diff --git a/crates/hermesllm/src/providers/response.rs b/crates/hermesllm/src/providers/response.rs
index c2497472..b1d387e3 100644
--- a/crates/hermesllm/src/providers/response.rs
+++ b/crates/hermesllm/src/providers/response.rs
@@ -104,6 +104,19 @@ impl Iterator for ProviderStreamResponseIter {
         }
     }
 }
+
+// Helper to serialize only the inner struct, not the enum wrapper.
+// This avoids the problem where serde serializes the enum variant as a wrapper object in JSON.
+impl ProviderResponseType {
+    /// Serialize the response as JSON bytes, omitting the enum wrapper.
+    pub fn as_json_bytes(&self) -> Result<Vec<u8>, serde_json::Error> {
+        match self {
+            ProviderResponseType::ChatCompletionsResponse(resp) => serde_json::to_vec(resp),
+            ProviderResponseType::MessagesResponse(resp) => serde_json::to_vec(resp),
+        }
+    }
+}
+
 pub trait ProviderResponse: Send + Sync {
     /// Get usage information if available - returns dynamic trait object
     fn usage(&self) -> Option<&dyn TokenUsage>;
diff --git a/crates/llm_gateway/src/stream_context.rs b/crates/llm_gateway/src/stream_context.rs
index ed13d09a..b41396fd 100644
--- a/crates/llm_gateway/src/stream_context.rs
+++ b/crates/llm_gateway/src/stream_context.rs
@@ -681,7 +681,7 @@ impl HttpContext for StreamContext {
             match (supported_api, self.resolved_api.as_ref()) {
                 (Some(supported_api), Some(_)) => {
                     match ProviderResponseType::try_from((&body[..], supported_api, &provider_id)) {
-                        Ok(response) => match serde_json::to_vec(&response) {
+                        Ok(response) => match response.as_json_bytes() {
                             Ok(bytes) => {
                                 self.set_http_response_body(0, bytes.len(), &bytes);
                             }