chore: bump version to 3.0.20

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

backend/Cargo.lock

@@ -1215,6 +1215,7 @@ dependencies = [
  "serde_json",
  "thiserror",
  "tokio",
+ "tokio-tungstenite 0.29.0",
  "tower",
  "tower-http 0.7.0",
  "tracing",

backend/crates/gateway-api/Cargo.toml

@@ -31,7 +31,9 @@ uuid = { version = "1.23.3", features = ["v7"] }
 
 [dev-dependencies]
 async-trait = "0.1.89"
-tokio = { version = "1.52.3", default-features = false, features = ["macros", "rt"] }
+axum = { version = "0.8.9", default-features = false, features = ["http1", "tokio"] }
+tokio = { version = "1.52.3", default-features = false, features = ["macros", "net", "rt-multi-thread"] }
+tokio-tungstenite = "0.29.0"
 tower = { version = "0.5.3", features = ["util"] }
 
 [lints]

backend/crates/gateway-api/tests/openai/responses/http.rs

@@ -1055,6 +1055,86 @@ async fn streaming_upstream_wire_failure_should_not_be_rewritten_as_a_gateway_er
     );
 }
 
+#[tokio::test]
+async fn atomic_uncommitted_upstream_failure_batch_should_be_forwarded_once() {
+    let trace = Arc::new(Trace::default());
+    let raw_created = Bytes::from_static(
+        b"event: response.created\r\ndata: {\"type\":\"response.created\",\"response\":{\"id\":\"resp_atomic_api\",\"model\":\"public-model\",\"status\":\"in_progress\"}}\r\n\r\n",
+    );
+    let raw_failure = Bytes::from_static(
+        b"event: response.failed\r\ndata: {\"type\":\"response.failed\",\"response\":{\"id\":\"resp_atomic_api\",\"status\":\"failed\",\"error\":{\"code\":\"rate_limit_exceeded\",\"message\":\"atomic upstream marker\"}}}\r\n\r\n",
+    );
+    let started_wire = ProviderEvent::canonical_with_wire(
+        vec![started()],
+        ProtocolWireEvent::json_with_raw_sse_metadata(
+            "openai",
+            Some("response.created".to_owned()),
+            json!({
+                "type": "response.created",
+                "response": {
+                    "id": "resp_atomic_api",
+                    "model": "public-model",
+                    "status": "in_progress"
+                }
+            }),
+            raw_created.clone(),
+            None,
+            None,
+        )
+        .expect("valid upstream started wire"),
+    );
+    let failed_wire = ProviderEvent::wire(
+        ProtocolWireEvent::json_with_raw_sse_metadata(
+            "openai",
+            Some("response.failed".to_owned()),
+            json!({
+                "type": "response.failed",
+                "response": {
+                    "id": "resp_atomic_api",
+                    "status": "failed",
+                    "error": {
+                        "code": "rate_limit_exceeded",
+                        "message": "atomic upstream marker"
+                    }
+                }
+            }),
+            raw_failure.clone(),
+            None,
+            None,
+        )
+        .expect("valid upstream failure wire"),
+    );
+    let batch = CoordinatedEvent::try_batch(
+        vec![started_wire, failed_wire],
+        CommitRequirement::CommitBeforeDelivery,
+    )
+    .expect("atomic failure delivery");
+    let session = FakeSession::streaming(
+        Arc::clone(&trace),
+        vec![
+            NextStep::Event(batch),
+            NextStep::Error(EngineError::Provider(ProviderError::new(
+                ProviderErrorKind::RateLimited,
+                UpstreamSendState::Sent,
+            ))),
+        ],
+    );
+
+    let response = stream_execution_response(Box::new(session), None).await;
+    let body = to_bytes(response.into_body(), usize::MAX)
+        .await
+        .expect("read SSE body");
+    let body = String::from_utf8(body.to_vec()).expect("SSE is UTF-8");
+
+    assert_eq!(body.matches("atomic upstream marker").count(), 1);
+    assert_eq!(body.matches("event: response.failed").count(), 1);
+    assert!(body.contains(std::str::from_utf8(&raw_created).expect("created frame")));
+    assert!(body.contains(std::str::from_utf8(&raw_failure).expect("failure frame")));
+    assert!(!body.contains("upstream capacity is temporarily unavailable"));
+    assert!(body.ends_with("data: [DONE]\n\n"));
+    assert_eq!(trace.snapshot(), vec!["next_event", "commit", "next_error"]);
+}
+
 #[tokio::test]
 async fn streaming_second_commit_request_should_fail_and_finalize_once() {
     let trace = Arc::new(Trace::default());

backend/crates/gateway-api/tests/openai/responses/websocket.rs

@@ -1,14 +1,34 @@
+use std::sync::{
+    Arc,
+    atomic::{AtomicBool, AtomicUsize, Ordering},
+};
+use std::time::{Duration, SystemTime};
+
 use axum::{
     body::Body,
     http::{Request, StatusCode, header::AUTHORIZATION},
 };
+use futures::{SinkExt, StreamExt, future::BoxFuture};
 use gateway_api::openai::responses::ResponseCreateFrameError;
+use gateway_core::engine::execution::{
+    AuthenticatedClient, ClientAuthenticationError, ExecutionService, ExecutionSession,
+    StartExecution, StartedExecution,
+};
+use gateway_core::engine::{CommitRequirement, CoordinatedEvent, EngineError, UpstreamSendState};
+use gateway_core::error::{GatewayError, ProviderError, ProviderErrorKind};
+use gateway_core::event::{
+    GatewayEvent, ProtocolWireEvent, ProviderEvent, ProviderResponseHeader, ResponseMeta,
+};
 use gateway_core::operation::Operation;
-use serde_json::json;
+use gateway_core::routing::PublicModelId;
+use serde_json::{Value, json};
+use tokio_tungstenite::connect_async;
+use tokio_tungstenite::tungstenite::Message as ClientMessage;
+use tokio_tungstenite::tungstenite::client::IntoClientRequest;
 use tower::ServiceExt;
 
 use super::decode_response_create;
-use crate::openai::{api_router, models::ModelsExecution};
+use crate::openai::{api_router, authenticated_client, models::ModelsExecution};
 
 #[test]
 fn response_create_should_default_to_the_websocket_streaming_contract() {
@@ -148,11 +168,233 @@ fn response_create_should_reject_non_boolean_stream_without_disclosing_body_valu
     assert!(!rendered.contains(opaque_stream_value));
 }
 
-// `serve_responses_websocket`/`forward_execution` 只有拿到 `hyper::upgrade::OnUpgrade`
-// 的真实 HTTP/1.1 升级连接才会运行;本 crate 的依赖图按 axum `ws` feature 编译
-// hyper(无 `http1`),测试进程内不存在能完成升级的服务端,`hyper::upgrade` 也不
-// 提供公开构造器。集成测试因此止步于 upgrade 边界:会话循环内的事件转发、错误帧
-// 与断连取消仍未被覆盖,需要 dev 依赖开启 `http1` 或 socket 泛化后才能落地。
+#[derive(Default)]
+struct AtomicFailureTrace {
+    next_calls: AtomicUsize,
+    committed: AtomicBool,
+    finalized: AtomicBool,
+}
+
+struct AtomicFailureSession {
+    trace: Arc<AtomicFailureTrace>,
+}
+
+impl ExecutionSession for AtomicFailureSession {
+    fn next_event(&mut self) -> BoxFuture<'_, Result<Option<CoordinatedEvent>, EngineError>> {
+        Box::pin(async move {
+            match self.trace.next_calls.fetch_add(1, Ordering::AcqRel) {
+                0 => Ok(Some(atomic_failure_batch())),
+                1 => {
+                    self.trace.finalized.store(true, Ordering::Release);
+                    Err(EngineError::Provider(ProviderError::new(
+                        ProviderErrorKind::RateLimited,
+                        UpstreamSendState::Sent,
+                    )))
+                }
+                _ => Ok(None),
+            }
+        })
+    }
+
+    fn collect_uncommitted(&mut self) -> BoxFuture<'_, Result<Vec<ProviderEvent>, EngineError>> {
+        Box::pin(async { Err(EngineError::InvalidDeliveryState) })
+    }
+
+    fn response_headers(&self) -> &[ProviderResponseHeader] {
+        &[]
+    }
+
+    fn commit_downstream(&mut self, _: Option<u16>) -> BoxFuture<'_, Result<(), EngineError>> {
+        Box::pin(async move {
+            self.trace.committed.store(true, Ordering::Release);
+            Ok(())
+        })
+    }
+
+    fn record_client_status(&mut self, _: u16) -> BoxFuture<'_, Result<(), EngineError>> {
+        Box::pin(async { Ok(()) })
+    }
+
+    fn is_finalized(&self) -> bool {
+        self.trace.finalized.load(Ordering::Acquire)
+    }
+
+    fn cancel(&self) {}
+
+    fn detach_finalize(self: Box<Self>) -> BoxFuture<'static, ()> {
+        Box::pin(async move {
+            self.trace.finalized.store(true, Ordering::Release);
+        })
+    }
+}
+
+struct AtomicFailureExecution {
+    client: AuthenticatedClient,
+    trace: Arc<AtomicFailureTrace>,
+}
+
+impl ExecutionService for AtomicFailureExecution {
+    fn authenticate(
+        &self,
+        plaintext: &str,
+    ) -> Result<AuthenticatedClient, ClientAuthenticationError> {
+        (plaintext == "sk_ws_atomic")
+            .then(|| self.client.clone())
+            .ok_or(ClientAuthenticationError::InvalidKey)
+    }
+
+    fn public_models(&self, _: &AuthenticatedClient) -> Vec<PublicModelId> {
+        vec![PublicModelId::new("model-a").expect("public model")]
+    }
+
+    fn contains_public_model(&self, _: &AuthenticatedClient, model: &PublicModelId) -> bool {
+        model.as_str() == "model-a"
+    }
+
+    fn start(
+        &self,
+        request: StartExecution,
+    ) -> BoxFuture<'_, Result<StartedExecution, GatewayError>> {
+        Box::pin(async move {
+            Ok(StartedExecution {
+                request_id: gateway_core::engine::ModelRequestId::new("req_ws_atomic")
+                    .expect("request id"),
+                created_at: SystemTime::now(),
+                stream: request.metadata.stream,
+                session: Box::new(AtomicFailureSession {
+                    trace: Arc::clone(&self.trace),
+                }),
+            })
+        })
+    }
+}
+
+fn atomic_failure_batch() -> CoordinatedEvent {
+    let started = ProviderEvent::canonical_with_wire(
+        vec![GatewayEvent::Started(ResponseMeta::new(
+            "resp_ws_atomic",
+            "model-a",
+        ))],
+        ProtocolWireEvent::json(
+            "openai",
+            Some("response.created".to_owned()),
+            json!({
+                "type": "response.created",
+                "response": {
+                    "id": "resp_ws_atomic",
+                    "model": "model-a",
+                    "status": "in_progress"
+                }
+            }),
+        )
+        .expect("created wire"),
+    );
+    let failed = ProviderEvent::wire(
+        ProtocolWireEvent::json(
+            "openai",
+            Some("response.failed".to_owned()),
+            json!({
+                "type": "response.failed",
+                "response": {
+                    "id": "resp_ws_atomic",
+                    "status": "failed",
+                    "error": {
+                        "code": "rate_limit_exceeded",
+                        "message": "websocket atomic upstream marker"
+                    }
+                }
+            }),
+        )
+        .expect("failed wire"),
+    );
+    CoordinatedEvent::try_batch(
+        vec![started, failed],
+        CommitRequirement::CommitBeforeDelivery,
+    )
+    .expect("atomic WebSocket batch")
+}
+
+#[tokio::test]
+async fn websocket_atomic_upstream_failure_batch_should_be_forwarded_once() {
+    let trace = Arc::new(AtomicFailureTrace::default());
+    let execution = Arc::new(AtomicFailureExecution {
+        client: authenticated_client("sk_ws_atomic"),
+        trace: Arc::clone(&trace),
+    });
+    let app = api_router(execution).await;
+    let listener = tokio::net::TcpListener::bind("127.0.0.1:0")
+        .await
+        .expect("bind WebSocket test server");
+    let address = listener.local_addr().expect("test server address");
+    let server = tokio::spawn(async move {
+        axum::serve(listener, app)
+            .await
+            .expect("serve WebSocket test router");
+    });
+    let mut request = format!("ws://{address}/v1/responses")
+        .into_client_request()
+        .expect("WebSocket request");
+    request.headers_mut().insert(
+        AUTHORIZATION,
+        "Bearer sk_ws_atomic".parse().expect("authorization"),
+    );
+    let (mut socket, response) = connect_async(request).await.expect("upgrade WebSocket");
+    assert_eq!(response.status(), StatusCode::SWITCHING_PROTOCOLS);
+    socket
+        .send(ClientMessage::Text(
+            json!({
+                "type": "response.create",
+                "model": "model-a",
+                "input": "hello"
+            })
+            .to_string()
+            .into(),
+        ))
+        .await
+        .expect("send response.create");
+
+    let mut messages = Vec::<Value>::new();
+    loop {
+        let message = tokio::time::timeout(Duration::from_secs(2), socket.next())
+            .await
+            .expect("WebSocket response timeout")
+            .expect("WebSocket remains open")
+            .expect("valid WebSocket frame");
+        let ClientMessage::Text(text) = message else {
+            continue;
+        };
+        let value: Value = serde_json::from_str(&text).expect("JSON WebSocket event");
+        let failed = value.get("type").and_then(Value::as_str) == Some("response.failed");
+        messages.push(value);
+        if failed {
+            break;
+        }
+    }
+
+    assert_eq!(
+        messages
+            .iter()
+            .filter(|message| message.get("type").and_then(Value::as_str) == Some("response.failed"))
+            .count(),
+        1
+    );
+    let failed = messages
+        .iter()
+        .find(|message| message.get("type").and_then(Value::as_str) == Some("response.failed"))
+        .expect("upstream failed event");
+    assert_eq!(
+        failed
+            .pointer("/response/error/message")
+            .and_then(Value::as_str),
+        Some("websocket atomic upstream marker")
+    );
+    assert!(trace.committed.load(Ordering::Acquire));
+    assert_eq!(trace.next_calls.load(Ordering::Acquire), 2);
+    assert!(trace.finalized.load(Ordering::Acquire));
+
+    socket.close(None).await.expect("close WebSocket");
+    server.abort();
+}
 
 const TEST_WEBSOCKET_KEY: &str = "AAAAAAAAAAAAAAAAAAAAAA==";
 

backend/crates/gateway-core/src/engine/coordinator.rs

@@ -165,7 +165,9 @@ where
             client_response_id: None,
             upstream_response_id: None,
             last_account_exhaustion: None,
+            last_account_exhaustion_events: Vec::new(),
             provider_attempt_outcomes: Vec::new(),
+            pending_terminal_failure: None,
         };
 
         if session.cancellation.is_cancelled() {
@@ -243,7 +245,16 @@ pub struct ResponseExecutionSession<S: ?Sized> {
     client_response_id: Option<String>,
     upstream_response_id: Option<String>,
     last_account_exhaustion: Option<ProviderError>,
+    /// 最近一次账号耗尽在重试时被丢弃的原始失败批次；只用于后续空选路终态。
+    last_account_exhaustion_events: Vec<ProviderEvent>,
     provider_attempt_outcomes: Vec<ProviderAttemptOutcome>,
+    /// 原子失败批次已交给协议层、但尚待下游提交后收敛的原 Provider 错误。
+    pending_terminal_failure: Option<PendingTerminalFailure>,
+}
+
+struct PendingTerminalFailure {
+    error: ProviderError,
+    send_state: UpstreamSendState,
 }
 
 impl<S: ?Sized> ResponseExecutionSession<S>
@@ -259,6 +270,12 @@ where
         if self.delivery_pending {
             return Err(EngineError::DownstreamCommitRequired);
         }
+        if let Some(pending) = self.pending_terminal_failure.take() {
+            let engine_error = provider_engine_error(&pending.error);
+            self.finish_provider_error_with_send_state(&pending.error, pending.send_state)
+                .await?;
+            return Err(engine_error);
+        }
         if self.finalized {
             return Ok(None);
         }
@@ -278,6 +295,21 @@ where
                     )));
                 }
                 PullOutcome::AttemptDiscarded => {}
+                PullOutcome::TerminalFailure {
+                    events,
+                    error,
+                    send_state,
+                } => {
+                    let requirement = if self.downstream_committed_at.is_some() {
+                        CommitRequirement::AlreadyCommitted
+                    } else {
+                        self.delivery_pending = true;
+                        CommitRequirement::CommitBeforeDelivery
+                    };
+                    self.pending_terminal_failure =
+                        Some(PendingTerminalFailure { error, send_state });
+                    return CoordinatedEvent::try_batch(events, requirement).map(Some);
+                }
                 PullOutcome::End => {
                     if self.downstream_committed_at.is_some() {
                         self.finish_success().await?;
@@ -306,6 +338,14 @@ where
             match self.pull().await? {
                 PullOutcome::Event(event) => events.push(event),
                 PullOutcome::AttemptDiscarded => events.clear(),
+                PullOutcome::TerminalFailure {
+                    error, send_state, ..
+                } => {
+                    let engine_error = provider_engine_error(&error);
+                    self.finish_provider_error_with_send_state(&error, send_state)
+                        .await?;
+                    return Err(engine_error);
+                }
                 PullOutcome::End => {
                     if events.is_empty() {
                         return Err(EngineError::InvalidDeliveryState);
@@ -434,13 +474,20 @@ where
             return Ok(());
         }
         self.delivery_pending = false;
+        if let Some(pending) = self.pending_terminal_failure.take() {
+            return self
+                .finish_provider_error_with_send_state(&pending.error, pending.send_state)
+                .await;
+        }
         self.finish_interruption(EngineError::Cancelled).await
     }
 
     async fn pull(&mut self) -> Result<PullOutcome, EngineError> {
         loop {
-            if self.current.is_none() {
-                self.prepare_attempt().await?;
+            if self.current.is_none()
+                && let Some(outcome) = self.prepare_attempt().await?
+            {
+                return Ok(outcome);
             }
 
             let boundary = {
@@ -470,10 +517,14 @@ where
                     if let Some(observation) = event.take_observation()
                         && let Err(error) = self.observe_response(observation)
                     {
-                        if self.handle_stream_error(error).await? {
-                            return Ok(PullOutcome::AttemptDiscarded);
+                        match self.handle_stream_error(error).await? {
+                            StreamErrorOutcome::AttemptDiscarded => {
+                                return Ok(PullOutcome::AttemptDiscarded);
+                            }
+                            terminal @ StreamErrorOutcome::TerminalFailure { .. } => {
+                                return Ok(terminal.into_pull_outcome());
+                            }
                         }
-                        continue;
                     }
                     let mut identity_error = None;
                     for fact in event.canonical_facts() {
@@ -483,10 +534,14 @@ where
                         }
                     }
                     if let Some(error) = identity_error {
-                        if self.handle_stream_error(error).await? {
-                            return Ok(PullOutcome::AttemptDiscarded);
+                        match self.handle_stream_error(error).await? {
+                            StreamErrorOutcome::AttemptDiscarded => {
+                                return Ok(PullOutcome::AttemptDiscarded);
+                            }
+                            terminal @ StreamErrorOutcome::TerminalFailure { .. } => {
+                                return Ok(terminal.into_pull_outcome());
+                            }
                         }
-                        continue;
                     }
                     for fact in event.canonical_facts() {
                         self.observe_event(fact).await?;
@@ -500,8 +555,13 @@ where
                     return Ok(PullOutcome::Event(event));
                 }
                 PollBoundary::Item(Some(Err(error))) => {
-                    if self.handle_stream_error(error).await? {
-                        return Ok(PullOutcome::AttemptDiscarded);
+                    match self.handle_stream_error(error).await? {
+                        StreamErrorOutcome::AttemptDiscarded => {
+                            return Ok(PullOutcome::AttemptDiscarded);
+                        }
+                        terminal @ StreamErrorOutcome::TerminalFailure { .. } => {
+                            return Ok(terminal.into_pull_outcome());
+                        }
                     }
                 }
                 PollBoundary::Item(None) => {
@@ -513,7 +573,7 @@ where
         }
     }
 
-    async fn prepare_attempt(&mut self) -> Result<(), EngineError> {
+    async fn prepare_attempt(&mut self) -> Result<Option<PullOutcome>, EngineError> {
         if self.attempts >= self.plan.max_attempts().get() {
             return Err(EngineError::EmptyRoutingPlan);
         }
@@ -615,6 +675,16 @@ where
                     if error.kind() == ProviderErrorKind::NoEligibleAccount
                         && let Some(exhaustion) = self.last_account_exhaustion.clone()
                     {
+                        let send_state = self.current_send_state();
+                        let mut events = std::mem::take(&mut self.last_account_exhaustion_events);
+                        if !events.is_empty() {
+                            self.observe_atomic_terminal_events(&mut events)?;
+                            return Ok(Some(PullOutcome::TerminalFailure {
+                                events,
+                                error: exhaustion,
+                                send_state,
+                            }));
+                        }
                         let engine_error = provider_engine_error(&exhaustion);
                         self.finish_provider_error(&exhaustion).await?;
                         return Err(engine_error);
@@ -739,10 +809,24 @@ where
             send_observed: false,
             response_observation: None,
         });
-        Ok(())
+        Ok(None)
     }
 
     async fn observe_event(&mut self, event: &GatewayEvent) -> Result<(), EngineError> {
+        self.observe_event_facts(event);
+        let current = self.current.as_mut().ok_or(EngineError::NoActiveAttempt)?;
+        if !current.send_observed {
+            self.engine
+                .store()
+                .mark_send_state(&self.request_id, UpstreamSendState::Sent)
+                .await?;
+            current.send_observed = true;
+            self.send_state_watermark = UpstreamSendState::Sent;
+        }
+        Ok(())
+    }
+
+    fn observe_event_facts(&mut self, event: &GatewayEvent) {
         let elapsed = elapsed_ms(self.timing_started_at);
         observe_event_timing(&mut self.timings, event, elapsed);
         if let GatewayEvent::Usage(observed) = event {
@@ -756,16 +840,6 @@ where
         if let GatewayEvent::ProviderCost(observed) = event {
             self.cost = observed.into_estimate();
         }
-        let current = self.current.as_mut().ok_or(EngineError::NoActiveAttempt)?;
-        if !current.send_observed {
-            self.engine
-                .store()
-                .mark_send_state(&self.request_id, UpstreamSendState::Sent)
-                .await?;
-            current.send_observed = true;
-            self.send_state_watermark = UpstreamSendState::Sent;
-        }
-        Ok(())
     }
 
     async fn observe_wire_event(&mut self) -> Result<(), EngineError> {
@@ -855,14 +929,22 @@ where
         Ok(())
     }
 
