Wire realtime api to core (#12268)

- Introduce `RealtimeConversationManager` for realtime API management 
- Add `op::conversation` to start conversation, insert audio, insert
text, and close conversation.
- emit conversation lifecycle and realtime events.
- Move shared realtime payload types into codex-protocol and add core
e2e websocket tests for start/replace/transport-close paths.

Things to consider:
- Should we use the same `op::` and `Events` channel to carry audio? I
think we should try this simple approach and later we can create
separate one if the channels got congested.
- Sending text updates to the client: we can start simple and later
restrict that.
- Provider auth isn't wired for now intentionally
This commit is contained in:
Ahmed Ibrahim
2026-02-20 19:06:35 -08:00
committed by GitHub
Unverified
parent 936e744c93
commit 6817f0be8a
28 changed files with 2102 additions and 42 deletions
+41 -2
View File
@@ -30,6 +30,11 @@ use crate::features::maybe_push_unstable_features_warning;
use crate::models_manager::manager::ModelsManager;
use crate::parse_command::parse_command;
use crate::parse_turn_item;
use crate::realtime_conversation::RealtimeConversationManager;
use crate::realtime_conversation::handle_audio as handle_realtime_conversation_audio;
use crate::realtime_conversation::handle_close as handle_realtime_conversation_close;
use crate::realtime_conversation::handle_start as handle_realtime_conversation_start;
use crate::realtime_conversation::handle_text as handle_realtime_conversation_text;
use crate::rollout::session_index;
use crate::stream_events_utils::HandleOutputCtx;
use crate::stream_events_utils::handle_non_tool_response_item;
@@ -282,7 +287,7 @@ pub struct CodexSpawnOk {
}
pub(crate) const INITIAL_SUBMIT_ID: &str = "";
pub(crate) const SUBMISSION_CHANNEL_CAPACITY: usize = 64;
pub(crate) const SUBMISSION_CHANNEL_CAPACITY: usize = 512;
const CYBER_VERIFY_URL: &str = "https://chatgpt.com/cyber";
const CYBER_SAFETY_URL: &str = "https://developers.openai.com/codex/concepts/cyber-safety";
@@ -523,11 +528,13 @@ pub(crate) struct Session {
/// session.
features: Features,
pending_mcp_server_refresh_config: Mutex<Option<McpServerRefreshConfig>>,
pub(crate) conversation: Arc<RealtimeConversationManager>,
pub(crate) active_turn: Mutex<Option<ActiveTurn>>,
pub(crate) services: SessionServices,
js_repl: Arc<JsReplHandle>,
next_internal_sub_id: AtomicU64,
}
/// The context needed for a single turn of the thread.
#[derive(Debug)]
pub(crate) struct TurnContext {
@@ -1357,6 +1364,7 @@ impl Session {
state: Mutex::new(state),
features: config.features.clone(),
pending_mcp_server_refresh_config: Mutex::new(None),
conversation: Arc::new(RealtimeConversationManager::new()),
active_turn: Mutex::new(None),
services,
js_repl,
@@ -1961,7 +1969,7 @@ impl Session {
state.take_startup_regular_task()
}
async fn get_config(&self) -> std::sync::Arc<Config> {
pub(crate) async fn get_config(&self) -> std::sync::Arc<Config> {
let state = self.state.lock().await;
state
.session_configuration
@@ -1969,6 +1977,11 @@ impl Session {
.clone()
}
pub(crate) async fn provider(&self) -> ModelProviderInfo {
let state = self.state.lock().await;
state.session_configuration.provider.clone()
}
pub(crate) async fn reload_user_config_layer(&self) {
let config_toml_path = {
let state = self.state.lock().await;
@@ -3170,6 +3183,29 @@ async fn submission_loop(sess: Arc<Session>, config: Arc<Config>, rx_sub: Receiv
Op::CleanBackgroundTerminals => {
