@sligqoer/dsh-usage-stats 0.2.0 → 0.2.2

package/lib/model-detection.js

@@ -0,0 +1,255 @@
+/** Bounded LM Fingerprint detection. No conversation text is sent or retained. */
+import { randomUUID } from 'node:crypto'
+import { createFingerprintEngine } from './lm-fingerprint-engine.js'
+
+const usageKeys = ['inputTokens', 'outputTokens', 'cacheReadTokens', 'cacheWriteTokens', 'reasoningTokens']
+const safeText = (value, max = 256) => typeof value === 'string' ? value.replace(/[\u0000-\u001f\u007f]/g, '').trim().slice(0, max) : ''
+const nonnegative = value => Number.isFinite(value) && value >= 0 ? value : 0
+export function safeUsage(value) {
+  return Object.fromEntries(usageKeys.map(key => [key, nonnegative(value?.[key])]))
+}
+export function modelSelection(value) {
+  const provider = safeText(value?.provider), model = safeText(value?.model)
+  if (!provider || !model) return null
+  const reasoningEffort = safeText(value.reasoningEffort, 64)
+  return { provider, model, ...(reasoningEffort ? { reasoningEffort } : {}) }
+}
+const stopReason = value => ['stop', 'max-tokens', 'tool-calls', 'error', 'aborted', 'unknown'].includes(value) ? value : 'unknown'
+
+export class ModelDetectionError extends Error {}
+export async function readDetectionModels(ctx, provider = '') {
+  const llm = ctx.get('llm')
+  if (typeof llm?.listProviders !== 'function' || typeof llm?.listModels !== 'function') throw new ModelDetectionError('当前宿主不支持读取模型列表，请更新 DSH。')
+  try {
+    const signal = AbortSignal.timeout(10_000)
+    const registered = await abortable(llm.listProviders(), signal)
+    if (!Array.isArray(registered)) throw new Error('invalid providers')
+    const seen = new Set()
+    const providers = registered.flatMap(item => {
+      // listProviders() returns ProviderInfo { id, name }; configurable-provider
+      // declarations use { provider, displayName } on older host integrations.
+      const id = safeText(item?.id || item?.provider)
+      if (!id || seen.has(id)) return []
+      seen.add(id)
+      return [{ provider: id, name: safeText(item.name || item.displayName || id, 512) }]
+    })
+    provider = safeText(provider)
+    if (!provider) return { providers, provider: '', models: [] }
+    if (!seen.has(provider)) throw new ModelDetectionError('请选择已配置的供应商。')
+    const listed = await abortable(llm.listModels(provider), signal)
+    if (!Array.isArray(listed)) throw new Error('invalid models')
+    const ids = new Set()
+    const models = listed.flatMap(item => {
+      const id = safeText(item?.id)
+      if (!id || item.provider !== provider || ids.has(id)) return []
+      ids.add(id)
+      return [{ id, name: safeText(item.name || id, 512) }]
+    })
+    return { providers, provider, models }
+  } catch (error) {
+    if (error instanceof ModelDetectionError) throw error
+    throw new ModelDetectionError('模型列表无法读取，请检查供应商配置后重试。')
+  }
+}
+export const PROBE_PLAN = Object.freeze({ samples: 3, maxTokens: 4096, timeoutMs: 60_000, budgetTokens: 12_000 })
+const sumTokens = usage => usage.inputTokens + usage.outputTokens + usage.cacheReadTokens + usage.cacheWriteTokens
+function modelKey(value) {
+  return safeText(value).toLowerCase().split('/').pop().replace(/[-_]20\d{6}$/, '').replace(/[-_]20\d{2}[-_]\d{2}[-_]\d{2}$/, '').replace(/(\d)-(\d)(?=$|[-.])/, '$1.$2')
+}
+function modelAliases(model) {
+  const first = modelKey(model.id), labels = String(model.name || '').split('/')
+  const prefix = first.match(/^\D+/)?.[0] || ''
+  return [...new Set([first, ...labels.map(v => modelKey(/^\d/.test(v) ? prefix + v : v))])]
+}
+export function normalizeFingerprintIdentity(raw) {
+  if (!raw || !Array.isArray(raw.candidates)) return null
+  const usedSamples = Number.isInteger(raw.usedSamples) && raw.usedSamples >= 0 && raw.usedSamples <= 3 ? raw.usedSamples : 0
+  const probability = v => usedSamples === 3 && raw.probabilityScope === 'reference-closed-set' && Number.isFinite(v) && v >= 0 && v <= 1 ? v : null
+  const candidates = raw.candidates.slice(0, 5).filter(c => safeText(c?.id) && Number.isFinite(c.score)).map(c => ({ id: safeText(c.id), name: safeText(c.name || c.id), family: safeText(c.family), score: c.score, probability: probability(c.probability) }))
+  return { method: safeText(raw.method), usedSamples, requestedSamples: raw.requestedSamples === 1 ? 1 : 3,
+    candidates, probability: probability(raw.probability), probabilityScope: 'reference-closed-set',
+    modelInBank: raw.modelInBank === true, sameCandidate: typeof raw.sameCandidate === 'boolean' ? raw.sameCandidate : null,
+    merged: raw.merged === true, unusualDistribution: raw.unusualDistribution === true }
+}
