chore: bump version to 3.0.7

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

README.md

@@ -12,7 +12,7 @@
 [![GHCR](https://img.shields.io/badge/GHCR-codex--proxy--rs-2496ED?logo=docker&logoColor=white&style=flat-square)](https://github.com/zyycn/codex-proxy-rs/pkgs/container/codex-proxy-rs)
 [![MIT](https://img.shields.io/badge/License-MIT-blue.svg?style=flat-square)](https://opensource.org/license/mit)
 
-[快速开始](#快速开始) · [客户端接入](#客户端接入) · [运维](#运维) · [部署文档](deploy/README.md) · [架构](docs/architecture.md)
+[快速开始](#快速开始) · [客户端接入](#客户端接入) · [运维](#运维) · [部署](deploy/README.md) · [接口](docs/api.md) · [架构](docs/architecture.md)
 
 </div>
 
@@ -21,14 +21,15 @@
 
 ## 能力
 
-| 领域     | 实现                                                         |
-| -------- | ------------------------------------------------------------ |
-| 协议     | OpenAI Responses JSON、SSE、WebSocket 与模型目录              |
-| Provider | OpenAI OAuth、xAI OAuth，多账号隔离与自动刷新                  |
-| 路由     | Client Key 平台绑定、Provider instance、模型映射与安全 fallback |
-| 延续     | OpenAI native/replay continuation；xAI 客户端完整历史         |
-| 管理     | Client Key、账号、模型目录、设置、观测、系统与 OAuth          |
-| 计量     | 模型请求 Token、费用、延迟、账号与 Provider 归因               |
+| 领域     | 实现                                                                        |
+| -------- | --------------------------------------------------------------------------- |
+| 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`         |
+| 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                               |
 
 ## 快速开始
 
@@ -47,15 +48,14 @@ sudo chown "$(id -u):10001" deploy/config.yaml
 chmod 0640 deploy/config.yaml
 ```
 
-分别生成三个密码：
+分别生成 PostgreSQL 与 Redis 密码：
 
 ```bash
 openssl rand -hex 24
 openssl rand -hex 24
-openssl rand -hex 24
 ```
 
-把结果写入 `deploy/config.yaml`：
+把两个结果写入 `deploy/config.yaml`，并单独设置管理员初始密码：
 
 | 配置                       | 约束                         |
 | -------------------------- | ---------------------------- |
@@ -64,7 +64,7 @@ openssl rand -hex 24
 | `admin.default_password`   | 至少 12 位，不能包含 `$`      |
 
 PostgreSQL、Redis 容器和应用通过同一份 `config.yaml` 使用这两个基础设施密码，不需要额外
-导出环境变量。
+导出环境变量。管理员密码只在首次创建管理员时生效。
 
 Linux 需要允许容器组写入运行目录：
 
@@ -88,9 +88,10 @@ curl -i http://127.0.0.1:8080/healthz
 ### 3. 初始化
 
 1. 使用 `admin@cpr.local` 与初始密码登录。
-2. 创建 Provider instance，并通过 OAuth 或 JSON 导入 OpenAI、xAI 账号。
-3. 按平台配置客户端模型到上游模型的映射；未配置的模型名原样透传。
-4. 创建 `sk_...` 客户端 Key，绑定平台并设置模型范围、速率与并发限制。
+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。
 
 > [!IMPORTANT]
 > xAI 使用 OAuth session，不支持把 xAI API Key 作为上游 credential。
@@ -125,18 +126,22 @@ curl http://127.0.0.1:8080/v1/responses \
 <details>
 <summary>API 路由</summary>
 
-| 路由                        | 用途                |
-| --------------------------- | ------------------- |
-| `POST /v1/responses`        | JSON 或 SSE 透明代理 |
-| `GET /v1/responses`         | WebSocket 透明代理   |
-| `POST /v1/responses/review` | OpenAI review 请求   |
-| `GET /v1/models`            | 启用的公开模型列表   |
-| `GET /v1/models/{model_id}` | 公开模型详情         |
+| 路由                               | 用途                           |
+| ---------------------------------- | ------------------------------ |
+| `POST /v1/responses`               | JSON 响应或 SSE Responses 流    |
+| `GET /v1/responses`                | Responses WebSocket 升级        |
+| `POST /v1/responses/review`        | review 子代理请求               |
+| `GET /v1/models`                   | 当前 Client Key 可用的模型列表   |
+| `GET /v1/models/catalog`           | Codex 客户端展示用模型目录      |
+| `GET /v1/models/{model_id}/info`   | Codex 客户端展示用模型详情      |
+| `GET /v1/models/{model_id}`        | OpenAI 兼容模型详情             |
 
 所有 `/v1/*` 路由都需要客户端 API Key。
 
 </details>
 
+完整的客户端与管理端路由、鉴权和 mutation 规则见 [接口文档](docs/api.md)。
+
 ## 运维
 
