@juspay/neurolink 9.81.3 → 9.82.0

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/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 };
+}

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@juspay/neurolink",
-  "version": "9.81.3",
+  "version": "9.82.0",
   "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": {
@@ -135,9 +135,10 @@
     "test:unit:vitest": "pnpm exec vitest run test/toolRouting.test.ts",
     "test:tool-routing": "pnpm run test:unit:vitest && npx tsx test/continuous-test-suite-tool-routing.ts",
     "test:classifier-router": "npx tsx test/continuous-test-suite-classifier-router.ts",
+    "test:system-messages:vitest": "pnpm exec vitest run test/systemMessages.test.ts",
     "test:tool-routing-semantic:vitest": "pnpm exec vitest run test/toolRoutingSemantic.test.ts",
     "test:tool-routing-semantic": "pnpm run test:tool-routing-semantic:vitest && npx tsx test/continuous-test-suite-tool-routing-semantic.ts",
-    "test:unit": "pnpm run test:envguard && pnpm run test:bugfixes && pnpm run test:mcp:infra && pnpm run test:mcp:bash && pnpm run test:mcp:limits && pnpm run test:mcp:spans && pnpm run test:autoresearch:redis && pnpm run test:unit:vitest && pnpm run test:tool-routing-cli:vitest && pnpm run test:tool-dedup:vitest && pnpm run test:model-pool:vitest && pnpm run test:tool-routing-semantic:vitest && pnpm run test:anthropic-tools-policy && pnpm run test:anthropic-multimodal && pnpm run test:excel-interop",
+    "test:unit": "pnpm run test:envguard && pnpm run test:bugfixes && pnpm run test:mcp:infra && pnpm run test:mcp:bash && pnpm run test:mcp:limits && pnpm run test:mcp:spans && pnpm run test:autoresearch:redis && pnpm run test:unit:vitest && pnpm run test:tool-routing-cli:vitest && pnpm run test:tool-dedup:vitest && pnpm run test:model-pool:vitest && pnpm run test:system-messages:vitest && pnpm run test:tool-routing-semantic:vitest && pnpm run test:anthropic-tools-policy && pnpm run test:anthropic-multimodal && pnpm run test:excel-interop",
     "// CI tier — live providers, runs only when API keys are present (test:credentials and test:dynamic make real provider calls when keys are set, so they live here, not in test:unit)": "",
     "test:live": "pnpm run test:providers && pnpm run test:mcp:http && pnpm run test:mcp:sdk && pnpm run test:mcp:cli && pnpm run test:observability && pnpm run test:context && pnpm run test:memory && pnpm run test:tool-reliability && pnpm run test:evaluation && pnpm run test:autoresearch && pnpm run test:credentials && pnpm run test:dynamic",
     "// CI tier — product output (image/video/TTS/PPT) — costs $$ per run": "",