-    /// 返回 `true` 表示调用方必须丢弃本 attempt 已收集的未提交事件。
-    async fn handle_stream_error(&mut self, error: ProviderError) -> Result<bool, EngineError> {
+    /// 失败可重试时要求丢弃本 attempt；预算耗尽时可返回最后一批原始失败事件。
+    async fn handle_stream_error(
+        &mut self,
+        mut error: ProviderError,
+    ) -> Result<StreamErrorOutcome, EngineError> {
+        // 原始 wire 只能活在本次决策栈内；后续 clone、attempt 记录与终态对象均只
+        // 接触已剥离的稳定错误字段。
+        let mut atomic_client_events = error.take_atomic_client_events();
         let current = self.current.take().ok_or(EngineError::NoActiveAttempt)?;
-        if matches!(
+        let account_exhausted = matches!(
             error.kind(),
             ProviderErrorKind::RateLimited | ProviderErrorKind::QuotaExhausted
-        ) {
+        );
+        if account_exhausted {
             self.last_account_exhaustion = Some(error.clone());
+            self.last_account_exhaustion_events.clear();
         }
         self.record_provider_failure(current.metadata.provider().clone(), error.kind());
         // attempt_send_state 是本 attempt 自身的发送事实，驱动重试门；
@@ -910,6 +992,9 @@ where
         let retryable = continuation_retry || same_account_retry || ordinary_retry;
 
         if retryable {
+            if account_exhausted {
+                self.last_account_exhaustion_events = atomic_client_events;
+            }
             if same_account_retry {
                 if let Some(account) = current.metadata.provider_account_id() {
                     self.credential_recovery_attempted_accounts
@@ -946,16 +1031,42 @@ where
                 })
                 .await?;
             self.reset_uncommitted_observations();
-            return Ok(true);
+            return Ok(StreamErrorOutcome::AttemptDiscarded);
         }
 
         let engine_error = provider_engine_error(&error);
         self.current = Some(current);
+        if !atomic_client_events.is_empty() {
+            self.observe_atomic_terminal_events(&mut atomic_client_events)?;
+            return Ok(StreamErrorOutcome::TerminalFailure {
+                events: atomic_client_events,
+                error,
+                send_state,
+            });
+        }
         self.finish_provider_error_with_send_state(&error, send_state)
             .await?;
         Err(engine_error)
     }
 
+    fn observe_atomic_terminal_events(
+        &mut self,
+        events: &mut [ProviderEvent],
+    ) -> Result<(), EngineError> {
+        for event in events {
+            if let Some(observation) = event.take_observation() {
+                self.observe_response(observation)
+                    .map_err(EngineError::Provider)?;
+            }
+            for fact in event.canonical_facts() {
+                self.observe_response_identity(fact)
+                    .map_err(EngineError::Provider)?;
+                self.observe_event_facts(fact);
+            }
+        }
+        Ok(())
+    }
+
     fn prepare_continuation_retry(
         &mut self,
         current: &CurrentAttempt,
@@ -1305,9 +1416,40 @@ where
 enum PullOutcome {
     Event(ProviderEvent),
     AttemptDiscarded,
+    TerminalFailure {
+        events: Vec<ProviderEvent>,
+        error: ProviderError,
+        send_state: UpstreamSendState,
+    },
     End,
 }
 
+enum StreamErrorOutcome {
+    AttemptDiscarded,
+    TerminalFailure {
+        events: Vec<ProviderEvent>,
+        error: ProviderError,
+        send_state: UpstreamSendState,
+    },
+}
+
+impl StreamErrorOutcome {
+    fn into_pull_outcome(self) -> PullOutcome {
+        match self {
+            Self::AttemptDiscarded => PullOutcome::AttemptDiscarded,
+            Self::TerminalFailure {
+                events,
+                error,
+                send_state,
+            } => PullOutcome::TerminalFailure {
+                events,
+                error,
+                send_state,
+            },
+        }
+    }
+}
+
 enum PollBoundary {
     Item(Option<Result<ProviderEvent, ProviderError>>),
     Cancelled,

backend/crates/gateway-core/src/engine/execution.rs

@@ -525,7 +525,8 @@ impl DefaultExecutionService {
         })
     }
 
-    /// 探测失败先落库再转成客户端错误；转换会丢掉上游状态码与 provider code。
+    /// 探测失败先记录完整分类事实，再仅把 Provider 显式标记为客户端可见的
+    /// 结构化 message/code/type 带入协议错误。
     async fn observe_probe_failure(
         &self,
         observed: &ProbeObservation,

backend/crates/gateway-core/src/error.rs

@@ -6,6 +6,7 @@ use std::time::Duration;
 use thiserror::Error;
 
 use crate::engine::UpstreamSendState;
+use crate::event::ProviderEvent;
 
 /// 应用层标识不满足约束。
 #[derive(Debug, Clone, PartialEq, Eq, Error)]
@@ -266,9 +267,13 @@ impl fmt::Debug for ClientVisibleUpstreamError {
 
 /// 单次 Provider 调用的稳定错误。
 ///
-/// 该类型不接收原始响应正文，也不会在 `Debug` 或 `Display` 中打印上游
+/// 稳定诊断字段不接收原始响应正文，也不会在 `Debug` 或 `Display` 中打印上游
 /// code/request ID/response ID。Adapter 若捕获到可能含 secret 的上下文，只能调用
 /// [`ProviderError::redact_sensitive_context`] 丢弃正文并留下脱敏标记。
+///
+/// 唯一例外是 [`ProviderError::with_atomic_client_events`]：它只承载尚未交付客户端、
+/// 必须与本错误一起完成重试判断的协议事件。Core 会在任何持久化或日志记录前取走
+/// 这些事件，不能把它们当作诊断上下文使用。
 #[derive(Clone)]
 pub struct ProviderError {
     kind: ProviderErrorKind,
@@ -284,8 +289,12 @@ pub struct ProviderError {
     retry_same_account: bool,
     sensitive_context_redacted: bool,
     client_visible_upstream_error: Option<Box<ClientVisibleUpstreamError>>,
+    atomic_client_events: Option<Box<AtomicClientEvents>>,
 }
 
+#[derive(Clone)]
+struct AtomicClientEvents(Vec<ProviderEvent>);
+
 impl ProviderError {
     /// 创建 Provider 错误。
     #[must_use]
@@ -304,6 +313,7 @@ impl ProviderError {
             retry_same_account: false,
             sensitive_context_redacted: false,
             client_visible_upstream_error: None,
+            atomic_client_events: None,
         }
     }
 
@@ -379,6 +389,17 @@ impl ProviderError {
         self
     }
 
+    /// 附加必须先由 Core 判断重试、再决定丢弃或交付的客户端协议事件。
+    ///
+    /// Provider 只能放入当前 attempt 尚未交付的事件；调用方应设置明确的有界缓冲，
+    /// 并保证失败事件本身位于这批事件的末尾。
+    #[must_use]
+    pub fn with_atomic_client_events(mut self, events: Vec<ProviderEvent>) -> Self {
+        self.atomic_client_events =
+            (!events.is_empty()).then(|| Box::new(AtomicClientEvents(events)));
+        self
+    }
+
     /// 丢弃敏感正文，只记录“发生过脱敏”这一事实。
     #[must_use]
     pub fn redact_sensitive_context(mut self, _sensitive_context: impl AsRef<str>) -> Self {
@@ -462,6 +483,23 @@ impl ProviderError {
     pub fn client_visible_upstream_error(&self) -> Option<&ClientVisibleUpstreamError> {
         self.client_visible_upstream_error.as_deref()
     }
+
+    /// 取走只供本次重试/提交决策使用的原子客户端事件。
+    ///
+    /// Core 必须在克隆错误、记录中间失败或构造终态前调用本方法，避免原始 wire
+    /// 进入诊断或持久化对象。
+    #[must_use]
+    pub fn take_atomic_client_events(&mut self) -> Vec<ProviderEvent> {
+        self.atomic_client_events
+            .take()
+            .map_or_else(Vec::new, |events| events.0)
+    }
+
+    /// 返回是否携带尚未提交的原子客户端事件；不暴露事件内容。
+    #[must_use]
+    pub fn has_atomic_client_events(&self) -> bool {
+        self.atomic_client_events.is_some()
+    }
 }
 
 impl fmt::Debug for ProviderError {
@@ -508,6 +546,13 @@ impl fmt::Debug for ProviderError {
                     .as_ref()
                     .map(|_| "<present>"),
             )
+            .field(
+                "atomic_client_events",
+                &self
+                    .atomic_client_events
+                    .as_ref()
+                    .map_or(0, |events| events.0.len()),
+            )
             .finish()
     }
 }

backend/crates/gateway-core/src/provider_ports/mod.rs

@@ -13,7 +13,7 @@ use crate::engine::credential::{
     OpaqueProviderData, ProviderAccountId, ProviderAccountStore,
 };
 use crate::error::{IdentifierError, validate_text};
-use crate::routing::ProviderKind;
+use crate::routing::{ProviderKind, UpstreamModelId};
 
 const MAX_PENDING_FLOW_TTL: Duration = Duration::from_secs(30 * 60);
 
@@ -527,6 +527,80 @@ impl ProviderCooldown {
     }
 }
 
+/// 可丢失 cooldown 的细粒度作用域。
+#[derive(Debug, Clone, PartialEq, Eq, PartialOrd, Ord, Hash)]
+pub enum ProviderCooldownScope {
+    /// 只阻止同一账号调用指定上游模型。
+    UpstreamModel(UpstreamModelId),
+}
+
+impl ProviderCooldownScope {
+    #[must_use]
+    pub const fn upstream_model(model: UpstreamModelId) -> Self {
+        Self::UpstreamModel(model)
+    }
+
+    #[must_use]
+    pub const fn kind(&self) -> &'static str {
+        match self {
+            Self::UpstreamModel(_) => "model",
+        }
+    }
+
+    #[must_use]
+    pub fn value(&self) -> &str {
+        match self {
+            Self::UpstreamModel(model) => model.as_str(),
+        }
+    }
+}
+
+/// 不进入账号持久状态的账号+作用域 cooldown。
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub struct ProviderScopedCooldown {
+    account_id: ProviderAccountId,
+    credential_revision: CredentialRevision,
+    scope: ProviderCooldownScope,
+    until: SystemTime,
+}
+
+impl ProviderScopedCooldown {
+    #[must_use]
+    pub const fn new(
+        account_id: ProviderAccountId,
+        credential_revision: CredentialRevision,
+        scope: ProviderCooldownScope,
+        until: SystemTime,
+    ) -> Self {
+        Self {
+            account_id,
+            credential_revision,
+            scope,
+            until,
+        }
+    }
+
+    #[must_use]
+    pub const fn account_id(&self) -> &ProviderAccountId {
+        &self.account_id
+    }
+
+    #[must_use]
+    pub const fn credential_revision(&self) -> CredentialRevision {
+        self.credential_revision
+    }
+
+    #[must_use]
+    pub const fn scope(&self) -> &ProviderCooldownScope {
+        &self.scope
+    }
+
+    #[must_use]
+    pub const fn until(&self) -> SystemTime {
+        self.until
+    }
+}
+
 pub trait ProviderCooldownPort: Send + Sync {
     fn put_if_later(
         &self,
@@ -543,6 +617,24 @@ pub trait ProviderCooldownPort: Send + Sync {
         account_id: &'a ProviderAccountId,
         through_revision: CredentialRevision,
     ) -> BoxFuture<'a, Result<bool, ProviderStoreError>>;
+
+    fn put_scoped_if_later(
+        &self,
+        cooldown: ProviderScopedCooldown,
+    ) -> BoxFuture<'_, Result<bool, ProviderStoreError>>;
+
+    fn read_scoped<'a>(
+        &'a self,
+        account_id: &'a ProviderAccountId,
+        scope: &'a ProviderCooldownScope,
+    ) -> BoxFuture<'a, Result<Option<ProviderScopedCooldown>, ProviderStoreError>>;
+
+    fn clear_scoped<'a>(
+        &'a self,
+        account_id: &'a ProviderAccountId,
+        scope: &'a ProviderCooldownScope,
+        through_revision: CredentialRevision,
+    ) -> BoxFuture<'a, Result<bool, ProviderStoreError>>;
 }
 
 #[derive(Debug, Clone, Copy, PartialEq, Eq)]

backend/crates/gateway-core/tests/engine/coordinator.rs

@@ -399,6 +399,69 @@ fn complete_stream(total_tokens: Option<u64>) -> Vec<Result<GatewayEvent, Provid
     events
 }
 
+fn atomic_response_failed(response_id: &str, marker: &str) -> ProviderError {
+    let started = ProviderEvent::canonical_with_wire(
+        vec![GatewayEvent::Started(ResponseMeta::new(
+            response_id,
+            "gpt-5",
+        ))],
+        ProtocolWireEvent::json(
+            "openai",
+            Some("response.created".to_owned()),
+            json!({
+                "type": "response.created",
+                "response": {"id": response_id, "model": "gpt-5", "status": "in_progress"}
+            }),
+        )
+        .expect("started wire"),
+    );
+    let failed = ProviderEvent::wire(
+        ProtocolWireEvent::json(
+            "openai",
+            Some("response.failed".to_owned()),
+            json!({
+                "type": "response.failed",
+                "response": {
+                    "id": response_id,
+                    "status": "failed",
+                    "error": {"code": "rate_limit_exceeded", "message": marker}
+                }
+            }),
+        )
+        .expect("failed wire"),
+    );
+    ProviderError::new(ProviderErrorKind::RateLimited, UpstreamSendState::Sent)
+        .with_status(429)
+        .with_replay_safe()
+        .with_atomic_client_events(vec![started, failed])
+}
+
+fn bare_atomic_response_failed(response_id: &str) -> ProviderError {
+    let failed = ProviderEvent::canonical_with_wire(
+        vec![GatewayEvent::Started(ResponseMeta::new(
+            response_id,
+            "fallback-model",
+        ))],
+        ProtocolWireEvent::json(
+            "openai",
+            Some("response.failed".to_owned()),
+            json!({
+                "type": "response.failed",
+                "response": {
+                    "id": response_id,
+                    "status": "failed",
+                    "error": {"code": "rate_limit_exceeded", "message": "bare failure"}
+                }
+            }),
+        )
+        .expect("bare failed wire"),
+    );
+    ProviderError::new(ProviderErrorKind::RateLimited, UpstreamSendState::Sent)
+        .with_status(429)
+        .with_replay_safe()
+        .with_atomic_client_events(vec![failed])
+}
+
 fn image_stream(image_output_tokens: Option<u64>) -> Vec<Result<GatewayEvent, ProviderError>> {
     let mut usage = Usage::new();
     usage.input_tokens = Some(12);
@@ -1837,6 +1900,277 @@ fn explicit_429_after_structural_event_should_retry_before_commit() {
     assert_eq!(state.finalizations[0].outcome, ExecutionOutcome::Succeeded);
 }
 
+#[test]
+fn atomic_response_failed_should_rotate_account_before_stream_commit() {
+    let operation = generate_operation();
+    let route_plan = plan(&operation);
+    let (coordinator, store, provider) = coordinator(vec![
+        Script::ObservedStream {
+            account_id: "acct_first",
+            items: vec![Err(atomic_response_failed(
+                "response-first-failed",
+                "first account failure",
+            ))],
+        },
+        Script::Stream {
+            account_id: "acct_second",
+            items: complete_stream(None),
+        },
+    ]);
+    let mut session = block_on(coordinator.start(
+        model_request(&operation, SystemTime::now() + Duration::from_secs(30)),
+        operation,
+        route_plan,
+        None,
+        None,
+        CancellationToken::new(),
+    ))
+    .expect("start execution");
+
+    let first = block_on(session.next_event())
+        .expect("second account should produce the first delivery")
+        .expect("started event");
+    let delivered = first.into_provider_events();
+    assert!(delivered.iter().flat_map(ProviderEvent::canonical_facts).any(
+        |event| matches!(event, GatewayEvent::Started(meta) if meta.response_id() == "response-1")
+    ));
+    assert!(
+        !delivered
+            .iter()
+            .filter_map(ProviderEvent::wire_event)
+            .any(|wire| {
+                wire.data().pointer("/response/id").and_then(Value::as_str)
+                    == Some("response-first-failed")
+            })
+    );
+    block_on(session.commit_downstream(Some(200))).expect("commit second account");
+    while block_on(session.next_event())
+        .expect("successful stream")
+        .is_some()
+    {}
+
+    assert_eq!(provider.contexts.lock().expect("contexts lock").len(), 2);
+    let state = store.state.lock().expect("store lock");
+    assert_eq!(state.commits, 1);
+    assert_eq!(state.intermediate_failures, 1);
+    assert_eq!(state.finalizations[0].outcome, ExecutionOutcome::Succeeded);
+}
+
+#[test]
+fn bare_atomic_response_failed_should_rotate_before_the_first_delivery() {
+    let operation = generate_operation();
+    let route_plan = plan(&operation);
+    let (coordinator, store, provider) = coordinator(vec![
+        Script::ObservedStream {
+            account_id: "acct_first",
+            items: vec![Err(bare_atomic_response_failed("response-bare-failed"))],
+        },
+        Script::Stream {
+            account_id: "acct_second",
+            items: complete_stream(None),
+        },
+    ]);
+    let mut session = block_on(coordinator.start(
+        model_request(&operation, SystemTime::now() + Duration::from_secs(30)),
+        operation,
+        route_plan,
+        None,
+        None,
+        CancellationToken::new(),
+    ))
+    .expect("start execution");
+
+    let first = block_on(session.next_event())
+        .expect("second account should replace the bare failure")
+        .expect("second account started event");
+    assert!(first
+        .into_provider_events()
+        .iter()
+        .flat_map(ProviderEvent::canonical_facts)
+        .any(|event| matches!(event, GatewayEvent::Started(meta) if meta.response_id() == "response-1")));
+    block_on(session.commit_downstream(Some(200))).expect("commit second account");
+    while block_on(session.next_event())
+        .expect("successful stream")
+        .is_some()
+    {}
+
+    assert_eq!(provider.contexts.lock().expect("contexts lock").len(), 2);
+    let state = store.state.lock().expect("store lock");
+    assert_eq!(state.intermediate_failures, 1);
+    assert_eq!(state.finalizations[0].outcome, ExecutionOutcome::Succeeded);
+}
+
+#[test]
+fn exhausted_atomic_response_failed_should_deliver_only_the_last_failure_once() {
+    let operation = generate_operation();
+    let route_plan = plan(&operation);
+    let scripts = (1..=route_plan.max_attempts().get())
+        .map(|attempt| {
+            let response_id = format!("response-attempt-{attempt}-failed");
+            let marker = if attempt == route_plan.max_attempts().get() {
+                "last upstream failure".to_owned()
+            } else {
+                format!("discarded failure {attempt}")
+            };
+            Script::ObservedStream {
+                account_id: "acct_failed",
+                items: vec![Err(atomic_response_failed(&response_id, &marker))],
+            }
+        })
+        .collect::<Vec<_>>();
+    let expected_attempts = route_plan.max_attempts().get();
+    let expected_last_response = format!("response-attempt-{expected_attempts}-failed");
+    let (coordinator, store, provider) = coordinator(scripts);
+    let mut session = block_on(coordinator.start(
+        model_request(&operation, SystemTime::now() + Duration::from_secs(30)),
+        operation,
+        route_plan,
+        None,
+        None,
+        CancellationToken::new(),
+    ))
+    .expect("start execution");
+
+    let terminal = block_on(session.next_event())
+        .expect("last failure remains deliverable")
+        .expect("terminal failure batch");
+    assert_eq!(
+        terminal.commit_requirement(),
+        CommitRequirement::CommitBeforeDelivery
+    );
+    let events = terminal.into_provider_events();
+    let failures = events
+        .iter()
+        .filter_map(ProviderEvent::wire_event)
+        .filter(|wire| wire.event_type() == Some("response.failed"))
+        .collect::<Vec<_>>();
+    assert_eq!(failures.len(), 1);
+    assert_eq!(
+        failures[0]
+            .data()
+            .pointer("/response/id")
+            .and_then(Value::as_str),
+        Some(expected_last_response.as_str())
+    );
+    assert!(
+        failures[0]
+            .data()
+            .to_string()
+            .contains("last upstream failure")
+    );
+    block_on(session.commit_downstream(Some(200))).expect("commit original terminal failure");
+    let error = block_on(session.next_event()).expect_err("typed failure finalizes the request");
+    assert!(matches!(error, EngineError::Provider(_)));
+    assert!(session.is_finalized());
+
+    assert_eq!(
+        provider.contexts.lock().expect("contexts lock").len(),
+        usize::try_from(expected_attempts).expect("attempt count fits usize")
+    );
+    let state = store.state.lock().expect("store lock");
+    assert_eq!(
+        state.intermediate_failures,
+        usize::try_from(expected_attempts - 1).expect("attempt count fits usize")
+    );
+    assert_eq!(state.commits, 1);
+    assert_eq!(state.finalizations[0].outcome, ExecutionOutcome::Incomplete);
+    assert!(state.finalizations[0].committed);
+}
+
+#[test]
+fn empty_selection_after_atomic_failure_should_deliver_the_last_upstream_wire() {
+    let operation = generate_operation();
+    let route_plan = plan(&operation);
+    let (coordinator, store, provider) = coordinator(vec![
+        Script::ObservedStream {
+            account_id: "acct_only",
+            items: vec![Err(bare_atomic_response_failed("response-only-failed"))],
+        },
+        Script::Error(ProviderError::new(
+            ProviderErrorKind::NoEligibleAccount,
+            UpstreamSendState::NotSent,
+        )),
+    ]);
+    let mut session = block_on(coordinator.start(
+        model_request(&operation, SystemTime::now() + Duration::from_secs(30)),
+        operation,
+        route_plan,
+        None,
+        None,
+        CancellationToken::new(),
+    ))
+    .expect("start execution");
+
+    let terminal = block_on(session.next_event())
+        .expect("last account failure remains deliverable")
+        .expect("terminal failure batch");
+    let events = terminal.into_provider_events();
+    let failure = events
+        .iter()
+        .filter_map(ProviderEvent::wire_event)
+        .find(|wire| wire.event_type() == Some("response.failed"))
+        .expect("last upstream response.failed");
+    assert_eq!(
+        failure
+            .data()
+            .pointer("/response/id")
+            .and_then(Value::as_str),
+        Some("response-only-failed")
+    );
+    block_on(session.commit_downstream(Some(200))).expect("commit terminal failure");
+    assert!(matches!(
+        block_on(session.next_event()),
+        Err(EngineError::Provider(_))
+    ));
+
+    assert_eq!(provider.contexts.lock().expect("contexts lock").len(), 2);
+    let state = store.state.lock().expect("store lock");
+    assert_eq!(state.attempts.len(), 1);
+    assert_eq!(state.intermediate_failures, 1);
+    assert_eq!(state.finalizations[0].outcome, ExecutionOutcome::Incomplete);
+    assert_eq!(
+        state.finalizations[0].client_response_id.as_deref(),
+        Some("response-only-failed")
+    );
+    assert_eq!(
+        state.finalizations[0].upstream_response_id.as_deref(),
+        Some("response-only-failed")
+    );
+}
+
+#[test]
+fn atomic_response_failed_keeps_collect_uncommitted_error_semantics() {
+    let operation = generate_operation();
+    let route_plan = plan(&operation);
+    let (coordinator, store, provider) = coordinator(vec![Script::ObservedStream {
+        account_id: "acct_required",
+        items: vec![Err(atomic_response_failed(
+            "response-required-failed",
+            "required account failure",
+        ))],
+    }]);
+    let required = ProviderAccountId::new("acct_required").expect("required account");
+    let mut session = block_on(coordinator.start(
+        model_request(&operation, SystemTime::now() + Duration::from_secs(30)),
+        operation,
+        route_plan,
+        Some(required),
+        None,
+        CancellationToken::new(),
+    ))
+    .expect("start execution");
+
+    let error = block_on(session.collect_uncommitted())
+        .expect_err("non-streaming collection must keep returning the typed failure");
+    assert!(matches!(error, EngineError::Provider(_)));
+    assert!(session.is_finalized());
+    assert_eq!(provider.contexts.lock().expect("contexts lock").len(), 1);
+    let state = store.state.lock().expect("store lock");
+    assert_eq!(state.commits, 0);
+    assert_eq!(state.intermediate_failures, 0);
+    assert_eq!(state.finalizations[0].outcome, ExecutionOutcome::Failed);
+}
+
 #[test]
 fn native_continuation_explicit_429_is_not_retried() {
     let continuation = NativeContinuationPin::new(

backend/crates/gateway-core/tests/error.rs

@@ -2,6 +2,8 @@ use gateway_core::engine::UpstreamSendState;
 use gateway_core::error::{
     ClientVisibleUpstreamError, GatewayError, ProviderError, ProviderErrorKind, SafeUpstreamValue,
 };
+use gateway_core::event::{ProtocolWireEvent, ProviderEvent};
+use serde_json::json;
 
