Cline rules

1# Project-Specific Rules (Computer Use Agent)
2 
3## Project Structure
4 
5This is a production-ready Python package using best practices architecture:
6 
7```
8computer-use/
9├── src/computer_use_agent/     # Main package
10│   ├── actions/                # Action execution layer
11│   ├── config/                 # Configuration & prompts
12│   ├── utils/                  # Utilities (retry, response handling)
13│   ├── agent.py                # Core orchestrator
14│   └── cli.py                  # CLI interface
15├── autonomous_agent.py         # Legacy PoC (maintained for compatibility)
16├── pyproject.toml              # Project metadata & dependencies
17└── README.md                   # Documentation
18```
19 
20## File Organization Rules
21 
22### Config Layer (`src/computer_use_agent/config/`)
23- `settings.py`: Configuration dataclasses
24- `prompts.py`: System prompts and instructions
25- No business logic, only configuration
26 
27### Actions Layer (`src/computer_use_agent/actions/`)
28- `executor.py`: Action execution logic
29- `screen.py`: Screen/coordinate management
30- Each action handler is a private method
31- No direct API calls, only desktop automation
32 
33### Utils Layer (`src/computer_use_agent/utils/`)
34- `response_handler.py`: Response processing utilities
35- `retry.py`: Retry logic with exponential backoff
36- `goal_rewriter.py`: Goal rewriting with inline context injection
37- Pure functions where possible
38- No application state
39 
40### Core Layer
41- `agent.py`: Main orchestrator, coordinates all components
42- `cli.py`: Command-line interface only
43- Minimal business logic in CLI
44 
45## Architecture Decisions
46 
47### Goal Rewriting with Inline Context Injection
48 
49**Design Pattern**: Context is injected inline during goal rewriting, not in the main system prompt.
50 
51**Rationale**:
52- **Lean system prompt**: Keep main prompt generic and universally useful
53- **Task-specific context**: Each goal gets exactly the contextual help it needs
54- **Scalability**: Add new apps by updating goal rewriter only, not system prompt
55- **Token efficiency**: Avoid bloating every API call with app-specific details
56 
57**Implementation**:
58- `goal_rewriter.py` uses Gemini 2.5 Pro to rewrite goals
59- Detects apps involved (Slack, Linear, VSCode, Cline, etc.)
60- Injects relevant shortcuts, tips, and workflow patterns INLINE at each step
61- Example: "press command+k to open command palette, type 'ENG-123' to search for the issue (Linear issues use format like ENG-123), then press Return to open it"
62 
63**What Goes in System Prompt**:
64- Generic macOS keyboard shortcuts (command+c/v, command+tab)
65- Keyboard-first interaction principles
66- Input field submission patterns
67- Cookie/privacy dialog handling
68- Safety decision guidelines
69 
70**What Goes in Goal Rewriter**:
71- App-specific keyboard shortcuts (Linear's command+shift+., Slack's command+k)
72- App-specific tips (Cline wait times, Slack search syntax)
73- Workflow patterns (Linear → VSCode git branch creation)
74- Context for specific actions being performed
75 
76**Benefits**:
77- Faster, cheaper API calls (less tokens in system prompt)
78- Goals are enriched with exactly what's needed for that task
79- Easy to add new apps without modifying core system prompt
80- Context adapts to actual task, not generic coverage
81 
82## Version Control
83 
84- Keep `autonomous_agent.py` for backward compatibility
85- Mark deprecated code with warnings
86- Use semantic versioning (MAJOR.MINOR.PATCH)
87- Update CHANGELOG for significant changes
88