chore: bump version to 3.0.11

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

.github/workflows/_quality.yml

@@ -64,6 +64,41 @@ jobs:
         uses: actions/checkout@9c091bb21b7c1c1d1991bb908d89e4e9dddfe3e0 # v7.0.0
         with:
           ref: ${{ inputs.ref }}
+          # PR 场景需要 base 提交做冻结清单的 append-only 比对。
+          fetch-depth: ${{ github.event_name == 'pull_request' && 0 || 1 }}
+
+      # 已应用迁移的字节被 sqlx 启动校验钉死：改动已发布迁移会让所有
+      # 存量部署重启即失败。冻结清单保证旧文件不被改、新文件必须入册。
+      - name: Migration freeze check
+        working-directory: backend/migrations
+        shell: bash
+        env:
+          BASE_SHA: ${{ github.event.pull_request.base.sha }}
+        run: |
+          set -euo pipefail
+          sha256sum --check --strict .frozen-sha256
+          for file in *.sql; do
+            if ! awk '{print $2}' .frozen-sha256 | grep -Fxq "$file"; then
+              echo "::error file=backend/migrations/$file::migration $file is not frozen; append its sha256 to backend/migrations/.frozen-sha256"
+              exit 1
+            fi
+          done
+          # 清单自证不足以防「改迁移 + 同 PR 刷新清单行」：校验对新哈希通过。
+          # 对 PR，断言 .frozen-sha256 相对 base 只增不改——base 里的每一行
+          # 必须原样保留（改哈希＝删行，删迁移＝删行，都会被挡）。
+          if [ -n "${BASE_SHA:-}" ]; then
+            base_manifest="$(git show "${BASE_SHA}:backend/migrations/.frozen-sha256" 2>/dev/null || true)"
+            if [ -n "$base_manifest" ]; then
+              while IFS= read -r line; do
+                [ -z "$line" ] && continue
+                if ! grep -Fxq "$line" .frozen-sha256; then
+                  file="$(printf '%s' "$line" | awk '{print $2}')"
+                  echo "::error file=backend/migrations/.frozen-sha256::frozen migration $file was modified or removed; released migrations are append-only — add a new numbered migration instead"
+                  exit 1
+                fi
+              done <<< "$base_manifest"
+            fi
+          fi
 
       - name: Setup Rust
         uses: dtolnay/rust-toolchain@1a3a6d54512beeaffd394f8d516ca16f2c506a20 # 1.97

.github/workflows/release.yml

@@ -629,6 +629,9 @@ jobs:
           prerelease: ${{ needs.release-info.outputs.prerelease == 'true' }}
           make_latest: ${{ needs.release-info.outputs.prerelease == 'true' && 'false' || 'true' }}
           body: |
+            > [!WARNING]
+            > **Upgrading across major versions?** Database schema and configuration are only compatible within a major version, and the one-click self-update refuses cross-major targets (releases before 3.0 lack this gate — do not use their built-in update to reach a newer major). Deploy the new major fresh (new `.runtime/` data directory) and re-import or re-authorize provider accounts and client keys. See the [upgrade notes](https://github.com/${{ github.repository }}/blob/main/deploy/README.md#镜像升级与源码构建).
+
             ## 📥 Installation
 
             **Docker:**

README.md

@@ -26,7 +26,7 @@
 | 协议     | OpenAI Responses JSON、SSE、WebSocket 与模型目录                             |
 | Provider | 固定 OpenAI、xAI 两个 Provider；各自管理 OAuth、账号、额度与模型目录          |
 | 路由     | Client Key 绑定 Provider、全局模型映射、会话亲和与同 Provider 安全 fallback   |
-| 透明边界 | OpenAI 保持 Responses 语义透明；xAI 的协议适配只驻留在 `providers/xai`         |
+| 透明边界 | OpenAI SSE 与 WebSocket 均原样转发上游字节；xAI 在 `providers/xai` 内做 Grok 与 Responses 协议转换 |
 | 延续     | OpenAI native/replay continuation；xAI 使用客户端提交的完整历史                |
 | 管理     | Client Key、账号、模型目录、设置、观测、系统更新与 OAuth                       |
 | 计量     | 模型请求 Token、费用、延迟、账号与 Provider 归因                               |
@@ -91,7 +91,7 @@ curl -i http://127.0.0.1:8080/healthz
 2. 在账号页选择对应 Provider：OpenAI 支持 OAuth、账号文件与 Agent Identity；xAI 支持 OAuth
    和兼容账号文件导入。
 3. 按需配置客户端模型到上游模型的全局映射；未命中映射的模型名原样透传。
-4. 创建 `sk_...` Client Key，绑定 `openai` 或 `xai`，并设置模型范围、速率与并发限制。
+4. 创建 `sk_...` Client Key，绑定 `openai` 或 `xai`，并设置速率与并发限制。
 
 > [!IMPORTANT]
 > xAI 使用 OAuth session，不支持把 xAI API Key 作为上游 credential。
@@ -158,9 +158,15 @@ docker compose -f deploy/compose.yaml up -d
 ```
 
 > [!IMPORTANT]
-> `.runtime/` 保存 PostgreSQL、Redis、日志、OpenAI 会话锚点和自更新状态。删除该目录会永久清除
+> `.runtime/` 保存 PostgreSQL、Redis、日志、OpenAI 会话锚点密钥和自更新状态。删除该目录会永久清除
 > 运行状态。
 
+> [!WARNING]
+> 跨大版本不是原地升级：数据库 schema 与配置只在同一大版本内兼容，内置一键自更新
+> 会拒绝跨大版本目标（3.0 之前的旧版没有这道门禁，不要用旧版界面的一键更新跨大版本）。
+> 升级到新的大版本需要全新部署（全新 `.runtime/` 数据目录），并重新导入或重新授权
+> Provider 账号与客户端 Key。
+
 Compose 默认只绑定 `127.0.0.1`。公网接入应使用 HTTPS 反向代理，转发 WebSocket
 upgrade 与真实客户端 IP；不要暴露 PostgreSQL 或 Redis。
 

backend/Cargo.lock

@@ -1189,7 +1189,6 @@ version = "0.1.0"
 dependencies = [
  "async-trait",
  "bytes",
- "chrono",
  "futures",
  "futures-timer",
  "serde",
@@ -1225,7 +1224,6 @@ dependencies = [
  "tracing",
  "tracing-appender",
  "tracing-subscriber",
- "uuid",
  "wiremock",
 ]
 
@@ -2151,7 +2149,6 @@ dependencies = [
  "url",
  "uuid",
  "wiremock",
- "zeroize",
 ]
 
 [[package]]
@@ -2161,6 +2158,7 @@ dependencies = [
  "async-stream",
  "async-trait",
  "base64",
+ "bytes",
  "chrono",
  "futures",
  "gateway-admin",

backend/crates/gateway-admin/src/model/auth.rs

@@ -39,7 +39,6 @@ impl AdminRequestContext {
 pub struct LoginCommand {
     pub username: Option<String>,
     pub password: String,
-    pub source: String,
 }
 
 impl std::fmt::Debug for LoginCommand {
@@ -48,7 +47,6 @@ impl std::fmt::Debug for LoginCommand {
             .debug_struct("LoginCommand")
             .field("username", &self.username)
             .field("password", &"[REDACTED]")
-            .field("source", &self.source)
             .finish()
     }
 }
@@ -65,8 +63,6 @@ pub struct LoginResult {
 pub enum LoginError {
     #[error("invalid administrator credentials")]
     InvalidCredentials,
-    #[error("administrator login is throttled")]
-    Throttled,
     #[error("administrator authentication is unavailable")]
     Unavailable,
 }

backend/crates/gateway-admin/src/ports/store.rs

@@ -156,7 +156,7 @@ pub trait AccountStore: Send + Sync {
     ) -> AdminStoreResult<()>;
 }
 
-/// 管理员密码、会话、登录限流和安全审计。
+/// 管理员密码、会话和安全审计。
 #[async_trait]
 pub trait AuthStore: Send + Sync {
     async fn load_password_hash(&self, admin_user_id: &str) -> AdminStoreResult<Option<String>>;
@@ -176,22 +176,6 @@ pub trait AuthStore: Send + Sync {
 
     async fn delete_session(&self, session_id: &str) -> AdminStoreResult<Option<AdminSession>>;
 
-    async fn login_source_is_throttled(
-        &self,
-        source: &str,
-        failure_limit: u32,
-        window_seconds: u64,
-    ) -> AdminStoreResult<bool>;
-
-    async fn record_login_failure(
-        &self,
-        source: &str,
-        failure_limit: u32,
-        window_seconds: u64,
-    ) -> AdminStoreResult<bool>;
-
-    async fn clear_login_failures(&self, source: &str) -> AdminStoreResult<()>;
-
     async fn append_audit_event(&self, event: AdminAuditEvent) -> AdminStoreResult<()>;
 }
 

backend/crates/gateway-admin/src/use_case/auth.rs

@@ -22,9 +22,6 @@ use crate::{
 
 use super::map_store_error;
 
-const LOGIN_FAILURE_LIMIT: u32 = 5;
-const LOGIN_FAILURE_WINDOW_SECONDS: u64 = 15 * 60;
-
 /// API 鉴权与管理员登录消费的统一服务。
 #[async_trait]
 pub trait AuthService: Send + Sync {
@@ -63,19 +60,6 @@ impl DefaultAuthService {
         }
     }
 
-    async fn record_failed_login(&self, source: &str) -> Result<LoginError, LoginError> {
-        let throttled = self
-            .store
-            .record_login_failure(source, LOGIN_FAILURE_LIMIT, LOGIN_FAILURE_WINDOW_SECONDS)
-            .await
-            .map_err(|_| LoginError::Unavailable)?;
-        Ok(if throttled {
-            LoginError::Throttled
-        } else {
-            LoginError::InvalidCredentials
-        })
-    }
-
     fn auth_audit(&self, action: &str, occurred_at: chrono::DateTime<Utc>) -> AdminAuditEvent {
         AdminAuditEvent {
             id: format!("audit_{}", Uuid::now_v7().simple()),
@@ -137,22 +121,13 @@ impl AuthService for DefaultAuthService {
     }
 
     async fn login(&self, command: LoginCommand) -> Result<LoginResult, LoginError> {
-        let source = normalized_login_source(&command.source);
-        if self
-            .store
-            .login_source_is_throttled(source, LOGIN_FAILURE_LIMIT, LOGIN_FAILURE_WINDOW_SECONDS)
-            .await
-            .map_err(|_| LoginError::Unavailable)?
-        {
-            return Err(LoginError::Throttled);
-        }
         if command
             .username
             .as_deref()
             .unwrap_or(&self.default_admin_user_id)
             != self.default_admin_user_id
         {
-            return Err(self.record_failed_login(source).await?);
+            return Err(LoginError::InvalidCredentials);
         }
         let hash = self
             .store
@@ -161,13 +136,9 @@ impl AuthService for DefaultAuthService {
             .map_err(|_| LoginError::Unavailable)?
             .ok_or(LoginError::InvalidCredentials)?;
         if !verify_admin_password(&command.password, &hash).map_err(|_| LoginError::Unavailable)? {
-            return Err(self.record_failed_login(source).await?);
+            return Err(LoginError::InvalidCredentials);
         }
 
-        self.store
-            .clear_login_failures(source)
-            .await
-            .map_err(|_| LoginError::Unavailable)?;
         let session_id = random_session_token();
         let expires_at = Utc::now() + self.session_ttl;
         self.store
@@ -246,8 +217,3 @@ fn valid_admin_api_key_shape(value: &str) -> bool {
         && value.starts_with("admin-")
         && value[6..].bytes().all(|byte| byte.is_ascii_hexdigit())
 }
-
-fn normalized_login_source(source: &str) -> &str {
-    let source = source.trim();
-    if source.is_empty() { "unknown" } else { source }
-}

backend/crates/gateway-admin/tests/main.rs

@@ -87,7 +87,6 @@ fn login_command_debug_should_redact_password() {
     let command = LoginCommand {
         username: None,
         password: "secret-password".to_owned(),
-        source: "127.0.0.1".to_owned(),
     };
     assert!(!format!("{command:?}").contains("secret-password"));
 }

backend/crates/gateway-admin/tests/use_case/auth.rs

@@ -15,7 +15,6 @@ struct MemoryAuthStore {
     password_hash: Mutex<Option<String>>,
     sessions: Mutex<BTreeMap<String, AdminSession>>,
     audits: Mutex<Vec<AdminAuditEvent>>,
-    failures: Mutex<u32>,
 }
 
 #[async_trait]
@@ -66,31 +65,6 @@ impl AuthStore for MemoryAuthStore {
         Ok(self.sessions.lock().expect("sessions").remove(session_id))
     }
 
-    async fn login_source_is_throttled(
-        &self,
-        _: &str,
-        failure_limit: u32,
-        _: u64,
-    ) -> AdminStoreResult<bool> {
-        Ok(*self.failures.lock().expect("failures") >= failure_limit)
-    }
-
-    async fn record_login_failure(
-        &self,
-        _: &str,
-        failure_limit: u32,
-        _: u64,
-    ) -> AdminStoreResult<bool> {
-        let mut failures = self.failures.lock().expect("failures");
-        *failures += 1;
-        Ok(*failures >= failure_limit)
-    }
-
-    async fn clear_login_failures(&self, _: &str) -> AdminStoreResult<()> {
-        *self.failures.lock().expect("failures") = 0;
-        Ok(())
-    }
-
     async fn append_audit_event(&self, event: AdminAuditEvent) -> AdminStoreResult<()> {
         self.audits.lock().expect("audits").push(event);
         Ok(())
@@ -107,7 +81,6 @@ async fn successful_login_should_create_expiring_session_and_audit() {
         .login(LoginCommand {
             username: Some("admin".to_owned()),
             password: "strong-test-password".to_owned(),
-            source: "127.0.0.1".to_owned(),
         })
         .await
         .expect("login");
@@ -142,7 +115,6 @@ async fn repeated_default_initialization_should_not_replace_password() {
             .login(LoginCommand {
                 username: Some("admin".to_owned()),
                 password: "first-strong-password".to_owned(),
-                source: "127.0.0.1".to_owned(),
             })
             .await
             .is_ok()

backend/crates/gateway-admin/tests/use_case/mod.rs

@@ -210,18 +210,6 @@ impl AuthStore for BootstrapAuthStore {
         Err(unavailable("admin session"))
     }
 
-    async fn login_source_is_throttled(&self, _: &str, _: u32, _: u64) -> AdminStoreResult<bool> {
-        Ok(false)
-    }
-
-    async fn record_login_failure(&self, _: &str, _: u32, _: u64) -> AdminStoreResult<bool> {
-        Err(unavailable("login failure"))
-    }
-
-    async fn clear_login_failures(&self, _: &str) -> AdminStoreResult<()> {
-        Ok(())
-    }
-
     async fn append_audit_event(&self, _: AdminAuditEvent) -> AdminStoreResult<()> {
         Ok(())
     }

backend/crates/gateway-api/src/admin/auth.rs

@@ -3,7 +3,7 @@
 use std::fmt;
 
 use axum::{
-    Extension, Json, Router,
+    Json, Router,
     extract::{FromRequestParts, State},
     http::{HeaderMap, HeaderValue, StatusCode, header::SET_COOKIE, request::Parts},
     response::{IntoResponse, Response},
@@ -14,6 +14,7 @@ use gateway_admin::{
     model::auth::{AdminPrincipal, AdminRequestContext, LoginCommand, LoginError},
 };
 use serde::{Deserialize, Serialize};
+use tower_http::request_id::RequestId;
 
 use super::{AdminEnvelope, AdminError, AdminResponse};
 
@@ -51,22 +52,32 @@ where
 
     async fn from_request_parts(parts: &mut Parts, state: &S) -> Result<Self, Self::Rejection> {
         let principal = require_admin_auth(state, &parts.headers).await?;
-        let request_id = parts
-            .headers
-            .get(REQUEST_ID_HEADER)
-            .and_then(|value| value.to_str().ok())
-            .map(str::trim)
-            .filter(|value| !value.is_empty())
+        let request_id = admin_request_id(parts)
             .ok_or_else(|| AdminError::internal("Missing admin request context"))?;
         Ok(Self {
             context: AdminRequestContext {
                 principal,
-                request_id: request_id.to_owned(),
+                request_id,
             },
         })
     }
 }
 
+/// request-id 层按配置的 header 名注入，同时写入与名字无关的扩展；
+/// 优先读扩展，使自定义 header 名不会让管理请求失去请求上下文。
+/// header 回退覆盖未装配该层的调用方。
+fn admin_request_id(parts: &Parts) -> Option<String> {
+    parts
+        .extensions
+        .get::<RequestId>()
+        .map(RequestId::header_value)
+        .or_else(|| parts.headers.get(REQUEST_ID_HEADER))
+        .and_then(|value| value.to_str().ok())
+        .map(str::trim)
+        .filter(|value| !value.is_empty())
+        .map(ToOwned::to_owned)
+}
+
 pub async fn require_admin_session<S>(state: &S, headers: &HeaderMap) -> Result<String, AdminError>
 where
     S: AdminSessionState + Send + Sync,
@@ -185,22 +196,6 @@ impl Default for AdminLogoutData {
     }
 }
 
-/// 请求上下文解析出的客户端登录来源。
-#[derive(Debug, Clone, PartialEq, Eq)]
-pub struct AdminLoginSource(String);
-
-impl AdminLoginSource {
-    #[must_use]
-    pub fn new(value: impl Into<String>) -> Self {
-        Self(value.into())
-    }
-
-    #[must_use]
-    pub fn as_str(&self) -> &str {
-        &self.0
-    }
-}
-
 /// 构造固定 GET/POST 管理员认证路由。
 pub fn router<S>() -> Router<S>
 where
@@ -214,24 +209,16 @@ where
 
 async fn login<S>(
     State(state): State<S>,
-    source: Option<Extension<AdminLoginSource>>,
     Json(payload): Json<AdminLoginRequest>,
 ) -> Result<Response, AdminError>
 where
     S: AdminSessionState + Send + Sync,
 {
-    let source = source
-        .as_ref()
-        .map_or("unknown", |Extension(source)| source.as_str());
     let (username, password) = payload.into_parts();
     let session = state
         .admin_services()
         .auth()
-        .login(LoginCommand {
-            username,
-            password,
-            source: source.to_owned(),
-        })
+        .login(LoginCommand { username, password })
         .await
         .map_err(map_login_error)?;
     let mut response = AdminResponse::new(
@@ -300,7 +287,6 @@ fn admin_session_cookie(headers: &HeaderMap) -> Option<String> {
 fn map_login_error(error: LoginError) -> AdminError {
     match error {
         LoginError::InvalidCredentials => AdminError::invalid_admin_credentials(),
-        LoginError::Throttled => AdminError::too_many_login_attempts(),
         LoginError::Unavailable => AdminError::internal("Failed to create admin session"),
     }
 }

backend/crates/gateway-api/src/lib.rs

@@ -16,7 +16,7 @@ use gateway_core::engine::execution::ExecutionService;
 use gateway_core::health::{HealthProbe, WorkerHealthSource};
 use gateway_core::lifecycle::ConnectionLifecycle;
 use serde::Deserialize;
-use tower_http::cors::{Any, CorsLayer};
+use tower_http::cors::CorsLayer;
 use tower_http::request_id::{MakeRequestUuid, PropagateRequestIdLayer, SetRequestIdLayer};
 use tower_http::services::{ServeDir, ServeFile};
 use tower_http::timeout::TimeoutLayer;
@@ -146,11 +146,17 @@ pub fn initialize(
                     .map_err(|_| ApiError::Config(ApiConfigError::InvalidCorsOrigin))
             })
             .collect::<Result<Vec<_>, _>>()?;
+        // credentials 模式禁止通配 allow_headers；显式列出鉴权与内容协商头。
         router = router.layer(
             CorsLayer::new()
                 .allow_origin(origins)
                 .allow_methods([Method::GET, Method::POST])
-                .allow_headers(Any)
+                .allow_headers([
+                    axum::http::header::AUTHORIZATION,
+                    axum::http::header::CONTENT_TYPE,
+                    HeaderName::from_static("x-api-key"),
+                    request_id_header.clone(),
+                ])
                 .allow_credentials(true),
         );
     }

backend/crates/gateway-api/src/openai/models.rs

@@ -2,11 +2,12 @@
 
 use axum::{
     Json,
-    extract::{Path, State},
+    extract::{Path, Query, State},
     http::{HeaderMap, StatusCode},
     response::{IntoResponse, Response},
 };
 use gateway_core::routing::{PublicModelId, PublicModelProfile};
+use serde::Deserialize;
 use serde_json::{Value, json};
 
 use crate::ApiState;
@@ -19,29 +20,48 @@ use super::{
 const MODEL_CREATED_TIMESTAMP: i64 = 1_700_000_000;
 const CODEX_BASE_INSTRUCTIONS: &str = "You are Codex, a coding agent. Follow the user's instructions and use the available tools to complete software engineering tasks. Inspect relevant files before editing, preserve unrelated changes, and verify the result.";
 
-/// `GET /v1/models`。
-pub(crate) async fn models(State(state): State<ApiState>, headers: HeaderMap) -> Response {
+#[derive(Debug, Deserialize)]
+pub(crate) struct ModelsQuery {
+    client_version: Option<String>,
+}
+
+/// `GET /v1/models`。Codex 携带 `client_version` 时返回其专用目录合同。
+pub(crate) async fn models(
+    State(state): State<ApiState>,
+    Query(query): Query<ModelsQuery>,
+    headers: HeaderMap,
+) -> Response {
     let service = state.openai();
     let client = match authenticate_client(service, &headers) {
         Ok(client) => client,
         Err(error) => return authentication_error_response(error),
     };
+
+    if query
+        .client_version
+        .as_deref()
+        .is_some_and(|version| !version.trim().is_empty())
+    {
+        let models = service
+            .public_model_profiles(&client)
+            .iter()
+            .enumerate()
+            .map(|(index, profile)| codex_model_json(profile, index))
+            .collect::<Vec<_>>();
+        return (
+            StatusCode::OK,
+            Json(json!({
+                "models": models,
+            })),
+        )
+            .into_response();
+    }
+
     let data = service
         .public_models(&client)
         .into_iter()
         .map(|model| openai_model_json(&model))
         .collect::<Vec<_>>();
-    let profiles = service.public_model_profiles(&client);
-
-    if !profiles.is_empty() {
-        let models = profiles
-            .iter()
-            .enumerate()
-            .map(|(index, profile)| codex_model_json(profile, index))
-            .collect::<Vec<_>>();
-        return (StatusCode::OK, Json(json!({ "models": models }))).into_response();
-    }
-
     (
         StatusCode::OK,
         Json(json!({
@@ -209,6 +229,7 @@ fn codex_model_json(profile: &PublicModelProfile, index: usize) -> Value {
 fn reasoning_effort_description(effort: &str) -> &'static str {
     match effort {
         "none" => "No reasoning",
+        "minimal" => "Minimal reasoning",
         "low" => "Fast responses with lighter reasoning",
         "medium" => "Balances speed and reasoning depth for everyday tasks",
         "high" => "Greater reasoning depth for complex problems",

backend/crates/gateway-api/src/openai/responses/request.rs

@@ -310,9 +310,18 @@ pub(super) fn decode_request_inner(
 ) -> Result<DecodedResponsesRequest, RequestDecodeError> {
     let value =
         serde_json::from_slice::<Value>(body).map_err(|_| RequestDecodeError::MalformedJson)?;
-    let Value::Object(mut object) = value else {
+    let Value::Object(object) = value else {
         return Err(RequestDecodeError::ExpectedObject);
     };
+    decode_request_object(object, review, request_headers)
+}
+
+/// 解码已解析的顶层 object；WebSocket 帧解码复用此内核，避免序列化往返。
+pub(super) fn decode_request_object(
+    mut object: Map<String, Value>,
+    review: bool,
+    request_headers: &OpenAiRequestHeaders,
+) -> Result<DecodedResponsesRequest, RequestDecodeError> {
     // `use_websocket` 只影响本地 transport。无论值是否可识别，都不得进入上游 body。
     let use_websocket = object
         .remove("use_websocket")

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

@@ -549,9 +549,7 @@ fn decode_response_create_inner(
     if matches!(body.get("stream"), Some(value) if value.as_bool() != Some(true)) {
         return Err(ResponseCreateFrameError::StreamingRequired);
     }
-    let encoded = serde_json::to_vec(&Value::Object(body))
-        .map_err(|_| ResponseCreateFrameError::InvalidJson)?;
-    super::request::decode_request_inner(&encoded, false, request_headers)
+    super::request::decode_request_object(body, false, request_headers)
         .map_err(ResponseCreateFrameError::Request)
 }
 

backend/crates/gateway-api/tests/admin/auth.rs

@@ -51,7 +51,6 @@ async fn default_auth_service_should_initialize_login_validate_and_logout() {
     let service = fixture.services.auth();
     let session = service
         .login(LoginCommand {
-            source: "127.0.0.1".to_owned(),
             username: Some("admin_1".to_owned()),
             password: "strong-admin-password".to_owned(),
         })
@@ -84,42 +83,6 @@ async fn default_auth_service_should_initialize_login_validate_and_logout() {
     assert_eq!(fixture.auth.audit_count(), 2);
 }
 
-#[tokio::test]
-async fn default_auth_service_should_throttle_repeated_invalid_identity() {
-    let fixture = AdminTestFixture::new().await;
-    for attempt in 1..=5 {
-        let error = fixture
-            .services
-            .auth()
-            .login(LoginCommand {
-                source: "shared-source".to_owned(),
-                username: Some("wrong-user".to_owned()),
-                password: "wrong-password".to_owned(),
-            })
-            .await
-            .expect_err("invalid login");
-        if attempt < 5 {
-            assert_eq!(error, LoginError::InvalidCredentials);
-        } else {
-            assert_eq!(error, LoginError::Throttled);
-        }
-    }
-    assert_eq!(
-        fixture
-            .services
-            .auth()
-            .login(LoginCommand {
-                source: "shared-source".to_owned(),
-                username: None,
-                password: "anything".to_owned(),
-            })
-            .await
-            .expect_err("source remains throttled"),
-        LoginError::Throttled
-    );
-    assert_eq!(fixture.auth.session_count(), 0);
-}
-
 #[tokio::test]
 async fn default_auth_service_should_verify_only_full_plaintext_admin_key() {
     let fixture = AdminTestFixture::new().await;
@@ -162,7 +125,6 @@ async fn audit_failure_should_revoke_new_session_before_returning_it() {
             .services
             .auth()
             .login(LoginCommand {
-                source: "source".to_owned(),
                 username: None,
                 password: "strong-admin-password".to_owned(),
             })

backend/crates/gateway-api/tests/admin/mod.rs

@@ -146,7 +146,6 @@ impl AdminSessionState for AdminTestState {
 pub(super) struct MemoryAuthStore {
     password_hash: Mutex<Option<String>>,
     sessions: Mutex<BTreeMap<String, AdminSession>>,
-    failures: Mutex<BTreeMap<String, u32>>,
     audits: Mutex<Vec<AdminAuditEvent>>,
     api_key: Arc<Mutex<Option<AdminApiKey>>>,
     fail_audit: AtomicBool,
@@ -157,7 +156,6 @@ impl MemoryAuthStore {
         Self {
             password_hash: Mutex::new(None),
             sessions: Mutex::new(BTreeMap::new()),
-            failures: Mutex::new(BTreeMap::new()),
             audits: Mutex::new(Vec::new()),
             api_key,
             fail_audit: AtomicBool::new(false),
@@ -239,37 +237,6 @@ impl AuthStore for MemoryAuthStore {
         Ok(self.sessions.lock().expect("sessions").remove(session_id))
     }
 
-    async fn login_source_is_throttled(
-        &self,
-        source: &str,
-        failure_limit: u32,
-        _: u64,
-    ) -> AdminStoreResult<bool> {
-        Ok(self
-            .failures
-            .lock()
-            .expect("failures")
-            .get(source)
-            .is_some_and(|count| *count >= failure_limit))
-    }
-
-    async fn record_login_failure(
-        &self,
-        source: &str,
-        failure_limit: u32,
-        _: u64,
-    ) -> AdminStoreResult<bool> {
-        let mut failures = self.failures.lock().expect("failures");
-        let count = failures.entry(source.to_owned()).or_default();
-        *count = count.saturating_add(1);
-        Ok(*count >= failure_limit)
-    }
-
-    async fn clear_login_failures(&self, source: &str) -> AdminStoreResult<()> {
-        self.failures.lock().expect("failures").remove(source);
-        Ok(())
-    }
-
     async fn append_audit_event(&self, event: AdminAuditEvent) -> AdminStoreResult<()> {
         if self.fail_audit.load(Ordering::SeqCst) {
             return Err(unavailable("auth audit"));

backend/crates/gateway-api/tests/admin/settings.rs

@@ -213,3 +213,26 @@ async fn settings_should_accept_admin_api_key_header() {
 
     assert_eq!(response.status(), StatusCode::OK);
 }
+
+#[tokio::test]
+async fn admin_auth_should_accept_a_configured_request_id_header_name() {
+    use axum::http::HeaderName;
+    use tower_http::request_id::{MakeRequestUuid, SetRequestIdLayer};
+
+    let fixture = AdminTestFixture::new().await;
+    fixture.auth.insert_session("valid-session");
+    // 部署把 api.request_id_header 改名后，注入的 header 不再叫 x-request-id；
+    // 管理请求仍须拿到请求上下文，而不是退化为 500。
+    let custom = HeaderName::from_static("x-trace-id");
+    let app = app(fixture.state()).layer(SetRequestIdLayer::new(custom, MakeRequestUuid));
+    let unlabelled = Request::builder()
+        .method(Method::GET)
+        .uri("/api/admin/settings")
+        .header(header::COOKIE, "cpr_admin_session=valid-session")
+        .body(Body::empty())
+        .expect("build settings request");
+
+    let response = app.oneshot(unlabelled).await.expect("settings response");
+
+    assert_eq!(response.status(), StatusCode::OK);
+}

backend/crates/gateway-api/tests/openai/mod.rs

@@ -39,11 +39,18 @@ use gateway_core::routing::{
 };
 
 pub(super) async fn api_router(execution: Arc<dyn ExecutionService>) -> axum::Router {
+    api_router_with_origins(execution, Vec::new()).await
+}
+
+pub(super) async fn api_router_with_origins(
+    execution: Arc<dyn ExecutionService>,
+    cors_allowed_origins: Vec<String>,
+) -> axum::Router {
     let admin = crate::admin::AdminTestFixture::new().await;
     gateway_api::initialize(
         gateway_api::ApiConfig {
             asset_directory: std::env::temp_dir(),
-            cors_allowed_origins: Vec::new(),
+            cors_allowed_origins,
             request_timeout_seconds: None,
             request_id_header: "x-request-id".to_owned(),
         },

backend/crates/gateway-api/tests/openai/models.rs

@@ -15,13 +15,13 @@ use tower::ServiceExt;
 
 use super::{api_router, authenticated_client};
 
-struct ModelsExecution {
+pub(super) struct ModelsExecution {
     client: AuthenticatedClient,
     profiles: bool,
 }
 
 impl ModelsExecution {
-    fn new() -> Arc<Self> {
+    pub(super) fn new() -> Arc<Self> {
         Arc::new(Self {
             client: authenticated_client("sk_models_test"),
             profiles: false,
@@ -67,7 +67,9 @@ impl ExecutionService for ModelsExecution {
             )
             .with_reasoning(
                 Some("medium".to_owned()),
-                ["low", "medium", "high"].map(str::to_owned).to_vec(),
+                ["low", "medium", "high", "xhigh"]
+                    .map(str::to_owned)
+                    .to_vec(),
             )
             .with_context_window_tokens(Some(500_000))
             .with_image_input(true)
@@ -133,11 +135,77 @@ async fn models_should_encode_provider_profiles_for_current_codex_clients() {
     let value: serde_json::Value = serde_json::from_slice(&body).expect("models JSON");
     let model = &value["models"][0];
 
+    assert_eq!(
+        value
+            .as_object()
+            .map(|catalog| catalog.keys().map(String::as_str).collect::<Vec<_>>()),
+        Some(vec!["models"])
+    );
     assert_eq!(model["slug"], "grok-4.5");
     assert_eq!(model["default_reasoning_level"], "medium");
     assert_eq!(model["context_window"], 500_000);
     assert_eq!(model["apply_patch_tool_type"], "freeform");
     assert_eq!(model["service_tiers"], serde_json::json!([]));
+    assert_eq!(
+        model["supported_reasoning_levels"],
+        serde_json::json!([
+            {
+                "effort": "low",
+                "description": "Fast responses with lighter reasoning"
+            },
+            {
+                "effort": "medium",
+                "description": "Balances speed and reasoning depth for everyday tasks"
+            },
+            {
+                "effort": "high",
+                "description": "Greater reasoning depth for complex problems"
+            },
+            {
+                "effort": "xhigh",
+                "description": "Extra high reasoning depth for complex problems"
+            }
+        ])
+    );
+}
+
+#[tokio::test]
+async fn models_should_keep_the_codex_contract_when_profiles_are_empty() {
+    let response = api_router(ModelsExecution::new())
+        .await
+        .oneshot(authorized_request("/v1/models?client_version=0.145.0"))
+        .await
+        .expect("list empty Codex models response");
+    let body = to_bytes(response.into_body(), usize::MAX)
+        .await
+        .expect("read empty Codex models body");
+    let value: serde_json::Value = serde_json::from_slice(&body).expect("Codex models JSON");
+
+    assert_eq!(value, serde_json::json!({ "models": [] }));
+}
+
+#[tokio::test]
+async fn models_with_provider_profiles_should_keep_the_openai_list_contract() {
+    let response = api_router(ModelsExecution::with_profiles())
+        .await
+        .oneshot(authorized_request("/v1/models"))
+        .await
+        .expect("list compatible models response");
+    let body = to_bytes(response.into_body(), usize::MAX)
+        .await
+        .expect("read compatible models body");
+    let value: serde_json::Value = serde_json::from_slice(&body).expect("models JSON");
+
+    assert_eq!(
+        value,
+        serde_json::json!({
+            "object": "list",
+            "data": [
+                {"id":"model-a","object":"model","created":1700000000_i64,"owned_by":"gateway"},
+                {"id":"model-b","object":"model","created":1700000000_i64,"owned_by":"gateway"}
+            ]
+        })
+    );
 }
 
 #[tokio::test]

backend/crates/gateway-api/tests/openai/router.rs

@@ -2,12 +2,15 @@ use axum::{
     Router,
     body::{Body, Bytes},
     extract::DefaultBodyLimit,
-    http::{Request, StatusCode},
+    http::{Method, Request, StatusCode},
     routing::post,
 };
 use gateway_api::openai::router::MAX_CLIENT_REQUEST_BODY_BYTES;
 use tower::ServiceExt;
 
+use super::api_router_with_origins;
+use super::models::ModelsExecution;
+
 fn body_limit_app() -> Router {
     Router::new()
         .route(
@@ -50,3 +53,50 @@ async fn responses_body_limit_should_reject_one_byte_over_sixteen_mibibytes() {
 
     assert_eq!(response.status(), StatusCode::PAYLOAD_TOO_LARGE);
 }
+
+#[tokio::test]
+async fn configured_cors_origin_assembles_router_and_answers_preflight() {
+    let router = api_router_with_origins(
+        ModelsExecution::new(),
+        vec!["https://app.example.com".into()],
+    )
+    .await;
+
+    let response = router
+        .oneshot(
+            Request::builder()
+                .method(Method::OPTIONS)
+                .uri("/v1/models")
+                .header("origin", "https://app.example.com")
+                .header("access-control-request-method", "GET")
+                .header("access-control-request-headers", "authorization")
+                .body(Body::empty())
+                .expect("build preflight request"),
+        )
+        .await
+        .expect("route preflight request");
+
+    assert_eq!(response.status(), StatusCode::OK);
