@juspay/neurolink 9.50.1 → 9.51.0

package/dist/cli/commands/proxy.js

@@ -26,7 +26,14 @@ const PROXY_TELEMETRY_SCRIPT_PATH = fileURLToPath(new URL("../../../scripts/obse
 // =============================================================================
 // STATE MANAGEMENT
 // =============================================================================
-const proxyStateManager = new StateFileManager("proxy-state.json");
+let proxyStateManager = new StateFileManager("proxy-state.json");
+/**
+ * Reinitialise the state manager with a custom base directory.
+ * Called when --dev redirects writable paths to .neurolink-dev/.
+ */
+function setProxyStateDir(baseDir) {
+    proxyStateManager = new StateFileManager("proxy-state.json", baseDir);
+}
 function saveProxyState(state) {
     proxyStateManager.save(state);
 }
@@ -333,12 +340,12 @@ async function loadProxyStartEnv(argv, spinner) {
         process.exit(1);
     }
 }
-async function createProxyNeurolinkRuntime() {
+async function createProxyNeurolinkRuntime(logsDir) {
     process.env.NEUROLINK_SKIP_MCP = "true";
     const { NeuroLink } = await import("../../lib/neurolink.js");
     const neurolink = new NeuroLink();
     const { initRequestLogger, cleanupLogs } = await import("../../lib/proxy/requestLogger.js");
-    initRequestLogger(true);
+    initRequestLogger(true, logsDir);
     cleanupLogs(7, 500);
     return { neurolink, cleanupLogs };
 }
@@ -701,7 +708,7 @@ function registerProxyShutdownHandlers(params) {
         catch {
             // non-fatal — proxy shutdown must not block on OTel
         }
-        if (signal === "SIGINT") {
+        if (signal === "SIGINT" && !params.isDev) {
             try {
                 const shutdownHost = params.host === "0.0.0.0" ? "localhost" : params.host;
                 await clearClaudeProxySettings(`http://${shutdownHost}:${params.port}`);
@@ -733,7 +740,11 @@ async function startProxyRuntime(params) {
         port: params.port,
         hostname: params.host,
     });
-    const guardPid = spawnFailOpenGuard(params.host, params.port, process.pid);
+    // Skip the fail-open guard in dev mode — it monitors the proxy and clears
+    // global Claude settings on exit, which is exactly what we want to avoid.
+    const guardPid = params.argv.dev
+        ? undefined
+        : spawnFailOpenGuard(params.host, params.port, process.pid);
     const readinessHost = params.host === "0.0.0.0" ? "127.0.0.1" : params.host;
     await waitForProxyReadiness({
         host: readinessHost,
@@ -767,10 +778,16 @@ async function startProxyRuntime(params) {
     if (params.spinner) {
         params.spinner.succeed(chalk.green("Claude proxy started successfully"));
     }
+    const isDev = params.argv.dev ?? false;
     const normalizedHost = params.host === "0.0.0.0" ? "localhost" : params.host;
     const url = `http://${normalizedHost}:${params.port}`;
     printProxyBanner(url, params.strategy);
-    logger.always(`  ${chalk.bold("Mode:")}       ${chalk.cyan(params.passthrough ? "passthrough" : "full")}`);
+    if (isDev) {
+        logger.always(`  ${chalk.bold("Mode:")}       ${chalk.magenta("dev (isolated — state in .neurolink-dev/)")}`);
+    }
+    else {
+        logger.always(`  ${chalk.bold("Mode:")}       ${chalk.cyan(params.passthrough ? "passthrough" : "full")}`);
+    }
     if (params.passthrough) {
         logger.always(chalk.yellow("  ! Passthrough mode forwards client auth directly to Anthropic"));
         logger.always(chalk.dim("    Stored proxy OAuth/API credentials are ignored; clients need their own valid Anthropic auth."));
