@juspay/neurolink 9.86.0 → 9.86.1

package/dist/lib/proxy/accountQuota.d.ts

@@ -35,5 +35,10 @@ export declare function loadAccountQuota(accountKey: string): Promise<AccountQuo
  * Update quota for a single account.
  * Updates in-memory cache immediately (non-blocking),
  * then debounces the disk write to every 5 seconds.
+ *
+ * Loads the persisted file into the cache before the first write so a save
+ * after a process restart merges with existing entries instead of rewriting
+ * the file with only the accounts used since boot (which silently erased
+ * other accounts' snapshots and blinded quota-aware routing to them).
  */
 export declare function saveAccountQuota(accountKey: string, quota: AccountQuota): Promise<void>;

package/dist/lib/proxy/accountQuota.js

@@ -174,10 +174,17 @@ export async function loadAccountQuota(accountKey) {
  * Update quota for a single account.
  * Updates in-memory cache immediately (non-blocking),
  * then debounces the disk write to every 5 seconds.
+ *
+ * Loads the persisted file into the cache before the first write so a save
+ * after a process restart merges with existing entries instead of rewriting
+ * the file with only the accounts used since boot (which silently erased
+ * other accounts' snapshots and blinded quota-aware routing to them).
  */
 export async function saveAccountQuota(accountKey, quota) {
+    if (!cacheLoaded) {
+        await loadAccountQuotas();
+    }
     memoryCache[accountKey] = quota;
-    cacheLoaded = true;
     dirty = true;
     scheduleFlush();
 }

package/dist/lib/server/routes/claudeProxyRoutes.d.ts

@@ -44,18 +44,32 @@ declare function resetEpochToMs(resetEpoch: number | undefined, now: number): nu
  * allow a couple of jittered same-account retries, then a short cooldown.
  */
 declare function planCooldownFor429(quota: AccountQuota | null, retryAfterMs: number, now: number): AccountCooldownPlan;
+/**
+ * Seed each account's runtime quota from the persisted snapshots in
+ * ~/.neurolink/account-quotas.json (keyed by label). Runtime state is
+ * in-memory only, so without this the quota-aware ordering is blind after a
+ * proxy restart: all accounts tie, selection falls back to token-store
+ * enumeration order, and the first account served becomes self-reinforcing
+ * (it alone has data) — starving the others regardless of their resets.
+ * Never overwrites fresher in-memory quota; stale disk snapshots degrade
+ * gracefully because past reset timestamps are ignored by resetEpochToMs.
+ */
+declare function seedRuntimeQuotasFromDisk(accounts: ProxyPassthroughAccount[]): Promise<void>;
 /**
  * Order accounts to MAXIMIZE quota utilization (fill-first, smart order):
  * spend the account whose window refreshes SOONEST first, so its about-to-reset
  * allowance isn't wasted, then move to accounts with longer-dated resets.
  *
  * Priority among usable accounts:
- *   1. soonest WEEKLY (7d) reset  — the scarce, use-it-or-lose-it ceiling
- *   2. soonest SESSION (5h) reset
- *   3. highest weekly utilization — finish off the one closest to done
- * Accounts with no quota data yet keep insertion order (stable sort) and sit
- * after those with a known soonest reset. Cooling/rejected accounts sort last,
- * soonest-back-to-service first, as last resort.
+ *   1. no quota data yet — probe first: one request reveals its windows and
+ *      self-corrects the ordering. (Ranking unknowns last would starve them
+ *      forever: never picked → never observed → never comparable.)
+ *   2. soonest WEEKLY (7d) reset  — the scarce, use-it-or-lose-it ceiling
+ *   3. soonest SESSION (5h) reset
+ *   4. highest weekly utilization — finish off the one closest to done
+ *   5. configured primary account, then insertion order
+ * Cooling/rejected accounts sort last, soonest-back-to-service first, as
+ * last resort.
  */
 declare function orderAccountsByQuota(accounts: ProxyPassthroughAccount[], now: number): ProxyPassthroughAccount[];
 /**
@@ -95,6 +109,8 @@ export declare const __testHooks: {
     planCooldownFor429: typeof planCooldownFor429;
     orderAccountsByQuota: typeof orderAccountsByQuota;
     resetEpochToMs: typeof resetEpochToMs;
+    seedRuntimeQuotasFromDisk: typeof seedRuntimeQuotasFromDisk;
+    getAccountRuntimeState: (key: string) => RuntimeAccountState | undefined;
     setConfiguredPrimaryAccountKey: (key: string | undefined) => void;
     getConfiguredPrimaryAccountKey: () => string | undefined;
     setPrimaryAccountIndex: (index: number) => void;

