@juspay/neurolink 10.9.1 → 10.10.1

package/dist/cli/commands/proxy.js

@@ -1497,9 +1497,34 @@ export async function createProxyStartApp(params) {
                 toolRegistry: params.neurolink.getToolRegistry(),
                 timestamp: Date.now(),
                 metadata: {},
+                // Route handlers publish limit/quota headers here. Only the streaming
+                // paths build their own Response (and set headers directly); every
+                // JSON and error path returns a plain object, so without this the
+                // headers had nowhere to go. Applied to each c.json() return below.
+                responseHeaders: {},
+            };
+            /** Copy handler-published headers onto the outgoing response. */
+            const applyResponseHeaders = () => {
+                for (const [key, value] of Object.entries(ctx.responseHeaders)) {
+                    c.header(key, value);
+                }
             };
             const result = await route.handler(ctx);
             if (result instanceof Response) {
+                // Streaming responses own their headers; merge in anything the
+                // handler published on the context that the Response lacks. A Response
+                // obtained from fetch() carries immutable headers, so this is
+                // best-effort — the body must still reach the client either way.
+                try {
+                    for (const [key, value] of Object.entries(ctx.responseHeaders)) {
+                        if (!result.headers.has(key)) {
+                            result.headers.set(key, value);
+                        }
+                    }
+                }
+                catch {
+                    // Immutable header guard — skip enrichment, keep the response.
+                }
                 return result;
             }
             if (result &&
@@ -1544,6 +1569,7 @@ export async function createProxyStartApp(params) {
                 });
                 return new Response(responseStream, {
                     headers: {
+                        ...ctx.responseHeaders,
                         "Content-Type": "text/event-stream",
                         "Cache-Control": "no-cache",
                         Connection: "keep-alive",
@@ -1556,6 +1582,7 @@ export async function createProxyStartApp(params) {
                 const httpResult = result;
                 const status = httpResult.httpStatus ?? 200;
                 delete httpResult.httpStatus;
+                applyResponseHeaders();
                 return c.json(result, status);
             }
             if (result &&
@@ -1564,8 +1591,10 @@ export async function createProxyStartApp(params) {
                 result.type === "error") {
                 const errorResult = result;
                 const status = mapClaudeErrorTypeToStatus(errorResult.error?.type);
+                applyResponseHeaders();
                 return c.json(result, status);
             }
+            applyResponseHeaders();
             return c.json(result ?? {});
         });
     }

