mirror of
https://github.com/pchuan98/codex.git
synced 2026-07-01 00:31:56 +08:00
[codex] Send request-scoped turn state over WebSocket (#27996)
## Context Turn state is scoped to one logical turn, but the WebSocket path currently exchanges it through upgrade headers, which are scoped to the physical connection. A connection may be reused across turns, so its handshake cannot represent the turn lifecycle reliably. ## Change Exchange turn state on each WebSocket response request instead: - send an established value in `response.create.client_metadata` - read the returned value from the existing `response.metadata` event - retain the first value in the turn-scoped `ModelClientSession` `OnceLock` - start the next logical turn without state, even when it reuses the same WebSocket connection This gives WebSocket requests the same first-value-wins contract as the existing HTTP path. ## Test plan Integration coverage verifies that: - WebSocket replays returned state on same-turn follow-ups - later response metadata does not replace the first value - state resets at the logical turn boundary without requiring a reconnect CI validates the full change. ## Stack This is 1/2. #28002 builds on this request-scoped transport to carry established state through compact requests.
This commit is contained in:
committed by
GitHub
Unverified
parent
9d938a46d9
commit
640d61b121
@@ -215,6 +215,7 @@ impl ResponsesWebsocketConnection {
|
||||
&self,
|
||||
request: ResponsesWsRequest,
|
||||
connection_reused: bool,
|
||||
turn_state: Option<Arc<OnceLock<String>>>,
|
||||
) -> Result<ResponseStream, ApiError> {
|
||||
let (tx_event, rx_event) =
|
||||
mpsc::channel::<std::result::Result<ResponseEvent, ApiError>>(1600);
|
||||
@@ -264,6 +265,7 @@ impl ResponsesWebsocketConnection {
|
||||
idle_timeout,
|
||||
telemetry,
|
||||
connection_reused,
|
||||
turn_state.as_deref(),
|
||||
)
|
||||
.await
|
||||
};
|
||||
@@ -631,6 +633,7 @@ async fn run_websocket_response_stream(
|
||||
idle_timeout: Duration,
|
||||
telemetry: Option<Arc<dyn WebsocketTelemetry>>,
|
||||
connection_reused: bool,
|
||||
turn_state: Option<&OnceLock<String>>,
|
||||
) -> Result<(), ApiError> {
|
||||
let mut last_server_model: Option<String> = None;
|
||||
send_websocket_request(
|
||||
@@ -682,6 +685,11 @@ async fn run_websocket_response_stream(
|
||||
continue;
|
||||
}
|
||||
};
|
||||
if let Some(response_turn_state) = event.turn_state()
|
||||
&& let Some(turn_state) = turn_state
|
||||
{
|
||||
let _ = turn_state.set(response_turn_state);
|
||||
}
|
||||
let model_verifications = event.model_verifications();
|
||||
let turn_moderation_metadata = event.turn_moderation_metadata();
|
||||
if event.kind() == "codex.rate_limits" {
|
||||
|
||||
@@ -23,6 +23,7 @@ use tracing::debug;
|
||||
use tracing::trace;
|
||||
|
||||
const X_REASONING_INCLUDED_HEADER: &str = "x-reasoning-included";
|
||||
const X_CODEX_TURN_STATE_HEADER: &str = "x-codex-turn-state";
|
||||
const OPENAI_MODEL_HEADER: &str = "openai-model";