 ```bash
@@ -153,14 +158,27 @@ docker compose -f deploy/compose.yaml up -d
 ```
 
 > [!IMPORTANT]
-> `.runtime/` 保存 PostgreSQL、Redis、自更新状态和日志。删除该目录会永久清除运行状态。
+> `.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
+> 运行状态。
 
 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。
 
+## 维护与发布
+
+发布仓库版本必须从干净、已同步上游的发布分支运行：
+
+```bash
+release/publish <version>
+```
+
+脚本会更新 `release/version.yaml`、创建约定提交与带注释的 `v<version>` tag，并原子推送分支和 tag。
+不要手工修改版本后单独打 tag。
+
 ## 文档
 
 - [部署](deploy/README.md)
+- [客户端与管理端接口](docs/api.md)
 - [架构与数据边界](docs/architecture.md)
 - [Release](https://github.com/zyycn/codex-proxy-rs/releases)
 - [容器镜像](https://github.com/zyycn/codex-proxy-rs/pkgs/container/codex-proxy-rs)

backend/Cargo.lock

@@ -1152,6 +1152,7 @@ dependencies = [
  "rand_core 0.6.4",
  "secrecy",
  "serde",
+ "serde_json",
  "subtle",
  "thiserror",
  "tokio",

backend/crates/gateway-admin/Cargo.toml

@@ -27,6 +27,7 @@ thiserror = "2.0.18"
 uuid = { version = "1.23.3", features = ["v7"] }
 
 [dev-dependencies]
+serde_json = { version = "1.0.150", features = ["preserve_order"] }
 tokio = { version = "1.52.3", default-features = false, features = ["macros", "rt"] }
 
 [lints]

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

@@ -29,7 +29,6 @@ pub enum AccountStatus {
     QuotaExhausted,
     Disabled,
     Banned,
-    Attention,
 }
 
 /// 账号列表排序字段。
@@ -184,7 +183,7 @@ pub struct AccountSummary {
     pub total: u64,
     pub active: u64,
     pub quota_exhausted: u64,
-    pub attention: u64,
+    pub unavailable: u64,
 }
 
 /// 账号启停写入命令。
@@ -222,6 +221,8 @@ pub enum AccountConnectionTestEvent {
     },
     Failed {
         message: String,
+        provider_error_code: Option<String>,
+        provider_error_type: Option<String>,
         account_status: AccountStatus,
     },
 }

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

@@ -718,7 +718,8 @@ pub struct OpsError {
     pub upstream_model_id: Option<String>,
     pub upstream_transport: Option<String>,
     pub failure_kind: String,
-    pub status_code: Option<u16>,
+    pub client_status_code: Option<u16>,
+    pub upstream_status_code: Option<u16>,
     pub provider_error_code: Option<String>,
     pub client_response_id: Option<String>,
     pub upstream_request_id: Option<String>,

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

@@ -10,7 +10,6 @@ use gateway_core::{
         credential::ProviderAccountId,
         probe::{AccountProbe, AccountProbeRequest},
     },
-    error::GatewayErrorKind,
     routing::{ProviderKind, UpstreamModelId, snapshot::SnapshotControl},
 };
 
@@ -454,12 +453,9 @@ impl AccountsService for DefaultAccountsService {
                     }))
                     .collect(),
                 Err(error) => vec![AccountConnectionTestEvent::Failed {
-                    message: match error.kind() {
-                        GatewayErrorKind::InvalidRequest
-                        | GatewayErrorKind::Unsupported
-                        | GatewayErrorKind::ModelNotFound => error.safe_message().to_owned(),
-                        _ => "Provider connection test failed".to_owned(),
-                    },
+                    message: error.client_message().to_owned(),
+                    provider_error_code: error.client_error_code().map(ToOwned::to_owned),
+                    provider_error_type: error.client_error_type().map(ToOwned::to_owned),
                     account_status: status,
                 }],
             }
@@ -485,8 +481,7 @@ fn account_status(account: &AccountRecord, now: chrono::DateTime<Utc>) -> Accoun
             {
                 AccountStatus::Expired
             }
-            crate::model::accounts::AccountAvailability::Ready => AccountStatus::Active,
-            _ => AccountStatus::Attention,
+            _ => AccountStatus::Active,
         }
     }
 }

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

@@ -2,13 +2,14 @@ use std::sync::{Arc, Mutex};
 
 use async_trait::async_trait;
 use chrono::{TimeDelta, Utc};
-use futures::future::BoxFuture;
+use futures::{StreamExt as _, future::BoxFuture};
 use gateway_core::{
     engine::{
         credential::{OpaqueProviderData, ProviderAccountId},
         probe::{AccountProbe, AccountProbeRequest, AccountProbeResult},
     },
-    error::GatewayError,
+    error::{ClientVisibleUpstreamError, GatewayError, GatewayErrorKind},
+    operation::{GenerateRequest, Operation, ProtocolPayload},
     routing::ProviderKind,
 };
 
@@ -17,9 +18,9 @@ use gateway_admin::{
     model::{
         MutationContext, Revision,
         accounts::{
-            AccountAvailability, AccountListQuery, AccountPage, AccountRecord, AccountSummary,
-            AccountUsage, AccountUsageWindowQuery, AccountUsageWindowResult, DeleteAccounts,
-            SetAccountEnabled,
+            AccountAvailability, AccountConnectionTestEvent, AccountListQuery, AccountPage,
+            AccountRecord, AccountSummary, AccountUsage, AccountUsageWindowQuery,
+            AccountUsageWindowResult, DeleteAccounts, SetAccountEnabled,
         },
         observability::TimeRange,
         provider_credentials::{
@@ -48,6 +49,7 @@ use gateway_admin::{
         },
     },
 };
+use serde_json::{Map, json};
 
 pub(super) type EventLog = Arc<Mutex<Vec<&'static str>>>;
 
@@ -159,10 +161,22 @@ impl ProviderAdmin for FakeProviderAdmin {
 
     fn connection_test_operation(
         &self,
-        _: &gateway_core::routing::UpstreamModelId,
-        _: &str,
+        model: &gateway_core::routing::UpstreamModelId,
+        input: &str,
     ) -> Result<gateway_core::operation::Operation, ProviderAdminError> {
-        Err(ProviderAdminError::new(ProviderAdminErrorKind::Unsupported))
+        let payload = ProtocolPayload::json_object(
+            "openai",
+            Map::from_iter([
+                ("model".to_owned(), json!(model.as_str())),
+                ("input".to_owned(), json!(input)),
+                ("stream".to_owned(), json!(true)),
+                ("store".to_owned(), json!(false)),
+            ]),
+        )
+        .map_err(|_| ProviderAdminError::new(ProviderAdminErrorKind::Invalid))?;
+        Ok(Operation::Generate(GenerateRequest::from_protocol_payload(
+            payload,
+        )))
     }
 
     fn dashboard_wire_profile(
@@ -395,7 +409,7 @@ impl AccountStore for FakeAccountStore {
                 total: 1,
                 active: 1,
                 quota_exhausted: 0,
-                attention: 0,
+                unavailable: 0,
             },
         })
     }
@@ -686,6 +700,37 @@ fn provider_registry_should_reject_duplicate_kind() {
     ));
 }
 
+#[tokio::test]
+async fn connection_test_should_preserve_safe_provider_error_details() {
+    let provider = FakeProviderAdmin::new("xai", events());
+    let store = FakeAccountStore::new("xai", events());
+    let services =
+        accounts_service_with_probe(provider, store, Arc::new(FailingAccountProbe)).await;
+
+    let events = services
+        .accounts()
+        .test_connection(
+            ProviderAccountId::new("acct_test").expect("account ID"),
+            gateway_core::routing::UpstreamModelId::new("grok-4.5").expect("model"),
+        )
+        .await
+        .expect("connection test stream")
+        .collect::<Vec<_>>()
+        .await;
+
+    assert!(matches!(
+        events.last(),
+        Some(AccountConnectionTestEvent::Failed {
+            message,
+            provider_error_code: Some(code),
+            provider_error_type: Some(error_type),
+            ..
+        }) if message == "included usage exhausted"
+            && code == "usage_exhausted"
+            && error_type == "invalid_request_error"
+    ));
+}
+
 #[tokio::test]
 async fn accounts_export_should_pass_store_loaded_timestamps_and_material_to_provider() {
     let provider = FakeProviderAdmin::new("openai", events());
@@ -1005,12 +1050,20 @@ fn rotation_result(command: CredentialRotationCommit) -> CredentialMutationResul
 async fn accounts_service(
     provider: Arc<FakeProviderAdmin>,
     store: Arc<FakeAccountStore>,
+) -> AdminServices {
+    accounts_service_with_probe(provider, store, Arc::new(SuccessfulAccountProbe)).await
+}
+
+async fn accounts_service_with_probe(
+    provider: Arc<FakeProviderAdmin>,
+    store: Arc<FakeAccountStore>,
+    probe: Arc<dyn AccountProbe>,
 ) -> AdminServices {
     super::AdminHarness::new()
         .accounts(store)
         .settings(Arc::new(StaticSettingsStore))
         .provider(provider)
-        .probe(Arc::new(SuccessfulAccountProbe))
+        .probe(probe)
         .build()
         .await
 }
@@ -1030,6 +1083,29 @@ impl AccountProbe for SuccessfulAccountProbe {
     }
 }
 
+struct FailingAccountProbe;
+
+impl AccountProbe for FailingAccountProbe {
+    fn probe(
+        &self,
+        _: AccountProbeRequest,
+    ) -> BoxFuture<'_, Result<AccountProbeResult, GatewayError>> {
+        Box::pin(async {
+            let detail = ClientVisibleUpstreamError::new(
+                "included usage exhausted",
+                Some("usage_exhausted".to_owned()),
+                Some("invalid_request_error".to_owned()),
+            )
+            .expect("safe provider error");
+            Err(GatewayError::new(
+                GatewayErrorKind::RateLimited,
+                "upstream capacity is temporarily unavailable",
+            )
+            .with_client_visible_upstream_error(detail))
+        })
+    }
+}
+
 fn store_unavailable() -> AdminStoreError {
     AdminStoreError::new(
         AdminStoreErrorKind::Unavailable,

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

@@ -174,7 +174,7 @@ pub struct AccountSummaryView {
     pub total: u64,
     pub active: u64,
     pub quota_exhausted: u64,
-    pub attention: u64,
+    pub unavailable: u64,
 }
 
 /// 一条安全账号视图。
@@ -543,10 +543,14 @@ impl From<DomainConnectionTestEvent> for AccountConnectionTestEvent {
             }),
             DomainConnectionTestEvent::Failed {
                 message,
+                provider_error_code,
+                provider_error_type,
                 account_status,
             } => serde_json::json!({
                 "type": "error",
                 "error": message,
+                "providerErrorCode": provider_error_code,
+                "providerErrorType": provider_error_type,
                 "accountStatus": account_status_name(account_status)
             }),
         };
@@ -1401,7 +1405,7 @@ fn account_page_data(
             total: result.summary.total,
             active: result.summary.active,
             quota_exhausted: result.summary.quota_exhausted,
-            attention: result.summary.attention,
+            unavailable: result.summary.unavailable,
         },
     }
 }
@@ -1798,7 +1802,6 @@ fn account_status_name(status: DomainAccountStatus) -> &'static str {
         DomainAccountStatus::QuotaExhausted => "quota_exhausted",
         DomainAccountStatus::Disabled => "disabled",
         DomainAccountStatus::Banned => "banned",
-        DomainAccountStatus::Attention => "attention",
     }
 }
 

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

@@ -801,8 +801,8 @@ pub struct DashboardCardsView {
 pub struct DashboardCredentialsCardView {
     pub total: String,
     pub total_value: u64,
-    pub enabled: String,
-    pub enabled_value: u64,
+    pub available: String,
+    pub available_value: u64,
     pub unavailable: String,
     pub unavailable_value: u64,
 }
@@ -1235,7 +1235,6 @@ pub struct OpsErrorView {
     pub account_id: Option<String>,
     pub route: String,
     pub model: Option<String>,
-    pub status_code: Option<i64>,
     pub client_status_code: Option<i64>,
     pub upstream_status_code: Option<i64>,
     pub transport: Option<String>,
@@ -1954,7 +1953,6 @@ fn usage_page_view(
 }
 
 fn ops_error_view(error: domain::OpsError) -> OpsErrorView {
-    let status = error.status_code.map(i64::from);
     OpsErrorView {
         id: error.event_id,
         request_id: error.request_id,
@@ -1964,9 +1962,8 @@ fn ops_error_view(error: domain::OpsError) -> OpsErrorView {
         account_id: error.provider_account_ref,
         route: error.operation,
         model: error.upstream_model_id,
-        status_code: status,
-        client_status_code: None,
-        upstream_status_code: None,
+        client_status_code: error.client_status_code.map(i64::from),
+        upstream_status_code: error.upstream_status_code.map(i64::from),
         transport: error.upstream_transport,
         attempt_index: error.attempt_index,
         failure_class: error.failure_kind,
@@ -2290,8 +2287,8 @@ fn dashboard_view(result: domain::DashboardResult, kind: TrendKind) -> Dashboard
             credentials: DashboardCredentialsCardView {
                 total: format_compact_number(provider_accounts.total),
                 total_value: provider_accounts.total,
-                enabled: format_compact_number(provider_accounts.enabled),
-                enabled_value: provider_accounts.enabled,
+                available: format_compact_number(provider_accounts.active),
+                available_value: provider_accounts.active,
                 unavailable: format_compact_number(provider_accounts.unavailable),
                 unavailable_value: provider_accounts.unavailable,
             },

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

@@ -201,11 +201,13 @@ struct RestartAcceptedView {
 #[derive(Debug, Serialize)]
 #[serde(rename_all = "camelCase")]
 struct SystemUpdateEventView {
+    id: String,
     operation_id: Option<String>,
     level: &'static str,
     step: Option<String>,
     message: String,
     terminal: bool,
+    progress_percent: Option<u8>,
     at: String,
 }
 
@@ -403,11 +405,13 @@ where
 impl From<SystemUpdateEvent> for SystemUpdateEventView {
     fn from(event: SystemUpdateEvent) -> Self {
         Self {
+            id: event.id,
             operation_id: event.operation_id,
             level: update_event_level_name(event.level),
             step: event.step,
             message: event.message,
             terminal: event.terminal,
+            progress_percent: event.progress_percent,
             at: event.occurred_at.to_rfc3339(),
         }
     }
@@ -416,11 +420,13 @@ impl From<SystemUpdateEvent> for SystemUpdateEventView {
 impl SystemUpdateEventView {
     fn into_json(self) -> serde_json::Value {
         serde_json::json!({
+            "id": self.id,
             "operationId": self.operation_id,
             "level": self.level,
             "step": self.step,
             "message": self.message,
             "terminal": self.terminal,
+            "progressPercent": self.progress_percent,
             "at": self.at,
         })
     }

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

@@ -6,7 +6,7 @@ use axum::{
     http::{HeaderMap, StatusCode},
     response::{IntoResponse, Response},
 };
-use gateway_core::routing::PublicModelId;
+use gateway_core::routing::{PublicModelId, PublicModelProfile};
 use serde_json::{Value, json};
 
 use crate::ApiState;
@@ -17,6 +17,7 @@ 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 {
@@ -30,6 +31,16 @@ pub(crate) async fn models(State(state): State<ApiState>, headers: HeaderMap) ->
         .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,
@@ -134,3 +145,75 @@ fn catalog_model_json(id: &str) -> Value {
         "source": "gateway",
     })
 }
+
+fn codex_model_json(profile: &PublicModelProfile, index: usize) -> Value {
+    let id = profile.model().as_str();
+    let presentation = profile.presentation();
+    let reasoning_levels = presentation
+        .supported_reasoning_efforts()
+        .iter()
+        .map(|effort| {
+            json!({
+                "effort": effort,
+                "description": reasoning_effort_description(effort),
+            })
+        })
+        .collect::<Vec<_>>();
+    let reasoning_supported = presentation
+        .supported_reasoning_efforts()
+        .iter()
+        .any(|effort| effort != "none");
+    let context_window = presentation.context_window_tokens();
+    let input_modalities = if presentation.image_input() {
+        vec!["text", "image"]
+    } else {
+        vec!["text"]
+    };
+
+    json!({
+        "slug": id,
+        "display_name": presentation.display_name().unwrap_or(id),
+        "description": presentation.description(),
+        "default_reasoning_level": presentation.default_reasoning_effort(),
+        "supported_reasoning_levels": reasoning_levels,
+        "shell_type": "shell_command",
+        "visibility": if presentation.hidden() { "hide" } else { "list" },
+        "supported_in_api": true,
+        "priority": index.saturating_add(1),
+        "additional_speed_tiers": [],
+        "service_tiers": [],
+        "availability_nux": Value::Null,
+        "upgrade": Value::Null,
+        "base_instructions": CODEX_BASE_INSTRUCTIONS,
+        "model_messages": Value::Null,
+        "include_skills_usage_instructions": false,
+        "supports_reasoning_summary_parameter": reasoning_supported,
+        "default_reasoning_summary": "auto",
+        "support_verbosity": presentation.verbosity(),
+        "default_verbosity": Value::Null,
+        "apply_patch_tool_type": presentation.agent_tools().then_some("freeform"),
+        "web_search_tool_type": "text",
+        "truncation_policy": { "mode": "tokens", "limit": 10_000 },
+        "supports_parallel_tool_calls": presentation.parallel_tool_calls(),
+        "supports_image_detail_original": presentation.image_detail_original(),
+        "context_window": context_window,
+        "max_context_window": context_window,
+        "effective_context_window_percent": 95,
+        "experimental_supported_tools": [],
+        "input_modalities": input_modalities,
+        "supports_search_tool": presentation.search_tool(),
+        "use_responses_lite": false,
+    })
+}
+
+fn reasoning_effort_description(effort: &str) -> &'static str {
+    match effort {
+        "none" => "No reasoning",
+        "low" => "Fast responses with lighter reasoning",
+        "medium" => "Balances speed and reasoning depth for everyday tasks",
+        "high" => "Greater reasoning depth for complex problems",
+        "xhigh" => "Extra high reasoning depth for complex problems",
+        "max" => "Maximum reasoning depth for the hardest problems",
+        _ => "Provider-supported reasoning level",
+    }
+}

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

@@ -9,7 +9,7 @@ use gateway_core::engine::execution::{
 };
 use gateway_core::error::{GatewayError, GatewayErrorKind};
 use gateway_core::lifecycle::{ConnectionDraining, ConnectionGuard, ConnectionLifecycle};
-use gateway_core::routing::PublicModelId;
+use gateway_core::routing::{PublicModelId, PublicModelProfile};
 use uuid::Uuid;
 
 use super::auth::ClientApiKeyAuthError;
@@ -51,6 +51,13 @@ impl OpenAiService {
             .collect()
     }
 
+    pub(crate) fn public_model_profiles(
+        &self,
+        client: &AuthenticatedClient,
+    ) -> Vec<PublicModelProfile> {
+        self.execution.public_model_profiles(client)
+    }
+
     pub(crate) fn contains_public_model(
         &self,
         client: &AuthenticatedClient,

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

@@ -276,6 +276,8 @@ mod actions {
             },
             DomainConnectionTestEvent::Failed {
                 message: "upstream unavailable".to_owned(),
+                provider_error_code: Some("usage_exhausted".to_owned()),
+                provider_error_type: Some("invalid_request_error".to_owned()),
                 account_status: AccountStatus::Active,
             },
         ]
@@ -302,7 +304,13 @@ mod actions {
                 }),
                 json!({ "type": "content", "text": "OK" }),
                 json!({ "type": "test_complete", "success": true, "accountStatus": "active" }),
-                json!({ "type": "error", "error": "upstream unavailable", "accountStatus": "active" }),
+                json!({
+                    "type": "error",
+                    "error": "upstream unavailable",
+                    "providerErrorCode": "usage_exhausted",
+                    "providerErrorType": "invalid_request_error",
+                    "accountStatus": "active"
+                }),
             ]
         );
     }

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

@@ -84,6 +84,10 @@ pub(super) struct AdminTestFixture {
 
 impl AdminTestFixture {
     pub async fn new() -> Self {
+        Self::with_system(Arc::new(UnusedSystem)).await
+    }
+
+    pub async fn with_system(system: Arc<dyn SystemOperations>) -> Self {
         let api_key = Arc::new(Mutex::new(None));
         let auth = Arc::new(MemoryAuthStore::new(api_key.clone()));
         let settings = Arc::new(MemorySettingsStore::new(api_key));
@@ -113,7 +117,7 @@ impl AdminTestFixture {
             providers,
             Arc::new(NoopSnapshot),
             Arc::new(NoopProbe),
-            Arc::new(UnusedSystem),
+            system,
         )
         .await
         .expect("initialize test admin services");

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

@@ -1,4 +1,24 @@
-use gateway_api::admin::system::{UpdateDetailQuery, UpdateRequest};
+use std::sync::Arc;
+
+use async_trait::async_trait;
+use axum::{
+    Router,
+    body::{Body, to_bytes},
+    http::{Request, header},
+};
+use chrono::Utc;
+use futures::stream;
+use gateway_admin::{
+    model::system::{
+        SystemOperationAccepted, SystemUpdateDetail, SystemUpdateEvent, SystemUpdateEventLevel,
+        SystemUpdateStatus, SystemVersion,
+    },
+    ports::system::{SystemOperationError, SystemOperations, SystemUpdateEventStream},
+};
+use gateway_api::admin::system::{self, UpdateDetailQuery, UpdateRequest};
+use tower::ServiceExt as _;
+
+use super::{AdminTestFixture, AdminTestState, unavailable_system};
 
 #[test]
 fn update_detail_query_should_default_to_no_refresh_and_reject_unknown_fields() {
@@ -16,3 +36,84 @@ fn update_request_should_preserve_target_version_for_domain_validation() {
         serde_json::from_str::<UpdateRequest>(r#"{"targetVersion":"v0.2.0","extra":1}"#).is_err()
     );
 }
+
+#[tokio::test]
+async fn update_event_stream_should_preserve_event_id_and_download_progress_percent() {
+    let fixture = AdminTestFixture::with_system(Arc::new(ProgressSystem)).await;
+    fixture.auth.insert_session("valid-session");
+    let response = app(fixture.state())
+        .oneshot(
+            Request::builder()
+                .uri("/api/admin/system/update/events")
+                .header(header::COOKIE, "cpr_admin_session=valid-session")
+                .header("x-request-id", "req_system_update_events")
+                .body(Body::empty())
+                .expect("update event request"),
+        )
+        .await
+        .expect("update event response");
+    let body = to_bytes(response.into_body(), 1024 * 1024)
+        .await
+        .expect("update event body");
+    let body = String::from_utf8(body.to_vec()).expect("UTF-8 event stream");
+
+    assert!(
+        body.contains(r#""id":"update-progress-1""#),
+        "unexpected update event stream: {body}"
+    );
+    assert!(
+        body.contains(r#""progressPercent":40"#),
+        "unexpected update event stream: {body}"
+    );
+}
+
+fn app(state: AdminTestState) -> Router {
+    system::router::<AdminTestState>().with_state(state)
+}
+
+struct ProgressSystem;
+
+#[async_trait]
+impl SystemOperations for ProgressSystem {
+    async fn version(&self) -> Result<SystemVersion, SystemOperationError> {
+        Err(unavailable_system())
+    }
+
+    async fn update_detail(&self, _: bool) -> Result<SystemUpdateDetail, SystemOperationError> {
+        Err(unavailable_system())
+    }
+
+    fn update_events(&self) -> SystemUpdateEventStream {
+        Box::pin(stream::once(async {
+            SystemUpdateEvent {
+                id: "update-progress-1".to_owned(),
+                operation_id: Some("update-1".to_owned()),
+                level: SystemUpdateEventLevel::Info,
+                step: Some("download".to_owned()),
+                message: "已下载 4.0 MiB / 10.0 MiB (40%)".to_owned(),
+                terminal: true,
+                progress_percent: Some(40),
+                occurred_at: Utc::now(),
+            }
+        }))
+    }
+
+    async fn perform_update(
+        &self,
+        _: Option<String>,
+    ) -> Result<SystemOperationAccepted, SystemOperationError> {
+        Err(unavailable_system())
+    }
+
+    async fn update_status(&self) -> Result<SystemUpdateStatus, SystemOperationError> {
+        Err(unavailable_system())
+    }
+
+    async fn rollback(&self) -> Result<SystemOperationAccepted, SystemOperationError> {
+        Err(unavailable_system())
+    }
+
+    async fn restart(&self) -> Result<SystemOperationAccepted, SystemOperationError> {
+        Err(unavailable_system())
+    }
+}

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

@@ -10,19 +10,28 @@ use gateway_core::engine::execution::{
     StartedExecution,
 };
 use gateway_core::error::{GatewayError, GatewayErrorKind};
-use gateway_core::routing::PublicModelId;
+use gateway_core::routing::{ModelPresentation, PublicModelId, PublicModelProfile};
 use tower::ServiceExt;
 
 use super::{api_router, authenticated_client};
 
 struct ModelsExecution {
     client: AuthenticatedClient,
+    profiles: bool,
 }
 
 impl ModelsExecution {
     fn new() -> Arc<Self> {
         Arc::new(Self {
             client: authenticated_client("sk_models_test"),
+            profiles: false,
+        })
+    }
+
+    fn with_profiles() -> Arc<Self> {
+        Arc::new(Self {
+            client: authenticated_client("sk_models_test"),
+            profiles: true,
         })
     }
 }
@@ -46,6 +55,26 @@ impl ExecutionService for ModelsExecution {
             .collect()
     }
 
+    fn public_model_profiles(&self, _: &AuthenticatedClient) -> Vec<PublicModelProfile> {
+        if !self.profiles {
+            return Vec::new();
+        }
+        vec![PublicModelProfile::new(
+            PublicModelId::new("grok-4.5").expect("model"),
+            ModelPresentation::new(
+                Some("Grok 4.5".to_owned()),
+                Some("xAI Grok 4.5 frontier model.".to_owned()),
+            )
+            .with_reasoning(
+                Some("medium".to_owned()),
+                ["low", "medium", "high"].map(str::to_owned).to_vec(),
+            )
+            .with_context_window_tokens(Some(500_000))
+            .with_image_input(true)
+            .with_agent_tools(true, true),
+        )]
+    }
+
     fn contains_public_model(&self, _: &AuthenticatedClient, model: &PublicModelId) -> bool {
         matches!(model.as_str(), "model-a" | "model-b")
     }
@@ -91,6 +120,26 @@ async fn models_should_encode_the_service_visible_catalog() {
     );
 }
 
+#[tokio::test]
+async fn models_should_encode_provider_profiles_for_current_codex_clients() {
+    let response = api_router(ModelsExecution::with_profiles())
+        .await
+        .oneshot(authorized_request("/v1/models?client_version=0.145.0"))
+        .await
+        .expect("list Codex models response");
+    let body = to_bytes(response.into_body(), usize::MAX)
+        .await
+        .expect("read Codex models body");
+    let value: serde_json::Value = serde_json::from_slice(&body).expect("models JSON");
+    let model = &value["models"][0];
+
+    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!([]));
+}
+
 #[tokio::test]
 async fn model_detail_should_keep_the_official_path_id_contract() {
     let response = api_router(ModelsExecution::new())

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

@@ -1,6 +1,6 @@
 //! 唯一的账号重试、发送与下游 commit barrier owner。
 
-use std::collections::{BTreeSet, VecDeque};
+use std::collections::BTreeSet;
 use std::num::NonZeroU32;
 use std::sync::Arc;
 use std::time::{Duration, Instant, SystemTime};
@@ -67,12 +67,6 @@ where
         let continuation_attempt =
             initial_continuation_attempt(&operation, &plan, continuation.as_ref());
         let image_generation_requested = operation.image_generation_requested();
-        let commit_policy = if continuation_attempt == ContinuationAttempt::None {
-            StreamCommitPolicy::FirstCanonicalEvent
-        } else {
-            StreamCommitPolicy::UntilOutputOrTerminal
-        };
-
         let mut session = ResponseExecutionSession {
             engine: Arc::clone(&self.engine),
             request_id,
@@ -86,7 +80,6 @@ where
             required_account,
             continuation,
             continuation_attempt,
-            commit_policy,
             account_state_owner,
             cancellation,
             attempts: 0,
@@ -106,7 +99,6 @@ where
             upstream_response_id: None,
             last_account_exhaustion: None,
             provider_attempt_outcomes: Vec::new(),
-            before_commit: VecDeque::new(),
         };
 
         if session.cancellation.is_cancelled() {
@@ -141,12 +133,6 @@ struct FailureFinalization {
     provider_response_id: Option<String>,
 }
 
-#[derive(Clone, Copy)]
-enum StreamCommitPolicy {
-    FirstCanonicalEvent,
-    UntilOutputOrTerminal,
-}
-
 /// API 可逐事件消费的 Core 执行会话。
 ///
 /// API 只能提交下游 delivery 边界；账号重试、断流终结与
@@ -164,7 +150,6 @@ pub struct ResponseExecutionSession<S: ?Sized> {
     required_account: Option<crate::engine::credential::ProviderAccountId>,
     continuation: Option<ContinuationBinding>,
     continuation_attempt: ContinuationAttempt,
-    commit_policy: StreamCommitPolicy,
     account_state_owner: Option<ProviderAccountStateOwner>,
     cancellation: CancellationToken,
     attempts: u32,
@@ -184,7 +169,6 @@ pub struct ResponseExecutionSession<S: ?Sized> {
     upstream_response_id: Option<String>,
     last_account_exhaustion: Option<ProviderError>,
     provider_attempt_outcomes: Vec<ProviderAttemptOutcome>,
-    before_commit: VecDeque<ProviderEvent>,
 }
 
 impl<S: ?Sized> ResponseExecutionSession<S>
@@ -203,15 +187,6 @@ where
         if self.finalized {
             return Ok(None);
         }
-        if self.downstream_committed_at.is_some()
-            && let Some(event) = self.before_commit.pop_front()
-        {
-            return Ok(Some(CoordinatedEvent::single(
-                event,
-                CommitRequirement::AlreadyCommitted,
-            )));
-        }
-
         loop {
             match self.pull().await? {
                 PullOutcome::Event(event) => {
@@ -221,25 +196,13 @@ where
                             CommitRequirement::AlreadyCommitted,
                         )));
                     }
-                    let commit_significant = match self.commit_policy {
-                        StreamCommitPolicy::FirstCanonicalEvent => event.has_canonical_facts(),
-                        StreamCommitPolicy::UntilOutputOrTerminal => event.is_commit_significant(),
-                    };
-                    self.before_commit.push_back(event);
-                    if !commit_significant {
-                        continue;
-                    }
                     self.delivery_pending = true;
-                    let events = self.before_commit.drain(..).collect();
-                    return CoordinatedEvent::try_batch(
-                        events,
+                    return Ok(Some(CoordinatedEvent::single(
+                        event,
                         CommitRequirement::CommitBeforeDelivery,
-                    )
-                    .map(Some);
-                }
-                PullOutcome::AttemptDiscarded => {
-                    self.before_commit.clear();
+                    )));
                 }
+                PullOutcome::AttemptDiscarded => {}
                 PullOutcome::End => {
                     if self.downstream_committed_at.is_some() {
                         self.finish_success().await?;
@@ -953,7 +916,6 @@ where
         self.timings.first_text_ms = None;
         self.timings.first_token_ms = None;
         self.upstream_complete = false;
-        self.before_commit.clear();
     }
 
     async fn finish_success(&mut self) -> Result<(), EngineError> {

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

@@ -29,7 +29,9 @@ use crate::event::{GatewayEvent, ProviderEvent, ProviderResponseHeader};
 use crate::operation::{Operation, ProviderSessionState};
 use crate::policy::{ClientApiKeyId, ClientPolicy};
 use crate::routing::snapshot::RuntimeSnapshotHandle;
-use crate::routing::{ProviderKind, PublicModelId, RoutingContext, RuntimeSnapshot};
+use crate::routing::{
+    ProviderKind, PublicModelId, PublicModelProfile, RoutingContext, RuntimeSnapshot,
+};
 
 const MODEL_REQUEST_DEADLINE: Duration = Duration::from_secs(10 * 60);
 const CONTINUATION_AFFINITY_TIMEOUT: Duration = Duration::from_millis(100);
@@ -139,6 +141,9 @@ pub trait ExecutionService: Send + Sync {
         plaintext: &str,
     ) -> Result<AuthenticatedClient, ClientAuthenticationError>;
     fn public_models(&self, client: &AuthenticatedClient) -> Vec<PublicModelId>;
+    fn public_model_profiles(&self, _client: &AuthenticatedClient) -> Vec<PublicModelProfile> {
+        Vec::new()
+    }
     fn contains_public_model(&self, client: &AuthenticatedClient, model: &PublicModelId) -> bool;
     fn start(
         &self,
@@ -568,6 +573,12 @@ impl ExecutionService for DefaultExecutionService {
             .public_models_for_provider(client.policy.provider_kind())
     }
 
+    fn public_model_profiles(&self, client: &AuthenticatedClient) -> Vec<PublicModelProfile> {
+        client
+            .snapshot
+            .public_model_profiles_for_provider(client.policy.provider_kind())
+    }
+
     fn contains_public_model(&self, client: &AuthenticatedClient, model: &PublicModelId) -> bool {
         client
             .snapshot

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

@@ -20,7 +20,9 @@ use crate::error::{
 };
 use crate::event::{EventSequenceValidator, ProviderEvent};
 use crate::operation::Operation;
-use crate::routing::{ModelCapabilities, ProviderCandidate, ProviderKind, UpstreamModelId};
+use crate::routing::{
+    ModelCapabilities, ModelPresentation, ProviderCandidate, ProviderKind, UpstreamModelId,
+};
 
 /// Box 只出现在 Provider Registry 的统一 event envelope 边界。
 pub type EventStream =
@@ -433,6 +435,7 @@ pub struct ProviderRequestObservation {
 pub struct ProviderModelCapabilities {
     upstream_model: UpstreamModelId,
     capabilities: ModelCapabilities,
+    presentation: Option<ModelPresentation>,
 }
 
 /// Provider 实时目录成功发布后的进程内单调代次。
@@ -459,9 +462,16 @@ impl ProviderModelCapabilities {
         Self {
             upstream_model,
             capabilities,
+            presentation: None,
         }
     }
 
+    #[must_use]
+    pub fn with_presentation(mut self, presentation: ModelPresentation) -> Self {
+        self.presentation = Some(presentation);
+        self
+    }
+
     #[must_use]
     pub const fn upstream_model(&self) -> &UpstreamModelId {
         &self.upstream_model
@@ -471,6 +481,11 @@ impl ProviderModelCapabilities {
     pub const fn capabilities(&self) -> &ModelCapabilities {
         &self.capabilities
     }
+
+    #[must_use]
+    pub const fn presentation(&self) -> Option<&ModelPresentation> {
+        self.presentation.as_ref()
+    }
 }
 
 impl ProviderRequest {

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

@@ -756,18 +756,6 @@ impl ProviderEvent {
     pub fn has_client_event(&self) -> bool {
         !self.canonical.is_empty() || self.wire.is_some()
     }
-
-    /// 返回是否包含足以冻结下游 commit barrier 的 canonical fact。
-    ///
-    /// 单独的 wire event 不能冻结边界：协议层在看到 `Started`
-    /// 携带的网关响应身份前无法安全编码它。Core 会保留这些
-    /// 事件，直到首个 canonical 可见增量或终态再一并交付。
-    #[must_use]
-    pub fn is_commit_significant(&self) -> bool {
-        self.canonical
-            .iter()
-            .any(GatewayEvent::is_commit_significant)
-    }
 }
 
 impl From<GatewayEvent> for ProviderEvent {
@@ -776,23 +764,6 @@ impl From<GatewayEvent> for ProviderEvent {
     }
 }
 
-impl GatewayEvent {
-    /// 返回该 fact 是否足以冻结下游 commit barrier。
-    ///
-    /// 生命周期、结构与结算 facts 可以在换号前安全丢弃；首个可见增量或
-    /// Provider 正常终态才会使客户端输出不可撤回。
-    #[must_use]
-    pub const fn is_commit_significant(&self) -> bool {
-        matches!(
-            self,
-            Self::TextDelta(_)
-                | Self::ReasoningDelta(_)
-                | Self::ToolCallDelta(_)
-                | Self::Completed(_)
-        )
-    }
-}
-
 /// Canonical event 顺序错误。
 #[derive(Debug, Clone, PartialEq, Eq, Error)]
 pub enum EventSequenceError {

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

@@ -476,7 +476,7 @@ pub trait ProviderCredentialStatePort: Send + Sync {
     ) -> BoxFuture<'a, Result<bool, ProviderStoreError>>;
 }
 
-/// 临时 cooldown 只保存调度截止时间；原因事实仍由 PostgreSQL 账号状态持有。
+/// 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。
 #[derive(Debug, Clone, PartialEq, Eq)]
 pub struct ProviderCooldown {
     account_id: ProviderAccountId,

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

@@ -136,6 +136,177 @@ pub struct ModelCapabilities {
     upstream_validates_features: bool,
 }
 
+/// Provider 为客户端模型目录提供的展示与交互能力。
+///
+/// 该值不参与路由，也不把任一 Provider 的 wire 类型带入 Core。
+#[derive(Debug, Clone, Default, PartialEq, Eq)]
+pub struct ModelPresentation {
+    display_name: Option<String>,
+    description: Option<String>,
+    default_reasoning_effort: Option<String>,
+    supported_reasoning_efforts: Vec<String>,
+    context_window_tokens: Option<u64>,
+    image_input: bool,
+    agent_tools: bool,
+    parallel_tool_calls: bool,
+    search_tool: bool,
+    image_detail_original: bool,
+    verbosity: bool,
+    hidden: bool,
+}
+
+impl ModelPresentation {
+    #[must_use]
+    pub fn new(display_name: Option<String>, description: Option<String>) -> Self {
+        Self {
+            display_name,
+            description,
+            ..Self::default()
+        }
+    }
+
+    #[must_use]
+    pub fn with_reasoning(
+        mut self,
+        default_effort: Option<String>,
+        supported_efforts: Vec<String>,
+    ) -> Self {
+        self.default_reasoning_effort = default_effort;
+        self.supported_reasoning_efforts = supported_efforts;
+        self
+    }
+
+    #[must_use]
+    pub const fn with_context_window_tokens(mut self, context_window_tokens: Option<u64>) -> Self {
+        self.context_window_tokens = context_window_tokens;
+        self
+    }
+
+    #[must_use]
+    pub const fn with_image_input(mut self, image_input: bool) -> Self {
+        self.image_input = image_input;
+        self
+    }
+
+    #[must_use]
+    pub const fn with_agent_tools(mut self, agent_tools: bool, parallel_tool_calls: bool) -> Self {
+        self.agent_tools = agent_tools;
+        self.parallel_tool_calls = parallel_tool_calls;
+        self
+    }
+
+    #[must_use]
+    pub const fn with_search_tool(mut self, search_tool: bool) -> Self {
+        self.search_tool = search_tool;
+        self
+    }
+
+    #[must_use]
+    pub const fn with_image_detail_original(mut self, image_detail_original: bool) -> Self {
+        self.image_detail_original = image_detail_original;
+        self
+    }
+
+    #[must_use]
+    pub const fn with_verbosity(mut self, verbosity: bool) -> Self {
+        self.verbosity = verbosity;
+        self
+    }
+
+    #[must_use]
+    pub const fn with_hidden(mut self, hidden: bool) -> Self {
+        self.hidden = hidden;
+        self
+    }
+
+    #[must_use]
+    pub fn display_name(&self) -> Option<&str> {
+        self.display_name.as_deref()
+    }
+
+    #[must_use]
+    pub fn description(&self) -> Option<&str> {
+        self.description.as_deref()
+    }
+
+    #[must_use]
+    pub fn default_reasoning_effort(&self) -> Option<&str> {
+        self.default_reasoning_effort.as_deref()
+    }
+
+    #[must_use]
+    pub fn supported_reasoning_efforts(&self) -> &[String] {
+        &self.supported_reasoning_efforts
+    }
+
+    #[must_use]
+    pub const fn context_window_tokens(&self) -> Option<u64> {
+        self.context_window_tokens
+    }
+
+    #[must_use]
+    pub const fn image_input(&self) -> bool {
+        self.image_input
+    }
+
+    #[must_use]
+    pub const fn agent_tools(&self) -> bool {
+        self.agent_tools
+    }
+
+    #[must_use]
+    pub const fn parallel_tool_calls(&self) -> bool {
+        self.parallel_tool_calls
+    }
+
+    #[must_use]
+    pub const fn search_tool(&self) -> bool {
+        self.search_tool
+    }
+
+    #[must_use]
+    pub const fn image_detail_original(&self) -> bool {
+        self.image_detail_original
+    }
+
+    #[must_use]
+    pub const fn verbosity(&self) -> bool {
+        self.verbosity
+    }
+
+    #[must_use]
+    pub const fn hidden(&self) -> bool {
+        self.hidden
+    }
+}
+
+/// 一个公开模型及其 Provider 编译后的客户端画像。
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub struct PublicModelProfile {
+    model: PublicModelId,
+    presentation: ModelPresentation,
+}
+
+impl PublicModelProfile {
+    #[must_use]
+    pub const fn new(model: PublicModelId, presentation: ModelPresentation) -> Self {
+        Self {
+            model,
+            presentation,
+        }
+    }
+
+    #[must_use]
+    pub const fn model(&self) -> &PublicModelId {
+        &self.model
+    }
+
+    #[must_use]
+    pub const fn presentation(&self) -> &ModelPresentation {
+        &self.presentation
+    }
+}
+
 impl ModelCapabilities {
     #[must_use]
     pub fn new(operations: BTreeSet<OperationKind>, max_output_tokens: Option<u64>) -> Self {
@@ -205,6 +376,7 @@ pub struct ProviderModel {
     provider: ProviderKind,
     upstream_model: UpstreamModelId,
     capabilities: ModelCapabilities,
+    presentation: Option<ModelPresentation>,
 }
 
 impl ProviderModel {
@@ -218,9 +390,16 @@ impl ProviderModel {
             provider,
             upstream_model,
             capabilities,
+            presentation: None,
         }
     }
 
+    #[must_use]
+    pub fn with_presentation(mut self, presentation: ModelPresentation) -> Self {
+        self.presentation = Some(presentation);
+        self
+    }
+
     #[must_use]
     pub const fn provider(&self) -> &ProviderKind {
         &self.provider
@@ -230,6 +409,11 @@ impl ProviderModel {
     pub const fn upstream_model(&self) -> &UpstreamModelId {
         &self.upstream_model
     }
+
+    #[must_use]
+    pub const fn presentation(&self) -> Option<&ModelPresentation> {
+        self.presentation.as_ref()
+    }
 }
 
 /// 本次请求选择 Provider 时使用的动态过滤事实。

backend/crates/gateway-core/src/routing/snapshot.rs

@@ -231,11 +231,15 @@ async fn compile_runtime_snapshot(
             continue;
         };
         provider_models.extend(models.into_iter().map(|model| {
-            ProviderModel::new(
+            let compiled = ProviderModel::new(
                 provider.clone(),
                 model.upstream_model().clone(),
                 model.capabilities().clone(),
-            )
+            );
+            match model.presentation().cloned() {
+                Some(presentation) => compiled.with_presentation(presentation),
+                None => compiled,
+            }
         }));
     }
 
@@ -287,6 +291,8 @@ pub struct RuntimeSnapshot {
     account_selection_policy: AccountSelectionPolicy,
     providers: Arc<BTreeSet<ProviderKind>>,
     provider_models: Arc<BTreeMap<ProviderKind, BTreeMap<UpstreamModelId, ModelCapabilities>>>,
+    provider_model_presentations:
+        Arc<BTreeMap<ProviderKind, BTreeMap<UpstreamModelId, super::ModelPresentation>>>,
     model_mappings: Arc<BTreeMap<String, String>>,
     provider_catalog_generations: Arc<BTreeMap<ProviderKind, ProviderCatalogGeneration>>,
     client_policies: Arc<BTreeMap<ClientApiKeyId, ClientPolicy>>,
@@ -313,17 +319,24 @@ impl RuntimeSnapshot {
 
         let mut model_map =
             BTreeMap::<ProviderKind, BTreeMap<UpstreamModelId, ModelCapabilities>>::new();
+        let mut presentation_map =
+            BTreeMap::<ProviderKind, BTreeMap<UpstreamModelId, super::ModelPresentation>>::new();
         for model in provider_models {
-            if !provider_set.contains(model.provider()) {
+            let ProviderModel {
+                provider,
+                upstream_model,
+                capabilities,
+                presentation,
+            } = model;
+            if !provider_set.contains(&provider) {
                 return Err(RoutingError::NotFound {
                     entity: "provider",
-                    id: model.provider().to_string(),
+                    id: provider.to_string(),
                 });
             }
-            let models = model_map.entry(model.provider).or_default();
-            let upstream_model = model.upstream_model;
+            let models = model_map.entry(provider.clone()).or_default();
             if models
-                .insert(upstream_model.clone(), model.capabilities)
+                .insert(upstream_model.clone(), capabilities)
                 .is_some()
             {
                 return Err(RoutingError::DuplicateEntity {
@@ -331,6 +344,12 @@ impl RuntimeSnapshot {
                     id: upstream_model.to_string(),
                 });
             }
+            if let Some(presentation) = presentation {
+                presentation_map
+                    .entry(provider)
+                    .or_default()
+                    .insert(upstream_model, presentation);
+            }
         }
 
         let mut client_policy_map = BTreeMap::new();
@@ -350,6 +369,7 @@ impl RuntimeSnapshot {
             account_selection_policy,
             providers: Arc::new(provider_set),
             provider_models: Arc::new(model_map),
+            provider_model_presentations: Arc::new(presentation_map),
             model_mappings: Arc::new(BTreeMap::new()),
             provider_catalog_generations: Arc::new(BTreeMap::new()),
             client_policies: Arc::new(client_policy_map),
@@ -402,6 +422,38 @@ impl RuntimeSnapshot {
         models.into_iter().collect()
     }
 
+    /// 返回 Provider 已明确声明画像的公开模型；没有画像时不猜测 Provider 语义。
+    #[must_use]
+    pub fn public_model_profiles_for_provider(
+        &self,
+        provider: &ProviderKind,
+    ) -> Vec<super::PublicModelProfile> {
+        let Some(presentations) = self.provider_model_presentations.get(provider) else {
+            return Vec::new();
+        };
+        let mut profiles = BTreeMap::new();
+        for (model, presentation) in presentations {
+            if let Ok(public_model) = PublicModelId::new(model.as_str().to_owned()) {
+                profiles.insert(public_model, presentation.clone());
+            }
+        }
+        for alias in self.model_mappings.keys() {
+            let target = self.mapped_model(alias);
+            let Some(presentation) = presentations.iter().find_map(|(model, presentation)| {
+                (model.as_str() == target).then_some(presentation)
+            }) else {
+                continue;
+            };
+            if let Ok(public_model) = PublicModelId::new(alias.clone()) {
+                profiles.insert(public_model, presentation.clone());
+            }
+        }
+        profiles
+            .into_iter()
+            .map(|(model, presentation)| super::PublicModelProfile::new(model, presentation))
+            .collect()
+    }
+
     #[must_use]
     pub fn public_models(&self) -> Vec<PublicModelId> {
         self.providers

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

@@ -1005,6 +1005,51 @@ fn authenticated_native_continuation_reaches_every_attempt_context() {
     ));
 }
 
+#[test]
+fn streaming_native_continuation_delivers_started_event_before_later_output() {
+    let operation = generate_operation();
+    let route_plan = plan(&operation);
+    let (coordinator, _, _) = coordinator(vec![Script::Stream {
+        account_id: "acct_one",
+        items: complete_stream(None),
+    }]);
+    let continuation = NativeContinuationPin::new(
+        PreviousResponseId::new("previous-response").expect("previous response"),
+        SafeUpstreamValue::new("upstream-response").expect("upstream response"),
+        ProviderKind::new("openai").expect("provider"),
+        ProviderAccountId::new("acct_one").expect("account"),
+    );
+    let mut session = block_on(coordinator.start(
+        model_request(&operation, SystemTime::now() + Duration::from_secs(30)),
+        operation,
+        route_plan,
+        None,
+        Some(ContinuationBinding::Pinned(continuation)),
+        CancellationToken::new(),
+    ))
+    .expect("start execution");
+
+    let first = block_on(session.next_event())
+        .expect("first event")
+        .expect("started event");
+    assert_eq!(
+        first.commit_requirement(),
+        CommitRequirement::CommitBeforeDelivery
+    );
+    let events = first.into_provider_events();
+    assert!(matches!(
+        events.as_slice(),
+        [event]
+            if matches!(event.canonical_facts(), [GatewayEvent::Started(_)])
+    ));
+
+    block_on(session.commit_downstream(Some(200))).expect("commit started event");
+    while block_on(session.next_event())
+        .expect("remaining event")
+        .is_some()
+    {}
+}
+
 #[test]
 fn native_continuation_replays_owner_before_safely_switching_account() {
     let Operation::Generate(generate) = generate_operation() else {

backend/crates/gateway-core/tests/routing/snapshot.rs

@@ -24,8 +24,8 @@ use gateway_core::routing::snapshot::{
     SnapshotSubscriptionError, SnapshotSubscriptionPort, runtime_revision_needs_refresh,
 };
 use gateway_core::routing::{
-    ConfigRevision, ModelCapabilities, ProviderKind, PublicModelId, RuntimeSnapshot,
-    UpstreamModelId,
+    ConfigRevision, ModelCapabilities, ModelPresentation, ProviderKind, PublicModelId,
+    RuntimeSnapshot, UpstreamModelId,
 };
 use gateway_core::task::WorkerKind;
 
@@ -135,13 +135,19 @@ impl Provider for PublishingCatalogProvider {
         if self.queries.fetch_add(1, Ordering::SeqCst) == 0 {
             self.generation.store(1, Ordering::SeqCst);
         }
-        Ok(vec![ProviderModelCapabilities::new(
-            UpstreamModelId::new("upstream-model").expect("model"),
-            ModelCapabilities::new(
-                std::collections::BTreeSet::from([OperationKind::Generate]),
+        Ok(vec![
+            ProviderModelCapabilities::new(
+                UpstreamModelId::new("upstream-model").expect("model"),
+                ModelCapabilities::new(
+                    std::collections::BTreeSet::from([OperationKind::Generate]),
+                    None,
+                ),
+            )
+            .with_presentation(ModelPresentation::new(
+                Some("Upstream Model".to_owned()),
                 None,
-            ),
-        )])
+            )),
+        ])
     }
 
     async fn execute(
@@ -215,6 +221,15 @@ fn compiler_retries_when_provider_publishes_catalog_during_compilation() {
             .map(|generation| generation.get()),
         Some(1),
     );
+    let profiles =
+        snapshot.public_model_profiles_for_provider(&ProviderKind::new("alpha").expect("provider"));
+    assert_eq!(
+        profiles
+            .iter()
+            .map(|profile| profile.model().as_str())
+            .collect::<Vec<_>>(),
+        vec!["public-model", "upstream-model"],
+    );
 }
 
 #[test]

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

@@ -49,7 +49,7 @@ pub(crate) async fn download_file(
         .map_err(|error| super::internal(format!("failed to create download file: {error}")))?;
     let mut stream = response.bytes_stream();
     let mut downloaded = 0_u64;
-    let mut last_percent = 0_u8;
+    let mut next_progress = 10_u8;
     while let Some(chunk) = stream.next().await {
         let chunk = chunk.map_err(|error| upstream(format!("download stream failed: {error}")))?;
         downloaded = downloaded.saturating_add(u64::try_from(chunk.len()).unwrap_or(u64::MAX));
@@ -59,9 +59,17 @@ pub(crate) async fn download_file(
         file.write_all(&chunk)
             .map_err(|error| super::internal(format!("failed to write download: {error}")))?;
         let percent = download_percent(downloaded, expected_size);
-        if percent > last_percent {
-            last_percent = percent;
-            events.emit_progress(operation_id, "downloading release archive", percent);
+        if percent >= next_progress || downloaded == expected_size {
+            events.emit_progress(
+                operation_id,
+                &format!(
+                    "已下载 {} / {} ({percent}%)",
+                    format_bytes(downloaded),
+                    format_bytes(expected_size)
+                ),
+                percent,
+            );
+            next_progress = percent.saturating_add(10);
         }
     }
     file.sync_all()
@@ -155,3 +163,15 @@ fn download_percent(downloaded: u64, total: u64) -> u8 {
     )
     .unwrap_or(100)
 }
+
+pub(crate) fn format_bytes(bytes: u64) -> String {
+    const KIB: u64 = 1024;
+    const MIB: u64 = KIB * 1024;
+    if bytes >= MIB {
+        return format!("{:.1} MiB", bytes as f64 / MIB as f64);
+    }
+    if bytes >= KIB {
+        return format!("{:.1} KiB", bytes as f64 / KIB as f64);
+    }
+    format!("{bytes} B")
+}

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

@@ -39,6 +39,8 @@ use self::state::{
 };
 use self::swap::{replace_release_files, rollback_release};
 
+pub use self::release::validate_download_url;
+
 const APP_BINARY_NAME: &str = "codex-proxy-rs";
 const DEFAULT_WEB_DIST_DIR: &str = "/app/web/dist";
 const DEFAULT_GITHUB_API_BASE: &str = "https://api.github.com/repos";
@@ -230,8 +232,15 @@ impl ProcessSystemOperations {
             .as_deref()
             .ok_or_else(|| conflict("update repository is not configured"))?;
         self.events
-            .info(None, Some("release"), "fetching latest release");
-        let release = fetch_latest(&self.config.github_api_base, repository).await?;
+            .info(None, Some("release"), "正在获取最新 Release 信息");
+        let release = match fetch_latest(&self.config.github_api_base, repository).await {
+            Ok(release) => release,
+            Err(error) => {
+                self.events
+                    .error_terminal(None, Some("release"), error.to_string());
+                return Err(error);
+            }
+        };
         let detail = detail_from_release(&self.config, &release);
         if detail.latest_version != target {
             self.events.warning_terminal(
@@ -256,12 +265,17 @@ impl ProcessSystemOperations {
             Some(&target),
             &self.config.version,
         )?;
+        self.events.info(
+            Some(&operation_id),
+            Some("prepare"),
+            format!("准备更新到 v{target}"),
+        );
         let result = self.install_release(&release, &target, &operation_id).await;
         match &result {
             Ok(()) => self.events.success_terminal(
                 Some(&operation_id),
                 Some("done"),
-                "release files replaced",
+                "更新文件已替换，等待服务重启生效",
             ),
             Err(error) => {
                 self.events
@@ -292,7 +306,23 @@ impl ProcessSystemOperations {
         version: &str,
         operation_id: &str,
     ) -> Result<(), OperationError> {
+        self.events.info(
+            Some(operation_id),
+            Some("asset"),
+            "正在选择匹配当前平台的更新包",
+        );
         let archive = select_archive(release, version)?;
+        self.events.info(
+            Some(operation_id),
+            Some("asset"),
+            format!(
+                "已选择更新包 {} ({})",
+                archive.name,
+                download::format_bytes(archive.size)
+            ),
+        );
+        self.events
+            .info(Some(operation_id), Some("verify"), "正在校验更新资源");
         if archive.size == 0 || archive.size > MAX_DOWNLOAD_SIZE {
             return Err(invalid("release archive size is invalid"));
         }
@@ -304,12 +334,16 @@ impl ProcessSystemOperations {
         if checksum.size == 0 || checksum.size > MAX_CHECKSUM_SIZE {
             return Err(invalid("release checksum size is invalid"));
         }
+        self.events
+            .info(Some(operation_id), Some("prepare"), "正在创建临时更新目录");
         fs::create_dir_all(&self.config.update_temp_dir)
             .map_err(|error| internal(format!("failed to prepare update temp dir: {error}")))?;
         let temp_root = fs::canonicalize(&self.config.update_temp_dir)
             .map_err(|error| internal(format!("failed to resolve update temp dir: {error}")))?;
         let temp = UpdateTempDir::create(&temp_root)?;
         let archive_path = temp.path().join(&archive.name);
+        self.events
+            .info(Some(operation_id), Some("download"), "开始下载更新包");
         download_file(
             &archive.browser_download_url,
             &archive_path,
@@ -319,6 +353,10 @@ impl ProcessSystemOperations {
             &self.events,
         )
         .await?;
+        self.events
+            .success(Some(operation_id), Some("download"), "更新包下载完成");
+        self.events
+            .info(Some(operation_id), Some("checksum"), "正在校验 checksum");
         verify_checksum(
             &archive_path,
             &archive.name,
@@ -327,12 +365,23 @@ impl ProcessSystemOperations {
             &self.config.github_api_base,
         )
         .await?;
+        self.events
+            .success(Some(operation_id), Some("checksum"), "checksum 校验通过");
+        self.events
+            .info(Some(operation_id), Some("extract"), "正在解压更新包");