@@ -778,29 +795,52 @@ async function startProxyRuntime(params) {
     if (params.loadedEnvFile) {
         logger.always(`  ${chalk.bold("Env File:")}   ${chalk.cyan(params.loadedEnvFile)}`);
     }
-    try {
-        await setClaudeProxySettings(url);
-        logger.always(chalk.green("  ✓ Auto-configured Claude Code settings"));
-        logger.always(chalk.dim("    Restart Claude Code to connect through proxy"));
+    if (!isDev) {
+        try {
+            await setClaudeProxySettings(url);
+            logger.always(chalk.green("  ✓ Auto-configured Claude Code settings"));
+            logger.always(chalk.dim("    Restart Claude Code to connect through proxy"));
+        }
+        catch (error) {
+            logger.debug("[proxy] Failed to auto-configure Claude Code: " +
+                (error instanceof Error ? error.message : String(error)));
+        }
     }
-    catch (error) {
-        logger.debug("[proxy] Failed to auto-configure Claude Code: " +
-            (error instanceof Error ? error.message : String(error)));
+    else {
+        logger.always(chalk.dim("  ⊘ Dev mode: skipping client auto-configuration"));
     }
     const maintenance = startProxyBackgroundMaintenance(params.cleanupLogs);
     registerProxyShutdownHandlers({
         server,
         host: params.host,
         port: params.port,
+        isDev,
         ...maintenance,
     });
 }
 async function startProxyCommandHandler(argv) {
     const spinner = argv.quiet ? null : ora("Starting Claude proxy...").start();
+    const isDev = argv.dev ?? false;
     try {
-        await ensureProxyStartAllowed(spinner);
+        // In dev mode: redirect writable state to .neurolink-dev/ and skip singleton check
+        let devPaths;
+        if (isDev) {
+            const { resolveProxyPaths } = await import("../../lib/proxy/proxyPaths.js");
+            devPaths = resolveProxyPaths(true);
+            setProxyStateDir(devPaths.stateDir);
+            const { initAccountQuota } = await import("../../lib/proxy/accountQuota.js");
+            initAccountQuota(devPaths.quotaFile);
+            // Ensure the dev state directory exists
+            const { mkdirSync, existsSync } = await import("fs");
+            if (!existsSync(devPaths.stateDir)) {
+                mkdirSync(devPaths.stateDir, { recursive: true, mode: 0o700 });
+            }
+        }
+        if (!isDev) {
+            await ensureProxyStartAllowed(spinner);
+        }
         const loadedEnvFile = await loadProxyStartEnv(argv, spinner);
-        const { neurolink, cleanupLogs } = await createProxyNeurolinkRuntime();
+        const { neurolink, cleanupLogs } = await createProxyNeurolinkRuntime(devPaths?.logsDir);
         const { proxyConfig, strategy, modelRouter, passthrough } = await loadProxyStartConfiguration(argv, spinner);
         if (spinner) {
             spinner.text = "Configuring server...";
@@ -904,6 +944,11 @@ export const proxyStartCommand = {
             type: "boolean",
             default: false,
             description: "Run in transparent passthrough mode (no retry, no rotation, no polyfill)",
+        })
+            .option("dev", {
+            type: "boolean",
+            default: false,
+            description: "Run in isolated dev mode — state files scoped to .neurolink-dev/ in cwd, no client auto-configuration, no singleton check",
         })
             .example("neurolink proxy start", "Start proxy on default port 55669 with fill-first strategy")
             .example("neurolink proxy start -p 8080 -s fill-first", "Start proxy on port 8080 with fill-first")