 #[test]
 fn provider_error_debug_should_not_expose_sensitive_context() {
@@ -37,6 +39,26 @@ fn provider_error_replay_proof_should_be_explicit() {
     assert!(error.replay_is_safe());
 }
 
+#[test]
+fn provider_error_atomic_client_events_should_be_take_only_and_debug_redacted() {
+    let marker = "atomic-wire-must-not-enter-debug";
+    let event = ProviderEvent::wire(
+        ProtocolWireEvent::json(
+            "openai",
+            Some("response.failed".to_owned()),
+            json!({"type": "response.failed", "message": marker}),
+        )
+        .expect("atomic wire"),
+    );
+    let mut error = ProviderError::new(ProviderErrorKind::RateLimited, UpstreamSendState::Sent)
+        .with_atomic_client_events(vec![event]);
+
+    assert!(error.has_atomic_client_events());
+    assert!(!format!("{error:?}").contains(marker));
+    assert_eq!(error.take_atomic_client_events().len(), 1);
+    assert!(!error.has_atomic_client_events());
+}
+
 #[test]
 fn gateway_error_should_keep_client_visible_upstream_fields_out_of_safe_diagnostics() {
     let message = "Your Codex quota is exhausted";

backend/crates/gateway-store/src/redis/credential_cooldown.rs

@@ -13,7 +13,8 @@ use gateway_core::{
     },
     error::StoreError as CoreStoreError,
     provider_ports::{
-        ProviderCooldown, ProviderCooldownPort, ProviderStoreError, ProviderStoreErrorKind,
+        ProviderCooldown, ProviderCooldownPort, ProviderCooldownScope, ProviderScopedCooldown,
+        ProviderStoreError, ProviderStoreErrorKind,
     },
     routing::ProviderKind,
 };
@@ -104,24 +105,35 @@ impl RedisCredentialCooldownRepository {
         let fingerprint = resource_fingerprint("credential cooldown", provider_account_id)?;
         Ok(format!("{}:account:{fingerprint}:cooldown", self.namespace))
     }
-}
 
-#[async_trait]
-impl CredentialCooldownRepository for RedisCredentialCooldownRepository {
-    async fn cache_credential_cooldown(&self, cooldown: &CredentialCooldown) -> StoreResult<bool> {
-        require_nonempty(
-            "credential cooldown",
-            "provider_account_id",
-            &cooldown.provider_account_id,
-        )?;
-        let until_ms = cooldown.cooldown_until.timestamp_millis();
+    fn scoped_key(
+        &self,
+        provider_account_id: &str,
+        scope: &ProviderCooldownScope,
+    ) -> StoreResult<String> {
+        let account_fingerprint = resource_fingerprint("credential cooldown", provider_account_id)?;
+        let scope_fingerprint = resource_fingerprint("credential cooldown scope", scope.value())?;
+        Ok(format!(
+            "{}:account:{account_fingerprint}:cooldown:{}:{scope_fingerprint}",
+            self.namespace,
+            scope.kind(),
+        ))
+    }
+
+    async fn cache_at_key(
+        &self,
+        key: String,
+        credential_revision: Revision,
+        cooldown_until: DateTime<Utc>,
+    ) -> StoreResult<bool> {
+        let until_ms = cooldown_until.timestamp_millis();
         if until_ms <= 0 {
             return Err(invalid("cooldown expiry must be positive"));
         }
         let mut connection = self.connection.clone();
         let written = Script::new(WRITE_SCRIPT)
-            .key(self.key(&cooldown.provider_account_id)?)
-            .arg(cooldown.credential_revision.get())
+            .key(key)
+            .arg(credential_revision.get())
             .arg(until_ms)
             .invoke_async::<i64>(&mut connection)
             .await
@@ -129,18 +141,10 @@ impl CredentialCooldownRepository for RedisCredentialCooldownRepository {
         Ok(written == 1)
     }
 
-    async fn read_credential_cooldown(
-        &self,
-        provider_account_id: &str,
-    ) -> StoreResult<Option<CredentialCooldown>> {
-        require_nonempty(
-            "credential cooldown",
-            "provider_account_id",
-            provider_account_id,
-        )?;
+    async fn read_at_key(&self, key: String) -> StoreResult<Option<(Revision, DateTime<Utc>)>> {
         let mut connection = self.connection.clone();
         let (present, revision, until_ms): (i64, String, String) = Script::new(READ_SCRIPT)
-            .key(self.key(provider_account_id)?)
+            .key(key)
             .invoke_async(&mut connection)
             .await
             .map_err(|_| redis_unavailable("read credential cooldown"))?;
@@ -155,11 +159,59 @@ impl CredentialCooldownRepository for RedisCredentialCooldownRepository {
             .map_err(|_| invalid("cached cooldown expiry is invalid"))?;
         let cooldown_until = DateTime::from_timestamp_millis(until_ms)
             .ok_or_else(|| invalid("cached cooldown expiry is invalid"))?;
-        Ok(Some(CredentialCooldown {
-            provider_account_id: provider_account_id.to_owned(),
-            credential_revision: Revision::new(revision)?,
-            cooldown_until,
-        }))
+        Ok(Some((Revision::new(revision)?, cooldown_until)))
+    }
+
+    async fn invalidate_at_key(
+        &self,
+        key: String,
+        through_revision: Revision,
+    ) -> StoreResult<bool> {
+        let mut connection = self.connection.clone();
+        let removed = Script::new(INVALIDATE_SCRIPT)
+            .key(key)
+            .arg(through_revision.get())
+            .invoke_async::<i64>(&mut connection)
+            .await
+            .map_err(|_| redis_unavailable("invalidate credential cooldown"))?;
+        Ok(removed == 1)
+    }
+}
+
+#[async_trait]
+impl CredentialCooldownRepository for RedisCredentialCooldownRepository {
+    async fn cache_credential_cooldown(&self, cooldown: &CredentialCooldown) -> StoreResult<bool> {
+        require_nonempty(
+            "credential cooldown",
+            "provider_account_id",
+            &cooldown.provider_account_id,
+        )?;
+        self.cache_at_key(
+            self.key(&cooldown.provider_account_id)?,
+            cooldown.credential_revision,
+            cooldown.cooldown_until,
+        )
+        .await
+    }
+
+    async fn read_credential_cooldown(
+        &self,
+        provider_account_id: &str,
+    ) -> StoreResult<Option<CredentialCooldown>> {
+        require_nonempty(
+            "credential cooldown",
+            "provider_account_id",
+            provider_account_id,
+        )?;
+        self.read_at_key(self.key(provider_account_id)?)
+            .await
+            .map(|value| {
+                value.map(|(credential_revision, cooldown_until)| CredentialCooldown {
+                    provider_account_id: provider_account_id.to_owned(),
+                    credential_revision,
+                    cooldown_until,
+                })
+            })
     }
 
     async fn invalidate_credential_cooldown(
@@ -172,14 +224,8 @@ impl CredentialCooldownRepository for RedisCredentialCooldownRepository {
             "provider_account_id",
             provider_account_id,
         )?;
-        let mut connection = self.connection.clone();
-        let removed = Script::new(INVALIDATE_SCRIPT)
-            .key(self.key(provider_account_id)?)
-            .arg(through_revision.get())
-            .invoke_async::<i64>(&mut connection)
+        self.invalidate_at_key(self.key(provider_account_id)?, through_revision)
             .await
-            .map_err(|_| redis_unavailable("invalidate credential cooldown"))?;
-        Ok(removed == 1)
     }
 }
 
@@ -241,6 +287,69 @@ impl ProviderCooldownPort for RedisCredentialCooldownRepository {
             .map_err(|_| provider_unavailable("clear credential cooldown"))
         })
     }
+
+    fn put_scoped_if_later(
+        &self,
+        cooldown: ProviderScopedCooldown,
+    ) -> futures::future::BoxFuture<'_, Result<bool, ProviderStoreError>> {
+        Box::pin(async move {
+            let revision = Revision::new(cooldown.credential_revision().get())
+                .map_err(|_| provider_invalid("encode scoped credential cooldown"))?;
+            self.cache_at_key(
+                self.scoped_key(cooldown.account_id().as_str(), cooldown.scope())
+                    .map_err(|_| provider_invalid("encode scoped credential cooldown"))?,
+                revision,
+                cooldown.until().into(),
+            )
+            .await
+            .map_err(|_| provider_unavailable("cache scoped credential cooldown"))
+        })
+    }
+
+    fn read_scoped<'a>(
+        &'a self,
+        account_id: &'a ProviderAccountId,
+        scope: &'a ProviderCooldownScope,
+    ) -> futures::future::BoxFuture<'a, Result<Option<ProviderScopedCooldown>, ProviderStoreError>>
+    {
+        Box::pin(async move {
+            self.read_at_key(
+                self.scoped_key(account_id.as_str(), scope)
+                    .map_err(|_| provider_invalid("encode scoped credential cooldown"))?,
+            )
+            .await
+            .map_err(|_| provider_unavailable("read scoped credential cooldown"))?
+            .map(|(revision, until)| {
+                Ok(ProviderScopedCooldown::new(
+                    account_id.clone(),
+                    CredentialRevision::new(revision.get())
+                        .map_err(|_| provider_invalid("decode scoped credential cooldown"))?,
+                    scope.clone(),
+                    until.into(),
+                ))
+            })
+            .transpose()
+        })
+    }
+
+    fn clear_scoped<'a>(
+        &'a self,
+        account_id: &'a ProviderAccountId,
+        scope: &'a ProviderCooldownScope,
+        through_revision: CredentialRevision,
+    ) -> futures::future::BoxFuture<'a, Result<bool, ProviderStoreError>> {
+        Box::pin(async move {
+            let revision = Revision::new(through_revision.get())
+                .map_err(|_| provider_invalid("encode scoped cooldown revision"))?;
+            self.invalidate_at_key(
+                self.scoped_key(account_id.as_str(), scope)
+                    .map_err(|_| provider_invalid("encode scoped credential cooldown"))?,
+                revision,
+            )
+            .await
+            .map_err(|_| provider_unavailable("clear scoped credential cooldown"))
+        })
+    }
 }
 
 /// 账号状态写入到可丢失 cooldown cache 的桥接 adapter。

backend/crates/gateway-store/tests/redis/credential_cooldown.rs

@@ -13,6 +13,7 @@ use gateway_core::{
         QuotaObservation, QuotaWriteOutcome,
     },
     error::{StoreError as CoreStoreError, StoreErrorKind},
+    provider_ports::{ProviderCooldownPort, ProviderCooldownScope, ProviderScopedCooldown},
     routing::ProviderKind,
 };
 use gateway_store::{
@@ -530,6 +531,109 @@ async fn credential_cooldown_read_removes_expired_grace_key() {
     assert!(namespace_keys(&mut connection, &namespace).await.is_empty());
 }
 
+#[tokio::test]
+async fn scoped_cooldown_isolated_by_model_and_revision_fenced() {
+    let Some((repository, mut connection, namespace)) = repository().await else {
+        return;
+    };
+    let account_id = ProviderAccountId::new("acct_scoped_cooldown").expect("valid account ID");
+    let revision = CredentialRevision::new(2).expect("positive revision");
+    let model_a = ProviderCooldownScope::upstream_model(
+        gateway_core::routing::UpstreamModelId::new("grok-4.5").expect("model"),
+    );
+    let model_b = ProviderCooldownScope::upstream_model(
+        gateway_core::routing::UpstreamModelId::new("grok-4.6").expect("model"),
+    );
+    let until_a: SystemTime = millisecond_precision(Utc::now() + Duration::seconds(30)).into();
+    let until_b: SystemTime = millisecond_precision(Utc::now() + Duration::seconds(60)).into();
+
+    assert!(
+        repository
+            .put_scoped_if_later(ProviderScopedCooldown::new(
+                account_id.clone(),
+                revision,
+                model_a.clone(),
+                until_a,
+            ))
+            .await
+            .expect("cache model A cooldown")
+    );
+    assert!(
+        !repository
+            .put_scoped_if_later(ProviderScopedCooldown::new(
+                account_id.clone(),
+                CredentialRevision::new(1).expect("stale revision"),
+                model_a.clone(),
+                until_b,
+            ))
+            .await
+            .expect("reject stale model A cooldown")
+    );
+    assert!(
+        repository
+            .put_scoped_if_later(ProviderScopedCooldown::new(
+                account_id.clone(),
+                revision,
+                model_b.clone(),
+                until_b,
+            ))
+            .await
+            .expect("cache model B cooldown")
+    );
+    assert_eq!(
+        repository
+            .read_scoped(&account_id, &model_a)
+            .await
+            .expect("read model A cooldown")
+            .expect("model A cooldown")
+            .until(),
+        until_a
+    );
+    assert_eq!(
+        repository
+            .read_scoped(&account_id, &model_b)
+            .await
+            .expect("read model B cooldown")
+            .expect("model B cooldown")
+            .until(),
+        until_b
+    );
+    assert!(
+        !repository
+            .clear_scoped(
+                &account_id,
+                &model_a,
+                CredentialRevision::new(1).expect("stale revision"),
+            )
+            .await
+            .expect("fence stale model A invalidation")
+    );
+    assert!(
+        repository
+            .clear_scoped(&account_id, &model_a, revision)
+            .await
+            .expect("clear model A cooldown")
+    );
+    assert!(
+        repository
+            .read_scoped(&account_id, &model_a)
+            .await
+            .expect("read cleared model A cooldown")
+            .is_none()
+    );
+    assert!(
+        repository
+            .read_scoped(&account_id, &model_b)
+            .await
+            .expect("read retained model B cooldown")
+            .is_some()
+    );
+    let keys = namespace_keys(&mut connection, &namespace).await;
+    assert_eq!(keys.len(), 1);
+    assert!(!keys[0].contains(account_id.as_str()));
+    assert!(!keys[0].contains(model_b.value()));
+}
+
 fn cooldown(account_id: &str, revision: u64, seconds: i64) -> CredentialCooldown {
     CredentialCooldown {
         provider_account_id: account_id.to_owned(),

backend/crates/providers/openai/src/credential/quota.rs

@@ -178,6 +178,14 @@ impl CodexQuotaFact {
     }
 }
 
+fn quota_fact_is_active_exhaustion(fact: CodexQuotaFact, now: SystemTime) -> bool {
+    fact.exhausted()
+        && fact
+            .resets_at()
+            .map(SystemTime::from)
+            .is_none_or(|reset_at| reset_at > now)
+}
+
 #[derive(Debug, Clone, Copy, Default, PartialEq, Eq)]
 pub struct CodexQuotaSyncSummary {
     pub updated: u64,
@@ -502,10 +510,20 @@ fn quota_refresh_candidate(
         return None;
     }
     let observation = observations.and_then(|observations| observations.get(account.id()));
+    let snapshot = quota_snapshot_for_account(&account, observation);
     let kind = match account.availability() {
         AccountAvailability::QuotaExhausted => CodexQuotaRefreshKind::Exhausted {
-            reset_at: quota_reset_at_from_observation(&account, observation),
+            reset_at: quota_reset_at_from_snapshot(snapshot.as_ref()),
         },
+        AccountAvailability::Cooldown
+            if snapshot
+                .as_ref()
+                .is_some_and(|snapshot| snapshot.fact().exhausted()) =>
+        {
+            CodexQuotaRefreshKind::Exhausted {
+                reset_at: quota_reset_at_from_snapshot(snapshot.as_ref()),
+            }
+        }
         AccountAvailability::Ready
             if observations.is_some() && quota_observation_is_stale(&account, observation, now) =>
         {
@@ -516,17 +534,32 @@ fn quota_refresh_candidate(
     Some(CodexQuotaRefreshCandidate { account, kind })
 }
 
-fn quota_reset_at_from_observation(
+fn quota_snapshot_for_account(
     account: &ProviderAccount,
     observation: Option<&QuotaObservation>,
-) -> Option<SystemTime> {
+) -> Option<CodexAccountQuotaSnapshot> {
     observation
         .filter(|observation| observation.expected_revision == account.revision())
         .and_then(quota_snapshot_from_observation)
+}
+
+fn quota_reset_at_from_snapshot(
+    snapshot: Option<&CodexAccountQuotaSnapshot>,
+) -> Option<SystemTime> {
+    snapshot
         .and_then(|snapshot| snapshot.fact().resets_at())
         .map(SystemTime::from)
 }
 
+fn quota_observation_confirms_active_exhaustion(
+    account: &ProviderAccount,
+    observation: Option<&QuotaObservation>,
+    now: SystemTime,
+) -> bool {
+    quota_snapshot_for_account(account, observation)
+        .is_some_and(|snapshot| quota_fact_is_active_exhaustion(snapshot.fact(), now))
+}
+
 fn quota_observation_is_stale(
     account: &ProviderAccount,
     observation: Option<&QuotaObservation>,
@@ -674,7 +707,7 @@ impl CodexCredentialQuotaService {
     }
 
     pub async fn synchronize(&self) -> Result<CodexQuotaSyncSummary, CodexCredentialQuotaError> {
-        let accounts = self.repository.list_for_provider().await?;
+        let mut accounts = self.repository.list_for_provider().await?;
         let mut summary = CodexQuotaSyncSummary::default();
         let now = SystemTime::now();
         let account_ids = accounts
@@ -694,6 +727,8 @@ impl CodexCredentialQuotaService {
                 None
             }
         };
+        self.reconcile_persisted_exhaustion(&mut accounts, observations.as_ref(), now)
+            .await;
         let accounts =
             self.scheduling
                 .reserve_periodic_refreshes(accounts, observations.as_ref(), now);
@@ -730,6 +765,9 @@ impl CodexCredentialQuotaService {
                 }
                 Err(CodexQuotaFetchError::Upstream { account, error }) => {
                     let (availability, reason, cooldown) = classify_error(&error, observed_at);
+                    let preserve_exhaustion = account.availability()
+                        == AccountAvailability::QuotaExhausted
+                        && availability == Some(AccountAvailability::Cooldown);
                     match availability {
                         Some(AccountAvailability::Invalid | AccountAvailability::Banned) => {
                             summary.invalid += 1;
@@ -741,6 +779,9 @@ impl CodexCredentialQuotaService {
                             continue;
                         }
                     }
+                    if preserve_exhaustion {
+                        continue;
+                    }
                     let _ = self
                         .repository
                         .apply_state(
@@ -757,6 +798,63 @@ impl CodexCredentialQuotaService {
         Ok(summary)
     }
 
+    async fn reconcile_persisted_exhaustion(
+        &self,
+        accounts: &mut [ProviderAccount],
+        observations: Option<&BTreeMap<ProviderAccountId, QuotaObservation>>,
+        observed_at: SystemTime,
+    ) {
+        let Some(observations) = observations else {
+            return;
+        };
+        for account in accounts {
+            if account.availability() != AccountAvailability::Cooldown
+                || !quota_observation_confirms_active_exhaustion(
+                    account,
+                    observations.get(account.id()),
+                    observed_at,
+                )
+            {
+                continue;
+            }
+            let Ok(Some(current)) = self.store.get_account(account.id()).await else {
+                continue;
+            };
+            if current.revision() != account.revision()
+                || current.availability() != AccountAvailability::Cooldown
+            {
+                continue;
+            }
+            if !matches!(
+                self.has_active_exhaustion(&current, observed_at).await,
+                Ok(true)
+            ) {
+                continue;
+            }
+            if let Err(error) = self
+                .repository
+                .apply_state(
+                    &current,
+                    AccountAvailability::QuotaExhausted,
+                    Some("quota_exhausted".to_owned()),
+                    None,
+                    observed_at,
+                )
+                .await
+            {
+                tracing::warn!(
+                    account_id = %account.id(),
+                    error = %error,
+                    "OpenAI persisted quota exhaustion reconciliation failed"
+                );
+                continue;
+            }
+            let enabled = current.enabled();
+            *account =
+                current.with_runtime_state(enabled, AccountAvailability::QuotaExhausted, None);
+        }
+    }
+
     /// 解析并 revision-fenced 落库单账号的 Provider quota JSON。
     async fn apply_fetched_quota(
         &self,
@@ -894,6 +992,30 @@ impl CodexCredentialQuotaService {
             .await?
             .filter(|account| account.provider().as_str() == "openai")
             .ok_or(CodexCredentialQuotaError::NotFound)?;
+        self.read_snapshot_for(&account).await
+    }
+
+    pub(super) async fn has_active_exhaustion(
+        &self,
+        account: &ProviderAccount,
+        now: SystemTime,
+    ) -> Result<bool, CodexCredentialQuotaError> {
+        match self.read_snapshot_for(account).await {
+            Ok(snapshot) => Ok(snapshot
+                .is_some_and(|snapshot| quota_fact_is_active_exhaustion(snapshot.fact(), now))),
+            Err(CodexCredentialQuotaError::RevisionConflict) => Ok(false),
+            Err(error) => Err(error),
+        }
+    }
+
+    async fn read_snapshot_for(
+        &self,
+        account: &ProviderAccount,
+    ) -> Result<Option<CodexAccountQuotaSnapshot>, CodexCredentialQuotaError> {
+        if account.provider().as_str() != "openai" {
+            return Err(CodexCredentialQuotaError::NotFound);
+        }
+        let account_id = account.id();
         let Some(observation) = self
             .store
             .get_quotas(std::slice::from_ref(account_id))

backend/crates/providers/openai/src/credential/selector.rs

@@ -552,13 +552,9 @@ impl CodexCredentialSelector {
                 CookieRecovery::None,
             ),
             CodexAccountFailure::RateLimited { retry_after } => {
-                let until = now.checked_add(retry_after.unwrap_or(DEFAULT_RATE_LIMIT_COOLDOWN));
-                (
-                    AccountAvailability::Cooldown,
-                    Some("rate_limited".to_owned()),
-                    until,
-                    CookieRecovery::None,
-                )
+                return self
+                    .record_rate_limit_failure(account, retry_after, now)
+                    .await;
             }
             CodexAccountFailure::CloudflareChallenge { retry_after } => {
                 let delay = self.cloudflare_challenge_delay(account.id(), now, retry_after);
@@ -596,6 +592,49 @@ impl CodexCredentialSelector {
         Ok(())
     }
 
+    async fn record_rate_limit_failure(
+        &self,
+        account: &ProviderAccount,
+        retry_after: Option<Duration>,
+        observed_at: SystemTime,
+    ) -> Result<(), CredentialSelectionError> {
+        let current = self.current_account(account.id()).await?;
+        if current.revision() != account.revision() {
+            return Err(CredentialRepositoryError::RevisionConflict.into());
+        }
+        if matches!(
+            current.availability(),
+            AccountAvailability::Invalid
+                | AccountAvailability::Expired
+                | AccountAvailability::Banned
+                | AccountAvailability::QuotaExhausted
+        ) {
+            return Ok(());
+        }
+        let quota_exhausted = self
+            .quota
+            .has_active_exhaustion(&current, observed_at)
+            .await
+            .map_err(|_| CredentialSelectionError::Store)?;
+        let (availability, reason, cooldown_until) = if quota_exhausted {
+            (
+                AccountAvailability::QuotaExhausted,
+                Some("quota_exhausted".to_owned()),
+                None,
+            )
+        } else {
+            (
+                AccountAvailability::Cooldown,
+                Some("rate_limited".to_owned()),
+                observed_at.checked_add(retry_after.unwrap_or(DEFAULT_RATE_LIMIT_COOLDOWN)),
+            )
+        };
+        self.repository
+            .apply_state(&current, availability, reason, cooldown_until, observed_at)
+            .await?;
+        Ok(())
+    }
+
     pub async fn record_success(
         &self,
         account: &ProviderAccount,

backend/crates/providers/openai/src/provider.rs

@@ -79,6 +79,8 @@ const PROVIDER_NAME: &str = "openai";
 const HTTP_SSE_TRANSPORT: &str = "http_sse";
 const WEBSOCKET_TRANSPORT: &str = "websocket";
 const MAX_COOKIE_HEADER_BYTES: usize = 16 * 1024;
+/// 与 2.x 预取器一致：提交边界前最多保留 64 KiB 原始上游 chunk。
+const MAX_STREAM_PREFETCH_BYTES: usize = 64 * 1024;
 pub const OFFICIAL_CODEX_BASE_PATH: &str = "/backend-api";
 pub const OFFICIAL_CODEX_BASE_URL: &str = "https://chatgpt.com/backend-api";
 
