Add WebRTC transport to realtime start (#16960)

Adds WebRTC startup to the experimental app-server
`thread/realtime/start` method with an optional transport enum. The
websocket path remains the default; WebRTC offers create the realtime
session through the shared start flow and emit the answer SDP via
`thread/realtime/sdp`.

---------

Co-authored-by: Codex <noreply@openai.com>

codex-rs/codex-api/src/endpoint/memories.rs

@@ -72,6 +72,7 @@ mod tests {
     use crate::provider::RetryConfig;
     use async_trait::async_trait;
     use codex_client::Request;
+    use codex_client::RequestBody;
     use codex_client::Response;
     use codex_client::StreamResponse;
     use codex_client::TransportError;
@@ -213,7 +214,11 @@ mod tests {
             request.url,
             "https://example.com/api/codex/memories/trace_summarize"
         );
-        let body = request.body.expect("request body should be present");
+        let body = request
+            .body
+            .as_ref()
+            .and_then(RequestBody::json)
+            .expect("request body should be JSON");
         assert_eq!(body["model"], "gpt-test");
         assert_eq!(body["traces"][0]["id"], "trace-1");
         assert_eq!(

codex-rs/codex-api/src/endpoint/mod.rs

@@ -1,6 +1,7 @@
 pub(crate) mod compact;
 pub(crate) mod memories;
 pub(crate) mod models;
+pub(crate) mod realtime_call;
 pub(crate) mod realtime_websocket;
 pub(crate) mod responses;
 pub(crate) mod responses_websocket;
@@ -9,6 +10,8 @@ mod session;
 pub use compact::CompactClient;
 pub use memories::MemoriesClient;
 pub use models::ModelsClient;
+pub use realtime_call::RealtimeCallClient;
+pub use realtime_call::RealtimeCallResponse;
 pub use realtime_websocket::RealtimeEventParser;
 pub use realtime_websocket::RealtimeSessionConfig;
 pub use realtime_websocket::RealtimeSessionMode;
@@ -16,6 +19,7 @@ pub use realtime_websocket::RealtimeWebsocketClient;
 pub use realtime_websocket::RealtimeWebsocketConnection;
 pub use realtime_websocket::RealtimeWebsocketEvents;
 pub use realtime_websocket::RealtimeWebsocketWriter;
+pub use realtime_websocket::session_update_session_json;
 pub use responses::ResponsesClient;
 pub use responses::ResponsesOptions;
 pub use responses_websocket::ResponsesWebsocketClient;