package/dist/cli/utils/serverUtils.d.ts

@@ -44,8 +44,9 @@ export declare class StateFileManager<T> {
     /**
      * Create a new state file manager
      * @param filename - Name of the state file (e.g., "serve-state.json")
+     * @param baseDir - Optional base directory (defaults to ~/.neurolink)
      */
-    constructor(filename: string);
+    constructor(filename: string, baseDir?: string);
     /**
      * Get the full path to the state file
      */

package/dist/cli/utils/serverUtils.js

@@ -92,9 +92,10 @@ export class StateFileManager {
     /**
      * Create a new state file manager
      * @param filename - Name of the state file (e.g., "serve-state.json")
+     * @param baseDir - Optional base directory (defaults to ~/.neurolink)
      */
-    constructor(filename) {
-        this.filePath = path.join(getNeuroLinkDir(), filename);
+    constructor(filename, baseDir) {
+        this.filePath = path.join(baseDir ?? getNeuroLinkDir(), filename);
     }
     /**
      * Get the full path to the state file
@@ -107,7 +108,10 @@ export class StateFileManager {
      * @param state - State object to save
      */
     save(state) {
-        ensureStateDir();
+        const dir = path.dirname(this.filePath);
+        if (!fs.existsSync(dir)) {
+            fs.mkdirSync(dir, { recursive: true, mode: 0o700 });
+        }
         fs.writeFileSync(this.filePath, JSON.stringify(state, null, 2));
     }
     /**

package/dist/context/contextCompactor.js

@@ -23,7 +23,7 @@ const DEFAULT_CONFIG = {
     enableSummarize: true,
     enableTruncate: true,
     pruneProtectTokens: 40_000,
-    pruneMinimumSavings: 20_000,
+    pruneMinimumSavings: 500,
     pruneProtectedTools: ["skill"],
     summarizationProvider: "vertex",
     summarizationModel: "gemini-2.5-flash",
@@ -151,7 +151,7 @@ export class ContextCompactor {
                     targetTokens: targetTokens,
                     provider: provider,
                     adaptiveBuffer: 0.15,
-                    maxIterations: 3,
+                    maxIterations: 6,
                 });
                 if (truncResult.truncated) {
                     currentMessages = truncResult.messages;

package/dist/context/stages/slidingWindowTruncator.d.ts

@@ -7,7 +7,7 @@
  *
  * Features:
  * - Adaptive truncation (PERF-001): calculates fraction from actual overage
- *   instead of fixed 50%, with iterative refinement up to 3 passes.
+ *   instead of fixed 50%, with iterative refinement up to 6 passes.
  * - Small conversation handling (BUG-005): for <= 4 messages, truncates
  *   message content proportionally instead of returning no-op.
  */

package/dist/context/stages/slidingWindowTruncator.js

@@ -7,7 +7,7 @@
  *
  * Features:
  * - Adaptive truncation (PERF-001): calculates fraction from actual overage
- *   instead of fixed 50%, with iterative refinement up to 3 passes.
+ *   instead of fixed 50%, with iterative refinement up to 6 passes.
  * - Small conversation handling (BUG-005): for <= 4 messages, truncates
  *   message content proportionally instead of returning no-op.
  */
@@ -153,8 +153,8 @@ export function truncateWithSlidingWindow(messages, config) {
                     messagesRemoved: evenRemoveCount,
                 };
             }
-            // Not enough -- increase fraction by 25% for next iteration
-            currentFraction = Math.min(0.95, currentFraction + 0.25);
+            // Not enough -- increase fraction by 10% for finer-grained escalation
+            currentFraction = Math.min(0.95, currentFraction + 0.1);
             continue;
         }
         // No token targets -- single-pass with calculated fraction

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

@@ -57,6 +57,11 @@ export declare class Utilities {
      * Supports number or string formats (e.g., '30s', '2m', '1h')
      */
     getTimeout(options: TextGenerationOptions | StreamOptions): number;
+    /**
+     * Get timeout scaled by estimated input token count.
+     * For large contexts (>100K tokens), increase timeout proportionally.
+     */
+    getContextAwareTimeout(options: TextGenerationOptions | StreamOptions, estimatedTokens?: number): number;
     /**
      * Check if a schema is a Zod schema
      */

