@juspay/neurolink 9.10.1 → 9.12.0

package/dist/context/toolOutputLimits.js

@@ -1,84 +1,78 @@
 /**
- * Tool Output Size Limits
- *
- * Truncates tool outputs exceeding size limits.
- * Can save full output to disk with a pointer.
- * Modeled on OpenCode's approach.
+ * Tool output preview generation.
+ * Generates head/tail previews of large tool outputs for context-efficient LLM calls.
+ * @module
  */
-import { writeFileSync, mkdirSync } from "fs";
-import { join } from "path";
-import { randomUUID } from "crypto";
-import { tmpdir } from "os";
-/** Maximum tool output size in bytes (50KB) */
-export const MAX_TOOL_OUTPUT_BYTES = 50 * 1024;
-/** Maximum tool output lines */
-export const MAX_TOOL_OUTPUT_LINES = 2_000;
+/** Default maximum preview size in bytes (50KB) */
+export const DEFAULT_MAX_PREVIEW_BYTES = 50 * 1024;
+/** Default maximum preview lines */
+export const DEFAULT_MAX_PREVIEW_LINES = 2_000;
+/** Default head ratio (25% of preview budget) */
+export const DEFAULT_HEAD_RATIO = 0.25;
+/** Tool name referenced in truncation notices for on-demand full-output access */
+export const RETRIEVE_CONTEXT_TOOL_NAME = "retrieve_context";
+/** Default tail ratio (75% of preview budget) */
+export const DEFAULT_TAIL_RATIO = 0.75;
 /**
- * Truncate tool output if it exceeds size limits.
+ * Generate a head/tail preview of a tool output string.
+ * If the output is within limits, returns it unchanged with truncated: false.
+ * If over limits, keeps the first 25% and last 75% with an omission notice.
+ *
+ * Industry pattern: 25/75 head/tail split. Head captures schema/headers/structure,
+ * tail captures the most recent and typically most relevant data.
  */
-export function truncateToolOutput(output, options) {
-    const maxBytes = options?.maxBytes ?? MAX_TOOL_OUTPUT_BYTES;
-    const maxLines = options?.maxLines ?? MAX_TOOL_OUTPUT_LINES;
-    const direction = options?.direction ?? "tail";
-    const saveToDisk = options?.saveToDisk ?? false;
+export function generateToolOutputPreview(output, options) {
+    const maxBytes = options?.maxBytes ?? DEFAULT_MAX_PREVIEW_BYTES;
+    const maxLines = options?.maxLines ?? DEFAULT_MAX_PREVIEW_LINES;
+    const rawHeadRatio = options?.headRatio ?? DEFAULT_HEAD_RATIO;
+    const rawTailRatio = options?.tailRatio ?? DEFAULT_TAIL_RATIO;
+    // Clamp ratios to valid range to avoid negative omittedBytes
+    const headRatio = Math.max(0, Math.min(1, rawHeadRatio));
+    const tailRatio = Math.max(0, Math.min(1, rawTailRatio));
     const originalSize = Buffer.byteLength(output, "utf-8");
-    // Check byte limit
-    const exceedsBytes = originalSize > maxBytes;
-    // Check line limit
     const lines = output.split("\n");
+    const exceedsBytes = originalSize > maxBytes;
     const exceedsLines = lines.length > maxLines;
     if (!exceedsBytes && !exceedsLines) {
-        return { content: output, truncated: false, originalSize };
+        return { preview: output, truncated: false, originalSize };
     }
-    // Save to disk if requested
-    let savedPath;
-    if (saveToDisk) {
-        try {
-            const saveDir = options?.saveDir ?? join(tmpdir(), "neurolink-tool-output");
-            mkdirSync(saveDir, { recursive: true });
-            savedPath = join(saveDir, `tool-output-${randomUUID()}.txt`);
-            writeFileSync(savedPath, output, "utf-8");
-        }
-        catch {
-            // Silently fail disk save
-        }
-    }
-    // Apply truncation
-    let truncated;
+    // Line-based split
+    const headLineCount = Math.max(1, Math.floor(maxLines * headRatio));
+    const tailLineCount = Math.max(1, maxLines - headLineCount);
+    let head;
+    let tail;
     if (exceedsLines) {
-        if (direction === "head") {
-            truncated = lines.slice(0, maxLines).join("\n");
-        }
-        else {
-            truncated = lines.slice(-maxLines).join("\n");
-        }
+        head = lines.slice(0, headLineCount).join("\n");
+        tail = lines.slice(-tailLineCount).join("\n");
     }
     else {
-        truncated = output;
-    }
-    // Apply byte limit
-    if (Buffer.byteLength(truncated, "utf-8") > maxBytes) {
-        if (direction === "head") {
-            truncated = truncated.slice(0, maxBytes);
-        }
-        else {
-            truncated = truncated.slice(-maxBytes);
-        }
+        head = lines
+            .slice(0, Math.max(1, Math.floor(lines.length * headRatio)))
+            .join("\n");
+        tail = lines
+            .slice(-Math.max(1, Math.ceil(lines.length * tailRatio)))
+            .join("\n");
     }
-    // Add truncation notice
-    const notice = savedPath
-        ? `\n\n[Output truncated from ${originalSize} bytes to ${Buffer.byteLength(truncated, "utf-8")} bytes. Full output saved to: ${savedPath}]`
-        : `\n\n[Output truncated from ${originalSize} bytes to ${Buffer.byteLength(truncated, "utf-8")} bytes]`;
-    if (direction === "head") {
-        truncated = truncated + notice;
+    // Byte-based cap on each portion
+    const headMaxBytes = Math.floor(maxBytes * headRatio);
+    const tailMaxBytes = maxBytes - headMaxBytes;
+    if (Buffer.byteLength(head, "utf-8") > headMaxBytes) {
+        head = Buffer.from(head, "utf-8")
+            .subarray(0, headMaxBytes)
+            .toString("utf-8");
     }
-    else {
-        truncated = notice + "\n" + truncated;
+    if (Buffer.byteLength(tail, "utf-8") > tailMaxBytes) {
+        const tailBuf = Buffer.from(tail, "utf-8");
+        tail = tailBuf.subarray(tailBuf.length - tailMaxBytes).toString("utf-8");
     }
+    const omittedBytes = Math.max(0, originalSize -
+        Buffer.byteLength(head, "utf-8") -
+        Buffer.byteLength(tail, "utf-8"));
+    const notice = `\n\n[... ${omittedBytes} bytes omitted. ` +
+        `Use ${RETRIEVE_CONTEXT_TOOL_NAME} tool to access full output ...]\n\n`;
     return {
-        content: truncated,
+        preview: head + notice + tail,
         truncated: true,
-        savedPath,
         originalSize,
     };
 }