         let extracted = extract_release(&archive_path, temp.path())?;
+        self.events
+            .success(Some(operation_id), Some("extract"), "更新包解压完成");
+        self.events
+            .info(Some(operation_id), Some("replace"), "正在替换应用文件");
         replace_release_files(
             &self.config.executable_path()?,
             &self.config.web_dist_dir,
             extracted,
-        )
+        )?;
+        self.events
+            .success(Some(operation_id), Some("replace"), "应用文件替换完成");
+        Ok(())
     }
 }
 
@@ -513,6 +562,17 @@ impl UpdateEvents {
         );
     }
 
+    fn success(&self, operation_id: Option<&str>, step: Option<&str>, message: impl Into<String>) {
+        self.emit(
+            SystemUpdateEventLevel::Success,
+            operation_id,
+            step,
+            message,
+            None,
+            false,
+        );
+    }
+
     fn warning_terminal(
         &self,
         operation_id: Option<&str>,

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

@@ -177,7 +177,8 @@ pub(crate) fn select_archive<'a>(
         })
 }
 
-pub(crate) fn validate_download_url(raw: &str, api_base: &str) -> Result<(), OperationError> {
+/// 校验 GitHub Release 资产 URL 及其重定向目标是否位于受信任边界内。
+pub fn validate_download_url(raw: &str, api_base: &str) -> Result<(), OperationError> {
     let url = reqwest::Url::parse(raw)
         .map_err(|error| invalid(format!("invalid download URL: {error}")))?;
     if local_download_allowed(&url, api_base) {
@@ -294,6 +295,7 @@ fn github_download_host_allowed(host: &str) -> bool {
         || host.ends_with(".github.com")
         || host == "objects.githubusercontent.com"
         || host.ends_with(".objects.githubusercontent.com")
+        || host == "release-assets.githubusercontent.com"
 }
 
 fn local_download_allowed(url: &reqwest::Url, api_base: &str) -> bool {

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

@@ -10,7 +10,9 @@ use gateway_admin::model::system::{
 };
 use gateway_admin::ports::system::{SystemOperationErrorKind, SystemOperations};
 use gateway_core::engine::CancellationToken;
-use gateway_host::system_update::{ProcessSystemOperations, SystemUpdateConfig};
+use gateway_host::system_update::{
+    ProcessSystemOperations, SystemUpdateConfig, validate_download_url,
+};
 use sha2::{Digest as _, Sha256};
 use tar::{Builder, EntryType, Header};
 use wiremock::matchers::{method, path};
@@ -202,6 +204,68 @@ async fn update_events_should_open_authenticated_sse_stream() {
     assert!(stream.next().await.is_some());
 }
 
+#[tokio::test]
+async fn update_events_should_preserve_complete_release_stage_sequence() {
+    let server = MockServer::start().await;
+    let fixture = Fixture::new();
+    fixture
+        .mount_release(
+            &server,
+            TARGET_VERSION,
+            ArchiveKind::Safe,
+            ChecksumKind::Valid,
+        )
+        .await;
+    let service = fixture.service(&server);
+    let events = service.update_events();
+    service
+        .perform_update(Some(TARGET_VERSION.to_owned()))
+        .await
+        .expect("update");
+    let steps = events
+        .map(|event| event.step.expect("release update step"))
+        .collect::<Vec<_>>()
+        .await;
+
+    assert_eq!(
+        steps,
+        [
+            "release", "prepare", "asset", "asset", "verify", "prepare", "download", "download",
+            "download", "checksum", "checksum", "extract", "extract", "replace", "replace", "done",
+        ]
+    );
+}
+
+#[tokio::test]
+async fn update_events_should_report_download_bytes_and_percent() {
+    let server = MockServer::start().await;
+    let fixture = Fixture::new();
+    fixture
+        .mount_release(
+            &server,
+            TARGET_VERSION,
+            ArchiveKind::Safe,
+            ChecksumKind::Valid,
+        )
+        .await;
+    let service = fixture.service(&server);
+    let events = service.update_events();
+    service
+        .perform_update(Some(TARGET_VERSION.to_owned()))
+        .await
+        .expect("update");
+    let progress = events
+        .filter_map(
+            |event| async move { event.progress_percent.map(|value| (value, event.message)) },
+        )
+        .collect::<Vec<_>>()
+        .await;
+
+    assert!(progress.iter().any(|(percent, message)| {
+        *percent == 100 && message.starts_with("已下载 ") && message.ends_with(" (100%)")
+    }));
+}
+
 #[tokio::test]
 async fn update_should_fail_when_release_checksum_is_missing() {
     let server = MockServer::start().await;
@@ -280,6 +344,17 @@ async fn update_should_reject_release_archive_from_untrusted_host() {
     );
 }
 
+#[test]
+fn update_should_trust_github_release_asset_redirect_host() {
+    assert!(
+        validate_download_url(
+            "https://release-assets.githubusercontent.com/github-production-release-asset/archive",
+            "https://api.github.com/repos",
+        )
+        .is_ok()
+    );
+}
+
 #[tokio::test]
 async fn update_should_reject_release_archive_with_unsafe_path() {
     let server = MockServer::start().await;

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

@@ -23,7 +23,8 @@ const MAX_PAGE_SIZE: u16 = 100;
 const MAX_FILTER_BYTES: usize = 256;
 const MAX_SEARCH_BYTES: usize = 512;
 const MAX_ACCOUNT_IDS: usize = 200;
-const DASHBOARD_ACCOUNT_LIMIT: u16 = 50;
+/// 概览卡只展示最近使用的四个账号；完整账号用量由账号管理页单独查询。
+const DASHBOARD_ACCOUNT_LIMIT: u16 = 4;
 const DIAGNOSTIC_LIMIT: i64 = 100;
 const ACCOUNT_USAGE_TIMELINE_HOURS: i64 = 24;
 
@@ -660,7 +661,8 @@ pub struct OpsErrorRecord {
     pub upstream_model_id: Option<String>,
     pub upstream_transport: Option<String>,
     pub failure_kind: String,
-    pub status_code: Option<u16>,
+    pub client_status_code: Option<u16>,
+    pub upstream_status_code: Option<u16>,
     pub provider_error_code: Option<String>,
     pub client_response_id: Option<String>,
     pub upstream_request_id: Option<String>,
@@ -1626,7 +1628,8 @@ fn admin_ops_error(error: OpsErrorRecord) -> admin_observability::OpsError {
         upstream_model_id: error.upstream_model_id,
         upstream_transport: error.upstream_transport,
         failure_kind: error.failure_kind,
-        status_code: error.status_code,
+        client_status_code: error.client_status_code,
+        upstream_status_code: error.upstream_status_code,
         provider_error_code: error.provider_error_code,
         client_response_id: error.client_response_id,
         upstream_request_id: error.upstream_request_id,
@@ -1977,8 +1980,12 @@ async fn provider_account_metrics(
     let row = sqlx::query(
         "select count(*)::bigint as total,
                 count(*) filter (where enabled)::bigint as enabled,
-                count(*) filter (where not enabled or availability <> 'ready')::bigint
-                  as unavailable,
+                count(*) filter (
+                  where not (
+                    enabled and availability = 'ready'
+                      and (access_token_expires_at is null or access_token_expires_at > $1)
+                  )
+                )::bigint as unavailable,
                 count(*) filter (
                   where enabled and availability = 'ready'
                     and (access_token_expires_at is null or access_token_expires_at > $1)
@@ -2921,7 +2928,7 @@ const OPS_ERRORS_CTE: &str = "with errors as (
               mr.provider_kind,
               mr.provider_account_ref, mr.upstream_model_id, mr.upstream_transport,
               coalesce(mr.error_kind, 'failed') as failure_kind,
-              coalesce(mr.upstream_status_code, mr.client_status_code) as status_code,
+              mr.client_status_code, mr.upstream_status_code,
               mr.provider_error_code, mr.client_response_id, mr.upstream_request_id,
               mr.latency_ms, coalesce(mr.error_message, mr.error_kind, 'request failed') as message,
               1::integer as occurrence_count,
@@ -2932,7 +2939,8 @@ const OPS_ERRORS_CTE: &str = "with errors as (
        select 'ops_event'::text, oe.id, oe.model_request_id, oe.attempt_index,
               mr.client_api_key_ref, oe.component, oe.operation,
               oe.provider_kind, oe.provider_account_ref,
-              oe.upstream_model_id, null::text, oe.failure_kind, oe.status_code,
+              oe.upstream_model_id, null::text, oe.failure_kind,
+              null::integer as client_status_code, oe.status_code as upstream_status_code,
               oe.provider_error_code, mr.client_response_id, oe.upstream_request_id,
               oe.latency_ms, oe.message, oe.occurrence_count, oe.created_at,
               'ops_event:' || oe.id
@@ -2947,7 +2955,8 @@ async fn list_ops_errors(pool: &PgPool, query: OpsErrorQuery) -> StoreResult<Ops
         " select source, event_id, request_id, attempt_index, client_api_key_ref,
                  component, operation, provider_kind, provider_account_ref,
                  upstream_model_id, upstream_transport,
-                 failure_kind, status_code, provider_error_code, client_response_id,
+                 failure_kind, client_status_code, upstream_status_code,
+                 provider_error_code, client_response_id,
                  upstream_request_id, latency_ms, message, occurrence_count, occurred_at,
                  stable_sort_id
           from errors e where e.occurred_at >= ",
@@ -3046,7 +3055,7 @@ fn push_ops_filter(statement: &mut QueryBuilder<Postgres>, filter: &OpsErrorFilt
         statement.push_bind(i32::try_from(index).unwrap_or(i32::MAX));
     }
     if let Some(status) = filter.status_code {
-        statement.push(" and e.status_code = ");
+        statement.push(" and e.upstream_status_code = ");
         statement.push_bind(i32::from(status));
     }
     if let Some(search) = &filter.search {
@@ -3147,7 +3156,8 @@ fn ops_error_from_row(row: &sqlx::postgres::PgRow) -> StoreResult<OpsErrorRecord
         upstream_model_id: get(row, "upstream_model_id")?,
         upstream_transport: get(row, "upstream_transport")?,
         failure_kind: get(row, "failure_kind")?,
-        status_code: optional_status(row, "status_code")?,
+        client_status_code: optional_status(row, "client_status_code")?,
+        upstream_status_code: optional_status(row, "upstream_status_code")?,
         provider_error_code: get(row, "provider_error_code")?,
         client_response_id: get(row, "client_response_id")?,
         upstream_request_id: get(row, "upstream_request_id")?,

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

@@ -1700,10 +1700,8 @@ fn admin_account_status(
             .is_some_and(|expires_at| expires_at <= now)
     {
         AdminAccountStatus::Expired
-    } else if account.availability == ProviderAccountAvailability::Ready {
-        AdminAccountStatus::Active
     } else {
-        AdminAccountStatus::Attention
+        AdminAccountStatus::Active
     }
 }
 
@@ -1728,7 +1726,7 @@ fn admin_account_summary(
             .count(),
     )
     .unwrap_or(u64::MAX);
-    let attention = u64::try_from(
+    let unavailable = u64::try_from(
         accounts
             .iter()
             .filter(|account| {
@@ -1746,7 +1744,7 @@ fn admin_account_summary(
         total,
         active,
         quota_exhausted,
-        attention,
+        unavailable,
     }
 }
 
@@ -1797,7 +1795,6 @@ const fn admin_account_status_name(status: AdminAccountStatus) -> &'static str {
         AdminAccountStatus::QuotaExhausted => "quota_exhausted",
         AdminAccountStatus::Disabled => "disabled",
         AdminAccountStatus::Banned => "banned",
-        AdminAccountStatus::Attention => "attention",
     }
 }
 

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

@@ -1,4 +1,4 @@
-//! PostgreSQL 账号 cooldown 事实的可丢失 Redis 热缓存。
+//! 请求调度用的可丢失账号 cooldown Redis 存储。
 
 use std::{sync::Arc, time::SystemTime};
 
@@ -243,7 +243,7 @@ impl ProviderCooldownPort for RedisCredentialCooldownRepository {
     }
 }
 
-/// PostgreSQL 账号事实到可丢失 cooldown cache 的 write-through adapter。
+/// 账号状态写入到可丢失 cooldown cache 的桥接 adapter。
 pub struct CooldownCachingProviderAccountStore {
     authoritative: Arc<dyn ProviderAccountStore>,
     cooldowns: Arc<dyn CredentialCooldownRepository>,

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

@@ -62,6 +62,57 @@ fn postgres_admin_observability_adapter_implements_terminal_port() {
     assert_port::<PgAdminObservabilityStore>();
 }
 
+#[tokio::test]
+async fn dashboard_account_metrics_should_partition_available_and_unavailable_accounts() {
+    let Some(database) = TestDatabase::create("dashboard_account_metrics").await else {
+        return;
+    };
+    let now = Utc::now();
+    sqlx::query(
+        "insert into provider_accounts (
+           id, provider_kind, name, upstream_user_id, 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_available', 'openai', 'available', 'user-available', 'oauth',
+            '{}'::jsonb, 1, false, $1 + interval '1 day', true, 'ready', $1, $1, $1),
+           ('acct_expired', 'openai', 'expired', 'user-expired', 'oauth',
+            '{}'::jsonb, 1, false, $1 - interval '1 day', true, 'ready', $1, $1, $1),
+           ('acct_banned', 'xai', 'banned', 'user-banned', 'oauth',
+            '{}'::jsonb, 1, false, $1 + interval '1 day', true, 'banned', $1, $1, $1),
+           ('acct_disabled', 'xai', 'disabled', 'user-disabled', 'oauth',
+            '{}'::jsonb, 1, false, $1 + interval '1 day', false, 'ready', $1, $1, $1)",
+    )
+    .bind(now)
+    .execute(&database.pool)
+    .await
+    .expect("seed account metric states");
+    let repository = PgObservabilityRepository::new(database.pool.clone());
+    let range = ObservabilityRange::new(now - TimeDelta::hours(1), now + TimeDelta::hours(1))
+        .expect("dashboard range");
+
+    let metrics = repository
+        .dashboard_summary(range)
+        .await
+        .expect("dashboard summary")
+        .provider_accounts;
+
+    assert_eq!(
+        (
+            metrics.total,
+            metrics.enabled,
+            metrics.active,
+            metrics.unavailable,
+            metrics.expired,
+            metrics.disabled,
+            metrics.banned,
+        ),
+        (4, 3, 1, 3, 1, 1, 1),
+    );
+    database.close().await;
+}
+
 #[tokio::test]
 async fn calculated_usage_billing_facts_keep_only_completed_calculated_costs() {
     let Some(database) = TestDatabase::create("calculated_usage_billing_facts").await else {
@@ -589,6 +640,20 @@ async fn observability_queries_preserve_request_account_cost_and_diagnostic_fact
         .await
         .expect("ops errors");
     assert_eq!(errors.total, 2);
+    let request_error = errors
+        .items
+        .iter()
+        .find(|error| error.source == "model_request")
+        .expect("request error");
+    assert_eq!(request_error.client_status_code, Some(502));
+    assert_eq!(request_error.upstream_status_code, Some(429));
+    let attempt_error = errors
+        .items
+        .iter()
+        .find(|error| error.source == "ops_event")
+        .expect("attempt error");
+    assert_eq!(attempt_error.client_status_code, None);
+    assert_eq!(attempt_error.upstream_status_code, Some(429));
 
     database.close().await;
 }

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

@@ -126,13 +126,13 @@ async fn terminal_admin_list_filters_and_sorts_before_pagination_with_retained_u
     beta.availability = ProviderAccountAvailability::Banned;
     let mut charlie = account("acct_charlie", "user-charlie");
     charlie.email = Some("charlie@example.invalid".to_owned());
-    let mut attention = account("acct_attention", "user-attention");
-    attention.email = Some("attention@example.invalid".to_owned());
-    attention.availability = ProviderAccountAvailability::Invalid;
+    let mut invalid = account("acct_invalid", "user-invalid");
+    invalid.email = Some("invalid@example.invalid".to_owned());
+    invalid.availability = ProviderAccountAvailability::Invalid;
     let mut quota_exhausted = account("acct_quota_exhausted", "user-quota-exhausted");
     quota_exhausted.email = Some("quota-exhausted@example.invalid".to_owned());
     quota_exhausted.availability = ProviderAccountAvailability::QuotaExhausted;
-    for account in [alpha, beta, charlie, attention, quota_exhausted] {
+    for account in [alpha, beta, charlie, invalid, quota_exhausted] {
         repository
             .insert_provider_account(account)
             .await
@@ -214,9 +214,8 @@ async fn terminal_admin_list_filters_and_sorts_before_pagination_with_retained_u
     assert_eq!(usage_page.config_revision.get(), 1);
     assert_eq!(usage_page.total, 5);
     assert_eq!(usage_page.summary.total, 5);
-    assert_eq!(usage_page.summary.active, 2);
+    assert_eq!(usage_page.summary.active, 3);
     assert_eq!(usage_page.summary.quota_exhausted, 1);
-    assert_eq!(usage_page.summary.attention, 1);
     assert_eq!(
         usage_page
             .items
@@ -277,19 +276,6 @@ async fn terminal_admin_list_filters_and_sorts_before_pagination_with_retained_u
         .await
         .expect("filter banned accounts");
     assert_eq!(banned.items[0].id, "acct_beta");
-    let attention = store
-        .list_accounts(AccountListQuery {
-            page: 1,
-            page_size: PageSize::new(10).expect("page size"),
-            provider_kind: None,
-            search: None,
-            status: Some(AccountStatus::Attention),
-            sort: None,
-        })
-        .await
-        .expect("filter attention accounts");
-    assert_eq!(attention.items[0].id, "acct_attention");
-
     database.close().await;
 }
 

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

@@ -215,7 +215,7 @@ impl fmt::Debug for ExportManagedCodexCredential {
     }
 }
 
-/// CPR canonical 账号导出文档；只允许显式序列化，Debug 永不输出 token。
+/// CPR canonical 账号导出文档；只允许显式序列化，Debug 永不输出 credential secret。
 #[derive(Serialize)]
 #[serde(rename_all = "camelCase")]
 pub struct CodexCprExportDocument {
@@ -252,21 +252,47 @@ impl fmt::Debug for CodexCprExportDocument {
 
 #[derive(Serialize)]
 #[serde(rename_all = "camelCase")]
-struct CodexCprExportAccount {
+struct CodexCprExportCommon {
     id: String,
     email: Option<String>,
     account_id: Option<String>,
     user_id: Option<String>,
     label: Option<String>,
     plan_type: Option<String>,
-    token: String,
-    refresh_token: Option<String>,
-    access_token_expires_at: Option<String>,
     status: &'static str,
     added_at: String,
     updated_at: String,
 }
 
+#[derive(Serialize)]
+#[serde(rename_all = "camelCase")]
+struct CodexCprOAuthExportAccount {
+    #[serde(flatten)]
+    common: CodexCprExportCommon,
+    token: String,
+    refresh_token: Option<String>,
+    access_token_expires_at: Option<String>,
+}
+
+#[derive(Serialize)]
+#[serde(rename_all = "camelCase")]
+struct CodexCprAgentIdentityExportAccount {
+    #[serde(flatten)]
+    common: CodexCprExportCommon,
+    auth_mode: &'static str,
+    agent_runtime_id: String,
+    agent_private_key: String,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    task_id: Option<String>,
+}
+
+#[derive(Serialize)]
+#[serde(untagged)]
+enum CodexCprExportAccount {
+    OAuth(CodexCprOAuthExportAccount),
+    AgentIdentity(CodexCprAgentIdentityExportAccount),
+}
+
 /// App 已从 Store 读取的当前账号、revision 与明文 Provider JSON。
 pub struct RotateManagedCodexCredential {
     pub current: LoadedCredential,
@@ -444,31 +470,51 @@ impl CodexCredentialAdmin {
             }
             let data = CodexCredentialCodec::decode_complete(&item.current.credential)
                 .map_err(|_| CodexCredentialAdminError::InvalidCredential)?;
-            let CodexCredentialData::OAuth(data) = data else {
-                return Err(CodexCredentialAdminError::InvalidCredential);
-            };
-            if account.authentication_kind() != CODEX_AUTHENTICATION_KIND_OAUTH
-                || account.has_refresh_token() != data.refresh_token.is_some()
-            {
-                return Err(CodexCredentialAdminError::InvalidCredential);
-            }
-            accounts.push(CodexCprExportAccount {
+            let common = CodexCprExportCommon {
                 id: account.id().as_str().to_owned(),
                 email: account.email().map(str::to_owned),
                 account_id: account.upstream_account_id().map(str::to_owned),
                 user_id: Some(account.upstream_user_id().to_owned()),
                 label: Some(account.name().to_owned()),
                 plan_type: account.plan_type().map(str::to_owned),
-                token: data.access_token,
-                refresh_token: data.refresh_token,
-                access_token_expires_at: account
-                    .access_token_expires_at()
-                    .map(DateTime::<Utc>::from)
-                    .map(|value| value.to_rfc3339()),
                 status: cpr_status(&account),
                 added_at: china_rfc3339(item.added_at),
                 updated_at: china_rfc3339(item.updated_at),
-            });
+            };
+            let exported = match data {
+                CodexCredentialData::OAuth(data) => {
+                    if account.authentication_kind() != CODEX_AUTHENTICATION_KIND_OAUTH
+                        || account.has_refresh_token() != data.refresh_token.is_some()
+                    {
+                        return Err(CodexCredentialAdminError::InvalidCredential);
+                    }
+                    CodexCprExportAccount::OAuth(CodexCprOAuthExportAccount {
+                        common,
+                        token: data.access_token,
+                        refresh_token: data.refresh_token,
+                        access_token_expires_at: account
+                            .access_token_expires_at()
+                            .map(DateTime::<Utc>::from)
+                            .map(|value| value.to_rfc3339()),
+                    })
+                }
+                CodexCredentialData::AgentIdentity(data) => {
+                    if account.authentication_kind() != CODEX_AUTHENTICATION_KIND_AGENT_IDENTITY
+                        || account.has_refresh_token()
+                        || account.access_token_expires_at().is_some()
+                    {
+                        return Err(CodexCredentialAdminError::InvalidCredential);
+                    }
+                    CodexCprExportAccount::AgentIdentity(CodexCprAgentIdentityExportAccount {
+                        common,
+                        auth_mode: "agentIdentity",
+                        agent_runtime_id: data.agent_runtime_id,
+                        agent_private_key: data.agent_private_key,
+                        task_id: data.task_id,
+                    })
+                }
+            };
+            accounts.push(exported);
         }
         Ok(CodexCprExportDocument {
             source_format: "cpr",
@@ -1034,7 +1080,25 @@ fn parse_import_document(
     let values = import_account_values(payload)?;
     let mut accounts = Vec::new();
     for value in values {
-        let parsed = match shape {
+        // CPR 导出允许 OAuth 与 Agent Identity 共存；按账号选择 Provider
+        // 解析器，避免一个 Agent Identity 条目把整份文档误判成单一格式。
+        let account_shape = if looks_like_agent_identity_account(value) {
+            ImportDocumentShape::AgentIdentity
+        } else if shape == ImportDocumentShape::AgentIdentity {
+            if value.get("credentials").is_some() || value.get("tokens").is_some() {
+                ImportDocumentShape::CredentialBundle
+            } else if value
+                .as_object()
+                .is_some_and(auth_document_provider_is_openai)
+            {
+                ImportDocumentShape::AuthDocument
+            } else {
+                ImportDocumentShape::Native
+            }
+        } else {
+            shape
+        };
+        let parsed = match account_shape {
             ImportDocumentShape::Native => Some(parse_native_account(value)?),
             ImportDocumentShape::AgentIdentity => Some(parse_agent_identity_account(value)?),
             ImportDocumentShape::CredentialBundle => parse_credential_bundle_account(value)?,
@@ -1076,26 +1140,37 @@ fn parse_agent_identity_account(
         .get("agent_identity")
         .or_else(|| value.get("credentials"))
         .unwrap_or(value);
-    let auth_mode =
-        first_string(value, &["auth_mode"]).or_else(|| first_string(identity, &["auth_mode"]));
+    let auth_mode = first_string(value, &["auth_mode", "authMode"])
+        .or_else(|| first_string(identity, &["auth_mode", "authMode"]));
     if auth_mode.as_deref() != Some("agentIdentity") {
         return Err(CodexCredentialAdminError::InvalidCredential);
     }
-    let runtime_id = first_string(identity, &["agent_runtime_id"])
+    let runtime_id = first_string(identity, &["agent_runtime_id", "agentRuntimeId"])
         .ok_or(CodexCredentialAdminError::InvalidCredential)?;
-    let private_key = first_string(identity, &["agent_private_key"])
+    let private_key = first_string(identity, &["agent_private_key", "agentPrivateKey"])
         .ok_or(CodexCredentialAdminError::InvalidCredential)?;
     Ok(ParsedCodexImportAccount {
         id: first_string(value, &["id"]),
         name: first_string(value, &["name", "label"]),
         email: first_string(identity, &["email"]),
-        plan_type: first_string(identity, &["plan_type"]),
-        chatgpt_account_id: first_string(identity, &["chatgpt_account_id", "account_id"]),
-        chatgpt_user_id: first_string(identity, &["chatgpt_user_id"]),
+        plan_type: first_string(identity, &["plan_type", "planType"]),
+        chatgpt_account_id: first_string(
+            identity,
+            &[
+                "chatgpt_account_id",
+                "chatgptAccountId",
+                "account_id",
+                "accountId",
+            ],
+        ),
+        chatgpt_user_id: first_string(
+            identity,
+            &["chatgpt_user_id", "chatgptUserId", "user_id", "userId"],
+        ),
         authentication: ParsedCodexAuthentication::AgentIdentity {
             runtime_id,
             private_key,
-            task_id: first_string(identity, &["task_id"]),
+            task_id: first_string(identity, &["task_id", "taskId"]),
         },
         status: first_string(value, &["status"]),
     })
@@ -1379,24 +1454,23 @@ fn looks_like_credential_bundle(value: &Value) -> bool {
 
 fn looks_like_agent_identity_document(value: &Value) -> bool {
     import_account_values(value).is_ok_and(|accounts| {
-        accounts.iter().any(|account| {
-            account
-                .get("auth_mode")
-                .and_then(Value::as_str)
-                .is_some_and(|mode| mode == "agentIdentity")
-                || account
-                    .get("agent_identity")
-                    .and_then(|identity| identity.get("agent_runtime_id"))
-                    .is_some()
-                || account
-                    .get("credentials")
-                    .and_then(|credentials| credentials.get("auth_mode"))
-                    .and_then(Value::as_str)
-                    .is_some_and(|mode| mode == "agentIdentity")
-        })
+        accounts
+            .iter()
+            .any(|account| looks_like_agent_identity_account(account))
     })
 }
 
+fn looks_like_agent_identity_account(value: &Value) -> bool {
+    let identity = value
+        .get("agent_identity")
+        .or_else(|| value.get("credentials"))
+        .unwrap_or(value);
+    first_string(value, &["auth_mode", "authMode"])
+        .or_else(|| first_string(identity, &["auth_mode", "authMode"]))
+        .is_some_and(|mode| mode == "agentIdentity")
+        || first_string(identity, &["agent_runtime_id", "agentRuntimeId"]).is_some()
+}
+
 fn looks_like_auth_document(value: &Value) -> bool {
     import_account_values(value).is_ok_and(|accounts| {
         accounts.iter().any(|account| {

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

@@ -165,6 +165,9 @@ pub enum CodexCatalogVisibility {
     Hide,
     /// 不进入 picker。
     None,
+    /// 上游新增的可见性值；保留模型条目但不作本地可见性判断。
+    #[serde(other)]
+    Unknown,
 }
 
 /// 一个已完整校验的 Codex 真实模型。

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

@@ -1,6 +1,7 @@
 //! 核心 Generate operation 到 Codex Responses wire request 的严格编码。
 
 use gateway_core::operation::GenerateRequest;
+use gateway_protocol::openai::WS_REQUEST_HEADER_RESPONSES_LITE_CLIENT_METADATA_KEY;
 use serde_json::{Map, Value};
 use sha2::{Digest, Sha256};
 
@@ -127,64 +128,37 @@ struct ExtractedRequestContext {
 impl ExtractedRequestContext {
     fn from_body(body: &Map<String, Value>) -> Self {
         Self {
-            turn_state: body_or_metadata_string(
-                body,
-                &["turnState", "turn_state", "x-codex-turn-state"],
-            ),
-            turn_metadata: body_or_metadata_string(
-                body,
-                &["turnMetadata", "turn_metadata", "x-codex-turn-metadata"],
-            ),
-            beta_features: body_or_metadata_string(body, &["betaFeatures", "beta_features"]),
-            version: body_or_metadata_string(body, &["version"]),
-            include_timing_metrics: body_or_metadata_string(
-                body,
-                &["includeTimingMetrics", "include_timing_metrics"],
-            ),
-            codex_window_id: body_or_metadata_string(
-                body,
-                &["codexWindowId", "codex_window_id", "x-codex-window-id"],
-            ),
-            parent_thread_id: body_or_metadata_string(
-                body,
-                &[
-                    "parentThreadId",
-                    "parent_thread_id",
-                    "x-codex-parent-thread-id",
-                ],
-            ),
-            conversation_id: body_or_metadata_string(body, &["conversation_id", "conversationId"]),
-            session_id: body_or_metadata_string(body, &["session_id", "sessionId"]),
-            thread_id: body_or_metadata_string(body, &["thread_id", "threadId"]),
-            client_request_id: body_or_metadata_string(
-                body,
-                &[
-                    "x-client-request-id",
-                    "client_request_id",
-                    "clientRequestId",
-                ],
-            ),
-            turn_id: body_or_metadata_string(body, &["turn_id", "turnId", "x-codex-turn-id"]),
-            responses_lite: body_or_metadata_string(
-                body,
-                &["ws_request_header_x_openai_internal_codex_responses_lite"],
-            ),
-            memgen_request: body_or_metadata_string(
-                body,
-                &["x-openai-memgen-request", "memgen_request"],
-            ),
+            turn_state: body_string(body, "turnState"),
+            turn_metadata: body_string(body, "turnMetadata"),
+            beta_features: body_string(body, "betaFeatures"),
+            version: body_string(body, "version"),
+            include_timing_metrics: body_string(body, "includeTimingMetrics"),
+            codex_window_id: body_string(body, "codexWindowId"),
+            parent_thread_id: body_string(body, "parentThreadId"),
+            conversation_id: body_string(body, "conversation_id"),
+            session_id: body_string(body, "session_id"),
+            thread_id: body_string(body, "thread_id"),
+            client_request_id: body_string(body, "x-client-request-id"),
+            turn_id: body_string(body, "turn_id"),
+            // Responses Lite 在 WebSocket body 中的这个键是官方 header 投影，
+            // 不是普通上下文字段的 metadata 回退。
+            responses_lite: body
+                .get("client_metadata")
+                .and_then(Value::as_object)
+                .and_then(|metadata| {
+                    non_empty_string(
+                        metadata.get(WS_REQUEST_HEADER_RESPONSES_LITE_CLIENT_METADATA_KEY),
+                    )
+                }),
+            // Memory consolidation 只由官方请求头提供；body 中同名字段不参与
+            // transport 事实提取。
+            memgen_request: None,
         }
     }
 }
 
-fn body_or_metadata_string(body: &Map<String, Value>, keys: &[&str]) -> Option<String> {
-    keys.iter()
-        .find_map(|key| non_empty_string(body.get(*key)))
-        .or_else(|| {
-            let metadata = body.get("client_metadata")?.as_object()?;
-            keys.iter()
-                .find_map(|key| non_empty_string(metadata.get(*key)))
-        })
+fn body_string(body: &Map<String, Value>, key: &str) -> Option<String> {
+    non_empty_string(body.get(key))
 }
 
 fn non_empty_string(value: Option<&Value>) -> Option<String> {
@@ -528,10 +502,12 @@ fn apply_protocol_context(request: &mut CodexResponsesRequest, context: &Map<Str
         .parent_thread_id
         .take()
         .or_else(|| context_string(context, "parent_thread_id"));
+    let prompt_cache_key = request.prompt_cache_key().map(ToOwned::to_owned);
     request.client_conversation_id = request
         .client_conversation_id
         .take()
-        .or_else(|| context_string(context, "conversation_id"));
+        .or_else(|| context_string(context, "conversation_id"))
+        .or(prompt_cache_key);
     request.client_session_id = request
         .client_session_id
         .take()

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

@@ -3,23 +3,31 @@ use std::sync::{Arc, Mutex};
 use std::time::Duration;
 
 use async_trait::async_trait;
+use base64::Engine as _;
+use base64::engine::general_purpose::STANDARD;
 use chrono::{TimeZone, Utc};
+use ed25519_dalek::SigningKey;
+use ed25519_dalek::pkcs8::EncodePrivateKey as _;
 use futures::future::BoxFuture;
 use gateway_core::engine::credential::{
-    CredentialRevision, LoadedCredential, PlaintextCredential, ProviderAccountId,
+    AccountAvailability, CredentialRevision, LoadedCredential, PlaintextCredential,
+    ProviderAccount, ProviderAccountId,
 };
 use gateway_core::provider_ports::{
     ProviderLeaseAcquisition, ProviderLeasePort, ProviderLeaseRequest, ProviderRefreshLeaseRequest,
     ProviderStoreError,
 };
+use gateway_core::routing::ProviderKind;
 use provider_openai::OFFICIAL_CODEX_BASE_URL;
 use provider_openai::credential::token_client::{RefreshFailure, TokenPair, TokenRefresher};
 use provider_openai::credential::{
-    CodexAccountIdentityService, CodexAccountIdentityVerifier, CodexCredentialAdmin,
-    CodexCredentialAdminError, CodexCredentialAdminService, CodexCredentialCodec,
-    CodexIdentityExpectation, CodexIdentityVerification, CodexIdentityVerificationError,
-    CodexJwtIdentityVerifier, ExportManagedCodexCredential, ImportCodexOAuthCredential,
-    ReqwestCodexAuthenticatedAccountSource, ReqwestOpenAiJwksSource, RotateManagedCodexCredential,
+    CODEX_AUTHENTICATION_KIND_AGENT_IDENTITY, CodexAccountIdentityService,
+    CodexAccountIdentityVerifier, CodexAgentIdentityAuthMode, CodexAgentIdentityCredentialData,
+    CodexCredentialAdmin, CodexCredentialAdminError, CodexCredentialAdminService,
+    CodexCredentialCodec, CodexIdentityExpectation, CodexIdentityVerification,
+    CodexIdentityVerificationError, CodexJwtIdentityVerifier, ExportManagedCodexCredential,
+    ImportCodexOAuthCredential, ReqwestCodexAuthenticatedAccountSource, ReqwestOpenAiJwksSource,
+    RotateManagedCodexCredential,
 };
 use provider_openai::transport::profile::{CodexWireProfile, CodexWireProfileState};
 use secrecy::{ExposeSecret, SecretString};
@@ -137,6 +145,60 @@ fn export_item(id: &str, token: &str) -> ExportManagedCodexCredential {
     }
 }
 
+fn agent_export_item(id: &str) -> (ExportManagedCodexCredential, String) {
+    let signing_key = SigningKey::from_bytes(&[7; 32]);
+    let private_key = STANDARD.encode(
+        signing_key
+            .to_pkcs8_der()
+            .expect("encode Agent Identity key")
+            .as_bytes(),
+    );
+    let credential =
+        CodexCredentialCodec::encode_agent_identity(CodexAgentIdentityCredentialData {
+            schema_version: 1,
+            auth_mode: CodexAgentIdentityAuthMode::AgentIdentity,
+            installation_id: "00000000-0000-4000-8000-000000000007".to_owned(),
+            agent_runtime_id: "runtime-export".to_owned(),
+            agent_private_key: private_key.clone(),
+            task_id: Some("task-export".to_owned()),
+            cookies: Vec::new(),
+        })
+        .expect("encode Agent Identity credential");
+    let account = ProviderAccount::new(
+        ProviderAccountId::new(id.to_owned()).expect("account id"),
+        ProviderKind::new("openai").expect("provider"),
+        format!("Agent Identity {id}"),
+        "agent-user-export".to_owned(),
+        CODEX_AUTHENTICATION_KIND_AGENT_IDENTITY.to_owned(),
+        CredentialRevision::new(1).expect("revision"),
+        None,
+    )
+    .with_profile(
+        Some("agent-export@example.com".to_owned()),
+        Some("agent-account-export".to_owned()),
+        Some("pro".to_owned()),
+    )
+    .with_runtime_state(true, AccountAvailability::Ready, None)
+    .with_refresh_schedule(false, None);
+    (
+        ExportManagedCodexCredential {
+            current: LoadedCredential {
+                account,
+                credential,
+            },
+            added_at: Utc
+                .with_ymd_and_hms(2026, 7, 18, 2, 47, 1)
+                .single()
+                .expect("added at"),
+            updated_at: Utc
+                .with_ymd_and_hms(2026, 7, 19, 3, 0, 0)
+                .single()
+                .expect("updated at"),
+        },
+        private_key,
+    )
+}
+
 #[test]
 fn cpr_export_matches_the_canonical_real_document_shape() {
     let document = CodexCredentialAdmin
@@ -211,6 +273,57 @@ fn cpr_export_batch_validation_and_debug_are_secret_safe() {
     assert!(debug.contains("account_count: 1"));
 }
 
+#[test]
+fn cpr_export_preserves_agent_identity_material_without_oauth_fields() {
+    let (item, private_key) = agent_export_item("acct_agent_export");
+    let value = CodexCredentialAdmin
+        .format_cpr_export(vec![item])
+        .expect("Agent Identity export")
+        .into_json()
+        .expect("serialize Agent Identity export");
+    let account = &value["accounts"][0];
+
+    assert_eq!(account["authMode"], "agentIdentity");
+    assert_eq!(account["agentRuntimeId"], "runtime-export");
+    assert!(
+        account["agentPrivateKey"]
+            .as_str()
+            .is_some_and(|value| value == private_key)
+    );
+    assert_eq!(account["taskId"], "task-export");
+    assert!(account.get("token").is_none());
+    assert!(account.get("refreshToken").is_none());
+    assert!(account.get("accessTokenExpiresAt").is_none());
+}
+
+#[tokio::test]
+async fn cpr_export_and_import_support_mixed_oauth_and_agent_identity_accounts() {
+    let (agent, _) = agent_export_item("acct_mixed_agent");
+    let value = CodexCredentialAdmin
+        .format_cpr_export(vec![
+            export_item("acct_mixed_oauth", "token-acct_mixed_oauth"),
+            agent,
+        ])
+        .expect("mixed export")
+        .into_json()
+        .expect("serialize mixed export");
+
+    let prepared = import_service(unused_import_refresher())
+        .prepare_import_document(value)
+        .await
+        .expect("mixed CPR import");
+    assert_eq!(prepared.accounts().len(), 2);
+    assert!(prepared.accounts().iter().any(|account| {
+        account.account.authentication_kind() == CODEX_AUTHENTICATION_KIND_AGENT_IDENTITY
+    }));
+    assert!(
+        prepared
+            .accounts()
+            .iter()
+            .any(|account| { account.account.authentication_kind() == "oauth" })
+    );
+}
+
 pub(super) struct ManualRefresher {
     outcome: Mutex<Option<Result<TokenPair, RefreshFailure>>>,
     seen: Mutex<Vec<String>>,

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

@@ -1,9 +1,9 @@
 use chrono::{TimeZone, Utc};
 use provider_openai::transport::profile::{CodexWireProfile, CodexWireProfileState};
 use provider_openai::transport::{
-    CodexBackendClient, CodexCatalogCapabilityEvidence, CodexClientError, CodexModelCatalogError,
-    CodexRequestContext, MAX_CODEX_MODEL_CATALOG_BYTES, build_reqwest_client,
-    parse_codex_model_catalog,
+    CodexBackendClient, CodexCatalogCapabilityEvidence, CodexCatalogVisibility, CodexClientError,
+    CodexModelCatalogError, CodexRequestContext, MAX_CODEX_MODEL_CATALOG_BYTES,
+    build_reqwest_client, parse_codex_model_catalog,
 };
 use wiremock::matchers::{header, method, path, query_param};
 use wiremock::{Mock, MockServer, ResponseTemplate};
@@ -172,6 +172,20 @@ fn unknown_input_modality_should_degrade_capability_evidence_to_unknown() {
     );
 }
 
+#[test]
+fn unknown_visibility_should_degrade_without_rejecting_the_snapshot() {
+    let snapshot = parse_codex_model_catalog(
+        br#"{"models":[{"slug":"gpt-future","display_name":"Future","visibility":"preview_only"}]}"#,
+        None,
+    )
+    .expect("future visibility should not reject the snapshot");
+
+    assert_eq!(
+        snapshot.models()[0].metadata().visibility(),
+        Some(CodexCatalogVisibility::Unknown)
+    );
+}
+
 #[test]
 fn invalid_request_slug_should_fail_the_entire_snapshot() {
     let result = parse_codex_model_catalog(

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

@@ -126,6 +126,87 @@ fn encoder_should_forward_the_client_prompt_cache_key() {
     );
 }
 
+#[test]
+fn encoder_should_restore_conversation_fallback_after_body_and_context_values() {
+    let request = request(Map::from_iter([
+        ("model".to_owned(), json!("client-model")),
+        ("input".to_owned(), json!("prompt")),
+        ("prompt_cache_key".to_owned(), json!("cache-conversation")),
+    ]));
+
+    let encoded = encode_generate_request(&request, "gpt-test").expect("encode");
+
+    assert_eq!(
+        encoded.client_conversation_id.as_deref(),
+        Some("cache-conversation")
+    );
+}
+
+#[test]
+fn encoder_should_ignore_legacy_context_aliases_and_metadata_fallbacks() {
+    let request = request(Map::from_iter([
+        ("model".to_owned(), json!("client-model")),
+        ("input".to_owned(), json!("prompt")),
+        ("turn_state".to_owned(), json!("legacy-turn-state")),
+        (
+            "x-codex-turn-state".to_owned(),
+            json!("legacy-header-turn-state"),
+        ),
+        ("turn_metadata".to_owned(), json!("legacy-turn-metadata")),
+        ("beta_features".to_owned(), json!("legacy-beta")),
+        ("include_timing_metrics".to_owned(), json!("legacy-timing")),
+        ("codex_window_id".to_owned(), json!("legacy-window")),
+        (
+            "x-codex-window-id".to_owned(),
+            json!("legacy-header-window"),
+        ),
+        ("parent_thread_id".to_owned(), json!("legacy-parent")),
+        (
+            "x-codex-parent-thread-id".to_owned(),
+            json!("legacy-header-parent"),
+        ),
+        ("conversationId".to_owned(), json!("legacy-conversation")),
+        ("sessionId".to_owned(), json!("legacy-session")),
+        ("threadId".to_owned(), json!("legacy-thread")),
+        ("client_request_id".to_owned(), json!("legacy-request")),
+        ("clientRequestId".to_owned(), json!("legacy-camel-request")),
+        ("turnId".to_owned(), json!("legacy-turn")),
+        ("x-codex-turn-id".to_owned(), json!("legacy-header-turn")),
+        (
+            "client_metadata".to_owned(),
+            json!({
+                "turnState": "metadata-turn-state",
+                "turnMetadata": "metadata-turn-metadata",
+                "conversation_id": "metadata-conversation",
+                "session_id": "metadata-session",
+                "thread_id": "metadata-thread",
+                "x-codex-window-id": "metadata-window"
+            }),
+        ),
+    ]));
+
+    let encoded = encode_generate_request(&request, "gpt-test").expect("encode");
+
+    assert_eq!(
+        (
+            encoded.turn_state,
+            encoded.turn_metadata,
+            encoded.beta_features,
+            encoded.include_timing_metrics,
+            encoded.codex_window_id,
+            encoded.parent_thread_id,
+            encoded.client_conversation_id,
+            encoded.client_session_id,
+            encoded.client_thread_id,
+            encoded.client_request_id,
+            encoded.client_turn_id,
+        ),
+        (
+            None, None, None, None, None, None, None, None, None, None, None,
+        )
+    );
+}
+
 #[test]
 fn encoder_should_project_explicit_websocket_transport_without_touching_body() {
     let payload = ProtocolPayload::json_object(

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

@@ -11,8 +11,9 @@ use gateway_admin::model::accounts::{
     AccountAvailability as AdminAccountAvailability, AccountRecord,
 };
 use gateway_admin::model::observability::{
-    CalculatedBillingBreakdown, CurrencyCost, DashboardWireAttribute, DashboardWireProfile,
-    DashboardWireTarget, DecimalAmount, ProviderBillingInput,
+    CalculatedBillingBreakdown, CurrencyCost, DashboardDesktopRelease, DashboardWireAttribute,
+    DashboardWireProfile, DashboardWireTarget, DecimalAmount, DesktopReleaseStatus,
+    ProviderBillingInput,
 };
 use gateway_admin::model::provider_credentials::{
     AuthorizationMutationTarget, AuthorizationOwner, AuthorizationOwnerBinding,
@@ -57,6 +58,7 @@ use crate::credential::{
     PreparedGrokCredentialRotation, PreparedGrokCredentialRotationGuard, RedirectUriAllowlist,
     RotateManagedGrokCredential, SecretValue, VerifiedGrokAccount, VerifiedTokenSet,
 };
+use crate::transport::profile::{GrokCliReleaseSnapshot, GrokCliReleaseStatus};
 use crate::transport::{GROK_CLI_BASE_URL, XAI_PROVIDER_NAME, grok_billing_breakdown};
 
 const PENDING_SCHEMA_VERSION: u64 = 2;
@@ -66,6 +68,7 @@ const MAX_PENDING_TEXT_BYTES: usize = 512;
 pub(crate) struct XaiAdminProvider {
     provider_kind: ProviderKind,
     wire_profile: XaiWireProfileState,
+    cli_release: GrokCliReleaseStatus,
     accounts: Arc<dyn ProviderAccountStore>,
     repository: GrokCredentialRepository,
     oauth_config: GrokOAuthConfig,
@@ -95,10 +98,12 @@ impl XaiAdminProvider {
         wire_profile: XaiWireProfileState,
         accounts: Arc<dyn ProviderAccountStore>,
         services: XaiAdminServices,
+        cli_release: GrokCliReleaseStatus,
     ) -> Self {
         Self {
             provider_kind,
             wire_profile,
+            cli_release,
             accounts,
             repository: services.repository,
             oauth_config: services.oauth_config,
@@ -216,34 +221,42 @@ impl ProviderAdmin for XaiAdminProvider {
     }
 
     fn dashboard_wire_profile(&self) -> Option<DashboardWireProfile> {
+        let profile = self.wire_profile.snapshot();
+        let release = dashboard_cli_release(&profile.client_version, self.cli_release.snapshot());
         Some(DashboardWireProfile {
             provider: self.provider_kind.as_str().to_owned(),
             product: "Grok Build".to_owned(),
-            version: self.wire_profile.client_version(),
+            version: profile.client_version.clone(),
             build: None,
             target: DashboardWireTarget {
-                os_type: self.wire_profile.target_os(),
+                os_type: profile.target_os.clone(),
                 os_version: "—".to_owned(),
-                arch: self.wire_profile.target_arch(),
-                terminal: self.wire_profile.client_mode(),
+                arch: profile.target_arch.clone(),
+                terminal: profile.client_mode.clone(),
             },
-            user_agent: self.wire_profile.user_agent(),
+            user_agent: format!(
+                "{}/{} ({}; {})",
+                profile.client_identifier,
+                profile.client_version,
+                profile.target_os,
+                profile.target_arch
+            ),
             attributes: vec![
                 DashboardWireAttribute {
                     label: "客户端标识".to_owned(),
-                    value: self.wire_profile.client_identifier(),
+                    value: profile.client_identifier,
                 },
                 DashboardWireAttribute {
                     label: "运行模式".to_owned(),
-                    value: self.wire_profile.client_mode(),
+                    value: profile.client_mode,
                 },
                 DashboardWireAttribute {
                     label: "Token 认证".to_owned(),
                     value: "xai-grok-cli".to_owned(),
                 },
             ],
-            verified_at: Some(self.wire_profile.verified_at()),
-            release: None,
+            verified_at: Some(profile.verified_at),
+            release: Some(release),
         })
     }
 
@@ -771,6 +784,36 @@ fn validate_account_record(
     Ok(())
 }
 
+fn dashboard_cli_release(
+    profile_version: &str,
+    snapshot: GrokCliReleaseSnapshot,
+) -> DashboardDesktopRelease {
+    let status = if snapshot.checked_at.is_none() {
+        DesktopReleaseStatus::Unchecked
+    } else if snapshot.last_error.is_some() {
+        DesktopReleaseStatus::Failed
+    } else if snapshot.latest_version.as_deref() == Some(profile_version) {
+        DesktopReleaseStatus::Current
+    } else if snapshot.latest_version.is_some() {
+        DesktopReleaseStatus::UpdateAvailable
+    } else {
+        DesktopReleaseStatus::Failed
+    };
+    DashboardDesktopRelease {
+        status,
+        checked_at: snapshot.checked_at,
+        latest_version: snapshot.latest_version,
+        latest_build: None,
+        published_at: None,
+        minimum_system_version: None,
+        hardware_requirements: None,
+        download_url: None,
+        download_size: None,
+        signature_present: None,
+        error: snapshot.last_error,
+    }
+}
+
 fn account_from_record(account: &AccountRecord) -> Result<ProviderAccount, ProviderAdminError> {
     let account_id = ProviderAccountId::new(account.id.clone())
         .map_err(|_| provider_error(ProviderAdminErrorKind::Invalid))?;

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

@@ -12,9 +12,10 @@ use gateway_core::engine::credential::{
     AccountAvailability, CredentialRevision, ProviderAccountId,
 };
 use gateway_core::provider_ports::{
-    ProviderLeaseAcquisition, ProviderLeasePort, ProviderLeaseRequest,
-    ProviderRefreshCapacityRequest, ProviderRefreshLeaseRequest, ProviderRefreshPolicy,
-    ProviderRuntimePolicyPort, ProviderStoreError, provider_refresh_retry_at,
+    ProviderCooldown, ProviderCooldownPort, ProviderLeaseAcquisition, ProviderLeasePort,
+    ProviderLeaseRequest, ProviderRefreshCapacityRequest, ProviderRefreshLeaseRequest,
+    ProviderRefreshPolicy, ProviderRuntimePolicyPort, ProviderStoreError,
+    provider_refresh_retry_at,
 };
 
 use super::catalog::GrokCredentialCatalogService;
@@ -36,6 +37,7 @@ 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);
+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。
 pub struct DueGrokCredential {
@@ -274,6 +276,7 @@ pub struct GrokCredentialRefreshService {
     refresher: Arc<dyn GrokCredentialRefresher>,
     catalog: Arc<GrokCredentialCatalogService>,
     leases: Arc<dyn ProviderLeasePort>,
+    cooldowns: Arc<dyn ProviderCooldownPort>,
     runtime_policy: Arc<dyn ProviderRuntimePolicyPort>,
 }
 
@@ -283,6 +286,7 @@ impl GrokCredentialRefreshService {
         refresher: Arc<dyn GrokCredentialRefresher>,
         catalog: Arc<GrokCredentialCatalogService>,
         leases: Arc<dyn ProviderLeasePort>,
+        cooldowns: Arc<dyn ProviderCooldownPort>,
         runtime_policy: Arc<dyn ProviderRuntimePolicyPort>,
     ) -> Self {
         Self {
@@ -290,6 +294,7 @@ impl GrokCredentialRefreshService {
             refresher,
             catalog,
             leases,
+            cooldowns,
             runtime_policy,
         }
     }
@@ -297,9 +302,6 @@ impl GrokCredentialRefreshService {
     pub async fn refresh_due(
         &self,
     ) -> Result<Vec<GrokCredentialRefreshOutcome>, GrokCredentialRefreshError> {
-        if !self.prepare_cycle_if_due().await? {
-            return Ok(Vec::new());
-        }
         self.refresh_due_excluding(&[]).await
     }
 
@@ -455,8 +457,21 @@ impl GrokCredentialRefreshService {
             .into_iter()
             .map(|account_id| GrokCredentialRefreshOutcome::Failed { account_id })
             .collect::<Vec<_>>();
-        outcomes.reserve(batch.credentials.len());
-        let refreshed = stream::iter(batch.credentials.into_iter().map(|credential| async move {
+        let mut credentials = Vec::with_capacity(batch.credentials.len());
+        for credential in batch.credentials {
+            if !self.runtime_cooldown_active(&credential).await {
+                credentials.push(credential);
+            }
+        }
+        if credentials.is_empty() {
+            return Ok(outcomes);
+        }
+        self.refresher
+            .prepare_cycle()
+            .await
+            .map_err(|_| GrokCredentialRefreshError::Preparation)?;
+        outcomes.reserve(credentials.len());
+        let refreshed = stream::iter(credentials.into_iter().map(|credential| async move {
             let account_id = credential.account_id.clone();
             (
                 account_id,
@@ -539,14 +554,11 @@ impl GrokCredentialRefreshService {
                 )
                 .await
             }
-            Err(GrokRefreshFailure::Ambiguous) => self
-                .persist_terminal_failure(
-                    credential,
-                    GrokCredentialAvailability::Invalid,
-                    "refresh_ambiguous",
-                )
-                .await
-                .map(|_| GrokCredentialRefreshOutcome::Ambiguous { account_id }),
+            Err(GrokRefreshFailure::Ambiguous) => {
+                self.record_runtime_cooldown(&credential, REFRESH_AMBIGUITY_COOLDOWN)
+                    .await;
+                Ok(GrokCredentialRefreshOutcome::Ambiguous { account_id })
+            }
             Err(GrokRefreshFailure::Transient) => {
                 if self
                     .persist_backoff(&credential, "refresh-transient")
@@ -644,6 +656,36 @@ impl GrokCredentialRefreshService {
         })
     }
 
+    async fn runtime_cooldown_active(&self, credential: &DueGrokCredential) -> bool {
+        let Ok(Some(cooldown)) = self.cooldowns.read(credential.account_id()).await else {
+            return false;
+        };
+        if cooldown.credential_revision() != credential.credential_revision() {
+            if cooldown.credential_revision() < credential.credential_revision() {
+                let _ = self
+                    .cooldowns
+                    .clear(credential.account_id(), credential.credential_revision())
+                    .await;
+            }
+            return false;
+        }
+        cooldown.until() > SystemTime::now()
+    }
+
+    async fn record_runtime_cooldown(&self, credential: &DueGrokCredential, duration: Duration) {
+        let cooldown_until = SystemTime::now()
+            .checked_add(duration)
+            .unwrap_or_else(SystemTime::now);
+        let _ = self
+            .cooldowns
+            .put_if_later(ProviderCooldown::new(
+                credential.account_id.clone(),
+                credential.credential_revision,
+                cooldown_until,
+            ))
+            .await;
+    }
+
     async fn persist_terminal_failure(
         &self,
         credential: DueGrokCredential,
@@ -721,10 +763,12 @@ impl GrokCredentialRecovery for GrokCredentialRefreshService {
                 GrokCredentialRefreshOutcome::Refreshed { .. }
                 | GrokCredentialRefreshOutcome::Stale { .. },
             ) => GrokCredentialRecoveryOutcome::Recovered,
-            Ok(
-                GrokCredentialRefreshOutcome::Invalidated { .. }
-                | GrokCredentialRefreshOutcome::Ambiguous { .. },
-            ) => GrokCredentialRecoveryOutcome::Rejected,
+            Ok(GrokCredentialRefreshOutcome::Invalidated { .. }) => {
+                GrokCredentialRecoveryOutcome::Rejected
+            }
+            Ok(GrokCredentialRefreshOutcome::Ambiguous { .. }) => {
+                GrokCredentialRecoveryOutcome::Unavailable
+            }
             Ok(
                 GrokCredentialRefreshOutcome::Transient { .. }
                 | GrokCredentialRefreshOutcome::Rejected { .. }

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

@@ -7,10 +7,11 @@ use std::time::{Duration, SystemTime};
 use chrono::Utc;
 use gateway_core::engine::credential::{
     AccountCandidate, AccountFeedbackStats, AccountSelectionContext, AccountSelector,
+    ProviderAccount, ProviderAccountId,
 };
 use gateway_core::provider_ports::{
-    ProviderLeaseAcquisition, ProviderLeasePort, ProviderLeaseRequest,
-    ProviderSchedulingLeaseRequest,
+    ProviderCooldown, ProviderCooldownPort, ProviderLeaseAcquisition, ProviderLeasePort,
+    ProviderLeaseRequest, ProviderSchedulingLeaseRequest,
 };
 use gateway_core::routing::ProviderKind;
 
@@ -26,6 +27,7 @@ use crate::{
 const DEFAULT_RATE_LIMIT_COOLDOWN: Duration = Duration::from_secs(60);
 const MAX_RATE_LIMIT_COOLDOWN: Duration = Duration::from_secs(24 * 60 * 60);
 const STREAM_INTERRUPTION_COOLDOWN: Duration = Duration::from_secs(30);
+const UNAUTHORIZED_COOLDOWN: Duration = Duration::from_secs(60);
 
 /// 仅经 Core account port、TTL catalog cache 和 Redis lease 选择一个 OAuth session。
 pub struct GrokAccountSessionSelector {
@@ -34,6 +36,7 @@ pub struct GrokAccountSessionSelector {
     catalog_cache: Arc<dyn GrokCredentialCatalogCache>,
     quota: Arc<GrokCredentialQuotaService>,
     scheduling: Arc<dyn ProviderLeasePort>,
+    cooldowns: Arc<dyn ProviderCooldownPort>,
