Диагностика ошибок в исследовательском коде

testing

diagnose is a Claude Code skill that root-causes a failing or wrong empirical result through a structured reproduce → minimise → hypothesise → instrument → fix loop, replacing guesswork with disciplined debugging. It targets research code in R, Stata, and Python, where bugs are usually silent wrong numbers rather than crashes — covering type coercion, NA and merge blow-ups, factor levels, clustering and standard-error choices, weighting, collinearity, convergence failures, seed issues, and package-version drift. The skill is symptom-driven and single-target: one wrong number or one failing run at a time; for auditing every numeric claim in a manuscript, the companion skill `/audit-reproducibility` is the right starting point. The `--no-fix` flag localizes the root cause without touching shared or load-bearing files. Ideal for researchers and analysts asking "why did my regression change", "this result won't reproduce", or "my script errors out with no clear cause".