@juspay/neurolink 9.54.0 → 9.54.2

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

@@ -20,7 +20,7 @@ import { createRawStreamCapture } from "../../proxy/rawStreamCapture.js";
 import { logBodyCapture, logRequest, logRequestAttempt, logStreamError, } from "../../proxy/requestLogger.js";
 import { createSSEInterceptor } from "../../proxy/sseInterceptor.js";
 import { needsRefresh, persistTokens, refreshToken, } from "../../proxy/tokenRefresh.js";
-import { applyRateLimitCooldownScope, buildProxyTranslationPlan, classifyClaudeProxyRequest, getActiveCooldownScope, partitionAccountsByCooldown, summarizeSkippedFallbacks, } from "../../proxy/routingPolicy.js";
+import { applyRateLimitCooldown, buildProxyTranslationPlan, clearAccountCooldown, getAccountCooldownUntil, partitionAccountsByCooldown, } from "../../proxy/routingPolicy.js";
 import { writeJsonSnapshotAtomically } from "../../proxy/snapshotPersistence.js";
 import { recordAttempt, recordAttemptError, recordCooldown, recordFinalError, recordFinalSuccess, } from "../../proxy/usageStats.js";
 import { logger } from "../../utils/logger.js";
@@ -440,9 +440,7 @@ function logProxyRoutingPlan(logProxyBody, stage, plan) {
         contentType: "application/json",
         body: {
             stage,
-            requestProfile: plan.profile,
             attempts: plan.attempts,
-            skipped: plan.skipped,
         },
     });
 }
@@ -1347,30 +1345,21 @@ async function executeClaudeFallbackTranslation(args) {
     return clientResponse;
 }
 async function tryConfiguredClaudeFallbackChain(args) {
-    const { ctx, body, parsedFallbackRequest, requestProfile, modelRouter, tracer, requestStartTime, logProxyBody, logFinalRequest, } = args;
+    const { ctx, body, parsedFallbackRequest, modelRouter, tracer, requestStartTime, logProxyBody, logFinalRequest, } = args;
     const chain = modelRouter?.getFallbackChain() ?? [];
     const fallbackPlan = buildProxyTranslationPlan({ provider: "anthropic", model: body.model }, chain, body.model, parsedFallbackRequest);
-    const fallbackPolicyReason = summarizeSkippedFallbacks(fallbackPlan);
     logProxyBody({
         phase: "routing_decision",
         contentType: "application/json",
         body: {
             stage: "anthropic_fallback",
-            requestProfile,
             attempts: fallbackPlan.attempts.slice(1),
-            skipped: fallbackPlan.skipped,
         },
     });
-    for (const skipped of fallbackPlan.skipped) {
-        const label = skipped.provider
-            ? `${skipped.provider}/${skipped.model ?? "unknown"}`
-            : "auto-provider";
-        logger.always(`[proxy] skipping fallback ${label}: ${skipped.reason}`);
-    }
     tracer?.setFallbackInfo({
         triggered: true,
         attemptCount: fallbackPlan.attempts.slice(1).length,
-        reason: fallbackPolicyReason ?? "all_anthropic_accounts_exhausted",
+        reason: "all_anthropic_accounts_exhausted",
     });
     for (const fallback of fallbackPlan.attempts.slice(1)) {
         if (!fallback.provider || !fallback.model) {
@@ -1410,10 +1399,7 @@ async function tryConfiguredClaudeFallbackChain(args) {
                 attemptCount: fallbackPlan.attempts.slice(1).length,
                 reason: "fallback_success",
             });
-            return {
-                response,
-                fallbackPolicyReason,
-            };
+            return { response };
         }
         catch (fallbackErr) {
             const errMsg = fallbackErr instanceof Error
@@ -1450,10 +1436,7 @@ async function tryConfiguredClaudeFallbackChain(args) {
             });
         }
     }
-    return {
-        response: null,
-        fallbackPolicyReason,
-    };
+    return { response: null };
 }
 async function tryAutoClaudeFallback(args) {
     const { ctx, body, tracer, requestStartTime, logProxyBody, logFinalRequest } = args;
@@ -1484,7 +1467,7 @@ async function tryAutoClaudeFallback(args) {
     }
 }
 function buildClaudeAnthropicFailureResponse(args) {
-    const { tracer, requestStartTime, authFailureMessage, invalidRequestFailure, sawNetworkError, sawTransientFailure, sawRateLimit, lastError, orderedAccounts, requestProfile, fallbackPolicyReason, buildLoggedClaudeError, logProxyBody, logFinalRequest, } = args;
+    const { tracer, requestStartTime, authFailureMessage, invalidRequestFailure, sawNetworkError, sawTransientFailure, sawRateLimit, lastError, orderedAccounts, buildLoggedClaudeError, logProxyBody, logFinalRequest, } = args;
     if (authFailureMessage && !sawRateLimit) {
         tracer?.setError("authentication_error", authFailureMessage);
         tracer?.end(401, Date.now() - requestStartTime);
@@ -1531,21 +1514,21 @@ function buildClaudeAnthropicFailureResponse(args) {
         return buildLoggedClaudeError(502, msg);
     }
     const earliestRecovery = orderedAccounts.reduce((min, account) => {
-        const cooldown = getActiveCooldownScope(getOrCreateRuntimeState(account.key), requestProfile);
-        return cooldown ? Math.min(min, cooldown.until) : min;
+        const until = getAccountCooldownUntil(getOrCreateRuntimeState(account.key));
+        return until !== null ? Math.min(min, until) : min;
     }, Infinity);
+    // If no active cooldown remains (expired while retries ran), use 1s
+    // instead of fabricating a long retry-after.
