@goodandready/dsh-subscriptions 0.4.16 → 0.4.17

package/README.md

@@ -136,6 +136,22 @@ const res = await ctx.subscriptions.request('codex', '/backend-api/codex/images/
 
 ---
 
+### 9. 🦛 Local Ollama Gateway & Seamless Fallback (`v0.4.17`)
+* **Native Provider (`ollama`)**: When a local Ollama is reachable at `ollamaBaseUrl` (default `http://127.0.0.1:11434`), it appears in the native DSH model picker with the models discovered from `/api/tags`. No API key needed.
+* **Seamless Quota Fallback (`ollamaFallback`, on by default)**: When every account of a provider is exhausted (or unreachable) and nothing has been streamed yet, the chat continues on a local model (`ollamaFallbackModel`, or the first model from `/api/tags`). The fallback is logged and recorded in request history as `kind: fallback`.
+* **Free ($0) Emergency Path**: Works with no internet and no quota — ideal for offline demos.
+
+### 10. ⚡ Reasoning Effort, Verbosity & Fast Mode (`v0.4.17`)
+* **Reasoning Effort**: Codex models advertise their supported effort levels from the live catalog; the native picker validates and the chosen effort is transmitted as `reasoning.effort` in the Codex `/responses` protocol. Grok forwards effort with its own catalog-aware filtering.
+* **Verbosity (`codexVerbosity`)**: `low` / `medium` / `high` is sent as `text.verbosity` for Codex reasoning models. Empty = protocol default.
+* **Fast Mode (`codexFastMode`)**: Sends `service_tier: priority` (1.5x speed billing tier) with every Codex request. The active-subscription chip shows a `⚡` prefix while enabled.
+
+### 11. 🚦 Family-Scoped Cooldowns & Model Filtering (`v0.4.17`)
+* **Reasoning vs Standard**: A 429 on a reasoning model (claude `*thinking*`, grok `*reasoning*`, all codex models) cools down only the reasoning family of that account — standard models on the same account keep working immediately. Legacy cooldowns (from older versions) still block the whole account until expiry.
+* **Hide Deprecated Models (`hideDeprecatedModels`)**: Filters `test`/`preview`/`dev`/`alpha`/`beta`/`legacy` model ids out of the native picker (applies to live catalogs and the static fallback).
+
+---
+
 ## 📦 Quick Installation
 
 ```bash
@@ -155,6 +171,12 @@ dsh-subscriptions:
   cooldownMs: 60000
   autoLoopback: true        # v0.4.9: catch loopback OAuth callbacks automatically
   privacyMask: false        # v0.4.9: mask emails and account identifiers in the UI
+  ollamaBaseUrl: http://127.0.0.1:11434  # v0.4.17: local Ollama gateway
+  ollamaFallback: true      # v0.4.17: seamless fallback when all accounts are exhausted
+  ollamaFallbackModel: ''   # v0.4.17: e.g. qwen2.5-coder; empty = first model from /api/tags
+  hideDeprecatedModels: false # v0.4.17: filter test/preview/beta/legacy model ids
+  codexVerbosity: ''        # v0.4.17: low | medium | high (text.verbosity)
+  codexFastMode: false      # v0.4.17: service_tier priority (1.5x speed tier)
   # Per-slot fields (v0.4.9): expiresAt (ms epoch), proxyUrl (http/https/socks5://)
   accounts:
     codex:

package/README.ru.md

@@ -136,6 +136,22 @@ const res = await ctx.subscriptions.request('codex', '/backend-api/codex/images/
 
 ---
 
+### 9. 🦛 Локальный шлюз Ollama и бесшовный фолбэк (`v0.4.17`)
+* **Нативный провайдер (`ollama`)**: если локальный Ollama доступен по `ollamaBaseUrl` (по умолчанию `http://127.0.0.1:11434`), он появляется в нативном селекторе моделей DSH с моделями из `/api/tags`. Ключ не нужен.
+* **Бесшовный фолбэк (`ollamaFallback`, включён по умолчанию)**: когда все аккаунты провайдера исчерпаны (или недоступны) и ещё ничего не выведено, чат продолжается на локальной модели (`ollamaFallbackModel`, либо первая из `/api/tags`). Фолбэк пишется в журнал и в историю запросов как `kind: fallback`.
+* **Бесплатный ($0) аварийный путь**: работает без интернета и без квот — удобно для офлайн-демо.
+
+### 10. ⚡ Effort, Verbosity и Fast Mode (`v0.4.17`)
+* **Уровень рассуждений (Effort)**: модели Codex объявляют поддерживаемые уровни из живого каталога; нативный селектор валидирует выбор, и выбранный уровень передаётся как `reasoning.effort` в протокол `/responses`. Grok передаёт effort со своей каталог-осознанной фильтрацией.
+* **Детализация (`codexVerbosity`)**: `low` / `medium` / `high` передаётся как `text.verbosity` для reasoning-моделей Codex. Пусто = дефолт протокола.
+* **Fast Mode (`codexFastMode`)**: c каждым Codex-запросом передаётся `service_tier: priority` (платный скоростной тир 1.5x). В чипе активной подписки при включённом режиме отображается `⚡`.
+
+### 11. 🚦 Кулдауны по семействам и фильтрация моделей (`v0.4.17`)
+* **Reasoning vs Standard**: 429 на reasoning-модели (claude `*thinking*`, grok `*reasoning*`, все codex-модели) охлаждает только семейство reasoning этого аккаунта — standard-модели того же аккаунта продолжают работать. Legacy-кулдауны (от старых версий) по-прежнему блокируют весь аккаунт до истечения.
+* **Скрыть устаревшие модели (`hideDeprecatedModels`)**: убирает из нативного селектора id с `test`/`preview`/`dev`/`alpha`/`beta`/`legacy` (действует и на живые каталоги, и на статичный фолбэк).
+
+---
+
 ## 📦 Быстрая установка
 