codex-rs/codex-api/src/endpoint/realtime_call.rs

@@ -0,0 +1,415 @@
+use crate::auth::AuthProvider;
+use crate::endpoint::realtime_websocket::RealtimeSessionConfig;
+use crate::endpoint::realtime_websocket::session_update_session_json;
+use crate::endpoint::session::EndpointSession;
+use crate::error::ApiError;
+use crate::provider::Provider;
+use bytes::Bytes;
+use codex_client::HttpTransport;
+use codex_client::RequestBody;
+use codex_client::RequestTelemetry;
+use http::HeaderMap;
+use http::HeaderValue;
+use http::Method;
+use http::header::CONTENT_TYPE;
+use serde::Serialize;
+use serde_json::Value;
+use serde_json::to_string;
+use serde_json::to_value;
+use std::sync::Arc;
+use tracing::instrument;
+
+const MULTIPART_BOUNDARY: &str = "codex-realtime-call-boundary";
+const MULTIPART_CONTENT_TYPE: &str = "multipart/form-data; boundary=codex-realtime-call-boundary";
+
+pub struct RealtimeCallClient<T: HttpTransport, A: AuthProvider> {
+    session: EndpointSession<T, A>,
+}
+
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub struct RealtimeCallResponse {
+    pub sdp: String,
+}
+
+#[derive(Serialize)]
+struct BackendRealtimeCallRequest<'a> {
+    sdp: &'a str,
+    session: &'a Value,
+}
+
+impl<T: HttpTransport, A: AuthProvider> RealtimeCallClient<T, A> {
+    pub fn new(transport: T, provider: Provider, auth: A) -> Self {
+        Self {
+            session: EndpointSession::new(transport, provider, auth),
+        }
+    }
+
+    pub fn with_telemetry(self, request: Option<Arc<dyn RequestTelemetry>>) -> Self {
+        Self {
+            session: self.session.with_request_telemetry(request),
+        }
+    }
+
+    fn path() -> &'static str {
+        "realtime/calls"
+    }
+
+    fn uses_backend_request_shape(&self) -> bool {
+        self.session.provider().base_url.contains("/backend-api")
+    }
+
+    #[instrument(
+        name = "realtime_call.create",
+        level = "info",
+        skip_all,
+        fields(
+            http.method = "POST",
+            api.path = "realtime/calls"
+        )
+    )]
+    pub async fn create(&self, sdp: String) -> Result<RealtimeCallResponse, ApiError> {
+        self.create_with_headers(sdp, HeaderMap::new()).await
+    }
+
+    pub async fn create_with_session(
+        &self,
+        sdp: String,
+        session_config: RealtimeSessionConfig,
+    ) -> Result<RealtimeCallResponse, ApiError> {
+        self.create_with_session_and_headers(sdp, session_config, HeaderMap::new())
+            .await
+    }
+
+    pub async fn create_with_headers(
+        &self,
+        sdp: String,
+        extra_headers: HeaderMap,
+    ) -> Result<RealtimeCallResponse, ApiError> {
+        let resp = self
+            .session
+            .execute_with(
+                Method::POST,
+                Self::path(),
+                extra_headers,
+                /*body*/ None,
+                |req| {
+                    req.headers
+                        .insert(CONTENT_TYPE, HeaderValue::from_static("application/sdp"));
+                    req.body = Some(RequestBody::Raw(Bytes::from(sdp.clone())));
+                },
+            )
+            .await?;
+
+        let sdp = decode_sdp_response(resp.body.as_ref())?;
+
+        Ok(RealtimeCallResponse { sdp })
+    }
+
+    pub async fn create_with_session_and_headers(
+        &self,
+        sdp: String,
+        session_config: RealtimeSessionConfig,
+        extra_headers: HeaderMap,
+    ) -> Result<RealtimeCallResponse, ApiError> {
+        let mut session = realtime_session_json(session_config)?;
+        if let Some(session) = session.as_object_mut() {
+            session.remove("id");
+        }
+        // TODO(aibrahim): Align the SIWC route with the API multipart shape and remove this branch.
+        if self.uses_backend_request_shape() {
+            let body = to_value(BackendRealtimeCallRequest {
+                sdp: &sdp,
+                session: &session,
+            })
+            .map_err(|err| ApiError::Stream(format!("failed to encode realtime call: {err}")))?;
+            let resp = self
+                .session
+                .execute(Method::POST, Self::path(), extra_headers, Some(body))
+                .await?;
+            let sdp = decode_sdp_response(resp.body.as_ref())?;
+            return Ok(RealtimeCallResponse { sdp });
+        }
+
+        let session = to_string(&session).map_err(|err| ApiError::InvalidRequest {
+            message: err.to_string(),
+        })?;
+        let mut body = Vec::new();
+        body.extend_from_slice(format!("--{MULTIPART_BOUNDARY}\r\n").as_bytes());
+        body.extend_from_slice(b"Content-Disposition: form-data; name=\"sdp\"\r\n");
