Act on reasoning-included per turn (#9402)

- Reset reasoning-included flag each turn and update compaction test

codex-rs/codex-api/src/common.rs

@@ -42,6 +42,10 @@ pub enum ResponseEvent {
     Created,
     OutputItemDone(ResponseItem),
     OutputItemAdded(ResponseItem),
+    /// Emitted when `X-Reasoning-Included: true` is present on the response,
+    /// meaning the server already accounted for past reasoning tokens and the
+    /// client should not re-estimate them.
+    ServerReasoningIncluded(bool),
     Completed {
         response_id: String,
         token_usage: Option<TokenUsage>,

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

@@ -157,6 +157,9 @@ impl Stream for AggregatedStream {
 
                     return Poll::Ready(Some(Ok(ResponseEvent::OutputItemDone(item))));
                 }
+                Poll::Ready(Some(Ok(ResponseEvent::ServerReasoningIncluded(included)))) => {
+                    return Poll::Ready(Some(Ok(ResponseEvent::ServerReasoningIncluded(included))));
+                }
                 Poll::Ready(Some(Ok(ResponseEvent::RateLimits(snapshot)))) => {
                     return Poll::Ready(Some(Ok(ResponseEvent::RateLimits(snapshot))));
                 }

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

@@ -29,18 +29,21 @@ use url::Url;
 
 type WsStream = WebSocketStream<MaybeTlsStream<TcpStream>>;
 const X_CODEX_TURN_STATE_HEADER: &str = "x-codex-turn-state";
+const X_REASONING_INCLUDED_HEADER: &str = "x-reasoning-included";
 
 pub struct ResponsesWebsocketConnection {
     stream: Arc<Mutex<Option<WsStream>>>,
     // TODO (pakrym): is this the right place for timeout?
     idle_timeout: Duration,
+    server_reasoning_included: bool,
 }
 
 impl ResponsesWebsocketConnection {
-    fn new(stream: WsStream, idle_timeout: Duration) -> Self {
+    fn new(stream: WsStream, idle_timeout: Duration, server_reasoning_included: bool) -> Self {
         Self {
             stream: Arc::new(Mutex::new(Some(stream))),
             idle_timeout,
+            server_reasoning_included,
         }
     }
 
@@ -56,11 +59,17 @@ impl ResponsesWebsocketConnection {
             mpsc::channel::<std::result::Result<ResponseEvent, ApiError>>(1600);
         let stream = Arc::clone(&self.stream);
         let idle_timeout = self.idle_timeout;
+        let server_reasoning_included = self.server_reasoning_included;
         let request_body = serde_json::to_value(&request).map_err(|err| {
             ApiError::Stream(format!("failed to encode websocket request: {err}"))
         })?;
 
         tokio::spawn(async move {
+            if server_reasoning_included {
+                let _ = tx_event
+                    .send(Ok(ResponseEvent::ServerReasoningIncluded(true)))
+                    .await;
+            }
             let mut guard = stream.lock().await;
             let Some(ws_stream) = guard.as_mut() else {
                 let _ = tx_event
@@ -111,10 +120,12 @@ impl<A: AuthProvider> ResponsesWebsocketClient<A> {
         headers.extend(extra_headers);
         apply_auth_headers(&mut headers, &self.auth);
 
-        let stream = connect_websocket(ws_url, headers, turn_state).await?;
+        let (stream, server_reasoning_included) =
+            connect_websocket(ws_url, headers, turn_state).await?;
         Ok(ResponsesWebsocketConnection::new(
             stream,
             self.provider.stream_idle_timeout,
+            server_reasoning_included,
         ))
     }
 }
@@ -137,7 +148,7 @@ async fn connect_websocket(
     url: Url,
     headers: HeaderMap,
     turn_state: Option<Arc<OnceLock<String>>>,
-) -> Result<WsStream, ApiError> {
+) -> Result<(WsStream, bool), ApiError> {
     let mut request = url
         .clone()
         .into_client_request()
@@ -147,6 +158,7 @@ async fn connect_websocket(
     let (stream, response) = tokio_tungstenite::connect_async(request)
         .await
         .map_err(|err| map_ws_error(err, &url))?;
+    let reasoning_included = response.headers().contains_key(X_REASONING_INCLUDED_HEADER);
     if let Some(turn_state) = turn_state
         && let Some(header_value) = response
             .headers()
@@ -155,7 +167,7 @@ async fn connect_websocket(
     {
         let _ = turn_state.set(header_value.to_string());
     }
-    Ok(stream)
+    Ok((stream, reasoning_included))
 }
 
 fn map_ws_error(err: WsError, url: &Url) -> ApiError {

codex-rs/codex-api/src/sse/responses.rs

@@ -25,6 +25,8 @@ use tokio_util::io::ReaderStream;
 use tracing::debug;
 use tracing::trace;
 
+const X_REASONING_INCLUDED_HEADER: &str = "x-reasoning-included";
+
 /// Streams SSE events from an on-disk fixture for tests.
 pub fn stream_from_fixture(
     path: impl AsRef<Path>,
@@ -58,6 +60,10 @@ pub fn spawn_response_stream(
         .get("X-Models-Etag")
         .and_then(|v| v.to_str().ok())
         .map(ToString::to_string);
+    let reasoning_included = stream_response
+        .headers
+        .get(X_REASONING_INCLUDED_HEADER)
+        .is_some();
     if let Some(turn_state) = turn_state.as_ref()
         && let Some(header_value) = stream_response
             .headers
@@ -74,6 +80,11 @@ pub fn spawn_response_stream(
         if let Some(etag) = models_etag {
             let _ = tx_event.send(Ok(ResponseEvent::ModelsEtag(etag))).await;
         }
+        if reasoning_included {
+            let _ = tx_event
+                .send(Ok(ResponseEvent::ServerReasoningIncluded(true)))
+                .await;
+        }
         process_sse(stream_response.bytes, tx_event, idle_timeout, telemetry).await;
     });