package/dist/lib/server/routes/claudeProxyRoutes.js

@@ -13,7 +13,7 @@ import { access, readFile } from "node:fs/promises";
 import { homedir } from "node:os";
 import { join } from "node:path";
 import { buildStableClaudeCodeBillingHeader, CLAUDE_CLI_USER_AGENT, CLAUDE_CODE_OAUTH_BETAS, getOrCreateClaudeCodeIdentity, parseClaudeCodeUserId, } from "../../auth/anthropicOAuth.js";
-import { parseQuotaHeaders, saveAccountQuota, } from "../../proxy/accountQuota.js";
+import { loadAccountQuotas, parseQuotaHeaders, saveAccountQuota, } from "../../proxy/accountQuota.js";
 import { buildClaudeError, ClaudeStreamSerializer, generateToolUseId, parseClaudeRequest, serializeClaudeResponse, } from "../../proxy/claudeFormat.js";
 import { buildAnthropicModelsListResponse, buildTranslationOptions, extractText, extractToolArgs, extractUsageFromStreamResult, handleTranslatedJsonRequest, handleTranslatedStreamRequest, hasTranslatedOutput, } from "../../proxy/proxyTranslationEngine.js";
 import { tracers } from "../../telemetry/tracers.js";
@@ -242,6 +242,30 @@ function maybeCoolFromQuota(state, quota, now) {
         logger.always(`[proxy] proactively cooling account (${reason}) ~${minutesUntil(clamped, now)}m from success-response quota (status rejected)`);
     }
 }
+/**
+ * Seed each account's runtime quota from the persisted snapshots in
+ * ~/.neurolink/account-quotas.json (keyed by label). Runtime state is
+ * in-memory only, so without this the quota-aware ordering is blind after a
+ * proxy restart: all accounts tie, selection falls back to token-store
+ * enumeration order, and the first account served becomes self-reinforcing
+ * (it alone has data) — starving the others regardless of their resets.
+ * Never overwrites fresher in-memory quota; stale disk snapshots degrade
+ * gracefully because past reset timestamps are ignored by resetEpochToMs.
+ */
+async function seedRuntimeQuotasFromDisk(accounts) {
+    try {
+        const persisted = await loadAccountQuotas();
+        for (const account of accounts) {
+            const state = getOrCreateRuntimeState(account.key);
+            if (!state.quota && persisted[account.label]) {
+                state.quota = persisted[account.label];
+            }
+        }
+    }
+    catch {
+        // Non-fatal: seeding is best-effort; ordering falls back to probe-first.
+    }
+}
 /** Quota-aware selection is on by default; disable with
  *  NEUROLINK_PROXY_QUOTA_ROUTING=off|false|0. Only affects the fill-first
  *  strategy (round-robin keeps strict rotation). */
@@ -260,6 +284,7 @@ function accountSortMetrics(accountKey, now) {
     const sessionRejected = q?.sessionStatus === "rejected" && sessionReset !== undefined;
     return {
         usable: !coolingActive && !weeklyRejected && !sessionRejected,
+        hasQuota: !!q,
         coolingUntil: st?.coolingUntil ?? 0,
         weeklyReset: weeklyReset ?? Number.POSITIVE_INFINITY,
         sessionReset: sessionReset ?? Number.POSITIVE_INFINITY,
@@ -272,14 +297,18 @@ function accountSortMetrics(accountKey, now) {
  * allowance isn't wasted, then move to accounts with longer-dated resets.
  *
  * Priority among usable accounts:
- *   1. soonest WEEKLY (7d) reset  — the scarce, use-it-or-lose-it ceiling
- *   2. soonest SESSION (5h) reset
- *   3. highest weekly utilization — finish off the one closest to done
