@goodandready/dsh-cost-meter 0.7.8 → 0.8.0

package/lib/index.js

@@ -1,378 +1,388 @@
-// dsh-cost-meter — host half.
-//
-// The harness prices nothing on purpose: pi-ai carries a `cost` field but
-// zeroes it and no consumer reports spend. So this half owns the rate card and
-// serves it to the browser over GET /dsh-cost-meter/state.
-//
-// Rates come from OpenRouter's public catalog (https://openrouter.ai/api/v1/models,
-// no auth). Besides the four token buckets it also publishes `pricing.overrides`
-// — time-of-day discount windows keyed by UTC HHMM — which is exactly the
-// peak/off-peak schedule DeepSeek runs. So the whole tariff, windows included,
-// is derived rather than typed in; `prices` in config stays as a manual
-// override for routes the catalog does not know (self-hosted, subscriptions).
-//
-// No agent hooks: the widget only reads, so it cannot corrupt a session.
-
-import { mkdirSync, readFileSync, writeFileSync } from 'node:fs'
+// dsh-cost-meter — host half.
+//
+// The harness prices nothing on purpose: pi-ai carries a `cost` field but
+// zeroes it and no consumer reports spend. So this half owns the rate card and
+// serves it to the browser over GET /dsh-cost-meter/state.
+//
+// Rates come from OpenRouter's public catalog (https://openrouter.ai/api/v1/models,
+// no auth). Besides the four token buckets it also publishes `pricing.overrides`
+// — time-of-day discount windows keyed by UTC HHMM — which is exactly the
+// peak/off-peak schedule DeepSeek runs. So the whole tariff, windows included,
+// is derived rather than typed in; `prices` in config stays as a manual
+// override for routes the catalog does not know (self-hosted, subscriptions).
+//
+// No agent hooks: the widget only reads, so it cannot corrupt a session.
+
+import { mkdirSync, readFileSync, writeFileSync } from 'node:fs'
 import { homedir } from 'node:os'
 import { dirname, join } from 'node:path'
-import z from '@deepseek-ai/schemastery'
-import { z as zod } from 'zod'
-
-export const name = 'dsh-cost-meter'
-export const inject = ['webServer']
-
-const NS = 'dsh-cost-meter'
-const STATE_PATH = '/dsh-cost-meter/state'
-const CATALOG_URL = 'https://openrouter.ai/api/v1/models'
+import z from '@deepseek-ai/schemastery'
+import { z as zod } from 'zod'
+
+export const name = 'dsh-cost-meter'
+export const inject = ['webServer']
+
+const NS = 'dsh-cost-meter'
+const STATE_PATH = '/dsh-cost-meter/state'
+const CATALOG_URL = 'https://openrouter.ai/api/v1/models'
 const CACHE_FILE = join(process.env.DSH_HOME || join(homedir(), '.dsh'), 'storages', 'dsh-cost-meter-catalog.json')
-
-/** One route's rates, in USD per 1M tokens. */
-const PriceRow = z.object({
-  cacheHit: z.number().description('Input tokens served from the provider cache.').default(0),
-  input: z.number().description('Input tokens that missed the cache.').default(0),
-  cacheWrite: z.number().description('Input tokens written into the cache.').default(0),
-  output: z.number().description('Output tokens (reasoning included).').default(0),
-})
-
-/**
- * DeepSeek's own tariff, from https://api-docs.deepseek.com/quick_start/pricing.
- * Their API publishes no prices (/models lists ids, there is no pricing route),
- * and OpenRouter's `deepseek/*` entries quote different numbers, so this table
- * is the authoritative source for the official route. Values are PEAK, USD per
- * 1M tokens; off-peak is exactly half. Peak: 01:00-04:00 and 06:00-10:00 UTC.
- */
-const DEEPSEEK_PEAK_RATES = {
-  'deepseek-v4-flash': { cacheHit: 0.014, input: 0.44, cacheWrite: 0.44, output: 1.32 },
-  'deepseek-v4-pro': { cacheHit: 0.044, input: 1.32, cacheWrite: 1.32, output: 3.96 },
-}
-const DEEPSEEK_PEAK_WINDOWS = [{ start: 60, end: 240 }, { start: 360, end: 600 }]
-
-// ponytail: a DeepSeek model reached THROUGH OpenRouter is still priced from
-// this table, though the bill comes from OpenRouter. Fix by classifying the
-// route (provider id / baseURL) before the model, and for a `:floor` slug take
-// the minimum over /models/{id}/endpoints rather than the base listing.
-
-export const Config = z.object({
-  currency: z.string().description('Currency symbol shown in the widget.').default('$'),
-  usdRate: z.number().description('Units of `currency` per USD. 1 keeps everything in USD.').default(1),
-  displayTimeZone: z
-    .string()
-    .description('IANA zone the widget shows tariff hours in.')
-    .default('Europe/Moscow'),
-  deepseekPeakPrices: z
-    .dict(PriceRow)
-    .description('Override the built-in DeepSeek peak rates, by model id.')
-    .default({}),
-  useOpenRouter: z
-    .boolean()
-    .description('Pull rates and discount windows from the public OpenRouter catalog.')
-    .default(true),
-  refreshHours: z.number().description('How often to refetch the catalog.').default(24),
-  modelMap: z
-    .dict(z.string())
-    .description('Route override: "provider/model" -> OpenRouter model id.')
-    .default({}),
-  prices: z
-    .dict(PriceRow)
-    .description('Manual rates by "provider/model" (or "*"). Win over the catalog.')
-    .default({}),
-  manualPeakWindowsUtc: z
-    .array(z.string())
-    .description('Peak windows HH:MM-HH:MM UTC, used only with manual prices.')
