@juspay/neurolink 9.51.2 → 9.51.4

package/dist/memory/memoryRetrievalTools.js

@@ -9,6 +9,7 @@
 import { tool } from "ai";
 import { z } from "zod";
 import { logger } from "../utils/logger.js";
+import { withTimeout } from "../utils/errorHandling.js";
 import { SpanSerializer, SpanType, SpanStatus, } from "../observability/index.js";
 import { getMetricsAggregator } from "../observability/index.js";
 /** Maximum characters returned per retrieval request */
@@ -19,18 +20,36 @@ const MAX_RETRIEVAL_LIMIT = 200_000;
 const MAX_SEARCH_MATCHES = 50;
 /**
  * Factory function that creates memory retrieval tools bound to a memory manager.
- * @param memoryManager - The Redis conversation memory manager instance
+ *
+ * @param memoryManager  Redis conversation memory manager instance.
+ * @param artifactStore  Optional artifact store for externalized MCP outputs.
+ *                       When provided, retrieve_context gains an `artifactId`
+ *                       parameter that fetches the full payload written by
+ *                       McpOutputNormalizer under strategy="externalize".
  * @returns Record of tool name to Vercel AI SDK tool definition
  */
-export function createMemoryRetrievalTools(memoryManager) {
+export function createMemoryRetrievalTools(memoryManager, artifactStore) {
     return {
         retrieve_context: tool({
-            description: "Retrieve messages from conversation memory. Use this to access full tool " +
-                "outputs when a result was truncated, review previous assistant responses, " +
-                "or search through conversation history. Supports filtering by role, " +
-                "pagination for large content, and regex search within messages.",
+            description: "Retrieve messages from conversation memory, or fetch the full payload of " +
+                "an externalized MCP tool output by artifact ID. Use this to:\n" +
+                "• Access full tool outputs when a result was truncated or externalized\n" +
+                "• Review previous assistant responses\n" +
+                "• Search through conversation history\n" +
+                "Supports filtering by role, pagination for large content, and regex search.\n" +
+                "To fetch an externalized artifact, provide `artifactId` (omit sessionId).",
             inputSchema: z.object({
-                sessionId: z.string().describe("Session ID for the conversation"),
+                sessionId: z
+                    .string()
+                    .optional()
+                    .describe("Session ID for conversation history retrieval. " +
+                    "Required unless artifactId is provided."),
+                artifactId: z
+                    .string()
+                    .optional()
+                    .describe("Artifact ID from an externalized MCP tool output " +
+                    "(visible in the tool output as neurolinkArtifactId=<id>). " +
+                    "When provided, returns the full stored payload directly."),
                 messageId: z
                     .string()
                     .optional()
@@ -64,19 +83,68 @@ export function createMemoryRetrievalTools(memoryManager) {
                     "Returns matching lines with line numbers."),
             }),
             execute: async (args) => {
+                // ── Artifact resolution path ────────────────────────────────────────
+                // When the caller supplies an artifactId we short-circuit to the
+                // artifact store (bypassing Redis) and return the full payload with
+                // optional offset/limit pagination.
+                if (args.artifactId) {
+                    if (!artifactStore) {
+                        logger.warn("[MemoryRetrievalTools] retrieve_context called with artifactId " +
+                            "but no ArtifactStore is configured");
+                        return {
+                            error: "Artifact store not configured — " +
+                                "mcp.outputLimits.strategy must be set to 'externalize' to use artifactId retrieval",
+                            artifactId: args.artifactId,
+                        };
+                    }
+                    const content = await withTimeout(artifactStore.retrieve(args.artifactId), 10_000, new Error(`ArtifactStore.retrieve() timed out for artifact "${args.artifactId}"`));
+                    if (content === null) {
+                        return {
+                            error: "Artifact not found or has expired",
+                            artifactId: args.artifactId,
+                        };
+                    }
+                    const charLimit = Math.min(args.limit ?? DEFAULT_RETRIEVAL_LIMIT, MAX_RETRIEVAL_LIMIT);
+                    const start = args.offset ?? 0;
+                    const slice = content.slice(start, start + charLimit);
+                    return {
+                        artifactId: args.artifactId,
+                        content: slice,
+                        totalSize: content.length,
+                        hasMore: start + charLimit < content.length,
+                        offset: start,
+                        limit: charLimit,
+                    };
+                }
+                // ── End artifact resolution ─────────────────────────────────────────
+                if (!args.sessionId) {
+                    return {
+                        error: "sessionId is required when artifactId is not provided",
+                    };
+                }
+                if (!memoryManager) {
+                    return {
+                        error: "Session history retrieval requires Redis conversation memory — " +
+                            "enable mcp.conversationMemory with a Redis backend, or use " +
+                            "artifactId to retrieve an externalized MCP tool output.",
+                    };
+                }
                 const span = SpanSerializer.createSpan(SpanType.MEMORY, "memory.retrieve", {
                     "memory.operation": "retrieve",
                     "memory.store": "redis",
                     "memory.query": args.search || args.messageId || `lastN:${args.lastN ?? "all"}`,
                 });
                 const startTime = Date.now();
+                // args.sessionId is guaranteed non-null here — we returned early above
+                // when it was missing. Cast via string coercion to satisfy eslint.
+                const sessionId = String(args.sessionId);
                 try {
-                    const conversation = await memoryManager.getSessionRaw(args.sessionId);
+                    const conversation = await withTimeout(memoryManager.getSessionRaw(sessionId), 10_000, new Error(`getSessionRaw() timed out for session "${sessionId}"`));
                     if (!conversation) {
                         span.durationMs = Date.now() - startTime;
                         const endedSpan = SpanSerializer.endSpan(span, SpanStatus.OK);
                         getMetricsAggregator().recordSpan(endedSpan);
-                        return { error: "Session not found", sessionId: args.sessionId };
+                        return { error: "Session not found", sessionId };
                     }
                     let messages = conversation.messages;
                     // Filter by specific messageId