     const retryAfterSec = Number.isFinite(earliestRecovery)
         ? Math.max(1, Math.ceil((earliestRecovery - Date.now()) / 1000))
-        : 60;
-    const contractMessage = fallbackPolicyReason
-        ? ` ${fallbackPolicyReason}`
-        : "";
-    logger.always(`[proxy] all accounts rate-limited for request-class=${requestProfile.primaryClass}, retry in ${retryAfterSec}s`);
-    const errorBody = buildClaudeError(429, `All accounts rate-limited. Earliest recovery in ${retryAfterSec}s.${contractMessage}`, "overloaded_error");
-    tracer?.setError("rate_limit_error", `All accounts rate-limited. Retry in ${retryAfterSec}s.${contractMessage}`);
+        : 1;
+    const errorMessage = `All accounts rate-limited. Earliest recovery in ${retryAfterSec}s.`;
+    logger.always(`[proxy] all accounts rate-limited, retry in ${retryAfterSec}s`);
+    const errorBody = buildClaudeError(429, errorMessage, "overloaded_error");
+    tracer?.setError("rate_limit_error", errorMessage);
     tracer?.end(429, Date.now() - requestStartTime);
     recordFinalError(429);
-    logFinalRequest(429, "", "final", "rate_limit_error", `All accounts rate-limited. Retry in ${retryAfterSec}s.${contractMessage}`);
+    logFinalRequest(429, "", "final", "rate_limit_error", errorMessage);
     const errorBodyText = JSON.stringify(errorBody);
     logProxyBody({
         phase: "client_response",
@@ -1568,22 +1551,9 @@ function buildClaudeAnthropicFailureResponse(args) {
     });
 }
 async function handleAnthropicSuccessfulResponse(args) {
-    const { ctx, body, account, accountState, requestProfile, response, tracer, requestStartTime, fetchStartMs, attemptNumber, finalBodyStr, upstreamSpan, logProxyBody, logFinalRequest, } = args;
-    accountState.backoffLevel = 0;
-    accountState.coolingUntil = undefined;
+    const { ctx, body, account, accountState, response, tracer, requestStartTime, fetchStartMs, attemptNumber, finalBodyStr, upstreamSpan, logProxyBody, logFinalRequest, } = args;
+    clearAccountCooldown(accountState);
     accountState.consecutiveRefreshFailures = 0;
-    if (accountState.requestClassCooldowns) {
-        delete accountState.requestClassCooldowns[`${requestProfile.primaryClass}:${requestProfile.requestedModel.toLowerCase()}`];
-    }
-    if (accountState.modelTierCooldowns) {
-        delete accountState.modelTierCooldowns[requestProfile.modelTier];
-    }
-    if (accountState.requestClassBackoffLevels) {
-        delete accountState.requestClassBackoffLevels[`${requestProfile.primaryClass}:${requestProfile.requestedModel.toLowerCase()}`];
-    }
-    if (accountState.modelTierBackoffLevels) {
-        delete accountState.modelTierBackoffLevels[requestProfile.modelTier];
-    }
     logger.always(`[proxy] ← ${response.status} account=${account.label}`);
     const quota = parseQuotaHeaders(response.headers);
     if (quota) {
@@ -2173,7 +2143,7 @@ async function handleAnthropicSuccessfulRetryResponse(args) {
     return retryJson;
 }
 async function handleAnthropicAuthRetry(args) {
-    const { ctx, body, account, accountState, requestProfile, headers, buildUpstreamBody, enabledAccounts, orderedAccounts, response: _response, tracer, requestStartTime, fetchStartMs, attemptNumber, finalBodyStr, upstreamSpan, logAttempt, logProxyBody, logFinalRequest, lastError, authFailureMessage, sawRateLimit, sawTransientFailure, sawNetworkError, } = args;
