@juspay/neurolink 9.82.0 → 9.84.0

package/dist/lib/providers/googleVertex.js

@@ -24,7 +24,7 @@ import { convertZodToJsonSchema, inlineJsonSchema, ensureNestedSchemaTypes, } fr
 import { createNativeThinkingConfig } from "../utils/thinkingConfig.js";
 import { TimeoutError, raceWithAbort, withTimeout, } from "../utils/async/index.js";
 import { parseTimeout } from "../utils/timeout.js";
-import { appendStepText, buildAbortedTurnMessage, buildContextCapMessage, buildToolLoopCapMessage, buildTurnStalledMessage, buildTurnTimeoutMessage, buildWrapupNudgeText, createContextGuard, createTextChannel, createTurnClock, extractThoughtSignature, isAbortError, mapGeminiFinishReason, prependConversationMessages, resolveTurnStopReason, } from "./googleNativeGemini3.js";
+import { appendStepText, buildAbortedTurnMessage, buildContextCapMessage, buildToolLoopCapMessage, buildTurnStalledMessage, buildTurnTimeoutMessage, buildWrapupNudgeText, createContextGuard, createTextChannel, createTurnClock, extractThoughtSignature, isAbortError, mapGeminiFinishReason, prependConversationMessages, resolveTurnStopReason, DedupExecuteMap, } from "./googleNativeGemini3.js";
 import { getContextWindowSize } from "../constants/contextWindows.js";
 import { ATTR, LANGFUSE_ATTR, spanJsonAttribute, tracers, withClientSpan, withClientStreamSpan, withSpan, } from "../telemetry/index.js";
 import { SpanKind, SpanStatusCode, context as otelContext, trace as otelTrace, } from "@opentelemetry/api";
@@ -1138,7 +1138,7 @@ export class GoogleVertexProvider extends BaseProvider {
         });
         // Convert Vercel AI SDK tools to @google/genai FunctionDeclarations
         let tools;
-        const executeMap = new Map();
+        const executeMap = new DedupExecuteMap();
         if (options.tools &&
             Object.keys(options.tools).length > 0 &&
             !options.disableTools) {
@@ -1325,6 +1325,7 @@ export class GoogleVertexProvider extends BaseProvider {
         // promptTokenCount is typically in the final chunk, candidatesTokenCount accumulates
         let totalInputTokens = 0;
         let totalOutputTokens = 0;
+        let totalCacheReadTokens = 0;
         // Track text parts as they arrive from the SDK so the returned async
         // iterable yields multiple chunks instead of a single buffered chunk.
         // The CLI's chunk-count smoke test asserts > 1 stream chunks for any
@@ -1440,6 +1441,10 @@ export class GoogleVertexProvider extends BaseProvider {
                                 totalInputTokens = usageMetadata.promptTokenCount;
                                 // Feed the context guard the REAL prompt size of this call.
                                 contextGuard.noteUsage(usageMetadata.promptTokenCount, usageMetadata.candidatesTokenCount ?? 0);
+                                // cachedContentTokenCount is OVERLAPPING (a subset already inside
+                                // promptTokenCount). Clamp to the prompt count so a later uncached
+                                // step resets it to 0 instead of leaving a stale cached value.
+                                totalCacheReadTokens = Math.min(usageMetadata.cachedContentTokenCount ?? 0, usageMetadata.promptTokenCount);
                             }
                             // Take the latest candidatesTokenCount (accumulates through chunks)
                             if (usageMetadata.candidatesTokenCount !== undefined &&
@@ -1891,6 +1896,10 @@ export class GoogleVertexProvider extends BaseProvider {
         // Filter out final_result from tool calls as it's an internal pattern
         const externalToolCalls = allToolCalls.filter((tc) => tc.toolName !== "final_result");
         const externalToolExecutions = toolExecutions.filter((te) => te.name !== "final_result");
+        // Gemini promptTokenCount is OVERLAPPING (already includes
+        // cachedContentTokenCount). Subtract once so the cached portion is billed at
+        // the cheaper cacheRead rate without double-counting; total is conserved.