@@ -684,6 +686,8 @@ fn cold_response_stream(response: ColdResponse) -> EventStream {
             })?;
         let request_id = context.request_id().as_str().to_owned();
         let cancellation = context.cancellation().clone();
+        let stream_commit_policy =
+            StreamCommitPolicy::for_continuation(context.continuation_attempt());
         let mut response = create_response_attempt(
             &client,
             &request,
@@ -824,6 +828,7 @@ fn cold_response_stream(response: ColdResponse) -> EventStream {
         // （transport/protocol/websocket.rs），push_frames 抽出的 data 即上游原文。
         let mut decoder =
             CodexCanonicalDecoder::new(upstream_model.as_str()).with_raw_sse_passthrough();
+        let mut pre_commit_events = PreCommitClientEvents::new(stream_commit_policy);
         loop {
             let Some(stream_deadline) = remaining(context.deadline()) else {
                 Err(provider_error(ProviderErrorKind::Timeout, UpstreamSendState::Sent))?;
@@ -882,30 +887,41 @@ fn cold_response_stream(response: ColdResponse) -> EventStream {
             }
             let first_event_changed =
                 observation_state.observe_stream_chunk(&chunk, context.timing_started_at());
-            let (mut events, terminal_failure) = match decoder.push(&chunk) {
+            let chunk_len = chunk.len();
+            let (mut events, canonical_failure) = match decoder.push(&chunk) {
                 CodexCanonicalOutcome::Events(events) => (events, None),
                 CodexCanonicalOutcome::Failed(failure) => {
                     let (events, error, semantic_output_seen) = failure.into_parts();
-                    let failure = map_canonical_error(
+                    (events, Some((error, semantic_output_seen)))
+                }
+            };
+            pre_commit_events.observe_chunk(chunk_len);
+            let terminal_failure = canonical_failure.map(|(error, semantic_output_seen)| {
+                let atomic_upstream_failure = matches!(&error, CodexCanonicalError::Upstream(_));
+                (
+                    map_canonical_error(
                         error,
                         &failure_diagnostics,
                         &failure_set_cookie_headers,
                         &failure_rate_limit_headers,
-                        ReplayBoundary::from_semantic_output(semantic_output_seen),
-                    );
-                    (events, Some(failure))
-                }
-            };
+                        ReplayBoundary::from_semantic_output(
+                            semantic_output_seen || pre_commit_events.is_committed(),
+                        ),
+                    ),
+                    atomic_upstream_failure,
+                )
+            });
+            let timing_signals = decoder.take_timing_signals();
             let timing_changed = first_event_changed
                 || observation_state
-                    .observe_timing_signals(decoder.take_timing_signals(), context.timing_started_at());
+                    .observe_timing_signals(timing_signals, context.timing_started_at());
             let completed = events
                 .iter()
                 .flat_map(ProviderEvent::canonical_facts)
                 .any(|event| matches!(event, GatewayEvent::Completed(_)));
             let terminal_changed = completed
                 && observation_state.mark_completed(terminal_response_is_incomplete(&events));
-            if let Some(failure) = terminal_failure.as_ref() {
+            if let Some((failure, _)) = terminal_failure.as_ref() {
                 apply_failure(&failure_context, &active_account, failure)
                 .await;
             }
@@ -933,33 +949,54 @@ fn cold_response_stream(response: ColdResponse) -> EventStream {
             if rate_limits_changed || timing_changed || terminal_changed {
                 yield ProviderEvent::observation(observation_state.observation()?);
             }
-            for event in events {
-                yield event;
-            }
-            if let Some(failure) = terminal_failure {
+            if let Some((mut failure, atomic_upstream_failure)) = terminal_failure {
+                let failure_after_commit =
+                    timing_signals.semantic_output || pre_commit_events.is_committed();
+                if failure_after_commit {
+                    for event in pre_commit_events.commit(events) {
+                        yield event;
+                    }
+                } else if atomic_upstream_failure {
+                    failure.error = failure
+                        .error
+                        .with_atomic_client_events(pre_commit_events.take_for_failure(events));
+                }
                 Err(failure.error)?;
                 return;
             }
+            let events = pre_commit_events.stage(events, timing_signals, completed)?;
+            for event in events {
+                yield event;
+            }
             if completed {
                 return;
             }
         }
-        let (mut events, terminal_failure) = match decoder.finish() {
+        let (mut events, canonical_failure) = match decoder.finish() {
             CodexCanonicalOutcome::Events(events) => (events, None),
             CodexCanonicalOutcome::Failed(failure) => {
                 let (events, error, semantic_output_seen) = failure.into_parts();
-                let failure = map_canonical_error(
+                (events, Some((error, semantic_output_seen)))
+            }
+        };
+        let terminal_failure = canonical_failure.map(|(error, semantic_output_seen)| {
+            let atomic_upstream_failure = matches!(&error, CodexCanonicalError::Upstream(_));
+            (
+                map_canonical_error(
                     error,
                     &failure_diagnostics,
                     &failure_set_cookie_headers,
                     &failure_rate_limit_headers,
-                    ReplayBoundary::from_semantic_output(semantic_output_seen),
-                );
-                (events, Some(failure))
-            }
-        };
+                    ReplayBoundary::from_semantic_output(
+                        semantic_output_seen || pre_commit_events.is_committed(),
+                    ),
+                ),
+                atomic_upstream_failure,
+            )
+        });
+        let timing_signals = decoder.take_timing_signals();
         let timing_changed = observation_state
-            .observe_timing_signals(decoder.take_timing_signals(), context.timing_started_at());
+            .observe_timing_signals(timing_signals, context.timing_started_at());
         let updates = take_rate_limit_updates(rate_limit_updates.as_ref()).await;
         let rate_limits_changed = if updates.is_empty() {
             false
@@ -969,7 +1006,7 @@ fn cold_response_stream(response: ColdResponse) -> EventStream {
             }
             observation_state.merge_rate_limit_headers(&updates)
         };
-        if let Some(failure) = terminal_failure.as_ref() {
+        if let Some((failure, _)) = terminal_failure.as_ref() {
             apply_failure(&failure_context, &active_account, failure)
             .await;
         }
@@ -1008,13 +1045,25 @@ fn cold_response_stream(response: ColdResponse) -> EventStream {
         if timing_changed || rate_limits_changed || terminal_changed {
             yield ProviderEvent::observation(observation_state.observation()?);
         }
-        for event in events {
-            yield event;
-        }
-        if let Some(failure) = terminal_failure {
+        if let Some((mut failure, atomic_upstream_failure)) = terminal_failure {
+            let failure_after_commit =
+                timing_signals.semantic_output || pre_commit_events.is_committed();
+            if failure_after_commit {
+                for event in pre_commit_events.commit(events) {
+                    yield event;
+                }
+            } else if atomic_upstream_failure {
+                failure.error = failure
+                    .error
+                    .with_atomic_client_events(pre_commit_events.take_for_failure(events));
+            }
             Err(failure.error)?;
             return;
         }
+        let events = pre_commit_events.finish(events, timing_signals, completed)?;
+        for event in events {
+            yield event;
+        }
     })
 }
 
@@ -1735,6 +1784,128 @@ impl MappedProviderFailure {
     }
 }
 
+/// 与 2.x 一致的流提交策略：普通请求优先首事件，续写/重放优先可恢复性。
+#[derive(Clone, Copy)]
+enum StreamCommitPolicy {
+    FirstForwardableEvent,
+    UntilOutputOrTerminal,
+}
+
+impl StreamCommitPolicy {
+    const fn for_continuation(attempt: ContinuationAttempt) -> Self {
+        match attempt {
+            ContinuationAttempt::None => Self::FirstForwardableEvent,
+            ContinuationAttempt::Native
+            | ContinuationAttempt::ReplayOwner
+            | ContinuationAttempt::ReplayAny => Self::UntilOutputOrTerminal,
+        }
+    }
+
+    fn boundary_reached(
+        self,
+        events: &[ProviderEvent],
+        timing_signals: ResponseEventSignals,
+        completed: bool,
+    ) -> bool {
+        match self {
+            Self::FirstForwardableEvent => events.iter().any(ProviderEvent::has_client_event),
+            Self::UntilOutputOrTerminal => timing_signals.semantic_output || completed,
+        }
+    }
+}
+
+/// 提交边界前的上游事件预取。
+///
+/// 原始 chunk 计数而不是重编码后的 event 大小，保持 2.x 的 64 KiB 资源边界；一旦
+/// 提交，后续事件不再具备无痕重放资格。
+struct PreCommitClientEvents {
+    policy: StreamCommitPolicy,
+    pending: Vec<ProviderEvent>,
+    prefetched_bytes: usize,
+    committed: bool,
+}
+
+impl PreCommitClientEvents {
+    const fn new(policy: StreamCommitPolicy) -> Self {
+        Self {
+            policy,
+            pending: Vec::new(),
+            prefetched_bytes: 0,
+            committed: false,
+        }
+    }
+
+    fn observe_chunk(&mut self, bytes: usize) {
+        if !self.committed {
+            self.prefetched_bytes = self.prefetched_bytes.saturating_add(bytes);
+        }
+    }
+
+    fn stage(
+        &mut self,
+        incoming: Vec<ProviderEvent>,
+        timing_signals: ResponseEventSignals,
+        completed: bool,
+    ) -> Result<Vec<ProviderEvent>, ProviderError> {
+        if self.committed {
+            return Ok(incoming);
+        }
+        let boundary_reached = self
+            .policy
+            .boundary_reached(&incoming, timing_signals, completed);
+        self.pending.extend(incoming);
+        if boundary_reached {
+            return Ok(self.commit_pending());
+        }
+        if self.prefetched_bytes > MAX_STREAM_PREFETCH_BYTES {
+            return Err(prefetch_protocol_error());
+        }
+        Ok(Vec::new())
+    }
+
+    fn finish(
+        &mut self,
+        incoming: Vec<ProviderEvent>,
+        timing_signals: ResponseEventSignals,
+        completed: bool,
+    ) -> Result<Vec<ProviderEvent>, ProviderError> {
+        let events = self.stage(incoming, timing_signals, completed)?;
+        if self.committed {
+            Ok(events)
+        } else {
+            Err(prefetch_protocol_error())
+        }
+    }
+
+    fn commit(&mut self, incoming: Vec<ProviderEvent>) -> Vec<ProviderEvent> {
+        if self.committed {
+            return incoming;
+        }
+        self.pending.extend(incoming);
+        self.commit_pending()
+    }
+
+    fn take_for_failure(&mut self, incoming: Vec<ProviderEvent>) -> Vec<ProviderEvent> {
+        self.pending.extend(incoming);
+        self.prefetched_bytes = 0;
+        std::mem::take(&mut self.pending)
+    }
+
+    const fn is_committed(&self) -> bool {
+        self.committed
+    }
+
+    fn commit_pending(&mut self) -> Vec<ProviderEvent> {
+        self.committed = true;
+        self.prefetched_bytes = 0;
+        std::mem::take(&mut self.pending)
+    }
+}
+
+fn prefetch_protocol_error() -> ProviderError {
+    provider_error(ProviderErrorKind::Protocol, UpstreamSendState::Sent)
+}
+
 #[derive(Clone, Copy)]
 enum ReplayBoundary {
     BeforeSemanticOutput,

backend/crates/providers/openai/src/transport/canonical.rs

@@ -273,8 +273,20 @@ impl CodexCanonicalDecoder {
         self.merge_timing_signals(signals);
         if matches!(event_type, Some("response.failed" | "error")) {
             let failure = ResponsesSseFailure::from_event(event_type.unwrap_or_default(), &value);
+            let mut canonical = Vec::new();
+            if !self.started {
+                // 失败首帧可能是唯一携带 response_id 的上游事实。身份投影只做
+                // 最佳努力：缺少 response/id 时仍必须保留原 typed failure。
+                let _ = self.start(&value, &mut canonical);
+            }
             if let Some(wire) = Self::wire_for_event(event, value, raw_sse_frame) {
-                output.push(ProviderEvent::wire(wire));
+                output.push(if canonical.is_empty() {
+                    ProviderEvent::wire(wire)
+                } else {
+                    ProviderEvent::canonical_with_wire(canonical, wire)
+                });
+            } else {
+                output.extend(canonical.into_iter().map(ProviderEvent::canonical));
             }
             return Err(CodexCanonicalError::Upstream(failure));
         }

backend/crates/providers/openai/src/transport/diagnostics.rs

@@ -415,15 +415,6 @@ fn classify_upstream_failure(
     if status == Some(StatusCode::NOT_FOUND) && body.trim().is_empty() {
         return CodexFailureCategory::CloudflarePathBlocked;
     }
-    if [code.as_str(), error_type.as_str()]
-        .into_iter()
-        .any(is_rate_limit_signal)
-        || is_rate_limit_message(&message)
-        || is_rate_limit_message(&body)
-        || status == Some(StatusCode::TOO_MANY_REQUESTS)
-    {
-        return CodexFailureCategory::RateLimited;
-    }
     if [code.as_str(), error_type.as_str()]
         .into_iter()
         .any(is_quota_signal)
@@ -433,6 +424,15 @@ fn classify_upstream_failure(
     {
         return CodexFailureCategory::QuotaExhausted;
     }
+    if [code.as_str(), error_type.as_str()]
+        .into_iter()
+        .any(is_rate_limit_signal)
+        || is_rate_limit_message(&message)
+        || is_rate_limit_message(&body)
+        || status == Some(StatusCode::TOO_MANY_REQUESTS)
+    {
+        return CodexFailureCategory::RateLimited;
+    }
     if is_upstream_overload(&code) || is_upstream_overload(&message) {
         return CodexFailureCategory::Unavailable;
     }

backend/crates/providers/openai/tests/admin.rs

@@ -36,8 +36,9 @@ use gateway_core::provider_ports::{
     NewOAuthPendingFlow, OAuthPendingClaimOutcome, OAuthPendingConsumeOutcome,
     OAuthPendingFlowPort, OAuthPendingPutOutcome, OAuthPendingReleaseOutcome,
     ProviderCatalogCacheKey, ProviderCatalogCachePort, ProviderCooldown, ProviderCooldownPort,
-    ProviderCredentialState, ProviderCredentialStatePort, ProviderRefreshPolicy,
-    ProviderRuntimePolicyPort, ProviderStoreError, ProviderStorePorts,
+    ProviderCooldownScope, ProviderCredentialState, ProviderCredentialStatePort,
+    ProviderRefreshPolicy, ProviderRuntimePolicyPort, ProviderScopedCooldown, ProviderStoreError,
+    ProviderStorePorts,
 };
 use gateway_core::routing::{
     ConfigRevision, ModelCapabilities, ProviderKind, ProviderModel, PublicModelId, RoutingContext,
@@ -901,6 +902,30 @@ impl ProviderCooldownPort for TestCooldown {
     ) -> BoxFuture<'a, Result<bool, ProviderStoreError>> {
         Box::pin(async { Ok(false) })
     }
+
+    fn put_scoped_if_later(
+        &self,
+        _cooldown: ProviderScopedCooldown,
+    ) -> BoxFuture<'_, Result<bool, ProviderStoreError>> {
+        Box::pin(async { Ok(false) })
+    }
+
+    fn read_scoped<'a>(
+        &'a self,
+        _account_id: &'a ProviderAccountId,
+        _scope: &'a ProviderCooldownScope,
+    ) -> BoxFuture<'a, Result<Option<ProviderScopedCooldown>, ProviderStoreError>> {
+        Box::pin(async { Ok(None) })
+    }
+
+    fn clear_scoped<'a>(
+        &'a self,
+        _account_id: &'a ProviderAccountId,
+        _scope: &'a ProviderCooldownScope,
+        _through_revision: CredentialRevision,
+    ) -> BoxFuture<'a, Result<bool, ProviderStoreError>> {
+        Box::pin(async { Ok(false) })
+    }
 }
 
 struct TestRuntimePolicy;

backend/crates/providers/openai/tests/credential/contract.rs

@@ -7,7 +7,8 @@ use chrono::Utc;
 use futures::executor::block_on;
 use gateway_core::engine::credential::{
     AccountAttemptFeedback, AccountAvailability, AccountFeedbackStats, AccountSelectionPolicy,
-    ProviderAccountId, RotationStrategy,
+    AccountStateChange, OpaqueProviderData, ProviderAccountId, ProviderAccountStore as _,
+    QuotaObservation, RotationStrategy,
 };
 use gateway_core::engine::{
     AccountAttemptContext, AttemptContext, CancellationToken, ModelRequestId, RequestAttemptContext,
@@ -23,6 +24,7 @@ use provider_openai::credential::{
 };
 use provider_openai::transport::profile::{CodexWireProfile, CodexWireProfileState};
 use secrecy::ExposeSecret;
+use serde_json::json;
 use url::Url;
 
 use crate::support::{
@@ -637,6 +639,171 @@ fn rate_limited_failure_marks_account_cooldown_with_a_cooldown_deadline() {
     );
 }
 
+#[test]
+fn rate_limited_failure_does_not_downgrade_persisted_quota_exhaustion() {
+    let store = Arc::new(MemoryAccountStore::default());
+    create_account(&store, "acct_primary", "at-primary");
+    let account = store.account("acct_primary").expect("account");
+    let reset_at = Utc::now().timestamp() + 24 * 60 * 60;
+    block_on(
+        store.compare_and_swap_quota(QuotaObservation {
+            account_id: account.id().clone(),
+            expected_revision: account.revision(),
+            quota: Some(OpaqueProviderData::new(
+                json!({
+                    "rate_limit": {
+                        "allowed": false,
+                        "primary_window": {"used_percent": 100, "reset_at": reset_at}
+                    }
+                })
+                .as_object()
+                .expect("quota object")
+                .clone(),
+            )),
+            observed_at: Some(SystemTime::now()),
+        }),
+    )
+    .expect("persist exhausted quota");
+    let selector = selector(&store, Arc::new(TestLeaseCoordinator::default()));
+
+    block_on(selector.record_failure(
+        &account,
+        CodexAccountFailure::RateLimited {
+            retry_after: Some(Duration::from_secs(30)),
+        },
+    ))
+    .expect("record rate-limit failure");
+
+    assert_eq!(
+        store
+            .account("acct_primary")
+            .expect("persisted account")
+            .availability(),
+        AccountAvailability::QuotaExhausted
+    );
+}
+
+#[test]
+fn rate_limited_failure_ignores_exhaustion_after_its_reset_time() {
+    let store = Arc::new(MemoryAccountStore::default());
+    create_account(&store, "acct_primary", "at-primary");
+    let account = store.account("acct_primary").expect("account");
+    let reset_at = Utc::now().timestamp() - 1;
+    block_on(
+        store.compare_and_swap_quota(QuotaObservation {
+            account_id: account.id().clone(),
+            expected_revision: account.revision(),
+            quota: Some(OpaqueProviderData::new(
+                json!({
+                    "rate_limit": {
+                        "allowed": false,
+                        "primary_window": {"used_percent": 100, "reset_at": reset_at}
+                    }
+                })
+                .as_object()
+                .expect("quota object")
+                .clone(),
+            )),
+            observed_at: Some(SystemTime::now()),
+        }),
+    )
+    .expect("persist expired quota exhaustion");
+    let selector = selector(&store, Arc::new(TestLeaseCoordinator::default()));
+
+    block_on(selector.record_failure(
+        &account,
+        CodexAccountFailure::RateLimited {
+            retry_after: Some(Duration::from_secs(30)),
+        },
+    ))
+    .expect("record rate-limit failure");
+
+    assert_eq!(
+        store
+            .account("acct_primary")
+            .expect("persisted account")
+            .availability(),
+        AccountAvailability::Cooldown
+    );
+}
+
+#[test]
+fn rate_limited_failure_preserves_terminal_account_state() {
+    let store = Arc::new(MemoryAccountStore::default());
+    create_account(&store, "acct_primary", "at-primary");
+    let account = store.account("acct_primary").expect("account");
+    block_on(store.apply_state_change(AccountStateChange {
+        account_id: account.id().clone(),
+        expected_revision: account.revision(),
+        availability: AccountAvailability::Invalid,
+        reason: Some("identity_verification_required".to_owned()),
+        cooldown_until: None,
+        observed_at: SystemTime::now(),
+    }))
+    .expect("mark account invalid");
+    let current = store.account("acct_primary").expect("invalid account");
+    let selector = selector(&store, Arc::new(TestLeaseCoordinator::default()));
+
+    block_on(selector.record_failure(
+        &current,
+        CodexAccountFailure::RateLimited {
+            retry_after: Some(Duration::from_secs(30)),
+        },
+    ))
+    .expect("record rate-limit failure");
+
+    assert_eq!(
+        store
+            .account("acct_primary")
+            .expect("persisted account")
+            .availability(),
+        AccountAvailability::Invalid
+    );
+}
+
+#[test]
+fn rate_limited_failures_for_distinct_accounts_do_not_conflict() {
+    let store = Arc::new(MemoryAccountStore::default());
+    create_account(&store, "acct_first", "at-first");
+    create_account(&store, "acct_second", "at-second");
+    let first = store.account("acct_first").expect("first account");
+    let second = store.account("acct_second").expect("second account");
+    let selector = selector(&store, Arc::new(TestLeaseCoordinator::default()));
+
+    block_on(async {
+        let (first_result, second_result) = futures::join!(
+            selector.record_failure(
+                &first,
+                CodexAccountFailure::RateLimited {
+                    retry_after: Some(Duration::from_secs(30)),
+                },
+            ),
+            selector.record_failure(
+                &second,
+                CodexAccountFailure::RateLimited {
+                    retry_after: Some(Duration::from_secs(30)),
+                },
+            ),
+        );
+        first_result.expect("record first account failure");
+        second_result.expect("record second account failure");
+    });
+
+    assert_eq!(
+        [
+            store
+                .account("acct_first")
+                .expect("persisted first account")
+                .availability(),
+            store
+                .account("acct_second")
+                .expect("persisted second account")
+                .availability(),
+        ],
+        [AccountAvailability::Cooldown; 2]
+    );
+}
+
 #[test]
 fn selector_keeps_a_quota_exhausted_account_eligible_when_configured_not_to_skip_it() {
     let store = Arc::new(MemoryAccountStore::default());

backend/crates/providers/openai/tests/credential/quota.rs

@@ -480,6 +480,148 @@ async fn periodic_quota_synchronization_skips_fresh_ready_accounts() {
     assert_eq!(summary, CodexQuotaSyncSummary::default());
 }
 