|
||||
const REQUEST_ID_HEADER: &str = "x-request-id";
|
||||
const TRUSTED_ACCESS_FOR_CYBER_VERIFICATION: &str = "trusted_access_for_cyber";
|
||||
@@ -56,8 +57,8 @@ pub fn spawn_response_stream(
|
||||
if let Some(turn_state) = turn_state.as_ref()
|
||||
&& let Some(header_value) = stream_response
|
||||
.headers
|
||||
.get("x-codex-turn-state")
|
||||
.and_then(|v| v.to_str().ok())
|
||||
.get(X_CODEX_TURN_STATE_HEADER)
|
||||
.and_then(|value| value.to_str().ok())
|
||||
{
|
||||
let _ = turn_state.set(header_value.to_string());
|
||||
}
|
||||
@@ -184,6 +185,16 @@ impl ResponsesStreamEvent {
|
||||
}
|
||||
}
|
||||
|
||||
pub(crate) fn turn_state(&self) -> Option<String> {
|
||||
if self.kind() != "response.metadata" {
|
||||
return None;
|
||||
}
|
||||
|
||||
self.headers
|
||||
.as_ref()
|
||||
.and_then(header_turn_state_value_from_json)
|
||||
}
|
||||
|
||||
pub(crate) fn model_verifications(&self) -> Option<Vec<ModelVerification>> {
|
||||
if self.kind() != "response.metadata" {
|
||||
return None;
|
||||
@@ -220,6 +231,17 @@ fn header_openai_model_value_from_json(value: &Value) -> Option<String> {
|
||||
})
|
||||
}
|
||||
|
||||
fn header_turn_state_value_from_json(value: &Value) -> Option<String> {
|
||||
let headers = value.as_object()?;
|
||||
headers.iter().find_map(|(name, value)| {
|
||||
if name.eq_ignore_ascii_case(X_CODEX_TURN_STATE_HEADER) {
|
||||
json_value_as_string(value)
|
||||
} else {
|
||||
None
|
||||
}
|
||||
})
|
||||
}
|
||||
|
||||
fn model_verifications_from_json_value(value: &Value) -> Option<Vec<ModelVerification>> {
|
||||
let verifications = value
|
||||
.as_array()
|
||||
|
||||
+18
-34
@@ -806,13 +806,10 @@ impl ModelClient {
|
||||
api_provider: codex_api::Provider,
|
||||
api_auth: SharedAuthProvider,
|
||||
responses_metadata: &CodexResponsesMetadata,
|
||||
turn_state: Option<Arc<OnceLock<String>>>,
|
||||
auth_context: AuthRequestTelemetryContext,
|
||||
request_route_telemetry: RequestRouteTelemetry,
|
||||
) -> std::result::Result<ApiWebSocketConnection, ApiError> {
|
||||
let headers = self
|
||||
.build_websocket_headers(responses_metadata, turn_state.as_ref())
|
||||
.await;
|
||||
let headers = self.build_websocket_headers(responses_metadata).await;
|
||||
let websocket_telemetry = ModelClientSession::build_websocket_telemetry(
|
||||
session_telemetry,
|
||||
auth_context,
|
||||
@@ -826,7 +823,7 @@ impl ModelClient {
|
||||
ApiWebSocketResponsesClient::new(api_provider, api_auth).connect(
|
||||
headers,
|
||||
codex_login::default_client::default_headers(),
|
||||
turn_state,
|
||||
/*turn_state*/ None,
|
||||
Some(websocket_telemetry),
|
||||
),
|
||||
)
|
||||
@@ -886,16 +883,14 @@ impl ModelClient {
|
||||
}
|
||||
|
||||
/// Builds websocket handshake headers for both prewarm and turn-time reconnect.
|
||||
///
|
||||
/// Callers should pass the current turn-state lock when available so sticky-routing state is
|
||||
/// replayed on reconnect within the same turn.