package/dist/neurolink.d.ts

@@ -47,6 +47,8 @@ export declare class NeuroLink {
     private mcpToolBatcher?;
     private mcpEnhancedDiscovery?;
     private mcpToolMiddlewares;
+    /** Artifact store for externalized MCP tool outputs (set when strategy=externalize). */
+    private mcpArtifactStore?;
     private _disableToolCacheForCurrentRequest;
     private mcpEnhancementsConfig?;
     private toolCircuitBreakers;

package/dist/neurolink.js

@@ -40,6 +40,8 @@ import { ToolCallBatcher } from "./mcp/batching/index.js";
 import { ToolResultCache } from "./mcp/caching/index.js";
 import { EnhancedToolDiscovery } from "./mcp/enhancedToolDiscovery.js";
 import { ExternalServerManager } from "./mcp/externalServerManager.js";
+import { McpOutputNormalizer, DEFAULT_MAX_MCP_OUTPUT_BYTES, DEFAULT_WARN_MCP_OUTPUT_BYTES, } from "./mcp/mcpOutputNormalizer.js";
+import { LocalTempArtifactStore } from "./artifacts/artifactStore.js";
 import { ToolRouter } from "./mcp/routing/index.js";
 // Import direct tools server for automatic registration
 import { directToolsServer } from "./mcp/servers/agent/directToolsServer.js";
@@ -216,6 +218,8 @@ export class NeuroLink {
     mcpToolBatcher;
     mcpEnhancedDiscovery;
     mcpToolMiddlewares = [];
+    /** Artifact store for externalized MCP tool outputs (set when strategy=externalize). */
+    mcpArtifactStore;
     _disableToolCacheForCurrentRequest = false;
     mcpEnhancementsConfig;
     // Enhanced error handling support
@@ -775,17 +779,18 @@ export class NeuroLink {
     initializeMCPEnhancements(config) {
         const mcpConfig = config?.mcp;
         this.mcpEnhancementsConfig = mcpConfig;
-        // ToolCache — disabled by default, opt-in
-        if (mcpConfig?.cache?.enabled) {
+        // BZ-664: ToolCache — enabled by default to prevent duplicate tool calls.
+        // Callers can explicitly opt out via mcp.cache.enabled = false.
+        if (mcpConfig?.cache?.enabled !== false) {
             this.mcpToolResultCache = new ToolResultCache({
-                ttl: mcpConfig.cache.ttl ?? 300_000,
-                maxSize: mcpConfig.cache.maxSize ?? 500,
-                strategy: mcpConfig.cache.strategy ?? "lru",
+                ttl: mcpConfig?.cache?.ttl ?? 300_000,
+                maxSize: mcpConfig?.cache?.maxSize ?? 500,
+                strategy: mcpConfig?.cache?.strategy ?? "lru",
             });
             logger.debug("[NeuroLink] MCP tool result cache initialized", {
-                ttl: mcpConfig.cache.ttl ?? 300_000,
-                maxSize: mcpConfig.cache.maxSize ?? 500,
-                strategy: mcpConfig.cache.strategy ?? "lru",
+                ttl: mcpConfig?.cache?.ttl ?? 300_000,
+                maxSize: mcpConfig?.cache?.maxSize ?? 500,
+                strategy: mcpConfig?.cache?.strategy ?? "lru",
             });
         }
         // ToolCallBatcher — disabled by default, opt-in
@@ -817,6 +822,25 @@ export class NeuroLink {
             });
         }
         // ToolRouter — lazy-initialized when 2+ external servers exist (see addExternalMCPServer)
+        // McpOutputNormalizer — active when mcp.outputLimits is configured
+        if (mcpConfig?.outputLimits) {
+            const strategy = mcpConfig.outputLimits.strategy ?? "externalize";
+            const maxBytes = mcpConfig.outputLimits.maxBytes ?? DEFAULT_MAX_MCP_OUTPUT_BYTES;
+            const warnBytes = mcpConfig.outputLimits.warnBytes ?? DEFAULT_WARN_MCP_OUTPUT_BYTES;
+            let artifactStore;
+            if (strategy === "externalize") {
+                artifactStore = new LocalTempArtifactStore();
+                this.mcpArtifactStore = artifactStore;
+                logger.debug("[NeuroLink] MCP artifact store initialized (local-temp)");
+            }
+            const normalizer = new McpOutputNormalizer({ strategy, maxBytes, warnBytes }, artifactStore);
+            this.externalServerManager.setOutputNormalizer(normalizer);
+            logger.debug("[NeuroLink] MCP output normalizer initialized", {
+                strategy,
+                maxBytes,
+                warnBytes,
+            });
+        }
     }
     /**
      * Register file reference tools with the MCP tool registry.
@@ -936,90 +960,46 @@ export class NeuroLink {
                 "redis" in memConfig &&
                 !!memConfig.redis) ||
             process.env.STORAGE_TYPE === "redis";
-        if (!memConfig?.enabled || !hasRedisConfig) {
-            logger.debug("[NeuroLink] Skipping memory retrieval tools — requires Redis conversation memory");
+        const hasArtifactStore = !!this.mcpArtifactStore;
