@posthog/ai 8.3.0 → 8.3.1

package/dist/openai/index.mjs

@@ -530,7 +530,7 @@ function formatOpenAIResponsesInput(input, instructions) {
   return messages;
 }
 
-var version = "8.3.0";
+var version = "8.3.1";
 
 const DEFAULT_MAX_DEPTH = 3;
 const MAX_STACK_LINES = 20;
@@ -640,7 +640,7 @@ const warnIfPostHogAiGateway = baseURL => {
  * object, it is mutated in place to set `__posthog_previously_captured_error`
  * so callers can re-throw the original error reference safely.
  */
-const captureAiGeneration = async (client, options) => {
+const captureAiGeneration$1 = async (client, options) => {
   if (!client.capture) {
     return;
   }
@@ -758,6 +758,43 @@ const captureAiGeneration = async (client, options) => {
   }
 };
 
+/**
+ * The declared convention only describes the wrapper's own usage numbers, so
+ * when the caller passes any of these through posthogProperties the wrapper no
+ * longer knows the convention of the reported counts (callers working around
+ * the double-billing already pass exclusive ones) and must not declare it.
+ * Output/reasoning token overrides don't affect the input/cache relationship,
+ * so they don't suppress the declaration. Subset of the input/cache keys in
+ * `TOKEN_PROPERTY_KEYS` (../utils.ts) — keep in sync if that taxonomy grows.
+ */
+const INPUT_OR_CACHE_TOKEN_KEYS = ['$ai_input_tokens', '$ai_cache_read_input_tokens', '$ai_cache_creation_input_tokens'];
+
+/**
+ * OpenAI-compatible usage reports `prompt_tokens` INCLUSIVE of cached tokens
+ * (`prompt_tokens_details.cached_tokens` is a subset of it), unlike Anthropic's
+ * exclusive convention. Ingestion auto-classifies Claude-shaped models as
+ * exclusive regardless of provider, so events for Claude served through
+ * OpenAI-compatible hosts (e.g. OpenRouter) would get cache reads billed twice.
+ * Declaring the convention on every event from this wrapper lets ingestion
+ * normalize correctly (see PostHog/posthog#49252); for non-Claude models the
+ * flag is a no-op. Callers can still override it via posthogProperties, and
+ * when they pass through input or cache token counts themselves the flag stays
+ * unset unless they declare it explicitly.
+ */
+const captureAiGeneration = (client, options) => {
+  const props = options.properties;
+  // Own-property check, matching how getTokensSource detects passthrough and
+  // how the properties spread actually copies values into the event.
+  const callerReportsTokens = props !== undefined && INPUT_OR_CACHE_TOKEN_KEYS.some(key => Object.prototype.hasOwnProperty.call(props, key));
+  return captureAiGeneration$1(client, {
+    ...options,
+    properties: callerReportsTokens ? props : {
+      $ai_cache_reporting_exclusive: false,
+      ...props
+    }
+  });
+};
+
 /**
  * Checks if a ResponseStreamEvent chunk represents the first token/content from the model.
  * This includes various content types like text, reasoning, audio, and refusals.