+    const { ctx, body, account, accountState, headers, buildUpstreamBody, enabledAccounts, orderedAccounts, response: _response, tracer, requestStartTime, fetchStartMs, attemptNumber, finalBodyStr, upstreamSpan, logAttempt, logProxyBody, logFinalRequest, lastError, authFailureMessage, sawRateLimit, sawTransientFailure, sawNetworkError, } = args;
     recordAttemptError(account.label, account.type, 401);
     let currentLastError = lastError;
     let currentAuthFailureMessage = authFailureMessage;
@@ -2257,9 +2227,8 @@ async function handleAnthropicAuthRetry(args) {
                 const cooldownMs = Number.isNaN(parsedRetryAfter)
                     ? 60_000
                     : Math.max(1, parsedRetryAfter) * 1000;
-                const cooldown = applyRateLimitCooldownScope({
+                const cooldown = applyRateLimitCooldown({
                     state: accountState,
-                    profile: requestProfile,
                     retryAfterMs: cooldownMs,
                     capMs: RATE_LIMIT_BACKOFF_CAP_MS,
                 });
@@ -2836,7 +2805,7 @@ async function prepareAnthropicAccountAttempt(args) {
     };
 }
 async function fetchAnthropicAccountResponse(args) {
-    const { url, headers, finalBodyStr, account, accountState, requestProfile, enabledAccounts, orderedAccounts, tracer, logAttempt, currentLastError, currentSawRateLimit, currentSawNetworkError, upstreamSpan, } = args;
+    const { url, headers, finalBodyStr, account, accountState, enabledAccounts, orderedAccounts, tracer, logAttempt, currentLastError, currentSawRateLimit, currentSawNetworkError, upstreamSpan, } = args;
     let lastError = currentLastError;
     let sawRateLimit = currentSawRateLimit;
     let sawNetworkError = currentSawNetworkError;
@@ -2889,9 +2858,8 @@ async function fetchAnthropicAccountResponse(args) {
                 }
             }
         }
-        const cooldown = applyRateLimitCooldownScope({
+        const cooldown = applyRateLimitCooldown({
             state: accountState,
-            profile: requestProfile,
             retryAfterMs: cooldownMs > 0 ? cooldownMs : undefined,
             capMs: RATE_LIMIT_BACKOFF_CAP_MS,
         });
@@ -2899,7 +2867,7 @@ async function fetchAnthropicAccountResponse(args) {
         recordAttemptError(account.label, account.type, 429);
         recordCooldown(account.label, account.type, Date.now() + cooldown.backoffMs, accountState.backoffLevel);
         lastError = await response.text();
-        logger.always(`[proxy] ← 429 account=${account.label} backoff-level=${accountState.backoffLevel} cooldown=${Math.round(cooldown.backoffMs / 1000)}s request-class=${cooldown.requestClassKey} model-tier=${cooldown.modelTierKey}`);
+        logger.always(`[proxy] ← 429 account=${account.label} backoff-level=${accountState.backoffLevel} cooldown=${Math.round(cooldown.backoffMs / 1000)}s`);
         logAttempt(429, "rate_limit_error", String(lastError));
         tracer?.setError("rate_limit_error", String(lastError).slice(0, 500));
         tracer?.recordRetry(account.label, "rate_limit");
@@ -2924,7 +2892,6 @@ async function fetchAnthropicAccountResponse(args) {
 async function handleAnthropicRoutedClaudeRequest(args) {
     const { ctx, body, modelRouter, tracer, requestStartTime, accountStrategy, buildLoggedClaudeError, logProxyBody, logFinalRequest, } = args;
     const parsedRequest = parseClaudeRequest(body);
-    const requestProfile = classifyClaudeProxyRequest(body.model, parsedRequest);
     const loadedAccounts = await loadClaudeProxyAccounts({
         ctx,
         body,