     account_feedback: Arc<AccountFeedbackStats>,
 }
 
@@ -45,6 +48,7 @@ impl GrokAccountSessionSelector {
         catalog_cache: Arc<dyn GrokCredentialCatalogCache>,
         quota: Arc<GrokCredentialQuotaService>,
         scheduling: Arc<dyn ProviderLeasePort>,
+        cooldowns: Arc<dyn ProviderCooldownPort>,
         account_feedback: Arc<AccountFeedbackStats>,
     ) -> Self {
         Self {
@@ -53,6 +57,7 @@ impl GrokAccountSessionSelector {
             catalog_cache,
             quota,
             scheduling,
+            cooldowns,
             account_feedback,
         }
     }
@@ -94,8 +99,10 @@ impl GrokAccountSessionSelector {
             .load_state(&self.provider_kind, &account_ids)
             .await
             .map_err(|_| GrokSessionSelectorError::Unavailable)?;
+        let runtime_cooldowns = self.runtime_cooldowns(&catalog_eligible).await;
         let mut candidates = catalog_eligible
             .into_iter()
+            .filter(|account| !runtime_cooldowns.contains(account.id()))
             .map(|account| {
                 let health = self
                     .account_feedback
@@ -195,6 +202,43 @@ impl GrokAccountSessionSelector {
             Err(GrokSessionSelectorError::NoEligibleSession)
         }
     }
+
+    async fn runtime_cooldowns(
+        &self,
+        accounts: &[ProviderAccount],
+    ) -> std::collections::BTreeSet<ProviderAccountId> {
+        let now = SystemTime::now();
+        let mut cooled = std::collections::BTreeSet::new();
+        for account in accounts {
+            let Ok(Some(cooldown)) = self.cooldowns.read(account.id()).await else {
+                continue;
+            };
+            if cooldown.credential_revision() != account.revision() {
+                if cooldown.credential_revision() < account.revision() {
+                    let _ = self.cooldowns.clear(account.id(), account.revision()).await;
+                }
+                continue;
+            }
+            if cooldown.until() > now {
+                cooled.insert(account.id().clone());
+            }
+        }
+        cooled
+    }
+
+    async fn record_runtime_cooldown(&self, session: &SelectedGrokSession, duration: Duration) {
+        let cooldown_until = SystemTime::now()
+            .checked_add(duration)
+            .unwrap_or_else(SystemTime::now);
+        let _ = self
+            .cooldowns
+            .put_if_later(ProviderCooldown::new(
+                session.account_id().clone(),
+                session.credential_revision(),
+                cooldown_until,
+            ))
+            .await;
+    }
 }
 
 impl fmt::Debug for GrokAccountSessionSelector {
@@ -205,6 +249,7 @@ impl fmt::Debug for GrokAccountSessionSelector {
             .field("catalog_cache", &"[TTL_CACHE]")
             .field("quota", &self.quota)
             .field("scheduling", &"[SCHEDULING_PORT]")
+            .field("cooldowns", &"[COOLDOWN_PORT]")
             .finish()
     }
 }
@@ -221,12 +266,12 @@ impl GrokSessionSelector for GrokAccountSessionSelector {
     ) -> GrokCredentialFeedbackFuture<'a> {
         Box::pin(async move {
             let observed_at = Utc::now();
-            let (availability, reason, cooldown_until) = match failure {
-                GrokCredentialFailure::Unauthorized => (
-                    GrokCredentialAvailability::Invalid,
-                    "upstream_unauthorized",
-                    None,
-                ),
+            let persistent = match failure {
+                GrokCredentialFailure::Unauthorized => {
+                    self.record_runtime_cooldown(session, UNAUTHORIZED_COOLDOWN)
+                        .await;
+                    return;
+                }
                 GrokCredentialFailure::QuotaExhausted => (
                     GrokCredentialAvailability::QuotaExhausted,
                     "upstream_quota_exhausted",
@@ -236,26 +281,16 @@ impl GrokSessionSelector for GrokAccountSessionSelector {
                     let retry_after = retry_after
                         .unwrap_or(DEFAULT_RATE_LIMIT_COOLDOWN)
                         .min(MAX_RATE_LIMIT_COOLDOWN);
-                    let cooldown_until = chrono::Duration::from_std(retry_after)
-                        .ok()
-                        .and_then(|duration| observed_at.checked_add_signed(duration));
-                    (
-                        GrokCredentialAvailability::Cooldown,
-                        "upstream_rate_limited",
-                        cooldown_until,
-                    )
+                    self.record_runtime_cooldown(session, retry_after).await;
+                    return;
                 }
                 GrokCredentialFailure::StreamInterrupted => {
-                    let cooldown_until = chrono::Duration::from_std(STREAM_INTERRUPTION_COOLDOWN)
-                        .ok()
-                        .and_then(|duration| observed_at.checked_add_signed(duration));
-                    (
-                        GrokCredentialAvailability::Cooldown,
-                        "upstream_stream_interrupted",
-                        cooldown_until,
-                    )
+                    self.record_runtime_cooldown(session, STREAM_INTERRUPTION_COOLDOWN)
+                        .await;
+                    return;
                 }
             };
+            let (availability, reason, cooldown_until) = persistent;
             let _ = self
                 .repository
                 .update_state(&UpdateGrokCredentialState {

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

@@ -60,12 +60,13 @@ pub use transport::{
     GrokCatalogModel, GrokClientIdentity, GrokCompactionDecodeError, GrokCompactionRequest,
     GrokCompactionSummaryDecoder, GrokCredentialFailure, GrokCredentialFeedbackFuture,
     GrokDnsResolutionError, GrokDnsResolutionPlan, GrokDnsResolutionPolicy, GrokEndpointPolicy,
-    GrokHeader, GrokHeaderValue, GrokInferenceChunkStream, GrokInferenceRequest,
+    GrokHeader, GrokHeaderValue, GrokInferenceChunkStream, GrokInferenceClientCacheStatus,
+    GrokInferenceDnsObservation, GrokInferenceDnsSource, GrokInferenceRequest,
     GrokInferenceResponse, GrokInferenceTransport, GrokInferenceTransportError,
-    GrokInferenceTransportErrorKind, GrokInferenceTransportFuture, GrokModelCatalogClient,
-    GrokModelCatalogError, GrokModelCatalogRequest, GrokModelCatalogSession,
-    GrokModelCatalogSessionError, GrokModelCatalogSnapshot, GrokModelCatalogTransport,
-    GrokModelCatalogTransportError, GrokModelCatalogTransportErrorKind,
+    GrokInferenceTransportErrorKind, GrokInferenceTransportFuture, GrokInferenceTransportMetrics,
+    GrokModelCatalogClient, GrokModelCatalogError, GrokModelCatalogRequest,
+    GrokModelCatalogSession, GrokModelCatalogSessionError, GrokModelCatalogSnapshot,
+    GrokModelCatalogTransport, GrokModelCatalogTransportError, GrokModelCatalogTransportErrorKind,
     GrokModelCatalogTransportFuture, GrokModelCatalogTransportResponse, GrokProviderConfigError,
     GrokProviderTransport, GrokRequestEncodeError, GrokReqwestTransportBuildError,
     GrokResponsesRequest, GrokSessionAffinityKey, GrokSessionBinding, GrokSessionDataError,
@@ -99,10 +100,12 @@ pub async fn initialize(
             OfficialGrokCliReleaseTransport::new().map_err(|_| XaiInitializeError::Transport)?,
         ),
     ));
+    let cli_release_status = cli_release.status();
     let provider_kind =
         ProviderKind::new(XAI_PROVIDER_NAME).map_err(|_| XaiInitializeError::ProviderKind)?;
     let accounts: Arc<dyn ProviderAccountStore> = ports.accounts();
     let leases = ports.leases();
+    let cooldowns = ports.cooldowns();
     let account_feedback = ports.account_feedback();
     let runtime_policy = ports.runtime_policy();
     let repository = GrokCredentialRepository::new(Arc::clone(&accounts));
@@ -135,6 +138,7 @@ pub async fn initialize(
         catalog_cache,
         Arc::clone(&quota),
         Arc::clone(&leases),
+        Arc::clone(&cooldowns),
         Arc::clone(&account_feedback),
     ));
     let inference: Arc<dyn GrokInferenceTransport> = Arc::new(
@@ -162,6 +166,7 @@ pub async fn initialize(
         refresher,
         Arc::clone(&catalog),
         Arc::clone(&leases),
+        cooldowns,
         Arc::clone(&runtime_policy),
     ));
     let credential_recovery: Arc<dyn GrokCredentialRecovery> = refresh.clone();
@@ -190,6 +195,7 @@ pub async fn initialize(
             catalog: Arc::clone(&catalog),
             runtime_policy,
         },
+        cli_release_status,
     ));
     let worker_contributions = provider::worker_contributions(
         refresh,

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

@@ -18,13 +18,19 @@ use gateway_core::engine::provider::{
     UpstreamTransport,
 };
 use gateway_core::engine::{AttemptContext, ContinuationAttempt, UpstreamSendState};
-use gateway_core::error::{ContinuationFailure, ProviderError, ProviderErrorKind};
-use gateway_core::event::{GatewayEvent, ProviderEvent, ProviderResponseObservation, ResponseMeta};
+use gateway_core::error::{
+    ClientVisibleUpstreamError, ContinuationFailure, ProviderError, ProviderErrorKind,
+};
+use gateway_core::event::{
+    GatewayEvent, ProviderEvent, ProviderResponseMetadata, ProviderResponseObservation,
+    ProviderResponseTimings, ResponseMeta,
+};
 use gateway_core::operation::{
     Feature, GenerateRequest, Operation, OperationKind, ProviderSessionState,
 };
 use gateway_core::routing::{
-    ModelCapabilities, ProviderCandidate, ProviderKind, SupportLevel, UpstreamModelId,
+    ModelCapabilities, ModelPresentation, ProviderCandidate, ProviderKind, SupportLevel,
+    UpstreamModelId,
 };
 use gateway_core::task::{
     ScheduledTask, WorkerContribution, WorkerCycleContext, WorkerDefinitionError, WorkerId,
@@ -35,7 +41,6 @@ use serde::{Deserialize, Serialize};
 use serde_json::{Map, Value};
 use url::Url;
 
-use crate::GrokCatalogCapabilityEvidence;
 use crate::XaiWireProfileState;
 use crate::credential::{
     GrokCredentialCatalogError, GrokCredentialCatalogService, GrokCredentialQuotaService,
@@ -50,12 +55,14 @@ use crate::transport::{
     GROK_RESPONSES_URL, GrokCompactionDecodeError, GrokCompactionRequest,
     GrokCompactionSummaryDecoder, GrokCredentialFailure, GrokInferenceChunkStream,
     GrokInferenceRequest, GrokInferenceResponse, GrokInferenceTransport,
-    GrokInferenceTransportError, GrokInferenceTransportErrorKind, GrokProviderConfigError,
-    GrokRequestEncodeError, GrokResponsesRequest, GrokSessionAffinityKey, GrokSessionSelection,
-    GrokSessionSelector, GrokSessionSelectorError, SelectedGrokSession,
+    GrokInferenceTransportError, GrokInferenceTransportErrorKind, GrokInferenceTransportMetrics,
+    GrokProviderConfigError, GrokRequestEncodeError, GrokResponsesRequest, GrokSessionAffinityKey,
+    GrokSessionSelection, GrokSessionSelector, GrokSessionSelectorError, SelectedGrokSession,
 };
+use crate::{GrokCatalogCapabilityEvidence, GrokCatalogModel};
 
 const HTTP_SSE_TRANSPORT: &str = "http_sse";
+const DEFAULT_GROK_MODEL: &str = "grok-4.5";
 const XAI_SESSION_STATE_MAX_BYTES: usize = 8 * 1024 * 1024;
 const XAI_SESSION_OUTPUT_LIMIT: usize = 4_096;
 const MIN_REASONING_CIPHERTEXT_BYTES: usize = 50;
@@ -138,38 +145,15 @@ impl Provider for GrokBuildProvider {
     async fn query_model_capabilities(
         &self,
     ) -> Result<Vec<ProviderModelCapabilities>, ProviderError> {
-        let models = self.catalog.query_models().await.map_err(|_| {
-            provider_error(ProviderErrorKind::Unavailable, UpstreamSendState::NotSent)
-        })?;
+        let Ok(models) = self.catalog.query_models().await else {
+            return Ok(vec![default_grok_model_capabilities()?]);
+        };
+        if models.is_empty() {
+            return Ok(vec![default_grok_model_capabilities()?]);
+        }
         Ok(models
             .into_iter()
-            .map(|model| {
-                let mut operations = BTreeSet::new();
-                if model.capabilities().responses_api()
-                    == GrokCatalogCapabilityEvidence::DeclaredNative
-                {
-                    operations.insert(OperationKind::Generate);
-                }
-                let capabilities = ModelCapabilities::new(
-                    operations,
-                    model
-                        .limits()
-                        .max_output_tokens()
-                        .map(std::num::NonZeroU64::get),
-                )
-                .with_feature(
-                    Feature::Reasoning,
-                    support(model.capabilities().reasoning_effort()),
-                )
-                .with_feature(
-                    Feature::Tools,
-                    tool_support(model.capabilities().streaming_tool_calls()),
-                )
-                .with_feature(Feature::Vision, SupportLevel::Unknown)
-                .with_feature(Feature::JsonSchema, SupportLevel::Unknown)
-                .with_feature(Feature::NativeContinuation, SupportLevel::Native);
-                ProviderModelCapabilities::new(model.request_model().clone(), capabilities)
-            })
+            .map(compile_grok_model_capabilities)
             .collect())
     }
 
@@ -403,6 +387,112 @@ fn tool_support(evidence: GrokCatalogCapabilityEvidence) -> SupportLevel {
     }
 }
 
+fn compile_grok_model_capabilities(model: GrokCatalogModel) -> ProviderModelCapabilities {
+    let mut operations = BTreeSet::new();
+    if model.capabilities().responses_api() == GrokCatalogCapabilityEvidence::DeclaredNative {
+        operations.insert(OperationKind::Generate);
+    }
+    let capabilities = ModelCapabilities::new(
+        operations,
+        model
+            .limits()
+            .max_output_tokens()
+            .map(std::num::NonZeroU64::get),
+    )
+    .with_feature(
+        Feature::Reasoning,
+        support(model.capabilities().reasoning_effort()),
+    )
+    .with_feature(
+        Feature::Tools,
+        tool_support(model.capabilities().streaming_tool_calls()),
+    )
+    .with_feature(Feature::Vision, SupportLevel::Unknown)
+    .with_feature(Feature::JsonSchema, SupportLevel::Unknown)
+    .with_feature(Feature::NativeContinuation, SupportLevel::Native);
+    ProviderModelCapabilities::new(model.request_model().clone(), capabilities)
+        .with_presentation(grok_model_presentation(&model))
+}
+
+fn default_grok_model_capabilities() -> Result<ProviderModelCapabilities, ProviderError> {
+    let model = UpstreamModelId::new(DEFAULT_GROK_MODEL.to_owned())
+        .map_err(|_| provider_error(ProviderErrorKind::Protocol, UpstreamSendState::NotSent))?;
+    let capabilities = ModelCapabilities::new(BTreeSet::from([OperationKind::Generate]), None)
+        .with_upstream_feature_validation();
+    Ok(ProviderModelCapabilities::new(model, capabilities)
+        .with_presentation(default_grok_model_presentation()))
+}
+
+fn default_grok_model_presentation() -> ModelPresentation {
+    ModelPresentation::new(
+        Some("Grok 4.5".to_owned()),
+        Some("xAI Grok 4.5 frontier model with reasoning and vision.".to_owned()),
+    )
+    .with_reasoning(
+        Some("medium".to_owned()),
+        ["low", "medium", "high"].map(str::to_owned).to_vec(),
+    )
+    .with_context_window_tokens(Some(500_000))
+    .with_image_input(true)
+    .with_agent_tools(true, true)
+}
+
+fn grok_model_presentation(model: &GrokCatalogModel) -> ModelPresentation {
+    let slug = model.request_model().as_str();
+    let known_grok_4_5 = matches!(
+        slug,
+        DEFAULT_GROK_MODEL | "grok-4.5-latest" | "grok-4.5-build-free" | "grok-build-latest"
+    );
+    let reasoning_evidence = model.capabilities().reasoning_effort();
+    let (default_reasoning, reasoning_efforts) = match reasoning_evidence {
+        GrokCatalogCapabilityEvidence::DeclaredNative => (
+            Some("medium".to_owned()),
+            ["low", "medium", "high"].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(),
+        ),
+        GrokCatalogCapabilityEvidence::Unknown => (None, Vec::new()),
+    };
+    let context_window_tokens = model
+        .limits()
+        .context_window_tokens()
+        .map(std::num::NonZeroU64::get)
+        .or(known_grok_4_5.then_some(500_000));
+    let tool_evidence = model.capabilities().streaming_tool_calls();
+
+    ModelPresentation::new(
+        model
+            .display_name()
+            .map(str::to_owned)
+            .or_else(|| known_grok_4_5.then(|| "Grok 4.5".to_owned())),
+        model
+            .metadata()
+            .description()
+            .map(str::to_owned)
+            .or_else(|| {
+                known_grok_4_5
+                    .then(|| "xAI Grok 4.5 frontier model with reasoning and vision.".to_owned())
+            }),
+    )
+    .with_reasoning(default_reasoning, reasoning_efforts)
+    .with_context_window_tokens(context_window_tokens)
+    .with_image_input(known_grok_4_5)
+    .with_agent_tools(
+        tool_evidence != GrokCatalogCapabilityEvidence::DeclaredUnsupported,
+        tool_evidence == GrokCatalogCapabilityEvidence::DeclaredNative
+            || (known_grok_4_5 && tool_evidence == GrokCatalogCapabilityEvidence::Unknown),
+    )
+    .with_search_tool(
+        model.capabilities().backend_search() == GrokCatalogCapabilityEvidence::DeclaredNative,
+    )
+    .with_hidden(model.metadata().hidden().unwrap_or(false))
+}
+
 #[derive(Clone, Serialize, Deserialize)]
 struct XaiSessionState {
     account_id: String,
@@ -692,16 +782,12 @@ fn attach_xai_session_update(
             .into_iter()
             .map(XaiReplayItem::ClientInput),
     );
-    transcript.extend(
-        capture
-            .output_items
-            .into_values()
-            .filter_map(|item| portable_output_item(item, false))
-            .map(|item| XaiReplayItem::AccountOutput {
-                account_id: capture.account_id.clone(),
-                item,
-            }),
-    );
+    transcript.extend(capture.output_items.into_values().map(|item| {
+        XaiReplayItem::AccountOutput {
+            account_id: capture.account_id.clone(),
+            item,
+        }
+    }));
     let state = XaiSessionState {
         account_id: capture.account_id,
         session_id: capture.session_id,
@@ -722,7 +808,11 @@ fn capture_output_item(event: &ProviderEvent, capture: &mut GrokSessionCapture)
         .or_else(|| wire.data().get("type").and_then(Value::as_str));
     if event_type == Some("response.output_item.done")
         && capture.output_items.len() < XAI_SESSION_OUTPUT_LIMIT
-        && let Some(item) = wire.data().get("item").cloned()
+        && let Some(item) = wire
+            .data()
+            .get("item")
+            .cloned()
+            .and_then(|item| portable_output_item(item, false))
     {
         let index = wire
             .data()
@@ -743,16 +833,17 @@ fn capture_output_item(event: &ProviderEvent, capture: &mut GrokSessionCapture)
         .and_then(Value::as_array)
         .filter(|output| !output.is_empty())
     {
-        capture.output_items.clear();
-        capture.output_items.extend(
-            output
-                .iter()
-                .take(XAI_SESSION_OUTPUT_LIMIT)
-                .enumerate()
-                .filter_map(|(index, item)| {
-                    u32::try_from(index).ok().map(|index| (index, item.clone()))
-                }),
-        );
+        for (index, item) in output.iter().take(XAI_SESSION_OUTPUT_LIMIT).enumerate() {
+            let Some(index) = u32::try_from(index).ok() else {
+                break;
+            };
+            if capture.output_items.contains_key(&index) {
+                continue;
+            }
+            if let Some(item) = portable_output_item(item.clone(), false) {
+                capture.output_items.insert(index, item);
+            }
+        }
     }
 }
 
@@ -862,17 +953,11 @@ async fn start_grok_inference(
             return Err(GrokInferenceStartFailure { observation, error });
         }
     };
-    let mut observation = ProviderResponseObservation::new(
-        UpstreamTransport::new(HTTP_SSE_TRANSPORT).map_err(|_| GrokInferenceStartFailure {
+    let observation =
+        xai_response_observation(&response).map_err(|error| GrokInferenceStartFailure {
             observation: None,
-            error: protocol_sent(),
-        })?,
-    )
-    .with_http_version(response.http_version())
-    .with_status_code(response.status_code());
-    if let Some(request_id) = response.request_id().cloned() {
-        observation = observation.with_request_id(request_id);
-    }
+            error,
+        })?;
     Ok(AcceptedGrokInference {
         response,
         observation,
@@ -1189,17 +1274,7 @@ fn cold_http_sse_stream(
             }
         };
 
-        let mut observation = ProviderResponseObservation::new(
-            UpstreamTransport::new(HTTP_SSE_TRANSPORT).map_err(|_| provider_error(
-                ProviderErrorKind::Protocol,
-                UpstreamSendState::Sent,
-            ))?,
-        )
-        .with_http_version(response.http_version())
-        .with_status_code(response.status_code());
-        if let Some(request_id) = response.request_id().cloned() {
-            observation = observation.with_request_id(request_id);
-        }
+        let observation = xai_response_observation(&response)?;
         yield ProviderEvent::observation(observation);
 
         let mut body = response.into_body();
@@ -1295,7 +1370,68 @@ fn xai_error_observation(
     if let Some(request_id) = error.request_id().cloned() {
         observation = observation.with_request_id(request_id);
     }
-    Ok(observation)
+    Ok(with_xai_transport_metrics(
+        observation,
+        error.transport_metrics(),
+    ))
+}
+
+fn xai_response_observation(
+    response: &GrokInferenceResponse,
+) -> Result<ProviderResponseObservation, ProviderError> {
+    let mut observation = ProviderResponseObservation::new(
+        UpstreamTransport::new(HTTP_SSE_TRANSPORT)
+            .map_err(|_| provider_error(ProviderErrorKind::Protocol, UpstreamSendState::Sent))?,
+    )
+    .with_http_version(response.http_version())
+    .with_status_code(response.status_code());
+    if let Some(request_id) = response.request_id().cloned() {
+        observation = observation.with_request_id(request_id);
+    }
+    Ok(with_xai_transport_metrics(
+        observation,
+        response.transport_metrics(),
+    ))
+}
+
+fn with_xai_transport_metrics(
+    mut observation: ProviderResponseObservation,
+    metrics: GrokInferenceTransportMetrics,
+) -> ProviderResponseObservation {
+    observation = observation.with_timings(ProviderResponseTimings {
+        headers_ms: metrics.headers_ms(),
+        ..ProviderResponseTimings::default()
+    });
+    if let Some(metadata) = xai_transport_metadata(metrics) {
+        observation = observation.with_provider_metadata(metadata);
+    }
+    observation
+}
+
+fn xai_transport_metadata(
+    metrics: GrokInferenceTransportMetrics,
+) -> Option<ProviderResponseMetadata> {
+    let mut metadata = Map::new();
+    if let Some(status) = metrics.client_cache_status() {
+        metadata.insert(
+            "clientCache".to_owned(),
+            Value::String(status.as_str().to_owned()),
+        );
+    }
+    if let Some(dns) = metrics.dns() {
+        metadata.insert(
+            "dnsSource".to_owned(),
+            Value::String(dns.source().as_str().to_owned()),
+        );
+        metadata.insert("dnsMs".to_owned(), Value::from(dns.duration_ms()));
+    }
+    if let Some(headers_ms) = metrics.headers_ms() {
+        metadata.insert("upstreamHeadersMs".to_owned(), Value::from(headers_ms));
+    }
+    if metadata.is_empty() {
+        return None;
+    }
+    ProviderResponseMetadata::new(serde_json::to_string(&Value::Object(metadata)).ok()?)
 }
 
 async fn record_failure(
@@ -1351,8 +1487,10 @@ async fn recover_or_record_failure(
             .await
         {
             GrokCredentialRecoveryOutcome::Recovered => error.with_same_account_retry(),
-            GrokCredentialRecoveryOutcome::Rejected
-            | GrokCredentialRecoveryOutcome::Unavailable => error,
+            GrokCredentialRecoveryOutcome::Rejected => error,
+            GrokCredentialRecoveryOutcome::Unavailable => {
+                record_failure(selector, session, error).await
+            }
         };
     }
     record_failure(selector, session, error).await
@@ -1415,10 +1553,23 @@ fn ensure_sent_context(context: &AttemptContext) -> Result<(), ProviderError> {
 fn map_request_error(error: GrokRequestEncodeError) -> ProviderError {
     let kind = match error {
         GrokRequestEncodeError::InvalidProtocolPayload
-        | GrokRequestEncodeError::InvalidRequestNormalization => ProviderErrorKind::InvalidRequest,
+        | GrokRequestEncodeError::InvalidRequestNormalization
+        | GrokRequestEncodeError::InvalidRequestField { .. } => ProviderErrorKind::InvalidRequest,
         GrokRequestEncodeError::Serialization => ProviderErrorKind::Protocol,
     };
-    provider_error(kind, UpstreamSendState::NotSent)
+    let provider_error = provider_error(kind, UpstreamSendState::NotSent);
+    if kind != ProviderErrorKind::InvalidRequest {
+        return provider_error;
+    }
+    let detail = ClientVisibleUpstreamError::new(
+        error.to_string(),
+        Some("invalid_request_normalization".to_owned()),
+        Some("invalid_request_error".to_owned()),
+    );
+    match detail {
+        Ok(detail) => provider_error.with_client_visible_upstream_error(detail),
+        Err(_) => provider_error,
+    }
 }
 
 fn map_selection_error(error: GrokSessionSelectorError) -> ProviderError {

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

@@ -49,6 +49,8 @@ pub use session::{
     GrokSessionSelector, GrokSessionSelectorError, GrokSessionSelectorFuture, SelectedGrokSession,
 };
 pub use transport::{
-    GrokInferenceChunkStream, GrokInferenceRequest, GrokInferenceResponse, GrokInferenceTransport,
+    GrokInferenceChunkStream, GrokInferenceClientCacheStatus, GrokInferenceDnsObservation,
+    GrokInferenceDnsSource, GrokInferenceRequest, GrokInferenceResponse, GrokInferenceTransport,
     GrokInferenceTransportError, GrokInferenceTransportErrorKind, GrokInferenceTransportFuture,
+    GrokInferenceTransportMetrics,
 };

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

@@ -4,8 +4,8 @@ use std::collections::{HashMap, VecDeque};
 use std::fmt;
 use std::io;
 use std::net::{IpAddr, Ipv4Addr, Ipv6Addr, SocketAddr};
-use std::sync::{Arc, Mutex};
-use std::time::Duration;
+use std::sync::{Arc, Mutex, OnceLock};
+use std::time::{Duration, Instant};
 
 use futures::StreamExt;
 use reqwest::dns::{Addrs, Name, Resolve, Resolving};
@@ -20,13 +20,14 @@ use crate::credential::discovery::MAX_OAUTH_RESPONSE_BYTES;
 use crate::{
     GrokBillingRequest, GrokBillingTransport, GrokBillingTransportError,
     GrokBillingTransportErrorKind, GrokBillingTransportFuture, GrokBillingTransportResponse,
+    GrokInferenceClientCacheStatus, GrokInferenceDnsObservation, GrokInferenceDnsSource,
     GrokInferenceRequest, GrokInferenceResponse, GrokInferenceTransport,
     GrokInferenceTransportError, GrokInferenceTransportErrorKind, GrokInferenceTransportFuture,
-    GrokModelCatalogRequest, GrokModelCatalogTransport, GrokModelCatalogTransportError,
-    GrokModelCatalogTransportErrorKind, GrokModelCatalogTransportFuture,
-    GrokModelCatalogTransportResponse, GrokSessionBinding, HttpMethod, MAX_GROK_BILLING_BYTES,
-    MAX_GROK_MODEL_CATALOG_BYTES, OAuthHttpRequest, OAuthHttpResponse, OAuthHttpTransport,
-    TransportFailure, TransportFailureKind, TransportFuture,
+    GrokInferenceTransportMetrics, GrokModelCatalogRequest, GrokModelCatalogTransport,
+    GrokModelCatalogTransportError, GrokModelCatalogTransportErrorKind,
+    GrokModelCatalogTransportFuture, GrokModelCatalogTransportResponse, GrokSessionBinding,
+    HttpMethod, MAX_GROK_BILLING_BYTES, MAX_GROK_MODEL_CATALOG_BYTES, OAuthHttpRequest,
+    OAuthHttpResponse, OAuthHttpTransport, TransportFailure, TransportFailureKind, TransportFuture,
 };
 use gateway_core::engine::UpstreamSendState;
 use gateway_core::error::SafeUpstreamValue;
@@ -42,6 +43,9 @@ pub(crate) const OFFICIAL_USERINFO_PATH: &str = "/oauth2/userinfo";
 const CONNECT_TIMEOUT: Duration = Duration::from_secs(10);
 const OAUTH_REQUEST_TIMEOUT: Duration = Duration::from_secs(30);
 const POOL_IDLE_TIMEOUT: Duration = Duration::from_secs(90);
+const POOL_MAX_IDLE_PER_HOST: usize = 2;
+const HTTP2_KEEP_ALIVE_INTERVAL: Duration = Duration::from_secs(15);
+const HTTP2_KEEP_ALIVE_TIMEOUT: Duration = Duration::from_secs(5);
 const MAX_ERROR_BODY_BYTES: usize = 64 * 1024;
 const MAX_INFERENCE_BODY_BYTES: usize = 256 * 1024 * 1024;
 const MAX_RETRY_AFTER_SECONDS: u64 = 120;
@@ -55,6 +59,10 @@ const TRUSTED_DOH_BOOTSTRAP: [SocketAddr; 2] = [
     SocketAddr::new(IpAddr::V4(Ipv4Addr::new(8, 8, 4, 4)), 443),
 ];
 
+tokio::task_local! {
+    static REQUEST_DNS_OBSERVER: Arc<RequestDnsObserver>;
+}
+
 /// 构建严格 reqwest transport 失败。
 #[derive(Debug, thiserror::Error, Clone, Copy, PartialEq, Eq)]
 pub enum GrokReqwestTransportBuildError {
@@ -314,12 +322,31 @@ impl ReqwestGrokInferenceTransport {
     fn client_for(
         &self,
         binding: &GrokSessionBinding,
-    ) -> Result<Client, GrokInferenceTransportError> {
-        self.clients
+    ) -> Result<(Client, GrokInferenceClientCacheStatus), GrokInferenceTransportError> {
+        let unbound_client = {
+            let mut clients = self
+                .clients
+                .lock()
+                .map_err(|_| inference_client_pool_unavailable())?;
+            if let Some(client) = clients.get(binding) {
+                return Ok((client, GrokInferenceClientCacheStatus::Hit));
+            }
+            clients.take_unbound()
+        };
+
+        let client = match unbound_client {
+            Some(client) => client,
+            None => self
+                .endpoint_policy
+                .build_inference_client(None)
+                .map_err(|_| inference_client_pool_unavailable_for_cache_miss())?,
+        };
+        let client = self
+            .clients
             .lock()
-            .map_err(|_| inference_client_pool_unavailable())?
-            .get_or_insert(binding, self.endpoint_policy.as_ref())
-            .map_err(|_| inference_client_pool_unavailable())
+            .map_err(|_| inference_client_pool_unavailable_for_cache_miss())?
+            .insert_if_absent(binding, client);
+        Ok((client, GrokInferenceClientCacheStatus::Miss))
     }
 }
 
@@ -342,53 +369,62 @@ impl GrokInferenceTransport for ReqwestGrokInferenceTransport {
                     UpstreamSendState::NotSent,
                 ));
             }
-            let client = self.client_for(request.binding())?;
+            let (client, client_cache_status) = self.client_for(request.binding())?;
             let mut builder = client
                 .post(request.endpoint().clone())
                 .body(request.body().to_vec());
             for header in request.headers() {
                 builder = builder.header(header.name(), header.value().expose());
             }
-            let response = builder
-                .send()
+            let dns_observer = Arc::new(RequestDnsObserver::default());
+            let headers_started_at = Instant::now();
+            let response = REQUEST_DNS_OBSERVER
+                .scope(Arc::clone(&dns_observer), builder.send())
                 .await
-                .map_err(classify_inference_reqwest_error)?;
+                .map_err(|error| {
+                    classify_inference_reqwest_error(error).with_transport_metrics(
+                        inference_transport_metrics(client_cache_status, None, &dns_observer),
+                    )
+                })?;
+            let transport_metrics = inference_transport_metrics(
+                client_cache_status,
+                Some(elapsed_millis(headers_started_at.elapsed())),
+                &dns_observer,
+            );
             if !response.status().is_success() {
-                return Err(classify_inference_status(response).await);
-            }
-            if !is_event_stream(&response) {
-                return Err(GrokInferenceTransportError::new(
-                    GrokInferenceTransportErrorKind::Protocol,
-                    UpstreamSendState::Sent,
-                ));
+                return Err(classify_inference_status(response)
+                    .await
+                    .with_transport_metrics(transport_metrics));
             }
             let http_version = upstream_http_version(response.version());
             let status_code = response.status().as_u16();
             let request_id = upstream_request_id(&response);
-            let body = response.bytes_stream().scan(0_usize, |observed, chunk| {
-                let item = match chunk {
-                    Ok(chunk)
-                        if observed
-                            .checked_add(chunk.len())
-                            .is_some_and(|total| total <= MAX_INFERENCE_BODY_BYTES) =>
-                    {
-                        *observed += chunk.len();
-                        Ok(chunk.to_vec())
-                    }
-                    Ok(_) => Err(GrokInferenceTransportError::new(
-                        GrokInferenceTransportErrorKind::Protocol,
-                        UpstreamSendState::Sent,
-                    )),
-                    Err(error) => Err(classify_inference_stream_error(&error)),
-                };
-                std::future::ready(Some(item))
-            });
-            Ok(GrokInferenceResponse::new(
-                Box::pin(body),
-                http_version,
-                status_code,
-                request_id,
-            ))
+            let body = response
+                .bytes_stream()
+                .scan(0_usize, move |observed, chunk| {
+                    let item = match chunk {
+                        Ok(chunk)
+                            if observed
+                                .checked_add(chunk.len())
+                                .is_some_and(|total| total <= MAX_INFERENCE_BODY_BYTES) =>
+                        {
+                            *observed += chunk.len();
+                            Ok(chunk.to_vec())
+                        }
+                        Ok(_) => Err(GrokInferenceTransportError::new(
+                            GrokInferenceTransportErrorKind::Protocol,
+                            UpstreamSendState::Sent,
+                        )
+                        .with_transport_metrics(transport_metrics)),
+                        Err(error) => Err(classify_inference_stream_error(&error)
+                            .with_transport_metrics(transport_metrics)),
+                    };
+                    std::future::ready(Some(item))
+                });
+            Ok(
+                GrokInferenceResponse::new(Box::pin(body), http_version, status_code, request_id)
+                    .with_transport_metrics(transport_metrics),
+            )
         })
     }
 }
@@ -412,22 +448,24 @@ impl BoundInferenceClients {
         }
     }
 
-    fn get_or_insert(
-        &mut self,
-        binding: &GrokSessionBinding,
-        endpoint_policy: &dyn GrokEndpointPolicy,
-    ) -> Result<Client, GrokReqwestTransportBuildError> {
+    fn get(&mut self, binding: &GrokSessionBinding) -> Option<Client> {
         if let Some(client) = self.by_binding.get(binding).cloned() {
             self.record_use(binding);
-            return Ok(client);
+            return Some(client);
         }
+        None
+    }
 
-        let client = match self.unbound_client.take() {
-            Some(client) => client,
-            None => endpoint_policy.build_inference_client(None)?,
-        };
+    fn take_unbound(&mut self) -> Option<Client> {
+        self.unbound_client.take()
+    }
+
+    fn insert_if_absent(&mut self, binding: &GrokSessionBinding, client: Client) -> Client {
+        if let Some(existing) = self.get(binding) {
+            return existing;
+        }
         self.insert(binding.clone(), client.clone());
-        Ok(client)
+        client
     }
 
     fn insert(&mut self, binding: GrokSessionBinding, client: Client) {
@@ -454,6 +492,29 @@ fn inference_client_pool_unavailable() -> GrokInferenceTransportError {
     )
 }
 
+fn inference_client_pool_unavailable_for_cache_miss() -> GrokInferenceTransportError {
+    inference_client_pool_unavailable().with_transport_metrics(
+        GrokInferenceTransportMetrics::default()
+            .with_client_cache_status(GrokInferenceClientCacheStatus::Miss),
+    )
+}
+
+fn inference_transport_metrics(
+    client_cache_status: GrokInferenceClientCacheStatus,
+    headers_ms: Option<u64>,
+    dns_observer: &RequestDnsObserver,
+) -> GrokInferenceTransportMetrics {
+    let mut metrics =
+        GrokInferenceTransportMetrics::default().with_client_cache_status(client_cache_status);
+    if let Some(headers_ms) = headers_ms {
+        metrics = metrics.with_headers_ms(headers_ms);
+    }
+    if let Some(dns) = dns_observer.observation() {
+        metrics = metrics.with_dns(dns);
+    }
+    metrics
+}
+
 /// 官方 Grok CLI proxy 模型目录 GET transport。
 pub struct ReqwestGrokModelCatalogTransport {
     client: Client,
@@ -587,8 +648,11 @@ fn build_official_client(
         .no_proxy()
         .connect_timeout(CONNECT_TIMEOUT)
         .pool_idle_timeout(POOL_IDLE_TIMEOUT)
-        .pool_max_idle_per_host(128)
+        .pool_max_idle_per_host(POOL_MAX_IDLE_PER_HOST)
         .http2_adaptive_window(true)
+        .http2_keep_alive_interval(HTTP2_KEEP_ALIVE_INTERVAL)
+        .http2_keep_alive_timeout(HTTP2_KEEP_ALIVE_TIMEOUT)
+        .http2_keep_alive_while_idle(true)
         .tcp_nodelay(true)
         .https_only(true)
         .dns_resolver(Arc::new(StrictDnsResolver::new(dns_policy)?));
@@ -632,6 +696,24 @@ struct StrictDnsResolver {
     trusted_doh: TrustedDohResolver,
 }
 
+#[derive(Debug, Default)]
+struct RequestDnsObserver {
+    observation: OnceLock<GrokInferenceDnsObservation>,
+}
+
+impl RequestDnsObserver {
+    fn record(&self, source: GrokInferenceDnsSource, duration: Duration) {
+        let _ = self.observation.set(GrokInferenceDnsObservation::new(
+            source,
+            elapsed_millis(duration),
+        ));
+    }
+
+    fn observation(&self) -> Option<GrokInferenceDnsObservation> {
+        self.observation.get().copied()
+    }
+}
+
 impl StrictDnsResolver {
     fn new(policy: GrokDnsResolutionPolicy) -> Result<Self, GrokReqwestTransportBuildError> {
         Ok(Self {
@@ -653,7 +735,9 @@ impl Resolve for StrictDnsResolver {
         }
         let policy = self.policy;
         let trusted_doh = self.trusted_doh.clone();
+        let observer = REQUEST_DNS_OBSERVER.try_with(Arc::clone).ok();
         Box::pin(async move {
+            let started_at = Instant::now();
             let system_addresses = match tokio::net::lookup_host((requested_host.as_str(), 0)).await
             {
                 Ok(addresses) => addresses.collect::<Vec<_>>(),
@@ -663,29 +747,43 @@ impl Resolve for StrictDnsResolver {
                 .iter()
                 .map(SocketAddr::ip)
                 .collect::<Vec<_>>();
-            match policy
+            let (source, result) = match policy
                 .plan_system_resolution(&requested_host, &system_ips)
                 .map_err(|_| safe_dns_error("DNS resolution rejected"))?
             {
-                GrokDnsResolutionPlan::System => {
-                    Ok(Box::new(system_addresses.into_iter()) as Addrs)
-                }
+                GrokDnsResolutionPlan::System => (
+                    GrokInferenceDnsSource::System,
+                    Ok(Box::new(system_addresses.into_iter()) as Addrs),
+                ),
                 GrokDnsResolutionPlan::TrustedDoh => {
-                    let addresses = trusted_doh.resolve(&requested_host).await?;
-                    policy
-                        .validate_trusted_doh_resolution(&requested_host, &addresses)
-                        .map_err(|_| safe_dns_error("trusted DNS result rejected"))?;
-                    Ok(Box::new(
-                        addresses
-                            .into_iter()
-                            .map(|address| SocketAddr::new(address, 0)),
-                    ) as Addrs)
+                    let result = match trusted_doh.resolve(&requested_host).await {
+                        Ok(addresses) => policy
+                            .validate_trusted_doh_resolution(&requested_host, &addresses)
+                            .map_err(|_| safe_dns_error("trusted DNS result rejected"))
+                            .map(|()| {
+                                Box::new(
+                                    addresses
+                                        .into_iter()
+                                        .map(|address| SocketAddr::new(address, 0)),
+                                ) as Addrs
+                            }),
+                        Err(error) => Err(error),
+                    };
+                    (GrokInferenceDnsSource::TrustedDoh, result)
                 }
+            };
+            if let Some(observer) = observer {
+                observer.record(source, started_at.elapsed());
             }
+            result
         })
     }
 }
 
+fn elapsed_millis(duration: Duration) -> u64 {
+    u64::try_from(duration.as_millis()).unwrap_or(u64::MAX)
+}
+
 #[derive(Clone)]
 struct TrustedDohResolver {
     client: Client,
@@ -1109,15 +1207,6 @@ fn retry_after(response: &Response) -> Option<Duration> {
         .map(Duration::from_secs)
 }
 
-fn is_event_stream(response: &Response) -> bool {
-    response
-        .headers()
-        .get(CONTENT_TYPE)
-        .and_then(|value| value.to_str().ok())
-        .and_then(|value| value.split(';').next())
-        .is_some_and(|value| value.trim().eq_ignore_ascii_case("text/event-stream"))
-}
-
 fn is_json_response(response: &Response) -> bool {
     response
         .headers()

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

@@ -145,18 +145,82 @@ impl GrokCliReleaseTransport for OfficialGrokCliReleaseTransport {
 pub struct GrokCliReleaseService {
     profile: XaiWireProfileState,
     transport: Arc<dyn GrokCliReleaseTransport>,
+    status: GrokCliReleaseStatus,
 }
 
 impl GrokCliReleaseService {
     #[must_use]
     pub fn new(profile: XaiWireProfileState, transport: Arc<dyn GrokCliReleaseTransport>) -> Self {
-        Self { profile, transport }
+        Self {
+            profile,
+            transport,
+            status: GrokCliReleaseStatus::default(),
+        }
+    }
+
+    #[must_use]
+    pub(crate) fn status(&self) -> GrokCliReleaseStatus {
+        self.status.clone()
     }
 
     pub async fn refresh(&self) -> Result<String, GrokCliReleaseError> {
-        let version = self.transport.fetch().await?;
-        self.profile.update_client_version(&version);
-        Ok(version)
+        let checked_at = Utc::now();
+        let result = self.transport.fetch().await;
+        match result {
+            Ok(version) => {
+                self.profile.update_client_version(&version);
+                self.status.record_success(checked_at, version.clone());
+                Ok(version)
+            }
+            Err(error) => {
+                self.status.record_failure(checked_at, &error);
+                Err(error)
+            }
+        }
+    }
+}
+
+/// 最近一次官方 Grok CLI 发布检查结果。
+#[derive(Debug, Clone, Default, PartialEq, Eq)]
+pub(crate) struct GrokCliReleaseSnapshot {
+    pub(crate) checked_at: Option<DateTime<Utc>>,
+    pub(crate) latest_version: Option<String>,
+    pub(crate) last_error: Option<String>,
+}
+
+/// Provider 内共享的 Grok CLI 发布检查观察状态。
+#[derive(Debug, Clone, Default)]
+pub(crate) struct GrokCliReleaseStatus {
+    snapshot: Arc<RwLock<GrokCliReleaseSnapshot>>,
+}
+
+impl GrokCliReleaseStatus {
+    #[must_use]
+    pub(crate) fn snapshot(&self) -> GrokCliReleaseSnapshot {
+        self.snapshot
+            .read()
+            .unwrap_or_else(std::sync::PoisonError::into_inner)
+            .clone()
+    }
+
+    fn record_success(&self, checked_at: DateTime<Utc>, latest_version: String) {
+        *self
+            .snapshot
+            .write()
+            .unwrap_or_else(std::sync::PoisonError::into_inner) = GrokCliReleaseSnapshot {
+            checked_at: Some(checked_at),
+            latest_version: Some(latest_version),
+            last_error: None,
+        };
+    }
+
+    fn record_failure(&self, checked_at: DateTime<Utc>, error: &GrokCliReleaseError) {
+        let mut snapshot = self
+            .snapshot
+            .write()
+            .unwrap_or_else(std::sync::PoisonError::into_inner);
+        snapshot.checked_at = Some(checked_at);
+        snapshot.last_error = Some(error.to_string());
     }
 }
 
@@ -191,3 +255,91 @@ pub enum GrokCliReleaseError {
     #[error("Grok CLI release version is invalid")]
     InvalidVersion,
 }
+
+#[cfg(test)]
+mod tests {
+    use std::collections::VecDeque;
+    use std::sync::Mutex;
+
+    use super::*;
+
+    struct ReleaseTransport {
+        outcomes: Mutex<VecDeque<Result<String, GrokCliReleaseError>>>,
+    }
+
+    impl ReleaseTransport {
+        fn new(outcomes: impl IntoIterator<Item = Result<String, GrokCliReleaseError>>) -> Self {
+            Self {
+                outcomes: Mutex::new(outcomes.into_iter().collect()),
+            }
+        }
+    }
+
+    impl GrokCliReleaseTransport for ReleaseTransport {
+        fn fetch(&self) -> BoxFuture<'_, Result<String, GrokCliReleaseError>> {
+            let outcome = self
+                .outcomes
+                .lock()
+                .expect("release outcomes lock")
+                .pop_front()
+                .expect("release outcome");
+            Box::pin(async move { outcome })
+        }
+    }
+
+    #[tokio::test]
+    async fn release_refresh_should_update_profile_and_observation() {
+        let profile = wire_profile();
+        let service = GrokCliReleaseService::new(
+            profile.clone(),
+            Arc::new(ReleaseTransport::new([Ok("0.2.112".to_owned())])),
+        );
+        let status = service.status();
+
+        assert_eq!(service.refresh().await.expect("release"), "0.2.112");
+        assert_eq!(profile.client_version(), "0.2.112");
+        let snapshot = status.snapshot();
+        assert!(snapshot.checked_at.is_some());
+        assert_eq!(snapshot.latest_version.as_deref(), Some("0.2.112"));
+        assert!(snapshot.last_error.is_none());
+    }
+
+    #[tokio::test]
+    async fn failed_release_refresh_should_keep_last_successful_version() {
+        let profile = wire_profile();
+        let service = GrokCliReleaseService::new(
+            profile.clone(),
+            Arc::new(ReleaseTransport::new([
+                Ok("0.2.112".to_owned()),
+                Err(GrokCliReleaseError::InvalidDocument),
+            ])),
+        );
+        let status = service.status();
+
+        service.refresh().await.expect("first release");
+        assert!(matches!(
+            service.refresh().await,
+            Err(GrokCliReleaseError::InvalidDocument)
+        ));
+
+        assert_eq!(profile.client_version(), "0.2.112");
+        let snapshot = status.snapshot();
+        assert!(snapshot.checked_at.is_some());
+        assert_eq!(snapshot.latest_version.as_deref(), Some("0.2.112"));
+        assert_eq!(
+            snapshot.last_error.as_deref(),
+            Some("Grok CLI release document is invalid")
+        );
+    }
+
+    fn wire_profile() -> XaiWireProfileState {
+        XaiWireProfileState::new(XaiWireProfileConfig {
+            client_identifier: "grok-shell".to_owned(),
+            client_version: "0.2.106".to_owned(),
+            client_mode: "headless".to_owned(),
+            target_os: "linux".to_owned(),
+            target_arch: "x86_64".to_owned(),
+            verified_at: Utc::now(),
+        })
+    }
+}

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

@@ -1,4 +1,4 @@
-use std::collections::BTreeMap;
+use std::collections::{BTreeMap, BTreeSet};
 use std::fmt;
 
 use gateway_core::operation::GenerateRequest;
@@ -279,6 +279,18 @@ pub enum GrokRequestEncodeError {
     /// Responses compatibility fields could not be normalized safely.
     #[error("Grok Build request normalization failed")]
     InvalidRequestNormalization,
+    /// 请求中的具体字段无法安全转换为 Grok Build 接受的形态。
+    #[error("Grok Build request field `{field}` could not be normalized safely")]
+    InvalidRequestField { field: &'static str },
+}
+
+impl GrokRequestEncodeError {
+    fn at_field(self, field: &'static str) -> Self {
+        match self {
+            Self::InvalidRequestNormalization => Self::InvalidRequestField { field },
+            error => error,
+        }
+    }
 }
 
 fn sanitize_account_identity(body: &mut Map<String, Value>) {
@@ -548,11 +560,12 @@ fn normalize_responses_request(
         }
         let text = text
             .as_object_mut()
-            .ok_or(GrokRequestEncodeError::InvalidRequestNormalization)?;
+            .ok_or(GrokRequestEncodeError::InvalidRequestField { field: "text" })?;
         if text.get("format").is_none_or(Value::is_null) {
             text.insert(
                 "format".to_owned(),
-                normalize_response_format(response_format)?,
+                normalize_response_format(response_format)
+                    .map_err(|error| error.at_field("response_format"))?,
             );
         }
     }
@@ -711,20 +724,22 @@ impl GrokResponseTransform {
                 .stream_calls
                 .get_mut(&primary)
                 .ok_or(GrokRequestEncodeError::InvalidRequestNormalization)?;
-            if matches!(
-                state.identity.kind,
-                ToolKind::ToolSearch | ToolKind::Custom | ToolKind::ApplyPatch
-            ) {
-                state.arguments.push_str(
-                    value
-                        .get("delta")
-                        .and_then(Value::as_str)
-                        .unwrap_or_default(),
-                );
-                if state.identity.kind == ToolKind::Custom {
-                    state.last_delta = value.as_object().cloned();
+            match state.identity.kind {
+                ToolKind::Custom => {
+                    state.arguments.push_str(
+                        value
+                            .get("delta")
+                            .and_then(Value::as_str)
+                            .unwrap_or_default(),
+                    );
+                    state.last_delta = value.as_object().cloned().map(|mut payload| {
+                        payload.remove("delta");
+                        payload
+                    });
+                    return Ok(Vec::new());
                 }
-                return Ok(Vec::new());
+                ToolKind::ToolSearch | ToolKind::ApplyPatch => return Ok(Vec::new()),
+                ToolKind::Function => {}
             }
         }
 
@@ -767,6 +782,10 @@ impl GrokResponseTransform {
                             &input,
                         ),
                     });
+                    if let Some(state) = self.stream_calls.get_mut(&primary) {
+                        state.arguments.clear();
+                        state.last_delta = None;
+                    }
                     return Ok(output);
                 }
                 ToolKind::Function => {}
@@ -783,10 +802,17 @@ impl GrokResponseTransform {
             return self.rewrite_apply_patch_done_event(value);
         }
 
+        let completed_call = (event_type == "response.output_item.done")
+            .then(|| value.get("item").and_then(Value::as_object))
+            .flatten()
+            .and_then(|item| self.primary_stream_call(item));
         self.rewrite_response_value(&mut value)?;
         if let Some(response) = value.get_mut("response").and_then(Value::as_object_mut) {
             self.restore_visible_tools(response);
         }
+        if let Some(primary) = completed_call {
+            self.take_stream_call(&primary);
+        }
         Ok(vec![GrokTransformedWireEvent {
             event_type: event_type.to_owned(),
             value,
@@ -847,6 +873,18 @@ impl GrokResponseTransform {
         Some(primary)
     }
 
+    fn primary_stream_call(&self, item: &Map<String, Value>) -> Option<String> {
+        ["id", "call_id"]
+            .into_iter()
+            .filter_map(|field| item.get(field).and_then(Value::as_str))
+            .find_map(|key| self.stream_keys.get(key).cloned())
+    }
+
+    fn take_stream_call(&mut self, primary: &str) -> Option<StreamCallState> {
+        self.stream_keys.retain(|_, owner| owner != primary);
+        self.stream_calls.remove(primary)
+    }
+
     fn rewrite_response_value(&self, value: &mut Value) -> Result<(), GrokRequestEncodeError> {
         match value {
             Value::Array(values) => {
@@ -953,15 +991,11 @@ impl GrokResponseTransform {
             .and_then(Value::as_object)
             .cloned()
             .ok_or(GrokRequestEncodeError::InvalidRequestNormalization)?;
-        let primary = ["id", "call_id"]
-            .into_iter()
-            .find_map(|field| original_item.get(field).and_then(Value::as_str))
-            .and_then(|key| self.stream_keys.get(key))
-            .cloned();
-        let mut added = primary
-            .as_ref()
-            .and_then(|key| self.stream_calls.get(key))
-            .and_then(|state| state.added_payload.clone())
+        let state = self
+            .primary_stream_call(&original_item)
+            .and_then(|primary| self.take_stream_call(&primary));
+        let mut added = state
+            .and_then(|state| state.added_payload)
             .unwrap_or_else(|| {
                 Map::from_iter([(
                     "type".to_owned(),
@@ -1148,20 +1182,27 @@ impl ToolNormalizer {
         mut self,
         payload: &mut Map<String, Value>,
     ) -> Result<GrokResponseTransform, GrokRequestEncodeError> {
-        let (tools, had_tools) = optional_array(payload.get("tools"))?;
+        let (tools, had_tools) =
+            optional_array(payload.get("tools")).map_err(|error| error.at_field("tools"))?;
         if had_tools {
             self.response.visible_tools.clone_from(&tools);
         }
-        let client_search = inspect_tool_search(&tools)?;
-        self.normalize_client_search_parallel(payload, client_search)?;
+        let client_search = inspect_tool_search(&tools).map_err(|error| error.at_field("tools"))?;
+        self.normalize_client_search_parallel(payload, client_search)
+            .map_err(|error| error.at_field("parallel_tool_calls"))?;
 
         let mut normalized_tools = Vec::with_capacity(tools.len());
         for raw_tool in &tools {
-            normalized_tools.extend(self.normalize_tool(raw_tool, "", client_search, false)?);
+            normalized_tools.extend(
+                self.normalize_tool(raw_tool, "", client_search, false)
+                    .map_err(|error| error.at_field("tools"))?,
+            );
         }
 
         if let Some(Value::Array(items)) = payload.get("input") {
-            let normalized = self.normalize_input_items(items)?;
+            let normalized = self
+                .normalize_input_items(items)
+                .map_err(|error| error.at_field("input"))?;
             normalized_tools.extend(normalized.loaded_tools);
             self.response.visible_tools.extend(normalized.visible_tools);
             payload.insert("input".to_owned(), Value::Array(normalized.items));
@@ -1169,11 +1210,14 @@ impl ToolNormalizer {
             .get("input")
             .is_some_and(|input| !input.is_null() && !input.is_string())
         {
-            return Err(GrokRequestEncodeError::InvalidRequestNormalization);
+            return Err(GrokRequestEncodeError::InvalidRequestField { field: "input" });
         }
 
         if self.client_search_tool.is_some() {
-            normalized_tools.push(Value::Object(self.build_client_search_function()?));
+            normalized_tools.push(Value::Object(
+                self.build_client_search_function()
+                    .map_err(|error| error.at_field("tools"))?,
+            ));
         }
         normalized_tools = dedupe_normalized_tools(normalized_tools);
         if normalized_tools.is_empty() {
@@ -1184,7 +1228,8 @@ impl ToolNormalizer {
         } else {
             payload.insert("tools".to_owned(), Value::Array(normalized_tools.clone()));
         }
-        self.normalize_tool_choice(payload, &normalized_tools)?;
+        self.normalize_tool_choice(payload, &normalized_tools)
+            .map_err(|error| error.at_field("tool_choice"))?;
         Ok(self.response)
     }
 
@@ -1299,6 +1344,11 @@ impl ToolNormalizer {
             return Ok(Vec::new());
         }
         let mut converted = without_defer_loading(tool);
+        if let Some(parameters) = converted.get("parameters").cloned()
+            && let Some(normalized) = normalize_function_parameters_root(&parameters)?
+        {
+            converted.insert("parameters".to_owned(), normalized);
+        }
         let alias = self.alias(ToolIdentity::new(ToolKind::Function, namespace, name));
         converted.insert("name".to_owned(), Value::String(alias));
         Ok(vec![Value::Object(converted)])
@@ -1448,6 +1498,136 @@ fn optional_array(value: Option<&Value>) -> Result<(Vec<Value>, bool), GrokReque