+export function summarizeModelDetection(job) {
+  if (job.status === 'running') return { verdict: 'running', reason: 'running' }
+  const identity = job.identity
+  const invalid = Object.fromEntries(['incomplete', 'unparseable', 'unavailable', 'timeout'].map(code => [code, (job.results || []).filter(r => r.outcome === code).length]))
+  const reason = job.status === 'cancelled' ? 'cancelled' : job.status === 'interrupted' ? 'interrupted' : job.status === 'budget-exhausted' ? 'budget' :
+    identity?.unusualDistribution ? 'distribution' : invalid.timeout ? 'timeout' : invalid.unavailable ? 'request-failed' : invalid.incomplete ? 'truncated' : invalid.unparseable ? 'format' : 'unfinished'
+  if (['failed', 'budget-exhausted'].includes(job.status) && identity?.usedSamples > 0 && identity.candidates.length && !identity.unusualDistribution) return { verdict: 'ranking-only', reason, invalid }
+  if (job.status === 'complete' && identity?.candidates?.length && !identity.unusualDistribution) {
+    if (identity.usedSamples < 3 || identity.probability === null) return { verdict: 'ranking-only', reason: 'partial-ranking' }
+    if (!identity.modelInBank) return { verdict: 'ranking-only', reason: 'model-not-enrolled' }
+    if (identity.probability < .75) return { verdict: 'ranking-only', reason: 'low-confidence' }
+    return { verdict: identity.sameCandidate ? 'candidate-consistent' : 'candidate-different', reason: identity.merged ? 'merged-class' : 'closed-set-ranking' }
+  }
+  return { verdict: 'inconclusive', reason, invalid }
+}
+function jobView(job) {
+  return { id: job.id, sessionId: job.sessionId, selection: job.selection, testSelection: job.testSelection,
+    status: job.status, phase: job.phase, startedAt: job.startedAt, finishedAt: job.finishedAt, progress: job.results.length,
+    engine: job.engineInfo, plan: job.plan, usage: safeUsage(job.usage), usageReported: job.results.some(r => r.usageReported),
+    usageIncomplete: job.status === 'running' || job.results.some(r => !r.usageReported), usageSaved: job.usageSaved === true,
+    results: job.results.map(({ time, selection, phase, outcome, stop, usage, usageReported, valid, expected, ttftMs, decodeMs }) => ({ time, selection, phase, outcome, stop, usage, usageReported, valid, expected, ttftMs, decodeMs })),
+    identity: job.identity || null, conclusion: summarizeModelDetection(job), message: job.message || '',
+    channel: { ttftMs: job.results.some(r => r.ttftMs !== null) ? job.results.filter(r => r.ttftMs !== null).reduce((n, r) => n + r.ttftMs, 0) / job.results.filter(r => r.ttftMs !== null).length : null,
+      tps: job.results.some(r => r.decodeMs > 0 && r.usage.outputTokens > 0) ? job.results.filter(r => r.decodeMs > 0 && r.usage.outputTokens > 0).reduce((n, r) => n + r.usage.outputTokens, 0) * 1000 / job.results.filter(r => r.decodeMs > 0 && r.usage.outputTokens > 0).reduce((n, r) => n + r.decodeMs, 0) : null } }
+}
+function abortable(promise, signal) {
+  return new Promise((resolve, reject) => {
+    const abort = () => reject(new Error('aborted'))
+    if (signal.aborted) { abort(); return }
+    signal.addEventListener('abort', abort, { once: true })
+    Promise.resolve(promise).then(resolve, reject).finally(() => signal.removeEventListener('abort', abort))
+  })
+}
+export function createModelDetectionRunner(ctx, options = {}) {
+  const engine = options.engine || createFingerprintEngine({ path: options.enginePath, fetch: options.fetch })
+  const jobs = new Map(), quotes = new Map()
+  let disposed = false
+  async function checkpoint(job, patch = {}) {
+    if (!options.onUpdate) return
+    try { await options.onUpdate({ ...jobView({ ...job, ...patch }), usageSaved: true }); job.usageSaved = true }
+    catch { job.usageSaved = false; job.halt = true; job.message = '检测用量保存失败，已停止后续请求；请检查本地存储。' }
+  }
+  function discardQuote(id) { const quote = quotes.get(id); if (quote) { quote.engine.dispose(); clearTimeout(quote.timer); quotes.delete(id) } }
+  async function request(job, challenge) {
+    const timeout = new AbortController(), timer = setTimeout(() => timeout.abort(), options.timeoutMs || PROBE_PLAN.timeoutMs)
+    const signal = AbortSignal.any([job.controller.signal, timeout.signal])
+    const item = { time: 0, selection: job.testSelection, phase: 'fingerprint', outcome: 'unavailable', stop: 'unknown',
+      usage: safeUsage(null), usageReported: false, valid: 0, expected: challenge.expected_count, ttftMs: null, decodeMs: null }
+    let iterator, text = '', firstToken = 0, output = null
+    const started = Date.now()
+    try {
+      const llm = ctx.get('llm')
+      iterator = llm.stream({ ...job.testSelection, maxTokens: job.plan.maxTokens,
