@juspay/neurolink 9.81.2 → 9.82.0

package/dist/proxy/proxyDispatcher.d.ts

@@ -0,0 +1,5 @@
+/**
+ * Install the tuned global undici dispatcher. Idempotent — safe to call once at
+ * proxy startup. No-op when `NEUROLINK_PROXY_KEEPALIVE` is off/false/0.
+ */
+export declare function configureProxyKeepAliveDispatcher(): void;

package/dist/proxy/proxyDispatcher.js

@@ -0,0 +1,60 @@
+/**
+ * Tuned global undici dispatcher for the proxy's upstream forwards.
+ *
+ * The Claude passthrough (`claudeProxyRoutes.ts`) forwards every request to
+ * Anthropic via the global `fetch` → undici global dispatcher. undici keep-alives
+ * by default, but with a ~4s idle timeout: between Claude Code turns (the user
+ * reads output, a tool runs, the model thinks) the idle socket closes, so the
+ * next request opens a brand-new TCP connection. Under sustained use that is a
+ * high rate of short-lived outbound flows.
+ *
+ * On hosts running a socket content-filter (CFIL) — e.g. SentinelOne or
+ * GlobalProtect network extensions — every new flow allocates per-flow kernel
+ * state, so a high flow rate amplifies any leak in that path. Reusing connections
+ * cuts the flow rate sharply, so we install a dispatcher with a longer keep-alive
+ * and a bounded, reused connection pool.
+ *
+ * Everything is overridable via env so it can be tuned (or disabled) per host:
+ *   NEUROLINK_PROXY_KEEPALIVE=off       disable; keep undici defaults
+ *   NEUROLINK_PROXY_KEEPALIVE_MS        idle keep-alive timeout (default 30000)
+ *   NEUROLINK_PROXY_KEEPALIVE_MAX_MS    keep-alive upper bound  (default 600000)
+ *   NEUROLINK_PROXY_MAX_CONNECTIONS     max pooled connections per origin (default 64)
+ */
+import { Agent, setGlobalDispatcher } from "undici";
+import { logger } from "../utils/logger.js";
+function readPositiveInt(name, fallback) {
+    const raw = process.env[name];
+    if (!raw) {
+        return fallback;
+    }
+    const parsed = Number.parseInt(raw, 10);
+    return Number.isFinite(parsed) && parsed > 0 ? parsed : fallback;
+}
+let configured = false;
+/**
+ * Install the tuned global undici dispatcher. Idempotent — safe to call once at
+ * proxy startup. No-op when `NEUROLINK_PROXY_KEEPALIVE` is off/false/0.
+ */
+export function configureProxyKeepAliveDispatcher() {
+    if (configured) {
+        return;
+    }
+    configured = true;
+    const toggle = (process.env.NEUROLINK_PROXY_KEEPALIVE ?? "").toLowerCase();
+    if (toggle === "off" || toggle === "false" || toggle === "0") {
+        logger.debug("[proxy] keep-alive dispatcher disabled via env");
+        return;
+    }
+    const keepAliveTimeout = readPositiveInt("NEUROLINK_PROXY_KEEPALIVE_MS", 30_000);
+    const keepAliveMaxTimeout = readPositiveInt("NEUROLINK_PROXY_KEEPALIVE_MAX_MS", 600_000);
+    const connections = readPositiveInt("NEUROLINK_PROXY_MAX_CONNECTIONS", 64);
+    setGlobalDispatcher(new Agent({
+        keepAliveTimeout,
+        keepAliveMaxTimeout,
+        connections,
+        pipelining: 1,
+    }));
+    logger.debug(`[proxy] tuned undici dispatcher installed ` +
+        `(keepAliveTimeout=${keepAliveTimeout}ms, ` +
+        `keepAliveMaxTimeout=${keepAliveMaxTimeout}ms, connections=${connections})`);
+}

package/dist/types/generate.d.ts

@@ -602,13 +602,17 @@ export type AdditionalMemoryUser = {
  *
  * - `completed` — the model finished on its own (text answer or final_result)
  * - `step-cap` — the `maxSteps` budget ran out while the model still wanted tools
+ * - `context-cap` — the in-loop context guard stopped the tool loop because the
+ *   accumulated conversation approached the model's context window (and the
+ *   terminal synthesis could not produce an answer); without the guard these
+ *   turns died mid-loop on a provider 400 "prompt is too long"
  * - `time-limit` — the `turnTimeoutMs` wall-clock deadline passed
  * - `stalled` — no progress (no chunk, no tool start/finish) for `stallTimeoutMs`
  * - `aborted` — the caller's `abortSignal` ended the turn
  * - `provider-error` — the provider/model failed the turn (e.g. persistent
  *   MALFORMED_FUNCTION_CALL after retry); usually worth a caller-side retry
  */
-export type GenerateStopReason = "completed" | "step-cap" | "time-limit" | "stalled" | "aborted" | "provider-error";
+export type GenerateStopReason = "completed" | "step-cap" | "context-cap" | "time-limit" | "stalled" | "aborted" | "provider-error";
 /**
  * Generate function result type - Primary output format
  * Future-ready for multi-modal outputs while maintaining text focus

