Add CrewAI + EvalView integration — regression testing for crews

[hidai25]

Summary

Adds a self-contained integration example showing how to use EvalView for regression testing CrewAI crews.

EvalView complements crewai test — while crewai test runs crews N times and shows scores, EvalView snapshots the full execution trace (which agent called which tool, with what parameters, in what order) and diffs it against a golden baseline on every change.

This addresses the use case described in issue #4174 — deterministic CI regression checks for tool-using agents.

What's included

• integrations/CrewAI-EvalView/
  • README.md — setup guide, test examples, CI config, watch mode
  • crew.py — example research + writing crew with tools
  • main.py — runnable entry point using EvalView's Python API
  • tests/research-report.yaml — tool-calling regression test
  • tests/safety-check.yaml — forbidden tool safety test
  • requirements.txt — crewai + evalview dependencies
  • .env.example

How it works

The native adapter calls crew.kickoff() directly (no HTTP server), captures tool calls via CrewAI's event bus (ToolUsageFinishedEvent), and returns structured traces for diffing.

from evalview.adapters.crewai_native_adapter import CrewAINativeAdapter

adapter = CrewAINativeAdapter(crew=crew)
# Then: evalview snapshot → evalview check → CI gate

Test plan

• Example runs with python main.py
• evalview snapshot captures baseline
• evalview check detects regressions after changes
• CI workflow blocks PRs on regression