-    .default([]),
-  manualOffPeakMultiplier: z
-    .number()
-    .description('Multiplier outside the manual peak windows.')
-    .default(1),
-})
-
-/** OpenRouter prices are USD per token; the widget speaks per 1M tokens. */
-function perMillion(value) {
-  const n = Number(value)
-  return Number.isFinite(n) ? n * 1e6 : 0
-}
-
-/** Pull the four buckets out of one OpenRouter pricing object. */
-function bucketsOf(pricing) {
-  return {
-    cacheHit: perMillion(pricing.input_cache_read),
-    input: perMillion(pricing.prompt),
-    cacheWrite: perMillion(pricing.input_cache_write),
-    output: perMillion(pricing.completion),
-  }
-}
-
-/** OpenRouter encodes window edges as UTC HHMM integers (1000 = 10:00). */
-export function hhmmToMinutes(value) {
-  const n = Number(value)
-  if (!Number.isFinite(n) || n < 0 || n > 2400) return undefined
-  return Math.floor(n / 100) * 60 + (n % 100)
-}
-
-/** "HH:MM-HH:MM" -> {start,end} in minutes; undefined when malformed. */
-export function parseWindow(spec) {
-  const m = /^(\d{1,2}):(\d{2})-(\d{1,2}):(\d{2})$/.exec(String(spec).trim())
-  if (!m) return undefined
-  return { start: Number(m[1]) * 60 + Number(m[2]), end: Number(m[3]) * 60 + Number(m[4]) }
-}
-
-/**
- * Normalize one catalog entry into the schedule the browser consumes:
- * a base rate card plus the discount windows that shadow it.
- */
-function scheduleOf(entry) {
-  const pricing = entry.pricing || {}
-  const base = bucketsOf(pricing)
-  const windows = []
-  for (const override of pricing.overrides || []) {
-    const start = hhmmToMinutes(override.utc_start)
-    const end = hhmmToMinutes(override.utc_end)
-    if (start === undefined || end === undefined) continue
-    // An override omits buckets it does not change; fall back to the base.
-    const rates = bucketsOf(override)
-    windows.push({
-      start,
-      end,
-      rates: {
-        cacheHit: rates.cacheHit || base.cacheHit,
-        input: rates.input || base.input,
-        cacheWrite: rates.cacheWrite || base.cacheWrite,
-        output: rates.output || base.output,
-      },
-    })
-  }
-  return { base, windows }
-}
-
-/**
- * Map a harness route onto an OpenRouter id. Exact id first, then the vendor
- * suffix (`deepseek-official/deepseek-v4-flash` -> `deepseek/deepseek-v4-flash`),
- * which is what the shared model names make possible.
- */
-export function matchModel(index, route, modelMap) {
-  if (!route) return undefined
-  const key = route.provider + '/' + route.model
-  const mapped = modelMap && modelMap[key]
-  if (mapped && index.byId[mapped]) return mapped
-  if (index.byId[key]) return key
-  if (index.byId[route.model]) return route.model
-  // Harness routes name models in their own casing (`MiniMax-M3`), the catalog
-  // is lower-case (`minimax/minimax-m3`), so matching is case-insensitive.
-  const lower = String(route.model).toLowerCase()
-  if (index.byLowerId[lower]) return index.byLowerId[lower]
-  if (index.byLowerId[key.toLowerCase()]) return index.byLowerId[key.toLowerCase()]
-  const bySuffix = index.bySuffix[lower]
-  return bySuffix && bySuffix.length ? bySuffix[0] : undefined
-}
-
-/** Build lookup tables once per fetch; the raw catalog is ~700 KB. */
+
+/** One route's rates, in USD per 1M tokens. */
+const PriceRow = z.object({
+  cacheHit: z.number().description('Input tokens served from the provider cache.').default(0),
+  input: z.number().description('Input tokens that missed the cache.').default(0),
+  cacheWrite: z.number().description('Input tokens written into the cache.').default(0),
+  output: z.number().description('Output tokens (reasoning included).').default(0),
+})
+
+/**
+ * DeepSeek's own tariff, from https://api-docs.deepseek.com/quick_start/pricing.
+ * Their API publishes no prices (/models lists ids, there is no pricing route),
+ * and OpenRouter's `deepseek/*` entries quote different numbers, so this table
+ * is the authoritative source for the official route. Values are PEAK, USD per
+ * 1M tokens; off-peak is exactly half. Peak: 01:00-04:00 and 06:00-10:00 UTC.
+ */
+const DEEPSEEK_PEAK_RATES = {
+  'deepseek-v4-flash': { cacheHit: 0.014, input: 0.44, cacheWrite: 0.44, output: 1.32 },
+  'deepseek-v4-pro': { cacheHit: 0.044, input: 1.32, cacheWrite: 1.32, output: 3.96 },
+}
+const DEEPSEEK_PEAK_WINDOWS = [{ start: 60, end: 240 }, { start: 360, end: 600 }]
+
+// ponytail: a DeepSeek model reached THROUGH OpenRouter is still priced from
+// this table, though the bill comes from OpenRouter. Fix by classifying the
+// route (provider id / baseURL) before the model, and for a `:floor` slug take
+// the minimum over /models/{id}/endpoints rather than the base listing.
+
+export const Config = z.object({
+  currency: z.string().description('Currency symbol shown in the widget.').default('$'),
+  usdRate: z.number().description('Units of `currency` per USD. 1 keeps everything in USD.').default(1),
+  displayTimeZone: z
+    .string()
+    .description('IANA zone the widget shows tariff hours in.')
+    .default('Europe/Moscow'),
+  deepseekPeakPrices: z
+    .dict(PriceRow)
+    .description('Override the built-in DeepSeek peak rates, by model id.')
