RACSP v2.0.0 — Ryou AI Cross SDD Protocol

Version: 2.0.0
Status: Production-Grade Architecture Specification
Evolution: RASS v3.0.0 → RACSP v1.0.0 → RACSP v2.0.0

---

What is RACSP?

RACSP is a project-native protocol, not a tool you use. It lives inside your repository in a .racsp/ folder and becomes invisible infrastructure that AI agents detect and use automatically.

You do not run RACSP commands during daily work. You simply open your AI agents normally in the project folder, and each agent detects .racsp/, reads its own SDD profile, and self-configures on the spot. No wrappers. No special commands. Just open agents and talk to them.

Practical Workflow

1. Run ryou setup once in your project — creates .racsp/ with protocol files, agent profiles, and shared brain
2. Open Claude Code normally → detects .racsp/ → reads its profile → self-configures
3. Open OpenCode normally → detects .racsp/ → reads its profile → self-configures
4. Open Antigravity normally → detects .racsp/ → reads its profile → self-configures
5. All agents work in parallel, sharing memory through .racsp/brain/
6. No special commands needed — agents update the Brain automatically as they work

---

Key Capabilities

• Transparent Multi-Agent Collaboration — Claude, OpenCode, Antigravity, Gemini, DeepSeek, and others work together without manual coordination
• Project-Native Protocol — Everything lives in .racsp/ inside the repository; no global state, no external services
• SDD Adaptive Profiles per Agent — Each agent has its own per-phase model mapping; self-configures from .racsp/agents/<agent>/. Default profile is RyouSet (Kimi K2.6 for implementation, GLM-5.1 for reasoning, DeepSeek V4 Pro for verification)
• Shared Project Brain — All agents read and write .racsp/brain/ automatically for persistent, synchronized project state
• Automatic Conflict Prevention — LOCK.md tracks active file locks; agents check before editing to avoid collisions
• Precision Subtask Delegation — Flagship models decompose work and delegate atomic subtasks to flash models with validation gates
• Anti-Hallucination — Agents read the Brain before acting; no phantom files, no broken contracts, no dependency hallucinations
• Token Efficiency — 80% token reduction through semantic compression and adaptive model selection per phase
• Session Survival — Project state survives model switching, provider changes, conversation resets, and session loss

---

Philosophy

THE PROJECT HOLDS THE MEMORY. NOT THE AI MODEL.

OPEN YOUR AGENTS NORMALLY. RACSP HANDLES THE REST.

---

Quick Start

One-Time Setup

# Clone RACSP
git clone https://github.com/kevinkeor/RACSP.git
cd RACSP

# Build and link CLI
cd ryou-cli
npm install
npm run build
npm link

Install in Your Project

cd your-project
ryou setup

This creates .racsp/ with:

• RACSP.md — protocol rules
• agents/ — per-agent SDD profiles
• brain/ — shared project memory (BRAIN.md, AGENTS.md, LOCK.md, CONTEXT.md)

Use Agents Normally

# Open any agent in the project folder — no special commands needed
cd your-project && claude
cd your-project && opencode
cd your-project && antigravity

Each agent detects .racsp/, reads its profile, and works with the right models per phase. The Brain updates automatically.

Optional Diagnostics

# Check full system status when needed
ryou diagnostics --full

---

How Agents Self-Configure

When you open an AI agent in a project with .racsp/, the following happens automatically:

1. Agent detects .racsp/ folder — presence signals that RACSP protocol is active
2. Reads RACSP.md — loads protocol rules, phase definitions, and coordination conventions
3. Reads agents/<agent>/ — loads its own SDD profile with per-phase model mappings
4. Reads BRAIN.md — loads current project state, recent decisions, and open tasks
5. Reads AGENTS.md — sees what other agents are working on right now
6. Reads LOCK.md — checks which files are locked to avoid conflicts
7. Works normally — uses the right models per phase, updates the Brain automatically

No manual configuration. No global profiles. The project tells the agent exactly how to behave.