@@ -2947,14 +2914,16 @@ async function handleAnthropicRoutedClaudeRequest(args) {
         attemptNumber: 0,
     };
     const acctSelectionSpan = tracer?.startAccountSelection();
-    const accountPartition = partitionAccountsByCooldown(orderedAccounts, (account) => getOrCreateRuntimeState(account.key), requestProfile);
+    const accountPartition = partitionAccountsByCooldown(orderedAccounts, (account) => getOrCreateRuntimeState(account.key));
     for (const skippedAccount of accountPartition.skipped) {
-        if (skippedAccount.cooldown.scope === "request_class" ||
-            skippedAccount.cooldown.scope === "model_tier") {
-            loopState.sawRateLimit = true;
-        }
-        loopState.lastError = `Skipped account=${skippedAccount.account.label} due to ${skippedAccount.cooldown.scope} cooldown ${skippedAccount.cooldown.key}`;
-        logger.always(`[proxy] skipping account=${skippedAccount.account.label} due to ${skippedAccount.cooldown.scope} cooldown=${skippedAccount.cooldown.key} remaining=${Math.max(1, Math.ceil((skippedAccount.cooldown.until - Date.now()) / 1000))}s`);
+        logger.always(`[proxy] skipping account=${skippedAccount.account.label} cooldown remaining=${Math.max(1, Math.ceil((skippedAccount.cooldown.until - Date.now()) / 1000))}s`);
+    }
+    // Only flag rate-limit when ALL accounts are cooling — if some are eligible,
+    // let the actual attempt results determine sawRateLimit via real 429 responses.
+    if (accountPartition.skipped.length > 0 &&
+        accountPartition.eligible.length === 0) {
+        loopState.sawRateLimit = true;
+        loopState.lastError = `All ${accountPartition.skipped.length} accounts are cooling down`;
     }
     accountLoop: for (const account of accountPartition.eligible) {
         const accountState = getOrCreateRuntimeState(account.key);
@@ -3009,7 +2978,6 @@ async function handleAnthropicRoutedClaudeRequest(args) {
                 finalBodyStr: preparedAttempt.finalBodyStr,
                 account,
                 accountState,
-                requestProfile,
                 enabledAccounts,
                 orderedAccounts,
                 tracer,
@@ -3046,7 +3014,6 @@ async function handleAnthropicRoutedClaudeRequest(args) {
                     body,
                     account,
                     accountState,
-                    requestProfile,
                     headers: preparedAttempt.headers,
                     buildUpstreamBody: preparedAttempt.buildUpstreamBody,
                     enabledAccounts,
@@ -3126,7 +3093,6 @@ async function handleAnthropicRoutedClaudeRequest(args) {
                 body,
                 account,
                 accountState,
-                requestProfile,
                 response,
                 tracer,
                 requestStartTime,
@@ -3150,7 +3116,6 @@ async function handleAnthropicRoutedClaudeRequest(args) {
         ctx,
         body,
         parsedFallbackRequest: parsedRequest,
-        requestProfile,
         modelRouter,
         tracer,
         requestStartTime,
@@ -3160,8 +3125,9 @@ async function handleAnthropicRoutedClaudeRequest(args) {
     if (configuredFallbackResult.response) {
         return configuredFallbackResult.response;
     }
-    const configuredChain = modelRouter?.getFallbackChain() ?? [];
-    if (configuredChain.length === 0 && !loopState.sawRateLimit) {
+    // Try auto-provider fallback when the configured chain didn't produce a
+    // response (either no chain configured, or all entries failed/deduped).