+        // Register when Redis is configured OR when an artifact store exists.
+        // Artifact store alone is sufficient for the artifactId retrieval path —
+        // session history retrieval just returns a clear error when Redis is absent.
+        if ((!memConfig?.enabled || !hasRedisConfig) && !hasArtifactStore) {
+            logger.debug("[NeuroLink] Skipping memory retrieval tools — requires Redis conversation memory or an artifact store");
             return;
         }
-        const tools = {
-            retrieve_context: {
-                description: "Retrieve messages from conversation memory. Use this to access full tool " +
-                    "outputs when a result was truncated, review previous assistant responses, " +
-                    "or search through conversation history.",
-                execute: async (params) => {
-                    // Lazy access: conversationMemory is initialized on first generate() call
-                    const memoryManager = this.conversationMemory;
-                    if (!memoryManager || !("getSessionRaw" in memoryManager)) {
-                        return {
-                            success: false,
-                            error: "Memory retrieval not available — Redis memory manager not initialized",
-                            metadata: {
-                                toolName: "retrieve_context",
-                                serverId: "direct",
-                                executionTime: 0,
-                            },
-                        };
-                    }
-                    const actualTools = createMemoryRetrievalTools(memoryManager);
-                    const result = await actualTools.retrieve_context.execute(params, {
-                        toolCallId: "memory-retrieval",
-                        messages: [],
-                    });
-                    // Check if the tool itself reported an error
-                    const hasError = result &&
-                        typeof result === "object" &&
-                        "error" in result &&
-                        !("messages" in result);
-                    const errorMsg = hasError
-                        ? result.error
-                        : undefined;
-                    return {
-                        success: !hasError,
-                        data: result,
-                        ...(errorMsg ? { error: errorMsg } : {}),
-                        metadata: {
-                            toolName: "retrieve_context",
-                            serverId: "direct",
-                            executionTime: 0,
-                        },
-                    };
-                },
+        // Extract the canonical tool definition (schema + description) from the
+        // memoryRetrievalTools factory. We pass undefined as the memoryManager here
+        // because we only need the Zod inputSchema and description at registration
+        // time — the actual manager is resolved lazily at execution time.
+        const canonicalTools = createMemoryRetrievalTools(undefined, this.mcpArtifactStore);
+        const retrieveContextDef = canonicalTools.retrieve_context;
+        // Register via this.registerTool() so the tool ends up in the "user-defined"
+        // category inside toolRegistry. getCustomTools() returns that category, which
+        // is what ToolsManager reads to build the tool schema sent to the LLM.
+        // (Tools registered via toolRegistry.registerTool() directly land in the
+        // "built-in" category and are never included in the LLM's tool schema.)
+        this.registerTool("retrieve_context", {
+            name: "retrieve_context",
+            description: retrieveContextDef.description ?? "Retrieve context or artifacts",
+            // Pass the Zod schema so ToolsManager gives the LLM full parameter types.
+            // registerTool() detects isZodSchema on inputSchema and preserves it.
+            inputSchema: retrieveContextDef
+                .inputSchema,
+            execute: async (params) => {
+                // Lazy: conversationMemory is initialized on the first generate() call.
+                // When only an artifact store is present (no Redis), memoryManager is
+                // undefined — createMemoryRetrievalTools handles that via an explicit guard.
+                const memoryManager = this.conversationMemory;
+                const tools = createMemoryRetrievalTools(memoryManager, this.mcpArtifactStore);