---

Architecture

┌─────────────────────────────────────────────────────────────────────────────┐
│                         AI AGENTS                                            │
│  ┌─────────┐ ┌─────────┐ ┌─────────┐ ┌─────────┐ ┌─────────┐             │
│  │  Claude │ │OpenCode │ │Antigravity│ │ Gemini  │ │ DeepSeek│             │
│  │  Code   │ │   Go    │ │         │ │  CLI    │ │   V4    │             │
│  └─────────┘ └─────────┘ └─────────┘ └─────────┘ └─────────┘             │
│  ┌─────────┐ ┌─────────┐ ┌─────────┐                                     │
│  │  Kimi   │ │   GLM   │ │  Local  │                                     │
│  │         │ │         │ │ Models  │                                     │
│  └─────────┘ └─────────┘ └─────────┘                                     │
├─────────────────────────────────────────────────────────────────────────────┤
│                    PROJECT DETECTION LAYER                                   │
│  ┌─────────────────────────────────────────────────────────────────────┐   │
│  │  Detect .racsp/ folder → Read RACSP.md → Load protocol rules       │   │
│  └─────────────────────────────────────────────────────────────────────┘   │
├─────────────────────────────────────────────────────────────────────────────┤
│              AGENT-SPECIFIC SDD PROFILES LAYER                               │
│  ┌─────────┐ ┌─────────┐ ┌─────────┐ ┌─────────┐ ┌─────────┐           │   │
│  │ Claude  │ │OpenCode │ │Antigravity│ │ Gemini  │ │ DeepSeek│           │   │
│  │ Profile │ │ Profile │ │ Profile │ │ Profile │ │ Profile │           │   │
│  │(8 phases│ │(8 phases│ │(8 phases│ │(8 phases│ │(8 phases│           │   │
│  │ models) │ │ models) │ │ models) │ │ models) │ │ models) │           │   │
│  └─────────┘ └─────────┘ └─────────┘ └─────────┘ └─────────┘           │   │
├─────────────────────────────────────────────────────────────────────────────┤
│                   SHARED PROJECT BRAIN LAYER                                 │
│  ┌─────────┐ ┌─────────┐ ┌─────────┐ ┌─────────┐                         │
│  │ BRAIN.md│ │AGENTS.md│ │ LOCK.md │ │CONTEXT.md│                         │
│  │ Project │ │ Active  │ │ File    │ │ Shared  │                         │
│  │ State   │ │ Agents  │ │ Locks   │ │ Context │                         │
│  └─────────┘ └─────────┘ └─────────┘ └─────────┘                         │
├─────────────────────────────────────────────────────────────────────────────┤
│                         PROJECT CODE                                         │
│  ┌─────────────────────────────────────────────────────────────────────┐   │
│  │  Your source files, tests, documentation, configuration               │   │
│  └─────────────────────────────────────────────────────────────────────┘   │
└─────────────────────────────────────────────────────────────────────────────┘

---

RACSP vs RASS

Dimension	RASS v3.0	RACSP v1.0	RACSP v2.0
Usage	Single-agent	Multi-agent with commands	Open agents normally, protocol in project
Configuration	Global	Global	Project-native .racsp/ folder
Memory	Runtime	Persistent brain	Shared Brain v2 with auto-sync
Agent Setup	Manual	Global config injection	Self-configuring from .racsp/
User Workflow	Chat with one agent	Use ryou commands	Just open agents and talk
Parallel Work	No	Yes with orchestration	Yes, transparently
Conflict Prevention	None	Manual locks	Automatic via LOCK.md

---

Documentation

• ARCHITECTURE.md — Complete technical specification
• Guia-de-Uso.html — Interactive usage guide
• docs/SETUP.md — Detailed setup instructions
• docs/specs/ — Component specifications
• docs/examples/ — Usage examples and templates

---

License

KeorSoft Open Development License (KODL) v1.0

Free and open source for use in developing any commercial or non-commercial tool or software. RACSP itself may not be sold, resold, or commercialized as a standalone product. See LICENSE.md for full terms.

Built with precision by Kevin Keor.

---

RACSP v2.0.0 — The transparent, project-native evolution of distributed AI collaboration.