dsh-cost-meter 1.7.19 → 1.7.21

package/lib/index.js

@@ -22,6 +22,7 @@ import { resolveDshHome } from '@deepseek-ai/dsh-home-paths'
 import { Ledger, applyConfigPatch, localDayKey, pickBalanceInfo, reconcileBalanceDelta, zeroDay, splitLedgerApiCost, repairLedgerPricing, dedupeWrapperProviderDays, unpriceLocalOriginModels, stripSecrets, stripSecretPatch, secretRefOf, readSecret, writeSecret, SECRET_TARGETS, looksLikeSecretHeaderValue } from './store.js'
 import { backfillLegacyLedger, importLegacyHistory, repairForkSeed, repairProviderDupes, recomputeLedgerPricingBasis } from './backfill.js'
 import { createLlmStreamBilling } from './billing-stream.js'
+import { installNativeSearchBilling, nativeSearchReconcile, isNativeSearchUsageEvent } from './native-search-billing.js'
 import { OFFICIAL_PRICING_URL, OFFICIAL_PRICING_URL_ZH, LEGACY_BASE_BOUNDARY, FLASH_PRICE_EFFECTIVE_AT, PRO_FLASH_ROUTING_EFFECTIVE_AT, normalizePrice, parsePricingHtml, repairDefaultPeakPrice, upgradeDeepSeekPriceTable, costOf, providerPriceEntryFor, buildPriceCatalog, usdFromCost, isWrapperProviderId, wrapperUpstreamProvider } from './pricing.js'
 import { createUsageDeduper, USAGE_DEDUP_WINDOW_MS, usageFingerprint } from './usage-dedup.js'
 import { CODING_PLAN_PROVIDERS, CODING_PLAN_PROVIDER_IDS, queryCodingPlan, scnetTokenPlanWindows, qwenTokenPlanWindows, emptyCustomBalance, queryCustomBalance, normalizeVolcengineKey } from './coding-plans.js'
@@ -31,6 +32,7 @@ import { queryGatewayQuota, emptyGatewayQuota, managementKeyVarOf, gatewaySource
 import { stateSchema } from './typert.host.js'
 import { customBalanceCredentialVars } from './custom-balance.js'
 import { readScnetSnapshot } from './scnet-snapshot.js'
+import { getSessionCost as readSessionCost } from './session-tree.js'
 
 export const name = 'cost-meter'
 
@@ -281,6 +283,7 @@ const usageProjectionStateSchema = z.object({
     at: z.number(),
     w: z.number().optional(),
   })).max(32).optional(),
+  nativeSearchIds: z.array(z.string()).max(128).optional(),
 })
 
