Usage dashboard missing ~3 months of Code session history; input token counts implausibly low

Describe the bug

The claude.ai usage dashboard is missing the large majority of my Code session history, and the token counts for sessions it does show are internally inconsistent. I am not estimating — I have a complete local database built from my JSONL session files and can count records precisely.

Session count discrepancy

• Dashboard (Overview tab, All filter): 183 sessions
• Local JSONL database (all Code sessions): 612 sessions, starting December 13, 2025
• The Models tab chart shows near-zero activity before approximately March 11, 2026. My records show active Code sessions from December 13 onward with no gap.

Token count discrepancy (Opus 4.7 as example)

• Dashboard: 28.7k input / 14.0M output tokens across 43 sessions
• Local JSONL records: 3.591M total tokens across 43 sessions
• The dashboard's 28.7k total input tokens implies 667 input tokens per session average. A typical session opening prompt exceeds this. This is not a rounding error.
• The output-to-input ratio of ~487:1 is not plausible for interactive coding sessions.

External corroboration

My GitHub contribution graph shows consistent daily commits from November 2025 through April 2026. I almost never use Code without making at least one commit, and almost never make a commit without using Code. The two signals should be tightly correlated; they are not — which independently confirms that the dashboard is missing sessions, not that the sessions didn't happen.

On the meaning of "All"

The "All" time filter should mean all sessions for this account. My account's first activity was November 20, 2025. "All" does not mean "since January 1" or "since March 11" — those are not defensible interpretations of the label. The chart should begin at account creation; it does not. This same misrepresentation of "All" also affects the Favorite model stat, reported separately.

What this is not

This is not a display or calculation error on correct underlying data. The sessions are absent. Three months of JSONL files exist locally; they are not reflected in the dashboard.

Additional context

Related to but distinct from #53469 (favorite model miscalculation). Both symptoms are downstream of the same ingestion failure — if sessions are missing, all derived stats are wrong.

The usage dashboard was released approximately April 13, 2026.