Add GEPA optimization for MCP Tools

[dandye]

Integrate GEPA Optimization and Refactor MCP Tool Routing

References

• https://github.com/gepa-ai/gepa
• GEPA: Reflective Prompt Evolution Can Outperform Reinforcement Learning
  

This pull request integrates GEPA (Gradient-based Engine for Prompt Ad-hoc Optimization) to automate the optimization of MCP tool docstrings and descriptions, improving model-routing accuracy across Google Threat Intelligence (GTI), SecOps SIEM, and SecOps SOAR MCP servers.

Key Changes

1. GEPA Optimization & Tool Routing

• GEPA Integration: Added optimization packages and execution scripts (optimize_gti_mcp.py, optimize_soar_mcp.py, and optimize_secops_mcp.py) that evaluate routing accuracy and iterate on docstring formulations using Vertex AI models.
• Evaluation Datasets: Added curated natural language queries to mcp_dataset.json for each server package to test tool resolution, parameter extraction, and baseline routing.
• Docstring Updates: Refactored tool descriptions based on optimized outputs. Long markdown relationship tables (e.g., in netloc, files, and urls tools) were converted to concise, comma-separated lists to minimize context size and focus routing.

2. Google SecOps SOAR Improvements

• Offline Fallback: Updated connection bindings to handle failed scope retrieval gracefully, allowing the server to continue in offline mode when API credentials are not provided.
• Live Connection Test: Introduced test_live_soar.py to facilitate local verification of SOAR API bindings.
• Enhanced Testing: Configured conftest.py to support automatic .env loading, and added new dynamic tests for retrieving alert events and entities.

3. Setup & Configuration

• Environment Variables: Added a .env.example template containing required configuration variables for local server execution and GEPA optimization.
• Git Configuration: Deleted obsolete .gitmodules and ensured local .env remains properly ignored.