[BUG] /insights report samples only ~0.3% of sessions, producing unrepresentative analysis

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report (please file separate reports for different bugs)
• [x] I am using the latest version of Claude Code

What's Wrong?

The /insights command generates a usage report, but the underlying analysis only samples a tiny fraction of actual sessions. In my case, only 7 out of 2,443 sessions (0.3%) were captured in the facets data (~/.claude/usage-data/facets/) that feeds the report.

This causes the report to heavily over-index on whichever project happens to appear most in the small sample. 3 of my 7 sampled sessions were from one project (a workout dashboard), so the entire report narrative — friction analysis, suggestions, interaction style — was dominated by that single project while multiple other projects were barely mentioned or completely absent.

The "At a Glance" summary, friction analysis, and suggestions all read as if I only work on one project, which is misleading.

What Should Happen?

The insights report should sample a representative cross-section of sessions so the analysis reflects the user's actual breadth of work. Suggestions:

• Increase the sample size significantly, or stratify sampling across distinct project directories so no single project dominates
• - At minimum, surface a disclaimer about sample size so users know the report may not be comprehensive
• - - Consider grouping/weighting by project so the narrative covers all areas of work proportionally

Error Messages/Logs

Steps to Reproduce

1. Use Claude Code extensively across multiple different projects (2,400+ sessions over ~25 days in my case)
2. 2. Run /insights
3. 3. Check the facets directory (~/.claude/usage-data/facets/) — only 7 JSON files were generated
4. 4. Observe that the report narrative is heavily skewed toward whichever project got the most representation in the tiny sample
5. 5. Note that the report's suggestions, friction analysis, and interaction style description don't reflect the full breadth of work

Claude Model

Opus

Is this a regression?

No, this never worked

Last Working Version

_No response_

Claude Code Version

2.1.31 (Claude Code)

Platform

Anthropic API

Operating System

macOS

Terminal/Shell

Terminal.app (macOS)

Additional Information

Facets directory (~/.claude/usage-data/facets/) contained only 7 JSON files for 2,443 total sessions. The 7 sampled sessions broke down as:

• 3 sessions: Workout dashboard / Strava+Whoop integration
• - 1 session: Google Docs template automation (Apps Script)
• - - 1 session: School quality data visualization (3K site)
• - - - 1 session: File transfer service with Slack reporting
• - - - - 1 session: Viewing the insights report itself

Projects that were completely unrepresented likely number in the dozens given the total session count. The report read as a single-project analysis rather than a comprehensive usage report.