||
+        (text.matches('*').count() >= 2 && !text.contains("/*") && !text.contains("*/"))
+    }
+
+    pub fn print(&self, text: &str) {
+        println!("{}", text);
+    }
+
+    /// Smart print that automatically detects and renders markdown
+    pub fn print_smart(&self, text: &str) {
+        if self.has_markdown(text) {
+            self.print_markdown(text);
+        } else {
+            self.print(text);
+        }
+    }
+
+    pub fn print_markdown(&self, markdown: &str) {
+        self.mad_skin.print_text(markdown);
+    }
+
+    pub fn _print_status(&self, status: &str) {
+        println!("📊 {}", status);
+    }
+
+    pub fn print_context(&self, used: u32, total: u32, percentage: f32) {
+        let total_dots = 10;
+        let filled_dots = ((percentage / 100.0) * total_dots as f32) as usize;
+        let empty_dots = total_dots.saturating_sub(filled_dots);
+
+        let filled_str = "●".repeat(filled_dots);
+        let empty_str = "○".repeat(empty_dots);
+        
+        // Determine color based on percentage
+        let color = if percentage < 40.0 {
+            crossterm::style::Color::Green
+        } else if percentage < 60.0 {
+            crossterm::style::Color::Yellow
+        } else if percentage < 80.0 {
+            crossterm::style::Color::Rgb { r: 255, g: 165, b: 0 } // Orange
+        } else {
+            crossterm::style::Color::Red
+        };
+
+        // Print with colored progress bar
+        print!("Context: ");
+        print!("{}", SetForegroundColor(color));
+        print!("{}{}", filled_str, empty_str);
+        print!("{}", ResetColor);
+        println!(" {:.0}% ({}/{} tokens)", percentage, used, total);
+    }
+
+    pub fn print_context_thinning(&self, message: &str) {
+        // Animated highlight for context thinning
+        // Use bright cyan/green with a quick flash animation
+        
+        // Flash animation: print with bright background, then normal
+        let frames = vec![
+            "\x1b[1;97;46m",  // Frame 1: Bold white on cyan background
+            "\x1b[1;97;42m",  // Frame 2: Bold white on green background
+            "\x1b[1;96;40m",  // Frame 3: Bold cyan on black background
+        ];
+        
+        println!();
+        
+        // Quick flash animation
+        for frame in &frames {
+            print!("\r{} ✨ {} ✨\x1b[0m", frame, message);
+            let _ = io::stdout().flush();
+            std::thread::sleep(std::time::Duration::from_millis(80));
+        }
+        
+        // Final display with bright cyan and sparkle emojis
+        print!("\r\x1b[1;96m✨ {} ✨\x1b[0m", message);
+        println!();
+        
+        // Add a subtle "success" indicator line
+        println!("\x1b[2;36m   └─ Context optimized successfully\x1b[0m");
+        println!();
+        
+        let _ = io::stdout().flush();
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn test_markdown_detection() {
+        let output = SimpleOutput::new();
+        
+        // Should detect markdown
+        assert!(output.has_markdown("**bold text**"));
+        assert!(output.has_markdown("`code`"));
+        assert!(output.has_markdown("```\ncode block\n```"));
+        assert!(output.has_markdown("# Header"));
+        assert!(output.has_markdown("- list item"));
+        assert!(output.has_markdown("* list item"));
+        assert!(output.has_markdown("+ list item"));
+        assert!(output.has_markdown("1. numbered item"));
+        assert!(output.has_markdown("[link](url)"));
+        assert!(output.has_markdown("*italic* text"));
+        
+        // Should NOT detect markdown
+        assert!(!output.has_markdown("plain text"));
+        assert!(!output.has_markdown("file.txt"));
+        assert!(!output.has_markdown("/* comment */"));
+        assert!(!output.has_markdown("just one * asterisk"));
+        assert!(!output.has_markdown("📁 Workspace: /path/to/dir"));
+        assert!(!output.has_markdown("✅ Success message"));
+    }
+}

crates/g3-cli/src/ui_writer_impl.rs

@@ -0,0 +1,347 @@
+use g3_core::ui_writer::UiWriter;
+use std::io::{self, Write};
+use std::sync::Mutex;
+
+/// Console implementation of UiWriter that prints to stdout
+pub struct ConsoleUiWriter {
+    current_tool_name: Mutex<Option<String>>,
+    current_tool_args: Mutex<Vec<(String, String)>>,
+    current_output_line: Mutex<Option<String>>,
+    output_line_printed: Mutex<bool>,
+    in_todo_tool: Mutex<bool>,
+}
+
+impl ConsoleUiWriter {
+    pub fn new() -> Self {
+        Self {
+            current_tool_name: Mutex::new(None),
+            current_tool_args: Mutex::new(Vec::new()),
+            current_output_line: Mutex::new(None),
+            output_line_printed: Mutex::new(false),
+            in_todo_tool: Mutex::new(false),
+        }
+    }
+
+    fn print_todo_line(&self, line: &str) {
+        // Transform and print todo list lines elegantly
+        let trimmed = line.trim();
+        
+        // Skip the "📝 TODO list:" prefix line
+        if trimmed.starts_with("📝 TODO list:") || trimmed == "📝 TODO list is empty" {
+            return;
+        }
+        
+        // Handle empty lines
+        if trimmed.is_empty() {
+            println!();
+            return;
+        }
+        
+        // Detect indentation level
+        let indent_count = line.chars().take_while(|c| c.is_whitespace()).count();
+        let indent = "  ".repeat(indent_count / 2); // Convert spaces to visual indent
+        
+        // Format based on line type
+        if trimmed.starts_with("- [ ]") {
+            // Incomplete task
+            let task = trimmed.strip_prefix("- [ ]").unwrap_or(trimmed).trim();
+            println!("{}☐ {}", indent, task);
+        } else if trimmed.starts_with("- [x]") || trimmed.starts_with("- [X]") {
+            // Completed task
+            let task = trimmed.strip_prefix("- [x]")
+                .or_else(|| trimmed.strip_prefix("- [X]"))
+                .unwrap_or(trimmed)
+                .trim();
+            println!("{}\x1b[2m☑ {}\x1b[0m", indent, task);
+        } else if trimmed.starts_with("- ") {
+            // Regular bullet point
+            let item = trimmed.strip_prefix("- ").unwrap_or(trimmed).trim();
+            println!("{}• {}", indent, item);
+        } else if trimmed.starts_with("# ") {
+            // Heading
+            let heading = trimmed.strip_prefix("# ").unwrap_or(trimmed).trim();
+            println!("\n\x1b[1m{}\x1b[0m", heading);
+        } else if trimmed.starts_with("## ") {
+            // Subheading
+            let subheading = trimmed.strip_prefix("## ").unwrap_or(trimmed).trim();
+            println!("\n\x1b[1m{}\x1b[0m", subheading);
+        } else if trimmed.starts_with("**") && trimmed.ends_with("**") {
+            // Bold text (section marker)
+            let text = trimmed.trim_start_matches("**").trim_end_matches("**");
+            println!("{}\x1b[1m{}\x1b[0m", indent, text);
+        } else {
+            // Regular text or note
+            println!("{}{}", indent, trimmed);
+        }
+    }
+}
+
+impl UiWriter for ConsoleUiWriter {
+    fn print(&self, message: &str) {
+        print!("{}", message);
+    }
+
+    fn println(&self, message: &str) {
+        println!("{}", message);
+    }
+
+    fn print_inline(&self, message: &str) {
+        print!("{}", message);
+        let _ = io::stdout().flush();
+    }
+
+    fn print_system_prompt(&self, prompt: &str) {
+        println!("🔍 System Prompt:");
+        println!("================");
+        println!("{}", prompt);
+        println!("================");
+        println!();
+    }
+
+    fn print_context_status(&self, message: &str) {
+        println!("{}", message);
+    }
+
+    fn print_context_thinning(&self, message: &str) {
+        // Animated highlight for context thinning
+        // Use bright cyan/green with a quick flash animation
+        
+        // Flash animation: print with bright background, then normal
+        let frames = vec![
+            "\x1b[1;97;46m",  // Frame 1: Bold white on cyan background
+            "\x1b[1;97;42m",  // Frame 2: Bold white on green background
+            "\x1b[1;96;40m",  // Frame 3: Bold cyan on black background
+        ];
+        
+        println!();
+        
+        // Quick flash animation
+        for frame in &frames {
+            print!("\r{} ✨ {} ✨\x1b[0m", frame, message);
+            let _ = io::stdout().flush();
+            std::thread::sleep(std::time::Duration::from_millis(80));
+        }
+        
+        // Final display with bright cyan and sparkle emojis
+        print!("\r\x1b[1;96m✨ {} ✨\x1b[0m", message);
+        println!();
+        
+        // Add a subtle "success" indicator line
+        println!("\x1b[2;36m   └─ Context optimized successfully\x1b[0m");
+        println!();
+        
+        let _ = io::stdout().flush();
+    }
+
+    fn print_tool_header(&self, tool_name: &str) {
+        // Store the tool name and clear args for collection
+        *self.current_tool_name.lock().unwrap() = Some(tool_name.to_string());
+        self.current_tool_args.lock().unwrap().clear();
+        
+        // Check if this is a todo tool call
+        let is_todo = tool_name == "todo_read" || tool_name == "todo_write";
+        *self.in_todo_tool.lock().unwrap() = is_todo;
+        
+        // For todo tools, we'll skip the normal header and print a custom one later
+        if is_todo {
+        }
+    }
+
+    fn print_tool_arg(&self, key: &str, value: &str) {
+        // Collect arguments instead of printing immediately
+        // Filter out any keys that look like they might be agent message content
+        // (e.g., keys that are suspiciously long or contain message-like content)
+        let is_valid_arg_key = key.len() < 50
+            && !key.contains('\n')
+            && !key.contains("I'll")
+            && !key.contains("Let me")
+            && !key.contains("Here's")
+            && !key.contains("I can");
+
+        if is_valid_arg_key {
+            self.current_tool_args
+                .lock()
+                .unwrap()
+                .push((key.to_string(), value.to_string()));
+        }
+    }
+
+    fn print_tool_output_header(&self) {
+        // Skip normal header for todo tools
+        if *self.in_todo_tool.lock().unwrap() {
+            println!(); // Just add a newline
+            return;
+        }
+        
+        println!();
+        // Now print the tool header with the most important arg in bold green
+        if let Some(tool_name) = self.current_tool_name.lock().unwrap().as_ref() {
+            let args = self.current_tool_args.lock().unwrap();
+
+            // Find the most important argument - prioritize file_path if available
+            let important_arg = args
+                .iter()
+                .find(|(k, _)| k == "file_path")
+                .or_else(|| args.iter().find(|(k, _)| k == "command" || k == "path"))
+                .or_else(|| args.first());
+
+            if let Some((_, value)) = important_arg {
+                // For multi-line values, only show the first line
+                let first_line = value.lines().next().unwrap_or("");
+
+                // Truncate long values for display
+                let display_value = if first_line.len() > 80 {
+                    // Use char_indices to safely truncate at character boundary
+                    let truncate_at = first_line.char_indices()
+                        .nth(77)
+                        .map(|(i, _)| i)
+                        .unwrap_or(first_line.len());
+                    format!("{}...", &first_line[..truncate_at])
+                } else {
+                    first_line.to_string()
+                };
+
+                // Add range information for read_file tool calls
+                let header_suffix = if tool_name == "read_file" {
+                    // Check if start or end parameters are present
+                    let has_start = args.iter().any(|(k, _)| k == "start");
+                    let has_end = args.iter().any(|(k, _)| k == "end");
+                    
+                    if has_start || has_end {
+                        let start_val = args.iter().find(|(k, _)| k == "start").map(|(_, v)| v.as_str()).unwrap_or("0");
+                        let end_val = args.iter().find(|(k, _)| k == "end").map(|(_, v)| v.as_str()).unwrap_or("end");
+                        format!(" [{}..{}]", start_val, end_val)
+                    } else {
+                        String::new()
+                    }
+                } else {
+                    String::new()
+                };
+
+                // Print with bold green tool name, purple (non-bold) for pipe and args
+                println!("┌─\x1b[1;32m {}\x1b[0m\x1b[35m | {}{}\x1b[0m", tool_name, display_value, header_suffix);
+            } else {
+                // Print with bold green formatting using ANSI escape codes
+                println!("┌─\x1b[1;32m {}\x1b[0m", tool_name);
+            }
+        }
+    }
+
+    fn update_tool_output_line(&self, line: &str) {
+        let mut current_line = self.current_output_line.lock().unwrap();
+        let mut line_printed = self.output_line_printed.lock().unwrap();
+
+        // If we've already printed a line, clear it first
+        if *line_printed {
+            // Move cursor up one line and clear it
+            print!("\x1b[1A\x1b[2K");
+        }
+
+        // Print the new line
+        println!("│ \x1b[2m{}\x1b[0m", line);
+        let _ = io::stdout().flush();
+
+        // Update state
+        *current_line = Some(line.to_string());
+        *line_printed = true;
+    }
+
+    fn print_tool_output_line(&self, line: &str) {
+        // Special handling for todo tools
+        if *self.in_todo_tool.lock().unwrap() {
+            self.print_todo_line(line);
+            return;
+        }
+        
+        println!("│ \x1b[2m{}\x1b[0m", line);
+    }
+
+    fn print_tool_output_summary(&self, count: usize) {
+        // Skip for todo tools
+        if *self.in_todo_tool.lock().unwrap() {
+            return;
+        }
+        
+        println!(
+            "│ \x1b[2m({} line{})\x1b[0m",
+            count,
+            if count == 1 { "" } else { "s" }
+        );
+    }
+
+    fn print_tool_timing(&self, duration_str: &str) {
+        // For todo tools, just print a simple completion message
+        if *self.in_todo_tool.lock().unwrap() {
+            println!();
+            *self.in_todo_tool.lock().unwrap() = false;
+            return;
+        }
+        
+        // Parse the duration string to determine color
+        // Format is like "1.5s", "500ms", "2m 30.0s"
+        let color_code = if duration_str.ends_with("ms") {
+            // Milliseconds - use default color (< 1s)
+            ""
+        } else if duration_str.contains('m') {
+            // Contains minutes
+            // Extract minutes value
+            if let Some(m_pos) = duration_str.find('m') {
+                if let Ok(minutes) = duration_str[..m_pos].trim().parse::<u32>() {
+                    if minutes >= 5 {
+                        "\x1b[31m" // Red for >= 5 minutes
+                    } else {
+                        "\x1b[38;5;208m" // Orange for >= 1 minute but < 5 minutes
+                    }
+                } else {
+                    "" // Default color if parsing fails
+                }
+            } else {
+                "" // Default color if 'm' not found (shouldn't happen)
+            }
+        } else if duration_str.ends_with('s') {
+            // Seconds only
+            if let Some(s_value) = duration_str.strip_suffix('s') {
+                if let Ok(seconds) = s_value.trim().parse::<f64>() {
+                    if seconds >= 1.0 {
+                        "\x1b[33m" // Yellow for >= 1 second
+                    } else {
+                        "" // Default color for < 1 second
+                    }
+                } else {
+                    "" // Default color if parsing fails
+                }
+            } else {
+                "" // Default color
+            }
+        } else {
+            // Milliseconds or other format - use default color
+            ""
+        };
+
+        println!("└─ ⚡️ {}{}\x1b[0m", color_code, duration_str);
+        println!();
+        // Clear the stored tool info
+        *self.current_tool_name.lock().unwrap() = None;
+        self.current_tool_args.lock().unwrap().clear();
+        *self.current_output_line.lock().unwrap() = None;
+        *self.output_line_printed.lock().unwrap() = false;
+    }
+
+    fn print_agent_prompt(&self) {
+        let _ = io::stdout().flush();
+    }
+
+    fn print_agent_response(&self, content: &str) {
+        print!("{}", content);
+        let _ = io::stdout().flush();
+    }
+
+    fn notify_sse_received(&self) {
+        // No-op for console - we don't track SSEs in console mode
+    }
+
+    fn flush(&self) {
+        let _ = io::stdout().flush();
+    }
+}
+

crates/g3-computer-control/Cargo.toml

@@ -0,0 +1,47 @@
+[package]
+name = "g3-computer-control"
+version = "0.1.0"
+edition = "2021"
+
+[build-dependencies]
+# Only needed for building Swift bridge on macOS
+
+[dependencies]
+# Workspace dependencies
+tokio = { workspace = true }
+anyhow = { workspace = true }
+thiserror = { workspace = true }
+serde = { workspace = true }
+serde_json = { workspace = true }
+tracing = { workspace = true }
+uuid = { workspace = true }
+
+shellexpand = "3.1"
+# Async trait support
+async-trait = "0.1"
+
+# WebDriver support
+fantoccini = "0.21"
+
+# macOS dependencies
+[target.'cfg(target_os = "macos")'.dependencies]
+core-graphics = "0.23"
+core-foundation = "0.10"
+cocoa = "0.25"
+objc = "0.2"
+accessibility = "0.2"
+image = "0.24"
+
+# Linux dependencies
+[target.'cfg(target_os = "linux")'.dependencies]
+x11 = { version = "2.21", features = ["xlib", "xtest"] }
+image = "0.24"
+
+# Windows dependencies
+[target.'cfg(target_os = "windows")'.dependencies]
+windows = { version = "0.52", features = [
+    "Win32_Foundation",
+    "Win32_UI_WindowsAndMessaging",
+    "Win32_UI_Input_KeyboardAndMouse",
+    "Win32_Graphics_Gdi",
+] }

crates/g3-computer-control/build.rs

@@ -0,0 +1,63 @@
+use std::env;
+use std::path::PathBuf;
+use std::process::Command;
+
+fn main() {
+    // Only build Vision bridge on macOS
+    if env::var("CARGO_CFG_TARGET_OS").unwrap() != "macos" {
+        return;
+    }
+
+    println!("cargo:rerun-if-changed=vision-bridge/Sources/VisionBridge/VisionOCR.swift");
+    println!("cargo:rerun-if-changed=vision-bridge/Sources/VisionBridge/VisionBridge.h");
+    println!("cargo:rerun-if-changed=vision-bridge/Package.swift");
+
+    let manifest_dir = PathBuf::from(env::var("CARGO_MANIFEST_DIR").unwrap());
+    let vision_bridge_dir = manifest_dir.join("vision-bridge");
+
+    // Build Swift package
+    println!("cargo:warning=Building VisionBridge Swift package...");
+    let build_status = Command::new("swift")
+        .args(&["build", "-c", "release"])
+        .current_dir(&vision_bridge_dir)
+        .status()
+        .expect("Failed to build Swift package");
+
+    if !build_status.success() {
+        panic!("Swift build failed");
+    }
+
+    // Find the built library
+    let lib_path = vision_bridge_dir
+        .join(".build/release")
+        .canonicalize()
+        .expect("Failed to find .build/release directory");
+
+    // Copy the dylib to the output directory so it can be found at runtime
+    let target_dir = manifest_dir.parent().unwrap().parent().unwrap().join("target");
+    let profile = env::var("PROFILE").unwrap_or_else(|_| "debug".to_string());
+    let output_dir = target_dir.join(&profile);
+    
+    let dylib_src = lib_path.join("libVisionBridge.dylib");
+    let dylib_dst = output_dir.join("libVisionBridge.dylib");
+    
+    std::fs::copy(&dylib_src, &dylib_dst)
+        .expect(&format!("Failed to copy dylib from {} to {}", dylib_src.display(), dylib_dst.display()));
+    
+    println!("cargo:warning=Copied libVisionBridge.dylib to {}", dylib_dst.display());
+    
+    // Add rpath so the dylib can be found at runtime
+    println!("cargo:rustc-link-arg=-Wl,-rpath,@executable_path");
+    println!("cargo:rustc-link-arg=-Wl,-rpath,@loader_path");
+    println!("cargo:rustc-link-search=native={}", lib_path.display());
+    println!("cargo:rustc-link-lib=dylib=VisionBridge");
+
+    // Link required frameworks
+    println!("cargo:rustc-link-lib=framework=Vision");
+    println!("cargo:rustc-link-lib=framework=AppKit");
+    println!("cargo:rustc-link-lib=framework=Foundation");
+    println!("cargo:rustc-link-lib=framework=CoreGraphics");
+    println!("cargo:rustc-link-lib=framework=CoreImage");
+
+    println!("cargo:warning=VisionBridge built successfully at {}", lib_path.display());
+}

crates/g3-computer-control/examples/debug_screenshot.rs

@@ -0,0 +1,46 @@
+use core_graphics::display::CGDisplay;
+
+fn main() {
+    let display = CGDisplay::main();
+    let image = display.image().expect("Failed to capture screen");
+    
+    println!("CGImage properties:");
+    println!("  Width: {}", image.width());
+    println!("  Height: {}", image.height());
+    println!("  Bits per component: {}", image.bits_per_component());
+    println!("  Bits per pixel: {}", image.bits_per_pixel());
+    println!("  Bytes per row: {}", image.bytes_per_row());
+    
+    let data = image.data();
+    let expected_size = image.width() * image.height() * 4;
+    println!("  Data length: {}", data.len());
+    println!("  Expected (w*h*4): {}", expected_size);
+    
+    // Check if there's padding in rows
+    let bytes_per_row = image.bytes_per_row();
+    let width = image.width();
+    let expected_bytes_per_row = width * 4;
+    println!("\nRow alignment:");
+    println!("  Actual bytes per row: {}", bytes_per_row);
+    println!("  Expected (width * 4): {}", expected_bytes_per_row);
+    println!("  Padding per row: {}", bytes_per_row - expected_bytes_per_row);
+    
+    // Sample some pixels from different locations
+    println!("\nFirst 3 pixels (raw bytes):");
+    for i in 0..3 {
+        let offset = i * 4;
+        println!("  Pixel {}: [{:3}, {:3}, {:3}, {:3}]", 
+                 i, data[offset], data[offset+1], data[offset+2], data[offset+3]);
+    }
+    
+    // Check a pixel from the middle
+    let mid_row = image.height() / 2;
+    let mid_col = image.width() / 2;
+    let mid_offset = (mid_row * bytes_per_row + mid_col * 4) as usize;
+    println!("\nMiddle pixel (row {}, col {}):", mid_row, mid_col);
+    println!("  Offset: {}", mid_offset);
+    if mid_offset + 3 < data.len() as usize {
+        println!("  Bytes: [{:3}, {:3}, {:3}, {:3}]", 
+                 data[mid_offset], data[mid_offset+1], data[mid_offset+2], data[mid_offset+3]);
+    }
+}

crates/g3-computer-control/examples/list_windows.rs

@@ -0,0 +1,56 @@
+use core_graphics::window::{kCGWindowListOptionOnScreenOnly, kCGNullWindowID, CGWindowListCopyWindowInfo};
+use core_foundation::dictionary::CFDictionary;
+use core_foundation::string::CFString;
+use core_foundation::base::{TCFType, ToVoid};
+
+fn main() {
+    println!("Listing all on-screen windows...");
+    println!("{:<10} {:<25} {}", "Window ID", "Owner", "Title");
+    println!("{}", "-".repeat(80));
+    
+    unsafe {
+        let window_list = CGWindowListCopyWindowInfo(
+            kCGWindowListOptionOnScreenOnly,
+            kCGNullWindowID
+        );
+        
+        let count = core_foundation::array::CFArray::<CFDictionary>::wrap_under_create_rule(window_list).len();
+        let array = core_foundation::array::CFArray::<CFDictionary>::wrap_under_create_rule(window_list);
+        
+        for i in 0..count {
+            let dict = array.get(i).unwrap();
+            
+            // Get window ID
+            let window_id_key = CFString::from_static_string("kCGWindowNumber");
+            let window_id: i64 = if let Some(value) = dict.find(window_id_key.to_void()) {
+                let num: core_foundation::number::CFNumber = TCFType::wrap_under_get_rule(*value as *const _);
+                num.to_i64().unwrap_or(0)
+            } else {
+                0
+            };
+            
+            // Get owner name
+            let owner_key = CFString::from_static_string("kCGWindowOwnerName");
+            let owner: String = if let Some(value) = dict.find(owner_key.to_void()) {
+                let s: CFString = TCFType::wrap_under_get_rule(*value as *const _);
+                s.to_string()
+            } else {
+                "Unknown".to_string()
+            };
+            
+            // Get window name/title
+            let name_key = CFString::from_static_string("kCGWindowName");
+            let title: String = if let Some(value) = dict.find(name_key.to_void()) {
+                let s: CFString = TCFType::wrap_under_get_rule(*value as *const _);
+                s.to_string()
+            } else {
+                "".to_string()
+            };
+            
+            // Show all windows
+            if !owner.is_empty() {
+                println!("{:<10} {:<25} {}", window_id, owner, title);
+            }
+        }
+    }
+}

crates/g3-computer-control/examples/macax_demo.rs

@@ -0,0 +1,74 @@
+//! Example demonstrating macOS Accessibility API tools
+//!
+//! This example shows how to use the macax tools to control macOS applications.
+//!
+//! Run with: cargo run --example macax_demo
+
+use anyhow::Result;
+use g3_computer_control::MacAxController;
+
+#[tokio::main]
+async fn main() -> Result<()> {
+    println!("🍎 macOS Accessibility API Demo\n");
+    println!("This demo shows how to control macOS applications using the Accessibility API.\n");
+    
+    // Create controller
+    let controller = MacAxController::new()?;
+    println!("✅ MacAxController initialized\n");
+    
+    // List running applications
+    println!("📱 Listing running applications:");
+    match controller.list_applications() {
+        Ok(apps) => {
+            for app in apps.iter().take(10) {
+                println!("  - {}", app.name);
+            }
+            if apps.len() > 10 {
+                println!("  ... and {} more", apps.len() - 10);
+            }
+        }
+        Err(e) => println!("  ❌ Error: {}", e),
+    }
+    println!();
+    
+    // Get frontmost app
+    println!("🎯 Getting frontmost application:");
+    match controller.get_frontmost_app() {
+        Ok(app) => println!("  Current: {}", app.name),
+        Err(e) => println!("  ❌ Error: {}", e),
+    }
+    println!();
+    
+    // Example: Activate Finder and get its UI tree
+    println!("📂 Activating Finder and inspecting UI:");
+    match controller.activate_app("Finder") {
+        Ok(_) => {
+            println!("  ✅ Finder activated");
+            
+            // Wait a moment for activation
+            tokio::time::sleep(tokio::time::Duration::from_millis(500)).await;
+            
+            // Get UI tree
+            match controller.get_ui_tree("Finder", 2) {
+                Ok(tree) => {
+                    println!("\n  UI Tree:");
+                    for line in tree.lines().take(10) {
+                        println!("    {}", line);
+                    }
+                }
+                Err(e) => println!("  ❌ Error getting UI tree: {}", e),
+            }
+        }
+        Err(e) => println!("  ❌ Error: {}", e),
+    }
+    println!();
+    
+    println!("✨ Demo complete!\n");
+    println!("💡 Tips:");
+    println!("  - Use --macax flag with g3 to enable these tools");
+    println!("  - Grant accessibility permissions in System Preferences");
+    println!("  - Add accessibility identifiers to your apps for easier automation");
+    println!("  - See docs/macax-tools.md for full documentation\n");
+    
+    Ok(())
+}

crates/g3-computer-control/examples/safari_demo.rs

@@ -0,0 +1,64 @@
+use g3_computer_control::SafariDriver;
+use g3_computer_control::webdriver::WebDriverController;
+use anyhow::Result;
+
+#[tokio::main]
+async fn main() -> Result<()> {
+    println!("Safari WebDriver Demo");
+    println!("=====================\n");
+    
+    println!("Make sure to:");
+    println!("1. Enable 'Allow Remote Automation' in Safari's Develop menu");
+    println!("2. Run: /usr/bin/safaridriver --enable");
+    println!("3. Start safaridriver in another terminal: safaridriver --port 4444\n");
+    
+    println!("Connecting to SafariDriver...");
+    let mut driver = SafariDriver::new().await?;
+    println!("✅ Connected!\n");
+    
+    // Navigate to a website
+    println!("Navigating to example.com...");
+    driver.navigate("https://example.com").await?;
+    println!("✅ Navigated\n");
+    
+    // Get page title
+    let title = driver.title().await?;
+    println!("Page title: {}\n", title);
+    
+    // Get current URL
+    let url = driver.current_url().await?;
+    println!("Current URL: {}\n", url);
+    
+    // Find an element
+    println!("Finding h1 element...");
+    let h1 = driver.find_element("h1").await?;
+    let h1_text = h1.text().await?;
+    println!("H1 text: {}\n", h1_text);
+    
+    // Find all paragraphs
+    println!("Finding all paragraphs...");
+    let paragraphs = driver.find_elements("p").await?;
+    println!("Found {} paragraphs\n", paragraphs.len());
+    
+    // Get page source
+    println!("Getting page source...");
+    let source = driver.page_source().await?;
+    println!("Page source length: {} bytes\n", source.len());
+    
+    // Execute JavaScript
+    println!("Executing JavaScript...");
+    let result = driver.execute_script("return document.title", vec![]).await?;
+    println!("JS result: {:?}\n", result);
+    
+    // Take a screenshot
+    println!("Taking screenshot...");
+    driver.screenshot("/tmp/safari_demo.png").await?;
+    println!("✅ Screenshot saved to /tmp/safari_demo.png\n");
+    
+    // Close the browser
+    println!("Closing browser...");
+    driver.quit().await?;
+    println!("✅ Done!");
+    
+    Ok(())
+}

crates/g3-computer-control/examples/test_permission_prompt.rs

@@ -0,0 +1,21 @@
+use g3_computer_control::create_controller;
+
+#[tokio::main]
+async fn main() {
+    println!("Testing screenshot with permission prompt...");
+    
+    let controller = create_controller().expect("Failed to create controller");
+    
+    match controller.take_screenshot("/tmp/test_with_prompt.png", None, None).await {
+        Ok(_) => {
+            println!("\n✅ Screenshot saved to /tmp/test_with_prompt.png");
+            println!("Opening screenshot...");
+            let _ = std::process::Command::new("open")
+                .arg("/tmp/test_with_prompt.png")
+                .spawn();
+        }
+        Err(e) => {
+            println!("❌ Screenshot failed: {}", e);
+        }
+    }
+}

crates/g3-computer-control/examples/test_screencapture_direct.rs

@@ -0,0 +1,39 @@
+use std::process::Command;
+
+fn main() {
+    let path = "/tmp/rust_screencapture_test.png";
+    
+    println!("Testing screencapture command from Rust...");
+    
+    let mut cmd = Command::new("screencapture");
+    cmd.arg("-x"); // No sound
+    cmd.arg(path);
+    
+    println!("Command: {:?}", cmd);
+    
+    match cmd.output() {
+        Ok(output) => {
+            println!("Exit status: {}", output.status);
+            println!("Stdout: {}", String::from_utf8_lossy(&output.stdout));
+            println!("Stderr: {}", String::from_utf8_lossy(&output.stderr));
+            
+            if output.status.success() {
+                println!("\n✅ Screenshot saved to: {}", path);
+                
+                // Check file exists and size
+                if let Ok(metadata) = std::fs::metadata(path) {
+                    println!("File size: {} bytes ({:.1} MB)", metadata.len(), metadata.len() as f64 / 1_000_000.0);
+                }
+                
+                // Open it
+                let _ = Command::new("open").arg(path).spawn();
+                println!("\nOpened screenshot - please verify it looks correct!");
+            } else {
+                println!("\n❌ Screenshot failed!");
+            }
+        }
+        Err(e) => {
+            println!("❌ Failed to execute screencapture: {}", e);
+        }
+    }
+}

crates/g3-computer-control/examples/test_screenshot_fix.rs

@@ -0,0 +1,68 @@
+use core_graphics::display::CGDisplay;
+use image::{ImageBuffer, RgbaImage};
+
+fn main() {
+    let display = CGDisplay::main();
+    let image = display.image().expect("Failed to capture screen");
+    
+    let width = image.width() as u32;
+    let height = image.height() as u32;
+    let bytes_per_row = image.bytes_per_row() as usize;
+    let data = image.data();
+    
+    println!("Testing screenshot fix...");
+    println!("Image: {}x{}, bytes_per_row: {}", width, height, bytes_per_row);
+    println!("Expected bytes per row: {}", width * 4);
+    println!("Padding per row: {} bytes", bytes_per_row - (width as usize * 4));
+    
+    // OLD METHOD (broken) - treating data as continuous
+    println!("\n=== OLD METHOD (BROKEN) ===");
+    let mut old_rgba = Vec::with_capacity(data.len() as usize);
+    for chunk in data.chunks_exact(4) {
+        old_rgba.push(chunk[2]); // R
+        old_rgba.push(chunk[1]); // G
+        old_rgba.push(chunk[0]); // B
+        old_rgba.push(chunk[3]); // A
+    }
+    println!("Converted {} pixels", old_rgba.len() / 4);
+    println!("Expected {} pixels", width * height);
+    
+    // NEW METHOD (fixed) - handling row padding
+    println!("\n=== NEW METHOD (FIXED) ===");
+    let mut new_rgba = Vec::with_capacity((width * height * 4) as usize);
+    for row in 0..height as usize {
+        let row_start = row * bytes_per_row;
+        let row_end = row_start + (width as usize * 4);
+        
+        for chunk in data[row_start..row_end].chunks_exact(4) {
+            new_rgba.push(chunk[2]); // R
+            new_rgba.push(chunk[1]); // G
+            new_rgba.push(chunk[0]); // B
+            new_rgba.push(chunk[3]); // A
+        }
+    }
+    println!("Converted {} pixels", new_rgba.len() / 4);
+    println!("Expected {} pixels", width * height);
+    
+    // Save a small crop from both methods
+    let crop_size = 200;
+    
+    // Old method crop
+    let old_crop: Vec<u8> = old_rgba.iter().take((crop_size * crop_size * 4) as usize).copied().collect();
+    if let Some(old_img) = ImageBuffer::from_raw(crop_size, crop_size, old_crop) {
+        let old_img: RgbaImage = old_img;
+        old_img.save("/tmp/screenshot_old_method.png").unwrap();
+        println!("\nSaved OLD method crop to: /tmp/screenshot_old_method.png");
+    }
+    
+    // New method crop
+    let new_crop: Vec<u8> = new_rgba.iter().take((crop_size * crop_size * 4) as usize).copied().collect();
+    if let Some(new_img) = ImageBuffer::from_raw(crop_size, crop_size, new_crop) {
+        let new_img: RgbaImage = new_img;
+        new_img.save("/tmp/screenshot_new_method.png").unwrap();
+        println!("Saved NEW method crop to: /tmp/screenshot_new_method.png");
+    }
+    
+    println!("\nOpen both images to compare:");
+    println!("  open /tmp/screenshot_old_method.png /tmp/screenshot_new_method.png");
+}

crates/g3-computer-control/examples/test_type_text.rs

@@ -0,0 +1,48 @@
+//! Test the new type_text functionality
+
+use anyhow::Result;
+use g3_computer_control::MacAxController;
+
+#[tokio::main]
+async fn main() -> Result<()> {
+    println!("🧪 Testing macax type_text functionality\n");
+    
+    let controller = MacAxController::new()?;
+    println!("✅ Controller initialized\n");
+    
+    // Test 1: Type simple text
+    println!("Test 1: Typing simple text into TextEdit");
+    println!("  Please open TextEdit and create a new document...");
+    std::thread::sleep(std::time::Duration::from_secs(3));
+    
+    match controller.type_text("TextEdit", "Hello, World!") {
+        Ok(_) => println!("  ✅ Successfully typed simple text\n"),
+        Err(e) => println!("  ❌ Failed: {}\n", e),
+    }
+    
+    std::thread::sleep(std::time::Duration::from_secs(1));
+    
+    // Test 2: Type unicode and emojis
+    println!("Test 2: Typing unicode and emojis");
+    match controller.type_text("TextEdit", "\n🌟 Unicode test: café, naïve, 日本語 🎉") {
+        Ok(_) => println!("  ✅ Successfully typed unicode text\n"),
+        Err(e) => println!("  ❌ Failed: {}\n", e),
+    }
+    
+    std::thread::sleep(std::time::Duration::from_secs(1));
+    
+    // Test 3: Type special characters
+    println!("Test 3: Typing special characters");
+    match controller.type_text("TextEdit", "\nSpecial: @#$%^&*()_+-=[]{}|;':,.<>?/") {
+        Ok(_) => println!("  ✅ Successfully typed special characters\n"),
+        Err(e) => println!("  ❌ Failed: {}\n", e),
+    }
+    
+    println!("\n✨ Tests complete!");
+    println!("\n💡 Now try with Things3:");
+    println!("   1. Open Things3");
+    println!("   2. Press Cmd+N to create a new task");
+    println!("   3. Run: g3 --macax 'type \"🌟 My awesome task\" into Things'");
+    
+    Ok(())
+}

crates/g3-computer-control/examples/test_vision.rs

@@ -0,0 +1,85 @@
+use g3_computer_control::ocr::{OCREngine, DefaultOCR};
+use anyhow::Result;
+
+#[tokio::main]
+async fn main() -> Result<()> {
+    println!("🧪 Testing Apple Vision OCR");
+    println!("===========================\n");
+    
+    // Initialize OCR engine
+    println!("📦 Initializing OCR engine...");
+    let ocr = DefaultOCR::new()?;
+    println!("✅ OCR engine: {}\n", ocr.name());
+    
+    // Check if test image exists
+    let test_image = "/tmp/safari_test.png";
+    if !std::path::Path::new(test_image).exists() {
+        println!("⚠️  Test image not found: {}", test_image);
+        println!("   Creating a screenshot...");
+        
+        let status = std::process::Command::new("screencapture")
+            .arg("-x")
+            .arg("-R")
+            .arg("0,0,1200,800")
+            .arg(test_image)
+            .status()?;
+        
+        if !status.success() {
+            anyhow::bail!("Failed to create screenshot");
+        }
+        
+        println!("✅ Screenshot created\n");
+    }
+    
+    // Run OCR
+    println!("🔍 Running Apple Vision OCR on {}...", test_image);
+    let start = std::time::Instant::now();
+    let locations = ocr.extract_text_with_locations(test_image).await?;
+    let duration = start.elapsed();
+    
+    println!("✅ OCR completed in {:.3}s\n", duration.as_secs_f64());
+    
+    // Display results
+    println!("📊 Results:");
+    println!("   Found {} text elements\n", locations.len());
+    
+    if locations.is_empty() {
+        println!("⚠️  No text found in image");
+    } else {
+        println!("   Top 20 results:");
+        println!("   {:<4} {:<40} {:<15} {:<12} {:<8}", "#", "Text", "Position", "Size", "Conf");
+        println!("   {}", "-".repeat(85));
+        
+        for (i, loc) in locations.iter().take(20).enumerate() {
+            let text = if loc.text.len() > 37 {
+                format!("{}...", &loc.text[..37])
+            } else {
+                loc.text.clone()
+            };
+            
+            println!("   {:<4} {:<40} ({:>4},{:>4})    {:>4}x{:<4}  {:.2}",
+                i + 1,
+                text,
+                loc.x,
+                loc.y,
+                loc.width,
+                loc.height,
+                loc.confidence
+            );
+        }
+        
+        if locations.len() > 20 {
+            println!("\n   ... and {} more", locations.len() - 20);
+        }
+        
+        // Performance comparison
+        println!("\n📈 Performance:");
+        println!("   OCR Speed: {:.3}s", duration.as_secs_f64());
+        println!("   Text elements: {}", locations.len());
+        println!("   Avg per element: {:.1}ms", duration.as_millis() as f64 / locations.len() as f64);
+    }
+    
+    println!("\n✅ Test complete!");
+    
+    Ok(())
+}

crates/g3-computer-control/examples/test_window_capture.rs

@@ -0,0 +1,45 @@
+use g3_computer_control::create_controller;
+
+#[tokio::main]
+async fn main() {
+    println!("Testing window-specific screenshot capture...");
+    
+    let controller = create_controller().expect("Failed to create controller");
+    
+    // Test 1: Capture iTerm2 window
+    println!("\n1. Capturing iTerm2 window...");
+    match controller.take_screenshot("/tmp/iterm_window.png", None, Some("iTerm2")).await {
+        Ok(_) => {
+            println!("   ✅ iTerm2 window captured to /tmp/iterm_window.png");
+            let _ = std::process::Command::new("open").arg("/tmp/iterm_window.png").spawn();
+        }
+        Err(e) => println!("   ❌ Failed: {}", e),
+    }
+    
+    // Wait a moment for the image to open
+    tokio::time::sleep(tokio::time::Duration::from_secs(2)).await;
+    
+    // Test 2: Full screen capture for comparison
+    println!("\n2. Capturing full screen for comparison...");
+    match controller.take_screenshot("/tmp/fullscreen.png", None, None).await {
+        Ok(_) => {
+            println!("   ✅ Full screen captured to /tmp/fullscreen.png");
+            let _ = std::process::Command::new("open").arg("/tmp/fullscreen.png").spawn();
+        }
+        Err(e) => println!("   ❌ Failed: {}", e),
+    }
+    
+    println!("\n=== Comparison ===");
+    println!("iTerm window:  /tmp/iterm_window.png (should show ONLY iTerm window)");
+    println!("Full screen:   /tmp/fullscreen.png (should show entire desktop)");
+    
+    // Show file sizes
+    if let Ok(meta1) = std::fs::metadata("/tmp/iterm_window.png") {
+        if let Ok(meta2) = std::fs::metadata("/tmp/fullscreen.png") {
+            println!("\nFile sizes:");
+            println!("  iTerm window: {:.1} MB", meta1.len() as f64 / 1_000_000.0);
+            println!("  Full screen:  {:.1} MB", meta2.len() as f64 / 1_000_000.0);
+            println!("\nWindow capture should be smaller than full screen.");
+        }
+    }
+}

crates/g3-computer-control/src/lib.rs

@@ -0,0 +1,49 @@
+// Suppress warnings from objc crate macros
+#![allow(unexpected_cfgs)]
+
+pub mod types;
+pub mod platform;
+pub mod ocr;
+pub mod webdriver;
+pub mod macax;
+
+// Re-export webdriver types for convenience
+pub use webdriver::{WebDriverController, WebElement, safari::SafariDriver};
+
+// Re-export macax types for convenience
+pub use macax::{MacAxController, AXElement, AXApplication};
+
+use anyhow::Result;
+use async_trait::async_trait;
+use types::*;
+
+#[async_trait]
+pub trait ComputerController: Send + Sync {
+    // Screen capture
+    async fn take_screenshot(&self, path: &str, region: Option<Rect>, window_id: Option<&str>) -> Result<()>;
+    
+    // OCR operations
+    async fn extract_text_from_screen(&self, region: Rect, window_id: &str) -> Result<String>;
+    async fn extract_text_from_image(&self, path: &str) -> Result<String>;
+    async fn extract_text_with_locations(&self, path: &str) -> Result<Vec<TextLocation>>;
+    async fn find_text_in_app(&self, app_name: &str, search_text: &str) -> Result<Option<TextLocation>>;
+    
+    // Mouse operations
+    fn move_mouse(&self, x: i32, y: i32) -> Result<()>;
+    fn click_at(&self, x: i32, y: i32, app_name: Option<&str>) -> Result<()>;
+}
+
+// Platform-specific constructor
+pub fn create_controller() -> Result<Box<dyn ComputerController>> {
+    #[cfg(target_os = "macos")]
+    return Ok(Box::new(platform::macos::MacOSController::new()?));
+    
+    #[cfg(target_os = "linux")]
+    return Ok(Box::new(platform::linux::LinuxController::new()?));
+    
+    #[cfg(target_os = "windows")]
+    return Ok(Box::new(platform::windows::WindowsController::new()?));
+    
+    #[cfg(not(any(target_os = "macos", target_os = "linux", target_os = "windows")))]
+    anyhow::bail!("Unsupported platform")
+}