 ```bash
@@ -152,6 +168,12 @@ dsh-subscriptions:
   cooldownMs: 60000
   autoLoopback: true        # v0.4.9: автоматически ловить loopback-callback
   privacyMask: false        # v0.4.9: маскировать email и учётные записи в UI
+  ollamaBaseUrl: http://127.0.0.1:11434  # v0.4.17: локальный шлюз Ollama
+  ollamaFallback: true      # v0.4.17: бесшовный фолбэк при исчерпании всех аккаунтов
+  ollamaFallbackModel: ''   # v0.4.17: например qwen2.5-coder; пусто = первая модель из /api/tags
+  hideDeprecatedModels: false # v0.4.17: фильтровать test/preview/beta/legacy id моделей
+  codexVerbosity: ''        # v0.4.17: low | medium | high (text.verbosity)
+  codexFastMode: false      # v0.4.17: service_tier priority (скоростной тир 1.5x)
   # Поля слота (v0.4.9): expiresAt (ms), proxyUrl (http/https/socks5://)
   accounts:
     codex:

package/README.zh.md

@@ -89,6 +89,20 @@ graph LR
 * 每个账号可配置独立代理（`http://`、`https://`、`socks5://`）；令牌刷新、供应商检查、模型请求均走该代理。
 * 账号卡片内的 **Check proxy** 按钮一键检测代理延迟。
 
+### 本地 Ollama 网关与无缝回退（`v0.4.17`）
+* **原生 provider (`ollama`)**：本地 Ollama 可达时（`ollamaBaseUrl`，默认 `http://127.0.0.1:11434`），自动出现在 DSH 原生模型选择器中（模型来自 `/api/tags`），无需密钥。
+* **无缝配额回退 (`ollamaFallback`，默认开启)**：某供应商的所有账号耗尽或不可达且尚未输出任何内容时，对话自动切换到本地模型（`ollamaFallbackModel` 或 `/api/tags` 第一个模型），并记录到请求历史（`kind: fallback`）。
+* 免费（$0）应急通道：无网络、无配额也能用。
+
+### Effort、Verbosity 与 Fast Mode（`v0.4.17`）
+* **推理力度 (Effort)**：Codex 模型从实时目录上报支持的力度等级，原生选择器校验后以 `reasoning.effort` 传入 `/responses` 协议；Grok 按自身目录过滤转发。
+* **详略程度 (`codexVerbosity`)**：`low`/`medium`/`high` 以 `text.verbosity` 传给 Codex 推理模型；留空为协议默认。
+* **Fast Mode (`codexFastMode`)**：每次 Codex 请求携带 `service_tier: priority`（1.5x 速度计费档）；启用时活动订阅芯片显示 `⚡`。
+
+### 按模型家族的冷却与模型过滤（`v0.4.17`）
+* **Reasoning 与 Standard 分离**：推理模型（claude `*thinking*`、grok `*reasoning*`、全部 codex 模型）触发 429 时，只冷却该账号的 reasoning 家族——同账号的 standard 模型立即可用；旧版冷却仍按整账号生效。
+* **隐藏过时模型 (`hideDeprecatedModels`)**：从原生选择器中过滤 `test`/`preview`/`dev`/`alpha`/`beta`/`legacy` 模型 id（同时作用于实时目录与静态回退）。
+
 ### 隐私模式与诊断报告（`v0.4.9`）
 * **`privacyMask`**：一个开关即可在整个界面隐藏个人数据（邮箱显示为 `j***n@example.com`），服务端遮蔽，适合屏幕共享。
 * **匿名诊断报告**：设置卡片内一键生成（插件/运行时版本、系统、各供应商健康状态、HTTP 状态聚合、最近错误与耗时、非敏感配置）并自动复制到剪贴板；令牌、邮箱、凭据名与代理地址严格排除。