+        messages: [{ role: 'user', content: [{ type: 'text', text: challenge.prompt }] }], signal })[Symbol.asyncIterator]()
+      for (;;) {
+        const next = await abortable(iterator.next(), signal)
+        if (signal.aborted) throw new Error('aborted')
+        if (next.done) break
+        const chunk = next.value
+        if (chunk?.type === 'text-delta') {
+          if (chunk.text && !firstToken) firstToken = Date.now()
+          text += String(chunk.text || '')
+          if (text.length > 32_768) { item.stop = 'max-tokens'; break }
+        }
+        if (chunk?.type === 'usage') {
+          for (const key of usageKeys) if (Number.isFinite(chunk.usage?.[key]) && chunk.usage[key] >= 0) { item.usage[key] = chunk.usage[key]; item.usageReported = true }
+        }
+        if (chunk?.type === 'finish') { item.stop = stopReason(chunk.reason?.kind); break }
+      }
+      if (['stop', 'max-tokens'].includes(item.stop)) {
+        const numbers = await job.engine.call('parse', text, signal)
+        item.valid = Math.min(numbers.length, challenge.expected_count)
+        if (item.stop === 'max-tokens' && numbers.length < challenge.expected_count) item.outcome = 'incomplete'
+        else if (numbers.length < Math.max(80, Math.ceil(challenge.expected_count * .55))) item.outcome = 'unparseable'
+        else { item.outcome = 'scored'; output = { text: numbers.slice(0, challenge.expected_count).join(','), expected_count: challenge.expected_count } }
+        if (firstToken) { item.ttftMs = firstToken - started; item.decodeMs = Date.now() - firstToken }
+      }
+    } catch { item.outcome = job.controller.signal.aborted ? 'cancelled' : timeout.signal.aborted ? 'timeout' : 'unavailable'; item.stop = signal.aborted ? 'aborted' : 'error' }
+    finally {
+      clearTimeout(timer); timeout.abort()
+      try { Promise.resolve(iterator?.return?.()).catch(() => {}) } catch {}
+      text = ''
+      item.time = Date.now(); job.results.push(item)
+      for (const key of usageKeys) job.usage[key] += item.usage[key]
+      if (item.outcome !== 'scored') job.halt = true
+      if (sumTokens(job.usage) >= job.plan.budgetTokens && job.results.length < job.plan.requests) { job.halt = true; job.budgetExhausted = true }
+      await checkpoint(job)
+    }
+    return output
+  }
+  async function run(job, quote) {
+    const outputs = []
+    try {
+      await checkpoint(job)
+      for (const challenge of quote.challenges) {
+        if (disposed || job.controller.signal.aborted || job.halt) break
+        job.phase = 'sampling'
+        const output = await request(job, challenge)
+        if (output) outputs.push(output)
+      }
+      // Score the valid samples even when a later call fails. The official
+      // scorer withholds calibrated probabilities for fewer than 3 samples.
+      if (!disposed && !job.controller.signal.aborted && job.usageSaved !== false && outputs.length > 0) {
+        job.phase = 'scoring'
+        const analysis = await job.engine.call('score', outputs, job.controller.signal)
+        const requested = modelKey(job.selection.model)
+        const claimed = job.engine.models.find(m => modelAliases(m).includes(requested))
+        job.identity = normalizeFingerprintIdentity({ method: analysis.method, usedSamples: analysis.used_outputs, requestedSamples: job.plan.samples,
+          candidates: (analysis.results || []).map(c => ({ id: c.model, name: c.display_name, family: c.family, score: c.score, probability: c.probability })),
+          probability: analysis.probability, probabilityScope: analysis.probability_scope,
+          modelInBank: !!claimed, sameCandidate: claimed ? claimed.id === analysis.prediction : null,
+          merged: !!claimed && modelAliases(claimed).length > 1,
+          unusualDistribution: outputs.some(o => Math.min(...o.text.split(',').map(Number)) >= 200) })
+      }
+    } catch { job.halt = true; job.message = '指纹检测未完成，请检查通道或引擎状态。' }
+    finally {
+      outputs.length = 0; quote.challenges.length = 0; job.engine.dispose()
+      const status = disposed || job.controller.signal.aborted ? 'cancelled' : job.budgetExhausted ? 'budget-exhausted' : job.halt ? 'failed' : 'complete'
+      if (status === 'cancelled' || job.usageSaved === false) job.identity = null
+      job.finishedAt = Date.now(); job.phase = 'finished'
+      await checkpoint(job, { status })
+      job.status = job.usageSaved === false && options.onUpdate ? 'failed' : status
+      if (job.usageSaved === false) job.identity = null
+    }
+  }
+  return {
+    catalog: () => engine.catalog(), updateEngine: () => engine.update(true),
+    async plan(body) {