+                // Return the result directly so the LLM receives clean output instead
+                // of a nested { success, data, metadata } wrapper.
+                // Bounded by TOOL_TIMEOUTS.EXECUTION_DEFAULT_MS so a stalled Redis or
+                // filesystem backend never hangs the tool call indefinitely.
+                return await withTimeout(tools.retrieve_context.execute(params, { toolCallId: "memory-retrieval", messages: [] }), TOOL_TIMEOUTS.EXECUTION_DEFAULT_MS, ErrorFactory.toolTimeout("retrieve_context", TOOL_TIMEOUTS.EXECUTION_DEFAULT_MS));
             },
-        };
-        const registrations = Object.entries(tools).map(async ([toolName, toolDef]) => {
-            const toolId = `direct.${toolName}`;
-            const toolInfo = {
-                name: toolName,
-                description: toolDef.description,
-                inputSchema: {},
-                serverId: "direct",
-                category: "built-in",
-            };
-            await this.toolRegistry.registerTool(toolId, toolInfo, {
-                execute: async (params) => {
-                    try {
-                        return await toolDef.execute(params);
-                    }
-                    catch (error) {
-                        // Known limitation: this non-throwing error path returns
-                        // { success: false } without recording errorCategories in
-                        // toolExecutionMetrics. These are internal memory-tool failures
-                        // (low frequency), so the risk of metric gaps is minimal.
-                        // A full fix would require access to the metrics map here,
-                        // which is not available in the registration closure.
-                        return {
-                            success: false,
-                            error: error instanceof Error ? error.message : String(error),
-                            metadata: { toolName, serverId: "direct", executionTime: 0 },
-                        };
-                    }
-                },
-                description: toolDef.description,
-                inputSchema: {},
-            });
-        });
-        void Promise.all(registrations).then(() => {
-            logger.info("[NeuroLink] Memory retrieval tools registered");
         });
+        logger.info("[NeuroLink] Memory retrieval tools registered");
     }
     /** Format memory context for prompt inclusion */
     formatMemoryContext(memoryContext, currentInput) {
@@ -7628,7 +7608,36 @@ Current user's request: ${currentInput}`;
     async executeExternalMCPTool(serverId, toolName, parameters, options) {
         try {
             mcpLogger.debug(`[NeuroLink] Executing external MCP tool: ${toolName} on ${serverId}`);
+            // BZ-664: Check existing ToolResultCache before executing to avoid
+            // duplicate identical calls within the same session.
+            //
+            // Safety guards aligned with executeToolInternal():
+            // - Skip destructive tools (destructiveHint annotation)
+            // - Scope cache key by serverId (two servers can expose same tool name)
+            //   and toolExecutionContext (prevents cross-session/user leaks)
+            const toolAnnotations = this.getToolAnnotationsForExecution(toolName);
+            const cacheEnabled = !!this.mcpToolResultCache &&
+                !this._disableToolCacheForCurrentRequest &&
+                !toolAnnotations?.destructiveHint;
+            const cacheKeyArgs = {
+                __serverId: serverId,
+                __args: parameters,
+                ...(this.toolExecutionContext
+                    ? { __ctx: this.toolExecutionContext }
+                    : {}),
+            };
+            if (cacheEnabled && this.mcpToolResultCache) {
+                const cached = this.mcpToolResultCache.getCachedResult(toolName, cacheKeyArgs);
+                if (cached !== undefined) {
+                    mcpLogger.debug(`[NeuroLink] Tool result cache HIT: ${toolName} on ${serverId}`);
+                    return cached;
+                }
+            }
             const result = await this.externalServerManager.executeTool(serverId, toolName, parameters, options);
+            // BZ-664: Store result in cache after successful execution
+            if (cacheEnabled && this.mcpToolResultCache) {
+                this.mcpToolResultCache.cacheResult(toolName, cacheKeyArgs, result);
+            }
             mcpLogger.debug(`[NeuroLink] External MCP tool executed successfully: ${toolName}`);
             return result;
         }