package/lib/accounts.js

@@ -42,6 +42,9 @@ export function vendorConfig(provider, cfg) {
     originator: pick('Originator', d.originator || ''),
     systemPrefix: pick('SystemPrefix', d.systemPrefix || ''),
     clientVersion: pick('ClientVersion', d.clientVersion || ''),
+    // #93/#92: per-vendor request shaping (codex only consumes these today).
+    verbosity: pick('Verbosity', ''),
+    fastMode: !!(cfg && cfg[`${provider}FastMode`]),
     models: Array.isArray(cfg && cfg[modelsKey]) && cfg[modelsKey].length
       ? cfg[modelsKey]
       : (d.models || []),
@@ -124,7 +127,8 @@ export function createAccountStore({ credentials, getConfig, fetchImpl, fetchFor
     }
     return {
       ...base,
-      cooldownUntil: cooldowns.get(ref) || 0,
+      cooldownUntil: cooldowns.get(ref) ? cooldowns.get(ref).until : 0,
+      cooldownFamilies: cooldowns.get(ref) && cooldowns.get(ref).families ? cooldowns.get(ref).families : null,
       quota: quotas.get(ref) || null,
       usage: windows.get(ref) || base.usage || null,
       health: health.get(ref) || null,
@@ -272,7 +276,15 @@ export function createAccountStore({ credentials, getConfig, fetchImpl, fetchFor
     resolveRaw,
     ensureFresh,
     refreshUsage,
-    rememberCooldown(ref, until) { cooldowns.set(ref, until) },
+    // #87: family-scoped cooldowns. families=null blocks the whole account
+    // (legacy behavior); a list blocks only those model families.
+    rememberCooldown(ref, until, families) {
+      if (!families || !families.length) { cooldowns.set(ref, { until: Number(until) || 0, families: null }); return }
+      const prev = cooldowns.get(ref)
+      const merged = new Set(families)
+      if (prev && Array.isArray(prev.families)) for (const f of prev.families) merged.add(f)
+      cooldowns.set(ref, { until: Math.max(Number(until) || 0, prev ? prev.until : 0), families: Array.from(merged) })
+    },
     rememberQuota(ref, snap) { if (snap) quotas.set(ref, snap); else quotas.delete(ref) },
     shouldSkipRefresh(ref, now, retryMs) { const f = refreshFailures.get(ref); if (!f) return false; return (Number(f.at) + Number(retryMs)) > Number(now) },
     getQuota(ref) { return quotas.get(ref) || null },

package/lib/adapter.js

@@ -5,6 +5,11 @@ import { modelCatalog } from './messages.js'
 import { streamWithRotation } from './stream-rotate.js'
 import { pickFetch } from './proxy.js'
 
+// #94: heuristic for test/preview/beta/legacy model ids.
+function isDeprecatedId(id) {
+  return /(^|[-_:])(test|preview|dev|alpha|beta|snapshot|experimental|legacy|deprecated)([-_:]|$)/i.test(id)
+}
+
 function asLlmError(err) {
   if (err instanceof LlmError) return err
   const code = (err && err.code) || 'VENDOR'
@@ -37,16 +42,22 @@ export class SubscriptionAdapter extends LlmAdapter {
     if (!accounts.some((a) => a.hasToken)) return []
     const cfg = this.deps.vendorConfig(provider)
     const fetchImpl = this.deps.fetchImpl || fetch
+    let models = null
     try {
       const blob = await this.deps.ensureFresh(
         provider,
         await this.deps.loadBlob(accounts.find((a) => a.hasToken).ref),
         accounts.find((a) => a.hasToken).ref,
       )
-      const models = await getVendor(provider).listModels(blob, cfg, fetchImpl)
-      if (Array.isArray(models) && models.length) return models
+      const live = await getVendor(provider).listModels(blob, cfg, fetchImpl)
+      if (Array.isArray(live) && live.length) models = live
     } catch { /* use built-in catalog */ }
-    return modelCatalog(provider, cfg.models)
+    if (!models) models = modelCatalog(provider, cfg.models)
+    // #94: optionally hide test/preview/beta/legacy ids from the picker.
+    if (typeof this.deps.hideDeprecatedModels === 'function' && this.deps.hideDeprecatedModels()) {
+      models = models.filter((m) => !isDeprecatedId(String((m && m.id) || '')))
+    }
+    return models
   }
 