+#[tokio::test]
+async fn periodic_quota_synchronization_repairs_cooldown_with_active_exhausted_quota() {
+    let store = Arc::new(MemoryAccountStore::default());
+    create_account(&store, "acct_periodic_stuck_cooldown").await;
+    let account = store
+        .account("acct_periodic_stuck_cooldown")
+        .expect("created account");
+    let reset_at = Utc::now().timestamp() + 24 * 60 * 60;
+    store
+        .apply_state_change(AccountStateChange {
+            account_id: account.id().clone(),
+            expected_revision: account.revision(),
+            availability: AccountAvailability::Cooldown,
+            reason: Some("rate_limited".to_owned()),
+            cooldown_until: SystemTime::now().checked_add(Duration::from_secs(24 * 60 * 60)),
+            observed_at: SystemTime::now(),
+        })
+        .await
+        .expect("mark account cooldown");
+    store
+        .compare_and_swap_quota(QuotaObservation {
+            account_id: account.id().clone(),
+            expected_revision: account.revision(),
+            quota: Some(OpaqueProviderData::new(
+                json!({
+                    "rate_limit": {
+                        "allowed": false,
+                        "primary_window": {"used_percent": 100, "reset_at": reset_at}
+                    }
+                })
+                .as_object()
+                .expect("quota object")
+                .clone(),
+            )),
+            observed_at: Some(SystemTime::now()),
+        })
+        .await
+        .expect("persist exhausted quota");
+    let server = MockServer::start().await;
+    Mock::given(method("GET"))
+        .and(path("/api/codex/usage"))
+        .respond_with(ResponseTemplate::new(429))
+        .mount(&server)
+        .await;
+    let service = quota_service_with_base_url(
+        &store,
+        reqwest::Client::builder()
+            .no_proxy()
+            .build()
+            .expect("client"),
+        server.uri(),
+    );
+
+    service.synchronize().await.expect("quota synchronization");
+
+    assert_eq!(
+        store
+            .account("acct_periodic_stuck_cooldown")
+            .expect("reconciled account")
+            .availability(),
+        AccountAvailability::QuotaExhausted
+    );
+}
+
+#[tokio::test]
+async fn periodic_quota_synchronization_does_not_downgrade_exhaustion_on_rate_limit() {
+    let store = Arc::new(MemoryAccountStore::default());
+    create_account(&store, "acct_periodic_exhausted").await;
+    let account = store
+        .account("acct_periodic_exhausted")
+        .expect("created account");
+    store
+        .apply_state_change(AccountStateChange {
+            account_id: account.id().clone(),
+            expected_revision: account.revision(),
+            availability: AccountAvailability::QuotaExhausted,
+            reason: Some("quota_exhausted".to_owned()),
+            cooldown_until: None,
+            observed_at: SystemTime::now(),
+        })
+        .await
+        .expect("mark account exhausted");
+    let server = MockServer::start().await;
+    Mock::given(method("GET"))
+        .and(path("/api/codex/usage"))
+        .respond_with(ResponseTemplate::new(429))
+        .mount(&server)
+        .await;
+    let service = quota_service_with_base_url(
+        &store,
+        reqwest::Client::builder()
+            .no_proxy()
+            .build()
+            .expect("client"),
+        server.uri(),
+    );
+
+    service.synchronize().await.expect("quota synchronization");
+
+    assert_eq!(
+        store
+            .account("acct_periodic_exhausted")
+            .expect("persisted account")
+            .availability(),
+        AccountAvailability::QuotaExhausted
+    );
+}
+
+#[tokio::test]
+async fn periodic_quota_synchronization_skips_ordinary_cooldown_without_exhausted_quota() {
+    let store = Arc::new(MemoryAccountStore::default());
+    create_account(&store, "acct_periodic_temporary_cooldown").await;
+    let account = store
+        .account("acct_periodic_temporary_cooldown")
+        .expect("created account");
+    store
+        .apply_state_change(AccountStateChange {
+            account_id: account.id().clone(),
+            expected_revision: account.revision(),
+            availability: AccountAvailability::Cooldown,
+            reason: Some("rate_limited".to_owned()),
+            cooldown_until: SystemTime::now().checked_add(Duration::from_secs(60)),
+            observed_at: SystemTime::now(),
+        })
+        .await
+        .expect("mark account cooldown");
+
+    let summary = blocked_network_quota_service(&store)
+        .synchronize()
+        .await
+        .expect("ordinary cooldown must not query upstream");
+
+    assert_eq!(summary, CodexQuotaSyncSummary::default());
+    assert_eq!(
+        store
+            .account("acct_periodic_temporary_cooldown")
+            .expect("cooldown account")
+            .availability(),
+        AccountAvailability::Cooldown
+    );
+}
+
 #[tokio::test]
 async fn periodic_quota_synchronization_limits_stale_ready_account_scans() {
     let store = Arc::new(MemoryAccountStore::default());

backend/crates/providers/openai/tests/provider/contract.rs

@@ -10,7 +10,7 @@ use gateway_core::engine::credential::{
 };
 use gateway_core::engine::provider::{Provider as _, ProviderRequest};
 use gateway_core::engine::{
-    AccountAttemptContext, AttemptContext, CancellationToken, ModelRequestId,
+    AccountAttemptContext, AttemptContext, CancellationToken, ContinuationAttempt, ModelRequestId,
     RequestAttemptContext, UpstreamSendState,
 };
 use gateway_core::error::ProviderErrorKind;
@@ -29,6 +29,12 @@ use provider_openai::transport::CodexWebSocketPool;
 use provider_openai::transport::profile::{CodexWireProfile, CodexWireProfileState};
 use provider_openai::{CodexProvider, OFFICIAL_CODEX_BASE_URL};
 use serde_json::{Map, json};
+use tokio::{
+    io::{AsyncReadExt, AsyncWriteExt},
+    net::{TcpListener, TcpStream},
+    sync::oneshot,
+    time::timeout,
+};
 use wiremock::matchers::{method, path};
 use wiremock::{Mock, MockServer, ResponseTemplate};
 
@@ -225,6 +231,78 @@ fn diagnostic_context(request_id: &str, account_id: &str) -> AttemptContext {
     )
 }
 
+async fn paused_chunked_sse_server(
+    first_chunk: String,
+    second_chunk: String,
+) -> (
+    String,
+    oneshot::Sender<()>,
+    oneshot::Receiver<()>,
+    tokio::task::JoinHandle<()>,
+) {
+    let listener = TcpListener::bind("127.0.0.1:0")
+        .await
+        .expect("bind chunked SSE listener");
+    let base_url = format!(
+        "http://{}",
+        listener.local_addr().expect("listener address")
+    );
+    let (release_sender, release_receiver) = oneshot::channel();
+    let (first_chunk_sender, first_chunk_sent) = oneshot::channel();
+    let server = tokio::spawn(async move {
+        let (mut stream, _) = listener.accept().await.expect("accept chunked SSE request");
+        read_http_request(&mut stream).await;
+        stream
+            .write_all(
+                b"HTTP/1.1 200 OK\r\ncontent-type: text/event-stream\r\ntransfer-encoding: chunked\r\n\r\n",
+            )
+            .await
+            .expect("write chunked SSE headers");
+        write_http_chunk(&mut stream, &first_chunk).await;
+        let _ = first_chunk_sender.send(());
+        let _ = release_receiver.await;
+        if !second_chunk.is_empty() {
+            write_http_chunk(&mut stream, &second_chunk).await;
+        }
+        stream
+            .write_all(b"0\r\n\r\n")
+            .await
+            .expect("terminate chunked SSE response");
+    });
+    (base_url, release_sender, first_chunk_sent, server)
+}
+
+async fn read_http_request(stream: &mut TcpStream) {
+    let mut request = Vec::new();
+    let mut buffer = [0_u8; 1024];
+    loop {
+        let read = stream.read(&mut buffer).await.expect("read HTTP request");
+        if read == 0 {
+            return;
+        }
+        request.extend_from_slice(&buffer[..read]);
+        if request.windows(4).any(|window| window == b"\r\n\r\n") {
+            return;
+        }
+    }
+}
+
+async fn write_http_chunk(stream: &mut TcpStream, body: &str) {
+    stream
+        .write_all(format!("{:X}\r\n", body.len()).as_bytes())
+        .await
+        .expect("write HTTP chunk size");
+    stream
+        .write_all(body.as_bytes())
+        .await
+        .expect("write HTTP chunk body");
+    stream
+        .write_all(b"\r\n")
+        .await
+        .expect("terminate HTTP chunk");
+    stream.flush().await.expect("flush HTTP chunk");
+}
+
 #[tokio::test]
 async fn openai_provider_rejects_a_foreign_provider_candidate_before_account_selection() {
     let store = Arc::new(MemoryAccountStore::default());
@@ -449,6 +527,355 @@ async fn completed_response_persists_session_affinity_before_stream_consumer_sto
     assert_eq!(affinity.binding_count(), 1);
 }
 
+#[tokio::test]
+async fn response_failed_before_semantic_output_is_atomic_and_persists_cooldown() {
+    let store = Arc::new(MemoryAccountStore::default());
+    let account_id = "acct_atomic_failure";
+    create_account(&store, account_id).await;
+    let (base_url, release, _first_chunk_sent, server) = paused_chunked_sse_server(
+        concat!(
+            "event: response.created\n",
+            "data: {\"type\":\"response.created\",\"response\":{\"id\":\"resp_atomic_failure\",\"model\":\"gpt-5.4\",\"status\":\"in_progress\"}}\n\n",
+            "event: response.failed\n",
+            "data: {\"type\":\"response.failed\",\"status_code\":429,\"retry_after_seconds\":17,\"response\":{\"id\":\"resp_atomic_failure\",\"status\":\"failed\",\"error\":{\"code\":\"rate_limit_exceeded\",\"message\":\"atomic upstream failure\"}}}\n\n"
+        )
+        .to_owned(),
+        String::new(),
+    )
+    .await;
+    let mut stream = provider_with_base_url(&store, base_url)
+        .execute(
+            planned_request("openai", http_generate_operation()),
+            context("req_atomic_failure", CancellationToken::new()),
+        )
+        .await
+        .expect("prepare provider stream");
+    let mut visible_before_error = 0;
+    let mut failure = loop {
+        match stream.next().await {
+            Some(Ok(event)) => visible_before_error += usize::from(event.has_client_event()),
+            Some(Err(error)) => break error,
+            None => panic!("response.failed must produce a typed failure"),
+        }
+    };
+
+    assert_eq!(visible_before_error, 0);
+    assert_eq!(failure.kind(), ProviderErrorKind::RateLimited);
+    assert_eq!(failure.send_state(), UpstreamSendState::Sent);
+    assert_eq!(failure.upstream_status(), Some(429));
+    assert!(failure.replay_is_safe());
+    let events = failure.take_atomic_client_events();
+    assert_eq!(
+        events
+            .iter()
+            .filter_map(|event| event.wire_event()?.event_type())
+            .collect::<Vec<_>>(),
+        vec!["response.created", "response.failed"]
+    );
+    let account = store.account(account_id).expect("rate-limited account");
+    assert_eq!(account.availability(), AccountAvailability::Cooldown);
+    assert!(
+        account
+            .cooldown_until()
+            .is_some_and(|until| until > SystemTime::now())
+    );
+    let _ = release.send(());
+    server.abort();
+    let _ = server.await;
+}
+
+#[tokio::test]
+async fn ordinary_request_should_forward_created_before_later_failed_chunk() {
+    let store = Arc::new(MemoryAccountStore::default());
+    create_account(&store, "acct_first_event_latency").await;
+    let (base_url, release, _first_chunk_sent, server) = paused_chunked_sse_server(
+        concat!(
+            "event: response.created\n",
+            "data: {\"type\":\"response.created\",\"response\":{\"id\":\"resp_first_event\",\"model\":\"gpt-5.4\",\"status\":\"in_progress\"}}\n\n"
+        )
+        .to_owned(),
+        concat!(
+            "event: response.failed\n",
+            "data: {\"type\":\"response.failed\",\"status_code\":429,\"response\":{\"id\":\"resp_first_event\",\"status\":\"failed\",\"error\":{\"code\":\"rate_limit_exceeded\",\"message\":\"later failure\"}}}\n\n"
+        )
+        .to_owned(),
+    )
+    .await;
+    let mut stream = provider_with_base_url(&store, base_url)
+        .execute(
+            planned_request("openai", http_generate_operation()),
+            context("req_first_event_latency", CancellationToken::new()),
+        )
+        .await
+        .expect("prepare provider stream");
+
+    let first_event = loop {
+        let next = timeout(Duration::from_secs(1), stream.next())
+            .await
+            .expect("first chunk must not wait for the withheld failure")
+            .expect("provider stream must stay open")
+            .expect("provider event");
+        if next.has_client_event() {
+            break next;
+        }
+    };
+    assert_eq!(
+        first_event.wire_event().and_then(|wire| wire.event_type()),
+        Some("response.created")
+    );
+
+    release.send(()).expect("release second upstream chunk");
+    let mut later_wire_types = Vec::new();
+    let failure = loop {
+        let next = timeout(Duration::from_secs(1), stream.next())
+            .await
+            .expect("later failure must arrive after release")
+            .expect("provider stream must return typed failure");
+        match next {
+            Ok(event) => {
+                if let Some(event_type) = event.wire_event().and_then(|wire| wire.event_type()) {
+                    later_wire_types.push(event_type.to_owned());
+                }
+            }
+            Err(error) => break error,
+        }
+    };
+
+    assert_eq!(later_wire_types, vec!["response.failed"]);
+    assert!(!failure.replay_is_safe());
+    assert!(!failure.has_atomic_client_events());
+    server.await.expect("chunked SSE server");
+}
+
+#[tokio::test]
+async fn continuation_should_hold_created_until_semantic_output() {
+    let store = Arc::new(MemoryAccountStore::default());
+    create_account(&store, "acct_continuation_prefetch").await;
+    let (base_url, release, first_chunk_sent, server) = paused_chunked_sse_server(
+        concat!(
+            "event: response.created\n",
+            "data: {\"type\":\"response.created\",\"response\":{\"id\":\"resp_continuation\",\"model\":\"gpt-5.4\",\"status\":\"in_progress\"}}\n\n"
+        )
+        .to_owned(),
+        concat!(
+            "event: response.content_part.added\n",
+            "data: {\"type\":\"response.content_part.added\",\"output_index\":0,\"content_index\":0,\"part\":{\"type\":\"output_text\"}}\n\n",
+            "event: response.output_text.delta\n",
+            "data: {\"type\":\"response.output_text.delta\",\"output_index\":0,\"content_index\":0,\"delta\":\"hello\"}\n\n",
+            "event: response.completed\n",
+            "data: {\"type\":\"response.completed\",\"response\":{\"id\":\"resp_continuation\",\"model\":\"gpt-5.4\",\"status\":\"completed\",\"output\":[],\"usage\":{\"input_tokens\":1,\"output_tokens\":1,\"total_tokens\":2}}}\n\n"
+        )
+        .to_owned(),
+    )
+    .await;
+    let mut stream = provider_with_base_url(&store, base_url)
+        .execute(
+            planned_request("openai", http_generate_operation()),
+            context("req_continuation_prefetch", CancellationToken::new())
+                .with_continuation_attempt(ContinuationAttempt::Native),
+        )
+        .await
+        .expect("prepare provider stream");
+
+    let mut first_chunk_sent = Box::pin(first_chunk_sent);
+    loop {
+        tokio::select! {
+            sent = &mut first_chunk_sent => {
+                sent.expect("first upstream chunk");
+                break;
+            }
+            next = stream.next() => {
+                let event = next
+                    .expect("provider stream must stay open")
+                    .expect("provider event");
+                assert!(!event.has_client_event(), "created was exposed before the first chunk barrier");
+            }
+        }
+    }
+
+    let blocked = timeout(Duration::from_millis(100), async {
+        loop {
+            let next = stream
+                .next()
+                .await
+                .expect("provider stream must stay open")
+                .expect("provider event");
+            if next.has_client_event() {
+                return next;
+            }
+        }
+    })
+    .await;
+    assert!(blocked.is_err());
+
+    release.send(()).expect("release semantic output chunk");
+    let first_event = loop {
+        let next = timeout(Duration::from_secs(1), stream.next())
+            .await
+            .expect("semantic output must release the prefetched batch")
+            .expect("provider stream must stay open")
+            .expect("provider event");
+        if next.has_client_event() {
+            break next;
+        }
+    };
+    assert_eq!(
+        first_event.wire_event().and_then(|wire| wire.event_type()),
+        Some("response.created")
+    );
+    while let Some(event) = stream.next().await {
+        event.expect("continuation response must complete");
+    }
+    server.await.expect("chunked SSE server");
+}
+
+#[tokio::test]
+async fn bare_response_failed_should_remain_an_atomic_replay_safe_failure() {
+    let store = Arc::new(MemoryAccountStore::default());
+    create_account(&store, "acct_bare_atomic_failure").await;
+    let (base_url, release, _first_chunk_sent, server) = paused_chunked_sse_server(
+        concat!(
+            "event: response.failed\n",
+            "data: {\"type\":\"response.failed\",\"status_code\":429,\"response\":{\"id\":\"resp_bare_failure\",\"status\":\"failed\",\"error\":{\"code\":\"rate_limit_exceeded\",\"message\":\"bare failure\"}}}\n\n"
+        )
+        .to_owned(),
+        String::new(),
+    )
+    .await;
+    let mut stream = provider_with_base_url(&store, base_url)
+        .execute(
+            planned_request("openai", http_generate_operation()),
+            context("req_bare_atomic_failure", CancellationToken::new()),
+        )
+        .await
+        .expect("prepare provider stream");
+    let mut visible_before_error = 0;
+    let mut failure = loop {
+        match stream.next().await {
+            Some(Ok(event)) => visible_before_error += usize::from(event.has_client_event()),
+            Some(Err(error)) => break error,
+            None => panic!("bare response.failed must produce a typed failure"),
+        }
+    };
+
+    assert_eq!(visible_before_error, 0);
+    assert!(failure.replay_is_safe());
+    let events = failure.take_atomic_client_events();
+    assert_eq!(events.len(), 1);
+    assert_eq!(
+        events[0].wire_event().and_then(|wire| wire.event_type()),
+        Some("response.failed")
+    );
+    assert!(events[0]
+        .canonical_facts()
+        .iter()
+        .any(|event| matches!(event, GatewayEvent::Started(meta) if meta.response_id() == "resp_bare_failure")));
+    let _ = release.send(());
+    server.abort();
+    let _ = server.await;
+}
+
+#[tokio::test]
+async fn continuation_prefetch_over_64_kib_without_a_boundary_should_fail_as_protocol_error() {
+    let store = Arc::new(MemoryAccountStore::default());
+    create_account(&store, "acct_prefetch_limit").await;
+    let padding = "x".repeat(64 * 1024);
+    let body = format!(
+        "event: response.created\ndata: {}\n\n",
+        json!({
+            "type": "response.created",
+            "response": {
+                "id": "resp_prefetch_limit",
+                "model": "gpt-5.4",
+                "status": "in_progress",
+                "padding": padding,
+            }
+        })
+    );
+    assert!(body.len() > 64 * 1024);
+    let (base_url, release, _first_chunk_sent, server) =
+        paused_chunked_sse_server(body, String::new()).await;
+    let mut stream = provider_with_base_url(&store, base_url)
+        .execute(
+            planned_request("openai", http_generate_operation()),
+            context("req_prefetch_limit", CancellationToken::new())
+                .with_continuation_attempt(ContinuationAttempt::Native),
+        )
+        .await
+        .expect("prepare provider stream");
+    let mut visible_before_error = 0;
+    let failure = loop {
+        match stream.next().await {
+            Some(Ok(event)) => visible_before_error += usize::from(event.has_client_event()),
+            Some(Err(error)) => break error,
+            None => panic!("oversized prefetch must fail"),
+        }
+    };
+
+    assert_eq!(visible_before_error, 0);
+    assert_eq!(failure.kind(), ProviderErrorKind::Protocol);
+    assert!(!failure.has_atomic_client_events());
+    let _ = release.send(());
+    server.abort();
+    let _ = server.await;
+}
+
+#[tokio::test]
+async fn response_failed_after_semantic_output_is_exposed_and_not_replay_safe() {
+    let store = Arc::new(MemoryAccountStore::default());
+    create_account(&store, "acct_semantic_failure").await;
+    let server = MockServer::start().await;
+    Mock::given(method("POST"))
+        .and(path("/codex/responses"))
+        .respond_with(
+            ResponseTemplate::new(200)
+                .insert_header("content-type", "text/event-stream")
+                .set_body_string(concat!(
+                    "event: response.created\n",
+                    "data: {\"type\":\"response.created\",\"response\":{\"id\":\"resp_semantic_failure\",\"model\":\"gpt-5.4\",\"status\":\"in_progress\"}}\n\n",
+                    "event: response.content_part.added\n",
+                    "data: {\"type\":\"response.content_part.added\",\"output_index\":0,\"content_index\":0,\"part\":{\"type\":\"output_text\"}}\n\n",
+                    "event: response.output_text.delta\n",
+                    "data: {\"type\":\"response.output_text.delta\",\"output_index\":0,\"content_index\":0,\"delta\":\"hello\"}\n\n",
+                    "event: response.failed\n",
+                    "data: {\"type\":\"response.failed\",\"status_code\":429,\"response\":{\"id\":\"resp_semantic_failure\",\"status\":\"failed\",\"error\":{\"code\":\"rate_limit_exceeded\",\"message\":\"failure after output\"}}}\n\n"
+                )),
+        )
+        .mount(&server)
+        .await;
+    let mut stream = provider_with_base_url(&store, server.uri())
+        .execute(
+            planned_request("openai", http_generate_operation()),
+            context("req_semantic_failure", CancellationToken::new()),
+        )
+        .await
+        .expect("prepare provider stream");
+    let mut wire_types = Vec::new();
+    let failure = loop {
+        match stream.next().await {
+            Some(Ok(event)) => {
+                if let Some(event_type) = event.wire_event().and_then(|wire| wire.event_type()) {
+                    wire_types.push(event_type.to_owned());
+                }
+            }
+            Some(Err(error)) => break error,
+            None => panic!("response.failed must produce a typed failure"),
+        }
+    };
+
+    assert_eq!(
+        wire_types,
+        vec![
+            "response.created",
+            "response.content_part.added",
+            "response.output_text.delta",
+            "response.failed"
+        ]
+    );
+    assert!(!failure.replay_is_safe());
+    assert!(!failure.has_atomic_client_events());
+}
+
 #[tokio::test]
 async fn disabled_account_diagnostic_uses_upstream_without_persisting_account_state() {
     let store = Arc::new(MemoryAccountStore::default());

backend/crates/providers/openai/tests/transport/canonical.rs

@@ -528,6 +528,40 @@ fn raw_sse_failure_should_keep_the_original_frame_before_reporting_the_typed_fai
     assert!(!format!("{failure:?}").contains("raw failure marker"));
 }
 