crates/g3-computer-control/src/macax/controller.rs

@@ -0,0 +1,822 @@
+use super::{AXApplication, AXElement};
+use anyhow::{Context, Result};
+use std::collections::HashMap;
+
+#[cfg(target_os = "macos")]
+use accessibility::{AXUIElement, AXUIElementAttributes, ElementFinder, TreeVisitor, TreeWalker, TreeWalkerFlow};
+
+#[cfg(target_os = "macos")]
+use core_foundation::base::TCFType;
+
+#[cfg(target_os = "macos")]
+use core_foundation::string::CFString;
+
+/// macOS Accessibility API controller using native APIs
+pub struct MacAxController {
+    // Cache for application elements
+    app_cache: std::sync::Mutex<HashMap<String, AXUIElement>>,
+}
+
+impl MacAxController {
+    pub fn new() -> Result<Self> {
+        #[cfg(target_os = "macos")]
+        {
+            // Check if we have accessibility permissions by trying to get system-wide element
+            let _system = AXUIElement::system_wide();
+            
+            Ok(Self {
+                app_cache: std::sync::Mutex::new(HashMap::new()),
+            })
+        }
+        
+        #[cfg(not(target_os = "macos"))]
+        {
+            anyhow::bail!("macOS Accessibility API is only available on macOS")
+        }
+    }
+    
+    /// List all running applications
+    #[cfg(target_os = "macos")]
+    pub fn list_applications(&self) -> Result<Vec<AXApplication>> {
+        let apps = Self::get_running_applications()?;
+        Ok(apps)
+    }
+    
+    #[cfg(not(target_os = "macos"))]
+    pub fn list_applications(&self) -> Result<Vec<AXApplication>> {
+        anyhow::bail!("Not supported on this platform")
+    }
+    
+    #[cfg(target_os = "macos")]
+    fn get_running_applications() -> Result<Vec<AXApplication>> {
+        use cocoa::appkit::NSApplicationActivationPolicy;
+        use cocoa::base::{id, nil};
+        use objc::{class, msg_send, sel, sel_impl};
+        
+        unsafe {
+            let workspace: id = msg_send![class!(NSWorkspace), sharedWorkspace];
+            let running_apps: id = msg_send![workspace, runningApplications];
+            let count: usize = msg_send![running_apps, count];
+            
+            let mut apps = Vec::new();
+            
+            for i in 0..count {
+                let app: id = msg_send![running_apps, objectAtIndex: i];
+                
+                // Get app name
+                let localized_name: id = msg_send![app, localizedName];
+                if localized_name == nil {
+                    continue;
+                }
+                let name_ptr: *const i8 = msg_send![localized_name, UTF8String];
+                let name = if !name_ptr.is_null() {
+                    std::ffi::CStr::from_ptr(name_ptr)
+                        .to_string_lossy()
+                        .to_string()
+                } else {
+                    continue;
+                };
+                
+                // Get bundle ID
+                let bundle_id_obj: id = msg_send![app, bundleIdentifier];
+                let bundle_id = if bundle_id_obj != nil {
+                    let bundle_id_ptr: *const i8 = msg_send![bundle_id_obj, UTF8String];
+                    if !bundle_id_ptr.is_null() {
+                        Some(
+                            std::ffi::CStr::from_ptr(bundle_id_ptr)
+                                .to_string_lossy()
+                                .to_string(),
+                        )
+                    } else {
+                        None
+                    }
+                } else {
+                    None
+                };
+                
+                // Get PID
+                let pid: i32 = msg_send![app, processIdentifier];
+                
+                // Skip background-only apps
+                let activation_policy: i64 = msg_send![app, activationPolicy];
+                if activation_policy == NSApplicationActivationPolicy::NSApplicationActivationPolicyRegular as i64 {
+                    apps.push(AXApplication {
+                        name,
+                        bundle_id,
+                        pid,
+                    });
+                }
+            }
+            
+            Ok(apps)
+        }
+    }
+    
+    /// Get the frontmost (active) application
+    #[cfg(target_os = "macos")]
+    pub fn get_frontmost_app(&self) -> Result<AXApplication> {
+        use cocoa::base::{id, nil};
+        use objc::{class, msg_send, sel, sel_impl};
+        
+        unsafe {
+            let workspace: id = msg_send![class!(NSWorkspace), sharedWorkspace];
+            let frontmost_app: id = msg_send![workspace, frontmostApplication];
+            
+            if frontmost_app == nil {
+                anyhow::bail!("No frontmost application");
+            }
+            
+            // Get app name
+            let localized_name: id = msg_send![frontmost_app, localizedName];
+            let name_ptr: *const i8 = msg_send![localized_name, UTF8String];
+            let name = std::ffi::CStr::from_ptr(name_ptr)
+                .to_string_lossy()
+                .to_string();
+            
+            // Get bundle ID
+            let bundle_id_obj: id = msg_send![frontmost_app, bundleIdentifier];
+            let bundle_id = if bundle_id_obj != nil {
+                let bundle_id_ptr: *const i8 = msg_send![bundle_id_obj, UTF8String];
+                if !bundle_id_ptr.is_null() {
+                    Some(
+                        std::ffi::CStr::from_ptr(bundle_id_ptr)
+                            .to_string_lossy()
+                            .to_string(),
+                    )
+                } else {
+                    None
+                }
+            } else {
+                None
+            };
+            
+            // Get PID
+            let pid: i32 = msg_send![frontmost_app, processIdentifier];
+            
+            Ok(AXApplication {
+                name,
+                bundle_id,
+                pid,
+            })
+        }
+    }
+    
+    #[cfg(not(target_os = "macos"))]
+    pub fn get_frontmost_app(&self) -> Result<AXApplication> {
+        anyhow::bail!("Not supported on this platform")
+    }
+    
+    /// Get AXUIElement for an application by name or PID
+    #[cfg(target_os = "macos")]
+    fn get_app_element(&self, app_name: &str) -> Result<AXUIElement> {
+        // Check cache first
+        {
+            let cache = self.app_cache.lock().unwrap();
+            if let Some(element) = cache.get(app_name) {
+                return Ok(element.clone());
+            }
+        }
+        
+        // Find the app by name
+        let apps = Self::get_running_applications()?;
+        let app = apps
+            .iter()
+            .find(|a| a.name == app_name)
+            .ok_or_else(|| anyhow::anyhow!("Application '{}' not found", app_name))?;
+        
+        // Create AXUIElement for the app
+        let element = AXUIElement::application(app.pid);
+        
+        // Cache it
+        {
+            let mut cache = self.app_cache.lock().unwrap();
+            cache.insert(app_name.to_string(), element.clone());
+        }
+        
+        Ok(element)
+    }
+    
+    /// Activate (bring to front) an application
+    #[cfg(target_os = "macos")]
+    pub fn activate_app(&self, app_name: &str) -> Result<()> {
+        use cocoa::base::id;
+        use objc::{class, msg_send, sel, sel_impl};
+        
+        // Find the app
+        let apps = Self::get_running_applications()?;
+        let app = apps
+            .iter()
+            .find(|a| a.name == app_name)
+            .ok_or_else(|| anyhow::anyhow!("Application '{}' not found", app_name))?;
+        
+        unsafe {
+            let workspace: id = msg_send![class!(NSWorkspace), sharedWorkspace];
+            let running_apps: id = msg_send![workspace, runningApplications];
+            let count: usize = msg_send![running_apps, count];
+            
+            for i in 0..count {
+                let running_app: id = msg_send![running_apps, objectAtIndex: i];
+                let pid: i32 = msg_send![running_app, processIdentifier];
+                
+                if pid == app.pid {
+                    let _: bool = msg_send![running_app, activateWithOptions: 0];
+                    return Ok(());
+                }
+            }
+        }
+        
+        anyhow::bail!("Failed to activate application")
+    }
+    
+    #[cfg(not(target_os = "macos"))]
+    pub fn activate_app(&self, _app_name: &str) -> Result<()> {
+        anyhow::bail!("Not supported on this platform")
+    }
+    
+    /// Get the UI hierarchy of an application
+    #[cfg(target_os = "macos")]
+    pub fn get_ui_tree(&self, app_name: &str, max_depth: usize) -> Result<String> {
+        let app_element = self.get_app_element(app_name)?;
+        let mut output = format!("Application: {}\n", app_name);
+        
+        Self::build_ui_tree(&app_element, &mut output, 0, max_depth)?;
+        
+        Ok(output)
+    }
+    
+    #[cfg(not(target_os = "macos"))]
+    pub fn get_ui_tree(&self, _app_name: &str, _max_depth: usize) -> Result<String> {
+        anyhow::bail!("Not supported on this platform")
+    }
+    
+    #[cfg(target_os = "macos")]
+    fn build_ui_tree(
+        element: &AXUIElement,
+        output: &mut String,
+        depth: usize,
+        max_depth: usize,
+    ) -> Result<()> {
+        if depth >= max_depth {
+            return Ok(());
+        }
+        
+        let indent = "  ".repeat(depth);
+        
+        // Get role
+        let role = element.role().ok().map(|s| s.to_string())
+            .unwrap_or_else(|| "Unknown".to_string());
+        
+        // Get title
+        let title = element.title().ok()
+            .map(|s| s.to_string());
+        
+        // Get identifier
+        let identifier = element.identifier().ok()
+            .map(|s| s.to_string());
+        
+        // Format output
+        output.push_str(&format!("{}Role: {}", indent, role));
+        if let Some(t) = title {
+            output.push_str(&format!(", Title: {}", t));
+        }
+        if let Some(id) = identifier {
+            output.push_str(&format!(", ID: {}", id));
+        }
+        output.push('\n');
+        
+        // Get children
+        if let Ok(children) = element.children() {
+            for i in 0..children.len() {
+                if let Some(child) = children.get(i) {
+                    let _ = Self::build_ui_tree(&child, output, depth + 1, max_depth);
+                }
+            }
+        }
+        
+        Ok(())
+    }
+    
+    /// Find UI elements in an application
+    #[cfg(target_os = "macos")]
+    pub fn find_elements(
+        &self,
+        app_name: &str,
+        role: Option<&str>,
+        title: Option<&str>,
+        identifier: Option<&str>,
+    ) -> Result<Vec<AXElement>> {
+        let app_element = self.get_app_element(app_name)?;
+        let mut found_elements = Vec::new();
+        
+        let visitor = ElementCollector {
+            role_filter: role.map(|s| s.to_string()),
+            title_filter: title.map(|s| s.to_string()),
+            identifier_filter: identifier.map(|s| s.to_string()),
+            results: std::cell::RefCell::new(&mut found_elements),
+            depth: std::cell::Cell::new(0),
+        };
+        
+        let walker = TreeWalker::new();
+        walker.walk(&app_element, &visitor);
+        
+        Ok(found_elements)
+    }
+    
+    #[cfg(not(target_os = "macos"))]
+    pub fn find_elements(
+        &self,
+        _app_name: &str,
+        _role: Option<&str>,
+        _title: Option<&str>,
+        _identifier: Option<&str>,
+    ) -> Result<Vec<AXElement>> {
+        anyhow::bail!("Not supported on this platform")
+    }
+    
+    /// Find a single element (helper for click, set_value, etc.)
+    #[cfg(target_os = "macos")]
+    fn find_element(
+        &self,
+        app_name: &str,
+        role: &str,
+        title: Option<&str>,
+        identifier: Option<&str>,
+    ) -> Result<AXUIElement> {
+        let app_element = self.get_app_element(app_name)?;
+        
+        let role_str = role.to_string();
+        let title_str = title.map(|s| s.to_string());
+        let identifier_str = identifier.map(|s| s.to_string());
+        
+        let finder = ElementFinder::new(
+            &app_element,
+            move |element| {
+                // Check role
+                let elem_role = element.role()
+                    .ok()
+                    .map(|s| s.to_string());
+                
+                if let Some(r) = elem_role {
+                    if !r.contains(&role_str) {
+                        return false;
+                    }
+                } else {
+                    return false;
+                }
+                
+                // Check title if specified
+                if let Some(ref title_filter) = title_str {
+                    let elem_title = element.title()
+                        .ok()
+                        .map(|s| s.to_string());
+                    
+                    if let Some(t) = elem_title {
+                        if !t.contains(title_filter) {
+                            return false;
+                        }
+                    } else {
+                        return false;
+                    }
+                }
+                
+                // Check identifier if specified
+                if let Some(ref id_filter) = identifier_str {
+                    let elem_id = element.identifier()
+                        .ok()
+                        .map(|s| s.to_string());
+                    
+                    if let Some(id) = elem_id {
+                        if !id.contains(id_filter) {
+                            return false;
+                        }
+                    } else {
+                        return false;
+                    }
+                }
+                
+                true
+            },
+            Some(std::time::Duration::from_secs(2)),
+        );
+        
+        finder.find().context("Element not found")
+    }
+    
+    /// Click on a UI element
+    #[cfg(target_os = "macos")]
+    pub fn click_element(
+        &self,
+        app_name: &str,
+        role: &str,
+        title: Option<&str>,
+        identifier: Option<&str>,
+    ) -> Result<()> {
+        let element = self.find_element(app_name, role, title, identifier)?;
+        
+        // Perform the press action
+        let action_name = CFString::new("AXPress");
+        element
+            .perform_action(&action_name)
+            .map_err(|e| anyhow::anyhow!("Failed to perform press action: {:?}", e))?;
+        
+        Ok(())
+    }
+    
+    #[cfg(not(target_os = "macos"))]
+    pub fn click_element(
+        &self,
+        _app_name: &str,
+        _role: &str,
+        _title: Option<&str>,
+        _identifier: Option<&str>,
+    ) -> Result<()> {
+        anyhow::bail!("Not supported on this platform")
+    }
+    
+    /// Set the value of a UI element
+    #[cfg(target_os = "macos")]
+    pub fn set_value(
+        &self,
+        app_name: &str,
+        role: &str,
+        value: &str,
+        title: Option<&str>,
+        identifier: Option<&str>,
+    ) -> Result<()> {
+        let element = self.find_element(app_name, role, title, identifier)?;
+        
+        // Set the value - convert CFString to CFType
+        let cf_value = CFString::new(value);
+        
+        element.set_value(cf_value.as_CFType())
+            .map_err(|e| anyhow::anyhow!("Failed to set value: {:?}", e))?;
+        
+        Ok(())
+    }
+    
+    #[cfg(not(target_os = "macos"))]
+    pub fn set_value(
+        &self,
+        _app_name: &str,
+        _role: &str,
+        _value: &str,
+        _title: Option<&str>,
+        _identifier: Option<&str>,
+    ) -> Result<()> {
+        anyhow::bail!("Not supported on this platform")
+    }
+    
+    /// Get the value of a UI element
+    #[cfg(target_os = "macos")]
+    pub fn get_value(
+        &self,
+        app_name: &str,
+        role: &str,
+        title: Option<&str>,
+        identifier: Option<&str>,
+    ) -> Result<String> {
+        let element = self.find_element(app_name, role, title, identifier)?;
+        
+        // Get the value
+        let value_type = element.value()
+            .map_err(|e| anyhow::anyhow!("Failed to get value: {:?}", e))?;
+        
+        // Try to downcast to CFString
+        if let Some(cf_string) = value_type.downcast::<CFString>() {
+            Ok(cf_string.to_string())
+        } else {
+            // For non-string values, try to get a description
+            Ok(format!("<non-string value>"))
+        }
+    }
+    
+    #[cfg(not(target_os = "macos"))]
+    pub fn get_value(
+        &self,
+        _app_name: &str,
+        _role: &str,
+        _title: Option<&str>,
+        _identifier: Option<&str>,
+    ) -> Result<String> {
+        anyhow::bail!("Not supported on this platform")
+    }
+    
+    /// Type text into the currently focused element (uses system text input)
+    #[cfg(target_os = "macos")]
+    pub fn type_text(&self, app_name: &str, text: &str) -> Result<()> {
+        use cocoa::base::{id, nil};
+        use cocoa::foundation::NSString;
+        use objc::{class, msg_send, sel, sel_impl};
+        
+        // First, make sure the app is active
+        self.activate_app(app_name)?;
+        
+        // Wait for app to fully activate
+        std::thread::sleep(std::time::Duration::from_millis(500));
+        
+        // Send a Tab key to try to focus on a text field
+        // This helps ensure something is focused before we paste
+        let _ = self.press_key(app_name, "tab", vec![]);
+        std::thread::sleep(std::time::Duration::from_millis(800));
+        
+        // Save old clipboard, set new content, paste, then restore
+        let old_content: id;
+        unsafe {
+            // Get the general pasteboard
+            let pasteboard: id = msg_send![class!(NSPasteboard), generalPasteboard];
+            
+            // Save current clipboard content
+            let ns_string_type = NSString::alloc(nil).init_str("public.utf8-plain-text");
+            old_content = msg_send![pasteboard, stringForType: ns_string_type];
+            
+            // Clear and set new content
+            let _: () = msg_send![pasteboard, clearContents];
+            
+            let ns_string = NSString::alloc(nil).init_str(text);
+            let ns_type = NSString::alloc(nil).init_str("public.utf8-plain-text");
+            let _: bool = msg_send![pasteboard, setString:ns_string forType:ns_type];
+        }
+        
+        // Wait a moment for clipboard to update
+        std::thread::sleep(std::time::Duration::from_millis(200));
+        
+        // Paste using Cmd+V (outside unsafe block)
+        self.press_key(app_name, "v", vec!["command"])?;
+        
+        // Wait for paste to complete
+        std::thread::sleep(std::time::Duration::from_millis(300));
+        
+        // Restore old clipboard content if it existed
+        unsafe {
+            if old_content != nil {
+                let pasteboard: id = msg_send![class!(NSPasteboard), generalPasteboard];
+                let _: () = msg_send![pasteboard, clearContents];
+                let ns_type = NSString::alloc(nil).init_str("public.utf8-plain-text");
+                let _: bool = msg_send![pasteboard, setString:old_content forType:ns_type];
+            }
+        }
+        
+        Ok(())
+    }
+    
+    #[cfg(not(target_os = "macos"))]
+    pub fn type_text(&self, _app_name: &str, _text: &str) -> Result<()> {
+        anyhow::bail!("Not supported on this platform")
+    }
+    
+    /// Focus on a text field or text area element
+    #[cfg(target_os = "macos")]
+    pub fn focus_element(
+        &self,
+        app_name: &str,
+        role: &str,
+        title: Option<&str>,
+        identifier: Option<&str>,
+    ) -> Result<()> {
+        let element = self.find_element(app_name, role, title, identifier)?;
+        
+        // Set focused attribute to true
+        use core_foundation::boolean::CFBoolean;
+        let cf_true = CFBoolean::true_value();
+        
+        element.set_attribute(&accessibility::AXAttribute::focused(), cf_true)
+            .map_err(|e| anyhow::anyhow!("Failed to focus element: {:?}", e))?;
+        
+        Ok(())
+    }
+    
+    /// Press a keyboard shortcut
+    #[cfg(target_os = "macos")]
+    pub fn press_key(
+        &self,
+        app_name: &str,
+        key: &str,
+        modifiers: Vec<&str>,
+    ) -> Result<()> {
+        use core_graphics::event::{
+            CGEvent, CGEventFlags, CGEventTapLocation,
+        };
+        use core_graphics::event_source::{CGEventSource, CGEventSourceStateID};
+        
+        // First, make sure the app is active
+        self.activate_app(app_name)?;
+        
+        // Wait a bit for activation
+        std::thread::sleep(std::time::Duration::from_millis(100));
+        
+        // Map key string to key code
+        let key_code = Self::key_to_keycode(key)
+            .ok_or_else(|| anyhow::anyhow!("Unknown key: {}", key))?;
+        
+        // Map modifiers to flags
+        let mut flags = CGEventFlags::CGEventFlagNull;
+        for modifier in modifiers {
+            match modifier.to_lowercase().as_str() {
+                "command" | "cmd" => flags |= CGEventFlags::CGEventFlagCommand,
+                "option" | "alt" => flags |= CGEventFlags::CGEventFlagAlternate,
+                "control" | "ctrl" => flags |= CGEventFlags::CGEventFlagControl,
+                "shift" => flags |= CGEventFlags::CGEventFlagShift,
+                _ => {}
+            }
+        }
+        
+        // Create event source
+        let source = CGEventSource::new(CGEventSourceStateID::HIDSystemState)
+            .ok().context("Failed to create event source")?;
+        
+        // Create key down event
+        let key_down = CGEvent::new_keyboard_event(source.clone(), key_code, true)
+            .ok().context("Failed to create key down event")?;
+        key_down.set_flags(flags);
+        
+        // Create key up event
+        let key_up = CGEvent::new_keyboard_event(source, key_code, false)
+            .ok().context("Failed to create key up event")?;
+        key_up.set_flags(flags);
+        
+        // Post events
+        key_down.post(CGEventTapLocation::HID);
+        std::thread::sleep(std::time::Duration::from_millis(50));
+        key_up.post(CGEventTapLocation::HID);
+        
+        Ok(())
+    }
+    
+    #[cfg(not(target_os = "macos"))]
+    pub fn press_key(
+        &self,
+        _app_name: &str,
+        _key: &str,
+        _modifiers: Vec<&str>,
+    ) -> Result<()> {
+        anyhow::bail!("Not supported on this platform")
+    }
+    
+    #[cfg(target_os = "macos")]
+    fn key_to_keycode(key: &str) -> Option<u16> {
+        // Map common keys to keycodes
+        // See: https://eastmanreference.com/complete-list-of-applescript-key-codes
+        match key.to_lowercase().as_str() {
+            "a" => Some(0x00),
+            "s" => Some(0x01),
+            "d" => Some(0x02),
+            "f" => Some(0x03),
+            "h" => Some(0x04),
+            "g" => Some(0x05),
+            "z" => Some(0x06),
+            "x" => Some(0x07),
+            "c" => Some(0x08),
+            "v" => Some(0x09),
+            "b" => Some(0x0B),
+            "q" => Some(0x0C),
+            "w" => Some(0x0D),
+            "e" => Some(0x0E),
+            "r" => Some(0x0F),
+            "y" => Some(0x10),
+            "t" => Some(0x11),
+            "1" => Some(0x12),
+            "2" => Some(0x13),
+            "3" => Some(0x14),
+            "4" => Some(0x15),
+            "6" => Some(0x16),
+            "5" => Some(0x17),
+            "=" => Some(0x18),
+            "9" => Some(0x19),
+            "7" => Some(0x1A),
+            "-" => Some(0x1B),
+            "8" => Some(0x1C),
+            "0" => Some(0x1D),
+            "]" => Some(0x1E),
+            "o" => Some(0x1F),
+            "u" => Some(0x20),
+            "[" => Some(0x21),
+            "i" => Some(0x22),
+            "p" => Some(0x23),
+            "return" | "enter" => Some(0x24),
+            "l" => Some(0x25),
+            "j" => Some(0x26),
+            "'" => Some(0x27),
+            "k" => Some(0x28),
+            ";" => Some(0x29),
+            "\\" => Some(0x2A),
+            "," => Some(0x2B),
+            "/" => Some(0x2C),
+            "n" => Some(0x2D),
+            "m" => Some(0x2E),
+            "." => Some(0x2F),
+            "tab" => Some(0x30),
+            "space" => Some(0x31),
+            "`" => Some(0x32),
+            "delete" | "backspace" => Some(0x33),
+            "escape" | "esc" => Some(0x35),
+            "f1" => Some(0x7A),
+            "f2" => Some(0x78),
+            "f3" => Some(0x63),
+            "f4" => Some(0x76),
+            "f5" => Some(0x60),
+            "f6" => Some(0x61),
+            "f7" => Some(0x62),
+            "f8" => Some(0x64),
+            "f9" => Some(0x65),
+            "f10" => Some(0x6D),
+            "f11" => Some(0x67),
+            "f12" => Some(0x6F),
+            "left" => Some(0x7B),
+            "right" => Some(0x7C),
+            "down" => Some(0x7D),
+            "up" => Some(0x7E),
+            _ => None,
+        }
+    }
+}
+
+#[cfg(target_os = "macos")]
+struct ElementCollector<'a> {
+    role_filter: Option<String>,
+    title_filter: Option<String>,
+    identifier_filter: Option<String>,
+    results: std::cell::RefCell<&'a mut Vec<AXElement>>,
+    depth: std::cell::Cell<usize>,
+}
+
+#[cfg(target_os = "macos")]
+impl<'a> TreeVisitor for ElementCollector<'a> {
+    fn enter_element(&self, element: &AXUIElement) -> TreeWalkerFlow {
+        self.depth.set(self.depth.get() + 1);
+        
+        if self.depth.get() > 20 {
+            return TreeWalkerFlow::SkipSubtree;
+        }
+        
+        // Get element properties
+        let role = element.role()
+            .ok()
+            .map(|s| s.to_string())
+            .unwrap_or_else(|| "Unknown".to_string());
+        
+        let title = element.title()
+            .ok()
+            .map(|s| s.to_string());
+        
+        let identifier = element.identifier()
+            .ok()
+            .map(|s| s.to_string());
+        
+        // Check if this element matches the filters
+        let role_matches = self.role_filter.as_ref().map_or(true, |r| role.contains(r));
+        let title_matches = self.title_filter.as_ref().map_or(true, |t| {
+            title.as_ref().map_or(false, |title_str| title_str.contains(t))
+        });
+        let identifier_matches = self.identifier_filter.as_ref().map_or(true, |id| {
+            identifier.as_ref().map_or(false, |id_str| id_str.contains(id))
+        });
+        
+        if role_matches && title_matches && identifier_matches {
+            // Get additional properties
+            let value = element.value()
+                .ok()
+                .and_then(|v| {
+                    v.downcast::<CFString>().map(|s| s.to_string())
+                });
+            
+            let label = element.description()
+                .ok()
+                .map(|s| s.to_string());
+            
+            let enabled = element.enabled()
+                .ok()
+                .map(|b| b.into())
+                .unwrap_or(false);
+            
+            let focused = element.focused()
+                .ok()
+                .map(|b| b.into())
+                .unwrap_or(false);
+            
+            // Count children
+            let children_count = element.children()
+                .ok()
+                .map(|arr| arr.len() as usize)
+                .unwrap_or(0);
+            
+            self.results.borrow_mut().push(AXElement {
+                role,
+                title,
+                value,
+                label,
+                identifier,
+                enabled,
+                focused,
+                position: None,
+                size: None,
+                children_count,
+            });
+        }
+        
+        TreeWalkerFlow::Continue
+    }
+    
+    fn exit_element(&self, _element: &AXUIElement) {
+        self.depth.set(self.depth.get() - 1);
+    }
+}

crates/g3-computer-control/src/macax/mod.rs

@@ -0,0 +1,65 @@
+pub mod controller;
+
+pub use controller::MacAxController;
+
+use serde::{Deserialize, Serialize};
+
+#[cfg(test)]
+mod tests;
+
+/// Represents an accessibility element in the UI hierarchy
+#[derive(Debug, Clone, Serialize, Deserialize)]
+pub struct AXElement {
+    pub role: String,
+    pub title: Option<String>,
+    pub value: Option<String>,
+    pub label: Option<String>,
+    pub identifier: Option<String>,
+    pub enabled: bool,
+    pub focused: bool,
+    pub position: Option<(f64, f64)>,
+    pub size: Option<(f64, f64)>,
+    pub children_count: usize,
+}
+
+/// Represents a macOS application
+#[derive(Debug, Clone)]
+pub struct AXApplication {
+    pub name: String,
+    pub bundle_id: Option<String>,
+    pub pid: i32,
+}
+
+impl AXElement {
+    /// Convert to a human-readable string representation
+    pub fn to_string(&self) -> String {
+        let mut parts = vec![format!("Role: {}", self.role)];
+        
+        if let Some(ref title) = self.title {
+            parts.push(format!("Title: {}", title));
+        }
+        if let Some(ref value) = self.value {
+            parts.push(format!("Value: {}", value));
+        }
+        if let Some(ref label) = self.label {
+            parts.push(format!("Label: {}", label));
+        }
+        if let Some(ref id) = self.identifier {
+            parts.push(format!("ID: {}", id));
+        }
+        
+        parts.push(format!("Enabled: {}", self.enabled));
+        parts.push(format!("Focused: {}", self.focused));
+        
+        if let Some((x, y)) = self.position {
+            parts.push(format!("Position: ({:.0}, {:.0})", x, y));
+        }
+        if let Some((w, h)) = self.size {
+            parts.push(format!("Size: ({:.0}, {:.0})", w, h));
+        }
+        
+        parts.push(format!("Children: {}", self.children_count));
+        
+        parts.join(", ")
+    }
+}

crates/g3-computer-control/src/macax/tests.rs

@@ -0,0 +1,37 @@
+#[cfg(test)]
+mod tests {
+    use crate::{AXElement, MacAxController};
+
+    #[test]
+    fn test_ax_element_to_string() {
+        let element = AXElement {
+            role: "button".to_string(),
+            title: Some("Click Me".to_string()),
+            value: None,
+            label: Some("Submit Button".to_string()),
+            identifier: Some("submitBtn".to_string()),
+            enabled: true,
+            focused: false,
+            position: Some((100.0, 200.0)),
+            size: Some((80.0, 30.0)),
+            children_count: 0,
+        };
+
+        let string_repr = element.to_string();
+        assert!(string_repr.contains("Role: button"));
+        assert!(string_repr.contains("Title: Click Me"));
+        assert!(string_repr.contains("Label: Submit Button"));
+        assert!(string_repr.contains("ID: submitBtn"));
+        assert!(string_repr.contains("Enabled: true"));
+        assert!(string_repr.contains("Position: (100, 200)"));
+        assert!(string_repr.contains("Size: (80, 30)"));
+    }
+
+    #[test]
+    fn test_controller_creation() {
+        // Just test that we can create a controller
+        // Actual functionality requires macOS and permissions
+        let result = MacAxController::new();
+        assert!(result.is_ok());
+    }
+}

crates/g3-computer-control/src/ocr/mod.rs

@@ -0,0 +1,26 @@
+use crate::types::TextLocation;
+use anyhow::Result;
+use async_trait::async_trait;
+
+/// OCR engine trait for text recognition with bounding boxes
+#[async_trait]
+pub trait OCREngine: Send + Sync {
+    /// Extract text with locations from an image file
+    async fn extract_text_with_locations(&self, path: &str) -> Result<Vec<TextLocation>>;
+    
+    /// Get the name of the OCR engine
+    fn name(&self) -> &str;
+}
+
+// Platform-specific modules
+#[cfg(target_os = "macos")]
+pub mod vision;
+
+pub mod tesseract;
+
+// Re-export the default OCR engine for the platform
+#[cfg(target_os = "macos")]
+pub use vision::AppleVisionOCR as DefaultOCR;
+
+#[cfg(not(target_os = "macos"))]
+pub use tesseract::TesseractOCR as DefaultOCR;

crates/g3-computer-control/src/ocr/tesseract.rs

@@ -0,0 +1,84 @@
+use super::OCREngine;
+use crate::types::TextLocation;
+use anyhow::Result;
+use async_trait::async_trait;
+
+/// Tesseract OCR engine (fallback/cross-platform)
+pub struct TesseractOCR;
+
+impl TesseractOCR {
+    pub fn new() -> Result<Self> {
+        // Check if tesseract is available
+        let tesseract_check = std::process::Command::new("which")
+            .arg("tesseract")
+            .output();
+        
+        if tesseract_check.is_err() || !tesseract_check.as_ref().unwrap().status.success() {
+            anyhow::bail!("Tesseract OCR is not installed on your system.\n\n\
+                To install tesseract:\n  macOS:   brew install tesseract\n  \
+                Linux:   sudo apt-get install tesseract-ocr (Ubuntu/Debian)\n           \
+                sudo yum install tesseract (RHEL/CentOS)\n  \
+                Windows: Download from https://github.com/UB-Mannheim/tesseract/wiki\n\n\
+                After installation, restart your terminal and try again.");
+        }
+        
+        Ok(Self)
+    }
+}
+
+#[async_trait]
+impl OCREngine for TesseractOCR {
+    async fn extract_text_with_locations(&self, path: &str) -> Result<Vec<TextLocation>> {
+        // Use tesseract CLI with TSV output to get bounding boxes
+        let output = std::process::Command::new("tesseract")
+            .arg(path)
+            .arg("stdout")
+            .arg("tsv")
+            .output()
+            .map_err(|e| anyhow::anyhow!("Failed to run tesseract: {}", e))?;
+        
+        if !output.status.success() {
+            anyhow::bail!("Tesseract failed: {}", String::from_utf8_lossy(&output.stderr));
+        }
+        
+        let tsv_text = String::from_utf8_lossy(&output.stdout);
+        let mut locations = Vec::new();
+        
+        // Parse TSV output (skip header line)
+        for (i, line) in tsv_text.lines().enumerate() {
+            if i == 0 { continue; } // Skip header
+            
+            let parts: Vec<&str> = line.split('\t').collect();
+            if parts.len() >= 12 {
+                // TSV format: level, page_num, block_num, par_num, line_num, word_num,
+                //             left, top, width, height, conf, text
+                if let (Ok(x), Ok(y), Ok(w), Ok(h), Ok(conf), text) = (
+                    parts[6].parse::<i32>(),
+                    parts[7].parse::<i32>(),
+                    parts[8].parse::<i32>(),
+                    parts[9].parse::<i32>(),
+                    parts[10].parse::<f32>(),
+                    parts[11],
+                ) {
+                    let trimmed = text.trim();
+                    if !trimmed.is_empty() && conf > 0.0 {
+                        locations.push(TextLocation {
+                            text: trimmed.to_string(),
+                            x,
+                            y,
+                            width: w,
+                            height: h,
+                            confidence: conf / 100.0, // Convert from 0-100 to 0-1
+                        });
+                    }
+                }
+            }
+        }
+        
+        Ok(locations)
+    }
+    
+    fn name(&self) -> &str {
+        "Tesseract OCR"
+    }
+}

crates/g3-computer-control/src/ocr/vision.rs

@@ -0,0 +1,103 @@
+use super::OCREngine;
+use crate::types::TextLocation;
+use anyhow::{Result, Context};
+use async_trait::async_trait;
+use std::ffi::{CStr, CString};
+use std::os::raw::{c_char, c_float, c_uint};
+
+// FFI bindings to Swift VisionBridge
+#[repr(C)]
+struct VisionTextBox {
+    text: *const c_char,
+    text_len: c_uint,
+    x: i32,
+    y: i32,
+    width: i32,
+    height: i32,
+    confidence: c_float,
+}
+
+extern "C" {
+    fn vision_recognize_text(
+        image_path: *const c_char,
+        image_path_len: c_uint,
+        out_boxes: *mut *mut std::ffi::c_void,
+        out_count: *mut c_uint,
+    ) -> bool;
+    
+    fn vision_free_boxes(boxes: *mut std::ffi::c_void, count: c_uint);
+}
+
+/// Apple Vision Framework OCR engine
+pub struct AppleVisionOCR;
+
+impl AppleVisionOCR {
+    pub fn new() -> Result<Self> {
+        Ok(Self)
+    }
+}
+
+#[async_trait]
+impl OCREngine for AppleVisionOCR {
+    async fn extract_text_with_locations(&self, path: &str) -> Result<Vec<TextLocation>> {
+        // Convert path to C string
+        let c_path = CString::new(path)
+            .context("Failed to convert path to C string")?;
+        
+        let mut boxes_ptr: *mut std::ffi::c_void = std::ptr::null_mut();
+        let mut count: c_uint = 0;
+        
+        // Call Swift Vision API
+        let success = unsafe {
+            vision_recognize_text(
+                c_path.as_ptr(),
+                path.len() as c_uint,
+                &mut boxes_ptr,
+                &mut count,
+            )
+        };
+        
+        if !success || boxes_ptr.is_null() {
+            anyhow::bail!("Apple Vision OCR failed");
+        }
+        
+        // Convert C array to Rust Vec
+        let mut locations = Vec::new();
+        
+        unsafe {
+            let typed_boxes = boxes_ptr as *const VisionTextBox;
+            let boxes_slice = std::slice::from_raw_parts(typed_boxes, count as usize);
+            
+            for box_data in boxes_slice {
+                // Convert C string to Rust String
+                let text = if !box_data.text.is_null() {
+                    CStr::from_ptr(box_data.text)
+                        .to_string_lossy()
+                        .into_owned()
+                } else {
+                    String::new()
+                };
+                
+                if !text.is_empty() {
+                    locations.push(TextLocation {
+                        text,
+                        x: box_data.x,
+                        y: box_data.y,
+                        width: box_data.width,
+                        height: box_data.height,
+                        confidence: box_data.confidence,
+                    });
+                }
+            }
+            
+            // Free the C array
+            vision_free_boxes(boxes_ptr, count);
+        }
+        
+        Ok(locations)
+    }
+    
+    fn name(&self) -> &str {
+        "Apple Vision Framework"
+    }
+}

crates/g3-computer-control/src/platform/linux.rs

@@ -0,0 +1,166 @@
+use crate::{ComputerController, types::*};
+use anyhow::Result;
+use async_trait::async_trait;
+use tesseract::Tesseract;
+use uuid::Uuid;
+
+pub struct LinuxController {
+    // Placeholder for X11 connection or other state
+}
+
+impl LinuxController {
+    pub fn new() -> Result<Self> {
+        // Initialize X11 connection
+        tracing::warn!("Linux computer control not fully implemented");
+        Ok(Self {})
+    }
+}
+
+#[async_trait]
+impl ComputerController for LinuxController {
+    async fn move_mouse(&self, _x: i32, _y: i32) -> Result<()> {
+        anyhow::bail!("Linux implementation not yet available")
+    }
+    
+    async fn click(&self, _button: MouseButton) -> Result<()> {
+        anyhow::bail!("Linux implementation not yet available")
+    }
+    
+    async fn double_click(&self, _button: MouseButton) -> Result<()> {
+        anyhow::bail!("Linux implementation not yet available")
+    }
+    
+    async fn type_text(&self, _text: &str) -> Result<()> {
+        anyhow::bail!("Linux implementation not yet available")
+    }
+    
+    async fn press_key(&self, _key: &str) -> Result<()> {
+        anyhow::bail!("Linux implementation not yet available")
+    }
+    
+    async fn list_windows(&self) -> Result<Vec<Window>> {
+        anyhow::bail!("Linux implementation not yet available")
+    }
+    
+    async fn focus_window(&self, _window_id: &str) -> Result<()> {
+        anyhow::bail!("Linux implementation not yet available")
+    }
+    
+    async fn get_window_bounds(&self, _window_id: &str) -> Result<Rect> {
+        anyhow::bail!("Linux implementation not yet available")
+    }
+    
+    async fn find_element(&self, _selector: &ElementSelector) -> Result<Option<UIElement>> {
+        anyhow::bail!("Linux implementation not yet available")
+    }
+    
+    async fn get_element_text(&self, _element_id: &str) -> Result<String> {
+        anyhow::bail!("Linux implementation not yet available")
+    }
+    
+    async fn get_element_bounds(&self, _element_id: &str) -> Result<Rect> {
+        anyhow::bail!("Linux implementation not yet available")
+    }
+    
+    async fn take_screenshot(&self, _path: &str, _region: Option<Rect>, _window_id: Option<&str>) -> Result<()> {
+        // Enforce that window_id must be provided
+        if _window_id.is_none() {
+            anyhow::bail!("window_id is required. You must specify which window to capture (e.g., 'Firefox', 'Terminal', 'gedit'). Use list_windows to see available windows.");
+        }
+
+        anyhow::bail!("Linux implementation not yet available")
+    }
+    
+    async fn extract_text_from_screen(&self, _region: Rect, _window_id: &str) -> Result<String> {
+        anyhow::bail!("Linux implementation not yet available")
+    }
+    
+    async fn extract_text_from_image(&self, _path: &str) -> Result<OCRResult> {
+        // Check if tesseract is available on the system
+        let tesseract_check = std::process::Command::new("which")
+            .arg("tesseract")
+            .output();
+        
+        if tesseract_check.is_err() || !tesseract_check.as_ref().unwrap().status.success() {
+            anyhow::bail!("Tesseract OCR is not installed on your system.\n\n\
+                To install tesseract:\n  \
+                Ubuntu/Debian: sudo apt-get install tesseract-ocr\n  \
+                RHEL/CentOS:   sudo yum install tesseract\n  \
+                Arch Linux:    sudo pacman -S tesseract\n\n\
+                After installation, restart your terminal and try again.");
+        }
+        
+        // Initialize Tesseract
+        let tess = Tesseract::new(None, Some("eng"))
+            .map_err(|e| {
+                anyhow::anyhow!("Failed to initialize Tesseract: {}\n\n\
+                    This usually means:\n1. Tesseract is not properly installed\n\
+                    2. Language data files are missing\n\nTo fix:\n  \
+                    Ubuntu/Debian: sudo apt-get install tesseract-ocr-eng\n  \
+                    RHEL/CentOS:   sudo yum install tesseract-langpack-eng\n  \
+                    Arch Linux:    sudo pacman -S tesseract-data-eng", e)
+            })?;
+        
+        let text = tess.set_image(_path)
+            .map_err(|e| anyhow::anyhow!("Failed to load image '{}': {}", _path, e))?
+            .get_text()
+            .map_err(|e| anyhow::anyhow!("Failed to extract text from image: {}", e))?;
+        
+        // Get confidence (simplified - would need more complex API calls for per-word confidence)
+        let confidence = 0.85; // Placeholder
+        
+        Ok(OCRResult {
+            text,
+            confidence,
+            bounds: Rect { x: 0, y: 0, width: 0, height: 0 }, // Would need image dimensions
+        })
+    }
+    
+    async fn find_text_on_screen(&self, _text: &str) -> Result<Option<Point>> {
+        // Check if tesseract is available on the system
+        let tesseract_check = std::process::Command::new("which")
+            .arg("tesseract")
+            .output();
+        
+        if tesseract_check.is_err() || !tesseract_check.as_ref().unwrap().status.success() {
+            anyhow::bail!("Tesseract OCR is not installed on your system.\n\n\
+                To install tesseract:\n  \
+                Ubuntu/Debian: sudo apt-get install tesseract-ocr\n  \
+                RHEL/CentOS:   sudo yum install tesseract\n  \
+                Arch Linux:    sudo pacman -S tesseract\n\n\
+                After installation, restart your terminal and try again.");
+        }
+        
+        // Take full screen screenshot
+        let temp_path = format!("/tmp/g3_ocr_search_{}.png", uuid::Uuid::new_v4());
+        self.take_screenshot(&temp_path, None, None).await?;
+        
+        // Use Tesseract to find text with bounding boxes
+        let tess = Tesseract::new(None, Some("eng"))
+            .map_err(|e| {
+                anyhow::anyhow!("Failed to initialize Tesseract: {}\n\n\
+                    This usually means:\n1. Tesseract is not properly installed\n\
+                    2. Language data files are missing\n\nTo fix:\n  \
+                    Ubuntu/Debian: sudo apt-get install tesseract-ocr-eng\n  \
+                    RHEL/CentOS:   sudo yum install tesseract-langpack-eng\n  \
+                    Arch Linux:    sudo pacman -S tesseract-data-eng", e)
+            })?;
+        
+        let full_text = tess.set_image(temp_path.as_str())
+            .map_err(|e| anyhow::anyhow!("Failed to load screenshot: {}", e))?
+            .get_text()
+            .map_err(|e| anyhow::anyhow!("Failed to extract text from screen: {}", e))?;
+        
+        // Clean up temp file
+        let _ = std::fs::remove_file(&temp_path);
+        
+        // Simple text search - full implementation would use get_component_images
+        // to get bounding boxes for each word
+        if full_text.contains(_text) {
+            tracing::warn!("Text found but precise coordinates not available in simplified implementation");
+            Ok(Some(Point { x: 0, y: 0 }))
+        } else {
+            Ok(None)
+        }
+    }
+}

