dsh-batch-regression

You run the same command a few times, the numbers jump around, and you wonder which one to trust. A single result is a fragile basis for judgment, especially on Apple Silicon where dynamic P/E core scheduling and sustained thermal throttling can quietly double a single-sample duration, while cpulimit, nice, and taskpolicy all fail in practice on that hardware. dsh-batch-regression hands the job to statistics: it runs the same command N rounds (at least 5 recommended), reports the median instead of the mean to neutralize outliers, and prints the full distribution so your conclusion has teeth.

It also bakes in execution discipline. For before-and-after comparisons, a delta must exceed 20 percent to be flagged as a reliable conclusion, keeping normal jitter from being mistaken for regression. For intermittent bugs, switch to a reproduction-rate mode that quantifies how many of N runs actually hit the issue. A failed round does not abort the batch; it is logged as FAIL and the run continues, so the final summary stays complete.

Built for engineers doing performance work, writing benchmarks, or chasing flaky bugs: describe the command and the metric, and the plugin handles repetition, median extraction, and confidence labeling out of the box. It is deliberately out of scope for one-shot decisions, commit-level regression bisection (use git bisect), or UI visual regression (use a dedicated screenshot-diff tool). Stay in your lane, and the tool stays in its lane.