+        const adjustedInputTokens = Math.max(0, totalInputTokens - totalCacheReadTokens);
         const result = {
             stream: createTextStream(),
             provider: this.providerName,
@@ -1899,9 +1908,12 @@ export class GoogleVertexProvider extends BaseProvider {
             stopReason,
             rawFinishReason: lastFinishReason,
             usage: {
-                input: totalInputTokens,
+                input: adjustedInputTokens,
                 output: totalOutputTokens,
-                total: totalInputTokens + totalOutputTokens,
+                total: adjustedInputTokens + totalCacheReadTokens + totalOutputTokens,
+                ...(totalCacheReadTokens > 0 && {
+                    cacheReadTokens: totalCacheReadTokens,
+                }),
             },
             toolCalls: externalToolCalls.map((tc) => ({
                 toolName: tc.toolName,
@@ -2073,7 +2085,7 @@ export class GoogleVertexProvider extends BaseProvider {
         const combinedTools = !options.disableTools ? options.tools || {} : {};
         // Convert Vercel AI SDK tools to @google/genai FunctionDeclarations
         let tools;
-        const executeMap = new Map();
+        const executeMap = new DedupExecuteMap();
         if (Object.keys(combinedTools).length > 0) {
             const functionDeclarations = [];
             for (const [name, tool] of Object.entries(combinedTools)) {
@@ -2254,6 +2266,7 @@ export class GoogleVertexProvider extends BaseProvider {
         // promptTokenCount is typically in the final chunk, candidatesTokenCount accumulates
         let totalInputTokens = 0;
         let totalOutputTokens = 0;
+        let totalCacheReadTokens = 0;
         // Abort scaffolding (mirrors executeNativeAnthropicStream). The native
         // Gemini SDK cancels via config.abortSignal, so drive an internal
         // AbortController: the caller's signal and the turn clock's watchdogs
@@ -2359,6 +2372,10 @@ export class GoogleVertexProvider extends BaseProvider {
                                 totalInputTokens = usageMetadata.promptTokenCount;
                                 // Feed the context guard the REAL prompt size of this call.
                                 contextGuard.noteUsage(usageMetadata.promptTokenCount, usageMetadata.candidatesTokenCount ?? 0);
+                                // cachedContentTokenCount is OVERLAPPING (a subset already inside
+                                // promptTokenCount). Clamp to the prompt count so a later uncached
+                                // step resets it to 0 instead of leaving a stale cached value.
+                                totalCacheReadTokens = Math.min(usageMetadata.cachedContentTokenCount ?? 0, usageMetadata.promptTokenCount);
                             }
                             // Take the latest candidatesTokenCount (accumulates through chunks)
                             if (usageMetadata.candidatesTokenCount !== undefined &&
@@ -2789,6 +2806,10 @@ export class GoogleVertexProvider extends BaseProvider {
         const externalToolCalls = allToolCalls.filter((tc) => tc.toolName !== "final_result");
         const externalToolExecutions = toolExecutions.filter((te) => te.name !== "final_result");
         // Build EnhancedGenerateResult
+        // Gemini promptTokenCount is OVERLAPPING (already includes
+        // cachedContentTokenCount). Subtract once so the cached portion is billed at
+        // the cheaper cacheRead rate without double-counting; total is conserved.
+        const adjustedInputTokens = Math.max(0, totalInputTokens - totalCacheReadTokens);
         const result = {
             content: finalText,
             provider: this.providerName,
@@ -2798,9 +2819,12 @@ export class GoogleVertexProvider extends BaseProvider {
             rawFinishReason: lastFinishReason,
             stepsUsed: step,
             usage: {
-                input: totalInputTokens,
+                input: adjustedInputTokens,
                 output: totalOutputTokens,
-                total: totalInputTokens + totalOutputTokens,
+                total: adjustedInputTokens + totalCacheReadTokens + totalOutputTokens,
+                ...(totalCacheReadTokens > 0 && {
+                    cacheReadTokens: totalCacheReadTokens,
+                }),
             },
             responseTime,
             toolsUsed: externalToolCalls.map((tc) => tc.toolName),
@@ -3092,7 +3116,7 @@ export class GoogleVertexProvider extends BaseProvider {
             : userContentParts);
         // Convert tools to Anthropic format if present
         let tools;
-        const executeMap = new Map();
+        const executeMap = new DedupExecuteMap();
         if (options.tools &&
             Object.keys(options.tools).length > 0 &&
             !options.disableTools) {
@@ -4401,7 +4425,7 @@ export class GoogleVertexProvider extends BaseProvider {
             : userContentParts);
         // Convert tools to Anthropic format if present
         let tools;
-        const executeMap = new Map();
+        const executeMap = new DedupExecuteMap();
         const toolExecutions = [];
         // Defensive guard: Vertex generate() bypasses BaseProvider.generate(), so
         // disableTools must be checked here before native tool declarations are