chore(release): v3.28.0

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

.github/actions/cargo-cache/cargo-cache.sh

@@ -42,8 +42,11 @@ else
   mkdir -p "$(dirname "$directory")"
   temporary="$(mktemp -d "${directory}.XXXXXX")"
   trap 'rm -rf "$temporary"' EXIT
-  mkdir "$temporary/context"
-  "${command[@]}" --output "type=local,dest=$temporary/snapshot" "$temporary/context"
+  mkdir "$temporary"/{context,snapshot}
+  # tar 流避开 local exporter 的逐文件同步阻塞，直接解包以免额外落盘整份归档。
+  # https://github.com/moby/buildkit/issues/2950
+  "${command[@]}" --output type=tar,dest=- "$temporary/context" \
+    | tar -xf - -C "$temporary/snapshot"
   if usable "$temporary/snapshot"; then
     rm -rf "$directory"
     mv "$temporary/snapshot" "$directory"

backend/Cargo.lock

@@ -145,6 +145,18 @@ dependencies = [
  "serde_json",
 ]
 
+[[package]]
+name = "async-compression"
+version = "0.4.50"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "ee19bd99b43e3691acbad4e840420a4881cea6c0b66a208125a824f8fd53f5a1"
+dependencies = [
+ "compression-codecs",
+ "compression-core",
+ "pin-project-lite",
+ "tokio",
+]
+
 [[package]]
 name = "async-lock"
 version = "3.4.2"
@@ -931,6 +943,23 @@ dependencies = [
  "tokio-util",
 ]
 
+[[package]]
+name = "compression-codecs"
+version = "0.4.45"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "98fc98460ba0ad5317075d3632b8dfc45d0be8c4a49347c2a38272019717614a"
+dependencies = [
+ "compression-core",
+ "flate2",
+ "memchr",
+]
+
+[[package]]
+name = "compression-core"
+version = "0.4.33"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "6e8ccc4ea9f6acc32d102c0f6d471d11d913ad15f20c04de743374861fa1d414"
+
 [[package]]
 name = "config"
 version = "0.15.27"
@@ -1717,6 +1746,7 @@ dependencies = [
  "http",
  "http-body",
  "http-body-util",
+ "regex",
  "secrecy",
  "serde",
  "serde_json",
@@ -3750,9 +3780,9 @@ dependencies = [
 
 [[package]]
 name = "rustls"
-version = "0.23.45"
+version = "0.23.46"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "0d41d731c7d2f962d1ccc364cec258de3c0e93b38c2fb3ba97ac74513048d634"
+checksum = "48e13bd8c0e9365c43cfa5c9e8f9ad49d3c8444926c9aac819e0e4dc503c8fdf"
 dependencies = [
  "aws-lc-rs",
  "log",
@@ -4743,6 +4773,7 @@ version = "0.7.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "08a05a66a4fdd61cbbe0a1d755ffe0ca6aba159dd4820936a0ff8a8278245b9c"
 dependencies = [
+ "async-compression",
  "bitflags",
  "bytes",
  "futures-core",

backend/apps/gateway/src/bootstrap/command.rs

@@ -107,7 +107,7 @@ async fn execute(
     let settings = gateway_admin::initialize_settings(
         store.admin_ports().settings(),
         core.snapshot_control(),
-        providers.admin,
+        providers.admin.clone(),
         std::sync::Arc::new(gateway_host::pricing::ModelsDevPricing),
     );
     let mut services = gateway_admin::public_service::Registry::new(service_middleware(
@@ -115,6 +115,12 @@ async fn execute(
         &plugin_runtime,
     ));
     or_shutdown!(plugin_runtime, services.register_settings(&settings));
+    let usage_data = gateway_admin::initialize_usage_data(
+        store.admin_ports().usage_data(),
+        store.admin_ports().settings(),
+        providers.admin,
+    );
+    or_shutdown!(plugin_runtime, services.register_usage_data(&usage_data));
     let services = std::sync::Arc::new(services);
     or_shutdown!(plugin_runtime, plugin_runtime.bind_services(&services));
     let commands = or_shutdown!(plugin_runtime, plugin_runtime.prepare_command_line().await);

backend/apps/gateway/tests/architecture.rs

@@ -4,6 +4,7 @@ use std::{
     collections::{BTreeMap, BTreeSet},
     fs,
     path::{Path, PathBuf},
+    sync::OnceLock,
 };
 
 use syn::{Item, visit::Visit};
@@ -26,22 +27,24 @@ pub(super) const WORKSPACE_MEMBERS: &[&str] = &[
 