package/dist/utils/async/index.d.ts

@@ -20,4 +20,4 @@
  */
 export { delay, sleep } from "./delay.js";
 export { calculateBackoff, createRetry, DEFAULT_RETRY_OPTIONS, RetryExhaustedError, retry, } from "./retry.js";
-export { TimeoutError, withTimeout, withTimeoutFn } from "./withTimeout.js";
+export { TimeoutError, raceWithAbort, withTimeout, withTimeoutFn, } from "./withTimeout.js";

package/dist/utils/async/index.js

@@ -20,4 +20,4 @@
  */
 export { delay, sleep } from "./delay.js";
 export { calculateBackoff, createRetry, DEFAULT_RETRY_OPTIONS, RetryExhaustedError, retry, } from "./retry.js";
-export { TimeoutError, withTimeout, withTimeoutFn } from "./withTimeout.js";
+export { TimeoutError, raceWithAbort, withTimeout, withTimeoutFn, } from "./withTimeout.js";

package/dist/utils/async/withTimeout.d.ts

@@ -49,6 +49,18 @@ export declare class TimeoutError extends Error {
  * ```
  */
 export declare function withTimeout<T>(promise: Promise<T>, ms: number, message?: string): Promise<T>;
+/**
+ * Race a promise against an AbortSignal so the caller observes an abort
+ * IMMEDIATELY instead of waiting for the operation to settle (or for a
+ * separate timeout to expire). Rejects with an abort-shaped error
+ * (`name === "AbortError"`) so provider loops route it to their existing
+ * cancellation handling.
+ *
+ * The underlying operation is NOT cancelled — it continues as a bounded
+ * ghost (the same tradeoff as {@link withTimeout}); its eventual settlement
+ * is swallowed so it can never surface as an unhandled rejection.
+ */
+export declare function raceWithAbort<T>(promise: Promise<T>, signal: AbortSignal | undefined): Promise<T>;
 /**
  * Execute a function with timeout protection.
  *

package/dist/utils/async/withTimeout.js

@@ -72,6 +72,51 @@ export async function withTimeout(promise, ms, message) {
         }
     }
 }
+/**
+ * Race a promise against an AbortSignal so the caller observes an abort
+ * IMMEDIATELY instead of waiting for the operation to settle (or for a
+ * separate timeout to expire). Rejects with an abort-shaped error
+ * (`name === "AbortError"`) so provider loops route it to their existing
+ * cancellation handling.
+ *
+ * The underlying operation is NOT cancelled — it continues as a bounded
+ * ghost (the same tradeoff as {@link withTimeout}); its eventual settlement
+ * is swallowed so it can never surface as an unhandled rejection.
+ */
+export async function raceWithAbort(promise, signal) {
+    if (!signal || typeof signal.addEventListener !== "function") {
+        return promise;
+    }
+    const makeAbortError = () => {
+        const e = new Error("Operation aborted");
+        e.name = "AbortError";
+        return e;
+    };
+    if (signal.aborted) {
+        promise.catch(() => {
+            // Swallow the abandoned settlement — it must never surface as an
+            // unhandled rejection after the race has already been decided.
+        });
+        return Promise.reject(makeAbortError());
+    }
+    return new Promise((resolve, reject) => {
+        const onAbort = () => {
+            promise.catch(() => {
+                // Swallow the abandoned settlement — it must never surface as an
+                // unhandled rejection after the race has already been decided.
+            });
+            reject(makeAbortError());
+        };
+        signal.addEventListener("abort", onAbort, { once: true });
+        promise.then((value) => {
+            signal.removeEventListener("abort", onAbort);
+            resolve(value);
+        }, (error) => {
+            signal.removeEventListener("abort", onAbort);
+            reject(error);
+        });
+    });
+}
 /**
  * Execute a function with timeout protection.
  *

package/dist/utils/mcpErrorText.d.ts

@@ -24,3 +24,14 @@ export declare function extractMcpErrorText(raw: unknown): string;
  * but no readable text is present, falls back to a generic message.
  */
 export declare function extractMcpToolErrorMessage(result: unknown): string | undefined;
+/**
+ * Detect a tool result that REPORTS failure without throwing, for the native
+ * loops' consecutive-failure breaker. Covers the two shapes NeuroLink itself
+ * produces or forwards:
+ * - MCP CallToolResult with `isError: true` (e.g. proxy-blocked tools) —
+ *   delegated to {@link extractMcpToolErrorMessage}
+ * - our own `{ error: "..." }` payloads
+ * Plain strings are never classified (too false-positive-prone).
+ * Returns the error text, or null when the result looks like a success.
+ */
+export declare function extractToolFailureText(result: unknown): string | null;

package/dist/utils/mcpErrorText.js

@@ -69,3 +69,26 @@ export function extractMcpToolErrorMessage(result) {
