|
| 1 | +# Development Conventions |
| 2 | + |
| 3 | +## Project Overview |
| 4 | + |
| 5 | +ComputerUseAgent is an AI-powered CLI tool that enables controlled execution of |
| 6 | +system operations through LLM guidance. At its core, the project uses the Claude |
| 7 | +3 API to interpret user requests and execute them through predefined tools. |
| 8 | + |
| 9 | +## Key Components |
| 10 | + |
| 11 | +- [`HybridSession`](src/modules/hybrid/hybrid_session.ts): The primary execution |
| 12 | + engine |
| 13 | +- [`ToolHandler`](src/utils/tool_handler.ts): Manages tool registration and |
| 14 | + execution |
| 15 | +- [`PlannerModule`](src/modules/planner/planner.ts): Creates execution plans for |
| 16 | + complex operations |
| 17 | + |
| 18 | +## Core Conventions |
| 19 | + |
| 20 | +### Session Management |
| 21 | + |
| 22 | +- All operations must go through the planning phase first |
| 23 | +- Each session generates a unique ID for tracking and logging |
| 24 | +- Sessions maintain their own message history and token usage |
| 25 | + |
| 26 | +### Tool Implementation |
| 27 | + |
| 28 | +```typescript |
| 29 | +interface ToolConfig { |
| 30 | + toolName: string; |
| 31 | + command: string; |
| 32 | + output: string; |
| 33 | + inputs: ToolInputConfig[]; |
| 34 | + description: string; |
| 35 | + enabled: boolean; |
| 36 | +} |
| 37 | +``` |
| 38 | + |
| 39 | +### Command Implementation |
| 40 | + |
| 41 | +When adding new commands: |
| 42 | + |
| 43 | +1. Create a new file in `src/commands/` following existing patterns (see |
| 44 | + `history.ts`, `export.ts`) |
| 45 | +2. Define command flags using `ParseOptions` interface |
| 46 | +3. Update help text in `parseFlagForHelp` function |
| 47 | +4. Register command in `main.ts` command handler section |
| 48 | + |
| 49 | +Example command structure: |
| 50 | + |
| 51 | +```typescript |
| 52 | +export async function handleNewCommand(args: string[]) { |
| 53 | + const commandFlags = { |
| 54 | + string: ["option1", "option2"], |
| 55 | + boolean: ["flag1"], |
| 56 | + default: { |
| 57 | + option1: "default", |
| 58 | + flag1: false, |
| 59 | + }, |
| 60 | + }; |
| 61 | + const flags = parseArgs(args, commandFlags); |
| 62 | + |
| 63 | + if (flags.help || flags._[1] === "help") { |
| 64 | + console.log(parseFlagForHelp(commandFlags)); |
| 65 | + return; |
| 66 | + } |
| 67 | + // Command implementation |
| 68 | +} |
| 69 | +``` |
| 70 | + |
| 71 | +### Error Handling |
| 72 | + |
| 73 | +- Tool errors should be captured and logged, not thrown |
| 74 | +- All errors must be logged with proper context |
| 75 | +- Sessions should gracefully handle tool failures |
| 76 | + |
| 77 | +### Logging |
| 78 | + |
| 79 | +- Use structured logging via `log` |
| 80 | +- Token usage must be tracked for all API calls |
| 81 | +- Session steps must be logged for audit trails |
| 82 | + |
| 83 | + |
| 84 | +### LLM Interactions |
| 85 | + |
| 86 | +- Keep system prompts in `constants.ts` |
| 87 | +- Messages must maintain conversation context |
| 88 | +- Tool calls should be processed sequentially |
| 89 | + |
| 90 | +## Best Practices |
| 91 | + |
| 92 | +1. Always use typed interfaces for tool inputs/outputs |
| 93 | +2. Log execution steps and tool usage |
| 94 | +3. Validate tool configurations before use |
| 95 | +4. Handle API rate limits gracefully |
| 96 | +5. Maintain clear session boundaries |
| 97 | + |
| 98 | +## Anti-Patterns |
| 99 | + |
| 100 | +- Directly executing shell commands without planning |
| 101 | +- Mixing session contexts |
| 102 | +- Bypassing the tool handler interface |
| 103 | +- Hardcoding system prompts in code |
| 104 | +- Ignoring token usage tracking |
| 105 | + |
| 106 | +Note: Editor and Bash sessions are legacy components scheduled for deprecation. |
| 107 | +New features should focus on extending the HybridSession functionality. |
0 commit comments