+      if (disposed) throw new ModelDetectionError('检测服务已停止。')
+      const selection = modelSelection(body?.selection), sessionId = safeText(body?.sessionId)
+      if (!selection) throw new ModelDetectionError('请选择供应商与检测模型。')
+      const samples = body.sampleCount ?? 3
+      if (![1, 3].includes(samples)) throw new ModelDetectionError('请选择 1 次初筛或 3 次标准检测。')
+      const budgetTokens = body.budgetTokens ?? PROBE_PLAN.budgetTokens
+      if (!Number.isInteger(budgetTokens) || budgetTokens < 1024 || budgetTokens > 50_000) throw new ModelDetectionError('Token 预算应为 1,024–50,000 的整数。')
+      const maxTokens = body.maxTokens ?? PROBE_PLAN.maxTokens
+      if (!Number.isInteger(maxTokens) || maxTokens < 1024 || maxTokens > 16_384) throw new ModelDetectionError('单次输出上限应为 1,024–16,384 的整数。')
+      const llm = ctx.get('llm')
+      if (typeof llm?.stream !== 'function' || typeof llm?.resolveModelInfo !== 'function') throw new ModelDetectionError('当前宿主不支持模型参数校验，请更新 DSH。')
+      let info
+      try { info = await llm.resolveModelInfo(selection.provider, selection.model, AbortSignal.timeout(10_000)) }
+      catch { throw new ModelDetectionError('模型信息无法读取，请检查供应商配置后重试。') }
+      const efforts = info.reasoning?.efforts?.map(e => e.id)
+      const effort = efforts ? ['off', 'none', 'minimal', 'low'].find(e => efforts.includes(e)) || selection.reasoningEffort || info.reasoning.defaultEffort : undefined
+      if (effort && !efforts.includes(effort)) throw new ModelDetectionError('模型推理参数不受支持。')
+      const testSelection = { provider: selection.provider, model: selection.model, ...(effort ? { reasoningEffort: effort } : {}) }
+      let lease
+      try {
+        lease = await engine.acquire()
+        const challenges = await lease.call('challenges', samples)
+        const plan = { requests: samples, samples, maxTokens, numbers: challenges.map(c => c.expected_count), maxOutputTokens: samples * maxTokens,
+          estimatedInputTokens: Math.ceil(challenges.reduce((n, c) => n + c.prompt.length, 0) / 2), budgetTokens, timeoutMs: options.timeoutMs || PROBE_PLAN.timeoutMs }
+        for (const [id, quote] of quotes) if (Date.now() - quote.createdAt > 5 * 60_000) discardQuote(id)
+        // Only the newest budget can start detection; keep one idle worker.
+        for (const id of quotes.keys()) discardQuote(id)
+        const id = randomUUID(), timer = setTimeout(() => discardQuote(id), 5 * 60_000); timer.unref?.()
+        const engineInfo = { name: 'lmfpd', version: lease.manifest.version, revision: lease.manifest.revision, builtAt: lease.manifest.builtAt, modelCount: lease.models.length }
+        quotes.set(id, { id, timer, sessionId, selection, testSelection, plan, engineInfo, engine: lease, challenges, createdAt: Date.now() })
+        return { id, selection, testSelection, plan, engine: engineInfo }
+      } catch { lease?.dispose(); throw new ModelDetectionError('指纹引擎不可用，请检查更新状态。') }
+    },
+    start(body) {
+      if (disposed) throw new ModelDetectionError('检测服务已停止。')
+      if (body?.ackCost !== true) throw new ModelDetectionError('请确认检测会消耗 Token 后再开始。')
+      const quote = quotes.get(body.planId)
+      if (!quote || Date.now() - quote.createdAt > 5 * 60_000) throw new ModelDetectionError('检测预算已过期，请刷新预算后开始。')
+      if (safeText(body.sessionId) !== quote.sessionId) throw new ModelDetectionError('当前会话已变化，请重新生成预算。')
+      if ([...jobs.values()].some(j => j.status === 'running')) throw new ModelDetectionError('已有检测正在运行，请先等待或取消。')
+      quotes.delete(quote.id); clearTimeout(quote.timer)
+      while (jobs.size >= 10) jobs.delete(jobs.keys().next().value)
+      const job = { id: randomUUID(), sessionId: quote.sessionId, selection: quote.selection, testSelection: quote.testSelection,
+        plan: quote.plan, engineInfo: quote.engineInfo, engine: quote.engine, status: 'running', phase: 'queued', startedAt: Date.now(), finishedAt: 0,
+        controller: new AbortController(), results: [], usage: safeUsage(null), identity: null }
+      jobs.set(job.id, job)
+      queueMicrotask(() => { void run(job, quote) })
+      return jobView(job)
+    },
+    status(id, sessionId) { const job = id ? jobs.get(id) : [...jobs.values()].reverse().find(j => !sessionId || j.sessionId === sessionId); if (id && !job) throw new ModelDetectionError('检测记录已过期。'); return job ? jobView(job) : null },
+    cancel(id) { const job = jobs.get(id); if (!job) throw new ModelDetectionError('检测记录已过期。'); if (job.status === 'running') job.controller.abort(); return jobView(job) },
+    dispose() { disposed = true; for (const job of jobs.values()) job.controller.abort(); for (const id of quotes.keys()) discardQuote(id); engine.dispose() }
+  }
+}