+        body.extend_from_slice(b"Content-Type: application/sdp\r\n\r\n");
+        body.extend_from_slice(sdp.as_bytes());
+        body.extend_from_slice(b"\r\n");
+        body.extend_from_slice(format!("--{MULTIPART_BOUNDARY}\r\n").as_bytes());
+        body.extend_from_slice(b"Content-Disposition: form-data; name=\"session\"\r\n");
+        body.extend_from_slice(b"Content-Type: application/json\r\n\r\n");
+        body.extend_from_slice(session.as_bytes());
+        body.extend_from_slice(b"\r\n");
+        body.extend_from_slice(format!("--{MULTIPART_BOUNDARY}--\r\n").as_bytes());
+
+        let resp = self
+            .session
+            .execute_with(
+                Method::POST,
+                Self::path(),
+                extra_headers,
+                /*body*/ None,
+                |req| {
+                    req.headers.insert(
+                        CONTENT_TYPE,
+                        HeaderValue::from_static(MULTIPART_CONTENT_TYPE),
+                    );
+                    req.body = Some(RequestBody::Raw(Bytes::from(body.clone())));
+                },
+            )
+            .await?;
+
+        let sdp = decode_sdp_response(resp.body.as_ref())?;
+
+        Ok(RealtimeCallResponse { sdp })
+    }
+}
+
+fn realtime_session_json(session_config: RealtimeSessionConfig) -> Result<Value, ApiError> {
+    session_update_session_json(session_config)
+        .map_err(|err| ApiError::Stream(format!("failed to encode realtime call session: {err}")))
+}
+
+fn decode_sdp_response(body: &[u8]) -> Result<String, ApiError> {
+    String::from_utf8(body.to_vec()).map_err(|err| {
+        ApiError::Stream(format!(
+            "failed to decode realtime call SDP response: {err}"
+        ))
+    })
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::endpoint::realtime_websocket::RealtimeEventParser;
+    use crate::endpoint::realtime_websocket::RealtimeSessionMode;
+    use crate::provider::RetryConfig;
+    use async_trait::async_trait;
+    use codex_client::Request;
+    use codex_client::Response;
+    use codex_client::StreamResponse;
+    use codex_client::TransportError;
+    use http::StatusCode;
+    use pretty_assertions::assert_eq;
+    use std::sync::Mutex;
+    use std::time::Duration;
+
+    #[derive(Clone)]
+    struct CapturingTransport {
+        last_request: Arc<Mutex<Option<Request>>>,
+    }
+
+    impl CapturingTransport {
+        fn new() -> Self {
+            Self {
+                last_request: Arc::new(Mutex::new(None)),
+            }
+        }
+    }
+
+    #[async_trait]
+    impl HttpTransport for CapturingTransport {
+        async fn execute(&self, req: Request) -> Result<Response, TransportError> {
+            *self.last_request.lock().unwrap() = Some(req);
+            Ok(Response {
+                status: StatusCode::OK,
+                headers: HeaderMap::new(),
+                body: Bytes::from_static(b"v=0\r\n"),
+            })
+        }
+
+        async fn stream(&self, _req: Request) -> Result<StreamResponse, TransportError> {
+            Err(TransportError::Build("stream should not run".to_string()))
+        }
+    }
+
+    #[derive(Clone, Default)]
+    struct DummyAuth;
+
+    impl AuthProvider for DummyAuth {
+        fn bearer_token(&self) -> Option<String> {
+            Some("test-token".to_string())
+        }
+    }
+
+    fn provider(base_url: &str) -> Provider {
+        Provider {
+            name: "test".to_string(),
+            base_url: base_url.to_string(),
+            query_params: None,
+            headers: HeaderMap::new(),
+            retry: RetryConfig {
+                max_attempts: 1,
+                base_delay: Duration::from_millis(1),
+                retry_429: false,
+                retry_5xx: true,
+                retry_transport: true,
+            },
+            stream_idle_timeout: Duration::from_secs(1),
+        }
+    }
+
+    fn realtime_session_config(session_id: &str) -> RealtimeSessionConfig {
+        RealtimeSessionConfig {
+            instructions: "hi".to_string(),
+            model: Some("gpt-realtime".to_string()),
+            session_id: Some(session_id.to_string()),
+            event_parser: RealtimeEventParser::RealtimeV2,
+            session_mode: RealtimeSessionMode::Conversational,
+        }
+    }
+
+    #[tokio::test]
+    async fn sends_sdp_offer_as_raw_body() {
+        let transport = CapturingTransport::new();
+        let client = RealtimeCallClient::new(
+            transport.clone(),
+            provider("https://api.openai.com/v1"),
+            DummyAuth,
+        );
+
+        let response = client
+            .create("v=offer\r\n".to_string())
+            .await
+            .expect("request should succeed");
+
