| name | mofa-eval |
| description | An LLM-as-a-Judge agent evaluation skill built in Rust for MoFA IDE Testing. |
mofa-eval (Agent Testing Skill)
This skill brings a robust agentic testing platform to MoFA. Utilizing OpenAI (via async-openai) as an LLM Judge, this skill can autonomously grade the outputs of other MoFA agents against standard rubrics and track regressions in a SQLite database.
Capabilities
- LLM-as-a-Judge: Uses OpenAI's GPT models to determine if an actual output logically entails the expected output without strict string matching.
- Evaluation Persistence: Stores every evaluated interaction categorized by
run_id into a SQLite database.
- Regression Testing: Provides an automated
compare_runs tool to explicitly notify if Agent Prompts have regressed performance between run iterations.
Usage Examples
Trigger a single evaluation:
echo '{"run_id":"test-run-01", "expected":"The capital is Paris.", "actual":"Paris is the capital city of France."}' | mofa-eval evaluate_response
Get a summary of a test run:
echo '{"run_id":"test-run-01"}' | mofa-eval score_summary