[caigee-cmd]

Backend

• DB engine is SQLite (modernc). SQLite has no percentile aggregate, so p50/p95 use the existing approach: read latencies sorted by key, then pick the nearest-rank value in Go.
• requestSeries: each bucket adds prompt_tokens, completion_tokens, cache_read_tokens, latency_p50_ms and latency_p95_ms.
• Grouped stats (models / accounts / providers): each row adds latency_p95_ms, prompt_tokens, completion_tokens, cache_read_tokens, cache_write_tokens and credits.
• Totals: credits is added to totals, and cache_write to tokens.
• Credits come from credits, falling back to request_usage_details.credit, the same source the request list uses.
• /api/logs/stats?dim=model,account,provider&limit=N:
  • dim limits which grouped lists are computed. Unknown names are ignored, and an empty dim means all three.
  • limit defaults to 6 and is capped at 500.
  • The recorder's stats cache key now includes dim and limit.
  • Calls without these parameters work exactly as before. SummarizeRequestLogs is kept as a wrapper around the new SummarizeRequestLogsWith.
• /api/logs/requests?provider=: matches the same effective provider the ranking groups by (the logged provider, else the account's provider).
• SPA fallback: /usage is added to the route list in internal/server/router.go.

Frontend

• TrafficChart has a HeroUI segmented switch (the existing FilterToggle) for:
  • requests: ok / failed
  • tokens: input / output / cache read
  • latency: p95 / p50
• The tooltip, axis formatting and legend follow the selected metric. Drag-to-logs still works.
• Usage page (/usage, sidebar → Workspace → Usage, Phosphor ChartBar icon). It is a separate page rather than an Overview tab because:
  • the sidebar already has one page per task (Overview, Accounts, Logs…);
  • the tables are full-width and need their own sort state, row-count and dimension tabs;
  • the Overview is already dense.
• On the Usage page:
  • Tabs switch between Models, Accounts and Providers. The tables use HeroUI Table with allowsSorting and SortableColumnHeader.
  • Columns: requests, success rate, p95, tokens (input/output), cache write, credits.
  • Each tab fetches dim=<tab>&limit=500, so the hard-coded limit of 6 no longer applies there.
  • Rows link to filtered logs.
• The window controls are pulled into StatsWindowControls and lib/statsWindow.ts and shared with the Overview (1h / 24h / 7d / 30d / custom from-to).
• Provider ranking rows on the Overview now drill down to /logs?provider=. LogsPage has a provider select and syncs provider with the URL.
• logsHref moved to lib/logsLink.ts.

Tests

• store:
  • per-bucket tokens and p50/p95
  • grouped p95, tokens, cache_write and credits (including the usage-detail fallback)
  • dim + limit
  • default limit of 6
  • provider filter
• api: dim/limit over HTTP, the default contract, provider on requests, and the /usage SPA fallback.
• console: ParseStatsDims.
• The frontend has no test runner (no vitest/jest; CI's npm test is for the worker), so there are no frontend unit tests. I didn't add a test framework.