package/dist/core/modules/Utilities.js

@@ -19,7 +19,7 @@
 import { z } from "zod";
 import { logger } from "../../utils/logger.js";
 import { getSafeMaxTokens } from "../../utils/tokenLimits.js";
-import { TimeoutError } from "../../utils/timeout.js";
+import { TimeoutError, getDefaultTimeout, parseTimeout, } from "../../utils/timeout.js";
 import { validateStreamOptions as validateStreamOpts, validateTextGenerationOptions, ValidationError, createValidationSummary, } from "../../utils/parameterValidation.js";
 import { STEP_LIMITS } from "../constants.js";
 /**
@@ -157,25 +157,36 @@ export class Utilities {
      * Supports number or string formats (e.g., '30s', '2m', '1h')
      */
     getTimeout(options) {
-        if (!options.timeout) {
-            return this.defaultTimeout;
-        }
-        if (typeof options.timeout === "number") {
-            return options.timeout;
-        }
-        // Parse string timeout (e.g., '30s', '2m', '1h')
-        const timeoutStr = options.timeout.toLowerCase();
-        const value = parseInt(timeoutStr);
-        if (timeoutStr.includes("h")) {
-            return value * 60 * 60 * 1000;
-        }
-        else if (timeoutStr.includes("m")) {
-            return value * 60 * 1000;
+        // If caller specified a timeout, use it (supports number ms and string formats)
+        if (options.timeout !== undefined && options.timeout !== null) {
+            const parsed = parseTimeout(options.timeout);
+            if (parsed !== undefined) {
+                return parsed;
+            }
         }
-        else if (timeoutStr.includes("s")) {
-            return value * 1000;
+        // Use per-provider default (e.g., vertex=60s, ollama=5m) instead of global 30s.
+        // Always use "generate" operation here — streaming operations have their own
+        // longer timeout (DEFAULT_TIMEOUTS.streaming = 2m) applied by the streaming
+        // infrastructure in BaseProvider.stream(). Both TextGenerationOptions and
+        // StreamOptions share the same `input` property, so there is no reliable
+        // discriminator to detect streaming at this level.
+        const providerDefault = parseTimeout(getDefaultTimeout(this.providerName, "generate"));
+        return providerDefault ?? this.defaultTimeout;
+    }
+    /**
+     * Get timeout scaled by estimated input token count.
+     * For large contexts (>100K tokens), increase timeout proportionally.
+     */
+    getContextAwareTimeout(options, estimatedTokens) {
+        const baseTimeout = this.getTimeout(options);
+        if (!estimatedTokens || estimatedTokens <= 100_000) {
+            return baseTimeout;
         }
-        return this.defaultTimeout;
+        // Scale: >100K → 1.5x, >200K → 2x, >300K → 2.5x (capped at 4x)
+        // Use (estimatedTokens - 1) so exact multiples stay in the lower tier
+        // (e.g., 100_000 → 1x, 100_001 → 1.5x)
+        const scale = 1 + Math.floor((estimatedTokens - 1) / 100_000) * 0.5;
+        return Math.round(baseTimeout * Math.min(scale, 4));
     }
     /**
      * Check if a schema is a Zod schema

package/dist/lib/context/contextCompactor.js

@@ -23,7 +23,7 @@ const DEFAULT_CONFIG = {
     enableSummarize: true,
     enableTruncate: true,
     pruneProtectTokens: 40_000,
-    pruneMinimumSavings: 20_000,
+    pruneMinimumSavings: 500,
     pruneProtectedTools: ["skill"],
     summarizationProvider: "vertex",
     summarizationModel: "gemini-2.5-flash",
@@ -151,7 +151,7 @@ export class ContextCompactor {
                     targetTokens: targetTokens,
                     provider: provider,
                     adaptiveBuffer: 0.15,
-                    maxIterations: 3,
+                    maxIterations: 6,
                 });
                 if (truncResult.truncated) {
                     currentMessages = truncResult.messages;

