From a5420779c4464a128b27e6b5c6942405fdc86ab5 Mon Sep 17 00:00:00 2001
From: Ahmed Ibrahim <aibrahim@openai.com>
Date: Wed, 4 Mar 2026 17:14:23 -0800
Subject: [PATCH] realtime: disable output interruption from turn detection

---
 .../src/endpoint/realtime_websocket/methods.rs | 18 ++++++++++++++++++
 .../endpoint/realtime_websocket/protocol.rs    |  9 +++++++++
 .../codex-api/tests/realtime_websocket_e2e.rs  | 12 ++++++++++++
 3 files changed, 39 insertions(+)

diff --git a/codex-rs/codex-api/src/endpoint/realtime_websocket/methods.rs b/codex-rs/codex-api/src/endpoint/realtime_websocket/methods.rs
index 7d4f760a7a..82a1770c92 100644
--- a/codex-rs/codex-api/src/endpoint/realtime_websocket/methods.rs
+++ b/codex-rs/codex-api/src/endpoint/realtime_websocket/methods.rs
@@ -9,6 +9,7 @@ use crate::endpoint::realtime_websocket::protocol::SessionAudioFormat;
 use crate::endpoint::realtime_websocket::protocol::SessionAudioInput;
 use crate::endpoint::realtime_websocket::protocol::SessionAudioOutput;
 use crate::endpoint::realtime_websocket::protocol::SessionAudioOutputFormat;
+use crate::endpoint::realtime_websocket::protocol::SessionTurnDetection;
 use crate::endpoint::realtime_websocket::protocol::SessionTool;
 use crate::endpoint::realtime_websocket::protocol::SessionToolParameters;
 use crate::endpoint::realtime_websocket::protocol::SessionToolProperties;
@@ -343,6 +344,11 @@ impl RealtimeWebsocketWriter {
                             kind: "audio/pcm".to_string(),
                             rate: 24_000,
                         },
+                        turn_detection: SessionTurnDetection {
+                            kind: "semantic_vad".to_string(),
+                            interrupt_response: false,
+                            create_response: true,
+                        },
                     },
                     output: SessionAudioOutput {
                         format: SessionAudioOutputFormat {
@@ -899,6 +905,18 @@ mod tests {
                 first_json["session"]["audio"]["input"]["format"]["rate"],
                 Value::from(24_000)
             );
+            assert_eq!(
+                first_json["session"]["audio"]["input"]["turn_detection"]["type"],
+                Value::String("semantic_vad".to_string())
+            );
+            assert_eq!(
+                first_json["session"]["audio"]["input"]["turn_detection"]["interrupt_response"],
+                Value::Bool(false)
+            );
+            assert_eq!(
+                first_json["session"]["audio"]["input"]["turn_detection"]["create_response"],
+                Value::Bool(true)
+            );
             assert_eq!(
                 first_json["session"]["audio"]["output"]["format"]["type"],
                 Value::String("audio/pcm".to_string())
diff --git a/codex-rs/codex-api/src/endpoint/realtime_websocket/protocol.rs b/codex-rs/codex-api/src/endpoint/realtime_websocket/protocol.rs
index ca88c5d91f..467d9dca95 100644
--- a/codex-rs/codex-api/src/endpoint/realtime_websocket/protocol.rs
+++ b/codex-rs/codex-api/src/endpoint/realtime_websocket/protocol.rs
@@ -48,6 +48,7 @@ pub(super) struct SessionAudio {
 #[derive(Debug, Clone, Serialize)]
 pub(super) struct SessionAudioInput {
     pub(super) format: SessionAudioFormat,
+    pub(super) turn_detection: SessionTurnDetection,
 }
 
 #[derive(Debug, Clone, Serialize)]
@@ -57,6 +58,14 @@ pub(super) struct SessionAudioFormat {
     pub(super) rate: u32,
 }
 
+#[derive(Debug, Clone, Serialize)]
+pub(super) struct SessionTurnDetection {
+    #[serde(rename = "type")]
+    pub(super) kind: String,
+    pub(super) interrupt_response: bool,
+    pub(super) create_response: bool,
+}
+
 #[derive(Debug, Clone, Serialize)]
 pub(super) struct SessionAudioOutput {
     pub(super) format: SessionAudioOutputFormat,
diff --git a/codex-rs/codex-api/tests/realtime_websocket_e2e.rs b/codex-rs/codex-api/tests/realtime_websocket_e2e.rs
index 66b99e286c..e8e90702c4 100644
--- a/codex-rs/codex-api/tests/realtime_websocket_e2e.rs
+++ b/codex-rs/codex-api/tests/realtime_websocket_e2e.rs
@@ -95,6 +95,18 @@ async fn realtime_ws_e2e_session_create_and_event_flow() {
             first_json["session"]["audio"]["input"]["format"]["rate"],
             Value::from(24_000)
         );
+        assert_eq!(
+            first_json["session"]["audio"]["input"]["turn_detection"]["type"],
+            Value::String("semantic_vad".to_string())
+        );
+        assert_eq!(
+            first_json["session"]["audio"]["input"]["turn_detection"]["interrupt_response"],
+            Value::Bool(false)
+        );
+        assert_eq!(
+            first_json["session"]["audio"]["input"]["turn_detection"]["create_response"],
+            Value::Bool(true)
+        );
         assert_eq!(
             first_json["session"]["tool_choice"],
             Value::String("auto".to_string())