crates/g3-computer-control/src/platform/macos.rs

@@ -0,0 +1,507 @@
+use crate::{ComputerController, types::{Rect, TextLocation}};
+use crate::ocr::{OCREngine, DefaultOCR};
+use anyhow::{Result, Context};
+use async_trait::async_trait;
+use std::path::Path;
+use core_graphics::window::{kCGWindowListOptionOnScreenOnly, kCGNullWindowID, CGWindowListCopyWindowInfo};
+use core_foundation::dictionary::CFDictionary;
+use core_foundation::string::CFString;
+use core_foundation::base::{TCFType, ToVoid};
+use core_foundation::array::CFArray;
+
+pub struct MacOSController {
+    ocr_engine: Box<dyn OCREngine>,
+    #[allow(dead_code)]
+    ocr_name: String,
+}
+
+impl MacOSController {
+    pub fn new() -> Result<Self> {
+        let ocr = Box::new(DefaultOCR::new()?);
+        let ocr_name = ocr.name().to_string();
+        tracing::info!("Initialized macOS controller with OCR engine: {}", ocr_name);
+        Ok(Self { ocr_engine: ocr, ocr_name })
+    }
+}
+
+#[async_trait]
+impl ComputerController for MacOSController {
+    async fn take_screenshot(&self, path: &str, region: Option<Rect>, window_id: Option<&str>) -> Result<()> {
+        // Enforce that window_id must be provided
+        if window_id.is_none() {
+            return Err(anyhow::anyhow!("window_id is required. You must specify which window to capture (e.g., 'Safari', 'Terminal', 'Google Chrome'). Use list_windows to see available windows."));
+        }
+
+        // Determine the temporary directory for screenshots
+        let temp_dir = std::env::var("TMPDIR")
+            .or_else(|_| std::env::var("HOME").map(|h| format!("{}/tmp", h)))
+            .unwrap_or_else(|_| "/tmp".to_string());
+        
+        // Ensure temp directory exists
+        std::fs::create_dir_all(&temp_dir)?;
+        
+        // If path is relative or doesn't specify a directory, use temp_dir
+        let final_path = if path.starts_with('/') {
+            path.to_string()
+        } else {
+            format!("{}/{}", temp_dir.trim_end_matches('/'), path)
+        };
+        
+        let path_obj = Path::new(&final_path);
+        if let Some(parent) = path_obj.parent() {
+            std::fs::create_dir_all(parent)?;
+        }
+        
+        let app_name = window_id.unwrap(); // Safe because we checked is_none() above
+        
+        // Get the window ID for the specified application
+        let cg_window_id = unsafe {
+            let window_list = CGWindowListCopyWindowInfo(
+                kCGWindowListOptionOnScreenOnly,
+                kCGNullWindowID
+            );
+            
+            let array = CFArray::<CFDictionary>::wrap_under_create_rule(window_list);
+            let count = array.len();
+            
+            let mut found_window_id: Option<(u32, String)> = None; // (id, owner)
+            let app_name_lower = app_name.to_lowercase();
+            
+            for i in 0..count {
+                let dict = array.get(i).unwrap();
+                
+                // Get owner name
+                let owner_key = CFString::from_static_string("kCGWindowOwnerName");
+                let owner: String = if let Some(value) = dict.find(owner_key.to_void()) {
+                    let s: CFString = TCFType::wrap_under_get_rule(*value as *const _);
+                    s.to_string()
+                } else {
+                    continue;
+                };
+                
+                tracing::debug!("Checking window: owner='{}', looking for '{}'", owner, app_name);
+                let owner_lower = owner.to_lowercase();
+                
+                // Normalize by removing spaces for exact matching
+                let app_name_normalized = app_name_lower.replace(" ", "");
+                let owner_normalized = owner_lower.replace(" ", "");
+                
+                // ONLY accept exact matches (case-insensitive, with or without spaces)
+                // This prevents "Goose" from matching "GooseStudio"
+                let is_match = owner_lower == app_name_lower || owner_normalized == app_name_normalized;
+                
+                if is_match {
+                    // Get window ID
+                    let window_id_key = CFString::from_static_string("kCGWindowNumber");
+                    if let Some(value) = dict.find(window_id_key.to_void()) {
+                        let num: core_foundation::number::CFNumber = TCFType::wrap_under_get_rule(*value as *const _);
+                        if let Some(id) = num.to_i64() {
+                            // Get window layer to filter out menu bar windows
+                            let layer_key = CFString::from_static_string("kCGWindowLayer");
+                            let layer: i32 = if let Some(value) = dict.find(layer_key.to_void()) {
+                                let num: core_foundation::number::CFNumber = TCFType::wrap_under_get_rule(*value as *const _);
+                                num.to_i32().unwrap_or(0)
+                            } else {
+                                0
+                            };
+                            
+                            // Get window bounds to verify it's a real window
+                            let bounds_key = CFString::from_static_string("kCGWindowBounds");
+                            let has_real_bounds = if let Some(value) = dict.find(bounds_key.to_void()) {
+                                let bounds_dict: CFDictionary = TCFType::wrap_under_get_rule(*value as *const _);
+                                let width_key = CFString::from_static_string("Width");
+                                let height_key = CFString::from_static_string("Height");
+                                
+                                if let (Some(w_val), Some(h_val)) = (
+                                    bounds_dict.find(width_key.to_void()),
+                                    bounds_dict.find(height_key.to_void()),
+                                ) {
+                                    let w_num: core_foundation::number::CFNumber = TCFType::wrap_under_get_rule(*w_val as *const _);
+                                    let h_num: core_foundation::number::CFNumber = TCFType::wrap_under_get_rule(*h_val as *const _);
+                                    let width = w_num.to_f64().unwrap_or(0.0);
+                                    let height = h_num.to_f64().unwrap_or(0.0);
+                                    // Real windows should be at least 100x100 pixels
+                                    width >= 100.0 && height >= 100.0
+                                } else {
+                                    false
+                                }
+                            } else {
+                                false
+                            };
+                            
+                            // Only accept windows that are:
+                            // 1. At layer 0 (normal windows, not menu bar)
+                            // 2. Have real bounds (width and height >= 100)
+                            if layer == 0 && has_real_bounds {
+                                tracing::info!("Found valid window: ID {} for app '{}' (layer={}, bounds valid)", id, owner, layer);
+                                found_window_id = Some((id as u32, owner.clone()));
+                                break;
+                            } else {
+                                tracing::debug!("Skipping window ID {} for '{}': layer={}, has_real_bounds={}", id, owner, layer, has_real_bounds);
+                            }
+                        }
+                    }
+                }
+            }
+            
+            found_window_id
+        };
+        
+        let (cg_window_id, matched_owner) = cg_window_id.ok_or_else(|| {
+            anyhow::anyhow!("Could not find window for application '{}'. Use list_windows to see available windows.", app_name)
+        })?;
+            tracing::info!("Taking screenshot of window ID {} for app '{}'", cg_window_id, matched_owner);
+        
+        // Use screencapture with the window ID for now
+        // TODO: Implement direct CGWindowListCreateImage approach with proper image saving
+        let mut cmd = std::process::Command::new("screencapture");
+        cmd.arg("-x"); // No sound
+        cmd.arg("-l");
+        cmd.arg(cg_window_id.to_string());
+        
+        if let Some(region) = region {
+            cmd.arg("-R");
+            cmd.arg(format!("{},{},{},{}", region.x, region.y, region.width, region.height));
+        }
+        
+        cmd.arg(&final_path);
+        
+        let screenshot_result = cmd.output()?;
+        
+        if !screenshot_result.status.success() {
+            let stderr = String::from_utf8_lossy(&screenshot_result.stderr);
+            return Err(anyhow::anyhow!("screencapture failed for window {}: {}", cg_window_id, stderr));
+        }
+        
+        Ok(())
+    }
+    
+    async fn extract_text_from_screen(&self, region: Rect, window_id: &str) -> Result<String> {
+        // Take screenshot of region first
+        let temp_path = format!("/tmp/g3_ocr_{}.png", uuid::Uuid::new_v4());
+        self.take_screenshot(&temp_path, Some(region), Some(window_id)).await?;
+        
+        // Extract text from the screenshot
+        let result = self.extract_text_from_image(&temp_path).await?;
+        
+        // Clean up temp file
+        let _ = std::fs::remove_file(&temp_path);
+        
+        Ok(result)
+    }
+    
+    async fn extract_text_from_image(&self, path: &str) -> Result<String> {
+        // Extract all text and concatenate
+        let locations = self.ocr_engine.extract_text_with_locations(path).await?;
+        Ok(locations.iter().map(|loc| loc.text.as_str()).collect::<Vec<_>>().join(" "))
+    }
+    
+    async fn extract_text_with_locations(&self, path: &str) -> Result<Vec<TextLocation>> {
+        // Use the OCR engine
+        self.ocr_engine.extract_text_with_locations(path).await
+    }
+    
+    async fn find_text_in_app(&self, app_name: &str, search_text: &str) -> Result<Option<TextLocation>> {
+        // Take screenshot of specific app window
+        let home = std::env::var("HOME").unwrap_or_else(|_| "/tmp".to_string());
+        let temp_path = format!("{}/tmp/g3_find_text_{}_{}.png", home, app_name, uuid::Uuid::new_v4());
+        self.take_screenshot(&temp_path, None, Some(app_name)).await?;
+        
+        // Get screenshot dimensions before we delete it
+        let screenshot_dims = get_image_dimensions(&temp_path)?;
+        
+        // Extract all text with locations
+        let locations = self.extract_text_with_locations(&temp_path).await?;
+        
+        // Get window bounds to calculate coordinate transformation
+        let window_bounds = self.get_window_bounds(app_name)?;
+        
+        // Clean up temp file
+        let _ = std::fs::remove_file(&temp_path);
+        
+        // Find matching text (case-insensitive)
+        let search_lower = search_text.to_lowercase();
+        for location in locations {
+            if location.text.to_lowercase().contains(&search_lower) {
+                // Transform coordinates from screenshot space to screen space
+                let transformed = transform_screenshot_to_screen_coords(
+                    location,
+                    window_bounds,
+                    screenshot_dims,
+                );
+                return Ok(Some(transformed));
+            }
+        }
+        
+        Ok(None)
+    }
+    
+    fn move_mouse(&self, x: i32, y: i32) -> Result<()> {
+        use core_graphics::event::{
+            CGEvent, CGEventTapLocation, CGEventType, CGMouseButton,
+        };
+        use core_graphics::event_source::{
+            CGEventSource, CGEventSourceStateID,
+        };
+        use core_graphics::geometry::CGPoint;
+        
+        let source = CGEventSource::new(CGEventSourceStateID::HIDSystemState)
+            .ok().context("Failed to create event source")?;
+        
+        let event = CGEvent::new_mouse_event(
+            source,
+            CGEventType::MouseMoved,
+            CGPoint::new(x as f64, y as f64),
+            CGMouseButton::Left,
+        ).ok().context("Failed to create mouse event")?;
+        
+        event.post(CGEventTapLocation::HID);
+        
+        Ok(())
+    }
+    
+    fn click_at(&self, x: i32, y: i32, _app_name: Option<&str>) -> Result<()> {
+        use core_graphics::event::{
+            CGEvent, CGEventTapLocation, CGEventType, CGMouseButton,
+        };
+        use core_graphics::event_source::{
+            CGEventSource, CGEventSourceStateID,
+        };
+        use core_graphics::geometry::CGPoint;
+        use core_graphics::display::CGDisplay;
+        
+        // IMPORTANT: Coordinates passed here are in NSScreen/CGWindowListCopyWindowInfo space
+        // (Y=0 at BOTTOM, increases UPWARD)
+        // But CGEvent uses a different coordinate system (Y=0 at TOP, increases DOWNWARD)
+        // We need to convert: CGEvent.y = screenHeight - NSScreen.y
+        
+        let screen_height = CGDisplay::main().pixels_high() as i32;
+        let cgevent_x = x;
+        let cgevent_y = screen_height - y;
+        
+        tracing::debug!("click_at: NSScreen coords ({}, {}) -> CGEvent coords ({}, {}) [screen_height={}]",
+            x, y, cgevent_x, cgevent_y, screen_height);
+        
+        let (global_x, global_y) = (cgevent_x, cgevent_y);
+        
+        let point = CGPoint::new(global_x as f64, global_y as f64);
+        
+        let source = CGEventSource::new(CGEventSourceStateID::HIDSystemState)
+            .ok().context("Failed to create event source")?;
+        
+        // Move mouse to position first
+        let move_event = CGEvent::new_mouse_event(
+            source.clone(),
+            CGEventType::MouseMoved,
+            point,
+            CGMouseButton::Left,
+        ).ok().context("Failed to create mouse move event")?;
+        move_event.post(CGEventTapLocation::HID);
+        
+        std::thread::sleep(std::time::Duration::from_millis(100));
+        
+        // Mouse down
+        let mouse_down = CGEvent::new_mouse_event(
+            source.clone(),
+            CGEventType::LeftMouseDown,
+            point,
+            CGMouseButton::Left,
+        ).ok().context("Failed to create mouse down event")?;
+        mouse_down.post(CGEventTapLocation::HID);
+        
+        std::thread::sleep(std::time::Duration::from_millis(50));
+        
+        // Mouse up
+        let mouse_up = CGEvent::new_mouse_event(
+            source,
+            CGEventType::LeftMouseUp,
+            point,
+            CGMouseButton::Left,
+        ).ok().context("Failed to create mouse up event")?;
+        mouse_up.post(CGEventTapLocation::HID);
+        
+        Ok(())
+    }
+}
+
+impl MacOSController {
+    /// Get window bounds for an application (helper method)
+    fn get_window_bounds(&self, app_name: &str) -> Result<(i32, i32, i32, i32)> {
+        unsafe {
+            let window_list = CGWindowListCopyWindowInfo(
+                kCGWindowListOptionOnScreenOnly,
+                kCGNullWindowID
+            );
+            
+            let array = CFArray::<CFDictionary>::wrap_under_create_rule(window_list);
+            let count = array.len();
+            
+            let app_name_lower = app_name.to_lowercase();
+            
+            for i in 0..count {
+                let dict = array.get(i).unwrap();
+                
+                // Get owner name
+                let owner_key = CFString::from_static_string("kCGWindowOwnerName");
+                let owner: String = if let Some(value) = dict.find(owner_key.to_void()) {
+                    let s: CFString = TCFType::wrap_under_get_rule(*value as *const _);
+                    s.to_string()
+                } else {
+                    continue;
+                };
+                
+                let owner_lower = owner.to_lowercase();
+                
+                // Normalize by removing spaces for exact matching
+                let app_name_normalized = app_name_lower.replace(" ", "");
+                let owner_normalized = owner_lower.replace(" ", "");
+                
+                // ONLY accept exact matches (case-insensitive, with or without spaces)
+                // This prevents "Goose" from matching "GooseStudio"
+                let is_match = owner_lower == app_name_lower || owner_normalized == app_name_normalized;
+                
+                if is_match {
+                    // Get window layer to filter out menu bar windows
+                    let layer_key = CFString::from_static_string("kCGWindowLayer");
+                    let layer: i32 = if let Some(value) = dict.find(layer_key.to_void()) {
+                        let num: core_foundation::number::CFNumber = TCFType::wrap_under_get_rule(*value as *const _);
+                        num.to_i32().unwrap_or(0)
+                    } else {
+                        0
+                    };
+                    
+                    // Skip menu bar windows (layer >= 20)
+                    if layer >= 20 {
+                        tracing::debug!("Skipping window for '{}' at layer {} (menu bar)", owner, layer);
+                        continue;
+                    }
+                    
+                    // Get window bounds to verify it's a real window
+                    let bounds_key = CFString::from_static_string("kCGWindowBounds");
+                    if let Some(value) = dict.find(bounds_key.to_void()) {
+                        let bounds_dict: CFDictionary = TCFType::wrap_under_get_rule(*value as *const _);
+                        
+                        let x_key = CFString::from_static_string("X");
+                        let y_key = CFString::from_static_string("Y");
+                        let width_key = CFString::from_static_string("Width");
+                        let height_key = CFString::from_static_string("Height");
+                        
+                        if let (Some(x_val), Some(y_val), Some(w_val), Some(h_val)) = (
+                            bounds_dict.find(x_key.to_void()),
+                            bounds_dict.find(y_key.to_void()),
+                            bounds_dict.find(width_key.to_void()),
+                            bounds_dict.find(height_key.to_void()),
+                        ) {
+                            let x_num: core_foundation::number::CFNumber = TCFType::wrap_under_get_rule(*x_val as *const _);
+                            let y_num: core_foundation::number::CFNumber = TCFType::wrap_under_get_rule(*y_val as *const _);
+                            let w_num: core_foundation::number::CFNumber = TCFType::wrap_under_get_rule(*w_val as *const _);
+                            let h_num: core_foundation::number::CFNumber = TCFType::wrap_under_get_rule(*h_val as *const _);
+                            
+                            let x: i32 = x_num.to_i64().unwrap_or(0) as i32;
+                            let y: i32 = y_num.to_i64().unwrap_or(0) as i32;
+                            let w: i32 = w_num.to_i64().unwrap_or(0) as i32;
+                            let h: i32 = h_num.to_i64().unwrap_or(0) as i32;
+                            
+                            // Only accept windows with real bounds (>= 100x100 pixels)
+                            if w >= 100 && h >= 100 {
+                                tracing::info!("Found valid window bounds for '{}': x={}, y={}, w={}, h={} (layer={})", owner, x, y, w, h, layer);
+                                return Ok((x, y, w, h));
+                            } else {
+                                tracing::debug!("Skipping window for '{}': too small ({}x{})", owner, w, h);
+                                continue;
+                            }
+                        } else {
+                            continue;
+                        }
+                    }
+                }
+            }
+        }
+        
+        Err(anyhow::anyhow!("Could not find window bounds for '{}'", app_name))
+    }
+}
+
+/// Get image dimensions from a PNG file
+fn get_image_dimensions(path: &str) -> Result<(i32, i32)> {
+    use std::fs::File;
+    use std::io::Read;
+    
+    let mut file = File::open(path)?;
+    let mut buffer = vec![0u8; 24];
+    file.read_exact(&mut buffer)?;
+    
+    // PNG signature check
+    if &buffer[0..8] != b"\x89PNG\r\n\x1a\n" {
+        anyhow::bail!("Not a valid PNG file");
+    }
+    
+    // Read IHDR chunk (width and height are at bytes 16-23)
+    let width = u32::from_be_bytes([buffer[16], buffer[17], buffer[18], buffer[19]]) as i32;
+    let height = u32::from_be_bytes([buffer[20], buffer[21], buffer[22], buffer[23]]) as i32;
+    
+    Ok((width, height))
+}
+
+/// Transform coordinates from screenshot space to screen space
+/// 
+/// The screenshot is taken of a window, and Vision OCR returns coordinates
+/// relative to the screenshot image. We need to transform these to actual
+/// screen coordinates for clicking.
+/// 
+/// On Retina displays, screenshots are taken at 2x resolution, so we need
+/// to account for this scaling factor.
+fn transform_screenshot_to_screen_coords(
+    location: TextLocation,
+    window_bounds: (i32, i32, i32, i32), // (x, y, width, height) in screen space
+    screenshot_dims: (i32, i32), // (width, height) in pixels
+) -> TextLocation {
+    let (win_x, win_y, win_width, win_height) = window_bounds;
+    let (screenshot_width, screenshot_height) = screenshot_dims;
+    
+    // Calculate scale factors
+    // On Retina displays, screenshot is typically 2x the window size
+    let scale_x = win_width as f64 / screenshot_width as f64;
+    let scale_y = win_height as f64 / screenshot_height as f64;
+    
+    tracing::debug!("Transform: screenshot={}x{}, window={}x{} at ({},{}), scale=({:.2},{:.2})",
+        screenshot_width, screenshot_height, win_width, win_height, win_x, win_y, scale_x, scale_y);
+    
+    // Transform coordinates from image space to screen space
+    // IMPORTANT: macOS screen coordinates have origin at BOTTOM-LEFT (Y increases upward)
+    // Image coordinates have origin at TOP-LEFT (Y increases downward)
+    // win_y is the BOTTOM of the window in screen coordinates
+    // So we need to: (win_y + win_height) to get window TOP, then subtract screenshot_y
+    let window_top_y = win_y + win_height;
+    
+    tracing::debug!("[transform] Input location in image space: x={}, y={}, width={}, height={}",
+        location.x, location.y, location.width, location.height);
+    tracing::debug!("[transform] Scale factors: scale_x={:.4}, scale_y={:.4}", scale_x, scale_y);
+    
+    let transformed_x = win_x + (location.x as f64 * scale_x) as i32;
+    let transformed_y = window_top_y - (location.y as f64 * scale_y) as i32;
+    let transformed_width = (location.width as f64 * scale_x) as i32;
+    let transformed_height = (location.height as f64 * scale_y) as i32;
+    
+    tracing::debug!("[transform] Calculation details:");
+    tracing::debug!("  - transformed_x = {} + ({} * {:.4}) = {} + {:.2} = {}", win_x, location.x, scale_x, win_x, location.x as f64 * scale_x, transformed_x);
+    tracing::debug!("  - transformed_width = ({} * {:.4}) = {:.2} -> {}", location.width, scale_x, location.width as f64 * scale_x, transformed_width);
+    tracing::debug!("  - transformed_height = ({} * {:.4}) = {:.2} -> {}", location.height, scale_y, location.height as f64 * scale_y, transformed_height);
+    
+    tracing::debug!("Transformed location: screenshot=({},{}) {}x{} -> screen=({},{}) {}x{}",
+        location.x, location.y, location.width, location.height,
+        transformed_x, transformed_y, transformed_width, transformed_height);
+    
+    TextLocation {
+        text: location.text,
+        x: transformed_x,
+        y: transformed_y,
+        width: transformed_width,
+        height: transformed_height,
+        confidence: location.confidence,
+    }
+}
+
+#[path = "macos_window_matching_test.rs"]
+#[cfg(test)]
+mod tests;

crates/g3-computer-control/src/platform/macos_window_matching_test.rs

@@ -0,0 +1,45 @@
+#[cfg(test)]
+mod window_matching_tests {
+    /// Test that window name matching handles spaces correctly
+    /// 
+    /// Issue: When a user requests a screenshot of "Goose Studio" but the actual
+    /// application name is "GooseStudio" (no space), the fuzzy matching should
+    /// still find the window.
+    /// 
+    /// The fix normalizes both names by removing spaces before comparing.
+    #[test]
+    fn test_space_normalization() {
+        let test_cases = vec![
+            // (user_input, actual_app_name, should_match)
+            ("Goose Studio", "GooseStudio", true),
+            ("GooseStudio", "Goose Studio", true),
+            ("Visual Studio Code", "VisualStudioCode", true),
+            ("Google Chrome", "Google Chrome", true),
+            ("Safari", "Safari", true),
+            ("iTerm", "iTerm2", true), // fuzzy match
+            ("Code", "Visual Studio Code", true), // fuzzy match
+        ];
+
+        for (user_input, app_name, should_match) in test_cases {
+            let user_lower = user_input.to_lowercase();
+            let app_lower = app_name.to_lowercase();
+            
+            let user_normalized = user_lower.replace(" ", "");
+            let app_normalized = app_lower.replace(" ", "");
+            
+            let is_exact = app_lower == user_lower || app_normalized == user_normalized;
+            let is_fuzzy = app_lower.contains(&user_lower) 
+                || user_lower.contains(&app_lower)
+                || app_normalized.contains(&user_normalized)
+                || user_normalized.contains(&app_normalized);
+            
+            let matches = is_exact || is_fuzzy;
+            
+            assert_eq!(
+                matches, should_match,
+                "Expected '{}' vs '{}' to match={}, but got match={}",
+                user_input, app_name, should_match, matches
+            );
+        }
+    }
+}

crates/g3-computer-control/src/platform/mod.rs

@@ -0,0 +1,8 @@
+#[cfg(target_os = "macos")]
+pub mod macos;
+
+#[cfg(target_os = "linux")]
+pub mod linux;
+
+#[cfg(target_os = "windows")]
+pub mod windows;

crates/g3-computer-control/src/platform/windows.rs

@@ -0,0 +1,167 @@
+use crate::{ComputerController, types::*};
+use anyhow::Result;
+use async_trait::async_trait;
+use tesseract::Tesseract;
+use uuid::Uuid;
+
+pub struct WindowsController {
+    // Placeholder for Windows-specific state
+}
+
+impl WindowsController {
+    pub fn new() -> Result<Self> {
+        tracing::warn!("Windows computer control not fully implemented");
+        Ok(Self {})
+    }
+}
+
+#[async_trait]
+impl ComputerController for WindowsController {
+    async fn move_mouse(&self, _x: i32, _y: i32) -> Result<()> {
+        anyhow::bail!("Windows implementation not yet available")
+    }
+    
+    async fn click(&self, _button: MouseButton) -> Result<()> {
+        anyhow::bail!("Windows implementation not yet available")
+    }
+    
+    async fn double_click(&self, _button: MouseButton) -> Result<()> {
+        anyhow::bail!("Windows implementation not yet available")
+    }
+    
+    async fn type_text(&self, _text: &str) -> Result<()> {
+        anyhow::bail!("Windows implementation not yet available")
+    }
+    
+    async fn press_key(&self, _key: &str) -> Result<()> {
+        anyhow::bail!("Windows implementation not yet available")
+    }
+    
+    async fn list_windows(&self) -> Result<Vec<Window>> {
+        anyhow::bail!("Windows implementation not yet available")
+    }
+    
+    async fn focus_window(&self, _window_id: &str) -> Result<()> {
+        anyhow::bail!("Windows implementation not yet available")
+    }
+    
+    async fn get_window_bounds(&self, _window_id: &str) -> Result<Rect> {
+        anyhow::bail!("Windows implementation not yet available")
+    }
+    
+    async fn find_element(&self, _selector: &ElementSelector) -> Result<Option<UIElement>> {
+        anyhow::bail!("Windows implementation not yet available")
+    }
+    
+    async fn get_element_text(&self, _element_id: &str) -> Result<String> {
+        anyhow::bail!("Windows implementation not yet available")
+    }
+    
+    async fn get_element_bounds(&self, _element_id: &str) -> Result<Rect> {
+        anyhow::bail!("Windows implementation not yet available")
+    }
+    
+    async fn take_screenshot(&self, _path: &str, _region: Option<Rect>, _window_id: Option<&str>) -> Result<()> {
+        // Enforce that window_id must be provided
+        if _window_id.is_none() {
+            anyhow::bail!("window_id is required. You must specify which window to capture (e.g., 'Chrome', 'Terminal', 'Notepad'). Use list_windows to see available windows.");
+        }
+
+        anyhow::bail!("Windows implementation not yet available")
+    }
+    
+    async fn extract_text_from_screen(&self, _region: Rect, _window_id: &str) -> Result<String> {
+        anyhow::bail!("Windows implementation not yet available")
+    }
+    
+    async fn extract_text_from_image(&self, _path: &str) -> Result<OCRResult> {
+        // Check if tesseract is available on the system
+        let tesseract_check = std::process::Command::new("where")
+            .arg("tesseract")
+            .output();
+        
+        if tesseract_check.is_err() || !tesseract_check.as_ref().unwrap().status.success() {
+            anyhow::bail!("Tesseract OCR is not installed on your system.\n\n\
+                To install tesseract on Windows:\n  \
+                1. Download the installer from: https://github.com/UB-Mannheim/tesseract/wiki\n  \
+                2. Run the installer and follow the instructions\n  \
+                3. Add tesseract to your PATH environment variable\n  \
+                4. Restart your terminal/command prompt\n\n\
+                After installation, restart your terminal and try again.");
+        }
+        
+        // Initialize Tesseract
+        let tess = Tesseract::new(None, Some("eng"))
+            .map_err(|e| {
+                anyhow::anyhow!("Failed to initialize Tesseract: {}\n\n\
+                    This usually means:\n1. Tesseract is not properly installed\n\
+                    2. Language data files are missing\n\nTo fix:\n  \
+                    1. Reinstall tesseract from https://github.com/UB-Mannheim/tesseract/wiki\n  \
+                    2. Make sure to select 'Additional language data' during installation\n  \
+                    3. Ensure tesseract is in your PATH", e)
+            })?;
+        
+        let text = tess.set_image(_path)
+            .map_err(|e| anyhow::anyhow!("Failed to load image '{}': {}", _path, e))?
+            .get_text()
+            .map_err(|e| anyhow::anyhow!("Failed to extract text from image: {}", e))?;
+        
+        // Get confidence (simplified - would need more complex API calls for per-word confidence)
+        let confidence = 0.85; // Placeholder
+        
+        Ok(OCRResult {
+            text,
+            confidence,
+            bounds: Rect { x: 0, y: 0, width: 0, height: 0 }, // Would need image dimensions
+        })
+    }
+    
+    async fn find_text_on_screen(&self, _text: &str) -> Result<Option<Point>> {
+        // Check if tesseract is available on the system
+        let tesseract_check = std::process::Command::new("where")
+            .arg("tesseract")
+            .output();
+        
+        if tesseract_check.is_err() || !tesseract_check.as_ref().unwrap().status.success() {
+            anyhow::bail!("Tesseract OCR is not installed on your system.\n\n\
+                To install tesseract on Windows:\n  \
+                1. Download the installer from: https://github.com/UB-Mannheim/tesseract/wiki\n  \
+                2. Run the installer and follow the instructions\n  \
+                3. Add tesseract to your PATH environment variable\n  \
+                4. Restart your terminal/command prompt\n\n\
+                After installation, restart your terminal and try again.");
+        }
+        
+        // Take full screen screenshot
+        let temp_path = format!("C:\\\\Temp\\\\g3_ocr_search_{}.png", uuid::Uuid::new_v4());
+        self.take_screenshot(&temp_path, None, None).await?;
+        
+        // Use Tesseract to find text with bounding boxes
+        let tess = Tesseract::new(None, Some("eng"))
+            .map_err(|e| {
+                anyhow::anyhow!("Failed to initialize Tesseract: {}\n\n\
+                    This usually means:\n1. Tesseract is not properly installed\n\
+                    2. Language data files are missing\n\nTo fix:\n  \
+                    1. Reinstall tesseract from https://github.com/UB-Mannheim/tesseract/wiki\n  \
+                    2. Make sure to select 'Additional language data' during installation\n  \
+                    3. Ensure tesseract is in your PATH", e)
+            })?;
+        
+        let full_text = tess.set_image(temp_path.as_str())
+            .map_err(|e| anyhow::anyhow!("Failed to load screenshot: {}", e))?
+            .get_text()
+            .map_err(|e| anyhow::anyhow!("Failed to extract text from screen: {}", e))?;
+        
+        // Clean up temp file
+        let _ = std::fs::remove_file(&temp_path);
+        
+        // Simple text search - full implementation would use get_component_images
+        // to get bounding boxes for each word
+        if full_text.contains(_text) {
+            tracing::warn!("Text found but precise coordinates not available in simplified implementation");
+            Ok(Some(Point { x: 0, y: 0 }))
+        } else {
+            Ok(None)
+        }
+    }
+}

crates/g3-computer-control/src/types.rs

@@ -0,0 +1,19 @@
+use serde::{Deserialize, Serialize};
+
+#[derive(Debug, Clone, Copy, Serialize, Deserialize)]
+pub struct Rect {
+    pub x: i32,
+    pub y: i32,
+    pub width: i32,
+    pub height: i32,
+}
+
+#[derive(Debug, Clone, Serialize, Deserialize)]
+pub struct TextLocation {
+    pub text: String,
+    pub x: i32,
+    pub y: i32,
+    pub width: i32,
+    pub height: i32,
+    pub confidence: f32,
+}

crates/g3-computer-control/src/webdriver/mod.rs

@@ -0,0 +1,111 @@
+pub mod safari;
+
+use anyhow::Result;
+use async_trait::async_trait;
+use serde_json::Value;
+
+/// WebDriver controller for browser automation
+#[async_trait]
+pub trait WebDriverController: Send + Sync {
+    /// Navigate to a URL
+    async fn navigate(&mut self, url: &str) -> Result<()>;
+    
+    /// Get the current URL
+    async fn current_url(&self) -> Result<String>;
+    
+    /// Get the page title
+    async fn title(&self) -> Result<String>;
+    
+    /// Find an element by CSS selector
+    async fn find_element(&mut self, selector: &str) -> Result<WebElement>;
+    
+    /// Find multiple elements by CSS selector
+    async fn find_elements(&mut self, selector: &str) -> Result<Vec<WebElement>>;
+    
+    /// Execute JavaScript in the browser
+    async fn execute_script(&mut self, script: &str, args: Vec<Value>) -> Result<Value>;
+    
+    /// Get the page source (HTML)
+    async fn page_source(&self) -> Result<String>;
+    
+    /// Take a screenshot and save to path
+    async fn screenshot(&mut self, path: &str) -> Result<()>;
+    
+    /// Close the current window/tab
+    async fn close(&mut self) -> Result<()>;
+    
+    /// Quit the browser session
+    async fn quit(self) -> Result<()>;
+}
+
+/// Represents a web element in the DOM
+pub struct WebElement {
+    pub(crate) inner: fantoccini::elements::Element,
+}
+
+impl WebElement {
+    /// Click the element
+    pub async fn click(&mut self) -> Result<()> {
+        self.inner.click().await?;
+        Ok(())
+    }
+    
+    /// Send keys/text to the element
+    pub async fn send_keys(&mut self, text: &str) -> Result<()> {
+        self.inner.send_keys(text).await?;
+        Ok(())
+    }
+    
+    /// Clear the element's content (for input fields)
+    pub async fn clear(&mut self) -> Result<()> {
+        self.inner.clear().await?;
+        Ok(())
+    }
+    
+    /// Get the element's text content
+    pub async fn text(&self) -> Result<String> {
+        Ok(self.inner.text().await?)
+    }
+    
+    /// Get an attribute value
+    pub async fn attr(&self, name: &str) -> Result<Option<String>> {
+        Ok(self.inner.attr(name).await?)
+    }
+    
+    /// Get a property value
+    pub async fn prop(&self, name: &str) -> Result<Option<String>> {
+        Ok(self.inner.prop(name).await?)
+    }
+    
+    /// Get the element's HTML
+    pub async fn html(&self, inner: bool) -> Result<String> {
+        Ok(self.inner.html(inner).await?)
+    }
+    
+    /// Check if element is displayed
+    pub async fn is_displayed(&self) -> Result<bool> {
+        Ok(self.inner.is_displayed().await?)
+    }
+    
+    /// Check if element is enabled
+    pub async fn is_enabled(&self) -> Result<bool> {
+        Ok(self.inner.is_enabled().await?)
+    }
+    
+    /// Check if element is selected (for checkboxes/radio buttons)
+    pub async fn is_selected(&self) -> Result<bool> {
+        Ok(self.inner.is_selected().await?)
+    }
+    
+    /// Find a child element by CSS selector
+    pub async fn find_element(&mut self, selector: &str) -> Result<WebElement> {
+        let elem = self.inner.find(fantoccini::Locator::Css(selector)).await?;
+        Ok(WebElement { inner: elem })
+    }
+    
+    /// Find multiple child elements by CSS selector
+    pub async fn find_elements(&mut self, selector: &str) -> Result<Vec<WebElement>> {
+        let elems = self.inner.find_all(fantoccini::Locator::Css(selector)).await?;
+        Ok(elems.into_iter().map(|inner| WebElement { inner }).collect())
+    }
+}

crates/g3-computer-control/src/webdriver/safari.rs

@@ -0,0 +1,212 @@
+use super::{WebDriverController, WebElement};
+use anyhow::{Context, Result};
+use async_trait::async_trait;
+use fantoccini::{Client, ClientBuilder};
+use serde_json::Value;
+use std::time::Duration;
+
+/// SafariDriver WebDriver controller
+pub struct SafariDriver {
+    client: Client,
+}
+
+impl SafariDriver {
+    /// Create a new SafariDriver instance
+    /// 
+    /// This will connect to SafariDriver running on the default port (4444).
+    /// Make sure to enable "Allow Remote Automation" in Safari's Develop menu first.
+    /// 
+    /// You can start SafariDriver manually with:
+    /// ```bash
+    /// /usr/bin/safaridriver --enable
+    /// ```
+    pub async fn new() -> Result<Self> {
+        Self::with_port(4444).await
+    }
+    
+    /// Create a new SafariDriver instance with a custom port
+    pub async fn with_port(port: u16) -> Result<Self> {
+        let url = format!("http://localhost:{}", port);
+        
+        let mut caps = serde_json::Map::new();
+        caps.insert("browserName".to_string(), Value::String("safari".to_string()));
+        
+        let client = ClientBuilder::native()
+            .capabilities(caps)
+            .connect(&url)
+            .await
+            .context("Failed to connect to SafariDriver. Make sure SafariDriver is running and 'Allow Remote Automation' is enabled in Safari's Develop menu.")?;
+        
+        Ok(Self { client })
+    }
+    
+    /// Go back in browser history
+    pub async fn back(&mut self) -> Result<()> {
+        self.client.back().await?;
+        Ok(())
+    }
+    
+    /// Go forward in browser history
+    pub async fn forward(&mut self) -> Result<()> {
+        self.client.forward().await?;
+        Ok(())
+    }
+    
+    /// Refresh the current page
+    pub async fn refresh(&mut self) -> Result<()> {
+        self.client.refresh().await?;
+        Ok(())
+    }
+    
+    /// Get all window handles
+    pub async fn window_handles(&mut self) -> Result<Vec<String>> {
+        let handles = self.client.windows().await?;
+        Ok(handles.into_iter()
+            .map(|h| h.into())
+            .collect())
+    }
+    
+    /// Switch to a window by handle
+    pub async fn switch_to_window(&mut self, handle: &str) -> Result<()> {
+        let window_handle: fantoccini::wd::WindowHandle = handle.to_string().try_into()?;
+        self.client.switch_to_window(window_handle).await?;
+        Ok(())
+    }
+    
+    /// Get the current window handle
+    pub async fn current_window_handle(&mut self) -> Result<String> {
+        Ok(self.client.window().await?.into())
+    }
+    
+    /// Close the current window
+    pub async fn close_window(&mut self) -> Result<()> {
+        self.client.close_window().await?;
+        Ok(())
+    }
+    
+    /// Create a new window/tab
+    pub async fn new_window(&mut self, is_tab: bool) -> Result<String> {
+        let window_type = if is_tab { "tab" } else { "window" };
+        let response = self.client.new_window(window_type == "tab").await?;
+        Ok(response.handle.into())
+    }
+    
+    /// Get cookies
+    pub async fn get_cookies(&mut self) -> Result<Vec<fantoccini::cookies::Cookie<'static>>> {
+        Ok(self.client.get_all_cookies().await?)
+    }
+    
+    /// Add a cookie
+    pub async fn add_cookie(&mut self, cookie: fantoccini::cookies::Cookie<'static>) -> Result<()> {
+        self.client.add_cookie(cookie).await?;
+        Ok(())
+    }
+    
+    /// Delete all cookies
+    pub async fn delete_all_cookies(&mut self) -> Result<()> {
+        self.client.delete_all_cookies().await?;
+        Ok(())
+    }
+    
+    /// Wait for an element to appear (with timeout)
+    pub async fn wait_for_element(&mut self, selector: &str, timeout: Duration) -> Result<WebElement> {
+        let start = std::time::Instant::now();
+        let poll_interval = Duration::from_millis(100);
+        
+        loop {
+            if let Ok(elem) = self.find_element(selector).await {
+                return Ok(elem);
+            }
+            
+            if start.elapsed() >= timeout {
+                anyhow::bail!("Timeout waiting for element: {}", selector);
+            }
+            
+            tokio::time::sleep(poll_interval).await;
+        }
+    }
+    
+    /// Wait for an element to be visible (with timeout)
+    pub async fn wait_for_visible(&mut self, selector: &str, timeout: Duration) -> Result<WebElement> {
+        let start = std::time::Instant::now();
+        let poll_interval = Duration::from_millis(100);
+        
+        loop {
+            if let Ok(elem) = self.find_element(selector).await {
+                if elem.is_displayed().await.unwrap_or(false) {
+                    return Ok(elem);
+                }
+            }
+            
+            if start.elapsed() >= timeout {
+                anyhow::bail!("Timeout waiting for element to be visible: {}", selector);
+            }
+            
+            tokio::time::sleep(poll_interval).await;
+        }
+    }
+}
+
+#[async_trait]
+impl WebDriverController for SafariDriver {
+    async fn navigate(&mut self, url: &str) -> Result<()> {
+        self.client.goto(url).await?;
+        Ok(())
+    }
+    
+    async fn current_url(&self) -> Result<String> {
+        Ok(self.client.current_url().await?.to_string())
+    }
+    
+    async fn title(&self) -> Result<String> {
+        Ok(self.client.title().await?)
+    }
+    
+    async fn find_element(&mut self, selector: &str) -> Result<WebElement> {
+        let elem = self.client.find(fantoccini::Locator::Css(selector)).await
+            .context(format!("Failed to find element with selector: {}", selector))?;
+        Ok(WebElement { inner: elem })
+    }
+    
+    async fn find_elements(&mut self, selector: &str) -> Result<Vec<WebElement>> {
+        let elems = self.client.find_all(fantoccini::Locator::Css(selector)).await?;
+        Ok(elems.into_iter().map(|inner| WebElement { inner }).collect())
+    }
+    
+    async fn execute_script(&mut self, script: &str, args: Vec<Value>) -> Result<Value> {
+        Ok(self.client.execute(script, args).await?)
+    }
+    
+    async fn page_source(&self) -> Result<String> {
+        Ok(self.client.source().await?)
+    }
+    
+    async fn screenshot(&mut self, path: &str) -> Result<()> {
+        let screenshot_data = self.client.screenshot().await?;
+        
+        // Expand tilde in path
+        let expanded_path = shellexpand::tilde(path);
+        let path_str = expanded_path.as_ref();
+        
+        // Create parent directories if needed
+        if let Some(parent) = std::path::Path::new(path_str).parent() {
+            std::fs::create_dir_all(parent)
+                .context("Failed to create parent directories for screenshot")?;
+        }
+        
+        std::fs::write(path_str, screenshot_data)
+            .context("Failed to write screenshot to file")?;
+        
+        Ok(())
+    }
+    
+    async fn close(&mut self) -> Result<()> {
+        self.client.close_window().await?;
+        Ok(())
+    }
+    
+    async fn quit(mut self) -> Result<()> {
+        self.client.close().await?;
+        Ok(())
+    }
+}

crates/g3-computer-control/tests/integration_test.rs

@@ -0,0 +1,17 @@
+use g3_computer_control::*;
+
+#[tokio::test]
+async fn test_screenshot() {
+    let controller = create_controller().expect("Failed to create controller");
+    
+    // Take screenshot
+    let path = "/tmp/test_screenshot.png";
+    let result = controller.take_screenshot(path, None, None).await;
+    assert!(result.is_ok(), "Failed to take screenshot: {:?}", result.err());
+    
+    // Verify file exists
+    assert!(std::path::Path::new(path).exists(), "Screenshot file was not created");
+    
+    // Clean up
+    let _ = std::fs::remove_file(path);
+}

crates/g3-computer-control/vision-bridge/Package.swift

@@ -0,0 +1,24 @@
+// swift-tools-version:5.9
+import PackageDescription
+
+let package = Package(
+    name: "VisionBridge",
+    platforms: [
+        .macOS(.v11)
+    ],
+    products: [
+        .library(
+            name: "VisionBridge",
+            type: .dynamic,
+            targets: ["VisionBridge"]
+        ),
+    ],
+    targets: [
+        .target(
+            name: "VisionBridge",
+            dependencies: [],
+            path: "Sources/VisionBridge",
+            publicHeadersPath: "."
+        ),
+    ]
+)

crates/g3-computer-control/vision-bridge/Sources/VisionBridge/VisionBridge.h

@@ -0,0 +1,39 @@
+#ifndef VisionBridge_h
+#define VisionBridge_h
+
+#include <stdint.h>
+#include <stdbool.h>
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+// Text box structure for FFI
+typedef struct {
+    const char* text;
+    uint32_t text_len;
+    int32_t x;
+    int32_t y;
+    int32_t width;
+    int32_t height;
+    float confidence;
+} VisionTextBox;
+
+// Recognize text in an image and return bounding boxes
+// Returns true on success, false on failure
+// Caller must free the returned boxes using vision_free_boxes
+bool vision_recognize_text(
+    const char* image_path,
+    uint32_t image_path_len,
+    VisionTextBox** out_boxes,
+    uint32_t* out_count
+);
+
+// Free memory allocated by vision_recognize_text
+void vision_free_boxes(VisionTextBox* boxes, uint32_t count);
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif /* VisionBridge_h */

crates/g3-computer-control/vision-bridge/Sources/VisionBridge/VisionOCR.swift

@@ -0,0 +1,145 @@
+import Foundation
+import Vision
+import AppKit
+import CoreGraphics
+
+// MARK: - C Bridge Functions
+
+@_cdecl("vision_recognize_text")
+public func vision_recognize_text(
+    _ imagePath: UnsafePointer<CChar>,
+    _ imagePathLen: UInt32,
+    _ outBoxes: UnsafeMutablePointer<UnsafeMutableRawPointer?>,
+    _ outCount: UnsafeMutablePointer<UInt32>
+) -> Bool {
+    // Convert C string to Swift String
+    guard let pathData = Data(bytes: imagePath, count: Int(imagePathLen)).withUnsafeBytes({
+        String(bytes: $0, encoding: .utf8)
+    }) else {
+        return false
+    }
+    
+    let path = pathData.trimmingCharacters(in: .whitespaces)
+    
+    // Load image
+    guard let image = NSImage(contentsOfFile: path),
+          let cgImage = image.cgImage(forProposedRect: nil, context: nil, hints: nil) else {
+        return false
+    }
+    
+    // Perform OCR
+    var textBoxes: [CTextBox] = []
+    let semaphore = DispatchSemaphore(value: 0)
+    var success = false
+    
+    let request = VNRecognizeTextRequest { request, error in
+        defer { semaphore.signal() }
+        
+        if let error = error {
+            print("Vision OCR error: \(error.localizedDescription)")
+            return
+        }
+        
+        guard let observations = request.results as? [VNRecognizedTextObservation] else {
+            return
+        }
+        
+        let imageSize = CGSize(width: cgImage.width, height: cgImage.height)
+        
+        for observation in observations {
+            guard let candidate = observation.topCandidates(1).first else { continue }
+            
+            let text = candidate.string
+            let boundingBox = observation.boundingBox
+            
+            // Convert normalized coordinates (bottom-left origin) to pixel coordinates (top-left origin)
+            let x = Int32(boundingBox.origin.x * imageSize.width)
+            let y = Int32((1.0 - boundingBox.origin.y - boundingBox.height) * imageSize.height)
+            let width = Int32(boundingBox.width * imageSize.width)
+            let height = Int32(boundingBox.height * imageSize.height)
+            
+            // Allocate C string for text
+            let cString = strdup(text)
+            
+            textBoxes.append(CTextBox(
+                text: cString,
+                text_len: UInt32(text.utf8.count),
+                x: x,
+                y: y,
+                width: width,
+                height: height,
+                confidence: observation.confidence
+            ))
+        }
+        
+        success = true
+    }
+    
+    // Configure request for best accuracy
+    request.recognitionLevel = .accurate
+    request.usesLanguageCorrection = true
+    request.recognitionLanguages = ["en-US"]
+    
+    // Perform request
+    let handler = VNImageRequestHandler(cgImage: cgImage, options: [:])
+    do {
+        try handler.perform([request])
+    } catch {
+        print("Vision request failed: \(error.localizedDescription)")
+        return false
+    }
+    
+    // Wait for completion
+    semaphore.wait()
+    
+    if !success {
+        return false
+    }
+    
+    // Allocate array for results
+    let boxesPtr = UnsafeMutablePointer<CTextBox>.allocate(capacity: textBoxes.count)
+    for (index, box) in textBoxes.enumerated() {
+        boxesPtr[index] = box
+    }
+    
+    outBoxes.pointee = UnsafeMutableRawPointer(boxesPtr)
+    outCount.pointee = UInt32(textBoxes.count)
+    
+    return true
+}
+
+@_cdecl("vision_free_boxes")
+public func vision_free_boxes(
+    _ boxes: UnsafeMutableRawPointer,
+    _ count: UInt32
+) {
+    let typedBoxes = boxes.assumingMemoryBound(to: CTextBox.self)
+    for i in 0..<Int(count) {
+        if let text = typedBoxes[i].text {
+            free(UnsafeMutableRawPointer(mutating: text))
+        }
+    }
+    typedBoxes.deallocate()
+}
+
+// MARK: - C-Compatible Structure
+
+public struct CTextBox {
+    public let text: UnsafePointer<CChar>?
+    public let text_len: UInt32
+    public let x: Int32
+    public let y: Int32
+    public let width: Int32
+    public let height: Int32
+    public let confidence: Float
+    
+    public init(text: UnsafePointer<CChar>?, text_len: UInt32, x: Int32, y: Int32, width: Int32, height: Int32, confidence: Float) {
+        self.text = text
+        self.text_len = text_len
+        self.x = x
+        self.y = y
+        self.width = width
+        self.height = height
+        self.confidence = confidence
+    }
+}

crates/g3-config/Cargo.toml

@@ -12,3 +12,6 @@ thiserror = { workspace = true }
 toml = "0.8"
 shellexpand = "3.0"
 dirs = "5.0"
+
+[dev-dependencies]
+tempfile = "3.8"

crates/g3-config/src/lib.rs

@@ -6,14 +6,21 @@ use std::path::Path;
 pub struct Config {
     pub providers: ProvidersConfig,
     pub agent: AgentConfig,
+    pub computer_control: ComputerControlConfig,
+    pub webdriver: WebDriverConfig,
+    pub macax: MacAxConfig,
 }
 