+    .default({}),
+  useOpenRouter: z
+    .boolean()
+    .description('Pull rates and discount windows from the public OpenRouter catalog.')
+    .default(true),
+  refreshHours: z.number().description('How often to refetch the catalog.').default(24),
+  modelMap: z
+    .dict(z.string())
+    .description('Route override: "provider/model" -> OpenRouter model id.')
+    .default({}),
+  prices: z
+    .dict(PriceRow)
+    .description('Manual rates by "provider/model" (or "*"). Win over the catalog.')
+    .default({}),
+  manualPeakWindowsUtc: z
+    .array(z.string())
+    .description('Peak windows HH:MM-HH:MM UTC, used only with manual prices.')
+    .default([]),
+  manualOffPeakMultiplier: z
+    .number()
+    .description('Multiplier outside the manual peak windows.')
+    .default(1),
+})
+
+/** OpenRouter prices are USD per token; the widget speaks per 1M tokens. */
+function perMillion(value) {
+  const n = Number(value)
+  return Number.isFinite(n) ? n * 1e6 : 0
+}
+
+/** Pull the four buckets out of one OpenRouter pricing object. */
+function bucketsOf(pricing) {
+  return {
+    cacheHit: perMillion(pricing.input_cache_read),
+    input: perMillion(pricing.prompt),
+    cacheWrite: perMillion(pricing.input_cache_write),
+    output: perMillion(pricing.completion),
+  }
+}
+
+/** OpenRouter encodes window edges as UTC HHMM integers (1000 = 10:00). */
+export function hhmmToMinutes(value) {
+  const n = Number(value)
+  if (!Number.isFinite(n) || n < 0 || n > 2400) return undefined
+  return Math.floor(n / 100) * 60 + (n % 100)
+}
+
+/** "HH:MM-HH:MM" -> {start,end} in minutes; undefined when malformed. */
+export function parseWindow(spec) {
+  const m = /^(\d{1,2}):(\d{2})-(\d{1,2}):(\d{2})$/.exec(String(spec).trim())
+  if (!m) return undefined
+  return { start: Number(m[1]) * 60 + Number(m[2]), end: Number(m[3]) * 60 + Number(m[4]) }
+}
+
+/**
+ * Normalize one catalog entry into the schedule the browser consumes:
+ * a base rate card plus the discount windows that shadow it.
+ */
+function scheduleOf(entry) {
+  const pricing = entry.pricing || {}
+  const base = bucketsOf(pricing)
+  const windows = []
+  for (const override of pricing.overrides || []) {
+    const start = hhmmToMinutes(override.utc_start)
+    const end = hhmmToMinutes(override.utc_end)
+    if (start === undefined || end === undefined) continue
+    // An override omits buckets it does not change; fall back to the base.
+    const rates = bucketsOf(override)
+    windows.push({
+      start,
+      end,
+      rates: {
+        cacheHit: rates.cacheHit || base.cacheHit,
+        input: rates.input || base.input,
+        cacheWrite: rates.cacheWrite || base.cacheWrite,
+        output: rates.output || base.output,
+      },
+    })
+  }
+  return { base, windows }
+}
+
+/**
+ * Map a harness route onto an OpenRouter id. Exact id first, then the vendor
+ * suffix (`deepseek-official/deepseek-v4-flash` -> `deepseek/deepseek-v4-flash`),
+ * which is what the shared model names make possible.
+ */
+export function matchModel(index, route, modelMap) {
+  if (!route || !index) return undefined
+  const key = route.provider ? route.provider + '/' + route.model : route.model
+  const mapped = modelMap && (modelMap[key] || modelMap[route.model])
+  if (mapped && index.byId && index.byId[mapped]) return mapped
+  if (index.byId && index.byId[key]) return key
+  if (index.byId && index.byId[route.model]) return route.model
+  // Harness routes name models in their own casing (`MiniMax-M3`), the catalog
+  // is lower-case (`minimax/minimax-m3`), so matching is case-insensitive.
+  const lower = String(route.model).toLowerCase()
+  const byLowerId = index.byLowerId || {}
+  if (byLowerId[lower]) return byLowerId[lower]
+  if (byLowerId[key.toLowerCase()]) return byLowerId[key.toLowerCase()]
+  const bySuffix = index.bySuffix || {}
+  const list = bySuffix[lower]
+  return list && list.length ? list[0] : undefined
+}
+
+/** Build lookup tables once per fetch; the raw catalog is ~700 KB. */
 export function indexCatalog(models) {
-  const byId = {}
-  const byLowerId = {}
-  const bySuffix = {}
-  for (const entry of models) {
-    if (!entry || typeof entry.id !== 'string' || !entry.pricing) continue
-    // Variants like ":batch" or ":free" are different products; skip them so a
-    // suffix match cannot silently price against a cheaper batch tier.
-    if (entry.id.includes(':')) continue
-    byId[entry.id] = scheduleOf(entry)
-    byLowerId[entry.id.toLowerCase()] = entry.id
-    const suffix = entry.id.includes('/') ? entry.id.slice(entry.id.indexOf('/') + 1) : entry.id
-    ;(bySuffix[suffix.toLowerCase()] = bySuffix[suffix.toLowerCase()] || []).push(entry.id)
-  }
-  return { byId, byLowerId, bySuffix }
-}
-
-// ---------------------------------------------------------------------------
-// Per-model usage projection
-//
-// `tokenUsage` from dsh-token-meter totals the whole session, which prices
-// wrongly the moment the model changes mid-session. The log has what is
-// needed: `request/header` carries `header.config = {provider, model}`, and
-// usage arrives afterwards for that same step. So this unit keeps the last
-// header's route and files every usage sample under it.
-//
-// The replace-if-same-step guard mirrors token-meter: a streamed usage chunk
-// and the finalized assistant message report the same numbers for one step.
