@goodandready/dsh-subscriptions 0.7.33 → 0.7.34

package/lib/accounts.js

@@ -130,6 +130,7 @@ export function createAccountStore({ credentials, getConfig, fetchImpl, fetchFor
   const notifyThresholds = new Map()
   const health = new Map()
   const requestCounts = new Map()
+  const lastUsedAt = new Map()
   const windows = new Map()
   const usageFetched = new Map()
   const pings = new Map()
@@ -278,6 +279,7 @@ export function createAccountStore({ credentials, getConfig, fetchImpl, fetchFor
     // Counters and health must not outlive the account, otherwise a
     // re-added slot inherits stale rotation statistics.
     requestCounts.delete(ref)
+    lastUsedAt.delete(ref)
     health.delete(ref)
     windows.delete(ref)
   }
@@ -427,6 +429,7 @@ export function createAccountStore({ credentials, getConfig, fetchImpl, fetchFor
         isDefault: !!slot.isDefault,
         ...(Array.isArray(slot.poolModels) ? { poolModels: slot.poolModels } : {}),
         credentialError: info.credentialError || null,
+        lastUsedAt: lastUsedAt.get(slot.ref) || 0,
         duplicateOf,
       })
     }
@@ -679,6 +682,7 @@ export function createAccountStore({ credentials, getConfig, fetchImpl, fetchFor
     recordExhaust(ref) { health.set(ref, recordExhaust(health.get(ref))) },
     recordBroken(ref) { health.set(ref, recordBroken(health.get(ref))) },
     recordSuccess(ref) {
+      if (ref) lastUsedAt.set(ref, Date.now())
       // successful request clears the cooldown and puts the account back in rotation
       cooldowns.delete(ref)
       quarantines.delete(ref)
@@ -739,7 +743,10 @@ export function createAccountStore({ credentials, getConfig, fetchImpl, fetchFor
       quarantines.delete(ref)
       saveQuarantinesToFile(quarantines)
     },
-    rememberRequest(ref) { requestCounts.set(ref, (requestCounts.get(ref) || 0) + 1) },
+    rememberRequest(ref) {
+      requestCounts.set(ref, (requestCounts.get(ref) || 0) + 1)
+      if (ref) lastUsedAt.set(ref, Date.now())
+    },
     getRequestCount(ref) { return requestCounts.get(ref) || 0 },
     rememberPing(ref, { latencyMs, latencyGrade }) { pings.set(ref, { latencyMs, latencyGrade, at: Date.now() }) },
     getPing(ref) { return pings.get(ref) || null },

package/lib/adapter-manager.js

@@ -48,6 +48,9 @@ export function createAdapterManager({
     vendorConfig: (provider) => vendorConfig(provider, live()),
     cooldownMs: () => live().cooldownMs,
     autoPacing: () => live().autoPacing,
+    promptCacheTtlMs: () => Number(live().promptCacheTtlMinutes || 0) * 60 * 1000,
+    proactiveGuard: () => Boolean(live().proactiveGuard),
+    proactiveMinRemaining: () => Number(live().proactiveMinRemaining || 1),
     drainBeforeReset: () => live().drainBeforeReset,
     drainSafetyPercent: () => live().drainSafetyPercent,
     switchAtRemaining: () => live().switchAtRemaining,
@@ -273,6 +276,26 @@ export function createAdapterManager({
     }
   }
 
+  async function selfHealCoolingAccounts() {
+    const providers = await store.loggedInProviders()
+    for (const provider of providers) {
+      const accounts = await store.listAccounts(provider).catch(() => [])
+      for (const account of accounts) {
+        if (!account || !account.cooldownUntil || Number(account.cooldownUntil) <= Date.now()) continue
+        try {
+          const blob = await store.loadBlob(account.ref)
+          if (!blob) continue
+          const vendor = getVendor(provider)
+          const cfg = vendorConfig(provider, live())
+          if (typeof vendor.check === 'function') await vendor.check(blob, cfg, fetch)
+          else if (typeof vendor.listModels === 'function') await vendor.listModels(blob, cfg, fetch)
+          else continue
+          store.clearCooldown(account.ref)
+        } catch { /* keep cooling on probe failure */ }
+      }
+    }
+  }
+
   function registerAdapterLifecycle() {
     return ctx.effect(() => {
       syncAdapter().catch((err) => {
@@ -333,6 +356,9 @@ export function createAdapterManager({
       }
       eager()
 