package/dist/session/globalSessionState.js

@@ -1,6 +1,33 @@
 import { nanoid } from "nanoid";
 import { NeuroLink } from "../neurolink.js";
 import { buildObservabilityConfigFromEnv } from "../utils/observabilityHelpers.js";
+/**
+ * Build mcp.outputLimits config from environment variables.
+ * Reads NEUROLINK_MCP_OUTPUT_STRATEGY and NEUROLINK_MCP_MAX_OUTPUT_BYTES.
+ * Returns undefined when neither variable is set (no overhead).
+ */
+function buildMcpOutputLimitsFromEnv() {
+    const strategyRaw = process.env.NEUROLINK_MCP_OUTPUT_STRATEGY;
+    const maxBytesRaw = process.env.NEUROLINK_MCP_MAX_OUTPUT_BYTES;
+    const warnBytesRaw = process.env.NEUROLINK_MCP_WARN_OUTPUT_BYTES;
+    if (!strategyRaw && !maxBytesRaw) {
+        return undefined;
+    }
+    const strategy = strategyRaw === "inline" || strategyRaw === "externalize"
+        ? strategyRaw
+        : "externalize"; // safe default when only maxBytes is set
+    const maxBytes = maxBytesRaw ? parseInt(maxBytesRaw, 10) : undefined;
+    const warnBytes = warnBytesRaw ? parseInt(warnBytesRaw, 10) : undefined;
+    return {
+        strategy,
+        ...(maxBytes !== undefined && Number.isFinite(maxBytes) && maxBytes >= 0
+            ? { maxBytes }
+            : {}),
+        ...(warnBytes !== undefined && Number.isFinite(warnBytes) && warnBytes >= 0
+            ? { warnBytes }
+            : {}),
+    };
+}
 export class GlobalSessionManager {
     static instance;
     loopSession = null;
@@ -25,6 +52,14 @@ export class GlobalSessionManager {
         if (observabilityConfig) {
             neurolinkOptions.observability = observabilityConfig;
         }
+        // Add MCP output limits from environment variables (CLI usage)
+        const mcpOutputLimits = buildMcpOutputLimitsFromEnv();
+        if (mcpOutputLimits) {
+            neurolinkOptions.mcp = {
+                ...neurolinkOptions.mcp,
+                outputLimits: mcpOutputLimits,
+            };
+        }
         this.loopSession = {
             neurolinkInstance: new NeuroLink(neurolinkOptions),
             sessionId,
@@ -99,7 +134,15 @@ export class GlobalSessionManager {
         }
         // Create new NeuroLink with observability config from environment (CLI usage)
         const observabilityConfig = buildObservabilityConfigFromEnv();
-        return new NeuroLink(observabilityConfig ? { observability: observabilityConfig } : undefined);
+        const mcpOutputLimits = buildMcpOutputLimitsFromEnv();
+        const options = {};
+        if (observabilityConfig) {
+            options.observability = observabilityConfig;
+        }
+        if (mcpOutputLimits) {
+            options.mcp = { outputLimits: mcpOutputLimits };
+        }
+        return new NeuroLink(Object.keys(options).length ? options : undefined);
     }
     getCurrentSessionId() {
         return this.getLoopSession()?.sessionId;

