DifferiX is an open, model-agnostic multi-agent runtime. It doesn't just let LLMs write code; it forces them to write verified, executable code inside a deterministic Python shell.

Most "autonomous agents" fail because they hallucinate success and loop endlessly. DifferiX solves this with Zero Trust Architecture, Tiered Context Assembly, and a Local Experience Engine.

🧠 Core Architectural Patterns

1. Zero Trust to Agent: A stochastic LLM is never trusted with state mutation. Success is verified by compilers, ASTs, and fs-hashes, not by the agent's words.
2. Probabilistic in a Deterministic Shell: The LLM acts purely as a translator. Python scripts act as the supreme judge, sandbox, and state machine.
3. Scarcity by Design (TCA): Artificial token budgets force the LLM to focus on surgical context (reducing input tokens by 80%).
4. Evolutionary Memory (A-Score): Experience is stored as abstract math. Unhelpful lessons fade out; helpful ones propagate without retraining the LLM.

✨ Features

• Verification Firewall: Blocks agents from faking success. If the file on disk didn't change, the patch is rejected.
• Behavior Guard: Detects prompt-context bloat and repeated actions in real-time. Kills the loop before it burns your API budget.
• Tiered Context Assembly (TCA): Waterfall token budgeting. AST-slicing for Python, Class Inventory extraction for CSS.
• Dynamic Skill Injection: The Researcher analyzes the task and injects only the relevant skills (e.g., style_brutalist or style_3d_webgl).
• Recursive Self-Improvement (RSI): The framework can safely modify its own source code, protected by a critical-file gate.
• Model Agnostic: Works with OpenRouter, Anthropic, Google, or local Ollama models.

🚀 Quick Start

# Clone the repository
git clone https://github.com/Tdahqu8/DifferiX.git
cd DifferiX

# Setup environment
python3 -m venv venv
source venv/bin/activate
pip install -r requirements.txt

# Add your API key
cp .env.example .env
# Edit .env and add OPENROUTER_API_KEY=sk-or-v1-...

# Run DifferiX
export VEGEX_RSI_MODE=1
python main.py

🎨 Showcase (WebDev Pro Plugin)

DifferiX dynamically injects design skills to generate production-ready UIs.

"Create a portfolio landing page for a Web3 graffiti artist. Make it look raw and bold." — Generated in 4 cycles using the style_brutalist skill.

🏗️ Architecture

[ User ] <--> [ LLM Interpreter ]
                     |
      +--------------+--------------+
      |                             |
[ Symbolic Verifier ]       [ Experience Engine ]
(Python / AST / Compilers)  (A-Score / SQLite / Math)
      |                             |
      +--------------+--------------+
                     |
             [ Verified Code ]

🔌 Ecosystem & Plugins

DifferiX supports .differixplugin files. Install plugins to give agents new superpowers (WebDev, EdTech, Crypto Auditing).

Note: Executable code is strictly forbidden inside plugins to prevent RCE. Plugins are data-only (.md, .json, .css).

# Install the official WebDev Pro plugin
differix plugin install https://github.com/Tdahqu8/DifferiX/releases/download/v0.1.0-alpha/webdev_pro.differixplugin

📜 License

MIT License. Build, fork, and create.