+      // #820: plugin network calls use a per-address connect floor.
+      process.env.DSH_CONNECT_ATTEMPT_TIMEOUT_MS = String(Number(live().connectAttemptTimeoutMs) || 1500)
+
       // #799: refresh at startup and hourly. Per-ref snapshots younger than
       // five minutes are skipped, and failures keep last-known-good.
       refreshModels().catch((err) => {
@@ -345,10 +371,22 @@ export function createAdapterManager({
       }, 60 * 60 * 1000)
       if (modelTimer && typeof modelTimer.unref === 'function') modelTimer.unref()
 
+      let selfHealTimer
+      const selfHealMin = Number(live().selfHealIntervalMin) || 0
+      if (selfHealMin > 0) {
+        selfHealTimer = setInterval(() => {
+          selfHealCoolingAccounts().catch((err) => {
+            if (logger && logger.debug) logger.debug('[dsh-subscriptions] self-heal error: ' + (err && err.message || err))
+          })
+        }, selfHealMin * 60 * 1000)
+        if (typeof selfHealTimer.unref === 'function') selfHealTimer.unref()
+      }
+
       return () => {
         adapterDisposed = true
         if (retryTimer) clearTimeout(retryTimer)
         clearInterval(modelTimer)
+        if (selfHealTimer) clearInterval(selfHealTimer)
         if (handle) {
           disposeAdapterHandle(handle)
           handle = undefined

package/lib/adapter.js

@@ -1,5 +1,6 @@
 import { resolveEffectiveReasoning, getModelReasoningCapabilities } from './reasoning-effort.js'
 import { inferModelCapabilities, isDeprecatedId, mergeAccountCatalogs, resolveCatalogSource } from './models-catalog.js'
+import { replayStateKey, unwrapReplayState, wrapReplayState } from './routing-reliability.js'
 import { fetchWithTimeout } from "./http.js"
 import { LlmAdapter, LlmError, attributionHeaders } from '@deepseek-ai/dsh-llm'
 import { displayName } from './refs.js'
@@ -206,6 +207,9 @@ export class SubscriptionAdapter extends LlmAdapter {
         ...(!poolEnabled ? { onlyRef: defaultRef || '__no_default_account__' } : {}),
         ...(defaultRef ? { defaultRef } : {}),
         _decisionTrace: decisionTrace,
+        promptCacheTtlMs: typeof deps.promptCacheTtlMs === 'function' ? deps.promptCacheTtlMs() : 0,
+        proactiveGuard: typeof deps.proactiveGuard === 'function' ? deps.proactiveGuard() : false,
+        proactiveMinRemaining: typeof deps.proactiveMinRemaining === 'function' ? deps.proactiveMinRemaining() : 1,
         autoPacing: effectiveAutoPacing,
         drainBeforeReset: effectiveDrainBeforeReset,
         drainSafetyPercent: effectiveDrainSafetyPercent,
@@ -390,10 +394,22 @@ export class SubscriptionAdapter extends LlmAdapter {
 
             blob = await deps.ensureFresh(provider, await deps.loadBlob(account.ref), account.ref)
             if (typeof deps.rememberRequest === 'function') deps.rememberRequest(account.ref)
+            const replayKey = replayStateKey(provider, account.ref, opts && opts.model)
+            const safeOptions = {
+              ...opts,
+              messages: (Array.isArray(opts.messages) ? opts.messages : []).map((message) => {
+                if (!message || !message.replayState) return message
+                const replay = unwrapReplayState(message.replayState, replayKey)
+                if (replay.matched) return { ...message, replayState: replay.state }
+                const copy = { ...message }
+                delete copy.replayState
+                return copy
+              }),
+            }
             for await (const chunk of vendor.streamOnce({
               blob,
               accountId: account.id || account.ref,
-              options: opts,
+              options: safeOptions,
               fetchImpl: quotaFetch,
               headers: attributionHeaders(),
               config: deps.vendorConfig(provider),
@@ -407,6 +423,7 @@ export class SubscriptionAdapter extends LlmAdapter {
                 if (chunk.type === 'finish') {
                   sawFinish = true
                   finishReason = chunk.reason
+                  if (chunk.replayState) chunk.replayState = wrapReplayState(chunk.replayState, replayKey)
                 } else if (chunk.type === 'usage' && chunk.usage) {
                   const u = toTokenUsage(chunk.usage)
                   if (u && typeof u.outputTokens === 'number' && u.outputTokens > 0) {

