tkn cnt fix #3

json.cpp

@@ -680,6 +680,14 @@ bool parse_nonstream_response(std::string_view body, Message &out, Usage &usage)
     if (usage.cached_tokens == 0) {
       usage.cached_tokens = u["cached_tokens"].as_int();
     }
+    if (usage.cached_tokens > usage.prompt_tokens) {
+      /* Some providers (e.g. Cohere models on OpenRouter) report
+       * prompt_tokens as the *uncached* portion and cached_tokens as an
+       * *additional* amount, instead of the OpenAI convention where
+       * cached_tokens is a subset of prompt_tokens. Normalize so
+       * prompt_tokens always holds the true total input. */
+      usage.prompt_tokens += usage.cached_tokens;
+    }
   }
   usage.model = root["model"].str();
   return true;