- * Accounts with no quota data yet keep insertion order (stable sort) and sit
- * after those with a known soonest reset. Cooling/rejected accounts sort last,
- * soonest-back-to-service first, as last resort.
+ *   1. no quota data yet — probe first: one request reveals its windows and
+ *      self-corrects the ordering. (Ranking unknowns last would starve them
+ *      forever: never picked → never observed → never comparable.)
+ *   2. soonest WEEKLY (7d) reset  — the scarce, use-it-or-lose-it ceiling
+ *   3. soonest SESSION (5h) reset
+ *   4. highest weekly utilization — finish off the one closest to done
+ *   5. configured primary account, then insertion order
+ * Cooling/rejected accounts sort last, soonest-back-to-service first, as
+ * last resort.
  */
 function orderAccountsByQuota(accounts, now) {
+    const primaryKey = configuredPrimaryAccountKey;
     return [...accounts].sort((a, b) => {
         const ma = accountSortMetrics(a.key, now);
         const mb = accountSortMetrics(b.key, now);
@@ -291,13 +320,22 @@ function orderAccountsByQuota(accounts, now) {
             const bu = mb.coolingUntil || Number.POSITIVE_INFINITY;
             return au - bu;
         }
+        if (ma.hasQuota !== mb.hasQuota) {
+            return ma.hasQuota ? 1 : -1;
+        }
         if (ma.weeklyReset !== mb.weeklyReset) {
             return ma.weeklyReset - mb.weeklyReset;
         }
         if (ma.sessionReset !== mb.sessionReset) {
             return ma.sessionReset - mb.sessionReset;
         }
-        return mb.weeklyUsed - ma.weeklyUsed;
+        if (ma.weeklyUsed !== mb.weeklyUsed) {
+            return mb.weeklyUsed - ma.weeklyUsed;
+        }
+        if (primaryKey && (a.key === primaryKey) !== (b.key === primaryKey)) {
+            return a.key === primaryKey ? -1 : 1;
+        }
+        return 0;
     });
 }
 // ---------------------------------------------------------------------------
@@ -1280,6 +1318,7 @@ async function loadClaudeProxyAccounts(args) {
         state.lastToken = account.token;
         state.lastRefreshToken = account.refreshToken;
     }
+    await seedRuntimeQuotasFromDisk(accounts);
     const enabledAccounts = accounts.filter((account) => {
         return !getOrCreateRuntimeState(account.key).permanentlyDisabled;
     });
@@ -3777,6 +3816,11 @@ export const __testHooks = {
     planCooldownFor429,
     orderAccountsByQuota,
     resetEpochToMs,
+    seedRuntimeQuotasFromDisk,
+    getAccountRuntimeState: (key) => {
+        const state = accountRuntimeState.get(key);
+        return state ? { ...state } : undefined;
+    },
     setConfiguredPrimaryAccountKey: (key) => {
         configuredPrimaryAccountKey = key;
     },

package/dist/proxy/accountQuota.d.ts

@@ -35,5 +35,10 @@ export declare function loadAccountQuota(accountKey: string): Promise<AccountQuo
  * Update quota for a single account.
  * Updates in-memory cache immediately (non-blocking),
  * then debounces the disk write to every 5 seconds.
+ *
+ * Loads the persisted file into the cache before the first write so a save
+ * after a process restart merges with existing entries instead of rewriting
+ * the file with only the accounts used since boot (which silently erased
+ * other accounts' snapshots and blinded quota-aware routing to them).
  */
 export declare function saveAccountQuota(accountKey: string, quota: AccountQuota): Promise<void>;