package/dist/core/baseProvider.d.ts

@@ -218,9 +218,18 @@ export declare abstract class BaseProvider implements AIProvider {
      */
     setSessionContext(sessionId?: string, userId?: string): void;
     /**
-     * Provider-specific error handling
+     * Provider-specific error formatting.
+     * Subclasses implement this to produce human-readable error messages
+     * (e.g., "❌ Google Vertex AI Provider Error\n\n...").
      */
-    protected abstract handleProviderError(error: unknown): Error;
+    protected abstract formatProviderError(error: unknown): Error;
+    /**
+     * Handle provider errors with abort passthrough.
+     * AbortErrors are never wrapped — they must propagate with their
+     * original identity so that isAbortError() can detect them in
+     * retry/fallback loops (directProviderGeneration, performMCPGenerationRetries).
+     */
+    protected handleProviderError(error: unknown): Error;
     /**
      * Image generation method. Providers that support it should override this.
      * By default, it throws an error indicating that the functionality is not supported.

package/dist/core/baseProvider.js

@@ -174,7 +174,7 @@ export class BaseProvider {
                 temperature: options.temperature,
                 maxTokens: options.maxTokens,
                 tools: options.tools, // 🔧 FIX: Pass user-provided tools (including RAG tools) to generation pipeline
-                disableTools: false,
+                disableTools: !!options.disableTools,
                 maxSteps: options.maxSteps || 5,
                 provider: options.provider,
                 model: options.model,
@@ -767,6 +767,21 @@ export class BaseProvider {
         this.userId = userId;
         this.toolsManager.setSessionContext(sessionId, userId);
     }
+    /**
+     * Handle provider errors with abort passthrough.
+     * AbortErrors are never wrapped — they must propagate with their
+     * original identity so that isAbortError() can detect them in
+     * retry/fallback loops (directProviderGeneration, performMCPGenerationRetries).
+     */
+    handleProviderError(error) {
+        if (isAbortError(error)) {
+            // Preserve AbortError identity — never wrap in provider-specific formatting
+            return error instanceof Error
+                ? error
+                : new DOMException("The operation was aborted", "AbortError");
+        }
+        return this.formatProviderError(error);
+    }
     /**
      * Image generation method. Providers that support it should override this.
      * By default, it throws an error indicating that the functionality is not supported.