   async resolveModel(provider, model, _signal) {
@@ -69,18 +80,19 @@ export class SubscriptionAdapter extends LlmAdapter {
   async *stream(options) {
     const provider = options.provider
     const deps = this.deps
+    let yieldedAny = false
     try {
       if (typeof deps.refreshUsage === 'function') {
         await deps.refreshUsage(provider)
       }
-      yield* streamWithRotation({
+      for await (const event of streamWithRotation({
         accounts: await deps.listAccounts(provider),
         nowMs: () => Date.now(),
         cooldownMs: deps.cooldownMs(),
         switchAtRemaining: typeof deps.switchAtRemaining === 'function' ? deps.switchAtRemaining() : (deps.switchAtRemaining ?? 0),
         options,
         onCooldown: (account) => {
-          deps.rememberCooldown(account.ref, account.cooldownUntil)
+          deps.rememberCooldown(account.ref, account.cooldownUntil, account.cooldownFamilies || null)
           if (typeof deps.recordSwitch === 'function') deps.recordSwitch(account.ref)
         },
         streamOnce: async function* (account, opts) {
@@ -148,8 +160,15 @@ export class SubscriptionAdapter extends LlmAdapter {
             throw err
           }
         },
-      })
+      })) { yieldedAny = true; yield event }
+      return
     } catch (err) {
+      // #91: when the whole pool is exhausted and nothing has been streamed
+      // yet, continue seamlessly on local Ollama instead of failing the chat.
+      if (!yieldedAny && typeof deps.ollamaFallback === 'function') {
+        yield* deps.ollamaFallback({ options, provider, err })
+        return
+      }
       throw asLlmError(err)
     }
   }

package/lib/client.js

@@ -788,7 +788,7 @@ function PluginCard(props) {
           expiryNotifyDays: (data && data.expiryNotifyDays) || 7,
           active: (data && data.active) || null,
         }
-      } catch { return { logged: [], usage: 0, expiresAt: {}, labels: {}, expiryNotifyDays: 7, active: null } }
+      } catch { return { logged: [], usage: 0, expiresAt: {}, labels: {}, expiryNotifyDays: 7, fastMode: false, active: null } }
     }
 
     // #52: компактный чип в шапке сессии со статусом подключённых подписок.
@@ -833,7 +833,7 @@ function PluginCard(props) {
         const idx = a.index != null ? ('#' + a.index) : ''
         const plan = a.plan ? (' ' + a.plan) : ''
         // #77: без provider-префикса
-        const label = idx + plan
+        const label = (a.provider === 'codex' && state.fastMode ? '\u26a1' : '') + idx + plan
         // windows: "5h 12%  7d 8%"
         const winLabels = wins.map((w) => (w.label || w.id) + ' ' + Math.round(w.usedPercent) + '%').join('  ')
         return React.createElement('span', {

