Mide-Lite

Mide-Lite (Gen 2) is a Markdown-native multi-agent framework designed for modern LLMs (Claude 3.5 Sonnet, GPT-4o).

It succeeds the Gen 1 implementation by shifting from ad-hoc prompt templates to structured, index-driven Markdown definitions with a reasoning-first kernel.

Project Structure

/
├── AGENTS.md              # Entry point: thin router to all framework components
├── README.md
├── src/
│   ├── system_prompt.md   # Kernel: agents, orchestration, memory protocol, reasoning
│   ├── memory/
│   │   ├── context.md     # Working memory: current phase, focus, blockers
│   │   └── decisions.md   # Append-only decision log
│   ├── agents/            # Agent personas (Markdown). Read directly for roles/capabilities.
│   ├── workflows/         # Workflow definitions. See index.md for catalog.
│   ├── rules/             # Coding standards. See base_rules.md for rules index.
│   ├── contracts/         # Artifact schema + content conventions
│   └── artifacts/         # Temporal outputs (gitignored)

Quickstart

1. Add this directory to your Cursor/LLM workspace.
2. Start a chat. The framework loads via AGENTS.md → src/system_prompt.md.
3. Engage an agent or workflow:

Direct Agent Invocation

Act as the Architect agent. Design a REST API for user authentication
with JWT tokens. Consider security boundaries and data flow.

Act as the Builder agent. Implement the login endpoint based on the
design above. Follow the rules in src/rules/typescript.md.

Act as the Critic agent with a security focus. Review the authentication
implementation for OWASP vulnerabilities.

Workflow-Based Orchestration

Act as the Supervisor agent. Run the "feature-dev" workflow to add
a password reset feature to the user service.

Supervisor: Execute the "bug-fix" workflow on src/auth/login.ts -
users are getting 401 errors on valid credentials.

Supervisor: Run "parallel-review" on the payment processing module.

Key References

• Framework entry point: AGENTS.md (routes to all components)
• Workflow catalog: src/workflows/index.md
• Agent definitions: src/agents/*.md
• Coding rules index: src/rules/base_rules.md
• Artifact contracts: src/contracts/Artifact.schema.json + src/contracts/content_conventions.md

Key Features

• Cognition-First: Agents must perform a <reasoning> block to plan strategies before generating code.
• Structured Markdown: Prompts and configuration are structured in Markdown to reduce instruction bleeding and improve adherence in modern models.
• Persistent Context: src/memory/context.md (working state) and src/memory/decisions.md (decision log) track state across sessions.