+    assert_eq!(
+        response
+            .headers()
+            .get("access-control-allow-origin")
+            .and_then(|value| value.to_str().ok()),
+        Some("https://app.example.com")
+    );
+    assert_eq!(
+        response
+            .headers()
+            .get("access-control-allow-credentials")
+            .and_then(|value| value.to_str().ok()),
+        Some("true")
+    );
+    let allow_headers = response
+        .headers()
+        .get("access-control-allow-headers")
+        .and_then(|value| value.to_str().ok())
+        .expect("allow-headers present");
+    assert!(allow_headers.contains("authorization"));
+    assert!(allow_headers.contains("x-api-key"));
+    assert!(allow_headers.contains("x-request-id"));
+}

backend/crates/gateway-core/Cargo.toml

@@ -14,7 +14,6 @@ path = "tests/main.rs"
 [dependencies]
 async-trait = "0.1.89"
 bytes = "1.11.1"
-chrono = "0.4.45"
 futures = "0.3.32"
 futures-timer = "3.0.3"
 serde = { version = "1.0.228", features = ["derive"] }

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

@@ -23,6 +23,7 @@ use crate::event::{
 };
 use crate::operation::{Operation, ProviderSessionState, RetrySafety};
 use crate::routing::RoutingPlan;
+use futures::future::Fuse;
 use futures::{FutureExt, StreamExt, pin_mut, select_biased};
 use futures_timer::Delay;
 
@@ -73,6 +74,12 @@ where
             client_api_key_ref,
             timing_started_at,
             deadline,
+            deadline_timer: Delay::new(
+                deadline
+                    .duration_since(SystemTime::now())
+                    .unwrap_or(Duration::ZERO),
+            )
+            .fuse(),
             pending_request: Some(request),
             request_persisted: false,
             operation,
@@ -86,6 +93,7 @@ where
             excluded_accounts: BTreeSet::new(),
             credential_recovery_attempted_accounts: BTreeSet::new(),
             current: None,
+            send_state_watermark: UpstreamSendState::NotSent,
             downstream_committed_at: None,
             client_status_code: None,
             delivery_pending: false,
@@ -143,6 +151,8 @@ pub struct ResponseExecutionSession<S: ?Sized> {
     client_api_key_ref: crate::policy::ClientApiKeyId,
     timing_started_at: Instant,
     deadline: SystemTime,
+    /// 会话级 deadline 计时器；deadline 固定，帧循环内复用而非逐事件新建。
+    deadline_timer: Fuse<Delay>,
     pending_request: Option<NewModelRequest>,
     request_persisted: bool,
     operation: Operation,
@@ -156,6 +166,8 @@ pub struct ResponseExecutionSession<S: ?Sized> {
     excluded_accounts: BTreeSet<crate::engine::credential::ProviderAccountId>,
     credential_recovery_attempted_accounts: BTreeSet<crate::engine::credential::ProviderAccountId>,
     current: Option<CurrentAttempt>,
+    /// 请求级发送状态水位；跨 attempt 单调不降，终态写回不得低于此档。
+    send_state_watermark: UpstreamSendState,
     downstream_committed_at: Option<SystemTime>,
     client_status_code: Option<u16>,
     delivery_pending: bool,
@@ -374,6 +386,7 @@ where
                     &mut current.stream,
                     self.cancellation.clone(),
                     self.deadline,
+                    &mut self.deadline_timer,
                 )
                 .await
             };
@@ -614,21 +627,30 @@ where
                 account.clone(),
             ));
         }
-        self.persist_request().await?;
-        self.engine
-            .store()
-            .record_attempt(AttemptRecord {
-                request_id: self.request_id.clone(),
-                attempt_count: next_attempt,
-                trigger,
-                provider_kind: metadata.provider().clone(),
-                provider_account_id: metadata.provider_account_id().cloned(),
-                provider_account_ref: metadata.provider_account_id().cloned(),
-                upstream_model_id: metadata.upstream_model().clone(),
-                upstream_transport: metadata.transport().as_str().to_owned(),
-                http_version: None,
-            })
-            .await?;
+        let attempt_record = AttemptRecord {
+            request_id: self.request_id.clone(),
+            attempt_count: next_attempt,
+            trigger,
+            provider_kind: metadata.provider().clone(),
+            provider_account_id: metadata.provider_account_id().cloned(),
+            provider_account_ref: metadata.provider_account_id().cloned(),
+            upstream_model_id: metadata.upstream_model().clone(),
+            upstream_transport: metadata.transport().as_str().to_owned(),
+            http_version: None,
+        };
+        if self.request_persisted {
+            self.engine.store().record_attempt(attempt_record).await?;
+        } else {
+            let request = self
+                .pending_request
+                .take()
+                .ok_or(EngineError::InvalidDeliveryState)?;
+            self.engine
+                .store()
+                .create_model_request_with_attempt(request, attempt_record)
+                .await?;
+            self.request_persisted = true;
+        }
         self.attempts = next_attempt.get();
         self.current = Some(CurrentAttempt {
             stream,
@@ -642,19 +664,6 @@ where
         Ok(())
     }
 
-    async fn persist_request(&mut self) -> Result<(), EngineError> {
-        if self.request_persisted {
-            return Ok(());
-        }
-        let request = self
-            .pending_request
-            .take()
-            .ok_or(EngineError::InvalidDeliveryState)?;
-        self.engine.store().create_model_request(request).await?;
-        self.request_persisted = true;
-        Ok(())
-    }
-
     async fn observe_event(&mut self, event: &GatewayEvent) -> Result<(), EngineError> {
         let elapsed = elapsed_ms(self.timing_started_at);
         observe_event_timing(&mut self.timings, event, elapsed);
@@ -676,6 +685,7 @@ where
                 .mark_send_state(&self.request_id, UpstreamSendState::Sent)
                 .await?;
             current.send_observed = true;
+            self.send_state_watermark = UpstreamSendState::Sent;
         }
         Ok(())
     }
@@ -688,6 +698,7 @@ where
                 .mark_send_state(&self.request_id, UpstreamSendState::Sent)
                 .await?;
             current.send_observed = true;
+            self.send_state_watermark = UpstreamSendState::Sent;
         }
         Ok(())
     }
@@ -776,11 +787,15 @@ where
             self.last_account_exhaustion = Some(error.clone());
         }
         self.record_provider_failure(current.metadata.provider().clone(), error.kind());
-        let send_state = if current.send_observed {
+        // attempt_send_state 是本 attempt 自身的发送事实，驱动重试门；
+        // 持久化与终态用请求级水位，二者不可混用（水位会把早先 attempt 的
+        // sent 传染给本 attempt，从而错误放行/拦截重试）。
+        let attempt_send_state = if current.send_observed {
             UpstreamSendState::Sent
         } else {
             error.send_state()
         };
+        let send_state = self.raise_send_watermark(attempt_send_state);
         self.engine
             .store()
             .mark_send_state(&self.request_id, send_state)
@@ -789,14 +804,14 @@ where
         let continuation_retry = self.prepare_continuation_retry(
             &current,
             &error,
-            send_state,
+            attempt_send_state,
             provider_proved_replay_safe,
         );
         let ordinary_retry = self.required_account.is_none()
             && self.continuation_attempt == ContinuationAttempt::None
             && self.downstream_committed_at.is_none()
             && !self.delivery_pending
-            && send_state != UpstreamSendState::Ambiguous
+            && attempt_send_state != UpstreamSendState::Ambiguous
             && (self.operation.retry_safety() == RetrySafety::Idempotent
                 || provider_proved_replay_safe)
             && self.attempts < self.plan.max_attempts().get();
@@ -804,7 +819,7 @@ where
             && provider_proved_replay_safe
             && self.downstream_committed_at.is_none()
             && !self.delivery_pending
-            && send_state != UpstreamSendState::Ambiguous
+            && attempt_send_state != UpstreamSendState::Ambiguous
             && self.attempts < self.plan.max_attempts().get()
             && current
                 .metadata
@@ -990,6 +1005,7 @@ where
         error: &ProviderError,
         send_state: UpstreamSendState,
     ) -> Result<(), EngineError> {
+        let send_state = self.raise_send_watermark(send_state);
         let outcome = if error.kind() == ProviderErrorKind::Cancelled {
             ExecutionOutcome::Cancelled
         } else if self.downstream_committed_at.is_some() {
@@ -1042,6 +1058,7 @@ where
         } else {
             UpstreamSendState::Ambiguous
         };
+        let send_state = self.raise_send_watermark(send_state);
         if self.attempts > 0 {
             self.engine
                 .store()
@@ -1162,9 +1179,7 @@ where
     }
 
     fn current_send_state(&self) -> UpstreamSendState {
-        if self.attempts == 0 {
-            UpstreamSendState::NotSent
-        } else if self
+        let observed = if self
             .current
             .as_ref()
             .is_some_and(|current| current.send_observed)
@@ -1172,7 +1187,15 @@ where
             UpstreamSendState::Sent
         } else {
             UpstreamSendState::NotSent
-        }
+        };
+        escalate_send_state(self.send_state_watermark, observed)
+    }
+
+    /// 抬升并返回请求级发送水位；attempt 间切换（`current` 被取走）后，
+    /// 后续终态沿用已达到的最高档，不会把已落库的 `sent` 写回 `not_sent`。
+    fn raise_send_watermark(&mut self, observed: UpstreamSendState) -> UpstreamSendState {
+        self.send_state_watermark = escalate_send_state(self.send_state_watermark, observed);
+        self.send_state_watermark
     }
 
     fn record_current_provider_success(&mut self) {
@@ -1225,17 +1248,17 @@ async fn poll_stream_item(
     stream: &mut ProviderStream,
     cancellation: CancellationToken,
     deadline: SystemTime,
+    mut deadline_timer: &mut Fuse<Delay>,
 ) -> PollBoundary {
-    let Ok(remaining) = deadline.duration_since(SystemTime::now()) else {
+    if SystemTime::now() >= deadline {
         return PollBoundary::Deadline;
-    };
+    }
     let next = stream.next().fuse();
     let cancelled = cancellation.cancelled().fuse();
-    let timeout = Delay::new(remaining).fuse();
-    pin_mut!(next, cancelled, timeout);
+    pin_mut!(next, cancelled);
     select_biased! {
         _ = cancelled => PollBoundary::Cancelled,
-        _ = timeout => PollBoundary::Deadline,
+        _ = deadline_timer => PollBoundary::Deadline,
         item = next => PollBoundary::Item(item),
     }
 }
@@ -1300,6 +1323,18 @@ fn provider_proved_replay_safe(error: &ProviderError) -> bool {
         || (error.send_state() != UpstreamSendState::Ambiguous && error.replay_is_safe())
 }
 
+/// 发送状态合并档位：`Sent` > `Ambiguous` > `NotSent`。
+/// 只要任一 attempt 达到过高档，请求整体就不允许回落到低档。
+const fn escalate_send_state(a: UpstreamSendState, b: UpstreamSendState) -> UpstreamSendState {
+    match (a, b) {
+        (UpstreamSendState::Sent, _) | (_, UpstreamSendState::Sent) => UpstreamSendState::Sent,
+        (UpstreamSendState::Ambiguous, _) | (_, UpstreamSendState::Ambiguous) => {
+            UpstreamSendState::Ambiguous
+        }
+        (UpstreamSendState::NotSent, UpstreamSendState::NotSent) => UpstreamSendState::NotSent,
+    }
+}
+
 fn observe_event_timing(timings: &mut ModelRequestTimings, event: &GatewayEvent, elapsed_ms: u64) {
     timings.first_event_ms.get_or_insert(elapsed_ms);
     match event {

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

@@ -22,7 +22,8 @@ use crate::engine::provider::ProviderRegistry;
 use crate::engine::{
     AttemptCoordinator, AttemptRecord, CancellationToken, CommitRequirement, CoordinatedEvent,
     EngineError, ExecutionStore, GatewayEngine, IntermediateFailure, ModelRequestFinalization,
-    ModelRequestId, NewModelRequest, ProviderAttemptOutcome, RecoveryReport, UpstreamSendState,
+    ModelRequestId, NewModelRequest, ProbeFailure, ProviderAccountId, ProviderAttemptOutcome,
+    RecoveryReport, UpstreamSendState,
 };
 use crate::error::{GatewayError, GatewayErrorKind, ProviderErrorKind, StoreError};
 use crate::event::{GatewayEvent, ProviderEvent, ProviderResponseHeader};
@@ -31,6 +32,7 @@ use crate::policy::{ClientApiKeyId, ClientPolicy};
 use crate::routing::snapshot::RuntimeSnapshotHandle;
 use crate::routing::{
     ProviderKind, PublicModelId, PublicModelProfile, RoutingContext, RuntimeSnapshot,
+    UpstreamModelId,
 };
 
 const MODEL_REQUEST_DEADLINE: Duration = Duration::from_secs(10 * 60);
@@ -187,6 +189,8 @@ pub struct DefaultExecutionService {
     snapshots: RuntimeSnapshotHandle,
     coordinator: Arc<AttemptCoordinator<dyn ExecutionStore>>,
     probe_coordinator: Arc<AttemptCoordinator<dyn ExecutionStore>>,
+    /// probe 自身走 transient store，探测失败仍写入持久 store 的 ops_events。
+    observations: Arc<dyn ExecutionStore>,
     providers: ProviderRegistry,
     admissions: Arc<dyn ClientAdmissionPort>,
     circuits: Arc<dyn ProviderCircuitPort>,
@@ -205,6 +209,7 @@ impl DefaultExecutionService {
         continuation: Arc<dyn NativeContinuationPort>,
         client_api_key_usage: Arc<dyn ClientApiKeyUsageSink>,
     ) -> Self {
+        let observations = Arc::clone(&execution);
         let engine = GatewayEngine::<dyn ExecutionStore>::new(execution, providers.clone());
         let transient: Arc<dyn ExecutionStore> = Arc::new(TransientExecutionStore);
         let probe_engine = GatewayEngine::<dyn ExecutionStore>::new(transient, providers.clone());
@@ -212,6 +217,7 @@ impl DefaultExecutionService {
             snapshots,
             coordinator: Arc::new(AttemptCoordinator::new(engine)),
             probe_coordinator: Arc::new(AttemptCoordinator::new(probe_engine)),
+            observations,
             providers,
             admissions,
             circuits,
@@ -417,6 +423,11 @@ impl DefaultExecutionService {
             upstream_model,
             operation,
         } = request;
+        let observed = ProbeObservation {
+            provider_kind: provider_kind.clone(),
+            account_id: account_id.clone(),
+            upstream_model: upstream_model.clone(),
+        };
         let snapshot = self.snapshots.acquire().map_err(|_| {
             GatewayError::new(
                 GatewayErrorKind::Internal,
@@ -464,7 +475,7 @@ impl DefaultExecutionService {
             started_at,
             deadline_at,
         };
-        let mut session = self
+        let mut session = match self
             .probe_coordinator
             .start(
                 new_request,
@@ -475,18 +486,33 @@ impl DefaultExecutionService {
                 CancellationToken::new(),
             )
             .await
-            .map_err(|error| gateway_error_from_engine(&error))?;
+        {
+            Ok(session) => session,
+            Err(error) => {
+                return Err(self
+                    .observe_probe_failure(&observed, started_at, &error)
+                    .await);
+            }
+        };
         let events = session.collect_uncommitted().await;
         publish_provider_attempt_outcomes(
             self.circuits.as_ref(),
             session.provider_attempt_outcomes(),
         )
         .await;
-        let events = events.map_err(|error| gateway_error_from_engine(&error))?;
-        session
-            .commit_downstream(Some(200))
-            .await
-            .map_err(|error| gateway_error_from_engine(&error))?;
+        let events = match events {
+            Ok(events) => events,
+            Err(error) => {
+                return Err(self
+                    .observe_probe_failure(&observed, started_at, &error)
+                    .await);
+            }
+        };
+        if let Err(error) = session.commit_downstream(Some(200)).await {
+            return Err(self
+                .observe_probe_failure(&observed, started_at, &error)
+                .await);
+        }
         Ok(AccountProbeResult {
             text: events
                 .into_iter()
@@ -498,6 +524,34 @@ impl DefaultExecutionService {
                 .collect(),
         })
     }
+
+    /// 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。
+    async fn observe_probe_failure(
+        &self,
+        observed: &ProbeObservation,
+        started_at: SystemTime,
+        error: &EngineError,
+    ) -> GatewayError {
+        if let EngineError::Provider(provider_error) = error {
+            let _ = self
+                .observations
+                .record_probe_failure(ProbeFailure {
+                    provider_kind: observed.provider_kind.clone(),
+                    account_id: observed.account_id.clone(),
+                    upstream_model_id: observed.upstream_model.clone(),
+                    error: provider_error.clone(),
+                    latency: started_at.elapsed().unwrap_or_default(),
+                })
+                .await;
+        }
+        gateway_error_from_engine(error)
+    }
+}
+
+struct ProbeObservation {
+    provider_kind: ProviderKind,
+    account_id: ProviderAccountId,
+    upstream_model: UpstreamModelId,
 }
 
 struct TransientExecutionStore;
@@ -868,7 +922,7 @@ fn attach_continuation_session_state(operation: &mut Operation, pin: &NativeCont
     {
         return;
     }
-    *operation = operation.clone().with_provider_session_state(state.clone());
+    operation.set_provider_session_state(state.clone());
 }
 
 pub fn gateway_error_from_engine(error: &EngineError) -> GatewayError {

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

@@ -543,6 +543,16 @@ pub struct IntermediateFailure {
     pub latency: Duration,
 }
 
+/// 不属于任何 `model_requests` 行的管理端账号探测失败。
+#[derive(Debug)]
+pub struct ProbeFailure {
+    pub provider_kind: ProviderKind,
+    pub account_id: ProviderAccountId,
+    pub upstream_model_id: UpstreamModelId,
+    pub error: ProviderError,
+    pub latency: Duration,
+}
+
 /// `model_requests` 可用的毫秒级阶段耗时。
 #[derive(Debug, Clone, Default, PartialEq, Eq)]
 pub struct ModelRequestTimings {
@@ -596,6 +606,19 @@ pub struct RecoveryReport {
 pub trait ExecutionStore: Send + Sync {
     async fn create_model_request(&self, request: NewModelRequest) -> Result<(), StoreError>;
     async fn record_attempt(&self, attempt: AttemptRecord) -> Result<(), StoreError>;
+    /// 请求插入与首次 attempt 合并持久化；两者写同一行，支持合并写的
+    /// store 可覆写为单次往返，缩短首 token 前的关键路径。
+    ///
+    /// 覆写实现应保证原子性：默认的两步实现里第二步失败会留下一条
+    /// `running` 请求行，由 deadline 回收器收敛。
+    async fn create_model_request_with_attempt(
+        &self,
+        request: NewModelRequest,
+        attempt: AttemptRecord,
+    ) -> Result<(), StoreError> {
+        self.create_model_request(request).await?;
+        self.record_attempt(attempt).await
+    }
     async fn mark_send_state(
         &self,
         request_id: &ModelRequestId,
@@ -616,6 +639,10 @@ pub trait ExecutionStore: Send + Sync {
         &self,
         failure: IntermediateFailure,
     ) -> Result<(), StoreError>;
+    /// 记录不挂在 `model_requests` 上的账号探测失败；默认丢弃。
+    async fn record_probe_failure(&self, _failure: ProbeFailure) -> Result<(), StoreError> {
+        Ok(())
+    }
     async fn finalize_model_request(
         &self,
         finalization: ModelRequestFinalization,

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

@@ -715,12 +715,6 @@ impl ProviderEvent {
         &self.canonical
     }
 
-    /// 返回 Core 可改写客户端响应身份的全部 facts。
-    #[must_use]
-    pub fn canonical_facts_mut(&mut self) -> &mut [GatewayEvent] {
-        &mut self.canonical
-    }
-
     /// 返回协议原生表达。
     #[must_use]
     pub fn wire_event(&self) -> Option<&ProtocolWireEvent> {

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

@@ -269,10 +269,15 @@ impl GenerateRequest {
     /// 附着同一客户端连接上一轮由 Provider 返回的不透明状态。
     #[must_use]
     pub fn with_provider_session_state(mut self, state: ProviderSessionState) -> Self {
-        Arc::make_mut(&mut self.payload).provider_session_state = Some(state);
+        self.set_provider_session_state(state);
         self
     }
 
+    /// 原地附着会话状态；payload 独占时不复制正文。
+    pub fn set_provider_session_state(&mut self, state: ProviderSessionState) {
+        Arc::make_mut(&mut self.payload).provider_session_state = Some(state);
+    }
+
     /// 返回最大输出 token 数。
     #[must_use]
     pub fn max_output_tokens(&self) -> Option<u64> {
@@ -597,6 +602,13 @@ impl Operation {
         }
     }
 
+    /// 原地附着 Provider 私有状态；payload 独占时不复制正文。
+    pub fn set_provider_session_state(&mut self, state: ProviderSessionState) {
+        if let Self::Generate(request) = self {
+            request.set_provider_session_state(state);
+        }
+    }
+
     /// 返回稳定 operation 分类。
     #[must_use]
     pub const fn kind(&self) -> OperationKind {

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

@@ -474,6 +474,19 @@ pub trait ProviderCredentialStatePort: Send + Sync {
         &'a self,
         account_id: &'a ProviderAccountId,
     ) -> BoxFuture<'a, Result<bool, ProviderStoreError>>;
+
+    /// 记录一次瞬态刷新失败并返回窗口内累计失败次数；每次失败刷新 TTL。
+    fn record_refresh_backoff<'a>(
+        &'a self,
+        account_id: &'a ProviderAccountId,
+        window: Duration,
+    ) -> BoxFuture<'a, Result<u32, ProviderStoreError>>;
+
+    /// 凭据完整成功轮换后清零失败计数，退避窗口重新从 base 起步。
+    fn clear_refresh_backoff<'a>(
+        &'a self,
+        account_id: &'a ProviderAccountId,
+    ) -> BoxFuture<'a, Result<(), ProviderStoreError>>;
 }
 
 /// 临时 cooldown 只保存可丢失的调度截止时间，不进入账号持久状态。
@@ -592,6 +605,33 @@ impl ProviderRefreshPolicy {
     }
 }
 
+/// 指数退避基准延迟；attempt=1 即为该值。
+const REFRESH_BACKOFF_BASE_DELAY: Duration = Duration::from_secs(5);
+/// 每多一次连续失败，基准延迟乘以该因子。
+const REFRESH_BACKOFF_FACTOR: u32 = 3;
+/// 退避延迟上限，避免连续失败时无限增长。
+const REFRESH_BACKOFF_CAP: Duration = Duration::from_secs(300);
+
+/// 基于连续失败计数的指数退避重试时刻；复用 `provider_refresh_retry_at` 的稳定扰动。
+///
+/// `attempt` 为窗口内累计失败次数（0 与 1 等价，均取基准延迟）。延迟按
+/// `base * factor^(attempt-1)` 增长并封顶到 `REFRESH_BACKOFF_CAP`。
+pub fn provider_refresh_backoff_at(
+    account_id: &ProviderAccountId,
+    observed_at: SystemTime,
+    attempt: u32,
+    reason: &'static str,
+) -> Result<SystemTime, ProviderStoreError> {
+    let exponent = attempt.saturating_sub(1);
+    let multiplier = REFRESH_BACKOFF_FACTOR.saturating_pow(exponent);
+    let scaled_seconds = REFRESH_BACKOFF_BASE_DELAY
+        .as_secs()
+        .saturating_mul(u64::from(multiplier))
+        .min(REFRESH_BACKOFF_CAP.as_secs());
+    let base_delay = Duration::from_secs(scaled_seconds);
+    provider_refresh_retry_at(account_id, observed_at, base_delay, reason)
+}
+
 /// 临时失败后的持久重试时刻；稳定扰动避免多实例同频重试。
 pub fn provider_refresh_retry_at(
     account_id: &ProviderAccountId,

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

@@ -1620,6 +1620,8 @@ fn rate_limited_account_exhaustion_survives_a_later_empty_selection() {
     assert_eq!(state.attempts.len(), 1);
     assert_eq!(state.intermediate_failures, 1);
     assert_eq!(state.finalizations.len(), 1);
+    // attempt-1 已把 sent 落库；attempt-2 空选路终态不得降级回 not_sent。
+    assert_eq!(state.finalizations[0].send_state, UpstreamSendState::Sent);
     assert_eq!(state.finalizations[0].upstream_status_code, Some(429));
     assert_eq!(
         state.finalizations[0].provider_error_code.as_deref(),

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

@@ -24,12 +24,15 @@ use gateway_core::engine::execution::{
     ProviderCircuitError, ProviderCircuitPort, provider_failure_affects_circuit,
 };
 use gateway_core::engine::probe::{AccountProbe, AccountProbeRequest};
-use gateway_core::engine::provider::ProviderRegistry;
-use gateway_core::engine::{
-    AttemptRecord, ExecutionStore, IntermediateFailure, ModelRequestFinalization, ModelRequestId,
-    NewModelRequest, RecoveryReport, UpstreamSendState,
+use gateway_core::engine::provider::{
+    Provider, ProviderCatalogGeneration, ProviderModelCapabilities, ProviderRegistry,
+    ProviderRequest, ProviderStream,
 };
-use gateway_core::error::{GatewayErrorKind, ProviderErrorKind, StoreError};
+use gateway_core::engine::{
+    AttemptContext, AttemptRecord, ExecutionStore, IntermediateFailure, ModelRequestFinalization,
+    ModelRequestId, NewModelRequest, ProbeFailure, RecoveryReport, UpstreamSendState,
+};
+use gateway_core::error::{GatewayErrorKind, ProviderError, ProviderErrorKind, StoreError};
 use gateway_core::operation::{GenerateRequest, Operation, OperationKind, ProtocolPayload};
 use gateway_core::policy::{ClientApiKeyId, ClientPolicy, PlaintextClientApiKey, RateLimits};
 use gateway_core::routing::snapshot::RuntimeSnapshotHandle;
@@ -78,6 +81,64 @@ fn account_probe_should_not_write_to_the_persistent_execution_store() {
     assert!(!store.touched.load(Ordering::SeqCst));
 }
 
+#[test]
+fn probe_failures_should_be_observable_without_a_model_request_row() {
+    let store = Arc::new(TrackingExecutionStore::default());
+    let providers = ProviderRegistry::new([Arc::new(FailingProvider) as Arc<dyn Provider>])
+        .expect("provider registry");
+    let service = DefaultExecutionService::new(
+        RuntimeSnapshotHandle::new(probe_snapshot()),
+        store.clone(),
+        providers,
+        Arc::new(UnusedAdmissions),
+        Arc::new(UnusedCircuits),
+        Arc::new(UnusedContinuation),
+        Arc::new(RecordingClientApiKeyUsage::default()),
+    );
+
+    let error = block_on(service.probe(AccountProbeRequest {
+        account_id: ProviderAccountId::new("acct_probe").expect("account ID"),
+        provider_kind: ProviderKind::new("openai").expect("provider kind"),
+        upstream_model: UpstreamModelId::new("gpt-probe").expect("model ID"),
+        operation: probe_operation(),
+    }))
+    .expect_err("the provider rejects every probe");
+
+    assert_eq!(error.kind(), GatewayErrorKind::UpstreamUnavailable);
+    assert!(!store.touched.load(Ordering::SeqCst));
+    assert_eq!(store.probe_failures(), vec!["transport".to_owned()]);
+}
+
+struct FailingProvider;
+
+#[async_trait]
+impl Provider for FailingProvider {
+    fn name(&self) -> &'static str {
+        "openai"
+    }
+
+    fn catalog_generation(&self) -> ProviderCatalogGeneration {
+        ProviderCatalogGeneration::default()
+    }
+
+    async fn query_model_capabilities(
+        &self,
+    ) -> Result<Vec<ProviderModelCapabilities>, ProviderError> {
+        Ok(Vec::new())
+    }
+
+    async fn execute(
+        &self,
+        _request: ProviderRequest,
+        _context: AttemptContext,
+    ) -> Result<ProviderStream, ProviderError> {
+        Err(ProviderError::new(
+            ProviderErrorKind::Transport,
+            UpstreamSendState::NotSent,
+        ))
+    }
+}
+
 #[test]
 fn successful_authentication_should_record_client_key_usage() {
     let usage = Arc::new(RecordingClientApiKeyUsage::default());
@@ -121,12 +182,20 @@ impl ClientApiKeyUsageSink for RecordingClientApiKeyUsage {
 #[derive(Default)]
 struct TrackingExecutionStore {
     touched: AtomicBool,
+    probe_failures: Mutex<Vec<String>>,
 }
 
 impl TrackingExecutionStore {
     fn touch(&self) {
         self.touched.store(true, Ordering::SeqCst);
     }
+
+    fn probe_failures(&self) -> Vec<String> {
+        self.probe_failures
+            .lock()
+            .expect("probe failures lock")
+            .clone()
+    }
 }
 
 #[async_trait]
@@ -170,6 +239,14 @@ impl ExecutionStore for TrackingExecutionStore {
         Ok(())
     }
 
+    async fn record_probe_failure(&self, failure: ProbeFailure) -> Result<(), StoreError> {
+        self.probe_failures
+            .lock()
+            .expect("probe failures lock")
+            .push(failure.error.kind().as_str().to_owned());
+        Ok(())
+    }
+
     async fn finalize_model_request(&self, _: ModelRequestFinalization) -> Result<(), StoreError> {
         self.touch();
         Ok(())

backend/crates/gateway-host/Cargo.toml

@@ -28,7 +28,6 @@ tokio = { version = "1.52.3", default-features = false, features = ["macros", "r
 tracing = "0.1.44"
 tracing-appender = "0.2.5"
 tracing-subscriber = { version = "0.3.23", features = ["env-filter", "json"] }
-uuid = { version = "1.23.3", features = ["v7"] }
 
 [dev-dependencies]
 filetime = "0.2.27"

backend/crates/gateway-host/src/serve.rs

@@ -1,7 +1,7 @@
 //! HTTP 监听、OS signal、原子连接注册与优雅 drain。
 
-use std::sync::Arc;
 use std::sync::atomic::{AtomicUsize, Ordering};
+use std::sync::{Arc, OnceLock};
 use std::time::Duration;
 
 use axum::Router;
@@ -121,16 +121,44 @@ pub(crate) async fn serve_router(
         wait_for_shutdown(&shutdown_cancellation).await;
         shutdown_connections.begin_draining();
     };
-    axum::serve(
+    // axum 的优雅关闭会无限等待存量连接结束，而 SSE 等长响应体不观察
+    // 取消信号：慢客户端可以把关闭拖住直到被 SIGKILL。整个 drain（axum
+    // 优雅关闭 + 游离 guard 等待）共享同一个从取消信号起算的绝对截止点，
+    // 逾期即放弃等待，存量连接随进程退出终止。
+    let drain_deadline_at = Arc::new(OnceLock::new());
+    let drain_deadline = {
+        let cancellation = cancellation.clone();
+        let deadline_at = Arc::clone(&drain_deadline_at);
+        async move {
+            cancellation.cancelled().await;
+            let deadline = tokio::time::Instant::now() + drain_timeout;
+            let _ = deadline_at.set(deadline);
+            tokio::time::sleep_until(deadline).await;
+        }
+    };
+    let serve = axum::serve(
         listener,
         router.into_make_service_with_connect_info::<std::net::SocketAddr>(),
     )
-    .with_graceful_shutdown(shutdown)
-    .await
-    .map_err(ServeError::Serve)?;
-
-    connections.begin_draining();
-    connections.wait_until_idle(drain_timeout).await;
+    .with_graceful_shutdown(shutdown);
+    tokio::select! {
+        biased;
+        result = serve => {
+            result.map_err(ServeError::Serve)?;
+            connections.begin_draining();
+            let remaining = drain_deadline_at.get().map_or(drain_timeout, |deadline| {
+                deadline.saturating_duration_since(tokio::time::Instant::now())
+            });
+            connections.wait_until_idle(remaining).await;
+        }
+        () = drain_deadline => {
+            tracing::warn!(
+                target: "gateway_shutdown",
+                timeout_secs = drain_timeout.as_secs(),
+                "优雅 drain 超时，放弃等待存量连接"
+            );
+        }
+    }
     Ok(())
 }
 

backend/crates/gateway-host/src/system_update/mod.rs

@@ -33,6 +33,7 @@ use self::download::{MAX_CHECKSUM_SIZE, MAX_DOWNLOAD_SIZE, download_file, verify
 use self::process::{environment_value, spawn_replacement};
 use self::release::{
     ReleaseCache, confirmed_target, detail_from_release, fetch_latest, select_archive,
+    versions_cross_major,
 };
 use self::state::{
     OperationFileLock, UpdateTempDir, finish, operation_id, read_status, set_running,
@@ -226,6 +227,11 @@ impl ProcessSystemOperations {
             return Err(conflict(reason));
         }
         let target = confirmed_target(target_version)?;
+        if versions_cross_major(&self.config.version, &target).unwrap_or(false) {
+            let reason = "跨大版本升级不提供自动更新，请按发布说明手动迁移";
+            self.events.error_terminal(None, Some("preflight"), reason);
+            return Err(conflict(reason));
+        }
         let repository = self
             .config
             .update_repository

backend/crates/gateway-host/src/system_update/release.rs

@@ -121,12 +121,13 @@ pub(crate) fn detail_from_release(
     release: &GitHubRelease,
 ) -> SystemUpdateDetail {
     let latest = normalize_version(&release.tag_name);
-    let available = (config.update_channel != "stable" || !release.prerelease)
+    let newer = (config.update_channel != "stable" || !release.prerelease)
         && version_is_newer(&config.version, &latest).unwrap_or(false);
+    let cross_major = versions_cross_major(&config.version, &latest).unwrap_or(false);
     SystemUpdateDetail {
         current_version: config.version.clone(),
-        latest_version: latest,
-        has_update: available,
+        latest_version: latest.clone(),
+        has_update: newer && !cross_major,
         deployment_mode: config.deployment_mode.clone(),
         build_type: config.build_type.clone(),
         release_url: release.html_url.clone(),
@@ -134,7 +135,9 @@ pub(crate) fn detail_from_release(
         cached: false,
         update_supported: config.update_support_error().is_none(),
         unsupported_reason: config.update_support_error(),
-        warning: None,
+        warning: (newer && cross_major).then(|| {
+            format!("检测到新的大版本 v{latest}：跨大版本升级不提供自动更新，请按发布说明手动迁移")
+        }),
     }
 }
 
@@ -265,6 +268,13 @@ fn version_is_newer(current: &str, latest: &str) -> Option<bool> {
     Some(latest > current)
 }
 
+/// 自动更新只在同 major 内提供；跨大版本的配置与数据迁移契约不保证兼容。
+pub(crate) fn versions_cross_major(current: &str, latest: &str) -> Option<bool> {
+    let current = semver::Version::parse(&normalize_version(current)).ok()?;
+    let latest = semver::Version::parse(&normalize_version(latest)).ok()?;
+    Some(current.major != latest.major)
+}
+
 fn repository_character(character: char) -> bool {
     character.is_ascii_alphanumeric() || matches!(character, '-' | '_' | '.')