package/dist/types/artifactTypes.d.ts

@@ -0,0 +1,63 @@
+/**
+ * Artifact Store Types (canonical location)
+ *
+ * Types for the MCP large-output artifact storage system.
+ * When mcp.outputLimits.strategy = "externalize", oversized MCP tool outputs
+ * are stored as artifacts and the model receives a compact surrogate instead.
+ *
+ * @module types/artifactTypes
+ */
+/** Metadata recorded alongside a stored artifact. */
+export type ArtifactMeta = {
+    /** Tool name that produced the output. */
+    toolName: string;
+    /** MCP server ID. */
+    serverId: string;
+    /** Session that triggered the tool call (optional). */
+    sessionId?: string;
+    /** Serialized byte size of the full payload. */
+    sizeBytes: number;
+    /** Whether the payload is valid JSON or plain text. */
+    contentType: "json" | "text";
+    /** Unix epoch ms when the artifact was created. */
+    createdAt: number;
+};
+/** Lightweight descriptor returned after a successful ArtifactStore.store(). */
+export type ArtifactRef = {
+    /** UUID v4 — stable identifier used in surrogate results and metadata. */
+    id: string;
+    /** First N characters of the payload (for surrogate headers). */
+    preview: string;
+    /** Full serialized byte size. */
+    sizeBytes: number;
+    /** Stored metadata. */
+    meta: ArtifactMeta;
+};
+/**
+ * Pluggable storage contract for externalized MCP tool outputs.
+ *
+ * Default backend: LocalTempArtifactStore (filesystem, single-process).
+ * Future backends can implement this interface for S3, Redis blobs, etc.
+ */
+export interface ArtifactStore {
+    /**
+     * Persist a payload and return a lightweight reference.
+     * @param payload  Serialized tool output (JSON string or plain text).
+     * @param meta     Descriptor without `createdAt` (assigned internally).
+     */
+    store(payload: string, meta: Omit<ArtifactMeta, "createdAt">): Promise<ArtifactRef>;
+    /**
+     * Retrieve the full payload by artifact ID.
+     * Returns `null` if the artifact is not found or has been cleaned up.
+     */
+    retrieve(id: string): Promise<string | null>;
+    /** Delete a single artifact. No-op if the ID does not exist. */
+    delete(id: string): Promise<void>;
+    /**
+     * Delete all artifacts older than `olderThanMs` milliseconds.
+     * Returns the number of artifacts deleted.
+     */
+    cleanup(olderThanMs: number): Promise<number>;
+    /** Generate a short preview string from a serialized payload. */
+    generatePreview(payload: string): string;
+}