-// ---------------------------------------------------------------------------
-
-const ZERO = { uncachedInputTokens: 0, outputTokens: 0, cacheReadTokens: 0, cacheWriteTokens: 0 }
-
-// Projection schemas are validated with ZOD, not schemastery: the registry
-// calls `schema.parse(view)` at the boundary, and a schemastery schema has no
-// `.parse`, which fails every history read with
-// "registration.def.schema.parse is not a function".
-const bucketSchema = zod.object({
-  uncachedInputTokens: zod.number().nonnegative(),
-  outputTokens: zod.number().nonnegative(),
-  cacheReadTokens: zod.number().nonnegative(),
-  cacheWriteTokens: zod.number().nonnegative(),
-}).strict()
-
-/** Pull the four buckets out of one provider usage record. */
-function bucketsFromUsage(usage) {
-  return {
-    uncachedInputTokens: Number(usage.inputTokens) || 0,
-    outputTokens: Number(usage.outputTokens) || 0,
-    cacheReadTokens: Number(usage.cacheReadTokens) || 0,
-    cacheWriteTokens: Number(usage.cacheWriteTokens) || 0,
-  }
-}
-
-function addBuckets(base, remove, add) {
-  const out = {}
-  for (const k of Object.keys(ZERO)) {
-    out[k] = (base[k] || 0) - ((remove && remove[k]) || 0) + (add[k] || 0)
-  }
-  return out
-}
-
-function sameBuckets(a, b) {
-  return a && b && Object.keys(ZERO).every((k) => (a[k] || 0) === (b[k] || 0))
-}
-
-/** The usage a streamed chunk or a finalized assistant message reports. */
-function usageOf(event) {
-  if (event.type === 'assistant/chunk' && event.data.chunk && event.data.chunk.type === 'usage') {
-    return { turn: event.data.turn, step: event.data.step, usage: event.data.chunk.usage }
-  }
-  if (event.type === 'assistant/message' && event.data.usage !== undefined) {
-    return { turn: event.data.turn, step: event.data.step, usage: event.data.usage }
-  }
-  return undefined
-}
-
-/**
- * Half-hour slot of the UTC day an event falls into (0..47). Tariff windows are
- * expressed in UTC clock time, and every real schedule so far lands on a whole
- * or half hour, so 48 slots reproduce the rate exactly while keeping the state
- * small enough to ship on every projection frame.
- */
-export function slotOf(timeMs) {
-  const d = new Date(timeMs)
-  return Math.floor((d.getUTCHours() * 60 + d.getUTCMinutes()) / 30)
-}
-
-const usageByModelProjection = {
-  key: 'costByModel',
-  schema: zod.object({
-    models: zod.array(zod.object({
-      provider: zod.string(),
-      model: zod.string(),
-      // Usage split by the half-hour of the UTC day it was spent in. Spend is
-      // priced at the rate in force THEN; a later tariff switch must not
-      // reprice tokens that are already paid for.
-      slots: zod.record(zod.string(), bucketSchema),
-    })),
-  }).strict(),
-  init: () => ({ route: null, byKey: {}, last: null }),
-  apply: (state, event) => {
-    if (event.type === 'request/header') {
-      const config = event.data && event.data.header && event.data.header.config
-      const provider = config && typeof config.provider === 'string' ? config.provider : undefined
-      const model = config && typeof config.model === 'string' ? config.model : undefined
-      if (provider === undefined || model === undefined) return state
-      if (state.route && state.route.provider === provider && state.route.model === model) return state
-      return { ...state, route: { provider, model } }
-    }
-    const sample = usageOf(event)
-    if (sample === undefined) return state
-    const route = state.route || { provider: '', model: 'unknown' }
-    const key = route.provider + '/' + route.model
-    const slot = String(slotOf(event.time))
-    const buckets = bucketsFromUsage(sample.usage)
-    // A streamed usage chunk and the finalized assistant message report the
-    // same step twice; the second one replaces the first instead of adding.
-    const previous =
-      state.last !== null &&
-      state.last.turn === sample.turn &&
-      state.last.step === sample.step &&
-      state.last.key === key &&
-      state.last.slot === slot
-        ? state.last.buckets
-        : undefined
-    if (previous !== undefined && sameBuckets(previous, buckets)) return state
-    const slots = state.byKey[key] || {}
-    return {
-      route: state.route,
-      byKey: {
-        ...state.byKey,
-        [key]: { ...slots, [slot]: addBuckets(slots[slot] || ZERO, previous, buckets) },
-      },
-      last: { turn: sample.turn, step: sample.step, key, slot, buckets },
-    }
-  },
-  view: (state) => ({
-    models: Object.keys(state.byKey).map((key) => {
-      const slash = key.indexOf('/')
-      return {
-        provider: key.slice(0, slash),
-        model: key.slice(slash + 1),
-        slots: state.byKey[key],
-      }
-    }),
-  }),
-  stateVersion: 2,
-}
-
-/** The agent's default route, read off settings descriptors when available. */
-function defaultRoute(ctx) {
-  const settings = ctx.get('settings')
-  if (settings === undefined) return undefined
-  try {
-    const found = settings
-      .describe({ redactSecrets: true })
-      .find((descriptor) => descriptor.ns === 'agent-default-model')
-    const value = found && found.value
-    if (!value || typeof value.provider !== 'string' || typeof value.model !== 'string') return undefined
-    return { provider: value.provider, model: value.model }
-  } catch {
-    return undefined
-  }
-}
-
-/**
- * Resolve one route's tariff: manual config wins, then DeepSeek's own table,
- * then the OpenRouter catalog. Returns `schedule: null` when nothing knows it,
- * so the widget can say so instead of pricing at zero.
