/insights: Inflated stats and recency bias in analysis

Summary

The /insights command generates reports with inflated statistics and over-represents recent sessions, making the analysis less useful for understanding actual usage patterns.

Issues

1. Impossibly inflated hours

The report shows 9,303 hours for a one-month period (2026-01-04 to 2026-02-04).

• Maximum possible: 744 hours (24h × 31 days)
• Reported: 9,303 hours (12.5× the maximum possible)

This suggests a unit conversion bug, double-counting of overlapping sessions, or incorrect aggregation.

2. Recency bias within the analysis window

Sessions from the last few days appear to be over-weighted compared to earlier sessions in the same month. The narrative and "project areas" descriptions seem heavily influenced by very recent work rather than representing the full month evenly.

3. One-month window is too short

A one-month snapshot captures users at a single point in their development cycle. For example, if a user is at the end of a project cycle (documentation, bug fixes, polish), the report concludes that's their primary work pattern—missing earlier phases like architecture, heavy feature development, and refactoring.

Suggestions

1. Fix the hours calculation
2. Weight sessions more evenly across the analysis period
3. Extend the default analysis window to 6 months (or make it configurable) to capture full development cycles

Environment

• Claude Code CLI
• macOS (Darwin 25.2.0)
• Report generated: 2026-02-05