|
||||
async fn build_websocket_headers(
|
||||
&self,
|
||||
responses_metadata: &CodexResponsesMetadata,
|
||||
turn_state: Option<&Arc<OnceLock<String>>>,
|
||||
) -> ApiHeaderMap {
|
||||
let mut headers =
|
||||
build_responses_headers(self.state.beta_features_header.as_deref(), turn_state);
|
||||
let mut headers = build_responses_headers(
|
||||
self.state.beta_features_header.as_deref(),
|
||||
/*turn_state*/ None,
|
||||
);
|
||||
if let Ok(header_value) = HeaderValue::from_str(&responses_metadata.thread_id) {
|
||||
headers.insert("x-client-request-id", header_value);
|
||||
}
|
||||
@@ -1088,7 +1083,6 @@ impl ModelClientSession {
|
||||
client_setup.api_provider,
|
||||
client_setup.api_auth,
|
||||
responses_metadata,
|
||||
Some(Arc::clone(&self.turn_state)),
|
||||
auth_context,
|
||||
RequestRouteTelemetry::for_endpoint(RESPONSES_ENDPOINT),
|
||||
)
|
||||
@@ -1120,7 +1114,6 @@ impl ModelClientSession {
|
||||
api_provider,
|
||||
api_auth,
|
||||
responses_metadata,
|
||||
options,
|
||||
auth_context,
|
||||
request_route_telemetry,
|
||||
} = params;
|
||||
@@ -1133,10 +1126,6 @@ impl ModelClientSession {
|
||||
self.websocket_session.last_request = None;
|
||||
self.websocket_session.last_response_rx = None;
|
||||
self.websocket_session.last_response_from_untraced_warmup = false;
|
||||
let turn_state = options
|
||||
.turn_state
|
||||
.clone()
|
||||
.unwrap_or_else(|| Arc::clone(&self.turn_state));
|
||||
let new_conn = match self
|
||||
.client
|
||||
.connect_websocket(
|
||||
@@ -1144,7 +1133,6 @@ impl ModelClientSession {
|
||||
api_provider,
|
||||
api_auth,
|
||||
responses_metadata,
|
||||
Some(turn_state),
|
||||
auth_context,
|
||||
request_route_telemetry,
|
||||
)
|
||||
@@ -1346,15 +1334,6 @@ impl ModelClientSession {
|
||||
client_setup.api_auth.as_ref(),
|
||||
pending_retry,
|
||||
);
|
||||
let compression = self.responses_request_compression(client_setup.auth.as_ref());
|
||||
|
||||
let options = self
|
||||
.build_responses_options(
|
||||
responses_metadata,
|
||||
compression,
|
||||
model_info.use_responses_lite,
|
||||
)
|
||||
.await;
|
||||
let request = self.client.build_responses_request(
|
||||
&client_setup.api_provider,
|
||||
prompt,
|
||||
@@ -1364,12 +1343,15 @@ impl ModelClientSession {
|
||||
service_tier.clone(),
|
||||
responses_metadata,
|
||||
)?;
|
||||
let mut client_metadata = self
|
||||
.client
|
||||
.build_ws_client_metadata(responses_metadata, model_info.use_responses_lite);
|
||||
if let Some(turn_state) = self.turn_state.get() {
|
||||
client_metadata.insert(X_CODEX_TURN_STATE_HEADER.to_string(), turn_state.clone());
|
||||
}
|
||||
let mut ws_payload = ResponseCreateWsRequest {
|
||||
client_metadata: response_create_client_metadata(
|
||||
Some(self.client.build_ws_client_metadata(
|
||||
responses_metadata,
|
||||
model_info.use_responses_lite,
|
||||
)),
|
||||
Some(client_metadata),
|
||||
request_trace.as_ref(),
|
||||
),
|
||||
..ResponseCreateWsRequest::from(&request)
|
||||
@@ -1384,7 +1366,6 @@ impl ModelClientSession {
|
||||
api_provider: client_setup.api_provider,
|
||||
api_auth: client_setup.api_auth,
|
||||
responses_metadata,
|
||||
options: &options,
|
||||
auth_context: request_auth_context,
|
||||
request_route_telemetry: RequestRouteTelemetry::for_endpoint(
|
||||
RESPONSES_ENDPOINT,
|
||||
@@ -1441,7 +1422,11 @@ impl ModelClientSession {
|
||||
))
|
||||
})?;
|
||||
let stream_result = websocket_connection
|