+#[test]
+fn bare_response_failed_should_project_started_identity_with_the_fallback_model() {
+    let raw = concat!(
+        "event: response.failed\n",
+        "data: {\"type\":\"response.failed\",\"response\":{\"id\":\"resp_bare_failed\",\"status\":\"failed\",\"error\":{\"code\":\"rate_limit_exceeded\",\"message\":\"bare failure\"}}}\n\n",
+    );
+
+    let failure = CodexCanonicalDecoder::new("fallback-model")
+        .push(raw.as_bytes())
+        .expect_err("bare response.failed remains a typed lifecycle failure");
+
+    assert!(matches!(
+        canonical_facts(failure.events()).as_slice(),
+        [GatewayEvent::Started(metadata)]
+            if metadata.response_id() == "resp_bare_failed"
+                && metadata.model() == "fallback-model"
+    ));
+}
+
+#[test]
+fn response_failed_without_identity_should_keep_the_upstream_typed_failure() {
+    let raw = concat!(
+        "event: response.failed\n",
+        "data: {\"type\":\"response.failed\",\"response\":{\"status\":\"failed\",\"error\":{\"code\":\"rate_limit_exceeded\",\"message\":\"missing identity\"}}}\n\n",
+    );
+
+    let failure = CodexCanonicalDecoder::new("fallback")
+        .push(raw.as_bytes())
+        .expect_err("missing identity must not replace the upstream lifecycle failure");
+
+    assert!(matches!(failure.error(), CodexCanonicalError::Upstream(_)));
+    assert!(canonical_facts(failure.events()).is_empty());
+}
+
 #[test]
 fn decoder_should_preserve_same_chunk_output_before_typed_failure() {
     let marker = "same-chunk-secret-marker";

backend/crates/providers/openai/tests/transport/diagnostics.rs

@@ -143,3 +143,19 @@ fn structured_expiry_code_classifies_without_an_auth_status() {
         CodexFailureCategory::CredentialExpired
     );
 }
+
+#[test]
+fn structured_quota_signal_takes_precedence_over_http_429() {
+    assert_eq!(
+        classify(
+            429,
+            r#"{"error":{"code":"insufficient_quota","message":"rate limit reached"}}"#
+        ),
+        CodexFailureCategory::QuotaExhausted
+    );
+}
+
+#[test]
+fn bare_http_429_remains_a_temporary_rate_limit() {
+    assert_eq!(classify(429, ""), CodexFailureCategory::RateLimited);
+}

backend/crates/providers/xai/src/credential/selector.rs

@@ -10,10 +10,11 @@ use gateway_core::engine::credential::{
     AccountSelectionContext, AccountSelector, ProviderAccount, ProviderAccountId,
 };
 use gateway_core::provider_ports::{
-    ProviderCooldown, ProviderCooldownPort, ProviderLeaseAcquisition, ProviderLeasePort,
-    ProviderLeaseRequest, ProviderSchedulingLeaseRequest,
+    ProviderCooldown, ProviderCooldownPort, ProviderCooldownScope, ProviderLeaseAcquisition,
+    ProviderLeasePort, ProviderLeaseRequest, ProviderSchedulingLeaseRequest,
+    ProviderScopedCooldown,
 };
-use gateway_core::routing::ProviderKind;
+use gateway_core::routing::{ProviderKind, UpstreamModelId};
 
 use super::catalog::{GrokCatalogScope, GrokCredentialCatalogCache, GrokCredentialQuotaService};
 use super::repository::GrokCredentialRepository;
@@ -28,6 +29,22 @@ const DEFAULT_RATE_LIMIT_COOLDOWN: Duration = Duration::from_secs(60);
 const MAX_RATE_LIMIT_COOLDOWN: Duration = Duration::from_secs(24 * 60 * 60);
 const STREAM_INTERRUPTION_COOLDOWN: Duration = Duration::from_secs(30);
 const UNAUTHORIZED_COOLDOWN: Duration = Duration::from_secs(60);
+const PAYMENT_REQUIRED_COOLDOWN: Duration = Duration::from_secs(60);
+const MODEL_ACCESS_DENIED_COOLDOWN: Duration = Duration::from_secs(5 * 60);
+const MODEL_QUOTA_COOLDOWN: Duration = Duration::from_secs(24 * 60 * 60);
+const MAX_MODEL_COOLDOWN: Duration = Duration::from_secs(24 * 60 * 60);
+
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+enum RuntimeCooldownScope {
+    Account,
+    Model,
+}
+
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+struct RuntimeCooldown {
+    until: SystemTime,
+    scope: RuntimeCooldownScope,
+}
 
 /// 仅经 Core account port、TTL catalog cache 和 Redis lease 选择一个 OAuth session。
 pub struct GrokAccountSessionSelector {
@@ -124,7 +141,8 @@ impl GrokAccountSessionSelector {
         let runtime_cooldowns = if diagnostic {
             std::collections::BTreeMap::new()
         } else {
-            self.runtime_cooldowns(&catalog_eligible).await
+            self.runtime_cooldowns(&catalog_eligible, request.upstream_model())
+                .await
         };
         let mut candidates = catalog_eligible
             .into_iter()
@@ -234,16 +252,26 @@ impl GrokAccountSessionSelector {
             return Err(GrokSessionSelectorError::CapacityUnavailable { retry_after });
         }
         // 钉死账号时只看该账号，否则取最早退出 cooldown 的候选。
-        let cooled_until = match request.required_account() {
+        let cooled = match request.required_account() {
             Some(required) => runtime_cooldowns.get(required).copied(),
-            None => runtime_cooldowns.values().copied().min(),
+            None => runtime_cooldowns
+                .values()
+                .copied()
+                .min_by_key(|item| item.until),
         };
-        if let Some(until) = cooled_until {
-            return Err(GrokSessionSelectorError::AccountCoolingDown {
-                retry_after: until
-                    .duration_since(SystemTime::now())
-                    .ok()
-                    .filter(|remaining| !remaining.is_zero()),
+        if let Some(cooled) = cooled {
+            let retry_after = cooled
+                .until
+                .duration_since(SystemTime::now())
+                .ok()
+                .filter(|remaining| !remaining.is_zero());
+            return Err(match cooled.scope {
+                RuntimeCooldownScope::Account => {
+                    GrokSessionSelectorError::AccountCoolingDown { retry_after }
+                }
+                RuntimeCooldownScope::Model => {
+                    GrokSessionSelectorError::ModelCoolingDown { retry_after }
+                }
             });
         }
         Err(GrokSessionSelectorError::NoEligibleSession)
@@ -252,28 +280,56 @@ impl GrokAccountSessionSelector {
     async fn runtime_cooldowns(
         &self,
         accounts: &[ProviderAccount],
-    ) -> std::collections::BTreeMap<ProviderAccountId, SystemTime> {
+        upstream_model: &UpstreamModelId,
+    ) -> std::collections::BTreeMap<ProviderAccountId, RuntimeCooldown> {
         let now = SystemTime::now();
-        // 冷却状态并发批量读取，避免逐账号串行往返。
-        let reads = futures::future::join_all(
+        let model_scope = ProviderCooldownScope::upstream_model(upstream_model.clone());
+        // 账号级和当前模型级冷却都并发批量读取，避免逐账号串行往返。
+        let account_reads = futures::future::join_all(
             accounts
                 .iter()
                 .map(|account| self.cooldowns.read(account.id())),
-        )
-        .await;
+        );
+        let model_reads = futures::future::join_all(
+            accounts
+                .iter()
+                .map(|account| self.cooldowns.read_scoped(account.id(), &model_scope)),
+        );
+        let (account_reads, model_reads) = futures::join!(account_reads, model_reads);
         let mut cooled = std::collections::BTreeMap::new();
-        for (account, read) in accounts.iter().zip(reads) {
-            let Ok(Some(cooldown)) = read else {
-                continue;
-            };
-            if cooldown.credential_revision() != account.revision() {
-                if cooldown.credential_revision() < account.revision() {
-                    let _ = self.cooldowns.clear(account.id(), account.revision()).await;
+        for ((account, account_read), model_read) in
+            accounts.iter().zip(account_reads).zip(model_reads)
+        {
+            if let Ok(Some(cooldown)) = account_read {
+                if cooldown.credential_revision() != account.revision() {
+                    if cooldown.credential_revision() < account.revision() {
+                        let _ = self.cooldowns.clear(account.id(), account.revision()).await;
+                    }
+                } else if cooldown.until() > now {
+                    insert_runtime_cooldown(
+                        &mut cooled,
+                        account.id(),
+                        cooldown.until(),
+                        RuntimeCooldownScope::Account,
+                    );
                 }
-                continue;
             }
-            if cooldown.until() > now {
-                cooled.insert(account.id().clone(), cooldown.until());
+            if let Ok(Some(cooldown)) = model_read {
+                if cooldown.credential_revision() != account.revision() {
+                    if cooldown.credential_revision() < account.revision() {
+                        let _ = self
+                            .cooldowns
+                            .clear_scoped(account.id(), &model_scope, account.revision())
+                            .await;
+                    }
+                } else if cooldown.until() > now {
+                    insert_runtime_cooldown(
+                        &mut cooled,
+                        account.id(),
+                        cooldown.until(),
+                        RuntimeCooldownScope::Model,
+                    );
+                }
             }
         }
         cooled
@@ -292,6 +348,26 @@ impl GrokAccountSessionSelector {
             ))
             .await;
     }
+
+    async fn record_model_runtime_cooldown(
+        &self,
+        session: &SelectedGrokSession,
+        upstream_model: UpstreamModelId,
+        duration: Duration,
+    ) {
+        let cooldown_until = SystemTime::now()
+            .checked_add(duration)
+            .unwrap_or_else(SystemTime::now);
+        let _ = self
+            .cooldowns
+            .put_scoped_if_later(ProviderScopedCooldown::new(
+                session.account_id().clone(),
+                session.credential_revision(),
+                ProviderCooldownScope::upstream_model(upstream_model),
+                cooldown_until,
+            ))
+            .await;
+    }
 }
 
 impl fmt::Debug for GrokAccountSessionSelector {
@@ -333,6 +409,43 @@ impl GrokSessionSelector for GrokAccountSessionSelector {
                     "upstream_quota_exhausted",
                     None,
                 ),
+                GrokCredentialFailure::FreeQuotaExhausted => (
+                    GrokCredentialAvailability::QuotaExhausted,
+                    "upstream_free_quota_exhausted",
+                    None,
+                ),
+                GrokCredentialFailure::PaymentRequired { retry_after } => {
+                    let retry_after = bounded_cooldown(
+                        retry_after,
+                        PAYMENT_REQUIRED_COOLDOWN,
+                        MAX_RATE_LIMIT_COOLDOWN,
+                    );
+                    self.record_runtime_cooldown(session, retry_after).await;
+                    return;
+                }
+                GrokCredentialFailure::ModelQuotaExhausted {
+                    upstream_model,
+                    retry_after,
+                } => {
+                    let retry_after =
+                        bounded_cooldown(retry_after, MODEL_QUOTA_COOLDOWN, MAX_MODEL_COOLDOWN);
+                    self.record_model_runtime_cooldown(session, upstream_model, retry_after)
+                        .await;
+                    return;
+                }
+                GrokCredentialFailure::ModelAccessDenied {
+                    upstream_model,
+                    retry_after,
+                } => {
+                    let retry_after = bounded_cooldown(
+                        retry_after,
+                        MODEL_ACCESS_DENIED_COOLDOWN,
+                        MAX_MODEL_COOLDOWN,
+                    );
+                    self.record_model_runtime_cooldown(session, upstream_model, retry_after)
+                        .await;
+                    return;
+                }
                 GrokCredentialFailure::RateLimited { retry_after } => {
                     let retry_after = retry_after
                         .unwrap_or(DEFAULT_RATE_LIMIT_COOLDOWN)
@@ -414,3 +527,29 @@ fn minimum_retry_after(left: Option<Duration>, right: Option<Duration>) -> Optio
         (None, None) => None,
     }
 }
+
+fn bounded_cooldown(
+    retry_after: Option<Duration>,
+    default: Duration,
+    maximum: Duration,
+) -> Duration {
+    retry_after
+        .filter(|duration| !duration.is_zero())
+        .unwrap_or(default)
+        .min(maximum)
+}
+
+fn insert_runtime_cooldown(
+    cooldowns: &mut std::collections::BTreeMap<ProviderAccountId, RuntimeCooldown>,
+    account_id: &ProviderAccountId,
+    until: SystemTime,
+    scope: RuntimeCooldownScope,
+) {
+    let replacement = RuntimeCooldown { until, scope };
+    if cooldowns
+        .get(account_id)
+        .is_none_or(|current| current.until < until)
+    {
+        cooldowns.insert(account_id.clone(), replacement);
+    }
+}

backend/crates/providers/xai/src/provider.rs

@@ -927,6 +927,7 @@ async fn start_grok_inference(
     transport: &dyn GrokInferenceTransport,
     credential_recovery: &dyn GrokCredentialRecovery,
     request: GrokInferenceRequest,
+    upstream_model: &UpstreamModelId,
     context: &AttemptContext,
     session: &SelectedGrokSession,
 ) -> Result<AcceptedGrokInference, GrokInferenceStartFailure> {
@@ -965,6 +966,7 @@ async fn start_grok_inference(
         InferenceBoundary::Response(Ok(response)) => response,
         InferenceBoundary::Response(Err(error)) => {
             let observation = xai_error_observation(&error).ok();
+            let credential_failure = transport_credential_failure(&error, upstream_model);
             let error =
                 map_continuation_failure(context, map_transport_error_for_context(error, context));
             let error = recover_or_record_failure(
@@ -972,6 +974,7 @@ async fn start_grok_inference(
                 credential_recovery,
                 session,
                 error,
+                credential_failure,
                 context.credential_recovery_attempted(),
             )
             .await;
@@ -993,6 +996,7 @@ async fn next_grok_chunk(
     body: &mut GrokInferenceChunkStream,
     selector: &dyn GrokSessionSelector,
     session: &SelectedGrokSession,
+    upstream_model: &UpstreamModelId,
     context: &AttemptContext,
 ) -> Result<Option<bytes::Bytes>, ProviderError> {
     let Some(stream_deadline) = remaining(context.deadline()) else {
@@ -1014,10 +1018,12 @@ async fn next_grok_chunk(
         )),
         chunk = body.next() => match chunk {
             Some(Ok(chunk)) => Ok(Some(chunk)),
-            Some(Err(error)) => {
-                let error = map_stream_error(error);
-                Err(record_stream_failure(selector, session, error).await)
-            }
+            Some(Err(error)) => Err(map_and_record_stream_transport_failure(
+                selector,
+                session,
+                error,
+                upstream_model,
+            ).await),
             None => Ok(None),
         },
     }
@@ -1067,6 +1073,7 @@ fn cold_compaction_http_sse_stream(
             transport.as_ref(),
             credential_recovery.as_ref(),
             inference_request,
+            &upstream_model,
             &context,
             &session,
         )
@@ -1092,6 +1099,7 @@ fn cold_compaction_http_sse_stream(
             &mut body,
             selector.as_ref(),
             &session,
+            &upstream_model,
             &context,
         )
         .await
@@ -1287,6 +1295,7 @@ fn cold_http_sse_stream(
             InferenceBoundary::Response(Ok(response)) => response,
             InferenceBoundary::Response(Err(error)) => {
                 let observation = xai_error_observation(&error)?;
+                let credential_failure = transport_credential_failure(&error, &upstream_model);
                 let error =
                     map_continuation_failure(&context, map_transport_error_for_context(error, &context));
                 let error = recover_or_record_failure(
@@ -1294,6 +1303,7 @@ fn cold_http_sse_stream(
                     credential_recovery.as_ref(),
                     &session,
                     error,
+                    credential_failure,
                     context.credential_recovery_attempted(),
                 )
                 .await;
@@ -1328,10 +1338,12 @@ fn cold_http_sse_stream(
                 )),
                 chunk = body.next() => match chunk {
                     Some(Ok(chunk)) => Ok(Some(chunk)),
-                    Some(Err(error)) => {
-                        let error = map_stream_error(error);
-                        Err(record_stream_failure(selector.as_ref(), &session, error).await)
-                    },
+                    Some(Err(error)) => Err(map_and_record_stream_transport_failure(
+                        selector.as_ref(),
+                        &session,
+                        error,
+                        &upstream_model,
+                    ).await),
                     None => Ok(None),
                 },
             }?;
@@ -1495,6 +1507,18 @@ async fn record_failure(
     error
 }
 
+async fn record_credential_failure(
+    selector: &dyn GrokSessionSelector,
+    session: &SelectedGrokSession,
+    error: ProviderError,
+    failure: GrokCredentialFailure,
+) -> ProviderError {
+    if session.allows_account_state_mutation() {
+        selector.record_failure(session, failure).await;
+    }
+    error
+}
+
 async fn record_stream_failure(
     selector: &dyn GrokSessionSelector,
     session: &SelectedGrokSession,
@@ -1524,6 +1548,7 @@ async fn recover_or_record_failure(
     recovery: &dyn GrokCredentialRecovery,
     session: &SelectedGrokSession,
     error: ProviderError,
+    credential_failure: Option<GrokCredentialFailure>,
     recovery_attempted: bool,
 ) -> ProviderError {
     if !session.allows_account_state_mutation() {
@@ -1536,12 +1561,68 @@ async fn recover_or_record_failure(
         {
             GrokCredentialRecoveryOutcome::Recovered => error.with_same_account_retry(),
             GrokCredentialRecoveryOutcome::Rejected => error,
-            GrokCredentialRecoveryOutcome::Unavailable => {
-                record_failure(selector, session, error).await
-            }
+            GrokCredentialRecoveryOutcome::Unavailable => match credential_failure {
+                Some(failure) => record_credential_failure(selector, session, error, failure).await,
+                None => error,
+            },
         };
     }
-    record_failure(selector, session, error).await
+    match credential_failure {
+        Some(failure) => record_credential_failure(selector, session, error, failure).await,
+        None => error,
+    }
+}
+
+fn transport_credential_failure(
+    error: &GrokInferenceTransportError,
+    upstream_model: &UpstreamModelId,
+) -> Option<GrokCredentialFailure> {
+    match error.kind() {
+        GrokInferenceTransportErrorKind::Unauthorized => Some(GrokCredentialFailure::Unauthorized),
+        GrokInferenceTransportErrorKind::RateLimited => Some(GrokCredentialFailure::RateLimited {
+            retry_after: error.retry_after(),
+        }),
+        GrokInferenceTransportErrorKind::QuotaExhausted => {
+            Some(GrokCredentialFailure::QuotaExhausted)
+        }
+        GrokInferenceTransportErrorKind::FreeQuotaExhausted => {
+            Some(GrokCredentialFailure::FreeQuotaExhausted)
+        }
+        GrokInferenceTransportErrorKind::PaymentRequired => {
+            Some(GrokCredentialFailure::PaymentRequired {
+                retry_after: error.retry_after(),
+            })
+        }
+        GrokInferenceTransportErrorKind::ModelQuotaExhausted => {
+            Some(GrokCredentialFailure::ModelQuotaExhausted {
+                upstream_model: upstream_model.clone(),
+                retry_after: error.retry_after(),
+            })
+        }
+        GrokInferenceTransportErrorKind::ModelAccessDenied => {
+            Some(GrokCredentialFailure::ModelAccessDenied {
+                upstream_model: upstream_model.clone(),
+                retry_after: error.retry_after(),
+            })
+        }
+        _ => None,
+    }
+}
+
+async fn map_and_record_stream_transport_failure(
+    selector: &dyn GrokSessionSelector,
+    session: &SelectedGrokSession,
+    error: GrokInferenceTransportError,
+    upstream_model: &UpstreamModelId,
+) -> ProviderError {
+    let request_scoped = error.kind() == GrokInferenceTransportErrorKind::SafetyRejected;
+    let credential_failure = transport_credential_failure(&error, upstream_model);
+    let error = map_stream_error(error);
+    match credential_failure {
+        Some(failure) => record_credential_failure(selector, session, error, failure).await,
+        None if request_scoped => error,
+        None => record_stream_failure(selector, session, error).await,
+    }
 }
 
 fn map_continuation_failure(context: &AttemptContext, error: ProviderError) -> ProviderError {
@@ -1628,6 +1709,11 @@ fn map_selection_error(error: GrokSessionSelectorError) -> ProviderError {
             cooling_down_message(retry_after),
             "account_cooling_down",
         ),
+        GrokSessionSelectorError::ModelCoolingDown { retry_after } => (
+            retry_after,
+            model_cooling_down_message(retry_after),
+            "model_cooling_down",
+        ),
         GrokSessionSelectorError::CapacityUnavailable { retry_after } => (
             retry_after,
             "account is at its concurrency or request-interval limit".to_owned(),
@@ -1665,10 +1751,20 @@ fn map_selection_error(error: GrokSessionSelectorError) -> ProviderError {
 fn cooling_down_message(retry_after: Option<Duration>) -> String {
     match retry_after {
         Some(retry_after) => format!(
-            "account is cooling down after an interrupted upstream stream; retry in {}s",
+            "account is cooling down after an upstream failure; retry in {}s",
             retry_after.as_secs().saturating_add(1)
         ),
-        None => "account is cooling down after an interrupted upstream stream".to_owned(),
+        None => "account is cooling down after an upstream failure".to_owned(),
+    }
+}
+
+fn model_cooling_down_message(retry_after: Option<Duration>) -> String {
+    match retry_after {
+        Some(retry_after) => format!(
+            "all eligible accounts are cooling down for this model; retry in {}s",
+            retry_after.as_secs().saturating_add(1)
+        ),
+        None => "all eligible accounts are cooling down for this model".to_owned(),
     }
 }
 
@@ -1690,13 +1786,19 @@ fn map_transport_error_with_state(
     forced_send_state: Option<UpstreamSendState>,
     allow_explicit_replay: bool,
 ) -> ProviderError {
-    let kind = match error.kind() {
+    let transport_kind = error.kind();
+    let kind = match transport_kind {
         GrokInferenceTransportErrorKind::InvalidRequest => ProviderErrorKind::InvalidRequest,
         GrokInferenceTransportErrorKind::Unsupported => ProviderErrorKind::Unsupported,
         GrokInferenceTransportErrorKind::Unauthorized => ProviderErrorKind::Unauthorized,
-        GrokInferenceTransportErrorKind::PermissionDenied => ProviderErrorKind::PermissionDenied,
+        GrokInferenceTransportErrorKind::PermissionDenied
+        | GrokInferenceTransportErrorKind::ModelAccessDenied
+        | GrokInferenceTransportErrorKind::PaymentRequired
+        | GrokInferenceTransportErrorKind::SafetyRejected => ProviderErrorKind::PermissionDenied,
         GrokInferenceTransportErrorKind::RateLimited => ProviderErrorKind::RateLimited,
-        GrokInferenceTransportErrorKind::QuotaExhausted => ProviderErrorKind::QuotaExhausted,
+        GrokInferenceTransportErrorKind::QuotaExhausted
+        | GrokInferenceTransportErrorKind::FreeQuotaExhausted
+        | GrokInferenceTransportErrorKind::ModelQuotaExhausted => ProviderErrorKind::QuotaExhausted,
         GrokInferenceTransportErrorKind::Timeout => ProviderErrorKind::Timeout,
         GrokInferenceTransportErrorKind::Transport => ProviderErrorKind::Transport,
         GrokInferenceTransportErrorKind::Protocol => ProviderErrorKind::Protocol,
@@ -1711,7 +1813,7 @@ fn map_transport_error_with_state(
         mapped = mapped.with_status(status);
         if allow_explicit_replay
             && forced_send_state.is_none()
-            && explicit_rejection_is_replay_safe(kind, status)
+            && explicit_rejection_is_replay_safe(transport_kind, status)
         {
             mapped = mapped.with_replay_safe();
         }
@@ -1725,6 +1827,9 @@ fn map_transport_error_with_state(
     if let Some(code) = error.upstream_code().cloned() {
         mapped = mapped.with_upstream_code(code);
     }
+    if let Some(detail) = error.client_visible_upstream_error().cloned() {
+        mapped = mapped.with_client_visible_upstream_error(detail);
+    }
     if error.requires_credential_recovery() {
         mapped = mapped.with_credential_recovery().with_replay_safe();
     }
@@ -1734,14 +1839,25 @@ fn map_transport_error_with_state(
     mapped
 }
 