package/dist/proxy/accountQuota.js

@@ -174,10 +174,17 @@ export async function loadAccountQuota(accountKey) {
  * Update quota for a single account.
  * Updates in-memory cache immediately (non-blocking),
  * then debounces the disk write to every 5 seconds.
+ *
+ * Loads the persisted file into the cache before the first write so a save
+ * after a process restart merges with existing entries instead of rewriting
+ * the file with only the accounts used since boot (which silently erased
+ * other accounts' snapshots and blinded quota-aware routing to them).
  */
 export async function saveAccountQuota(accountKey, quota) {
+    if (!cacheLoaded) {
+        await loadAccountQuotas();
+    }
     memoryCache[accountKey] = quota;
-    cacheLoaded = true;
     dirty = true;
     scheduleFlush();
 }

package/dist/server/routes/claudeProxyRoutes.d.ts

@@ -44,18 +44,32 @@ declare function resetEpochToMs(resetEpoch: number | undefined, now: number): nu
  * allow a couple of jittered same-account retries, then a short cooldown.
  */
 declare function planCooldownFor429(quota: AccountQuota | null, retryAfterMs: number, now: number): AccountCooldownPlan;
+/**
+ * Seed each account's runtime quota from the persisted snapshots in
+ * ~/.neurolink/account-quotas.json (keyed by label). Runtime state is
+ * in-memory only, so without this the quota-aware ordering is blind after a
+ * proxy restart: all accounts tie, selection falls back to token-store
+ * enumeration order, and the first account served becomes self-reinforcing
+ * (it alone has data) — starving the others regardless of their resets.
+ * Never overwrites fresher in-memory quota; stale disk snapshots degrade
+ * gracefully because past reset timestamps are ignored by resetEpochToMs.
+ */
+declare function seedRuntimeQuotasFromDisk(accounts: ProxyPassthroughAccount[]): Promise<void>;
 /**
  * Order accounts to MAXIMIZE quota utilization (fill-first, smart order):
  * spend the account whose window refreshes SOONEST first, so its about-to-reset
  * allowance isn't wasted, then move to accounts with longer-dated resets.
  *
  * Priority among usable accounts:
- *   1. soonest WEEKLY (7d) reset  — the scarce, use-it-or-lose-it ceiling
- *   2. soonest SESSION (5h) reset
- *   3. highest weekly utilization — finish off the one closest to done
- * Accounts with no quota data yet keep insertion order (stable sort) and sit
- * after those with a known soonest reset. Cooling/rejected accounts sort last,
- * soonest-back-to-service first, as last resort.
+ *   1. no quota data yet — probe first: one request reveals its windows and
+ *      self-corrects the ordering. (Ranking unknowns last would starve them
+ *      forever: never picked → never observed → never comparable.)
+ *   2. soonest WEEKLY (7d) reset  — the scarce, use-it-or-lose-it ceiling
+ *   3. soonest SESSION (5h) reset
+ *   4. highest weekly utilization — finish off the one closest to done
+ *   5. configured primary account, then insertion order
+ * Cooling/rejected accounts sort last, soonest-back-to-service first, as
+ * last resort.
  */
 declare function orderAccountsByQuota(accounts: ProxyPassthroughAccount[], now: number): ProxyPassthroughAccount[];
 /**
@@ -95,6 +109,8 @@ export declare const __testHooks: {
     planCooldownFor429: typeof planCooldownFor429;
     orderAccountsByQuota: typeof orderAccountsByQuota;
     resetEpochToMs: typeof resetEpochToMs;
+    seedRuntimeQuotasFromDisk: typeof seedRuntimeQuotasFromDisk;
+    getAccountRuntimeState: (key: string) => RuntimeAccountState | undefined;
     setConfiguredPrimaryAccountKey: (key: string | undefined) => void;
     getConfiguredPrimaryAccountKey: () => string | undefined;
     setPrimaryAccountIndex: (index: number) => void;