package/lib/index.js

@@ -15,6 +15,7 @@ import { encryptWithPassphrase, decryptWithPassphrase } from './crypto.js'
 import { quotaSnapshot } from './ratelimit.js'
 import { createAccountStore, normalizeSlots, vendorConfig } from './accounts.js'
 import { startLoopback } from './loopback.js'
+import { OllamaAdapter, ollamaAlive, ollamaModels, ollamaBase } from './ollama.js'
 import { maskEmail, maskLabel, maskText } from './mask.js'
 import { proxyFetch, pickFetch } from './proxy.js'
 import { HistoryStore } from './history.js'
@@ -71,7 +72,19 @@ export const Config = z.object({
     .description('When on, redirect_uri is this Web UI origin + /dsh-subscriptions/oauth/callback. When off, the vendor CLI registered redirect is used and you paste the redirected URL.'),
   autoLoopback: z.boolean().default(true)
     .description('#89 When on and the vendor redirect_uri is a loopback address (codex :1455, grok :56121), a temporary local server catches the OAuth callback automatically - no paste needed. Paste fallback stays available.'),
+  ollamaBaseUrl: z.string().default('http://127.0.0.1:11434')
+    .description('#91 Local Ollama base URL. Served as the ollama provider in the native model picker when reachable.'),
+  ollamaFallback: z.boolean().default(true)
+    .description('#91 When every account of a provider is exhausted, continue the chat on local Ollama instead of failing.'),
+  ollamaFallbackModel: z.string().default('')
+    .description('#91 Ollama model used for the fallback (for example qwen2.5-coder). Empty = first model from /api/tags.'),
+  hideDeprecatedModels: z.boolean().default(false)
+    .description('#94 Hide test/preview/beta/legacy model ids from the native model picker.'),
   codexClientId: z.string().default(''),
+  codexVerbosity: z.string().default('')
+    .description('#93 Response verbosity for Codex reasoning models: low, medium or high. Empty = protocol default.'),
+  codexFastMode: z.boolean().default(false)
+    .description('#92 Fast Mode for Codex: sends service_tier priority (1.5x speed billing tier) with every request.'),
   codexRedirectUri: z.string().default(''),
   codexBaseUrl: z.string().default(''),
   claudeClientId: z.string().default(''),
@@ -165,6 +178,44 @@ export function apply(ctx, config) {
   const history = new HistoryStore()
   const recordHistory = (entry) => history.add(entry)
 
+  // #91: local Ollama - native provider + seamless fallback when the whole
+  // pool is exhausted and nothing has been streamed yet.
+  const ollamaAdapter = new OllamaAdapter({
+    baseUrl: () => ollamaBase(live()),
+    fallbackModel: () => live().ollamaFallbackModel || '',
+  })
+  let ollamaHandle
+  async function syncOllama() {
+    const cfg = live()
+    const alive = !!cfg.ollamaFallback && await ollamaAlive(ollamaBase(cfg), fetch)
+    if (alive && !ollamaHandle) {
+      try { ollamaHandle = ctx.llm.registerAdapter(['ollama'], ollamaAdapter) } catch { /* already registered elsewhere */ }
+    } else if (!alive && ollamaHandle) {
+      try { ollamaHandle() } catch { /* already gone */ }
+      ollamaHandle = undefined
+    }
+  }
+  async function* ollamaFallbackStream({ options, provider, err }) {
+    const cfg = live()
+    const models = await ollamaModels(ollamaBase(cfg), fetch).catch(() => [])
+    if (!cfg.ollamaFallback || !models.length) throw err
+    const model = cfg.ollamaFallbackModel || models[0].id
+    try { ctx.emit && ctx.emit('subscriptions.ollama-fallback', { provider, model, reason: err && err.code || 'EXHAUSTED' }) } catch {}
+    try { ctx.log && ctx.log.warn && ctx.log.warn(`[dsh-subscriptions] ${provider}: все аккаунты исчерпаны (${err && err.code || 'EXHAUSTED'}), откат на ollama/${model}`) } catch {}
+    try {
+      recordHistory({
+        provider: 'ollama',
+        ref: 'OLLAMA_FALLBACK',
+        model,
+        path: '/v1/chat/completions',
+        method: 'POST',
+        status: 200,
+        kind: 'fallback',
+      })
+    } catch {}
+    yield* ollamaAdapter.stream({ ...options, provider: 'ollama', model })
+  }
+
   const subscriptions = createSubscriptionsService({
     listAccounts: (provider) => store.listAccounts(provider),
     loadBlob: (ref) => store.loadBlob(ref),
@@ -172,7 +223,7 @@ export function apply(ctx, config) {
     vendorConfig: (provider) => vendorConfig(provider, live()),
     cooldownMs: () => live().cooldownMs,
     switchAtRemaining: () => live().switchAtRemaining,
-    rememberCooldown: (ref, until) => store.rememberCooldown(ref, until),
+    rememberCooldown: (ref, until, families) => store.rememberCooldown(ref, until, families),
     recordSuccess: (ref) => store.recordSuccess(ref),
     getHealth: (ref) => store.getHealth(ref),
     recordSwitch: (ref) => store.recordSwitch(ref),
@@ -185,6 +236,8 @@ export function apply(ctx, config) {
     recordHistory,
     fetchImpl: fetch,
     fetchForRef,
+    ollamaFallback: ollamaFallbackStream,
+    hideDeprecatedModels: () => !!live().hideDeprecatedModels,
   })
 