         ? `MCP tool returned isError: ${text}`
         : "MCP tool returned isError: true";
 }
+/**
+ * Detect a tool result that REPORTS failure without throwing, for the native
+ * loops' consecutive-failure breaker. Covers the two shapes NeuroLink itself
+ * produces or forwards:
+ * - MCP CallToolResult with `isError: true` (e.g. proxy-blocked tools) —
+ *   delegated to {@link extractMcpToolErrorMessage}
+ * - our own `{ error: "..." }` payloads
+ * Plain strings are never classified (too false-positive-prone).
+ * Returns the error text, or null when the result looks like a success.
+ */
+export function extractToolFailureText(result) {
+    const mcpError = extractMcpToolErrorMessage(result);
+    if (mcpError) {
+        return mcpError;
+    }
+    if (result !== null &&
+        typeof result === "object" &&
+        typeof result.error === "string" &&
+        result.error.length > 0) {
+        return result.error;
+    }
+    return null;
+}

package/dist/utils/systemMessages.d.ts

@@ -0,0 +1,29 @@
+import type { ModelMessage, SystemModelMessage } from "../types/index.js";
+/**
+ * Partition a built message array into system messages and the rest, so the
+ * system prompt can ride `generateText`'s top-level `system` option instead of
+ * the `messages` array.
+ *
+ * The AI SDK deprecates system-role entries inside `messages` (warns by default
+ * from ai@6.0.170 via the `allowSystemInMessages` option, rejected by default
+ * in v7, flagged as a prompt-injection risk). The `system` option accepts full
+ * SystemModelMessage objects, so per-message providerOptions (e.g. the anthropic
+ * cacheControl breakpoint set by buildMessagesArray) survive the move. See
+ * issue #1024.
+ *
+ * Order is preserved within both partitions and the input is not mutated.
+ *
+ * Guard: the partition only runs when it leaves a non-empty `messages` array.
+ * If every message is a system message (a system-only priming call) — or there
+ * are no system messages at all — the original array is returned untouched with
+ * `system: undefined`. The AI SDK rejects an empty `messages` array, and a
+ * system-only call has no hoisted form (the SDK also requires a non-system
+ * message), so that degenerate case is deliberately left on the old
+ * system-in-`messages` path rather than made to throw — it keeps whatever
+ * behaviour it had before #1024. The normal system-plus-conversation path is
+ * always hoisted.
+ */
+export declare function extractSystemMessages(messages: ModelMessage[]): {
+    system: SystemModelMessage[] | undefined;
+    messages: ModelMessage[];
+};

package/dist/utils/systemMessages.js

@@ -0,0 +1,44 @@
+/**
+ * Partition a built message array into system messages and the rest, so the
+ * system prompt can ride `generateText`'s top-level `system` option instead of
+ * the `messages` array.
+ *
+ * The AI SDK deprecates system-role entries inside `messages` (warns by default
+ * from ai@6.0.170 via the `allowSystemInMessages` option, rejected by default
+ * in v7, flagged as a prompt-injection risk). The `system` option accepts full
+ * SystemModelMessage objects, so per-message providerOptions (e.g. the anthropic
+ * cacheControl breakpoint set by buildMessagesArray) survive the move. See
+ * issue #1024.
+ *
+ * Order is preserved within both partitions and the input is not mutated.
+ *
+ * Guard: the partition only runs when it leaves a non-empty `messages` array.
+ * If every message is a system message (a system-only priming call) — or there
+ * are no system messages at all — the original array is returned untouched with
+ * `system: undefined`. The AI SDK rejects an empty `messages` array, and a
+ * system-only call has no hoisted form (the SDK also requires a non-system
+ * message), so that degenerate case is deliberately left on the old
+ * system-in-`messages` path rather than made to throw — it keeps whatever
+ * behaviour it had before #1024. The normal system-plus-conversation path is
+ * always hoisted.
+ */
+export function extractSystemMessages(messages) {
+    const system = [];
+    const rest = [];
+    for (const message of messages) {
+        if (message.role === "system") {
+            system.push(message);
+        }
+        else {
+            rest.push(message);
+        }
+    }
+    // Only hoist when a non-empty messages array remains. A system-only call (or
+    // a no-system array) passes through unchanged: hoisting would empty
+    // `messages`, which the SDK rejects, and a system-only call has no compliant
+    // hoisted form.
+    if (system.length === 0 || rest.length === 0) {
+        return { system: undefined, messages };
+    }
+    return { system, messages: rest };
+}