+    if (!loopState.sawRateLimit) {
         const autoFallbackResponse = await tryAutoClaudeFallback({
             ctx,
             body,
@@ -3184,8 +3150,6 @@ async function handleAnthropicRoutedClaudeRequest(args) {
         sawRateLimit: loopState.sawRateLimit,
         lastError: loopState.lastError,
         orderedAccounts,
-        requestProfile,
-        fallbackPolicyReason: configuredFallbackResult.fallbackPolicyReason,
         buildLoggedClaudeError,
         logProxyBody,
         logFinalRequest,
@@ -3406,8 +3370,6 @@ function getOrCreateRuntimeState(accountKey) {
         backoffLevel: 0,
         consecutiveRefreshFailures: 0,
         permanentlyDisabled: false,
-        requestClassCooldowns: {},
-        modelTierCooldowns: {},
     };
     accountRuntimeState.set(accountKey, initial);
     return initial;

package/dist/lib/types/proxyTypes.d.ts

@@ -641,10 +641,6 @@ export type RuntimeAccountState = {
     backoffLevel: number;
     consecutiveRefreshFailures: number;
     permanentlyDisabled: boolean;
-    requestClassCooldowns?: Record<string, number>;
-    modelTierCooldowns?: Record<string, number>;
-    requestClassBackoffLevels?: Record<string, number>;
-    modelTierBackoffLevels?: Record<string, number>;
     lastToken?: string;
     lastRefreshToken?: string;
 };
@@ -697,63 +693,26 @@ export type CachedSession = {
 };
 /** Model tier classification for proxy routing decisions. */
 export type ClaudeProxyModelTier = "opus" | "sonnet" | "haiku" | "other";
-/** Request class for proxy routing policy. */
-export type ClaudeProxyRequestClass = "multimodal" | "high-tool-count-non-stream-structured" | "strong-tool-fidelity" | "streaming-conversational" | "standard";
-/** Full classification profile for a proxy request. */
-export type ClaudeProxyRequestProfile = {
-    requestedModel: string;
-    modelTier: ClaudeProxyModelTier;
-    primaryClass: ClaudeProxyRequestClass;
-    classes: ClaudeProxyRequestClass[];
-    stream: boolean;
-    toolCount: number;
-    hasImages: boolean;
-    hasThinking: boolean;
-    hasToolHistory: boolean;
-    requiresToolUse: boolean;
-    requiresSpecificTool: boolean;
-    requiresStrongToolFidelity: boolean;
-    isHighToolCountNonStream: boolean;
-    isStreamingConversational: boolean;
-    isMultimodal: boolean;
-};
-/** Outcome of evaluating a single fallback candidate. */
-export type FallbackEligibilityDecision = {
-    provider?: string;
-    model?: string;
-    eligible: boolean;
-    reason: string;
-};
 /** A single provider attempt in the proxy translation plan. */
 export type ProxyTranslationAttempt = {
     provider?: string;
     model?: string;
     label: string;
 };
-/** Ordered plan of provider attempts and skipped candidates. */
+/** Ordered plan of provider attempts for a proxy request. */
 export type ProxyTranslationPlan = {
-    profile: ClaudeProxyRequestProfile;
+    requestedModel: string;
+    modelTier: ClaudeProxyModelTier;
     attempts: ProxyTranslationAttempt[];
-    skipped: FallbackEligibilityDecision[];
+    skipped: never[];
 };
-/** Discriminated union describing why a cooldown is active. */
-export type CooldownScope = {
-    scope: "request_class";
-    key: string;
-    until: number;
-} | {
-    scope: "model_tier";
-    key: string;
-    until: number;
-} | {
-    scope: "generic";
-    key: "generic";
-    until: number;
-};
-/** An account skipped during partitioning, with the cooldown that caused it. */
+/** An account skipped during partitioning, with its cooldown info. */
 export type CooldownSkippedAccount<T> = {
     account: T;
-    cooldown: CooldownScope;
+    cooldown: {
+        until: number;
+        backoffLevel: number;
+    };
 };
 /** Mutable readiness state tracked by the proxy process. */
 export type ProxyReadinessState = {

package/dist/lib/utils/messageBuilder.js

@@ -350,11 +350,9 @@ export function convertToModelMessages(messages) {
                 // Assistant messages only support text content, filter out images
                 const textOnlyContent = validContent.filter((item) => item.type === "text");
                 if (textOnlyContent.length === 0) {
-                    // If no text content, convert to empty string
-                    return {
-                        role: "assistant",
-                        content: "",
-                    };
+                    // No text content (e.g., only images/files) — skip message
+                    // to avoid sending empty content to providers like Claude
+                    return null;
                 }
                 else if (textOnlyContent.length === 1) {
                     // Single text item, use string content
@@ -1073,9 +1071,44 @@ export async function buildMultimodalMessagesArray(options, provider, model) {
                 msg.role === "assistant" ||
                 msg.role === "system") {
                 const providerOptions = msg.providerOptions;
+                // Sanitize assistant array content: strip tool_use/tool_result blocks
+                // that providers cannot handle. If an assistant message ends up empty
+                // after stripping, skip it to avoid sending content: "" to Claude.