package/lib/quota-controls.js

@@ -1,5 +1,5 @@
 /** Persistent quota controls, bounded custom queries and host-side scheduling. */
-import { readFile, mkdir, writeFile, rename } from 'node:fs/promises'
+import { readFile, mkdir, writeFile, rename, stat as fileStat } from 'node:fs/promises'
 import { dirname } from 'node:path'
 import { randomUUID } from 'node:crypto'
 import { quotaTemplate, hasBuiltinQuotaQuery, officialQuotaSpec, customQuotaAuth, quotaErrorCode } from './provider-quotas.js'
@@ -20,7 +20,7 @@ export class QuotaControlError extends Error {
 const fail = (message, code, status) => { throw new QuotaControlError(message, code, status) }
 
 function defaults() {
-  return { autoQuota: false, intervalMinutes: 60, showModelDetails: true, showModelMatch: false, showTokenSpeed: true, advancedModelSelect: false, customQueries: [] }
+  return { autoQuota: false, intervalMinutes: 10, showModelDetection: true, advancedModelSelect: true, modelReasoning: true, customQueries: [] }
 }
 
 /**
@@ -213,7 +213,7 @@ export async function executeQuotaQuery(ctx, provider, rawTemplate, request = fe
 }
 
 /** Source reads are explicit; cached page reads never spend account credit. */
-export function createQuotaControls(ctx, { path, sources, request = fetch, now = Date.now }) {
+export function createQuotaControls(ctx, { path, legacyPath, sources, request = fetch, now = Date.now }) {
   let settings = defaults()
   let snapshots = {}
   let lastRefreshAt = 0
@@ -246,14 +246,24 @@ export function createQuotaControls(ctx, { path, sources, request = fetch, now =
 
   const load = () => loaded ||= (async () => {
     try {
-      const stat = JSON.parse(await readFile(path, 'utf8'))
+      let text
+      try { text = await readFile(path, 'utf8') }
+      catch (error) {
+        if (error.code !== 'ENOENT' || !legacyPath) throw error
+        const candidates = await Promise.all((Array.isArray(legacyPath) ? legacyPath : [legacyPath]).map(async file => {
+          try { return { file, modified: (await fileStat(file)).mtimeMs } } catch { return null }
+        }))
+        const latest = candidates.filter(Boolean).sort((a, b) => b.modified - a.modified)[0]
+        if (!latest) throw error
+        text = await readFile(latest.file, 'utf8')
+      }
+      const stat = JSON.parse(text)
       const saved = stat.settings
       if (plain(saved)) {
         settings.autoQuota = saved.autoQuota === true
-        settings.showModelDetails = saved.showModelDetails !== false
-        settings.showModelMatch = saved.showModelMatch === true
-        settings.showTokenSpeed = saved.showTokenSpeed !== false
-        settings.advancedModelSelect = saved.advancedModelSelect === true
+        settings.showModelDetection = saved.showModelDetection !== false
+        settings.advancedModelSelect = saved.advancedModelSelect !== false
+        settings.modelReasoning = saved.modelReasoning !== false
         if (QUOTA_INTERVALS.includes(saved.intervalMinutes)) settings.intervalMinutes = saved.intervalMinutes
         if (Array.isArray(saved.customQueries)) {
           settings.customQueries = saved.customQueries.slice(0, 50).flatMap((query) => {
@@ -264,7 +274,7 @@ export function createQuotaControls(ctx, { path, sources, request = fetch, now =
       }
       if (plain(stat.snapshots)) snapshots = Object.fromEntries(Object.keys(sources).filter((key) => plain(stat.snapshots[key])).map((key) => [key, stat.snapshots[key]]))
       if (Number.isFinite(stat.lastRefreshAt) && stat.lastRefreshAt > 0) lastRefreshAt = Math.min(stat.lastRefreshAt, now())
-    } catch {} // Missing/corrupt settings use conservative defaults.
+    } catch {} // Missing/corrupt settings use defaults.
   })()
 