     }
 }
 
+/// Grok Build 要求 function parameters 的根节点必为 object；Codex 会为可选参数
+/// 生成 `object | null`，这里只移除根节点 nullability，嵌套字段保持原样。
+fn normalize_function_parameters_root(
+    value: &Value,
+) -> Result<Option<Value>, GrokRequestEncodeError> {
+    let Some(schema) = value.as_object() else {
+        return Ok(None);
+    };
+    let mut normalized = schema.clone();
+    let mut changed = false;
+
+    if let Some(types) = normalized.get("type").and_then(Value::as_array).cloned() {
+        let removed_null = types.iter().any(|value| value.as_str() == Some("null"));
+        if removed_null {
+            let remaining = types
+                .into_iter()
+                .filter(|value| value.as_str() != Some("null"))
+                .collect::<Vec<_>>();
+            if remaining.len() != 1 || remaining[0].as_str() != Some("object") {
+                return Err(invalid_function_parameters_root());
+            }
+            normalized.insert("type".to_owned(), Value::String("object".to_owned()));
+            changed = true;
+        }
+    }
+
+    for keyword in ["anyOf", "oneOf"] {
+        let Some(branches) = normalized.get(keyword).and_then(Value::as_array).cloned() else {
+            continue;
+        };
+        let removed_null = branches.iter().any(is_null_only_schema);
+        if !removed_null {
+            continue;
+        }
+        let remaining = branches
+            .into_iter()
+            .filter(|branch| !is_null_only_schema(branch))
+            .collect::<Vec<_>>();
+        if remaining.is_empty()
+            || remaining.iter().any(|branch| {
+                branch.as_object().is_none_or(|branch| {
+                    !is_object_root_schema(branch, &normalized, &mut BTreeSet::new())
+                })
+            })
+        {
+            return Err(invalid_function_parameters_root());
+        }
+        if remaining.len() == 1 && normalized.len() == 1 {
+            normalized = remaining[0]
+                .as_object()
+                .cloned()
+                .ok_or_else(invalid_function_parameters_root)?;
+            normalized.insert("type".to_owned(), Value::String("object".to_owned()));
+        } else {
+            normalized.insert(keyword.to_owned(), Value::Array(remaining));
+            normalized.insert("type".to_owned(), Value::String("object".to_owned()));
+        }
+        changed = true;
+    }
+
+    Ok(changed.then_some(Value::Object(normalized)))
+}
+
+fn is_null_only_schema(value: &Value) -> bool {
+    let Some(schema) = value.as_object() else {
+        return false;
+    };
+    match schema.get("type") {
+        Some(Value::String(kind)) => kind == "null",
+        Some(Value::Array(types)) if !types.is_empty() => {
+            types.iter().all(|kind| kind.as_str() == Some("null"))
+        }
+        _ => false,
+    }
+}
+
+fn is_object_root_schema(
+    schema: &Map<String, Value>,
+    root: &Map<String, Value>,
+    visited: &mut BTreeSet<String>,
+) -> bool {
+    match schema.get("type") {
+        Some(Value::String(kind)) => return kind == "object",
+        Some(Value::Array(types)) if !types.is_empty() => {
+            return types.iter().all(|kind| kind.as_str() == Some("object"));
+        }
+        Some(_) => return false,
+        None => {}
+    }
+    if schema.contains_key("properties") {
+        return true;
+    }
+    let Some(reference) = schema.get("$ref").and_then(Value::as_str) else {
+        return false;
+    };
+    if !visited.insert(reference.to_owned()) {
+        return false;
+    }
+    resolve_local_schema_ref(root, reference)
+        .is_some_and(|resolved| is_object_root_schema(resolved, root, visited))
+}
+
+fn resolve_local_schema_ref<'a>(
+    root: &'a Map<String, Value>,
+    reference: &str,
+) -> Option<&'a Map<String, Value>> {
+    if reference == "#" {
+        return Some(root);
+    }
+    let path = reference.strip_prefix("#/")?;
+    let mut segments = path.split('/');
+    let first = decode_json_pointer_segment(segments.next()?);
+    let mut current = root.get(&first)?;
+    for segment in segments {
+        let segment = decode_json_pointer_segment(segment);
+        current = current.as_object()?.get(&segment)?;
+    }
+    current.as_object()
+}
+
+fn decode_json_pointer_segment(segment: &str) -> String {
+    segment.replace("~1", "/").replace("~0", "~")
+}
+
+const fn invalid_function_parameters_root() -> GrokRequestEncodeError {
+    GrokRequestEncodeError::InvalidRequestField {
+        field: "tools[].parameters",
+    }
+}
+
 fn inspect_tool_search(tools: &[Value]) -> Result<bool, GrokRequestEncodeError> {
     let mut client_search = false;
     let mut server_search = false;

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

@@ -79,6 +79,117 @@ impl fmt::Debug for GrokInferenceRequest {
 pub type GrokInferenceChunkStream =
     Pin<Box<dyn Stream<Item = Result<Vec<u8>, GrokInferenceTransportError>> + Send + 'static>>;
 
+/// Whether the account-isolated inference client was already cached.
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub enum GrokInferenceClientCacheStatus {
+    /// The binding reused its existing account-isolated client.
+    Hit,
+    /// The binding was absent when the request first checked the cache.
+    Miss,
+}
+
+impl GrokInferenceClientCacheStatus {
+    #[must_use]
+    pub const fn as_str(self) -> &'static str {
+        match self {
+            Self::Hit => "hit",
+            Self::Miss => "miss",
+        }
+    }
+}
+
+/// Resolver that supplied the addresses used by the request.
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub enum GrokInferenceDnsSource {
+    /// The system resolver returned an entirely public address set.
+    System,
+    /// The system result was unusable and the trusted DoH fallback supplied addresses.
+    TrustedDoh,
+}
+
+impl GrokInferenceDnsSource {
+    #[must_use]
+    pub const fn as_str(self) -> &'static str {
+        match self {
+            Self::System => "system",
+            Self::TrustedDoh => "trusted_doh",
+        }
+    }
+}
+
+/// DNS work observed while opening an upstream connection.
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub struct GrokInferenceDnsObservation {
+    source: GrokInferenceDnsSource,
+    duration_ms: u64,
+}
+
+impl GrokInferenceDnsObservation {
+    #[must_use]
+    pub const fn new(source: GrokInferenceDnsSource, duration_ms: u64) -> Self {
+        Self {
+            source,
+            duration_ms,
+        }
+    }
+
+    #[must_use]
+    pub const fn source(self) -> GrokInferenceDnsSource {
+        self.source
+    }
+
+    #[must_use]
+    pub const fn duration_ms(self) -> u64 {
+        self.duration_ms
+    }
+}
+
+/// Low-cardinality transport timings and pool facts for one inference request.
+#[derive(Debug, Clone, Copy, Default, PartialEq, Eq)]
+pub struct GrokInferenceTransportMetrics {
+    headers_ms: Option<u64>,
+    client_cache_status: Option<GrokInferenceClientCacheStatus>,
+    dns: Option<GrokInferenceDnsObservation>,
+}
+
+impl GrokInferenceTransportMetrics {
+    #[must_use]
+    pub const fn with_headers_ms(mut self, headers_ms: u64) -> Self {
+        self.headers_ms = Some(headers_ms);
+        self
+    }
+
+    #[must_use]
+    pub const fn with_client_cache_status(
+        mut self,
+        status: GrokInferenceClientCacheStatus,
+    ) -> Self {
+        self.client_cache_status = Some(status);
+        self
+    }
+
+    #[must_use]
+    pub const fn with_dns(mut self, dns: GrokInferenceDnsObservation) -> Self {
+        self.dns = Some(dns);
+        self
+    }
+
+    #[must_use]
+    pub const fn headers_ms(self) -> Option<u64> {
+        self.headers_ms
+    }
+
+    #[must_use]
+    pub const fn client_cache_status(self) -> Option<GrokInferenceClientCacheStatus> {
+        self.client_cache_status
+    }
+
+    #[must_use]
+    pub const fn dns(self) -> Option<GrokInferenceDnsObservation> {
+        self.dns
+    }
+}
+
 /// Accepted inference response. Non-success HTTP responses must be returned as
 /// [`GrokInferenceTransportError`] instead.
 pub struct GrokInferenceResponse {
@@ -86,6 +197,7 @@ pub struct GrokInferenceResponse {
     http_version: UpstreamHttpVersion,
     status_code: u16,
     request_id: Option<SafeUpstreamValue>,
+    transport_metrics: GrokInferenceTransportMetrics,
 }
 
 impl GrokInferenceResponse {
@@ -102,9 +214,19 @@ impl GrokInferenceResponse {
             http_version,
             status_code,
             request_id,
+            transport_metrics: GrokInferenceTransportMetrics::default(),
         }
     }
 
+    #[must_use]
+    pub const fn with_transport_metrics(
+        mut self,
+        transport_metrics: GrokInferenceTransportMetrics,
+    ) -> Self {
+        self.transport_metrics = transport_metrics;
+        self
+    }
+
     #[must_use]
     pub const fn http_version(&self) -> UpstreamHttpVersion {
         self.http_version
@@ -120,6 +242,11 @@ impl GrokInferenceResponse {
         self.request_id.as_ref()
     }
 
+    #[must_use]
+    pub const fn transport_metrics(&self) -> GrokInferenceTransportMetrics {
+        self.transport_metrics
+    }
+
     pub fn into_body(self) -> GrokInferenceChunkStream {
         self.body
     }
@@ -132,6 +259,7 @@ impl fmt::Debug for GrokInferenceResponse {
             .field("http_version", &self.http_version)
             .field("status_code", &self.status_code)
             .field("request_id", &self.request_id)
+            .field("transport_metrics", &self.transport_metrics)
             .field("body", &"[SSE STREAM]")
             .finish()
     }
@@ -174,6 +302,7 @@ pub struct GrokInferenceTransportError {
     http_version: Option<UpstreamHttpVersion>,
     request_id: Option<SafeUpstreamValue>,
     upstream_code: Option<SafeUpstreamValue>,
+    transport_metrics: GrokInferenceTransportMetrics,
     credential_recovery_required: bool,
     sensitive_context_redacted: bool,
 }
@@ -190,6 +319,11 @@ impl GrokInferenceTransportError {
             http_version: None,
             request_id: None,
             upstream_code: None,
+            transport_metrics: GrokInferenceTransportMetrics {
+                headers_ms: None,
+                client_cache_status: None,
+                dns: None,
+            },
             credential_recovery_required: false,
             sensitive_context_redacted: false,
         }
@@ -229,6 +363,15 @@ impl GrokInferenceTransportError {
         self
     }
 
+    #[must_use]
+    pub const fn with_transport_metrics(
+        mut self,
+        transport_metrics: GrokInferenceTransportMetrics,
+    ) -> Self {
+        self.transport_metrics = transport_metrics;
+        self
+    }
+
     #[must_use]
     pub const fn with_credential_recovery(mut self) -> Self {
         self.credential_recovery_required = true;
@@ -281,6 +424,11 @@ impl GrokInferenceTransportError {
         self.upstream_code.as_ref()
     }
 
+    #[must_use]
+    pub const fn transport_metrics(&self) -> GrokInferenceTransportMetrics {
+        self.transport_metrics
+    }
+
     #[must_use]
     pub const fn requires_credential_recovery(&self) -> bool {
         self.credential_recovery_required
@@ -304,6 +452,7 @@ impl fmt::Debug for GrokInferenceTransportError {
             .field("http_version", &self.http_version)
             .field("request_id", &self.request_id)
             .field("upstream_code", &self.upstream_code)
+            .field("transport_metrics", &self.transport_metrics)
             .field(
                 "credential_recovery_required",
                 &self.credential_recovery_required,

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

@@ -9,7 +9,9 @@ use futures::{StreamExt as _, future::BoxFuture};
 use gateway_admin::model::accounts::{
     AccountAvailability as AdminAccountAvailability, AccountRecord,
 };
-use gateway_admin::model::observability::{CurrencyCost, ProviderBillingInput};
+use gateway_admin::model::observability::{
+    CurrencyCost, DesktopReleaseStatus, ProviderBillingInput,
+};
 use gateway_admin::model::provider_credentials::{
     CompleteAuthorization, PrepareCredentialImport, PrepareCredentialRefresh,
     PrepareCredentialRotation, ProviderDocument, ProviderExportCredentialInput,
@@ -281,7 +283,9 @@ async fn xai_admin_provider_validates_known_billing_breakdown() {
     assert_eq!(profile.product, "Grok Build");
     assert_eq!(profile.version, "0.2.106");
     assert_eq!(profile.user_agent, "grok-shell/0.2.106 (linux; x86_64)");
-    assert!(profile.release.is_none());
+    let release = profile.release.expect("release status");
+    assert_eq!(release.status, DesktopReleaseStatus::Unchecked);
+    assert!(release.checked_at.is_none());
 
     let billing = admin
         .calculated_billing(&ProviderBillingInput {

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

@@ -20,8 +20,8 @@ use provider_xai::{
 };
 
 use crate::support::{
-    MemoryGrokCatalogCache, MemoryProviderAccountStore, create_input, credential_object,
-    runtime_policy, seed_input,
+    MemoryCooldownPort, MemoryGrokCatalogCache, MemoryProviderAccountStore, create_input,
+    credential_object, runtime_policy, seed_input,
 };
 
 const OFFICIAL_FIXTURE: &[u8] =
@@ -169,11 +169,13 @@ async fn fixture_many(
         available: lease_available,
         calls: AtomicUsize::new(0),
     });
+    let cooldowns = Arc::new(MemoryCooldownPort::default());
     let service = GrokCredentialRefreshService::new(
         repository.clone(),
         refresher_port,
         catalog,
         leases,
+        cooldowns,
         runtime_policy(),
     );
     (store, repository, refresher, service)
@@ -380,10 +382,11 @@ async fn banned_failure_marks_account_banned() {
 }
 
 #[tokio::test]
-async fn ambiguous_refresh_fails_closed_and_is_not_retried() {
+async fn ambiguous_refresh_uses_runtime_cooldown_without_persisting_account_state() {
     let input = due_input("ambiguous");
     let id = input.account_id.clone();
-    let (store, _, _, service) = fixture(input, [Err(GrokRefreshFailure::Ambiguous)], true).await;
+    let (store, _, refresher, service) =
+        fixture(input, [Err(GrokRefreshFailure::Ambiguous)], true).await;
     let outcomes = service.refresh_due().await.expect("refresh");
     assert!(matches!(
         outcomes.as_slice(),
@@ -391,8 +394,9 @@ async fn ambiguous_refresh_fails_closed_and_is_not_retried() {
     ));
     assert_eq!(
         store.account(&id).expect("account").availability(),
-        AccountAvailability::Invalid
+        AccountAvailability::Unknown
     );
+    assert_eq!(store.account(&id).expect("account").revision().get(), 1);
     assert!(
         service
             .refresh_due()
@@ -400,6 +404,7 @@ async fn ambiguous_refresh_fails_closed_and_is_not_retried() {
             .expect("second cycle")
             .is_empty()
     );
+    assert_eq!(refresher.prepare_calls.load(Ordering::SeqCst), 1);
 }
 
 #[tokio::test]

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

@@ -23,7 +23,8 @@ use provider_xai::{
 };
 
 use crate::support::{
-    MemoryGrokCatalogCache, MemoryProviderAccountStore, account_id, create_input, seed_input,
+    MemoryCooldownPort, MemoryGrokCatalogCache, MemoryProviderAccountStore, account_id,
+    create_input, seed_input,
 };
 
 struct SchedulingCoordinator {
@@ -88,6 +89,7 @@ struct SelectorFixture {
     cache: Arc<MemoryGrokCatalogCache>,
     selector: GrokAccountSessionSelector,
     coordinator: Arc<SchedulingCoordinator>,
+    cooldowns: Arc<MemoryCooldownPort>,
     feedback: Arc<AccountFeedbackStats>,
 }
 
@@ -137,6 +139,7 @@ impl SelectorFixture {
             signals: Mutex::new(signals),
             denied: Mutex::new(BTreeSet::new()),
         });
+        let cooldowns = Arc::new(MemoryCooldownPort::default());
         let catalog_cache: Arc<dyn GrokCredentialCatalogCache> = cache.clone();
         let lease_port: Arc<dyn ProviderLeasePort> = coordinator.clone();
         let quota = Arc::new(crate::support::grok_quota_service(
@@ -150,6 +153,7 @@ impl SelectorFixture {
             catalog_cache,
             quota,
             lease_port,
+            cooldowns.clone(),
             Arc::clone(&feedback),
         );
         Self {
@@ -157,6 +161,7 @@ impl SelectorFixture {
             cache,
             selector,
             coordinator,
+            cooldowns,
             feedback,
         }
     }
@@ -257,7 +262,7 @@ async fn required_account_overrides_smart_selection_without_fallback() {
 }
 
 #[tokio::test]
-async fn unauthorized_feedback_marks_only_selected_account_invalid() {
+async fn unauthorized_feedback_records_runtime_cooldown_without_persisting_account_state() {
     let fixture = SelectorFixture::new(&["feedback-a", "feedback-b"]).await;
     let session = fixture
         .selector
@@ -275,7 +280,7 @@ async fn unauthorized_feedback_marks_only_selected_account_invalid() {
             .account(&selected)
             .expect("selected")
             .availability(),
-        AccountAvailability::Invalid
+        AccountAvailability::Ready
     );
     let other = [account_id("feedback-a"), account_id("feedback-b")]
         .into_iter()
@@ -285,16 +290,78 @@ async fn unauthorized_feedback_marks_only_selected_account_invalid() {
         fixture.store.account(&other).expect("other").availability(),
         AccountAvailability::Ready
     );
+    assert!(
+        fixture
+            .cooldowns
+            .cooldown(&selected)
+            .is_some_and(|cooldown| cooldown.until() > SystemTime::now())
+    );
 }
 
 #[tokio::test]
-async fn rate_limit_feedback_applies_bounded_cooldown() {
-    let fixture = SelectorFixture::new(&["rate-limit"]).await;
+async fn cooldown_for_superseded_credential_does_not_block_rotated_account() {
+    let fixture = SelectorFixture::new(&["revision-fence"]).await;
+    let id = account_id("revision-fence");
     let session = fixture
         .selector
         .select(fixture.request(BTreeSet::new()))
         .await
         .expect("session");
+    fixture
+        .selector
+        .record_failure(
+            &session,
+            GrokCredentialFailure::RateLimited {
+                retry_after: Some(Duration::from_secs(60)),
+            },
+        )
+        .await;
+
+    let current = fixture
+        .store
+        .load_credential(&id, CredentialRevision::new(1).expect("revision"))
+        .await
+        .expect("current credential");
+    let prepared = GrokCredentialAdmin
+        .prepare_rotation(&RotateManagedGrokCredential {
+            current,
+            secret: provider_xai::GrokOAuthSecret {
+                access_token: provider_xai::SecretValue::new("new-access"),
+                refresh_token: provider_xai::SecretValue::new("new-refresh"),
+                id_token: None,
+                scope: provider_xai::OFFICIAL_SCOPES.join(" "),
+            },
+            verified_account: crate::support::profile("subject-revision-fence"),
+            next_refresh_at: chrono::Utc::now() + chrono::Duration::minutes(30),
+        })
+        .expect("rotate");
+    assert!(matches!(
+        fixture
+            .store
+            .compare_and_swap_credential(prepared.credential)
+            .await
+            .expect("persist rotation"),
+        CredentialCasOutcome::Updated(revision) if revision.get() == 2
+    ));
+
+    let next = fixture
+        .selector
+        .select(fixture.request(BTreeSet::new()))
+        .await
+        .expect("replacement credential should not inherit cooldown");
+    assert_eq!(next.account_id(), &id);
+    assert!(fixture.cooldowns.cooldown(&id).is_none());
+}
+
+#[tokio::test]
+async fn rate_limit_feedback_records_runtime_cooldown_without_persisting_account_state() {
+    let fixture = SelectorFixture::new(&["rate-limit", "available"]).await;
+    let session = fixture
+        .selector
+        .select(fixture.request(BTreeSet::new()))
+        .await
+        .expect("session");
+    let selected = session.account_id().clone();
     fixture
         .selector
         .record_failure(
@@ -304,20 +371,24 @@ async fn rate_limit_feedback_applies_bounded_cooldown() {
             },
         )
         .await;
-    let account = fixture
-        .store
-        .account(session.account_id())
-        .expect("account");
-    assert_eq!(account.availability(), AccountAvailability::Cooldown);
+    let account = fixture.store.account(&selected).expect("account");
+    assert_eq!(account.availability(), AccountAvailability::Ready);
     assert!(
-        account
-            .cooldown_until()
-            .is_some_and(|until| until > SystemTime::now())
+        fixture
+            .cooldowns
+            .cooldown(&selected)
+            .is_some_and(|cooldown| cooldown.until() > SystemTime::now())
     );
+    let next = fixture
+        .selector
+        .select(fixture.request(BTreeSet::new()))
+        .await
+        .expect("runtime cooldown should leave the other account available");
+    assert_ne!(next.account_id(), &selected);
 }
 
 #[tokio::test]
-async fn interrupted_stream_feedback_applies_account_cooldown() {
+async fn interrupted_stream_feedback_records_runtime_cooldown_without_persisting_account_state() {
     let fixture = SelectorFixture::new(&["stream-interrupted"]).await;
     let session = fixture
         .selector
@@ -333,12 +404,20 @@ async fn interrupted_stream_feedback_applies_account_cooldown() {
         .store
         .account(session.account_id())
         .expect("account");
-    assert_eq!(account.availability(), AccountAvailability::Cooldown);
+    assert_eq!(account.availability(), AccountAvailability::Ready);
     assert!(
-        account
-            .cooldown_until()
-            .is_some_and(|until| until > SystemTime::now())
+        fixture
+            .cooldowns
+            .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)
+    ));
 }
 
 #[tokio::test]

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

@@ -31,12 +31,14 @@ use gateway_core::routing::{
 use provider_xai::{
     GrokBuildProvider, GrokCredentialCatalogCache, GrokCredentialFailure,
     GrokCredentialFeedbackFuture, GrokCredentialRecovery, GrokCredentialRecoveryOutcome,
-    GrokCredentialRepository, GrokInferenceRequest, GrokInferenceResponse, GrokInferenceTransport,
+    GrokCredentialRepository, GrokInferenceClientCacheStatus, GrokInferenceDnsObservation,
+    GrokInferenceDnsSource, GrokInferenceRequest, GrokInferenceResponse, GrokInferenceTransport,
     GrokInferenceTransportError, GrokInferenceTransportErrorKind, GrokInferenceTransportFuture,
-    GrokModelCatalogRequest, GrokModelCatalogTransport, GrokModelCatalogTransportFuture,
-    GrokModelCatalogTransportResponse, GrokSessionBinding, GrokSessionSelection,
-    GrokSessionSelector, GrokSessionSelectorError, GrokSessionSelectorFuture, SecretValue,
-    SelectedGrokSession,
+    GrokInferenceTransportMetrics, GrokModelCatalogRequest, GrokModelCatalogTransport,
+    GrokModelCatalogTransportError, GrokModelCatalogTransportErrorKind,
+    GrokModelCatalogTransportFuture, GrokModelCatalogTransportResponse, GrokSessionBinding,
+    GrokSessionSelection, GrokSessionSelector, GrokSessionSelectorError, GrokSessionSelectorFuture,
+    SecretValue, SelectedGrokSession,
 };
 use serde_json::{Map, json};
 
@@ -47,6 +49,17 @@ use crate::support::{
 const MODEL: &str = "grok-4.5";
 const CATALOG_FIXTURE: &[u8] =
     include_bytes!("../transport/catalog/fixtures/official_grok_models_snapshot.json");
+
+fn observed_transport_metrics() -> GrokInferenceTransportMetrics {
+    GrokInferenceTransportMetrics::default()
+        .with_headers_ms(42)
+        .with_client_cache_status(GrokInferenceClientCacheStatus::Miss)
+        .with_dns(GrokInferenceDnsObservation::new(
+            GrokInferenceDnsSource::System,
+            7,
+        ))
+}
+
 const CATALOG_WITHOUT_TOOL_METADATA: &[u8] = br#"{
   "object": "list",
   "data": [{
@@ -245,6 +258,7 @@ impl GrokSessionSelector for StubSelector {
 
 enum InferenceMode {
     Success,
+    SuccessWithMetrics(GrokInferenceTransportMetrics),
     SuccessBody(Vec<u8>),
     Error(GrokInferenceTransportError),
     StreamError(GrokInferenceTransportError),
@@ -265,6 +279,14 @@ impl StubInferenceTransport {
         })
     }
 
+    fn success_with_metrics(metrics: GrokInferenceTransportMetrics) -> Arc<Self> {
+        Arc::new(Self {
+            calls: AtomicUsize::new(0),
+            requests: Mutex::new(Vec::new()),
+            modes: Mutex::new(VecDeque::from([InferenceMode::SuccessWithMetrics(metrics)])),
+        })
+    }
+
     fn error(error: GrokInferenceTransportError) -> Arc<Self> {
         Arc::new(Self {
             calls: AtomicUsize::new(0),
@@ -308,6 +330,13 @@ impl GrokInferenceTransport for StubInferenceTransport {
                     200,
                     None,
                 )),
+                InferenceMode::SuccessWithMetrics(metrics) => Ok(GrokInferenceResponse::new(
+                    Box::pin(stream::iter([Ok(SUCCESS_SSE.to_vec())])),
+                    UpstreamHttpVersion::Http2,
+                    200,
+                    None,
+                )
+                .with_transport_metrics(metrics)),
                 InferenceMode::SuccessBody(body) => Ok(GrokInferenceResponse::new(
                     Box::pin(stream::iter([Ok(body)])),
                     UpstreamHttpVersion::Http2,
@@ -352,6 +381,18 @@ impl GrokModelCatalogTransport for CatalogWithoutToolMetadataTransport {
     }
 }
 
+struct UnavailableCatalogTransport;
+
+impl GrokModelCatalogTransport for UnavailableCatalogTransport {
+    fn execute(&self, _: GrokModelCatalogRequest) -> GrokModelCatalogTransportFuture<'_> {
+        Box::pin(async {
+            Err(GrokModelCatalogTransportError::new(
+                GrokModelCatalogTransportErrorKind::Unavailable,
+            ))
+        })
+    }
+}
+
 struct StubRecovery {
     calls: AtomicUsize,
     outcome: GrokCredentialRecoveryOutcome,
@@ -506,6 +547,19 @@ fn operation_with_reasoning_effort(effort: &str) -> Operation {
     Operation::Generate(GenerateRequest::from_protocol_payload(payload))
 }
 
+fn operation_with_invalid_tools() -> Operation {
+    let payload = ProtocolPayload::json_object(
+        "openai",
+        Map::from_iter([
+            ("model".to_owned(), json!("client-model")),
+            ("input".to_owned(), json!("hello")),
+            ("tools".to_owned(), json!({"not": "an array"})),
+        ]),
+    )
+    .expect("OpenAI payload");
+    Operation::Generate(GenerateRequest::from_protocol_payload(payload))
+}
+
 fn compaction_operation() -> Operation {
     let payload = ProtocolPayload::json_object(
         "openai",
@@ -693,7 +747,7 @@ async fn execute_forwards_required_account_to_selector() {
 #[tokio::test]
 async fn execute_returns_cold_stream_and_records_selected_account() {
     let selector = StubSelector::success();
-    let transport = StubInferenceTransport::success();
+    let transport = StubInferenceTransport::success_with_metrics(observed_transport_metrics());
     let provider = provider(selector, transport.clone()).await;
     let mut stream = provider
         .execute(
@@ -718,6 +772,23 @@ async fn execute_returns_cold_stream_and_records_selected_account() {
     assert_eq!(observation.transport().as_str(), "http_sse");
     assert_eq!(observation.http_version(), Some(UpstreamHttpVersion::Http2));
     assert_eq!(observation.status_code(), Some(200));
+    assert_eq!(observation.timings().headers_ms, Some(42));
+    let metadata: serde_json::Value = serde_json::from_str(
+        observation
+            .provider_metadata()
+            .expect("xAI transport metadata")
+            .as_json(),
+    )
+    .expect("valid metadata JSON");
+    assert_eq!(
+        metadata,
+        json!({
+            "clientCache": "miss",
+            "dnsSource": "system",
+            "dnsMs": 7,
+            "upstreamHeadersMs": 42
+        })
+    );
     assert!(
         events
             .iter()
@@ -727,6 +798,48 @@ async fn execute_returns_cold_stream_and_records_selected_account() {
     );
 }
 
+#[tokio::test]
+async fn transport_error_observation_should_retain_available_metrics() {
+    let transport = StubInferenceTransport::error(
+        GrokInferenceTransportError::new(
+            GrokInferenceTransportErrorKind::Unavailable,
+            UpstreamSendState::Sent,
+        )
+        .with_status(503)
+        .with_transport_metrics(observed_transport_metrics()),
+    );
+    let provider = provider(StubSelector::success(), transport).await;
+    let mut stream = provider
+        .execute(
+            provider_request("xai"),
+            context(CancellationToken::new(), None),
+        )
+        .await
+        .expect("provider stream");
+
+    let event = stream
+        .next()
+        .await
+        .expect("observation event")
+        .expect("observation must succeed");
+    let observation = event.response_observation().expect("transport observation");
+    assert_eq!(observation.status_code(), Some(503));
+    assert_eq!(observation.timings().headers_ms, Some(42));
+    let metadata: serde_json::Value = serde_json::from_str(
+        observation
+            .provider_metadata()
+            .expect("xAI transport metadata")
+            .as_json(),
+    )
+    .expect("valid metadata JSON");
+    assert_eq!(metadata["clientCache"], "miss");
+    assert_eq!(metadata["dnsSource"], "system");
+    assert_eq!(metadata["dnsMs"], 7);
+
+    let error = next_provider_error(&mut stream).await;
+    assert_eq!(error.kind(), ProviderErrorKind::Unavailable);
+}
+
 #[tokio::test]
 async fn compaction_stream_should_emit_openai_wire_and_keep_only_accounting_canonical() {
     let summary = valid_compaction_summary("validated summary");
@@ -1047,6 +1160,38 @@ async fn first_unauthorized_should_refresh_and_request_one_same_account_retry()
     assert!(selector.feedback.lock().expect("feedback").is_empty());
 }
 
+#[tokio::test]
+async fn unavailable_unauthorized_recovery_records_temporary_credential_feedback() {
+    let selector = StubSelector::success();
+    let transport = StubInferenceTransport::error(
+        GrokInferenceTransportError::new(
+            GrokInferenceTransportErrorKind::Unauthorized,
+            UpstreamSendState::Sent,
+        )
+        .with_status(401)
+        .with_credential_recovery(),
+    );
+    let recovery = StubRecovery::new(GrokCredentialRecoveryOutcome::Unavailable);
+    let provider = provider_with_recovery(selector.clone(), transport, recovery.clone()).await;
+    let mut stream = provider
+        .execute(
+            provider_request("xai"),
+            context(CancellationToken::new(), None),
+        )
+        .await
+        .expect("stream");
+
+    let error = next_provider_error(&mut stream).await;
+
+    assert_eq!(error.kind(), ProviderErrorKind::Unauthorized);
+    assert!(!error.retries_same_account());
+    assert_eq!(recovery.calls.load(Ordering::SeqCst), 1);
+    assert_eq!(
+        selector.feedback.lock().expect("feedback").as_slice(),
+        &[GrokCredentialFailure::Unauthorized]
+    );
+}
+
 #[tokio::test]
 async fn explicit_http_429_marks_provider_error_replay_safe() {
     let error = mapped_transport_error(
@@ -1628,6 +1773,37 @@ async fn provider_rejects_target_owned_by_other_provider() {
     assert_eq!(selector.calls.load(Ordering::SeqCst), 0);
 }
 
+#[tokio::test]
+async fn provider_should_expose_safe_request_normalization_field_before_selection() {
+    let selector = StubSelector::success();
+    let provider = provider(selector.clone(), StubInferenceTransport::success()).await;
+    let error = match provider
+        .execute(
+            provider_request_with_operation("xai", operation_with_invalid_tools()),
+            context(CancellationToken::new(), None),
+        )
+        .await
+    {
+        Ok(_) => panic!("invalid tools must fail"),
+        Err(error) => error,
+    };
+
+    assert_eq!(error.kind(), ProviderErrorKind::InvalidRequest);
+    assert_eq!(
+        error.client_visible_upstream_error().map(|detail| (
+            detail.message(),
+            detail.code(),
+            detail.error_type()
+        )),
+        Some((
+            "Grok Build request field `tools` could not be normalized safely",
+            Some("invalid_request_normalization"),
+            Some("invalid_request_error"),
+        )),
+    );
+    assert_eq!(selector.calls.load(Ordering::SeqCst), 0);
+}
+
 #[tokio::test]
 async fn provider_compiles_realtime_catalog_capabilities() {
     let provider = provider(StubSelector::success(), StubInferenceTransport::success()).await;
@@ -1645,6 +1821,41 @@ async fn provider_compiles_realtime_catalog_capabilities() {
             ))
             .is_some()
     );
+    let presentation = capabilities[0]
+        .presentation()
+        .expect("Grok model presentation");
+    assert_eq!(presentation.display_name(), Some("Grok 4.5"));
+    assert_eq!(
+        presentation.supported_reasoning_efforts(),
+        ["low", "medium", "high"]
+    );
+    assert_eq!(presentation.context_window_tokens(), Some(1_000_000));
+    assert!(presentation.agent_tools());
+}
+
+#[tokio::test]
+async fn provider_publishes_default_codex_profile_when_catalog_is_unavailable() {
+    let provider = provider_with_catalog_transport(
+        StubSelector::success(),
+        StubInferenceTransport::success(),
+        StubRecovery::new(GrokCredentialRecoveryOutcome::Unavailable),
+        Arc::new(UnavailableCatalogTransport),
+    )
+    .await;
+    let capabilities = provider
+        .query_model_capabilities()
+        .await
+        .expect("fallback capabilities");
+
+    assert_eq!(capabilities.len(), 1);
+    assert_eq!(capabilities[0].upstream_model().as_str(), MODEL);
+    let presentation = capabilities[0]
+        .presentation()
+        .expect("fallback Grok model presentation");
+    assert_eq!(presentation.display_name(), Some("Grok 4.5"));
+    assert_eq!(presentation.context_window_tokens(), Some(500_000));
+    assert!(presentation.agent_tools());
+    assert!(presentation.parallel_tool_calls());
 }
 
 #[tokio::test]

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

@@ -15,9 +15,9 @@ use gateway_core::engine::credential::{
 };
 use gateway_core::error::{StoreError, StoreErrorKind};
 use gateway_core::provider_ports::{
-    ProviderRefreshPolicy, ProviderRuntimePolicyPort, ProviderSessionAffinityKey,
-    ProviderSessionAffinityPort, ProviderSessionExclusionPort, ProviderSessionExclusions,
-    ProviderStoreError,
+    ProviderCooldown, ProviderCooldownPort, ProviderRefreshPolicy, ProviderRuntimePolicyPort,
+    ProviderSessionAffinityKey, ProviderSessionAffinityPort, ProviderSessionExclusionPort,
+    ProviderSessionExclusions, ProviderStoreError,
 };
 use gateway_core::routing::ProviderKind;
 use provider_xai::{
@@ -550,6 +550,17 @@ pub struct TestSessionAffinity;
 
 pub struct TestSessionExclusions;
 
+#[derive(Default)]
+pub struct MemoryCooldownPort {
+    cooldowns: Mutex<BTreeMap<ProviderAccountId, ProviderCooldown>>,
+}
+
+impl MemoryCooldownPort {
+    pub fn cooldown(&self, account_id: &ProviderAccountId) -> Option<ProviderCooldown> {
+        lock(&self.cooldowns).get(account_id).cloned()
+    }
+}
+
 impl ProviderSessionAffinityPort for TestSessionAffinity {
     fn load<'a>(
         &'a self,
@@ -613,6 +624,50 @@ impl ProviderSessionExclusionPort for TestSessionExclusions {
     }
 }
 
+impl ProviderCooldownPort for MemoryCooldownPort {
+    fn put_if_later(
+        &self,
+        cooldown: ProviderCooldown,
+    ) -> futures::future::BoxFuture<'_, Result<bool, ProviderStoreError>> {
+        Box::pin(async move {
+            let mut cooldowns = lock(&self.cooldowns);
+            let should_write = cooldowns.get(cooldown.account_id()).is_none_or(|current| {
+                current.credential_revision() < cooldown.credential_revision()
+                    || (current.credential_revision() == cooldown.credential_revision()
+                        && current.until() < cooldown.until())
+            });
+            if should_write {
+                cooldowns.insert(cooldown.account_id().clone(), cooldown);
+            }
+            Ok(should_write)
+        })
+    }
+
+    fn read<'a>(
+        &'a self,
+        account_id: &'a ProviderAccountId,
+    ) -> futures::future::BoxFuture<'a, Result<Option<ProviderCooldown>, ProviderStoreError>> {
+        Box::pin(async move { Ok(self.cooldown(account_id)) })
+    }
+
+    fn clear<'a>(
+        &'a self,
+        account_id: &'a ProviderAccountId,
+        through_revision: CredentialRevision,
+    ) -> futures::future::BoxFuture<'a, Result<bool, ProviderStoreError>> {
+        Box::pin(async move {
+            let mut cooldowns = lock(&self.cooldowns);
+            let should_remove = cooldowns
+                .get(account_id)
+                .is_some_and(|current| current.credential_revision() <= through_revision);
+            if should_remove {
+                cooldowns.remove(account_id);
+            }
+            Ok(should_remove)
+        })
+    }
+}
+
 impl ProviderRuntimePolicyPort for StaticRuntimePolicy {
     fn load_refresh_policy(
         &self,

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

@@ -114,6 +114,42 @@ fn decoder_should_normalize_text_usage_and_completion() {
     ));
 }
 
+#[test]
+fn decoder_should_emit_each_complete_text_event_without_waiting_for_the_stream_end() {
+    let mut decoder = GrokCanonicalDecoder::new("grok-4.5");
+    let created = decoder
+        .push(
+            concat!(
+                "event: response.created\n",
+                "data: {\"type\":\"response.created\",\"response\":{\"id\":\"resp_live\",\"model\":\"grok-4.5\"}}\n\n",
+            )
+            .as_bytes(),
+        )
+        .expect("created event");
+    let delta = decoder
+        .push(
+            concat!(
+                "event: response.output_text.delta\n",
+                "data: {\"type\":\"response.output_text.delta\",\"output_index\":0,\"content_index\":0,\"delta\":\"live\"}\n\n",
+            )
+            .as_bytes(),
+        )
+        .expect("text delta");
+
+    assert!(
+        created
+            .iter()
+            .filter_map(ProviderEvent::wire_event)
+            .any(|event| event.event_type() == Some("response.created"))
+    );
+    assert!(delta.iter().any(|event| {
+        event.wire_event().is_some_and(|wire| {
+            wire.event_type() == Some("response.output_text.delta")
+                && wire.data().get("delta") == Some(&serde_json::json!("live"))
+        })
+    }));
+}
+
 #[test]
 fn decoder_should_preserve_image_tool_tokens_in_canonical_usage() {
     let body = concat!(
@@ -416,6 +452,72 @@ fn decoder_should_restore_custom_apply_patch_arguments_as_raw_input() {
     }));
 }
 
+#[test]
+fn decoder_should_buffer_only_custom_arguments_until_the_done_event() {
+    let request = tool_request(serde_json::json!({
+        "model": "client",
+        "input": "render",
+        "tools": [{"type": "custom", "name": "render"}]
+    }));
+    let mut decoder = GrokCanonicalDecoder::for_request("grok-4.5", &request);
+    decoder
+        .push(
+            concat!(
+                "event: response.created\n",
+                "data: {\"type\":\"response.created\",\"response\":{\"id\":\"resp_render\",\"model\":\"grok-4.5\"}}\n\n",
+            )
+            .as_bytes(),
+        )
+        .expect("custom response created");
+    let added = decoder
+        .push(
+            concat!(
+                "event: response.output_item.added\n",
+                "data: {\"type\":\"response.output_item.added\",\"output_index\":0,\"item\":{\"id\":\"item_render\",\"type\":\"function_call\",\"call_id\":\"call_render\",\"name\":\"render\"}}\n\n",
+            )
+            .as_bytes(),
+        )
+        .expect("custom item added");
+    let first_delta = decoder
+        .push(
+            concat!(
+                "event: response.function_call_arguments.delta\n",
+                "data: {\"type\":\"response.function_call_arguments.delta\",\"item_id\":\"item_render\",\"call_id\":\"call_render\",\"output_index\":0,\"delta\":\"{\\\"input\\\":\\\"hel\"}\n\n",
+            )
+            .as_bytes(),
+        )
+        .expect("first custom argument delta");
+    let second_delta = decoder
+        .push(
+            concat!(
+                "event: response.function_call_arguments.delta\n",
+                "data: {\"type\":\"response.function_call_arguments.delta\",\"item_id\":\"item_render\",\"call_id\":\"call_render\",\"output_index\":0,\"delta\":\"lo\\\"}\"}\n\n",
+            )
+            .as_bytes(),
+        )
+        .expect("second custom argument delta");
+    let done = decoder
+        .push(
+            concat!(
+                "event: response.function_call_arguments.done\n",
+                "data: {\"type\":\"response.function_call_arguments.done\",\"item_id\":\"item_render\",\"call_id\":\"call_render\",\"output_index\":0,\"arguments\":\"{\\\"input\\\":\\\"hello\\\"}\"}\n\n",
+            )
+            .as_bytes(),
+        )
+        .expect("custom argument done");
+
+    assert!(wire_events(&added).iter().any(|event| {
+        event.event_type() == Some("response.output_item.added")
+            && event.data().pointer("/item/type") == Some(&serde_json::json!("custom_tool_call"))
+    }));
+    assert!(first_delta.is_empty());
+    assert!(second_delta.is_empty());
+    assert!(wire_events(&done).iter().any(|event| {
+        event.event_type() == Some("response.custom_tool_call_input.done")
+            && event.data().get("input") == Some(&serde_json::json!("hello"))
+    }));
+}
+
 #[test]
 fn decoder_should_coalesce_reasoning_item_part_and_summary_index() {
     let body = concat!(

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

@@ -16,11 +16,12 @@ use wiremock::{Mock, MockServer, ResponseTemplate};
 
 use provider_xai::{
     FailClosedTokenVerifier, FormField, GrokBillingClient, GrokDnsResolutionPlan,
-    GrokDnsResolutionPolicy, GrokEndpointPolicy, GrokInferenceRequest, GrokInferenceTransport,
-    GrokInferenceTransportErrorKind, GrokModelCatalogSession, GrokOAuthClient, GrokOAuthConfig,
-    GrokReqwestTransportBuildError, GrokSessionBinding, HttpMethod, OAuthHttpRequest,
-    OAuthHttpTransport, OfficialGrokEndpointPolicy, ReqwestGrokInferenceTransport,
-    ReqwestGrokModelCatalogTransport, ReqwestOAuthTransport, SecretValue,
+    GrokDnsResolutionPolicy, GrokEndpointPolicy, GrokInferenceClientCacheStatus,
+    GrokInferenceRequest, GrokInferenceTransport, GrokInferenceTransportErrorKind,
+    GrokModelCatalogSession, GrokOAuthClient, GrokOAuthConfig, GrokReqwestTransportBuildError,
+    GrokSessionBinding, HttpMethod, OAuthHttpRequest, OAuthHttpTransport,
+    OfficialGrokEndpointPolicy, ReqwestGrokInferenceTransport, ReqwestGrokModelCatalogTransport,
+    ReqwestOAuthTransport, SecretValue,
 };
 
 use crate::support::loopback_endpoint_policy;
@@ -89,6 +90,43 @@ async fn inference_transport_should_stream_one_official_shape_response() {
     assert_eq!(chunks.len(), 1);
 }
 
+#[tokio::test]
+async fn inference_transport_should_report_account_client_cache_miss_then_hit() {
+    let server = MockServer::start().await;
+    Mock::given(method("POST"))
+        .and(path("/v1/responses"))
+        .respond_with(
+            ResponseTemplate::new(200).set_body_raw("data: [DONE]\n\n", "text/event-stream"),
+        )
+        .expect(2)
+        .mount(&server)
+        .await;
+    let origin = Url::parse(&server.uri()).expect("wiremock origin");
+    let transport = inference_transport(&origin);
+
+    let first = transport
+        .execute(inference_request(&origin))
+        .await
+        .expect("first response");
+    let first_metrics = first.transport_metrics();
+    assert_eq!(
+        first_metrics.client_cache_status(),
+        Some(GrokInferenceClientCacheStatus::Miss)
+    );
+    assert!(first_metrics.headers_ms().is_some());
+    first.into_body().collect::<Vec<_>>().await;
+
+    let second = transport
+        .execute(inference_request(&origin))
+        .await
+        .expect("second response");
+    assert_eq!(
+        second.transport_metrics().client_cache_status(),
+        Some(GrokInferenceClientCacheStatus::Hit)
+    );
+    second.into_body().collect::<Vec<_>>().await;
+}
+
 #[tokio::test]
 async fn inference_transport_should_reuse_one_client_only_within_the_same_binding() {
     let server = MockServer::start().await;
@@ -105,6 +143,7 @@ async fn inference_transport_should_reuse_one_client_only_within_the_same_bindin
     let endpoint_policy: Arc<dyn GrokEndpointPolicy> = Arc::new(CountingEndpointPolicy {
         inner: loopback_endpoint_policy(&origin),
         inference_client_builds: Arc::clone(&client_builds),
+        concurrency: None,
     });
     let transport = ReqwestGrokInferenceTransport::new(endpoint_policy).expect("transport");
 
@@ -132,6 +171,7 @@ async fn inference_transport_should_evict_the_least_recent_binding_at_the_fixed_
     let endpoint_policy: Arc<dyn GrokEndpointPolicy> = Arc::new(CountingEndpointPolicy {
         inner: loopback_endpoint_policy(&origin),
         inference_client_builds: Arc::clone(&client_builds),
+        concurrency: None,
     });
     let transport = ReqwestGrokInferenceTransport::new(endpoint_policy).expect("transport");
 
@@ -156,6 +196,62 @@ async fn inference_transport_should_evict_the_least_recent_binding_at_the_fixed_
     );
 }
 
+#[tokio::test]
+async fn inference_transport_should_build_distinct_cold_account_clients_concurrently() {
+    let server = MockServer::start().await;
+    Mock::given(method("POST"))
+        .and(path("/v1/responses"))
+        .respond_with(
+            ResponseTemplate::new(200).set_body_raw("data: [DONE]\n\n", "text/event-stream"),
+        )
+        .expect(3)
+        .mount(&server)
+        .await;
+    let origin = Url::parse(&server.uri()).expect("wiremock origin");
+    let concurrency = Arc::new(BuildConcurrency::new(Duration::from_millis(100)));
+    let endpoint_policy: Arc<dyn GrokEndpointPolicy> = Arc::new(CountingEndpointPolicy {
+        inner: loopback_endpoint_policy(&origin),
+        inference_client_builds: Arc::new(AtomicUsize::new(0)),
+        concurrency: Some(Arc::clone(&concurrency)),
+    });
+    let transport = Arc::new(
+        ReqwestGrokInferenceTransport::new(endpoint_policy).expect("concurrent transport"),
+    );
+    execute_inference(&transport, &origin, "account-primer").await;
+
+    let start = Arc::new(std::sync::Barrier::new(2));
+    let first = {
+        let start = Arc::clone(&start);
+        let transport = Arc::clone(&transport);
+        let origin = origin.clone();
+        tokio::task::spawn_blocking(move || {
+            let runtime = tokio::runtime::Builder::new_current_thread()
+                .enable_all()
+                .build()
+                .expect("first request runtime");
+            start.wait();
+            runtime.block_on(execute_inference(&transport, &origin, "account-cold-a"));
+        })
+    };
+    let second = {
+        let start = Arc::clone(&start);
+        let transport = Arc::clone(&transport);
+        let origin = origin.clone();
+        tokio::task::spawn_blocking(move || {
+            let runtime = tokio::runtime::Builder::new_current_thread()
+                .enable_all()
+                .build()
+                .expect("second request runtime");
+            start.wait();
+            runtime.block_on(execute_inference(&transport, &origin, "account-cold-b"));
+        })
+    };
+    first.await.expect("first task");
+    second.await.expect("second task");
+
+    assert!(concurrency.max_active.load(Ordering::SeqCst) >= 2);
+}
+
 #[tokio::test]
 async fn inference_transport_should_classify_http_failures_without_retaining_bodies() {
     let cases = [
@@ -201,6 +297,11 @@ async fn inference_transport_should_classify_http_failures_without_retaining_bod
                 Some(gateway_core::event::UpstreamHttpVersion::Http11),
             )
         );
+        assert_eq!(
+            error.transport_metrics().client_cache_status(),
+            Some(GrokInferenceClientCacheStatus::Miss)
+        );
+        assert!(error.transport_metrics().headers_ms().is_some());
         assert!(!rendered.contains(&secret));
     }
 }
@@ -318,27 +419,28 @@ async fn inference_transport_should_classify_reasoning_decode_rejections() {
 }
 
 #[tokio::test]
-async fn inference_transport_should_reject_success_with_non_sse_content_type() {
+async fn inference_transport_should_not_reject_sse_bytes_only_because_content_type_is_wrong() {
     let server = MockServer::start().await;
     Mock::given(method("POST"))
         .and(path("/v1/responses"))
-        .respond_with(ResponseTemplate::new(200).set_body_raw("{}", "application/json"))
+        .respond_with(ResponseTemplate::new(200).set_body_raw(
+            "data: {\"type\":\"response.created\",\"response\":{\"id\":\"resp_1\"}}\n\n",
+            "application/json",
+        ))
         .expect(1)
         .mount(&server)
         .await;
     let origin = Url::parse(&server.uri()).expect("wiremock origin");
-    let error = inference_transport(&origin)
+    let chunks = inference_transport(&origin)
         .execute(inference_request(&origin))
         .await
-        .expect_err("successful non-SSE response must fail closed");
+        .expect("body syntax, rather than response metadata, determines the stream")
+        .into_body()
+        .collect::<Vec<_>>()
+        .await;
 
-    assert_eq!(
-        (error.kind(), error.send_state()),
-        (
-            GrokInferenceTransportErrorKind::Protocol,
-            UpstreamSendState::Sent,
-        )
-    );
+    assert_eq!(chunks.len(), 1);
+    assert!(chunks[0].is_ok());
 }
 
 #[tokio::test]
@@ -534,6 +636,24 @@ async fn execute_inference(transport: &ReqwestGrokInferenceTransport, origin: &U
 struct CountingEndpointPolicy {
     inner: Arc<dyn GrokEndpointPolicy>,
     inference_client_builds: Arc<AtomicUsize>,
+    concurrency: Option<Arc<BuildConcurrency>>,
+}
+
+#[derive(Debug)]
+struct BuildConcurrency {
+    delay: Duration,
+    active: AtomicUsize,
+    max_active: AtomicUsize,
+}
+
+impl BuildConcurrency {
+    fn new(delay: Duration) -> Self {
+        Self {
+            delay,
+            active: AtomicUsize::new(0),
+            max_active: AtomicUsize::new(0),
+        }
+    }
 }
 
 impl GrokEndpointPolicy for CountingEndpointPolicy {
@@ -549,7 +669,15 @@ impl GrokEndpointPolicy for CountingEndpointPolicy {
         timeout: Option<Duration>,
     ) -> Result<Client, GrokReqwestTransportBuildError> {
         self.inference_client_builds.fetch_add(1, Ordering::SeqCst);
-        self.inner.build_inference_client(timeout)
+        let Some(concurrency) = &self.concurrency else {
+            return self.inner.build_inference_client(timeout);
+        };
+        let active = concurrency.active.fetch_add(1, Ordering::SeqCst) + 1;
+        concurrency.max_active.fetch_max(active, Ordering::SeqCst);
+        std::thread::sleep(concurrency.delay);
+        let result = self.inner.build_inference_client(timeout);
+        concurrency.active.fetch_sub(1, Ordering::SeqCst);
+        result
     }
 
     fn validate_oauth(&self, url: &Url) -> bool {

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

@@ -273,6 +273,104 @@ fn explicit_session_should_merge_client_tools_into_the_native_cache_route() {
     assert_eq!(body.pointer("/tool_choice"), Some(&json!("auto")));
 }
 
+#[test]
+fn function_parameters_should_remove_only_nullable_object_roots() {
+    let request = raw_request(json!({
+        "model": "client",
+        "input": "hello",
+        "tools": [{
+            "type": "function",
+            "name": "automation_update",
+            "parameters": {
+                "anyOf": [
+                    {
+                        "type": "object",
+                        "properties": {
+                            "title": {
+                                "anyOf": [{"type": "string"}, {"type": "null"}]
+                            }
+                        },
+                        "required": ["title"],
+                        "additionalProperties": false
+                    },
+                    {"type": "null"}
+                ]
+            }
+        }]
+    }));
+
+    let encoded = GrokResponsesRequest::encode(&request, "grok-4.5", &client_key())
+        .expect("nullable object root");
+    let body = Value::Object(encoded.body().clone());
+
+    assert_eq!(
+        body.pointer("/tools/0/parameters/type"),
+        Some(&json!("object"))
+    );
+    assert_eq!(body.pointer("/tools/0/parameters/anyOf"), None);
+    assert_eq!(
+        body.pointer("/tools/0/parameters/properties/title/anyOf"),
+        Some(&json!([{"type": "string"}, {"type": "null"}]))
+    );
+}
+
+#[test]
+fn function_parameters_should_keep_local_object_refs_after_removing_null() {
+    let request = raw_request(json!({
+        "model": "client",
+        "input": "hello",
+        "tools": [{
+            "type": "function",
+            "name": "lookup",
+            "parameters": {
+                "$defs": {
+                    "Args": {
+                        "type": "object",
+                        "properties": {"query": {"type": "string"}}
+                    }
+                },
+                "anyOf": [{"$ref": "#/$defs/Args"}, {"type": "null"}]
+            }
+        }]
+    }));
+
+    let encoded = GrokResponsesRequest::encode(&request, "grok-4.5", &client_key())
+        .expect("nullable local object ref");
+    let body = Value::Object(encoded.body().clone());
+
+    assert_eq!(
+        body.pointer("/tools/0/parameters/type"),
+        Some(&json!("object"))
+    );
+    assert_eq!(
+        body.pointer("/tools/0/parameters/anyOf"),
+        Some(&json!([{"$ref": "#/$defs/Args"}]))
+    );
+}
+
+#[test]
+fn function_parameters_should_report_the_invalid_nullable_root_field() {
+    let request = raw_request(json!({
+        "model": "client",
+        "input": "hello",
+        "tools": [{
+            "type": "function",
+            "name": "invalid",
+            "parameters": {
+                "anyOf": [{"type": "object"}, {"type": "string"}, {"type": "null"}]
+            }
+        }]
+    }));
+
+    assert_eq!(
+        GrokResponsesRequest::encode(&request, "grok-4.5", &client_key())
+            .expect_err("nullable non-object root"),
+        GrokRequestEncodeError::InvalidRequestField {
+            field: "tools[].parameters"
+        }
+    );
+}
+
 #[test]
 fn explicit_session_should_enable_cache_after_codex_additional_tools_normalization() {
     let request = raw_request(json!({
@@ -721,7 +819,7 @@ fn compaction_history_should_reject_missing_empty_or_non_string_plaintext() {
         assert_eq!(
             GrokResponsesRequest::encode(&request, "grok-4.5", &client_key())
                 .expect_err("invalid compaction plaintext"),
-            GrokRequestEncodeError::InvalidRequestNormalization
+            GrokRequestEncodeError::InvalidRequestField { field: "input" }
         );
     }
 }
@@ -765,27 +863,42 @@ fn tool_search_history_should_load_returned_tools_at_the_original_turn() {
 
 #[test]
 fn unsupported_or_ambiguous_tool_contracts_should_fail_before_upstream_io() {
-    for body in [
-        json!({"model": "client", "input": "x", "tools": [{"type": "future_tool"}]}),
-        json!({"model": "client", "input": "x", "tools": [
-            {"type": "tool_search", "execution": "client"},
-            {"type": "tool_search", "execution": "server"}
-        ]}),
-        json!({"model": "client", "input": "x", "tools": [
-            {"type": "shell"}, {"type": "local_shell"}
-        ]}),
-        json!({"model": "client", "input": [{
-            "type": "apply_patch_call",
-            "call_id": "patch_1",
-            "operation": {"type": "update_file", "path": "a.txt"}
-        }]}),
-        json!({"model": "client", "input": [{"type": "compaction_trigger"}]}),
+    for (body, field) in [
+        (
+            json!({"model": "client", "input": "x", "tools": [{"type": "future_tool"}]}),
+            "tools",
+        ),
+        (
+            json!({"model": "client", "input": "x", "tools": [
+                {"type": "tool_search", "execution": "client"},
+                {"type": "tool_search", "execution": "server"}
+            ]}),
+            "tools",
+        ),
+        (
+            json!({"model": "client", "input": "x", "tools": [
+                {"type": "shell"}, {"type": "local_shell"}
+            ]}),
+            "tools",
+        ),
+        (
+            json!({"model": "client", "input": [{
+                "type": "apply_patch_call",
+                "call_id": "patch_1",
+                "operation": {"type": "update_file", "path": "a.txt"}
+            }]}),
+            "input",
+        ),
+        (
+            json!({"model": "client", "input": [{"type": "compaction_trigger"}]}),
+            "input",
+        ),
     ] {
         let request = raw_request(body);
         assert_eq!(
             GrokResponsesRequest::encode(&request, "grok-4.5", &client_key())
                 .expect_err("invalid tool contract"),
-            GrokRequestEncodeError::InvalidRequestNormalization
+            GrokRequestEncodeError::InvalidRequestField { field }
         );
     }
 }

deploy/README.md

@@ -19,21 +19,21 @@ sudo chown "$(id -u):10001" deploy/config.yaml
 chmod 0640 deploy/config.yaml
 ```
 
