Weekly AI Coding Reports Track Activity, Not Productivity, Researchers Warn

A weekly AI coding report generated from tools like Claude Code and Codex captures session activity, token usage, provider splits, and model appearances — but cannot measure code quality or actual productivity. Experts caution that token counts, including input, output, cache creation, and cache reads, reflect recorded activity rather than work value, since a failed run can consume more tokens than a high-impact fix. Consistent time boundaries across all report sections are essential, as mixing calendar-day and rolling-window calculations can cause figures to contradict each other. Unknown models should remain labeled as such rather than being approximated, to preserve the auditability of historical records. The report's primary value lies in prompting one concrete workflow decision each week, such as reducing unattended runs or fixing authentication failures, rather than serving as a performance score.