package/dist/lib/context/stages/slidingWindowTruncator.d.ts

@@ -7,7 +7,7 @@
  *
  * Features:
  * - Adaptive truncation (PERF-001): calculates fraction from actual overage
- *   instead of fixed 50%, with iterative refinement up to 3 passes.
+ *   instead of fixed 50%, with iterative refinement up to 6 passes.
  * - Small conversation handling (BUG-005): for <= 4 messages, truncates
  *   message content proportionally instead of returning no-op.
  */

package/dist/lib/context/stages/slidingWindowTruncator.js

@@ -7,7 +7,7 @@
  *
  * Features:
  * - Adaptive truncation (PERF-001): calculates fraction from actual overage
- *   instead of fixed 50%, with iterative refinement up to 3 passes.
+ *   instead of fixed 50%, with iterative refinement up to 6 passes.
  * - Small conversation handling (BUG-005): for <= 4 messages, truncates
  *   message content proportionally instead of returning no-op.
  */
@@ -153,8 +153,8 @@ export function truncateWithSlidingWindow(messages, config) {
                     messagesRemoved: evenRemoveCount,
                 };
             }
-            // Not enough -- increase fraction by 25% for next iteration
-            currentFraction = Math.min(0.95, currentFraction + 0.25);
+            // Not enough -- increase fraction by 10% for finer-grained escalation
+            currentFraction = Math.min(0.95, currentFraction + 0.1);
             continue;
         }
         // No token targets -- single-pass with calculated fraction

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

@@ -57,6 +57,11 @@ export declare class Utilities {
      * Supports number or string formats (e.g., '30s', '2m', '1h')
      */
     getTimeout(options: TextGenerationOptions | StreamOptions): number;
+    /**
+     * Get timeout scaled by estimated input token count.
+     * For large contexts (>100K tokens), increase timeout proportionally.
+     */
+    getContextAwareTimeout(options: TextGenerationOptions | StreamOptions, estimatedTokens?: number): number;
     /**
      * Check if a schema is a Zod schema
      */

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

@@ -19,7 +19,7 @@
 import { z } from "zod";
 import { logger } from "../../utils/logger.js";
 import { getSafeMaxTokens } from "../../utils/tokenLimits.js";
-import { TimeoutError } from "../../utils/timeout.js";
+import { TimeoutError, getDefaultTimeout, parseTimeout, } from "../../utils/timeout.js";
 import { validateStreamOptions as validateStreamOpts, validateTextGenerationOptions, ValidationError, createValidationSummary, } from "../../utils/parameterValidation.js";
 import { STEP_LIMITS } from "../constants.js";
 /**
@@ -157,25 +157,36 @@ export class Utilities {
      * Supports number or string formats (e.g., '30s', '2m', '1h')
      */
     getTimeout(options) {
-        if (!options.timeout) {
-            return this.defaultTimeout;
-        }
-        if (typeof options.timeout === "number") {
-            return options.timeout;
-        }
-        // Parse string timeout (e.g., '30s', '2m', '1h')
-        const timeoutStr = options.timeout.toLowerCase();
-        const value = parseInt(timeoutStr);
-        if (timeoutStr.includes("h")) {
-            return value * 60 * 60 * 1000;
-        }
-        else if (timeoutStr.includes("m")) {
-            return value * 60 * 1000;
+        // If caller specified a timeout, use it (supports number ms and string formats)
+        if (options.timeout !== undefined && options.timeout !== null) {
+            const parsed = parseTimeout(options.timeout);
+            if (parsed !== undefined) {
+                return parsed;
+            }
         }
-        else if (timeoutStr.includes("s")) {
-            return value * 1000;
+        // Use per-provider default (e.g., vertex=60s, ollama=5m) instead of global 30s.
+        // Always use "generate" operation here — streaming operations have their own
+        // longer timeout (DEFAULT_TIMEOUTS.streaming = 2m) applied by the streaming
+        // infrastructure in BaseProvider.stream(). Both TextGenerationOptions and
+        // StreamOptions share the same `input` property, so there is no reliable
+        // discriminator to detect streaming at this level.
+        const providerDefault = parseTimeout(getDefaultTimeout(this.providerName, "generate"));
+        return providerDefault ?? this.defaultTimeout;
+    }
+    /**
+     * Get timeout scaled by estimated input token count.
+     * For large contexts (>100K tokens), increase timeout proportionally.
+     */
+    getContextAwareTimeout(options, estimatedTokens) {
+        const baseTimeout = this.getTimeout(options);
+        if (!estimatedTokens || estimatedTokens <= 100_000) {
+            return baseTimeout;
         }
-        return this.defaultTimeout;
+        // Scale: >100K → 1.5x, >200K → 2x, >300K → 2.5x (capped at 4x)
+        // Use (estimatedTokens - 1) so exact multiples stay in the lower tier
+        // (e.g., 100_000 → 1x, 100_001 → 1.5x)
+        const scale = 1 + Math.floor((estimatedTokens - 1) / 100_000) * 0.5;
+        return Math.round(baseTimeout * Math.min(scale, 4));
     }
     /**
      * Check if a schema is a Zod schema