- */
+  const byId = {}
+  const byLowerId = {}
+  const bySuffix = {}
+  for (const entry of models) {
+    if (!entry || typeof entry.id !== 'string' || !entry.pricing) continue
+    // Variants like ":batch" or ":free" are different products; skip them so a
+    // suffix match cannot silently price against a cheaper batch tier.
+    if (entry.id.includes(':')) continue
+    byId[entry.id] = scheduleOf(entry)
+    byLowerId[entry.id.toLowerCase()] = entry.id
+    const suffix = entry.id.includes('/') ? entry.id.slice(entry.id.indexOf('/') + 1) : entry.id
+    ;(bySuffix[suffix.toLowerCase()] = bySuffix[suffix.toLowerCase()] || []).push(entry.id)
+  }
+  return { byId, byLowerId, bySuffix }
+}
+
+// ---------------------------------------------------------------------------
+// Per-model usage projection
+//
+// `tokenUsage` from dsh-token-meter totals the whole session, which prices
+// wrongly the moment the model changes mid-session. The log has what is
+// needed: `request/header` carries `header.config = {provider, model}`, and
+// usage arrives afterwards for that same step. So this unit keeps the last
+// header's route and files every usage sample under it.
+//
+// The replace-if-same-step guard mirrors token-meter: a streamed usage chunk
+// and the finalized assistant message report the same numbers for one step.
+// ---------------------------------------------------------------------------
+
+const ZERO = { uncachedInputTokens: 0, outputTokens: 0, cacheReadTokens: 0, cacheWriteTokens: 0 }
+
+// Projection schemas are validated with ZOD, not schemastery: the registry
+// calls `schema.parse(view)` at the boundary, and a schemastery schema has no
+// `.parse`, which fails every history read with
+// "registration.def.schema.parse is not a function".
+const bucketSchema = zod.object({
+  uncachedInputTokens: zod.number().nonnegative(),
+  outputTokens: zod.number().nonnegative(),
+  cacheReadTokens: zod.number().nonnegative(),
+  cacheWriteTokens: zod.number().nonnegative(),
+}).strict()
+
+/** Pull the four buckets out of one provider usage record. */
+function bucketsFromUsage(usage) {
+  return {
+    uncachedInputTokens: Number(usage.inputTokens) || 0,
+    outputTokens: Number(usage.outputTokens) || 0,
+    cacheReadTokens: Number(usage.cacheReadTokens) || 0,
+    cacheWriteTokens: Number(usage.cacheWriteTokens) || 0,
+  }
+}
+
+function addBuckets(base, remove, add) {
+  const out = {}
+  for (const k of Object.keys(ZERO)) {
+    out[k] = (base[k] || 0) - ((remove && remove[k]) || 0) + (add[k] || 0)
+  }
+  return out
+}
+
+function sameBuckets(a, b) {
+  return a && b && Object.keys(ZERO).every((k) => (a[k] || 0) === (b[k] || 0))
+}
+
+/** The usage a streamed chunk or a finalized assistant message reports. */
+function usageOf(event) {
+  if (event.type === 'assistant/chunk' && event.data.chunk && event.data.chunk.type === 'usage') {
+    return { turn: event.data.turn, step: event.data.step, usage: event.data.chunk.usage }
+  }
+  if (event.type === 'assistant/message' && event.data.usage !== undefined) {
+    return { turn: event.data.turn, step: event.data.step, usage: event.data.usage }
+  }
+  return undefined
+}
+
+/**
+ * Half-hour slot of the UTC day an event falls into (0..47). Tariff windows are
+ * expressed in UTC clock time, and every real schedule so far lands on a whole
+ * or half hour, so 48 slots reproduce the rate exactly while keeping the state
+ * small enough to ship on every projection frame.
+ */
+export function slotOf(timeMs) {
+  let ms = typeof timeMs === 'number' && Number.isFinite(timeMs) ? timeMs : NaN
+  if (Number.isNaN(ms)) {
+    if (timeMs instanceof Date && !Number.isNaN(timeMs.getTime())) {
+      ms = timeMs.getTime()
+    } else {
+      ms = Date.now()
+    }
+  }
+  const d = new Date(ms)
+  return Math.floor((d.getUTCHours() * 60 + d.getUTCMinutes()) / 30)
+}
+
+const usageByModelProjection = {
+  key: 'costByModel',
+  schema: zod.object({
+    models: zod.array(zod.object({
+      provider: zod.string(),
+      model: zod.string(),
+      // Usage split by the half-hour of the UTC day it was spent in. Spend is
+      // priced at the rate in force THEN; a later tariff switch must not
+      // reprice tokens that are already paid for.
+      slots: zod.record(zod.string(), bucketSchema),
+    })),
+  }).strict(),
+  init: () => ({ route: null, byKey: {}, last: null }),
+  apply: (state, event) => {
+    if (event.type === 'request/header') {
+      const config = event.data && event.data.header && event.data.header.config
+      const provider = config && typeof config.provider === 'string' ? config.provider : undefined
+      const model = config && typeof config.model === 'string' ? config.model : undefined
+      if (provider === undefined || model === undefined) return state
+      if (state.route && state.route.provider === provider && state.route.model === model) return state
+      return { ...state, route: { provider, model } }
+    }
+    const sample = usageOf(event)
+    if (sample === undefined) return state
+    const route = state.route || { provider: '', model: 'unknown' }
+    const key = route.provider + '/' + route.model
+    const slot = String(slotOf(event.time))
+    const buckets = bucketsFromUsage(sample.usage)
+    // A streamed usage chunk and the finalized assistant message report the
+    // same step twice; the second one replaces the first instead of adding.
