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;

llm.cpp

@@ -297,14 +297,23 @@ static void stream_handle_data(StreamCtx *sc, std::string_view data) {
       sc->last_usage.prompt_tokens = pt;
       sc->last_usage.completion_tokens = u["completion_tokens"].as_int();
       sc->last_usage.total_tokens = tt;
+      /* Resolve this chunk's cached-token count only when the chunk actually
+       * reports it, so a later chunk that omits the field can never clobber a
+       * prior value (nor can a stale value spuriously inflate the total). */
       const Json &ptd = u["prompt_tokens_details"];
-      if (ptd.is_object()) {
+      int chunk_cached = -1;
-        int ct_val = ptd["cached_tokens"].as_int();
+      if (ptd.is_object() && ptd.contains("cached_tokens"))
-        if (ct_val > 0) sc->last_usage.cached_tokens = ct_val;
+        chunk_cached = ptd["cached_tokens"].as_int();
-      }
+      else if (u.contains("cached_tokens"))
-      if (sc->last_usage.cached_tokens == 0) {
+        chunk_cached = u["cached_tokens"].as_int();
-        int ct_val = u["cached_tokens"].as_int();
+      if (chunk_cached >= 0) {
-        if (ct_val > 0) sc->last_usage.cached_tokens = ct_val;
+        sc->last_usage.cached_tokens = chunk_cached;
+        /* Some providers (e.g. Cohere on OpenRouter) report prompt_tokens as the
+         * *uncached* portion and cached_tokens as an *additional* amount rather
+         * than a subset. Fold cached into the true total input so the breakdown
+         * and the context percentage stay correct. */
+        if (chunk_cached > sc->last_usage.prompt_tokens)
+          sc->last_usage.prompt_tokens += chunk_cached;
       }
     }
   }