   const persist = () => {
@@ -301,8 +311,9 @@ export function createQuotaControls(ctx, { path, sources, request = fetch, now =
     if (key !== 'providers') return clone(body)
     const entries = directory(ctx)
     const quotas = (Array.isArray(body.quotas) ? body.quotas : []).filter((quota) => {
-      if (!officialQuotaSpec(quota.provider)?.configuredOnly) return true
       const entry = entries.find((item) => item.provider === quota.provider)
+      if (!entry) return false
+      if (!officialQuotaSpec(quota.provider)?.configuredOnly) return true
       return !!entry?.profile && (settings.customQueries.some((query) => query.provider === quota.provider) || hasBuiltinQuotaQuery(entry.provider, entry.profile))
     })
     const shown = new Set(quotas.map((quota) => quota.provider))
@@ -322,7 +333,12 @@ export function createQuotaControls(ctx, { path, sources, request = fetch, now =
     if (!Object.hasOwn(sources, key)) fail('未知配额来源。')
     const source = sources[key]
     const adapter = plain(source)
-    if (provider && (key !== 'providers' || !directory(ctx).some((entry) => entry.provider === provider))) fail('所选供应商已不存在，请重新选择。')
+    if (provider && key !== 'providers') fail('所选供应商已不存在，请重新选择。')
+    // A card can outlive the provider's settings. Return the current display
+    // snapshot so the client removes it instead of retaining its old balance.
+    if (provider && !directory(ctx).some((entry) => entry.provider === provider)) {
+      return displaySnapshot(key, snapshots[key] || { ok: true, quotas: [], cached: true })
+    }
     if (!sourceEnabled(key)) return { ok: true, available: false, status: 'not-configured', cached: true }
     if (!fresh) {
       const body = snapshots[key] || (key === 'providers' ? { ok: true, quotas: [], cached: true } : { ok: true, available: adapter, status: adapter ? 'unqueried' : 'unavailable', cached: true })
@@ -426,22 +442,18 @@ export function createQuotaControls(ctx, { path, sources, request = fetch, now =
           if (!QUOTA_INTERVALS.includes(patch.intervalMinutes)) fail('仅支持 10 分钟、1 小时、5 小时或每天。')
           next.intervalMinutes = patch.intervalMinutes
         }
-        if (patch.showModelDetails != null) {
-          if (typeof patch.showModelDetails !== 'boolean') fail('模型明细开关值无效。')
-          next.showModelDetails = patch.showModelDetails
-        }
-        if (patch.showModelMatch != null) {
-          if (typeof patch.showModelMatch !== 'boolean') fail('检测模型匹配开关值无效。')
-          next.showModelMatch = patch.showModelMatch
-        }
-        if (patch.showTokenSpeed != null) {
-          if (typeof patch.showTokenSpeed !== 'boolean') fail('Token 输出速度开关值无效。')
-          next.showTokenSpeed = patch.showTokenSpeed
+        if (patch.showModelDetection != null) {
+          if (typeof patch.showModelDetection !== 'boolean') fail('模型检测开关值无效。')
+          next.showModelDetection = patch.showModelDetection
         }
         if (patch.advancedModelSelect != null) {
           if (typeof patch.advancedModelSelect !== 'boolean') fail('高级模型选择器开关值无效。')
           next.advancedModelSelect = patch.advancedModelSelect
         }
+        if (Object.prototype.hasOwnProperty.call(patch, 'modelReasoning')) {
+          if (typeof patch.modelReasoning !== 'boolean') fail('模型推理强度开关值无效。')
+          next.modelReasoning = patch.modelReasoning
+        }
       } else fail('未知设置操作。')
       const oldSnapshots = clone(snapshots)
       if (patch.action) {
@@ -465,7 +477,7 @@ export function createQuotaControls(ctx, { path, sources, request = fetch, now =
   return {
     get, update, test, read, tick,
     providers: async () => ({ ok: true, providers: providerDirectory() }),
-    dispose: () => { disposed = true },
+    dispose: async () => { disposed = true; await updateQueue.catch(() => {}); await Promise.allSettled([allFlight, ...flights.values()]); await saveQueue.catch(() => {}) },
   }
 }
 
@@ -474,6 +486,35 @@ const WORKBUDDY_SOURCES = {
   'workbuddy-ai': { name: 'WorkBuddy AI', path: '/plugins/dsh-workbuddy-connect/ai/status' },
 }
 