+                // Only assistant messages need this — user messages may contain valid
+                // image/file blocks that must pass through unchanged.
+                let sanitizedContent = msg.content;
+                if (msg.role === "assistant" && Array.isArray(msg.content)) {
+                    const textParts = msg.content.filter((item) => !!item &&
+                        typeof item === "object" &&
+                        item.type === "text" &&
+                        typeof item.text === "string");
+                    if (textParts.length === 0) {
+                        // All content was tool_use/tool_result/non-text — skip message
+                        continue;
+                    }
+                    // Check if any retained text part carries providerOptions
+                    // (e.g. Anthropic cache_control). If so, preserve them as
+                    // array content to avoid losing per-block metadata.
+                    const hasItemProviderOptions = textParts.some((item) => !!item.providerOptions);
+                    if (hasItemProviderOptions) {
+                        sanitizedContent = textParts;
+                    }
+                    else {
+                        sanitizedContent =
+                            textParts.length === 1
+                                ? textParts[0].text
+                                : textParts
+                                    .map((p) => p.text)
+                                    .join(" ");
+                    }
+                }
+                // Skip empty string content to avoid Claude API rejection
+                if (sanitizedContent === "") {
+                    continue;
+                }
                 messages.push({
                     role: msg.role,
-                    content: msg.content,
+                    content: sanitizedContent,
                     ...(providerOptions && { providerOptions }),
                 });
             }

package/dist/providers/googleVertex.d.ts

@@ -5,6 +5,20 @@ import { BaseProvider } from "../core/baseProvider.js";
 import type { EnhancedGenerateResult, TextGenerationOptions } from "../types/generateTypes.js";
 import type { NeurolinkCredentials } from "../types/providers.js";
 import type { StreamOptions, StreamResult } from "../types/streamTypes.js";
+/**
+ * Resolve the correct Vertex AI location for a given model.
+ *
+ * Google-published models (gemini-*) require the global endpoint
+ * (`aiplatform.googleapis.com`), not regional endpoints like
+ * `us-east5-aiplatform.googleapis.com`. Regional endpoints return
+ * "model not found" for these models.
+ *
+ * Anthropic-on-Vertex models (claude-*) require regional endpoints
+ * and are handled separately by `createVertexAnthropicSettings`.
+ *
+ * Embedding models and custom models use the configured location as-is.
+ */
+export declare const resolveVertexLocation: (modelName: string | undefined, configuredLocation: string) => string;
 /**
  * Vertex Model Aliases
  *

package/dist/providers/googleVertex.js

@@ -79,7 +79,36 @@ const getVertexLocation = () => {
     return (process.env.GOOGLE_CLOUD_LOCATION ||
         process.env.VERTEX_LOCATION ||
         process.env.GOOGLE_VERTEX_LOCATION ||
-        "us-central1");
+        "global");
+};
+/**
+ * Resolve the correct Vertex AI location for a given model.
+ *
+ * Google-published models (gemini-*) require the global endpoint
+ * (`aiplatform.googleapis.com`), not regional endpoints like
+ * `us-east5-aiplatform.googleapis.com`. Regional endpoints return
+ * "model not found" for these models.
+ *
+ * Anthropic-on-Vertex models (claude-*) require regional endpoints
+ * and are handled separately by `createVertexAnthropicSettings`.
+ *
+ * Embedding models and custom models use the configured location as-is.
+ */
+export const resolveVertexLocation = (modelName, configuredLocation) => {
+    if (!modelName) {
+        return configuredLocation;
+    }
+    const normalized = modelName.toLowerCase();
+    // Google-published models always use the global endpoint.
+    // Hardcoded because Google's Vertex AI serves Gemini models exclusively
+    // from the global endpoint — regional endpoints like us-east5 return
+    // "Publisher Model was not found" errors. The env var GOOGLE_VERTEX_LOCATION
+    // is typically set for Anthropic-on-Vertex (which needs regional), so we
+    // cannot rely on it for Gemini routing.