package/dist/core/modules/GenerationHandler.js

@@ -26,6 +26,7 @@ import { DEFAULT_MAX_STEPS, DEFAULT_WRAPUP_TIME_LEAD_MS, } from "../constants.js
 import { createStepBudgetGuard, estimateFixedOverheadTokens, } from "../../context/stepBudgetGuard.js";
 import { isTemperatureDeprecatedError, isSchemaComplexityError, isToolsSchemaConflictError, isToolsSchemaExclusionInForce, } from "./structuredOutputPolicy.js";
 import { coerceJsonToSchema } from "../../utils/json/coerce.js";
+import { convertZodToJsonSchema } from "../../utils/schemaConversion.js";
 import { NoObjectGeneratedError } from "../../utils/generationErrors.js";
 import { Output, stepCountIs } from "../../utils/tool.js";
 import { generateText } from "../../utils/generation.js";
@@ -105,11 +106,14 @@ export function resolveTurnBudget(options, turnStartMs) {
  * `providerOptions:` spreads in the args literal would silently clobber each
  * other (object spread does not deep-merge).
  */
-function buildProviderOptions(options, isGoogleProvider, callerTimeoutMs) {
+function buildProviderOptions(options, isGoogleProvider, callerTimeoutMs, finalResultSchema) {
     const providerOptions = {};
     if (callerTimeoutMs !== undefined) {
         providerOptions.neurolink = { timeoutMs: callerTimeoutMs };
     }
+    if (finalResultSchema) {
+        providerOptions.anthropic = { finalResultSchema };
+    }
     if (options.thinkingConfig?.enabled && isGoogleProvider) {
         // Gemini 3 uses thinkingLevel; Gemini 2.5 uses thinkingBudget.
         providerOptions.google = {
@@ -236,7 +240,22 @@ export class GenerationHandler {
         const prepareStep = options.prepareStep;
         const { callerTimeoutMs, turnBudgetMs, wrapupLeadMs, turnDeadline } = resolveTurnBudget(options, turnStartMs);
         let wrapupForced = false;
-        const providerOptions = buildProviderOptions(options, isGoogleProvider, callerTimeoutMs);
+        // The native Anthropic Messages surface cannot combine AI-SDK structured
+        // output with tools (see structuredOutputPolicy — experimental_output
+        // replaces the tools array), so `useStructuredOutput` is false above and
+        // the schema would simply be dropped for every agent/MCP turn. Hand the
+        // JSON Schema to the provider instead: it appends an additive
+        // `final_result` tool and returns the answer as that tool's arguments,
+        // keeping the real tools callable. Bedrock is deliberately excluded — it
+        // runs on the third-party @ai-sdk/amazon-bedrock model, which has no such
+        // handling.
+        const finalResultSchema = this.providerName === "anthropic" &&
+            !!options.schema &&
+            shouldUseTools &&
+            Object.keys(tools).length > 0
+            ? convertZodToJsonSchema(options.schema)
+            : undefined;
+        const providerOptions = buildProviderOptions(options, isGoogleProvider, callerTimeoutMs, finalResultSchema);
         // Hoist system-role messages into generateText's top-level `system` option
         // rather than passing them inside `messages` (deprecated by the AI SDK,
         // rejected in v7). See extractSystemMessages for the rationale. (#1024)

package/dist/core/modules/structuredOutputPolicy.d.ts

@@ -23,6 +23,14 @@ export declare function isGeminiProvider(providerName: string, modelName: string
  * experimental_output + tools silently drops tool_use blocks on this surface, so
  * structured output must be disabled when tools are active. Vertex+Claude is NOT
  * matched here (different transport, no conflict).
+ *
+ * Being excluded here no longer means the schema is LOST for provider
+ * "anthropic": GenerationHandler forwards the JSON Schema to the provider via
+ * `providerOptions.anthropic.finalResultSchema`, and the provider appends an
+ * additive `final_result` tool (see providers/anthropic/structuredOutput.ts) —
+ * schema enforcement without giving up tool calling. "bedrock" has no such
+ * handling (it runs on the third-party @ai-sdk/amazon-bedrock model) and still
+ * falls back to text-mode coercion.
  */
 export declare function isNativeAnthropicProvider(providerName: string): boolean;
 /**

package/dist/core/modules/structuredOutputPolicy.js

@@ -33,6 +33,14 @@ export function isGeminiProvider(providerName, modelName) {
  * experimental_output + tools silently drops tool_use blocks on this surface, so
  * structured output must be disabled when tools are active. Vertex+Claude is NOT
  * matched here (different transport, no conflict).
+ *
+ * Being excluded here no longer means the schema is LOST for provider
+ * "anthropic": GenerationHandler forwards the JSON Schema to the provider via
+ * `providerOptions.anthropic.finalResultSchema`, and the provider appends an
+ * additive `final_result` tool (see providers/anthropic/structuredOutput.ts) —
+ * schema enforcement without giving up tool calling. "bedrock" has no such
+ * handling (it runs on the third-party @ai-sdk/amazon-bedrock model) and still
+ * falls back to text-mode coercion.
  */
 export function isNativeAnthropicProvider(providerName) {
     return providerName === "anthropic" || providerName === "bedrock";