handlers::clean_background_terminals(&sess).await;
}
Op::RealtimeConversationStart(params) => {
if let Err(err) =
handle_realtime_conversation_start(&sess, sub.id.clone(), params).await
{
sess.send_event_raw(Event {
id: sub.id.clone(),
msg: EventMsg::Error(ErrorEvent {
message: err.to_string(),
codex_error_info: Some(CodexErrorInfo::Other),
}),
})
.await;
}
}
Op::RealtimeConversationAudio(params) => {
handle_realtime_conversation_audio(&sess, sub.id.clone(), params).await;
}
Op::RealtimeConversationText(params) => {
handle_realtime_conversation_text(&sess, sub.id.clone(), params).await;
}
Op::RealtimeConversationClose => {
handle_realtime_conversation_close(&sess, sub.id.clone()).await;
}
Op::OverrideTurnContext {
cwd,
approval_policy,
@@ -4015,6 +4051,7 @@ mod handlers {
pub async fn shutdown(sess: &Arc<Session>, sub_id: String) -> bool {
sess.abort_all_tasks(TurnAbortReason::Interrupted).await;
let _ = sess.conversation.shutdown().await;
sess.services
.unified_exec_manager
.terminate_all_processes()
@@ -7467,6 +7504,7 @@ mod tests {
state: Mutex::new(state),
features: config.features.clone(),
pending_mcp_server_refresh_config: Mutex::new(None),
conversation: Arc::new(RealtimeConversationManager::new()),
active_turn: Mutex::new(None),
services,
js_repl,
@@ -7623,6 +7661,7 @@ mod tests {
state: Mutex::new(state),
features: config.features.clone(),
pending_mcp_server_refresh_config: Mutex::new(None),
conversation: Arc::new(RealtimeConversationManager::new()),
active_turn: Mutex::new(None),
services,
js_repl,
+1
View File
@@ -13,6 +13,7 @@ pub mod auth;
mod client;
mod client_common;
pub mod codex;
mod realtime_conversation;
pub use codex::SteerInputError;
mod codex_thread;
mod compact_remote;
+354
View File
@@ -0,0 +1,354 @@
use crate::CodexAuth;
use crate::api_bridge::map_api_error;
use crate::codex::Session;
use crate::default_client::default_headers;
use crate::error::CodexErr;
use crate::error::Result as CodexResult;
use async_channel::Receiver;
use async_channel::Sender;
use async_channel::TrySendError;
use codex_api::Provider as ApiProvider;
use codex_api::RealtimeAudioFrame;
use codex_api::RealtimeEvent;
use codex_api::RealtimeSessionConfig;
use codex_api::RealtimeWebsocketClient;
use codex_api::endpoint::realtime_websocket::RealtimeWebsocketEvents;
use codex_api::endpoint::realtime_websocket::RealtimeWebsocketWriter;
use codex_protocol::protocol::CodexErrorInfo;
use codex_protocol::protocol::ConversationAudioParams;
use codex_protocol::protocol::ConversationStartParams;
use codex_protocol::protocol::ConversationTextParams;
use codex_protocol::protocol::ErrorEvent;
use codex_protocol::protocol::Event;
use codex_protocol::protocol::EventMsg;
use codex_protocol::protocol::RealtimeConversationClosedEvent;
use codex_protocol::protocol::RealtimeConversationRealtimeEvent;
use codex_protocol::protocol::RealtimeConversationStartedEvent;
use http::HeaderMap;
use std::sync::Arc;
use tokio::sync::Mutex;
use tokio::task::JoinHandle;
use tracing::error;
use tracing::warn;
const AUDIO_IN_QUEUE_CAPACITY: usize = 256;
const TEXT_IN_QUEUE_CAPACITY: usize = 64;
const OUTPUT_EVENTS_QUEUE_CAPACITY: usize = 256;
pub(crate) struct RealtimeConversationManager {
state: Mutex<Option<ConversationState>>,
}
#[allow(dead_code)]
struct ConversationState {
audio_tx: Sender<RealtimeAudioFrame>,
text_tx: Sender<String>,
task: JoinHandle<()>,
}
#[allow(dead_code)]
impl RealtimeConversationManager {
pub(crate) fn new() -> Self {