 }

backend/crates/gateway-host/src/workers/mod.rs

@@ -108,22 +108,19 @@ impl WorkerSupervisor {
         Ok(())
     }
 
-    /// 先协作取消；超时后 abort 不合作任务，lease 依赖 TTL 兜底。
+    /// 先协作取消；所有任务共享同一绝对截止点，逾期任务被 abort，
+    /// lease 依赖 TTL 兜底。每个句柄恰好 join 一次：已在截止点前退出的
+    /// 不会被二次 poll，逾期的在 abort 后等待其真正终止。
     pub async fn shutdown(&self, timeout: Duration) {
         self.cancellation.cancel();
-        let mut handles = std::mem::take(&mut *lock_unpoisoned(&self.handles));
-        let joined = tokio::time::timeout(timeout, async {
-            for (_, handle) in &mut handles {
-                let _ = handle.await;
-            }
-        })
-        .await
-        .is_ok();
-        if !joined {
-            for (_, handle) in &handles {
+        let handles = std::mem::take(&mut *lock_unpoisoned(&self.handles));
+        let deadline = tokio::time::Instant::now() + timeout;
+        for (_, mut handle) in handles {
+            if tokio::time::timeout_at(deadline, &mut handle)
+                .await
+                .is_err()
+            {
                 handle.abort();
-            }
-            for (_, handle) in handles {
                 let _ = handle.await;
             }
         }

backend/crates/gateway-host/tests/system_update/mod.rs

@@ -18,7 +18,8 @@ use tar::{Builder, EntryType, Header};
 use wiremock::matchers::{method, path};
 use wiremock::{Mock, MockServer, ResponseTemplate};
 
-const TARGET_VERSION: &str = "9.9.9";
+const TARGET_VERSION: &str = "1.9.9";
+const CROSS_MAJOR_VERSION: &str = "2.0.0";
 
 #[tokio::test]
 async fn restart_should_not_shutdown_when_replacement_spawn_fails() {
@@ -174,6 +175,43 @@ async fn update_detail_should_use_cached_release_when_not_refreshed() {
     assert!(service.update_detail(false).await.is_ok());
 }
 
+#[tokio::test]
+async fn update_detail_should_withhold_cross_major_release() {
+    let server = MockServer::start().await;
+    let fixture = Fixture::new();
+    fixture
+        .mount_release_once(&server, CROSS_MAJOR_VERSION)
+        .await;
+
+    let detail = fixture
+        .service(&server)
+        .update_detail(true)
+        .await
+        .expect("detail");
+    assert_eq!(detail.latest_version, CROSS_MAJOR_VERSION);
+    assert!(!detail.has_update);
+    assert!(detail.warning.is_some());
+}
+
+#[tokio::test]
+async fn update_should_reject_cross_major_target_before_fetching_release() {
+    let fixture = Fixture::new();
+    let service = ProcessSystemOperations::new(
+        CancellationToken::new(),
+        fixture.config("http://127.0.0.1:1/repos"),
+    );
+
+    let error = service
+        .perform_update(Some(CROSS_MAJOR_VERSION.to_owned()))
+        .await
+        .expect_err("cross-major target must be rejected");
+    assert_eq!(error.kind(), SystemOperationErrorKind::Conflict);
+    assert_eq!(
+        fs::read(fixture.executable()).expect("binary"),
+        b"old-binary"
+    );
+}
+
 #[tokio::test]
 async fn update_events_should_close_after_terminal_update_log() {
     let fixture = Fixture::new();
@@ -397,7 +435,7 @@ async fn update_should_reject_when_confirmed_target_differs_from_remote_latest()
     let server = MockServer::start().await;
     let fixture = Fixture::new();
     fixture
-        .mount_release(&server, "9.9.8", ArchiveKind::Safe, ChecksumKind::Valid)
+        .mount_release(&server, "1.9.8", ArchiveKind::Safe, ChecksumKind::Valid)
         .await;
 
     assert!(

backend/crates/gateway-host/tests/workers/mod.rs

@@ -626,6 +626,55 @@ async fn shutdown_timeout_aborts_uncooperative_task_and_drops_guard() {
     );
 }
 
+#[tokio::test]
+async fn shutdown_timeout_must_not_rejoin_workers_that_already_exited() {
+    // 挂死任务放在 join 顺序最末的 kind（Retention），保证超时触发时
+    // 排在它前面的合作任务句柄都已被 join 过一次。
+    let hung = HungTask {
+        entered: Arc::new(AtomicBool::new(false)),
+        notification: Arc::new(Notify::new()),
+    };
+    let cooperative = CancellationTask {
+        entered: Arc::new(AtomicUsize::new(0)),
+        observed: Arc::new(AtomicUsize::new(0)),
+    };
+    let mut plan = ACTIVE_KINDS
+        .into_iter()
+        .map(|kind| {
+            if kind == WorkerKind::Retention {
+                registration(kind, kind.as_str(), hung.clone(), test_schedule(), true)
+            } else {
+                registration(
+                    kind,
+                    kind.as_str(),
+                    cooperative.clone(),
+                    test_schedule(),
+                    true,
+                )
+            }
+        })
+        .collect::<Vec<_>>();
+    plan.push(WorkerContribution::Disabled {
+        kind: WorkerKind::OpsFlush,
+        reason: WorkerDisabledReason::NoBufferedOpsEvents,
+    });
+    let leases = Arc::new(FakeLeasePort::default());
+    let supervisor = supervisor();
+    supervisor
+        .start(plan, leases.clone())
+        .expect("complete plan");
+    hung.notification.notified().await;
+    wait_until(|| cooperative.entered.load(Ordering::SeqCst) == ACTIVE_KINDS.len() - 1).await;
+
+    supervisor.shutdown(Duration::from_millis(100)).await;
+
+    assert_eq!(
+        cooperative.observed.load(Ordering::SeqCst),
+        ACTIVE_KINDS.len() - 1
+    );
+    assert!(lock_unpoisoned(&leases.active_guards).is_empty());
+}
+
 #[tokio::test]
 async fn running_worker_becomes_unhealthy_when_last_success_is_stale() {
     let task = SucceedsThenHangsTask {

backend/crates/gateway-protocol/src/openai/sse.rs

@@ -72,26 +72,50 @@ pub const DONE_SSE_FRAME: &str = "data: [DONE]\n\n";
 #[derive(Debug, Default)]
 pub struct SseEventDecoder {
     pending: Vec<u8>,
+    /// `pending` 中已确认不含帧分隔符的前缀长度。新数据到达时从该边界
+    /// 回退分隔符最大跨界宽度继续扫描，大帧跨多个 chunk 不会被反复重扫。
+    scanned: usize,
 }
 
 impl SseEventDecoder {
+    /// 返回从 `consumed` 起的下一个完整帧结束位置（含分隔符），并推进扫描边界。
+    fn next_frame_end(&mut self, consumed: usize) -> Option<usize> {
+        // `\r\n\r\n` 分隔符可能跨 chunk 边界，扫描起点回退 3 字节。
+        let from = self.scanned.saturating_sub(3).max(consumed);
+        match sse_frame_separator_bytes(&self.pending[from..]) {
+            Some((position, separator_len)) => {
+                let end = from + position + separator_len;
+                self.scanned = end;
+                Some(end)
+            }
+            None => {
+                self.scanned = self.pending.len();
+                None
+            }
+        }
+    }
+
+    fn commit_consumed(&mut self, consumed: usize) {
+        if consumed != 0 {
+            self.pending.drain(..consumed);
+            self.scanned = self.scanned.saturating_sub(consumed);
+        }
+    }
+
     /// 追加一个字节块并返回其中已经完整的事件。
     pub fn push(&mut self, chunk: &[u8]) -> Result<Vec<SseEvent>, SseError> {
         self.pending.extend_from_slice(chunk);
         let mut events = Vec::new();
         let mut consumed = 0usize;
 
-        while let Some(frame_len) = sse_frame_end(&self.pending[consumed..]) {
-            let end = consumed.saturating_add(frame_len);
+        while let Some(end) = self.next_frame_end(consumed) {
             let frame = std::str::from_utf8(&self.pending[consumed..end])
                 .map_err(|error| SseError::ParseError(error.to_string()))?;
             events.extend(parse_sse_events(frame)?);
             consumed = end;
         }
 
-        if consumed != 0 {
-            self.pending.drain(..consumed);
-        }
+        self.commit_consumed(consumed);
         if self.pending.len() > MAX_SSE_EVENT_BUFFER_BYTES {
             return Err(SseError::BufferExceeded {
                 max_bytes: MAX_SSE_EVENT_BUFFER_BYTES,
@@ -110,8 +134,7 @@ impl SseEventDecoder {
         let mut frames = Vec::new();
         let mut consumed = 0usize;
 
-        while let Some(frame_len) = sse_frame_end(&self.pending[consumed..]) {
-            let end = consumed.saturating_add(frame_len);
+        while let Some(end) = self.next_frame_end(consumed) {
             let raw = self.pending[consumed..end].to_vec();
             let events = std::str::from_utf8(&raw)
                 .ok()
@@ -121,9 +144,7 @@ impl SseEventDecoder {
             consumed = end;
         }
 
-        if consumed != 0 {
-            self.pending.drain(..consumed);
-        }
+        self.commit_consumed(consumed);
         if self.pending.len() > MAX_SSE_EVENT_BUFFER_BYTES {
             return Err(SseError::BufferExceeded {
                 max_bytes: MAX_SSE_EVENT_BUFFER_BYTES,
@@ -134,6 +155,7 @@ impl SseEventDecoder {
 
     /// 流结束时解析尚未带空行分隔符的最后一帧。
     pub fn finish(&mut self) -> Result<Vec<SseEvent>, SseError> {
+        self.scanned = 0;
         if self.pending.is_empty() {
             return Ok(Vec::new());
         }
@@ -147,6 +169,7 @@ impl SseEventDecoder {
     ///
     /// 与 [`Self::push_frames`] 一致，旁路解析失败不会丢弃原始帧。
     pub fn finish_frames(&mut self) -> Result<Vec<SseFrame>, SseError> {
+        self.scanned = 0;
         if self.pending.is_empty() {
             return Ok(Vec::new());
         }
@@ -235,9 +258,6 @@ pub enum SseError {
     /// 解析错误。
     #[error("SSE parse error: {0}")]
     ParseError(String),
-    /// 用量提取错误。
-    #[error("Usage extraction error: {0}")]
-    UsageExtraction(String),
 }
 
 #[derive(Debug, Default)]

backend/crates/gateway-protocol/tests/openai/sse.rs

@@ -107,6 +107,29 @@ fn incremental_decoder_should_decode_across_arbitrary_chunk_boundaries() {
     );
 }
 
+#[test]
+fn incremental_decoder_should_find_separators_split_at_every_byte_boundary() {
+    // CRLF 分隔符按单字节到达时必须仍被识别；帧内容跨任意多次 push 不丢失。
+    let body = b"event: one\r\ndata: first\r\n\r\nevent: two\ndata: second\n\ndata: third\r\n\r\n";
+    let mut decoder = SseEventDecoder::default();
+    let mut events = Vec::new();
+    for byte in body {
+        events.extend(decoder.push(&[*byte]).expect("single-byte chunk"));
+    }
+
+    assert_eq!(
+        events
+            .iter()
+            .map(|event| (event.event.as_deref(), event.data.as_str()))
+            .collect::<Vec<_>>(),
+        vec![
+            (Some("one"), "first"),
+            (Some("two"), "second"),
+            (None, "third"),
+        ]
+    );
+}
+
 #[test]
 fn raw_frame_decoder_should_preserve_exact_frame_across_chunk_boundaries() {
     let raw = b": keep-alive\r\nid: evt_1\r\nevent: response.future\r\nretry: 250\r\ndata: { \"type\": \"response.future\", \"opaque\": true }\r\n\r\n";

backend/crates/gateway-store/src/lib.rs

@@ -86,6 +86,37 @@ pub type StoreResult<T> = Result<T, StoreError>;
 pub struct StoreConfig {
     database: StoreConnectionConfig,
     redis: StoreConnectionConfig,
+    #[serde(default)]
+    pool: StorePoolConfig,
+}
+
+/// PostgreSQL 连接池预算；acquire 超时决定池耗尽时快速失败而非排队积压。
+#[derive(Debug, Clone, Copy, PartialEq, Eq, Deserialize)]
+#[serde(deny_unknown_fields, default)]
+pub struct StorePoolConfig {
+    pub max_connections: u32,
+    pub acquire_timeout_seconds: u64,
+}
+
+impl Default for StorePoolConfig {
+    fn default() -> Self {
+        Self {
+            max_connections: 10,
+            acquire_timeout_seconds: 5,
+        }
+    }
+}
+
+impl StorePoolConfig {
+    fn validate(&self) -> StoreResult<()> {
+        if self.max_connections == 0 || self.acquire_timeout_seconds == 0 {
+            return Err(StoreError::InvalidData {
+                entity: "store config",
+                message: "pool limits must be positive".to_owned(),
+            });
+        }
+        Ok(())
+    }
 }
 
 impl StoreConfig {
@@ -104,6 +135,7 @@ impl StoreConfig {
         }
         self.database.validate("database")?;
         self.redis.validate("redis")?;
+        self.pool.validate()?;
         Ok(())
     }
 
@@ -227,7 +259,7 @@ pub async fn initialize(mut config: StoreConfig) -> StoreResult<StoreBundle> {
     const REDIS_NAMESPACE: &str = "codex-proxy-rs";
 
     config.resolve_and_validate(std::path::Path::new("."))?;
-    let pool = postgres::connect_and_migrate(&config.database_url()?).await?;
+    let pool = postgres::connect_and_migrate(&config.database_url()?, config.pool).await?;
     let redis_client = ::redis::Client::open(config.redis_url()?)
         .map_err(|_| redis_unavailable("create Redis client"))?;
     let redis_connection = redis_client
@@ -823,44 +855,6 @@ impl AuthStore for AdminAuthStoreAdapter {
             .map_err(|error| admin_store_error("admin session", error))
     }
 
-    async fn login_source_is_throttled(
-        &self,
-        source: &str,
-        failure_limit: u32,
-        window_seconds: u64,
-    ) -> AdminStoreResult<bool> {
-        redis::AdminAuthStateRepository::login_source_is_throttled(
-            &self.state,
-            source,
-            failure_limit,
-            window_seconds,
-        )
-        .await
-        .map_err(|error| admin_store_error("admin login throttle", error))
-    }
-
-    async fn record_login_failure(
-        &self,
-        source: &str,
-        failure_limit: u32,
-        window_seconds: u64,
-    ) -> AdminStoreResult<bool> {
-        redis::AdminAuthStateRepository::record_login_failure(
-            &self.state,
-            source,
-            failure_limit,
-            window_seconds,
-        )
-        .await
-        .map_err(|error| admin_store_error("admin login throttle", error))
-    }
-
-    async fn clear_login_failures(&self, source: &str) -> AdminStoreResult<()> {
-        redis::AdminAuthStateRepository::clear_login_failures(&self.state, source)
-            .await
-            .map_err(|error| admin_store_error("admin login throttle", error))
-    }
-
     async fn append_audit_event(&self, event: AdminAuditModel) -> AdminStoreResult<()> {
         let config_revision = event
             .config_revision

backend/crates/gateway-store/src/postgres/client_keys.rs

@@ -310,17 +310,12 @@ impl UpdateClientApiKeyDetails {
 
 #[async_trait]
 pub trait ClientApiKeyRepository: Send + Sync {
-    async fn authenticate_client_api_key(
-        &self,
-        key: &str,
-    ) -> StoreResult<Option<ClientApiKeySecret>>;
     async fn list_client_api_keys(
         &self,
         query: ClientApiKeyListQuery,
     ) -> StoreResult<ClientApiKeyPage>;
     async fn get_client_api_key(&self, id: &str) -> StoreResult<Option<ClientApiKeyRecord>>;
     async fn reveal_client_api_key(&self, id: &str) -> StoreResult<Option<ClientApiKeySecret>>;
-    async fn insert_client_api_key(&self, key: NewClientApiKey) -> StoreResult<()>;
     async fn update_client_api_key(&self, key: UpdateClientApiKey) -> StoreResult<bool>;
     async fn touch_client_api_keys(
         &self,
@@ -342,22 +337,6 @@ impl PgClientApiKeyRepository {
 
 #[async_trait]
 impl ClientApiKeyRepository for PgClientApiKeyRepository {
-    async fn authenticate_client_api_key(
-        &self,
-        key: &str,
-    ) -> StoreResult<Option<ClientApiKeySecret>> {
-        validate_key(key)?;
-        let row = sqlx::query_as::<_, (String, String, String, bool, i64, i64)>(
-            "select id, key, provider_kind, enabled, max_concurrency, requests_per_minute
-             from client_api_keys where key = $1",
-        )
-        .bind(key)
-        .fetch_optional(&self.pool)
-        .await
-        .map_err(|_| postgres_unavailable("authenticate client API key"))?;
-        row.map(client_secret_from_row).transpose()
-    }
-
     async fn list_client_api_keys(
         &self,
         query: ClientApiKeyListQuery,
@@ -432,27 +411,6 @@ impl ClientApiKeyRepository for PgClientApiKeyRepository {
         .transpose()
     }
 
-    async fn insert_client_api_key(&self, key: NewClientApiKey) -> StoreResult<()> {
-        key.validate()?;
-        sqlx::query(
-            "insert into client_api_keys (
-               id, name, label, provider_kind, key, enabled, max_concurrency, requests_per_minute,
-               last_used_at, created_at, updated_at
-             ) values ($1, $2, $3, $4, $5, true, $6, $7, null, now(), now())",
-        )
-        .bind(key.id)
-        .bind(key.name)
-        .bind(key.label)
-        .bind(key.provider_kind)
-        .bind(key.key)
-        .bind(to_i64(key.max_concurrency)?)
-        .bind(to_i64(key.requests_per_minute)?)
-        .execute(&self.pool)
-        .await
-        .map_err(|_| postgres_unavailable("insert client API key"))?;
-        Ok(())
-    }
-
     async fn update_client_api_key(&self, key: UpdateClientApiKey) -> StoreResult<bool> {
         require_nonempty(ENTITY, "id", &key.id)?;
         require_nonempty(ENTITY, "name", &key.name)?;

backend/crates/gateway-store/src/postgres/execution.rs

@@ -10,7 +10,7 @@ use gateway_core::accounting::{CostSource as CoreCostSource, Usage as CoreUsage}
 use gateway_core::engine::{
     AttemptRecord as CoreAttemptRecord, ExecutionStore, IntermediateFailure,
     ModelRequestFinalization as CoreModelRequestFinalization, ModelRequestId,
-    NewModelRequest as CoreNewModelRequest, RecoveryReport as CoreRecoveryReport,
+    NewModelRequest as CoreNewModelRequest, ProbeFailure, RecoveryReport as CoreRecoveryReport,
     UpstreamSendState as CoreUpstreamSendState,
 };
 use gateway_core::error::{StoreError as CoreStoreError, StoreErrorKind as CoreStoreErrorKind};
@@ -295,6 +295,12 @@ pub struct ModelRequestRecoveryReport {
 #[async_trait]
 pub trait ModelRequestRepository: Send + Sync {
     async fn insert_model_request(&self, request: NewModelRequest) -> StoreResult<()>;
+    /// 请求行与首次 attempt 列一次插入；`attempt.attempt_count` 必须为 1。
+    async fn insert_model_request_with_first_attempt(
+        &self,
+        request: NewModelRequest,
+        attempt: ModelRequestAttemptStart,
+    ) -> StoreResult<()>;
     async fn begin_model_request_attempt(
         &self,
         attempt: ModelRequestAttemptStart,
@@ -383,11 +389,71 @@ impl ModelRequestRepository for PgExecutionStore {
         Ok(())
     }
 
+    async fn insert_model_request_with_first_attempt(
+        &self,
+        request: NewModelRequest,
+        attempt: ModelRequestAttemptStart,
+    ) -> StoreResult<()> {
+        request.validate()?;
+        attempt.validate()?;
+        if attempt.attempt_count != 1 || attempt.model_request_id != request.id {
+            return Err(invalid("first attempt must target the inserted request"));
+        }
+        sqlx::query(
+            "insert into model_requests (
+               id, client_api_key_id, client_api_key_ref, config_revision, protocol,
+               operation, endpoint, client_transport, requested_model_id,
+               client_ip, user_agent, reasoning_effort,
+               reasoning_preset, request_kind, subagent_kind, compact,
+               image_generation_requested, started_at, deadline_at,
+               provider_kind, provider_account_id, provider_account_ref,
+               upstream_model_id, upstream_transport, http_version,
+               attempt_count, upstream_send_state
+             ) values (
+               $1, $2, $3, $4, $5, $6, $7, $8, $9, $10::inet, $11, $12,
+               $13, $14, $15, $16, $17, $18, $19,
+               $20, $21, $22, $23, $24, $25, 1, 'not_sent'
+             )",
+        )
+        .bind(request.id)
+        .bind(request.client_api_key_id)
+        .bind(request.client_api_key_ref)
+        .bind(to_i64(request.config_revision, "config_revision")?)