+    const previous =
+      state.last !== null &&
+      state.last.turn === sample.turn &&
+      state.last.step === sample.step &&
+      state.last.key === key &&
+      state.last.slot === slot
+        ? state.last.buckets
+        : undefined
+    if (previous !== undefined && sameBuckets(previous, buckets)) return state
+    const slots = state.byKey[key] || {}
+    return {
+      route: state.route,
+      byKey: {
+        ...state.byKey,
+        [key]: { ...slots, [slot]: addBuckets(slots[slot] || ZERO, previous, buckets) },
+      },
+      last: { turn: sample.turn, step: sample.step, key, slot, buckets },
+    }
+  },
+  view: (state) => ({
+    models: Object.keys(state.byKey).map((key) => {
+      const slash = key.indexOf('/')
+      return {
+        provider: key.slice(0, slash),
+        model: key.slice(slash + 1),
+        slots: state.byKey[key],
+      }
+    }),
+  }),
+  stateVersion: 2,
+}
+
+/** The agent's default route, read off settings descriptors when available. */
+function defaultRoute(ctx) {
+  const settings = (ctx?.get && typeof ctx.get === 'function' ? ctx.get('settings') : null) || ctx?.settings
+  if (settings === undefined) return undefined
+  try {
+    const found = settings
+      .describe({ redactSecrets: true })
+      .find((descriptor) => descriptor.ns === 'agent-default-model')
+    const value = found && found.value
+    if (!value || typeof value.provider !== 'string' || typeof value.model !== 'string') return undefined
+    return { provider: value.provider, model: value.model }
+  } catch {
+    return undefined
+  }
+}
+
+/**
+ * Resolve one route's tariff: manual config wins, then DeepSeek's own table,
+ * then the OpenRouter catalog. Returns `schedule: null` when nothing knows it,
+ * so the widget can say so instead of pricing at zero.
+ */
 export function resolveTariff(cfg, index, route) {
-  if (!route) return { schedule: null, source: 'none', matchedId: null }
-  const key = route.provider + '/' + route.model
-  const prices = cfg.prices || {}
-  const manual = prices[key] || prices[route.model] || prices['*']
-
-  if (manual) {
-    const base = {
-      cacheHit: Number(manual.cacheHit) || 0,
-      input: Number(manual.input) || 0,
-      cacheWrite: Number(manual.cacheWrite) || 0,
-      output: Number(manual.output) || 0,
-    }
-    const factor = Number(cfg.manualOffPeakMultiplier) || 1
-    const windows = []
-    for (const spec of cfg.manualPeakWindowsUtc || []) {
-      const parsed = parseWindow(spec)
-      if (parsed) windows.push({ start: parsed.start, end: parsed.end, rates: base })
-    }
-    const schedule = windows.length
-      ? {
-          base: {
-            cacheHit: base.cacheHit * factor,
-            input: base.input * factor,
-            cacheWrite: base.cacheWrite * factor,
-            output: base.output * factor,
-          },
-          windows,
-        }
-      : { base, windows: [] }
-    return { schedule, source: 'manual', matchedId: key }
-  }
-
-  // Keyed by MODEL, not provider: a DeepSeek model is billed at DeepSeek's
-  // published rate whoever routes it (deepseek-official, opencode-go, ollama).
-  // Date/variant suffixes (`deepseek-v4-flash:0731`) name the same product.
+  if (!route) return { schedule: null, source: 'none', matchedId: null }
+  const key = route.provider + '/' + route.model
+  const prices = cfg.prices || {}
+  const manual = prices[key] || prices[route.model] || prices['*']
+
+  if (manual) {
+    const base = {
+      cacheHit: Number(manual.cacheHit) || 0,
+      input: Number(manual.input) || 0,
+      cacheWrite: Number(manual.cacheWrite) || 0,
+      output: Number(manual.output) || 0,
+    }
+    const factor = Number(cfg.manualOffPeakMultiplier) || 1
+    const windows = []
+    for (const spec of cfg.manualPeakWindowsUtc || []) {
+      const parsed = parseWindow(spec)
+      if (parsed) windows.push({ start: parsed.start, end: parsed.end, rates: base })
+    }
+    const schedule = windows.length
+      ? {
+          base: {
+            cacheHit: base.cacheHit * factor,
+            input: base.input * factor,
+            cacheWrite: base.cacheWrite * factor,
+            output: base.output * factor,
+          },
+          windows,
+        }
+      : { base, windows: [] }
+    return { schedule, source: 'manual', matchedId: key }
+  }
+
+  // Keyed by MODEL, not provider: a DeepSeek model is billed at DeepSeek's
+  // published rate whoever routes it (deepseek-official, opencode-go, ollama).
+  // Date/variant suffixes (`deepseek-v4-flash:0731`) name the same product.
   const providerLower = String(route.provider || "").toLowerCase();
   // #3: Provider-aware routing: if routed explicitly through OpenRouter,
   // consult OpenRouter catalog first to honor actual provider billing.