package/dist/core/conversationMemoryManager.d.ts

@@ -50,7 +50,7 @@ export declare class ConversationMemoryManager implements IConversationMemoryMan
      * Returns messages from pointer onwards (or all if no pointer)
      * Now consistently async to match Redis implementation
      */
-    buildContextMessages(sessionId: string): Promise<ChatMessage[]>;
+    buildContextMessages(sessionId: string, _userId?: string, _enableSummarization?: boolean, requestId?: string): Promise<ChatMessage[]>;
     getSession(sessionId: string, _userId?: string): SessionMemory | undefined;
     createSummarySystemMessage(content: string, summarizesFrom?: string, summarizesTo?: string): ChatMessage;
     private ensureInitialized;
@@ -59,4 +59,16 @@ export declare class ConversationMemoryManager implements IConversationMemoryMan
     getStats(): Promise<ConversationMemoryStats>;
     clearSession(sessionId: string): Promise<boolean>;
     clearAllSessions(): Promise<void>;
+    /**
+     * Get the raw messages array for a session.
+     * Returns the full messages list without context filtering or summarization.
+     * Returns a deep copy to prevent external mutation of internal state.
+     */
+    getSessionMessages(sessionId: string, _userId?: string): Promise<ChatMessage[]>;
+    /**
+     * Replace the entire messages array for a session.
+     * Creates the session if it does not exist.
+     * Resets summary pointers since old pointers may reference messages that no longer exist.
+     */
+    setSessionMessages(sessionId: string, messages: ChatMessage[], userId?: string): Promise<void>;
 }

package/dist/core/conversationMemoryManager.js

@@ -92,11 +92,12 @@ export class ConversationMemoryManager {
                 if (!this.summarizationInProgress.has(options.sessionId)) {
                     setImmediate(async () => {
                         try {
-                            await this.checkAndSummarize(session, tokenThreshold);
+                            await this.checkAndSummarize(session, tokenThreshold, options.requestId);
                         }
                         catch (error) {
                             logger.error("Background summarization failed", {
                                 sessionId: session.sessionId,
+                                requestId: options.requestId,
                                 error: error instanceof Error ? error.message : String(error),
                             });
                         }
@@ -154,7 +155,7 @@ export class ConversationMemoryManager {
     /**
      * Check if summarization is needed based on token count
      */
-    async checkAndSummarize(session, threshold) {
+    async checkAndSummarize(session, threshold, requestId) {
         // Acquire lock - if already in progress, skip
         if (this.summarizationInProgress.has(session.sessionId)) {
             logger.debug("[ConversationMemoryManager] Summarization already in progress, skipping", {
@@ -164,7 +165,7 @@ export class ConversationMemoryManager {
         }
         this.summarizationInProgress.add(session.sessionId);
         try {
-            await this.summarizationEngine.checkAndSummarize(session, threshold, this.config, "[ConversationMemory]");
+            await this.summarizationEngine.checkAndSummarize(session, threshold, this.config, "[ConversationMemory]", requestId);
         }
         catch (error) {
             logger.error("Token counting or summarization failed", {
@@ -195,9 +196,9 @@ export class ConversationMemoryManager {
      * Returns messages from pointer onwards (or all if no pointer)
      * Now consistently async to match Redis implementation
      */
-    async buildContextMessages(sessionId) {
+    async buildContextMessages(sessionId, _userId, _enableSummarization, requestId) {
         const session = this.sessions.get(sessionId);
-        return session ? buildContextFromPointer(session) : [];
+        return session ? buildContextFromPointer(session, requestId) : [];
     }
     getSession(sessionId, _userId) {
         return this.sessions.get(sessionId);
@@ -263,4 +264,34 @@ export class ConversationMemoryManager {
         this.sessions.clear();
         logger.info("All sessions cleared", { clearedCount: sessionIds.length });
     }
+    /**
+     * Get the raw messages array for a session.
+     * Returns the full messages list without context filtering or summarization.
+     * Returns a deep copy to prevent external mutation of internal state.
+     */
+    async getSessionMessages(sessionId, _userId) {
+        await this.ensureInitialized();
+        const session = this.sessions.get(sessionId);
+        return session ? session.messages.map((msg) => ({ ...msg })) : [];
+    }
+    /**
+     * Replace the entire messages array for a session.
+     * Creates the session if it does not exist.
+     * Resets summary pointers since old pointers may reference messages that no longer exist.
+     */
+    async setSessionMessages(sessionId, messages, userId) {
+        await this.ensureInitialized();
+        let session = this.sessions.get(sessionId);
+        if (!session) {
+            session = this.createNewSession(sessionId, userId);
+            this.sessions.set(sessionId, session);
+            this.enforceSessionLimit();
+        }
+        session.messages = [...messages];
+        session.summarizedUpToMessageId = undefined;
+        session.summarizedMessage = undefined;
+        session.lastTokenCount = undefined;
+        session.lastCountedAt = undefined;
+        session.lastActivity = Date.now();
+    }
 }