+        assert_eq!(
+            response,
+            RealtimeCallResponse {
+                sdp: "v=0\r\n".to_string()
+            }
+        );
+
+        let request = transport.last_request.lock().unwrap().clone().unwrap();
+        assert_eq!(request.method, Method::POST);
+        assert_eq!(request.url, "https://api.openai.com/v1/realtime/calls");
+        assert_eq!(
+            request.headers.get(CONTENT_TYPE).unwrap(),
+            HeaderValue::from_static("application/sdp")
+        );
+        assert_eq!(
+            request
+                .headers
+                .get(http::header::AUTHORIZATION)
+                .and_then(|value| value.to_str().ok()),
+            Some("Bearer test-token")
+        );
+        assert_eq!(
+            request.body,
+            Some(RequestBody::Raw(Bytes::from_static(b"v=offer\r\n")))
+        );
+    }
+
+    #[tokio::test]
+    async fn sends_api_session_call_as_multipart_body() {
+        let transport = CapturingTransport::new();
+        let client = RealtimeCallClient::new(
+            transport.clone(),
+            provider("https://api.openai.com/v1"),
+            DummyAuth,
+        );
+
+        let response = client
+            .create_with_session(
+                "v=offer\r\n".to_string(),
+                realtime_session_config("sess-api"),
+            )
+            .await
+            .expect("request should succeed");
+
+        assert_eq!(
+            response,
+            RealtimeCallResponse {
+                sdp: "v=0\r\n".to_string()
+            }
+        );
+
+        let request = transport.last_request.lock().unwrap().clone().unwrap();
+        assert_eq!(request.method, Method::POST);
+        assert_eq!(request.url, "https://api.openai.com/v1/realtime/calls");
+        assert_eq!(
+            request.headers.get(CONTENT_TYPE).unwrap(),
+            HeaderValue::from_static(MULTIPART_CONTENT_TYPE)
+        );
+        let Some(RequestBody::Raw(body)) = request.body else {
+            panic!("multipart body should be raw");
+        };
+        let body = std::str::from_utf8(&body).expect("multipart body should be utf-8");
+        let mut session = realtime_session_json(realtime_session_config("sess-api"))
+            .expect("session should encode");
+        session
+            .as_object_mut()
+            .expect("session should be an object")
+            .remove("id");
+        let session = to_string(&session).expect("session should serialize");
+        assert_eq!(
+            body,
+            format!(
+                "--codex-realtime-call-boundary\r\n\
+                 Content-Disposition: form-data; name=\"sdp\"\r\n\
+                 Content-Type: application/sdp\r\n\
+                 \r\n\
+                 v=offer\r\n\
+                 \r\n\
+                 --codex-realtime-call-boundary\r\n\
+                 Content-Disposition: form-data; name=\"session\"\r\n\
+                 Content-Type: application/json\r\n\
+                 \r\n\
+                 {session}\r\n\
+                 --codex-realtime-call-boundary--\r\n"
+            )
+        );
+    }
+
+    #[tokio::test]
+    async fn sends_backend_session_call_as_json_body() {
+        let transport = CapturingTransport::new();
+        let client = RealtimeCallClient::new(
+            transport.clone(),
+            provider("https://chatgpt.com/backend-api/codex"),
+            DummyAuth,
+        );
+
+        let response = client
+            .create_with_session(
+                "v=offer\r\n".to_string(),
+                realtime_session_config("sess-backend"),
+            )
+            .await
+            .expect("request should succeed");
+
+        assert_eq!(
+            response,
+            RealtimeCallResponse {
+                sdp: "v=0\r\n".to_string()
+            }
+        );
+
+        let request = transport.last_request.lock().unwrap().clone().unwrap();
+        assert_eq!(request.method, Method::POST);
+        assert_eq!(
+            request.url,
+            "https://chatgpt.com/backend-api/codex/realtime/calls"
+        );
+        let mut expected_session = realtime_session_json(realtime_session_config("sess-backend"))
+            .expect("session should encode");
+        expected_session
+            .as_object_mut()
+            .expect("session should be an object")
+            .remove("id");
+        assert_eq!(
+            request.body,
+            Some(RequestBody::Json(
+                to_value(BackendRealtimeCallRequest {
+                    sdp: "v=offer\r\n",
+                    session: &expected_session,
+                })
+                .expect("request should encode")
+            ))
+        );
+    }
+}

