Gate merges on a golden-dataset eval run — 200–500 cases, PR-triggered, regression below baseline blocks the merge

Logiciel·medium signal

Treat agent quality like tests: curate a golden dataset of 200–500 examples spanning happy, recoverable, unrecoverable, and adversarial cases, then have every pull request trigger an automated judge run, with any score below the established baseline blocking the merge. Tools like Promptfoo and DeepEval run these unit-style during local iteration. This converts 'it feels better' prompt changes into a CI quality gate that catches silent regressions.