+        .bind(request.protocol)
+        .bind(request.operation)
+        .bind(request.endpoint)
+        .bind(request.client_transport)
+        .bind(request.requested_model_id)
+        .bind(request.client_ip)
+        .bind(request.user_agent)
+        .bind(request.reasoning_effort)
+        .bind(request.reasoning_preset)
+        .bind(request.request_kind)
+        .bind(request.subagent_kind)
+        .bind(request.compact)
+        .bind(request.image_generation_requested)
+        .bind(request.started_at)
+        .bind(request.deadline_at)
+        .bind(attempt.provider_kind)
+        .bind(attempt.provider_account_id)
+        .bind(attempt.provider_account_ref)
+        .bind(attempt.upstream_model_id)
+        .bind(attempt.upstream_transport)
+        .bind(attempt.http_version)
+        .execute(&self.pool)
+        .await
+        .map_err(|_| postgres_unavailable("insert model request with first attempt"))?;
+        Ok(())
+    }
+
     async fn begin_model_request_attempt(
         &self,
         attempt: ModelRequestAttemptStart,
     ) -> StoreResult<u32> {
         attempt.validate()?;
+        // upstream_send_state 是请求级单调水位（sent > ambiguous > not_sent），
+        // 由 mark/finalize 抬升；开启新 attempt 不得把已持久化的 sent 重置回
+        // not_sent——崩溃恢复的终态写回会原样继承本列。
         let count = sqlx::query_scalar::<_, i32>(
             "update model_requests
              set provider_kind = $2,
@@ -396,8 +462,7 @@ impl ModelRequestRepository for PgExecutionStore {
                  upstream_model_id = $5,
                  upstream_transport = $6,
                  http_version = $7,
-                 attempt_count = $8,
-                 upstream_send_state = 'not_sent'
+                 attempt_count = $8
              where id = $1 and outcome = 'running' and downstream_committed_at is null
                and $8 = attempt_count + 1
              returning attempt_count",
@@ -632,53 +697,15 @@ impl ExecutionStore for PgExecutionStore {
         &self,
         request: CoreNewModelRequest,
     ) -> Result<(), CoreStoreError> {
-        self.insert_model_request(NewModelRequest {
-            id: request.id.as_str().to_owned(),
-            client_api_key_id: request
-                .client_api_key_id
-                .as_ref()
-                .map(|id| id.as_str().to_owned()),
-            client_api_key_ref: request.client_api_key_ref.as_str().to_owned(),
-            config_revision: request.config_revision.get(),
-            protocol: request.protocol,
-            operation: request.operation.as_str().to_owned(),
-            endpoint: request.endpoint,
-            client_transport: request.client_transport,
-            requested_model_id: request.requested_model.as_str().to_owned(),
-            client_ip: request.client_ip.map(|address| address.to_string()),
-            user_agent: request.user_agent,
-            reasoning_effort: request.reasoning_effort,
-            reasoning_preset: request.reasoning_preset,
-            request_kind: request.request_kind,
-            subagent_kind: request.subagent_kind,
-            compact: request.compact,
-            image_generation_requested: request.image_generation_requested,
-            started_at: DateTime::<Utc>::from(request.started_at),
-            deadline_at: DateTime::<Utc>::from(request.deadline_at),
-        })
-        .await
-        .map_err(core_store_error)
+        self.insert_model_request(new_model_request_row(request))
+            .await
+            .map_err(core_store_error)
     }
 
     async fn record_attempt(&self, attempt: CoreAttemptRecord) -> Result<(), CoreStoreError> {
         let expected_count = attempt.attempt_count.get();
         let persisted = self
-            .begin_model_request_attempt(ModelRequestAttemptStart {
-                model_request_id: attempt.request_id.as_str().to_owned(),
-                attempt_count: expected_count,
-                provider_kind: attempt.provider_kind.as_str().to_owned(),
-                provider_account_id: attempt
-                    .provider_account_id
-                    .as_ref()
-                    .map(|id| id.as_str().to_owned()),
-                provider_account_ref: attempt
-                    .provider_account_ref
-                    .as_ref()
-                    .map(|id| id.as_str().to_owned()),
-                upstream_model_id: attempt.upstream_model_id.as_str().to_owned(),
-                upstream_transport: attempt.upstream_transport,
-                http_version: attempt.http_version,
-            })
+            .begin_model_request_attempt(attempt_start_row(attempt))
             .await
             .map_err(core_store_error)?;
         if persisted == expected_count {
@@ -688,6 +715,24 @@ impl ExecutionStore for PgExecutionStore {
         }
     }
 
+    async fn create_model_request_with_attempt(
+        &self,
+        request: CoreNewModelRequest,
+        attempt: CoreAttemptRecord,
+    ) -> Result<(), CoreStoreError> {
+        // 合并写只对首次 attempt 成立（插入即带 attempt 列）；其余走基础两步。
+        if attempt.attempt_count.get() != 1 {
+            self.create_model_request(request).await?;
+            return self.record_attempt(attempt).await;
+        }
+        self.insert_model_request_with_first_attempt(
+            new_model_request_row(request),
+            attempt_start_row(attempt),
+        )
+        .await
+        .map_err(core_store_error)
+    }
+
     async fn mark_send_state(
         &self,
         request_id: &ModelRequestId,
@@ -776,6 +821,56 @@ impl ExecutionStore for PgExecutionStore {
         .map_err(core_store_error)
     }
 
+    async fn record_probe_failure(&self, failure: ProbeFailure) -> Result<(), CoreStoreError> {
+        let error = failure.error;
+        let retry_after_ms = error
+            .retry_after()
+            .map(|duration| u64::try_from(duration.as_millis()))
+            .transpose()
+            .map_err(|_| CoreStoreError::new(CoreStoreErrorKind::InvalidData))?;
+        let latency_ms = u64::try_from(failure.latency.as_millis())
+            .map_err(|_| CoreStoreError::new(CoreStoreErrorKind::InvalidData))?;
+        let provider_error_code = error.upstream_code().map(|code| code.as_str().to_owned());
+        tracing::warn!(
+            target: "gateway_probe",
+            provider_kind = failure.provider_kind.as_str(),
+            account_id = failure.account_id.as_str(),
+            upstream_model = failure.upstream_model_id.as_str(),
+            failure_kind = error.kind().as_str(),
+            send_state = ?error.send_state(),
+            upstream_status = ?error.upstream_status(),
+            provider_error_code = ?provider_error_code,
+            latency_ms,
+            "账号连接测试失败"
+        );
+        super::OpsEventRepository::append_ops_event(
+            &super::PgOpsEventRepository::new(self.pool.clone()),
+            super::OpsEvent {
+                id: Uuid::now_v7().to_string(),
+                model_request_id: None,
+                attempt_index: None,
+                level: super::OpsEventLevel::Warning,
+                component: "account_probe".to_owned(),
+                operation: "connection_test".to_owned(),
+                provider_kind: Some(failure.provider_kind.as_str().to_owned()),
+                provider_account_id: Some(failure.account_id.as_str().to_owned()),
+                provider_account_ref: Some(failure.account_id.as_str().to_owned()),
+                upstream_model_id: Some(failure.upstream_model_id.as_str().to_owned()),
+                failure_kind: error.kind().as_str().to_owned(),
+                status_code: error.upstream_status(),
+                provider_error_code,
+                retry_after_ms,
+                upstream_request_id: error.upstream_request_id().map(|id| id.as_str().to_owned()),
+                latency_ms: Some(latency_ms),
+                message: "account connection test failed".to_owned(),
+                occurrence_count: 1,
+                created_at: Utc::now(),
+            },
+        )
+        .await
+        .map_err(core_store_error)
+    }
+
     async fn finalize_model_request(
         &self,
         finalization: CoreModelRequestFinalization,
@@ -937,6 +1032,52 @@ fn core_store_error(error: StoreError) -> CoreStoreError {
     CoreStoreError::new(kind)
 }
 
+fn new_model_request_row(request: CoreNewModelRequest) -> NewModelRequest {
+    NewModelRequest {
+        id: request.id.as_str().to_owned(),
+        client_api_key_id: request
+            .client_api_key_id
+            .as_ref()
+            .map(|id| id.as_str().to_owned()),
+        client_api_key_ref: request.client_api_key_ref.as_str().to_owned(),
+        config_revision: request.config_revision.get(),
+        protocol: request.protocol,
+        operation: request.operation.as_str().to_owned(),
+        endpoint: request.endpoint,
+        client_transport: request.client_transport,
+        requested_model_id: request.requested_model.as_str().to_owned(),
+        client_ip: request.client_ip.map(|address| address.to_string()),
+        user_agent: request.user_agent,
+        reasoning_effort: request.reasoning_effort,
+        reasoning_preset: request.reasoning_preset,
+        request_kind: request.request_kind,
+        subagent_kind: request.subagent_kind,
+        compact: request.compact,
+        image_generation_requested: request.image_generation_requested,
+        started_at: DateTime::<Utc>::from(request.started_at),
+        deadline_at: DateTime::<Utc>::from(request.deadline_at),
+    }
+}
+
+fn attempt_start_row(attempt: CoreAttemptRecord) -> ModelRequestAttemptStart {
+    ModelRequestAttemptStart {
+        model_request_id: attempt.request_id.as_str().to_owned(),
+        attempt_count: attempt.attempt_count.get(),
+        provider_kind: attempt.provider_kind.as_str().to_owned(),
+        provider_account_id: attempt
+            .provider_account_id
+            .as_ref()
+            .map(|id| id.as_str().to_owned()),
+        provider_account_ref: attempt
+            .provider_account_ref
+            .as_ref()
+            .map(|id| id.as_str().to_owned()),
+        upstream_model_id: attempt.upstream_model_id.as_str().to_owned(),
+        upstream_transport: attempt.upstream_transport,
+        http_version: attempt.http_version,
+    }
+}
+
 fn optional_i64(value: Option<u64>, field: &'static str) -> StoreResult<Option<i64>> {
     value.map(|value| to_i64(value, field)).transpose()
 }

backend/crates/gateway-store/src/postgres/history.rs

@@ -1,111 +0,0 @@
-//! 不含正文或会话状态的模型请求历史查询。
-
-use async_trait::async_trait;
-use chrono::{DateTime, Utc};
-use sqlx::PgPool;
-
-use crate::{StoreResult, postgres_unavailable, require_nonempty};
-
-#[derive(Debug, Clone, PartialEq, Eq)]
-pub struct ModelRequestHistoryRecord {
-    pub id: String,
-    pub client_api_key_ref: String,
-    pub requested_model_id: String,
-    pub provider_kind: Option<String>,
-    pub provider_account_ref: Option<String>,
-    pub upstream_model_id: Option<String>,
-    pub outcome: String,
-    pub client_response_id: Option<String>,
-    pub upstream_response_id: Option<String>,
-    pub started_at: DateTime<Utc>,
-    pub completed_at: Option<DateTime<Utc>>,
-}
-
-#[async_trait]
-pub trait ModelRequestHistoryRepository: Send + Sync {
-    async fn find_model_request_history(
-        &self,
-        model_request_id: &str,
-    ) -> StoreResult<Option<ModelRequestHistoryRecord>>;
-}
-
-#[derive(Clone)]
-pub struct PgHistoryRepository {
-    pool: PgPool,
-}
-
-impl PgHistoryRepository {
-    #[must_use]
-    pub const fn new(pool: PgPool) -> Self {
-        Self { pool }
-    }
-}
-
-#[async_trait]
-impl ModelRequestHistoryRepository for PgHistoryRepository {
-    async fn find_model_request_history(
-        &self,
-        model_request_id: &str,
-    ) -> StoreResult<Option<ModelRequestHistoryRecord>> {
-        require_nonempty(
-            "model request history",
-            "model_request_id",
-            model_request_id,
-        )?;
-        fetch_history(&self.pool, "id", model_request_id).await
-    }
-}
-
-async fn fetch_history(
-    pool: &PgPool,
-    lookup: &'static str,
-    value: &str,
-) -> StoreResult<Option<ModelRequestHistoryRecord>> {
-    let row = match lookup {
-        "id" => {
-            sqlx::query_as::<_, HistoryRow>(
-                "select id, client_api_key_ref, requested_model_id, provider_kind,
-                    provider_account_ref, upstream_model_id, outcome,
-                    client_response_id, upstream_response_id,
-                    started_at, completed_at
-             from model_requests where id = $1",
-            )
-            .bind(value)
-            .fetch_optional(pool)
-            .await
-        }
-        _ => return Ok(None),
-    }
-    .map_err(|_| postgres_unavailable("read model request history"))?;
-    Ok(row.map(history_from_row))
-}
-
-type HistoryRow = (
-    String,
-    String,
-    String,
-    Option<String>,
-    Option<String>,
-    Option<String>,
-    String,
-    Option<String>,
-    Option<String>,
-    DateTime<Utc>,
-    Option<DateTime<Utc>>,
-);
-
-fn history_from_row(row: HistoryRow) -> ModelRequestHistoryRecord {
-    ModelRequestHistoryRecord {
-        id: row.0,
-        client_api_key_ref: row.1,
-        requested_model_id: row.2,
-        provider_kind: row.3,
-        provider_account_ref: row.4,
-        upstream_model_id: row.5,
-        outcome: row.6,
-        client_response_id: row.7,
-        upstream_response_id: row.8,
-        started_at: row.9,
-        completed_at: row.10,
-    }
-}

backend/crates/gateway-store/src/postgres/mod.rs

@@ -3,13 +3,14 @@
 use async_trait::async_trait;
 use sqlx::{PgPool, postgres::PgPoolOptions};
 
-use crate::{Revision, StoreBackend, StoreError, StoreResult, postgres_unavailable};
+use crate::{
+    Revision, StoreBackend, StoreError, StorePoolConfig, StoreResult, postgres_unavailable,
+};
 
 mod admin_security_audit;
 mod admission_recovery;
 mod client_keys;
 mod execution;
-mod history;
 mod observability;
 mod ops_events;
 mod provider_accounts;
@@ -21,7 +22,6 @@ pub use admin_security_audit::*;
 pub use admission_recovery::*;
 pub use client_keys::*;
 pub use execution::*;
-pub use history::*;
 pub use observability::*;
 pub use ops_events::*;
 pub use provider_accounts::*;
@@ -31,12 +31,19 @@ pub use snapshot::*;
 
 static MIGRATOR: sqlx::migrate::Migrator = sqlx::migrate!("../../migrations");
 
-/// 建立 PostgreSQL pool 并只执行冻结的 `0001_initial.sql` migration 集。
-pub async fn connect_and_migrate(database_url: &str) -> StoreResult<PgPool> {
+/// 建立 PostgreSQL pool 并只执行冻结的 migration 集。
+pub async fn connect_and_migrate(
+    database_url: &str,
+    pool_config: StorePoolConfig,
+) -> StoreResult<PgPool> {
     if database_url.trim().is_empty() {
         return Err(postgres_unavailable("connect PostgreSQL"));
     }
     let pool = PgPoolOptions::new()
+        .max_connections(pool_config.max_connections)
+        .acquire_timeout(std::time::Duration::from_secs(
+            pool_config.acquire_timeout_seconds,
+        ))
         .connect(database_url)
         .await
         .map_err(|_| postgres_unavailable("connect PostgreSQL"))?;

backend/crates/gateway-store/src/postgres/observability.rs

@@ -759,35 +759,35 @@ impl ObservabilityRepository for PgObservabilityRepository {
         range: ObservabilityRange,
     ) -> StoreResult<DashboardObservation> {
         let filter = UsageRecordFilter::default();
-        let requests = request_metrics(&self.pool, range, &filter).await?;
-        let attempts = attempt_metrics(&self.pool, range, &filter).await?;
-        let provider_accounts = provider_account_metrics(&self.pool, range.end).await?;
-        let trend = request_metric_series(&self.pool, range, &filter).await?;
         let account_usage_range = ObservabilityRange::new(
             range.end - TimeDelta::hours(ACCOUNT_USAGE_TIMELINE_HOURS),
             range.end,
         )?;
-        let account_usage = provider_account_usage(
-            &self.pool,
+        let account_usage_query =
             ProviderAccountUsageQuery::recent(account_usage_range, DASHBOARD_ACCOUNT_LIMIT)?
-                .with_hourly_request_buckets()?,
-        )
-        .await?;
-        let recent_requests = list_usage_records(
-            &self.pool,
-            UsageRecordQuery {
-                range,
-                filter: UsageRecordFilter {
-                    outcome: Some("succeeded".to_owned()),
-                    ..UsageRecordFilter::default()
-                },
-                cursor: None,
-                page: ObservabilityPageNumber::new(1)?,
-                page_size: ObservabilityPageSize::new(10)?,
+                .with_hourly_request_buckets()?;
+        let recent_query = UsageRecordQuery {
+            range,
+            filter: UsageRecordFilter {
+                outcome: Some("succeeded".to_owned()),
+                ..UsageRecordFilter::default()
             },
-        )
-        .await?
-        .items;
+            cursor: None,
+            page: ObservabilityPageNumber::new(1)?,
+            page_size: ObservabilityPageSize::new(10)?,
+        };
+        // 分两批并发：页面延迟从六路之和降为两批各自最慢者之和，同时单次
+        // 渲染最多占用 3 个池连接，不与数据面写路径争抢整个连接池。
+        let (requests, attempts, provider_accounts) = futures::try_join!(
+            request_metrics(&self.pool, range, &filter),
+            attempt_metrics(&self.pool, range, &filter),
+            provider_account_metrics(&self.pool, range.end),
+        )?;
+        let (trend, account_usage, recent_requests) = futures::try_join!(
+            request_metric_series(&self.pool, range, &filter),
+            provider_account_usage(&self.pool, account_usage_query),
+            async { Ok(list_usage_records(&self.pool, recent_query).await?.items) },
+        )?;
         Ok(DashboardObservation {
             range,
             requests,
@@ -1651,7 +1651,10 @@ async fn request_metrics(
     filter: &UsageRecordFilter,
 ) -> StoreResult<RequestMetrics> {
     filter.validate()?;
-    let mut query = QueryBuilder::<Postgres>::new(
+    // 结果计数覆盖范围内全部请求（成功率/失败率的分母分子）；用量、缓存、
+    // 延迟与成本聚合仅统计用量事实（已完整交付客户端的成功响应）。
+    let fact = completed_usage_fact_predicate("mr");
+    let mut query = QueryBuilder::<Postgres>::new(format!(
         "select count(*)::bigint as request_count,
                 count(*) filter (where outcome = 'succeeded')::bigint as success_count,
                 count(*) filter (where outcome = 'failed')::bigint as failure_count,
@@ -1659,40 +1662,43 @@ async fn request_metrics(
                 count(*) filter (where outcome = 'incomplete')::bigint as incomplete_count,
                 count(*) filter (where client_status_code between 400 and 499)::bigint
                   as caller_error_count,
-                coalesce(sum(input_tokens), 0)::bigint as input_tokens,
-                coalesce(sum(output_tokens), 0)::bigint as output_tokens,
-                coalesce(sum(cached_tokens), 0)::bigint as cached_tokens,
-                coalesce(sum(cache_write_tokens), 0)::bigint as cache_write_tokens,
-                coalesce(sum(reasoning_tokens), 0)::bigint as reasoning_tokens,
-                coalesce(sum(total_tokens), 0)::bigint as total_tokens,
-                coalesce(sum(first_token_ms), 0)::bigint as first_token_latency_sum,
-                count(first_token_ms)::bigint as first_token_latency_count,
-                coalesce(sum(latency_ms), 0)::bigint as latency_sum,
-                count(latency_ms)::bigint as latency_count,
-                max(latency_ms)::bigint as max_latency_ms,
-                min(latency_ms)::bigint as min_latency_ms,
-                count(*) filter (where input_tokens is not null)::bigint
+                coalesce(sum(input_tokens) filter (where {fact}), 0)::bigint as input_tokens,
+                coalesce(sum(output_tokens) filter (where {fact}), 0)::bigint as output_tokens,
+                coalesce(sum(cached_tokens) filter (where {fact}), 0)::bigint as cached_tokens,
+                coalesce(sum(cache_write_tokens) filter (where {fact}), 0)::bigint
+                  as cache_write_tokens,
+                coalesce(sum(reasoning_tokens) filter (where {fact}), 0)::bigint
+                  as reasoning_tokens,
+                coalesce(sum(total_tokens) filter (where {fact}), 0)::bigint as total_tokens,
+                coalesce(sum(first_token_ms) filter (where {fact}), 0)::bigint
+                  as first_token_latency_sum,
+                count(first_token_ms) filter (where {fact})::bigint as first_token_latency_count,
+                coalesce(sum(latency_ms) filter (where {fact}), 0)::bigint as latency_sum,
+                count(latency_ms) filter (where {fact})::bigint as latency_count,
+                max(latency_ms) filter (where {fact})::bigint as max_latency_ms,
+                min(latency_ms) filter (where {fact})::bigint as min_latency_ms,
+                count(*) filter (where {fact} and input_tokens is not null)::bigint
                   as cache_eligible_request_count,
-                count(*) filter (where input_tokens is not null and cached_tokens > 0)::bigint
+                count(*) filter (where {fact} and input_tokens is not null
+                                   and cached_tokens > 0)::bigint
                   as cache_hit_request_count,
                 percentile_cont(0.50) within group (order by latency_ms)
-                  as latency_p50_ms,
+                  filter (where {fact}) as latency_p50_ms,
                 percentile_cont(0.95) within group (order by latency_ms)
-                  as latency_p95_ms,
+                  filter (where {fact}) as latency_p95_ms,
                 percentile_cont(0.99) within group (order by latency_ms)
-                  as latency_p99_ms,
+                  filter (where {fact}) as latency_p99_ms,
                 percentile_cont(0.50) within group (order by first_token_ms)
-                  as first_token_p50_ms,
+                  filter (where {fact}) as first_token_p50_ms,
                 percentile_cont(0.95) within group (order by first_token_ms)
-                  as first_token_p95_ms,
+                  filter (where {fact}) as first_token_p95_ms,
                 percentile_cont(0.99) within group (order by first_token_ms)
-                  as first_token_p99_ms
-         from model_requests mr where mr.started_at >= ",
-    );
+                  filter (where {fact}) as first_token_p99_ms
+         from model_requests mr where mr.started_at >= "
+    ));
     query.push_bind(range.start);
     query.push(" and mr.started_at < ");
     query.push_bind(range.end);
-    push_completed_usage_fact_filter(&mut query, "mr");
     push_usage_filter(&mut query, filter, "mr");
     let row = query
         .build()
@@ -1709,9 +1715,12 @@ async fn request_metric_series(
 ) -> StoreResult<Vec<RequestMetricPoint>> {
     filter.validate()?;
     let granularity = granularity_for(range);
+    // 与 request_metrics 同一契约：结果计数覆盖全部请求，用量/延迟/成本
+    // 聚合仅统计用量事实。
+    let fact = completed_usage_fact_predicate("mr");
     let mut query = QueryBuilder::<Postgres>::new("select date_bin(");
     query.push_bind(granularity.sql_interval());
-    query.push(
+    query.push(format!(
         "::interval, mr.started_at, timestamptz '1970-01-01 00:00:00+00') as bucket_start,
                 count(*)::bigint as request_count,
                 count(*) filter (where outcome = 'succeeded')::bigint as success_count,
@@ -1720,46 +1729,49 @@ async fn request_metric_series(
                 count(*) filter (where outcome = 'incomplete')::bigint as incomplete_count,
                 count(*) filter (where client_status_code between 400 and 499)::bigint
                   as caller_error_count,
-                coalesce(sum(input_tokens), 0)::bigint as input_tokens,
-                coalesce(sum(output_tokens), 0)::bigint as output_tokens,
-                coalesce(sum(cached_tokens), 0)::bigint as cached_tokens,
-                coalesce(sum(cache_write_tokens), 0)::bigint as cache_write_tokens,
-                coalesce(sum(reasoning_tokens), 0)::bigint as reasoning_tokens,
-                coalesce(sum(total_tokens), 0)::bigint as total_tokens,
-                coalesce(sum(first_token_ms), 0)::bigint as first_token_latency_sum,
-                count(first_token_ms)::bigint as first_token_latency_count,
-                coalesce(sum(latency_ms), 0)::bigint as latency_sum,
-                count(latency_ms)::bigint as latency_count,
-                max(latency_ms)::bigint as max_latency_ms,
-                min(latency_ms)::bigint as min_latency_ms,
-                count(*) filter (where input_tokens is not null)::bigint
+                coalesce(sum(input_tokens) filter (where {fact}), 0)::bigint as input_tokens,
+                coalesce(sum(output_tokens) filter (where {fact}), 0)::bigint as output_tokens,
+                coalesce(sum(cached_tokens) filter (where {fact}), 0)::bigint as cached_tokens,
+                coalesce(sum(cache_write_tokens) filter (where {fact}), 0)::bigint
+                  as cache_write_tokens,
+                coalesce(sum(reasoning_tokens) filter (where {fact}), 0)::bigint
+                  as reasoning_tokens,
+                coalesce(sum(total_tokens) filter (where {fact}), 0)::bigint as total_tokens,
+                coalesce(sum(first_token_ms) filter (where {fact}), 0)::bigint
+                  as first_token_latency_sum,
+                count(first_token_ms) filter (where {fact})::bigint as first_token_latency_count,
+                coalesce(sum(latency_ms) filter (where {fact}), 0)::bigint as latency_sum,
+                count(latency_ms) filter (where {fact})::bigint as latency_count,
+                max(latency_ms) filter (where {fact})::bigint as max_latency_ms,
+                min(latency_ms) filter (where {fact})::bigint as min_latency_ms,
+                count(*) filter (where {fact} and input_tokens is not null)::bigint
                   as cache_eligible_request_count,
-                count(*) filter (where input_tokens is not null and cached_tokens > 0)::bigint
+                count(*) filter (where {fact} and input_tokens is not null
+                                   and cached_tokens > 0)::bigint
                   as cache_hit_request_count,
                 percentile_cont(0.50) within group (order by latency_ms)
-                  as latency_p50_ms,
+                  filter (where {fact}) as latency_p50_ms,
                 percentile_cont(0.95) within group (order by latency_ms)
-                  as latency_p95_ms,
+                  filter (where {fact}) as latency_p95_ms,
                 percentile_cont(0.99) within group (order by latency_ms)
-                  as latency_p99_ms,
+                  filter (where {fact}) as latency_p99_ms,
                 percentile_cont(0.50) within group (order by first_token_ms)
-                  as first_token_p50_ms,
+                  filter (where {fact}) as first_token_p50_ms,
                 percentile_cont(0.95) within group (order by first_token_ms)
-                  as first_token_p95_ms,
+                  filter (where {fact}) as first_token_p95_ms,
                 percentile_cont(0.99) within group (order by first_token_ms)
-                  as first_token_p99_ms,
-                count(*) filter (where cost_source = 'provider_reported')::bigint
+                  filter (where {fact}) as first_token_p99_ms,
+                count(*) filter (where {fact} and cost_source = 'provider_reported')::bigint
                   as provider_reported_count,
-                count(*) filter (where cost_source = 'calculated')::bigint
+                count(*) filter (where {fact} and cost_source = 'calculated')::bigint
                   as calculated_count,
-                count(*) filter (where cost_source = 'unavailable')::bigint
+                count(*) filter (where {fact} and cost_source = 'unavailable')::bigint
                   as unavailable_count
-         from model_requests mr where mr.started_at >= ",
-    );
+         from model_requests mr where mr.started_at >= "
+    ));
     query.push_bind(range.start);
     query.push(" and mr.started_at < ");
     query.push_bind(range.end);
-    push_completed_usage_fact_filter(&mut query, "mr");
     push_usage_filter(&mut query, filter, "mr");
     query.push(" group by bucket_start order by bucket_start");
     let rows = query
@@ -2249,8 +2261,14 @@ fn push_usage_filter(query: &mut QueryBuilder<Postgres>, filter: &UsageRecordFil
 /// `model_requests` 同时承担执行审计：包括上游已发送、但尚未收到首个事件就断开的
 /// WebSocket 请求。这些失败仍必须留给 Ops Errors 和调度健康度分析，不能混入用量、
 /// 成本、账号使用次数或请求明细。
+fn completed_usage_fact_predicate(alias: &str) -> String {
+    format!(
+        "{alias}.outcome = 'succeeded' and {alias}.downstream_committed_at is not null and {alias}.client_status_code between 200 and 399"
+    )
+}
+
 fn push_completed_usage_fact_filter(query: &mut QueryBuilder<Postgres>, alias: &str) {
-    query.push(format!(" and {alias}.outcome = 'succeeded' and {alias}.downstream_committed_at is not null and {alias}.client_status_code between 200 and 399"));
+    query.push(format!(" and {}", completed_usage_fact_predicate(alias)));
 }
 
 async fn provider_account_usage(

backend/crates/gateway-store/src/postgres/provider_accounts.rs

@@ -460,10 +460,6 @@ pub trait ProviderAccountRepository: Send + Sync {
         &self,
         update: ProviderCredentialUpdate,
     ) -> StoreResult<Revision>;
-    async fn update_provider_account_observation(
-        &self,
-        observation: ProviderAccountObservation,
-    ) -> StoreResult<bool>;
     async fn apply_provider_account_state(
         &self,
         update: ProviderAccountStateUpdate,
@@ -654,36 +650,6 @@ impl ProviderAccountRepository for PgProviderAccountRepository {
         Revision::new(to_u64(next)?)
     }
 
-    async fn update_provider_account_observation(
-        &self,
-        observation: ProviderAccountObservation,
-    ) -> StoreResult<bool> {
-        observation.validate()?;
-        let result = sqlx::query(
-            "update provider_accounts
-             set availability = $2, availability_reason = $3, cooldown_until = $4,
-                 provider_quota_json = $5, availability_observed_at = $6,
-                 quota_observed_at = $7,
-                 updated_at = greatest(now(), $6, coalesce($7, $6))
-             where id = $1",
-        )
-        .bind(observation.account_id)
-        .bind(observation.availability.as_str())
-        .bind(observation.availability_reason)
-        .bind(observation.cooldown_until)
-        .bind(
-            observation
-                .provider_quota_json
-                .map(|quota| quota.as_value()),
-        )
-        .bind(observation.availability_observed_at)
-        .bind(observation.quota_observed_at)
-        .execute(&self.pool)
-        .await
-        .map_err(|_| postgres_unavailable("update provider account observation"))?;
-        Ok(result.rows_affected() == 1)
-    }
-
     async fn apply_provider_account_state(
         &self,
         update: ProviderAccountStateUpdate,

backend/crates/gateway-store/src/postgres/retention.rs

@@ -78,47 +78,49 @@ impl RetentionRepository for PgRetentionRepository {
             });
         }
 
-        let mut transaction = self
-            .pool
-            .begin()
-            .await
-            .map_err(|_| postgres_unavailable("begin retention transaction"))?;
-        let model_requests = sqlx::query(
+        // 分批删除，每批独立提交：清理是幂等的，无需跨表原子性；
+        // 有界批量避免百万行单事务的长锁与 WAL 峰值。
+        let model_requests = purge_in_batches(
+            &self.pool,
             "delete from model_requests
-             where outcome <> 'running'
-               and completed_at < $1 - ($2 * interval '1 day')",
+             where ctid in (
+               select ctid from model_requests
+                where outcome <> 'running'
+                  and completed_at < $1 - ($2 * interval '1 day')
+                limit $3
+             )",
+            now,
+            settings.usage_retention_days,
+            "delete expired model requests",
         )
-        .bind(now)
-        .bind(i64::from(settings.usage_retention_days))
-        .execute(&mut *transaction)
-        .await
-        .map_err(|_| postgres_unavailable("delete expired model requests"))?
-        .rows_affected();
-        let ops_events = sqlx::query(
+        .await?;
+        let ops_events = purge_in_batches(
+            &self.pool,
             "delete from ops_events
-             where model_request_id is null
-               and created_at < $1 - ($2 * interval '1 day')",
+             where ctid in (
+               select ctid from ops_events
+                where model_request_id is null
+                  and created_at < $1 - ($2 * interval '1 day')
+                limit $3
+             )",
+            now,
+            settings.ops_event_retention_days,
+            "delete expired ops events",
         )
-        .bind(now)
-        .bind(i64::from(settings.ops_event_retention_days))
-        .execute(&mut *transaction)
-        .await
-        .map_err(|_| postgres_unavailable("delete expired ops events"))?
-        .rows_affected();
-        let admin_audit_events = sqlx::query(
+        .await?;
+        let admin_audit_events = purge_in_batches(
+            &self.pool,
             "delete from admin_audit_events
-             where created_at < $1 - ($2 * interval '1 day')",
+             where ctid in (
+               select ctid from admin_audit_events
+                where created_at < $1 - ($2 * interval '1 day')
+                limit $3
+             )",
+            now,
+            settings.audit_retention_days,
+            "delete expired admin audit events",
         )
-        .bind(now)
-        .bind(i64::from(settings.audit_retention_days))
-        .execute(&mut *transaction)
-        .await
-        .map_err(|_| postgres_unavailable("delete expired admin audit events"))?
-        .rows_affected();
-        transaction
-            .commit()
-            .await
-            .map_err(|_| postgres_unavailable("commit retention transaction"))?;
+        .await?;
         Ok(RetentionReport {
             model_requests,
             ops_events,
@@ -127,6 +129,32 @@ impl RetentionRepository for PgRetentionRepository {
     }
 }
 
+const RETENTION_DELETE_BATCH_ROWS: i64 = 10_000;
+
+async fn purge_in_batches(
+    pool: &PgPool,
+    delete_sql: &'static str,
+    now: DateTime<Utc>,
+    retention_days: u32,
+    label: &'static str,
+) -> StoreResult<u64> {
+    let mut total = 0u64;
+    loop {
+        let deleted = sqlx::query(delete_sql)
+            .bind(now)
+            .bind(i64::from(retention_days))
+            .bind(RETENTION_DELETE_BATCH_ROWS)
+            .execute(pool)
+            .await
+            .map_err(|_| postgres_unavailable(label))?