 #[test]
 fn workspace_member_list_matches_the_frozen_dag_scope() {
-    let manifest =
-        fs::read_to_string(backend_root().join("Cargo.toml")).expect("read workspace manifest");
-    for member in WORKSPACE_MEMBERS {
-        assert!(
-            manifest.contains(&format!("\"{member}\"")),
-            "workspace must include {member}"
-        );
-    }
-    let member_count = manifest
-        .lines()
-        .map(str::trim)
-        .filter(|line| line.starts_with("\"apps/") || line.starts_with("\"crates/"))
-        .count();
+    let metadata = cargo_metadata_json();
+    let root = backend_root();
+    let actual: BTreeSet<_> = metadata["packages"]
+        .as_array()
+        .expect("Cargo packages")
+        .iter()
+        .map(|package| {
+            Path::new(package["manifest_path"].as_str().expect("package manifest"))
+                .parent()
+                .expect("package directory")
+                .strip_prefix(&root)
+                .expect("workspace member path")
+                .to_path_buf()
+        })
+        .collect();
     assert_eq!(
-        member_count,
-        WORKSPACE_MEMBERS.len(),
+        actual,
+        WORKSPACE_MEMBERS.iter().map(PathBuf::from).collect(),
         "new workspace members must extend the dependency DAG rules"
     );
 }
@@ -378,7 +381,7 @@ const ALLOWED_INTERNAL_EDGES: &[(&str, &str)] = &[
 #[test]
 fn workspace_internal_dependency_edges_match_frozen_dag() {
     let metadata = cargo_metadata_json();
-    let actual = workspace_runtime_dependency_edges(&metadata);
+    let actual = workspace_runtime_dependency_edges(metadata);
     let mut expected: Vec<(String, String)> = ALLOWED_INTERNAL_EDGES
         .iter()
         .map(|(from, to)| ((*from).to_owned(), (*to).to_owned()))
@@ -474,24 +477,6 @@ fn adapter_public_module_surfaces_match_allowlist() {
     }
 }
 
-#[test]
-fn production_sources_do_not_host_tests() {
-    for member in WORKSPACE_MEMBERS {
-        let src = backend_root().join(member).join("src");
-        let sources = super::rust_files(&src);
-        assert!(!sources.is_empty(), "{member} has no production sources");
-        for relative in sources {
-            let path = src.join(relative);
-            let source = fs::read_to_string(&path).expect("read production source");
-            assert!(
-                !source.contains("#[cfg(test)]"),
-                "{} hosts tests in production src",
-                path.display()
-            );
-        }
-    }
-}
-
 #[test]
 fn workspace_modules_follow_conventional_file_layout() {
     let targets = test_source_roots();
@@ -748,40 +733,46 @@ pub(super) fn backend_root() -> PathBuf {
         .expect("backend workspace root")
 }
 
-fn cargo_metadata_json() -> serde_json::Value {
-    let output = std::process::Command::new("cargo")
-        .args(["metadata", "--format-version", "1", "--no-deps"])
-        .current_dir(backend_root())
-        .output()
-        .expect("run cargo metadata");
-    assert!(
-        output.status.success(),
-        "cargo metadata failed: {}",
-        String::from_utf8_lossy(&output.stderr)
-    );
-    serde_json::from_slice(&output.stdout).expect("parse cargo metadata")
+fn cargo_metadata_json() -> &'static serde_json::Value {
+    static METADATA: OnceLock<serde_json::Value> = OnceLock::new();
+    METADATA.get_or_init(|| {
+        let output = std::process::Command::new("cargo")
+            .args(["metadata", "--format-version", "1", "--no-deps", "--locked"])
+            .current_dir(backend_root())
+            .output()
+            .expect("run cargo metadata");
+        assert!(
+            output.status.success(),
+            "cargo metadata failed: {}",
+            String::from_utf8_lossy(&output.stderr)
+        );
+        serde_json::from_slice(&output.stdout).expect("parse cargo metadata")
+    })
 }
 
