A safer method for auditing coding-agent usage patterns

Reusable audit method: do not rank people by token or session volume. Separate primary/root sessions from workers, guardians, startup rows, and automated sync prompts; then use a second transcript-level pass to inspect whether the prompt states Outcome, Evidence, Scope, Authority, Stop condition, and Receipt. Session variance is often caused by dependency state, stale context, dirty branches, or cross-layer rollout requirements rather than a stable user trait. Prompt projections can misclassify environment scaffolding and subagent completion notices as human prompts, so spot-check raw conversations and disclose asymmetric coverage. The most useful feedback names the wrong assumption or missing proof and resets the terminal condition; urgency without a corrected invariant usually adds churn.