Self {
state: Mutex::new(None),
}
}
pub(crate) async fn running_state(&self) -> Option<()> {
let state = self.state.lock().await;
state.as_ref().map(|_| ())
}
pub(crate) async fn start(
&self,
api_provider: ApiProvider,
extra_headers: Option<HeaderMap>,
prompt: String,
session_id: Option<String>,
) -> CodexResult<Receiver<RealtimeEvent>> {
let previous_state = {
let mut guard = self.state.lock().await;
guard.take()
};
if let Some(state) = previous_state {
state.task.abort();
let _ = state.task.await;
}
let session_config = RealtimeSessionConfig { prompt, session_id };
let client = RealtimeWebsocketClient::new(api_provider);
let connection = client
.connect(
session_config,
extra_headers.unwrap_or_default(),
default_headers(),
)
.await
.map_err(map_api_error)?;
let writer = connection.writer();
let events = connection.events();
let (audio_tx, audio_rx) =
async_channel::bounded::<RealtimeAudioFrame>(AUDIO_IN_QUEUE_CAPACITY);
let (text_tx, text_rx) = async_channel::bounded::<String>(TEXT_IN_QUEUE_CAPACITY);
let (events_tx, events_rx) =
async_channel::bounded::<RealtimeEvent>(OUTPUT_EVENTS_QUEUE_CAPACITY);
let task = spawn_realtime_input_task(writer, events, text_rx, audio_rx, events_tx);
let mut guard = self.state.lock().await;
*guard = Some(ConversationState {
audio_tx,
text_tx,
task,
});
Ok(events_rx)
}
pub(crate) async fn audio_in(&self, frame: RealtimeAudioFrame) -> CodexResult<()> {
let sender = {
let guard = self.state.lock().await;
guard.as_ref().map(|state| state.audio_tx.clone())
};
let Some(sender) = sender else {
return Err(CodexErr::InvalidRequest(
"conversation is not running".to_string(),
));
};
match sender.try_send(frame) {
Ok(()) => Ok(()),
Err(TrySendError::Full(_)) => {
warn!("dropping input audio frame due to full queue");
Ok(())
}
Err(TrySendError::Closed(_)) => Err(CodexErr::InvalidRequest(
"conversation is not running".to_string(),
)),
}
}
pub(crate) async fn text_in(&self, text: String) -> CodexResult<()> {
let sender = {
let guard = self.state.lock().await;
guard.as_ref().map(|state| state.text_tx.clone())
};
let Some(sender) = sender else {
return Err(CodexErr::InvalidRequest(
"conversation is not running".to_string(),
));
};
sender
.send(text)
.await
.map_err(|_| CodexErr::InvalidRequest("conversation is not running".to_string()))?;
Ok(())
}
pub(crate) async fn shutdown(&self) -> CodexResult<()> {
let state = {
let mut guard = self.state.lock().await;
guard.take()
};
if let Some(state) = state {
state.task.abort();
let _ = state.task.await;
}
Ok(())
}
}
pub(crate) async fn handle_start(
sess: &Arc<Session>,
sub_id: String,
params: ConversationStartParams,
) -> CodexResult<()> {
let provider = sess.provider().await;
let auth = sess.services.auth_manager.auth().await;
let api_provider = provider.to_api_provider(auth.as_ref().map(CodexAuth::auth_mode))?;
let requested_session_id = params
.session_id
.or_else(|| Some(sess.conversation_id.to_string()));
let events_rx = match sess
.conversation
.start(
api_provider,
None,
params.prompt,
requested_session_id.clone(),
)
.await
{
Ok(events_rx) => events_rx,
Err(err) => {
send_conversation_error(sess, sub_id, err.to_string(), CodexErrorInfo::Other).await;
return Ok(());
}
};
sess.send_event_raw(Event {
id: sub_id.clone(),
msg: EventMsg::RealtimeConversationStarted(RealtimeConversationStartedEvent {
session_id: requested_session_id,
}),
})
.await;
let sess_clone = Arc::clone(sess);
tokio::spawn(async move {
let ev = |msg| Event {
id: sub_id.clone(),