package/lib/client.js

@@ -80,7 +80,7 @@ window.__ModuleLoader__.load({
     
     const en = {
       "syncModels": "Sync Models",
-      "syncModelsTitle": "Discover and synchronize available models from vendor API",
+      "syncModelsTitle": "Discover and synchronize available models from provider API",
       "modelsSynced": "Models synchronized",
       "recommendation": "Recommendation",
       "notConnected": "Not connected",
@@ -89,9 +89,9 @@ window.__ModuleLoader__.load({
       "usageFull": "Usage 100%",
       "connected": "Connected",
       "title": "Subscriptions",
-      "intro": "Log in with a consumer subscription. Tokens stay on the host credentials store. The browser never reads them back. After Connect, if the provider lands on localhost or a vendor page, paste the redirected URL or the code here.",
+      "intro": "Log in with a consumer subscription. Tokens stay on the host credentials store. The browser never reads them back. After Connect, if the provider lands on localhost or a provider page, paste the redirected URL or the code here.",
       "useOrigin": "Use this Web UI origin as OAuth redirect_uri",
-      "useOriginHint": "Leave off unless you registered your own OAuth client for this origin. Vendor CLI clients usually require their published redirect URI plus the paste step.",
+      "useOriginHint": "Leave off unless you registered your own OAuth client for this origin. Provider CLI clients usually require their published redirect URI plus the paste step.",
       "closeModal": "Close",
       "groupAuth": "Sign-in and redirect",
       "groupPrivacy": "Privacy and model list",
@@ -152,7 +152,7 @@ window.__ModuleLoader__.load({
       "smokeFail": "Smoke test failed",
       "testConnection": "Test",
       "testConnectionSuccess": "Connection verified",
-      "cascadingFallback": "Cross-vendor fallback",
+      "cascadingFallback": "Cross-provider fallback",
       "cascadingFallbackHint": "Automatically route queries to alternative providers when all accounts of a primary provider are exhausted.",
       "cascadingFallbackOrder": "Fallback Provider Order",
       "cascadingFallbackOrderHint": "Sequence of alternative providers to try when primary accounts are exhausted.",

package/lib/config-schema.js

@@ -181,6 +181,16 @@ export const Config = z.object({
     .description('#510 In-app toast alerts for quarantine and quota threshold events.'),
   autoPacing: z.boolean().volatile().default(true)
     .description('#352 Predictive burn-rate load balancing between accounts.'),
+  promptCacheTtlMinutes: z.number().volatile().default(5)
+    .description('#810 Prompt-cache affinity TTL for session scoring. 0 disables cache affinity.'),
+  proactiveGuard: z.boolean().volatile().default(false)
+    .description('#792 Block an account before a request when its remaining quota is at the configured floor.'),
+  proactiveMinRemaining: z.number().volatile().default(1)
+    .description('#792 Remaining-quota floor used by the proactive guard.'),
+  selfHealIntervalMin: z.number().volatile().default(0)
+    .description('#792 Probe cooling accounts and release their cooldown on success. 0 disables.'),
+  connectAttemptTimeoutMs: z.number().volatile().default(1500)
+    .description('#820 Per-address connect timeout floor for plugin network calls; larger values are preserved.'),
   drainBeforeReset: z.boolean().volatile().default(false)
     .description('#485 Prioritize accounts with closest reset window to burn expiring quota first.'),
   drainSafetyPercent: z.number().volatile().default(5)

package/lib/http.js

@@ -1,4 +1,6 @@
 import { timingSafeEqual } from 'node:crypto'
+import { Agent } from 'undici'
+import { effectiveConnectAttemptTimeout } from './routing-reliability.js'
 
 export function writeJson(res, code, body) {
   try {
@@ -341,6 +343,15 @@ export function isTimeoutError(err) {
     || String(err.message || '').toLowerCase().includes('operation was aborted')
 }
 