codex-rs/codex-api/src/endpoint/realtime_websocket/methods_common.rs

@@ -8,8 +8,12 @@ use crate::endpoint::realtime_websocket::methods_v2::session_update_session as v
 use crate::endpoint::realtime_websocket::methods_v2::websocket_intent as v2_websocket_intent;
 use crate::endpoint::realtime_websocket::protocol::RealtimeEventParser;
 use crate::endpoint::realtime_websocket::protocol::RealtimeOutboundMessage;
+use crate::endpoint::realtime_websocket::protocol::RealtimeSessionConfig;
 use crate::endpoint::realtime_websocket::protocol::RealtimeSessionMode;
 use crate::endpoint::realtime_websocket::protocol::SessionUpdateSession;
+use serde_json::Result as JsonResult;
+use serde_json::Value;
+use serde_json::to_value;
 
 pub(super) const REALTIME_AUDIO_SAMPLE_RATE: u32 = 24_000;
 const AGENT_FINAL_MESSAGE_PREFIX: &str = "\"Agent Final Message\":\n\n";
@@ -60,6 +64,17 @@ pub(super) fn session_update_session(
     }
 }
 
+pub fn session_update_session_json(config: RealtimeSessionConfig) -> JsonResult<Value> {
+    let mut session = session_update_session(
+        config.event_parser,
+        config.instructions,
+        config.session_mode,
+    );
+    session.id = config.session_id;
+    session.model = config.model;
+    to_value(session)
+}
+
 pub(super) fn websocket_intent(event_parser: RealtimeEventParser) -> Option<&'static str> {
     match event_parser {
         RealtimeEventParser::V1 => v1_websocket_intent(),

codex-rs/codex-api/src/endpoint/realtime_websocket/methods_v1.rs

@@ -40,7 +40,9 @@ pub(super) fn conversation_handoff_append_message(
 
 pub(super) fn session_update_session(instructions: String) -> SessionUpdateSession {
     SessionUpdateSession {
+        id: None,
         r#type: SessionType::Quicksilver,
+        model: None,
         instructions: Some(instructions),
         output_modalities: None,
         audio: SessionAudio {

codex-rs/codex-api/src/endpoint/realtime_websocket/methods_v2.rs

@@ -62,7 +62,9 @@ pub(super) fn session_update_session(
 ) -> SessionUpdateSession {
     match session_mode {
         RealtimeSessionMode::Conversational => SessionUpdateSession {
+            id: None,
             r#type: SessionType::Realtime,
+            model: None,
             instructions: Some(instructions),
             output_modalities: Some(vec![REALTIME_V2_OUTPUT_MODALITY_AUDIO.to_string()]),
             audio: SessionAudio {
@@ -107,7 +109,9 @@ pub(super) fn session_update_session(
             tool_choice: Some(REALTIME_V2_TOOL_CHOICE.to_string()),
         },
         RealtimeSessionMode::Transcription => SessionUpdateSession {
+            id: None,
             r#type: SessionType::Transcription,
+            model: None,
             instructions: None,
             output_modalities: None,
             audio: SessionAudio {

codex-rs/codex-api/src/endpoint/realtime_websocket/mod.rs

@@ -11,6 +11,7 @@ pub use methods::RealtimeWebsocketClient;
 pub use methods::RealtimeWebsocketConnection;
 pub use methods::RealtimeWebsocketEvents;
 pub use methods::RealtimeWebsocketWriter;
+pub use methods_common::session_update_session_json;
 pub use protocol::RealtimeEventParser;
 pub use protocol::RealtimeSessionConfig;
 pub use protocol::RealtimeSessionMode;

codex-rs/codex-api/src/endpoint/realtime_websocket/protocol.rs

@@ -48,9 +48,13 @@ pub(super) enum RealtimeOutboundMessage {
 
 #[derive(Debug, Clone, Serialize)]
 pub(super) struct SessionUpdateSession {
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub(super) id: Option<String>,
     #[serde(rename = "type")]
     pub(super) r#type: SessionType,
     #[serde(skip_serializing_if = "Option::is_none")]
+    pub(super) model: Option<String>,
+    #[serde(skip_serializing_if = "Option::is_none")]
     pub(super) instructions: Option<String>,
     #[serde(skip_serializing_if = "Option::is_none")]
     pub(super) output_modalities: Option<Vec<String>>,

codex-rs/codex-api/src/endpoint/session.rs

@@ -5,6 +5,7 @@ use crate::provider::Provider;
 use crate::telemetry::run_with_request_telemetry;
 use codex_client::HttpTransport;
 use codex_client::Request;
+use codex_client::RequestBody;
 use codex_client::RequestTelemetry;
 use codex_client::Response;
 use codex_client::StreamResponse;
@@ -53,7 +54,7 @@ impl<T: HttpTransport, A: AuthProvider> EndpointSession<T, A> {
         let mut req = self.provider.build_request(method.clone(), path);
         req.headers.extend(extra_headers.clone());
         if let Some(body) = body {
-            req.body = Some(body.clone());
+            req.body = Some(RequestBody::Json(body.clone()));
         }
         add_auth_headers(&self.auth, req)
     }

codex-rs/codex-api/src/lib.rs

@@ -37,6 +37,8 @@ pub use crate::common::response_create_client_metadata;
 pub use crate::endpoint::CompactClient;
 pub use crate::endpoint::MemoriesClient;
 pub use crate::endpoint::ModelsClient;
+pub use crate::endpoint::RealtimeCallClient;
+pub use crate::endpoint::RealtimeCallResponse;
 pub use crate::endpoint::RealtimeEventParser;
 pub use crate::endpoint::RealtimeSessionConfig;
 pub use crate::endpoint::RealtimeSessionMode;
@@ -48,6 +50,7 @@ pub use crate::endpoint::ResponsesClient;
 pub use crate::endpoint::ResponsesOptions;
 pub use crate::endpoint::ResponsesWebsocketClient;
 pub use crate::endpoint::ResponsesWebsocketConnection;
+pub use crate::endpoint::session_update_session_json;
 pub use crate::error::ApiError;
 pub use crate::provider::Provider;
 pub use crate::provider::RetryConfig;