 /** state → wire payload 读侧投影(新版 wire.view 与旧版 view 共用同一实现)。 */
@@ -377,8 +380,16 @@ function makeCostUsageProjection(ledger) {
     // v7→v8(issue #77):折叠计入 compaction/summary(压缩摘要调用,此前漏计);
     // 升版本触发宿主对旧 checkpoint 全量重放,历史摘要调用量随之补齐。
     // v8→v9(issue #109):默认回退补齐峰谷档位，旧金额 checkpoint 需按事件重放。
-    stateVersion: 9,
-    init: () => ({ provider: 'deepseek', model: 'default', totals: zeroBuckets(), byModel: {}, byProviderModel: {}, last: null, createdAt: 0, seedEndSeq: -1, shadow: emptyShadow(), seedLength: -1, seedDeducted: false, recent: [] }),
+    // v10 (#130):新版宿主 init(header, inheritedEventCount) 区分恢复和继承。
+    // 普通重启也追加 end-seed，不能把恢复前已付费的调用当作 fork 种子扣掉。
+    stateVersion: 10,
+    init: (header, inheritedEventCount) => {
+      const known = Number.isInteger(inheritedEventCount) && inheritedEventCount >= 0
+      const boundary = known && inheritedEventCount > 0 ? inheritedEventCount : -1
+      return { provider: 'deepseek', model: 'default', totals: zeroBuckets(), byModel: {}, byProviderModel: {}, last: null,
+        createdAt: Number(header?.createdAt) || 0, seedEndSeq: boundary, shadow: emptyShadow(),
+        seedLength: boundary, seedDeducted: known, recent: [] }
+    },
     apply(state, event) {
       // 兼容旧 checkpoint 的缺字段(版本升级前持久化的 v5 状态):缺省回落。
       if (state.seedDeducted === undefined) state.seedDeducted = false
@@ -478,6 +489,8 @@ function makeCostUsageProjection(ledger) {
       let eventProvider = null
       let eventModel = null
       let keyOverride = null
+      const nativeSearch = isNativeSearchUsageEvent(event)
+      let nativeSearchIds = state.nativeSearchIds ?? []
       // 判空用 != null:usage === null 时 !== undefined 会放行,随后读
       // usage.inputTokens 直接抛 TypeError 打断投影;billing-stream 侧同处
       // 还会让 null 覆盖先前捕获的有效 usage 快照导致整次调用漏计。
@@ -489,6 +502,13 @@ function makeCostUsageProjection(ledger) {
         usage = event.data.usage
         turn = event.data.turn ?? 0
         step = event.data.step ?? 0
+      } else if (nativeSearch && event.data?.usage != null && typeof event.data.requestId === 'string') {
+        if (nativeSearchIds.includes(event.data.requestId)) return state
+        nativeSearchIds = [...nativeSearchIds.slice(-127), event.data.requestId]
+        usage = event.data.usage
+        eventProvider = event.data.provider
+        eventModel = event.data.model
+        keyOverride = `search:${event.data.requestId}`
       } else if (event.type === 'compaction/summary' && event.data?.usage != null) {
         usage = event.data.usage
         const source = event.data.message?.source ?? {}
@@ -528,7 +548,8 @@ function makeCostUsageProjection(ledger) {
         return state
       }
       // 按事件时刻计费(历史正确):峰谷时代前用 legacyBase,之后按峰谷两档。
-      const atMs = Number.isFinite(Number(event.time)) && Number(event.time) > 0 ? Number(event.time) : Date.now()
+      const billingTime = nativeSearch ? event.data.startedAtMs : event.time
+      const atMs = Number.isFinite(Number(billingTime)) && Number(billingTime) > 0 ? Number(billingTime) : Date.now()
       // 指纹窗口去重(与 lib/usage-dedup.js 同语义的序列化形态):先按窗口清扫
       // 滚动列表,重复转发跳过(清扫结果仍落回 state,保持窗口推进)。有界:
       // 超过 24 条先截断再判定,checkpoint 体积恒定。compaction/summary 不参与:
@@ -603,7 +624,7 @@ function makeCostUsageProjection(ledger) {
       }
       // createdAt/seedLength/seedDeducted 必须随状态携带:usage 样本更新不能丢掉 fork 过滤基准。
       // recent 随行:包装层转发对去重的滚动窗口状态(有界 ≤25 条)。
-      return { provider: state.provider, model: state.model, totals, byModel, byProviderModel, createdAt: state.createdAt, seedEndSeq: state.seedEndSeq, seedLength: state.seedLength, seedDeducted: state.seedDeducted, shadow: shadowAgg, recent, last: { key, provider: effectiveProvider, model, buckets, cost: billed } }
+      return { provider: state.provider, model: state.model, totals, byModel, byProviderModel, createdAt: state.createdAt, seedEndSeq: state.seedEndSeq, seedLength: state.seedLength, seedDeducted: state.seedDeducted, shadow: shadowAgg, recent, nativeSearchIds, last: nativeSearch ? state.last : { key, provider: effectiveProvider, model, buckets, cost: billed } }
     },
     view: projectionView,
     // DSH 0.1.1-rc.1 起会话投影需声明 wire 才会向客户端推送(PR #39 by
@@ -1095,7 +1116,7 @@ async function buildState(ledger, balance = emptyBalance(), goQuota = emptyGoQuo
     // 自定义余额 {{VAR}} 占位符凭据状态(v1.7.6,issue #86):变量名 → { configured, source }。
     customVarStatus,
     // 余额差交叉校验提示(issue #18):本地今日合计与官方余额当日变动偏差超阈时 ok=false。
-    reconcile,
+    reconcile: nativeSearchReconcile(ledger, localeOf(ledger.config), reconcile),
     codingPlans,
     planStats,
     history: ledger.history(90),
@@ -2128,6 +2149,9 @@ function createService(ctx, ledger) {
 
     // 跨全部日期返回前 N 个会话(issue #22 按会话视角,不分日期)。
     // sort:cost(费用) | time(会话创建时间) | recent(实时顺序,即账本/侧边栏顺序);dir:asc | desc。
+    async getSessionCost(sessionId) {
+      return readSessionCost(ledger, ctx, sessionId)
+    },
     async getTopSessions(limit, sort = 'cost', dir = 'desc') {
       const n = Math.max(1, Math.min(500, Math.floor(Number(limit)) || 100))
       const sortKey = sort === 'time' || sort === 'recent' ? sort : 'cost'
@@ -2800,6 +2824,8 @@ export function apply(ctx) {
   }), { global: true })
 