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

@@ -38,6 +38,12 @@ export declare class GenerationHandler {
      * Execute the generation with AI SDK
      */
     executeGeneration(model: LanguageModelV1, messages: CoreMessage[], tools: Record<string, Tool>, options: TextGenerationOptions): Promise<Awaited<ReturnType<typeof generateText>>>;
+    /**
+     * Extract cache metrics from provider metadata (e.g. Anthropic's providerMetadata.anthropic)
+     * The Vercel AI SDK's LanguageModelUsage only has promptTokens/completionTokens/totalTokens.
+     * Cache metrics are surfaced via providerMetadata by provider-specific SDK adapters.
+     */
+    private extractCacheMetricsFromProviderMetadata;
     /**
      * Log generation completion information
      */

package/dist/core/modules/GenerationHandler.js

@@ -14,8 +14,24 @@
  */
 import { generateText, Output, NoObjectGeneratedError } from "ai";
 import { logger } from "../../utils/logger.js";
-import { extractTokenUsage } from "../../utils/tokenUtils.js";
+import { extractTokenUsage, extractCacheCreationTokens, extractCacheReadTokens, calculateCacheSavingsPercent, } from "../../utils/tokenUtils.js";
 import { DEFAULT_MAX_STEPS } from "../constants.js";
+/**
+ * Safely preview-serialize a value for debug logging.
+ * Handles undefined, circular references, and non-serializable values.
+ */
+function safePreview(v) {
+    if (v === undefined) {
+        return "";
+    }
+    try {
+        const text = typeof v === "string" ? v : JSON.stringify(v);
+        return (text ?? "").substring(0, 200);
+    }
+    catch {
+        return "[unserializable]";
+    }
+}
 /**
  * GenerationHandler class - Handles text generation operations for AI providers
  */
@@ -39,16 +55,41 @@ export class GenerationHandler {
     async callGenerateText(model, messages, tools, options, shouldUseTools, includeStructuredOutput) {
         // Check if this is a Google provider (for provider-specific options)
         const isGoogleProvider = this.providerName === "google-ai" || this.providerName === "vertex";
-        // Check if this is an Anthropic provider
-        const isAnthropicProvider = this.providerName === "anthropic" || this.providerName === "bedrock";
+        // Check if this is an Anthropic provider (includes Vertex+Claude)
+        const isAnthropicProvider = this.providerName === "anthropic" ||
+            this.providerName === "bedrock" ||
+            (this.providerName === "vertex" && this.modelName?.startsWith("claude-"));
         const useStructuredOutput = includeStructuredOutput &&
             !!options.schema &&
             (options.output?.format === "json" ||
                 options.output?.format === "structured");
+        // Annotate the last tool with cache_control so the full tool-definition
+        // block becomes a cache breakpoint for Anthropic-family providers.
+        // Non-Anthropic providers harmlessly ignore unknown providerOptions.
+        // Note: The AI SDK Tool type doesn't yet include providerOptions, so we
+        // use a type assertion. The Anthropic adapter reads this at runtime.
+        const toolsWithCache = { ...tools };
+        if (isAnthropicProvider &&
+            shouldUseTools &&
+            Object.keys(toolsWithCache).length > 0) {
+            const toolNames = Object.keys(toolsWithCache);
+            const lastToolName = toolNames[toolNames.length - 1];
+            if (lastToolName && toolsWithCache[lastToolName]) {
+                const lastTool = toolsWithCache[lastToolName];
+                toolsWithCache[lastToolName] = {
+                    ...lastTool,
+                    providerOptions: {
+                        ...(lastTool.providerOptions ?? {}),
+                        anthropic: { cacheControl: { type: "ephemeral" } },
+                    },
+                };
+            }
+        }
         return await generateText({
             model,
             messages,
-            ...(shouldUseTools && Object.keys(tools).length > 0 && { tools }),
+            ...(shouldUseTools &&
+                Object.keys(toolsWithCache).length > 0 && { tools: toolsWithCache }),
             maxSteps: options.maxSteps ?? DEFAULT_MAX_STEPS,
             ...(shouldUseTools &&
                 options.toolChoice && { toolChoice: options.toolChoice }),
@@ -116,8 +157,90 @@ export class GenerationHandler {
         const useStructuredOutput = !!options.schema &&
             (options.output?.format === "json" ||
                 options.output?.format === "structured");
+        const requestId = options.requestId ||
+            options.context?.requestId ||