@@ -381,132 +391,136 @@ export function resolveTariff(cfg, index, route) {
     if (matchedId) return { schedule: index.byId[matchedId], source: "openrouter", matchedId };
   }
 
-  const base = String(route.model).toLowerCase().split(':')[0]
-  const peak = (cfg.deepseekPeakPrices || {})[base] || DEEPSEEK_PEAK_RATES[base]
-  if (peak) {
-    const full = {
-      cacheHit: Number(peak.cacheHit) || 0,
-      input: Number(peak.input) || 0,
-      cacheWrite: Number(peak.cacheWrite) || 0,
-      output: Number(peak.output) || 0,
-    }
-    return {
-      schedule: {
-        base: {
-          cacheHit: full.cacheHit / 2,
-          input: full.input / 2,
-          cacheWrite: full.cacheWrite / 2,
-          output: full.output / 2,
-        },
-        windows: DEEPSEEK_PEAK_WINDOWS.map((w) => ({ start: w.start, end: w.end, rates: full })),
-      },
-      source: 'deepseek',
-      matchedId: base,
-    }
-  }
-
-  if (cfg.useOpenRouter !== false) {
-    const matchedId = matchModel(index, route, cfg.modelMap)
-    if (matchedId) return { schedule: index.byId[matchedId], source: 'openrouter', matchedId }
-  }
-  return { schedule: null, source: 'none', matchedId: null }
-}
-
-/** "provider/model" -> {provider, model}; the model half may contain slashes. */
-function splitRoute(key) {
-  const slash = String(key).indexOf('/')
-  if (slash < 0) return undefined
-  return { provider: key.slice(0, slash), model: key.slice(slash + 1) }
-}
-
-export function apply(ctx, config) {
-  let current = config
-  let catalog = { index: { byId: {}, bySuffix: {} }, fetchedAt: 0, error: null }
-
-  // Disk cache survives restarts, so a boot without network still prices.
-  try {
-    const cached = JSON.parse(readFileSync(CACHE_FILE, 'utf8'))
-    if (cached && Array.isArray(cached.models)) {
-      catalog = { index: indexCatalog(cached.models), fetchedAt: cached.fetchedAt || 0, error: null }
-    }
-  } catch {
-    // absent or unreadable cache is normal on first run
-  }
-
-  async function refresh(signal) {
-    const cfg = current ?? config
-    if (cfg.useOpenRouter === false) return
-    const ageMs = Date.now() - catalog.fetchedAt
-    if (catalog.fetchedAt && ageMs < (Number(cfg.refreshHours) || 24) * 3600e3) return
-    try {
-      const response = await fetch(CATALOG_URL, { signal, headers: { accept: 'application/json' } })
-      if (!response.ok) throw new Error('HTTP ' + response.status)
-      const body = await response.json()
-      const models = Array.isArray(body && body.data) ? body.data : []
-      if (models.length === 0) throw new Error('empty catalog')
-      // Keep only what pricing needs — the raw payload is mostly prose.
-      const slim = models.map((m) => ({ id: m.id, pricing: m.pricing }))
-      catalog = { index: indexCatalog(slim), fetchedAt: Date.now(), error: null }
-      try {
-        mkdirSync(dirname(CACHE_FILE), { recursive: true })
-        writeFileSync(CACHE_FILE, JSON.stringify({ fetchedAt: catalog.fetchedAt, models: slim }))
-      } catch {
-        // a read-only storages dir only costs us the restart-time cache
-      }
-    } catch (error) {
-      catalog = { ...catalog, error: String((error && error.message) || error) }
-    }
-  }
-
-  // Live config from the GUI, with the composition entry as the base layer.
-  ctx.inject(['settings'], (sctx) => {
-    const scope = sctx.settings.register(NS, Config, { base: config })
-    current = scope.get() ?? config
-    const stop = scope.watch(() => {
-      current = scope.get() ?? config
-      void refresh()
-    })
-    sctx.effect(() => () => {
-      stop()
-      current = config
-    })
-  })
-
-  // Optional capability: headless assemblies without the registry keep working,
-  // the widget just falls back to the session-wide tokenUsage projection.
-  ctx.inject(['sessionProjections'], (pctx) => {
-    pctx.sessionProjections.register(usageByModelProjection)
-  })
-
-  ctx.effect(() => {
-    const controller = new AbortController()
-    void refresh(controller.signal)
-    const timer = setInterval(() => void refresh(controller.signal), 3600e3)
-    return () => {
-      clearInterval(timer)
-      controller.abort()
-    }
-  }, 'dsh-cost-meter: catalog refresh')
-
-  ctx.effect(
-    () =>
-      ctx.webServer.register({
-        kind: 'exact',
-        path: STATE_PATH,
-        handler: (req, res) => {
-          if (req.method !== 'GET') {
-            res.writeHead(405, { 'Content-Type': 'application/json' })
-            res.end(JSON.stringify({ error: 'method not allowed' }))
-            return
-          }
-          const cfg = current ?? config
-          const route = defaultRoute(ctx) ?? null
-          const own = resolveTariff(cfg, catalog.index, route)
-
-          // The client asks for every model the session actually used, so a
-          // session that switched models can price each part on its own tariff.
-          const schedules = {}
-          const asked = new URL(req.url, 'http://localhost').searchParams.get('models')
+  const base = String(route.model).toLowerCase().split(':')[0]
+  const peak = (cfg.deepseekPeakPrices || {})[base] || DEEPSEEK_PEAK_RATES[base]
+  if (peak) {
+    const full = {
+      cacheHit: Number(peak.cacheHit) || 0,
+      input: Number(peak.input) || 0,
+      cacheWrite: Number(peak.cacheWrite) || 0,
+      output: Number(peak.output) || 0,
+    }
+    return {
+      schedule: {
+        base: {
+          cacheHit: full.cacheHit / 2,
+          input: full.input / 2,
+          cacheWrite: full.cacheWrite / 2,
+          output: full.output / 2,
+        },
+        windows: DEEPSEEK_PEAK_WINDOWS.map((w) => ({ start: w.start, end: w.end, rates: full })),
+      },
+      source: 'deepseek',
+      matchedId: base,
+    }
+  }
+
+  if (cfg.useOpenRouter !== false) {
+    const matchedId = matchModel(index, route, cfg.modelMap)
+    if (matchedId) return { schedule: index.byId[matchedId], source: 'openrouter', matchedId }
+  }
+  return { schedule: null, source: 'none', matchedId: null }
+}
+
+/** "provider/model" -> {provider, model}; the model half may contain slashes. */
+export function splitRoute(key) {
+  const str = String(key || '').trim()
+  if (!str) return undefined
+  const slash = str.indexOf('/')
+  if (slash < 0) return { provider: '', model: str }
+  return { provider: str.slice(0, slash), model: str.slice(slash + 1) }
+}
+
+export function apply(ctx, config) {
+  let current = config
+  let catalog = { index: { byId: {}, byLowerId: {}, bySuffix: {} }, fetchedAt: 0, error: null }
+
+  // Disk cache survives restarts, so a boot without network still prices.