   // costUsage 投影:向会话历史页/推送帧提供 token 桶(客户端计价)。
+  installNativeSearchBilling(ctx, ledger)
+
   ctx.inject(['sessionProjections'], (projectionCtx) => {
     projectionCtx.sessionProjections.register(makeCostUsageProjection(ledger))
   })

package/lib/native-search-billing.js

@@ -0,0 +1,278 @@
+/**
+ * DeepSeek 原生搜索不经过 llm/stream，且宿主会丢弃 Messages 响应的 usage。
+ * 只在 web.search 的异步作用域内订阅官方 Messages 响应诊断；不替换 fetch、
+ * 不读取请求 headers/body，也不改变搜索返回值、取消或错误语义。
+ */
+import { AsyncLocalStorage } from 'node:async_hooks'
+import { channel } from 'node:diagnostics_channel'
+import { randomUUID } from 'node:crypto'
+import { readFileSync, writeFileSync, renameSync, mkdirSync, statSync } from 'node:fs'
+import { dirname } from 'node:path'
+import { gunzipSync, inflateSync, brotliDecompressSync } from 'node:zlib'
+
+export const NATIVE_SEARCH_USAGE_EVENT = 'cost-meter/native-search-usage'
+export function isNativeSearchUsageEvent(event) {
+  const data = event?.data
+  return event?.type === NATIVE_SEARCH_USAGE_EVENT && data?.provider === 'deepseek-official'
+    && typeof data.model === 'string' && /^deepseek-[a-z\d._:-]{1,120}$/i.test(data.model)
+    && typeof data.requestId === 'string' && /^[a-f\d]{8}(?:-[a-f\d]{4}){3}-[a-f\d]{12}$/i.test(data.requestId)
+    && Number.isFinite(data.startedAtMs) && data.startedAtMs > 0
+    && ['inputTokens', 'outputTokens', 'cacheReadTokens', 'cacheWriteTokens', 'reasoningTokens'].every(key => Number.isSafeInteger(data.usage?.[key]) && data.usage[key] >= 0)
+}
+const MAX_BODY_BYTES = 4 * 1024 * 1024
+const monitors = new WeakMap()
+
+const dayKey = (at) => {
+  const date = new Date(at)
+  return `${date.getFullYear()}-${String(date.getMonth() + 1).padStart(2, '0')}-${String(date.getDate()).padStart(2, '0')}`
+}
+const count = (value) => Number.isSafeInteger(value) && value >= 0
+
+/** Anthropic 的 input_tokens 不含两类 cache token；五桶不能重复加总。 */
+export function nativeSearchUsage(response) {
+  const source = response?.usage
+  if (!source || !count(source.input_tokens) || !count(source.output_tokens)) return null
+  for (const key of ['cache_read_input_tokens', 'cache_creation_input_tokens']) {
+    if (source[key] !== undefined && !count(source[key])) return null
+  }
+  if (typeof response.model !== 'string' || !/^deepseek-[a-z\d._:-]{1,120}$/i.test(response.model)) return null
+  return {
+    model: response.model,
+    usage: {
+      inputTokens: source.input_tokens,
+      outputTokens: source.output_tokens,
+      cacheReadTokens: source.cache_read_input_tokens ?? 0,
+      cacheWriteTokens: source.cache_creation_input_tokens ?? 0,
+      reasoningTokens: 0,
+    },
+  }
+}
+
+function officialRequest(request) {
+  return request?.method === 'POST' && String(request.origin) === 'https://api.deepseek.com'
+    && request.path === '/anthropic/v1/messages'
+}
+
+function responseEncoding(headers) {
+  if (Array.isArray(headers)) {
+    for (let i = 0; i + 1 < headers.length; i += 2) {
+      if (String(headers[i]).toLowerCase() === 'content-encoding') return String(headers[i + 1]).trim().toLowerCase()
+    }
+    return ''
+  }
+  return String(headers?.['content-encoding'] ?? '').trim().toLowerCase()
+}
+
+function parseBody(request) {
+  let bytes = Buffer.concat(request.chunks, request.bytes)
+  const options = { maxOutputLength: MAX_BODY_BYTES }
+  if (request.encoding === 'gzip') bytes = gunzipSync(bytes, options)
+  else if (request.encoding === 'deflate') bytes = inflateSync(bytes, options)
+  else if (request.encoding === 'br') bytes = brotliDecompressSync(bytes, options)
+  else if (request.encoding !== '' && request.encoding !== 'identity') return null
+  return nativeSearchUsage(JSON.parse(bytes.toString('utf8')))
+}
+
+/**
+ * 可注入诊断通道/目标判定仅用于不付费的本地回归；生产固定使用官方端点。
+ * 每个 HTTP request 独立计数，相同 token 的并发/重复搜索不会相互去重。
+ */
+export function createNativeSearchBilling({ account, record = () => {}, uncovered = () => {}, now = Date.now, targetRequest = officialRequest, diagnostics = channel }) {
+  const scope = new AsyncLocalStorage()
+  const requests = new WeakMap()
+  const pending = new Set()
+  const subscriptions = []
+  let active = true
+  const subscribe = (name, listener) => {
