feat(gateway): Codex/OpenAI Responses WebSocket 代理模式

在 /v1/responses 上支持 WebSocket 升级,把客户端帧中继到上游 Codex /
OpenAI Responses WebSocket 端点,同时保持既有的路由、鉴权、配额与用量
语义:

- 路由与准入:control/route/ai.rs 识别 WebSocket 升级请求;
  websocket/ingress.rs 复用 API Key 鉴权、IP 规则与并发许可,并引入
  独立的 WebSocket 连接许可
- 中继:websocket/responses/* 按 connection / session / turn 分层,
  帧解析归一化、socket 写入有界、continuation 保持调度亲和性
- 配额:orchestration/codex_quota_breaker.rs 在账号配额耗尽时熔断并
  自动恢复,不再直接断开客户端连接
- 用量:每个 turn 的终态用量落库,request_metadata 记录
  websocket_mode / websocket_transport,管理端与 usage 视图暴露
  is_websocket
- 管理端:provider 可配置 Responses WebSocket 开关
This commit is contained in:
AAEE86
2026-08-17 14:50:33 +08:00
committed by ZheFox
parent 9a0d346ff3
commit 71b54070e8
72 changed files with 10441 additions and 108 deletions
+2
View File
@@ -377,11 +377,13 @@ pub struct AppState {
pub(crate) frontdoor_runtime_guards: Arc<FrontdoorRuntimeGuardConfig>,
pub(crate) request_body_buffer_budget: Arc<Semaphore>,
pub(crate) request_gate: Option<Arc<ConcurrencyGate>>,
pub(crate) websocket_connection_gate: Option<Arc<ConcurrencyGate>>,
pub(crate) auth_snapshot_load_gate: Option<Arc<ConcurrencyGate>>,
pub(crate) candidate_planning_gate: Option<Arc<ConcurrencyGate>>,
pub(crate) upstream_execution_gate: Option<Arc<ConcurrencyGate>>,
pub(crate) upstream_target_admission: Arc<crate::upstream_admission::UpstreamTargetAdmission>,
pub(crate) distributed_request_gate: Option<Arc<RuntimeSemaphore>>,
pub(crate) distributed_websocket_connection_gate: Option<Arc<RuntimeSemaphore>>,
pub(crate) client: reqwest::Client,
pub(crate) owner_forward_client: reqwest::Client,
pub(crate) auth_context_cache: Arc<AuthContextCache>,
+84 -2
View File
@@ -325,6 +325,7 @@ impl AppState {
frontdoor_runtime_guards.request_body_buffer_budget_permits,
)),
request_gate: None,
websocket_connection_gate: None,
auth_snapshot_load_gate: frontdoor_runtime_guards
.auth_snapshot_load_gate_limit
.map(|limit| Arc::new(ConcurrencyGate::new("gateway_auth_snapshot_load", limit))),
@@ -341,6 +342,7 @@ impl AppState {
),
),
distributed_request_gate: None,
distributed_websocket_connection_gate: None,
client,
owner_forward_client,
auth_context_cache: Arc::new(AuthContextCache::default()),
@@ -574,8 +576,20 @@ impl AppState {
}
pub fn with_request_concurrency_limit(mut self, limit: usize) -> Self {
self.request_gate = Some(Arc::new(ConcurrencyGate::new(
"gateway_requests",
let limit = limit.max(1);
self.request_gate = Some(Arc::new(ConcurrencyGate::new("gateway_requests", limit)));
if self.websocket_connection_gate.is_none() {
self.websocket_connection_gate = Some(Arc::new(ConcurrencyGate::new(
"gateway_websocket_connections",
limit,
)));
}
self
}
pub fn with_websocket_connection_limit(mut self, limit: usize) -> Self {
self.websocket_connection_gate = Some(Arc::new(ConcurrencyGate::new(
"gateway_websocket_connections",
limit.max(1),
)));
self
@@ -646,6 +660,11 @@ impl AppState {
self
}
pub fn with_distributed_websocket_connection_gate(mut self, gate: RuntimeSemaphore) -> Self {
self.distributed_websocket_connection_gate = Some(Arc::new(gate));
self
}
pub fn with_frontdoor_cors_config(mut self, config: FrontdoorCorsConfig) -> Self {
self.frontdoor_cors = Some(Arc::new(config));
self
@@ -1224,6 +1243,12 @@ impl AppState {
self.request_gate.as_ref().map(|gate| gate.snapshot())
}
pub(crate) fn websocket_connection_concurrency_snapshot(&self) -> Option<ConcurrencySnapshot> {