+            .rows_affected();
+        total += deleted;
+        if deleted < RETENTION_DELETE_BATCH_ROWS.unsigned_abs() {
+            return Ok(total);
+        }
+    }
+}
+
 fn to_u32(value: i64) -> StoreResult<u32> {
     u32::try_from(value).map_err(|_| StoreError::InvalidData {
         entity: "retention settings",

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

@@ -1,5 +1,7 @@
 //! Provider account availability 的可重建 Redis cache fence。
 
+use std::time::Duration;
+
 use async_trait::async_trait;
 use chrono::{DateTime, Utc};
 use redis::{Script, aio::ConnectionManager};
@@ -28,6 +30,12 @@ redis.call('PEXPIRE', KEYS[1], ARGV[5])
 return 1
 "#;
 
+const RECORD_REFRESH_BACKOFF_SCRIPT: &str = r#"
+local n = redis.call('INCR', KEYS[1])
+redis.call('PEXPIRE', KEYS[1], ARGV[1])
+return n
+"#;
+
 const MAX_PROVIDER_CATALOG_BYTES: usize = 1024 * 1024;
 const CREDENTIAL_STATE_TTL_SECONDS: u64 = 24 * 60 * 60;
 
@@ -132,6 +140,14 @@ impl RedisCredentialStateRepository {
         Ok(format!("{}:account:{fingerprint}:state", self.namespace))
     }
 
+    fn backoff_key(&self, provider_account_id: &str) -> StoreResult<String> {
+        let fingerprint = resource_fingerprint("credential refresh backoff", provider_account_id)?;
+        Ok(format!(
+            "{}:account:{fingerprint}:refresh-backoff",
+            self.namespace
+        ))
+    }
+
     fn catalog_key(&self, key: &RedisProviderCatalogCacheKey) -> StoreResult<String> {
         key.validate()?;
         let fingerprint = resource_fingerprint("provider catalog cache", &key.catalog_scope)?;
@@ -340,6 +356,48 @@ impl ProviderCredentialStatePort for RedisCredentialStateRepository {
                 .map_err(|_| provider_unavailable("clear credential state"))
         })
     }
+
+    fn record_refresh_backoff<'a>(
+        &'a self,
+        account_id: &'a ProviderAccountId,
+        window: Duration,
+    ) -> futures::future::BoxFuture<'a, Result<u32, ProviderStoreError>> {
+        Box::pin(async move {
+            let window_millis = u64::try_from(window.as_millis())
+                .ok()
+                .filter(|millis| *millis > 0)
+                .ok_or_else(|| provider_invalid("record refresh backoff"))?;
+            let key = self
+                .backoff_key(account_id.as_str())
+                .map_err(|_| provider_unavailable("record refresh backoff"))?;
+            let mut connection = self.connection.clone();
+            let count = Script::new(RECORD_REFRESH_BACKOFF_SCRIPT)
+                .key(key)
+                .arg(window_millis)
+                .invoke_async::<i64>(&mut connection)
+                .await
+                .map_err(|_| provider_unavailable("record refresh backoff"))?;
+            Ok(u32::try_from(count).unwrap_or(u32::MAX))
+        })
+    }
+
+    fn clear_refresh_backoff<'a>(
+        &'a self,
+        account_id: &'a ProviderAccountId,
+    ) -> futures::future::BoxFuture<'a, Result<(), ProviderStoreError>> {
+        Box::pin(async move {
+            let key = self
+                .backoff_key(account_id.as_str())
+                .map_err(|_| provider_unavailable("clear refresh backoff"))?;
+            let mut connection = self.connection.clone();
+            redis::cmd("DEL")
+                .arg(key)
+                .query_async::<i64>(&mut connection)
+                .await
+                .map_err(|_| provider_unavailable("clear refresh backoff"))?;
+            Ok(())
+        })
+    }
 }
 
 impl ProviderCatalogCachePort for RedisCredentialStateRepository {

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

@@ -1,7 +1,7 @@
 //! 可丢失、可从 PostgreSQL 或 Provider 重建的 Redis 协调状态。
 
 use chrono::{DateTime, SecondsFormat, Utc};
-use redis::{Script, aio::ConnectionManager};
+use redis::aio::ConnectionManager;
 use serde::{Deserialize, Serialize};
 use sha2::{Digest, Sha256};
 
@@ -33,17 +33,6 @@ pub use runtime_change::*;
 use crate::{StoreError, StoreResult, redis_unavailable, require_nonempty};
 
 pub(crate) const MAX_REDIS_EXACT_INTEGER: u64 = (1_u64 << 53) - 1;
-const RECORD_ADMIN_LOGIN_FAILURE_SCRIPT: &str = r#"
-local count = redis.call('INCR', KEYS[1])
-local ttl = redis.call('PTTL', KEYS[1])
-if count == 1 or ttl < 0 then
-  redis.call('PEXPIRE', KEYS[1], ARGV[2])
-end
-if count >= tonumber(ARGV[1]) then
-  return 1
-end
-return 0
-"#;
 
 /// Redis 中可丢失的管理员会话事实；认证秘密不属于该结构。
 #[derive(Debug, Clone, PartialEq, Eq)]
@@ -71,7 +60,7 @@ impl AdminSessionRecord {
     }
 }
 
-/// 管理员会话与登录失败窗口的 Redis 基础设施端口。
+/// 管理员会话的 Redis 基础设施端口。
 #[async_trait]
 pub trait AdminAuthStateRepository: Send + Sync {
     async fn load_admin_session(&self, session_id: &str)
@@ -85,22 +74,9 @@ pub trait AdminAuthStateRepository: Send + Sync {
         &self,
         session_id: &str,
     ) -> StoreResult<Option<AdminSessionRecord>>;
-    async fn login_source_is_throttled(
-        &self,
-        source: &str,
-        failure_limit: u32,
-        window_seconds: u64,
-    ) -> StoreResult<bool>;
-    async fn record_login_failure(
-        &self,
-        source: &str,
-        failure_limit: u32,
-        window_seconds: u64,
-    ) -> StoreResult<bool>;
-    async fn clear_login_failures(&self, source: &str) -> StoreResult<()>;
 }
 
-/// Redis 管理员会话与登录失败窗口 adapter。
+/// Redis 管理员会话 adapter。
 #[derive(Clone)]
 pub struct RedisAdminAuthStateRepository {
     connection: ConnectionManager,
@@ -119,11 +95,6 @@ impl RedisAdminAuthStateRepository {
         let fingerprint = resource_fingerprint("admin session", session_id)?;
         Ok(format!("{}:session:{{{fingerprint}}}", self.namespace))
     }
-
-    fn failure_key(&self, source: &str) -> StoreResult<String> {
-        let fingerprint = resource_fingerprint("admin login failure", source)?;
-        Ok(format!("{}:failure:{{{fingerprint}}}", self.namespace))
-    }
 }
 
 #[async_trait]
@@ -179,54 +150,6 @@ impl AdminAuthStateRepository for RedisAdminAuthStateRepository {
             .map(|value| decode_admin_session(&value))
             .transpose()
     }
-
-    async fn login_source_is_throttled(
-        &self,
-        source: &str,
-        failure_limit: u32,
-        window_seconds: u64,
-    ) -> StoreResult<bool> {
-        validate_failure_policy(failure_limit, window_seconds)?;
-        let key = self.failure_key(source)?;
-        let mut connection = self.connection.clone();
-        let failures = redis::cmd("GET")
-            .arg(key)
-            .query_async::<Option<u64>>(&mut connection)
-            .await
-            .map_err(|_| redis_unavailable("load admin login failure window"))?
-            .unwrap_or_default();
-        Ok(failures >= u64::from(failure_limit))
-    }
-
-    async fn record_login_failure(
-        &self,
-        source: &str,
-        failure_limit: u32,
-        window_seconds: u64,
-    ) -> StoreResult<bool> {
-        let window_millis = validate_failure_policy(failure_limit, window_seconds)?;
-        let key = self.failure_key(source)?;
-        let mut connection = self.connection.clone();
-        let throttled = Script::new(RECORD_ADMIN_LOGIN_FAILURE_SCRIPT)
-            .key(key)
-            .arg(failure_limit)
-            .arg(window_millis)
-            .invoke_async::<i64>(&mut connection)
-            .await
-            .map_err(|_| redis_unavailable("record admin login failure"))?;
-        Ok(throttled == 1)
-    }
-
-    async fn clear_login_failures(&self, source: &str) -> StoreResult<()> {
-        let key = self.failure_key(source)?;
-        let mut connection = self.connection.clone();
-        redis::cmd("DEL")
-            .arg(key)
-            .query_async::<i64>(&mut connection)
-            .await
-            .map_err(|_| redis_unavailable("clear admin login failures"))?;
-        Ok(())
-    }
 }
 
 #[derive(Debug, Serialize, Deserialize)]
@@ -259,17 +182,6 @@ fn decode_admin_session(value: &str) -> StoreResult<AdminSessionRecord> {
     })
 }
 
-fn validate_failure_policy(failure_limit: u32, window_seconds: u64) -> StoreResult<u64> {
-    if failure_limit == 0 {
-        return Err(admin_auth_invalid("login failure limit must be positive"));
-    }
-    let window_millis = window_seconds
-        .checked_mul(1_000)
-        .filter(|value| *value > 0 && *value <= MAX_REDIS_EXACT_INTEGER)
-        .ok_or_else(|| admin_auth_invalid("login failure window is outside the supported range"))?;
-    Ok(window_millis)
-}
-
 fn admin_auth_invalid(message: &str) -> StoreError {
     StoreError::InvalidData {
         entity: "admin authentication state",

backend/crates/gateway-store/tests/main.rs

@@ -1,2 +1,3 @@
 mod postgres;
 mod redis;
+mod support;

backend/crates/gateway-store/tests/postgres/execution.rs

@@ -4,7 +4,9 @@ use gateway_core::engine::{
     ExecutionOutcome, ExecutionStore, ModelRequestFinalization as CoreModelRequestFinalization,
     ModelRequestId, ModelRequestTimings as CoreModelRequestTimings, UpstreamSendState,
 };
-use gateway_store::postgres::{ModelRequestRepository, NewModelRequest, PgExecutionStore};
+use gateway_store::postgres::{
+    ModelRequestAttemptStart, ModelRequestRepository, NewModelRequest, PgExecutionStore,
+};
 
 use super::TestDatabase;
 
@@ -41,6 +43,116 @@ fn model_request_rejects_mismatched_client_key_live_id() {
     assert!(request.validate().is_err());
 }
 
+#[tokio::test]
+async fn merged_first_attempt_insert_should_match_sequential_semantics() {
+    let Some(database) = TestDatabase::create("execution_merged_insert").await else {
+        return;
+    };
+    let repository = PgExecutionStore::new(database.pool.clone());
+    let started_at = Utc::now();
+    sqlx::query(
+        "insert into provider_accounts (
+           id, provider_kind, name, email, upstream_user_id,
+           upstream_account_id, plan_type, authentication_kind,
+           provider_credentials_json, credential_revision,
+           has_refresh_token, access_token_expires_at, enabled, availability,
+           availability_observed_at, created_at, updated_at
+         ) values (
+           'acct_merged', 'openai', 'merged', null, 'user-merged', null, null, 'oauth',
+           '{}'::jsonb, 1, false, $1 + interval '1 day', true, 'ready', $1, $1, $1
+         )",
+    )
+    .bind(started_at)
+    .execute(&database.pool)
+    .await
+    .expect("seed provider account");
+    let request = NewModelRequest {
+        id: "req_merged".to_owned(),
+        client_api_key_id: None,
+        client_api_key_ref: "key_merged".to_owned(),
+        config_revision: 1,
+        protocol: "openai".to_owned(),
+        operation: "responses".to_owned(),
+        endpoint: "/v1/responses".to_owned(),
+        client_transport: "http_sse".to_owned(),
+        requested_model_id: "coding".to_owned(),
+        client_ip: None,
+        user_agent: None,
+        reasoning_effort: None,
+        reasoning_preset: None,
+        request_kind: None,
+        subagent_kind: None,
+        compact: false,
+        image_generation_requested: false,
+        started_at,
+        deadline_at: started_at + Duration::seconds(30),
+    };
+    let attempt = ModelRequestAttemptStart {
+        model_request_id: "req_merged".to_owned(),
+        attempt_count: 1,
+        provider_kind: "openai".to_owned(),
+        provider_account_id: Some("acct_merged".to_owned()),
+        provider_account_ref: Some("acct_merged".to_owned()),
+        upstream_model_id: "gpt-test".to_owned(),
+        upstream_transport: "http_sse".to_owned(),
+        http_version: None,
+    };
+
+    repository
+        .insert_model_request_with_first_attempt(request, attempt)
+        .await
+        .expect("merged insert");
+
+    let (attempt_count, send_state, provider_kind, outcome): (i32, String, String, String) =
+        sqlx::query_as(
+            "select attempt_count, upstream_send_state, provider_kind, outcome
+             from model_requests where id = 'req_merged'",
+        )
+        .fetch_one(&database.pool)
+        .await
+        .expect("load merged request");
+    assert_eq!(
+        (
+            attempt_count,
+            send_state.as_str(),
+            provider_kind.as_str(),
+            outcome.as_str()
+        ),
+        (1, "not_sent", "openai", "running")
+    );
+
+    // 后续 attempt 沿用常规 CAS 递增路径；已持久化的 sent 水位不被重试重置。
+    repository
+        .mark_upstream_send_state(
+            "req_merged",
+            gateway_store::postgres::UpstreamSendState::Sent,
+        )
+        .await
+        .expect("mark sent before retry");
+    let second = repository
+        .begin_model_request_attempt(ModelRequestAttemptStart {
+            model_request_id: "req_merged".to_owned(),
+            attempt_count: 2,
+            provider_kind: "openai".to_owned(),
+            provider_account_id: Some("acct_merged".to_owned()),
+            provider_account_ref: Some("acct_merged".to_owned()),
+            upstream_model_id: "gpt-test".to_owned(),
+            upstream_transport: "http_sse".to_owned(),
+            http_version: None,
+        })
+        .await
+        .expect("second attempt");
+    assert_eq!(second, 2);
+    let send_state: String = sqlx::query_scalar(
+        "select upstream_send_state from model_requests where id = 'req_merged'",
+    )
+    .fetch_one(&database.pool)
+    .await
+    .expect("load send state after retry");
+    assert_eq!(send_state, "sent");
+    database.close().await;
+}
+
 #[tokio::test]
 async fn downstream_commit_should_atomically_record_http_status_once() {
     let Some(database) = TestDatabase::create("execution_commit_status").await else {

backend/crates/gateway-store/tests/postgres/history.rs

@@ -1,21 +0,0 @@
-use chrono::Utc;
-use gateway_store::postgres::ModelRequestHistoryRecord;
-
-#[test]
-fn history_contains_no_payload_field() {
-    let history = ModelRequestHistoryRecord {
-        id: "request-1".to_owned(),
-        client_api_key_ref: "key-1".to_owned(),
-        requested_model_id: "coding".to_owned(),
-        provider_kind: None,
-        provider_account_ref: None,
-        upstream_model_id: None,
-        outcome: "failed".to_owned(),
-        client_response_id: None,
-        upstream_response_id: None,
-        started_at: Utc::now(),
-        completed_at: None,
-    };
-
-    assert_eq!(history.id, "request-1");
-}

backend/crates/gateway-store/tests/postgres/mod.rs

@@ -1,4 +1,4 @@
-use std::{env, str::FromStr};
+use std::str::FromStr;
 
 use gateway_store::postgres::connect_and_migrate;
 use sqlx::{
@@ -11,7 +11,6 @@ mod admin_security_audit;
 mod admission_recovery;
 mod client_keys;
 mod execution;
-mod history;
 mod observability;
 mod ops_events;
 mod provider_accounts;
@@ -27,7 +26,7 @@ pub(super) struct TestDatabase {
 
 impl TestDatabase {
     pub(super) async fn create(label: &str) -> Option<Self> {
-        let database_url = env::var("CPR_TEST_DATABASE_URL").ok()?;
+        let database_url = crate::support::test_env("CPR_TEST_DATABASE_URL")?;
         let schema = format!("cpr_store_{label}_{}", Uuid::new_v4().simple());
         let admin = PgPoolOptions::new()
             .max_connections(1)
@@ -59,6 +58,12 @@ impl TestDatabase {
             .execute(&mut *migration)
             .await
             .expect("apply terminal migration");
+        sqlx::raw_sql(include_str!(
+            "../../../../migrations/0002_retention_delete_indexes.sql"
+        ))
+        .execute(&mut *migration)
+        .await
+        .expect("apply retention index migration");
         migration.commit().await.expect("commit terminal migration");
         Some(Self {
             admin,
@@ -82,7 +87,7 @@ impl TestDatabase {
 
 #[tokio::test]
 async fn connect_and_migrate_should_apply_0001_once_and_reopen_cleanly() {
-    let Ok(database_url) = env::var("CPR_TEST_DATABASE_URL") else {
+    let Some(database_url) = crate::support::test_env("CPR_TEST_DATABASE_URL") else {
         return;
     };
     let database = format!("cpr_store_migrator_{}", Uuid::new_v4().simple());
@@ -103,7 +108,7 @@ async fn connect_and_migrate_should_apply_0001_once_and_reopen_cleanly() {
         .database(&database)
         .to_url_lossy()
         .to_string();
-    let first = connect_and_migrate(&isolated_url)
+    let first = connect_and_migrate(&isolated_url, gateway_store::StorePoolConfig::default())
         .await
         .expect("apply 0001 through production migrator");
     let first_table_count = sqlx::query_scalar::<_, i64>(
@@ -114,7 +119,7 @@ async fn connect_and_migrate_should_apply_0001_once_and_reopen_cleanly() {
     .expect("count migrated tables");
     first.close().await;
 
-    let second = connect_and_migrate(&isolated_url)
+    let second = connect_and_migrate(&isolated_url, gateway_store::StorePoolConfig::default())
         .await
         .expect("reopen database through production migrator");
     let migration_count =
@@ -132,7 +137,7 @@ async fn connect_and_migrate_should_apply_0001_once_and_reopen_cleanly() {
     .expect("drop migration test database");
     admin.close().await;
 
-    assert_eq!((first_table_count, migration_count), (8, 1));
+    assert_eq!((first_table_count, migration_count), (8, 2));
 }
 
 #[test]

backend/crates/gateway-store/tests/postgres/observability.rs

@@ -174,7 +174,9 @@ async fn admin_observability_adapter_preserves_utc_queries_metrics_costs_and_det
         .await
         .expect("admin dashboard summary");
     assert_eq!(dashboard.range, range);
-    assert_eq!(dashboard.requests.request_count, 1);
+    assert_eq!(dashboard.requests.request_count, 3);
+    assert_eq!(dashboard.requests.success_count, 2);
+    assert_eq!(dashboard.requests.failure_count, 1);
     assert_eq!(dashboard.requests.first_token_latency_sum_ms, 120);
     assert_eq!(dashboard.requests.latency_sum_ms, 900);
     assert_eq!(dashboard.requests.min_latency_ms, Some(900));
@@ -217,6 +219,13 @@ async fn admin_observability_adapter_preserves_utc_queries_metrics_costs_and_det
             .iter()
             .map(|point| point.metrics.request_count)
             .sum::<u64>(),
+        3,
+    );
+    assert_eq!(
+        dashboard_trend
+            .iter()
+            .map(|point| point.metrics.failure_count)
+            .sum::<u64>(),
         1,
     );
 
@@ -235,7 +244,7 @@ async fn admin_observability_adapter_preserves_utc_queries_metrics_costs_and_det
             .iter()
             .map(|point| point.metrics.request_count)
             .sum::<u64>(),
-        1,
+        2,
     );
     assert_eq!(
         trend
@@ -442,7 +451,10 @@ async fn observability_queries_preserve_request_account_cost_and_diagnostic_fact
         .dashboard_summary(range)
         .await
         .expect("dashboard summary");
-    assert_eq!(dashboard.requests.request_count, 1);
+    assert_eq!(dashboard.requests.request_count, 3);
+    assert_eq!(dashboard.requests.success_count, 2);
+    assert_eq!(dashboard.requests.failure_count, 1);
+    assert_eq!(dashboard.requests.caller_error_count, 0);
     assert_eq!(dashboard.requests.cache_eligible_request_count, 1);
     assert_eq!(dashboard.requests.cache_hit_request_count, 1);
     assert_eq!(dashboard.requests.cache_hit_request_rate(), Some(1.0));
@@ -501,6 +513,14 @@ async fn observability_queries_preserve_request_account_cost_and_diagnostic_fact
             .iter()
             .map(|point| point.metrics.request_count)
             .sum::<u64>(),
+        3
+    );
+    assert_eq!(
+        dashboard
+            .trend
+            .iter()
+            .map(|point| point.metrics.failure_count)
+            .sum::<u64>(),
         1
     );
 
@@ -592,7 +612,8 @@ async fn observability_queries_preserve_request_account_cost_and_diagnostic_fact
         .expect("usage summary");
     assert_eq!(overview.attempts.attempt_count, 4);
     assert_eq!(overview.attempts.failure_count, 2);
-    assert_eq!(overview.requests.request_count, 1);
+    assert_eq!(overview.requests.request_count, 3);
+    assert_eq!(overview.requests.failure_count, 1);
 
     let succeeded = repository
         .usage_summary(

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

@@ -283,7 +283,7 @@ fn running_request(
 }
 
 async fn repository() -> Option<(RedisClientAdmissionRepository, ConnectionManager, String)> {
-    let redis_url = std::env::var("CPR_TEST_REDIS_URL").ok()?;
+    let redis_url = crate::support::test_env("CPR_TEST_REDIS_URL")?;
     let client = redis::Client::open(redis_url).expect("valid CPR_TEST_REDIS_URL");
     let connection = client
         .get_connection_manager()

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

@@ -536,7 +536,7 @@ fn millisecond_precision(value: DateTime<Utc>) -> DateTime<Utc> {
 }
 
 async fn repository() -> Option<(RedisCredentialCooldownRepository, ConnectionManager, String)> {
-    let redis_url = std::env::var("CPR_TEST_REDIS_URL").ok()?;
+    let redis_url = crate::support::test_env("CPR_TEST_REDIS_URL")?;
     let client = redis::Client::open(redis_url).expect("valid CPR_TEST_REDIS_URL");
     let connection = client
         .get_connection_manager()

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

@@ -404,7 +404,7 @@ fn acquired(
 }
 
 async fn repository() -> Option<(RedisCredentialLeaseRepository, ConnectionManager, String)> {
-    let redis_url = std::env::var("CPR_TEST_REDIS_URL").ok()?;
+    let redis_url = crate::support::test_env("CPR_TEST_REDIS_URL")?;
     let client = redis::Client::open(redis_url).expect("valid CPR_TEST_REDIS_URL");
     let connection = client
         .get_connection_manager()

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

@@ -32,25 +32,14 @@ fn admin_auth_state_rejects_invalid_ttl_boundaries() {
             .block_on(repository.store_admin_session("expired-session", &session))
             .is_err()
     );
-    assert!(
-        runtime
-            .block_on(repository.record_login_failure("source", 0, 60))
-            .is_err()
-    );
-    assert!(
-        runtime
-            .block_on(repository.record_login_failure("source", 5, 0))
-            .is_err()
-    );
 }
 
 #[tokio::test]
-async fn admin_auth_state_keeps_fixed_ttl_and_atomic_failure_window() {
+async fn admin_auth_state_keeps_fixed_ttl_and_opaque_keys() {
     let Some((repository, mut connection, namespace)) = admin_auth_repository().await else {
         return;
     };
     let session_id = "session-real-secret-id";
-    let source = "203.0.113.7";
     let password = "must-never-enter-redis";
     let admin_api_key = "admin-must-never-enter-redis";
     let session = AdminSessionRecord {
@@ -70,29 +59,14 @@ async fn admin_auth_state_keeps_fixed_ttl_and_atomic_failure_window() {
         Some(session.clone())
     );
 
-    for attempt in 1..=5 {
-        let throttled = repository
-            .record_login_failure(source, 5, 60)
-            .await
-            .expect("record failure");
-        assert_eq!(throttled, attempt == 5);
-    }
-    assert!(
-        repository
-            .login_source_is_throttled(source, 5, 60)
-            .await
-            .expect("read failure window")
-    );
-
     let keys = redis::cmd("KEYS")
         .arg(format!("{namespace}:*"))
         .query_async::<Vec<String>>(&mut connection)
         .await
         .expect("list isolated test keys");
-    assert_eq!(keys.len(), 2);
+    assert_eq!(keys.len(), 1);
     for key in &keys {
         assert!(!key.contains(session_id));
-        assert!(!key.contains(source));
         let ttl = redis::cmd("PTTL")
             .arg(key)
             .query_async::<i64>(&mut connection)
@@ -105,21 +79,10 @@ async fn admin_auth_state_keeps_fixed_ttl_and_atomic_failure_window() {
             .await
             .expect("read isolated test value");
         assert!(!value.contains(session_id));
-        assert!(!value.contains(source));
         assert!(!value.contains(password));
         assert!(!value.contains(admin_api_key));
     }
 
-    repository
-        .clear_login_failures(source)
-        .await
-        .expect("clear failure window");
-    assert!(
-        !repository
-            .login_source_is_throttled(source, 5, 60)
-            .await
-            .expect("read cleared failure window")
-    );
     assert_eq!(
         repository
             .delete_admin_session(session_id)
@@ -138,7 +101,7 @@ async fn admin_auth_state_keeps_fixed_ttl_and_atomic_failure_window() {
 
 async fn admin_auth_repository()
 -> Option<(RedisAdminAuthStateRepository, ConnectionManager, String)> {
-    let redis_url = std::env::var("CPR_TEST_REDIS_URL").ok()?;
+    let redis_url = crate::support::test_env("CPR_TEST_REDIS_URL")?;
     let client = redis::Client::open(redis_url).expect("valid CPR_TEST_REDIS_URL");
     let connection = client
         .get_connection_manager()

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

@@ -91,7 +91,7 @@ async fn native_continuation_rejects_provider_state_for_a_different_provider() {
 }
 
 async fn repository() -> Option<(RedisNativeContinuationRepository, ConnectionManager, String)> {
-    let redis_url = std::env::var("CPR_TEST_REDIS_URL").ok()?;
+    let redis_url = crate::support::test_env("CPR_TEST_REDIS_URL")?;
     let client = redis::Client::open(redis_url).expect("valid CPR_TEST_REDIS_URL");
     let connection = client
         .get_connection_manager()

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

@@ -132,7 +132,7 @@ async fn repository() -> Option<(
     ConnectionManager,
     String,
 )> {
-    let redis_url = std::env::var("CPR_TEST_REDIS_URL").ok()?;
+    let redis_url = crate::support::test_env("CPR_TEST_REDIS_URL")?;
     let client = redis::Client::open(redis_url).ok()?;
     let connection: ConnectionManager = client.get_connection_manager().await.ok()?;
     let namespace = format!("gateway-store-test-{}", Uuid::new_v4());

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

@@ -157,7 +157,7 @@ async fn repository_with_policy(
     threshold: u32,
     open_duration: Duration,
 ) -> Option<(RedisProviderCircuitRepository, ConnectionManager, String)> {
-    let redis_url = std::env::var("CPR_TEST_REDIS_URL").ok()?;
+    let redis_url = crate::support::test_env("CPR_TEST_REDIS_URL")?;
     let client = redis::Client::open(redis_url).expect("valid CPR_TEST_REDIS_URL");
     let connection = client
         .get_connection_manager()

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

@@ -113,7 +113,7 @@ async fn affinity_repository() -> Option<(
     ConnectionManager,
     String,
 )> {
-    let redis_url = std::env::var("CPR_TEST_REDIS_URL").ok()?;
+    let redis_url = crate::support::test_env("CPR_TEST_REDIS_URL")?;
     let client = redis::Client::open(redis_url).expect("valid CPR_TEST_REDIS_URL");
     let connection = client
         .get_connection_manager()

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

@@ -92,7 +92,7 @@ async fn repository() -> Option<(
     ConnectionManager,
     String,
 )> {
-    let redis_url = std::env::var("CPR_TEST_REDIS_URL").ok()?;
+    let redis_url = crate::support::test_env("CPR_TEST_REDIS_URL")?;
     let client = redis::Client::open(redis_url).expect("valid CPR_TEST_REDIS_URL");
     let connection = client
         .get_connection_manager()

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

@@ -15,8 +15,8 @@ use uuid::Uuid;
 #[tokio::test]
 async fn store_bundle_worker_plan_and_leader_lease_are_single_use_and_fenced() {
     let (Some(database_url), Some(redis_url)) = (
-        std::env::var("CPR_TEST_DATABASE_URL").ok(),
-        std::env::var("CPR_TEST_REDIS_URL").ok(),
+        crate::support::test_env("CPR_TEST_DATABASE_URL"),
+        crate::support::test_env("CPR_TEST_REDIS_URL"),
     ) else {
         return;
     };

backend/crates/gateway-store/tests/support.rs

@@ -0,0 +1,20 @@
+/// 读取集成测试基础设施地址（PostgreSQL / Redis）。
+///
+/// 本地未配置时返回 `None`，测试跳过；CI 中缺失则直接失败——
+/// 静默跳过产生的绿灯与真实回归不可区分。
+pub(crate) fn test_env(name: &str) -> Option<String> {
+    let value = std::env::var(name).ok();
+    assert!(
+        value.is_some() || !running_in_ci(),
+        "{name} must be set in CI: store integration tests are not allowed to skip silently"
+    );
+    value
+}
+
+/// 按 is-ci 约定判定 CI：`CI` 存在且非 falsey（`0`/`false`/空）。
+/// 开发者用 `CI=false` 显式声明「非 CI」时不应把跳过变成 panic。
+fn running_in_ci() -> bool {
+    std::env::var("CI")
+        .ok()
+        .is_some_and(|value| !matches!(value.as_str(), "" | "0" | "false"))
+}

backend/crates/providers/openai/Cargo.toml

@@ -43,7 +43,6 @@ tracing = "0.1.44"
 tungstenite = { git = "https://github.com/openai-oss-forks/tungstenite-rs", rev = "4fffad30fe373adbdcffab9545e9e9bf4f2fc19f", features = ["deflate", "proxy"] }
 url = "2.5.8"
 uuid = { version = "1.23.3", features = ["v4", "v7"] }
-zeroize = { version = "1.8.2", features = ["derive"] }
 
 [dev-dependencies]
 tempfile = "3.27.0"

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

@@ -658,6 +658,8 @@ impl CodexCredentialQuotaService {
         if outcome == QuotaWriteOutcome::Conflict {
             return Ok(false);
         }
+        // 可用性写回必须基于最新账号状态：revision fence 只覆盖凭据轮换，
+        // 不覆盖其他路径的仅状态变化（Cooldown/QuotaExhausted 转换）。
         let Some(current) = self.store.get_account(account.id()).await? else {
             return Ok(false);
         };

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

@@ -8,9 +8,9 @@ use chrono::{DateTime, Utc};
 use futures::{StreamExt as _, stream};
 use gateway_core::engine::credential::{AccountAvailability, ProviderAccount, ProviderAccountId};
 use gateway_core::provider_ports::{
-    ProviderLeaseAcquisition, ProviderLeasePort, ProviderLeaseRequest,
+    ProviderCredentialStatePort, ProviderLeaseAcquisition, ProviderLeasePort, ProviderLeaseRequest,
     ProviderRefreshCapacityRequest, ProviderRefreshLeaseRequest, ProviderRefreshPolicy,
-    ProviderRuntimePolicyPort, ProviderStoreError, provider_refresh_retry_at,
+    ProviderRuntimePolicyPort, ProviderStoreError, provider_refresh_backoff_at,
 };
 use secrecy::{ExposeSecret, SecretString};
 use thiserror::Error;
@@ -25,7 +25,8 @@ use super::types::{CodexCredentialPrincipal, CodexOAuthSecret, RotateCodexCreden
 
 const PROVIDER_NAME: &str = "openai";
 const MAX_REFRESH_BATCH: u32 = 1_000;
-const REFRESH_RETRY_DELAY: Duration = Duration::from_secs(10 * 60);
+/// 连续失败计数窗口；每次瞬态失败刷新该 TTL，静默满窗后计数过期归零。
+const REFRESH_BACKOFF_WINDOW: Duration = Duration::from_secs(30 * 60);
 
 pub struct DueCodexCredential {
     pub account: ProviderAccount,
@@ -59,9 +60,6 @@ pub enum CodexCredentialRefreshOutcome {
     Transient {
         account_id: String,
     },
-    Ambiguous {
-        account_id: String,
-    },
     LeaseUnavailable {
         account_id: String,
     },
@@ -88,6 +86,7 @@ pub struct CodexCredentialRefreshService {
     refresher: Arc<dyn TokenRefresher>,
     identity: Arc<dyn CodexAccountIdentityVerifier>,
     leases: Arc<dyn ProviderLeasePort>,
+    credential_state: Arc<dyn ProviderCredentialStatePort>,
     runtime_policy: Arc<dyn ProviderRuntimePolicyPort>,
 }
 
@@ -97,6 +96,7 @@ impl CodexCredentialRefreshService {
         refresher: Arc<dyn TokenRefresher>,
         identity: Arc<dyn CodexAccountIdentityVerifier>,
         leases: Arc<dyn ProviderLeasePort>,
+        credential_state: Arc<dyn ProviderCredentialStatePort>,
         runtime_policy: Arc<dyn ProviderRuntimePolicyPort>,
     ) -> Self {
         Self {
@@ -104,6 +104,7 @@ impl CodexCredentialRefreshService {
             refresher,
             identity,
             leases,
+            credential_state,
             runtime_policy,
         }
     }
@@ -206,13 +207,16 @@ impl CodexCredentialRefreshService {
                 }
             }
             Err(RefreshFailure::Transport) => {
-                self.persist_terminal(
-                    &due.account,
-                    AccountAvailability::Invalid,
-                    "refresh_ambiguous",
-                    CodexCredentialRefreshOutcome::Ambiguous { account_id },
-                )
-                .await
+                // 上游瞬态（429/5xx/超时/畸形响应等，发送状态未知）保留现有凭据、
+                // 推进退避重试；仅 invalid_grant/banned 才终态失效账号。
+                if self
+                    .defer_refresh(&due.account, "transport-ambiguous")
+                    .await?
