@danhachuel/thunderbolt 0.5.47 → 0.5.49

package/app/main.py

@@ -7043,7 +7043,7 @@ def _render_llm_card(settings: dict[str, Any], cards: list[dict[str, Any]], inde
                     "meta/llama-3.1-70b-instruct",
                     "meta/llama-3.3-70b-instruct",
                     "nvidia/llama-3.1-nemotron-70b-instruct",
-                ] if definition.code == "openai" else []
+                ] if definition.code == "openai" else (["google/paligemma"] if definition.code == "nvidia_nim_paligemma" else [])
                 options = ["__select_model__", *list(dict.fromkeys([*discovered_models, *default_models]))]
                 # Mantém modelos guardados mesmo quando a descoberta ainda não foi
                 # executada ou quando o provider deixou de os devolver temporariamente.
@@ -7128,11 +7128,17 @@ def _render_llm_card(settings: dict[str, Any], cards: list[dict[str, Any]], inde
             try:
                 from integrations.openai_model_discovery import fetch_openai_compatible_models
                 discovered = fetch_openai_compatible_models(edited.get("api_key", ""), edited.get("base_url", ""))
+                if definition.code == "nvidia_nim_paligemma" and "google/paligemma" not in discovered:
+                    discovered.insert(0, "google/paligemma")
                 st.session_state[f"llm_model_catalog_{card_id}"] = discovered
                 _persist_llm_cards(settings, cards)
                 st.success(f"{len(discovered)} modelo(s) disponíveis neste endpoint.")
             except Exception:
-                st.error("Não foi possível consultar os modelos deste endpoint.")
+                if definition.code == "nvidia_nim_paligemma" and str(edited.get("api_key") or "").strip():
+                    st.session_state[f"llm_model_catalog_{card_id}"] = ["google/paligemma"]
+                    st.warning("O catálogo NIM não expôs modelos, mas o modelo Paligemma foi disponibilizado para este provider.")
+                else:
+                    st.error("Não foi possível consultar os modelos deste endpoint.")
         elif save_clicked:
             _persist_llm_cards(settings, cards)
             st.success("Cartão LLM guardado.")

package/hermes_ui/growth_youtube.py

@@ -161,7 +161,9 @@ def _vision_prompt(video: Mapping[str, Any]) -> str:
 def analyse_thumbnail_with_paligemma(path: str, card: Mapping[str, Any]) -> dict[str, Any]:
     key = str(card.get("api_key") or "").strip()
     base = str(card.get("base_url") or "https://integrate.api.nvidia.com/v1").rstrip("/")
-    endpoint = base if base.endswith("/chat/completions") else f"{base}/chat/completions"
+    if "integrate.api.nvidia.com" in base:
+        base = base.replace("integrate.api.nvidia.com", "ai.api.nvidia.com")
+    endpoint = base if "/vlm/google/paligemma" in base else f"{base}/vlm/google/paligemma"
     model = str(card.get("model") or "google/paligemma").strip()
     if not key:
         return {"score": 50, "diagnosis": "API key NVIDIA NIM Paligemma não configurada.", "status": "sem_dados"}

package/hermes_ui/llm_providers.py

@@ -13,6 +13,7 @@ from integrations.openai_model_discovery import (
     DEFAULT_NVIDIA_NIM_BASE_URL,
     OpenAICompatibleAPIError,
     fetch_openai_compatible_models,
+    validate_paligemma_api_key,
     validate_openai_compatible_api_key,
     validate_openrouter_api_key,
 )
@@ -431,7 +432,9 @@ def test_llm_provider_card(card: Mapping[str, Any]) -> dict[str, Any]:
     if not model:
         return {"ok": False, "status": "error", "message": "Missing configuration — introduza o modelo antes do teste."}
     try:
-        if normalized["provider"] == "openrouter":
+        if normalized["provider"] == "nvidia_nim_paligemma":
+            validate_paligemma_api_key(api_key, base_url)
+        elif normalized["provider"] == "openrouter":
             validate_openrouter_api_key(api_key, base_url, model)
         else:
             validate_openai_compatible_api_key(api_key, base_url, model)

package/integrations/openai_model_discovery.py

@@ -258,3 +258,39 @@ def fetch_replicate_models(
     if not models:
         raise ModelDiscoveryError("O endpoint Replicate não devolveu modelos com versão disponível.")
     return sorted(models, key=str.casefold)[:MAX_MODEL_IDS]
+
+
+def validate_paligemma_api_key(
+    api_key: str,
+    base_url: str,
+    *,
+    timeout: float = DEFAULT_TIMEOUT_SECONDS,
+) -> None:
+    """Validate the Paligemma VLM route instead of the text-only chat route."""
+    token = str(api_key or "").strip()
+    if not token:
+        raise OpenAICompatibleAPIError("Informe a API key antes do teste.")
+    configured = str(base_url or DEFAULT_NVIDIA_NIM_BASE_URL).rstrip("/")
+    if configured.endswith("/chat/completions"):
+        configured = configured.removesuffix("/chat/completions")
+    if "integrate.api.nvidia.com" in configured:
+        configured = configured.replace("integrate.api.nvidia.com", "ai.api.nvidia.com")
+    endpoint = configured if "/vlm/google/paligemma" in configured else f"{configured}/vlm/google/paligemma"
+    payload = {
+        "messages": [{"role": "user", "content": "Describe this image in one word."}],
+        "max_tokens": 8,
+        "temperature": 0,
+    }
+    try:
+        response = requests.post(
+            endpoint,
+            headers={"Accept": "application/json", "Authorization": f"Bearer {token}", "Content-Type": "application/json"},
+            json=payload,
+            timeout=timeout,
+        )
+    except requests.RequestException as exc:
+        raise OpenAICompatibleAPIError(f"Não foi possível testar o endpoint Paligemma: {exc}") from exc
+    if response.status_code >= 400:
+        if response.status_code in (401, 403):
+            raise OpenAICompatibleAPIError(f"O endpoint Paligemma recusou a API key (HTTP {response.status_code}).")
+        raise OpenAICompatibleAPIError(f"O endpoint Paligemma devolveu HTTP {response.status_code}.")

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@danhachuel/thunderbolt",
-  "version": "0.5.47",
+  "version": "0.5.49",
   "description": "Thunderbolt — interface local para operação de canais faceless e motor MoneyPrinterTurbo",
   "license": "MIT",
   "main": "scripts/cli.mjs",