fix: preserve parallel tool call setting

crates/agentic-server-core/benches/executor_throughput.rs

@@ -145,6 +145,7 @@ fn make_request(input: &str, stream: bool, prev_id: Option<String>) -> RequestPa
         max_output_tokens: None,
         truncation: None,
         metadata: None,
+        parallel_tool_calls: None,
     }
 }
 

crates/agentic-server-core/src/executor/modes/conversation.rs

@@ -141,6 +141,7 @@ mod tests {
             max_output_tokens: None,
             truncation: None,
             metadata: None,
+            parallel_tool_calls: None,
         };
         RequestContext {
             enriched_request: req.clone(),

crates/agentic-server-core/src/executor/modes/response.rs

@@ -120,6 +120,7 @@ mod tests {
             max_output_tokens: None,
             truncation: None,
             metadata: None,
+            parallel_tool_calls: None,
         };
         RequestContext {
             enriched_request: req.clone(),

crates/agentic-server-core/src/types/request_response.rs

@@ -28,6 +28,7 @@ pub struct RequestPayload {
     pub max_output_tokens: Option<u32>,
     pub truncation: Option<String>,
     pub metadata: Option<Value>,
+    pub parallel_tool_calls: Option<bool>,
 }
 
 fn default_true() -> bool {
@@ -61,6 +62,8 @@ pub struct UpstreamRequest<'a> {
     pub truncation: Option<&'a str>,
     #[serde(skip_serializing_if = "Option::is_none")]
     pub metadata: Option<&'a Value>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub parallel_tool_calls: Option<bool>,
 }
 
 // serde's `skip_serializing_if` requires a `&Option<T>` receiver, so the
@@ -109,6 +112,7 @@ impl RequestPayload {
             max_output_tokens: self.max_output_tokens,
             truncation: self.truncation.as_deref(),
             metadata: self.metadata.as_ref(),
+            parallel_tool_calls: self.parallel_tool_calls,
         })
     }
 }
@@ -214,6 +218,20 @@ mod tests {
         assert_eq!(value["input"], "hi");
     }
 
+    #[test]
+    fn to_upstream_request_preserves_parallel_tool_calls() {
+        let payload: RequestPayload = serde_json::from_value(serde_json::json!({
+            "model": "test",
+            "input": "hi",
+            "parallel_tool_calls": false
+        }))
+        .unwrap();
+
+        let upstream = payload.to_upstream_request(false).expect("valid upstream request");
+        let value = serde_json::to_value(upstream).unwrap();
+        assert_eq!(value["parallel_tool_calls"], false);
+    }
+
     #[test]
     fn to_upstream_request_flattens_namespace_and_skips_unknown_tools() {
         let payload: RequestPayload = serde_json::from_value(serde_json::json!({

crates/agentic-server-core/tests/dispatch_loop_cassette_test.rs

@@ -127,6 +127,7 @@ fn request(text: &str, tools: Option<Vec<ResponsesTool>>) -> RequestPayload {
         max_output_tokens: Some(1024),
         truncation: None,
         metadata: None,
+        parallel_tool_calls: None,
     }
 }
 

crates/agentic-server-core/tests/support/mod.rs

@@ -369,6 +369,7 @@ pub fn make_request(
         max_output_tokens: None,
         truncation: None,
         metadata: None,
+        parallel_tool_calls: None,
     }
 }
 

crates/agentic-server-core/tests/web_search_tool_test.rs

@@ -584,6 +584,7 @@ async fn execute_runs_web_search_and_sends_tool_output_back_to_model() {
         max_output_tokens: Some(1024),
         truncation: None,
         metadata: None,
+        parallel_tool_calls: None,
     };
 
     let result = ExecuteRequest::new(payload, Arc::clone(&exec_ctx)).run().await.unwrap();
@@ -667,6 +668,7 @@ async fn execute_relaxes_forced_tool_choice_after_web_search_result() {
         max_output_tokens: Some(1024),
         truncation: None,
         metadata: None,
+        parallel_tool_calls: None,
     };
 
     let result = ExecuteRequest::new(payload, Arc::clone(&exec_ctx)).run().await.unwrap();
@@ -714,6 +716,7 @@ async fn execute_returns_mixed_client_tool_calls_without_followup_model_request(
         max_output_tokens: Some(1024),
         truncation: None,
         metadata: None,
+        parallel_tool_calls: None,
     };
 
     let result = ExecuteRequest::new(payload, Arc::clone(&exec_ctx)).run().await.unwrap();
@@ -760,6 +763,7 @@ async fn execute_returns_mixed_client_tool_calls_without_followup_model_request(
         max_output_tokens: Some(1024),
         truncation: None,
         metadata: None,
+        parallel_tool_calls: None,
     };
     let continuation = ExecuteRequest::new(continuation_payload, exec_ctx).run().await.unwrap();
     assert!(matches!(continuation, Either::Left(_)));
@@ -829,6 +833,7 @@ async fn execute_accumulates_usage_across_web_search_model_rounds() {
         max_output_tokens: Some(1024),
         truncation: None,
         metadata: None,
+        parallel_tool_calls: None,
     };
 