+    if (normalized.startsWith("gemini-")) {
+        return "global";
+    }
+    return configuredLocation;
 };
 const getDefaultVertexModel = () => {
     // Use gemini-2.5-flash as default - latest and best price-performance model
@@ -96,8 +125,9 @@ const hasGoogleCredentials = () => {
 // Module-level cache for runtime-created credentials file to avoid per-request writes
 let cachedCredentialsPath = null;
 // Enhanced Vertex settings creation with authentication fallback and proxy support
-const createVertexSettings = async (region, credentials) => {
-    const location = credentials?.location || region || getVertexLocation();
+const createVertexSettings = async (region, credentials, modelName) => {
+    const configuredLocation = credentials?.location || region || getVertexLocation();
+    const location = resolveVertexLocation(modelName, configuredLocation);
     const project = credentials?.projectId || getVertexProjectId();
     const baseSettings = {
         project,
@@ -326,7 +356,12 @@ const createVertexAnthropicSettings = async (region, credentials) => {
     // which is invalid. The correct global endpoint omits the region prefix entirely.
     // Since the SDK doesn't handle this, redirect "global" to "us-east5" for Anthropic.
     const anthropicRegion = !region || region === "global" ? "us-east5" : region;
-    const baseVertexSettings = await createVertexSettings(anthropicRegion, credentials);
+    // Override credentials.location so it cannot conflict with the redirected
+    // region — createVertexSettings checks credentials.location first.
+    const anthropicCredentials = credentials?.location
+        ? { ...credentials, location: anthropicRegion }
+        : credentials;
+    const baseVertexSettings = await createVertexSettings(anthropicRegion, anthropicCredentials);
     // GoogleVertexAnthropicProviderSettings extends GoogleVertexProviderSettings
     // so we can use the same settings with proper typing
     return {
@@ -570,7 +605,9 @@ export class GoogleVertexProvider extends BaseProvider {
             networkConfig: {
                 projectId: this.projectId,
                 location: this.location,
-                expectedEndpoint: `https://${this.location}-aiplatform.googleapis.com`,
+                expectedEndpoint: this.location === "global"
+                    ? "https://aiplatform.googleapis.com"
+                    : `https://${this.location}-aiplatform.googleapis.com`,
                 httpProxy: process.env.HTTP_PROXY || process.env.http_proxy,
                 httpsProxy: process.env.HTTPS_PROXY || process.env.https_proxy,
                 noProxy: process.env.NO_PROXY || process.env.no_proxy,
@@ -582,7 +619,7 @@ export class GoogleVertexProvider extends BaseProvider {
             message: "Starting Vertex settings creation with network configuration analysis",
         });
         try {
-            const vertexSettings = await createVertexSettings(this.location, this.credentials);
+            const vertexSettings = await createVertexSettings(this.location, this.credentials, modelName);
             const vertexSettingsEndTime = process.hrtime.bigint();
             const vertexSettingsDurationNs = vertexSettingsEndTime - vertexSettingsStartTime;
             logger.debug(`[GoogleVertexProvider] ✅ LOG_POINT_V009_VERTEX_SETTINGS_SUCCESS`, {
@@ -1116,12 +1153,13 @@ export class GoogleVertexProvider extends BaseProvider {
     /**
      * Create @google/genai client configured for Vertex AI
      */
-    async createVertexGenAIClient(regionOverride) {
+    async createVertexGenAIClient(regionOverride, modelName) {
         const project = this.credentials?.projectId || getVertexProjectId();
-        const location = this.credentials?.location ||
+        const configuredLocation = this.credentials?.location ||
             regionOverride ||
             this.location ||
             getVertexLocation();
+        const location = resolveVertexLocation(modelName, configuredLocation);
         const mod = await import("@google/genai");
         const ctor = mod.GoogleGenAI;
         if (!ctor) {
@@ -1308,8 +1346,8 @@ export class GoogleVertexProvider extends BaseProvider {
         }, (span) => this.executeNativeGemini3StreamWithSpan(options, modelName, span));
     }
     async executeNativeGemini3StreamWithSpan(options, modelName, span) {
-        const client = await this.createVertexGenAIClient(options.region);
-        const effectiveLocation = options.region || this.location || getVertexLocation();
+        const client = await this.createVertexGenAIClient(options.region, modelName);