package/dist/lib/mcp/externalServerManager.d.ts

@@ -36,6 +36,11 @@ export declare class ExternalServerManager extends EventEmitter {
      * Get current HITL manager
      */
     getHITLManager(): HITLManager | undefined;
+    /**
+     * Resolve the human-readable server name for an event payload.
+     * Falls back to serverId if the instance or config.name isn't available.
+     */
+    getServerName(serverId: string): string;
     /**
      * Load MCP server configurations from .mcp-config.json file with parallel loading support
      * Automatically registers servers found in the configuration

package/dist/lib/mcp/externalServerManager.js

@@ -194,10 +194,16 @@ export class ExternalServerManager extends EventEmitter {
         this.toolDiscovery = new ToolDiscoveryService();
         // Forward tool discovery events
         this.toolDiscovery.on("toolRegistered", (event) => {
-            this.emit("toolDiscovered", event);
+            this.emit("toolDiscovered", {
+                ...event,
+                serverName: this.getServerName(event.serverId),
+            });
         });
         this.toolDiscovery.on("toolUnregistered", (event) => {
-            this.emit("toolRemoved", event);
+            this.emit("toolRemoved", {
+                ...event,
+                serverName: this.getServerName(event.serverId),
+            });
         });
         // Handle process cleanup
         process.on("SIGINT", () => this.shutdown());
@@ -223,6 +229,14 @@ export class ExternalServerManager extends EventEmitter {
     getHITLManager() {
         return this.hitlManager;
     }
+    /**
+     * Resolve the human-readable server name for an event payload.
+     * Falls back to serverId if the instance or config.name isn't available.
+     */
+    getServerName(serverId) {
+        const instance = this.servers.get(serverId);
+        return instance?.config?.name || serverId;
+    }
     /**
      * Load MCP server configurations from .mcp-config.json file with parallel loading support
      * Automatically registers servers found in the configuration
@@ -712,6 +726,8 @@ export class ExternalServerManager extends EventEmitter {
                 };
             }
             mcpLogger.info(`[ExternalServerManager] Removing server: ${serverId}`);
+            // Capture name before deletion removes the instance
+            const serverName = this.getServerName(serverId);
             // Stop the server
             await this.stopServer(serverId);
             // Remove from registry
@@ -719,6 +735,7 @@ export class ExternalServerManager extends EventEmitter {
             // Emit event
             this.emit("disconnected", {
                 serverId,
+                serverName,
                 reason: "Manually removed",
                 timestamp: new Date(),
             });
@@ -816,6 +833,7 @@ export class ExternalServerManager extends EventEmitter {
             // Emit connected event
             this.emit("connected", {
                 serverId,
+                serverName: this.getServerName(serverId),
                 toolCount: instance.toolsMap.size,
                 timestamp: new Date(),
             });
@@ -921,6 +939,7 @@ export class ExternalServerManager extends EventEmitter {
         // Emit status change event
         this.emit("statusChanged", {
             serverId,
+            serverName: this.getServerName(serverId),
             oldStatus,
             newStatus,
             timestamp: new Date(),
@@ -941,6 +960,7 @@ export class ExternalServerManager extends EventEmitter {
         // Emit failed event
         this.emit("failed", {
             serverId,
+            serverName: this.getServerName(serverId),
             error: error.message,
             timestamp: new Date(),
         });
@@ -965,6 +985,7 @@ export class ExternalServerManager extends EventEmitter {
         // Emit disconnected event
         this.emit("disconnected", {
             serverId,
+            serverName: this.getServerName(serverId),
             reason,
             timestamp: new Date(),
         });
@@ -1078,6 +1099,7 @@ export class ExternalServerManager extends EventEmitter {
             // Emit health check event
             this.emit("healthCheck", {
                 serverId,
+                serverName: this.getServerName(serverId),
                 health,
                 timestamp: new Date(),
             });

