/insights: two reports generated 74s apart from identical data differ ~30%; totals don't reconcile with session-meta
Environment: Claude Code 2.1.216 · macOS 26.5.2
What happened: Running /insights twice in quick succession produced two report snapshots 74 seconds apart from the same on-disk ~/.claude/usage-data/:
report-2026-07-20-182844.html— 149 active hours, 279 commits, 1,775 header messagesreport-2026-07-20-183002.html(=report.html) — 210 active hours, 331 commits, 2,406 header messages
Ground truth from disk: summing duration_minutes and git_commits across all session-meta/*.json gives 210.9h / 338 commits — the later snapshot is disk-consistent; the earlier one undercounts ~30%.
Also: the rendered "773 total sessions" matches neither the 599 session-meta files on disk nor the report's own "analyzed" counts (386/286). I could find no documentation of what these denominators represent.
Expected behaviour: deterministic totals from identical input, or a visible completeness/partial indicator on the rendered report; a documented definition of the session-count denominators.
Impact: the report can silently render a materially incomplete aggregation with no way to tell — anyone quoting the first snapshot inherits ~30% error in headline figures.
Related: #23209 (impossible hours count) — same symptom family: numeric unreliability in /insights aggregation.