package/dist/server/routes/claudeProxyRoutes.js

@@ -13,7 +13,7 @@ import { access, readFile } from "node:fs/promises";
 import { homedir } from "node:os";
 import { join } from "node:path";
 import { buildStableClaudeCodeBillingHeader, CLAUDE_CLI_USER_AGENT, CLAUDE_CODE_OAUTH_BETAS, getOrCreateClaudeCodeIdentity, parseClaudeCodeUserId, } from "../../auth/anthropicOAuth.js";
-import { parseQuotaHeaders, saveAccountQuota, } from "../../proxy/accountQuota.js";
+import { loadAccountQuotas, parseQuotaHeaders, saveAccountQuota, } from "../../proxy/accountQuota.js";
 import { buildClaudeError, ClaudeStreamSerializer, generateToolUseId, parseClaudeRequest, serializeClaudeResponse, } from "../../proxy/claudeFormat.js";
 import { buildAnthropicModelsListResponse, buildTranslationOptions, extractText, extractToolArgs, extractUsageFromStreamResult, handleTranslatedJsonRequest, handleTranslatedStreamRequest, hasTranslatedOutput, } from "../../proxy/proxyTranslationEngine.js";
 import { tracers } from "../../telemetry/tracers.js";
@@ -242,6 +242,30 @@ function maybeCoolFromQuota(state, quota, now) {
         logger.always(`[proxy] proactively cooling account (${reason}) ~${minutesUntil(clamped, now)}m from success-response quota (status rejected)`);
     }
 }
+/**
+ * Seed each account's runtime quota from the persisted snapshots in
+ * ~/.neurolink/account-quotas.json (keyed by label). Runtime state is
+ * in-memory only, so without this the quota-aware ordering is blind after a
+ * proxy restart: all accounts tie, selection falls back to token-store
+ * enumeration order, and the first account served becomes self-reinforcing
+ * (it alone has data) — starving the others regardless of their resets.
+ * Never overwrites fresher in-memory quota; stale disk snapshots degrade
+ * gracefully because past reset timestamps are ignored by resetEpochToMs.
+ */
+async function seedRuntimeQuotasFromDisk(accounts) {
+    try {
+        const persisted = await loadAccountQuotas();
+        for (const account of accounts) {
+            const state = getOrCreateRuntimeState(account.key);
+            if (!state.quota && persisted[account.label]) {
+                state.quota = persisted[account.label];
+            }
+        }
+    }
+    catch {
+        // Non-fatal: seeding is best-effort; ordering falls back to probe-first.
+    }
+}
 /** Quota-aware selection is on by default; disable with
  *  NEUROLINK_PROXY_QUOTA_ROUTING=off|false|0. Only affects the fill-first
  *  strategy (round-robin keeps strict rotation). */
@@ -260,6 +284,7 @@ function accountSortMetrics(accountKey, now) {
     const sessionRejected = q?.sessionStatus === "rejected" && sessionReset !== undefined;
     return {
         usable: !coolingActive && !weeklyRejected && !sessionRejected,
+        hasQuota: !!q,
         coolingUntil: st?.coolingUntil ?? 0,
         weeklyReset: weeklyReset ?? Number.POSITIVE_INFINITY,
         sessionReset: sessionReset ?? Number.POSITIVE_INFINITY,
@@ -272,14 +297,18 @@ function accountSortMetrics(accountKey, now) {
  * allowance isn't wasted, then move to accounts with longer-dated resets.
  *
  * Priority among usable accounts:
- *   1. soonest WEEKLY (7d) reset  — the scarce, use-it-or-lose-it ceiling
- *   2. soonest SESSION (5h) reset
- *   3. highest weekly utilization — finish off the one closest to done
- * Accounts with no quota data yet keep insertion order (stable sort) and sit
