Feature request: per-model token usage and cost statistics

[ffftttft]

Problem

pi-web already has the two raw ingredients for cost tracking, but nothing that aggregates them:

• Usage data: per-message token counts (input/output, cache read/write after 每条信息显示 Cache Write token 数量或者累加到 Input token 数量 #194) are shown in the chat and stored in session .jsonl files.
• Pricing data: models.dev pricing presets were integrated into the models config in feat(models): add pricing presets and upstream model discovery #256.

What's missing is the aggregation layer: a view that answers "how much have I spent per model this week/month, and where?"

Proposal

A usage statistics page/panel that aggregates across sessions:

• Per model (provider/model id): total input / output / cache-read / cache-write tokens, message count, session count.
• Estimated cost: tokens × the configured pricing preset (from the models.dev catalog or user-overridden prices in models.json); clearly marked as an estimate, hidden or marked N/A for models without pricing info.
• Time range filtering: today / 7d / 30d / all.
• Data source: scan session .jsonl files (same read path as the session list) and aggregate assistant message usage blocks — no changes to the agent runtime needed. Results can be cached server-side and invalidated on session file mtime changes.

Nice-to-haves (could be follow-ups): per-project breakdown, CSV export, daily trend chart.

Notes

• Happy to contribute a PR for this. I'd start with a read-only /api/usage endpoint + a stats panel, no schema changes to session files.
• If maintainers would rather see this as an extension/plugin instead of core UI, let me know and I'll scope it that way.