+    const source = diagnostics(`undici:request:${name}`)
+    // diagnostics_channel 的监听器不得抛错，否则 Node 会作为 uncaughtException 处理。
+    const safe = (event) => { try { if (active) listener(event) } catch {} }
+    source.subscribe(safe)
+    subscriptions.push(() => source.unsubscribe(safe))
+  }
+  const miss = (item, reason) => {
+    if (item.coverageReported) return
+    item.coverageReported = true
+    try { uncovered({ startedAtMs: item.startedAtMs, reason }) } catch {}
+  }
+  const capture = (item, chunk, source) => {
+    if (!active || item.done || item.status < 200 || item.status >= 300 || item.tooLarge) return
+    // 兼容层与诊断通道可能同时出现，只选先到达的一条，不重复累计响应字节。
+    if (item.captureSource && item.captureSource !== source) return
+    item.captureSource = source
+    item.bytes += chunk.byteLength
+    if (item.bytes > MAX_BODY_BYTES) { item.tooLarge = true; item.chunks = []; return }
+    item.chunks.push(Buffer.from(chunk))
+  }
+  const release = (request, item) => {
+    item.chunks = []
+    try { item.restore?.() } catch {}
+    item.restore = undefined
+    item.release = undefined
+    pending.delete(item)
+    requests.delete(request)
+  }
+  subscribe('create', ({ request }) => {
+    const current = scope.getStore()
+    if (!current || !targetRequest(request)) return
+    const item = { session: current.session, startedAtMs: now(), requestId: randomUUID(), bytes: 0, chunks: [], encoding: '', status: 0, done: false }
+    requests.set(request, item)
+    current.requests.push(item)
+    pending.add(item)
+    item.release = () => release(request, item)
+    // Node 20/22 的 undici 6 尚无响应 body 诊断通道。只观察这一请求的
+    // onData，不更换 dispatcher/fetch；保留 this/返回值，并在结束/卸载时恢复。
+    // 这是有能力检测的旧运行时兼容层；宿主更换网络实现后退回缺口提示。
+    const own = Object.getOwnPropertyDescriptor(request, 'onData')
+    const original = request.onData
+    if (typeof original === 'function' && own?.configurable !== false && Object.isExtensible(request)) {
+      const wrapped = function (...args) {
+        try { capture(item, args[0], 'onData') } catch {}
+        return Reflect.apply(original, this, args)
+      }
+      Object.defineProperty(request, 'onData', { configurable: true, writable: true, value: wrapped })
+      item.restore = () => {
+        if (Object.getOwnPropertyDescriptor(request, 'onData')?.value !== wrapped) return
+        if (own) Object.defineProperty(request, 'onData', own)
+        else delete request.onData
+      }
+    }
+  })
+  subscribe('headers', ({ request, response }) => {
+    const item = requests.get(request)
+    if (!item) return
+    item.status = response.statusCode
+    item.encoding = responseEncoding(response.headers)
+  })
+  subscribe('bodyChunkReceived', ({ request, chunk }) => {
+    const item = requests.get(request)
+    if (item) capture(item, chunk, 'diagnostic')
+  })
+  subscribe('trailers', ({ request }) => {
+    const item = requests.get(request)
+    if (!item || item.done) return
+    item.done = true
+    try {
+      if (item.status < 200 || item.status >= 300) return
+      const parsed = item.tooLarge || item.bytes === 0 ? null : parseBody(item)
+      if (!parsed) { miss(item, 'usage-unavailable'); return }
+      const event = { ...parsed, provider: 'deepseek-official', startedAtMs: item.startedAtMs, requestId: item.requestId }
+      try { account(event, item.session) } catch { miss(item, 'account-failed'); return }
+      if (item.session) {
+        try { record(item.session, event) } catch { miss(item, 'history-unavailable') }
+      } else miss(item, 'history-unavailable')
+    } catch {
+      miss(item, 'usage-unavailable')
+    } finally {
+      release(request, item)
+    }
+  })
+  subscribe('error', ({ request }) => {
+    const item = requests.get(request)
+    if (!item || item.done) return
+    item.done = true
+    // 请求已派发后的中断可能发生费用，但没有完整 usage 时不能猜测金额。
+    miss(item, 'response-incomplete')
+    release(request, item)
+  })
+  return {
+    async run(operation, session) {
+      if (!active || scope.getStore()) return operation()
+      const current = { session, requests: [] }
+      try { return await scope.run(current, operation) }
+      finally {
+        for (const item of current.requests) {
+          if (!item.done) {