   // Служба генерации картинок на подписке.
@@ -235,7 +288,7 @@ export function apply(ctx, config) {
     vendorConfig: (provider) => vendorConfig(provider, live()),
     cooldownMs: () => live().cooldownMs,
     switchAtRemaining: () => live().switchAtRemaining,
-    rememberCooldown: (ref, until) => store.rememberCooldown(ref, until),
+    rememberCooldown: (ref, until, families) => store.rememberCooldown(ref, until, families),
     rememberQuota: (ref, snap) => store.rememberQuota(ref, snap),
     getQuota: (ref) => store.getQuota(ref),
     refreshUsage: (provider) => store.refreshUsage(provider),
@@ -470,6 +523,7 @@ export function apply(ctx, config) {
 
   ctx.effect(() => {
     syncAdapter().catch(() => { /* first paint */ })
+    syncOllama().catch(() => { /* first paint */ })
     // #75: eager refresh usage на старте, чтобы windows (5h/7d) появились в blob сразу
     // и активная подписка в чипе сразу показывала 5h/7d/..., а не ждала probeInterval.
     const eager = async () => {
@@ -560,7 +614,7 @@ export function apply(ctx, config) {
       await tick()
     }
     wrapped().catch(() => {})
-    const timer = setInterval(() => { wrapped().catch(() => {}) }, 60 * 1000)
+    const timer = setInterval(() => { wrapped().catch(() => {}); syncOllama().catch(() => {}) }, 60 * 1000)
     return () => clearInterval(timer)
   }, 'dsh-subscriptions: probe loop')
 
@@ -676,6 +730,7 @@ export function apply(ctx, config) {
         expiresAt: expires,
         labels,
         expiryNotifyDays: live().expiryNotifyDays,
+        fastMode: !!live().codexFastMode,
         active,
       })
     },

package/lib/messages.js

@@ -65,7 +65,7 @@ export function openaiTools(options) {
   }))
 }
 
-export function codexResponsesBody(options, fallbackInstructions) {
+export function codexResponsesBody(options, fallbackInstructions, vendorCfg) {
   const systemParts = []
   if (options.system) systemParts.push(options.system)
   const input = []
@@ -134,6 +134,13 @@ export function codexResponsesBody(options, fallbackInstructions) {
     instructions,
     input,
     ...(responsesTools && responsesTools.length ? { tools: responsesTools } : {}),
+    // #93: reasoning effort chosen in the native picker flows to the protocol.
+    ...(options.reasoningEffort ? { reasoning: { effort: String(options.reasoningEffort) } } : {}),
+    // #93: verbosity comes from the codexVerbosity setting (low/medium/high).
+    ...(vendorCfg && /^(low|medium|high)$/.test(String(vendorCfg.verbosity || ''))
+      ? { text: { verbosity: String(vendorCfg.verbosity) } } : {}),
+    // #92: Fast Mode = 1.5x speed billing tier on the Codex backend.
+    ...(vendorCfg && vendorCfg.fastMode ? { service_tier: 'priority' } : {}),
     ...(options.maxTokens != null ? { max_output_tokens: options.maxTokens } : {}),
     ...(options.temperature != null ? { temperature: options.temperature } : {}),
   }