||||
.stream_request(ws_request, self.websocket_session.connection_reused())
|
||||
.stream_request(
|
||||
ws_request,
|
||||
self.websocket_session.connection_reused(),
|
||||
Some(Arc::clone(&self.turn_state)),
|
||||
)
|
||||
.await
|
||||
.map_err(|err| {
|
||||
let response_debug_context =
|
||||
@@ -1920,7 +1905,6 @@ struct WebsocketConnectParams<'a> {
|
||||
api_provider: codex_api::Provider,
|
||||
api_auth: SharedAuthProvider,
|
||||
responses_metadata: &'a CodexResponsesMetadata,
|
||||
options: &'a ApiResponsesOptions,
|
||||
auth_context: AuthRequestTelemetryContext,
|
||||
request_route_telemetry: RequestRouteTelemetry,
|
||||
}
|
||||
|
||||
@@ -585,7 +585,7 @@ async fn websocket_handshake_includes_attestation_for_chatgpt_codex_responses()
|
||||
);
|
||||
|
||||
let headers = model_client
|
||||
.build_websocket_headers(&responses_metadata, /*turn_state*/ None)
|
||||
.build_websocket_headers(&responses_metadata)
|
||||
.await;
|
||||
|
||||
assert_eq!(
|
||||
|
||||
@@ -16,6 +16,7 @@ use core_test_support::skip_if_no_network;
|
||||
use core_test_support::test_codex::test_codex;
|
||||
use pretty_assertions::assert_eq;
|
||||
use serde_json::Value;
|
||||
use serde_json::json;
|
||||
|
||||
const TURN_STATE_HEADER: &str = "x-codex-turn-state";
|
||||
|
||||
@@ -91,56 +92,111 @@ async fn responses_turn_state_persists_within_turn_and_resets_after() -> Result<
|
||||
async fn websocket_turn_state_persists_within_turn_and_resets_after() -> Result<()> {
|
||||
skip_if_no_network!(Ok(()));
|
||||
|
||||
let call_id = "ws-shell-turn-state";
|
||||
// First connection delivers turn_state; second (same turn) must send it; third (new turn) must not.
|
||||
let server = start_websocket_server_with_headers(vec![
|
||||
WebSocketConnectionConfig {
|
||||
requests: vec![vec![
|
||||
let server = start_websocket_server_with_headers(vec![WebSocketConnectionConfig {
|
||||
requests: vec![
|
||||
vec![
|
||||
json!({
|
||||
"type": "response.metadata",
|
||||
"headers": {(TURN_STATE_HEADER): "ts-1"},
|
||||
}),
|
||||
ev_response_created("resp-1"),
|
||||
ev_reasoning_item("rsn-1", &["thinking"], &[]),
|
||||
ev_shell_command_call(call_id, "echo websocket"),
|
||||
ev_shell_command_call("ws-shell-turn-state", "echo websocket"),
|
||||
ev_completed("resp-1"),
|
||||
]],
|
||||
response_headers: vec![(TURN_STATE_HEADER.to_string(), "ts-1".to_string())],
|
||||
accept_delay: None,
|
||||
close_after_requests: true,
|
||||
},
|
||||
WebSocketConnectionConfig {
|
||||
requests: vec![vec![
|
||||
],
|
||||
vec![
|
||||
ev_response_created("resp-2"),
|
||||
ev_assistant_message("msg-1", "done"),
|
||||
ev_completed("resp-2"),
|
||||
]],
|
||||
response_headers: Vec::new(),
|
||||
accept_delay: None,
|
||||
close_after_requests: true,
|
||||
},
|
||||
WebSocketConnectionConfig {
|
||||
requests: vec![vec![
|
||||
],
|
||||
vec![
|
||||
ev_response_created("resp-3"),
|
||||
ev_assistant_message("msg-2", "done"),
|
||||
ev_completed("resp-3"),
|
||||
]],
|
||||
response_headers: Vec::new(),
|
||||
accept_delay: None,
|
||||
close_after_requests: true,
|
||||
},
|
||||
])
|
||||
],
|
||||
],
|
||||
response_headers: Vec::new(),
|
||||
accept_delay: None,
|
||||
close_after_requests: false,
|
||||
}])
|
||||
.await;
|
||||
|
||||
let mut builder = test_codex();
|
||||
let test = builder.build_with_websocket_server(&server).await?;
|
||||
// Phase 1: the first response mints state for its same-turn tool follow-up.