+            item.done = true
+            if (active) miss(item, 'response-incomplete')
+            item.release()
+          }
+        }
+      }
+    },
+    dispose() {
+      active = false
+      for (const unsubscribe of subscriptions) unsubscribe()
+      for (const item of pending) { item.done = true; item.release() }
+      pending.clear()
+      scope.disable()
+    },
+  }
+}
+
+/** 仅保存计数和日期，重启仍能解释未覆盖调用；不写查询、响应正文或凭据。 */
+export function createSearchCoverage(path, now = Date.now) {
+  let days = {}
+  try {
+    const raw = statSync(path).size < 64 * 1024 ? readFileSync(path, 'utf8') : ''
+    if (raw.length < 64 * 1024) {
+      const parsed = JSON.parse(raw)
+      for (const [day, total] of Object.entries(parsed.days ?? {})) {
+        if (/^\d{4}-\d{2}-\d{2}$/.test(day) && count(total)) days[day] = total
+      }
+    }
+  } catch {}
+  let timer
+  const flush = () => {
+    clearTimeout(timer)
+    timer = undefined
+    try {
+      mkdirSync(dirname(path), { recursive: true })
+      writeFileSync(`${path}.tmp`, JSON.stringify({ days }))
+      renameSync(`${path}.tmp`, path)
+    } catch { console.warn('[dsh-cost-meter] 原生搜索覆盖提示未能持久化') }
+  }
+  return {
+    add({ startedAtMs }) {
+      const day = dayKey(startedAtMs)
+      days[day] = Math.min(Number.MAX_SAFE_INTEGER, (days[day] ?? 0) + 1)
+      days = Object.fromEntries(Object.entries(days).sort(([a], [b]) => a.localeCompare(b)).slice(-90))
+      if (!timer) { timer = setTimeout(flush, 1000); timer.unref?.() }
+    },
+    today() { return days[dayKey(now())] ?? 0 },
+    close() { if (timer) flush() },
+  }
+}
+
+/** 保留已有对账偏差信息，将真实发生的搜索覆盖缺口追加到同一个提示。 */
+export function nativeSearchReconcile(ledger, locale, reconcile) {
+  const missed = monitors.get(ledger)?.coverage.today() ?? 0
+  if (!missed) return reconcile
+  const message = locale === 'en'
+    ? `Native search coverage: ${missed} official search request(s) today lack complete usage or a durable usage event; their tokens or history may be missing. The current host/Node transport must expose response usage for exact accounting; no per-search estimate was added.`
+    : `原生搜索统计缺口：今日 ${missed} 次官方搜索请求未取得完整 usage 或未能持久化用量事件，费用或历史可能漏记。精确计费需要宿主/Node 网络实现提供响应 usage；未按搜索次数虚构金额。`
+  return { ok: false, message: [reconcile?.message, message].filter(Boolean).join('\n') }
+}
+
+/** 可选 web 服务：未安装搜索能力时不影响插件启动。 */
+export function installNativeSearchBilling(ctx, ledger) {
+  if (monitors.has(ledger)) return
+  const coverage = createSearchCoverage(`${ledger.path}.native-search-coverage.json`)
+  const monitor = createNativeSearchBilling({
+    account(event, session) {
+      const usage = event.usage
+      ledger.account({ input: usage.inputTokens, output: usage.outputTokens, cacheRead: usage.cacheReadTokens, cacheWrite: usage.cacheWriteTokens, reasoning: usage.reasoningTokens }, event.model, session?.id, event.startedAtMs, event.provider)
+    },
+    record(session, event) { session.append(NATIVE_SEARCH_USAGE_EVENT, event) },
+    uncovered: (event) => coverage.add(event),
+  })
+  monitors.set(ledger, { coverage })
+  ctx.effect(() => () => { monitor.dispose(); coverage.close(); monitors.delete(ledger) }, 'cost-meter: native search billing')
+  ctx.inject(['web'], (webCtx) => {
+    const web = webCtx.web
+    const own = Object.getOwnPropertyDescriptor(web, 'search')
+    const original = own?.value ?? Object.getOwnPropertyDescriptor(Object.getPrototypeOf(web), 'search')?.value
+    if (typeof original !== 'function' || own?.configurable === false) return
+    const wrapped = function (...args) {
+      let session
+      try { session = this.ctx?.get?.('agents')?.currentInitiator?.()?.session ?? webCtx.get?.('agents')?.currentInitiator?.()?.session } catch {}
+      return monitor.run(() => Reflect.apply(original, this, args), session)
+    }
+    Object.defineProperty(web, 'search', { configurable: true, writable: true, value: wrapped })
+    webCtx.effect(() => () => {
+      if (Object.getOwnPropertyDescriptor(web, 'search')?.value !== wrapped) return
+      if (own) Object.defineProperty(web, 'search', own)
+      else delete web.search
+    }, 'cost-meter: web search observation')
+  })
+}