- * after those with a known soonest reset. Cooling/rejected accounts sort last,
- * soonest-back-to-service first, as last resort.
+ *   1. no quota data yet — probe first: one request reveals its windows and
+ *      self-corrects the ordering. (Ranking unknowns last would starve them
+ *      forever: never picked → never observed → never comparable.)
+ *   2. soonest WEEKLY (7d) reset  — the scarce, use-it-or-lose-it ceiling
+ *   3. soonest SESSION (5h) reset
+ *   4. highest weekly utilization — finish off the one closest to done
+ *   5. configured primary account, then insertion order
+ * Cooling/rejected accounts sort last, soonest-back-to-service first, as
+ * last resort.
  */
 function orderAccountsByQuota(accounts, now) {
+    const primaryKey = configuredPrimaryAccountKey;
     return [...accounts].sort((a, b) => {
         const ma = accountSortMetrics(a.key, now);
         const mb = accountSortMetrics(b.key, now);
@@ -291,13 +320,22 @@ function orderAccountsByQuota(accounts, now) {
             const bu = mb.coolingUntil || Number.POSITIVE_INFINITY;
             return au - bu;
         }
+        if (ma.hasQuota !== mb.hasQuota) {
+            return ma.hasQuota ? 1 : -1;
+        }
         if (ma.weeklyReset !== mb.weeklyReset) {
             return ma.weeklyReset - mb.weeklyReset;
         }
         if (ma.sessionReset !== mb.sessionReset) {
             return ma.sessionReset - mb.sessionReset;
         }
-        return mb.weeklyUsed - ma.weeklyUsed;
+        if (ma.weeklyUsed !== mb.weeklyUsed) {
+            return mb.weeklyUsed - ma.weeklyUsed;
+        }
+        if (primaryKey && (a.key === primaryKey) !== (b.key === primaryKey)) {
+            return a.key === primaryKey ? -1 : 1;
+        }
+        return 0;
     });
 }
 // ---------------------------------------------------------------------------
@@ -1280,6 +1318,7 @@ async function loadClaudeProxyAccounts(args) {
         state.lastToken = account.token;
         state.lastRefreshToken = account.refreshToken;
     }
+    await seedRuntimeQuotasFromDisk(accounts);
     const enabledAccounts = accounts.filter((account) => {
         return !getOrCreateRuntimeState(account.key).permanentlyDisabled;
     });
@@ -3777,6 +3816,11 @@ export const __testHooks = {
     planCooldownFor429,
     orderAccountsByQuota,
     resetEpochToMs,
+    seedRuntimeQuotasFromDisk,
+    getAccountRuntimeState: (key) => {
+        const state = accountRuntimeState.get(key);
+        return state ? { ...state } : undefined;
+    },
     setConfiguredPrimaryAccountKey: (key) => {
         configuredPrimaryAccountKey = key;
     },

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@juspay/neurolink",
-  "version": "9.86.0",
+  "version": "9.86.1",
   "packageManager": "pnpm@10.15.1",
   "description": "Universal AI Development Platform with working MCP integration, multi-provider support, voice (TTS/STT/realtime), and professional CLI. 58+ external MCP servers discoverable, multimodal file processing, RAG pipelines. Build, test, and deploy AI applications with 21+ providers: OpenAI, Anthropic, Google AI Studio, Google Vertex, AWS Bedrock, Azure OpenAI, Mistral, LiteLLM, SageMaker, Hugging Face, Ollama, OpenAI-compatible, OpenRouter, DeepSeek, NVIDIA NIM, LM Studio, llama.cpp, plus voice (OpenAI TTS, ElevenLabs, Deepgram, Azure Speech).",
   "author": {