-fn explicit_rejection_is_replay_safe(kind: ProviderErrorKind, status: u16) -> bool {
+fn explicit_rejection_is_replay_safe(kind: GrokInferenceTransportErrorKind, status: u16) -> bool {
     matches!(
         (kind, status),
-        (ProviderErrorKind::Unauthorized, 401)
-            | (ProviderErrorKind::QuotaExhausted, 402)
-            | (ProviderErrorKind::QuotaExhausted, 403)
-            | (ProviderErrorKind::QuotaExhausted, 429)
-            | (ProviderErrorKind::RateLimited, 429)
+        (GrokInferenceTransportErrorKind::Unauthorized, 401)
+            | (
+                GrokInferenceTransportErrorKind::QuotaExhausted,
+                402 | 403 | 429
+            )
+            | (
+                GrokInferenceTransportErrorKind::FreeQuotaExhausted,
+                402 | 403 | 429
+            )
+            | (
+                GrokInferenceTransportErrorKind::ModelQuotaExhausted,
+                402 | 403 | 429
+            )
+            | (GrokInferenceTransportErrorKind::PaymentRequired, 402)
+            | (GrokInferenceTransportErrorKind::ModelAccessDenied, 403)
+            | (GrokInferenceTransportErrorKind::RateLimited, 429)
     )
 }
 

backend/crates/providers/xai/src/transport/canonical.rs

@@ -19,6 +19,7 @@ use gateway_protocol::openai::sse::{SseEvent, SseEventDecoder};
 use serde_json::Value;
 
 use super::request::{GrokResponseTransform, GrokResponsesRequest};
+use super::scrub_account_fingerprints;
 
 const CONTENTS_PER_OUTPUT: u32 = 1_024;
 const LONG_CONTEXT_THRESHOLD: u64 = 200_000;
@@ -820,53 +821,6 @@ fn upstream_error_field<'a>(value: &'a Value, field: &str) -> Option<&'a str> {
         .and_then(Value::as_str)
 }
 
-const UUID_TEXT_LEN: usize = 36;
-
-/// 把 message 中的 UUID 替换为占位符，控制字符归一为空格，
-/// 使文本满足客户端可见错误的约束且不携带可定位账号的标识。
-fn scrub_account_fingerprints(message: &str) -> String {
-    let bytes = message.as_bytes();
-    let mut scrubbed = String::with_capacity(message.len());
-    let mut index = 0;
-    while index < message.len() {
-        if uuid_at(bytes, index) {
-            scrubbed.push_str("[redacted]");
-            index += UUID_TEXT_LEN;
-            continue;
-        }
-        let character = message[index..].chars().next().expect("char boundary");
-        scrubbed.push(if character.is_control() {
-            ' '
-        } else {
-            character
-        });
-        index += character.len_utf8();
-    }
-    scrubbed
-}
-
-fn uuid_at(bytes: &[u8], index: usize) -> bool {
-    let Some(candidate) = bytes.get(index..index + UUID_TEXT_LEN) else {
-        return false;
-    };
-    if index > 0 && bytes[index - 1].is_ascii_alphanumeric() {
-        return false;
-    }
-    if bytes
-        .get(index + UUID_TEXT_LEN)
-        .is_some_and(u8::is_ascii_alphanumeric)
-    {
-        return false;
-    }
-    candidate.iter().enumerate().all(|(offset, byte)| {
-        if matches!(offset, 8 | 13 | 18 | 23) {
-            *byte == b'-'
-        } else {
-            byte.is_ascii_hexdigit()
-        }
-    })
-}
-
 fn protocol_error(_error: impl std::fmt::Debug) -> ProviderError {
     protocol_error_marker()
 }

backend/crates/providers/xai/src/transport/mod.rs

@@ -54,3 +54,50 @@ pub use transport::{
     GrokInferenceTransportError, GrokInferenceTransportErrorKind, GrokInferenceTransportFuture,
     GrokInferenceTransportMetrics,
 };
+
+const UUID_TEXT_LEN: usize = 36;
+
+/// 把 message 中的 UUID 替换为占位符，控制字符归一为空格，
+/// 使文本满足客户端可见错误的约束且不携带可定位账号的标识。
+fn scrub_account_fingerprints(message: &str) -> String {
+    let bytes = message.as_bytes();
+    let mut scrubbed = String::with_capacity(message.len());
+    let mut index = 0;
+    while index < message.len() {
+        if uuid_at(bytes, index) {
+            scrubbed.push_str("[redacted]");
+            index += UUID_TEXT_LEN;
+            continue;
+        }
+        let character = message[index..].chars().next().expect("char boundary");
+        scrubbed.push(if character.is_control() {
+            ' '
+        } else {
+            character
+        });
+        index += character.len_utf8();
+    }
+    scrubbed
+}
+
+fn uuid_at(bytes: &[u8], index: usize) -> bool {
+    let Some(candidate) = bytes.get(index..index + UUID_TEXT_LEN) else {
+        return false;
+    };
+    if index > 0 && bytes[index - 1].is_ascii_alphanumeric() {
+        return false;
+    }
+    if bytes
+        .get(index + UUID_TEXT_LEN)
+        .is_some_and(u8::is_ascii_alphanumeric)
+    {
+        return false;
+    }
+    candidate.iter().enumerate().all(|(offset, byte)| {
+        if matches!(offset, 8 | 13 | 18 | 23) {
+            *byte == b'-'
+        } else {
+            byte.is_ascii_hexdigit()
+        }
+    })
+}

backend/crates/providers/xai/src/transport/network.rs

@@ -30,9 +30,11 @@ use crate::{
     OAuthHttpResponse, OAuthHttpTransport, TransportFailure, TransportFailureKind, TransportFuture,
 };
 use gateway_core::engine::UpstreamSendState;
-use gateway_core::error::SafeUpstreamValue;
+use gateway_core::error::{ClientVisibleUpstreamError, SafeUpstreamValue};
 use gateway_core::event::UpstreamHttpVersion;
 
+use super::scrub_account_fingerprints;
+
 pub(crate) const OFFICIAL_OAUTH_HOST: &str = "auth.x.ai";
 const OFFICIAL_INFERENCE_HOST: &str = "cli-chat-proxy.grok.com";
 const OFFICIAL_INFERENCE_PATH: &str = "/v1/responses";
@@ -990,29 +992,22 @@ async fn classify_inference_status(response: Response) -> GrokInferenceTransport
         Ok(BoundedBody::TooLarge) | Err(_) => Vec::new(),
     };
     let metadata = inference_error_metadata(&body);
-    let credential_recovery_required = status == StatusCode::UNAUTHORIZED
-        || (status == StatusCode::FORBIDDEN && forbidden_requires_recovery(&metadata));
     let kind = match status {
         StatusCode::BAD_REQUEST
         | StatusCode::NOT_FOUND
         | StatusCode::CONFLICT
         | StatusCode::UNPROCESSABLE_ENTITY => GrokInferenceTransportErrorKind::InvalidRequest,
         StatusCode::UNAUTHORIZED => GrokInferenceTransportErrorKind::Unauthorized,
-        StatusCode::PAYMENT_REQUIRED => GrokInferenceTransportErrorKind::QuotaExhausted,
-        StatusCode::FORBIDDEN => classify_forbidden(&metadata),
+        StatusCode::PAYMENT_REQUIRED => classify_payment_required(&metadata),
+        StatusCode::FORBIDDEN => classify_forbidden(&metadata, &body),
         StatusCode::REQUEST_TIMEOUT | StatusCode::GATEWAY_TIMEOUT => {
             GrokInferenceTransportErrorKind::Timeout
         }
-        StatusCode::TOO_MANY_REQUESTS => {
-            if quota_exhausted(&metadata) {
-                GrokInferenceTransportErrorKind::QuotaExhausted
-            } else {
-                GrokInferenceTransportErrorKind::RateLimited
-            }
-        }
+        StatusCode::TOO_MANY_REQUESTS => classify_too_many_requests(&metadata),
         status if status.is_server_error() => GrokInferenceTransportErrorKind::Unavailable,
         _ => GrokInferenceTransportErrorKind::Protocol,
     };
+    let credential_recovery_required = kind == GrokInferenceTransportErrorKind::Unauthorized;
     let mut error = GrokInferenceTransportError::new(kind, UpstreamSendState::Sent)
         .with_status(status_code)
         .with_response_facts(http_version, request_id)
@@ -1022,6 +1017,15 @@ async fn classify_inference_status(response: Response) -> GrokInferenceTransport
     } else {
         metadata.code.as_deref().and_then(normalize_failure_code)
     };
+    if let Some(message) = metadata.client_message.as_deref()
+        && let Ok(detail) = ClientVisibleUpstreamError::new(
+            scrub_account_fingerprints(message),
+            upstream_code.clone(),
+            metadata.error_type.clone(),
+        )
+    {
+        error = error.with_client_visible_upstream_error(detail);
+    }
     if let Some(code) = upstream_code.and_then(|code| SafeUpstreamValue::new(code).ok()) {
         error = error.with_upstream_code(code);
     }
@@ -1039,6 +1043,8 @@ struct InferenceErrorMetadata {
     code: Option<String>,
     error_type: Option<String>,
     message: Option<String>,
+    // 仅结构化 JSON 的 message 可进入客户端协议；纯文本正文只参与内部分类。
+    client_message: Option<String>,
 }
 
 fn inference_error_metadata(body: &[u8]) -> InferenceErrorMetadata {
@@ -1053,6 +1059,9 @@ fn inference_error_metadata(body: &[u8]) -> InferenceErrorMetadata {
         return InferenceErrorMetadata::default();
     };
     let nested = root.get("error").and_then(Value::as_object);
+    let message = nested
+        .and_then(|error| first_string(error, &["message", "error"]))
+        .or_else(|| first_string(root, &["message", "error"]));
     InferenceErrorMetadata {
         code: nested
             .and_then(|error| first_string(error, &["code", "error_code"]))
@@ -1060,9 +1069,8 @@ fn inference_error_metadata(body: &[u8]) -> InferenceErrorMetadata {
         error_type: nested
             .and_then(|error| first_string(error, &["type", "error_type"]))
             .or_else(|| first_string(root, &["type", "error_type"])),
-        message: nested
-            .and_then(|error| first_string(error, &["message", "error"]))
-            .or_else(|| first_string(root, &["message", "error"])),
+        client_message: message.clone(),
+        message,
     }
 }
 
@@ -1075,22 +1083,53 @@ fn first_string(object: &serde_json::Map<String, Value>, fields: &[&str]) -> Opt
         .map(ToOwned::to_owned)
 }
 
-fn classify_forbidden(metadata: &InferenceErrorMetadata) -> GrokInferenceTransportErrorKind {
-    let text = inference_metadata_text(metadata);
-    if credential_rejected(&text) {
-        GrokInferenceTransportErrorKind::Unauthorized
-    } else if quota_exhausted(metadata) {
+fn classify_payment_required(metadata: &InferenceErrorMetadata) -> GrokInferenceTransportErrorKind {
+    if model_quota_exhausted(metadata) {
+        GrokInferenceTransportErrorKind::ModelQuotaExhausted
+    } else if paid_quota_exhausted(metadata) {
         GrokInferenceTransportErrorKind::QuotaExhausted
+    } else if free_quota_exhausted(metadata) {
+        GrokInferenceTransportErrorKind::FreeQuotaExhausted
+    } else {
+        GrokInferenceTransportErrorKind::PaymentRequired
+    }
+}
+
+fn classify_forbidden(
+    metadata: &InferenceErrorMetadata,
+    body: &[u8],
+) -> GrokInferenceTransportErrorKind {
+    let text = inference_metadata_text(metadata);
+    let body_text = String::from_utf8_lossy(body).to_ascii_lowercase();
+    if safety_rejected(&text) || safety_rejected(&body_text) {
+        GrokInferenceTransportErrorKind::SafetyRejected
+    } else if credential_rejected(&text) {
+        GrokInferenceTransportErrorKind::Unauthorized
+    } else if model_quota_exhausted(metadata) {
+        GrokInferenceTransportErrorKind::ModelQuotaExhausted
+    } else if paid_quota_exhausted(metadata) {
+        GrokInferenceTransportErrorKind::QuotaExhausted
+    } else if free_quota_exhausted(metadata) {
+        GrokInferenceTransportErrorKind::FreeQuotaExhausted
+    } else if model_access_denied(metadata) {
+        GrokInferenceTransportErrorKind::ModelAccessDenied
     } else {
         GrokInferenceTransportErrorKind::PermissionDenied
     }
 }
 
-fn forbidden_requires_recovery(metadata: &InferenceErrorMetadata) -> bool {
-    let text = inference_metadata_text(metadata);
-    credential_rejected(&text)
-        || text.contains("access to the chat endpoint is denied")
-        || text.trim_matches([' ', '.', '!', '\t', '\r', '\n']) == "access denied"
+fn classify_too_many_requests(
+    metadata: &InferenceErrorMetadata,
+) -> GrokInferenceTransportErrorKind {
+    if model_quota_exhausted(metadata) {
+        GrokInferenceTransportErrorKind::ModelQuotaExhausted
+    } else if paid_quota_exhausted(metadata) {
+        GrokInferenceTransportErrorKind::QuotaExhausted
+    } else if free_quota_exhausted(metadata) {
+        GrokInferenceTransportErrorKind::FreeQuotaExhausted
+    } else {
+        GrokInferenceTransportErrorKind::RateLimited
+    }
 }
 
 fn reasoning_decode_failed(metadata: &InferenceErrorMetadata) -> bool {
@@ -1100,7 +1139,17 @@ fn reasoning_decode_failed(metadata: &InferenceErrorMetadata) -> bool {
     })
 }
 
-fn quota_exhausted(metadata: &InferenceErrorMetadata) -> bool {
+fn paid_quota_exhausted(metadata: &InferenceErrorMetadata) -> bool {
+    contains_any(
+        &inference_metadata_text(metadata),
+        &[
+            "personal-team-blocked:spending-limit",
+            "personal_team_blocked_spending_limit",
+        ],
+    )
+}
+
+fn free_quota_exhausted(metadata: &InferenceErrorMetadata) -> bool {
     contains_any(
         &inference_metadata_text(metadata),
         &[
@@ -1110,12 +1159,29 @@ fn quota_exhausted(metadata: &InferenceErrorMetadata) -> bool {
             "free_usage_exhausted",
             "used all the included free usage",
             "used all your free usage",
-            "personal-team-blocked:spending-limit",
-            "personal_team_blocked_spending_limit",
         ],
     )
 }
 
+fn model_quota_exhausted(metadata: &InferenceErrorMetadata) -> bool {
+    inference_metadata_text(metadata).contains("used all the included free usage for model")
+}
+
+fn model_access_denied(metadata: &InferenceErrorMetadata) -> bool {
+    metadata.message.as_deref().is_some_and(|message| {
+        let message = message.to_ascii_lowercase();
+        message.contains("access to the chat endpoint is denied")
+            || message.trim_matches([' ', '.', '!', '\t', '\r', '\n']) == "access denied"
+    })
+}
+
+fn safety_rejected(text: &str) -> bool {
+    contains_any(
+        text,
+        &["content violates usage guidelines", "safety_check_type_"],
+    )
+}
+
 fn inference_metadata_text(metadata: &InferenceErrorMetadata) -> String {
     [
         metadata.code.as_deref().unwrap_or_default(),

backend/crates/providers/xai/src/transport/session.rs

@@ -300,8 +300,8 @@ pub type GrokSessionSelectorFuture<'a> = Pin<
     Box<dyn Future<Output = Result<SelectedGrokSession, GrokSessionSelectorError>> + Send + 'a>,
 >;
 
-/// 凭据维度的上游失败，可安全落库为可用性反馈。
-#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+/// 已脱敏的上游失败反馈；selector 决定账号持久状态或细粒度运行时 cooldown。
+#[derive(Debug, Clone, PartialEq, Eq)]
 pub enum GrokCredentialFailure {
     /// 选中的 OAuth token 被拒绝。
     Unauthorized,
@@ -310,8 +310,25 @@ pub enum GrokCredentialFailure {
         /// 已解析并脱敏的延迟；不保留上游原始 header。
         retry_after: Option<Duration>,
     },
-    /// 选中的 OAuth 账号配额耗尽。
+    /// 选中的 OAuth 账号付费额度耗尽。
     QuotaExhausted,
+    /// 选中的 OAuth 账号订阅级免费额度耗尽。
+    FreeQuotaExhausted,
+    /// HTTP 402 没有给出可证明账号额度耗尽的稳定信号。
+    PaymentRequired {
+        /// 已解析并脱敏的延迟；缺失时由 selector 使用短暂默认值。
+        retry_after: Option<Duration>,
+    },
+    /// 选中账号只耗尽了指定模型的额度。
+    ModelQuotaExhausted {
+        upstream_model: UpstreamModelId,
+        retry_after: Option<Duration>,
+    },
+    /// 选中账号只缺少指定模型的访问权限。
+    ModelAccessDenied {
+        upstream_model: UpstreamModelId,
+        retry_after: Option<Duration>,
+    },
     /// 成功的 SSE 响应在终止 Responses 事件前结束。
     StreamInterrupted,
 }
@@ -325,7 +342,7 @@ pub trait GrokSessionSelector: Send + Sync {
     /// 执行一次选择，不做 Provider 内部回退。
     fn select(&self, request: GrokSessionSelection) -> GrokSessionSelectorFuture<'_>;
 
-    /// 持久化一条已分类的失败，不重试也不替换原错误。
+    /// 记录一条已分类的失败反馈，不重试也不替换原错误。
     fn record_failure<'a>(
         &'a self,
         session: &'a SelectedGrokSession,
@@ -361,6 +378,12 @@ pub enum GrokSessionSelectorError {
         /// 最早恢复的会话剩余冷却时长。
         retry_after: Option<Duration>,
     },
+    /// 每个可用会话都只针对当前模型处于运行时冷却窗口内。
+    #[error("Grok Build model is cooling down for the available accounts")]
+    ModelCoolingDown {
+        /// 最早恢复的账号+模型组合剩余冷却时长。
+        retry_after: Option<Duration>,
+    },
     /// 会话元数据或 Provider 持有的明文密钥非法。
     #[error("Grok Build session data is invalid")]
     InvalidSession,

backend/crates/providers/xai/src/transport/transport.rs

@@ -5,7 +5,7 @@ use std::time::Duration;
 
 use futures::Stream;
 use gateway_core::engine::UpstreamSendState;
-use gateway_core::error::SafeUpstreamValue;
+use gateway_core::error::{ClientVisibleUpstreamError, SafeUpstreamValue};
 use gateway_core::event::UpstreamHttpVersion;
 use url::Url;
 use zeroize::Zeroizing;
@@ -278,8 +278,18 @@ pub enum GrokInferenceTransportErrorKind {
     PermissionDenied,
     /// 会话级或 Provider 级限流。
     RateLimited,
-    /// 账号额度或配额耗尽。
+    /// 账号付费额度或 spending limit 耗尽。
     QuotaExhausted,
+    /// 账号订阅级免费额度耗尽。
+    FreeQuotaExhausted,
+    /// 当前账号只耗尽了指定模型的免费额度。
+    ModelQuotaExhausted,
+    /// 当前账号只缺少指定模型的访问权限。
+    ModelAccessDenied,
+    /// 上游要求付费，但没有给出可证明账号额度耗尽的稳定信号。
+    PaymentRequired,
+    /// 当前请求被内容安全策略拒绝，不代表账号或模型不可用。
+    SafetyRejected,
     /// 截止时间或 transport 超时。
     Timeout,
     /// 网络/TLS/连接失败。
@@ -301,7 +311,8 @@ pub struct GrokInferenceTransportError {
     retry_after: Option<Duration>,
     http_version: Option<UpstreamHttpVersion>,
     request_id: Option<SafeUpstreamValue>,
-    upstream_code: Option<SafeUpstreamValue>,
+    upstream_code: Option<Box<SafeUpstreamValue>>,
+    client_visible_upstream_error: Option<Box<ClientVisibleUpstreamError>>,
     transport_metrics: GrokInferenceTransportMetrics,
     credential_recovery_required: bool,
     sensitive_context_redacted: bool,
@@ -319,6 +330,7 @@ impl GrokInferenceTransportError {
             http_version: None,
             request_id: None,
             upstream_code: None,
+            client_visible_upstream_error: None,
             transport_metrics: GrokInferenceTransportMetrics {
                 headers_ms: None,
                 client_cache_status: None,
@@ -359,7 +371,14 @@ impl GrokInferenceTransportError {
     /// 附着从错误 JSON 中提取并清洗后的稳定机器码。
     #[must_use]
     pub fn with_upstream_code(mut self, code: SafeUpstreamValue) -> Self {
-        self.upstream_code = Some(code);
+        self.upstream_code = Some(Box::new(code));
+        self
+    }
+
+    /// 附着仅供原客户端协议展示的结构化上游错误。
+    #[must_use]
+    pub fn with_client_visible_upstream_error(mut self, error: ClientVisibleUpstreamError) -> Self {
+        self.client_visible_upstream_error = Some(Box::new(error));
         self
     }
 
@@ -420,8 +439,13 @@ impl GrokInferenceTransportError {
     }
 
     #[must_use]
-    pub const fn upstream_code(&self) -> Option<&SafeUpstreamValue> {
-        self.upstream_code.as_ref()
+    pub fn upstream_code(&self) -> Option<&SafeUpstreamValue> {
+        self.upstream_code.as_deref()
+    }
+
+    #[must_use]
+    pub fn client_visible_upstream_error(&self) -> Option<&ClientVisibleUpstreamError> {
+        self.client_visible_upstream_error.as_deref()
     }
 