 #[derive(Debug, Clone, Serialize, Deserialize)]
 pub struct ProvidersConfig {
     pub openai: Option<OpenAIConfig>,
     pub anthropic: Option<AnthropicConfig>,
+    pub databricks: Option<DatabricksConfig>,
     pub embedded: Option<EmbeddedConfig>,
+    pub ollama: Option<OllamaConfig>,
     pub default_provider: String,
+    pub coach: Option<String>,  // Provider to use for coach in autonomous mode
+    pub player: Option<String>, // Provider to use for player in autonomous mode
 }
 
 #[derive(Debug, Clone, Serialize, Deserialize)]
@@ -33,6 +40,16 @@ pub struct AnthropicConfig {
     pub temperature: Option<f32>,
 }
 
+#[derive(Debug, Clone, Serialize, Deserialize)]
+pub struct DatabricksConfig {
+    pub host: String,
+    pub token: Option<String>, // Optional - will use OAuth if not provided
+    pub model: String,
+    pub max_tokens: Option<u32>,
+    pub temperature: Option<f32>,
+    pub use_oauth: Option<bool>, // Default to true if token not provided
+}
+
 #[derive(Debug, Clone, Serialize, Deserialize)]
 pub struct EmbeddedConfig {
     pub model_path: String,
@@ -44,11 +61,65 @@ pub struct EmbeddedConfig {
     pub threads: Option<u32>,    // Number of CPU threads to use
 }
 
+#[derive(Debug, Clone, Serialize, Deserialize)]
+pub struct OllamaConfig {
+    pub model: String,
+    pub base_url: Option<String>, // Default: http://localhost:11434
+    pub max_tokens: Option<u32>,
+    pub temperature: Option<f32>,
+}
+
 #[derive(Debug, Clone, Serialize, Deserialize)]
 pub struct AgentConfig {
     pub max_context_length: usize,
     pub enable_streaming: bool,
     pub timeout_seconds: u64,
+    pub auto_compact: bool,
+}
+
+#[derive(Debug, Clone, Serialize, Deserialize)]
+pub struct ComputerControlConfig {
+    pub enabled: bool,
+    pub require_confirmation: bool,
+    pub max_actions_per_second: u32,
+}
+
+#[derive(Debug, Clone, Serialize, Deserialize)]
+pub struct WebDriverConfig {
+    pub enabled: bool,
+    pub safari_port: u16,
+}
+
+#[derive(Debug, Clone, Serialize, Deserialize)]
+pub struct MacAxConfig {
+    pub enabled: bool,
+}
+
+impl Default for MacAxConfig {
+    fn default() -> Self {
+        Self {
+            enabled: false,
+        }
+    }
+}
+
+impl Default for WebDriverConfig {
+    fn default() -> Self {
+        Self {
+            enabled: false,
+            safari_port: 4444,
+        }
+    }
+}
+
+impl Default for ComputerControlConfig {
+    fn default() -> Self {
+        Self {
+            enabled: false, // Disabled by default for safety
+            require_confirmation: true,
+            max_actions_per_second: 5,
+        }
+    }
 }
 
 impl Default for Config {
@@ -57,14 +128,29 @@ impl Default for Config {
             providers: ProvidersConfig {
                 openai: None,
                 anthropic: None,
+                databricks: Some(DatabricksConfig {
+                    host: "https://your-workspace.cloud.databricks.com".to_string(),
+                    token: None, // Will use OAuth by default
+                    model: "databricks-claude-sonnet-4".to_string(),
+                    max_tokens: Some(4096),
+                    temperature: Some(0.1),
+                    use_oauth: Some(true),
+                }),
                 embedded: None,
-                default_provider: "anthropic".to_string(),
+                ollama: None,
+                default_provider: "databricks".to_string(),
+                coach: None,  // Will use default_provider if not specified
+                player: None, // Will use default_provider if not specified
             },
             agent: AgentConfig {
                 max_context_length: 8192,
                 enable_streaming: true,
                 timeout_seconds: 60,
+                auto_compact: true,
             },
+            computer_control: ComputerControlConfig::default(),
+            webdriver: WebDriverConfig::default(),
+            macax: MacAxConfig::default(),
         }
     }
 }
@@ -88,9 +174,9 @@ impl Config {
             })
         };
         
-        // If no config exists, create and save a default Qwen config
+        // If no config exists, create and save a default Databricks config
         if !config_exists {
-            let qwen_config = Self::default_qwen_config();
+            let databricks_config = Self::default();
             
             // Save to default location
             let config_dir = dirs::home_dir()
@@ -105,13 +191,13 @@ impl Config {
             std::fs::create_dir_all(&config_dir).ok();
             
             let config_file = config_dir.join("config.toml");
-            if let Err(e) = qwen_config.save(config_file.to_str().unwrap()) {
+            if let Err(e) = databricks_config.save(config_file.to_str().unwrap()) {
                 eprintln!("Warning: Could not save default config: {}", e);
             } else {
-                println!("Created default Qwen configuration at: {}", config_file.display());
+                println!("Created default Databricks configuration at: {}", config_file.display());
             }
             
-            return Ok(qwen_config);
+            return Ok(databricks_config);
         }
         
         // Existing config loading logic
@@ -151,12 +237,14 @@ impl Config {
         let config = settings.build()?.try_deserialize()?;
         Ok(config)
     }
-    
+
+    #[allow(dead_code)]
     fn default_qwen_config() -> Self {
         Self {
             providers: ProvidersConfig {
                 openai: None,
                 anthropic: None,
+                databricks: None,
                 embedded: Some(EmbeddedConfig {
                     model_path: "~/.cache/g3/models/qwen2.5-7b-instruct-q3_k_m.gguf".to_string(),
                     model_type: "qwen".to_string(),
@@ -166,13 +254,20 @@ impl Config {
                     gpu_layers: Some(32),
                     threads: Some(8),
                 }),
+                ollama: None,
                 default_provider: "embedded".to_string(),
+                coach: None,  // Will use default_provider if not specified
+                player: None, // Will use default_provider if not specified
             },
             agent: AgentConfig {
                 max_context_length: 8192,
                 enable_streaming: true,
                 timeout_seconds: 60,
+                auto_compact: true,
             },
+            computer_control: ComputerControlConfig::default(),
+            webdriver: WebDriverConfig::default(),
+            macax: MacAxConfig::default(),
         }
     }
     
@@ -181,4 +276,127 @@ impl Config {
         std::fs::write(path, toml_string)?;
         Ok(())
     }
+    
+    pub fn load_with_overrides(
+        config_path: Option<&str>,
+        provider_override: Option<String>,
+        model_override: Option<String>,
+    ) -> Result<Self> {
+        // Load the base configuration
+        let mut config = Self::load(config_path)?;
+        
+        // Apply provider override
+        if let Some(provider) = provider_override {
+            config.providers.default_provider = provider;
+        }
+        
+        // Apply model override to the active provider
+        if let Some(model) = model_override {
+            match config.providers.default_provider.as_str() {
+                "anthropic" => {
+                    if let Some(ref mut anthropic) = config.providers.anthropic {
+                        anthropic.model = model;
+                    } else {
+                        return Err(anyhow::anyhow!(
+                            "Provider 'anthropic' is not configured. Please add anthropic configuration to your config file."
+                        ));
+                    }
+                }
+                "databricks" => {
+                    if let Some(ref mut databricks) = config.providers.databricks {
+                        databricks.model = model;
+                    } else {
+                        return Err(anyhow::anyhow!(
+                            "Provider 'databricks' is not configured. Please add databricks configuration to your config file."
+                        ));
+                    }
+                }
+                "embedded" => {
+                    if let Some(ref mut embedded) = config.providers.embedded {
+                        embedded.model_path = model;
+                    } else {
+                        return Err(anyhow::anyhow!(
+                            "Provider 'embedded' is not configured. Please add embedded configuration to your config file."
+                        ));
+                    }
+                }
+                "openai" => {
+                    if let Some(ref mut openai) = config.providers.openai {
+                        openai.model = model;
+                    } else {
+                        return Err(anyhow::anyhow!(
+                            "Provider 'openai' is not configured. Please add openai configuration to your config file."
+                        ));
+                    }
+                }
+                _ => return Err(anyhow::anyhow!("Unknown provider: {}", 
+                    config.providers.default_provider)),
+            }
+        }
+        
+        Ok(config)
+    }
+    
+    /// Get the provider to use for coach mode in autonomous execution
+    pub fn get_coach_provider(&self) -> &str {
+        self.providers.coach
+            .as_deref()
+            .unwrap_or(&self.providers.default_provider)
+    }
+    
+    /// Get the provider to use for player mode in autonomous execution
+    pub fn get_player_provider(&self) -> &str {
+        self.providers.player
+            .as_deref()
+            .unwrap_or(&self.providers.default_provider)
+    }
+    
+    /// Create a copy of the config with a different default provider
+    pub fn with_provider_override(&self, provider: &str) -> Result<Self> {
+        // Validate that the provider is configured
+        match provider {
+            "anthropic" if self.providers.anthropic.is_none() => {
+                return Err(anyhow::anyhow!(
+                    "Provider '{}' is specified but not configured. Please add {} configuration to your config file.",
+                    provider, provider
+                ));
+            }
+            "databricks" if self.providers.databricks.is_none() => {
+                return Err(anyhow::anyhow!(
+                    "Provider '{}' is specified but not configured. Please add {} configuration to your config file.",
+                    provider, provider
+                ));
+            }
+            "embedded" if self.providers.embedded.is_none() => {
+                return Err(anyhow::anyhow!(
+                    "Provider '{}' is specified but not configured. Please add {} configuration to your config file.",
+                    provider, provider
+                ));
+            }
+            "openai" if self.providers.openai.is_none() => {
+                return Err(anyhow::anyhow!(
+                    "Provider '{}' is specified but not configured. Please add {} configuration to your config file.",
+                    provider, provider
+                ));
+            }
+            _ => {} // Provider is configured or unknown (will be caught later)
+        }
+        
+        let mut config = self.clone();
+        config.providers.default_provider = provider.to_string();
+        Ok(config)
+    }
+    
+    /// Create a copy of the config for coach mode in autonomous execution
+    pub fn for_coach(&self) -> Result<Self> {
+        self.with_provider_override(self.get_coach_provider())
+    }
+    
+    /// Create a copy of the config for player mode in autonomous execution
+    pub fn for_player(&self) -> Result<Self> {
+        self.with_provider_override(self.get_player_provider())
+    }
 }
+
+#[cfg(test)]
+mod tests;

crates/g3-config/src/tests.rs

@@ -0,0 +1,131 @@
+#[cfg(test)]
+mod tests {
+    use crate::Config;
+    use std::fs;
+    use tempfile::TempDir;
+
+    #[test]
+    fn test_coach_player_providers() {
+        // Create a temporary directory for the test config
+        let temp_dir = TempDir::new().unwrap();
+        let config_path = temp_dir.path().join("test_config.toml");
+        
+        // Write a test configuration with coach and player providers
+        let config_content = r#"
+[providers]
+default_provider = "databricks"
+coach = "anthropic"
+player = "embedded"
+
+[providers.databricks]
+host = "https://test.databricks.com"
+token = "test-token"
+model = "test-model"
+
+[providers.anthropic]
+api_key = "test-key"
+model = "claude-3"
+
+[providers.embedded]
+model_path = "test.gguf"
+model_type = "llama"
+
+[agent]
+max_context_length = 8192
+enable_streaming = true
+timeout_seconds = 60
+"#;
+        
+        fs::write(&config_path, config_content).unwrap();
+        
+        // Load the configuration
+        let config = Config::load(Some(config_path.to_str().unwrap())).unwrap();
+        
+        // Test that the providers are correctly identified
+        assert_eq!(config.providers.default_provider, "databricks");
+        assert_eq!(config.get_coach_provider(), "anthropic");
+        assert_eq!(config.get_player_provider(), "embedded");
+        
+        // Test creating coach config
+        let coach_config = config.for_coach().unwrap();
+        assert_eq!(coach_config.providers.default_provider, "anthropic");
+        
+        // Test creating player config
+        let player_config = config.for_player().unwrap();
+        assert_eq!(player_config.providers.default_provider, "embedded");
+    }
+    
+    #[test]
+    fn test_coach_player_fallback_to_default() {
+        // Create a temporary directory for the test config
+        let temp_dir = TempDir::new().unwrap();
+        let config_path = temp_dir.path().join("test_config.toml");
+        
+        // Write a test configuration WITHOUT coach and player providers
+        let config_content = r#"
+[providers]
+default_provider = "databricks"
+
+[providers.databricks]
+host = "https://test.databricks.com"
+token = "test-token"
+model = "test-model"
+
+[agent]
+max_context_length = 8192
+enable_streaming = true
+timeout_seconds = 60
+"#;
+        
+        fs::write(&config_path, config_content).unwrap();
+        
+        // Load the configuration
+        let config = Config::load(Some(config_path.to_str().unwrap())).unwrap();
+        
+        // Test that coach and player fall back to default provider
+        assert_eq!(config.get_coach_provider(), "databricks");
+        assert_eq!(config.get_player_provider(), "databricks");
+        
+        // Test creating coach config (should use default)
+        let coach_config = config.for_coach().unwrap();
+        assert_eq!(coach_config.providers.default_provider, "databricks");
+        
+        // Test creating player config (should use default)
+        let player_config = config.for_player().unwrap();
+        assert_eq!(player_config.providers.default_provider, "databricks");
+    }
+    
+    #[test]
+    fn test_invalid_provider_error() {
+        // Create a temporary directory for the test config
+        let temp_dir = TempDir::new().unwrap();
+        let config_path = temp_dir.path().join("test_config.toml");
+        
+        // Write a test configuration with an unconfigured provider
+        let config_content = r#"
+[providers]
+default_provider = "databricks"
+coach = "openai"  # OpenAI is not configured
+
+[providers.databricks]
+host = "https://test.databricks.com"
+token = "test-token"
+model = "test-model"
+
+[agent]
+max_context_length = 8192
+enable_streaming = true
+timeout_seconds = 60
+"#;
+        
+        fs::write(&config_path, config_content).unwrap();
+        
+        // Load the configuration
+        let config = Config::load(Some(config_path.to_str().unwrap())).unwrap();
+        
+        // Test that trying to create a coach config with unconfigured provider fails
+        let result = config.for_coach();
+        assert!(result.is_err());
+        assert!(result.unwrap_err().to_string().contains("not configured"));
+    }
+}

crates/g3-core/Cargo.toml

@@ -8,6 +8,7 @@ description = "Core engine for G3 AI coding agent"
 g3-providers = { path = "../g3-providers" }
 g3-config = { path = "../g3-config" }
 g3-execution = { path = "../g3-execution" }
+g3-computer-control = { path = "../g3-computer-control" }
 tokio = { workspace = true }
 reqwest = { workspace = true }
 anyhow = { workspace = true }
@@ -18,7 +19,10 @@ serde_json = { workspace = true }
 uuid = { workspace = true }
 async-trait = "0.1"
 tokio-stream = "0.1"
-llama_cpp = { version = "0.3.2", features = ["metal"] }
-shellexpand = "3.1"
 tokio-util = "0.7"
 futures-util = "0.3"
+chrono = { version = "0.4", features = ["serde"] }
+rand = "0.8"
+regex = "1.0"
+shellexpand = "3.1"
+serde_yaml = "0.9"

crates/g3-core/src/code_search.rs

@@ -0,0 +1,787 @@
+//! Code search functionality using ast-grep for syntax-aware semantic searches
+
+use anyhow::{anyhow, Result};
+use serde::{Deserialize, Serialize};
+use serde_json::Value;
+use std::collections::HashMap;
+use std::process::Stdio;
+use std::time::{Duration, Instant};
+use tokio::io::{AsyncBufReadExt, BufReader};
+use tokio::process::Command;
+use tokio::sync::Semaphore;
+use tracing::{debug, error, info, warn};
+
+/// Maximum number of searches allowed per request
+const MAX_SEARCHES: usize = 20;
+
+/// Default timeout for individual searches in seconds
+const DEFAULT_TIMEOUT_SECS: u64 = 60;
+
+/// Default maximum concurrency
+const DEFAULT_MAX_CONCURRENCY: usize = 4;
+
+/// Default maximum matches per search
+const DEFAULT_MAX_MATCHES: usize = 500;
+
+/// Search specification for a single ast-grep search
+#[derive(Debug, Clone, Deserialize)]
+pub struct SearchSpec {
+    pub name: String,
+    pub mode: SearchMode,
+    
+    // Pattern mode fields
+    pub pattern: Option<String>,
+    pub language: Option<String>,
+    
+    // YAML mode fields
+    pub rule_yaml: Option<String>,
+    
+    // Common fields
+    pub paths: Option<Vec<String>>,
+    pub globs: Option<Vec<String>>,
+    pub json_style: Option<JsonStyle>,
+    pub context: Option<u32>,
+    pub threads: Option<u32>,
+    pub include_metadata: Option<bool>,
+    pub no_ignore: Option<Vec<NoIgnoreType>>,
+    pub severity: Option<HashMap<String, SeverityLevel>>,
+    pub timeout_secs: Option<u64>,
+}
+
+/// Search mode: pattern or yaml
+#[derive(Debug, Clone, Deserialize, PartialEq)]
+#[serde(rename_all = "lowercase")]
+pub enum SearchMode {
+    Pattern,
+    Yaml,
+}
+
+/// JSON output style
+#[derive(Debug, Clone, Deserialize, PartialEq)]
+#[serde(rename_all = "lowercase")]
+pub enum JsonStyle {
+    Pretty,
+    Stream,
+    Compact,
+}
+
+impl Default for JsonStyle {
+    fn default() -> Self {
+        JsonStyle::Stream
+    }
+}
+
+/// No-ignore types
+#[derive(Debug, Clone, Deserialize, PartialEq)]
+#[serde(rename_all = "lowercase")]
+pub enum NoIgnoreType {
+    Hidden,
+    Dot,
+    Exclude,
+    Global,
+    Parent,
+    Vcs,
+}
+
+/// Severity levels for YAML rules
+#[derive(Debug, Clone, Deserialize, PartialEq)]
+#[serde(rename_all = "lowercase")]
+pub enum SeverityLevel {
+    Error,
+    Warning,
+    Info,
+    Hint,
+    Off,
+}
+
+/// Request structure for code search
+#[derive(Debug, Deserialize)]
+pub struct CodeSearchRequest {
+    pub searches: Vec<SearchSpec>,
+    pub max_concurrency: Option<usize>,
+    pub max_matches_per_search: Option<usize>,
+}
+
+/// Result of a single search
+#[derive(Debug, Serialize)]
+pub struct SearchResult {
+    pub name: String,
+    pub mode: String,
+    pub status: String,
+    pub cmd: Vec<String>,
+    pub match_count: Option<usize>,
+    pub truncated: Option<bool>,
+    pub matches: Option<Vec<Value>>,
+    pub stderr: Option<String>,
+    pub exit_code: Option<i32>,
+    pub duration_ms: u64,
+}
+
+/// Summary of all searches
+#[derive(Debug, Serialize)]
+pub struct SearchSummary {
+    pub completed: usize,
+    pub total: usize,
+    pub total_matches: usize,
+    pub duration_ms: u64,
+}
+
+/// Complete response structure
+#[derive(Debug, Serialize)]
+pub struct CodeSearchResponse {
+    pub summary: SearchSummary,
+    pub searches: Vec<SearchResult>,
+}
+
+/// YAML rule structure for validation
+#[derive(Debug, Deserialize)]
+struct YamlRule {
+    pub id: String,
+    pub language: String,
+    pub rule: Value,
+}
+
+/// Execute a batch of code searches using ast-grep
+pub async fn execute_code_search(request: CodeSearchRequest) -> Result<CodeSearchResponse> {
+    let start_time = Instant::now();
+    
+    // Validate request
+    if request.searches.is_empty() {
+        return Err(anyhow!("No searches specified"));
+    }
+    
+    if request.searches.len() > MAX_SEARCHES {
+        return Err(anyhow!(
+            "Too many searches: {} (max: {})",
+            request.searches.len(),
+            MAX_SEARCHES
+        ));
+    }
+    
+    // Check if ast-grep is available
+    check_ast_grep_available().await?;
+    
+    let max_concurrency = request.max_concurrency.unwrap_or(DEFAULT_MAX_CONCURRENCY);
+    let max_matches = request.max_matches_per_search.unwrap_or(DEFAULT_MAX_MATCHES);
+    
+    // Create semaphore for concurrency control
+    let semaphore = std::sync::Arc::new(Semaphore::new(max_concurrency));
+    
+    // Execute searches concurrently
+    let mut tasks = Vec::new();
+
+    for search in request.searches {
+        let sem = semaphore.clone();
+        let task = tokio::spawn(async move {
+            let _permit = sem.acquire().await.unwrap();
+            execute_single_search(search, max_matches).await
+        });
+        tasks.push(task);
+    }
+    
+    // Wait for all searches to complete
+    let mut results = Vec::new();
+    let mut total_matches = 0;
+    let mut completed = 0;
+    
+    for task in tasks {
+        match task.await {
+            Ok(result) => {
+                if result.status == "ok" {
+                    completed += 1;
+                    if let Some(count) = result.match_count {
+                        total_matches += count;
+                    }
+                }
+                results.push(result);
+            }
+            Err(e) => {
+                error!("Task join error: {}", e);
+                // Create an error result
+                results.push(SearchResult {
+                    name: "unknown".to_string(),
+                    mode: "unknown".to_string(),
+                    status: "error".to_string(),
+                    cmd: vec![],
+                    match_count: None,
+                    truncated: None,
+                    matches: None,
+                    stderr: Some(format!("Task execution error: {}", e)),
+                    exit_code: None,
+                    duration_ms: 0,
+                });
+            }
+        }
+    }
+    
+    let total_duration = start_time.elapsed();
+    
+    Ok(CodeSearchResponse {
+        summary: SearchSummary {
+            completed,
+            total: results.len(),
+            total_matches,
+            duration_ms: total_duration.as_millis() as u64,
+        },
+        searches: results,
+    })
+}
+
+/// Execute a single search
+async fn execute_single_search(search: SearchSpec, max_matches: usize) -> SearchResult {
+    let start_time = Instant::now();
+    let timeout_secs = search.timeout_secs.unwrap_or(DEFAULT_TIMEOUT_SECS);
+    
+    // Validate the search specification
+    if let Err(e) = validate_search_spec(&search) {
+        return SearchResult {
+            name: search.name,
+            mode: format!("{:?}", search.mode).to_lowercase(),
+            status: "error".to_string(),
+            cmd: vec![],
+            match_count: None,
+            truncated: None,
+            matches: None,
+            stderr: Some(format!("Validation error: {}", e)),
+            exit_code: None,
+            duration_ms: start_time.elapsed().as_millis() as u64,
+        };
+    }
+    
+    // Build command
+    let cmd_args = match build_ast_grep_command(&search) {
+        Ok(args) => args,
+        Err(e) => {
+            return SearchResult {
+                name: search.name,
+                mode: format!("{:?}", search.mode).to_lowercase(),
+                status: "error".to_string(),
+                cmd: vec![],
+                match_count: None,
+                truncated: None,
+                matches: None,
+                stderr: Some(format!("Command build error: {}", e)),
+                exit_code: None,
+                duration_ms: start_time.elapsed().as_millis() as u64,
+            };
+        }
+    };
+    
+    debug!("Executing ast-grep command: {:?}", cmd_args);
+
+    // Execute with timeout
+    let timeout_duration = Duration::from_secs(timeout_secs);
+    
+    match tokio::time::timeout(timeout_duration, run_ast_grep_command(&cmd_args)).await {
+        Ok(Ok((stdout, stderr, exit_code))) => {
+            let duration_ms = start_time.elapsed().as_millis() as u64;
+            
+            if exit_code == 0 {
+                // Parse JSON output
+                match parse_ast_grep_output(&stdout, max_matches) {
+                    Ok((matches, truncated)) => {
+                        SearchResult {
+                            name: search.name,
+                            mode: format!("{:?}", search.mode).to_lowercase(),
+                            status: "ok".to_string(),
+                            cmd: cmd_args,
+                            match_count: Some(matches.len()),
+                            truncated: Some(truncated),
+                            matches: Some(matches),
+                            stderr: if stderr.is_empty() { None } else { Some(stderr) },
+                            exit_code: None,
+                            duration_ms,
+                        }
+                    }
+                    Err(e) => {
+                        SearchResult {
+                            name: search.name,
+                            mode: format!("{:?}", search.mode).to_lowercase(),
+                            status: "error".to_string(),
+                            cmd: cmd_args,
+                            match_count: None,
+                            truncated: None,
+                            matches: None,
+                            stderr: Some(format!("JSON parse error: {}\nRaw output: {}", e, stdout)),
+                            exit_code: Some(exit_code),
+                            duration_ms,
+                        }
+                    }
+                }
+            } else {
+                SearchResult {
+                    name: search.name,
+                    mode: format!("{:?}", search.mode).to_lowercase(),
+                    status: "error".to_string(),
+                    cmd: cmd_args,
+                    match_count: None,
+                    truncated: None,
+                    matches: None,
+                    stderr: Some(stderr),
+                    exit_code: Some(exit_code),
+                    duration_ms,
+                }
+            }
+        }
+        Ok(Err(e)) => {
+            SearchResult {
+                name: search.name,
+                mode: format!("{:?}", search.mode).to_lowercase(),
+                status: "error".to_string(),
+                cmd: cmd_args,
+                match_count: None,
+                truncated: None,
+                matches: None,
+                stderr: Some(format!("Execution error: {}", e)),
+                exit_code: None,
+                duration_ms: start_time.elapsed().as_millis() as u64,
+            }
+        }
+        Err(_) => {
+            SearchResult {
+                name: search.name,
+                mode: format!("{:?}", search.mode).to_lowercase(),
+                status: "timeout".to_string(),
+                cmd: cmd_args,
+                match_count: None,
+                truncated: None,
+                matches: None,
+                stderr: Some(format!("Search timed out after {} seconds", timeout_secs)),
+                exit_code: None,
+                duration_ms: start_time.elapsed().as_millis() as u64,
+            }
+        }
+    }
+}
+
+/// Validate a search specification
+fn validate_search_spec(search: &SearchSpec) -> Result<()> {
+    match search.mode {
+        SearchMode::Pattern => {
+            if search.pattern.is_none() || search.pattern.as_ref().unwrap().is_empty() {
+                return Err(anyhow!("Pattern mode requires non-empty 'pattern' field"));
+            }
+        }
+        SearchMode::Yaml => {
+            let rule_yaml = search.rule_yaml.as_ref()
+                .ok_or_else(|| anyhow!("YAML mode requires 'rule_yaml' field"))?;
+            
+            if rule_yaml.is_empty() {
+                return Err(anyhow!("YAML mode requires non-empty 'rule_yaml' field"));
+            }
+            
+            // Parse and validate YAML structure
+            let parsed: YamlRule = serde_yaml::from_str(rule_yaml)
+                .map_err(|e| anyhow!("Invalid YAML rule: {}", e))?;
+            
+            if parsed.id.is_empty() {
+                return Err(anyhow!("YAML rule must have non-empty 'id' field"));
+            }
+            
+            if parsed.language.is_empty() {
+                return Err(anyhow!("YAML rule must have non-empty 'language' field"));
+            }
+            
+            // Validate language is supported (basic check)
+            validate_language(&parsed.language)?;
+        }
+    }
+    
+    // Validate context range
+    if let Some(context) = search.context {
+        if context > 20 {
+            return Err(anyhow!("Context lines cannot exceed 20"));
+        }
+    }
+    
+    Ok(())
+}
+
+/// Validate that a language is supported by ast-grep
+fn validate_language(language: &str) -> Result<()> {
+    let supported_languages = [
+        "rust", "javascript", "typescript", "python", "java", "c", "cpp", "csharp",
+        "go", "html", "css", "json", "yaml", "xml", "bash", "kotlin", "swift",
+        "php", "ruby", "scala", "dart", "lua", "r", "sql", "dockerfile",
+        "Rust", "JavaScript", "TypeScript", "Python", "Java", "C", "Cpp", "CSharp",
+        "Go", "Html", "Css", "Json", "Yaml", "Xml", "Bash", "Kotlin", "Swift",
+        "Php", "Ruby", "Scala", "Dart", "Lua", "R", "Sql", "Dockerfile"
+    ];
+    
+    if !supported_languages.contains(&language) {
+        warn!("Language '{}' may not be supported by ast-grep", language);
+    }
+    
+    Ok(())
+}
+
+/// Build ast-grep command arguments
+fn build_ast_grep_command(search: &SearchSpec) -> Result<Vec<String>> {
+    let mut args = vec!["ast-grep".to_string()];
+    
+    match search.mode {
+        SearchMode::Pattern => {
+            args.push("run".to_string());
+            
+            // Add pattern
+            args.push("-p".to_string());
+            args.push(search.pattern.as_ref().unwrap().clone());
+            
+            // Add language if specified
+            if let Some(ref lang) = search.language {
+                args.push("-l".to_string());
+                args.push(lang.clone());
+            }
+        }
+        SearchMode::Yaml => {
+            args.push("scan".to_string());
+            
+            // Add inline rules
+            args.push("--inline-rules".to_string());
+            args.push(search.rule_yaml.as_ref().unwrap().clone());
+            
+            // Add include-metadata if requested
+            if search.include_metadata.unwrap_or(false) {
+                args.push("--include-metadata".to_string());
+            }
+            
+            // Add severity overrides
+            if let Some(ref severity_map) = search.severity {
+                for (rule_id, severity) in severity_map {
+                    match severity {
+                        SeverityLevel::Error => {
+                            args.push("--error".to_string());
+                            args.push(rule_id.clone());
+                        }
+                        SeverityLevel::Warning => {
+                            args.push("--warning".to_string());
+                            args.push(rule_id.clone());
+                        }
+                        SeverityLevel::Info => {
+                            args.push("--info".to_string());
+                            args.push(rule_id.clone());
+                        }
+                        SeverityLevel::Hint => {
+                            args.push("--hint".to_string());
+                            args.push(rule_id.clone());
+                        }
+                        SeverityLevel::Off => {
+                            args.push("--off".to_string());
+                            args.push(rule_id.clone());
+                        }
+                    }
+                }
+            }
+        }
+    }
+    
+    // Add common arguments
+    
+    // Add globs if specified
+    if let Some(ref globs) = search.globs {
+        if !globs.is_empty() {
+            args.push("--globs".to_string());
+            args.push(globs.join(","));
+        }
+    }
+    
+    // Add context
+    if let Some(context) = search.context {
+        args.push("-C".to_string());
+        args.push(context.to_string());
+    }
+    
+    // Add threads
+    if let Some(threads) = search.threads {
+        args.push("-j".to_string());
+        args.push(threads.to_string());
+    }
+    
+    // Add JSON output style
+    let json_style = search.json_style.as_ref().unwrap_or(&JsonStyle::Stream);
+    let json_arg = match json_style {
+        JsonStyle::Pretty => "--json=pretty",
+        JsonStyle::Stream => "--json=stream",
+        JsonStyle::Compact => "--json=compact",
+    };
+    args.push(json_arg.to_string());
+    
+    // Add no-ignore options
+    if let Some(ref no_ignore_list) = search.no_ignore {
+        for no_ignore_type in no_ignore_list {
+            let flag = match no_ignore_type {
+                NoIgnoreType::Hidden => "--no-ignore=hidden",
+                NoIgnoreType::Dot => "--no-ignore=dot",
+                NoIgnoreType::Exclude => "--no-ignore=exclude",
+                NoIgnoreType::Global => "--no-ignore=global",
+                NoIgnoreType::Parent => "--no-ignore=parent",
+                NoIgnoreType::Vcs => "--no-ignore=vcs",
+            };
+            args.push(flag.to_string());
+        }
+    }
+    
+    // Add paths (default to current directory if none specified)
+    if let Some(ref paths) = search.paths {
+        if !paths.is_empty() {
+            args.extend(paths.clone());
+        } else {
+            args.push(".".to_string());
+        }
+    } else {
+        args.push(".".to_string());
+    }
+    
+    Ok(args)
+}
+
+/// Run ast-grep command and capture output
+async fn run_ast_grep_command(args: &[String]) -> Result<(String, String, i32)> {
+    let mut cmd = Command::new(&args[0]);
+    cmd.args(&args[1..]);
+    cmd.stdout(Stdio::piped());
+    cmd.stderr(Stdio::piped());
+    
+    debug!("Running command: {:?}", args);
+    
+    let mut child = cmd.spawn()
+        .map_err(|e| anyhow!("Failed to spawn ast-grep process: {}", e))?;
+    
+    let stdout = child.stdout.take().unwrap();
+    let stderr = child.stderr.take().unwrap();
+    
+    let stdout_reader = BufReader::new(stdout);
+    let stderr_reader = BufReader::new(stderr);
+    
+    let stdout_task = tokio::spawn(async move {
+        let mut lines = stdout_reader.lines();
+        let mut output = String::new();
+        while let Ok(Some(line)) = lines.next_line().await {
+            if !output.is_empty() {
+                output.push('\n');
+            }
+            output.push_str(&line);
+        }
+        output
+    });
+    
+    let stderr_task = tokio::spawn(async move {
+        let mut lines = stderr_reader.lines();
+        let mut output = String::new();
+        while let Ok(Some(line)) = lines.next_line().await {
+            if !output.is_empty() {
+                output.push('\n');
+            }
+            output.push_str(&line);
+        }
+        output
+    });
+    
+    let status = child.wait().await
+        .map_err(|e| anyhow!("Failed to wait for ast-grep process: {}", e))?;
+    
+    let stdout_output = stdout_task.await
+        .map_err(|e| anyhow!("Failed to read stdout: {}", e))?;
+    let stderr_output = stderr_task.await
+        .map_err(|e| anyhow!("Failed to read stderr: {}", e))?;
+    
+    let exit_code = status.code().unwrap_or(-1);
+    
+    Ok((stdout_output, stderr_output, exit_code))
+}
+
+/// Parse ast-grep JSON output
+fn parse_ast_grep_output(output: &str, max_matches: usize) -> Result<(Vec<Value>, bool)> {
+    if output.trim().is_empty() {
+        return Ok((vec![], false));
+    }
+    
+    let mut matches = Vec::new();
+    let mut truncated = false;
+    
+    // Handle stream format (line-delimited JSON)
+    for line in output.lines() {
+        let line = line.trim();
+        if line.is_empty() {
+            continue;
+        }
+        
+        match serde_json::from_str::<Value>(line) {
+            Ok(match_obj) => {
+                if matches.len() >= max_matches {
+                    truncated = true;
+                    break;
+                }
+                matches.push(match_obj);
+            }
+            Err(e) => {
+                debug!("Failed to parse JSON line '{}': {}", line, e);
+                // Try to parse the entire output as a single JSON array
+                match serde_json::from_str::<Vec<Value>>(output) {
+                    Ok(array_matches) => {
+                        let take_count = array_matches.len().min(max_matches);
+                        let total_count = array_matches.len();
+                        matches = array_matches.into_iter().take(take_count).collect();
+                        truncated = take_count < total_count;
+                        break;
+                    }
+                    Err(e2) => {
+                        return Err(anyhow!(
+                            "Failed to parse ast-grep output as line-delimited JSON or JSON array. Line error: {}, Array error: {}",
+                            e, e2
+                        ));
+                    }
+                }
+            }
+        }
+    }
+    
+    Ok((matches, truncated))
+}
+
+/// Check if ast-grep is available and provide installation hints if not
+async fn check_ast_grep_available() -> Result<()> {
+    match Command::new("ast-grep")
+        .arg("--version")
+        .output()
+        .await
+    {
+        Ok(output) => {
+            if output.status.success() {
+                let version = String::from_utf8_lossy(&output.stdout);
+                info!("Found ast-grep: {}", version.trim());
+                Ok(())
+            } else {
+                Err(anyhow!("ast-grep command failed: {}", String::from_utf8_lossy(&output.stderr)))
+            }
+        }
+        Err(_) => {
+            Err(anyhow!(
+                "ast-grep not found. Please install it using one of these methods:\n\n\
+                • Homebrew (macOS): brew install ast-grep\n\
+                • MacPorts (macOS): sudo port install ast-grep\n\
+                • Nix: nix-env -iA nixpkgs.ast-grep\n\
+                • Cargo: cargo install ast-grep\n\
+                • npm: npm install -g @ast-grep/cli\n\
+                • pip: pip install ast-grep\n\n\
+                For more installation options, visit: https://ast-grep.github.io/guide/quick-start.html"
+            ))
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    
+    #[test]
+    fn test_validate_pattern_search() {
+        let search = SearchSpec {
+            name: "test".to_string(),
+            mode: SearchMode::Pattern,
+            pattern: Some("fn $NAME() {}".to_string()),
+            language: Some("rust".to_string()),
+            rule_yaml: None,
+            paths: None,
+            globs: None,
+            json_style: None,
+            context: None,
+            threads: None,
+            include_metadata: None,
+            no_ignore: None,
+            severity: None,
+            timeout_secs: None,
+        };
+        
+        assert!(validate_search_spec(&search).is_ok());
+    }
+    
+    #[test]
+    fn test_validate_yaml_search() {
+        let yaml_rule = r#"
+id: test-rule
+language: Rust
+rule:
+  pattern: "fn $NAME() {}"
+"#;
+        
+        let search = SearchSpec {
+            name: "test".to_string(),
+            mode: SearchMode::Yaml,
+            pattern: None,
+            language: None,
+            rule_yaml: Some(yaml_rule.to_string()),
+            paths: None,
+            globs: None,
+            json_style: None,
+            context: None,
+            threads: None,
+            include_metadata: None,
+            no_ignore: None,
+            severity: None,
+            timeout_secs: None,
+        };
+        
+        assert!(validate_search_spec(&search).is_ok());
+    }
+    
+    #[test]
+    fn test_build_pattern_command() {
+        let search = SearchSpec {
+            name: "test".to_string(),
+            mode: SearchMode::Pattern,
+            pattern: Some("fn $NAME() {}".to_string()),
+            language: Some("rust".to_string()),
+            rule_yaml: None,
+            paths: Some(vec!["src/".to_string()]),
+            globs: None,
+            json_style: Some(JsonStyle::Stream),
+            context: Some(2),
+            threads: Some(4),
+            include_metadata: None,
+            no_ignore: None,
+            severity: None,
+            timeout_secs: None,
+        };
+        
+        let cmd = build_ast_grep_command(&search).unwrap();
+        
+        assert_eq!(cmd[0], "ast-grep");
+        assert_eq!(cmd[1], "run");
+        assert!(cmd.contains(&"-p".to_string()));
+        assert!(cmd.contains(&"fn $NAME() {}".to_string()));
+        assert!(cmd.contains(&"-l".to_string()));
+        assert!(cmd.contains(&"rust".to_string()));
+        assert!(cmd.contains(&"--json=stream".to_string()));
+        assert!(cmd.contains(&"-C".to_string()));
+        assert!(cmd.contains(&"2".to_string()));
+        assert!(cmd.contains(&"-j".to_string()));
+        assert!(cmd.contains(&"4".to_string()));
+        assert!(cmd.contains(&"src/".to_string()));
+    }
+    
+    #[test]
+    fn test_parse_stream_json() {
+        let output = r#"{"file":"test.rs","text":"fn hello() {}"}
+{"file":"test2.rs","text":"fn world() {}"}"#;
+        
+        let (matches, truncated) = parse_ast_grep_output(output, 10).unwrap();
+        
+        assert_eq!(matches.len(), 2);
+        assert!(!truncated);
+        assert_eq!(matches[0]["file"], "test.rs");
+        assert_eq!(matches[1]["file"], "test2.rs");
+    }
+    
+    #[test]
+    fn test_parse_truncated_output() {
+        let output = r#"{"file":"test1.rs","text":"fn a() {}"}
+{"file":"test2.rs","text":"fn b() {}"}
+{"file":"test3.rs","text":"fn c() {}"}"#;
+        
+        let (matches, truncated) = parse_ast_grep_output(output, 2).unwrap();
+        
+        assert_eq!(matches.len(), 2);
+        assert!(truncated);
+    }
+}

