LAO

# LAO (Leeway AI Office) LAO turns rough product ideas into locked, AI-ready specifications before coding agents start implementation. --- ## Why LAO? Coding agents are powerful, but vague chats often become drifting implementations. LAO adds a planning layer between idea and code. ``` Idea ➔ Spec Sprout ➔ Decision Cards ➔ Spec Lock ➔ DevLoop ``` --- ## Who is it for? - **Developers** using Claude Code, Codex, Gemini CLI, Cursor, or other local AI agents (fully configurable). - **Solo builders** who want structured planning before implementation. - **Teams** that want reusable project specs instead of scattered AI chats. --- ## Core Concepts - **Local CLI AI Engine**: Harnesses local CLIs for ultra-fast runs with minimal orchestration overhead. - **Director + Specialized Agents**: Specialized step agents led by a routing Director. - **Golden Rules**: Programmatic constraints enforced on specifications. - **Decision Cards**: Micro-decisions agreed on before locking requirements. - **Compiled Spec**: The single source of truth for the project. - **Handover Gate**: The read-only lock reducing spec-code drift. - **DevLoop Console**: Local execution, building, and verification loop. --- ## Quick Start for Existing CLI Users ### Prerequisites - Node.js v18.0.0 or higher - Configurable local AI CLI clients (`gemini`, `claude`, `codex`, `cursor` etc.) authenticated ### 1. Install globally You can install LAO directly from GitHub: ```bash # Using yarn yarn global add git+https://github.com/naka98/LAO.git # OR Using npm npm install -g git+https://github.com/naka98/LAO.git ``` ### 2. Run the office Navigate to your project root folder and launch LAO: ```bash lao ``` LAO will start on port `4000` and automatically open the guided planning workspace in your browser (`http://localhost:4000`). --- ## Learn More - 📄 [Why LAO?](./docs/why-lao.md) - 📐 [Architecture Rationale](./docs/v0.9_architecture.md) - ⚙️ [Operating Principles](./docs/operating-principles.md) <details> <summary>🛠️ Local Development (Contributing)</summary> If you want to clone the repository and run it locally for development: ```bash # Clone the repository git clone https://github.com/naka98/LAO.git cd LAO # Install dependencies & start npm install # or yarn install npm start # or yarn start ``` </details> <details> <summary>⚙️ Deep Dive & Technical Highlights (v0.9.3)</summary> ### 1. Local CLI AI Engine & E2BIG Bypass - Executes local CLI tools (`gemini`, `claude`, etc.) via non-login shells (`-c`) to minimize startup latency down to 10ms. - Uses file redirection (`<`) instead of command substitution to bypass UNIX `ARG_MAX` limits (`E2BIG`), allowing massive specification prompts. ### 2. Concurrency Control (Spawn Queue) - Limits active AI tasks to `maxConcurrency = 2` to prevent CPU freeze and SQLite database locks (`SQLITE_BUSY`). - Evicts redundant requests and kills orphans automatically after 90 seconds. ### 3. Planning Harness & Self-Correction - Automated Linter checks specifications for formatting constraints (e.g., Given-When-Then criteria). - Performs up to 3 self-correction loops using AI feedback injection. - Integrates with a local `RULES.md` in the project root to enforce tech-stack boundaries. ### 4. Human-In-The-Loop (HITL) Interventions - Renders detailed warnings on the UI if self-correction fails 3 times. - Features a **[Force Commit]** bypass option to let developers manually override errors. ### 5. Multi-Agent & Context Budgeting - Centralized Director routes work to step agents. - Slices context to send only relevant parts of specifications to keep prompt tokens within the budget. ### 6. DevLoop Console & SSE Streaming - Runs `build`, `launch`, and `verify` commands in the console. - Streams live progress diagnostics directly to the browser UI via Server-Sent Events (SSE) to prevent user wait fatigue. </details>