package/dist/lib/core/modules/GenerationHandler.js

@@ -26,6 +26,7 @@ import { DEFAULT_MAX_STEPS, DEFAULT_WRAPUP_TIME_LEAD_MS, } from "../constants.js
 import { createStepBudgetGuard, estimateFixedOverheadTokens, } from "../../context/stepBudgetGuard.js";
 import { isTemperatureDeprecatedError, isSchemaComplexityError, isToolsSchemaConflictError, isToolsSchemaExclusionInForce, } from "./structuredOutputPolicy.js";
 import { coerceJsonToSchema } from "../../utils/json/coerce.js";
+import { convertZodToJsonSchema } from "../../utils/schemaConversion.js";
 import { NoObjectGeneratedError } from "../../utils/generationErrors.js";
 import { Output, stepCountIs } from "../../utils/tool.js";
 import { generateText } from "../../utils/generation.js";
@@ -105,11 +106,14 @@ export function resolveTurnBudget(options, turnStartMs) {
  * `providerOptions:` spreads in the args literal would silently clobber each
  * other (object spread does not deep-merge).
  */
-function buildProviderOptions(options, isGoogleProvider, callerTimeoutMs) {
+function buildProviderOptions(options, isGoogleProvider, callerTimeoutMs, finalResultSchema) {
     const providerOptions = {};
     if (callerTimeoutMs !== undefined) {
         providerOptions.neurolink = { timeoutMs: callerTimeoutMs };
     }
+    if (finalResultSchema) {
+        providerOptions.anthropic = { finalResultSchema };
+    }
     if (options.thinkingConfig?.enabled && isGoogleProvider) {
         // Gemini 3 uses thinkingLevel; Gemini 2.5 uses thinkingBudget.
         providerOptions.google = {
@@ -236,7 +240,22 @@ export class GenerationHandler {
         const prepareStep = options.prepareStep;
         const { callerTimeoutMs, turnBudgetMs, wrapupLeadMs, turnDeadline } = resolveTurnBudget(options, turnStartMs);
         let wrapupForced = false;
-        const providerOptions = buildProviderOptions(options, isGoogleProvider, callerTimeoutMs);
+        // The native Anthropic Messages surface cannot combine AI-SDK structured
+        // output with tools (see structuredOutputPolicy — experimental_output
+        // replaces the tools array), so `useStructuredOutput` is false above and
+        // the schema would simply be dropped for every agent/MCP turn. Hand the
+        // JSON Schema to the provider instead: it appends an additive
+        // `final_result` tool and returns the answer as that tool's arguments,
+        // keeping the real tools callable. Bedrock is deliberately excluded — it
+        // runs on the third-party @ai-sdk/amazon-bedrock model, which has no such
+        // handling.
+        const finalResultSchema = this.providerName === "anthropic" &&
+            !!options.schema &&
+            shouldUseTools &&
+            Object.keys(tools).length > 0
+            ? convertZodToJsonSchema(options.schema)
+            : undefined;
+        const providerOptions = buildProviderOptions(options, isGoogleProvider, callerTimeoutMs, finalResultSchema);
         // Hoist system-role messages into generateText's top-level `system` option
         // rather than passing them inside `messages` (deprecated by the AI SDK,
         // rejected in v7). See extractSystemMessages for the rationale. (#1024)

package/dist/lib/core/modules/structuredOutputPolicy.d.ts

@@ -23,6 +23,14 @@ export declare function isGeminiProvider(providerName: string, modelName: string
  * experimental_output + tools silently drops tool_use blocks on this surface, so
  * structured output must be disabled when tools are active. Vertex+Claude is NOT
  * matched here (different transport, no conflict).
+ *
+ * Being excluded here no longer means the schema is LOST for provider
+ * "anthropic": GenerationHandler forwards the JSON Schema to the provider via
+ * `providerOptions.anthropic.finalResultSchema`, and the provider appends an
+ * additive `final_result` tool (see providers/anthropic/structuredOutput.ts) —
+ * schema enforcement without giving up tool calling. "bedrock" has no such
+ * handling (it runs on the third-party @ai-sdk/amazon-bedrock model) and still
+ * falls back to text-mode coercion.
  */
 export declare function isNativeAnthropicProvider(providerName: string): boolean;
 /**