package/lib/pricing.js

@@ -109,6 +109,7 @@ export const DEFAULT_PEAK_WINDOWS = [
 export const DEFAULT_PROVIDER_PRICE_TABLE = {
   openai: {
     models: {
+      'gpt-6-astra': { input: 10, cachedInput: 1, cacheWrite: 12.5, output: 50, longContext: { aboveInputTokens: 272000, cacheMiss: 20, cacheHit: 2, cacheWrite: 25, output: 75 }, billingMode: 'flat', sourceUrl: 'https://developers.openai.com/api/docs/models/gpt-6-astra', checkedAt: '2026-09-11', notes: '标准 API 价，OpenCode Zen 同价；完整输入（含缓存读写）超过 272K 时整次请求按长上下文档计费；不含 Batch/Flex/Fast 服务档折扣或倍率' },
       'gpt-5.6-sol': { input: 2, cachedInput: 0.2, output: 10, billingMode: 'flat', sourceUrl: 'https://opencode.ai/docs/zen', checkedAt: '2026-08-25', notes: '≤272K 档;超过 272K 按 $4/$15 计(缓存读 $0.40、写入 $5);缓存写入 $2.50;目录标注 2026-09-18 前为五折促销价(issue #58)' },
       'gpt-5.6-terra': { input: 2, cachedInput: 0.2, output: 12, billingMode: 'flat', sourceUrl: 'https://opencode.ai/docs/zen', checkedAt: '2026-08-17', notes: '≤272K 档;超过 272K 按 $4/$18 计;缓存写入 $2.50' },
       'gpt-5.6-luna': { input: 0.2, cachedInput: 0.02, output: 1.2, billingMode: 'flat', sourceUrl: 'https://opencode.ai/docs/zen', checkedAt: '2026-08-17', notes: '≤272K 档;超过 272K 按 $0.40/$1.80 计;缓存写入 $0.25' },
@@ -285,6 +286,7 @@ export const DEFAULT_PROVIDER_PRICE_TABLE = {
 export const PROVIDER_MODEL_FAMILIES = {
   deepseek: { 'deepseek-v4-flash': 'DeepSeek v4', 'deepseek-v4-pro': 'DeepSeek v4', 'deepseek-v4-flash-vision-exp': 'DeepSeek v4', 'deepseek-v4.1-flash': 'DeepSeek v4.1' },
   openai: {
+    'gpt-6-astra': 'GPT-6 Astra',
     'gpt-5.6-sol': 'GPT-5.6', 'gpt-5.6-terra': 'GPT-5.6', 'gpt-5.6-luna': 'GPT-5.6',
     'gpt-5.5': 'GPT-5.5', 'gpt-5.5-pro': 'GPT-5.5',
     'gpt-5.4': 'GPT-5.4', 'gpt-5.4-pro': 'GPT-5.4', 'gpt-5.4-mini': 'GPT-5.4', 'gpt-5.4-nano': 'GPT-5.4',
@@ -542,7 +544,8 @@ function completeTier(raw) {
   const cacheHit = n('cacheHit') ?? n('cachedInput') ?? n('cacheRead') ?? cacheMiss
   const output = n('output') ?? 0
   const reasoning = n('reasoning')
-  return reasoning === undefined ? { cacheHit, cacheMiss, output } : { cacheHit, cacheMiss, output, reasoning }
+  const cacheWrite = n('cacheWrite')
+  return { cacheHit, cacheMiss, output, ...(reasoning === undefined ? {} : { reasoning }), ...(cacheWrite === undefined ? {} : { cacheWrite }) }
 }
 
 /**
@@ -559,6 +562,16 @@ export function normalizePrice(value) {
   }
   if (!('cacheHit' in value) && !('cacheMiss' in value) && !('output' in value) && !('input' in value)) return null
   const entry = completeTier(value)
+  const validRate = n => typeof n === 'number' && Number.isFinite(n) && n >= 0
+  if (value.cacheWrite !== undefined && !validRate(value.cacheWrite)) return null
+  if (value.longContext !== undefined) {
+    const long = value.longContext
+    if (!long || typeof long !== 'object' || Array.isArray(long)
+      || !validRate(long.aboveInputTokens) || long.aboveInputTokens === 0
+      || !['cacheMiss', 'cacheHit', 'output'].every(key => validRate(long[key]))
+      || (long.cacheWrite !== undefined && !validRate(long.cacheWrite))) return null
+    entry.longContext = { aboveInputTokens: long.aboveInputTokens, ...completeTier(long) }
+  }
   if (value.legacy === true) entry.legacy = true
   if (value.billingMode === 'flat' || value.billingMode === 'deepseek-peak' || value.billingMode === 'batch') entry.billingMode = value.billingMode
   for (const key of ['sourceUrl', 'checkedAt', 'notes']) if (typeof value[key] === 'string') entry[key] = value[key]
@@ -594,6 +607,7 @@ export function normalizePrice(value) {
 /** 全部价格为 0 的记录视为空记录。 */
 export function isZeroPrice(entry) {
   return entry !== null && entry.cacheHit === 0 && entry.cacheMiss === 0 && entry.output === 0
+    && !(entry.cacheWrite > 0) && (!entry.longContext || isZeroPrice(entry.longContext))
 }
 