+                {
+                    Ok(CodexCredentialRefreshOutcome::Transient { account_id })
+                } else {
+                    Ok(CodexCredentialRefreshOutcome::Stale { account_id })
+                }
             }
         }
     }
@@ -326,25 +330,34 @@ impl CodexCredentialRefreshService {
                     .await;
             }
             Err(CodexIdentityVerificationError::Unavailable) => {
-                return self
-                    .persist_terminal(
-                        &due.account,
-                        AccountAvailability::Invalid,
-                        "refreshed_identity_unavailable",
-                        CodexCredentialRefreshOutcome::Ambiguous {
-                            account_id: due.account.id().to_string(),
-                        },
-                    )
-                    .await;
+                // JWKS/签名边界短暂不可用属瞬态：保留现有凭据、推进退避重试，
+                // 不终态失效账号。
+                return if self
+                    .defer_refresh(&due.account, "identity-unavailable")
+                    .await?
+                {
+                    Ok(CodexCredentialRefreshOutcome::Transient {
+                        account_id: due.account.id().to_string(),
+                    })
+                } else {
+                    Ok(CodexCredentialRefreshOutcome::Stale {
+                        account_id: due.account.id().to_string(),
+                    })
+                };
             }
         };
         let profile = match verification {
             CodexIdentityVerification::Complete(profile) => profile,
             CodexIdentityVerification::SignedOnly(signed) => {
-                let retry_at = provider_refresh_retry_at(
+                let attempt = self
+                    .credential_state
+                    .record_refresh_backoff(due.account.id(), REFRESH_BACKOFF_WINDOW)
+                    .await
+                    .unwrap_or(1);
+                let retry_at = provider_refresh_backoff_at(
                     due.account.id(),
                     SystemTime::now(),
-                    REFRESH_RETRY_DELAY,
+                    attempt,
                     "identity-completion",
                 )?;
                 match self
@@ -388,10 +401,17 @@ impl CodexCredentialRefreshService {
             })
             .await;
         match result {
-            Ok(revision) => Ok(CodexCredentialRefreshOutcome::Refreshed {
-                account_id: due.account.id().to_string(),
-                credential_revision: revision.get(),
-            }),
+            Ok(revision) => {
+                // 凭据完整成功轮换后清零连续失败计数，退避窗口重新从 base 起步。
+                let _ = self
+                    .credential_state
+                    .clear_refresh_backoff(due.account.id())
+                    .await;
+                Ok(CodexCredentialRefreshOutcome::Refreshed {
+                    account_id: due.account.id().to_string(),
+                    credential_revision: revision.get(),
+                })
+            }
             Err(CredentialRepositoryError::RevisionConflict) => {
                 Ok(CodexCredentialRefreshOutcome::Stale {
                     account_id: due.account.id().to_string(),
@@ -443,12 +463,13 @@ impl CodexCredentialRefreshService {
         account: &ProviderAccount,
         reason: &'static str,
     ) -> Result<bool, CodexCredentialRefreshError> {
-        let retry_at = provider_refresh_retry_at(
-            account.id(),
-            SystemTime::now(),
-            REFRESH_RETRY_DELAY,
-            reason,
-        )?;
+        let attempt = self
+            .credential_state
+            .record_refresh_backoff(account.id(), REFRESH_BACKOFF_WINDOW)
+            .await
+            .unwrap_or(1);
+        let retry_at =
+            provider_refresh_backoff_at(account.id(), SystemTime::now(), attempt, reason)?;
         match self.repository.defer_refresh(account, retry_at).await {
             Ok(_) => Ok(true),
             Err(CredentialRepositoryError::RevisionConflict) => Ok(false),

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

@@ -307,7 +307,13 @@ fn parse_token_pair(body: &[u8]) -> Result<TokenPair, ()> {
     })
 }
 
-fn classify_refresh_failure(_status: StatusCode, body: &[u8]) -> RefreshFailure {
+fn classify_refresh_failure(status: StatusCode, body: &[u8]) -> RefreshFailure {
+    // 5xx 与 429 按状态码判为瞬态：其正文（CDN/网关页等）不是权威 OAuth 错误响应，
+    // 不据此终态失效账号。
+    if status.is_server_error() || status == StatusCode::TOO_MANY_REQUESTS {
+        return RefreshFailure::Transport;
+    }
+    // 4xx 正文才是权威 OAuth 错误响应，据其区分真过期/封禁与其余瞬态。
     let lower = String::from_utf8_lossy(body).to_ascii_lowercase();
     if lower.contains("account has been deactivated") {
         return RefreshFailure::Banned;

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

@@ -64,6 +64,7 @@ pub async fn initialize(
     let session_exclusions = ports.session_exclusions();
     let account_feedback = ports.account_feedback();
     let runtime_policy = ports.runtime_policy();
+    let credential_state = ports.credential_state();
     let profile = config.wire_profile_state();
     let session_identity = config
         .session_identity()
@@ -170,6 +171,7 @@ pub async fn initialize(
         refresher,
         Arc::clone(&identity),
         Arc::clone(&leases),
+        credential_state,
         Arc::clone(&runtime_policy),
     ));
     let pending = Arc::new(OpenAiOAuthPendingStore::new(

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

@@ -818,10 +818,12 @@ fn cold_response_stream(response: ColdResponse) -> EventStream {
         let failure_rate_limit_headers = response.rate_limit_headers.clone();
         let rate_limit_updates = response.rate_limit_header_updates;
         let turn_state_updates = response.turn_state_update;
-        let mut decoder = CodexCanonicalDecoder::new(upstream_model.as_str());
-        if response.transport == CodexBackendTransport::HttpSse {
-            decoder = decoder.with_raw_sse_passthrough();
-        }
+        // OpenAI 线路为透明代理：HTTP SSE 与 WebSocket 两条上游均启用 raw 透传，
+        // 下游按字节转发上游原文，避免 serde 往返改写数值/精度（大整数→f64、logprobs 等）。
+        // WS 帧由 reducer 以 encode_sse_event(&event, raw) 逐字节内嵌上游原始 JSON
+        // （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();
         loop {
             let Some(stream_deadline) = remaining(context.deadline()) else {
                 Err(provider_error(ProviderErrorKind::Timeout, UpstreamSendState::Sent))?;
@@ -877,7 +879,8 @@ fn cold_response_stream(response: ColdResponse) -> EventStream {
             };
             if let (Some(capture), Some(updates)) =
                 (session_capture.as_mut(), turn_state_updates.as_ref())
-                && let Some(turn_state) = updates.lock().await.clone()
+                // take：仅在上游写入过新 turn_state 时更新 capture，避免逐 chunk 克隆。
+                && let Some(turn_state) = updates.lock().await.take()
             {
                 capture.turn_state = Some(turn_state);
             }
@@ -994,7 +997,7 @@ fn cold_response_stream(response: ColdResponse) -> EventStream {
         }
         if let (Some(capture), Some(updates)) =
             (session_capture.as_mut(), turn_state_updates.as_ref())
-            && let Some(turn_state) = updates.lock().await.clone()
+            && let Some(turn_state) = updates.lock().await.take()
         {
             capture.turn_state = Some(turn_state);
         }
@@ -2194,7 +2197,6 @@ impl ScheduledTask for OpenAiOAuthRefreshTask {
                     matches!(
                         outcome,
                         CodexCredentialRefreshOutcome::Transient { .. }
-                            | CodexCredentialRefreshOutcome::Ambiguous { .. }
                             | CodexCredentialRefreshOutcome::Failed { .. }
                     )
                 })

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

@@ -31,7 +31,7 @@ use crate::transport::{
         CodexWebSocketConnection, CodexWebSocketPool, CodexWebSocketPoolKey,
         DEFAULT_INITIAL_EVENT_TIMEOUT, WEBSOCKET_FAST_PATH_BUDGET, WebSocketOriginBreaker,
         WebSocketPoolDecision, execute_prepared_response_create_request_stream,
-        post_send_ambiguous, prepare_response_create_request_with_pool,
+        post_send_ambiguous, prepare_response_create_request_with_pool, websocket_audit_dir,
         write_websocket_audit_artifact_from_env,
     },
 };
@@ -176,13 +176,17 @@ impl CodexBackendClient {
             &websocket_request,
         )
         .map_err(CodexClientError::WebSocketEncode)?;
-        let artifact = websocket_audit_artifact_from_attempt(
-            &websocket_request,
-            websocket_create.connection().opening_audit_snapshot(),
-            websocket_payload_audit_snapshot(&websocket_request),
-        );
-        if let Err(error) = write_websocket_audit_artifact_from_env(&artifact).await {
-            tracing::warn!(error = %error, "Failed to write Codex WebSocket audit artifact");
+        // 审计未启用时跳过 artifact 构造：payload 快照会深拷贝整个请求 body，
+        // 且位于首字节前的关键路径上。
+        if websocket_audit_dir().is_some() {
+            let artifact = websocket_audit_artifact_from_attempt(
+                &websocket_request,
+                websocket_create.connection().opening_audit_snapshot(),
+                websocket_payload_audit_snapshot(&websocket_request),
+            );
+            if let Err(error) = write_websocket_audit_artifact_from_env(&artifact).await {
+                tracing::warn!(error = %error, "Failed to write Codex WebSocket audit artifact");
+            }
         }
         let connection_profile = websocket_connection_profile(&headers);
         let pool_key =

backend/crates/providers/openai/src/transport/protocol/responses.rs

@@ -366,21 +366,6 @@ fn is_tool_execution_event(event_type: &str) -> bool {
         )
 }
 
-/// 从 Codex SSE 收集出的非流式 Responses 结果。
-#[derive(Debug, Clone, PartialEq, Eq)]
-pub enum CollectedResponse {
-    /// 收集到 `response.completed`。
-    Completed(Value),
-    /// 收集到 `response.incomplete`。
-    Incomplete(Value),
-    /// 收集到 `response.failed` 或 `error`。
-    Failed(ResponsesSseFailure),
-    /// SSE 未包含成功终止响应。
-    MissingCompleted,
-    /// 完成响应为空。
-    Empty,
-}
-
 /// Codex Responses SSE 失败事件。
 #[derive(Clone, PartialEq, Eq)]
 pub struct ResponsesSseFailure {

backend/crates/providers/openai/src/transport/protocol/websocket.rs

@@ -50,43 +50,32 @@ pub struct WebSocketAuditArtifact {
 
 /// 将一条公开 WebSocket JSON 事件编码为 SSE 帧。
 pub fn websocket_event_to_sse_frame(raw: &str) -> Option<String> {
-    let event = websocket_event_type(raw)?;
-    if is_internal_websocket_event(&event) {
-        return None;
-    }
-    if websocket_event_should_skip(raw) {
-        return None;
-    }
-    Some(encode_sse_event(&event, raw))
+    let value = serde_json::from_str::<Value>(raw).ok()?;
+    websocket_event_frame(&value, raw)
 }
 
-pub(crate) fn websocket_event_type(raw: &str) -> Option<String> {
-    serde_json::from_str::<Value>(raw).ok().and_then(|value| {
-        value
-            .get("type")
-            .and_then(Value::as_str)
-            .map(ToString::to_string)
-    })
+/// 同上，复用已解析的事件 JSON；`data` 段逐字节内嵌 `raw` 原文。
+///
+/// 仅剥离传输层内部帧：`codex.rate_limits` 与承载 turn_state 的
+/// `response.metadata` 不下发客户端；业务事件一律原样转发，不做 schema 校验。
+pub(crate) fn websocket_event_frame(value: &Value, raw: &str) -> Option<String> {
+    let event = websocket_event_type(value)?;
+    if is_internal_websocket_event(event) || event == "response.metadata" {
+        return None;
+    }
+    Some(encode_sse_event(event, raw))
+}
+
+pub(crate) fn websocket_event_type(value: &Value) -> Option<&str> {
+    value.get("type").and_then(Value::as_str)
 }
 
 fn is_internal_websocket_event(event: &str) -> bool {
     event == "codex.rate_limits"
 }
 
-/// 判断 WebSocket 事件是否应在转发前剥离。
-///
-/// 仅剥离传输层内部帧：`response.metadata` 承载 turn_state，由上层提取转存到
-/// 会话状态，不下发客户端。业务事件一律原样转发，不做 schema 校验。
-/// （`codex.rate_limits` 内部事件由 `is_internal_websocket_event` 单独剥离。）
-fn websocket_event_should_skip(raw: &str) -> bool {
-    websocket_event_type(raw).as_deref() == Some("response.metadata")
-}
-
 /// 提取 `response.metadata` 帧中的字符串响应头。
-pub fn websocket_metadata_headers(raw: &str) -> Vec<(String, String)> {
-    let Ok(value) = serde_json::from_str::<Value>(raw) else {
-        return Vec::new();
-    };
+pub fn websocket_metadata_headers(value: &Value) -> Vec<(String, String)> {
     if value.get("type").and_then(Value::as_str) != Some("response.metadata") {
         return Vec::new();
     }
@@ -100,9 +89,8 @@ pub fn websocket_metadata_headers(raw: &str) -> Vec<(String, String)> {
 }
 
 /// 从 `response.metadata` 帧中提取 `x-codex-turn-state`。
-pub fn websocket_metadata_turn_state(raw: &str) -> Option<String> {
-    let value = serde_json::from_str::<Value>(raw).ok()?;
-    websocket_metadata_headers(raw)
+pub fn websocket_metadata_turn_state(value: &Value) -> Option<String> {
+    websocket_metadata_headers(value)
         .into_iter()
         .find_map(|(name, value)| {
             name.eq_ignore_ascii_case("x-codex-turn-state")
@@ -128,8 +116,7 @@ fn json_value_as_string(value: &Value) -> Option<String> {
 /// 从 `response.completed` 旁路提取 response ID，供连接池记录续接能力。
 ///
 /// 该值不参与客户端 wire 的可交付性判断；无法读取时只是不记录连接内续接状态。
-pub fn websocket_response_completed_id(raw: &str) -> Option<String> {
-    let value = serde_json::from_str::<Value>(raw).ok()?;
+pub fn websocket_response_completed_id(value: &Value) -> Option<String> {
     if value.get("type").and_then(Value::as_str) != Some("response.completed") {
         return None;
     }
@@ -179,11 +166,39 @@ pub fn websocket_response_create_payload(request: &CodexResponsesRequest) -> Val
     Value::Object(payload)
 }
 
+/// 借用原始 body 序列化 `response.create` 帧，不复制字段。
+///
+/// 输出与序列化 [`websocket_response_create_payload`] 的合并 Map 逐字节一致：
+/// body 自带 `type` 键时以 body 值置于首位，其余字段按插入顺序原样透传。
+struct ResponseCreateFrame<'a> {
+    body: &'a Map<String, Value>,
+}
+
+impl serde::Serialize for ResponseCreateFrame<'_> {
+    fn serialize<S: serde::Serializer>(&self, serializer: S) -> Result<S::Ok, S::Error> {
+        use serde::ser::SerializeMap as _;
+        let mut map = serializer.serialize_map(Some(self.body.len() + 1))?;
+        match self.body.get("type") {
+            Some(body_type) => map.serialize_entry("type", body_type)?,
+            None => map.serialize_entry("type", "response.create")?,
+        }
+        for (key, value) in self.body {
+            if key == "type" {
+                continue;
+            }
+            map.serialize_entry(key, value)?;
+        }
+        map.end()
+    }
+}
+
 /// 生成 Responses WebSocket `response.create` 文本帧内容。
 pub fn websocket_response_create_payload_text(
     request: &CodexResponsesRequest,
 ) -> Result<String, serde_json::Error> {
-    serde_json::to_string(&websocket_response_create_payload(request))
+    serde_json::to_string(&ResponseCreateFrame {
+        body: request.body(),
+    })
 }
 
 fn websocket_payload_keys(request: &CodexResponsesRequest) -> Vec<String> {

backend/crates/providers/openai/src/transport/websocket/audit.rs

@@ -3,6 +3,7 @@
 use std::{
     io,
     path::{Path, PathBuf},
+    sync::OnceLock,
 };
 
 use chrono::Utc;
@@ -19,6 +20,19 @@ const REDACTED_HEADER_VALUE: &str = "<redacted>";
 /// WebSocket audit artifact 输出目录环境变量。
 pub const WS_AUDIT_DIR_ENV: &str = "CODEX_PROXY_WS_AUDIT_DIR";
 
+static WS_AUDIT_DIR: OnceLock<Option<PathBuf>> = OnceLock::new();
+
+/// 返回进程级审计输出目录；未配置时调用方应整体跳过 artifact 构造。
+pub fn websocket_audit_dir() -> Option<&'static Path> {
+    WS_AUDIT_DIR
+        .get_or_init(|| {
+            std::env::var_os(WS_AUDIT_DIR_ENV)
+                .filter(|value| !value.is_empty())
+                .map(PathBuf::from)
+        })
+        .as_deref()
+}
+
 /// 显式写入 WebSocket audit artifact。
 pub async fn write_websocket_audit_artifact_for_dir(
     dir: Option<&Path>,
@@ -39,14 +53,7 @@ pub async fn write_websocket_audit_artifact_for_dir(
 pub async fn write_websocket_audit_artifact_from_env(
     artifact: &WebSocketAuditArtifact,
 ) -> io::Result<Option<PathBuf>> {
-    let Some(dir) = std::env::var_os(WS_AUDIT_DIR_ENV)
-        .filter(|value| !value.is_empty())
-        .map(PathBuf::from)
-    else {
-        return Ok(None);
-    };
-
-    write_websocket_audit_artifact_for_dir(Some(&dir), artifact).await
+    write_websocket_audit_artifact_for_dir(websocket_audit_dir(), artifact).await
 }
 
 impl CodexWebSocketConnection {

backend/crates/providers/openai/src/transport/websocket/exchange/reducer.rs

@@ -1,9 +1,10 @@
 //! WebSocket aggregate/stream 共用事件归约器。
 
 use gateway_protocol::openai::events;
+use serde_json::Value;
 
 use crate::transport::protocol::websocket::{
-    websocket_event_to_sse_frame, websocket_event_type, websocket_metadata_headers,
+    websocket_event_frame, websocket_event_type, websocket_metadata_headers,
     websocket_metadata_turn_state, websocket_response_completed_id,
 };
 use crate::transport::response_meta;
@@ -33,40 +34,41 @@ pub(super) fn reduce_websocket_event(
     metadata: &mut CodexWebSocketConnectionMetadata,
     continuation: &mut WebSocketContinuationState,
 ) -> Result<ExchangeAction, CodexWebSocketExchangeError> {
-    if let Some(headers) = websocket_rate_limit_event_headers(raw) {
+    // 每帧只解析一次 JSON，后续提取全部复用同一 Value；
+    // 不可解析的帧不承载可路由的事件类型，忽略。
+    let Ok(value) = serde_json::from_str::<Value>(raw) else {
+        return Ok(ExchangeAction::Ignore);
+    };
+    if let Some(parsed) = events::parse_rate_limits_event(&value) {
+        let headers = events::rate_limits_to_header_pairs(&parsed);
         metadata.rate_limit_headers.extend(headers.iter().cloned());
         return Ok(ExchangeAction::RateLimits(headers));
     }
 
     response_meta::merge_response_metadata(
         &mut metadata.response_metadata,
-        websocket_metadata_headers(raw),
+        websocket_metadata_headers(&value),
     );
-    if let Some(turn_state) = websocket_metadata_turn_state(raw) {
+    if let Some(turn_state) = websocket_metadata_turn_state(&value) {
         metadata.turn_state = Some(turn_state.clone());
         return Ok(ExchangeAction::TurnState(turn_state));
     }
 
-    let event = websocket_event_type(raw);
-    if event.as_deref() == Some("response.completed")
-        && let Some(response_id) = websocket_response_completed_id(raw)
+    let event = websocket_event_type(&value);
+    if event == Some("response.completed")
+        && let Some(response_id) = websocket_response_completed_id(&value)
     {
         continuation.record_completed(response_id);
     }
 
-    let terminal = match event.as_deref() {
+    let terminal = match event {
         Some("response.completed") => Some(WebSocketTerminalKind::Completed),
         Some("response.incomplete") => Some(WebSocketTerminalKind::Incomplete),
         Some("response.failed" | "error") => Some(WebSocketTerminalKind::Failed),
         _ => None,
     };
-    Ok(match websocket_event_to_sse_frame(raw) {
+    Ok(match websocket_event_frame(&value, raw) {
         Some(frame) => ExchangeAction::Forward { frame, terminal },
         None => ExchangeAction::Ignore,
     })
 }
-
-fn websocket_rate_limit_event_headers(raw: &str) -> Option<Vec<(String, String)>> {
-    events::parse_rate_limits_event_raw(raw)
-        .map(|parsed| events::rate_limits_to_header_pairs(&parsed))
-}

backend/crates/providers/openai/src/transport/websocket/mod.rs

@@ -12,7 +12,7 @@ mod pump;
 
 pub use self::{
     audit::{
-        WS_AUDIT_DIR_ENV, write_websocket_audit_artifact_for_dir,
+        WS_AUDIT_DIR_ENV, websocket_audit_dir, write_websocket_audit_artifact_for_dir,
         write_websocket_audit_artifact_from_env,
     },
     breaker::{

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

@@ -764,6 +764,21 @@ impl ProviderCredentialStatePort for TestCredentialState {
     ) -> BoxFuture<'a, Result<bool, ProviderStoreError>> {
         Box::pin(async { Ok(false) })
     }
+
+    fn record_refresh_backoff<'a>(
+        &'a self,
+        _account_id: &'a ProviderAccountId,
+        _window: Duration,
+    ) -> BoxFuture<'a, Result<u32, ProviderStoreError>> {
+        Box::pin(async { Ok(1) })
+    }
+
+    fn clear_refresh_backoff<'a>(
+        &'a self,
+        _account_id: &'a ProviderAccountId,
+    ) -> BoxFuture<'a, Result<(), ProviderStoreError>> {
+        Box::pin(async { Ok(()) })
+    }
 }
 
 struct TestCooldown;

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

@@ -1,13 +1,13 @@
-use std::collections::VecDeque;
+use std::collections::{HashMap, VecDeque};
 use std::sync::{Arc, Mutex};
 use std::time::{Duration, SystemTime};
 
 use async_trait::async_trait;
 use futures::future::BoxFuture;
-use gateway_core::engine::credential::AccountAvailability;
+use gateway_core::engine::credential::{AccountAvailability, ProviderAccountId};
 use gateway_core::provider_ports::{
-    ProviderLeaseAcquisition, ProviderLeasePort, ProviderLeaseRequest, ProviderRefreshLeaseRequest,
-    ProviderStoreError,
+    ProviderCredentialState, ProviderCredentialStatePort, ProviderLeaseAcquisition,
+    ProviderLeasePort, ProviderLeaseRequest, ProviderRefreshLeaseRequest, ProviderStoreError,
 };
 use provider_openai::credential::token_client::{RefreshFailure, TokenPair, TokenRefresher};
 use provider_openai::credential::{
@@ -214,6 +214,85 @@ impl ProviderLeasePort for RefreshLeases {
     }
 }
 
+/// 真实累加连续失败计数的测试 double，用于断言退避的指数增长与成功清零。
+#[derive(Default)]
+struct CountingCredentialState {
+    counts: Mutex<HashMap<String, u32>>,
+    cleared: Mutex<Vec<String>>,
+}
+
+impl CountingCredentialState {
+    fn count(&self, account_id: &str) -> u32 {
+        self.counts
+            .lock()
+            .expect("backoff counts lock")
+            .get(account_id)
+            .copied()
+            .unwrap_or(0)
+    }
+
+    fn was_cleared(&self, account_id: &str) -> bool {
+        self.cleared
+            .lock()
+            .expect("cleared accounts lock")
+            .iter()
+            .any(|id| id == account_id)
+    }
+}
+
+impl ProviderCredentialStatePort for CountingCredentialState {
+    fn replace(
+        &self,
+        _state: ProviderCredentialState,
+    ) -> BoxFuture<'_, Result<(), ProviderStoreError>> {
+        Box::pin(async { Ok(()) })
+    }
+
+    fn read<'a>(
+        &'a self,
+        _account_id: &'a ProviderAccountId,
+    ) -> BoxFuture<'a, Result<Option<ProviderCredentialState>, ProviderStoreError>> {
+        Box::pin(async { Ok(None) })
+    }
+
+    fn clear<'a>(
+        &'a self,
+        _account_id: &'a ProviderAccountId,
+    ) -> BoxFuture<'a, Result<bool, ProviderStoreError>> {
+        Box::pin(async { Ok(false) })
+    }
+
+    fn record_refresh_backoff<'a>(
+        &'a self,
+        account_id: &'a ProviderAccountId,
+        _window: Duration,
+    ) -> BoxFuture<'a, Result<u32, ProviderStoreError>> {
+        Box::pin(async move {
+            let mut counts = self.counts.lock().expect("backoff counts lock");
+            let entry = counts.entry(account_id.as_str().to_owned()).or_insert(0);
+            *entry += 1;
+            Ok(*entry)
+        })
+    }
+
+    fn clear_refresh_backoff<'a>(
+        &'a self,
+        account_id: &'a ProviderAccountId,
+    ) -> BoxFuture<'a, Result<(), ProviderStoreError>> {
+        Box::pin(async move {
+            self.counts
+                .lock()
+                .expect("backoff counts lock")
+                .remove(account_id.as_str());
+            self.cleared
+                .lock()
+                .expect("cleared accounts lock")
+                .push(account_id.as_str().to_owned());
+            Ok(())
+        })
+    }
+}
+
 async fn setup(
     outcome: Result<TokenPair, RefreshFailure>,
     lease_available: bool,
@@ -242,6 +321,7 @@ async fn setup(
             available: lease_available,
             requests: Mutex::new(Vec::new()),
         }),
+        Arc::new(CountingCredentialState::default()),
         runtime_policy(),
     );
     (store, refresher, service)
@@ -307,6 +387,7 @@ async fn refreshed_identity_rejection_revision_fences_account_as_invalid() {
             available: true,
             requests: Mutex::new(Vec::new()),
         }),
+        Arc::new(CountingCredentialState::default()),
         runtime_policy(),
     );
 
@@ -336,6 +417,7 @@ async fn unavailable_signature_verification_persists_refresh_backoff() {
             available: true,
             requests: Mutex::new(Vec::new()),
         }),
+        Arc::new(CountingCredentialState::default()),
         runtime_policy(),
     );
     let before = SystemTime::now();
@@ -344,15 +426,16 @@ async fn unavailable_signature_verification_persists_refresh_backoff() {
 
     assert!(matches!(
         outcomes.as_slice(),
-        [CodexCredentialRefreshOutcome::Ambiguous { .. }]
+        [CodexCredentialRefreshOutcome::Transient { .. }]
     ));
     let account = store.account("acct_refresh").expect("deferred account");
-    assert_eq!(account.revision().get(), 1);
-    assert_eq!(account.availability(), AccountAvailability::Invalid);
+    assert_eq!(account.revision().get(), 2);
+    // JWKS/签名边界短暂不可用是瞬态：账号保持可用、不被永久失效，仅推进退避重试。
+    assert_eq!(account.availability(), AccountAvailability::Ready);
     assert!(
         account
             .next_refresh_at()
-            .is_some_and(|retry| retry < before)
+            .is_some_and(|retry| retry > before)
     );
 }
 
@@ -382,6 +465,7 @@ async fn unavailable_usage_preserves_rotated_tokens_and_persists_backoff() {
             available: true,
             requests: Mutex::new(Vec::new()),
         }),
+        Arc::new(CountingCredentialState::default()),
         runtime_policy(),
     );
     let before = SystemTime::now();
@@ -445,6 +529,7 @@ async fn permanent_refresh_failure_cannot_overwrite_a_newer_credential_revision(
             available: true,
             requests: Mutex::new(Vec::new()),
         }),
+        Arc::new(CountingCredentialState::default()),
         runtime_policy(),
     );
 
@@ -489,16 +574,18 @@ async fn proven_pre_send_transport_failure_uses_short_cooldown() {
 }
 
 #[tokio::test]
-async fn ambiguous_refresh_does_not_mutate_tokens_or_account_state() {
+async fn transient_refresh_defers_without_invalidating_account() {
     let (store, refresher, service) = setup(Err(RefreshFailure::Transport), true).await;
     let outcomes = service.refresh_due().await.expect("refresh due");
+    // 上游瞬态（含可能已发出的 ambiguous）不再永久失效账号，改为退避重试。
     assert!(matches!(
         outcomes.as_slice(),
-        [CodexCredentialRefreshOutcome::Ambiguous { .. }]
+        [CodexCredentialRefreshOutcome::Transient { .. }]
     ));
     let account = store.account("acct_refresh").expect("account");
-    assert_eq!(account.revision().get(), 1);
-    assert_eq!(account.availability(), AccountAvailability::Invalid);
+    assert_eq!(account.revision().get(), 2);
+    assert_eq!(account.availability(), AccountAvailability::Ready);
+    // 已退避到未来，本轮不再 due，且未重复兑换 token。
     assert!(
         service
             .refresh_due()
@@ -572,6 +659,7 @@ async fn malformed_account_refresh_does_not_stop_later_accounts() {
             available: true,
             requests: Mutex::new(Vec::new()),
         }),
+        Arc::new(CountingCredentialState::default()),
         runtime_policy(),
     );
 
@@ -596,3 +684,128 @@ async fn malformed_account_refresh_does_not_stop_later_accounts() {
         2
     );
 }
+
+async fn seed_due_account(store: &MemoryAccountStore) {
+    store
+        .seed_oauth_credential(ImportCodexOAuthCredential {
+            account_id: "acct_refresh".to_owned(),
+            name: "refresh".to_owned(),
+            secret: secret("old-access"),
+            verified_account: profile("chatgpt-acct_refresh"),
+            next_refresh_at: Some(chrono::Utc::now() - chrono::Duration::seconds(1)),
+            enabled: true,
+        })
+        .await;
+}
+
+#[tokio::test]
+async fn refresh_backoff_grows_exponentially_across_attempts() {
+    let store = Arc::new(MemoryAccountStore::default());
+    seed_due_account(&store).await;
+    let credential_state = Arc::new(CountingCredentialState::default());
+    let state_port: Arc<dyn ProviderCredentialStatePort> = credential_state.clone();
+    let service = CodexCredentialRefreshService::new(
+        store.repository(),
+        Arc::new(Refresher::scripted([
+            Err(RefreshFailure::Transport),
+            Err(RefreshFailure::Transport),
+        ])),
+        Arc::new(VerifiedIdentity),
+        Arc::new(RefreshLeases {
+            available: true,
+            requests: Mutex::new(Vec::new()),
+        }),
+        state_port,
+        runtime_policy(),
+    );
+
+    let before_first = SystemTime::now();
+    let first = service.refresh_due().await.expect("first refresh cycle");
+    assert!(matches!(
+        first.as_slice(),
+        [CodexCredentialRefreshOutcome::Transient { .. }]
+    ));
+    let first_account = store
+        .account("acct_refresh")
+        .expect("account after first defer");
+    let first_delay = first_account
+        .next_refresh_at()
+        .expect("first retry scheduled")
+        .duration_since(before_first)
+        .expect("first delay is in the future");
+    assert_eq!(credential_state.count("acct_refresh"), 1);
+
+    // 已退避到未来的账号真实调度不会再次到期，这里手动复位以触发第二次退避。
+    store
+        .repository()
+        .defer_refresh(&first_account, SystemTime::now() - Duration::from_secs(1))
+        .await