crates/g3-core/src/error_handling.rs

@@ -0,0 +1,501 @@
+//! Error handling module for G3 with retry logic and detailed logging
+//!
+//! This module provides:
+//! - Classification of errors as recoverable or non-recoverable
+//! - Retry logic with exponential backoff and jitter for recoverable errors
+//! - Detailed error logging with context information
+//! - Request/response capture for debugging
+
+use anyhow::Result;
+use serde::{Deserialize, Serialize};
+use std::time::Duration;
+use tracing::{error, info, warn};
+
+/// Maximum number of retry attempts for recoverable errors (default mode)
+const DEFAULT_MAX_RETRY_ATTEMPTS: u32 = 3;
+
+/// Maximum number of retry attempts for autonomous mode
+const AUTONOMOUS_MAX_RETRY_ATTEMPTS: u32 = 6;
+
+/// Base delay for exponential backoff (in milliseconds)
+const BASE_RETRY_DELAY_MS: u64 = 1000;
+
+/// Maximum delay between retries (in milliseconds) for default mode
+const DEFAULT_MAX_RETRY_DELAY_MS: u64 = 10000;
+
+/// Maximum delay between retries (in milliseconds) for autonomous mode
+/// Spread over 10 minutes (600 seconds) with 6 retries
+const AUTONOMOUS_MAX_RETRY_DELAY_MS: u64 = 120000; // 2 minutes max per retry
+
+// Removed unused constants AUTONOMOUS_RETRY_BUDGET_MS and DEFAULT_JITTER_FACTOR
+
+/// Jitter factor for autonomous mode (higher for better distribution)
+const JITTER_FACTOR: f64 = 0.3;
+
+/// Error context information for detailed logging
+#[derive(Debug, Clone, Serialize, Deserialize)]
+pub struct ErrorContext {
+    /// The operation that was being performed
+    pub operation: String,
+    /// The provider being used
+    pub provider: String,
+    /// The model being used
+    pub model: String,
+    /// The last prompt sent (truncated for logging)
+    pub last_prompt: String,
+    /// Raw request data (if available)
+    pub raw_request: Option<String>,
+    /// Raw response data (if available)
+    pub raw_response: Option<String>,
+    /// Stack trace
+    pub stack_trace: String,
+    /// Timestamp
+    pub timestamp: u64,
+    /// Number of tokens in context
+    pub context_tokens: u32,
+    /// Session ID if available
+    pub session_id: Option<String>,
+    /// Whether to skip file logging (quiet mode)
+    pub quiet: bool,
+}
+
+impl ErrorContext {
+    pub fn new(
+        operation: String,
+        provider: String,
+        model: String,
+        last_prompt: String,
+        session_id: Option<String>,
+        context_tokens: u32,
+        quiet: bool,
+    ) -> Self {
+        let timestamp = std::time::SystemTime::now()
+            .duration_since(std::time::UNIX_EPOCH)
+            .unwrap_or_default()
+            .as_secs();
+
+        // Capture stack trace
+        let stack_trace = std::backtrace::Backtrace::force_capture().to_string();
+
+        Self {
+            operation,
+            provider,
+            model,
+            last_prompt: truncate_for_logging(&last_prompt, 1000),
+            raw_request: None,
+            raw_response: None,
+            stack_trace,
+            timestamp,
+            context_tokens,
+            session_id,
+            quiet,
+        }
+    }
+
+    pub fn with_request(mut self, request: String) -> Self {
+        self.raw_request = Some(truncate_for_logging(&request, 5000));
+        self
+    }
+
+    pub fn with_response(mut self, response: String) -> Self {
+        self.raw_response = Some(truncate_for_logging(&response, 5000));
+        self
+    }
+
+    /// Log the error context with ERROR level
+    pub fn log_error(&self, error: &anyhow::Error) {
+        error!("=== G3 ERROR DETAILS ===");
+        error!("Operation: {}", self.operation);
+        error!("Provider: {} | Model: {}", self.provider, self.model);
+        error!("Error: {}", error);
+        error!("Timestamp: {}", self.timestamp);
+        error!("Session ID: {:?}", self.session_id);
+        error!("Context Tokens: {}", self.context_tokens);
+        error!("Last Prompt: {}", self.last_prompt);
+        
+        if let Some(ref req) = self.raw_request {
+            error!("Raw Request: {}", req);
+        }
+        
+        if let Some(ref resp) = self.raw_response {
+            error!("Raw Response: {}", resp);
+        }
+        
+        error!("Stack Trace:\n{}", self.stack_trace);
+        error!("=== END ERROR DETAILS ===");
+
+        // Also save to error log file
+        self.save_to_file();
+    }
+
+    /// Save error context to a file for later analysis
+    fn save_to_file(&self) {
+        // Skip file logging if quiet mode is enabled
+        if self.quiet {
+            return;
+        }
+
+        let logs_dir = std::path::Path::new("logs/errors");
+        if !logs_dir.exists() {
+            if let Err(e) = std::fs::create_dir_all(logs_dir) {
+                error!("Failed to create error logs directory: {}", e);
+                return;
+            }
+        }
+
+        let filename = format!(
+            "logs/errors/error_{}_{}.json",
+            self.timestamp,
+            self.session_id.as_deref().unwrap_or("unknown")
+        );
+
+        match serde_json::to_string_pretty(self) {
+            Ok(json_content) => {
+                if let Err(e) = std::fs::write(&filename, json_content) {
+                    error!("Failed to save error context to {}: {}", filename, e);
+                } else {
+                    info!("Error details saved to: {}", filename);
+                }
+            }
+            Err(e) => {
+                error!("Failed to serialize error context: {}", e);
+            }
+        }
+    }
+}
+
+/// Classification of error types
+#[derive(Debug, Clone, PartialEq)]
+pub enum ErrorType {
+    /// Recoverable errors that should be retried
+    Recoverable(RecoverableError),
+    /// Non-recoverable errors that should terminate execution
+    NonRecoverable,
+}
+
+/// Types of recoverable errors
+#[derive(Debug, Clone, PartialEq)]
+pub enum RecoverableError {
+    /// Rate limit exceeded
+    RateLimit,
+    /// Temporary network error
+    NetworkError,
+    /// Server error (5xx)
+    ServerError,
+    /// Model is busy/overloaded
+    ModelBusy,
+    /// Timeout
+    Timeout,
+    /// Token limit exceeded (might be recoverable with summarization)
+    TokenLimit,
+}
+
+/// Classify an error as recoverable or non-recoverable
+pub fn classify_error(error: &anyhow::Error) -> ErrorType {
+    let error_str = error.to_string().to_lowercase();
+
+    // Check for recoverable error patterns
+    if error_str.contains("rate limit") || error_str.contains("rate_limit") || error_str.contains("429") {
+        return ErrorType::Recoverable(RecoverableError::RateLimit);
+    }
+
+    if error_str.contains("network") || error_str.contains("connection") || 
+       error_str.contains("dns") || error_str.contains("refused") {
+        return ErrorType::Recoverable(RecoverableError::NetworkError);
+    }
+
+    if error_str.contains("500") || error_str.contains("502") || 
+       error_str.contains("503") || error_str.contains("504") ||
+       error_str.contains("server error") || error_str.contains("internal error") {
+        return ErrorType::Recoverable(RecoverableError::ServerError);
+    }
+
+    if error_str.contains("busy") || error_str.contains("overloaded") || 
+       error_str.contains("capacity") || error_str.contains("unavailable") {
+        return ErrorType::Recoverable(RecoverableError::ModelBusy);
+    }
+
+    // Enhanced timeout detection - check for various timeout patterns
+    if error_str.contains("timeout") || 
+       error_str.contains("timed out") || 
+       error_str.contains("operation timed out") ||
+       error_str.contains("request or response body error") ||  // Common timeout pattern
+       error_str.contains("stream error") && error_str.contains("timed out") {
+        return ErrorType::Recoverable(RecoverableError::Timeout);
+    }
+
+    if error_str.contains("token") && (error_str.contains("limit") || error_str.contains("exceeded")) {
+        return ErrorType::Recoverable(RecoverableError::TokenLimit);
+    }
+
+    // Default to non-recoverable for unknown errors
+    ErrorType::NonRecoverable
+}
+
+/// Calculate retry delay for autonomous mode with better distribution over 10 minutes
+fn calculate_autonomous_retry_delay(attempt: u32) -> Duration {
+    use rand::Rng;
+    let mut rng = rand::thread_rng();
+    
+    // Distribute 6 retries over 10 minutes (600 seconds)
+    // Base delays: 10s, 30s, 60s, 120s, 180s, 200s = 600s total
+    let base_delays_ms = [10000, 30000, 60000, 120000, 180000, 200000];
+    let base_delay = base_delays_ms.get(attempt.saturating_sub(1) as usize).unwrap_or(&200000);
+    
+    // Add jitter of ±30% to prevent thundering herd
+    let jitter = (*base_delay as f64 * 0.3 * rng.gen::<f64>()) as u64;
+    let final_delay = if rng.gen_bool(0.5) {
+        base_delay + jitter
+    } else {
+        base_delay.saturating_sub(jitter)
+    };
+    
+    Duration::from_millis(final_delay)
+}
+
+/// Calculate retry delay with exponential backoff and jitter
+pub fn calculate_retry_delay(attempt: u32, is_autonomous: bool) -> Duration {
+    if is_autonomous {
+        return calculate_autonomous_retry_delay(attempt);
+    }
+    
+    use rand::Rng;
+    let max_retry_delay_ms = if is_autonomous { AUTONOMOUS_MAX_RETRY_DELAY_MS } else { DEFAULT_MAX_RETRY_DELAY_MS };
+    
+    // Exponential backoff: delay = base * 2^attempt
+    let base_delay = BASE_RETRY_DELAY_MS * (2_u64.pow(attempt.saturating_sub(1)));
+    let capped_delay = base_delay.min(max_retry_delay_ms);
+    
+    // Add jitter to prevent thundering herd
+    let mut rng = rand::thread_rng();
+    let jitter = (capped_delay as f64 * JITTER_FACTOR * rng.gen::<f64>()) as u64;
+    let final_delay = if rng.gen_bool(0.5) {
+        capped_delay + jitter
+    } else {
+        capped_delay.saturating_sub(jitter)
+    };
+    
+    Duration::from_millis(final_delay)
+}
+
+/// Retry logic for async operations
+pub async fn retry_with_backoff<F, Fut, T>(
+    operation_name: &str,
+    mut operation: F,
+    context: &ErrorContext,
+    is_autonomous: bool,
+) -> Result<T>
+where
+    F: FnMut() -> Fut,
+    Fut: std::future::Future<Output = Result<T>>,
+{
+    let mut attempt = 0;
+    let mut _last_error = None;
+
+    loop {
+        attempt += 1;
+        
+        match operation().await {
+            Ok(result) => {
+                if attempt > 1 {
+                    info!(
+                        "Operation '{}' succeeded after {} attempts",
+                        operation_name, attempt
+                    );
+                }
+                return Ok(result);
+            }
+            Err(error) => {
+                let error_type = classify_error(&error);
+                let max_attempts = if is_autonomous { AUTONOMOUS_MAX_RETRY_ATTEMPTS } else { DEFAULT_MAX_RETRY_ATTEMPTS };
+                
+                match error_type {
+                    ErrorType::Recoverable(recoverable_type) => {
+                        if attempt >= max_attempts {
+                            error!(
+                                "Operation '{}' failed after {} attempts. Giving up.",
+                                operation_name, attempt
+                            );
+                            context.clone().log_error(&error);
+                            return Err(error);
+                        }
+                        
+                        let delay = calculate_retry_delay(attempt, is_autonomous);
+                        warn!(
+                            "Recoverable error ({:?}) in '{}' (attempt {}/{}). Retrying in {:?}...",
+                            recoverable_type, operation_name, attempt, max_attempts, delay
+                        );
+                        warn!("Error details: {}", error);
+                        
+                        // Special handling for token limit errors
+                        if matches!(recoverable_type, RecoverableError::TokenLimit) {
+                            info!("Token limit error detected. Consider triggering summarization.");
+                        }
+                        
+                        tokio::time::sleep(delay).await;
+                        _last_error = Some(error);
+                    }
+                    ErrorType::NonRecoverable => {
+                        error!(
+                            "Non-recoverable error in '{}' (attempt {}). Terminating.",
+                            operation_name, attempt
+                        );
+                        context.clone().log_error(&error);
+                        return Err(error);
+                    }
+                }
+            }
+        }
+    }
+}
+
+/// Helper function to truncate strings for logging
+fn truncate_for_logging(s: &str, max_len: usize) -> String {
+    if s.len() <= max_len {
+        s.to_string()
+    } else {
+        // Find a safe UTF-8 boundary to truncate at
+        // We need to ensure we don't cut in the middle of a multi-byte character
+        let mut truncate_at = max_len;
+        
+        // Walk backwards from max_len to find a character boundary
+        while truncate_at > 0 && !s.is_char_boundary(truncate_at) {
+            truncate_at -= 1;
+        }
+        
+        // If we couldn't find a boundary (shouldn't happen), use a safe default
+        if truncate_at == 0 {
+            truncate_at = max_len.min(s.len());
+        }
+        
+        format!("{}... (truncated, {} total bytes)", &s[..truncate_at], s.len())
+    }
+}
+
+/// Macro for creating error context easily
+#[macro_export]
+macro_rules! error_context {
+    ($operation:expr, $provider:expr, $model:expr, $prompt:expr, $session_id:expr, $tokens:expr) => {
+        $crate::error_handling::ErrorContext::new(
+            $operation.to_string(),
+            $provider.to_string(),
+            $model.to_string(),
+            $prompt.to_string(),
+            $session_id,
+            $tokens,
+        )
+    };
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use anyhow::anyhow;
+
+    #[test]
+    fn test_error_classification() {
+        // Rate limit errors
+        let error = anyhow!("Rate limit exceeded");
+        assert_eq!(classify_error(&error), ErrorType::Recoverable(RecoverableError::RateLimit));
+        
+        let error = anyhow!("HTTP 429 Too Many Requests");
+        assert_eq!(classify_error(&error), ErrorType::Recoverable(RecoverableError::RateLimit));
+        
+        // Network errors
+        let error = anyhow!("Network connection failed");
+        assert_eq!(classify_error(&error), ErrorType::Recoverable(RecoverableError::NetworkError));
+        
+        // Server errors
+        let error = anyhow!("HTTP 503 Service Unavailable");
+        assert_eq!(classify_error(&error), ErrorType::Recoverable(RecoverableError::ServerError));
+        
+        // Model busy
+        let error = anyhow!("Model is busy, please try again");
+        assert_eq!(classify_error(&error), ErrorType::Recoverable(RecoverableError::ModelBusy));
+        
+        // Timeout
+        let error = anyhow!("Request timed out");
+        assert_eq!(classify_error(&error), ErrorType::Recoverable(RecoverableError::Timeout));
+        
+        // Token limit
+        let error = anyhow!("Token limit exceeded");
+        assert_eq!(classify_error(&error), ErrorType::Recoverable(RecoverableError::TokenLimit));
+        
+        // Non-recoverable
+        let error = anyhow!("Invalid API key");
+        assert_eq!(classify_error(&error), ErrorType::NonRecoverable);
+        
+        let error = anyhow!("Malformed request");
+        assert_eq!(classify_error(&error), ErrorType::NonRecoverable);
+    }
+
+    #[test]
+    fn test_retry_delay_calculation() {
+        // Test that delays increase exponentially
+        let delay1 = calculate_retry_delay(1, false);
+        let delay2 = calculate_retry_delay(2, false);
+        let delay3 = calculate_retry_delay(3, false);
+        
+        // Due to jitter, we can't test exact values, but the base should increase
+        assert!(delay1.as_millis() >= (BASE_RETRY_DELAY_MS as f64 * 0.7) as u128);
+        assert!(delay1.as_millis() <= (BASE_RETRY_DELAY_MS as f64 * 1.3) as u128);
+        
+        // Delay 2 should be roughly 2x delay 1 (minus jitter)
+        assert!(delay2.as_millis() >= delay1.as_millis());
+        
+        // Delay 3 should be roughly 2x delay 2 (minus jitter)
+        assert!(delay3.as_millis() >= delay2.as_millis());
+        
+        // Test max cap
+        let delay_max = calculate_retry_delay(10, false);
+        assert!(delay_max.as_millis() <= (DEFAULT_MAX_RETRY_DELAY_MS as f64 * 1.3) as u128);
+    }
+
+    #[test]
+    fn test_autonomous_retry_delay_calculation() {
+        // Test autonomous mode delays are distributed over 10 minutes
+        let delay1 = calculate_retry_delay(1, true);
+        let delay2 = calculate_retry_delay(2, true);
+        let delay3 = calculate_retry_delay(3, true);
+        let delay4 = calculate_retry_delay(4, true);
+        let delay5 = calculate_retry_delay(5, true);
+        let delay6 = calculate_retry_delay(6, true);
+        
+        // Base delays should be around: 10s, 30s, 60s, 120s, 180s, 200s
+        // With ±30% jitter
+        assert!(delay1.as_millis() >= 7000 && delay1.as_millis() <= 13000);
+        assert!(delay2.as_millis() >= 21000 && delay2.as_millis() <= 39000);
+        assert!(delay3.as_millis() >= 42000 && delay3.as_millis() <= 78000);
+        assert!(delay4.as_millis() >= 84000 && delay4.as_millis() <= 156000);
+        assert!(delay5.as_millis() >= 126000 && delay5.as_millis() <= 234000);
+        assert!(delay6.as_millis() >= 140000 && delay6.as_millis() <= 260000);
+    }
+
+    #[test]
+    fn test_truncate_for_logging() {
+        let short_text = "Hello, world!";
+        assert_eq!(truncate_for_logging(short_text, 20), "Hello, world!");
+        
+        let long_text = "This is a very long text that should be truncated for logging purposes";
+        let truncated = truncate_for_logging(long_text, 20);
+        assert!(truncated.starts_with("This is a very long "));
+        assert!(truncated.contains("truncated"));
+        assert!(truncated.contains("total bytes"));
+    }
+    
+    #[test]
+    fn test_truncate_with_multibyte_chars() {
+        // Test with multi-byte UTF-8 characters
+        let text_with_emoji = "Hello 👋 World 🌍 Test ✨ More text here";
+        let truncated = truncate_for_logging(text_with_emoji, 10);
+        // Should truncate at a valid UTF-8 boundary
+        assert!(truncated.starts_with("Hello "));
+        
+        // Test with box-drawing characters like the one causing the panic
+        let text_with_box = "Some text ┌─────┐ more text";
+        let truncated = truncate_for_logging(text_with_box, 12);
+        // Should not panic and should truncate at a valid boundary
+        assert!(truncated.contains("Some text"));
+        assert!(truncated.contains("truncated"));
+    }
+}

crates/g3-core/src/error_handling_test.rs

@@ -0,0 +1,154 @@
+//! Integration tests for error handling with retry logic
+
+#[cfg(test)]
+mod tests {
+    use crate::error_handling::*;
+    use std::sync::atomic::{AtomicU32, Ordering};
+    use std::sync::Arc;
+
+    #[tokio::test]
+    async fn test_retry_with_recoverable_error() {
+        let attempt_count = Arc::new(AtomicU32::new(0));
+
+        let context = ErrorContext::new(
+            "test_operation".to_string(),
+            "test_provider".to_string(),
+            "test_model".to_string(),
+            "test prompt".to_string(),
+            None,
+            100,
+            false,  // quiet parameter
+        );
+
+        let result = retry_with_backoff(
+            "test_operation",
+            || {
+                let counter = Arc::clone(&attempt_count);
+                async move {
+                    let count = counter.fetch_add(1, Ordering::SeqCst);
+                    if count < 2 {
+                        // Fail with recoverable error on first two attempts
+                        Err(anyhow::anyhow!("Rate limit exceeded"))
+                    } else {
+                        // Succeed on third attempt
+                        Ok("Success")
+                    }
+                }
+            },
+            &context,
+            false, // not autonomous mode
+        )
+        .await;
+
+        assert!(result.is_ok());
+        assert_eq!(result.unwrap(), "Success");
+        assert_eq!(attempt_count.load(Ordering::SeqCst), 3);
+    }
+
+    #[tokio::test]
+    async fn test_retry_with_non_recoverable_error() {
+        let attempt_count = Arc::new(AtomicU32::new(0));
+
+        let context = ErrorContext::new(
+            "test_operation".to_string(),
+            "test_provider".to_string(),
+            "test_model".to_string(),
+            "test prompt".to_string(),
+            None,
+            100,
+            false,  // quiet parameter
+        );
+
+        let result: Result<&str, _> = retry_with_backoff(
+            "test_operation",
+            || {
+                let counter = Arc::clone(&attempt_count);
+                async move {
+                    counter.fetch_add(1, Ordering::SeqCst);
+                    // Always fail with non-recoverable error
+                    Err(anyhow::anyhow!("Invalid API key"))
+                }
+            },
+            &context,
+            false, // not autonomous mode
+        )
+        .await;
+
+        assert!(result.is_err());
+        assert_eq!(attempt_count.load(Ordering::SeqCst), 1); // Should only try once
+    }
+
+    #[tokio::test]
+    async fn test_retry_exhaustion() {
+        let attempt_count = Arc::new(AtomicU32::new(0));
+
+        let context = ErrorContext::new(
+            "test_operation".to_string(),
+            "test_provider".to_string(),
+            "test_model".to_string(),
+            "test prompt".to_string(),
+            None,
+            100,
+            false,  // quiet parameter
+        );
+
+        let result: Result<&str, _> = retry_with_backoff(
+            "test_operation",
+            || {
+                let counter = Arc::clone(&attempt_count);
+                async move {
+                    counter.fetch_add(1, Ordering::SeqCst);
+                    // Always fail with recoverable error
+                    Err(anyhow::anyhow!("Network connection failed"))
+                }
+            },
+            &context,
+            false, // not autonomous mode
+        )
+        .await;
+
+        assert!(result.is_err());
+        assert_eq!(attempt_count.load(Ordering::SeqCst), 3); // Should try MAX_RETRY_ATTEMPTS times
+    }
+
+    #[test]
+    fn test_error_context_truncation() {
+        let long_prompt = "a".repeat(2000);
+        let context = ErrorContext::new(
+            "test_op".to_string(),
+            "provider".to_string(),
+            "model".to_string(),
+            long_prompt,
+            None,
+            100,
+            false,  // quiet parameter
+        );
+
+        // The prompt should be truncated to 1000 chars
+        assert!(context.last_prompt.len() < 1100); // Some buffer for the truncation message
+        assert!(context.last_prompt.contains("truncated"));
+    }
+
+    #[test]
+    fn test_retry_delay_increases() {
+        let delay1 = calculate_retry_delay(1, false);
+        let delay2 = calculate_retry_delay(2, false);
+        let delay3 = calculate_retry_delay(3, false);
+
+        // Delays should generally increase (though jitter can affect this)
+        // We'll test the base delays without jitter
+        let base1 = 1000u64; // BASE_RETRY_DELAY_MS
+        let base2 = 1000u64 * 2;
+        let base3 = 1000u64 * 4;
+
+        // Check that delays are within expected ranges (accounting for jitter)
+        assert!(delay1.as_millis() >= (base1 as f64 * 0.7) as u128);
+        assert!(delay1.as_millis() <= (base1 as f64 * 1.3) as u128);
+
+        assert!(delay2.as_millis() >= (base2 as f64 * 0.7) as u128);
+        assert!(delay2.as_millis() <= (base2 as f64 * 1.3) as u128);
+
+        assert!(delay3.as_millis() >= (base3 as f64 * 0.7) as u128);
+        assert!(delay3.as_millis() <= (base3 as f64 * 1.3) as u128);
+    }
+}

crates/g3-core/src/fixed_filter_json.rs

@@ -0,0 +1,463 @@
+// FINAL CORRECTED implementation of filter_json_tool_calls function according to specification
+// 1. Detect tool call start with regex '\w*{\w*"tool"\w*:\w*"' on the very next newline
+// 2. Enter suppression mode and use brace counting to find complete JSON
+// 3. Only elide JSON content between first '{' and last '}' (inclusive)
+// 4. Return everything else as the final filtered string
+
+//! JSON tool call filtering for streaming LLM responses.
+//! 
+//! This module filters out JSON tool calls from LLM output streams while preserving
+//! regular text content. It uses a state machine to handle streaming chunks.
+
+use regex::Regex;
+use std::cell::RefCell;
+use tracing::debug;
+
+// Thread-local state for tracking JSON tool call suppression
+thread_local! {
+    static FIXED_JSON_TOOL_STATE: RefCell<FixedJsonToolState> = RefCell::new(FixedJsonToolState::new());
+}
+
+/// Internal state for tracking JSON tool call filtering across streaming chunks.
+#[derive(Debug, Clone)]
+struct FixedJsonToolState {
+    /// True when actively suppressing a confirmed tool call
+    suppression_mode: bool,
+    /// True when buffering potential JSON (saw { but not yet confirmed as tool call)
+    potential_json_mode: bool,
+    /// Tracks nesting depth of braces within JSON
+    brace_depth: i32,
+    buffer: String,
+    json_start_in_buffer: Option<usize>, // Position where confirmed JSON tool call starts
+    content_returned_up_to: usize, // Track how much content we've already returned
+    potential_json_start: Option<usize>, // Where the potential JSON started
+}
+
+impl FixedJsonToolState {
+    fn new() -> Self {
+        Self {
+            suppression_mode: false,
+            potential_json_mode: false,
+            brace_depth: 0,
+            buffer: String::new(),
+            json_start_in_buffer: None,
+            content_returned_up_to: 0,
+            potential_json_start: None,
+        }
+    }
+
+    fn reset(&mut self) {
+        self.suppression_mode = false;
+        self.potential_json_mode = false;
+        self.brace_depth = 0;
+        self.buffer.clear();
+        self.json_start_in_buffer = None;
+        self.content_returned_up_to = 0;
+        self.potential_json_start = None;
+    }
+}
+
+// FINAL CORRECTED implementation according to specification
+
+/// Filters JSON tool calls from streaming LLM content.
+///
+/// Processes content chunks and removes JSON tool calls while preserving regular text.
+/// Maintains state across calls to handle tool calls spanning multiple chunks.
+pub fn fixed_filter_json_tool_calls(content: &str) -> String {
+    if content.is_empty() {
+        return String::new();
+    }
+
+    FIXED_JSON_TOOL_STATE.with(|state| {
+        let mut state = state.borrow_mut();
+
+        // Add new content to buffer
+        state.buffer.push_str(content);
+
+        // If we're already in suppression mode, continue brace counting
+        if state.suppression_mode {
+            // Count braces in the new content only
+            for ch in content.chars() {
+                match ch {
+                    '{' => state.brace_depth += 1,
+                    '}' => {
+                        state.brace_depth -= 1;
+                        // Exit suppression mode when all braces are closed
+                        if state.brace_depth <= 0 {
+                            debug!("JSON tool call completed - exiting suppression mode");
+
+                            // Extract the complete result with JSON filtered out
+                            let result = extract_fixed_content(
+                                &state.buffer,
+                                state.json_start_in_buffer.unwrap_or(0),
+                            );
+
+                            // Return only the part we haven't returned yet
+                            let new_content = if result.len() > state.content_returned_up_to {
+                                result[state.content_returned_up_to..].to_string()
+                            } else {
+                                String::new()
+                            };
+
+                            state.reset();
+                            return new_content;
+                        }
+                    }
+                    _ => {}
+                }
+            }
+            
+            // CRITICAL FIX: After counting braces, if still in suppression mode,
+            // check if a new tool call pattern appears. This handles truncated JSON
+            // followed by complete JSON.
+            if state.suppression_mode {
+                let current_json_start = state.json_start_in_buffer.unwrap();
+                // Don't require newline - the new JSON might be concatenated directly
+                let tool_call_regex = Regex::new(r#"\{\s*"tool"\s*:\s*""#).unwrap();
+                
+                // Look for new tool call patterns after the current one
+                if let Some(captures) = tool_call_regex.find(&state.buffer[current_json_start + 1..]) {
+                    let new_json_start = current_json_start + 1 + captures.start() + captures.as_str().find('{').unwrap();
+                    
+                    debug!("Detected new tool call at position {} while processing incomplete one at {} - discarding old", new_json_start, current_json_start);
+                    
+                    // The previous JSON was incomplete/malformed
+                    // Return content before the old JSON (if any)
+                    let content_before_old_json = if current_json_start > state.content_returned_up_to {
+                        state.buffer[state.content_returned_up_to..current_json_start].to_string()
+                    } else {
+                        String::new()
+                    };
+                    
+                    // Update state to skip the incomplete JSON and position at the new one
+                    // We'll process the new JSON on the next call
+                    state.content_returned_up_to = new_json_start;
+                    state.suppression_mode = false;
+                    state.json_start_in_buffer = None;
+                    state.brace_depth = 0;
+                    
+                    return content_before_old_json;
+                }
+            }
+            
+            // Still in suppression mode, return empty string (content is being accumulated)
+            return String::new();
+        }
+
+        // Check if we're in potential JSON mode (saw { but waiting to confirm it's a tool call)
+        if state.potential_json_mode {
+            // Check if the buffer contains a confirmed tool call pattern
+            let tool_call_regex = Regex::new(r#"(?m)^\s*\{\s*"tool"\s*:\s*""#).unwrap();
+            
+            if let Some(captures) = tool_call_regex.find(&state.buffer) {
+                // Confirmed! This is a tool call - enter suppression mode
+                let match_text = captures.as_str();
+                if let Some(brace_offset) = match_text.find('{') {
+                    let json_start = captures.start() + brace_offset;
+                    
+                    debug!("Confirmed JSON tool call at position {} - entering suppression mode", json_start);
+                    
+                    state.potential_json_mode = false;
+                    state.suppression_mode = true;
+                    state.brace_depth = 0;
+                    state.json_start_in_buffer = Some(json_start);
+                    
+                    // Count braces from json_start to see if JSON is complete
+                    let buffer_slice = state.buffer[json_start..].to_string();
+                    for ch in buffer_slice.chars() {
+                        match ch {
+                            '{' => state.brace_depth += 1,
+                            '}' => {
+                                state.brace_depth -= 1;
+                                if state.brace_depth <= 0 {
+                                    debug!("JSON tool call completed immediately");
+                                    let result = extract_fixed_content(&state.buffer, json_start);
+                                    let new_content = if result.len() > state.content_returned_up_to {
+                                        result[state.content_returned_up_to..].to_string()
+                                    } else {
+                                        String::new()
+                                    };
+                                    state.reset();
+                                    return new_content;
+                                }
+                            }
+                            _ => {}
+                        }
+                    }
+                    // JSON incomplete, stay in suppression mode, return nothing
+                    return String::new();
+                }
+            }
+            
+            // Check if we can rule out this being a tool call
+            // If we have enough content after the { and it doesn't match the pattern, release it
+            if let Some(potential_start) = state.potential_json_start {
+                let content_after_brace = &state.buffer[potential_start..];
+                
+                // Rule out as a tool call if:
+                // 1. Closing } appears before we see the full pattern
+                // 2. Content clearly doesn't match the tool call pattern
+                // 3. Newline appears after the opening brace (tool calls should be compact)
+                
+                let has_closing_brace = content_after_brace.contains('}');
+                let has_newline = content_after_brace[1..].contains('\n'); // Skip first char which is {
+                let long_enough = content_after_brace.len() >= 10;
+                
+                // Detect non-tool JSON patterns:
+                // - { followed by " and a key that doesn't start with "tool"
+                // - { followed by "t" but not "to"
+                // - { followed by "to" but not "too", etc.
+                let not_tool_pattern = Regex::new(r#"^\{\s*"(?:[^t]|t(?:[^o]|o(?:[^o]|o(?:[^l]|l[^"\s:]))))"#).unwrap();
+                let definitely_not_tool = not_tool_pattern.is_match(content_after_brace);
+                
+                if has_closing_brace || has_newline || (long_enough && definitely_not_tool) {
+                    debug!("Potential JSON ruled out - not a tool call");
+                    state.potential_json_mode = false;
+                    state.potential_json_start = None;
+                    
+                    // Return the buffered content we've been holding
+                    let new_content = if state.buffer.len() > state.content_returned_up_to {
+                        state.buffer[state.content_returned_up_to..].to_string()
+                    } else {
+                        String::new()
+                    };
+                    state.content_returned_up_to = state.buffer.len();
+                    return new_content;
+                }
+            }
+            
+            // Still in potential mode, keep buffering
+            return String::new();
+        }
+
+        // Detect potential JSON start: { at the beginning of a line
+        let potential_json_regex = Regex::new(r"(?m)^\s*\{\s*").unwrap();
+        
+        if let Some(captures) = potential_json_regex.find(&state.buffer[state.content_returned_up_to..]) {
+            let match_start = state.content_returned_up_to + captures.start();
+            let brace_pos = match_start + captures.as_str().find('{').unwrap();
+            
+            debug!("Potential JSON detected at position {} - entering buffering mode", brace_pos);
+            
+            // Fast path: check if this is already a confirmed tool call
+            let tool_call_regex = Regex::new(r#"(?m)^\s*\{\s*"tool"\s*:\s*""#).unwrap();
+            if tool_call_regex.is_match(&state.buffer[brace_pos..]) {
+                // This is a confirmed tool call! Process it immediately
+                let json_start = brace_pos;
+                debug!("Immediately confirmed tool call at position {}", json_start);
+                
+                // Return content before JSON
+                let content_before = if json_start > state.content_returned_up_to {
+                    state.buffer[state.content_returned_up_to..json_start].to_string()
+                } else {
+                    String::new()
+                };
+                
+                state.content_returned_up_to = json_start;
+                state.suppression_mode = true;
+                state.brace_depth = 0;
+                state.json_start_in_buffer = Some(json_start);
+                
+                // Count braces to see if JSON is complete
+                let buffer_slice = state.buffer[json_start..].to_string();
+                for ch in buffer_slice.chars() {
+                    match ch {
+                        '{' => state.brace_depth += 1,
+                        '}' => {
+                            state.brace_depth -= 1;
+                            if state.brace_depth <= 0 {
+                                debug!("JSON tool call completed in same chunk");
+                                let result = extract_fixed_content(&state.buffer, json_start);
+                                let content_after = if result.len() > json_start {
+                                    &result[json_start..]
+                                } else {
+                                    ""
+                                };
+                                let final_result = format!("{}{}", content_before, content_after);
+                                state.reset();
+                                return final_result;
+                            }
+                        }
+                        _ => {}
+                    }
+                }
+                // JSON incomplete, return content before and stay in suppression mode
+                return content_before;
+            }
+            
+            // Return content before the potential JSON
+            let content_before = if brace_pos > state.content_returned_up_to {
+                state.buffer[state.content_returned_up_to..brace_pos].to_string()
+            } else {
+                String::new()
+            };
+            
+            state.content_returned_up_to = brace_pos;
+            state.potential_json_mode = true;
+            state.potential_json_start = Some(brace_pos);
+            
+            // Optimization: immediately check if we can rule this out for single-chunk processing
+            let content_after_brace = &state.buffer[brace_pos..];
+            let has_closing_brace = content_after_brace.contains('}');
+            let has_newline = content_after_brace.len() > 1 && content_after_brace[1..].contains('\n');
+            let long_enough = content_after_brace.len() >= 10;
+            
+            let not_tool_pattern = Regex::new(r#"^\{\s*"(?:[^t]|t(?:[^o]|o(?:[^o]|o(?:[^l]|l[^"\s:]))))"#).unwrap();
+            let definitely_not_tool = not_tool_pattern.is_match(content_after_brace);
+            
+            if has_closing_brace || has_newline || (long_enough && definitely_not_tool) {
+                debug!("Immediately ruled out as not a tool call");
+                state.potential_json_mode = false;
+                state.potential_json_start = None;
+                
+                // Return all the buffered content
+                let new_content = if state.buffer.len() > state.content_returned_up_to {
+                    state.buffer[state.content_returned_up_to..].to_string()
+                } else {
+                    String::new()
+                };
+                state.content_returned_up_to = state.buffer.len();
+                return format!("{}{}", content_before, new_content);
+            }
+            
+            return content_before;
+        }
+
+        // Check for tool call pattern using corrected regex
+        let tool_call_regex = Regex::new(r#"(?m)^\s*\{\s*"tool"\s*:\s*"[^"]*""#).unwrap();
+
+        if let Some(captures) = tool_call_regex.find(&state.buffer) {
+            let match_text = captures.as_str();
+
+            // Find the position of the opening brace in the match
+            if let Some(brace_offset) = match_text.find('{') {
+                let json_start = captures.start() + brace_offset;
+
+                debug!(
+                    "Detected JSON tool call at position {} - entering suppression mode",
+                    json_start
+                );
+
+                // Return content before JSON that we haven't returned yet
+                let content_before_json = if json_start >= state.content_returned_up_to {
+                    state.buffer[state.content_returned_up_to..json_start].to_string()
+                } else {
+                    String::new()
+                };
+
+                state.content_returned_up_to = json_start;
+
+                // Enter suppression mode
+                state.suppression_mode = true;
+                state.brace_depth = 0;
+                state.json_start_in_buffer = Some(json_start);
+
+                // Count braces from the JSON start to see if it's complete
+                let buffer_clone = state.buffer.clone();
+                for ch in buffer_clone[json_start..].chars() {
+                    match ch {
+                        '{' => state.brace_depth += 1,
+                        '}' => {
+                            state.brace_depth -= 1;
+                            if state.brace_depth <= 0 {
+                                // JSON is complete in this chunk
+                                debug!("JSON tool call completed in same chunk");
+                                let result = extract_fixed_content(&buffer_clone, json_start);
+
+                                // Return content before JSON plus content after JSON
+                                let content_after_json = if result.len() > json_start {
+                                    &result[json_start..]
+                                } else {
+                                    ""
+                                };
+
+                                let final_result =
+                                    format!("{}{}", content_before_json, content_after_json);
+                                state.reset();
+                                return final_result;
+                            }
+                        }
+                        _ => {}
+                    }
+                }
+
+                // JSON is incomplete, return only the content before JSON
+                return content_before_json;
+            }
+        }
+
+        // No JSON tool call detected, return only the new content we haven't returned yet
+        
+
+        if state.buffer.len() > state.content_returned_up_to {
+            let result = state.buffer[state.content_returned_up_to..].to_string();
+            state.content_returned_up_to = state.buffer.len();
+            result
+        } else {
+            String::new()
+        }
+    })
+}
+
+/// Extracts content from buffer, removing the JSON tool call.
+///
+/// Given a buffer and the start position of a JSON tool call, this function:
+/// 1. Extracts all content before the JSON
+/// 2. Finds the end of the JSON (matching closing brace)
+/// 3. Extracts all content after the JSON
+/// 4. Returns the concatenation of before + after (JSON removed)
+///
+/// # Arguments
+/// * `full_content` - The full content buffer
+/// * `json_start` - Position where the JSON tool call begins
+fn extract_fixed_content(full_content: &str, json_start: usize) -> String {
+    // Find the end of the JSON using proper brace counting with string handling
+    let mut brace_depth = 0;
+    let mut json_end = json_start;
+    let mut in_string = false;
+    let mut escape_next = false;
+
+    for (i, ch) in full_content[json_start..].char_indices() {
+        if escape_next {
+            escape_next = false;
+            continue;
+        }
+
+        match ch {
+            '\\' if in_string => escape_next = true,
+            '"' if !escape_next => in_string = !in_string,
+            '{' if !in_string => {
+                brace_depth += 1;
+            }
+            '}' if !in_string => {
+                brace_depth -= 1;
+                if brace_depth == 0 {
+                    json_end = json_start + i + 1; // +1 to include the closing brace
+                    break;
+                }
+            }
+            _ => {}
+        }
+    }
+
+    // Return content before and after the JSON (excluding the JSON itself)
+    let before = &full_content[..json_start];
+    let after = if json_end < full_content.len() {
+        &full_content[json_end..]
+    } else {
+        ""
+    };
+
+    format!("{}{}", before, after)
+}
+
+/// Resets the global JSON filtering state.
+///
+/// Call this between independent filtering sessions to ensure clean state.
+/// This is particularly important in tests and when starting new conversations.
+pub fn reset_fixed_json_tool_state() {
+    FIXED_JSON_TOOL_STATE.with(|state| {
+        let mut state = state.borrow_mut();
+        state.reset();
+    });
+}

crates/g3-core/src/fixed_filter_tests.rs

@@ -0,0 +1,384 @@
+//! Tests for JSON tool call filtering.
+//!
+//! These tests verify that the filter correctly identifies and removes JSON tool calls
+//! from LLM output streams while preserving all other content.
+
+#[cfg(test)]
+mod fixed_filter_tests {
+    use crate::fixed_filter_json::{fixed_filter_json_tool_calls, reset_fixed_json_tool_state};
+    use regex::Regex;
+
+    /// Test that regular text without tool calls passes through unchanged.
+    #[test]
+    fn test_no_tool_call_passthrough() {
+        reset_fixed_json_tool_state();
+        let input = "This is regular text without any tool calls.";
+        let result = fixed_filter_json_tool_calls(input);
+        assert_eq!(result, input);
+    }
+
+    /// Test detection and removal of a complete tool call in a single chunk.
+    #[test]
+    fn test_simple_tool_call_detection() {
+        reset_fixed_json_tool_state();
+        let input = r#"Some text before
+{"tool": "shell", "args": {"command": "ls"}}
+Some text after"#;
+
+        let result = fixed_filter_json_tool_calls(input);
+        let expected = "Some text before\n\nSome text after";
+        assert_eq!(result, expected);
+    }
+
+    /// Test handling of tool calls that arrive across multiple streaming chunks.
+    #[test]
+    fn test_streaming_chunks() {
+        reset_fixed_json_tool_state();
+
+        // Simulate streaming where the tool call comes in multiple chunks
+        let chunks = vec![
+            "Some text before\n",
+            "{\"tool\": \"",
+            "shell\", \"args\": {",
+            "\"command\": \"ls\"",
+            "}}\nText after",
+        ];
+
+        let mut results = Vec::new();
+        for chunk in chunks {
+            let result = fixed_filter_json_tool_calls(chunk);
+            results.push(result);
+        }
+
+        // The final accumulated result should have the JSON filtered out
+        let final_result: String = results.join("");
+        let expected = "Some text before\n\nText after";
+        assert_eq!(final_result, expected);
+    }
+
+    /// Test correct handling of nested braces within JSON strings.
+    #[test]
+    fn test_nested_braces_in_tool_call() {
+        reset_fixed_json_tool_state();
+
+        let input = r#"Text before
+{"tool": "write_file", "args": {"file_path": "test.json", "content": "{\"nested\": \"value\"}"}}
+Text after"#;
+
+        let result = fixed_filter_json_tool_calls(input);
+        let expected = "Text before\n\nText after";
+        assert_eq!(result, expected);
+    }
+
+    /// Verify the regex pattern matches the specification with flexible whitespace.
+    #[test]
+    fn test_regex_pattern_specification() {
+        // Test the corrected regex pattern that's more flexible with whitespace
+        let pattern = Regex::new(r#"(?m)^\s*\{\s*"tool"\s*:"#).unwrap();
+
+        let test_cases = vec![
+            (
+                r#"line
+{"tool":"#,
+                true,
+            ),
+            (
+                r#"line
+{"tool" :"#,
+                true,
+            ),
+            (
+                r#"line
+{ "tool":"#,
+                true,
+            ), // Space after { DOES match with \s*
+            (
+                r#"line
+{"tool123":"#,
+                false,
+            ), // "tool123" is not exactly "tool"
+            (
+                r#"line
+{"tool" : "#,
+                true,
+            ),
+        ];
+
+        for (input, should_match) in test_cases {
+            let matches = pattern.is_match(input);
+            assert_eq!(
+                matches, should_match,
+                "Pattern matching failed for: {}",
+                input
+            );
+        }
+    }
+
+    /// Test that tool calls must appear at the start of a line (after newline).
+    #[test]
+    fn test_newline_requirement() {
+        reset_fixed_json_tool_state();
+
+        // According to spec, tool call should be detected "on the very next newline"
+        // Our current regex matches any line that contains the pattern, not just after newlines
+        let input_with_newline = "Text\n{\"tool\": \"shell\", \"args\": {\"command\": \"ls\"}}";
+        let input_without_newline = "Text {\"tool\": \"shell\", \"args\": {\"command\": \"ls\"}}";
+
+        let result1 = fixed_filter_json_tool_calls(input_with_newline);
+        reset_fixed_json_tool_state();
+        let result2 = fixed_filter_json_tool_calls(input_without_newline);
+
+        // With the new aggressive filtering, only the newline case should trigger suppression
+        // The pattern requires { to be at the start of a line (after ^)
+        assert_eq!(result1, "Text\n");
+        // Without newline before {, it should pass through unchanged
+        assert_eq!(result2, input_without_newline);
+    }
+
+    /// Test handling of escaped quotes within JSON strings.
+    #[test]
+    fn test_json_with_escaped_quotes() {
+        reset_fixed_json_tool_state();
+
+        let input = r#"Text
+{"tool": "write_file", "args": {"content": "He said \"hello\" to me"}}
+More text"#;
+
+        let result = fixed_filter_json_tool_calls(input);
+        let expected = "Text\n\nMore text";
+        assert_eq!(result, expected);
+    }
+
+    /// Test graceful handling of incomplete/malformed JSON.
+    #[test]
+    fn test_edge_case_malformed_json() {
+        reset_fixed_json_tool_state();
+
+        // Test what happens with malformed JSON that starts like a tool call
+        let input = r#"Text
+{"tool": "shell", "args": {"command": "ls"
+More text"#;
+
+        let result = fixed_filter_json_tool_calls(input);
+        // Should handle gracefully - since JSON is incomplete, it should return content before JSON
+        let expected = "Text\n";
+        assert_eq!(result, expected);
+    }
+
+    /// Test processing multiple independent tool calls sequentially.
+    #[test]
+    fn test_multiple_tool_calls_sequential() {
+        reset_fixed_json_tool_state();
+
+        // Test processing multiple tool calls one at a time
+        let input1 = r#"First text
+{"tool": "shell", "args": {"command": "ls"}}
+Middle text"#;
+        let result1 = fixed_filter_json_tool_calls(input1);
+        let expected1 = "First text\n\nMiddle text";
+        assert_eq!(result1, expected1);
+
+        // Reset and process second tool call
+        reset_fixed_json_tool_state();
+        let input2 = r#"More text
+{"tool": "read_file", "args": {"file_path": "test.txt"}}
+Final text"#;
+        let result2 = fixed_filter_json_tool_calls(input2);
+        let expected2 = "More text\n\nFinal text";
+        assert_eq!(result2, expected2);
+    }
+
+    /// Test tool calls with complex multi-line arguments.
+    #[test]
+    fn test_tool_call_with_complex_args() {
+        reset_fixed_json_tool_state();
+
+        let input = r#"Before
+{"tool": "str_replace", "args": {"file_path": "test.rs", "diff": "--- old\n-old line\n+++ new\n+new line", "start": 0, "end": 100}}
+After"#;
+
+        let result = fixed_filter_json_tool_calls(input);
+        let expected = "Before\n\nAfter";
+        assert_eq!(result, expected);
+    }
+
+    /// Test input containing only a tool call with no surrounding text.
+    #[test]
+    fn test_tool_call_only() {
+        reset_fixed_json_tool_state();
+
+        let input = r#"
+{"tool": "final_output", "args": {"summary": "Task completed successfully"}}"#;
+
+        let result = fixed_filter_json_tool_calls(input);
+        let expected = "\n";
+        assert_eq!(result, expected);
+    }
+
+    /// Test accurate brace counting with deeply nested structures.
+    #[test]
+    fn test_brace_counting_accuracy() {
+        reset_fixed_json_tool_state();
+
+        // Test complex nested structure
+        let input = r#"Start
+{"tool": "write_file", "args": {"content": "function() { return {a: 1, b: {c: 2}}; }", "file_path": "test.js"}}
+End"#;
+
+        let result = fixed_filter_json_tool_calls(input);
+        let expected = "Start\n\nEnd";
+        assert_eq!(result, expected);
+    }
+
+    /// Test that braces within strings don't affect brace counting.
+    #[test]
+    fn test_string_escaping_in_json() {
+        reset_fixed_json_tool_state();
+
+        // Test JSON with escaped quotes and braces in strings
+        let input = r#"Text
+{"tool": "shell", "args": {"command": "echo \"Hello {world}\" > file.txt"}}
+More"#;
+
+        let result = fixed_filter_json_tool_calls(input);
+        let expected = "Text\n\nMore";
+        assert_eq!(result, expected);
+    }
+
+    /// Verify compliance with the exact specification requirements.
+    #[test]
+    fn test_specification_compliance() {
+        reset_fixed_json_tool_state();
+
+        // Test the exact specification requirements:
+        // 1. Detect start with regex '\w*{\w*"tool"\w*:\w*"' on newline
+        // 2. Enter suppression mode and use brace counting
+        // 3. Elide only JSON between first '{' and last '}' (inclusive)
+        // 4. Return everything else
+
+        let input = "Before text\nSome more text\n{\"tool\": \"test\", \"args\": {}}\nAfter text\nMore after";
+        let result = fixed_filter_json_tool_calls(input);
+        let expected = "Before text\nSome more text\n\nAfter text\nMore after";
+        assert_eq!(result, expected);
+    }
+
+    /// Test that non-tool JSON objects are not filtered.
+    #[test]
+    fn test_no_false_positives() {
+        reset_fixed_json_tool_state();
+
+        // Test that we don't incorrectly identify non-tool JSON as tool calls
+        let input = r#"Some text
+{"not_tool": "value", "other": "data"}
+More text"#;
+        let result = fixed_filter_json_tool_calls(input);
+        // Should pass through unchanged since it doesn't match the tool pattern
+        assert_eq!(result, input);
+    }
+
+    /// Test patterns that look similar to tool calls but aren't exact matches.
+    #[test]
+    fn test_partial_tool_patterns() {
+        reset_fixed_json_tool_state();
+
+        // Test patterns that look like tool calls but aren't complete
+        let test_cases = vec![
+            "Text\n{\"too\": \"value\"}",   // "too" not "tool"
+            "Text\n{\"tools\": \"value\"}", // "tools" not "tool"
+            "Text\n{\"tool\": }",           // Missing value after colon
+        ];
+
+        for input in test_cases {
+            reset_fixed_json_tool_state();
+            let result = fixed_filter_json_tool_calls(input);
+            // These should all pass through unchanged
+            assert_eq!(result, input, "Input should pass through: {}", input);
+        }
+    }
+
+    /// Test streaming with very small chunks (character-by-character).
+    #[test]
+    fn test_streaming_edge_cases() {
+        reset_fixed_json_tool_state();
+
+        // Test streaming with very small chunks
+        let chunks = vec![
+            "Text\n", "{", "\"", "tool", "\"", ":", " ", "\"", "test", "\"", "}", "\nAfter",
+        ];
+
+        let mut results = Vec::new();
+        for chunk in chunks {
+            let result = fixed_filter_json_tool_calls(chunk);
+            results.push(result);
+        }
+
+        let final_result: String = results.join("");
+        // With the new aggressive filtering, the JSON should be completely filtered out
+        // even when it arrives in very small chunks
+        let expected = "Text\n\nAfter";
+        assert_eq!(final_result, expected);
+    }
+
+    /// Debug test with detailed logging for streaming behavior.
+    #[test]
+    fn test_streaming_debug() {
+        reset_fixed_json_tool_state();
+
+        // Debug the exact failing case
+        let chunks = vec![
+            "Some text before\n",
+            "{\"tool\": \"",
+            "shell\", \"args\": {",
+            "\"command\": \"ls\"",
+            "}}\nText after",
+        ];
+
+        let mut results = Vec::new();
+        for (i, chunk) in chunks.iter().enumerate() {
+            let result = fixed_filter_json_tool_calls(chunk);
+            println!("Chunk {}: {:?} -> {:?}", i, chunk, result);
+            results.push(result);
+        }
+
+        let final_result: String = results.join("");
+        println!("Final result: {:?}", final_result);
+        println!("Expected: {:?}", "Some text before\n\nText after");
+
+        let expected = "Some text before\n\nText after";
+        assert_eq!(final_result, expected);
+    }
+
+    /// Test handling of truncated JSON followed by complete JSON (the json_err pattern)
+    #[test]
+    fn test_truncated_then_complete_json() {
+        reset_fixed_json_tool_state();
+
+        // Simulate the pattern from json_err trace:
+        // 1. Incomplete/truncated JSON appears
+        // 2. Then the same complete JSON appears
+        let chunks = vec![
+            "Some text\n",
+            r#"{"tool": "str_replace", "args": {"diff":"...","file_path":"./crates/g3-cli"#,  // Truncated
+            r#"{"tool": "str_replace", "args": {"diff":"...","file_path":"./crates/g3-cli/src/lib.rs"}}"#,  // Complete
+            "\nMore text",
+        ];
+
+        let mut results = Vec::new();
+        for (i, chunk) in chunks.iter().enumerate() {
+            let result = fixed_filter_json_tool_calls(chunk);
+            println!("Chunk {}: {:?} -> {:?}", i, chunk, result);
+            results.push(result);
+        }
+
+        let final_result: String = results.join("");
+        println!("Final result: {:?}", final_result);
+
+        // The truncated JSON should be discarded when the complete one appears
+        // Both JSONs should be filtered out, leaving only the text
+        let expected = "Some text\n\nMore text";
+        assert_eq!(
+            final_result, expected,
+            "Failed to handle truncated JSON followed by complete JSON"
+        );
+    }
+}