package/lib/ollama.js

@@ -0,0 +1,70 @@
+import { LlmAdapter, LlmError, attributionHeaders } from '@deepseek-ai/dsh-llm'
+import { openaiMessages, openaiTools } from './messages.js'
+import { openaiChatStream } from './wire.js'
+
+// #91: local Ollama gateway - a $0 emergency provider and seamless quota
+// fallback. Speaks the OpenAI-compatible subset Ollama serves at /v1.
+
+export function ollamaBase(cfg) {
+  return String(cfg.ollamaBaseUrl || 'http://127.0.0.1:11434').replace(/\/$/, '')
+}
+
+export async function ollamaAlive(baseUrl, fetchImpl, timeoutMs = 2000) {
+  try {
+    const res = await fetchImpl(baseUrl + '/api/tags', { signal: AbortSignal.timeout(timeoutMs) })
+    return !!res && res.ok
+  } catch { return false }
+}
+
+export async function ollamaModels(baseUrl, fetchImpl) {
+  const res = await fetchImpl(baseUrl + '/api/tags', { signal: AbortSignal.timeout(3000) })
+  if (!res || !res.ok) throw new LlmError('ollama /api/tags http ' + (res && res.status), 'VENDOR', { status: res && res.status })
+  const j = await res.json()
+  return (Array.isArray(j.models) ? j.models : []).map((m) => ({
+    id: m.name,
+    name: m.name,
+    ...(m.details && m.details.parameter_size ? { description: m.details.parameter_size } : {}),
+  }))
+}
+
+export class OllamaAdapter extends LlmAdapter {
+  constructor(deps) {
+    super()
+    this.deps = deps
+  }
+
+  providerInfo(provider) {
+    return { id: provider, name: 'Ollama (local)' }
+  }
+
+  providerRetryPolicy(_provider) {
+    return undefined
+  }
+
+  async listModels(provider) {
+    return ollamaModels(this.deps.baseUrl(), this.deps.fetchImpl || fetch)
+  }
+
+  async resolveModel(provider, model, _signal) {
+    return { provider, id: model, name: model }
+  }
+
+  async *stream(options) {
+    const base = this.deps.baseUrl()
+    const model = options.model || this.deps.fallbackModel() || ''
+    const body = { model, messages: openaiMessages(options), stream: true }
+    const tools = openaiTools(options)
+    if (tools) body.tools = tools
+    const res = await (this.deps.fetchImpl || fetch)(base + '/v1/chat/completions', {
+      method: 'POST',
+      headers: { 'Content-Type': 'application/json', ...attributionHeaders() },
+      body: JSON.stringify(body),
+      signal: options.signal,
+    })
+    if (!res || !res.ok) {
+      const txt = res && res.text ? await res.text().catch(() => '') : ''
+      throw new LlmError('ollama http ' + (res && res.status) + (txt ? ': ' + txt.slice(0, 200) : ''), 'VENDOR', { status: res && res.status })
+    }
+    yield* openaiChatStream(res.body)
+  }
+}

package/lib/rotate.js

@@ -12,6 +12,25 @@ export function isSwitchableError(err) {
   return status === 429
 }
 
+// #87: family classification. Cooldowns are scoped to the model family that
+// hit the limit, so a reasoning 429 does not block standard models.
+export function modelFamily(provider, model) {
+  const id = String(model || '')
+  if (provider === 'claude') return /thinking/i.test(id) ? 'reasoning' : 'standard'
+  if (provider === 'grok') return /reasoning/i.test(id) ? 'reasoning' : 'standard'
+  if (provider === 'codex') return 'reasoning'
+  return 'standard'
+}
+
+// Legacy cooldowns (no family list) block everything; scoped ones block only
+// their own family. An unknown family never blocks a differently-scoped cooldown.
+export function cooldownBlocks(acc, family) {
+  const fams = acc && acc.cooldownFamilies
+  if (!Array.isArray(fams) || !fams.length) return true
+  if (!family) return true
+  return fams.includes(family)
+}
+
 export function pickAccount(accounts, nowMs, opts) {
   const list = Array.isArray(accounts) ? accounts : []
   const now = Number(nowMs) || 0
@@ -47,7 +66,7 @@ export function pickAccount(accounts, nowMs, opts) {
   const fallback = []
   for (const acc of list) {
     if (!acc || !acc.hasToken) continue
-    const isCooldown = acc.cooldownUntil && Number(acc.cooldownUntil) > now
+    const isCooldown = acc.cooldownUntil && Number(acc.cooldownUntil) > now && cooldownBlocks(acc, opts && opts.family)
     const exhausted = isQuotaExhausted(acc) || isUsageExhausted(acc)
     if (isCooldown || exhausted) {
       tiers[2].push(acc)
@@ -68,8 +87,14 @@ export function pickAccount(accounts, nowMs, opts) {
   return null
 }
 
-export function markCooldown(account, nowMs, cooldownMs) {
+export function markCooldown(account, nowMs, cooldownMs, family) {
   const wait = Number(cooldownMs)
   const ms = Number.isFinite(wait) && wait > 0 ? wait : 30 * 60 * 1000
-  return { ...account, cooldownUntil: (Number(nowMs) || 0) + ms }
+  const prev = Array.isArray(account.cooldownFamilies) ? account.cooldownFamilies.slice() : []
+  const fams = family ? (prev.includes(family) ? prev : prev.concat(family)) : prev
+  return {
+    ...account,
+    cooldownUntil: (Number(nowMs) || 0) + ms,
+    ...(fams.length ? { cooldownFamilies: fams } : {}),
+  }
 }

package/lib/stream-rotate.js

@@ -1,4 +1,4 @@
-import { pickAccount, markCooldown, isSwitchableError } from './rotate.js'
+import { pickAccount, markCooldown, isSwitchableError, modelFamily } from './rotate.js'
 