+  try {
+    const cached = JSON.parse(readFileSync(CACHE_FILE, 'utf8'))
+    if (cached && Array.isArray(cached.models)) {
+      catalog = { index: indexCatalog(cached.models), fetchedAt: cached.fetchedAt || 0, error: null }
+    }
+  } catch {
+    // absent or unreadable cache is normal on first run
+  }
+
+  async function refresh(signal) {
+    const cfg = current ?? config
+    if (cfg.useOpenRouter === false) return
+    const ageMs = Date.now() - catalog.fetchedAt
+    if (catalog.fetchedAt && ageMs < (Number(cfg.refreshHours) || 24) * 3600e3) return
+    try {
+      const timeoutSignal = AbortSignal.timeout(15000)
+      const combinedSignal = signal ? (AbortSignal.any ? AbortSignal.any([signal, timeoutSignal]) : timeoutSignal) : timeoutSignal
+      const response = await fetch(CATALOG_URL, { signal: combinedSignal, headers: { accept: 'application/json' } })
+      if (!response.ok) throw new Error('HTTP ' + response.status)
+      const body = await response.json()
+      const models = Array.isArray(body && body.data) ? body.data : []
+      if (models.length === 0) throw new Error('empty catalog')
+      // Keep only what pricing needs — the raw payload is mostly prose.
+      const slim = models.map((m) => ({ id: m.id, pricing: m.pricing }))
+      catalog = { index: indexCatalog(slim), fetchedAt: Date.now(), error: null }
+      try {
+        mkdirSync(dirname(CACHE_FILE), { recursive: true })
+        writeFileSync(CACHE_FILE, JSON.stringify({ fetchedAt: catalog.fetchedAt, models: slim }))
+      } catch {
+        // a read-only storages dir only costs us the restart-time cache
+      }
+    } catch (error) {
+      catalog = { ...catalog, error: String((error && error.message) || error) }
+    }
+  }
+
+  // Live config from the GUI, with the composition entry as the base layer.
+  ctx.inject(['settings'], (sctx) => {
+    const scope = sctx.settings.register(NS, Config, { base: config })
+    current = scope.get() ?? config
+    const stop = scope.watch(() => {
+      current = scope.get() ?? config
+      void refresh()
+    })
+    sctx.effect(() => () => {
+      stop()
+      current = config
+    })
+  })
+
+  // Optional capability: headless assemblies without the registry keep working,
+  // the widget just falls back to the session-wide tokenUsage projection.
+  ctx.inject(['sessionProjections'], (pctx) => {
+    pctx.sessionProjections.register(usageByModelProjection)
+  })
+
+  ctx.effect(() => {
+    const controller = new AbortController()
+    void refresh(controller.signal)
+    const timer = setInterval(() => void refresh(controller.signal), 3600e3)
+    return () => {
+      clearInterval(timer)
+      controller.abort()
+    }
+  }, 'dsh-cost-meter: catalog refresh')
+
+  ctx.effect(
+    () =>
+      ctx.webServer.register({
+        kind: 'exact',
+        path: STATE_PATH,
+        handler: (req, res) => {
+          if (req.method !== 'GET') {
+            res.writeHead(405, { 'Content-Type': 'application/json' })
+            res.end(JSON.stringify({ error: 'method not allowed' }))
+            return
+          }
+          const cfg = current ?? config
+          const route = defaultRoute(ctx) ?? null
+          const own = resolveTariff(cfg, catalog.index, route)
+
+          // The client asks for every model the session actually used, so a
+          // session that switched models can price each part on its own tariff.
+          const schedules = {}
+          const asked = new URL(req.url, 'http://localhost').searchParams.get('models')
           for (const key of (asked ? asked.split(',') : []).slice(0, 24)) {
             const trimmed = String(key || '').trim()
             if (!trimmed || trimmed.length > 128) continue
@@ -519,22 +533,23 @@ export function apply(ctx, config) {
               matchedId: resolved.matchedId,
             }
           }
-
-          res.writeHead(200, { 'Content-Type': 'application/json', 'Cache-Control': 'no-store' })
-          res.end(JSON.stringify({
-            currency: cfg.currency,
-            usdRate: cfg.usdRate,
-            displayTimeZone: cfg.displayTimeZone || 'Europe/Moscow',
-            route,
-            source: own.source,
-            matchedId: own.matchedId,
-            schedules,
-            fetchedAt: catalog.fetchedAt || null,
-            catalogError: catalog.error,
-            schedule: own.schedule,
-          }))
-        },
-      }),
-    'dsh-cost-meter: state route',
-  )
-}
+
+          res.writeHead(200, { 'Content-Type': 'application/json', 'Cache-Control': 'no-store' })
+          res.end(JSON.stringify({
+            currency: cfg.currency,
+            usdRate: cfg.usdRate,
+            displayTimeZone: cfg.displayTimeZone || 'Europe/Moscow',
+            route,
+            source: own.source,
+            matchedId: own.matchedId,
+            schedules,
+            fetchedAt: catalog.fetchedAt || null,
+            catalogError: catalog.error,
+            catalogModelsCount: Object.keys((catalog.index && catalog.index.byId) || {}).length,
+            schedule: own.schedule,
+          }))
+        },
+      }),
+    'dsh-cost-meter: state route',
+  )
+}