crates/g3-core/src/project.rs

@@ -0,0 +1,184 @@
+use anyhow::Result;
+use serde::{Deserialize, Serialize};
+use std::path::{Path, PathBuf};
+
+/// Represents a G3 project with workspace configuration
+#[derive(Debug, Clone, Serialize, Deserialize)]
+pub struct Project {
+    /// The workspace directory for the project
+    pub workspace_dir: PathBuf,
+    
+    /// Path to the requirements document (for autonomous mode)
+    pub requirements_path: Option<PathBuf>,
+    
+    /// Override requirements text (takes precedence over requirements_path)
+    pub requirements_text: Option<String>,
+    
+    /// Whether the project is in autonomous mode
+    pub autonomous: bool,
+    
+    /// Project name (derived from workspace directory name)
+    pub name: String,
+    
+    /// Session ID for tracking
+    pub session_id: Option<String>,
+}
+
+impl Project {
+    /// Create a new project with the given workspace directory
+    pub fn new(workspace_dir: PathBuf) -> Self {
+        let name = workspace_dir
+            .file_name()
+            .and_then(|n| n.to_str())
+            .unwrap_or("unnamed")
+            .to_string();
+        
+        Self {
+            workspace_dir,
+            requirements_path: None,
+            requirements_text: None,
+            autonomous: false,
+            name,
+            session_id: None,
+        }
+    }
+    
+    /// Create a project for autonomous mode
+    pub fn new_autonomous(workspace_dir: PathBuf) -> Result<Self> {
+        let mut project = Self::new(workspace_dir.clone());
+        project.autonomous = true;
+        
+        // Look for requirements.md in the workspace directory
+        let requirements_path = workspace_dir.join("requirements.md");
+        if requirements_path.exists() {
+            project.requirements_path = Some(requirements_path);
+        }
+        
+        Ok(project)
+    }
+    
+    /// Create a project for autonomous mode with requirements text override
+    pub fn new_autonomous_with_requirements(workspace_dir: PathBuf, requirements_text: String) -> Result<Self> {
+        let mut project = Self::new(workspace_dir.clone());
+        project.autonomous = true;
+        project.requirements_text = Some(requirements_text);
+        
+        // Don't look for requirements.md file when text is provided
+        // The text override takes precedence
+        
+        Ok(project)
+    }
+    
+    /// Set the workspace directory and update related paths
+    pub fn set_workspace(&mut self, workspace_dir: PathBuf) {
+        self.workspace_dir = workspace_dir.clone();
+        self.name = workspace_dir
+            .file_name()
+            .and_then(|n| n.to_str())
+            .unwrap_or("unnamed")
+            .to_string();
+        
+        // Update requirements path if in autonomous mode
+        if self.autonomous {
+            let requirements_path = workspace_dir.join("requirements.md");
+            if requirements_path.exists() {
+                self.requirements_path = Some(requirements_path);
+            }
+        }
+    }
+    
+    /// Get the workspace directory
+    pub fn workspace(&self) -> &Path {
+        &self.workspace_dir
+    }
+    
+    /// Check if requirements file exists
+    pub fn has_requirements(&self) -> bool {
+        // Has requirements if either text override is provided or requirements file exists
+        self.requirements_text.is_some() || self.requirements_path.is_some()
+    }
+    
+    /// Check if implementation files exist in the workspace
+    pub fn has_implementation_files(&self) -> bool {
+        self.check_dir_for_implementation_files(&self.workspace_dir)
+    }
+    
+    /// Recursively check a directory for implementation files
+    #[allow(clippy::only_used_in_recursion)]
+    fn check_dir_for_implementation_files(&self, dir: &Path) -> bool {
+        // Common source file extensions
+        let extensions = vec![
+            "swift", "rs", "py", "js", "ts", "java", "cpp", "c",
+            "go", "rb", "php", "cs", "kt", "scala", "m", "h"
+        ];
+        
+        if let Ok(entries) = std::fs::read_dir(dir) {
+            for entry in entries.flatten() {
+                let path = entry.path();
+                
+                if path.is_file() {
+                    // Check if it's a source file
+                    if let Some(ext) = path.extension() {
+                        if let Some(ext_str) = ext.to_str() {
+                            if extensions.contains(&ext_str) {
+                                return true;
+                            }
+                        }
+                    }
+                } else if path.is_dir() {
+                    // Skip hidden directories and common non-source directories
+                    if let Some(name) = path.file_name().and_then(|n| n.to_str()) {
+                        if !name.starts_with('.') && name != "logs" && name != "target" && name != "node_modules" {
+                            // Recursively check subdirectories
+                            if self.check_dir_for_implementation_files(&path) {
+                                return true;
+                            }
+                        }
+                    }
+                }
+            }
+        }
+        false
+    }
+    
+    /// Read the requirements file content
+    pub fn read_requirements(&self) -> Result<Option<String>> {
+        // Prioritize requirements text override
+        if let Some(ref text) = self.requirements_text {
+            Ok(Some(text.clone()))
+        } else if let Some(ref path) = self.requirements_path {
+            // Fall back to reading from file
+            Ok(Some(std::fs::read_to_string(path)?))
+        } else {
+            Ok(None)
+        }
+    }
+    
+    /// Create the workspace directory if it doesn't exist
+    pub fn ensure_workspace_exists(&self) -> Result<()> {
+        if !self.workspace_dir.exists() {
+            std::fs::create_dir_all(&self.workspace_dir)?;
+        }
+        Ok(())
+    }
+    
+    /// Change to the workspace directory
+    pub fn enter_workspace(&self) -> Result<()> {
+        std::env::set_current_dir(&self.workspace_dir)?;
+        Ok(())
+    }
+    
+    /// Get the logs directory for the project
+    pub fn logs_dir(&self) -> PathBuf {
+        self.workspace_dir.join("logs")
+    }
+    
+    /// Ensure the logs directory exists
+    pub fn ensure_logs_dir(&self) -> Result<()> {
+        let logs_dir = self.logs_dir();
+        if !logs_dir.exists() {
+            std::fs::create_dir_all(&logs_dir)?;
+        }
+        Ok(())
+    }
+}

crates/g3-core/src/providers/mod.rs

@@ -1 +0,0 @@
-

crates/g3-core/src/take_screenshot_test.rs

@@ -0,0 +1,37 @@
+// Test to verify take_screenshot requires window_id
+
+#[cfg(test)]
+mod take_screenshot_tests {
+    use super::*;
+    use serde_json::json;
+
+    #[test]
+    fn test_take_screenshot_requires_window_id() {
+        // Create a tool call without window_id
+        let tool_call = ToolCall {
+            tool: "take_screenshot".to_string(),
+            args: json!({
+                "path": "test.png"
+            }),
+        };
+
+        // Verify that window_id is missing
+        assert!(tool_call.args.get("window_id").is_none());
+    }
+
+    #[test]
+    fn test_take_screenshot_with_window_id() {
+        // Create a tool call with window_id
+        let tool_call = ToolCall {
+            tool: "take_screenshot".to_string(),
+            args: json!({
+                "path": "test.png",
+                "window_id": "Safari"
+            }),
+        };
+
+        // Verify that window_id is present
+        assert!(tool_call.args.get("window_id").is_some());
+        assert_eq!(tool_call.args.get("window_id").unwrap().as_str().unwrap(), "Safari");
+    }
+}

crates/g3-core/src/task_result.rs

@@ -0,0 +1,168 @@
+use crate::ContextWindow;
+
+/// Result of a task execution containing both the response and the context window
+#[derive(Debug, Clone)]
+pub struct TaskResult {
+    /// The actual response content from the task execution
+    pub response: String,
+    /// The complete context window at the time of completion
+    pub context_window: ContextWindow,
+}
+
+impl TaskResult {
+    pub fn new(response: String, context_window: ContextWindow) -> Self {
+        Self {
+            response,
+            context_window,
+        }
+    }
+
+    /// Extract the final_output content from the response (for coach feedback in autonomous mode)
+    /// This looks for the complete final_output content, not just the last block
+    pub fn extract_final_output(&self) -> String {
+        // Remove any timing information at the end
+        let content_without_timing = if let Some(timing_pos) = self.response.rfind("\n⏱️") {
+            &self.response[..timing_pos]
+        } else {
+            &self.response
+        };
+
+        // Look for the final_output marker pattern
+        // The final_output content typically appears after the tool is called
+        // and is the substantive content that follows
+        
+        // First, try to find if there's a clear final_output section
+        // This would be the content after the last tool execution
+        if let Some(final_output_pos) = content_without_timing.rfind("final_output") {
+            // Find the content that follows the final_output call
+            // Skip past the tool call line and any immediate formatting
+            if let Some(content_start) = content_without_timing[final_output_pos..].find('\n') {
+                let start_pos = final_output_pos + content_start + 1;
+                let final_content = &content_without_timing[start_pos..];
+                
+                // Trim and return the complete content
+                let trimmed = final_content.trim();
+                if !trimmed.is_empty() {
+                    return trimmed.to_string();
+                }
+            }
+        }
+        
+        // Fallback to the original extract_last_block behavior if we can't find final_output
+        // This maintains backward compatibility
+        self.extract_last_block()
+    }
+
+    /// Extract the last block from the response (for coach feedback in autonomous mode)
+    /// This looks for the final_output content which is the last substantial block
+    pub fn extract_last_block(&self) -> String {
+        // Remove any timing information at the end
+        let content_without_timing = if let Some(timing_pos) = self.response.rfind("\n⏱️") {
+            &self.response[..timing_pos]
+        } else {
+            &self.response
+        };
+        
+        // Split by double newlines to find the last substantial block
+        let blocks: Vec<&str> = content_without_timing.split("\n\n").collect();
+        
+        // Find the last non-empty block that isn't just whitespace
+        blocks.iter()
+            .rev()
+            .find(|block| !block.trim().is_empty())
+            .map(|block| block.trim().to_string())
+            .unwrap_or_else(|| {
+                // Fallback: if we can't find a clear block, take the whole thing
+                content_without_timing.trim().to_string()
+            })
+    }
+
+    /// Check if the response contains an approval (for autonomous mode)
+    pub fn is_approved(&self) -> bool {
+        self.extract_final_output().contains("IMPLEMENTATION_APPROVED")
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn test_extract_last_block() {
+        // Test case 1: Response with timing info
+        let context_window = ContextWindow::new(1000);
+        let response_with_timing = "Some initial content\n\nFinal block content\n\n⏱️ 2.3s | 💭 1.2s".to_string();
+        let result = TaskResult::new(response_with_timing, context_window.clone());
+        assert_eq!(result.extract_last_block(), "Final block content");
+        
+        // Test case 2: Response without timing
+        let response_no_timing = "Some initial content\n\nFinal block content".to_string();
+        let result = TaskResult::new(response_no_timing, context_window.clone());
+        assert_eq!(result.extract_last_block(), "Final block content");
+        
+        // Test case 3: Response with IMPLEMENTATION_APPROVED
+        let response_approved = "Some content\n\nIMPLEMENTATION_APPROVED".to_string();
+        let result = TaskResult::new(response_approved, context_window.clone());
+        assert!(result.is_approved());
+        
+        // Test case 4: Response without approval
+        let response_not_approved = "Some content\n\nNeeds more work".to_string();
+        let result = TaskResult::new(response_not_approved, context_window);
+        assert!(!result.is_approved());
+    }
+
+    #[test]
+    fn test_extract_last_block_edge_cases() {
+        let context_window = ContextWindow::new(1000);
+        
+        // Test empty response
+        let empty_response = "".to_string();
+        let result = TaskResult::new(empty_response, context_window.clone());
+        assert_eq!(result.extract_last_block(), "");
+        
+        // Test single block
+        let single_block = "Just one block".to_string();
+        let result = TaskResult::new(single_block, context_window.clone());
+        assert_eq!(result.extract_last_block(), "Just one block");
+        
+        // Test multiple empty blocks
+        let multiple_empty = "\n\n\n\nSome content\n\n\n\n".to_string();
+        let result = TaskResult::new(multiple_empty, context_window);
+        assert_eq!(result.extract_last_block(), "Some content");
+    }
+
+    #[test]
+    fn test_extract_final_output() {
+        let context_window = ContextWindow::new(1000);
+        
+        // Test case 1: Response with final_output tool call
+        let response_with_final_output = "Analyzing files...\n\nCalling final_output\n\nThis is the complete feedback\nwith multiple lines\nand important details\n\n⏱️ 2.3s".to_string();
+        let result = TaskResult::new(response_with_final_output, context_window.clone());
+        assert_eq!(result.extract_final_output(), "This is the complete feedback\nwith multiple lines\nand important details");
+        
+        // Test case 2: Response with IMPLEMENTATION_APPROVED in final_output
+        let response_approved = "Review complete\n\nfinal_output called\n\nIMPLEMENTATION_APPROVED".to_string();
+        let result = TaskResult::new(response_approved, context_window.clone());
+        assert_eq!(result.extract_final_output(), "IMPLEMENTATION_APPROVED");
+        assert!(result.is_approved());
+        
+        // Test case 3: Response with detailed feedback in final_output
+        let response_feedback = "Checking implementation...\n\nfinal_output\n\nThe following issues need to be addressed:\n1. Missing error handling in main.rs\n2. Tests are not comprehensive\n3. Documentation needs improvement\n\nPlease fix these issues.".to_string();
+        let result = TaskResult::new(response_feedback, context_window.clone());
+        let extracted = result.extract_final_output();
+        assert!(extracted.contains("The following issues need to be addressed:"));
+        assert!(extracted.contains("1. Missing error handling"));
+        assert!(extracted.contains("Please fix these issues."));
+        assert!(!result.is_approved());
+        
+        // Test case 4: Response without final_output (fallback to extract_last_block)
+        let response_no_final_output = "Some analysis\n\nFinal thoughts here".to_string();
+        let result = TaskResult::new(response_no_final_output, context_window.clone());
+        assert_eq!(result.extract_final_output(), "Final thoughts here");
+        
+        // Test case 5: Empty response
+        let empty_response = "".to_string();
+        let result = TaskResult::new(empty_response, context_window);
+        assert_eq!(result.extract_final_output(), "");
+    }
+}

crates/g3-core/src/task_result_comprehensive_tests.rs

@@ -0,0 +1,247 @@
+use crate::{ContextWindow, TaskResult};
+use g3_providers::{Message, MessageRole};
+use std::sync::Arc;
+
+#[test]
+fn test_task_result_basic_functionality() {
+    // Create a context window with some messages
+    let mut context = ContextWindow::new(10000);
+    context.add_message(Message {
+        role: MessageRole::User,
+        content: "Test message 1".to_string(),
+    });
+    context.add_message(Message {
+        role: MessageRole::Assistant,
+        content: "Response 1".to_string(),
+    });
+    
+    // Create a TaskResult
+    let response = "This is the response\n\nFinal output block".to_string();
+    let result = TaskResult::new(response.clone(), context.clone());
+    
+    // Test basic properties
+    assert_eq!(result.response, response);
+    assert_eq!(result.context_window.conversation_history.len(), 2);
+    assert_eq!(result.context_window.total_tokens, 10000);
+}
+
+#[test]
+fn test_extract_last_block_various_formats() {
+    let context = ContextWindow::new(1000);
+    
+    // Test 1: Standard format with multiple blocks
+    let response1 = "First block\n\nSecond block\n\nThird block".to_string();
+    let result1 = TaskResult::new(response1, context.clone());
+    assert_eq!(result1.extract_last_block(), "Third block");
+    
+    // Test 2: With timing information
+    let response2 = "Content\n\nFinal block\n\n⏱️ 2.3s | 💭 1.2s".to_string();
+    let result2 = TaskResult::new(response2, context.clone());
+    assert_eq!(result2.extract_last_block(), "Final block");
+    
+    // Test 3: Single line response
+    let response3 = "Single line response".to_string();
+    let result3 = TaskResult::new(response3, context.clone());
+    assert_eq!(result3.extract_last_block(), "Single line response");
+    
+    // Test 4: Empty response
+    let response4 = "".to_string();
+    let result4 = TaskResult::new(response4, context.clone());
+    assert_eq!(result4.extract_last_block(), "");
+    
+    // Test 5: Only whitespace
+    let response5 = "\n\n\n   \n\n".to_string();
+    let result5 = TaskResult::new(response5, context.clone());
+    assert_eq!(result5.extract_last_block(), "");
+    
+    // Test 6: Multiple blocks with empty ones
+    let response6 = "First\n\n\n\n\n\nLast block here".to_string();
+    let result6 = TaskResult::new(response6, context.clone());
+    assert_eq!(result6.extract_last_block(), "Last block here");
+}
+
+#[test]
+fn test_is_approved_detection() {
+    let context = ContextWindow::new(1000);
+    
+    // Test approved cases
+    let approved_responses = vec![
+        "Analysis complete\n\nIMPLEMENTATION_APPROVED",
+        "Some content\n\nThe implementation is good. IMPLEMENTATION_APPROVED",
+        "IMPLEMENTATION_APPROVED",
+        "Review done\n\n✅ IMPLEMENTATION_APPROVED - All tests pass",
+    ];
+    
+    for response in approved_responses {
+        let result = TaskResult::new(response.to_string(), context.clone());
+        assert!(result.is_approved(), "Failed to detect approval in: {}", response);
+    }
+    
+    // Test not approved cases
+    let not_approved_responses = vec![
+        "Needs more work",
+        "Implementation needs fixes",
+        "IMPLEMENTATION_REJECTED",
+        "Almost there but not APPROVED",
+        "",
+    ];
+    
+    for response in not_approved_responses {
+        let result = TaskResult::new(response.to_string(), context.clone());
+        assert!(!result.is_approved(), "Incorrectly detected approval in: {}", response);
+    }
+}
+
+#[test]
+fn test_context_window_preservation() {
+    // Create a context window with specific state
+    let mut context = ContextWindow::new(5000);
+    context.used_tokens = 1234;
+    
+    // Add some messages
+    for i in 0..5 {
+        context.add_message(Message {
+            role: if i % 2 == 0 { MessageRole::User } else { MessageRole::Assistant },
+            content: format!("Message {}", i),
+        });
+    }
+    
+    // Create TaskResult
+    let result = TaskResult::new("Response".to_string(), context.clone());
+    
+    // Verify context is preserved
+    assert_eq!(result.context_window.total_tokens, 5000);
+    assert!(result.context_window.used_tokens > 1234); // Should have increased
+    assert_eq!(result.context_window.conversation_history.len(), 5);
+    
+    // Verify messages are preserved correctly
+    for i in 0..5 {
+        let is_user = matches!(result.context_window.conversation_history[i].role, MessageRole::User);
+        let expected_is_user = i % 2 == 0;
+        assert_eq!(is_user, expected_is_user, "Message {} has wrong role", i);
+        assert_eq!(result.context_window.conversation_history[i].content, format!("Message {}", i));
+    }
+}
+
+#[test]
+fn test_coach_feedback_extraction_scenarios() {
+    let context = ContextWindow::new(1000);
+    
+    // Scenario 1: Coach feedback with file operations and analysis
+    let coach_response = r#"Reading file: src/main.rs
+📄 File content (23 lines):
+fn main() {
+    println!("Hello");
+}
+
+Analyzing implementation...
+
+The implementation needs the following fixes:
+1. Add error handling
+2. Implement missing functions
+3. Add tests"#;
+    
+    let result = TaskResult::new(coach_response.to_string(), context.clone());
+    let feedback = result.extract_last_block();
+    assert!(feedback.contains("Add error handling"));
+    assert!(feedback.contains("Implement missing functions"));
+    assert!(feedback.contains("Add tests"));
+    
+    // Scenario 2: Coach approval
+    let approval_response = r#"Checking compilation...
+✅ Build successful
+
+Running tests...
+✅ All tests pass
+
+IMPLEMENTATION_APPROVED"#;
+    
+    let result = TaskResult::new(approval_response.to_string(), context.clone());
+    assert!(result.is_approved());
+    assert_eq!(result.extract_last_block(), "IMPLEMENTATION_APPROVED");
+    
+    // Scenario 3: Complex feedback with timing
+    let complex_response = r#"Tool execution log...
+
+Analysis complete.
+
+The following issues were found:
+- Memory leak in process_data()
+- Missing input validation
+
+⏱️ 5.2s | 💭 2.1s"#;
+    
+    let result = TaskResult::new(complex_response.to_string(), context.clone());
+    let feedback = result.extract_last_block();
+    assert!(feedback.contains("Memory leak"));
+    assert!(feedback.contains("Missing input validation"));
+    assert!(!feedback.contains("⏱️")); // Timing should be stripped
+}
+
+#[test]
+fn test_edge_cases_and_special_characters() {
+    let context = ContextWindow::new(1000);
+    
+    // Test with special characters and emojis
+    let response_with_emojis = "First part 🚀\n\n✅ Final part with emojis 🎉".to_string();
+    let result = TaskResult::new(response_with_emojis, context.clone());
+    assert_eq!(result.extract_last_block(), "✅ Final part with emojis 🎉");
+    
+    // Test with code blocks
+    let response_with_code = "Explanation\n\n```rust\nfn main() {}\n```\n\nFinal comment".to_string();
+    let result = TaskResult::new(response_with_code, context.clone());
+    assert_eq!(result.extract_last_block(), "Final comment");
+    
+    // Test with mixed newlines
+    let mixed_newlines = "Part 1\r\n\r\nPart 2\n\nPart 3".to_string();
+    let result = TaskResult::new(mixed_newlines, context.clone());
+    assert_eq!(result.extract_last_block(), "Part 3");
+}
+
+#[test]
+fn test_large_response_handling() {
+    let context = ContextWindow::new(100000);
+    
+    // Create a large response
+    let mut large_response = String::new();
+    for i in 0..100 {
+        large_response.push_str(&format!("Block {} with some content\n\n", i));
+    }
+    large_response.push_str("This is the final block after 100 other blocks");
+    
+    let result = TaskResult::new(large_response, context);
+    assert_eq!(result.extract_last_block(), "This is the final block after 100 other blocks");
+}
+
+#[test]
+fn test_concurrent_access() {
+    use std::thread;
+    
+    let context = ContextWindow::new(1000);
+    let result = Arc::new(TaskResult::new(
+        "Concurrent test\n\nFinal block".to_string(),
+        context,
+    ));
+    
+    let mut handles = vec![];
+    
+    // Spawn multiple threads to access the TaskResult
+    for _ in 0..10 {
+        let result_clone = Arc::clone(&result);
+        let handle = thread::spawn(move || {
+            // Each thread extracts the last block
+            let block = result_clone.extract_last_block();
+            assert_eq!(block, "Final block");
+            
+            // Check approval status
+            assert!(!result_clone.is_approved());
+        });
+        handles.push(handle);
+    }
+    
+    // Wait for all threads to complete
+    for handle in handles {
+        handle.join().unwrap();
+    }
+}
+

crates/g3-core/src/task_result_tests.rs

@@ -0,0 +1,48 @@
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn test_extract_last_block() {
+        // Test case 1: Response with timing info
+        let context_window = ContextWindow::new(1000);
+        let response_with_timing = "Some initial content\n\nFinal block content\n\n⏱️ 2.3s | 💭 1.2s".to_string();
+        let result = TaskResult::new(response_with_timing, context_window.clone());
+        assert_eq!(result.extract_last_block(), "Final block content");
+        
+        // Test case 2: Response without timing
+        let response_no_timing = "Some initial content\n\nFinal block content".to_string();
+        let result = TaskResult::new(response_no_timing, context_window.clone());
+        assert_eq!(result.extract_last_block(), "Final block content");
+        
+        // Test case 3: Response with IMPLEMENTATION_APPROVED
+        let response_approved = "Some content\n\nIMPLEMENTATION_APPROVED".to_string();
+        let result = TaskResult::new(response_approved, context_window.clone());
+        assert!(result.is_approved());
+        
+        // Test case 4: Response without approval
+        let response_not_approved = "Some content\n\nNeeds more work".to_string();
+        let result = TaskResult::new(response_not_approved, context_window);
+        assert!(!result.is_approved());
+    }
+
+    #[test]
+    fn test_extract_last_block_edge_cases() {
+        let context_window = ContextWindow::new(1000);
+        
+        // Test empty response
+        let empty_response = "".to_string();
+        let result = TaskResult::new(empty_response, context_window.clone());
+        assert_eq!(result.extract_last_block(), "");
+        
+        // Test single block
+        let single_block = "Just one block".to_string();
+        let result = TaskResult::new(single_block, context_window.clone());
+        assert_eq!(result.extract_last_block(), "Just one block");
+        
+        // Test multiple empty blocks
+        let multiple_empty = "\n\n\n\nSome content\n\n\n\n".to_string();
+        let result = TaskResult::new(multiple_empty, context_window);
+        assert_eq!(result.extract_last_block(), "Some content");
+    }
+}

crates/g3-core/src/tilde_expansion_tests.rs

@@ -0,0 +1,36 @@
+#[cfg(test)]
+mod tilde_expansion_tests {
+    use std::env;
+
+    #[test]
+    fn test_tilde_expansion() {
+        // Test that shellexpand works
+        let path_with_tilde = "~/test.txt";
+        let expanded = shellexpand::tilde(path_with_tilde);
+        
+        // Get the actual home directory
+        let home = env::var("HOME").expect("HOME environment variable not set");
+        
+        // Verify expansion happened
+        assert_eq!(expanded.as_ref(), format!("{}/test.txt", home));
+        assert!(!expanded.contains("~"));
+    }
+
+    #[test]
+    fn test_tilde_expansion_with_subdirs() {
+        let path_with_tilde = "~/Documents/test.txt";
+        let expanded = shellexpand::tilde(path_with_tilde);
+        
+        let home = env::var("HOME").expect("HOME environment variable not set");
+        
+        assert_eq!(expanded.as_ref(), format!("{}/Documents/test.txt", home));
+    }
+
+    #[test]
+    fn test_no_tilde_unchanged() {
+        let path_without_tilde = "/absolute/path/test.txt";
+        let expanded = shellexpand::tilde(path_without_tilde);
+        
+        assert_eq!(expanded.as_ref(), path_without_tilde);
+    }
+}

crates/g3-core/src/ui_writer.rs

@@ -0,0 +1,83 @@
+/// Interface for UI output operations
+/// This trait abstracts all UI operations to allow different implementations
+/// (console, TUI, web, etc.) without coupling the core logic to specific output methods.
+pub trait UiWriter: Send + Sync {
+    /// Print a simple message
+    fn print(&self, message: &str);
+    
+    /// Print a message with a newline
+    fn println(&self, message: &str);
+    
+    /// Print without newline (for progress indicators)
+    fn print_inline(&self, message: &str);
+    
+    /// Print a system prompt section
+    fn print_system_prompt(&self, prompt: &str);
+    
+    /// Print a context window status message
+    fn print_context_status(&self, message: &str);
+    
+    /// Print a context thinning success message with highlight and animation
+    fn print_context_thinning(&self, message: &str);
+    
+    /// Print a tool execution header
+    fn print_tool_header(&self, tool_name: &str);
+    
+    /// Print a tool argument
+    fn print_tool_arg(&self, key: &str, value: &str);
+    
+    /// Print tool output header
+    fn print_tool_output_header(&self);
+    
+    /// Update the current tool output line (replaces previous line)
+    fn update_tool_output_line(&self, line: &str);
+    
+    /// Print a tool output line
+    fn print_tool_output_line(&self, line: &str);
+    
+    /// Print tool output summary (when output is truncated)
+    fn print_tool_output_summary(&self, hidden_count: usize);
+    
+    /// Print tool execution timing
+    fn print_tool_timing(&self, duration_str: &str);
+    
+    /// Print the agent prompt indicator
+    fn print_agent_prompt(&self);
+    
+    /// Print agent response inline (for streaming)
+    fn print_agent_response(&self, content: &str);
+    
+    /// Notify that an SSE event was received (including pings)
+    fn notify_sse_received(&self);
+    
+    /// Flush any buffered output
+    fn flush(&self);
+    
+    /// Returns true if this UI writer wants full, untruncated output
+    /// Default is false (truncate for human readability)
+    fn wants_full_output(&self) -> bool { false }
+}
+
+/// A no-op implementation for when UI output is not needed
+pub struct NullUiWriter;
+
+impl UiWriter for NullUiWriter {
+    fn print(&self, _message: &str) {}
+    fn println(&self, _message: &str) {}
+    fn print_inline(&self, _message: &str) {}
+    fn print_system_prompt(&self, _prompt: &str) {}
+    fn print_context_status(&self, _message: &str) {}
+    fn print_context_thinning(&self, _message: &str) {}
+    fn print_tool_header(&self, _tool_name: &str) {}
+    fn print_tool_arg(&self, _key: &str, _value: &str) {}
+    fn print_tool_output_header(&self) {}
+    fn update_tool_output_line(&self, _line: &str) {}
+    fn print_tool_output_line(&self, _line: &str) {}
+    fn print_tool_output_summary(&self, _hidden_count: usize) {}
+    fn print_tool_timing(&self, _duration_str: &str) {}
+    fn print_agent_prompt(&self) {}
+    fn print_agent_response(&self, _content: &str) {}
+    fn notify_sse_received(&self) {}
+    fn flush(&self) {}
+    fn wants_full_output(&self) -> bool { false }
+}

crates/g3-core/tests/test_context_thinning.rs

@@ -0,0 +1,270 @@
+use g3_core::ContextWindow;
+use g3_providers::{Message, MessageRole};
+
+#[test]
+fn test_thinning_thresholds() {
+    let mut context = ContextWindow::new(10000);
+    
+    // At 0%, should not thin
+    assert!(!context.should_thin());
+    
+    // Simulate reaching 50% usage
+    context.used_tokens = 5000;
+    assert!(context.should_thin());
+    
+    // After thinning at 50%, should not thin again until next threshold
+    context.last_thinning_percentage = 50;
+    assert!(!context.should_thin());
+    
+    // At 60%, should thin again
+    context.used_tokens = 6000;
+    assert!(context.should_thin());
+    
+    // After thinning at 60%, should not thin
+    context.last_thinning_percentage = 60;
+    assert!(!context.should_thin());
+    
+    // At 70%, should thin
+    context.used_tokens = 7000;
+    assert!(context.should_thin());
+    
+    // At 80%, should thin
+    context.last_thinning_percentage = 70;
+    context.used_tokens = 8000;
+    assert!(context.should_thin());
+    
+    // After 80%, should not thin (compaction takes over)
+    context.last_thinning_percentage = 80;
+    context.used_tokens = 8500;
+    assert!(!context.should_thin());
+}
+
+#[test]
+fn test_thin_context_basic() {
+    let mut context = ContextWindow::new(10000);
+    
+    // Add some messages to the first third
+    for i in 0..9 {
+        if i % 2 == 0 {
+            context.add_message(Message {
+                role: MessageRole::Assistant,
+                content: format!("Assistant message {}", i),
+            });
+        } else {
+            // Add tool results with varying sizes
+            let content = if i == 1 {
+                // Large tool result (> 1000 chars)
+                format!("Tool result: {}", "x".repeat(1500))
+            } else if i == 3 {
+                // Another large tool result
+                format!("Tool result: {}", "y".repeat(2000))
+            } else {
+                // Small tool result (< 1000 chars)
+                format!("Tool result: small result {}", i)
+            };
+            
+            context.add_message(Message {
+                role: MessageRole::User,
+                content,
+            });
+        }
+    }
+    
+    // Trigger thinning at 50%
+    context.used_tokens = 5000;
+    let (summary, _chars_saved) = context.thin_context();
+    
+    println!("Thinning summary: {}", summary);
+    
+    // Should have thinned at least 1 large tool result in the first third
+    assert!(summary.contains("1 tool result"), "Summary was: {}", summary);
+    assert!(summary.contains("50%"));
+    
+    // Check that the large tool results were replaced
+    let first_third_end = context.conversation_history.len() / 3;
+    for i in 0..first_third_end {
+        if let Some(msg) = context.conversation_history.get(i) {
+            if matches!(msg.role, MessageRole::User) && msg.content.starts_with("Tool result:") {
+                if msg.content.len() > 1000 {
+                    panic!("Found un-thinned large tool result at index {}", i);
+                }
+            }
+        }
+    }
+}
+
+#[test]
+fn test_thin_write_file_tool_calls() {
+    let mut context = ContextWindow::new(10000);
+    
+    // Add some messages including a write_file tool call with large content
+    context.add_message(Message {
+        role: MessageRole::User,
+        content: "Please create a large file".to_string(),
+    });
+    
+    // Add an assistant message with a write_file tool call containing large content
+    let large_content = "x".repeat(1500);
+    let tool_call_json = format!(
+        r#"{{"tool": "write_file", "args": {{"file_path": "test.txt", "content": "{}"}}}}"#,
+        large_content
+    );
+    context.add_message(Message {
+        role: MessageRole::Assistant,
+        content: format!("I'll create that file.\n\n{}", tool_call_json),
+    });
+    
+    context.add_message(Message {
+        role: MessageRole::User,
+        content: "Tool result: ✅ Successfully wrote 1500 lines".to_string(),
+    });
+    
+    // Add more messages to ensure we have enough for "first third" logic
+    for i in 0..6 {
+        context.add_message(Message {
+            role: MessageRole::Assistant,
+            content: format!("Response {}", i),
+        });
+    }
+    
+    // Trigger thinning at 50%
+    context.used_tokens = 5000;
+    let (summary, _chars_saved) = context.thin_context();
+    
+    println!("Thinning summary: {}", summary);
+    
+    // Should have thinned the write_file tool call
+    assert!(summary.contains("tool call") || summary.contains("chars saved"));
+    
+    // Check that the large content was replaced with a file reference
+    let first_third_end = context.conversation_history.len() / 3;
+    for i in 0..first_third_end {
+        if let Some(msg) = context.conversation_history.get(i) {
+            if matches!(msg.role, MessageRole::Assistant) && msg.content.contains("write_file") {
+                // The content should now reference an external file
+                assert!(msg.content.contains("<content saved to"));
+                assert!(!msg.content.contains(&large_content));
+            }
+        }
+    }
+}
+
+#[test]
+fn test_thin_str_replace_tool_calls() {
+    let mut context = ContextWindow::new(10000);
+    
+    // Add some messages including a str_replace tool call with large diff
+    context.add_message(Message {
+        role: MessageRole::User,
+        content: "Please update the file".to_string(),
+    });
+    
+    // Add an assistant message with a str_replace tool call containing large diff
+    let large_diff = format!("--- old\n{}\n+++ new\n{}", "-old line\n".repeat(100), "+new line\n".repeat(100));
+    let tool_call_json = format!(
+        r#"{{"tool": "str_replace", "args": {{"file_path": "test.txt", "diff": "{}"}}}}"#,
+        large_diff.replace('\n', "\\n")
+    );
+    context.add_message(Message {
+        role: MessageRole::Assistant,
+        content: format!("I'll update that file.\n\n{}", tool_call_json),
+    });
+    
+    context.add_message(Message {
+        role: MessageRole::User,
+        content: "Tool result: ✅ applied unified diff".to_string(),
+    });
+    
+    // Add more messages to ensure we have enough for "first third" logic
+    for i in 0..6 {
+        context.add_message(Message {
+            role: MessageRole::Assistant,
+            content: format!("Response {}", i),
+        });
+    }
+    
+    // Trigger thinning at 50%
+    context.used_tokens = 5000;
+    let (summary, _chars_saved) = context.thin_context();
+    
+    println!("Thinning summary: {}", summary);
+    
+    // Should have thinned the str_replace tool call
+    assert!(summary.contains("tool call") || summary.contains("chars saved"));
+    
+    // Check that the large diff was replaced with a file reference
+    let first_third_end = context.conversation_history.len() / 3;
+    for i in 0..first_third_end {
+        if let Some(msg) = context.conversation_history.get(i) {
+            if matches!(msg.role, MessageRole::Assistant) && msg.content.contains("str_replace") {
+                // The diff should now reference an external file
+                assert!(msg.content.contains("<diff saved to"));
+                // Should not contain the large diff content
+                assert!(!msg.content.contains("old line"));
+            }
+        }
+    }
+}
+
+#[test]
+fn test_thin_context_no_large_results() {
+    let mut context = ContextWindow::new(10000);
+    
+    // Add only small messages
+    for i in 0..9 {
+        context.add_message(Message {
+            role: MessageRole::User,
+            content: format!("Tool result: small {}", i),
+        });
+    }
+    
+    context.used_tokens = 5000;
+    let (summary, _chars_saved) = context.thin_context();
+    
+    // Should report no large results found
+    assert!(summary.contains("no large tool results or tool calls found"));
+}
+
+#[test]
+fn test_thin_context_only_affects_first_third() {
+    let mut context = ContextWindow::new(10000);
+    
+    // Add 12 messages (first third = 4 messages)
+    for i in 0..12 {
+        let content = if i % 2 == 1 {
+            // All odd indices are large tool results
+            format!("Tool result: {}", "x".repeat(1500))
+        } else {
+            format!("Assistant message {}", i)
+        };
+        
+        let role = if i % 2 == 1 {
+            MessageRole::User
+        } else {
+            MessageRole::Assistant
+        };
+        
+        context.add_message(Message { role, content });
+    }
+    
+    context.used_tokens = 5000;
+    let (summary, _chars_saved) = context.thin_context();
+    
+    // First third is 4 messages (indices 0-3), so only indices 1 and 3 should be thinned
+    // That's 2 tool results
+    assert!(summary.contains("2 tool results"));
+    
+    // Check that messages after the first third are NOT thinned
+    let first_third_end = context.conversation_history.len() / 3;
+    for i in first_third_end..context.conversation_history.len() {
+        if let Some(msg) = context.conversation_history.get(i) {
+            if matches!(msg.role, MessageRole::User) && msg.content.starts_with("Tool result:") {
+                // These should still be large (not thinned)
+                if i % 2 == 1 {
+                    assert!(msg.content.len() > 1000, 
+                        "Message at index {} should not have been thinned", i);
+                }
+            }
+        }
+    }
+}