+        const effectiveLocation = resolveVertexLocation(modelName, options.region || this.location || getVertexLocation());
         logger.debug("[GoogleVertex] Using native @google/genai for Gemini 3", {
             model: modelName,
             hasTools: !!options.tools && Object.keys(options.tools).length > 0,
@@ -1503,8 +1541,8 @@ export class GoogleVertexProvider extends BaseProvider {
                 [ATTR.NL_PROVIDER]: this.providerName,
             },
         }, async (span) => {
-            const client = await this.createVertexGenAIClient(options.region);
-            const effectiveLocation = options.region || this.location || getVertexLocation();
+            const client = await this.createVertexGenAIClient(options.region, modelName);
+            const effectiveLocation = resolveVertexLocation(modelName, options.region || this.location || getVertexLocation());
             logger.debug("[GoogleVertex] Using native @google/genai for Gemini 3 generate", {
                 model: modelName,
                 project: this.projectId,

package/dist/proxy/routingPolicy.d.ts

@@ -1,33 +1,43 @@
-import type { ClaudeProxyModelTier, ClaudeProxyRequestClass, ClaudeProxyRequestProfile, CooldownScope, CooldownSkippedAccount, FallbackEligibilityDecision, FallbackEntry, ParsedClaudeRequest, ProxyTranslationAttempt, ProxyTranslationPlan, RuntimeAccountState } from "../types/index.js";
-export type { ClaudeProxyModelTier, ClaudeProxyRequestClass, ClaudeProxyRequestProfile, CooldownScope, CooldownSkippedAccount, FallbackEligibilityDecision, ProxyTranslationAttempt, ProxyTranslationPlan, };
+import type { ClaudeProxyModelTier, CooldownSkippedAccount, FallbackEntry, ParsedClaudeRequest, ProxyTranslationAttempt, ProxyTranslationPlan, RuntimeAccountState } from "../types/index.js";
+export type { ClaudeProxyModelTier, ProxyTranslationAttempt, ProxyTranslationPlan, };
 export declare function inferClaudeProxyModelTier(modelName: string): ClaudeProxyModelTier;
-export declare function classifyClaudeProxyRequest(requestedModel: string, parsed: ParsedClaudeRequest): ClaudeProxyRequestProfile;
-export declare function getRequestClassCooldownKey(profile: ClaudeProxyRequestProfile): string;
-export declare function getModelTierCooldownKey(profile: ClaudeProxyRequestProfile): string;
-export declare function evaluateFallbackEligibility(profile: ClaudeProxyRequestProfile, candidate: {
-    provider?: string;
-    model?: string;
-}): FallbackEligibilityDecision;
+/**
+ * Build a translation plan for a Claude-compatible proxy request.
+ * The plan lists the primary provider followed by eligible fallback targets.
+ * All configured fallback entries are always eligible — no contract-based gating.
+ * When no fallback chain is configured, an "auto-provider" entry is appended.
+ */
 export declare function buildProxyTranslationPlan(primary: {
     provider: string;
     model?: string;
-}, fallbackChain: FallbackEntry[], requestedModel: string, parsed: ParsedClaudeRequest): ProxyTranslationPlan;
-export declare function summarizeSkippedFallbacks(plan: Pick<ProxyTranslationPlan, "profile" | "skipped">): string | null;
-export declare function getActiveCooldownScope(state: RuntimeAccountState, profile: ClaudeProxyRequestProfile, now?: number): CooldownScope | null;
+}, fallbackChain: FallbackEntry[], requestedModel: string, _parsed: ParsedClaudeRequest): ProxyTranslationPlan;
+/**
+ * Check whether an account is currently cooling down.
+ * Returns the cooldown timestamp if active, null otherwise.
+ */
+export declare function getAccountCooldownUntil(state: RuntimeAccountState, now?: number): number | null;
+/**
+ * Partition accounts into eligible (no cooldown) and skipped (cooling down).
+ */
 export declare function partitionAccountsByCooldown<T extends {
     key: string;
-}>(accounts: T[], getState: (account: T) => RuntimeAccountState, profile: ClaudeProxyRequestProfile, now?: number): {
+}>(accounts: T[], getState: (account: T) => RuntimeAccountState, now?: number): {
     eligible: T[];
     skipped: CooldownSkippedAccount<T>[];
 };
-export declare function applyRateLimitCooldownScope(args: {
+/**
+ * Apply a rate-limit cooldown to an account.
+ * Uses simple exponential backoff with a floor and cap.
+ */
+export declare function applyRateLimitCooldown(args: {
     state: RuntimeAccountState;
-    profile: ClaudeProxyRequestProfile;
     retryAfterMs?: number;
     now?: number;
     capMs: number;
 }): {
     backoffMs: number;
-    requestClassKey: string;
-    modelTierKey: string;
 };
+/**
+ * Clear cooldown state for an account after a successful request.
+ */
+export declare function clearAccountCooldown(state: RuntimeAccountState): void;