-分别执行三次以下命令：
+为 PostgreSQL 与 Redis 分别生成一个密码：
 
 ```bash
 openssl rand -hex 24
 ```
 
-把结果写入 `deploy/config.yaml`：
+把两个 48 位十六进制结果分别写入 `deploy/config.yaml` 的：
 
 - `store.database.password`
 - `store.redis.password`
-- `admin.default_password`
 
-PostgreSQL 与 Redis 密码必须是 48 位十六进制字符。管理员初始化密码至少 12 位、不能是
-常见弱口令且不能包含 `$`。Compose 通过 `config.yaml` 的凭据桥接区引用同一密码，三个密码
-都不需要额外导出为环境变量，也不能嵌入连接 URL。
+另行设置 `admin.default_password`。它至少需要 12 个字符，不能是常见弱口令，也不能包含 `$`。
+
+PostgreSQL 与 Redis 密码必须是 48 位十六进制字符。Compose 通过 `config.yaml` 的凭据桥接区
+引用同一密码；三个值都不需要额外导出为环境变量，数据库和 Redis 密码也不能嵌入连接 URL。
 
 Linux 上应用容器以 `10001:10001` 运行，需要允许该组写入应用数据和日志目录：
 
@@ -45,6 +45,11 @@ chmod 0770 .runtime/data .runtime/logs
 `config.yaml` 通过 Compose `configs` 只读挂载。普通 Compose 对本地文件保留宿主机的
 UID/GID 和 mode，因此配置由当前用户持有，并只向容器组 `10001` 开放读取权限。
 
+模板中的 `openai` / `xai` 只保留请求画像启动基线。OpenAI 的上游地址、WebSocket 池、额度刷新
+与 OAuth 设置，以及 xAI 的 OAuth、额度和模型目录策略，均由各自 Provider 使用代码内默认值管理；
+模板不重复列出这些默认项。运行后，OpenAI CLI/Desktop 与 xAI CLI 的官方发布检查只更新进程内
+版本字段，不回写 `config.yaml`；检查失败时继续使用上一份有效画像。
+
 ## 启动
 
 ```bash
@@ -78,80 +83,27 @@ cargo run -p codex-proxy-rs
 后端会从当前目录向上查找 `deploy/config.yaml`。相对数据和日志目录以该文件所在目录解析；
 Compose 只把监听地址和数据库、Redis 地址固定覆盖为容器内部服务名。
 
-## 持久化
+## 持久化与备份
 
 Compose 使用以下绑定目录：
 
-- `.runtime/data` → 应用身份密钥、credential 加密 keyring 和更新状态
+- `.runtime/data` → OpenAI 会话锚点密钥、更新状态与临时更新目录
 - `.runtime/logs` → 应用文件日志
 - `.runtime/postgres` → PostgreSQL
 - `.runtime/redis` → Redis AOF
 
 普通 `docker compose down` 不会删除这些目录。删除 `.runtime` 会永久清除本地状态。
 
-PostgreSQL 由 `0001_initial.sql` 创建当前业务结构，Redis 保存可丢失的协调状态。
+PostgreSQL 是账号、Client Key、运行设置、请求记录与审计的权威存储；账号 credential 当前按
+Provider schema 以明文 JSON 保存于 PostgreSQL。Redis 只保存可重建、可过期的协调状态，例如
+会话亲和、lease、cooldown、OAuth pending flow 与套餐模型目录 cache。
+
+备份必须包含 `.runtime/postgres`。若希望保留短期 Redis 状态、OpenAI 会话亲和和更新状态，
+同时备份 `.runtime/redis` 与 `.runtime/data`；后两者丢失不会造成账号 credential 丢失，但会使
+缓存和会话锚点重新建立。
 
 完整运行时、Provider、revision 与恢复边界见 [架构文档](../docs/architecture.md)。
 
-## Credential 加密 keyring
-
-首次启动会在 `.runtime/data/credential_keyring` 生成：
-
-- `active_key_id`：新 envelope 唯一使用的 active key ID。
-- `envelope_keys/<key-id>.key`：active key 与只读历史解密 key。
-- `credential_fingerprint_hmac_key`：稳定 credential 去重域，不能随 envelope key 轮换。
-- `resource_pseudonym_hmac_key`：稳定资源伪名域，不能随 envelope key 轮换。
-
-目录权限固定为 `0700`，文件权限固定为 `0600`；符号链接、组/其他用户可读权限、无效 key ID
-或重复 key 材料都会让进程拒绝启动。Keyring 不进入 PostgreSQL，必须和 PostgreSQL 备份分开保存；
-丢失任何仍被数据库引用的 envelope key 后，对应 credential、Cookie 或 continuation 将无法恢复。
-
-Envelope key 使用“读旧写新”两阶段轮换。多节点必须为每个节点执行相同步骤：
-
-1. 生成相同的新 key ID 和 32-byte key，写入 `envelope_keys/<key-id>.key`，保持当前
-   `active_key_id` 不变。
-2. 滚动重启全部节点，使每个旧进程都先把新 key 加载为历史解密 key。
-3. 在全部节点原子替换 `active_key_id`，再滚动重启。已重启节点只用新 key 写入，尚未重启节点
-   仍能用第二步加载的 key 解密新 envelope。
-
-单机 Compose 可按以下方式准备和激活新 key；不要把命令输出或 key 文件内容写入日志：
-
-```bash
-new_key_id="ek-$(openssl rand -hex 16)"
-openssl rand -hex 32 | sudo sh -c \
-  'umask 077; cat > ".runtime/data/credential_keyring/envelope_keys/$1.key"' sh "$new_key_id"
-sudo chown 10001:10001 ".runtime/data/credential_keyring/envelope_keys/$new_key_id.key"
-
-# 先重启一次，让当前进程把新 key 作为历史 key 加载。
-docker compose -f deploy/compose.yaml restart codex-proxy-rs
-
-sudo env NEW_KEY_ID="$new_key_id" sh -c '
-  umask 077
-  printf "%s" "$NEW_KEY_ID" > .runtime/data/credential_keyring/active_key_id.next
-  chown 10001:10001 .runtime/data/credential_keyring/active_key_id.next
-  mv -f .runtime/data/credential_keyring/active_key_id.next \
-    .runtime/data/credential_keyring/active_key_id
-'
-docker compose -f deploy/compose.yaml restart codex-proxy-rs
-unset new_key_id
-```
-
-历史 key 只有在所有引用计数均为零、备份也满足同一退役策略后才能删除：
-
-```sql
-select 'upstream_credentials' as source, count(*) from upstream_credentials where secret_key_id = $1
-union all
-select 'codex_account_cookies', count(*) from codex_account_cookies where secret_key_id = $1
-union all
-select 'conversation_items', count(*) from conversation_items where secret_key_id = $1
-union all
-select 'continuation_bindings', count(*) from continuation_bindings where secret_key_id = $1;
-```
-
-删除仍被引用的历史 key 不会降级或尝试其他 key；解密会以 unknown key fail closed。确认零引用后，
-从全部节点删除对应 `.key` 文件并滚动重启，才算完成退役。Fingerprint 与 resource pseudonym 两个
-HMAC key 不属于此轮换流程，不能用 envelope key 替换。
-
 ## 密码语义
 
 - `admin.default_password` 只在首次创建管理员时使用。
@@ -159,9 +111,10 @@ HMAC key 不属于此轮换流程，不能用 envelope key 替换。
 - Redis 在每次容器创建时使用 `redis.password`。
 
 已有 PostgreSQL 数据目录后，直接修改 `database.password` 不会修改数据库用户密码，只会导致
-应用无法连接。轮换时必须先在 PostgreSQL 中修改用户密码，再同步更新 `config.yaml`。
+应用无法连接。轮换时必须先在 PostgreSQL 中修改用户密码，再同步更新 `config.yaml`。Redis
+密码变更后需要使用新密码重建或重新配置 Redis 数据目录。
 
-## 构建与升级
+## 镜像升级与源码构建
 
