@logan-luo/dsh-session-cost

English | 中文

Session cost accounting, node half: a versioned rate-card engine, the per-session costStats projection, and the durable ledger behind the cost Remote namespace. The surfaces (dock cost strip, cost view tab, usage dashboard) live in @logan-luo/dsh-client-ui-session-cost.

Shipped surfaces

• Rate-card engine (src/pricing.ts): versioned official DeepSeek CNY pricing with effective-date boundaries and Beijing peak windows; pure functions shared with the browser half.
• costStats projection unit (src/projection.ts): prices every usage-bearing assistant/message from its model, billing instant (the open step's start), and the rate-card version in force; folds per-model, per-day, and per-rate-version buckets. Registered on ctx.sessionProjections, so the value reaches the browser through the standard session-projection feed (useProjection('costStats')).
• Ledger service (src/index.ts, SessionCostService extends TypertRemoteService): reconciles the ledger over the ctx.sessionQuery corpus in the background — a warm-up pass at startup, then one incremental pass per reconcileIntervalMs tick (ctx.interval, fiber-bound) — folding each session's log past its scan watermark, pruning deleted sessions, and containing unreadable logs. cost.dashboard is a pure read over the latest successfully reconciled ledger: opening the dashboard or switching filters never waits on a corpus scan, a failed pass keeps the previous data readable and retries next tick, and concurrent passes never overlap. Rows persist in the session_cost storage domain.
• Rollups (src/ledger.ts): pure folds and aggregations — per-model, per-day/week/month (Beijing calendar), per-project, with billing-instant bounds.

Requests whose model has no official card are counted as unpriced rather than guessed.

Model Experience

None: the package assembles no model request. The projection unit and ledger only observe assistant/message usage and provenance from the session event stream and the persisted logs.

KV Cache effect

None; the package neither assembles nor sends a provider request.

Known Limitations and Deferred Work

• The rate table is hardcoded to the official DeepSeek CNY card published at https://api-docs.deepseek.com/zh-cn/quick_start/pricing/. A future config key can accept a deployment override without changing the engine.
• Peak windows are calendarized in a fixed Beijing offset (DEEPSEEK_BEIJING_OFFSET_MINUTES); a named timezone (Asia/Shanghai) stays a deferred config option.
• Reconciliation re-reads every session log whose events grew past its watermark (with bounded read concurrency), and each dashboard call materializes the row table in memory; a per-log fingerprint to skip untouched logs and an indexed rollup are deferred until large-corpus deployments need them. reconcileIntervalMs (default 5000) tunes the background reconcile period; the schema floor of 1000 ms guards against accidental high-frequency scans.
• Crash consistency: the startup snapshot serves durable rows covered by a written watermark, which is exactly the last committed ledger. A process killed mid-pass can leave a partially refolded reused-session id (new rows under an old watermark) that becomes visible until the first pass detects the lifecycle change and re-folds it; distinguishing it at startup would require a corpus scan, which the snapshot design deliberately avoids.
• Browser cache: switching back to a cached selection republishes its rollup immediately and revalidates; on a Beijing calendar rollover between the switch and the revalidate the previous period's cached value can show for one render frame before the fresh rollup replaces it.