@posthog/ai 7.16.2 → 7.16.4

package/dist/vercel/index.mjs

@@ -1,7 +1,7 @@
 import { v4 } from 'uuid';
 import { uuidv7 } from '@posthog/core';
 
-var version = "7.16.2";
+var version = "7.16.4";
 
 // Type guards for safer type checking
 
@@ -659,30 +659,6 @@ const extractWebSearchCount = (providerMetadata, usage) => {
   });
 };
 
-// Extract additional token values from provider metadata
-const extractAdditionalTokenValues = providerMetadata => {
-  if (providerMetadata && typeof providerMetadata === 'object' && 'anthropic' in providerMetadata && providerMetadata.anthropic && typeof providerMetadata.anthropic === 'object' && 'cacheCreationInputTokens' in providerMetadata.anthropic) {
-    return {
-      cacheCreationInputTokens: providerMetadata.anthropic.cacheCreationInputTokens
-    };
-  }
-  return {};
-};
-
-// For Anthropic providers in V3, inputTokens.total is the sum of all tokens (uncached + cache read + cache write).
-// Our cost calculation expects inputTokens to be only the uncached portion for Anthropic.
-// This helper subtracts cache tokens from inputTokens for Anthropic V3 models.
-const adjustAnthropicV3CacheTokens = (model, provider, usage) => {
-  if (isV3Model(model) && provider.toLowerCase().includes('anthropic')) {
-    const cacheReadTokens = usage.cacheReadInputTokens || 0;
-    const cacheWriteTokens = usage.cacheCreationInputTokens || 0;
-    const cacheTokens = cacheReadTokens + cacheWriteTokens;
-    if (usage.inputTokens && cacheTokens > 0) {
-      usage.inputTokens = Math.max(usage.inputTokens - cacheTokens, 0);
-    }
-  }
-};
-
 // Helper to extract numeric token value from V2 (number) or V3 (object with .total) usage formats
 const extractTokenCount = value => {
   if (typeof value === 'number') {
@@ -720,6 +696,63 @@ const extractCacheReadTokens = usage => {
   return undefined;
 };
 
+// Helper to extract cache write tokens from V3 (usage.inputTokens.cacheWrite). Providers like
+// Amazon Bedrock populate this standardized field instead of providerMetadata.anthropic.
+const extractCacheWriteTokens = usage => {
+  if ('inputTokens' in usage && usage.inputTokens && typeof usage.inputTokens === 'object' && 'cacheWrite' in usage.inputTokens) {
+    return usage.inputTokens.cacheWrite;
+  }
+  return undefined;
+};
+
+// Extract additional token values from provider metadata, with a V3 standardized fallback
+// (e.g. Amazon Bedrock exposes cache write tokens via usage.inputTokens.cacheWrite rather
+// than providerMetadata.anthropic.cacheCreationInputTokens). A cacheWrite of 0 is treated
+// as absent so we preserve the pre-fallback event shape on providers that simply omit the
+// field — consumers downstream saw `$ai_cache_creation_input_tokens` missing, not 0.
+const extractAdditionalTokenValues = (providerMetadata, usage) => {
+  if (providerMetadata && typeof providerMetadata === 'object' && 'anthropic' in providerMetadata && providerMetadata.anthropic && typeof providerMetadata.anthropic === 'object' && 'cacheCreationInputTokens' in providerMetadata.anthropic) {
+    return {
+      cacheCreationInputTokens: providerMetadata.anthropic.cacheCreationInputTokens
+    };
+  }
+  if (usage && typeof usage === 'object') {
+    const cacheWrite = extractCacheWriteTokens(usage);
+    if (typeof cacheWrite === 'number' && cacheWrite > 0) {
+      return {
+        cacheCreationInputTokens: cacheWrite
+      };
+    }
+  }
+  return {};
+};
+
+// Detects Anthropic Claude regardless of host (direct Anthropic, Amazon Bedrock, Google Vertex, etc.).
+// The server applies exclusive cache token accounting based on the model name, so any Claude model
+// needs its V3 input tokens adjusted to exclude cache tokens — not just those routed through a
+// provider whose name contains "anthropic". Accepts the resolved modelId string (not the raw model)
+// so it sees the same id the server does after posthogModelOverride / response.modelId fallbacks.
+const isAnthropicClaudeModel = (modelId, provider) => {
+  if (provider.toLowerCase().includes('anthropic')) {
+    return true;
+  }
+  return /claude|anthropic/i.test(modelId);
+};
+
+// For Anthropic providers in V3, inputTokens.total is the sum of all tokens (uncached + cache read + cache write).
+// Our cost calculation expects inputTokens to be only the uncached portion for Anthropic.
+// This helper subtracts cache tokens from inputTokens for Anthropic V3 models.
+const adjustAnthropicV3CacheTokens = (model, modelId, provider, usage) => {
+  if (isV3Model(model) && isAnthropicClaudeModel(modelId, provider)) {
+    const cacheReadTokens = usage.cacheReadInputTokens || 0;
+    const cacheWriteTokens = usage.cacheCreationInputTokens || 0;
+    const cacheTokens = cacheReadTokens + cacheWriteTokens;
+    if (usage.inputTokens && cacheTokens > 0) {
+      usage.inputTokens = Math.max(usage.inputTokens - cacheTokens, 0);
+    }
+  }
+};
+
 /**
  * Wraps a Vercel AI SDK language model (V2 or V3) with PostHog tracing.
  * Automatically detects the model version and applies appropriate instrumentation.
@@ -757,7 +790,7 @@ const wrapVercelLanguageModel = (model, phClient, options) => {
           const content = mapVercelOutput(result.content ?? []);
           const latency = (Date.now() - startTime) / 1000;
           const providerMetadata = result.providerMetadata;
-          const additionalTokenValues = extractAdditionalTokenValues(providerMetadata);
+          const additionalTokenValues = extractAdditionalTokenValues(providerMetadata, result.usage);
           const webSearchCount = extractWebSearchCount(providerMetadata, result.usage);
 
           // V2 usage has simple numbers, V3 has objects with .total - normalize both
@@ -788,7 +821,7 @@ const wrapVercelLanguageModel = (model, phClient, options) => {
             ...additionalTokenValues,
             rawUsage: rawUsageData
           };
-          adjustAnthropicV3CacheTokens(model, provider, usage);
+          adjustAnthropicV3CacheTokens(model, modelId, provider, usage);
 
           // Extract finish reason - V2 returns a string, V3 returns an object with .unified
           const rawFinishReason = result.finishReason;
@@ -915,8 +948,8 @@ const wrapVercelLanguageModel = (model, phClient, options) => {
               }
               if (chunk.type === 'finish') {
                 providerMetadata = chunk.providerMetadata;
-                const additionalTokenValues = extractAdditionalTokenValues(providerMetadata);
                 const chunkUsage = chunk.usage || {};
+                const additionalTokenValues = extractAdditionalTokenValues(providerMetadata, chunkUsage);
                 usage = {
                   inputTokens: extractTokenCount(chunk.usage?.inputTokens),
                   outputTokens: extractTokenCount(chunk.usage?.outputTokens),
@@ -983,7 +1016,7 @@ const wrapVercelLanguageModel = (model, phClient, options) => {
                   providerMetadata
                 }
               };
-              adjustAnthropicV3CacheTokens(model, provider, finalUsage);
+              adjustAnthropicV3CacheTokens(model, modelId, provider, finalUsage);
               await sendEventToPosthog({
                 client: phClient,
                 distinctId: mergedOptions.posthogDistinctId,