 /**
@@ -1021,7 +1035,7 @@ export function peakPhaseAt(atMs, windows) {
 
 /**
  * 为一次用量挑选价格档位:生效后峰时段 → peak;生效后谷时段 → offPeak;
- * 生效前(或禁用峰谷)→ 基础价格。cache write 与 cache hit 同价。
+ * 生效前(或禁用峰谷)→ 基础价格。保留显式缓存写价和单次输入长度档位。
  * @param entry - 模型价格记录。
  * @param atMs - 计费时刻。
  * @param peak - { enabled, effectiveAtMs, windows } 峰谷配置。
@@ -1029,9 +1043,10 @@ export function peakPhaseAt(atMs, windows) {
  */
 export function tierFor(entry, atMs, peak) {
   const base = priceAt(entry, atMs) ?? { cacheHit: 0, cacheMiss: 0, output: 0 }
-  const asTier = price => price.reasoning === undefined
-    ? { cacheHit: price.cacheHit, cacheMiss: price.cacheMiss, output: price.output }
-    : { cacheHit: price.cacheHit, cacheMiss: price.cacheMiss, output: price.output, reasoning: price.reasoning }
+  const asTier = price => ({ cacheHit: price.cacheHit, cacheMiss: price.cacheMiss, output: price.output,
+    ...(price.reasoning === undefined ? {} : { reasoning: price.reasoning }),
+    ...(price.cacheWrite === undefined ? {} : { cacheWrite: price.cacheWrite }),
+    ...(price.longContext === undefined ? {} : { longContext: price.longContext }) })
   // 峰谷时代之前(2026-08-16 16:00 UTC 前):按当时的基础价计费(历史正确性)。
   if (Number.isFinite(atMs) && atMs < Date.parse(LEGACY_BASE_BOUNDARY)) {
     const lb = base.legacyBase
@@ -1062,16 +1077,20 @@ export function tierFor(entry, atMs, peak) {
  * @returns 美元成本(非负)。
  */
 export function costOf(tokens, entry, atMs, peak) {
-  const tier = tierFor(entry, atMs, peak)
+  let tier = tierFor(entry, atMs, peak)
   const input = Math.max(0, Number(tokens?.input) || 0)
   const output = Math.max(0, Number(tokens?.output) || 0)
   const cacheRead = Math.max(0, Number(tokens?.cacheRead) || 0)
   const cacheWrite = Math.max(0, Number(tokens?.cacheWrite) || 0)
   const reasoning = Math.max(0, Number(tokens?.reasoning) || 0)
+  // 宿主的 input/cacheRead/cacheWrite 是互斥桶；输入长度包含缓存，输出不参与阈值。
+  // 此处必须传单次调用，日/会话聚合无法还原每次请求的上下文档位。
+  if (tier.longContext && input + cacheRead + cacheWrite > tier.longContext.aboveInputTokens) tier = tier.longContext
   const reasoningPrice = typeof tier.reasoning === 'number' ? tier.reasoning : 0
   const cost = (input * tier.cacheMiss
     + output * tier.output
-    + (cacheRead + cacheWrite) * tier.cacheHit
+    + cacheRead * tier.cacheHit
+    + cacheWrite * (tier.cacheWrite ?? tier.cacheHit)
     + reasoning * reasoningPrice) / 1_000_000
   // 终值防护:档位字段缺失/非法导致的 NaN/Infinity 与负值一律按 0 入账。
   return Number.isFinite(cost) && cost > 0 ? cost : 0