package/dist/types/artifactTypes.js

@@ -0,0 +1,10 @@
+/**
+ * Artifact Store Types (canonical location)
+ *
+ * Types for the MCP large-output artifact storage system.
+ * When mcp.outputLimits.strategy = "externalize", oversized MCP tool outputs
+ * are stored as artifacts and the model receives a compact surrogate instead.
+ *
+ * @module types/artifactTypes
+ */
+export {};

package/dist/types/configTypes.d.ts

@@ -44,7 +44,7 @@ export type NeurolinkConstructorConfig = {
  * Configuration for MCP enhancement modules wired into generate()/stream() paths.
  *
  * These modules are automatically applied during tool execution when configured:
- * - cache: Tool result caching (disabled by default)
+ * - cache: Tool result caching (enabled by default, opt out with enabled: false)
  * - annotations: Auto-infer tool safety metadata (enabled by default)
  * - router: Multi-server tool routing (auto-activates with 2+ servers)
  * - batcher: Batch programmatic tool calls (disabled by default)
@@ -52,7 +52,7 @@ export type NeurolinkConstructorConfig = {
  * - middleware: Global tool execution middleware chain (empty by default)
  */
 export type MCPEnhancementsConfig = {
-    /** Tool result caching. Default: disabled. Enable to cache read-only tool results. */
+    /** Tool result caching. Default: enabled. Set enabled: false to opt out. */
     cache?: {
         enabled?: boolean;
         /** Cache TTL in milliseconds. Default: 300000 (5 min) */
@@ -90,6 +90,38 @@ export type MCPEnhancementsConfig = {
     };
     /** Global tool middleware applied to every tool execution. Default: empty. */
     middleware?: ToolMiddleware[];
+    /**
+     * Large MCP tool output handling.
+     *
+     * MCP servers can return arbitrarily large payloads. Without limits these
+     * are loaded entirely into memory, cached in full, stored whole in Redis, and
+     * injected into the LLM context window — all of which silently fail at scale.
+     *
+     * When configured, NeuroLink intercepts oversized outputs at the tool boundary
+     * (before caching and before memory persistence) and applies the chosen
+     * strategy so the model receives a compact surrogate instead of a firehose.
+     *
+     * Two strategies:
+     *  - "inline"      Keep sending the full payload to the model regardless of
+     *                  size. A warning is emitted above warnBytes.
+     *  - "externalize" Store the full payload on disk as an artifact and return a
+     *                  compact surrogate with a head/tail preview and an artifact
+     *                  ID. The model uses `retrieve_context` with that ID to read
+     *                  the full output on demand, with offset/limit pagination.
+     *
+     * Defaults (when `outputLimits` is set):
+     *  strategy  = "externalize"
+     *  maxBytes  = 100 KB (100 * 1024)
+     *  warnBytes =  50 KB  (50 * 1024)
+     */
+    outputLimits?: {
+        /** What to do when output exceeds maxBytes. Default: "externalize". */
+        strategy?: "inline" | "externalize";
+        /** Byte ceiling above which the strategy fires. Default: 102400 (100 KB). */
+        maxBytes?: number;
+        /** Bytes at which a warning is emitted even when still inline. Default: 51200 (50 KB). */
+        warnBytes?: number;
+    };
 };
 /**
  * Authentication configuration for NeuroLink SDK

package/dist/types/conversation.d.ts

@@ -231,6 +231,13 @@ export type ChatMessageMetadata = {
     toolOutputPreview?: string;
     /** Original byte size of the full tool output before any truncation */
     originalSize?: number;
+    /**
+     * Artifact store ID for an externalized MCP tool output.
+     * Set when `mcp.outputLimits.strategy = "externalize"` and the tool output
+     * exceeded `maxBytes`. Use retrieve_context with this ID to fetch the full
+     * payload from the local artifact store.
+     */
+    artifactId?: string;
 };
 /**
  * Chat message format for conversation history

