[Task 06] Usage recorder

[lubobill1990]

Task 06 — Usage recorder

Depends on: 01, 05
Unblocks: 07, 08, 09, 10

Goal

Single-call API that writes one usage_events row and atomically updates the
usage_daily aggregate.

Scope

New file src/lib/usage-recorder.ts:

export interface RecordUsageInput {
  account: Account
  modelId: string
  endpoint: 'chat.completions' | 'messages' | 'embeddings'
  upstreamFormat: 'openai' | 'anthropic'
  isStreaming: boolean
  usage: NormalizedUsage
  durationMs: number
  status: 'ok' | 'error' | 'aborted'
  requestId?: string
  isInternal?: boolean   // gate from x-internal-pricing-sync
}

export function recordUsage(input: RecordUsageInput): void

Behavior:

• If isInternal === true, return immediately (no insert).
• Look up current model_pricing row by modelId. Missing → all six
  *_price_snapshot columns are NULL. premiumMultiplier likewise → 0.
• Compute premium_request_count = 1 * (premiumMultiplier ?? 0).
• In one transaction:
  1. INSERT INTO usage_events (...).
  2. INSERT INTO usage_daily (...) ON CONFLICT(day, account_name, model_id, endpoint) DO UPDATE SET req_count = req_count + 1, input_tokens = input_tokens + excluded.input_tokens, ....
• Wrap the entire body in try/catch; recorder errors must NOT propagate.
  Log via consola.error.

Day computation

date(?ts/1000, 'unixepoch', 'localtime')

Use SQLite's expression so the boundary matches the user's local timezone
without dragging Node's Intl into the hot path.

Definition of Done

• Unit test against in-memory DB: single insert produces 1 event row +
  1 daily row with the right counts.
• Unit test: a second insert into the same (day, account, model, endpoint) increments the daily row.
• Unit test: missing model_pricing row → snapshots are NULL, no throw.
• Unit test: isInternal: true → 0 rows inserted.
• Unit test: SQL error is swallowed; outer caller does not see exception.

---

• Source: docs/tasks/06-usage-recorder.md
• Design: docs/design/
• Phase: foundation
• Depends on tasks: 1, 5