package/lib/session-tree.js

@@ -0,0 +1,117 @@
+/**
+ * 会话费用展示聚合。父子会话仍分别入账;这里只读取账本,不把子代理费用再次写入父账。
+ * 宿主的 parentSession 也用于普通 fork,只有 origin=subagent 的连续链才属于子代理。
+ */
+
+const FIELDS = ['input', 'output', 'cacheRead', 'cacheWrite', 'reasoning', 'calls', 'cost', 'apiCost']
+const amount = value => typeof value === 'number' && Number.isFinite(value) && value >= 0 ? value : 0
+const validId = value => typeof value === 'string' && value.length > 0 && value.length <= 512
+const emptyBuckets = () => Object.fromEntries(FIELDS.map(key => [key, 0]))
+
+function empty(id) {
+  return { id, ...emptyBuckets(), byProviderModel: {} }
+}
+
+function add(target, source) {
+  for (const key of FIELDS) target[key] += amount(key === 'apiCost' ? (source.apiCost ?? source.cost) : source[key])
+}
+
+function addSession(target, source) {
+  add(target, source)
+  for (const [key, bucket] of Object.entries(source.byProviderModel ?? {})) {
+    if (bucket === null || typeof bucket !== 'object' || Array.isArray(bucket)) continue
+    // defineProperty 避免来源中的 __proto__ 等键改写普通对象原型。
+    if (!Object.hasOwn(target.byProviderModel, key)) {
+      Object.defineProperty(target.byProviderModel, key, { value: emptyBuckets(), enumerable: true })
+    }
+    add(target.byProviderModel[key], bucket)
+  }
+}
+
+function cleanHeader(value) {
+  if (value === null || typeof value !== 'object' || !validId(value.id)) return null
+  return { id: value.id, ...(value.origin === 'subagent' && validId(value.parentSession)
+    ? { origin: 'subagent', parentSession: value.parentSession } : {}) }
+}
+
+/** 只取宿主公开目录的身份字段;不读取会话正文、工具结果或请求凭据。 */
+export async function readSessionHeaders(ctx) {
+  const headers = new Map()
+  const accept = records => {
+    if (!Array.isArray(records)) return
+    for (const record of records) {
+      const header = cleanHeader(record?.header ?? record)
+      if (header !== null) headers.set(header.id, header)
+    }
+  }
+  const get = name => { try { return ctx?.get?.(name) } catch { return undefined } }
+  let listed = false
+  const query = get('sessionQuery')
+  if (typeof query?.listSessions === 'function') {
+    try {
+      const records = await query.listSessions()
+      if (Array.isArray(records)) { accept(records); listed = true }
+    } catch { /* 单次目录故障不影响主会话费用;尝试旧版持久化接口。 */ }
+  }
+  if (!listed) {
+    const persistence = get('sessionPersistence')
+    if (typeof persistence?.list === 'function') {
+      try { accept(await persistence.list()) } catch { /* 无持久化目录时仍可使用 live 会话。 */ }
+    }
+  }
+  const sessions = get('sessions') ?? ctx?.sessions
+  if (typeof sessions?.list === 'function') {
+    try { accept(sessions.list()) } catch { /* 兼容没有会话目录服务的旧宿主。 */ }
+  }
+  return [...headers.values()]
+}
+
+/** 返回确定属于目标会话的连续子代理后代;普通 fork、断链、自引用与环均不归并。 */
+export function subagentIds(sessionId, headers) {
+  const byId = new Map()
+  for (const raw of Array.isArray(headers) ? headers : []) {
+    const header = cleanHeader(raw?.header ?? raw)
+    if (header !== null) byId.set(header.id, header)
+  }
+  const descendants = new Set()
+  for (const candidate of byId.values()) {
+    if (candidate.id === sessionId || candidate.origin !== 'subagent') continue
+    const seen = new Set([candidate.id])
+    let current = candidate, belongs = false, cyclic = false
+    while (current?.origin === 'subagent' && validId(current.parentSession)) {
+      const parent = current.parentSession
+      if (seen.has(parent)) { cyclic = true; break }
+      seen.add(parent)
+      if (parent === sessionId) belongs = true
+      current = byId.get(parent)
+    }
+    if (belongs && !cyclic) descendants.add(candidate.id)
+  }
+  return descendants
+}
+
+/** own 与 subagents 分开返回,让客户端先选定主会话的完整费用口径后只加一次后代。 */
+export function aggregateSessionCost(days, sessionId, headers = [], includeSubagents = false) {
+  if (!validId(sessionId)) throw new Error('invalid session id')
+  const descendants = includeSubagents ? subagentIds(sessionId, headers) : new Set()
+  const own = empty(sessionId), subagents = empty(sessionId)
+  let found = false
+  const counted = new Set()
+  for (const day of Object.values(days ?? {})) {
+    if (!Array.isArray(day?.sessions)) continue
+    for (const row of day.sessions) {
+      if (row === null || typeof row !== 'object') continue
+      if (row.id === sessionId) { addSession(own, row); found = true }
+      else if (descendants.has(row.id)) { addSession(subagents, row); counted.add(row.id) }
+    }
+  }
+  return { own, subagents, subagentCount: counted.size, found }
+}
+
+/** 按需读取一个会话,避免 getState 广播全部历史会话明细。 */
+export async function getSessionCost(ledger, ctx, sessionId) {
+  if (!validId(sessionId)) throw new Error('invalid session id')
+  const include = ledger.config?.includeSubagentCost === true
+  const headers = include ? await readSessionHeaders(ctx) : []
+  return aggregateSessionCost(ledger.days, sessionId, headers, include)
+}