msg,
};
while let Ok(event) = events_rx.recv().await {
sess_clone
.send_event_raw(ev(EventMsg::RealtimeConversationRealtime(
RealtimeConversationRealtimeEvent { payload: event },
)))
.await;
}
if let Some(()) = sess_clone.conversation.running_state().await {
sess_clone
.send_event_raw(ev(EventMsg::RealtimeConversationClosed(
RealtimeConversationClosedEvent {
reason: Some("transport_closed".to_string()),
},
)))
.await;
}
});
Ok(())
}
pub(crate) async fn handle_audio(
sess: &Arc<Session>,
sub_id: String,
params: ConversationAudioParams,
) {
if let Err(err) = sess.conversation.audio_in(params.frame).await {
send_conversation_error(sess, sub_id, err.to_string(), CodexErrorInfo::BadRequest).await;
}
}
pub(crate) async fn handle_text(
sess: &Arc<Session>,
sub_id: String,
params: ConversationTextParams,
) {
if let Err(err) = sess.conversation.text_in(params.text).await {
send_conversation_error(sess, sub_id, err.to_string(), CodexErrorInfo::BadRequest).await;
}
}
pub(crate) async fn handle_close(sess: &Arc<Session>, sub_id: String) {
match sess.conversation.shutdown().await {
Ok(()) => {
sess.send_event_raw(Event {
id: sub_id,
msg: EventMsg::RealtimeConversationClosed(RealtimeConversationClosedEvent {
reason: Some("requested".to_string()),
}),
})
.await;
}
Err(err) => {
send_conversation_error(sess, sub_id, err.to_string(), CodexErrorInfo::Other).await;
}
}
}
fn spawn_realtime_input_task(
writer: RealtimeWebsocketWriter,
events: RealtimeWebsocketEvents,
text_rx: Receiver<String>,
audio_rx: Receiver<RealtimeAudioFrame>,
events_tx: Sender<RealtimeEvent>,
) -> JoinHandle<()> {
tokio::spawn(async move {
loop {
tokio::select! {
biased;
text = text_rx.recv() => {
match text {
Ok(text) => {
if let Err(err) = writer.send_conversation_item_create(text).await {
let mapped_error = map_api_error(err);
warn!("failed to send input text: {mapped_error}");
break;
}
}
Err(_) => break,
}
}
event = events.next_event() => {
match event {
Ok(Some(event)) => {
let should_stop = matches!(&event, RealtimeEvent::Error(_));
if events_tx.send(event).await.is_err() {
break;
}
if should_stop {
error!("realtime stream error event received");
break;
}
}
Ok(None) => {
let _ = events_tx
.send(RealtimeEvent::Error(
"realtime websocket connection is closed".to_string(),
))
.await;
break;
}
Err(err) => {
let mapped_error = map_api_error(err);
if events_tx
.send(RealtimeEvent::Error(mapped_error.to_string()))
.await
.is_err()
{
break;
}
error!("realtime stream closed: {mapped_error}");
break;
}
}
}
frame = audio_rx.recv() => {
match frame {
Ok(frame) => {
if let Err(err) = writer.send_audio_frame(frame).await {
let mapped_error = map_api_error(err);
error!("failed to send input audio: {mapped_error}");
break;
}
}
Err(_) => break,
}
}
}
}
})
}
async fn send_conversation_error(
sess: &Arc<Session>,
sub_id: String,
message: String,
codex_error_info: CodexErrorInfo,
) {
sess.send_event_raw(Event {
id: sub_id,
msg: EventMsg::Error(ErrorEvent {
message,
codex_error_info: Some(codex_error_info),
}),
})
.await;
}
+3
View File
@@ -122,6 +122,9 @@ fn event_msg_persistence_mode(ev: &EventMsg) -> Option<EventPersistenceMode> {
| EventMsg::CollabCloseEnd(_)
| EventMsg::CollabResumeEnd(_) => Some(EventPersistenceMode::Extended),
EventMsg::Warning(_)
| EventMsg::RealtimeConversationStarted(_)
| EventMsg::RealtimeConversationRealtime(_)
| EventMsg::RealtimeConversationClosed(_)
| EventMsg::ModelReroute(_)
| EventMsg::AgentMessageDelta(_)
| EventMsg::AgentReasoningDelta(_)