+        .expect("reset next_refresh_at to the past");
+
+    let before_second = SystemTime::now();
+    let second = service.refresh_due().await.expect("second refresh cycle");
+    assert!(matches!(
+        second.as_slice(),
+        [CodexCredentialRefreshOutcome::Transient { .. }]
+    ));
+    let second_delay = store
+        .account("acct_refresh")
+        .expect("account after second defer")
+        .next_refresh_at()
+        .expect("second retry scheduled")
+        .duration_since(before_second)
+        .expect("second delay is in the future");
+    assert_eq!(credential_state.count("acct_refresh"), 2);
+
+    // base=5s、factor=3：第二次（attempt=2）应比第一次（attempt=1）显著更久。
+    assert!(
+        second_delay > first_delay * 2,
+        "second backoff {second_delay:?} should grow well beyond first {first_delay:?}"
+    );
+}
+
+#[tokio::test]
+async fn successful_refresh_clears_backoff_counter() {
+    let store = Arc::new(MemoryAccountStore::default());
+    seed_due_account(&store).await;
+    let credential_state = Arc::new(CountingCredentialState::default());
+    let state_port: Arc<dyn ProviderCredentialStatePort> = credential_state.clone();
+    let service = CodexCredentialRefreshService::new(
+        store.repository(),
+        Arc::new(Refresher::scripted([
+            Err(RefreshFailure::Transport),
+            Ok(success_tokens()),
+        ])),
+        Arc::new(VerifiedIdentity),
+        Arc::new(RefreshLeases {
+            available: true,
+            requests: Mutex::new(Vec::new()),
+        }),
+        state_port,
+        runtime_policy(),
+    );
+
+    // 第一次瞬态失败 → 计数累加到 1。
+    let first = service.refresh_due().await.expect("first refresh cycle");
+    assert!(matches!(
+        first.as_slice(),
+        [CodexCredentialRefreshOutcome::Transient { .. }]
+    ));
+    assert_eq!(credential_state.count("acct_refresh"), 1);
+
+    // 复位到期后成功刷新，退避计数被清零。
+    let deferred = store.account("acct_refresh").expect("deferred account");
+    store
+        .repository()
+        .defer_refresh(&deferred, SystemTime::now() - Duration::from_secs(1))
+        .await
+        .expect("reset next_refresh_at to the past");
+
+    let second = service.refresh_due().await.expect("second refresh cycle");
+    assert!(matches!(
+        second.as_slice(),
+        [CodexCredentialRefreshOutcome::Refreshed { .. }]
+    ));
+    assert!(credential_state.was_cleared("acct_refresh"));
+    assert_eq!(credential_state.count("acct_refresh"), 0);
+}

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

@@ -221,6 +221,24 @@ async fn token_revoked_text_without_invalid_grant_should_remain_temporary() {
     assert_eq!(failure, RefreshFailure::Transport);
 }
 
+#[tokio::test]
+async fn server_error_or_rate_limit_must_stay_transient_even_with_oauth_error_body() {
+    // 5xx/429（含 CDN/网关页恰好嵌入 invalid_grant 字样）按状态码判瞬态，
+    // 不因正文子串把账号永久终态——正文只在 4xx 时才是权威 OAuth 错误。
+    for status in [500, 502, 503, 429] {
+        let failure = refresh_failure(
+            status,
+            r#"{"error":"invalid_grant","error_description":"refresh_token_expired"}"#,
+        )
+        .await;
+        assert_eq!(
+            failure,
+            RefreshFailure::Transport,
+            "status {status} must classify as transient"
+        );
+    }
+}
+
 async fn refresh_failure(status: u16, body: &str) -> RefreshFailure {
     let server = MockServer::start().await;
     Mock::given(method("POST"))

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

@@ -3,6 +3,7 @@ use gateway_core::event::{ContentKind, FinishReason, GatewayEvent, ProviderEvent
 use provider_openai::transport::canonical::{
     CodexCanonicalDecoder, CodexCanonicalError, CodexCanonicalFailure, CodexCanonicalOutcome,
 };
+use provider_openai::transport::protocol::websocket::websocket_event_to_sse_frame;
 
 const METADATA_PREFIX_FIXTURE: &str = include_str!(
     "../../../../gateway-api/tests/openai/responses/fixtures/http_sse/metadata_only_prefix.sse"
@@ -63,6 +64,53 @@ fn raw_sse_passthrough_should_keep_original_bytes_alongside_canonical_facts() {
     ));
 }
 
+#[test]
+fn websocket_raw_passthrough_should_preserve_upstream_number_bytes() {
+    // OpenAI 线路透明代理：WebSocket 上游帧经 reducer→SSE→decoder(raw 透传) 后，
+    // data 段必须与上游原文逐字节一致，不得经 serde 往返改写数值/精度。
+    // `1e3` 若被重序列化会变成 `1000.0`——用它作为字节改写的探针。
+    let upstream = r#"{"type":"response.created","response":{"id":"resp_ws_raw","model":"gpt-test","x_precision":1e3}}"#;
+    let frame = websocket_event_to_sse_frame(upstream)
+        .expect("client-visible WS event yields an SSE frame");
+
+    let events = CodexCanonicalDecoder::new("fallback")
+        .with_raw_sse_passthrough()
+        .push(frame.as_bytes())
+        .expect("WS-derived SSE frame should remain deliverable");
+    let wire = events
+        .iter()
+        .filter_map(ProviderEvent::wire_event)
+        .find(|wire| wire.has_json_data())
+        .expect("response.created wire");
+    let raw = wire
+        .raw_sse_frame()
+        .map(AsRef::as_ref)
+        .expect("raw frame bytes");
+    let raw = std::str::from_utf8(raw).expect("utf8 raw frame");
+
+    assert!(
+        raw.contains("1e3"),
+        "raw frame must keep upstream `1e3` verbatim: {raw}"
+    );
+    assert!(
+        !raw.contains("1000.0"),
+        "raw frame must not re-serialize the number via serde: {raw}"
+    );
+    assert!(matches!(
+        wire_for_response_created(&events),
+        [GatewayEvent::Started(metadata)] if metadata.response_id() == "resp_ws_raw"
+    ));
+}
+
+fn wire_for_response_created(events: &[ProviderEvent]) -> &[GatewayEvent] {
+    events
+        .iter()
+        .find(|event| {
+            event.wire_event().and_then(|wire| wire.event_type()) == Some("response.created")
+        })
+        .map_or(&[], ProviderEvent::canonical_facts)
+}
+
 #[test]
 fn raw_sse_passthrough_should_forward_unparseable_frames_without_failure() {
     let raw = b"retry: later\ndata: opaque\n\n";

backend/crates/providers/openai/tests/transport/protocol/websocket.rs

@@ -90,6 +90,32 @@ fn websocket_response_create_payload_should_keep_explicit_empty_instructions() {
     assert_eq!(value["instructions"], "");
 }
 
+#[test]
+fn websocket_response_create_payload_text_should_match_merged_map_serialization() {
+    // body 自带 type 键与未知字段：文本帧必须与「合并 Map 再序列化」逐字节一致。
+    let request = CodexResponsesRequest::from_body(
+        json!({
+            "model": "gpt-test",
+            "type": "client-supplied",
+            "x_unknown": {"nested": [1, 2.5, "3"]},
+            "input": []
+        })
+        .as_object()
+        .expect("request object")
+        .clone(),
+    );
+
+    let merged = serde_json::to_string(
+        &provider_openai::transport::protocol::websocket::websocket_response_create_payload(
+            &request,
+        ),
+    )
+    .expect("merged payload");
+    let streamed = websocket_response_create_payload_text(&request).expect("streamed payload");
+
+    assert_eq!(streamed, merged);
+}
+
 #[test]
 fn required_websocket_audit_should_forbid_http_fallback() {
     let request = CodexResponsesRequest::from_body(
@@ -138,8 +164,7 @@ fn websocket_metadata_turn_state_should_accept_case_insensitive_header() {
     let event = json!({
         "type": "response.metadata",
         "headers": {"X-Codex-Turn-State": ["turn-from-metadata"]}
-    })
-    .to_string();
+    });
 
     assert_eq!(
         websocket_metadata_turn_state(&event).as_deref(),
@@ -155,16 +180,14 @@ fn websocket_completed_id_should_read_the_id_without_validating_the_response_sha
             "id": "resp_valid",
             "usage": {"input_tokens": 1, "output_tokens": 1, "total_tokens": 2}
         }
-    })
-    .to_string();
+    });
     let invalid_usage = json!({
         "type": "response.completed",
         "response": {
             "id": "resp_invalid",
             "usage": {"input_tokens": "bad", "output_tokens": 1, "total_tokens": 1}
         }
-    })
-    .to_string();
+    });
 
     assert_eq!(
         websocket_response_completed_id(&valid),
@@ -226,16 +249,14 @@ fn websocket_completed_id_should_leave_unreadable_completion_untracked() {
         "response": {
             "usage": {"input_tokens": 1, "output_tokens": 1, "total_tokens": 2}
         }
-    })
-    .to_string();
+    });
     let incomplete_usage = json!({
         "type": "response.completed",
         "response": {
             "id": "resp_incomplete_usage",
             "usage": {"input_tokens": 1, "output_tokens": 1}
         }
-    })
-    .to_string();
+    });
 
     assert_eq!(websocket_response_completed_id(&missing_id), None);
     assert_eq!(

backend/crates/providers/xai/Cargo.toml

@@ -11,6 +11,7 @@ autotests = false
 async-stream = "0.3.6"
 async-trait = "0.1.89"
 base64 = "0.22.1"
+bytes = "1.9.0"
 chrono = { version = "0.4.45", features = ["clock", "serde"] }
 futures = "0.3.32"
 gateway-admin = { path = "../../gateway-admin" }

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

@@ -316,36 +316,81 @@ impl ProviderAdmin for XaiAdminProvider {
             .map_err(map_provider_store_error)?;
         let mut subjects = BTreeSet::new();
         let mut credentials = Vec::new();
+        // 逐条目独立成败：verify 会真实轮换 RT，后续条目的任何失败都不得
+        // 丢弃已轮换成功的凭据（旧 RT 已被上游作废，丢弃即不可逆报废）。
+        // 失败条目跳过；仅当没有任何条目成功时返回首个失败。响应只携带
+        // 成功计数，每个被跳过的条目在此逐条留日志——尤其 verify 成功后
+        // 才失败/被去重的条目，其源文件里的 RT 已被轮换作废，运维只能从
+        // 日志得知需要重新授权。
+        let mut first_failure: Option<ProviderAdminError> = None;
         for entry in document.into_entries() {
             let name = entry.name().to_owned();
             let email = entry.email().map(str::to_owned);
-            let tokens = self
+            let tokens = match self
                 .oauth
                 .verify_imported_credential(&discovery, entry.into_candidate())
                 .await
-                .map_err(|error| map_failure_class(error.class()))?;
+            {
+                Ok(tokens) => tokens,
+                Err(error) => {
+                    tracing::warn!(
+                        target: "xai_admin",
+                        account = name.as_str(),
+                        class = ?error.class(),
+                        "导入条目验证失败，已跳过"
+                    );
+                    first_failure.get_or_insert_with(|| map_failure_class(error.class()));
+                    continue;
+                }
+            };
+            // 同一账号的另一份 grant：账号已由更早条目入册，本条目跳过；
+            // 本条目若经历了 refresh，其源文件 RT 已作废。
             if !subjects.insert(tokens.evidence().subject().to_owned()) {
-                return Err(provider_error(ProviderAdminErrorKind::Invalid));
+                tracing::warn!(
+                    target: "xai_admin",
+                    account = name.as_str(),
+                    "导入条目与更早条目属同一账号，已跳过；该条目的 refresh token 可能已被轮换作废"
+                );
+                continue;
             }
-            let prepared = GrokCredentialAdmin
-                .prepare_verified_account(
-                    &VerifiedGrokAccount {
-                        account_id: ProviderAccountId::new(format!(
-                            "acct_{}",
-                            Uuid::now_v7().simple()
-                        ))
-                        .map_err(|_| provider_error(ProviderAdminErrorKind::Internal))?,
-                        name,
-                        email,
-                        upstream_account_id: None,
-                        plan_type: None,
-                        tokens,
-                        enabled: true,
-                    },
-                    policy,
-                )
-                .map_err(map_repository_error)?;
-            credentials.push(prepared_create(prepared, Utc::now())?);
+            let account = VerifiedGrokAccount {
+                account_id: ProviderAccountId::new(format!("acct_{}", Uuid::now_v7().simple()))
+                    .map_err(|_| provider_error(ProviderAdminErrorKind::Internal))?,
+                name,
+                email,
+                upstream_account_id: None,
+                plan_type: None,
+                tokens,
+                enabled: true,
+            };
+            let prepared = match GrokCredentialAdmin.prepare_verified_account(&account, policy) {
+                Ok(prepared) => prepared,
+                Err(error) => {
+                    tracing::warn!(
+                        target: "xai_admin",
+                        account = account.name.as_str(),
+                        "已验证条目本地入册失败，已跳过；其 refresh token 已被轮换，需重新授权该账号"
+                    );
+                    first_failure.get_or_insert_with(|| map_repository_error(error));
+                    continue;
+                }
+            };
+            match prepared_create(prepared, Utc::now()) {
+                Ok(credential) => credentials.push(credential),
+                Err(error) => {
+                    tracing::warn!(
+                        target: "xai_admin",
+                        account = account.name.as_str(),
+                        "已验证条目序列化失败，已跳过；其 refresh token 已被轮换，需重新授权该账号"
+                    );
+                    first_failure.get_or_insert(error);
+                }
+            }
+        }
+        if credentials.is_empty() {
+            return Err(
+                first_failure.unwrap_or_else(|| provider_error(ProviderAdminErrorKind::Invalid))
+            );
         }
         Ok(PreparedCredentialImport {
             provider_kind: self.provider_kind.clone(),

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

@@ -172,6 +172,12 @@ impl GrokOAuthImportDocument {
                 entries.push(entry);
             }
         }
+        // 同一 refresh token 的重复条目只保留首个：RT exchange 会作废旧
+        // grant，重复条目逐条交换会让先完成的轮换被后续条目作废。
+        let mut seen_refresh_tokens = HashSet::new();
+        entries.retain(|entry| {
+            seen_refresh_tokens.insert(entry.candidate.refresh_token.expose().to_owned())
+        });
         if entries.is_empty() {
             return Err(GrokOAuthImportError::InvalidField("document"));
         }

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

@@ -12,10 +12,10 @@ use gateway_core::engine::credential::{
     AccountAvailability, CredentialRevision, ProviderAccountId,
 };
 use gateway_core::provider_ports::{
-    ProviderCooldown, ProviderCooldownPort, ProviderLeaseAcquisition, ProviderLeasePort,
-    ProviderLeaseRequest, ProviderRefreshCapacityRequest, ProviderRefreshLeaseRequest,
-    ProviderRefreshPolicy, ProviderRuntimePolicyPort, ProviderStoreError,
-    provider_refresh_retry_at,
+    ProviderCooldown, ProviderCooldownPort, ProviderCredentialStatePort, ProviderLeaseAcquisition,
+    ProviderLeasePort, ProviderLeaseRequest, ProviderRefreshCapacityRequest,
+    ProviderRefreshLeaseRequest, ProviderRefreshPolicy, ProviderRuntimePolicyPort,
+    ProviderStoreError, provider_refresh_backoff_at,
 };
 
 use super::catalog::GrokCredentialCatalogService;
@@ -36,7 +36,8 @@ const MAX_REFRESH_BATCH: u32 = 100;
 const MAX_REFRESH_EXCLUSIONS: usize = 400;
 const MAX_SECRET_BYTES: usize = 64 * 1_024;
 const DISCOVERY_NEGATIVE_CACHE_TTL: Duration = Duration::from_secs(5);
-const REFRESH_RETRY_DELAY: Duration = Duration::from_secs(10 * 60);
+/// 连续失败计数窗口；静默满窗后计数过期归零。
+const REFRESH_BACKOFF_WINDOW: Duration = Duration::from_secs(30 * 60);
 const REFRESH_AMBIGUITY_COOLDOWN: Duration = Duration::from_secs(24 * 60 * 60);
 
 /// 一个到期且已按 revision 读取明文 RT 的 xAI account。
@@ -277,6 +278,7 @@ pub struct GrokCredentialRefreshService {
     catalog: Arc<GrokCredentialCatalogService>,
     leases: Arc<dyn ProviderLeasePort>,
     cooldowns: Arc<dyn ProviderCooldownPort>,
+    credential_state: Arc<dyn ProviderCredentialStatePort>,
     runtime_policy: Arc<dyn ProviderRuntimePolicyPort>,
 }
 
@@ -287,6 +289,7 @@ impl GrokCredentialRefreshService {
         catalog: Arc<GrokCredentialCatalogService>,
         leases: Arc<dyn ProviderLeasePort>,
         cooldowns: Arc<dyn ProviderCooldownPort>,
+        credential_state: Arc<dyn ProviderCredentialStatePort>,
         runtime_policy: Arc<dyn ProviderRuntimePolicyPort>,
     ) -> Self {
         Self {
@@ -295,6 +298,7 @@ impl GrokCredentialRefreshService {
             catalog,
             leases,
             cooldowns,
+            credential_state,
             runtime_policy,
         }
     }
@@ -339,17 +343,6 @@ impl GrokCredentialRefreshService {
         self.refresh_one_with_policy(credential, policy).await
     }
 
-    pub async fn prepare_cycle_if_due(&self) -> Result<bool, GrokCredentialRefreshError> {
-        if !self.repository.has_due_refresh().await? {
-            return Ok(false);
-        }
-        self.refresher
-            .prepare_cycle()
-            .await
-            .map_err(|_| GrokCredentialRefreshError::Preparation)?;
-        Ok(true)
-    }
-
     /// 手工刷新一个指定 revision；只返回 Provider 验证后的 CAS command，不写 Store。
     pub async fn prepare_manual_refresh(
         &self,
@@ -650,6 +643,10 @@ impl GrokCredentialRefreshService {
         {
             let _ = self.catalog.cache_seed(&account_id, seed).await;
         }
+        let _ = self
+            .credential_state
+            .clear_refresh_backoff(&account_id)
+            .await;
         Ok(GrokCredentialRefreshOutcome::Refreshed {
             account_id,
             credential_revision: record.credential_revision,
@@ -729,10 +726,15 @@ impl GrokCredentialRefreshService {
         credential: &DueGrokCredential,
         reason: &'static str,
     ) -> Result<bool, GrokCredentialRefreshError> {
-        let retry_at = provider_refresh_retry_at(
+        let attempt = self
+            .credential_state
+            .record_refresh_backoff(&credential.account_id, REFRESH_BACKOFF_WINDOW)
+            .await
+            .unwrap_or(1);
+        let retry_at = provider_refresh_backoff_at(
             &credential.account_id,
             SystemTime::now(),
-            REFRESH_RETRY_DELAY,
+            attempt,
             reason,
         )?;
         match self
@@ -781,22 +783,6 @@ impl GrokCredentialRecovery for GrokCredentialRefreshService {
 }
 
 impl GrokCredentialRepository {
-    pub async fn has_due_refresh(&self) -> Result<bool, GrokCredentialRepositoryError> {
-        Ok(self
-            .list_all_accounts()
-            .await?
-            .iter()
-            .any(|account| account_due(account, SystemTime::now())))
-    }
-
-    pub async fn list_due_refresh(
-        &self,
-    ) -> Result<Vec<DueGrokCredential>, GrokCredentialRepositoryError> {
-        self.list_due_refresh_batch_excluding(&[])
-            .await
-            .map(|batch| batch.credentials)
-    }
-
     async fn list_due_refresh_batch_excluding(
         &self,
         excluded_account_ids: &[ProviderAccountId],

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

@@ -102,7 +102,7 @@ impl GrokAccountSessionSelector {
         let runtime_cooldowns = self.runtime_cooldowns(&catalog_eligible).await;
         let mut candidates = catalog_eligible
             .into_iter()
-            .filter(|account| !runtime_cooldowns.contains(account.id()))
+            .filter(|account| !runtime_cooldowns.contains_key(account.id()))
             .map(|account| {
                 let health = self
                     .account_feedback
@@ -197,18 +197,30 @@ impl GrokAccountSessionSelector {
         }
 
         if capacity_denied {
-            Err(GrokSessionSelectorError::CapacityUnavailable { retry_after })
-        } else {
-            Err(GrokSessionSelectorError::NoEligibleSession)
+            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() {
+            Some(required) => runtime_cooldowns.get(required).copied(),
+            None => runtime_cooldowns.values().copied().min(),
+        };
+        if let Some(until) = cooled_until {
+            return Err(GrokSessionSelectorError::AccountCoolingDown {
+                retry_after: until
+                    .duration_since(SystemTime::now())
+                    .ok()
+                    .filter(|remaining| !remaining.is_zero()),
+            });
+        }
+        Err(GrokSessionSelectorError::NoEligibleSession)
     }
 
     async fn runtime_cooldowns(
         &self,
         accounts: &[ProviderAccount],
-    ) -> std::collections::BTreeSet<ProviderAccountId> {
+    ) -> std::collections::BTreeMap<ProviderAccountId, SystemTime> {
         let now = SystemTime::now();
-        let mut cooled = std::collections::BTreeSet::new();
+        let mut cooled = std::collections::BTreeMap::new();
         for account in accounts {
             let Ok(Some(cooldown)) = self.cooldowns.read(account.id()).await else {
                 continue;
@@ -220,7 +232,7 @@ impl GrokAccountSessionSelector {
                 continue;
             }
             if cooldown.until() > now {
-                cooled.insert(account.id().clone());
+                cooled.insert(account.id().clone(), cooldown.until());
             }
         }
         cooled

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

@@ -57,24 +57,25 @@ pub use transport::{
     GrokBillingTransportError, GrokBillingTransportErrorKind, GrokBillingTransportFuture,
     GrokBillingTransportResponse, GrokCanonicalDecoder, GrokCatalogApiBackend,
     GrokCatalogCapabilities, GrokCatalogCapabilityEvidence, GrokCatalogLimits, GrokCatalogMetadata,
-    GrokCatalogModel, GrokClientIdentity, GrokCompactionDecodeError, GrokCompactionRequest,
-    GrokCompactionSummaryDecoder, GrokCredentialFailure, GrokCredentialFeedbackFuture,
-    GrokDnsResolutionError, GrokDnsResolutionPlan, GrokDnsResolutionPolicy, GrokEndpointPolicy,
-    GrokHeader, GrokHeaderValue, GrokInferenceChunkStream, GrokInferenceClientCacheStatus,
-    GrokInferenceDnsObservation, GrokInferenceDnsSource, GrokInferenceRequest,
-    GrokInferenceResponse, GrokInferenceTransport, GrokInferenceTransportError,
-    GrokInferenceTransportErrorKind, GrokInferenceTransportFuture, GrokInferenceTransportMetrics,
-    GrokModelCatalogClient, GrokModelCatalogError, GrokModelCatalogRequest,
-    GrokModelCatalogSession, GrokModelCatalogSessionError, GrokModelCatalogSnapshot,
-    GrokModelCatalogTransport, GrokModelCatalogTransportError, GrokModelCatalogTransportErrorKind,
-    GrokModelCatalogTransportFuture, GrokModelCatalogTransportResponse, GrokProviderConfigError,
-    GrokProviderTransport, GrokRequestEncodeError, GrokReqwestTransportBuildError,
-    GrokResponsesRequest, GrokSessionAffinityKey, GrokSessionBinding, GrokSessionDataError,
-    GrokSessionLeaseGuard, GrokSessionSelection, GrokSessionSelector, GrokSessionSelectorError,
-    GrokSessionSelectorFuture, MAX_GROK_BILLING_BYTES, MAX_GROK_MODEL_CATALOG_BYTES,
-    OfficialGrokEndpointPolicy, ReqwestGrokInferenceTransport, ReqwestGrokModelCatalogTransport,
-    ReqwestOAuthTransport, SelectedGrokSession, XAI_PROVIDER_NAME, build_grok_headers,
-    grok_billing_breakdown, parse_grok_billing, parse_grok_model_catalog,
+    GrokCatalogModel, GrokCatalogReasoningEffort, GrokClientIdentity, GrokCompactionDecodeError,
+    GrokCompactionRequest, GrokCompactionSummaryDecoder, GrokCredentialFailure,
+    GrokCredentialFeedbackFuture, GrokDnsResolutionError, GrokDnsResolutionPlan,
+    GrokDnsResolutionPolicy, GrokEndpointPolicy, GrokHeader, GrokHeaderValue,
+    GrokInferenceChunkStream, GrokInferenceClientCacheStatus, GrokInferenceDnsObservation,
+    GrokInferenceDnsSource, GrokInferenceRequest, GrokInferenceResponse, GrokInferenceTransport,
+    GrokInferenceTransportError, GrokInferenceTransportErrorKind, GrokInferenceTransportFuture,
+    GrokInferenceTransportMetrics, GrokModelCatalogClient, GrokModelCatalogError,
+    GrokModelCatalogRequest, GrokModelCatalogSession, GrokModelCatalogSessionError,
+    GrokModelCatalogSnapshot, GrokModelCatalogTransport, GrokModelCatalogTransportError,
+    GrokModelCatalogTransportErrorKind, GrokModelCatalogTransportFuture,
+    GrokModelCatalogTransportResponse, GrokProviderConfigError, GrokProviderTransport,
+    GrokRequestEncodeError, GrokReqwestTransportBuildError, GrokResponsesRequest,
+    GrokSessionAffinityKey, GrokSessionBinding, GrokSessionDataError, GrokSessionLeaseGuard,
+    GrokSessionSelection, GrokSessionSelector, GrokSessionSelectorError, GrokSessionSelectorFuture,
+    MAX_GROK_BILLING_BYTES, MAX_GROK_MODEL_CATALOG_BYTES, OfficialGrokEndpointPolicy,
+    ReqwestGrokInferenceTransport, ReqwestGrokModelCatalogTransport, ReqwestOAuthTransport,
+    SelectedGrokSession, XAI_PROVIDER_NAME, build_grok_headers, grok_billing_breakdown,
+    parse_grok_billing, parse_grok_model_catalog,
 };
 
 /// xAI 初始化后交给组装根的最小能力集。
@@ -167,6 +168,7 @@ pub async fn initialize(
         Arc::clone(&catalog),
         Arc::clone(&leases),
         cooldowns,
+        ports.credential_state(),
         Arc::clone(&runtime_policy),
     ));
     let credential_recovery: Arc<dyn GrokCredentialRecovery> = refresh.clone();

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

@@ -430,7 +430,9 @@ fn default_grok_model_presentation() -> ModelPresentation {
     )
     .with_reasoning(
         Some("medium".to_owned()),
-        ["low", "medium", "high"].map(str::to_owned).to_vec(),
+        ["low", "medium", "high", "xhigh"]
+            .map(str::to_owned)
+            .to_vec(),
     )
     .with_context_window_tokens(Some(500_000))
     .with_image_input(true)
@@ -444,17 +446,35 @@ fn grok_model_presentation(model: &GrokCatalogModel) -> ModelPresentation {
         DEFAULT_GROK_MODEL | "grok-4.5-latest" | "grok-4.5-build-free" | "grok-build-latest"
     );
     let reasoning_evidence = model.capabilities().reasoning_effort();
+    let catalog_reasoning_efforts = model
+        .capabilities()
+        .reasoning_efforts()
+        .iter()
+        .map(|effort| effort.as_str().to_owned())
+        .collect::<Vec<_>>();
+    let catalog_default_reasoning = model
+        .capabilities()
+        .default_reasoning_effort()
+        .map(|effort| effort.as_str().to_owned());
     let (default_reasoning, reasoning_efforts) = match reasoning_evidence {
+        GrokCatalogCapabilityEvidence::DeclaredNative if !catalog_reasoning_efforts.is_empty() => (
+            catalog_default_reasoning.or_else(|| catalog_reasoning_efforts.first().cloned()),
+            catalog_reasoning_efforts,
+        ),
         GrokCatalogCapabilityEvidence::DeclaredNative => (
-            Some("medium".to_owned()),
-            ["low", "medium", "high"].map(str::to_owned).to_vec(),
+            catalog_default_reasoning.or_else(|| Some("medium".to_owned())),
+            ["low", "medium", "high", "xhigh"]
+                .map(str::to_owned)
+                .to_vec(),
         ),
         GrokCatalogCapabilityEvidence::DeclaredUnsupported => {
             (Some("none".to_owned()), vec!["none".to_owned()])
         }
         GrokCatalogCapabilityEvidence::Unknown if known_grok_4_5 => (
-            Some("medium".to_owned()),
-            ["low", "medium", "high"].map(str::to_owned).to_vec(),
+            catalog_default_reasoning.or_else(|| Some("medium".to_owned())),
+            ["low", "medium", "high", "xhigh"]
+                .map(str::to_owned)
+                .to_vec(),
         ),
         GrokCatalogCapabilityEvidence::Unknown => (None, Vec::new()),
     };
@@ -969,7 +989,7 @@ async fn next_grok_chunk(
     selector: &dyn GrokSessionSelector,
     session: &SelectedGrokSession,
     context: &AttemptContext,
-) -> Result<Option<Vec<u8>>, ProviderError> {
+) -> Result<Option<bytes::Bytes>, ProviderError> {
     let Some(stream_deadline) = remaining(context.deadline()) else {
         return Err(provider_error(
             ProviderErrorKind::Timeout,
@@ -1572,21 +1592,55 @@ fn map_request_error(error: GrokRequestEncodeError) -> ProviderError {
     }
 }
 
+/// 将选择阶段失败映射为带结构化 code 与 retry_after 的 Provider 错误。
 fn map_selection_error(error: GrokSessionSelectorError) -> ProviderError {
-    match error {
-        GrokSessionSelectorError::CapacityUnavailable { retry_after } => {
-            let error = provider_error(ProviderErrorKind::Unavailable, UpstreamSendState::NotSent);
-            match retry_after {
-                Some(retry_after) => error.with_retry_after(retry_after),
-                None => error,
-            }
-        }
-        GrokSessionSelectorError::NoEligibleSession | GrokSessionSelectorError::Unavailable => {
-            provider_error(ProviderErrorKind::Unavailable, UpstreamSendState::NotSent)
-        }
+    let (retry_after, message, code) = match error {
+        GrokSessionSelectorError::AccountCoolingDown { retry_after } => (
+            retry_after,
+            cooling_down_message(retry_after),
+            "account_cooling_down",
+        ),
+        GrokSessionSelectorError::CapacityUnavailable { retry_after } => (
+            retry_after,
+            "account is at its concurrency or request-interval limit".to_owned(),
+            "account_capacity_busy",
+        ),
+        GrokSessionSelectorError::NoEligibleSession => (
+            None,
+            "no account is eligible for the requested model".to_owned(),
+            "no_eligible_account",
+        ),
+        GrokSessionSelectorError::Unavailable => (
+            None,
+            "account scheduling state is temporarily unreadable".to_owned(),
+            "account_selector_unavailable",
+        ),
         GrokSessionSelectorError::InvalidSession => {
