agent-regression-evals

This plugin provides three deterministic scripts to detect regressions in agent behavior: a trajectory checker grades tool-call paths with multiple match modes and forbidden-tool enforcement; a flake detector classifies cases as stable or flaky and reports pass@k and pass^k with statistical confidence intervals; and a regression gate performs paired statistical testing (t-based CI, McNemar test) to block merges on significant drops, critical-case regressions, dropped cases, or invalid quarantine entries.