     let result = ExecuteRequest::new(payload, exec_ctx).run().await.unwrap();
@@ -871,6 +876,7 @@ async fn stream_emits_web_search_lifecycle_events_before_final_payload() {
         max_output_tokens: Some(1024),
         truncation: None,
         metadata: None,
+        parallel_tool_calls: None,
     };
 
     let result = ExecuteRequest::new(payload, Arc::clone(&exec_ctx)).run().await.unwrap();
@@ -952,6 +958,7 @@ async fn stream_hides_web_search_function_events_when_name_arrives_on_done() {
         max_output_tokens: Some(1024),
         truncation: None,
         metadata: None,
+        parallel_tool_calls: None,
     };
 
     let result = ExecuteRequest::new(payload, Arc::clone(&exec_ctx)).run().await.unwrap();
@@ -1017,6 +1024,7 @@ async fn execute_runs_multiple_web_search_calls_concurrently() {
         max_output_tokens: Some(1024),
         truncation: None,
         metadata: None,
+        parallel_tool_calls: None,
     };
 
     let result = tokio::time::timeout(Duration::from_secs(2), ExecuteRequest::new(payload, exec_ctx).run())
@@ -1064,6 +1072,7 @@ async fn execute_feeds_web_search_execution_errors_back_to_model() {
         max_output_tokens: Some(1024),
         truncation: None,
         metadata: None,
+        parallel_tool_calls: None,
     };
 
     let result = ExecuteRequest::new(payload, exec_ctx).run().await.unwrap();
@@ -1113,6 +1122,7 @@ async fn execute_returns_incomplete_after_max_gateway_tool_rounds() {
         max_output_tokens: Some(1024),
         truncation: None,
         metadata: None,
+        parallel_tool_calls: None,
     };
 
     // Budget exhausted while the model keeps requesting tools → the response is
@@ -1162,6 +1172,7 @@ async fn execute_feeds_invalid_web_search_arguments_back_to_model() {
         max_output_tokens: Some(1024),
         truncation: None,
         metadata: None,
+        parallel_tool_calls: None,
     };
 
     let result = ExecuteRequest::new(payload, exec_ctx).run().await.unwrap();
@@ -1218,6 +1229,7 @@ async fn execute_runs_large_gateway_fanout_without_hard_cap() {
         max_output_tokens: Some(1024),
         truncation: None,
         metadata: None,
+        parallel_tool_calls: None,
     };
 
     let result = ExecuteRequest::new(payload, exec_ctx)
@@ -1280,6 +1292,7 @@ async fn stream_error_events_escape_error_messages() {
         max_output_tokens: Some(1024),
         truncation: None,
         metadata: None,
+        parallel_tool_calls: None,
     };
 
     let result = ExecuteRequest::new(payload, exec_ctx).run().await.unwrap();
@@ -1354,6 +1367,7 @@ async fn incomplete_turn_persists_a_consistent_conversation_for_continuation() {
         max_output_tokens: Some(1024),
         truncation: None,
         metadata: None,
+        parallel_tool_calls: None,
     };
 
     let result = ExecuteRequest::new(payload, Arc::clone(&exec_ctx)).run().await.unwrap();
@@ -1379,6 +1393,7 @@ async fn incomplete_turn_persists_a_consistent_conversation_for_continuation() {
         max_output_tokens: Some(1024),
         truncation: None,
         metadata: None,
+        parallel_tool_calls: None,
     };
     let _ = ExecuteRequest::new(continuation_payload, exec_ctx).run().await.unwrap();
 
@@ -1448,6 +1463,7 @@ async fn stream_returns_incomplete_after_max_gateway_tool_rounds() {
         max_output_tokens: Some(1024),
         truncation: None,
         metadata: None,
+        parallel_tool_calls: None,
     };
 
     let result = ExecuteRequest::new(payload, exec_ctx).run().await.unwrap();

crates/agentic-server/tests/responses_test.rs

@@ -132,6 +132,34 @@ async fn test_store_false_with_web_search_reaches_executor() {
     assert_eq!(requests[0]["tools"][0]["name"], "web_search");
 }
 
+#[tokio::test]
+async fn test_gateway_normalization_preserves_parallel_tool_calls() {
+    // Arrange
+    let (llm_url, requests, _h1) = spawn_mock_vllm_json_capture().await;
+    let (gw_url, _h2) = spawn_gateway(test_state(&test_config(&llm_url))).await;
+
+    // Act
+    let resp = reqwest::Client::new()
+        .post(format!("{gw_url}/v1/responses"))
+        .json(&serde_json::json!({
+            "model": "test",
+            "input": [{"type": "message", "role": "user", "content": "hi"}],
+            "tools": [{"type": "web_search_preview"}],
+            "parallel_tool_calls": false,
+            "store": false,
+            "stream": false
+        }))
+        .send()
+        .await
+        .unwrap();
+
+    // Assert
+    assert_eq!(resp.status(), 200);
+    let requests = requests.lock().await;
+    assert_eq!(requests.len(), 1);
+    assert_eq!(requests[0]["parallel_tool_calls"], false);
+}
+
 #[tokio::test]
 async fn test_store_false_proxies_large_json_body_to_vllm() {
     // Arrange