# Project: WebDevSwarm
**Objective:** Build a Local Multi-Agent System (MAS) where a 26B Orchestrator manages specialized 7-8B "Worker" models to execute web development tasks.
## 🏗 Architecture Overview
- **Orchestrator (The Brain):** 26B Model (GPU-resident). Responsible for:
- Intent parsing.
- Task decomposition (Planning).
- Routing to specialists.
- Final synthesis and verification.
- **Specialists (The Workers):** 7-8B Models (CPU-resident). Domain-specific (Frontend, Backend, DevOps, etc.).
- **The Blackboard (Shared State):** A central JSON-based state file (`shared_state/manifest.json`) that acts as the single source of truth for the project's current architecture, file structure, and variable names.
## 📂 Directory Structure
- `config/`: System-wide settings, model paths, and API configurations.
- `orchestrator/`: Logic for the 26B model's planning and routing loops.
- `specialists/`: Individual directories for each specialist (e.g., `specialists/frontend/`).
- `shared_state/`: The `manifest.json` and other context files.
- `logs/`: Execution logs for debugging agent handoffs.
## 🚀 Implementation Roadmap
### Phase 1: Foundation (The Protocol)
- [ ] Define the **Blackboard Schema**: A strict JSON structure that all agents must read from and write to.
- [ ] Implement the **Orchestrator Loop**: A Python-based controller that manages the flow: `User -> Orchestrator -> Specialist -> Orchestrator -> User`.
### Phase 2: Specialist Alpha (Frontend/UI)
- [ ] **Target:** Build the "Frontend/UI" specialist.
- [ ] **Strategy:** Instead of full fine-tuning immediately, we will start with **Advanced System Prompting + RAG** (Retrieval-Augmented Generation) using a library of high-quality UI patterns.
- [ ] **Verification:** Use the browser/vision tools to visually confirm the specialist's output.
### Phase 3: Integration & The "Loop"
- [ ] Connect the Orchestrator to the Specialist via a local API (e.g., llama.cpp server).
- [ ] Test the "Plan $\rightarrow$ Execute $\rightarrow$ Verify" cycle.
### Phase 4: Scaling the Swarm
- [ ] Add "Backend Logic" specialist.
- [ ] Add "DevOps/Deployment" specialist.
- [ ] Implement "Testing/Quality" specialist.
## 🛠 Tech Stack
- **Inference:** `llama.cpp` (GGUF) for CPU/GPU hybrid execution.
- **Orchestration Logic:** Python.
- **Communication:** JSON-based Blackboard.
- **Verification:** Browser/Vision tools.