     #[must_use]
@@ -452,6 +476,13 @@ impl fmt::Debug for GrokInferenceTransportError {
             .field("http_version", &self.http_version)
             .field("request_id", &self.request_id)
             .field("upstream_code", &self.upstream_code)
+            .field(
+                "client_visible_upstream_error",
+                &self
+                    .client_visible_upstream_error
+                    .as_ref()
+                    .map(|_| "<present>"),
+            )
             .field("transport_metrics", &self.transport_metrics)
             .field(
                 "credential_recovery_required",

backend/crates/providers/xai/tests/admin.rs

@@ -33,9 +33,9 @@ use gateway_core::provider_ports::{
     NewOAuthPendingFlow, OAuthPendingBinding, OAuthPendingClaimOutcome, OAuthPendingConsumeOutcome,
     OAuthPendingFlowPort, OAuthPendingPutOutcome, OAuthPendingReleaseOutcome,
     ProviderCatalogCacheKey, ProviderCatalogCachePort, ProviderCooldown, ProviderCooldownPort,
-    ProviderCredentialState, ProviderCredentialStatePort, ProviderLeaseAcquisition,
-    ProviderLeasePort, ProviderLeaseRequest, ProviderRefreshPolicy, ProviderRuntimePolicyPort,
-    ProviderStoreError, ProviderStorePorts,
+    ProviderCooldownScope, ProviderCredentialState, ProviderCredentialStatePort,
+    ProviderLeaseAcquisition, ProviderLeasePort, ProviderLeaseRequest, ProviderRefreshPolicy,
+    ProviderRuntimePolicyPort, ProviderScopedCooldown, ProviderStoreError, ProviderStorePorts,
 };
 use gateway_core::routing::{
     ConfigRevision, ModelCapabilities, ProviderKind, ProviderModel, PublicModelId, RoutingContext,
@@ -796,6 +796,30 @@ impl ProviderCooldownPort for TestCooldown {
     ) -> BoxFuture<'a, Result<bool, ProviderStoreError>> {
         Box::pin(async { Ok(false) })
     }
+
+    fn put_scoped_if_later(
+        &self,
+        _cooldown: ProviderScopedCooldown,
+    ) -> BoxFuture<'_, Result<bool, ProviderStoreError>> {
+        Box::pin(async { Ok(false) })
+    }
+
+    fn read_scoped<'a>(
+        &'a self,
+        _account_id: &'a ProviderAccountId,
+        _scope: &'a ProviderCooldownScope,
+    ) -> BoxFuture<'a, Result<Option<ProviderScopedCooldown>, ProviderStoreError>> {
+        Box::pin(async { Ok(None) })
+    }
+
+    fn clear_scoped<'a>(
+        &'a self,
+        _account_id: &'a ProviderAccountId,
+        _scope: &'a ProviderCooldownScope,
+        _through_revision: CredentialRevision,
+    ) -> BoxFuture<'a, Result<bool, ProviderStoreError>> {
+        Box::pin(async { Ok(false) })
+    }
 }
 
 struct TestRuntimePolicy;

backend/crates/providers/xai/tests/credential/selector.rs

@@ -10,9 +10,10 @@ use gateway_core::engine::credential::{
     QuotaWriteOutcome, RotationStrategy,
 };
 use gateway_core::provider_ports::{
-    ProviderLeaseAcquisition, ProviderLeasePort, ProviderLeaseRequest, ProviderSchedulingState,
-    ProviderStoreError,
+    ProviderCooldownScope, ProviderLeaseAcquisition, ProviderLeasePort, ProviderLeaseRequest,
+    ProviderSchedulingState, ProviderStoreError,
 };
+use gateway_core::routing::UpstreamModelId;
 use provider_xai::{
     GrokAccountSessionSelector, GrokBillingRequest, GrokBillingTransport,
     GrokBillingTransportError, GrokBillingTransportErrorKind, GrokBillingTransportFuture,
@@ -119,7 +120,8 @@ impl SelectorFixture {
                 .replace(GrokPlanCatalog::new(
                     GrokCatalogScope::for_account(&account).expect("catalog scope"),
                     Utc::now(),
-                    GrokCredentialCatalogSeed::new(["grok-4.5"], None).expect("catalog"),
+                    GrokCredentialCatalogSeed::new(["grok-4.5", "grok-4.6"], None)
+                        .expect("catalog"),
                 ))
                 .await
                 .expect("cache catalog");
@@ -183,9 +185,32 @@ impl SelectorFixture {
         excluded: BTreeSet<ProviderAccountId>,
         required_account: Option<ProviderAccountId>,
         strategy: RotationStrategy,
+    ) -> GrokSessionSelection {
+        self.request_for_model_with_policy("grok-4.5", excluded, required_account, strategy)
+    }
+
+    fn request_for_model(
+        &self,
+        upstream_model: &str,
+        required_account: Option<ProviderAccountId>,
+    ) -> GrokSessionSelection {
+        self.request_for_model_with_policy(
+            upstream_model,
+            BTreeSet::new(),
+            required_account,
+            RotationStrategy::Smart,
+        )
+    }
+
+    fn request_for_model_with_policy(
+        &self,
+        upstream_model: &str,
+        excluded: BTreeSet<ProviderAccountId>,
+        required_account: Option<ProviderAccountId>,
+        strategy: RotationStrategy,
     ) -> GrokSessionSelection {
         GrokSessionSelection::new(
-            gateway_core::routing::UpstreamModelId::new("grok-4.5").expect("model"),
+            UpstreamModelId::new(upstream_model).expect("model"),
             excluded,
             required_account,
             AccountSelectionPolicy::new(
@@ -387,6 +412,131 @@ async fn rate_limit_feedback_records_runtime_cooldown_without_persisting_account
     assert_ne!(next.account_id(), &selected);
 }
 
+#[tokio::test]
+async fn unknown_payment_required_feedback_only_records_runtime_account_cooldown() {
+    let fixture = SelectorFixture::new(&["payment-required", "available"]).await;
+    let session = fixture
+        .selector
+        .select(fixture.request(BTreeSet::new()))
+        .await
+        .expect("session");
+    let selected = session.account_id().clone();
+    fixture
+        .selector
+        .record_failure(
+            &session,
+            GrokCredentialFailure::PaymentRequired {
+                retry_after: Some(Duration::from_secs(5)),
+            },
+        )
+        .await;
+
+    assert_eq!(
+        fixture
+            .store
+            .account(&selected)
+            .expect("selected account")
+            .availability(),
+        AccountAvailability::Ready
+    );
+    assert!(fixture.cooldowns.cooldown(&selected).is_some());
+    let next = fixture
+        .selector
+        .select(fixture.request(BTreeSet::new()))
+        .await
+        .expect("another account remains available");
+    assert_ne!(next.account_id(), &selected);
+}
+
+#[tokio::test]
+async fn model_quota_feedback_blocks_only_the_failed_account_and_model() {
+    let fixture = SelectorFixture::new(&["model-quota", "available"]).await;
+    let failed_model = UpstreamModelId::new("grok-4.5").expect("failed model");
+    let session = fixture
+        .selector
+        .select(fixture.request_for_model(failed_model.as_str(), None))
+        .await
+        .expect("session");
+    let selected = session.account_id().clone();
+    fixture
+        .selector
+        .record_failure(
+            &session,
+            GrokCredentialFailure::ModelQuotaExhausted {
+                upstream_model: failed_model.clone(),
+                retry_after: Some(Duration::from_secs(5)),
+            },
+        )
+        .await;
+
+    assert_eq!(
+        fixture
+            .store
+            .account(&selected)
+            .expect("selected account")
+            .availability(),
+        AccountAvailability::Ready
+    );
+    let scope = ProviderCooldownScope::upstream_model(failed_model);
+    assert!(
+        fixture
+            .cooldowns
+            .scoped_cooldown(&selected, &scope)
+            .is_some()
+    );
+    let unrelated_model = fixture
+        .selector
+        .select(fixture.request_for_model("grok-4.6", Some(selected.clone())))
+        .await
+        .expect("same account remains valid for another model");
+    assert_eq!(unrelated_model.account_id(), &selected);
+    let failed_model_retry = fixture
+        .selector
+        .select(fixture.request_for_model("grok-4.5", None))
+        .await
+        .expect("another account serves the failed model");
+    assert_ne!(failed_model_retry.account_id(), &selected);
+}
+
+#[tokio::test]
+async fn model_access_feedback_reports_model_cooldown_without_blocking_the_account() {
+    let fixture = SelectorFixture::new(&["model-access"]).await;
+    let model = UpstreamModelId::new("grok-4.5").expect("model");
+    let session = fixture
+        .selector
+        .select(fixture.request_for_model(model.as_str(), None))
+        .await
+        .expect("session");
+    fixture
+        .selector
+        .record_failure(
+            &session,
+            GrokCredentialFailure::ModelAccessDenied {
+                upstream_model: model,
+                retry_after: None,
+            },
+        )
+        .await;
+
+    assert_eq!(
+        fixture
+            .store
+            .account(session.account_id())
+            .expect("account")
+            .availability(),
+        AccountAvailability::Ready
+    );
+    assert!(matches!(
+        fixture
+            .selector
+            .select(fixture.request_for_model("grok-4.5", None))
+            .await,
+        Err(GrokSessionSelectorError::ModelCoolingDown {
+            retry_after: Some(_)
+        })
+    ));
+}
+
 #[tokio::test]
 async fn interrupted_stream_feedback_records_runtime_cooldown_without_persisting_account_state() {
     let fixture = SelectorFixture::new(&["stream-interrupted"]).await;

backend/crates/providers/xai/tests/provider/contract.rs

@@ -18,7 +18,8 @@ use gateway_core::engine::{
     UpstreamSendState,
 };
 use gateway_core::error::{
-    ContinuationFailure, ProviderError, ProviderErrorKind, SafeUpstreamValue,
+    ClientVisibleUpstreamError, ContinuationFailure, ProviderError, ProviderErrorKind,
+    SafeUpstreamValue,
 };
 use gateway_core::event::{GatewayEvent, UpstreamHttpVersion};
 use gateway_core::operation::{
@@ -1295,6 +1296,157 @@ async fn explicit_quota_http_429_marks_provider_error_replay_safe() {
     assert!(error.replay_is_safe());
 }
 
+#[tokio::test]
+async fn unknown_http_402_is_retryable_without_account_quota_feedback() {
+    let selector = StubSelector::success();
+    let transport = StubInferenceTransport::error(
+        GrokInferenceTransportError::new(
+            GrokInferenceTransportErrorKind::PaymentRequired,
+            UpstreamSendState::Sent,
+        )
+        .with_status(402),
+    );
+    let provider = provider(selector.clone(), transport).await;
+    let mut stream = provider
+        .execute(
+            provider_request("xai"),
+            context(CancellationToken::new(), None),
+        )
+        .await
+        .expect("stream");
+
+    let error = next_provider_error(&mut stream).await;
+
+    assert_eq!(error.kind(), ProviderErrorKind::PermissionDenied);
+    assert!(error.replay_is_safe());
+    assert_eq!(
+        selector.feedback.lock().expect("feedback").as_slice(),
+        &[GrokCredentialFailure::PaymentRequired { retry_after: None }]
+    );
+}
+
+#[tokio::test]
+async fn model_quota_transport_failure_preserves_model_scoped_feedback() {
+    let selector = StubSelector::success();
+    let transport = StubInferenceTransport::error(
+        GrokInferenceTransportError::new(
+            GrokInferenceTransportErrorKind::ModelQuotaExhausted,
+            UpstreamSendState::Sent,
+        )
+        .with_status(403),
+    );
+    let provider = provider(selector.clone(), transport).await;
+    let mut stream = provider
+        .execute(
+            provider_request("xai"),
+            context(CancellationToken::new(), None),
+        )
+        .await
+        .expect("stream");
+
+    let error = next_provider_error(&mut stream).await;
+
+    assert_eq!(error.kind(), ProviderErrorKind::QuotaExhausted);
+    assert!(error.replay_is_safe());
+    assert_eq!(
+        selector.feedback.lock().expect("feedback").as_slice(),
+        &[GrokCredentialFailure::ModelQuotaExhausted {
+            upstream_model: UpstreamModelId::new("grok-4.5").expect("model"),
+            retry_after: None,
+        }]
+    );
+}
+
+#[tokio::test]
+async fn model_access_denial_is_retryable_without_credential_recovery() {
+    let selector = StubSelector::success();
+    let transport = StubInferenceTransport::error(
+        GrokInferenceTransportError::new(
+            GrokInferenceTransportErrorKind::ModelAccessDenied,
+            UpstreamSendState::Sent,
+        )
+        .with_status(403),
+    );
+    let provider = provider(selector.clone(), transport).await;
+    let mut stream = provider
+        .execute(
+            provider_request("xai"),
+            context(CancellationToken::new(), None),
+        )
+        .await
+        .expect("stream");
+
+    let error = next_provider_error(&mut stream).await;
+
+    assert_eq!(error.kind(), ProviderErrorKind::PermissionDenied);
+    assert!(error.replay_is_safe());
+    assert_eq!(
+        selector.feedback.lock().expect("feedback").as_slice(),
+        &[GrokCredentialFailure::ModelAccessDenied {
+            upstream_model: UpstreamModelId::new("grok-4.5").expect("model"),
+            retry_after: None,
+        }]
+    );
+}
+
+#[tokio::test]
+async fn safety_rejection_does_not_rotate_or_mutate_account_state() {
+    let selector = StubSelector::success();
+    let transport = StubInferenceTransport::error(
+        GrokInferenceTransportError::new(
+            GrokInferenceTransportErrorKind::SafetyRejected,
+            UpstreamSendState::Sent,
+        )
+        .with_status(403),
+    );
+    let provider = provider(selector.clone(), transport).await;
+    let mut stream = provider
+        .execute(
+            provider_request("xai"),
+            context(CancellationToken::new(), None),
+        )
+        .await
+        .expect("stream");
+
+    let error = next_provider_error(&mut stream).await;
+
+    assert_eq!(error.kind(), ProviderErrorKind::PermissionDenied);
+    assert!(!error.replay_is_safe());
+    assert!(selector.feedback.lock().expect("feedback").is_empty());
+}
+
+#[tokio::test]
+async fn transport_error_should_preserve_client_visible_upstream_detail() {
+    let detail = ClientVisibleUpstreamError::new(
+        "You have run out of credits",
+        Some("personal_team_blocked_spending_limit".to_owned()),
+        Some("insufficient_quota".to_owned()),
+    )
+    .expect("safe upstream detail");
+    let error = mapped_transport_error(
+        GrokInferenceTransportError::new(
+            GrokInferenceTransportErrorKind::QuotaExhausted,
+            UpstreamSendState::Sent,
+        )
+        .with_status(402)
+        .with_client_visible_upstream_error(detail),
+        false,
+    )
+    .await;
+    let detail = error
+        .client_visible_upstream_error()
+        .expect("mapped client-visible detail");
+
+    assert_eq!(
+        (detail.message(), detail.code(), detail.error_type()),
+        (
+            "You have run out of credits",
+            Some("personal_team_blocked_spending_limit"),
+            Some("insufficient_quota"),
+        )
+    );
+}
+
 #[tokio::test]
 async fn explicit_http_408_does_not_mark_provider_error_replay_safe() {
     let error = mapped_transport_error(

backend/crates/providers/xai/tests/support.rs

@@ -15,9 +15,10 @@ use gateway_core::engine::credential::{
 };
 use gateway_core::error::{StoreError, StoreErrorKind};
 use gateway_core::provider_ports::{
-    ProviderCooldown, ProviderCooldownPort, ProviderRefreshPolicy, ProviderRuntimePolicyPort,
-    ProviderSessionAffinityKey, ProviderSessionAffinityPort, ProviderSessionExclusionPort,
-    ProviderSessionExclusions, ProviderStoreError,
+    ProviderCooldown, ProviderCooldownPort, ProviderCooldownScope, ProviderRefreshPolicy,
+    ProviderRuntimePolicyPort, ProviderScopedCooldown, ProviderSessionAffinityKey,
+    ProviderSessionAffinityPort, ProviderSessionExclusionPort, ProviderSessionExclusions,
+    ProviderStoreError,
 };
 use gateway_core::routing::ProviderKind;
 use provider_xai::{
@@ -561,12 +562,24 @@ pub struct TestSessionExclusions;
 #[derive(Default)]
 pub struct MemoryCooldownPort {
     cooldowns: Mutex<BTreeMap<ProviderAccountId, ProviderCooldown>>,
+    scoped_cooldowns:
+        Mutex<BTreeMap<(ProviderAccountId, ProviderCooldownScope), ProviderScopedCooldown>>,
 }
 
 impl MemoryCooldownPort {
     pub fn cooldown(&self, account_id: &ProviderAccountId) -> Option<ProviderCooldown> {
         lock(&self.cooldowns).get(account_id).cloned()
     }
+
+    pub fn scoped_cooldown(
+        &self,
+        account_id: &ProviderAccountId,
+        scope: &ProviderCooldownScope,
+    ) -> Option<ProviderScopedCooldown> {
+        lock(&self.scoped_cooldowns)
+            .get(&(account_id.clone(), scope.clone()))
+            .cloned()
+    }
 }
 
 impl ProviderSessionAffinityPort for TestSessionAffinity {
@@ -674,6 +687,53 @@ impl ProviderCooldownPort for MemoryCooldownPort {
             Ok(should_remove)
         })
     }
+
+    fn put_scoped_if_later(
+        &self,
+        cooldown: ProviderScopedCooldown,
+    ) -> futures::future::BoxFuture<'_, Result<bool, ProviderStoreError>> {
+        Box::pin(async move {
+            let key = (cooldown.account_id().clone(), cooldown.scope().clone());
+            let mut cooldowns = lock(&self.scoped_cooldowns);
+            let should_write = cooldowns.get(&key).is_none_or(|current| {
+                current.credential_revision() < cooldown.credential_revision()
+                    || (current.credential_revision() == cooldown.credential_revision()
+                        && current.until() < cooldown.until())
+            });
+            if should_write {
+                cooldowns.insert(key, cooldown);
+            }
+            Ok(should_write)
+        })
+    }
+
+    fn read_scoped<'a>(
+        &'a self,
+        account_id: &'a ProviderAccountId,
+        scope: &'a ProviderCooldownScope,
+    ) -> futures::future::BoxFuture<'a, Result<Option<ProviderScopedCooldown>, ProviderStoreError>>
+    {
+        Box::pin(async move { Ok(self.scoped_cooldown(account_id, scope)) })
+    }
+
+    fn clear_scoped<'a>(
+        &'a self,
+        account_id: &'a ProviderAccountId,
+        scope: &'a ProviderCooldownScope,
+        through_revision: CredentialRevision,
+    ) -> futures::future::BoxFuture<'a, Result<bool, ProviderStoreError>> {
+        Box::pin(async move {
+            let key = (account_id.clone(), scope.clone());
+            let mut cooldowns = lock(&self.scoped_cooldowns);
+            let should_remove = cooldowns
+                .get(&key)
+                .is_some_and(|current| current.credential_revision() <= through_revision);
+            if should_remove {
+                cooldowns.remove(&key);
+            }
+            Ok(should_remove)
+        })
+    }
 }
 
 impl ProviderRuntimePolicyPort for StaticRuntimePolicy {

backend/crates/providers/xai/tests/transport/network.rs

@@ -257,7 +257,7 @@ async fn inference_transport_should_classify_http_failures_without_retaining_bod
     let cases = [
         (400, GrokInferenceTransportErrorKind::InvalidRequest),
         (401, GrokInferenceTransportErrorKind::Unauthorized),
-        (402, GrokInferenceTransportErrorKind::QuotaExhausted),
+        (402, GrokInferenceTransportErrorKind::PaymentRequired),
         (403, GrokInferenceTransportErrorKind::PermissionDenied),
         (408, GrokInferenceTransportErrorKind::Timeout),
         (429, GrokInferenceTransportErrorKind::RateLimited),
@@ -302,10 +302,82 @@ async fn inference_transport_should_classify_http_failures_without_retaining_bod
             Some(GrokInferenceClientCacheStatus::Miss)
         );
         assert!(error.transport_metrics().headers_ms().is_some());
+        assert!(error.client_visible_upstream_error().is_none());
         assert!(!rendered.contains(&secret));
     }
 }
 
+#[tokio::test]
+async fn inference_transport_should_expose_safe_flat_json_error_details() {
+    let server = MockServer::start().await;
+    let message = "You have run out of credits or need a Grok subscription";
+    Mock::given(method("POST"))
+        .and(path("/v1/responses"))
+        .respond_with(ResponseTemplate::new(402).set_body_json(json!({
+            "code": "personal-team-blocked:spending-limit",
+            "error": message,
+        })))
+        .expect(1)
+        .mount(&server)
+        .await;
+    let origin = Url::parse(&server.uri()).expect("wiremock origin");
+    let error = inference_transport(&origin)
+        .execute(inference_request(&origin))
+        .await
+        .expect_err("spending limit response must fail");
+    let detail = error
+        .client_visible_upstream_error()
+        .expect("structured error detail");
+    let rendered = format!("{error:?}\n{error}");
+
+    assert_eq!(
+        error.kind(),
+        GrokInferenceTransportErrorKind::QuotaExhausted
+    );
+    assert_eq!(
+        (detail.message(), detail.code(), detail.error_type()),
+        (message, Some("personal_team_blocked_spending_limit"), None,)
+    );
+    assert!(!rendered.contains(message));
+}
+
+#[tokio::test]
+async fn inference_transport_should_scrub_nested_json_error_details() {
+    let server = MockServer::start().await;
+    let account_fingerprint = "123e4567-e89b-12d3-a456-426614174000";
+    Mock::given(method("POST"))
+        .and(path("/v1/responses"))
+        .respond_with(ResponseTemplate::new(429).set_body_json(json!({
+            "error": {
+                "code": "rate-limit:burst",
+                "type": "rate_limit_error",
+                "message": format!("team {account_fingerprint}\nrate limited"),
+            }
+        })))
+        .expect(1)
+        .mount(&server)
+        .await;
+    let origin = Url::parse(&server.uri()).expect("wiremock origin");
+    let error = inference_transport(&origin)
+        .execute(inference_request(&origin))
+        .await
+        .expect_err("rate limit response must fail");
+    let detail = error
+        .client_visible_upstream_error()
+        .expect("structured error detail");
+    let rendered = format!("{error:?}\n{error}");
+
+    assert_eq!(
+        (detail.message(), detail.code(), detail.error_type()),
+        (
+            "team [redacted] rate limited",
+            Some("rate_limit_burst"),
+            Some("rate_limit_error"),
+        )
+    );
+    assert!(!rendered.contains(account_fingerprint));
+}
+
 #[tokio::test]
 async fn inference_transport_should_scope_forbidden_failures_from_safe_metadata() {
     let cases = [
@@ -317,14 +389,20 @@ async fn inference_transport_should_scope_forbidden_failures_from_safe_metadata(
         ),
         (
             json!({"error": {"code": "usage_exhausted", "message": "used all the included free usage for model"}}),
-            GrokInferenceTransportErrorKind::QuotaExhausted,
+            GrokInferenceTransportErrorKind::ModelQuotaExhausted,
             false,
             Some("usage_exhausted"),
         ),
         (
             json!({"error": {"code": "permission_denied", "message": "access to the chat endpoint is denied"}}),
-            GrokInferenceTransportErrorKind::PermissionDenied,
-            true,
+            GrokInferenceTransportErrorKind::ModelAccessDenied,
+            false,
+            Some("permission_denied"),
+        ),
+        (
+            json!({"error": {"code": "permission_denied", "message": "request rejected", "details": {"check": "SAFETY_CHECK_TYPE_INPUT"}}}),
+            GrokInferenceTransportErrorKind::SafetyRejected,
+            false,
             Some("permission_denied"),
         ),
         (
@@ -359,7 +437,62 @@ async fn inference_transport_should_scope_forbidden_failures_from_safe_metadata(
 }
 
 #[tokio::test]
-async fn inference_transport_should_classify_free_usage_429_as_quota_exhausted() {
+async fn inference_transport_should_keep_unknown_402_out_of_account_quota_state() {
+    let server = MockServer::start().await;
+    Mock::given(method("POST"))
+        .and(path("/v1/responses"))
+        .respond_with(ResponseTemplate::new(402).set_body_json(json!({
+            "error": {
+                "code": "payment_required",
+                "type": "billing_error",
+                "message": "Payment is required for this request"
+            }
+        })))
+        .expect(1)
+        .mount(&server)
+        .await;
+    let origin = Url::parse(&server.uri()).expect("wiremock origin");
+    let error = inference_transport(&origin)
+        .execute(inference_request(&origin))
+        .await
+        .expect_err("payment-required response must fail");
+
+    assert_eq!(
+        (error.kind(), error.requires_credential_recovery()),
+        (GrokInferenceTransportErrorKind::PaymentRequired, false)
+    );
+}
+
+#[tokio::test]
+async fn inference_transport_should_classify_model_quota_for_403_and_429() {
+    for status in [403, 429] {
+        let server = MockServer::start().await;
+        Mock::given(method("POST"))
+            .and(path("/v1/responses"))
+            .respond_with(ResponseTemplate::new(status).set_body_json(json!({
+                "error": {
+                    "code": "subscription:free-usage-exhausted",
+                    "message": "You have used all the included free usage for model grok-4.5"
+                }
+            })))
+            .expect(1)
+            .mount(&server)
+            .await;
+        let origin = Url::parse(&server.uri()).expect("wiremock origin");
+        let error = inference_transport(&origin)
+            .execute(inference_request(&origin))
+            .await
+            .expect_err("model quota response must fail");
+
+        assert_eq!(
+            error.kind(),
+            GrokInferenceTransportErrorKind::ModelQuotaExhausted
+        );
+    }
+}
+
+#[tokio::test]
+async fn inference_transport_should_classify_free_usage_429_as_account_free_quota_exhausted() {
     let server = MockServer::start().await;
     Mock::given(method("POST"))
         .and(path("/v1/responses"))
@@ -380,7 +513,7 @@ async fn inference_transport_should_classify_free_usage_429_as_quota_exhausted()
 
     assert_eq!(
         error.kind(),
-        GrokInferenceTransportErrorKind::QuotaExhausted
+        GrokInferenceTransportErrorKind::FreeQuotaExhausted
     );
     assert_eq!(error.status(), Some(429));
     assert_eq!(

release/version.yaml

@@ -1 +1 @@
-version: "3.0.19"
+version: "3.0.20"