package/dist/lib/core/modules/structuredOutputPolicy.js

@@ -33,6 +33,14 @@ export function isGeminiProvider(providerName, modelName) {
  * experimental_output + tools silently drops tool_use blocks on this surface, so
  * structured output must be disabled when tools are active. Vertex+Claude is NOT
  * matched here (different transport, no conflict).
+ *
+ * Being excluded here no longer means the schema is LOST for provider
+ * "anthropic": GenerationHandler forwards the JSON Schema to the provider via
+ * `providerOptions.anthropic.finalResultSchema`, and the provider appends an
+ * additive `final_result` tool (see providers/anthropic/structuredOutput.ts) —
+ * schema enforcement without giving up tool calling. "bedrock" has no such
+ * handling (it runs on the third-party @ai-sdk/amazon-bedrock model) and still
+ * falls back to text-mode coercion.
  */
 export function isNativeAnthropicProvider(providerName) {
     return providerName === "anthropic" || providerName === "bedrock";

package/dist/lib/providers/anthropic/client.d.ts

@@ -1,6 +1,6 @@
 import { type AIProviderName } from "../../constants/enums.js";
 import { BaseProvider } from "../../core/baseProvider.js";
-import type { AnthropicProviderConfig, StreamOptions, StreamResult, ValidationSchema, EnhancedGenerateResult, TextGenerationOptions, AnthropicAuthMethod, AnthropicResponseMetadata, ClaudeSubscriptionTier, ClaudeUsageInfo } from "../../types/index.js";
+import type { AnthropicProviderConfig, StreamOptions, StreamResult, ValidationSchema, EnhancedGenerateResult, TextGenerationOptions, AnthropicAuthMethod, AnthropicRateLimitInfo, AnthropicResponseMetadata, ClaudeSubscriptionTier, ClaudeUsageInfo } from "../../types/index.js";
 import type { LanguageModel } from "../../types/index.js";
 /**
  * Anthropic Provider v2 - BaseProvider Implementation
@@ -103,12 +103,18 @@ export declare class AnthropicProvider extends BaseProvider {
      */
     getLastResponseMetadata(): AnthropicResponseMetadata | null;
     /**
-     * Update response metadata from API response headers.
-     * This should be called after each API request to track rate limits.
-     * @param headers - Response headers from the API
-     * @param requestId - Optional request ID
+     * Update response metadata from a captured limit snapshot.
+     *
+     * Takes already-parsed rate-limit info rather than raw headers: parsing now
+     * lives in `rateLimitCapture`, which is the only layer that sees the raw
+     * response and understands both header families (unified subscription
+     * windows and legacy per-tier counters).
+     *
+     * @param rateLimit - Parsed rate-limit figures
+     * @param requestId - Optional Anthropic request ID
+     * @param usageUpdate - Optional token counts to fold into usage tracking
      */
-    protected updateResponseMetadata(headers: Headers | Record<string, string>, requestId?: string, usageUpdate?: {
+    protected updateResponseMetadata(rateLimit: AnthropicRateLimitInfo, requestId?: string, usageUpdate?: {
         inputTokens?: number;
         outputTokens?: number;
     }): void;
@@ -127,7 +133,16 @@ export declare class AnthropicProvider extends BaseProvider {
      * BaseProvider so that expired tokens are renewed automatically.
      */
     generate(optionsOrPrompt: TextGenerationOptions | string, analysisSchema?: ValidationSchema): Promise<EnhancedGenerateResult | null>;
-    protected executeStream(options: StreamOptions, _analysisSchema?: ValidationSchema): Promise<StreamResult>;
+    /**
+     * Fold a captured snapshot into the provider's usage bookkeeping and log it.
+     *
+     * `updateResponseMetadata` had no callers before this — the metadata it
+     * maintains, and the public `getLastResponseMetadata()` / `getUsageInfo()`
+     * that read it, were never populated by anything.
+     */
+    private recordLimitSnapshot;
+    protected executeStream(options: StreamOptions, analysisSchema?: ValidationSchema): Promise<StreamResult>;
+    private executeStreamInCaptureScope;
     isAvailable(): Promise<boolean>;
     getModel(): LanguageModel;
 }