feat(providers): model-registry routing + Responses API support and streaming fallbacks for OpenAI/Gemini

src/server/mod.rs

@@ -304,7 +304,17 @@ async fn chat_completions(
         }
     } else {
         // Handle non-streaming response
-        let result = provider.chat_completion(unified_request).await;
+        // Allow provider-specific routing: for OpenAI, some models prefer the
+        // Responses API (/v1/responses). Use the model registry heuristic to
+        // choose chat_responses vs chat_completion automatically.
+        let use_responses = provider.name() == "openai"
+            && crate::utils::registry::model_prefers_responses(&state.model_registry, &unified_request.model);
+
+        let result = if use_responses {
+            provider.chat_responses(unified_request).await
+        } else {
+            provider.chat_completion(unified_request).await
+        };
 
         match result {
             Ok(response) => {