@javenlu233/dsh-session-cost

@javenlu233/dsh-session-cost

English | 中文

Function plugin registering the sessionCost projection unit: whole-log token cost folded from provider usage reports and priced by a configurable per-model, peak/off-peak table, served through the session-projection seam (registry snapshot, change feed, and every projection carrier: history tail page, session/projection push frames, session list rows). Clients render full-session cost figures that paging and compaction cannot change; the reference consumer is the web Cost view tab.

Pricing model

Each usage sample is attributed to the newest recorded model route (request/context) and to the peak/off-peak period of the sample's event time, then billed at that route's per-1M-token rate:

cost = uncachedInput × miss + cacheRead × hit + cacheWrite × miss + output × out

cacheWrite is billed at the miss rate because a cache write is a full-price prompt token. Prices live in view, never in state, so re-pricing a folded log after a config change needs no re-fold.

Fold semantics

• Mirrors token-meter's tokenUsage fold: a usage chunk provides an early sample that survives a later request failure, and an assembled assistant/message replaces that step's sample (same turn/step), so a chunk and its message never double count.
• Auxiliary DeepSeek web_search usage is captured from the Messages response and attached as opaque tool/result.meta.sessionCost, then folded additively into the triggering turn (it never replaces conversation usage). It is priced at the search model's own rate (typically deepseek-v4-flash).
• request/context is a last-wins route record for conversation samples; a sample attributes to the newest route, falling back to defaultRoute when none is recorded or a route has no configured price. Each sample also retains the route price revision active at its event time, so later price changes do not reprice history.
• Peak windows are [start, end) hours on Mon–Fri in a fixed-offset timezone (default Beijing 9:00–12:00 and 14:00–18:00, +480 minutes; weekends and other hours are off-peak).
• Every bucket is 0 until its first contributing event; total sums the four bucket costs, cacheHitPercent is cacheRead / billedInput rounded to an integer, and billedInput sums the three prompt-side buckets.

Composition

- id: session-cost
  name: '@javenlu233/dsh-session-cost'

All config fields default to the DeepSeek peak/off-peak table; override any field in a later patch layer. Prices are CNY per 1M tokens, keyed by provider-owned model id. Use updates for later route-specific peak/off-peak revisions:

- id: session-cost
  name: '@javenlu233/dsh-session-cost'
  config:
    currency: CNY
    defaultRoute: deepseek-flash
    peakWindows: [[9, 12], [14, 18]]
    timezoneOffsetMinutes: 480
    prices:
      deepseek-flash:
        flat: { cacheRead: 0.02, uncachedInput: 1.0, cacheWrite: 1.0, output: 2.0 }
        peak: { cacheRead: 0.10, uncachedInput: 3.0, cacheWrite: 3.0, output: 9.0 }
        offPeak: { cacheRead: 0.05, uncachedInput: 1.5, cacheWrite: 1.5, output: 4.5 }
        updates:
          - effectiveAt: 1789012800000 # 2026-09-10 12:00 Beijing time
            peak: { cacheRead: 0.04, uncachedInput: 2.0, cacheWrite: 2.0, output: 8.0 }
            offPeak: { cacheRead: 0.02, uncachedInput: 1.0, cacheWrite: 1.0, output: 4.0 }
      deepseek-v4-flash:
        flat: { cacheRead: 0.02, uncachedInput: 1.0, cacheWrite: 1.0, output: 2.0 }
        peak: { cacheRead: 0.10, uncachedInput: 3.0, cacheWrite: 3.0, output: 9.0 }
        offPeak: { cacheRead: 0.05, uncachedInput: 1.5, cacheWrite: 1.5, output: 4.5 }
        updates:
          - effectiveAt: 1789012800000
            peak: { cacheRead: 0.04, uncachedInput: 2.0, cacheWrite: 2.0, output: 8.0 }
            offPeak: { cacheRead: 0.02, uncachedInput: 1.0, cacheWrite: 1.0, output: 4.0 }
      deepseek-v4-flash-vision-exp:
        flat: { cacheRead: 0.02, uncachedInput: 1.0, cacheWrite: 1.0, output: 2.0 }
        peak: { cacheRead: 0.10, uncachedInput: 3.0, cacheWrite: 3.0, output: 9.0 }
        offPeak: { cacheRead: 0.05, uncachedInput: 1.5, cacheWrite: 1.5, output: 4.5 }
        updates:
          - effectiveAt: 1789012800000
            peak: { cacheRead: 0.04, uncachedInput: 2.0, cacheWrite: 2.0, output: 8.0 }
            offPeak: { cacheRead: 0.02, uncachedInput: 1.0, cacheWrite: 1.0, output: 4.0 }
      deepseek-v4-pro:
        flat: { cacheRead: 0.025, uncachedInput: 3.0, cacheWrite: 3.0, output: 6.0 }
        peak: { cacheRead: 0.30, uncachedInput: 9.0, cacheWrite: 9.0, output: 27.0 }
        offPeak: { cacheRead: 0.15, uncachedInput: 4.5, cacheWrite: 4.5, output: 13.5 }
        updates:
          - effectiveAt: 1789358400000 # 2026-09-14 12:00 Beijing time
            peak: { cacheRead: 0.04, uncachedInput: 2.0, cacheWrite: 2.0, output: 8.0 }
            offPeak: { cacheRead: 0.02, uncachedInput: 1.0, cacheWrite: 1.0, output: 4.0 }

The default table is DeepSeek's published peak/off-peak pricing (old V4 pricing effective 2026-08-17, V4.1-Flash pricing effective 2026-09-10, and Pro routing effective 2026-09-14); deployments that price another provider or a custom rate override prices wholesale and maintain updates themselves.

Injects sessionProjections — the plugin's whole purpose; in assemblies without the registry the fiber stays pending and nothing registers.

Model Experience

Conversation tokens: none; those are already-logged usage events.

Auxiliary web_search: the plugin reads the existing DeepSeek Messages response (the harness already performs this call) and records its usage onto tool/result meta. It does not add tools, change the search prompt, or alter the model-visible tool result text.

KV Cache effect

None for conversation requests. Search-call cache traffic, when the provider reports it, is billed in the search cut.

Known Limitations and Deferred Work

• Estimate, not an invoice — prices come from the configured table, the peak period uses each usage sample's event time (the assembled-message time, not the request start), and a mid-session model switch is priced only to the resolution of request/context; the figure may differ from a provider bill.
• Cumulative totals only — the fold publishes whole-session buckets, not a per-turn or per-step breakdown; the per-(route, period) keying keeps the door open for that later.
• Cache-write is provider-optional — DeepSeek reports no cache-write, so its sessions show a zero cache-write bucket; the bucket exists for providers that do report it.
• Search tokens need a live capture — sessions logged before this plugin, or web_search calls that never went through the web_search tool (direct ctx.web.search), have no tool/result.meta.sessionCost and cannot recover the flash search usage.