+/** Display the selected built-in account; never import credentials on a page read. */
+export function nativeWorkBuddyQuotaSource(getService, provider = 'workbuddy') {
+  if (!Object.hasOwn(WORKBUDDY_SOURCES, provider)) fail('未知 WorkBuddy 积分来源。')
+  const channelId = provider === 'workbuddy-ai' ? 'workbuddy' : 'buddy'
+  const name = provider === 'workbuddy-ai' ? 'WorkBuddy 国际版' : 'WorkBuddy 国内版'
+  const selected = service => service?.accountsView(channelId).accounts.find(account => account.selected && account.enabled && account.status === 'ready')
+  return {
+    provider, name,
+    description: '读取“模型接入 → 账号渠道”中已选定的 ' + name + ' 账号积分，无需其他插件。',
+    available: () => Boolean(selected(getService())),
+    async read() {
+      const service = getService()
+      const account = selected(service)
+      if (!account) return { ok: true, available: false, status: 'not-configured' }
+      const reading = (await service.refreshQuota({ channelId, accountId: account.id })).results?.[0]
+      const current = selected(service)
+      if (!reading || reading.error || !current || current.id !== account.id || current.version !== account.version) {
+        return { ok: true, available: true, status: reading?.error?.code === 'AUTH' ? 'signed-out' : 'unavailable' }
+      }
+      return {
+        ok: true, available: true, status: 'signed-in', syncedAt: reading.observedAt,
+        credits: { total: reading.total, ...reading.unlimited === true ? { unlimited: true } : {},
+          accounts: (reading.kinds ?? []).map(pool => ({ packageName: pool.name, remain: pool.balance, size: pool.total })) },
+        models: service.models(channelId).filter(model => model.accountId === account.id && model.billing === 'account-included' && model.callable === true).map(model => ({ id: model.modelId, name: model.displayName, free: true, badges: [] })),
+      }
+    },
+  }
+}
+
 /** Optional plugin registration is detected without probing its HTTP route. */
 export function workBuddyAvailable(ctx, provider = 'workbuddy') {
   if (!Object.hasOwn(WORKBUDDY_SOURCES, provider)) return false

package/lib/supplemental-usage.js

@@ -0,0 +1,77 @@
+/** Persist and deduplicate reported detection usage across plugin restarts. */
+import { readFile, mkdir, writeFile, rename } from 'node:fs/promises'
+import { dirname } from 'node:path'
+import { getBridge, HUB_SERVICE } from './bridge.js'
+
+const keys = ['inputTokens', 'outputTokens', 'cacheReadTokens', 'cacheWriteTokens', 'reasoningTokens']
+function sanitize(raw) {
+  if (!raw || typeof raw.id !== 'string' || !raw.id || !Number.isFinite(raw.time) || raw.time <= 0
+    || typeof raw.selection?.provider !== 'string' || typeof raw.selection?.model !== 'string') throw new Error('Invalid supplemental usage')
+  return { id: raw.id.slice(0, 320), time: raw.time,
+    selection: { provider: raw.selection.provider.slice(0, 256), model: raw.selection.model.slice(0, 256) },
+    usageReported: raw.usageReported === true,
+    usage: Object.fromEntries(keys.map(key => [key, Number.isFinite(raw.usage?.[key]) && raw.usage[key] >= 0 ? raw.usage[key] : 0])),
+    stop: ['stop', 'max-tokens'].includes(raw.stop) ? raw.stop : 'unknown',
+    ttftMs: Number.isFinite(raw.ttftMs) && raw.ttftMs >= 0 ? raw.ttftMs : null,
+    decodeMs: Number.isFinite(raw.decodeMs) && raw.decodeMs > 0 ? raw.decodeMs : null }
+}
+
+export function createSupplementalUsage(ctx, path, legacyPath, history) {
+  let records = new Map(), warning = '', loaded, flight, queue = Promise.resolve(), connected = false
+  const merge = rows => { const next = new Map(records); for (const row of rows) { const cleaned = sanitize(row); next.set(cleaned.id, cleaned) }; records = next }
+  const persist = () => {
+    const text = JSON.stringify({ version: 1, records: [...records.values()] })
+    return queue = queue.catch(() => {}).then(async () => {
+      await mkdir(dirname(path), { recursive: true })
+      await writeFile(path + '.tmp', text, { mode: 0o600 }); await rename(path + '.tmp', path)
+    })
+  }
+  const load = () => loaded ||= (async () => {
+    try {
+      let document
+      try { document = JSON.parse(await readFile(path, 'utf8')); if (document.version !== 1 || !Array.isArray(document.records)) throw new Error('Invalid cache'); merge(document.records) }
+      catch (error) {
+        if (error.code !== 'ENOENT') throw error
+        try { document = JSON.parse(await readFile(legacyPath, 'utf8')) } catch (legacyError) { if (legacyError.code !== 'ENOENT') throw legacyError; return }
+        if (document.version !== 1 || !Array.isArray(document.usage)) throw new Error('Invalid legacy history')
+        merge(document.usage.flatMap(job => (job.results || []).map((row, index) => ({ ...row, selection: row.selection || job.selection,
+          time: row.time || job.finishedAt || job.startedAt, id: `${job.id}:${index}` }))))
+        await persist()
+      }
+    } catch { warning = '检测用量记录无法读取，统计可能不完整。' }
+  })()
+  return {
+    refresh() {
+      if (flight) return flight
+      return flight = (async () => {
+        await load()
+        if (history) {
+          const before = JSON.stringify([...records.values()])
+          merge(history.usageRecords())
+          if (before !== JSON.stringify([...records.values()])) await persist()
+          warning = history.warning()
+        }
+        const hub = getBridge(ctx, HUB_SERVICE)
+        connected = !!hub
+        if (!hub || typeof hub.readSupplementalUsage !== 'function') return
+        let timer
+        try {
+          const snapshot = await Promise.race([
+            Promise.resolve().then(() => hub.readSupplementalUsage()),
+            new Promise((_, reject) => { timer = setTimeout(() => reject(new Error('Supplemental usage timeout')), 1500); timer.unref?.() }),
+          ])
+          if (snapshot.protocolVersion !== 1 || !Array.isArray(snapshot.records)) throw new Error('Invalid interface')
+          const before = JSON.stringify([...records.values()])
+          merge(snapshot.records)
+          if (before !== JSON.stringify([...records.values()])) await persist()
+          warning = snapshot.warning || ''
+        } catch { warning = '补充用量暂时不可用，保留上次成功记录。' }
+        finally { clearTimeout(timer) }
+      })().finally(() => { flight = null })
+    },
+    usage: () => [...records.values()],
+    warning: () => warning,
+    connected: () => connected,
+    flush: () => queue.catch(() => {}),
+  }
+}