self.websocket_connection_gate
.as_ref()
.map(|gate| gate.snapshot())
}
pub(crate) fn auth_snapshot_load_concurrency_snapshot(&self) -> Option<ConcurrencySnapshot> {
self.auth_snapshot_load_gate
.as_ref()
@@ -1251,6 +1276,15 @@ impl AppState {
}
}
pub(crate) async fn distributed_websocket_connection_concurrency_snapshot(
&self,
) -> Result<Option<RuntimeSemaphoreSnapshot>, RuntimeSemaphoreError> {
match self.distributed_websocket_connection_gate.as_ref() {
Some(gate) => gate.snapshot().await.map(Some),
None => Ok(None),
}
}
pub(crate) async fn metric_samples(&self) -> Vec<MetricSample> {
let now = std::time::Instant::now();
let snapshot = self.metric_snapshot.read().await.clone();
@@ -1557,6 +1591,9 @@ impl AppState {
if let Some(snapshot) = self.request_concurrency_snapshot() {
samples.extend(snapshot.to_metric_samples("gateway_requests"));
}
if let Some(snapshot) = self.websocket_connection_concurrency_snapshot() {
samples.extend(snapshot.to_metric_samples("gateway_websocket_connections"));
}
if let Some(snapshot) = self.auth_snapshot_load_concurrency_snapshot() {
samples.extend(snapshot.to_metric_samples("gateway_auth_snapshot_load"));
}
@@ -1600,6 +1637,28 @@ impl AppState {
)])],
}
};
let distributed_websocket_connection_metrics = async {
let Some(gate) = self.distributed_websocket_connection_gate.as_ref() else {
return Vec::new();
};
match tokio::time::timeout(DISTRIBUTED_CONCURRENCY_METRICS_TIMEOUT, gate.snapshot())
.await
{
Ok(Ok(snapshot)) => {
snapshot.to_metric_samples("gateway_websocket_connections_distributed")
}
Ok(Err(_)) | Err(_) => vec![MetricSample::new(
"concurrency_unavailable",
"Whether the distributed concurrency gate is currently unavailable.",
MetricKind::Gauge,
1,
)
.with_labels(vec![MetricLabel::new(
"gate",
"gateway_websocket_connections_distributed",
)])],
}
};
let postgres_observability_metrics = async {
match tokio::time::timeout(
POSTGRES_OBSERVABILITY_METRICS_TIMEOUT,
@@ -1649,6 +1708,7 @@ impl AppState {
);
let (
distributed_request_metrics,
distributed_websocket_connection_metrics,
postgres_observability_metrics,
postgres_activity_group_metrics,
redis_runtime_metrics,
@@ -1656,6 +1716,7 @@ impl AppState {
usage_counter_pending_health_metrics,
) = tokio::join!(
distributed_request_metrics,
distributed_websocket_connection_metrics,
postgres_observability_metrics,
postgres_activity_group_metrics,
redis_runtime_metrics,
@@ -1663,6 +1724,7 @@ impl AppState {
usage_counter_pending_health_metrics,
);
samples.extend(distributed_request_metrics);
samples.extend(distributed_websocket_connection_metrics);
samples.extend(postgres_observability_metrics);
samples.extend(postgres_activity_group_metrics);
samples.extend(redis_runtime_metrics);
@@ -1783,6 +1845,26 @@ impl AppState {
Ok(AdmissionPermit::from_parts(local, distributed))
}
pub(crate) async fn try_acquire_websocket_connection_permit(
&self,
) -> Result<Option<AdmissionPermit>, RequestAdmissionError> {
let local = self
.websocket_connection_gate
.as_ref()
.map(|gate| gate.try_acquire())
.transpose()
.map_err(RequestAdmissionError::Local)?;
let distributed = match self.distributed_websocket_connection_gate.as_ref() {
Some(gate) => Some(
gate.try_acquire()
.await
.map_err(RequestAdmissionError::Distributed)?,
),
None => None,
};
Ok(AdmissionPermit::from_parts(local, distributed))
}
pub fn has_auth_api_key_data_reader(&self) -> bool {
self.data.has_auth_api_key_reader()
}