 export async function* streamWithRotation({
   accounts,
@@ -13,7 +13,7 @@ export async function* streamWithRotation({
   let lastError = null
   const tried = new Set()
   while (true) {
-    const account = pickAccount(pool, nowMs(), { switchAtRemaining })
+    const account = pickAccount(pool, nowMs(), { switchAtRemaining, family: modelFamily(options && options.provider, options && options.model) })
     if (!account) {
       if (lastError) throw lastError
       const err = new Error('no usable subscription account for this provider')
@@ -33,8 +33,9 @@ export async function* streamWithRotation({
     } catch (err) {
       lastError = err
       if (!isSwitchableError(err)) throw err
-      const cooled = markCooldown(account, nowMs(), cooldownMs)
+      const cooled = markCooldown(account, nowMs(), cooldownMs, modelFamily(options && options.provider, options && options.model))
       account.cooldownUntil = cooled.cooldownUntil
+      if (cooled.cooldownFamilies) account.cooldownFamilies = cooled.cooldownFamilies
       if (onCooldown) onCooldown(account)
     }
   }

package/lib/subscriptions.js

@@ -1,5 +1,5 @@
 import { isProvider } from "./refs.js"
-import { pickAccount, markCooldown, isSwitchableError } from "./rotate.js"
+import { pickAccount, markCooldown, isSwitchableError, modelFamily } from "./rotate.js"
 import { quotaSnapshot } from "./ratelimit.js"
 import { getVendor } from "./vendors/index.js"
 import { pickFetch } from "./proxy.js"
@@ -92,7 +92,7 @@ export function createSubscriptionsService(deps) {
       const tried = new Set()
       let lastError = null
       while (true) {
-        const account = pickAccount(pool, Date.now(), { switchAtRemaining: thr })
+        const account = pickAccount(pool, Date.now(), { switchAtRemaining: thr, family: modelFamily(provider, body && body.model) })
         if (!account) {
           if (lastError) throw lastError
           const err = new Error("no usable subscription account for this provider")
@@ -180,9 +180,9 @@ export function createSubscriptionsService(deps) {
         } catch (err) {
           lastError = err
           if (!isSwitchableError(err)) throw err
-          const cooled = markCooldown(account, Date.now(), cooldownMs)
+          const cooled = markCooldown(account, Date.now(), cooldownMs, modelFamily(provider, body && body.model))
           account.cooldownUntil = cooled.cooldownUntil
-          if (typeof deps.rememberCooldown === "function") deps.rememberCooldown(account.ref, account.cooldownUntil)
+          if (typeof deps.rememberCooldown === "function") deps.rememberCooldown(account.ref, account.cooldownUntil, cooled.cooldownFamilies || null)
           // try next account
         }
       }

package/lib/vendors/codex.js

@@ -205,7 +205,7 @@ export async function usage(blob, cfg, fetchImpl) {
 
 export async function* streamOnce({ blob, options, fetchImpl, headers, config, signal }) {
   const base = (config.baseUrl || defaults().baseUrl).replace(/\/$/, '')
-  const body = codexResponsesBody(options, INSTRUCTIONS)
+  const body = codexResponsesBody(options, INSTRUCTIONS, config)
   const res = await fetchImpl(`${base}/responses`, {
     method: 'POST',
     headers: {

package/lib/vendors/grok.js

@@ -213,6 +213,8 @@ export async function* streamOnce({ blob, options, fetchImpl, headers, config, s
   const base = (config.baseUrl || defaults().baseUrl).replace(/\/$/, '')
   let body = codexResponsesBody(options, '')
   if (!body.instructions) delete body.instructions
+  // grok manages reasoning itself (cli catalog aware); drop the codex-level field
+  delete body.reasoning
   const catalog = await cliCatalogCached(blob, config, fetchImpl).catch(() => new Map())
   const reasoning = grokReasoningBody(options.model, options.reasoningEffort, catalog)
   if (reasoning) body.reasoning = reasoning

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@goodandready/dsh-subscriptions",
-  "version": "0.4.16",
+  "version": "0.4.17",
   "description": "Use ChatGPT Codex, Claude, Grok, and Antigravity subscriptions as DeepSeek Harness LLM providers via OAuth.",
   "license": "MIT",
   "type": "module",
@@ -63,5 +63,8 @@
   "dependencies": {
     "socks": "^2.8.9",
     "undici": "^8.10.1"
+  },
+  "devDependencies": {
+    "@deepseek-ai/dsh-llm": "0.1.0-rc.8"
   }
 }