mcp-server-langgraph copilot-instructions.md

GitHub Copilot Instructions for MCP Server with LangGraph

Project Context

You are working on a production-ready LangGraph agent with Model Context Protocol (MCP) implementation. This is enterprise-grade software with:

• Multi-LLM Support: 100+ providers via LiteLLM
• Security: JWT auth + OpenFGA fine-grained authorization
• Secrets: Infisical integration
• Observability: OpenTelemetry (traces, metrics, logs)
• Deployment: Kubernetes with Helm/Kustomize

Code Style

Python Standards

• Black formatter: 127 character line length
• Type hints: Always required for public APIs
• Docstrings: Google-style for all public functions
• Import order: Standard → Third-party → Local (via isort)

# Example
def process_message(message: str, user_id: str) -> AgentResponse:
    """
    Process a user message through the agent.

    Args:
        message: The user's input message
        user_id: Unique user identifier

    Returns:
        AgentResponse with the agent's reply

    Raises:
        AuthorizationError: If user lacks permissions
    """
    pass

Testing Standards

• Mark all tests: @pytest.mark.unit, @pytest.mark.integration, etc.
• Mock external services (LLMs, OpenFGA, Infisical)
• Aim for >80% coverage on critical paths

@pytest.mark.unit
def test_process_message_success():
    """Test message processing with valid input."""
    # Arrange
    message = "Hello, agent"
    user_id = "user-123"

    # Act
    response = process_message(message, user_id)

    # Assert
    assert response.status == "success"
    assert len(response.content) > 0

Security Requirements

Authentication

• Never hardcode secrets
• Use settings.jwt_secret_key from config
• Validate tokens on every request
• Log all auth failures

# ✅ Good
token = settings.jwt_secret_key

# ❌ Bad
token = "hardcoded-secret"

Authorization

• Check OpenFGA before protected operations
• Use principle of least privilege
• Log authorization failures with context