+            "unknown";
+        logger.info("[GenerationHandler] Calling generateText", {
+            requestId,
+            model: model.modelId || "unknown",
+            messageCount: messages.length,
+            toolCount: Object.keys(tools || {}).length,
+            maxSteps: options.maxSteps,
+            temperature: options.temperature,
+        });
+        if (logger.shouldLog("debug")) {
+            try {
+                logger.debug("[Observability] Full generateText parameters", {
+                    requestId,
+                    model: model.modelId || "unknown",
+                    messageCount: messages.length,
+                    messages: messages.map((msg, i) => ({
+                        index: i,
+                        role: msg.role,
+                        contentLength: typeof msg.content === "string"
+                            ? msg.content.length
+                            : safePreview(msg.content).length,
+                        contentPreview: typeof msg.content === "string"
+                            ? msg.content.substring(0, 200)
+                            : "[multimodal]",
+                    })),
+                    toolNames: Object.keys(tools || {}),
+                    toolCount: Object.keys(tools || {}).length,
+                    maxSteps: options.maxSteps,
+                    temperature: options.temperature,
+                    maxTokens: options.maxTokens,
+                });
+            }
+            catch {
+                // Ignore serialization errors in debug logging
+            }
+        }
+        const genStartTime = Date.now();
         try {
-            return await this.callGenerateText(model, messages, tools, options, shouldUseTools, true);
+            const result = await this.callGenerateText(model, messages, tools, options, shouldUseTools, true);
+            logger.info("[GenerationHandler] generateText returned", {
+                requestId,
+                durationMs: Date.now() - genStartTime,
+                finishReason: result.finishReason,
+                steps: result.steps?.length || 1,
+                toolCallsTotal: result.toolCalls?.length || 0,
+                responseChars: result.text?.length || 0,
+            });
+            if (logger.shouldLog("debug")) {
+                logger.debug("[Observability] Full LLM response", {
+                    requestId,
+                    finishReason: result.finishReason,
+                    responseTextPreview: result.text?.substring(0, 200) || "",
+                    responseTextLength: result.text?.length || 0,
+                    toolCalls: result.toolCalls?.map((tc) => ({
+                        toolName: tc.toolName,
+                        argsPreview: safePreview(tc.args),
+                    })),
+                    toolResults: result.toolResults?.map((tr) => ({
+                        toolName: tr.toolName,
+                        resultPreview: safePreview(tr.result),
+                    })),
+                    steps: result.steps?.map((step, i) => ({
+                        stepIndex: i,
+                        stepType: step.stepType,
+                        textPreview: step.text?.substring(0, 200),
+                        textLength: step.text?.length || 0,
+                        toolCalls: step.toolCalls?.map((tc) => ({
+                            toolName: tc.toolName,
+                            argsPreview: safePreview(tc.args),
+                        })),
+                        toolResults: step.toolResults?.map((tr) => ({
+                            toolName: tr.toolName,
+                            resultPreview: safePreview(tr.result),
+                        })),
+                        finishReason: step.finishReason,
+                    })),
+                    usage: result.usage,
+                    providerMetadata: result.experimental_providerMetadata ||
+                        result.providerMetadata,
+                });
+            }
+            return result;
         }
         catch (error) {
             // If NoObjectGeneratedError is thrown when using schema + tools together,
@@ -130,16 +253,50 @@ export class GenerationHandler {
                 });
                 // Retry without experimental_output - the formatEnhancedResult method
                 // will extract JSON from the text response
-                return await this.callGenerateText(model, messages, tools, options, shouldUseTools, false);
+                const result = await this.callGenerateText(model, messages, tools, options, shouldUseTools, false);