-            provider_error(ProviderErrorKind::Protocol, UpstreamSendState::NotSent)
+            return provider_error(ProviderErrorKind::Protocol, UpstreamSendState::NotSent);
         }
+    };
+    let error = provider_error(ProviderErrorKind::Unavailable, UpstreamSendState::NotSent);
+    let error = match retry_after {
+        Some(retry_after) => error.with_retry_after(retry_after),
+        None => error,
+    };
+    match ClientVisibleUpstreamError::new(
+        message,
+        Some(code.to_owned()),
+        Some("account_unavailable_error".to_owned()),
+    ) {
+        Ok(detail) => error.with_client_visible_upstream_error(detail),
+        Err(_) => error,
+    }
+}
+
+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",
+            retry_after.as_secs().saturating_add(1)
+        ),
+        None => "account is cooling down after an interrupted upstream stream".to_owned(),
     }
 }
 

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

@@ -7,7 +7,9 @@ use gateway_core::accounting::{
     CurrencyCode, Decimal, Money, ProviderReportedCost, Usage,
 };
 use gateway_core::engine::UpstreamSendState;
-use gateway_core::error::{ProviderError, ProviderErrorKind, SafeUpstreamValue};
+use gateway_core::error::{
+    ClientVisibleUpstreamError, ProviderError, ProviderErrorKind, SafeUpstreamValue,
+};
 use gateway_core::event::{
     ContentItem, ContentKind, FinishReason, GatewayEvent, ProtocolWireEvent, ProviderEvent,
     ReasoningDelta, ResponseMeta, TextDelta, ToolCallDelta,
@@ -211,10 +213,13 @@ impl GrokCanonicalDecoder {
                 continue;
             };
             let event_type = event_type.to_owned();
-            let transformed = self
+            // 无法转换的上游帧（多为畸形帧）直接丢弃，保持已开始的客户端流不中断。
+            let Ok(transformed) = self
                 .response_transform
                 .rewrite_stream_event(&event_type, value)
-                .map_err(protocol_error)?;
+            else {
+                continue;
+            };
             for (index, transformed) in transformed.into_iter().enumerate() {
                 let transformed_type = transformed.event_type().to_owned();
                 if !client_visible_event(&transformed_type) {
@@ -222,7 +227,23 @@ impl GrokCanonicalDecoder {
                 }
                 let value = transformed.into_value();
                 let mut canonical = Vec::new();
-                self.decode_event(&transformed_type, &value, &mut canonical)?;
+                // 终态事件（completed/incomplete）fail-closed：用量/计费校验失败即断流。
+                // 其余内容事件容忍字段校验失败——正常上游变体（空 delta、重复 index、
+                // 缺字段等）不打断已开始的客户端流：跳过 canonical 提取、wire 原样转发。
+                // 真·上游错误（response.failed/error）为非 Protocol 类别，按终态传播。
+                let terminal_event = matches!(
+                    transformed_type.as_str(),
+                    "response.completed" | "response.incomplete"
+                );
+                match self.decode_event(&transformed_type, &value, &mut canonical) {
+                    Ok(()) => {}
+                    Err(error)
+                        if !terminal_event && error.kind() == ProviderErrorKind::Protocol =>
+                    {
+                        canonical.clear();
+                    }
+                    Err(error) => return Err(error),
+                }
                 let wire_event = if transformed_type == event_type {
                     event.event.clone()
                 } else {
@@ -762,11 +783,7 @@ fn incomplete_finish_reason(response: &Value) -> FinishReason {
 }
 
 fn upstream_event_error(value: &Value) -> ProviderError {
-    let code = value
-        .pointer("/error/code")
-        .or_else(|| value.pointer("/response/error/code"))
-        .or_else(|| value.get("code"))
-        .and_then(Value::as_str);
+    let code = upstream_error_field(value, "code");
     let kind = match code {
         Some("invalid_request" | "invalid_prompt") => ProviderErrorKind::InvalidRequest,
         Some("unsupported" | "unsupported_feature") => ProviderErrorKind::Unsupported,
@@ -781,9 +798,75 @@ fn upstream_event_error(value: &Value) -> ProviderError {
     if let Some(code) = code.and_then(|code| SafeUpstreamValue::new(code.to_owned()).ok()) {
         error = error.with_upstream_code(code);
     }
+    // 结构化 message/code/type 供原客户端展示与重试分类；message 先脱去
+    // 账号指纹（上游限流文案内嵌 team UUID），非结构化正文不透出。
+    if let Some(message) = upstream_error_field(value, "message")
+        && let Ok(client_error) = ClientVisibleUpstreamError::new(
+            scrub_account_fingerprints(message),
+            code.map(str::to_owned),
+            upstream_error_field(value, "type").map(str::to_owned),
+        )
+    {
+        error = error.with_client_visible_upstream_error(client_error);
+    }
     error
 }
 
+fn upstream_error_field<'a>(value: &'a Value, field: &str) -> Option<&'a str> {
+    value
+        .pointer(&format!("/error/{field}"))
+        .or_else(|| value.pointer(&format!("/response/error/{field}")))
+        .or_else(|| value.get(field))
+        .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/catalog.rs

@@ -589,11 +589,46 @@ pub enum GrokCatalogApiBackend {
     Unknown,
 }
 
+/// Grok 官方目录允许声明的 reasoning effort。
+#[derive(Debug, Clone, Copy, PartialEq, Eq, Deserialize)]
+#[serde(rename_all = "lowercase")]
+pub enum GrokCatalogReasoningEffort {
+    None,
+    Minimal,
+    Low,
+    Medium,
+    High,
+    Xhigh,
+    Max,
+    /// 上游新增、当前 adapter 尚未识别的 effort；不进入能力投影。
+    #[serde(other)]
+    Unknown,
+}
+
+impl GrokCatalogReasoningEffort {
+    /// 返回 Responses API 使用的规范值。
+    #[must_use]
+    pub const fn as_str(self) -> &'static str {
+        match self {
+            Self::None => "none",
+            Self::Minimal => "minimal",
+            Self::Low => "low",
+            Self::Medium => "medium",
+            Self::High => "high",
+            Self::Xhigh => "xhigh",
+            Self::Max => "max",
+            Self::Unknown => "unknown",
+        }
+    }
+}
+
 /// Grok 目录中允许进入控制面的能力证据。
 #[derive(Debug, Clone, PartialEq, Eq)]
 pub struct GrokCatalogCapabilities {
     responses_api: GrokCatalogCapabilityEvidence,
     reasoning_effort: GrokCatalogCapabilityEvidence,
+    reasoning_efforts: Vec<GrokCatalogReasoningEffort>,
+    default_reasoning_effort: Option<GrokCatalogReasoningEffort>,
     backend_search: GrokCatalogCapabilityEvidence,
     streaming_tool_calls: GrokCatalogCapabilityEvidence,
     api_backend: Option<GrokCatalogApiBackend>,
@@ -612,6 +647,18 @@ impl GrokCatalogCapabilities {
         self.reasoning_effort
     }
 
+    /// 返回上游声明的可选 reasoning effort，顺序与目录一致。
+    #[must_use]
+    pub fn reasoning_efforts(&self) -> &[GrokCatalogReasoningEffort] {
+        &self.reasoning_efforts
+    }
+
+    /// 返回上游声明或菜单推导出的默认 reasoning effort。
+    #[must_use]
+    pub const fn default_reasoning_effort(&self) -> Option<GrokCatalogReasoningEffort> {
+        self.default_reasoning_effort
+    }
+
     /// 返回后端 search 支持证据。
     #[must_use]
     pub const fn backend_search(&self) -> GrokCatalogCapabilityEvidence {
@@ -850,6 +897,10 @@ struct GrokModelWire {
         alias = "supports_reasoning_effort"
     )]
     supports_reasoning_effort: Option<bool>,
+    #[serde(rename = "reasoningEffort", alias = "reasoning_effort")]
+    reasoning_effort: Option<GrokCatalogReasoningEffort>,
+    #[serde(default, rename = "reasoningEfforts", alias = "reasoning_efforts")]
+    reasoning_efforts: Vec<GrokReasoningEffortOptionWire>,
     #[serde(rename = "supportsBackendSearch", alias = "supports_backend_search")]
     supports_backend_search: Option<bool>,
     #[serde(rename = "streamToolCalls", alias = "stream_tool_calls")]
@@ -857,6 +908,26 @@ struct GrokModelWire {
     hidden: Option<bool>,
 }
 
+#[derive(Debug, Deserialize)]
+#[serde(untagged)]
+enum GrokReasoningEffortOptionWire {
+    Bare(GrokCatalogReasoningEffort),
+    Detailed {
+        value: GrokCatalogReasoningEffort,
+        #[serde(default)]
+        default: bool,
+    },
+}
+
+impl GrokReasoningEffortOptionWire {
+    fn into_parts(self) -> (GrokCatalogReasoningEffort, bool) {
+        match self {
+            Self::Bare(value) => (value, false),
+            Self::Detailed { value, default } => (value, default),
+        }
+    }
+}
+
 /// 解析唯一受支持的 Grok 官方完整快照。
 ///
 /// # Errors
@@ -922,15 +993,41 @@ fn normalize_model(wire: GrokModelWire) -> Result<GrokCatalogModel, GrokModelCat
         .api_backend
         .filter(|backend| *backend != GrokCatalogApiBackend::Unknown);
     let responses_api = responses_evidence(wire.supported_in_api, api_backend);
+    let mut menu_default = None;
+    // 未识别的 effort 值不进入能力投影，也不作为默认值候选。
+    let reasoning_efforts = wire
+        .reasoning_efforts
+        .into_iter()
+        .filter_map(|option| {
+            let (effort, is_default) = option.into_parts();
+            if effort == GrokCatalogReasoningEffort::Unknown {
+                return None;
+            }
+            if is_default && menu_default.is_none() {
+                menu_default = Some(effort);
+            }
+            Some(effort)
+        })
+        .collect::<Vec<_>>();
+    let default_reasoning_effort = wire
+        .reasoning_effort
+        .filter(|effort| *effort != GrokCatalogReasoningEffort::Unknown)
+        .or(menu_default)
+        .or_else(|| reasoning_efforts.first().copied());
+    let reasoning_effort = if reasoning_efforts.is_empty() {
+        GrokCatalogCapabilityEvidence::from_wire(wire.supports_reasoning_effort)
+    } else {
+        GrokCatalogCapabilityEvidence::DeclaredNative
+    };
 
     Ok(GrokCatalogModel {
         request_model,
         display_name: wire.name,
         capabilities: GrokCatalogCapabilities {
             responses_api,
-            reasoning_effort: GrokCatalogCapabilityEvidence::from_wire(
-                wire.supports_reasoning_effort,
-            ),
+            reasoning_effort,
+            reasoning_efforts,
+            default_reasoning_effort,
             backend_search: GrokCatalogCapabilityEvidence::from_wire(wire.supports_backend_search),
             streaming_tool_calls: GrokCatalogCapabilityEvidence::from_wire(wire.stream_tool_calls),
             api_backend,

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

@@ -21,10 +21,10 @@ pub use catalog::{
     GrokBillingRequest, GrokBillingSnapshot, GrokBillingTransport, GrokBillingTransportError,
     GrokBillingTransportErrorKind, GrokBillingTransportFuture, GrokBillingTransportResponse,
     GrokCatalogApiBackend, GrokCatalogCapabilities, GrokCatalogCapabilityEvidence,
-    GrokCatalogLimits, GrokCatalogMetadata, GrokCatalogModel, GrokModelCatalogClient,
-    GrokModelCatalogError, GrokModelCatalogRequest, GrokModelCatalogSession,
-    GrokModelCatalogSessionError, GrokModelCatalogSnapshot, GrokModelCatalogTransport,
-    GrokModelCatalogTransportError, GrokModelCatalogTransportErrorKind,
+    GrokCatalogLimits, GrokCatalogMetadata, GrokCatalogModel, GrokCatalogReasoningEffort,
+    GrokModelCatalogClient, GrokModelCatalogError, GrokModelCatalogRequest,
+    GrokModelCatalogSession, GrokModelCatalogSessionError, GrokModelCatalogSnapshot,
+    GrokModelCatalogTransport, GrokModelCatalogTransportError, GrokModelCatalogTransportErrorKind,
     GrokModelCatalogTransportFuture, GrokModelCatalogTransportResponse, MAX_GROK_BILLING_BYTES,
     MAX_GROK_MODEL_CATALOG_BYTES, parse_grok_billing, parse_grok_model_catalog,
 };

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

@@ -409,7 +409,8 @@ impl GrokInferenceTransport for ReqwestGrokInferenceTransport {
                                 .is_some_and(|total| total <= MAX_INFERENCE_BODY_BYTES) =>
                         {
                             *observed += chunk.len();
-                            Ok(chunk.to_vec())
+                            // Bytes 直传，长流不再逐 chunk 复制到 Vec。
+                            Ok(chunk)
                         }
                         Ok(_) => Err(GrokInferenceTransportError::new(
                             GrokInferenceTransportErrorKind::Protocol,

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

@@ -311,6 +311,12 @@ pub enum GrokSessionSelectorError {
         /// Optional selector-derived delay.
         retry_after: Option<Duration>,
     },
+    /// Every eligible session is inside its runtime cooldown window.
+    #[error("Grok Build account is cooling down")]
+    AccountCoolingDown {
+        /// Remaining cooldown of the earliest recovering session.
+        retry_after: Option<Duration>,
+    },
     /// Session metadata or Provider-owned plaintext secret is invalid.
     #[error("Grok Build session data is invalid")]
     InvalidSession,

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

@@ -77,7 +77,7 @@ impl fmt::Debug for GrokInferenceRequest {
 
 /// Stream of raw SSE byte chunks returned after the POST is accepted.
 pub type GrokInferenceChunkStream =
-    Pin<Box<dyn Stream<Item = Result<Vec<u8>, GrokInferenceTransportError>> + Send + 'static>>;
+    Pin<Box<dyn Stream<Item = Result<bytes::Bytes, GrokInferenceTransportError>> + Send + 'static>>;
 
 /// Whether the account-isolated inference client was already cached.
 #[derive(Debug, Clone, Copy, PartialEq, Eq)]

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

@@ -775,6 +775,21 @@ impl ProviderCredentialStatePort for TestCredentialState {
     ) -> BoxFuture<'a, Result<bool, ProviderStoreError>> {
         Box::pin(async { Ok(false) })
     }
+
+    fn record_refresh_backoff<'a>(
+        &'a self,
+        _account_id: &'a ProviderAccountId,
+        _window: Duration,
+    ) -> BoxFuture<'a, Result<u32, ProviderStoreError>> {
+        Box::pin(async { Ok(1) })
+    }
+
+    fn clear_refresh_backoff<'a>(
+        &'a self,
+        _account_id: &'a ProviderAccountId,
+    ) -> BoxFuture<'a, Result<(), ProviderStoreError>> {
+        Box::pin(async { Ok(()) })
+    }
 }
 
 struct TestCooldown;

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

@@ -283,7 +283,7 @@ fn minimal_oauth_credential_should_not_require_a_source_format() {
 }
 
 #[test]
-fn oauth_account_document_should_not_require_unique_source_names() {
+fn duplicate_refresh_token_entries_should_collapse_to_the_first() {
     let mut document: serde_json::Value =
         serde_json::from_slice(&oauth_account_document()).expect("fixture JSON");
     let duplicate = document["accounts"][0].clone();
@@ -292,10 +292,29 @@ fn oauth_account_document_should_not_require_unique_source_names() {
         .expect("accounts array")
         .push(duplicate);
 
+    // 同一 refresh token 重复出现时解析层去重，避免逐条 RT exchange
+    // 让先完成的轮换被后续重复条目作废。
     let parsed = GrokOAuthImportDocument::parse_json(
         &serde_json::to_vec(&document).expect("serialize fixture"),
     )
-    .expect("upstream identity verification owns duplicate detection");
+    .expect("duplicate refresh tokens collapse at parse");
+    assert_eq!(parsed.into_entries().len(), 1);
+
+    // 不同 refresh token 的同名条目不受影响，重复检测仍归上游身份验证。
+    let mut document: serde_json::Value =
+        serde_json::from_slice(&oauth_account_document()).expect("fixture JSON");
+    let mut distinct = document["accounts"][0].clone();
+    distinct["credentials"]["refresh_token"] =
+        serde_json::Value::String("fixture-refresh-token-second".to_owned());
+    document["accounts"]
+        .as_array_mut()
+        .expect("accounts array")
+        .push(distinct);
+
+    let parsed = GrokOAuthImportDocument::parse_json(
+        &serde_json::to_vec(&document).expect("serialize fixture"),
+    )
+    .expect("distinct refresh tokens keep their entries");
 
     assert_eq!(parsed.into_entries().len(), 2);
 }

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

@@ -1,15 +1,17 @@
-use std::collections::VecDeque;
+use std::collections::{BTreeMap, VecDeque};
 use std::sync::atomic::{AtomicUsize, Ordering};
 use std::sync::{Arc, Mutex};
-use std::time::Duration;
+use std::time::{Duration, SystemTime};
 
 use async_trait::async_trait;
 use chrono::Utc;
 use gateway_core::engine::credential::{
-    AccountAvailability, CredentialCasOutcome, CredentialRevision, ProviderAccountStore,
+    AccountAvailability, CredentialCasOutcome, CredentialCasUpdate, CredentialRevision,
+    ProviderAccountId, ProviderAccountStore, ProviderAccountUpdate,
 };
 use gateway_core::provider_ports::{
-    ProviderLeaseAcquisition, ProviderLeasePort, ProviderLeaseRequest, ProviderStoreError,
+    ProviderCredentialState, ProviderCredentialStatePort, ProviderLeaseAcquisition,
+    ProviderLeasePort, ProviderLeaseRequest, ProviderStoreError,
 };
 use provider_xai::{
     GrokCredentialCatalogCache, GrokCredentialRecovery, GrokCredentialRecoveryOutcome,
@@ -141,6 +143,86 @@ async fn fixture(
     fixture_many([input], responses, lease_available).await
 }
 
+/// 真实累加连续失败计数的测试 double，用于断言退避的指数增长与成功清零。
+#[derive(Default)]
+struct CountingCredentialState {
+    counts: Mutex<BTreeMap<String, u32>>,
+    cleared: Mutex<Vec<String>>,
+}
+
+impl CountingCredentialState {
+    fn count(&self, account_id: &ProviderAccountId) -> u32 {
+        self.counts
+            .lock()
+            .expect("backoff counts lock")
+            .get(account_id.as_str())
+            .copied()
+            .unwrap_or(0)
+    }
+
+    fn was_cleared(&self, account_id: &ProviderAccountId) -> bool {
+        self.cleared
+            .lock()
+            .expect("cleared accounts lock")
+            .iter()
+            .any(|id| id == account_id.as_str())
+    }
+}
+
+impl ProviderCredentialStatePort for CountingCredentialState {
+    fn replace(
+        &self,
+        _state: ProviderCredentialState,
+    ) -> futures::future::BoxFuture<'_, Result<(), ProviderStoreError>> {
+        Box::pin(async { Ok(()) })
+    }
+
+    fn read<'a>(
+        &'a self,
+        _account_id: &'a ProviderAccountId,
+    ) -> futures::future::BoxFuture<'a, Result<Option<ProviderCredentialState>, ProviderStoreError>>
+    {
+        Box::pin(async { Ok(None) })
+    }
+
+    fn clear<'a>(
+        &'a self,
+        _account_id: &'a ProviderAccountId,
+    ) -> futures::future::BoxFuture<'a, Result<bool, ProviderStoreError>> {
+        Box::pin(async { Ok(false) })
+    }
+
+    fn record_refresh_backoff<'a>(
+        &'a self,
+        account_id: &'a ProviderAccountId,
+        _window: Duration,
+    ) -> futures::future::BoxFuture<'a, Result<u32, ProviderStoreError>> {
+        Box::pin(async move {
+            let mut counts = self.counts.lock().expect("backoff counts lock");
+            let entry = counts.entry(account_id.as_str().to_owned()).or_insert(0);
+            *entry += 1;
+            Ok(*entry)
+        })
+    }
+
+    fn clear_refresh_backoff<'a>(
+        &'a self,
+        account_id: &'a ProviderAccountId,
+    ) -> futures::future::BoxFuture<'a, Result<(), ProviderStoreError>> {
+        Box::pin(async move {
+            self.counts
+                .lock()
+                .expect("backoff counts lock")
+                .remove(account_id.as_str());
+            self.cleared
+                .lock()
+                .expect("cleared accounts lock")
+                .push(account_id.as_str().to_owned());
+            Ok(())
+        })
+    }
+}
+
 async fn fixture_many(
     inputs: impl IntoIterator<Item = provider_xai::CreateGrokCredential>,
     responses: impl IntoIterator<Item = Result<GrokRefreshTokens, GrokRefreshFailure>>,
@@ -150,6 +232,26 @@ async fn fixture_many(
     GrokCredentialRepository,
     Arc<QueueRefresher>,
     GrokCredentialRefreshService,
+) {
+    fixture_many_with_state(
+        inputs,
+        responses,
+        lease_available,
+        Arc::new(CountingCredentialState::default()),
+    )
+    .await
+}
+
+async fn fixture_many_with_state(
+    inputs: impl IntoIterator<Item = provider_xai::CreateGrokCredential>,
+    responses: impl IntoIterator<Item = Result<GrokRefreshTokens, GrokRefreshFailure>>,
+    lease_available: bool,
+    credential_state: Arc<dyn ProviderCredentialStatePort>,
+) -> (
+    Arc<MemoryProviderAccountStore>,
+    GrokCredentialRepository,
+    Arc<QueueRefresher>,
+    GrokCredentialRefreshService,
 ) {
     let store = MemoryProviderAccountStore::shared();
     let account_store: Arc<dyn ProviderAccountStore> = store.clone();
@@ -176,11 +278,40 @@ async fn fixture_many(
         catalog,
         leases,
         cooldowns,
+        credential_state,
         runtime_policy(),
     );
     (store, repository, refresher, service)
 }
 
+/// 用 store 的 CAS 把 next_refresh_at 复位到过去，使已退避到未来的账号再次到期。
+async fn force_due(store: &MemoryProviderAccountStore, id: &ProviderAccountId) {
+    let account = store.account(id).expect("account to reset");
+    let credential = store.credential(id).expect("credential to reset");
+    let update = CredentialCasUpdate::new(
+        id.clone(),
+        account.revision(),
+        ProviderAccountUpdate {
+            account_id: id.clone(),
+            name: account.name().to_owned(),
+            email: account.email().map(str::to_owned),
+            plan_type: account.plan_type().map(str::to_owned),
+        },
+        credential,
+        true,
+        account.access_token_expires_at(),
+        Some(SystemTime::now() - Duration::from_secs(1)),
+    )
+    .expect("valid due-time reset");
+    assert!(matches!(
+        store
+            .compare_and_swap_credential(update)
+            .await
+            .expect("reset next_refresh_at to the past"),
+        CredentialCasOutcome::Updated(_)
+    ));
+}
+
 #[tokio::test]
 async fn successful_refresh_rotates_plaintext_tokens_once() {
     let input = due_input("success");
@@ -423,6 +554,95 @@ async fn pre_send_transient_failure_applies_bounded_cooldown() {
     );
 }
 
+#[tokio::test]
+async fn refresh_backoff_grows_exponentially_across_attempts() {
+    let input = due_input("backoff-growth");
+    let id = input.account_id.clone();
+    let counting = Arc::new(CountingCredentialState::default());
+    let (store, _, _, service) = fixture_many_with_state(
+        [input],
+        [
+            Err(GrokRefreshFailure::Transient),
+            Err(GrokRefreshFailure::Transient),
+        ],
+        true,
+        counting.clone(),
+    )
+    .await;
+
+    let before_first = SystemTime::now();
+    let first = service.refresh_due().await.expect("first refresh cycle");
+    assert!(matches!(
+        first.as_slice(),
+        [GrokCredentialRefreshOutcome::Transient { .. }]
+    ));
+    let first_delay = store
+        .account(&id)
+        .expect("account after first defer")
+        .next_refresh_at()
+        .expect("first retry scheduled")
+        .duration_since(before_first)
+        .expect("first delay is in the future");
+    assert_eq!(counting.count(&id), 1);
+
+    force_due(&store, &id).await;
+
+    let before_second = SystemTime::now();
+    let second = service.refresh_due().await.expect("second refresh cycle");
+    assert!(matches!(
+        second.as_slice(),
+        [GrokCredentialRefreshOutcome::Transient { .. }]
+    ));
+    let second_delay = store
+        .account(&id)
+        .expect("account after second defer")
+        .next_refresh_at()
+        .expect("second retry scheduled")
+        .duration_since(before_second)
+        .expect("second delay is in the future");
+    assert_eq!(counting.count(&id), 2);
+
+    // base=5s、factor=3：第二次（attempt=2）应比第一次（attempt=1）显著更久。
+    assert!(
+        second_delay > first_delay * 2,
+        "second backoff {second_delay:?} should grow well beyond first {first_delay:?}"
+    );
+}
+
+#[tokio::test]
+async fn successful_refresh_clears_backoff_counter() {
+    let input = due_input("backoff-clear");
+    let id = input.account_id.clone();
+    let counting = Arc::new(CountingCredentialState::default());
+    let (store, _, _, service) = fixture_many_with_state(
+        [input],
+        [
+            Err(GrokRefreshFailure::Transient),
+            Ok(success_tokens(Some("cleared-refresh"))),
+        ],
+        true,
+        counting.clone(),
+    )
+    .await;
+
+    let first = service.refresh_due().await.expect("first refresh cycle");
+    assert!(matches!(
+        first.as_slice(),
+        [GrokCredentialRefreshOutcome::Transient { .. }]
+    ));
+    assert_eq!(counting.count(&id), 1);
+
+    force_due(&store, &id).await;
+
+    let second = service.refresh_due().await.expect("second refresh cycle");
+    assert!(matches!(
+        second.as_slice(),
+        [GrokCredentialRefreshOutcome::Refreshed { .. }]
+    ));
+    assert!(counting.was_cleared(&id));
+    assert_eq!(counting.count(&id), 0);
+}
+
 #[tokio::test]
 async fn invalid_refresh_lifetime_is_rejected_without_cas_write() {
     let input = due_input("bad-lifetime");

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

@@ -411,13 +411,16 @@ async fn interrupted_stream_feedback_records_runtime_cooldown_without_persisting
             .cooldown(session.account_id())
             .is_some_and(|cooldown| cooldown.until() > SystemTime::now())
     );
-    assert!(matches!(
-        fixture
-            .selector
-            .select(fixture.request(BTreeSet::new()))
-            .await,
-        Err(GrokSessionSelectorError::NoEligibleSession)
-    ));
+    let retry_after = match fixture
+        .selector
+        .select(fixture.request(BTreeSet::new()))
+        .await
+    {
+        Err(GrokSessionSelectorError::AccountCoolingDown { retry_after }) => retry_after,
+        other => panic!("expected AccountCoolingDown, got {other:?}"),
+    };
+    let retry_after = retry_after.expect("retry_after");
+    assert!(retry_after <= Duration::from_secs(30));
 }
 
 #[tokio::test]

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

@@ -325,20 +325,20 @@ impl GrokInferenceTransport for StubInferenceTransport {
         Box::pin(async move {
             match mode {
                 InferenceMode::Success => Ok(GrokInferenceResponse::new(
-                    Box::pin(stream::iter([Ok(SUCCESS_SSE.to_vec())])),
+                    Box::pin(stream::iter([Ok(bytes::Bytes::from_static(SUCCESS_SSE))])),
                     UpstreamHttpVersion::Http2,
                     200,
                     None,
                 )),
                 InferenceMode::SuccessWithMetrics(metrics) => Ok(GrokInferenceResponse::new(
-                    Box::pin(stream::iter([Ok(SUCCESS_SSE.to_vec())])),
+                    Box::pin(stream::iter([Ok(bytes::Bytes::from_static(SUCCESS_SSE))])),
                     UpstreamHttpVersion::Http2,
                     200,
                     None,
                 )
                 .with_transport_metrics(metrics)),
                 InferenceMode::SuccessBody(body) => Ok(GrokInferenceResponse::new(
-                    Box::pin(stream::iter([Ok(body)])),
+                    Box::pin(stream::iter([Ok(bytes::Bytes::from(body))])),
                     UpstreamHttpVersion::Http2,
                     200,
                     None,
@@ -1330,6 +1330,60 @@ async fn selector_capacity_failure_occurs_before_visible_upstream_send() {
     assert_eq!(transport.calls.load(Ordering::SeqCst), 0);
 }
 
+#[tokio::test]
+async fn selection_failures_carry_a_distinguishable_client_error_code() {
+    let cases = [
+        (
+            GrokSessionSelectorError::AccountCoolingDown {
+                retry_after: Some(Duration::from_secs(12)),
+            },
+            "account_cooling_down",
+        ),
+        (
+            GrokSessionSelectorError::CapacityUnavailable {
+                retry_after: Some(Duration::from_millis(20)),
+            },
+            "account_capacity_busy",
+        ),
+        (
+            GrokSessionSelectorError::NoEligibleSession,
+            "no_eligible_account",
+        ),
+        (
+            GrokSessionSelectorError::Unavailable,
+            "account_selector_unavailable",
+        ),
+    ];
+    for (failure, expected_code) in cases {
+        let transport = StubInferenceTransport::success();
+        let provider = provider(StubSelector::failing(failure), transport.clone()).await;
+        let error = match provider
+            .execute(
+                provider_request("xai"),
+                context(CancellationToken::new(), None),
+            )
+            .await
+        {
+            Ok(_) => panic!("selection must fail"),
+            Err(error) => error,
+        };
+        let detail = error
+            .client_visible_upstream_error()
+            .expect("client visible detail");
+        assert_eq!(error.kind(), ProviderErrorKind::Unavailable);
+        assert_eq!(detail.code(), Some(expected_code));
+        assert_eq!(detail.error_type(), Some("account_unavailable_error"));
+        assert_eq!(transport.calls.load(Ordering::SeqCst), 0);
+        if expected_code == "account_cooling_down" {
+            assert!(
+                detail.message().contains("retry in 13s"),
+                "{}",
+                detail.message()
+            );
+        }
+    }
+}
+
 #[tokio::test]
 async fn native_previous_response_is_rejected_before_selection() {
     let selector = StubSelector::success();
@@ -1827,7 +1881,7 @@ async fn provider_compiles_realtime_catalog_capabilities() {
     assert_eq!(presentation.display_name(), Some("Grok 4.5"));
     assert_eq!(
         presentation.supported_reasoning_efforts(),
-        ["low", "medium", "high"]
+        ["low", "medium", "high", "xhigh"]
     );
     assert_eq!(presentation.context_window_tokens(), Some(1_000_000));
     assert!(presentation.agent_tools());