 ```bash
 docker compose -f deploy/compose.yaml build codex-proxy-rs
@@ -175,6 +128,14 @@ docker compose -f deploy/compose.yaml pull codex-proxy-rs
 docker compose -f deploy/compose.yaml up -d --no-build
 ```
 
+仓库维护者发布新版本时必须从干净且已同步上游的分支运行：
+
+```bash
+release/publish <version>
+```
+
+该脚本负责更新 `release/version.yaml`、创建版本提交和带注释 tag，并原子推送分支与 tag。
+
 构建元数据仍可作为一次性进程环境传入，不需要 `.env` 文件：
 
 ```bash
@@ -184,22 +145,22 @@ CPR_BUILD_TIME="$(date -u +%Y-%m-%dT%H:%M:%SZ)" \
 docker compose -f deploy/compose.yaml build codex-proxy-rs
 ```
 
-### 管理端 Release 更新
+### 管理端在线更新
 
-Compose 已显式装配正式更新所需的运行参数：
+Compose 已显式装配正式发布构建所需的运行参数：
 
 - `CPR_UPDATE_REPOSITORY`：只接受 `owner/repository`；默认 `zyycn/codex-proxy-rs`。
 - `CPR_GITHUB_API_BASE`：正式环境必须为 `https://api.github.com/repos`。
 - `CPR_UPDATE_CHANNEL`：`stable` 会拒绝 prerelease。
-- `CPR_UPDATE_EXE_PATH`、`CPR_WEB_DIST_DIR`：分别指向容器内二进制和旧前端静态目录。
+- `CPR_UPDATE_EXE_PATH`、`CPR_WEB_DIST_DIR`：分别指向容器内二进制和前端静态目录。
 - `CPR_UPDATE_TEMP_DIR`、`CPR_UPDATE_STATE_FILE`、`CPR_UPDATE_LOCK_FILE`：全部位于持久化的
   `.runtime/data`。
 - `CPR_ENABLE_SELF_RESTART=true`：更新或回滚完成后允许管理端请求重启；Docker 进程退出后由
   Compose 的 `restart: unless-stopped` 拉起新进程。
 
-Release 必须提供当前 OS/架构的 `codex-proxy-rs_<version>_<os>_<arch>.tar.gz` 和
-`checksums.txt`。服务会在替换前再次查询远端最新版本，校验下载 host、声明大小、SHA-256 和 tar
-路径；二进制或静态资源任一替换失败时恢复旧文件。成功后的旧二进制和旧静态目录分别保留为
+Release 必须提供当前 OS/架构的 `codex-proxy-rs_<version>_<os>_<arch>.tar.gz` 与
+`checksums.txt`。服务会在替换前再次查询远端最新版本，校验下载 host、声明大小、SHA-256 和
+归档路径；二进制或静态资源任一替换失败时恢复旧文件。成功后的旧二进制和旧静态目录分别保留为
 `*.backup`，管理端 rollback 会交换当前文件与这份备份。更新状态和跨进程锁可在以下位置排查：
 
 ```text

deploy/config.example.yaml

@@ -5,7 +5,7 @@
 #   sudo chown "$(id -u):10001" deploy/config.yaml
 #   chmod 640 deploy/config.yaml
 #
-# 三个密码建议分别执行 `openssl rand -hex 24` 生成。
+# PostgreSQL 与 Redis 密码分别执行 `openssl rand -hex 24` 生成；管理员初始密码单独设置。
 
 schema_version: 1
 
@@ -51,8 +51,9 @@ api:
   request_id_header: 'x-request-id'
 
 openai:
+  # 上游地址、WebSocket 池、额度刷新和 OAuth 使用 Provider 内置默认值；模板只保留画像启动基线。
   wire_profile:
-    # 该画像来自 Desktop 制品与 bundled Codex Core 的联合审计。
+    # Provider 每 24 小时读取官方 npm metadata 与 Desktop appcast，仅在进程内更新版本字段。
     originator: 'Codex Desktop'
     codex_version: '0.145.0'
     desktop_version: '26.721.31836'
@@ -64,8 +65,9 @@ openai:
     verified_at: '2026-07-25T00:00:00+08:00'
 
 xai:
+  # OAuth 端点、额度与模型目录策略由 xAI Provider 固定管理；这里只声明请求画像启动基线。
   wire_profile:
-    # 对照 grok2api 当前推荐构建核验；更新前应重新抓包确认。
+    # Provider 每 24 小时读取官方 npm metadata，仅在进程内更新 client_version。
     client_identifier: 'grok-shell'
     client_version: '0.2.106'
     client_mode: 'headless'

docs/api.md

@@ -0,0 +1,236 @@
+# Codex Proxy RS 接口
+
+本文列出 v3.0.0 当前公开 HTTP 接口。路由事实以
+`backend/crates/gateway-api/src` 中的 router 为准。
+
+## 1. 鉴权与公共约定
+
+### 客户端接口
+
+所有 `/v1/*` 请求都使用管理端创建的 Client Key：
+
+```http
+Authorization: Bearer sk_...
+```
+
+Client Key 固定绑定一个 Provider：`openai` 或 `xai`。同一次请求不会跨 Provider fallback。
+
+### 管理接口
+
+除登录、会话状态和登出外，所有 `/api/admin/*` 请求都需要以下任一鉴权方式：
+
+- 浏览器登录后得到的 `cpr_admin_session` Cookie；
+- `x-api-key: <admin-api-key>`。
+
+已鉴权的管理请求还必须携带非空 `x-request-id`。管理端响应统一带
+`Cache-Control: no-store`。
+
+普通成功响应使用以下信封：
+
+```json
+{
+  "code": 200,
+  "message": "OK",
+  "data": {}
+}
+```
+
+错误响应仍使用 `code`、`message`、`data`，其中 `data` 为 `null`。稳定业务码包括参数错误
+`40001`、会话缺失 `40101`、资源不存在 `40401`、revision 冲突 `40901`、上游失败 `50201`
+和依赖不可用 `50301`。
+
+### 配置 revision
+
+会改变路由快照或安全配置的 mutation 使用 `expectedConfigRevision` 做乐观并发控制。客户端应从
+账号、Client Key 或设置查询结果读取最新 `configRevision`；收到 `40901` 后重新读取，不应盲目重试
+旧 revision。
+
+## 2. 健康检查
+
+| 方法 | 路由 | 鉴权 | 说明 |
+| --- | --- | --- | --- |
+| `GET` | `/healthz` | 无 | Core、Store 和后台任务健康时返回 `204`，否则返回 `503` |
+
+## 3. Responses 与模型目录
+
+请求正文上限为 16 MiB。
+
+| 方法 | 路由 | 说明 |
+| --- | --- | --- |
+| `POST` | `/v1/responses` | OpenAI Responses JSON；`stream=true` 返回 SSE，否则返回完整 JSON |
+| `GET` | `/v1/responses` | 通过 HTTP Upgrade 建立 Responses WebSocket |
+| `POST` | `/v1/responses/review` | 使用同一 Responses 合同发起 review 子代理请求 |
+| `GET` | `/v1/models` | 返回当前 Client Key 所属 Provider 的可用公开模型 |
+| `GET` | `/v1/models/catalog` | 返回 Codex 客户端使用的模型目录 |
+| `GET` | `/v1/models/{model_id}/info` | 返回 Codex 客户端使用的单模型信息 |
+| `GET` | `/v1/models/{model_id}` | 返回 OpenAI 兼容的单模型详情 |
+
+OpenAI 路径保留客户端 Responses wire 语义；xAI 所需的协议转换只在 xAI Provider 内完成。模型映射
+是全局精确映射，未命中时模型名原样交给所属 Provider。
+
+## 4. 管理员认证
+
+| 方法 | 路由 | 请求 | 说明 |
+| --- | --- | --- | --- |
+| `POST` | `/api/admin/auth/login` | `{ username?, password }` | 创建管理员会话并设置 Cookie |
+| `GET` | `/api/admin/auth/status` | 无 | 返回当前 Cookie 是否已认证 |
+| `POST` | `/api/admin/auth/logout` | 无 | 删除当前会话并清除 Cookie |
+
+## 5. 账号
+
+账号 API 使用统一路由，不存在 Provider Instance 或 Provider 专属账号路由。需要 Provider 的请求只接受
+`provider: "openai" | "xai"`。
+
+| 方法 | 路由 | 主要 query/body | 说明 |
+| --- | --- | --- | --- |
+| `GET` | `/api/admin/accounts` | `page`、`pageSize`、`provider`、`search`、`status`、排序字段 | 分页查询账号与汇总 |
+| `GET` | `/api/admin/accounts/detail` | `accountId` | 查询账号详情、额度和本地用量 |
+| `GET` | `/api/admin/accounts/export` | `accountIds`、`confirm=export_sensitive_accounts` | 显式导出最多 200 个账号的敏感 Provider 文档 |
+| `POST` | `/api/admin/accounts/import` | `{ provider, expectedConfigRevision, data }` | 导入或按上游身份更新账号 |
+| `POST` | `/api/admin/accounts/refresh` | `{ accountId, expectedConfigRevision }` | 手工刷新账号 credential |
+| `POST` | `/api/admin/accounts/rotate` | OpenAI rotation 字段 | 手工替换 OpenAI OAuth token |
+| `POST` | `/api/admin/accounts/enable` | `{ provider, accountId, expectedConfigRevision }` | 启用账号 |
+| `POST` | `/api/admin/accounts/disable` | `{ provider, accountId, expectedConfigRevision }` | 禁用账号 |
+| `POST` | `/api/admin/accounts/delete` | `{ provider, accountIds, expectedConfigRevision }` | 批量删除 1–200 个账号 |
+| `GET` | `/api/admin/accounts/quota` | `accountId` | 读取当前额度，不强制访问上游 |
+| `POST` | `/api/admin/accounts/quota/refresh` | `{ accountId }` | 立即刷新该账号额度 |
+| `GET` | `/api/admin/accounts/models` | `accountId` | 优先读取该 Provider + 套餐的模型 cache，缺失时有限实时拉取 |
+| `POST` | `/api/admin/accounts/models/refresh` | `{ accountId }` | 强制拉取最新模型并覆盖 cache |
+| `GET` | `/api/admin/accounts/connection-test` | `accountId`、`modelId` | 通过 SSE 返回实时连接测试事件，不作为业务 Responses 用量记录 |
+| `POST` | `/api/admin/accounts/oauth/start` | OAuth start 字段 | 创建 OpenAI 或 xAI OAuth flow |
+| `POST` | `/api/admin/accounts/oauth/complete` | `{ provider, flowId, callbackUrl }` | 消费 OAuth callback 并写入账号 |
+
+账号列表支持以下稳定值：
+
+- `provider`: `all`、`openai`、`xai`；
+- `status`: `active`、`expired`、`quota_exhausted`、`disabled`、`banned`；
+- `sortBy`: `email`、`status`、`planType`、`usage`、`lastUsedAt`、`expiresAt`；
+- `sortDirection`: `asc`、`desc`。
+
+导入的 `data` 必须是 JSON object，内部 schema 由目标 Provider 独占解释：
+
+- OpenAI 接受 OAuth 账号文档、账号 bundle 和 Agent Identity；
+- xAI 从单账号 object 或 `accounts` 数组中提取 OAuth token；包装中的代理、并发、优先级等字段不参与认证；
+- xAI API Key 不是受支持的账号 credential；
+- 导入仍会通过 Provider 的身份或刷新链路校验，不能只凭文件外形写入账号。
+
+OpenAI 的 CPR 导出保持 OAuth 账号的既有字段；Agent Identity 账号则输出
+`authMode: "agentIdentity"`、`agentRuntimeId`、`agentPrivateKey` 和可选的
+`taskId`，不伪造 OAuth token 或过期时间。一个 `sourceFormat: "cpr"` 文档可以包含两种
+认证类型，导入时按每个账号的认证模式分别解析。
+
+```json
+{
+  "sourceFormat": "cpr",
+  "accounts": [{
+    "id": "acct_...",
+    "accountId": "...",
+    "userId": "...",
+    "authMode": "agentIdentity",
+    "agentRuntimeId": "...",
+    "agentPrivateKey": "...",
+    "taskId": "..."
+  }]
+}
+```
+
+OpenAI rotation 请求字段为：
+
+```json
+{
+  "provider": "openai",
+  "accountId": "acct_...",
+  "expectedConfigRevision": 1,
+  "expectedCredentialRevision": 1,
+  "accessToken": "...",
+  "refreshToken": "..."
+}
+```
+
+OAuth start 使用：
+
+```json
+{
+  "provider": "openai",
+  "expectedConfigRevision": 1,
+  "name": "account name",
+  "accountId": null,
+  "expectedCredentialRevision": null
+}
+```
+
+重新授权已有账号时，`accountId` 与 `expectedCredentialRevision` 必须同时提供。
+
+## 6. Client Key
+
+| 方法 | 路由 | 主要 query/body | 说明 |
+| --- | --- | --- | --- |
+| `GET` | `/api/admin/client-keys` | `cursor`、`limit`、`search`、`sortBy`、`sortDirection` | 游标分页查询 |
+| `POST` | `/api/admin/client-keys/create` | 创建字段 | 创建绑定 Provider 的 Client Key |
+| `GET` | `/api/admin/client-keys/reveal` | `id` | 显式读取完整明文 Key |
+| `POST` | `/api/admin/client-keys/update` | 更新字段 | 原子更新名称、Provider 和限额 |
+| `POST` | `/api/admin/client-keys/enable` | `{ id, expectedConfigRevision }` | 启用 |
+| `POST` | `/api/admin/client-keys/disable` | `{ id, expectedConfigRevision }` | 禁用 |
+| `POST` | `/api/admin/client-keys/delete` | `{ id, expectedConfigRevision }` | 删除 |
+
+创建字段为 `expectedConfigRevision`、`name`、可选 `label`、`providerKind`、`maxConcurrency`、
+`requestsPerMinute`。更新请求再增加 `id`。创建和 reveal 响应会返回完整明文 Key，调用方必须立即安全保存。
+
+## 7. 运行设置
+
+| 方法 | 路由 | 说明 |
+| --- | --- | --- |
+| `GET` | `/api/admin/settings` | 读取运行设置与 `configRevision` |
+| `POST` | `/api/admin/settings/update` | 原子替换全部运行设置 |
+| `GET` | `/api/admin/settings/admin-api-key` | 只返回管理 API Key 是否存在 |
+| `POST` | `/api/admin/settings/admin-api-key/delete` | 删除管理 API Key |
+| `POST` | `/api/admin/settings/admin-api-key/regenerate` | 重新生成并一次性返回完整管理 API Key |
+
+设置更新字段包括：
+
+```text
+expectedConfigRevision
+modelMappings
+refreshMarginSeconds
+refreshConcurrency
+maxConcurrentPerAccount
+requestIntervalMs
+rotationStrategy
+usageRetentionDays
+opsEventRetentionDays
+auditRetentionDays
+```
+
+`rotationStrategy` 可取 `smart`、`quota_reset_priority`、`round_robin`、`sticky`。
+
+## 8. Dashboard、用量与错误
+
+| 方法 | 路由 | 说明 |
+| --- | --- | --- |
+| `GET` | `/api/admin/dashboard/summary` | Dashboard 汇总；支持 `kind`、`startTime`、`endTime` |
+| `GET` | `/api/admin/dashboard/trend` | Dashboard 趋势；`kind=usage|latency|errors` |
+| `GET` | `/api/admin/usage/records` | 请求记录分页列表 |
+| `GET` | `/api/admin/usage/records/detail` | 按 `id` 查询请求详情 |
+| `GET` | `/api/admin/usage/records/summary` | 当前筛选条件的请求汇总 |
+| `GET` | `/api/admin/usage/insights/overview` | 用量、成本与成功率洞察 |
+| `GET` | `/api/admin/usage/insights/diagnostics` | 按维度聚合诊断 |
+| `GET` | `/api/admin/operations/errors` | 运维错误分页列表 |
+
+用量查询可组合页码/游标、时间范围、Provider、Client Key、账号、模型、route、transport、状态码、
+request/response/upstream ID、outcome 与搜索文本。诊断 `dimension` 可取 `model`、`account`、
+`apiKey`、`provider`、`transport`、`failureClass`、`status`。
+
+## 9. 版本、更新与重启
+
+| 方法 | 路由 | 主要 query/body | 说明 |
+| --- | --- | --- | --- |
+| `GET` | `/api/admin/system/version` | 无 | 当前构建、部署模式和可用更新 |
+| `GET` | `/api/admin/system/update/detail` | `refresh=true|false` | 读取或强制刷新 Release 详情 |
+| `GET` | `/api/admin/system/update/events` | 无 | SSE 更新事件流 |
+| `POST` | `/api/admin/system/update` | 可选 `{ targetVersion }` | 开始在线更新 |
+| `GET` | `/api/admin/system/update/status` | 无 | 查询当前更新或回滚状态 |
+| `POST` | `/api/admin/system/rollback` | 无 | 回滚到保留的上一版本 |
+| `POST` | `/api/admin/system/restart` | 无 | 请求进程重启 |
+
+在线更新仅在当前部署模式、Release 资产和进程重启能力都满足要求时可用。仓库发版流程见
+[部署文档](../deploy/README.md) 与根目录 [README](../README.md)。

docs/architecture.md

@@ -130,7 +130,7 @@ credential 更新使用 `credential_revision` CAS。认证永久失败、封禁
 | `admin_audit_events` | 管理 mutation 审计 |
 | `client_api_keys` | 客户端鉴权、限额与授权范围 |
 | `runtime_settings` | 全局配置与 `config_revision` |
-| `provider_accounts` | 账号资料、加密 credential、revision、quota、cooldown |
+| `provider_accounts` | 账号资料、Provider-owned 明文 credential JSON、revision、quota、cooldown |
 | `model_requests` | 请求、attempt、计费、交付与恢复事实 |
 | `ops_events` | 脱敏运行事件 |
 
@@ -138,7 +138,8 @@ credential 更新使用 `credential_revision` CAS。认证永久失败、封禁
 
 - 一个事实只存一次；可表达关系使用真实 FK 与支持索引。
 - Provider 差异只进入受 schema/version 校验的 JSONB 边界。
-- secret 只保存加密 envelope；日志、Debug、API 和 audit 不输出 secret。
+- credential JSON 以 Provider 自己的 schema 明文保存在 `provider_credentials_json`；日志、Debug、API 和 audit
+  不输出其中的 secret，Provider 只在需要发往上游时把值包装为运行时 secret。
 - stale recovery 只把超时 `running` 请求收敛为 `incomplete`，不重放业务请求。
 - `ops_events` 当前逐条同步持久化，`occurrence_count=1`；OpsFlush 可禁用，不存在隐藏内存聚合权威。
 - retention 只删除已满足保留规则的历史事实，不改变运行中请求。

frontend/index.html

@@ -6,7 +6,7 @@
       name="viewport"
       content="width=device-width, initial-scale=1.0, maximum-scale=1.0, user-scalable=no, viewport-fit=cover"
     />
-    <link rel="icon" type="image/svg+xml" href="/favicon.svg" />
+    <link rel="icon" type="image/svg+xml" href="/favicon.svg?v=20260726" />
     <title>Codex Proxy RS</title>
   </head>
   <body>

frontend/public/favicon.svg

@@ -1,9 +1,9 @@
-<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 64 64">
-  <rect width="64" height="64" rx="16" fill="#111827"/>
-  <g transform="translate(10 8) scale(1.85)" fill="none" stroke="#fff" stroke-width="2.2" stroke-linecap="round" stroke-linejoin="round">
-    <path d="M12 5c.67 0 1.35.09 2 .26 1.78-2 5.03-2.84 6.42-2.26 1.4.58-.42 7-.42 7 .57 1.07 1 2.24 1 3.44C21 17.9 16.97 21 12 21s-9-3.1-9-7.56c0-1.25.5-2.4 1-3.44 0 0-1.89-6.42-.5-7S8.22 3.26 10 5.26A8.3 8.3 0 0 1 12 5Z"/>
-    <path d="M8 14v.5"/>
-    <path d="M16 14v.5"/>
-    <path d="M11.25 16.25h1.5L12 17l-.75-.75Z"/>
+<svg xmlns="http://www.w3.org/2000/svg" viewBox="48 48 416 416" role="img" aria-labelledby="title">
+  <title id="title">Codex Proxy</title>
+  <rect x="48" y="48" width="416" height="416" rx="88" fill="#1c2a3d" />
+  <g fill="none" stroke="#fff" stroke-linecap="round" stroke-width="36">
+    <path d="M156 307A112 112 0 0 1 307 156" />
+    <path d="M356 205A112 112 0 0 1 205 356" />
   </g>
+  <path d="m407 105-119 153-183 149 124-157Z" fill="#fff" />
 </svg>

frontend/src/components/base/BaseButton.vue

@@ -108,7 +108,11 @@ const labelClasses = computed(() => [
     :title="title || (iconOnly ? label : undefined)"
     :aria-busy="loading"
   >
-    <LoaderCircle v-if="loading" class="animate-spin" :size="loadingIconSize[size]" />
+    <LoaderCircle
+      v-if="loading"
+      class="shrink-0 animate-spin origin-center transform-view will-change-transform"
+      :size="loadingIconSize[size]"
+    />
     <span v-if="$slots.icon && !loading" class="inline-flex shrink-0">
       <slot name="icon" />
     </span>

frontend/src/components/base/BaseInput.vue

@@ -24,6 +24,11 @@ const props = withDefaults(
 
 const model = defineModel<string>({ default: '' })
 
+// number 输入也保持字符串契约，避免原生 v-model 隐式转换为 number。
+function updateModel(event: Event) {
+  model.value = (event.target as HTMLInputElement).value
+}
+
 const containerClasses = computed(() => [
   'relative inline-flex h-(--cp-input-height-default) w-full min-w-0 items-center gap-2.5 overflow-visible rounded-(--cp-input-radius-base) border-0 px-3.5 text-[13px] text-(--cp-text-primary) shadow-(--cp-shadow-input) transition-[background-color,box-shadow,color] duration-[160ms]',
   props.disabled
@@ -64,7 +69,7 @@ const inputClasses = computed(() => [
         <slot name="prefix" />
       </span>
       <input
-        v-model="model"
+        :value="model"
         class="base-input__field"
         :class="inputClasses"
         :name="name"
@@ -74,6 +79,7 @@ const inputClasses = computed(() => [
         :autocomplete="autocomplete"
         :aria-label="ariaLabel"
         :aria-invalid="error ? 'true' : undefined"
+        @input="updateModel"
       >
       <span v-if="$slots.suffix" :class="iconClasses">
         <slot name="suffix" />

frontend/src/components/base/BasePageHeader.vue

@@ -6,7 +6,7 @@ defineProps<{
 </script>
 
 <template>
-  <header class="flex min-h-17 shrink-0 items-start justify-between gap-4">
+  <header class="flex min-h-17 shrink-0 flex-wrap items-start justify-between gap-4">
     <div class="min-w-0">
       <h1 class="m-0 text-[34px] leading-[1.15] font-extrabold text-(--cp-text-primary)">
         {{ title }}
@@ -20,7 +20,10 @@ defineProps<{
         </slot>
       </p>
     </div>
-    <div v-if="$slots.actions" class="mt-0.5 flex shrink-0 items-center gap-2">
+    <div
+      v-if="$slots.actions"
+      class="mt-0.5 flex shrink-0 items-center gap-2 max-sm:w-full max-sm:justify-end"
+    >
       <slot name="actions" />
     </div>
   </header>

frontend/src/components/base/BaseScrollbar.vue

@@ -1,7 +1,7 @@
 <script setup lang="ts">
-import { useEventListener, useResizeObserver, useScroll, useTimeoutFn } from '@vueuse/core'
+import { useEventListener, useResizeObserver, useTimeoutFn } from '@vueuse/core'
 import { clamp } from 'es-toolkit'
-import { computed, nextTick, onMounted, shallowRef, useTemplateRef } from 'vue'
+import { computed, nextTick, onBeforeUnmount, onMounted, shallowRef, useTemplateRef } from 'vue'
 
 const props = withDefaults(
   defineProps<{
@@ -42,11 +42,26 @@ const verticalTrackHovering = shallowRef(false)
 const horizontalTrackHovering = shallowRef(false)
 const dragging = shallowRef(false)
 const horizontalDragging = shallowRef(false)
+const dragDocument = computed(() =>
+  dragging.value || horizontalDragging.value ? document : null,
+)
+
+interface AxisMetrics {
+  scrollRange: number
+  thumbRange: number
+}
+
+const verticalMetrics: AxisMetrics = { scrollRange: 0, thumbRange: 0 }
+const horizontalMetrics: AxisMetrics = { scrollRange: 0, thumbRange: 0 }
 
 let dragStartY = 0
 let dragStartScrollTop = 0
 let horizontalDragStartX = 0
 let horizontalDragStartScrollLeft = 0
+let scrollTopPosition = 0
+let scrollLeftPosition = 0
+let scrollFrameId: number | undefined
+let rootHovering = false
 const { start: startHideTimer, stop: stopHideTimer } = useTimeoutFn(hideScrollbar, 1600, {
   immediate: false,
 })
@@ -88,30 +103,6 @@ const horizontalTrackInsetClass = computed(() =>
   props.trackInset === 'none' ? 'bottom-0 left-0' : 'bottom-1 left-1',
 )
 
-function trackHeight(wrap: HTMLElement) {
-  return clamp(wrap.clientHeight - 8, 0, Number.POSITIVE_INFINITY)
-}
-
-function maxScrollTop(wrap: HTMLElement) {
-  return clamp(wrap.scrollHeight - wrap.clientHeight, 0, Number.POSITIVE_INFINITY)
-}
-
-function trackWidth(wrap: HTMLElement) {
-  return clamp(wrap.clientWidth - 8, 0, Number.POSITIVE_INFINITY)
-}
-
-function maxScrollLeft(wrap: HTMLElement) {
-  return clamp(wrap.scrollWidth - wrap.clientWidth, 0, Number.POSITIVE_INFINITY)
-}
-
-function maxThumbTop(wrap: HTMLElement) {
-  return clamp(trackHeight(wrap) - thumbHeight.value, 0, Number.POSITIVE_INFINITY)
-}
-
-function maxHorizontalThumbLeft(wrap: HTMLElement) {
-  return clamp(trackWidth(wrap) - horizontalThumbWidth.value, 0, Number.POSITIVE_INFINITY)
-}
-
 function showScrollbar() {
   visible.value = true
 }
@@ -122,7 +113,9 @@ function clearHideTimer() {
 
 function scheduleHideScrollbar() {
   clearHideTimer()
-  startHideTimer()
+  if (!rootHovering) {
+    startHideTimer()
+  }
 }
 
 function hideScrollbar() {
@@ -135,6 +128,16 @@ function activateScrollbar() {
   scheduleHideScrollbar()
 }
 
+function handleRootMouseEnter() {
+  rootHovering = true
+  activateScrollbar()
+}
+
+function handleRootMouseLeave() {
+  rootHovering = false
+  hideScrollbar()
+}
+
 type ScrollbarAxis = 'vertical' | 'horizontal'
 
 function setTrackHovering(axis: ScrollbarAxis, hovering: boolean) {
@@ -163,48 +166,90 @@ function update() {
     return
   }
 
-  updateVerticalScrollbar(wrap)
-  updateHorizontalScrollbar(wrap)
+  measureVerticalScrollbar(wrap)
+  measureHorizontalScrollbar(wrap)
+  syncScrollbarPosition(wrap, true)
 }
 
-function updateVerticalScrollbar(wrap: HTMLElement) {
+function measureVerticalScrollbar(wrap: HTMLElement) {
   if (!props.vertical) {
+    verticalMetrics.scrollRange = 0
+    verticalMetrics.thumbRange = 0
     thumbHeight.value = 0
-    thumbTop.value = 0
     return
   }
 
-  const scrollRange = maxScrollTop(wrap)
-  const availableTrackHeight = trackHeight(wrap)
+  const clientHeight = wrap.clientHeight
+  const scrollHeight = wrap.scrollHeight
+  const scrollRange = clamp(scrollHeight - clientHeight, 0, Number.POSITIVE_INFINITY)
+  const availableTrackHeight = clamp(clientHeight - 8, 0, Number.POSITIVE_INFINITY)
   if (scrollRange <= overflowTolerance || availableTrackHeight <= 0) {
     wrap.scrollTop = 0
+    verticalMetrics.scrollRange = 0
+    verticalMetrics.thumbRange = 0
     thumbHeight.value = 0
-    thumbTop.value = 0
     return
   }
-  const ratio = wrap.clientHeight / wrap.scrollHeight
-  thumbHeight.value = clamp(availableTrackHeight * ratio, 32, availableTrackHeight)
-  thumbTop.value = (wrap.scrollTop / scrollRange) * maxThumbTop(wrap)
+
+  const ratio = clientHeight / scrollHeight
+  const nextThumbHeight = clamp(availableTrackHeight * ratio, 32, availableTrackHeight)
+  verticalMetrics.scrollRange = scrollRange
+  verticalMetrics.thumbRange = availableTrackHeight - nextThumbHeight
+  thumbHeight.value = nextThumbHeight
 }
 
-function updateHorizontalScrollbar(wrap: HTMLElement) {
+function measureHorizontalScrollbar(wrap: HTMLElement) {
   if (!props.horizontal) {
+    horizontalMetrics.scrollRange = 0
+    horizontalMetrics.thumbRange = 0
     horizontalThumbWidth.value = 0
-    horizontalThumbLeft.value = 0
     return
   }
 
-  const scrollRange = maxScrollLeft(wrap)
-  const availableTrackWidth = trackWidth(wrap)
+  const clientWidth = wrap.clientWidth
+  const scrollWidth = wrap.scrollWidth
+  const scrollRange = clamp(scrollWidth - clientWidth, 0, Number.POSITIVE_INFINITY)
+  const availableTrackWidth = clamp(clientWidth - 8, 0, Number.POSITIVE_INFINITY)
   if (scrollRange <= overflowTolerance || availableTrackWidth <= 0) {
     wrap.scrollLeft = 0
+    horizontalMetrics.scrollRange = 0
+    horizontalMetrics.thumbRange = 0
     horizontalThumbWidth.value = 0
-    horizontalThumbLeft.value = 0
     return
   }
-  const ratio = wrap.clientWidth / wrap.scrollWidth
-  horizontalThumbWidth.value = clamp(availableTrackWidth * ratio, 32, availableTrackWidth)
-  horizontalThumbLeft.value = (wrap.scrollLeft / scrollRange) * maxHorizontalThumbLeft(wrap)
+
+  const ratio = clientWidth / scrollWidth
+  const nextThumbWidth = clamp(availableTrackWidth * ratio, 32, availableTrackWidth)
+  horizontalMetrics.scrollRange = scrollRange
+  horizontalMetrics.thumbRange = availableTrackWidth - nextThumbWidth
+  horizontalThumbWidth.value = nextThumbWidth
+}
+
+function updateVerticalThumbPosition(scrollTop: number) {
+  const { scrollRange, thumbRange } = verticalMetrics
+  thumbTop.value
+    = scrollRange > 0 ? (clamp(scrollTop, 0, scrollRange) / scrollRange) * thumbRange : 0
+}
+
+function updateHorizontalThumbPosition(scrollLeft: number) {
+  const { scrollRange, thumbRange } = horizontalMetrics
+  horizontalThumbLeft.value
+    = scrollRange > 0 ? (clamp(scrollLeft, 0, scrollRange) / scrollRange) * thumbRange : 0
+}
+
+function syncScrollbarPosition(wrap: HTMLElement, force = false) {
+  const scrollTop = wrap.scrollTop
+  const scrollLeft = wrap.scrollLeft
+
+  if (force || scrollTop !== scrollTopPosition) {
+    updateVerticalThumbPosition(scrollTop)
+  }
+  if (force || scrollLeft !== scrollLeftPosition) {
+    updateHorizontalThumbPosition(scrollLeft)
+  }
+
+  scrollTopPosition = scrollTop
+  scrollLeftPosition = scrollLeft
 }
 
 async function scrollToTop() {
@@ -230,13 +275,14 @@ async function scrollToBottom() {
   update()
 }
 
-function handleScroll() {
+function flushScrollFrame() {
+  scrollFrameId = undefined
   const wrap = wrapRef.value
   if (!wrap) {
     return
   }
 
-  update()
+  syncScrollbarPosition(wrap)
   activateScrollbar()
   emit('scroll', {
     scrollTop: wrap.scrollTop,
@@ -244,6 +290,14 @@ function handleScroll() {
   })
 }
 
+function handleScroll() {
+  if (scrollFrameId !== undefined) {
+    return
+  }
+
+  scrollFrameId = requestAnimationFrame(flushScrollFrame)
+}
+
 function handleTrackPointerDown(event: PointerEvent) {
   if (event.target !== event.currentTarget) {
     return
@@ -254,11 +308,11 @@ function handleTrackPointerDown(event: PointerEvent) {
     return
   }
 
+  update()
   const rect = (event.currentTarget as HTMLElement).getBoundingClientRect()
   activateScrollbar()
   const nextThumbTop = event.clientY - rect.top - thumbHeight.value / 2
-  const scrollRange = maxScrollTop(wrap)
-  const thumbRange = maxThumbTop(wrap)
+  const { scrollRange, thumbRange } = verticalMetrics
   wrap.scrollTop
     = thumbRange > 0 ? (clamp(nextThumbTop, 0, thumbRange) / thumbRange) * scrollRange : 0
 }
@@ -273,11 +327,11 @@ function handleHorizontalTrackPointerDown(event: PointerEvent) {
     return
   }
 
+  update()
   const rect = (event.currentTarget as HTMLElement).getBoundingClientRect()
   activateScrollbar()
   const nextThumbLeft = event.clientX - rect.left - horizontalThumbWidth.value / 2
-  const scrollRange = maxScrollLeft(wrap)
-  const thumbRange = maxHorizontalThumbLeft(wrap)
+  const { scrollRange, thumbRange } = horizontalMetrics
   wrap.scrollLeft
     = thumbRange > 0 ? (clamp(nextThumbLeft, 0, thumbRange) / thumbRange) * scrollRange : 0
 }
@@ -289,6 +343,7 @@ function handleThumbPointerDown(event: PointerEvent) {
   }
 
   event.preventDefault()
+  update()
   dragging.value = true
   visible.value = true
   clearHideTimer()
@@ -303,6 +358,7 @@ function handleHorizontalThumbPointerDown(event: PointerEvent) {
   }
 
   event.preventDefault()
+  update()
   horizontalDragging.value = true
   visible.value = true
   clearHideTimer()
@@ -320,12 +376,11 @@ function handleThumbPointerMove(event: PointerEvent) {
     return
   }
 
-  const thumbRange = maxThumbTop(wrap)
+  const { scrollRange, thumbRange } = verticalMetrics
   if (thumbRange <= 0) {
     return
   }
 
-  const scrollRange = maxScrollTop(wrap)
   wrap.scrollTop = dragStartScrollTop + ((event.clientY - dragStartY) / thumbRange) * scrollRange
 }
 
@@ -339,12 +394,11 @@ function handleHorizontalThumbPointerMove(event: PointerEvent) {
     return
   }
 
-  const thumbRange = maxHorizontalThumbLeft(wrap)
+  const { scrollRange, thumbRange } = horizontalMetrics
   if (thumbRange <= 0) {
     return
   }
 
-  const scrollRange = maxScrollLeft(wrap)
   wrap.scrollLeft
     = horizontalDragStartScrollLeft
       + ((event.clientX - horizontalDragStartX) / thumbRange) * scrollRange
@@ -368,19 +422,33 @@ function handleHorizontalThumbPointerUp() {
   activateScrollbar()
 }
 
+function handleDocumentPointerMove(event: PointerEvent) {
+  handleThumbPointerMove(event)
+  handleHorizontalThumbPointerMove(event)
+}
+
+function handleDocumentPointerUp() {
+  handleThumbPointerUp()
+  handleHorizontalThumbPointerUp()
+}
+
 onMounted(async () => {
   await nextTick()
   update()
 })
 
+onBeforeUnmount(() => {
+  if (scrollFrameId !== undefined) {
+    cancelAnimationFrame(scrollFrameId)
+  }
+})
+
 useResizeObserver([wrapRef, viewRef], update)
-useScroll(wrapRef, { onScroll: handleScroll })
-useEventListener(document, 'pointermove', handleThumbPointerMove)
-useEventListener(document, 'pointerup', handleThumbPointerUp)
-useEventListener(document, 'pointermove', handleHorizontalThumbPointerMove)
-useEventListener(document, 'pointerup', handleHorizontalThumbPointerUp)
-useEventListener(rootRef, ['mouseenter', 'mousemove'], activateScrollbar)
-useEventListener(rootRef, 'mouseleave', hideScrollbar)
+useEventListener(wrapRef, 'scroll', handleScroll, { passive: true })
+useEventListener(dragDocument, 'pointermove', handleDocumentPointerMove)
+useEventListener(dragDocument, ['pointerup', 'pointercancel'], handleDocumentPointerUp)
+useEventListener(rootRef, 'mouseenter', handleRootMouseEnter)
+useEventListener(rootRef, 'mouseleave', handleRootMouseLeave)
 useEventListener(verticalTrackRef, 'mouseenter', () => handleTrackMouseEnter('vertical'))
 useEventListener(verticalTrackRef, 'mouseleave', () => handleTrackMouseLeave('vertical'))
 useEventListener(verticalTrackRef, 'pointerdown', handleTrackPointerDown)

frontend/src/components/base/BaseTable/useHorizontalStickyShadow.ts

@@ -15,6 +15,8 @@ interface UseHorizontalStickyShadowOptions {
 export function useHorizontalStickyShadow(options: UseHorizontalStickyShadowOptions) {
   const horizontalScrolled = shallowRef(false)
   const horizontalCanScrollRight = shallowRef(false)
+  let horizontalScrollRange = 0
+  let scrollLeftPosition = 0
 
   function maxScrollLeft(wrap: HTMLElement) {
     return Math.max(wrap.scrollWidth - wrap.clientWidth, 0)
@@ -25,11 +27,26 @@ export function useHorizontalStickyShadow(options: UseHorizontalStickyShadowOpti
   }
 
   function resetHorizontalScrollbar() {
+    horizontalScrollRange = 0
+    scrollLeftPosition = 0
     horizontalScrolled.value = false
     horizontalCanScrollRight.value = false
+    syncHeaderScroll(0)
   }
 