+                logger.info("[GenerationHandler] generateText returned (fallback)", {
+                    requestId,
+                    durationMs: Date.now() - genStartTime,
+                    finishReason: result.finishReason,
+                    steps: result.steps?.length || 1,
+                    toolCallsTotal: result.toolCalls?.length || 0,
+                    responseChars: result.text?.length || 0,
+                });
+                return result;
             }
             // Re-throw other errors
             throw error;
         }
     }
+    /**
+     * Extract cache metrics from provider metadata (e.g. Anthropic's providerMetadata.anthropic)
+     * The Vercel AI SDK's LanguageModelUsage only has promptTokens/completionTokens/totalTokens.
+     * Cache metrics are surfaced via providerMetadata by provider-specific SDK adapters.
+     */
+    extractCacheMetricsFromProviderMetadata(generateResult) {
+        const providerMeta = generateResult
+            .providerMetadata ||
+            generateResult.experimental_providerMetadata;
+        if (!providerMeta) {
+            return {};
+        }
+        // Anthropic surfaces cache metrics under providerMetadata.anthropic
+        const anthropicMeta = providerMeta.anthropic;
+        if (anthropicMeta) {
+            const cacheCreationTokens = extractCacheCreationTokens(anthropicMeta);
+            const cacheReadTokens = extractCacheReadTokens(anthropicMeta);
+            return {
+                ...(cacheCreationTokens !== undefined && { cacheCreationTokens }),
+                ...(cacheReadTokens !== undefined && { cacheReadTokens }),
+            };
+        }
+        return {};
+    }
     /**
      * Log generation completion information
      */
     logGenerationComplete(generateResult) {
+        const cacheMetrics = this.extractCacheMetricsFromProviderMetadata(generateResult);
         logger.debug(`generateText completed`, {
             provider: this.providerName,
             model: this.modelName,
@@ -147,6 +304,12 @@ export class GenerationHandler {
             toolResultsCount: generateResult.toolResults?.length || 0,
             finishReason: generateResult.finishReason,
             usage: generateResult.usage,
+            ...(cacheMetrics.cacheCreationTokens !== undefined && {
+                cacheCreationTokens: cacheMetrics.cacheCreationTokens,
+            }),
+            ...(cacheMetrics.cacheReadTokens !== undefined && {
+                cacheReadTokens: cacheMetrics.cacheReadTokens,
+            }),
             timestamp: Date.now(),
         });
     }
@@ -274,6 +437,28 @@ export class GenerationHandler {
         // Note: The AI SDK bundles thinking tokens into promptTokens for Google models.
         // Separate reasoningTokens tracking will work when/if the AI SDK adds support.
         const usage = extractTokenUsage(generateResult.usage);
+        // Merge cache metrics from providerMetadata if not already present in usage
+        // The AI SDK's LanguageModelUsage doesn't include cache tokens; they come from
+        // provider-specific metadata (e.g. Anthropic's providerMetadata.anthropic)
+        if (usage.cacheCreationTokens === undefined ||
+            usage.cacheReadTokens === undefined) {
+            const cacheMetrics = this.extractCacheMetricsFromProviderMetadata(generateResult);
+            if (usage.cacheCreationTokens === undefined &&
+                cacheMetrics.cacheCreationTokens !== undefined) {
+                usage.cacheCreationTokens = cacheMetrics.cacheCreationTokens;
+            }
+            if (usage.cacheReadTokens === undefined &&
+                cacheMetrics.cacheReadTokens !== undefined) {
+                usage.cacheReadTokens = cacheMetrics.cacheReadTokens;
+            }
+            // Recalculate cache savings if we added cache metrics
+            if (usage.cacheReadTokens !== undefined) {
+                const savingsPercent = calculateCacheSavingsPercent(usage.cacheReadTokens, usage.input);
+                if (savingsPercent !== undefined) {
+                    usage.cacheSavingsPercent = savingsPercent;
+                }
+            }
+        }
         return {
             content,
             usage,
@@ -309,7 +494,7 @@ export class GenerationHandler {
             provider: this.providerName,
             model: this.modelName,
             responseTextLength: result.text?.length || 0,
-            responsePreview: result.text?.substring(0, 500) + "...",
+            responsePreview: result.text?.substring(0, 500) ?? "",
             finishReason: result.finishReason,
             usage: result.usage,
         });