crates/g3-core/tests/test_token_counting.rs

@@ -0,0 +1,94 @@
+use g3_core::ContextWindow;
+use g3_providers::Usage;
+
+#[test]
+fn test_token_accumulation() {
+    let mut window = ContextWindow::new(10000);
+    
+    // First API call: 100 prompt + 50 completion = 150 total
+    let usage1 = Usage {
+        prompt_tokens: 100,
+        completion_tokens: 50,
+        total_tokens: 150,
+    };
+    window.update_usage_from_response(&usage1);
+    assert_eq!(window.used_tokens, 150, "First call should have 150 tokens");
+    assert_eq!(window.cumulative_tokens, 150, "Cumulative should be 150");
+
+    // Second API call: 200 prompt + 75 completion = 275 total
+    let usage2 = Usage {
+        prompt_tokens: 200,
+        completion_tokens: 75,
+        total_tokens: 275,
+    };
+    window.update_usage_from_response(&usage2);
+    assert_eq!(window.used_tokens, 425, "Second call should accumulate to 425 tokens");
+    assert_eq!(window.cumulative_tokens, 425, "Cumulative should be 425");
+
+    // Third API call with SMALLER token count: 50 prompt + 25 completion = 75 total
+    let usage3 = Usage {
+        prompt_tokens: 50,
+        completion_tokens: 25,
+        total_tokens: 75,
+    };
+    window.update_usage_from_response(&usage3);
+    assert_eq!(window.used_tokens, 500, "Third call should accumulate to 500 tokens");
+    assert_eq!(window.cumulative_tokens, 500, "Cumulative should be 500");
+    
+    // Verify tokens never decrease
+    assert!(window.used_tokens >= 425, "Token count should never decrease!");
+}
+
+#[test]
+fn test_add_streaming_tokens() {
+    let mut window = ContextWindow::new(10000);
+    
+    // Add some streaming tokens
+    window.add_streaming_tokens(100);
+    assert_eq!(window.used_tokens, 100);
+    assert_eq!(window.cumulative_tokens, 100);
+    
+    // Add more
+    window.add_streaming_tokens(50);
+    assert_eq!(window.used_tokens, 150);
+    assert_eq!(window.cumulative_tokens, 150);
+    
+    // Now update from provider response
+    let usage = Usage {
+        prompt_tokens: 80,
+        completion_tokens: 40,
+        total_tokens: 120,
+    };
+    window.update_usage_from_response(&usage);
+    
+    // Should ADD to existing, not replace
+    assert_eq!(window.used_tokens, 270, "Should add 120 to existing 150");
+    assert_eq!(window.cumulative_tokens, 270);
+}
+
+#[test]
+fn test_percentage_calculation() {
+    let mut window = ContextWindow::new(1000);
+    
+    // Add tokens via provider response
+    let usage = Usage {
+        prompt_tokens: 150,
+        completion_tokens: 100,
+        total_tokens: 250,
+    };
+    window.update_usage_from_response(&usage);
+    
+    assert_eq!(window.percentage_used(), 25.0);
+    assert_eq!(window.remaining_tokens(), 750);
+    
+    // Add more tokens
+    let usage2 = Usage {
+        prompt_tokens: 300,
+        completion_tokens: 200,
+        total_tokens: 500,
+    };
+    window.update_usage_from_response(&usage2);
+    
+    assert_eq!(window.percentage_used(), 75.0);
+    assert_eq!(window.remaining_tokens(), 250);
+}

crates/g3-execution/Cargo.toml

@@ -7,6 +7,7 @@ description = "Code execution engine for G3 AI agent"
 [dependencies]
 tokio = { workspace = true }
 anyhow = { workspace = true }
+futures = "0.3"
 thiserror = { workspace = true }
 tracing = { workspace = true }
 regex = "1.0"

crates/g3-execution/src/lib.rs

@@ -166,6 +166,31 @@ impl CodeExecutor {
     
     /// Execute Bash code
     async fn execute_bash(&self, code: &str) -> Result<ExecutionResult> {
+        // Check if this is a detached/daemon command that should run independently
+        let is_detached = code.trim_start().starts_with("setsid ") 
+            || code.trim_start().starts_with("nohup ")
+            || code.contains(" disown")
+            || (code.contains(" &") && (code.contains("nohup") || code.contains("setsid")));
+        
+        if is_detached {
+            // For detached commands, just spawn and return immediately
+            use std::process::Stdio;
+            Command::new("bash")
+                .arg("-c")
+                .arg(code)
+                .stdin(Stdio::null())
+                .stdout(Stdio::null())
+                .stderr(Stdio::null())
+                .spawn()?;
+            
+            return Ok(ExecutionResult {
+                stdout: "✅ Command launched in background (detached process)".to_string(),
+                stderr: String::new(),
+                exit_code: 0,
+                success: true,
+            });
+        }
+        
         let output = Command::new("bash")
             .arg("-c")
             .arg(code)
@@ -203,3 +228,105 @@ impl Default for CodeExecutor {
         Self::new()
     }
 }
+
+/// Trait for receiving streaming output from command execution
+pub trait OutputReceiver: Send + Sync {
+    /// Called when a new line of output is available
+    fn on_output_line(&self, line: &str);
+}
+
+impl CodeExecutor {
+    /// Execute bash command with streaming output
+    pub async fn execute_bash_streaming<R: OutputReceiver>(
+        &self, 
+        code: &str, 
+        receiver: &R
+    ) -> Result<ExecutionResult> {
+        use std::process::Stdio;
+        use tokio::io::{AsyncBufReadExt, BufReader};
+        use tokio::process::Command as TokioCommand;
+        
+        // Check if this is a detached/daemon command that should run independently
+        // Look for patterns like: setsid, nohup with &, or explicit backgrounding with disown
+        let is_detached = code.trim_start().starts_with("setsid ") 
+            || code.trim_start().starts_with("nohup ")
+            || code.contains(" disown")
+            || (code.contains(" &") && (code.contains("nohup") || code.contains("setsid")));
+        
+        if is_detached {
+            // For detached commands, just spawn and return immediately
+            TokioCommand::new("bash")
+                .arg("-c")
+                .arg(code)
+                .spawn()?;
+            
+            // Don't wait for the process - it's meant to run independently
+            return Ok(ExecutionResult {
+                stdout: "✅ Command launched in background (detached process)".to_string(),
+                stderr: String::new(),
+                exit_code: 0,
+                success: true,
+            });
+        }
+        
+        let mut child = TokioCommand::new("bash")
+            .arg("-c")
+            .arg(code)
+            .stdout(Stdio::piped())
+            .stderr(Stdio::piped())
+            .spawn()?;
+        
+        let stdout = child.stdout.take().unwrap();
+        let stderr = child.stderr.take().unwrap();
+        
+        let stdout_reader = BufReader::new(stdout);
+        let stderr_reader = BufReader::new(stderr);
+        
+        let mut stdout_lines = stdout_reader.lines();
+        let mut stderr_lines = stderr_reader.lines();
+        
+        let mut stdout_output = Vec::new();
+        let mut stderr_output = Vec::new();
+        
+        // Read output lines as they come
+        loop {
+            tokio::select! {
+                line = stdout_lines.next_line() => {
+                    match line {
+                        Ok(Some(line)) => {
+                            receiver.on_output_line(&line);
+                            stdout_output.push(line);
+                        }
+                        Ok(None) => break, // EOF
+                        Err(e) => {
+                            error!("Error reading stdout: {}", e);
+                            break;
+                        }
+                    }
+                }
+                line = stderr_lines.next_line() => {
+                    match line {
+                        Ok(Some(line)) => {
+                            receiver.on_output_line(&line.to_string());
+                            stderr_output.push(line);
+                        }
+                        Ok(None) => {}, // stderr EOF, continue
+                        Err(e) => {
+                            error!("Error reading stderr: {}", e);
+                        }
+                    }
+                }
+                else => break
+            }
+        }
+        
+        let status = child.wait().await?;
+        
+        Ok(ExecutionResult {
+            stdout: stdout_output.join("\n"),
+            stderr: stderr_output.join("\n"),
+            exit_code: status.code().unwrap_or(-1),
+            success: status.success(),
+        })
+    }
+}

crates/g3-providers/Cargo.toml

@@ -16,3 +16,16 @@ async-trait = "0.1"
 tokio-stream = "0.1"
 futures-util = "0.3"
 bytes = "1.0"
+# OAuth dependencies
+axum = "0.7"
+base64 = "0.22"
+chrono = { version = "0.4", features = ["serde"] }
+sha2 = "0.10"
+url = "2.5"
+webbrowser = "1.0"
+nanoid = "0.4"
+serde_urlencoded = "0.7"
+tokio-util = "0.7"
+dirs = "5.0"
+llama_cpp = { version = "0.3.2", features = ["metal"] }
+shellexpand = "3.1"

crates/g3-providers/src/anthropic.rs

@@ -41,6 +41,7 @@
 //!         max_tokens: Some(1000),
 //!         temperature: Some(0.7),
 //!         stream: false,
+//!         tools: None,
 //!     };
 //!
 //!     // Get a completion
@@ -74,6 +75,7 @@
 //!         max_tokens: Some(1000),
 //!         temperature: Some(0.7),
 //!         stream: true,
+//!         tools: None,
 //!     };
 //!
 //!     let mut stream = provider.stream(request).await?;
@@ -154,8 +156,9 @@ impl AnthropicProvider {
             .post(ANTHROPIC_API_URL)
             .header("x-api-key", &self.api_key)
             .header("anthropic-version", ANTHROPIC_VERSION)
+            // Anthropic beta 1m context window. Enable if needed. It costs extra, so check first.
+            // .header("anthropic-beta", "context-1m-2025-08-07")
             .header("content-type", "application/json");
-
         if streaming {
             builder = builder.header("accept", "text/event-stream");
         }
@@ -194,20 +197,6 @@ impl AnthropicProvider {
             .collect()
     }
 
-    fn convert_anthropic_tool_calls(&self, content: &[AnthropicContent]) -> Vec<ToolCall> {
-        content
-            .iter()
-            .filter_map(|c| match c {
-                AnthropicContent::ToolUse { id, name, input } => Some(ToolCall {
-                    id: id.clone(),
-                    tool: name.clone(),
-                    args: input.clone(),
-                }),
-                _ => None,
-            })
-            .collect()
-    }
-
     fn convert_messages(&self, messages: &[Message]) -> Result<(Option<String>, Vec<AnthropicMessage>)> {
         let mut system_message = None;
         let mut anthropic_messages = Vec::new();
@@ -281,26 +270,43 @@ impl AnthropicProvider {
         &self,
         mut stream: impl futures_util::Stream<Item = reqwest::Result<Bytes>> + Unpin,
         tx: mpsc::Sender<Result<CompletionChunk>>,
-    ) {
+    ) -> Option<Usage> {
         let mut buffer = String::new();
         let mut current_tool_calls: Vec<ToolCall> = Vec::new();
         let mut partial_tool_json = String::new(); // Accumulate partial JSON for tool calls
-
+        let mut accumulated_usage: Option<Usage> = None;
+        let mut byte_buffer = Vec::new(); // Buffer for incomplete UTF-8 sequences
+        let mut message_stopped = false; // Track if we've received message_stop
+        
         while let Some(chunk_result) = stream.next().await {
             match chunk_result {
                 Ok(chunk) => {
-                    let chunk_str = match std::str::from_utf8(&chunk) {
-                        Ok(s) => s,
+                    // Append new bytes to our buffer
+                    byte_buffer.extend_from_slice(&chunk);
+                    
+                    // Try to convert the entire buffer to UTF-8
+                    let chunk_str = match std::str::from_utf8(&byte_buffer) {
+                        Ok(s) => {
+                            // Successfully converted entire buffer, clear it and use the string
+                            let result = s.to_string();
+                            byte_buffer.clear();
+                            result
+                        }
                         Err(e) => {
-                            error!("Invalid UTF-8 in stream chunk: {}", e);
-                            let _ = tx
-                                .send(Err(anyhow!("Invalid UTF-8 in stream chunk: {}", e)))
-                                .await;
-                            return;
+                            // Check if this is an incomplete sequence at the end
+                            let valid_up_to = e.valid_up_to();
+                            if valid_up_to > 0 {
+                                // We have some valid UTF-8, extract it and keep the rest for next iteration
+                                let valid_bytes = byte_buffer.drain(..valid_up_to).collect::<Vec<_>>();
+                                std::str::from_utf8(&valid_bytes).unwrap().to_string()
+                            } else {
+                                // No valid UTF-8 at all, skip this chunk and continue
+                                continue;
+                            }
                         }
                     };
 
-                    buffer.push_str(chunk_str);
+                    buffer.push_str(&chunk_str);
 
                     // Process complete lines
                     while let Some(line_end) = buffer.find('\n') {
@@ -311,6 +317,12 @@ impl AnthropicProvider {
                             continue;
                         }
 
+                        // If we've already sent the final chunk, skip processing more events
+                        if message_stopped {
+                            debug!("Skipping event after message_stop: {}", line);
+                            continue;
+                        }
+
                         // Parse Server-Sent Events format
                         if let Some(data) = line.strip_prefix("data: ") {
                             if data == "[DONE]" {
@@ -318,20 +330,34 @@ impl AnthropicProvider {
                                 let final_chunk = CompletionChunk {
                                     content: String::new(),
                                     finished: true,
+                                    usage: accumulated_usage.clone(),
                                     tool_calls: if current_tool_calls.is_empty() { None } else { Some(current_tool_calls.clone()) },
                                 };
                                 if tx.send(Ok(final_chunk)).await.is_err() {
                                     debug!("Receiver dropped, stopping stream");
                                 }
-                                return;
+                                return accumulated_usage;
                             }
 
                             debug!("Raw Claude API JSON: {}", data);
 
                             match serde_json::from_str::<AnthropicStreamEvent>(data) {
                                 Ok(event) => {
-                                    debug!("Parsed event: {:?}", event);
+                                    debug!("Parsed event type: {}, event: {:?}", event.event_type, event);
                                     match event.event_type.as_str() {
+                                        "message_start" => {
+                                            // Extract usage data from message_start event
+                                            if let Some(message) = event.message {
+                                                if let Some(usage) = message.usage {
+                                                    accumulated_usage = Some(Usage {
+                                                        prompt_tokens: usage.input_tokens,
+                                                        completion_tokens: usage.output_tokens,
+                                                        total_tokens: usage.input_tokens + usage.output_tokens,
+                                                    });
+                                                    debug!("Captured usage from message_start: {:?}", accumulated_usage);
+                                                }
+                                            }
+                                        }
                                         "content_block_start" => {
                                             debug!("Received content_block_start event: {:?}", event);
                                             if let Some(content_block) = event.content_block {
@@ -354,11 +380,12 @@ impl AnthropicProvider {
                                                             let chunk = CompletionChunk {
                                                                 content: String::new(),
                                                                 finished: false,
+                                                                usage: None,
                                                                 tool_calls: Some(vec![tool_call]),
                                                             };
                                                             if tx.send(Ok(chunk)).await.is_err() {
                                                                 debug!("Receiver dropped, stopping stream");
-                                                                return;
+                                                                return accumulated_usage;
                                                             }
                                                         } else {
                                                             // Arguments are empty, we'll accumulate them from partial_json
@@ -380,11 +407,12 @@ impl AnthropicProvider {
                                                     let chunk = CompletionChunk {
                                                         content: text,
                                                         finished: false,
+                                                        usage: None,
                                                         tool_calls: None,
                                                     };
                                                     if tx.send(Ok(chunk)).await.is_err() {
                                                         debug!("Receiver dropped, stopping stream");
-                                                        return;
+                                                        return accumulated_usage;
                                                     }
                                                 }
                                                 // Handle partial JSON for tool calls
@@ -419,25 +447,29 @@ impl AnthropicProvider {
                                                 let chunk = CompletionChunk {
                                                     content: String::new(),
                                                     finished: false,
+                                                    usage: None,
                                                     tool_calls: Some(current_tool_calls.clone()),
                                                 };
                                                 if tx.send(Ok(chunk)).await.is_err() {
                                                     debug!("Receiver dropped, stopping stream");
-                                                    return;
+                                                    return accumulated_usage;
                                                 }
                                             }
                                         }
                                         "message_stop" => {
                                             debug!("Received message stop event");
+                                            message_stopped = true;
                                             let final_chunk = CompletionChunk {
                                                 content: String::new(),
                                                 finished: true,
+                                                usage: accumulated_usage.clone(),
                                                 tool_calls: if current_tool_calls.is_empty() { None } else { Some(current_tool_calls.clone()) },
                                             };
                                             if tx.send(Ok(final_chunk)).await.is_err() {
                                                 debug!("Receiver dropped, stopping stream");
                                             }
-                                            return;
+                                            // Don't return here - let the stream naturally exhaust
+                                            // This prevents dropping the sender prematurely
                                         }
                                         "error" => {
                                             if let Some(error) = event.error {
@@ -445,7 +477,7 @@ impl AnthropicProvider {
                                                 let _ = tx
                                                     .send(Err(anyhow!("Anthropic API error: {:?}", error)))
                                                     .await;
-                                                return;
+                                                break; // Break to let stream exhaust naturally
                                             }
                                         }
                                         _ => {
@@ -464,7 +496,10 @@ impl AnthropicProvider {
                 Err(e) => {
                     error!("Stream error: {}", e);
                     let _ = tx.send(Err(anyhow!("Stream error: {}", e))).await;
-                    return;
+                    // Don't return here either - let the stream exhaust naturally
+                    // The error has been sent to the receiver, so it will handle it
+                    // Breaking here ensures we clean up properly
+                    break;
                 }
             }
         }
@@ -473,9 +508,11 @@ impl AnthropicProvider {
         let final_chunk = CompletionChunk {
             content: String::new(),
             finished: true,
+            usage: accumulated_usage.clone(),
             tool_calls: if current_tool_calls.is_empty() { None } else { Some(current_tool_calls) },
         };
         let _ = tx.send(Ok(final_chunk)).await;
+        accumulated_usage
     }
 }
 
@@ -596,7 +633,14 @@ impl LLMProvider for AnthropicProvider {
         // Spawn task to process the stream
         let provider = self.clone();
         tokio::spawn(async move {
-            provider.parse_streaming_response(stream, tx).await;
+            let usage = provider.parse_streaming_response(stream, tx).await;
+            // Log the final usage if available
+            if let Some(usage) = usage {
+                debug!(
+                    "Stream completed with usage - prompt: {}, completion: {}, total: {}",
+                    usage.prompt_tokens, usage.completion_tokens, usage.total_tokens
+                );
+            }
         });
 
         Ok(ReceiverStream::new(rx))
@@ -668,14 +712,8 @@ enum AnthropicContent {
 
 #[derive(Debug, Deserialize)]
 struct AnthropicResponse {
-    id: String,
-    #[serde(rename = "type")]
-    response_type: String,
-    role: String,
     content: Vec<AnthropicContent>,
     model: String,
-    stop_reason: Option<String>,
-    stop_sequence: Option<String>,
     usage: AnthropicUsage,
 }
 
@@ -697,12 +735,18 @@ struct AnthropicStreamEvent {
     error: Option<AnthropicError>,
     #[serde(default)]
     content_block: Option<AnthropicContent>,
+    #[serde(default)]
+    message: Option<AnthropicStreamMessage>,
+}
+
+#[derive(Debug, Deserialize)]
+struct AnthropicStreamMessage {
+    #[serde(default)]
+    usage: Option<AnthropicUsage>,
 }
 
 #[derive(Debug, Deserialize)]
 struct AnthropicDelta {
-    #[serde(rename = "type")]
-    delta_type: Option<String>,
     text: Option<String>,
     partial_json: Option<String>,
 }
@@ -710,7 +754,9 @@ struct AnthropicDelta {
 #[derive(Debug, Deserialize)]
 struct AnthropicError {
     #[serde(rename = "type")]
+    #[allow(dead_code)]
     error_type: String,
+    #[allow(dead_code)]
     message: String,
 }
 
@@ -813,32 +859,4 @@ mod tests {
         assert!(anthropic_tools[0].input_schema.required.is_some());
         assert_eq!(anthropic_tools[0].input_schema.required.as_ref().unwrap()[0], "location");
     }
-
-    #[test]
-    fn test_tool_call_conversion() {
-        let provider = AnthropicProvider::new(
-            "test-key".to_string(),
-            None,
-            None,
-            None,
-        ).unwrap();
-
-        let content = vec![
-            AnthropicContent::Text {
-                text: "I'll help you get the weather.".to_string(),
-            },
-            AnthropicContent::ToolUse {
-                id: "toolu_123".to_string(),
-                name: "get_weather".to_string(),
-                input: serde_json::json!({"location": "San Francisco, CA"}),
-            },
-        ];
-
-        let tool_calls = provider.convert_anthropic_tool_calls(&content);
-
-        assert_eq!(tool_calls.len(), 1);
-        assert_eq!(tool_calls[0].id, "toolu_123");
-        assert_eq!(tool_calls[0].tool, "get_weather");
-        assert_eq!(tool_calls[0].args["location"], "San Francisco, CA");
-    }
 }

crates/g3-providers/src/embedded.rs

@@ -1,5 +1,5 @@
 use anyhow::Result;
-use g3_providers::{
+use crate::{
     CompletionChunk, CompletionRequest, CompletionResponse, CompletionStream, LLMProvider, Message,
     MessageRole, Usage,
 };
@@ -8,22 +8,18 @@ use llama_cpp::{
     LlamaModel, LlamaParams, LlamaSession, SessionParams,
 };
 use std::path::{Path, PathBuf};
-use std::sync::atomic::{AtomicBool, AtomicU64, Ordering};
 use std::sync::Arc;
-use std::time::Duration;
 use tokio::sync::mpsc;
 use tokio::sync::Mutex;
 use tokio_stream::wrappers::ReceiverStream;
-use tracing::{debug, error, info, warn};
+use tracing::{debug, error, info};
 
 pub struct EmbeddedProvider {
-    model: Arc<LlamaModel>,
     session: Arc<Mutex<LlamaSession>>,
     model_name: String,
     max_tokens: u32,
     temperature: f32,
     context_length: u32,
-    generation_active: Arc<AtomicBool>,
 }
 
 impl EmbeddedProvider {
@@ -71,8 +67,10 @@ impl EmbeddedProvider {
             .map_err(|e| anyhow::anyhow!("Failed to load model: {}", e))?;
 
         // Create session with parameters
-        let mut session_params = SessionParams::default();
-        session_params.n_ctx = context_size;
+        let mut session_params = SessionParams {
+            n_ctx: context_size,
+            ..Default::default()
+        };
         if let Some(threads) = threads {
             session_params.n_threads = threads;
         }
@@ -84,13 +82,11 @@ impl EmbeddedProvider {
         info!("Successfully loaded {} model", model_type);
 
         Ok(Self {
-            model: Arc::new(model),
             session: Arc::new(Mutex::new(session)),
             model_name: format!("embedded-{}", model_type),
             max_tokens: max_tokens.unwrap_or(2048),
             temperature: temperature.unwrap_or(0.1),
             context_length: context_size,
-            generation_active: Arc::new(AtomicBool::new(false)),
         })
     }
 
@@ -143,7 +139,7 @@ impl EmbeddedProvider {
                         in_conversation = false;
                     }
                     MessageRole::Assistant => {
-                        formatted.push_str(" ");
+                        formatted.push(' ');
                         formatted.push_str(&message.content);
                         formatted.push_str("</s> ");
                         in_conversation = false;
@@ -152,8 +148,8 @@ impl EmbeddedProvider {
             }
             
             // If the last message was from user, add a space for the assistant's response
-            if messages.last().map_or(false, |m| matches!(m.role, MessageRole::User)) {
-                formatted.push_str(" ");
+            if messages.last().is_some_and(|m| matches!(m.role, MessageRole::User)) {
+                formatted.push(' ');
             }
             
             formatted
@@ -429,7 +425,6 @@ impl EmbeddedProvider {
     // Download the Qwen 2.5 7B model if it doesn't exist
     fn download_qwen_model(model_path: &Path) -> Result<()> {
         use std::fs;
-        use std::io::Write;
         use std::process::Command;
         
         const MODEL_URL: &str = "https://huggingface.co/Qwen/Qwen2.5-7B-Instruct-GGUF/resolve/main/qwen2.5-7b-instruct-q3_k_m.gguf";
@@ -446,7 +441,7 @@ impl EmbeddedProvider {
         
         // Use curl with progress bar for download
         let output = Command::new("curl")
-            .args(&[
+            .args([
                 "-L",  // Follow redirects
                 "-#",  // Show progress bar
                 "-f",  // Fail on HTTP errors
@@ -662,6 +657,7 @@ impl LLMProvider for EmbeddedProvider {
                                     let chunk = CompletionChunk {
                                         content: remaining_to_send.to_string(),
                                         finished: false,
+                                        usage: None,
                                         tool_calls: None,
                                     };
                                     let _ = tx.blocking_send(Ok(chunk));
@@ -688,6 +684,7 @@ impl LLMProvider for EmbeddedProvider {
                             let chunk = CompletionChunk {
                                 content: remaining_to_send.to_string(),
                                 finished: false,
+                                usage: None,
                                 tool_calls: None,
                             };
                             let _ = tx.blocking_send(Ok(chunk));
@@ -721,6 +718,7 @@ impl LLMProvider for EmbeddedProvider {
                             let chunk = CompletionChunk {
                                 content: to_send.to_string(),
                                 finished: false,
+                                usage: None,
                                 tool_calls: None,
                             };
                             if tx.blocking_send(Ok(chunk)).is_err() {
@@ -736,6 +734,7 @@ impl LLMProvider for EmbeddedProvider {
                         let chunk = CompletionChunk {
                             content: unsent_tokens.clone(),
                             finished: false,
+                            usage: None,
                             tool_calls: None,
                         };
                         if tx.blocking_send(Ok(chunk)).is_err() {
@@ -756,6 +755,7 @@ impl LLMProvider for EmbeddedProvider {
             let final_chunk = CompletionChunk {
                 content: String::new(),
                 finished: true,
+                usage: None,  // Embedded models calculate usage differently
                 tool_calls: None,
             };
             let _ = tx.blocking_send(Ok(final_chunk));

crates/g3-providers/src/lib.rs

@@ -67,6 +67,7 @@ pub struct CompletionChunk {
     pub content: String,
     pub finished: bool,
     pub tool_calls: Option<Vec<ToolCall>>,
+    pub usage: Option<Usage>,  // Add usage tracking for streaming
 }
 
 #[derive(Debug, Clone, Serialize, Deserialize)]
@@ -84,8 +85,17 @@ pub struct Tool {
 }
 
 pub mod anthropic;
+pub mod databricks;
+pub mod embedded;
+pub mod oauth;
+pub mod ollama;
+pub mod openai;
 
 pub use anthropic::AnthropicProvider;
+pub use databricks::DatabricksProvider;
+pub use embedded::EmbeddedProvider;
+pub use ollama::OllamaProvider;
+pub use openai::OpenAIProvider;
 
 /// Provider registry for managing multiple LLM providers
 pub struct ProviderRegistry {

crates/g3-providers/src/oauth.rs

@@ -0,0 +1,463 @@
+use anyhow::Result;
+use axum::{extract::Query, response::Html, routing::get, Router};
+use base64::Engine;
+use chrono::{DateTime, Utc};
+use serde::{Deserialize, Serialize};
+use serde_json::Value;
+use sha2::Digest;
+use std::{collections::HashMap, fs, net::SocketAddr, path::PathBuf, sync::Arc};
+use tokio::sync::{oneshot, Mutex as TokioMutex};
+use url::Url;
+
+#[derive(Debug, Clone)]
+struct OidcEndpoints {
+    authorization_endpoint: String,
+    token_endpoint: String,
+}
+
+#[derive(Serialize, Deserialize)]
+struct TokenData {
+    /// The access token used to authenticate API requests
+    access_token: String,
+
+    /// Optional refresh token that can be used to obtain a new access token
+    /// when the current one expires, enabling offline access without user interaction
+    refresh_token: Option<String>,
+
+    /// When the access token expires (if known)
+    /// Used to determine when a token needs to be refreshed
+    expires_at: Option<DateTime<Utc>>,
+}
+
+struct TokenCache {
+    cache_path: PathBuf,
+}
+
+fn get_base_path() -> PathBuf {
+    // Use a similar pattern to Goose but for g3
+    // macOS/Linux: ~/.config/g3/databricks/oauth
+    // Windows: ~\AppData\Roaming\g3\config\databricks\oauth\
+    let mut path = dirs::config_dir().unwrap_or_else(|| PathBuf::from("."));
+    path.push("g3");
+    path.push("databricks");
+    path.push("oauth");
+    path
+}
+
+impl TokenCache {
+    fn new(host: &str, client_id: &str, scopes: &[String]) -> Self {
+        let mut hasher = sha2::Sha256::new();
+        hasher.update(host.as_bytes());
+        hasher.update(client_id.as_bytes());
+        hasher.update(scopes.join(",").as_bytes());
+        let hash = format!("{:x}", hasher.finalize());
+
+        fs::create_dir_all(get_base_path()).unwrap_or(());
+        let cache_path = get_base_path().join(format!("{}.json", hash));
+
+        Self { cache_path }
+    }
+
+    fn load_token(&self) -> Option<TokenData> {
+        if let Ok(contents) = fs::read_to_string(&self.cache_path) {
+            if let Ok(token_data) = serde_json::from_str::<TokenData>(&contents) {
+                // Only return tokens that have a refresh token
+                if token_data.refresh_token.is_some() {
+                    // If token is not expired, return it for immediate use
+                    if let Some(expires_at) = token_data.expires_at {
+                        if expires_at > Utc::now() {
+                            return Some(token_data);
+                        }
+                        // If token is expired but has refresh token, return it so we can refresh
+                        return Some(token_data);
+                    }
+                    // No expiration time but has refresh token, return it
+                    return Some(token_data);
+                }
+                // Token doesn't have a refresh token, ignore it to force a new OAuth flow
+            }
+        }
+        None
+    }
+
+    fn save_token(&self, token_data: &TokenData) -> Result<()> {
+        if let Some(parent) = self.cache_path.parent() {
+            fs::create_dir_all(parent)?;
+        }
+        let contents = serde_json::to_string(token_data)?;
+        fs::write(&self.cache_path, contents)?;
+        Ok(())
+    }
+}
+
+async fn get_workspace_endpoints(host: &str) -> Result<OidcEndpoints> {
+    let base_url = Url::parse(host).expect("Invalid host URL");
+    let oidc_url = base_url
+        .join("oidc/.well-known/oauth-authorization-server")
+        .expect("Invalid OIDC URL");
+
+    let client = reqwest::Client::new();
+    let resp = client.get(oidc_url.clone()).send().await?;
+
+    if !resp.status().is_success() {
+        return Err(anyhow::anyhow!(
+            "Failed to get OIDC configuration from {}",
+            oidc_url
+        ));
+    }
+
+    let oidc_config: Value = resp.json().await?;
+
+    let authorization_endpoint = oidc_config
+        .get("authorization_endpoint")
+        .and_then(|v| v.as_str())
+        .ok_or_else(|| anyhow::anyhow!("authorization_endpoint not found in OIDC configuration"))?
+        .to_string();
+
+    let token_endpoint = oidc_config
+        .get("token_endpoint")
+        .and_then(|v| v.as_str())
+        .ok_or_else(|| anyhow::anyhow!("token_endpoint not found in OIDC configuration"))?
+        .to_string();
+
+    Ok(OidcEndpoints {
+        authorization_endpoint,
+        token_endpoint,
+    })
+}
+
+struct OAuthFlow {
+    endpoints: OidcEndpoints,
+    client_id: String,
+    redirect_url: String,
+    scopes: Vec<String>,
+    state: String,
+    verifier: String,
+}
+
+impl OAuthFlow {
+    fn new(
+        endpoints: OidcEndpoints,
+        client_id: String,
+        redirect_url: String,
+        scopes: Vec<String>,
+    ) -> Self {
+        Self {
+            endpoints,
+            client_id,
+            redirect_url,
+            scopes,
+            state: nanoid::nanoid!(16),
+            verifier: nanoid::nanoid!(64),
+        }
+    }
+
+    /// Extracts token data from an OAuth 2.0 token response.
+    fn extract_token_data(
+        &self,
+        token_response: &Value,
+        old_refresh_token: Option<&str>,
+    ) -> Result<TokenData> {
+        // Extract access token (required)
+        let access_token = token_response
+            .get("access_token")
+            .and_then(|v| v.as_str())
+            .ok_or_else(|| anyhow::anyhow!("access_token not found in token response"))?
+            .to_string();
+
+        // Extract refresh token if available
+        let refresh_token = token_response
+            .get("refresh_token")
+            .and_then(|v| v.as_str())
+            .map(|s| s.to_string())
+            .or_else(|| old_refresh_token.map(|s| s.to_string()));
+
+        // Handle token expiration
+        let expires_at =
+            if let Some(expires_in) = token_response.get("expires_in").and_then(|v| v.as_u64()) {
+                // Traditional OAuth flow with expires_in seconds
+                Some(Utc::now() + chrono::Duration::seconds(expires_in as i64))
+            } else {
+                // If the server doesn't provide any expiration info, log it but don't set an expiration
+                tracing::debug!(
+                    "No expiration information provided by server, token expiration unknown."
+                );
+                None
+            };
+
+        Ok(TokenData {
+            access_token,
+            refresh_token,
+            expires_at,
+        })
+    }
+
+    fn get_authorization_url(&self) -> String {
+        let challenge = {
+            let digest = sha2::Sha256::digest(self.verifier.as_bytes());
+            base64::engine::general_purpose::URL_SAFE_NO_PAD.encode(digest)
+        };
+
+        let params = [
+            ("response_type", "code"),
+            ("client_id", &self.client_id),
+            ("redirect_uri", &self.redirect_url),
+            ("scope", &self.scopes.join(" ")),
+            ("state", &self.state),
+            ("code_challenge", &challenge),
+            ("code_challenge_method", "S256"),
+        ];
+
+        format!(
+            "{}?{}",
+            self.endpoints.authorization_endpoint,
+            serde_urlencoded::to_string(params).unwrap()
+        )
+    }
+
+    async fn exchange_code_for_token(&self, code: &str) -> Result<TokenData> {
+        let params = [
+            ("grant_type", "authorization_code"),
+            ("code", code),
+            ("redirect_uri", &self.redirect_url),
+            ("code_verifier", &self.verifier),
+            ("client_id", &self.client_id),
+        ];
+
+        let client = reqwest::Client::new();
+        let resp = client
+            .post(&self.endpoints.token_endpoint)
+            .header("Content-Type", "application/x-www-form-urlencoded")
+            .form(&params)
+            .send()
+            .await?;
+
+        if !resp.status().is_success() {
+            let err_text = resp.text().await?;
+            return Err(anyhow::anyhow!(
+                "Failed to exchange code for token: {}",
+                err_text
+            ));
+        }
+
+        let token_response: Value = resp.json().await?;
+        self.extract_token_data(&token_response, None)
+    }
+
+    async fn refresh_token(&self, refresh_token: &str) -> Result<TokenData> {
+        let params = [
+            ("grant_type", "refresh_token"),
+            ("refresh_token", refresh_token),
+            ("client_id", &self.client_id),
+        ];
+
+        tracing::debug!("Refreshing token using refresh_token");
+
+        let client = reqwest::Client::new();
+        let resp = client
+            .post(&self.endpoints.token_endpoint)
+            .header("Content-Type", "application/x-www-form-urlencoded")
+            .form(&params)
+            .send()
+            .await?;
+
+        if !resp.status().is_success() {
+            let err_text = resp.text().await?;
+            return Err(anyhow::anyhow!("Failed to refresh token: {}", err_text));
+        }
+
+        let token_response: Value = resp.json().await?;
+        self.extract_token_data(&token_response, Some(refresh_token))
+    }
+
+    async fn execute(&self) -> Result<TokenData> {
+        // Create a channel that will send the auth code from the app process
+        let (tx, rx) = oneshot::channel();
+        let state = self.state.clone();
+        let tx = Arc::new(TokioMutex::new(Some(tx)));
+
+        // Setup a server that will receive the redirect, capture the code, and display success/failure
+        let app = Router::new().route(
+            "/",
+            get(move |Query(params): Query<HashMap<String, String>>| {
+                let tx = Arc::clone(&tx);
+                let state = state.clone();
+                async move {
+                    let code = params.get("code").cloned();
+                    let received_state = params.get("state").cloned();
+
+                    if let (Some(code), Some(received_state)) = (code, received_state) {
+                        if received_state == state {
+                            if let Some(sender) = tx.lock().await.take() {
+                                if sender.send(code).is_ok() {
+                                    return Html(
+                                        "<h2>G3 Authentication Success</h2><p>You can close this window and return to your terminal.</p>",
+                                    );
+                                }
+                            }
+                            Html("<h2>Error</h2><p>Authentication already completed.</p>")
+                        } else {
+                            Html("<h2>Error</h2><p>State mismatch.</p>")
+                        }
+                    } else {
+                        Html("<h2>Error</h2><p>Authentication failed.</p>")
+                    }
+                }
+            }),
+        );
+
+        // Start the server to accept the oauth code
+        let redirect_url = Url::parse(&self.redirect_url)?;
+        let port = redirect_url.port().unwrap_or(80);
+        let addr = SocketAddr::from(([127, 0, 0, 1], port));
+
+        let listener = tokio::net::TcpListener::bind(addr).await?;
+
+        let server_handle = tokio::spawn(async move {
+            let server = axum::serve(listener, app);
+            server.await.unwrap();
+        });
+
+        // Open the browser which will redirect with the code to the server
+        let authorization_url = self.get_authorization_url();
+        if std::env::var("G3_RETRO_MODE").is_err() {
+            println!("🔐 Opening browser for Databricks authentication...");
+        }
+        if webbrowser::open(&authorization_url).is_err() {
+            println!(
+                "Please open this URL in your browser:\n{}",
+                authorization_url
+            );
+        }
+
+        // Wait for the authorization code with a timeout
+        let code = tokio::time::timeout(
+            std::time::Duration::from_secs(120), // 2 minute timeout
+            rx,
+        )
+        .await
+        .map_err(|_| anyhow::anyhow!("Authentication timed out after 2 minutes"))??;
+
+        // Stop the server
+        server_handle.abort();
+
+        if std::env::var("G3_RETRO_MODE").is_err() {
+            println!("✅ Authentication successful! Exchanging code for token...");
+        }
+
+        // Exchange the code for a token
+        self.exchange_code_for_token(&code).await
+    }
+}
+
+pub async fn get_oauth_token_async(
+    host: &str,
+    client_id: &str,
+    redirect_url: &str,
+    scopes: &[String],
+) -> Result<String> {
+    let token_cache = TokenCache::new(host, client_id, scopes);
+
+    // Try cache first
+    if let Some(token) = token_cache.load_token() {
+        // If token has an expiration time, check if it's expired
+        if let Some(expires_at) = token.expires_at {
+            if expires_at > Utc::now() {
+                tracing::debug!("Using cached token");
+                return Ok(token.access_token);
+            }
+            // Token is expired, will try to refresh below
+            tracing::debug!("Token is expired, attempting to refresh");
+        } else {
+            // No expiration time was provided by the server
+            tracing::debug!("Token has no expiration time, using cached token");
+            return Ok(token.access_token);
+        }
+
+        // Token is expired or has no expiration, try to refresh if we have a refresh token
+        if let Some(refresh_token) = token.refresh_token {
+            // Get endpoints for token refresh
+            match get_workspace_endpoints(host).await {
+                Ok(endpoints) => {
+                    let flow = OAuthFlow::new(
+                        endpoints,
+                        client_id.to_string(),
+                        redirect_url.to_string(),
+                        scopes.to_vec(),
+                    );
+
+                    // Try to refresh the token
+                    match flow.refresh_token(&refresh_token).await {
+                        Ok(new_token) => {
+                            if let Err(e) = token_cache.save_token(&new_token) {
+                                tracing::warn!("Failed to save refreshed token: {}", e);
+                            }
+                            tracing::info!("Successfully refreshed token");
+                            return Ok(new_token.access_token);
+                        }
+                        Err(e) => {
+                            tracing::warn!(
+                                "Failed to refresh token, will try new auth flow: {}",
+                                e
+                            );
+                            // Continue to new auth flow
+                        }
+                    }
+                }
+                Err(e) => {
+                    tracing::warn!("Failed to get endpoints for token refresh: {}", e);
+                    // Continue to new auth flow
+                }
+            }
+        }
+    }
+
+    // Get endpoints and execute flow for a new token
+    let endpoints = get_workspace_endpoints(host).await?;
+    let flow = OAuthFlow::new(
+        endpoints,
+        client_id.to_string(),
+        redirect_url.to_string(),
+        scopes.to_vec(),
+    );
+
+    // Execute the OAuth flow and get token
+    let token = flow.execute().await?;
+
+    // Cache and return
+    token_cache.save_token(&token)?;
+    if std::env::var("G3_RETRO_MODE").is_err() {
+        println!("🎉 Databricks authentication complete!");
+    }
+    Ok(token.access_token)
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn test_token_cache() -> Result<()> {
+        let cache = TokenCache::new(
+            "https://example.com",
+            "test-client",
+            &["scope1".to_string()],
+        );
+
+        // Test with expiration time
+        let token_data = TokenData {
+            access_token: "test-token".to_string(),
+            refresh_token: Some("test-refresh-token".to_string()),
+            expires_at: Some(Utc::now() + chrono::Duration::hours(1)),
+        };
+
+        cache.save_token(&token_data)?;
+
+        let loaded_token = cache.load_token().unwrap();
+        assert_eq!(loaded_token.access_token, token_data.access_token);
+        assert_eq!(loaded_token.refresh_token, token_data.refresh_token);
+        assert!(loaded_token.expires_at.is_some());
+
+        Ok(())
+    }
+}