|
||||
test.submit_turn("run the echo command").await?;
|
||||
test.submit_turn("second turn").await?;
|
||||
// Phase 2: the follow-up replays that state on the same physical connection.
|
||||
// Phase 3: the next logical turn reuses the connection but starts with empty state.
|
||||
test.submit_turn("start another turn").await?;
|
||||
|
||||
let handshakes = server.handshakes();
|
||||
assert_eq!(handshakes.len(), 3);
|
||||
assert_eq!(handshakes[0].header(TURN_STATE_HEADER), None);
|
||||
assert_eq!(server.handshakes().len(), 1);
|
||||
let requests = server.single_connection();
|
||||
assert_eq!(requests.len(), 3);
|
||||
assert_eq!(
|
||||
handshakes[1].header(TURN_STATE_HEADER),
|
||||
Some("ts-1".to_string())
|
||||
requests
|
||||
.iter()
|
||||
.map(|request| request.body_json()["client_metadata"][TURN_STATE_HEADER].clone())
|
||||
.collect::<Vec<_>>(),
|
||||
vec![json!(null), json!("ts-1"), json!(null)]
|
||||
);
|
||||
|
||||
server.shutdown().await;
|
||||
Ok(())
|
||||
}
|
||||
|
||||
#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
|
||||
async fn websocket_turn_state_is_stable_within_turn() -> Result<()> {
|
||||
skip_if_no_network!(Ok(()));
|
||||
|
||||
let server = start_websocket_server_with_headers(vec![WebSocketConnectionConfig {
|
||||
requests: vec![
|
||||
vec![
|
||||
json!({
|
||||
"type": "response.metadata",
|
||||
"headers": {(TURN_STATE_HEADER): "ts-1"},
|
||||
}),
|
||||
ev_response_created("resp-1"),
|
||||
ev_shell_command_call("ws-shell-1", "echo one"),
|
||||
ev_completed("resp-1"),
|
||||
],
|
||||
vec![
|
||||
json!({
|
||||
"type": "response.metadata",
|
||||
"headers": {(TURN_STATE_HEADER): "ts-2"},
|
||||
}),
|
||||
ev_response_created("resp-2"),
|
||||
ev_shell_command_call("ws-shell-2", "echo two"),
|
||||
ev_completed("resp-2"),
|
||||
],
|
||||
vec![
|
||||
ev_response_created("resp-3"),
|
||||
ev_assistant_message("msg-1", "done"),
|
||||
ev_completed("resp-3"),
|
||||
],
|
||||
],
|
||||
response_headers: Vec::new(),
|
||||
accept_delay: None,
|
||||
close_after_requests: false,
|
||||
}])
|
||||
.await;
|
||||
let mut builder = test_codex();
|
||||
let test = builder.build_with_websocket_server(&server).await?;
|
||||
|
||||
// Phase 1: the initial request starts empty and receives the first metadata value.
|
||||
// Phase 2: the first tool follow-up replays it and ignores a later value.
|
||||
// Phase 3: the second follow-up sends the original value on the same connection.
|
||||
test.submit_turn("run two echo commands").await?;
|
||||
|
||||
assert_eq!(server.handshakes().len(), 1);
|
||||
let requests = server.single_connection();
|
||||
assert_eq!(requests.len(), 3);
|
||||
assert_eq!(
|
||||
requests
|
||||
.iter()
|
||||
.map(|request| request.body_json()["client_metadata"][TURN_STATE_HEADER].clone())
|
||||
.collect::<Vec<_>>(),
|
||||
vec![json!(null), json!("ts-1"), json!("ts-1")]
|
||||
);
|
||||
assert_eq!(handshakes[2].header(TURN_STATE_HEADER), None);
|
||||
|
||||
server.shutdown().await;
|
||||
Ok(())
|
||||
|
||||
Reference in New Issue
Block a user