package/dist/lib/neurolink.js

@@ -50,7 +50,7 @@ import { createMemoryRetrievalTools } from "./memory/memoryRetrievalTools.js";
 import { getMetricsAggregator, MetricsAggregator, } from "./observability/metricsAggregator.js";
 import { SpanStatus, SpanType } from "./observability/types/spanTypes.js";
 import { SpanSerializer } from "./observability/utils/spanSerializer.js";
-import { flushOpenTelemetry, getLangfuseHealthStatus, initializeOpenTelemetry, isOpenTelemetryInitialized, setLangfuseContext, shutdownOpenTelemetry, } from "./services/server/ai/observability/instrumentation.js";
+import { flushOpenTelemetry, getLangfuseHealthStatus, initializeOpenTelemetry, isOpenTelemetryInitialized, runWithCurrentLangfuseContext, setLangfuseContext, shutdownOpenTelemetry, } from "./services/server/ai/observability/instrumentation.js";
 import { TaskManager } from "./tasks/taskManager.js";
 import { createTaskTools } from "./tasks/tools/taskTools.js";
 import { ATTR } from "./telemetry/attributes.js";
@@ -1129,7 +1129,10 @@ Current user's request: ${currentInput}`;
      * Supports additional users with per-user prompt and maxWords overrides.
      */
     storeMemoryInBackground(originalPrompt, responseContent, userId, additionalUsers) {
-        setImmediate(async () => {
+        // Preserve AsyncLocalStorage context across setImmediate boundary so that
+        // memory writes appear under the originating Langfuse trace instead of
+        // becoming orphan spans.
+        const wrappedMemoryWrite = runWithCurrentLangfuseContext(async () => {
             try {
                 const client = this.ensureMemoryReady();
                 if (!client) {
@@ -1145,12 +1148,18 @@ Current user's request: ${currentInput}`;
                         : undefined;
                     writeOps.push(client.add(user.userId, content, addOptions));
                 }
-                await Promise.all(writeOps);
+                // withTimeout races against Promise.all — if the timeout fires, the
+                // await resolves with an error but the underlying client.add() calls
+                // may still complete in the background. This is acceptable: the memory
+                // client API (Mem0) doesn't support AbortSignal, and these are
+                // fire-and-forget background writes where a stale completion is harmless.
+                await withTimeout(Promise.all(writeOps), 30_000, new Error("Background memory write timed out after 30s"));
             }
             catch (error) {
                 logger.warn("Memory storage failed:", error);
             }
         });
+        setImmediate(wrappedMemoryWrite);
     }
     /**
      * Set up HITL event forwarding to main emitter
@@ -3723,6 +3732,21 @@ Current user's request: ${currentInput}`;
             conversationMessageCount: conversationMessages.length,
             shouldCompact: budgetResult.shouldCompact,
         });
+        // Scale timeout for large contexts if caller didn't set one explicitly.
+        // Providers read options.timeout via getTimeout(), so setting it here
+        // propagates to any downstream provider call.
+        if (options.timeout === undefined &&
+            budgetResult.estimatedInputTokens > 100_000) {
+            // >100K → 1.5x, >200K → 2x, >300K → 2.5x (capped at 4x) of 60s base
+            const scale = 1 + Math.floor((budgetResult.estimatedInputTokens - 1) / 100_000) * 0.5;
+            const scaledMs = Math.round(60_000 * Math.min(scale, 4));
+            options.timeout = scaledMs;
+            logger.info("[TokenBudget] Scaled timeout for large context", {