package/dist/types/index.d.ts

@@ -6,6 +6,8 @@ export * from "./cli.js";
 export * from "./common.js";
 export type { AnalyticsConfig, BackupInfo, BackupMetadata, CacheConfig, ConfigUpdateOptions, ConfigValidationResult, FallbackConfig, MCPEnhancementsConfig, NeuroLinkConfig, PerformanceConfig, RetryConfig, ToolConfig, } from "./configTypes.js";
 export type { ExternalMCPConfigValidation, ExternalMCPManagerConfig, ExternalMCPOperationResult, ExternalMCPServerEvents, ExternalMCPServerHealth, ExternalMCPServerInstance, ExternalMCPServerStatus, ExternalMCPToolContext, ExternalMCPToolInfo, ExternalMCPToolResult, } from "./externalMcp.js";
+export type { ArtifactMeta, ArtifactRef, ArtifactStore, } from "./artifactTypes.js";
+export type { McpOutputContext, McpOutputNormalizerConfig, McpOutputStrategy, NormalizedMcpOutput, } from "./mcpOutputTypes.js";
 export type { AuthorizationUrlResult, CircuitBreakerConfig, CircuitBreakerEvents, CircuitBreakerState, CircuitBreakerStats, DiscoveredMcp, ExternalToolExecutionOptions, FlexibleValidationResult, HTTPRetryConfig, MCPClientResult, MCPConnectedServer, MCPDiscoveredServer, MCPExecutableTool, MCPOAuthConfig, MCPServerCategory, MCPServerConfig, MCPServerConnectionStatus, MCPServerInfo, MCPServerMetadata, MCPServerRegistryEntry, MCPServerStatus, MCPStatus, MCPToolInfo, MCPToolMetadata, MCPTransportType, McpMetadata, McpRegistry, NeuroLinkExecutionContext, NeuroLinkMCPServer, NeuroLinkMCPTool, OAuthClientInformation, OAuthTokens as McpOAuthTokens, RateLimitConfig, TokenBucketRateLimitConfig, TokenExchangeRequest, TokenStorage, ToolDiscoveryResult, ToolRegistryEvents, ToolValidationResult, } from "./mcpTypes.js";
 export type { ModelCapability, ModelFilter, ModelPricing, ModelResolutionContext, ModelStats, ModelUseCase, } from "./providers.js";
 export * from "./providers.js";

package/dist/types/mcpOutputTypes.d.ts

@@ -0,0 +1,40 @@
+/**
+ * MCP Output Normalizer Types (canonical location)
+ *
+ * Types for the large MCP response handling system.
+ *
+ * @module types/mcpOutputTypes
+ */
+/**
+ * Two honest strategies for oversized MCP tool outputs:
+ *  - "inline"      Full payload always sent to the model (warning logged above warnBytes).
+ *  - "externalize" Full payload stored as an artifact; model receives a compact
+ *                  surrogate with head/tail preview and an artifact ID it can
+ *                  resolve via retrieve_context with offset/limit pagination.
+ */
+export type McpOutputStrategy = "inline" | "externalize";
+/** Configuration for McpOutputNormalizer. */
+export type McpOutputNormalizerConfig = {
+    strategy: McpOutputStrategy;
+    /** Byte ceiling above which the strategy fires. */
+    maxBytes: number;
+    /** Bytes at which a warning is emitted while still inline. */
+    warnBytes: number;
+};
+/** Contextual info passed alongside the raw MCP callResult. */
+export type McpOutputContext = {
+    toolName: string;
+    serverId: string;
+    sessionId?: string;
+};
+/** Value returned by McpOutputNormalizer.normalize(). */
+export type NormalizedMcpOutput = {
+    /** The result to substitute for the raw callResult. May be a surrogate. */
+    result: unknown;
+    /** Whether the full payload was written to the artifact store. */
+    isExternalized: boolean;
+    /** Artifact ID when isExternalized === true. */
+    artifactId?: string;
+    /** Serialized byte size of the original payload. */
+    originalBytes: number;
+};

package/dist/types/mcpOutputTypes.js

@@ -0,0 +1,8 @@
+/**
+ * MCP Output Normalizer Types (canonical location)
+ *
+ * Types for the large MCP response handling system.
+ *
+ * @module types/mcpOutputTypes
+ */
+export {};