hidai25/eval-view

Captures tool-call sequences and parameter changes across multi-turn agent interactions, then compares against golden baselines to surface silent regressions that pass health checks. Detects coordinated drift across test runs to distinguish provider rollouts from system failures, using a four-layer scoring system from free offline tool analysis to optional LLM-as-judge evaluation. Integrates deeply with CI/CD via GitHub Actions with automatic PR comments and cost/latency tracking, while supporting agent auto-detection across LangGraph, CrewAI, and other frameworks without requiring API keys for baseline operations.