+                requestId,
+                estimatedTokens: budgetResult.estimatedInputTokens,
+                scaledTimeoutMs: scaledMs,
+            });
+        }
         const compactionSessionId = this.getCompactionSessionId(options);
         const lastCompactionCount = this.lastCompactionMessageCount.get(compactionSessionId) ?? 0;
         if (!budgetResult.shouldCompact ||
@@ -3798,6 +3822,8 @@ Current user's request: ${currentInput}`;
             toolDefinitions: availableTools,
         });
         if (!finalBudget.withinBudget) {
+            // Clear watermark so handleContextOverflow recovery can re-compact
+            this.lastCompactionMessageCount.delete(compactionSessionId);
             throw new ContextBudgetExceededError(`Context exceeds model budget after all compaction stages. ` +
                 `Estimated: ${finalBudget.estimatedInputTokens} tokens, ` +
                 `Budget: ${finalBudget.availableInputTokens} tokens. ` +
@@ -3993,6 +4019,8 @@ Current user's request: ${currentInput}`;
                                 : undefined,
                         });
                         if (!finalBudget.withinBudget) {
+                            // Clear watermark so handleContextOverflow recovery can re-compact
+                            this.lastCompactionMessageCount.delete(dpgCompactionSessionId);
                             throw new ContextBudgetExceededError(`Context exceeds model budget after all compaction stages. ` +
                                 `Estimated: ${finalBudget.estimatedInputTokens} tokens, ` +
                                 `Budget: ${finalBudget.availableInputTokens} tokens.`, {
@@ -5016,6 +5044,8 @@ Current user's request: ${currentInput}`;
                     toolDefinitions: availableTools,
                 });
                 if (!finalBudget.withinBudget) {
+                    // Clear watermark so handleContextOverflow recovery can re-compact
+                    this.lastCompactionMessageCount.delete(streamCompactionSessionId);
                     throw new ContextBudgetExceededError(`Stream context exceeds model budget after all compaction stages. ` +
                         `Estimated: ${finalBudget.estimatedInputTokens} tokens, ` +
                         `Budget: ${finalBudget.availableInputTokens} tokens.`, {
@@ -7508,6 +7538,7 @@ Current user's request: ${currentInput}`;
                 // Emit server added event
                 this.emitter.emit("externalMCP:serverAdded", {
                     serverId,
+                    serverName: config.name || serverId,
                     config,
                     toolCount: result.metadata?.toolsDiscovered || 0,
                     timestamp: Date.now(),
@@ -7535,12 +7566,15 @@ Current user's request: ${currentInput}`;
         this.invalidateToolCache(); // Invalidate cache when an external server is removed
         try {
             mcpLogger.info(`[NeuroLink] Removing external MCP server: ${serverId}`);
+            // Capture the configured name before removal destroys the instance
+            const serverName = this.externalServerManager.getServerName(serverId);
             const result = await this.externalServerManager.removeServer(serverId);
             if (result.success) {
                 mcpLogger.info(`[NeuroLink] External MCP server removed successfully: ${serverId}`);
                 // Emit server removed event
                 this.emitter.emit("externalMCP:serverRemoved", {
                     serverId,
+                    serverName,
                     timestamp: Date.now(),
                 });
             }