-  function updateHorizontalScrollbar() {
+  function syncHeaderScroll(scrollLeft: number) {
+    const headerWrap = options.headerWrapRef.value
+    if (headerWrap && headerWrap.scrollLeft !== scrollLeft) {
+      headerWrap.scrollLeft = scrollLeft
+    }
+  }
+
+  function updateHorizontalScrollPosition(scrollLeft: number) {
+    horizontalScrolled.value = scrollLeft > 0
+    horizontalCanScrollRight.value = scrollLeft < horizontalScrollRange - 1
+  }
+
+  function measureHorizontalScrollbar() {
     const wrap = scrollWrap()
     if (!wrap || !options.hasRows.value) {
       resetHorizontalScrollbar()
@@ -42,21 +59,25 @@ export function useHorizontalStickyShadow(options: UseHorizontalStickyShadowOpti
       return
     }
 
-    horizontalScrolled.value = wrap.scrollLeft > 0
-    horizontalCanScrollRight.value = wrap.scrollLeft < scrollRange - 1
+    horizontalScrollRange = scrollRange
+    scrollLeftPosition = wrap.scrollLeft
+    syncHeaderScroll(scrollLeftPosition)
+    updateHorizontalScrollPosition(scrollLeftPosition)
   }
 
-  function handleTableScroll() {
-    const wrap = scrollWrap()
-    if (wrap && options.headerWrapRef.value) {
-      options.headerWrapRef.value.scrollLeft = wrap.scrollLeft
+  function handleTableScroll(payload: { scrollTop: number, scrollLeft: number }) {
+    if (payload.scrollLeft === scrollLeftPosition) {
+      return
     }
-    updateHorizontalScrollbar()
+
+    scrollLeftPosition = payload.scrollLeft
+    syncHeaderScroll(scrollLeftPosition)
+    updateHorizontalScrollPosition(scrollLeftPosition)
   }
 
   onMounted(async () => {
     await nextTick()
-    updateHorizontalScrollbar()
+    measureHorizontalScrollbar()
   })
 
   useResizeObserver(
@@ -64,12 +85,12 @@ export function useHorizontalStickyShadow(options: UseHorizontalStickyShadowOpti
       [scrollWrap(), options.tableViewRef.value].filter(
         (element): element is HTMLDivElement | HTMLTableElement => Boolean(element),
       ),
-    updateHorizontalScrollbar,
+    measureHorizontalScrollbar,
   )
 
   watch(options.watchSources, async () => {
     await nextTick()
-    updateHorizontalScrollbar()
+    measureHorizontalScrollbar()
   })
 
   return {

frontend/src/layout/components/AppSidebar.vue

@@ -1,7 +1,6 @@
 <script setup lang="ts">
 import {
   ArrowUpCircle,
-  Cat,
   ChartNoAxesColumn,
   Info,
   KeyRound,
@@ -118,7 +117,11 @@ const sidebarEl = ref<HTMLElement | null>(null)
 const brandLabelEl = ref<HTMLElement | null>(null)
 const navSignalEl = useTemplateRef<HTMLElement>('navSignal')
 const isCollapsed = computed(() => !props.mobile && Boolean(props.collapsed))
-const sidebarWidth = computed(() => (isCollapsed.value ? 88 : 256))
+const collapsedSidebarWidth = 88
+const expandedSidebarWidth = 251
+const sidebarWidth = computed(() =>
+  isCollapsed.value ? collapsedSidebarWidth : expandedSidebarWidth,
+)
 const brandLabelVisible = shallowRef(!isCollapsed.value)
 const themeToggleLabel = computed(() =>
   effectiveTheme.value === 'dark' ? '切换浅色模式' : '切换暗黑模式',
@@ -209,7 +212,7 @@ function animateSidebarWidth(collapsed: boolean) {
     return
   }
 
-  const targetWidth = collapsed ? 88 : 256
+  const targetWidth = collapsed ? collapsedSidebarWidth : expandedSidebarWidth
   const currentWidth = sidebarEl.value.getBoundingClientRect().width
 
   if (prefersReducedMotion()) {
@@ -326,7 +329,7 @@ onBeforeUnmount(() => {
     class="z-20 h-dvh shrink-0 flex-col overflow-hidden bg-(--cp-bg-surface) px-4 shadow-(--cp-shadow-sidebar)"
     :class="[
       mobile ? 'flex' : 'hidden min-[961px]:flex',
-      isCollapsed ? 'w-22 basis-22 items-center' : 'w-64 basis-64',
+      isCollapsed ? 'w-22 basis-22 items-center' : 'w-[251px] basis-[251px]',
     ]"
   >
     <div
@@ -336,9 +339,14 @@ onBeforeUnmount(() => {
       <BaseMotionIcon
         aria-hidden="true"
         variant="brand"
-        class="inline-flex size-11 items-center justify-center relative -top-0.5 rounded-(--cp-icon-button-radius) bg-(--cp-bg-muted) text-(--cp-text-primary)"
+        class="inline-flex size-11 items-center justify-center relative -top-0.5 rounded-(--cp-icon-button-radius) bg-(--cp-brand-mark-bg)"
       >
-        <Cat :size="27" stroke-width="2" />
+        <img
+          src="/favicon.svg"
+          alt=""
+          class="block size-9 select-none"
+          draggable="false"
+        >
       </BaseMotionIcon>
       <span
         v-show="brandLabelVisible"

frontend/src/layout/components/FloatingSidebarToggle.vue

@@ -0,0 +1,240 @@
+<script setup lang="ts">
+import type { CSSProperties } from 'vue'
+import { Menu } from '@lucide/vue'
+import {
+  useDraggable,
+  useElementSize,
+  useScreenSafeArea,
+  useStorage,
+  useWindowSize,
+} from '@vueuse/core'
+import {
+  computed,
+  nextTick,
+  onMounted,
+  onScopeDispose,
+  shallowRef,
+  useTemplateRef,
+  watch,
+} from 'vue'
+
+interface StoredPosition {
+  initialized: boolean
+  side: 'left' | 'right'
+  x: number
+  y: number
+}
+
+const emit = defineEmits<{
+  open: []
+}>()
+
+const edgeGap = 12
+const defaultButtonSize = 44
+const dragClickThreshold = 6
+const storageKey = 'codex-proxy:mobile-sidebar-toggle-position'
+
+const toggleEl = useTemplateRef<HTMLElement>('toggleEl')
+const { width: viewportWidth, height: viewportHeight } = useWindowSize({
+  initialWidth: 390,
+  initialHeight: 760,
+  type: 'visual',
+})
+const { width: toggleWidth, height: toggleHeight } = useElementSize(toggleEl, {
+  width: defaultButtonSize,
+  height: defaultButtonSize,
+})
+const safeArea = useScreenSafeArea()
+const storedPosition = useStorage<StoredPosition>(
+  storageKey,
+  {
+    initialized: false,
+    side: 'left',
+    x: edgeGap,
+    y: 136,
+  },
+  undefined,
+  { mergeDefaults: true, shallow: true },
+)
+const pointerStart = shallowRef({ x: edgeGap, y: 136 })
+const ignoreNextClick = shallowRef(false)
+const positionReady = shallowRef(false)
+let ignoreClickTimer: number | undefined
+
+const bounds = computed(() => {
+  const buttonWidth = positiveNumber(toggleWidth.value, defaultButtonSize)
+  const buttonHeight = positiveNumber(toggleHeight.value, defaultButtonSize)
+  const windowWidth = positiveNumber(viewportWidth.value, 390)
+  const windowHeight = positiveNumber(viewportHeight.value, 760)
+  const minX = parseCssPixel(safeArea.left.value) + edgeGap
+  const maxX = Math.max(
+    minX,
+    windowWidth - parseCssPixel(safeArea.right.value) - edgeGap - buttonWidth,
+  )
+  const minY = parseCssPixel(safeArea.top.value) + edgeGap
+  const maxY = Math.max(
+    minY,
+    windowHeight - parseCssPixel(safeArea.bottom.value) - edgeGap - buttonHeight,
+  )
+
+  return { minX, maxX, minY, maxY }
+})
+
+const { x, y, isDragging } = useDraggable(toggleEl, {
+  initialValue: {
+    x: storedPosition.value.x,
+    y: storedPosition.value.y,
+  },
+  preventDefault: false,
+  stopPropagation: false,
+  onEnd(position) {
+    applyPosition(snapPosition(position.x, position.y))
+  },
+})
+
+const toggleStyle = computed<CSSProperties>(() => ({
+  transform: `translate3d(${x.value}px, ${y.value}px, 0)`,
+  touchAction: 'none',
+}))
+
+function handlePointerDown(event: PointerEvent) {
+  pointerStart.value = { x: event.clientX, y: event.clientY }
+  const target = event.currentTarget
+  if (target instanceof HTMLElement)
+    target.setPointerCapture?.(event.pointerId)
+}
+
+function handlePointerUp(event: PointerEvent) {
+  const target = event.currentTarget
+  if (target instanceof HTMLElement)
+    target.releasePointerCapture?.(event.pointerId)
+
+  const moved = Math.hypot(
+    event.clientX - pointerStart.value.x,
+    event.clientY - pointerStart.value.y,
+  )
+  ignoreNextClick.value = true
+  window.clearTimeout(ignoreClickTimer)
+  ignoreClickTimer = window.setTimeout(() => {
+    ignoreNextClick.value = false
+  }, 320)
+
+  if (moved <= dragClickThreshold) {
+    emit('open')
+  }
+}
+
+function handleClick() {
+  if (isDragging.value || ignoreNextClick.value) {
+    ignoreNextClick.value = false
+    return
+  }
+
+  emit('open')
+}
+
+function defaultPosition() {
+  const targetY = Math.round(positiveNumber(viewportHeight.value, 760) * 0.36)
+  return snapPosition(bounds.value.minX, targetY)
+}
+
+function restorePosition() {
+  if (!storedPosition.value.initialized) {
+    applyPosition(defaultPosition())
+    return
+  }
+
+  const sideX = storedPosition.value.side === 'right' ? bounds.value.maxX : bounds.value.minX
+  applyPosition({
+    initialized: true,
+    side: storedPosition.value.side,
+    x: sideX,
+    y: storedPosition.value.y,
+  })
+}
+
+function snapPosition(rawX: number, rawY: number) {
+  const { minX, maxX, minY, maxY } = bounds.value
+  const side = rawX <= (minX + maxX) / 2 ? 'left' : 'right'
+
+  return {
+    initialized: true,
+    side,
+    x: side === 'left' ? minX : maxX,
+    y: clamp(rawY, minY, maxY),
+  } satisfies StoredPosition
+}
+
+function applyPosition(position: StoredPosition) {
+  const { minX, maxX, minY, maxY } = bounds.value
+  const next = {
+    initialized: true,
+    side: position.side,
+    x: clamp(position.x, minX, maxX),
+    y: clamp(position.y, minY, maxY),
+  } satisfies StoredPosition
+
+  x.value = next.x
+  y.value = next.y
+  storedPosition.value = next
+}
+
+function clamp(value: number, min: number, max: number) {
+  return Math.min(Math.max(value, min), max)
+}
+
+function positiveNumber(value: number, fallback: number) {
+  return Number.isFinite(value) && value > 0 ? value : fallback
+}
+
+function parseCssPixel(value: string) {
+  const parsed = Number.parseFloat(value)
+  return Number.isFinite(parsed) ? parsed : 0
+}
+
+onMounted(async () => {
+  await nextTick()
+  safeArea.update()
+  restorePosition()
+  positionReady.value = true
+})
+
+watch(
+  bounds,
+  () => {
+    if (!positionReady.value)
+      return
+    restorePosition()
+  },
+  { flush: 'post' },
+)
+
+onScopeDispose(() => {
+  window.clearTimeout(ignoreClickTimer)
+})
+</script>
+
+<template>
+  <div
+    ref="toggleEl"
+    class="fixed top-0 left-0 z-40 min-[961px]:hidden"
+    :class="positionReady ? 'opacity-100' : 'opacity-0'"
+    :style="toggleStyle"
+  >
+    <button
+      type="button"
+      class="inline-flex size-11.5 items-center justify-center rounded-(--cp-icon-button-radius) border-0 bg-(--cp-bg-surface) text-(--cp-text-primary) shadow-[0_18px_34px_-18px_var(--cp-shadow-sticky)] outline-none transition-[background-color,box-shadow,color,opacity,transform] duration-150 hover:bg-(--cp-default-bg-hover) hover:text-(--cp-normal) active:scale-95 focus-visible:ring-2 focus-visible:ring-(--cp-info-border) focus-visible:ring-offset-2 focus-visible:ring-offset-(--cp-bg-page)"
+      :class="
+        isDragging
+          ? 'scale-95 cursor-grabbing shadow-[0_14px_28px_-20px_var(--cp-shadow-sticky)]'
+          : 'cursor-grab'
+      "
+      aria-label="打开侧边栏"
+      @pointerdown="handlePointerDown"
+      @pointerup="handlePointerUp"
+      @click="handleClick"
+    >
+      <Menu class="size-5" />
+    </button>
+  </div>
+</template>

frontend/src/layout/components/MobileSidebarToolbar.vue

@@ -1,26 +0,0 @@
-<script setup lang="ts">
-import { Menu } from '@lucide/vue'
-
-import BaseButton from '@/components/base/BaseButton.vue'
-
-const emit = defineEmits<{
-  open: []
-}>()
-</script>
-
-<template>
-  <div
-    class="sticky top-0 z-40 -mx-4 -mt-4 mb-4 flex h-15 items-center bg-(--cp-bg-page)/96 px-4 backdrop-blur-sm min-[961px]:hidden"
-  >
-    <BaseButton
-      icon-only
-      size="md"
-      variant="default"
-      label="打开侧边栏"
-      class="shadow-[0_12px_28px_-20px_var(--cp-shadow-sticky)]"
-      @click="emit('open')"
-    >
-      <Menu class="size-5" />
-    </BaseButton>
-  </div>
-</template>

frontend/src/layout/index.vue

@@ -9,7 +9,7 @@ import { useUiStore } from '@/stores/modules/ui'
 
 import AppAboutModal from './components/AppAboutModal.vue'
 import AppSidebar from './components/AppSidebar.vue'
-import MobileSidebarToolbar from './components/MobileSidebarToolbar.vue'
+import FloatingSidebarToggle from './components/FloatingSidebarToggle.vue'
 import SystemUpdateModal from './components/SystemUpdateModal.vue'
 
 const uiStore = useUiStore()
@@ -77,12 +77,12 @@ watch(
       @open-about="aboutOpen = true"
       @open-system-update="openSystemUpdate"
     />
+    <FloatingSidebarToggle v-if="!mobileSidebarOpen" @open="openMobileSidebar" />
     <main class="h-dvh min-w-0 flex-1 overflow-hidden">
       <BaseScrollbar
         ref="pageScrollbarRef"
         view-class="flex min-h-full min-w-0 flex-col p-4 min-[961px]:p-6"
       >
-        <MobileSidebarToolbar v-if="!mobileSidebarOpen" @open="openMobileSidebar" />
         <RouterView v-slot="{ Component }">
           <component :is="Component" class="min-h-0 flex-1" />
         </RouterView>

frontend/src/stores/modules/system-update.ts

@@ -1,4 +1,4 @@
-import { useEventSource } from '@vueuse/core'
+import { until, useEventSource } from '@vueuse/core'
 import { delay } from 'es-toolkit'
 import { defineStore } from 'pinia'
 
@@ -14,6 +14,7 @@ import { ApiError } from '@/api/request'
 import { errorMessage } from '@/utils/async'
 
 const maxUpdateLogs = 200
+const updateEventReadyTimeoutMs = 3_000
 const restartReadyTimeoutMs = 60_000
 const restartProbeTimeoutMs = 2_000
 const restartReadyPollIntervalMs = 500
@@ -106,19 +107,30 @@ export const useSystemUpdateStore = defineStore('system-update', () => {
     }
   })
 
-  function connectUpdateEvents(force = false) {
+  async function connectUpdateEvents(force = false) {
     if (force) {
       disconnectUpdateEvents()
     }
     else if (updateEventSource.value) {
-      return
+      return updateEventStatus.value === 'OPEN'
     }
 
     updateStreamError.value = ''
     openUpdateEventSource()
     if (!updateEventSource.value) {
       updateStreamError.value = '当前浏览器不支持实时更新日志'
+      return false
     }
+
+    if (updateEventStatus.value !== 'OPEN') {
+      await until(updateEventStatus).toBe('OPEN', {
+        timeout: updateEventReadyTimeoutMs,
+      })
+    }
+    const connected = updateEventStatus.value === 'OPEN'
+    if (!connected)
+      updateStreamError.value = '更新日志连接超时'
+    return connected
   }
 
   function disconnectUpdateEvents() {
@@ -186,7 +198,7 @@ export const useSystemUpdateStore = defineStore('system-update', () => {
       return null
 
     clearUpdateLogs()
-    connectUpdateEvents(true)
+    await connectUpdateEvents(true)
     updating.value = true
     updateError.value = ''
     updateSuccess.value = false

frontend/src/styles/tokens.css

@@ -8,6 +8,7 @@
   --cp-bg-tertiary: #eef6ff;
   --cp-bg-nav-active: #e9eef5;
   --cp-bg-dark: #111827;
+  --cp-brand-mark-bg: #1c2a3d;
   --cp-white: #ffffff;
   --cp-transparent: #ffffff00;
   --cp-overlay-scrim: #0e17264d;

frontend/src/views/accounts/components/AccountCreateModal.vue

@@ -2,7 +2,7 @@
 import type { useAccountOnboarding } from '../composables/useAccountOnboarding'
 import type { AccountRow } from '../constants'
 import { Openai, Xai } from '@boxicons/vue'
-import { Copy, KeyRound, Upload } from '@lucide/vue'
+import { Copy, KeyRound, LayoutGrid, Upload } from '@lucide/vue'
 
 import { useClipboard, useFileDialog } from '@vueuse/core'
 import { computed, ref } from 'vue'
@@ -51,7 +51,7 @@ const { open: openImportFile, onChange: onImportFileChange } = useFileDialog({
 })
 
 const modeOptions = computed(() => accountProviderModeOptions(form.value.provider))
-const isProviderSelected = computed(() => form.value.provider === 'openai' || form.value.provider === 'xai')
+const isProviderSelected = computed(() => ['openai', 'xai', 'batch'].includes(form.value.provider))
 const isChoosingProvider = computed(() => !props.reauthorizing && !isProviderSelected.value)
 
 const provider = computed({
@@ -91,10 +91,17 @@ const oauthCallback = computed({
 
 const oauthAuthUrl = computed(() => form.value.oauthAuthUrl || '')
 const isXai = computed(() => form.value.provider === 'xai')
-const importFileLabel = computed(() => mode.value === 'agent_identity' ? 'Agent 身份文件' : '账号文件')
+const isBatch = computed(() => form.value.provider === 'batch')
+const importFileLabel = computed(() => {
+  if (isBatch.value)
+    return '批量账号文件'
+  return mode.value === 'agent_identity' ? 'Agent 身份文件' : '账号文件'
+})
 const importFilePlaceholder = computed(() => mode.value === 'agent_identity'
   ? '粘贴 Agent 身份文件内容'
-  : '粘贴 CPR、Sub2API 或 CPA 账号文件内容')
+  : isBatch.value
+    ? '粘贴 CPR 多平台导出文件内容'
+    : '粘贴 CPR、Sub2API 或 CPA 账号文件内容')
 
 const accountName = computed(() => {
   return props.account?.email || props.account?.accountId || props.account?.id || '该账号'
@@ -140,6 +147,8 @@ const description = computed<string | undefined>(() => {
     return undefined
   if (props.reauthorizing)
     return '完成授权后粘贴回调地址，系统会更新账号凭据'
+  if (isBatch.value)
+    return '导入 CPR 多平台账号包，系统会按文档中的平台自动分流'
   if (isXai.value) {
     return mode.value === 'oauth'
       ? '复制 xAI 授权链接，完成后粘贴回调地址，不使用 xAI API Key'
@@ -152,7 +161,7 @@ const description = computed<string | undefined>(() => {
   return '导入 CPR、Sub2API 或 CPA 账号文件，已存在账号会更新'
 })
 
-function selectProvider(value: 'openai' | 'xai') {
+function selectProvider(value: 'openai' | 'xai' | 'batch') {
   provider.value = value
 }
 
@@ -198,7 +207,8 @@ async function copyText(value: string, successText: string) {
     :hide-footer="isChoosingProvider"
   >
     <template #icon>
-      <Xai v-if="isXai" class="text-(--cp-text-primary)" aria-hidden="true" :width="20" :height="20" />
+      <LayoutGrid v-if="isBatch" class="text-(--cp-text-primary)" aria-hidden="true" :width="20" :height="20" />
+      <Xai v-else-if="isXai" class="text-(--cp-text-primary)" aria-hidden="true" :width="20" :height="20" />
       <Openai v-else class="text-(--cp-text-primary)" aria-hidden="true" :width="20" :height="20" />
     </template>
 
@@ -209,7 +219,12 @@ async function copyText(value: string, successText: string) {
     />
 
     <div v-else class="flex flex-col gap-4">
-      <BaseSegmented v-if="!reauthorizing" v-model="mode" :options="modeOptions" class="w-full" />
+      <BaseSegmented
+        v-if="!reauthorizing && !isBatch"
+        v-model="mode"
+        :options="modeOptions"
+        class="w-full"
+      />
 
       <div v-if="mode === 'oauth'" class="flex flex-col gap-4">
         <div class="rounded-(--cp-input-radius-base) bg-(--cp-bg-subtle) px-4 py-3">
@@ -316,7 +331,7 @@ async function copyText(value: string, successText: string) {
         :disabled="!canSubmit"
         @click="emit('create')"
       >
-        {{ reauthorizing ? '完成重新授权' : mode === 'oauth' ? '完成授权导入' : '导入' }}
+        {{ reauthorizing ? '完成重新授权' : mode === 'oauth' ? '完成授权导入' : isBatch ? '批量导入' : '导入' }}
       </BaseButton>
     </template>
   </BaseModal>

frontend/src/views/accounts/components/AccountOverviewCards.vue

@@ -33,8 +33,8 @@ const overviewItems = computed(() => [
     icon: Gauge,
   },
   {
-    label: '需处理',
-    value: formatCount(props.summary.attention),
+    label: '不可用账号',
+    value: formatCount(props.summary.unavailable),
     caption: '过期 / 禁用 / 封禁',
     tone: 'danger',
     icon: AlertTriangle,

frontend/src/views/accounts/components/AccountProviderChooser.vue

@@ -1,5 +1,6 @@
 <script setup lang="ts">
 import { Openai, Xai } from '@boxicons/vue'
+import { LayoutGrid } from '@lucide/vue'
 
 withDefaults(
   defineProps<{
@@ -11,10 +12,15 @@ withDefaults(
 )
 
 const emit = defineEmits<{
-  select: [provider: 'openai' | 'xai']
+  select: [provider: 'openai' | 'xai' | 'batch']
 }>()
 
 const providers = [
+  {
+    value: 'batch' as const,
+    label: '批量导入',
+    icon: LayoutGrid,
+  },
   {
     value: 'openai' as const,
     label: 'OpenAI',

frontend/src/views/accounts/composables/useAccountConnectionTest.ts

@@ -273,7 +273,10 @@ export function useAccountConnectionTest(options: AccountConnectionTestOptions)
     if (event.type === 'error') {
       applyAccountStatus(event.accountStatus)
       connectionTestError.value = event.error || '测试连接失败'
-      appendConnectionTestLog(connectionTestError.value, 'danger')
+      const detail = event.providerErrorCode || event.providerErrorType
+        ? { code: event.providerErrorCode, type: event.providerErrorType }
+        : undefined
+      appendConnectionTestLog(connectionTestError.value, 'danger', detail)
       finishConnectionTest('error')
       clearConnectionTestRun()
     }

frontend/src/views/accounts/composables/useAccountMutations.ts

@@ -24,7 +24,6 @@ import { useAccountOnboarding } from './useAccountOnboarding'
 
 type AccountRow = Awaited<ReturnType<typeof getAccounts>>['items'][number]
 
-const attentionAccountStatuses = new Set(['expired', 'disabled', 'banned'])
 const accountStatusSortRank: Record<string, number> = {
   active: 0,
   quota_exhausted: 1,
@@ -250,15 +249,19 @@ export function useAccountMutations(options: {
         + Number(status === 'quota_exhausted')
         - Number(current.status === 'quota_exhausted'),
       ),
-      attention: Math.max(
+      unavailable: Math.max(
         0,
-        options.accountSummary.value.attention
-        + Number(attentionAccountStatuses.has(status))
-        - Number(attentionAccountStatuses.has(current.status)),
+        options.accountSummary.value.unavailable
+        + Number(isUnavailableAccountStatus(status))
+        - Number(isUnavailableAccountStatus(current.status)),
       ),
     }
   }
 
+  function isUnavailableAccountStatus(status: string) {
+    return status === 'expired' || status === 'disabled' || status === 'banned'
+  }
+
   function sortAccountsByStatus(rows: AccountRow[]) {
     if (options.sort.value?.key !== 'status')
       return rows

frontend/src/views/accounts/composables/useAccountOnboarding.ts

@@ -11,6 +11,12 @@ import { toast } from '@/components/base/BaseToast'
 import { useAsyncAction } from '@/composables/useAsyncAction'
 
 type AccountRow = Awaited<ReturnType<typeof getAccounts>>['items'][number]
+type ImportProvider = 'openai' | 'xai'
+
+interface MixedImportDocument {
+  provider: ImportProvider
+  document: Record<string, unknown>
+}
 
 export function useAccountOnboarding(options: {
   reload: () => Promise<unknown>
@@ -46,10 +52,10 @@ export function useAccountOnboarding(options: {
 
     await creatingAccountAction.run(
       async () => {
-        await importAccountDocument()
-        await finishCreate(
-          createForm.value.provider === 'xai' ? 'xAI OAuth 账号已导入' : 'OpenAI 账号已导入',
-        )
+        const message = createForm.value.provider === 'batch'
+          ? await importMixedAccountDocument()
+          : await importAccountDocument()
+        await finishCreate(message)
       },
       { errorText: '导入失败' },
     )
@@ -170,6 +176,41 @@ export function useAccountOnboarding(options: {
       data,
     }))
     commitConfigRevision(result.configRevision)
+    return createForm.value.provider === 'xai' ? 'xAI 账号已导入' : 'OpenAI 账号已导入'
+  }
+
+  async function importMixedAccountDocument() {
+    const documents = parseMixedImportDocuments(parseImportJson(createForm.value.importText))
+    let importedCount = 0
+    const failures: string[] = []
+
+    for (const entry of documents) {
+      try {
+        const expectedConfigRevision = await requireConfigRevision()
+        const result = await withConflictRefresh(() => importAccounts({
+          provider: entry.provider,
+          expectedConfigRevision,
+          data: entry.document,
+        }))
+        commitConfigRevision(result.configRevision)
+        importedCount += result.importedCount
+      }
+      catch (error) {
+        const message = error instanceof Error && error.message
+          ? error.message
+          : '导入失败'
+        failures.push(`${entry.provider === 'xai' ? 'xAI' : 'OpenAI'}：${message}`)
+      }
+    }
+
+    if (importedCount === 0) {
+      throw new Error(failures.length > 0 ? `批量导入失败：${failures.join('；')}` : '批量文件没有可导入的账号')
+    }
+
+    if (failures.length > 0) {
+      return `已导入 ${importedCount} 个账号，${failures.length} 个文档失败：${failures.join('；')}`
+    }
+    return `已导入 ${importedCount} 个账号`
   }
 
   async function finishCreate(message: string) {
@@ -195,7 +236,7 @@ export function useAccountOnboarding(options: {
     () => {
       createForm.value = {
         ...createForm.value,
-        mode: 'oauth',
+        mode: createForm.value.provider === 'batch' ? 'json' : 'oauth',
         oauthFlowId: '',
         oauthAuthUrl: '',
         oauthCallback: '',
@@ -255,3 +296,28 @@ function parseImportJson(value: string) {
     throw new Error('JSON 格式不正确')
   }
 }
+
+function parseMixedImportDocuments(value: unknown): MixedImportDocument[] {
+  if (!isRecord(value) || !Array.isArray(value.documents))
+    throw new Error('批量导入文件必须是 CPR 多平台导出文件')
+
+  const documents: MixedImportDocument[] = []
+  for (const entry of value.documents) {
+    if (!isRecord(entry))
+      throw new Error('批量导入文件包含无效的 Provider 文档')
+    const provider = entry.provider
+    if (provider !== 'openai' && provider !== 'xai')
+      throw new Error('批量导入文件包含无效的 Provider 文档')
+    if (!isRecord(entry.document))
+      throw new Error('批量导入文件包含无效的 Provider 文档')
+    documents.push({ provider, document: entry.document })
+  }
+
+  if (documents.length === 0)
+    throw new Error('批量文件没有可导入的账号文档')
+  return documents
+}
+
+function isRecord(value: unknown): value is Record<string, unknown> {
+  return typeof value === 'object' && value !== null && !Array.isArray(value)
+}

frontend/src/views/accounts/composables/useAccountsQuery.ts

@@ -17,7 +17,7 @@ export function useAccountsQuery() {
     total: 0,
     active: 0,
     quotaExhausted: 0,
-    attention: 0,
+    unavailable: 0,
   })
 
   const query = usePagedQuery({

frontend/src/views/accounts/constants.ts

@@ -98,7 +98,8 @@ export const accountColumns = [
     sortable: true,
     flex: 1.2,
     minWidth: '160px',
-    format: (_value: unknown, row: AccountRow) => row.usage.lastUsedAtDisplay,
+    format: (_value: unknown, row: AccountRow) => optionalAccountCell(row.usage.lastUsedAtDisplay),
+    emptyText: '',
     cellClass: `${relaxedCellClass} text-(--cp-text-secondary)`,
   },
   {
@@ -108,6 +109,8 @@ export const accountColumns = [
     sortKey: 'expiresAt',
     flex: 1.2,
     minWidth: '160px',
+    format: value => optionalAccountCell(value),
+    emptyText: '',
     cellClass: `${relaxedCellClass} text-(--cp-text-secondary)`,
   },
   {
@@ -201,3 +204,7 @@ function accountProviderLabel(value?: string | null) {
     return 'xAI'
   return value || '—'
 }
+
+function optionalAccountCell(value: unknown) {
+  return value === '—' || value === '-' ? '' : value
+}

frontend/src/views/dashboard/composables/presenter.ts

@@ -39,8 +39,8 @@ const emptyCards: DashboardSummary['cards'] = {
   credentials: {
     total: '0',
     totalValue: 0,
-    enabled: '0',
-    enabledValue: 0,
+    available: '0',
+    availableValue: 0,
     unavailable: '0',
     unavailableValue: 0,
   },
@@ -147,9 +147,9 @@ function metricCards(
       tone: 'normal',
       details: [
         {
-          label: '启用',
-          value: credentials.enabled,
-          tone: credentials.enabledValue > 0 ? 'success' : 'normal',
+          label: '可用',
+          value: credentials.available,
+          tone: credentials.availableValue > 0 ? 'success' : 'normal',
         },
         {
           label: '不可用',

frontend/src/views/login/components/LoginPanel.vue

@@ -1,5 +1,5 @@
 <script setup lang="ts">
-import { ArrowRight, Cat, CircleAlert, Eye, EyeOff, KeyRound, Mail, Moon, Sun } from '@lucide/vue'
+import { ArrowRight, CircleAlert, Eye, EyeOff, KeyRound, Mail, Moon, Sun } from '@lucide/vue'
 import { computed, shallowRef } from 'vue'
 
 import BaseButton from '@/components/base/BaseButton.vue'
@@ -58,7 +58,12 @@ function togglePasswordVisible(): void {
     <header class="flex min-w-0 items-center justify-between gap-4.5 max-[560px]:gap-3.5">
       <div class="flex min-w-0 items-center gap-3">
         <BaseMotionIcon aria-hidden="true" variant="brand" class="login-logo">
-          <Cat :size="20" :stroke-width="2.1" />
+          <img
+            src="/favicon.svg"
+            alt=""
+            class="block size-8 select-none"
+            draggable="false"
+          >
         </BaseMotionIcon>
         <span class="grid min-w-0 gap-1">
           <strong

frontend/src/views/login/index.vue

@@ -93,7 +93,7 @@ async function handleSubmit(): Promise<void> {
   --login-form-bg-c: color-mix(in srgb, var(--cp-bg-muted) 78%, var(--cp-info-bg));
   --login-form-shadow: color-mix(in srgb, var(--cp-text-primary) 20%, transparent);
   --login-form-line: color-mix(in srgb, var(--cp-info-border) 54%, transparent);
-  --login-logo-bg: color-mix(in srgb, var(--cp-info-bg) 74%, var(--cp-bg-surface));
+  --login-logo-bg: var(--cp-brand-mark-bg);
   --login-logo-text: var(--cp-accent-primary);
   --login-title: var(--cp-text-primary);
   --login-brand-title: var(--cp-text-strong);
@@ -159,7 +159,7 @@ async function handleSubmit(): Promise<void> {
   --login-form-bg-c: color-mix(in srgb, var(--cp-bg-page) 90%, transparent);
   --login-form-shadow: color-mix(in srgb, var(--cp-bg-dark) 70%, transparent);
   --login-form-line: color-mix(in srgb, var(--cp-info-text) 66%, transparent);
-  --login-logo-bg: color-mix(in srgb, var(--cp-bg-surface) 84%, transparent);
+  --login-logo-bg: var(--cp-brand-mark-bg);
   --login-logo-text: var(--cp-info-text);
   --login-title: var(--cp-text-strong);
   --login-brand-title: var(--cp-white);

frontend/src/views/usage/components/OpsErrorDetailModal.vue

@@ -57,7 +57,7 @@ function latencyDisplay(value: unknown) {
     <template v-if="record">
       <section class="rounded-(--cp-card-radius) bg-(--cp-bg-subtle) px-4 py-3.5">
         <div class="flex flex-wrap items-center gap-3">
-          <UsageStatusCodeBadge :status-code="record.statusCode" />
+          <UsageStatusCodeBadge :status-code="record.upstreamStatusCode" />
           <span class="text-[12px] font-bold text-(--cp-text-secondary)">
             客户端 {{ display(record.clientStatusCode) }}
           </span>

frontend/src/views/usage/components/OpsErrorPanel.vue

@@ -88,9 +88,9 @@ function showDetail(record: Awaited<ReturnType<typeof getOpsErrors>>['items'][nu
       @page-change="handlePageChange"
       @page-size-change="handlePageSizeChange"
     >
-      <template #statusCode="{ row }">
+      <template #upstreamStatusCode="{ row }">
         <UsageStatusCodeBadge
-          :status-code="typeof row.statusCode === 'number' ? row.statusCode : null"
+          :status-code="typeof row.upstreamStatusCode === 'number' ? row.upstreamStatusCode : null"
         />
       </template>
       <template #failureClass="{ row }">

frontend/src/views/usage/components/UsageSummaryCards.vue

@@ -9,6 +9,10 @@ const props = defineProps<{
   summary: Awaited<ReturnType<typeof getUsageRecordSummary>>
 }>()
 
+function averageLatencyDisplay(value: string) {
+  return !value || value === '—' || value === '-' ? '0 ms' : value
+}
+
 const items = computed(() => [
   {
     key: 'requests',
@@ -38,7 +42,7 @@ const items = computed(() => [
     key: 'latency',
     label: '平均耗时',
     icon: Timer,
-    value: props.summary.averageLatencyMs,
+    value: averageLatencyDisplay(props.summary.averageLatencyMs),
     detail: '成功请求平均值',
     tone: 'text-(--cp-normal-text) bg-(--cp-normal-bg)',
   },

frontend/src/views/usage/composables/useUsageRecordsTable.ts

@@ -247,7 +247,7 @@ function emptySummary() {
     cachedTokens: '0',
     cacheWriteTokens: '0',
     totalTokens: '0',
-    averageLatencyMs: '—',
+    averageLatencyMs: '0 ms',
   }
   return summary
 }

frontend/src/views/usage/constants.ts

@@ -108,7 +108,7 @@ export const opsErrorColumns = [
     cellClass:
       'whitespace-nowrap font-mono text-[12px] font-[650] tabular-nums text-(--cp-text-secondary)',
   },
-  { key: 'statusCode', label: '状态码', width: '96px', align: 'center' as const },
+  { key: 'upstreamStatusCode', label: '上游状态', width: '96px', align: 'center' as const },
   {
     key: 'failureClass',
     label: '失败分类',

release/version.yaml

@@ -1 +1 @@
-version: "3.0.0"
+version: "3.0.7"

skills/codex-proxy-rs/SKILL.md

@@ -19,9 +19,13 @@ description: Codex Proxy RS 仓库开发指南。Use when changing or auditing i
 - `gateway-api` 只做协议适配；`gateway-core` 编排；Provider 独占 credential、catalog 与 transport；`gateway-store` 实现持久化端口。
 - Provider 每次调用只选择一个 credential，不隐藏换号、业务 retry 或跨 Provider fallback。
 - PostgreSQL 是持久化权威；Redis 只保存可恢复协调状态和 OAuth pending flow。
+- 当前只有 `openai` 与 `xai` 两个 Provider；账号归属 Provider，不存在 Provider Instance 层。
+- OpenAI 的 OAuth 与 Agent Identity credential 由 `providers/openai` 独占解析；账号导出使用 CPR 文档，Agent Identity
+  只输出运行所需身份材料，不伪造 OAuth 字段。
 - 测试放在各 package 的 `tests/`。禁止在生产 `src/` 写 test-only 代码。
 - Vue 使用 `<script setup lang="ts">`、现有基础组件和主题 token。
 - README 面向使用者，保持简短；长期架构只写入 `docs/architecture.md`。
+- 完整 HTTP/Admin 路由与敏感导入/导出格式写入 `docs/api.md`。
 - 不添加兼容 shim、重复状态机、第二套配置或补丁式旁路。
 
 ## 工作流
@@ -54,4 +58,5 @@ docker compose -f deploy/compose.yaml config --quiet
 - 提交前检查 `git status --short`、cached/unstaged diff 和 `git diff --check`。
 - 提交带 `Co-authored-by: Codex <noreply@openai.com>`。
 - 发布以 `release/version.yaml` 和带注释的 `vX.Y.Z` tag 为准。
+- 正式发布使用 `release/publish <version>`，不要手工绕过发布脚本。
 - 发布完成后核对远端 `main`、tag、Actions、Release asset 与 GHCR。

skills/codex-proxy-rs/references/repo-guide.md

@@ -32,19 +32,19 @@
 - Core 冻结 RuntimeSnapshot、RoutePlan、retry/fallback 与 downstream commit 边界。
 - Provider 每次 `execute` 只选择一个 credential 并返回 cold canonical stream。
 - 换号只由 Core 在下游 commit 前按重放安全性决定。
-- fallback 只允许同 instance 账号和同 Provider kind instance；不跨 Provider kind。
+- fallback 只允许同一 Provider kind 内的账号；不跨 Provider，也不存在 Provider Instance 层。
 - OpenAI continuation 为 native → replay owner → replay any；xAI 使用客户端完整历史。
 - Provider wire profile 以配置为启动基线，并由共享运行时状态统一发布。OpenAI CLI 读取 `@openai/codex`、Desktop 读取官方 appcast、xAI CLI 读取 `@xai-official/grok`；发现新版本后自动更新对应版本字段，所有消费边界不得维护独立常量。
 - `downstream_committed_at` 是不可撤回交付承诺，不是首字节已经写达的证明。
 
 ## 存储
 
-- PostgreSQL 只有 `0001_initial.sql` 定义的八张终态业务表。
+- PostgreSQL 只有 `0001_initial.sql` 定义的七张终态业务表。
 - `config_revision` 只用于会改变调度快照或安全配置的管理 mutation。
 - quota、cooldown、catalog generation、自动 refresh 不推进全局 revision。
 - refresh 只推进账号 `credential_revision`；Redis cooldown 是可丢失热缓存。
 - OAuth pending 使用 Provider 域隔离 SHA-256 key、固定三字段 Hash 和原子一次消费。
-- secret 不进入日志、Debug、fixture、文档或 audit details。
+- 真实 secret 不进入日志、Debug、fixture、文档或 audit details；测试只能使用合成值。
 
 ## 前端