package/lib/usage-preferences.js

@@ -0,0 +1,51 @@
+/** Persistent usage display preferences. */
+import { readFile, mkdir, writeFile, rename } from 'node:fs/promises'
+import { dirname } from 'node:path'
+
+const defaults = () => ({ showModelDetails: false, showTokenSpeed: true })
+export class PreferenceError extends Error {}
+export function validateDisplayPatch(patch) {
+  if (!patch || typeof patch !== 'object' || Array.isArray(patch)) throw new PreferenceError('设置应为 JSON 对象。')
+  for (const [key, value] of Object.entries(patch)) {
+    if (!['showModelDetails', 'showTokenSpeed'].includes(key) || typeof value !== 'boolean') throw new PreferenceError('用量展示设置无效。')
+  }
+}
+
+export function createUsagePreferences(path, legacyPath) {
+  let settings = defaults(), loaded, queue = Promise.resolve()
+  const load = () => loaded ||= (async () => {
+    let text
+    try { text = await readFile(path, 'utf8') }
+    catch (error) {
+      if (error.code !== 'ENOENT') throw error
+      try { text = await readFile(legacyPath, 'utf8') } catch (legacyError) { if (legacyError.code !== 'ENOENT') throw legacyError }
+    }
+    if (text) {
+      const saved = JSON.parse(text).settings
+      if (!saved || typeof saved !== 'object') throw new Error('Invalid usage display preferences')
+      for (const key of Object.keys(settings)) if (typeof saved[key] === 'boolean') settings[key] = saved[key]
+    }
+  })()
+  return {
+    async get() { await load(); return { ok: true, settings: { ...settings } } },
+    update(patch) {
+      const operation = queue.catch(() => {}).then(async () => {
+        await load()
+        validateDisplayPatch(patch)
+        const next = { ...settings }
+        for (const [key, value] of Object.entries(patch)) {
+          if (!Object.hasOwn(next, key) || typeof value !== 'boolean') throw new PreferenceError('用量展示设置无效。')
+          next[key] = value
+        }
+        await mkdir(dirname(path), { recursive: true })
+        await writeFile(path + '.tmp', JSON.stringify({ version: 1, settings: next }), { mode: 0o600 })
+        await rename(path + '.tmp', path)
+        settings = next
+        return { ok: true, settings: { ...settings } }
+      })
+      queue = operation
+      return operation
+    },
+    flush: () => queue.catch(() => {}),
+  }
+}

package/locale/en.json

@@ -0,0 +1,6 @@
+{
+  "meta": {
+    "title": "Usage statistics",
+    "description": "Usage statistics, provider quotas, model detection and controls for DSH Desktop 0.2.1-alpha.2 and 0.2.0-rc.2."
+  }
+}

package/locale/ja.json

@@ -0,0 +1,6 @@
+{
+  "meta": {
+    "title": "利用状況",
+    "description": "利用統計、提供元の利用枠、モデル検査、操作設定。DSH Desktop 0.2.1-alpha.2 と 0.2.0-rc.2 に対応。"
+  }
+}

package/locale/zh.json

@@ -0,0 +1,6 @@
+{
+  "meta": {
+    "title": "用量统计",
+    "description": "会话用量统计、供应商配额、模型检测与控制设置。适用于 DSH Desktop 0.2.1-alpha.2 和 0.2.0-rc.2。"
+  }
+}