crates/g3-providers/src/ollama.rs

@@ -0,0 +1,751 @@
+//! Ollama LLM provider implementation for the g3-providers crate.
+//!
+//! This module provides an implementation of the `LLMProvider` trait for Ollama,
+//! supporting both completion and streaming modes with native tool calling.
+//!
+//! # Features
+//!
+//! - Support for any Ollama model (llama3.2, mistral, qwen, etc.)
+//! - Both completion and streaming response modes
+//! - Native tool calling support for compatible models
+//! - Configurable base URL (defaults to http://localhost:11434)
+//! - Simple configuration with no authentication required
+//!
+//! # Usage
+//!
+//! ```rust,no_run
+//! use g3_providers::{OllamaProvider, LLMProvider, CompletionRequest, Message, MessageRole};
+//!
+//! #[tokio::main]
+//! async fn main() -> anyhow::Result<()> {
+//!     // Create the provider with default settings (localhost:11434)
+//!     let provider = OllamaProvider::new(
+//!         "llama3.2".to_string(),
+//!         None, // Optional: base_url
+//!         None, // Optional: max tokens
+//!         None, // Optional: temperature
+//!     )?;
+//!
+//!     // Create a completion request
+//!     let request = CompletionRequest {
+//!         messages: vec![
+//!             Message {
+//!                 role: MessageRole::User,
+//!                 content: "Hello! How are you?".to_string(),
+//!             },
+//!         ],
+//!         max_tokens: Some(1000),
+//!         temperature: Some(0.7),
+//!         stream: false,
+//!         tools: None,
+//!     };
+//!
+//!     // Get a completion
+//!     let response = provider.complete(request).await?;
+//!     println!("Response: {}", response.content);
+//!
+//!     Ok(())
+//! }
+//! ```
+
+use anyhow::{anyhow, Result};
+use bytes::Bytes;
+use futures_util::stream::StreamExt;
+use reqwest::Client;
+use serde::{Deserialize, Serialize};
+use std::time::Duration;
+use tokio::sync::mpsc;
+use tokio_stream::wrappers::ReceiverStream;
+use tracing::{debug, error, info, warn};
+
+use crate::{
+    CompletionChunk, CompletionRequest, CompletionResponse, CompletionStream, LLMProvider, Message,
+    MessageRole, Tool, ToolCall, Usage,
+};
+
+const DEFAULT_BASE_URL: &str = "http://localhost:11434";
+const DEFAULT_TIMEOUT_SECS: u64 = 600;
+
+pub const OLLAMA_DEFAULT_MODEL: &str = "llama3.2";
+pub const OLLAMA_KNOWN_MODELS: &[&str] = &[
+    "llama3.2",
+    "llama3.2:1b",
+    "llama3.2:3b",
+    "llama3.1",
+    "llama3.1:8b",
+    "llama3.1:70b",
+    "mistral",
+    "mistral-nemo",
+    "mixtral",
+    "qwen2.5",
+    "qwen2.5:7b",
+    "qwen2.5:14b",
+    "qwen2.5:32b",
+    "qwen2.5-coder",
+    "qwen2.5-coder:7b",
+    "qwen3-coder",
+    "phi3",
+    "gemma2",
+];
+
+#[derive(Debug, Clone)]
+pub struct OllamaProvider {
+    client: Client,
+    base_url: String,
+    model: String,
+    max_tokens: Option<u32>,
+    temperature: f32,
+}
+
+impl OllamaProvider {
+    pub fn new(
+        model: String,
+        base_url: Option<String>,
+        max_tokens: Option<u32>,
+        temperature: Option<f32>,
+    ) -> Result<Self> {
+        let client = Client::builder()
+            .timeout(Duration::from_secs(DEFAULT_TIMEOUT_SECS))
+            .build()
+            .map_err(|e| anyhow!("Failed to create HTTP client: {}", e))?;
+
+        let base_url = base_url
+            .unwrap_or_else(|| DEFAULT_BASE_URL.to_string())
+            .trim_end_matches('/')
+            .to_string();
+
+        info!(
+            "Initialized Ollama provider with model: {} at {}",
+            model, base_url
+        );
+
+        Ok(Self {
+            client,
+            base_url,
+            model,
+            max_tokens,
+            temperature: temperature.unwrap_or(0.7),
+        })
+    }
+
+    fn convert_tools(&self, tools: &[Tool]) -> Vec<OllamaTool> {
+        tools
+            .iter()
+            .map(|tool| OllamaTool {
+                r#type: "function".to_string(),
+                function: OllamaFunction {
+                    name: tool.name.clone(),
+                    description: tool.description.clone(),
+                    parameters: tool.input_schema.clone(),
+                },
+            })
+            .collect()
+    }
+
+    fn convert_messages(&self, messages: &[Message]) -> Result<Vec<OllamaMessage>> {
+        let mut ollama_messages = Vec::new();
+
+        for message in messages {
+            let role = match message.role {
+                MessageRole::System => "system",
+                MessageRole::User => "user",
+                MessageRole::Assistant => "assistant",
+            };
+
+            ollama_messages.push(OllamaMessage {
+                role: role.to_string(),
+                content: message.content.clone(),
+                tool_calls: None, // Only used in responses
+            });
+        }
+
+        if ollama_messages.is_empty() {
+            return Err(anyhow!("At least one message is required"));
+        }
+
+        Ok(ollama_messages)
+    }
+
+    fn create_request_body(
+        &self,
+        messages: &[Message],
+        tools: Option<&[Tool]>,
+        streaming: bool,
+        max_tokens: Option<u32>,
+        temperature: f32,
+    ) -> Result<OllamaRequest> {
+        let ollama_messages = self.convert_messages(messages)?;
+        let ollama_tools = tools.map(|t| self.convert_tools(t));
+
+        let mut options = OllamaOptions {
+            temperature,
+            num_predict: max_tokens,
+        };
+
+        // If max_tokens is provided, use it; otherwise use the instance default
+        if max_tokens.is_none() {
+            options.num_predict = self.max_tokens;
+        }
+
+        let request = OllamaRequest {
+            model: self.model.clone(),
+            messages: ollama_messages,
+            tools: ollama_tools,
+            stream: streaming,
+            options,
+        };
+
+        Ok(request)
+    }
+
+    async fn parse_streaming_response(
+        &self,
+        mut stream: impl futures_util::Stream<Item = reqwest::Result<Bytes>> + Unpin,
+        tx: mpsc::Sender<Result<CompletionChunk>>,
+    ) -> Option<Usage> {
+        let mut buffer = String::new();
+        let mut accumulated_usage: Option<Usage> = None;
+        let mut current_tool_calls: Vec<OllamaToolCall> = Vec::new();
+        let mut byte_buffer = Vec::new();
+
+        while let Some(chunk_result) = stream.next().await {
+            match chunk_result {
+                Ok(chunk) => {
+                    // Append new bytes to our buffer
+                    byte_buffer.extend_from_slice(&chunk);
+
+                    // Try to convert the entire buffer to UTF-8
+                    let chunk_str = match std::str::from_utf8(&byte_buffer) {
+                        Ok(s) => {
+                            let result = s.to_string();
+                            byte_buffer.clear();
+                            result
+                        }
+                        Err(e) => {
+                            let valid_up_to = e.valid_up_to();
+                            if valid_up_to > 0 {
+                                let valid_bytes =
+                                    byte_buffer.drain(..valid_up_to).collect::<Vec<_>>();
+                                std::str::from_utf8(&valid_bytes).unwrap().to_string()
+                            } else {
+                                continue;
+                            }
+                        }
+                    };
+
+                    buffer.push_str(&chunk_str);
+
+                    // Process complete lines
+                    while let Some(line_end) = buffer.find('\n') {
+                        let line = buffer[..line_end].trim().to_string();
+                        buffer.drain(..line_end + 1);
+
+                        if line.is_empty() {
+                            continue;
+                        }
+
+                        // Ollama streaming sends JSON objects per line
+                        match serde_json::from_str::<OllamaStreamChunk>(&line) {
+                            Ok(chunk) => {
+                                // Handle text content
+                                if let Some(message) = &chunk.message {
+                                    let content = &message.content;
+                                    if !content.is_empty() {
+                                        debug!("Sending text chunk: '{}'", content);
+                                        let chunk = CompletionChunk {
+                                            content: content.clone(),
+                                            finished: false,
+                                            usage: None,
+                                            tool_calls: None,
+                                        };
+                                        if tx.send(Ok(chunk)).await.is_err() {
+                                            debug!("Receiver dropped, stopping stream");
+                                            return accumulated_usage;
+                                        }
+                                    }
+
+                                    // Handle tool calls
+                                    if let Some(tool_calls) = &message.tool_calls {
+                                        current_tool_calls.extend(tool_calls.clone());
+                                    }
+                                }
+
+                                // Check if stream is done
+                                if chunk.done.unwrap_or(false) {
+                                    debug!("Stream completed");
+
+                                    // Update usage if available
+                                    if let Some(eval_count) = chunk.eval_count {
+                                        accumulated_usage = Some(Usage {
+                                            prompt_tokens: chunk.prompt_eval_count.unwrap_or(0),
+                                            completion_tokens: eval_count,
+                                            total_tokens: chunk.prompt_eval_count.unwrap_or(0)
+                                                + eval_count,
+                                        });
+                                    }
+
+                                    // Send final chunk with tool calls if any
+                                    let final_tool_calls: Vec<ToolCall> = current_tool_calls
+                                        .iter()
+                                        .map(|tc| ToolCall {
+                                            id: tc.function.name.clone(), // Ollama doesn't provide IDs
+                                            tool: tc.function.name.clone(),
+                                            args: tc.function.arguments.clone(),
+                                        })
+                                        .collect();
+
+                                    let final_chunk = CompletionChunk {
+                                        content: String::new(),
+                                        finished: true,
+                                        usage: accumulated_usage.clone(),
+                                        tool_calls: if final_tool_calls.is_empty() {
+                                            None
+                                        } else {
+                                            Some(final_tool_calls)
+                                        },
+                                    };
+                                    if tx.send(Ok(final_chunk)).await.is_err() {
+                                        debug!("Receiver dropped, stopping stream");
+                                    }
+                                    return accumulated_usage;
+                                }
+                            }
+                            Err(e) => {
+                                debug!("Failed to parse Ollama stream chunk: {} - Line: {}", e, line);
+                                // Don't error out, just continue
+                            }
+                        }
+                    }
+                }
+                Err(e) => {
+                    error!("Stream error: {}", e);
+                    let error_msg = e.to_string();
+                    if error_msg.contains("unexpected EOF") || error_msg.contains("connection") {
+                        warn!("Connection terminated unexpectedly, treating as end of stream");
+                        break;
+                    } else {
+                        let _ = tx.send(Err(anyhow!("Stream error: {}", e))).await;
+                    }
+                    return accumulated_usage;
+                }
+            }
+        }
+
+        // Send final chunk if we haven't already
+        let final_tool_calls: Vec<ToolCall> = current_tool_calls
+            .iter()
+            .map(|tc| ToolCall {
+                id: tc.function.name.clone(),
+                tool: tc.function.name.clone(),
+                args: tc.function.arguments.clone(),
+            })
+            .collect();
+
+        let final_chunk = CompletionChunk {
+            content: String::new(),
+            finished: true,
+            usage: accumulated_usage.clone(),
+            tool_calls: if final_tool_calls.is_empty() {
+                None
+            } else {
+                Some(final_tool_calls)
+            },
+        };
+        let _ = tx.send(Ok(final_chunk)).await;
+        accumulated_usage
+    }
+
+    /// Fetch available models from the Ollama instance
+    pub async fn fetch_available_models(&self) -> Result<Vec<String>> {
+        let response = self
+            .client
+            .get(format!("{}/api/tags", self.base_url))
+            .send()
+            .await
+            .map_err(|e| anyhow!("Failed to fetch Ollama models: {}", e))?;
+
+        if !response.status().is_success() {
+            let status = response.status();
+            let error_text = response
+                .text()
+                .await
+                .unwrap_or_else(|_| "Unknown error".to_string());
+            return Err(anyhow!(
+                "Failed to fetch Ollama models: {} - {}",
+                status,
+                error_text
+            ));
+        }
+
+        let json: serde_json::Value = response.json().await?;
+        let models = json
+            .get("models")
+            .and_then(|v| v.as_array())
+            .ok_or_else(|| anyhow!("Unexpected response format: missing 'models' array"))?;
+
+        let model_names: Vec<String> = models
+            .iter()
+            .filter_map(|model| model.get("name").and_then(|n| n.as_str()).map(String::from))
+            .collect();
+
+        debug!("Found {} models in Ollama", model_names.len());
+        Ok(model_names)
+    }
+}
+
+#[async_trait::async_trait]
+impl LLMProvider for OllamaProvider {
+    async fn complete(&self, request: CompletionRequest) -> Result<CompletionResponse> {
+        debug!(
+            "Processing Ollama completion request with {} messages",
+            request.messages.len()
+        );
+
+        let max_tokens = request.max_tokens.or(self.max_tokens);
+        let temperature = request.temperature.unwrap_or(self.temperature);
+
+        let request_body = self.create_request_body(
+            &request.messages,
+            request.tools.as_deref(),
+            false,
+            max_tokens,
+            temperature,
+        )?;
+
+        debug!(
+            "Sending request to Ollama API: model={}, temperature={}",
+            self.model, request_body.options.temperature
+        );
+
+        let response = self
+            .client
+            .post(format!("{}/api/chat", self.base_url))
+            .json(&request_body)
+            .send()
+            .await
+            .map_err(|e| anyhow!("Failed to send request to Ollama API: {}", e))?;
+
+        let status = response.status();
+        if !status.is_success() {
+            let error_text = response
+                .text()
+                .await
+                .unwrap_or_else(|_| "Unknown error".to_string());
+            return Err(anyhow!("Ollama API error {}: {}", status, error_text));
+        }
+
+        let response_text = response.text().await?;
+        debug!("Raw Ollama API response: {}", response_text);
+
+        let ollama_response: OllamaResponse =
+            serde_json::from_str(&response_text).map_err(|e| {
+                anyhow!(
+                    "Failed to parse Ollama response: {} - Response: {}",
+                    e,
+                    response_text
+                )
+            })?;
+
+        let content = ollama_response.message.content.clone();
+
+        let usage = Usage {
+            prompt_tokens: ollama_response.prompt_eval_count.unwrap_or(0),
+            completion_tokens: ollama_response.eval_count.unwrap_or(0),
+            total_tokens: ollama_response.prompt_eval_count.unwrap_or(0)
+                + ollama_response.eval_count.unwrap_or(0),
+        };
+
+        debug!(
+            "Ollama completion successful: {} tokens generated",
+            usage.completion_tokens
+        );
+
+        Ok(CompletionResponse {
+            content,
+            usage,
+            model: self.model.clone(),
+        })
+    }
+
+    async fn stream(&self, request: CompletionRequest) -> Result<CompletionStream> {
+        debug!(
+            "Processing Ollama request (non-streaming) with {} messages",
+            request.messages.len()
+        );
+
+        if let Some(ref tools) = request.tools {
+            debug!("Request has {} tools", tools.len());
+            for tool in tools.iter().take(5) {
+                debug!("  Tool: {}", tool.name);
+            }
+        }
+
+        let max_tokens = request.max_tokens.or(self.max_tokens);
+        let temperature = request.temperature.unwrap_or(self.temperature);
+
+        let request_body = self.create_request_body(
+            &request.messages,
+            request.tools.as_deref(),
+            false, // Use non-streaming mode to avoid streaming bugs
+            max_tokens,
+            temperature,
+        )?;
+
+        debug!(
+            "Sending request to Ollama API (stream=false): model={}, temperature={}",
+            self.model, request_body.options.temperature
+        );
+
+        let response = self
+            .client
+            .post(format!("{}/api/chat", self.base_url))
+            .json(&request_body)
+            .send()
+            .await
+            .map_err(|e| anyhow!("Failed to send request to Ollama API: {}", e))?;
+
+        let status = response.status();
+        if !status.is_success() {
+            let error_text = response
+                .text()
+                .await
+                .unwrap_or_else(|_| "Unknown error".to_string());
+            return Err(anyhow!("Ollama API error {}: {}", status, error_text));
+        }
+
+        // For non-streaming, parse the complete JSON response
+        let response_text = response.text().await?;
+        debug!("Raw Ollama API response: {}", response_text);
+
+        let ollama_response: OllamaResponse =
+            serde_json::from_str(&response_text).map_err(|e| {
+                anyhow!(
+                    "Failed to parse Ollama response: {} - Response: {}",
+                    e,
+                    response_text
+                )
+            })?;
+
+        let (tx, rx) = mpsc::channel(100);
+
+        tokio::spawn(async move {
+            let content = ollama_response.message.content;
+            let usage = Usage {
+                prompt_tokens: ollama_response.prompt_eval_count.unwrap_or(0),
+                completion_tokens: ollama_response.eval_count.unwrap_or(0),
+                total_tokens: ollama_response.prompt_eval_count.unwrap_or(0)
+                    + ollama_response.eval_count.unwrap_or(0),
+            };
+
+            // Extract tool calls if present
+            let tool_calls: Option<Vec<ToolCall>> = ollama_response.message.tool_calls.map(|tcs| {
+                tcs.iter()
+                    .map(|tc| ToolCall {
+                        id: tc.function.name.clone(),
+                        tool: tc.function.name.clone(),
+                        args: tc.function.arguments.clone(),
+                    })
+                    .collect()
+            });
+
+            // Send content if any
+            if !content.is_empty() {
+                let _ = tx.send(Ok(CompletionChunk {
+                    content,
+                    finished: false,
+                    usage: None,
+                    tool_calls: None,
+                })).await;
+            }
+
+            // Send final chunk with usage and tool calls
+            let _ = tx.send(Ok(CompletionChunk {
+                content: String::new(),
+                finished: true,
+                usage: Some(usage),
+                tool_calls,
+            })).await;
+        });
+
+        Ok(ReceiverStream::new(rx))
+    }
+
+    fn name(&self) -> &str {
+        "ollama"
+    }
+
+    fn model(&self) -> &str {
+        &self.model
+    }
+
+    fn has_native_tool_calling(&self) -> bool {
+        // Most modern Ollama models support tool calling
+        // Models like llama3.2, qwen2.5, mistral, etc. have good tool support
+        true
+    }
+}
+
+// Ollama API request/response structures
+
+#[derive(Debug, Serialize)]
+struct OllamaRequest {
+    model: String,
+    messages: Vec<OllamaMessage>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    tools: Option<Vec<OllamaTool>>,
+    stream: bool,
+    options: OllamaOptions,
+}
+
+#[derive(Debug, Serialize)]
+struct OllamaOptions {
+    temperature: f32,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    num_predict: Option<u32>, // Ollama's equivalent of max_tokens
+}
+
+#[derive(Debug, Serialize)]
+struct OllamaTool {
+    r#type: String,
+    function: OllamaFunction,
+}
+
+#[derive(Debug, Serialize)]
+struct OllamaFunction {
+    name: String,
+    description: String,
+    parameters: serde_json::Value,
+}
+
+#[derive(Debug, Clone, Serialize, Deserialize)]
+struct OllamaMessage {
+    role: String,
+    content: String,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    tool_calls: Option<Vec<OllamaToolCall>>,
+}
+
+#[derive(Debug, Clone, Serialize, Deserialize)]
+struct OllamaToolCall {
+    function: OllamaToolCallFunction,
+}
+
+#[derive(Debug, Clone, Serialize, Deserialize)]
+struct OllamaToolCallFunction {
+    name: String,
+    arguments: serde_json::Value,
+}
+
+#[derive(Debug, Deserialize)]
+struct OllamaResponse {
+    message: OllamaMessage,
+    #[allow(dead_code)]
+    done: bool,
+    #[allow(dead_code)]
+    total_duration: Option<u64>,
+    #[allow(dead_code)]
+    load_duration: Option<u64>,
+    prompt_eval_count: Option<u32>,
+    eval_count: Option<u32>,
+}
+
+#[derive(Debug, Deserialize)]
+struct OllamaStreamChunk {
+    message: Option<OllamaMessage>,
+    done: Option<bool>,
+    prompt_eval_count: Option<u32>,
+    eval_count: Option<u32>,
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn test_provider_creation() {
+        let provider = OllamaProvider::new(
+            "llama3.2".to_string(),
+            None,
+            Some(1000),
+            Some(0.7),
+        )
+        .unwrap();
+
+        assert_eq!(provider.model(), "llama3.2");
+        assert_eq!(provider.name(), "ollama");
+        assert!(provider.has_native_tool_calling());
+    }
+
+    #[test]
+    fn test_message_conversion() {
+        let provider = OllamaProvider::new(
+            "llama3.2".to_string(),
+            None,
+            None,
+            None,
+        )
+        .unwrap();
+
+        let messages = vec![
+            Message {
+                role: MessageRole::System,
+                content: "You are a helpful assistant.".to_string(),
+            },
+            Message {
+                role: MessageRole::User,
+                content: "Hello!".to_string(),
+            },
+        ];
+
+        let ollama_messages = provider.convert_messages(&messages).unwrap();
+
+        assert_eq!(ollama_messages.len(), 2);
+        assert_eq!(ollama_messages[0].role, "system");
+        assert_eq!(ollama_messages[1].role, "user");
+    }
+
+    #[test]
+    fn test_tool_conversion() {
+        let provider = OllamaProvider::new(
+            "llama3.2".to_string(),
+            None,
+            None,
+            None,
+        )
+        .unwrap();
+
+        let tools = vec![Tool {
+            name: "get_weather".to_string(),
+            description: "Get the current weather".to_string(),
+            input_schema: serde_json::json!({
+                "type": "object",
+                "properties": {
+                    "location": {
+                        "type": "string",
+                        "description": "The city and state"
+                    }
+                },
+                "required": ["location"]
+            }),
+        }];
+
+        let ollama_tools = provider.convert_tools(&tools);
+
+        assert_eq!(ollama_tools.len(), 1);
+        assert_eq!(ollama_tools[0].r#type, "function");
+        assert_eq!(ollama_tools[0].function.name, "get_weather");
+    }
+
+    #[test]
+    fn test_custom_base_url() {
+        let provider = OllamaProvider::new(
+            "llama3.2".to_string(),
+            Some("http://custom:11434".to_string()),
+            None,
+            None,
+        )
+        .unwrap();
+
+        assert_eq!(provider.base_url, "http://custom:11434");
+    }
+}

crates/g3-providers/src/openai.rs

@@ -0,0 +1,495 @@
+use anyhow::Result;
+use async_trait::async_trait;
+use bytes::Bytes;
+use futures_util::stream::StreamExt;
+use reqwest::Client;
+use serde::Deserialize;
+use serde_json::json;
+use tokio::sync::mpsc;
+use tokio_stream::wrappers::ReceiverStream;
+use tracing::{debug, error};
+
+use crate::{
+    CompletionChunk, CompletionRequest, CompletionResponse, CompletionStream, LLMProvider,
+    Message, MessageRole, Tool, ToolCall, Usage,
+};
+
+#[derive(Clone)]
+pub struct OpenAIProvider {
+    client: Client,
+    api_key: String,
+    model: String,
+    base_url: String,
+    max_tokens: Option<u32>,
+    _temperature: Option<f32>,
+}
+
+impl OpenAIProvider {
+    pub fn new(
+        api_key: String,
+        model: Option<String>,
+        base_url: Option<String>,
+        max_tokens: Option<u32>,
+        temperature: Option<f32>,
+    ) -> Result<Self> {
+        Ok(Self {
+            client: Client::new(),
+            api_key,
+            model: model.unwrap_or_else(|| "gpt-4o".to_string()),
+            base_url: base_url.unwrap_or_else(|| "https://api.openai.com/v1".to_string()),
+            max_tokens,
+            _temperature: temperature,
+        })
+    }
+
+    fn create_request_body(
+        &self,
+        messages: &[Message],
+        tools: Option<&[Tool]>,
+        stream: bool,
+        max_tokens: Option<u32>,
+        _temperature: Option<f32>,
+    ) -> serde_json::Value {
+        let mut body = json!({
+            "model": self.model,
+            "messages": convert_messages(messages),
+            "stream": stream,
+        });
+
+        if let Some(max_tokens) = max_tokens.or(self.max_tokens) {
+            body["max_completion_tokens"] = json!(max_tokens);
+        }
+
+        // OpenAI calls with temp setting seem to fail, so don't send one.
+        // if let Some(temperature) = temperature.or(self.temperature) {
+        //     body["temperature"] = json!(temperature);
+        // }
+
+        if let Some(tools) = tools {
+            if !tools.is_empty() {
+                body["tools"] = json!(convert_tools(tools));
+            }
+        }
+
+        if stream {
+            body["stream_options"] = json!({
+                "include_usage": true,
+            });
+        }
+
+        body
+    }
+
+    async fn parse_streaming_response(
+        &self,
+        mut stream: impl futures_util::Stream<Item = reqwest::Result<Bytes>> + Unpin,
+        tx: mpsc::Sender<Result<CompletionChunk>>,
+    ) -> Option<Usage> {
+        let mut buffer = String::new();
+        let mut accumulated_content = String::new();
+        let mut accumulated_usage: Option<Usage> = None;
+        let mut current_tool_calls: Vec<OpenAIStreamingToolCall> = Vec::new();
+
+        while let Some(chunk_result) = stream.next().await {
+            match chunk_result {
+                Ok(chunk) => {
+                    let chunk_str = match std::str::from_utf8(&chunk) {
+                        Ok(s) => s,
+                        Err(e) => {
+                            error!("Failed to parse chunk as UTF-8: {}", e);
+                            continue;
+                        }
+                    };
+
+                    buffer.push_str(chunk_str);
+
+                    // Process complete lines
+                    while let Some(line_end) = buffer.find('\n') {
+                        let line = buffer[..line_end].trim().to_string();
+                        buffer.drain(..line_end + 1);
+
+                        if line.is_empty() {
+                            continue;
+                        }
+
+                        // Parse Server-Sent Events format
+                        if let Some(data) = line.strip_prefix("data: ") {
+                            if data == "[DONE]" {
+                                debug!("Received stream completion marker");
+
+                                // Send final chunk with accumulated content and tool calls
+                                if !accumulated_content.is_empty() || !current_tool_calls.is_empty() {
+                                    let tool_calls = if current_tool_calls.is_empty() {
+                                        None
+                                    } else {
+                                        Some(
+                                            current_tool_calls
+                                                .iter()
+                                                .filter_map(|tc| tc.to_tool_call())
+                                                .collect(),
+                                        )
+                                    };
+
+                                    let final_chunk = CompletionChunk {
+                                        content: accumulated_content.clone(),
+                                        finished: true,
+                                        tool_calls,
+                                        usage: accumulated_usage.clone(),
+                                    };
+                                    let _ = tx.send(Ok(final_chunk)).await;
+                                }
+
+                                return accumulated_usage;
+                            }
+
+                            // Parse the JSON data
+                            match serde_json::from_str::<OpenAIStreamChunk>(data) {
+                                Ok(chunk_data) => {
+                                    // Handle content
+                                    for choice in &chunk_data.choices {
+                                        if let Some(content) = &choice.delta.content {
+                                            accumulated_content.push_str(content);
+
+                                            let chunk = CompletionChunk {
+                                                content: content.clone(),
+                                                finished: false,
+                                                tool_calls: None,
+                                                usage: None,
+                                            };
+                                            if tx.send(Ok(chunk)).await.is_err() {
+                                                debug!("Receiver dropped, stopping stream");
+                                                return accumulated_usage;
+                                            }
+                                        }
+
+                                        // Handle tool calls
+                                        if let Some(delta_tool_calls) = &choice.delta.tool_calls {
+                                            for delta_tool_call in delta_tool_calls {
+                                                if let Some(index) = delta_tool_call.index {
+                                                    // Ensure we have enough tool calls in our vector
+                                                    while current_tool_calls.len() <= index {
+                                                        current_tool_calls
+                                                            .push(OpenAIStreamingToolCall::default());
+                                                    }
+
+                                                    let tool_call = &mut current_tool_calls[index];
+
+                                                    if let Some(id) = &delta_tool_call.id {
+                                                        tool_call.id = Some(id.clone());
+                                                    }
+
+                                                    if let Some(function) = &delta_tool_call.function {
+                                                        if let Some(name) = &function.name {
+                                                            tool_call.name = Some(name.clone());
+                                                        }
+                                                        if let Some(arguments) = &function.arguments {
+                                                            tool_call.arguments.push_str(arguments);
+                                                        }
+                                                    }
+                                                }
+                                            }
+                                        }
+                                    }
+
+                                    // Handle usage
+                                    if let Some(usage) = chunk_data.usage {
+                                        accumulated_usage = Some(Usage {
+                                            prompt_tokens: usage.prompt_tokens,
+                                            completion_tokens: usage.completion_tokens,
+                                            total_tokens: usage.total_tokens,
+                                        });
+                                    }
+                                }
+                                Err(e) => {
+                                    debug!("Failed to parse stream chunk: {} - Data: {}", e, data);
+                                }
+                            }
+                        }
+                    }
+                }
+                Err(e) => {
+                    error!("Stream error: {}", e);
+                    let _ = tx.send(Err(anyhow::anyhow!("Stream error: {}", e))).await;
+                    return accumulated_usage;
+                }
+            }
+        }
+
+        // Send final chunk if we haven't already
+        let tool_calls = if current_tool_calls.is_empty() {
+            None
+        } else {
+            Some(
+                current_tool_calls
+                    .iter()
+                    .filter_map(|tc| tc.to_tool_call())
+                    .collect(),
+            )
+        };
+        
+        let final_chunk = CompletionChunk {
+            content: String::new(),
+            finished: true,
+            tool_calls,
+            usage: accumulated_usage.clone(),
+        };
+        let _ = tx.send(Ok(final_chunk)).await;
+        
+        accumulated_usage
+    }
+}
+
+#[async_trait]
+impl LLMProvider for OpenAIProvider {
+    async fn complete(&self, request: CompletionRequest) -> Result<CompletionResponse> {
+        debug!(
+            "Processing OpenAI completion request with {} messages",
+            request.messages.len()
+        );
+
+        let body = self.create_request_body(
+            &request.messages,
+            request.tools.as_deref(),
+            false,
+            request.max_tokens,
+            request.temperature,
+        );
+
+        debug!("Sending request to OpenAI API: model={}", self.model);
+
+        let response = self
+            .client
+            .post(format!("{}/chat/completions", self.base_url))
+            .header("Authorization", format!("Bearer {}", self.api_key))
+            .json(&body)
+            .send()
+            .await?;
+
+        let status = response.status();
+        if !status.is_success() {
+            let error_text = response
+                .text()
+                .await
+                .unwrap_or_else(|_| "Unknown error".to_string());
+            return Err(anyhow::anyhow!("OpenAI API error {}: {}", status, error_text));
+        }
+
+        let openai_response: OpenAIResponse = response.json().await?;
+
+        let content = openai_response
+            .choices
+            .first()
+            .and_then(|choice| choice.message.content.clone())
+            .unwrap_or_default();
+
+        let usage = Usage {
+            prompt_tokens: openai_response.usage.prompt_tokens,
+            completion_tokens: openai_response.usage.completion_tokens,
+            total_tokens: openai_response.usage.total_tokens,
+        };
+
+        debug!(
+            "OpenAI completion successful: {} tokens generated",
+            usage.completion_tokens
+        );
+
+        Ok(CompletionResponse {
+            content,
+            usage,
+            model: self.model.clone(),
+        })
+    }
+
+    async fn stream(&self, request: CompletionRequest) -> Result<CompletionStream> {
+        debug!(
+            "Processing OpenAI streaming request with {} messages",
+            request.messages.len()
+        );
+
+        let body = self.create_request_body(
+            &request.messages,
+            request.tools.as_deref(),
+            true,
+            request.max_tokens,
+            request.temperature,
+        );
+
+        debug!("Sending streaming request to OpenAI API: model={}", self.model);
+
+        let response = self
+            .client
+            .post(format!("{}/chat/completions", self.base_url))
+            .header("Authorization", format!("Bearer {}", self.api_key))
+            .json(&body)
+            .send()
+            .await?;
+
+        let status = response.status();
+        if !status.is_success() {
+            let error_text = response
+                .text()
+                .await
+                .unwrap_or_else(|_| "Unknown error".to_string());
+            return Err(anyhow::anyhow!("OpenAI API error {}: {}", status, error_text));
+        }
+
+        let stream = response.bytes_stream();
+        let (tx, rx) = mpsc::channel(100);
+
+        // Spawn task to process the stream
+        let provider = self.clone();
+        tokio::spawn(async move {
+            let usage = provider.parse_streaming_response(stream, tx).await;
+            // Log the final usage if available
+            if let Some(usage) = usage {
+                debug!(
+                    "Stream completed with usage - prompt: {}, completion: {}, total: {}",
+                    usage.prompt_tokens, usage.completion_tokens, usage.total_tokens
+                );
+            }
+        });
+
+        Ok(ReceiverStream::new(rx))
+    }
+
+    fn name(&self) -> &str {
+        "openai"
+    }
+
+    fn model(&self) -> &str {
+        &self.model
+    }
+
+    fn has_native_tool_calling(&self) -> bool {
+        // OpenAI models support native tool calling
+        true
+    }
+}
+
+fn convert_messages(messages: &[Message]) -> Vec<serde_json::Value> {
+    messages
+        .iter()
+        .map(|msg| {
+            json!({
+                "role": match msg.role {
+                    MessageRole::System => "system",
+                    MessageRole::User => "user",
+                    MessageRole::Assistant => "assistant",
+                },
+                "content": msg.content,
+            })
+        })
+        .collect()
+}
+
+fn convert_tools(tools: &[Tool]) -> Vec<serde_json::Value> {
+    tools
+        .iter()
+        .map(|tool| {
+            json!({
+                "type": "function",
+                "function": {
+                    "name": tool.name,
+                    "description": tool.description,
+                    "parameters": tool.input_schema,
+                }
+            })
+        })
+        .collect()
+}
+
+// OpenAI API response structures
+#[derive(Debug, Deserialize)]
+struct OpenAIResponse {
+    choices: Vec<OpenAIChoice>,
+    usage: OpenAIUsage,
+}
+
+#[derive(Debug, Deserialize)]
+struct OpenAIChoice {
+    message: OpenAIMessage,
+}
+
+#[allow(dead_code)]
+#[derive(Debug, Deserialize)]
+struct OpenAIMessage {
+    content: Option<String>,
+    #[serde(default)]
+    tool_calls: Option<Vec<OpenAIToolCall>>,
+}
+
+#[allow(dead_code)]
+#[derive(Debug, Deserialize)]
+struct OpenAIToolCall {
+    id: String,
+    function: OpenAIFunction,
+}
+
+#[allow(dead_code)]
+#[derive(Debug, Deserialize)]
+struct OpenAIFunction {
+    name: String,
+    arguments: String,
+}
+
+// Streaming tool call accumulator
+#[derive(Debug, Default)]
+struct OpenAIStreamingToolCall {
+    id: Option<String>,
+    name: Option<String>,
+    arguments: String,
+}
+
+impl OpenAIStreamingToolCall {
+    fn to_tool_call(&self) -> Option<ToolCall> {
+        let id = self.id.as_ref()?;
+        let name = self.name.as_ref()?;
+        
+        let args = serde_json::from_str(&self.arguments).unwrap_or(serde_json::Value::Null);
+        
+        Some(ToolCall {
+            id: id.clone(),
+            tool: name.clone(),
+            args,
+        })
+    }
+}
+
+#[derive(Debug, Deserialize)]
+struct OpenAIUsage {
+    prompt_tokens: u32,
+    completion_tokens: u32,
+    total_tokens: u32,
+}
+
+// Streaming response structures
+#[derive(Debug, Deserialize)]
+struct OpenAIStreamChunk {
+    choices: Vec<OpenAIStreamChoice>,
+    usage: Option<OpenAIUsage>,
+}
+
+#[derive(Debug, Deserialize)]
+struct OpenAIStreamChoice {
+    delta: OpenAIDelta,
+}
+
+#[derive(Debug, Deserialize)]
+struct OpenAIDelta {
+    content: Option<String>,
+    #[serde(default)]
+    tool_calls: Option<Vec<OpenAIDeltaToolCall>>,
+}
+
+#[derive(Debug, Deserialize)]
+struct OpenAIDeltaToolCall {
+    index: Option<usize>,
+    id: Option<String>,
+    function: Option<OpenAIDeltaFunction>,
+}
+
+#[derive(Debug, Deserialize)]
+struct OpenAIDeltaFunction {
+    name: Option<String>,
+    arguments: Option<String>,
+}

docs/ACCUMULATIVE_MODE.md

@@ -0,0 +1,389 @@
+# Accumulative Autonomous Mode
+
+## Overview
+
+Accumulative Autonomous Mode is the **new default interactive mode** for G3. It combines the ease of interactive chat with the power of autonomous implementation, allowing you to build projects iteratively by describing what you want, one requirement at a time.
+
+## How It Works
+
+### The Flow
+
+1. **Start G3** in any directory (no arguments needed)
+2. **Describe** what you want to build
+3. **G3 automatically**:
+   - Adds your input to accumulated requirements
+   - Runs autonomous mode (coach-player feedback loop)
+   - Implements your requirements with quality checks
+4. **Continue** adding more requirements or refinements
+5. **Repeat** until your project is complete
+
+### Example Session
+
+```bash
+$ cd ~/projects/my-new-app
+$ g3
+
+🪿 G3 AI Coding Agent - Accumulative Mode
+      >> describe what you want, I'll build it iteratively
+
+📁 Workspace: /Users/you/projects/my-new-app
+
+💡 Each input you provide will be added to requirements
+   and I'll automatically work on implementing them.
+
+   Type 'exit' or 'quit' to stop, Ctrl+D to finish
+
+============================================================
+📝 What would you like me to build? (describe your requirements)
+============================================================
+requirement> create a simple web server in Python with Flask that serves a homepage
+
+📋 Current instructions and requirements (Turn 1):
+   create a simple web server in Python with Flask that serves a homepage
+
+🚀 Starting autonomous implementation...
+
+🤖 G3 AI Coding Agent - Autonomous Mode
+📁 Using workspace: /Users/you/projects/my-new-app
+📋 Requirements loaded from --requirements flag
+🔄 Starting coach-player feedback loop...
+📂 No existing implementation files detected
+🎯 Starting with player implementation
+
+=== TURN 1/5 - PLAYER MODE ===
+🎯 Starting player implementation...
+📋 Player starting initial implementation (no prior coach feedback)
+
+[Player creates files, writes code...]
+
+=== TURN 1/5 - COACH MODE ===
+🎓 Starting coach review...
+🎓 Coach review completed
+Coach feedback:
+The Flask server is implemented correctly with a homepage route. 
+The code follows best practices and meets the requirements.
+IMPLEMENTATION_APPROVED
+
+=== SESSION COMPLETED - IMPLEMENTATION APPROVED ===
+✅ Coach approved the implementation!
+
+============================================================
+📊 AUTONOMOUS MODE SESSION REPORT
+============================================================
+⏱️  Total Duration: 12.34s
+🔄 Turns Taken: 1/5
+📝 Final Status: ✅ APPROVED
+...
+============================================================
+
+✅ Autonomous run completed
+
+============================================================
+📝 Turn 2 - What's next? (add more requirements or refinements)
+============================================================
+requirement> add a /api/users endpoint that returns a list of users as JSON
+
+📋 Current instructions and requirements (Turn 2):
+   add a /api/users endpoint that returns a list of users as JSON
+
+🚀 Starting autonomous implementation...
+
+[Autonomous mode runs again with BOTH requirements...]
+
+============================================================
+📝 Turn 3 - What's next? (add more requirements or refinements)
+============================================================
+requirement> exit
+
+👋 Goodbye!
+```
+
+## Key Features
+
+### 1. Requirement Accumulation
+
+Each input you provide is:
+- **Numbered sequentially** (1, 2, 3, ...)
+- **Stored in memory** for the session
+- **Included in all subsequent runs**
+
+This means the agent always has the full context of what you've asked for.
+
+### 2. Automatic Requirements Document
+
+G3 automatically generates a structured requirements document:
+
+```markdown
+# Project Requirements
+
+## Current Instructions and Requirements:
+
+1. create a simple web server in Python with Flask that serves a homepage
+2. add a /api/users endpoint that returns a list of users as JSON
+3. add error handling for 404 and 500 errors
+
+## Latest Requirement (Turn 3):
+
+add error handling for 404 and 500 errors
+```
+
+This document is passed to autonomous mode, ensuring the agent:
+- Knows all previous requirements
+- Focuses on the latest addition
+- Maintains consistency across iterations
+
+### 3. Full Autonomous Quality
+
+Each requirement triggers a complete autonomous run with:
+- **Coach-Player Feedback Loop**: Quality assurance built-in
+- **Multiple Turns**: Up to 5 iterations per requirement (configurable with `--max-turns`)
+- **Compilation Checks**: Ensures code actually works
+- **Testing**: Coach can run tests to verify functionality
+
+### 4. Error Recovery
+
+If an autonomous run fails:
+- You're notified of the error
+- You can provide additional requirements to fix issues
+- The session continues (doesn't crash)
+
+### 5. Workspace Management
+
+- Uses **current directory** as workspace
+- All files created in current directory
+- No need to specify workspace path
+- Works with existing projects or empty directories
+
+## Command-Line Options
+
+### Default (Accumulative Mode)
+
+```bash
+g3
+```
+
+Starts accumulative autonomous mode in the current directory.
+
+### With Options
+
+```bash
+# Use a specific workspace
+g3 --workspace ~/projects/my-app
+
+# Limit autonomous turns per requirement
+g3 --max-turns 3
+
+# Enable macOS Accessibility tools
+g3 --macax
+
+# Enable WebDriver browser automation
+g3 --webdriver
+
+# Use a specific provider/model
+g3 --provider anthropic --model claude-3-5-sonnet-20241022
+
+# Show prompts and code during execution
+g3 --show-prompt --show-code
+
+# Disable log files
+g3 --quiet
+```
+
+### Disable Accumulative Mode
+
+To use the traditional chat mode (without automatic autonomous runs):
+
+```bash
+g3 --chat
+
+# Alternative: legacy flag also works
+g3 --accumulative
+```
+
+This gives you the old behavior where you chat with the agent without automatic autonomous runs.
+
+## Use Cases
+
+### 1. Rapid Prototyping
+
+```bash
+requirement> create a REST API for a todo app
+requirement> add SQLite database storage
+requirement> add authentication with JWT
+requirement> add rate limiting
+```
+
+### 2. Iterative Refinement
+
+```bash
+requirement> create a data visualization dashboard
+requirement> make the charts interactive
+requirement> add dark mode support
+requirement> optimize for mobile devices
+```
+
+### 3. Bug Fixing
+
+```bash
+requirement> fix the login form validation
+requirement> handle edge case when username is empty
+requirement> add better error messages
+```
+
+### 4. Feature Addition
+
+```bash
+requirement> add export to CSV functionality
+requirement> add email notifications
+requirement> add admin dashboard
+```
+
+## Tips and Best Practices
+
+### 1. Start Simple
+
+Begin with a basic requirement, let it be implemented, then add complexity:
+
+```bash
+✅ Good:
+requirement> create a basic Flask web server
+requirement> add a homepage with a form
+requirement> add form validation
+
+❌ Too Complex:
+requirement> create a full-stack web app with authentication, database, API, and frontend
+```
+
+### 2. Be Specific
+
+The more specific you are, the better the results:
+
+```bash
+✅ Good:
+requirement> add a /api/users endpoint that returns JSON with id, name, and email fields
+
+❌ Vague:
+requirement> add users
+```
+
+### 3. One Thing at a Time
+
+Focus each requirement on a single feature or fix:
+
+```bash
+✅ Good:
+requirement> add error handling for database connections
+requirement> add logging for all API requests
+
+❌ Multiple Things:
+requirement> add error handling and logging and monitoring and alerts
+```
+
+### 4. Review Between Turns
+
+After each autonomous run completes:
+- Check the generated files
+- Test the functionality
+- Decide what to add or fix next
+
+### 5. Use Exit Commands
+
+When done:
+- Type `exit` or `quit`
+- Press `Ctrl+D` (EOF)
+- Press `Ctrl+C` to cancel current input
+
+## Comparison with Other Modes
+
+| Feature | Accumulative (Default) | Traditional Interactive | Autonomous | Single-Shot |
+|---------|----------------------|------------------------|------------|-------------|
+| **Command** | `g3` | `g3 --accumulative` | `g3 --autonomous` | `g3 "task"` |
+| **Input Style** | Iterative prompts | Chat messages | requirements.md file | Command-line arg |
+| **Auto-Autonomous** | ✅ Yes | ❌ No | ✅ Yes | ❌ No |
+| **Coach-Player Loop** | ✅ Yes | ❌ No | ✅ Yes | ❌ No |
+| **Accumulates Requirements** | ✅ Yes | ❌ No | ❌ No | ❌ No |
+| **Multiple Iterations** | ✅ Yes | ✅ Yes | ✅ Yes | ❌ No |
+| **Best For** | Iterative development | Quick questions | Pre-planned projects | One-off tasks |
+
+## Technical Details
+
+### Requirements Storage
+
+- Stored in memory (not persisted to disk)
+- Numbered sequentially starting from 1
+- Formatted as markdown list
+- Passed to autonomous mode as `--requirements` override
+
+### History
+
+- Saved to `~/.g3_accumulative_history`
+- Separate from traditional interactive history
+- Persists across sessions
+- Uses rustyline for readline support
+
+### Workspace
+
+- Defaults to current directory
+- Can be overridden with `--workspace`
+- All files created in workspace
+- Logs saved to `workspace/logs/`
+
+### Autonomous Execution
+
+- Full coach-player feedback loop
+- Configurable max turns (default: 5)
+- Respects all CLI flags (--macax, --webdriver, etc.)
+- Error handling allows continuation
+
+## Troubleshooting
+
+### "No requirements provided"
+
+This shouldn't happen in accumulative mode, but if it does:
+- Check that you entered a requirement
+- Ensure the requirement isn't empty
+- Try restarting G3
+
+### "Autonomous run failed"
+
+If an autonomous run fails:
+- Read the error message
+- Provide a new requirement to fix the issue
+- Or type `exit` and investigate manually
+
+### "Context window full"
+
+If you hit token limits:
+- The agent will auto-summarize
+- Or you can start a new session
+- Consider using `--max-turns` to limit iterations
+
+### "Coach never approves"
+
+If the coach keeps rejecting:
+- Check the coach feedback for specific issues
+- Provide more specific requirements
+- Consider increasing `--max-turns`
+
+## Future Enhancements
+
+Planned improvements:
+
+1. **Persistence**: Save accumulated requirements to disk
+2. **Editing**: Edit or remove previous requirements
+3. **Branching**: Try different approaches
+4. **Templates**: Pre-defined requirement sets
+5. **Review**: Show all accumulated requirements
+6. **Export**: Save to requirements.md
+7. **Undo**: Remove last requirement
+8. **Replay**: Re-run with same requirements
+
+## Feedback
+
+This is a new feature! Please provide feedback:
+- What works well?
+- What's confusing?
+- What features would you like?
+- Any bugs or issues?
+
+Open an issue on GitHub or contribute improvements!

test-ai-requirements.sh

@@ -0,0 +1,39 @@
+#!/bin/bash
+# Test script for AI-enhanced interactive requirements mode
+
+echo "Testing AI-enhanced interactive requirements mode..."
+echo ""
+
+# Create a test workspace
+TEST_WORKSPACE="/tmp/g3-test-interactive-$(date +%s)"
+mkdir -p "$TEST_WORKSPACE"
+
+echo "Test workspace: $TEST_WORKSPACE"
+echo ""
+
+# Create sample brief input
+BRIEF_INPUT="build a calculator cli in rust with basic operations"
+
+echo "Brief input:"
+echo "---"
+echo "$BRIEF_INPUT"
+echo "---"
+echo ""
+
+echo "This will:"
+echo "1. Send brief input to AI"
+echo "2. AI generates structured requirements.md"
+echo "3. Show enhanced requirements"
+echo "4. Prompt for confirmation (y/e/n)"
+echo ""
+
+echo "To test manually, run:"
+echo "cargo run -- --autonomous --interactive-requirements --workspace $TEST_WORKSPACE"
+echo ""
+echo "Then type: $BRIEF_INPUT"
+echo "Press Ctrl+D"
+echo "Review the AI-generated requirements"
+echo "Choose 'y' to proceed, 'e' to edit, or 'n' to cancel"
+echo ""
+
+echo "Test workspace will be at: $TEST_WORKSPACE"