Use Evidra Bench for live infrastructure regression benchmarking

Hi kubectl-ai maintainers,

I wanted to point you to an open-source benchmark/regression testing project we are building for AI infrastructure agents:

https://github.com/vitas/evidra-bench

Evidra Bench runs agents against real live infrastructure scenarios, not just synthetic prompts. The open-source catalog currently includes 78 real-life scenarios, mostly Kubernetes, plus Helm, Argo CD, Terraform, and AWS/LocalStack cases.

It measures:

• final-state pass/fail from deterministic infrastructure checks;
• safe pass vs unsafe pass;
• diagnosis-before-mutation;
• forbidden or out-of-scope actions;
• retry loops and premature success claims;
• turns, duration, token usage, and cost where available;
• transcripts, tool calls, timelines, and failure autopsy artifacts.

kubectl-ai could be tested either as a CLI agent or through its MCP server mode. This would let maintainers compare releases, prompts, models, and tool changes on the same repeatable scenario set.

Example public report:

https://bench.evidra.cc/bench/reports/kubernetes-mcp-readiness-2026-05

This is not meant to replace your existing evals, just to offer an external open-source live-infra regression benchmark that may be useful for release checks.