+const connectDispatchers = new Map()
+function connectDispatcher() {
+  const timeout = effectiveConnectAttemptTimeout(process.env.DSH_CONNECT_ATTEMPT_TIMEOUT_MS)
+  if (!connectDispatchers.has(timeout)) {
+    connectDispatchers.set(timeout, new Agent({ connect: { timeout } }))
+  }
+  return connectDispatchers.get(timeout)
+}
+
 /**
  * @param {any} fetchImpl
  * @param {string} url
@@ -351,6 +362,7 @@ export async function fetchWithTimeout(fetchImpl, url, init = {}, { timeoutMs =
   const safeHeaders = sanitizeRequestHeaders(url, init, { allowedOrigins })
   init = { ...init, headers: safeHeaders }
   const impl = fetchImpl || fetch
+  if (!init.dispatcher && impl === globalThis.fetch) init.dispatcher = connectDispatcher()
   if (!timeoutMs || timeoutMs <= 0) return impl(url, init)
   const controller = new AbortController()
   let timer = setTimeout(() => {

package/lib/rotate.js

@@ -1,7 +1,7 @@
 import { computePacingRisk, calculateBurnRatePerHour, calculateBurnUrgency } from "./forecast.js"
 import { isQuarantined } from './quarantine.js'
 import { getPinnedAccountRef, pinSession, touchSessionPin, unpinSession, isCacheAwareProvider } from './session-pin.js'
-import { accountCatalogHasModel, proactiveQuotaDecision } from './routing-reliability.js'
+import { accountCatalogHasModel, proactiveQuotaDecision, promptCacheAffinity } from './routing-reliability.js'
 
 const SWITCH_CODES = new Set([
   'RATE_LIMIT',
@@ -390,6 +390,13 @@ export function pickAccountWithTrace(accounts, nowMs, opts) {
           if (Math.abs(aUrgency - bUrgency) > 0.01) return bUrgency - aUrgency
         }
 
+        // #810: keep session traffic on a warm prompt-cache account.
+        if (opts && Number(opts.promptCacheTtlMs) > 0) {
+          const aCache = promptCacheAffinity(a, now, opts.promptCacheTtlMs).score
+          const bCache = promptCacheAffinity(b, now, opts.promptCacheTtlMs).score
+          if (Math.abs(aCache - bCache) > 0.01) return bCache - aCache
+        }
+
         // 0. Auto-Pacing: lower pacing risk preferred (#352)
         const aRisk = Number(a.pacingRisk) || 0
         const bRisk = Number(b.pacingRisk) || 0
@@ -449,6 +456,7 @@ export function pickAccountWithTrace(accounts, nowMs, opts) {
             healthScore: chosen.healthScore ?? 100,
             urgency: calculateBurnUrgency(getAccountRemainingPercent(chosen), chosen.quota && chosen.quota.resetAt, now, opts && opts.drainSafetyPercent != null ? Number(opts.drainSafetyPercent) : 5),
             defaultAccount: Boolean(opts && opts.defaultRef && chosenRef === opts.defaultRef),
+            cacheAffinity: opts && Number(opts.promptCacheTtlMs) > 0 ? promptCacheAffinity(chosen, now, opts.promptCacheTtlMs) : null,
           },
         },
       }

package/lib/stream-rotate.js

@@ -54,6 +54,9 @@ export async function* streamWithRotation({
       model: options && options.model,
       onlyRef: options && options.onlyRef,
       defaultRef: options && options.defaultRef,
+      promptCacheTtlMs: options && options.promptCacheTtlMs,
+      proactiveGuard: options && options.proactiveGuard,
+      proactiveMinRemaining: options && options.proactiveMinRemaining,
       family: modelFamily(options && options.provider, options && options.model),
       sessionId: options && options.sessionId,
       tag: options && options.tag,

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@goodandready/dsh-subscriptions",
-  "version": "0.7.33",
+  "version": "0.7.34",
   "description": "Use ChatGPT Codex, Claude, Grok, Antigravity, Kimi, GLM, Cursor, Kiro, Copilot, Qwen, ERNIE, Spark, JetBrains, Perplexity, Replit, and Cody subscriptions as DeepSeek Harness LLM providers via OAuth.",
   "license": "MIT",
   "type": "module",