-/// 提取成员 `[dependencies]` 段内声明的依赖名;段落以下一个 `[` 表头结束
+/// 复用 Cargo 解析后的实际包名，覆盖重命名依赖与平台条件，排除测试和构建依赖
 fn dependency_names(member: &str) -> Vec<String> {
-    let manifest = fs::read_to_string(backend_root().join(member).join("Cargo.toml"))
-        .expect("read member manifest");
-    let mut in_dependencies = false;
-    let mut names = Vec::new();
-    for line in manifest.lines() {
-        let line = line.trim();
-        if line.starts_with('[') {
-            in_dependencies = line == "[dependencies]";
-            continue;
-        }
-        if !in_dependencies || line.is_empty() || line.starts_with('#') {
-            continue;
-        }
-        if let Some((name, _)) = line.split_once('=') {
-            names.push(name.trim().to_owned());
-        }
-    }
-    names
+    let manifest = backend_root().join(member).join("Cargo.toml");
+    let package = cargo_metadata_json()["packages"]
+        .as_array()
+        .expect("Cargo packages")
+        .iter()
+        .find(|package| {
+            Path::new(package["manifest_path"].as_str().expect("package manifest")) == manifest
+        })
+        .expect("workspace member");
+    package["dependencies"]
+        .as_array()
+        .expect("package dependencies")
+        .iter()
+        .filter(|dependency| dependency["kind"].is_null())
+        .map(|dependency| {
+            dependency["name"]
+                .as_str()
+                .expect("dependency name")
+                .to_owned()
+        })
+        .collect()
 }
 
 fn assert_no_dependency(member: &str, forbidden: &[&str]) {

backend/apps/gateway/tests/main.rs

@@ -59,9 +59,16 @@ fn cargo_library_root_is_conventional_lib() {
 fn workspace_production_files_have_no_hidden_modules_or_test_hooks() {
     for member in architecture::WORKSPACE_MEMBERS {
         let root = architecture::backend_root().join(member).join("src");
-        for relative in rust_files(&root) {
+        let sources = rust_files(&root);
+        assert!(!sources.is_empty(), "{member} has no production sources");
+        for relative in sources {
             let path = root.join(relative);
             let source = fs::read_to_string(&path).expect("read production source");
+            assert!(
+                !source.contains("#[cfg(test)]"),
+                "{} hosts tests in production src",
+                path.display()
+            );
             assert!(
                 !source.contains("include!("),
                 "{} uses include!",

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

@@ -39,6 +39,7 @@ pub use use_case::{
     proxies::ProxiesService,
     settings::SettingsService,
     system::SystemService,
+    usage_data::UsageDataService,
 };
 
 use model::AdminError;
@@ -217,6 +218,7 @@ pub struct AdminServices {
     client_distribution: Arc<dyn ClientDistributionService>,
     observability: Arc<dyn ObservabilityService>,
     settings: Arc<dyn SettingsService>,
+    usage_data: Arc<dyn UsageDataService>,
     system: Arc<dyn SystemService>,
     credentials: Arc<CredentialsService>,
     // Runtime 只持有 Weak，这个强引用保障账号端口与 Admin 服务同寿命
@@ -285,6 +287,11 @@ impl AdminServices {
         self.client_distribution.as_ref()
     }
 
+    #[must_use]
+    pub fn usage_data(&self) -> &dyn UsageDataService {
+        self.usage_data.as_ref()
+    }
+
     #[must_use]
     pub fn observability(&self) -> &dyn ObservabilityService {
         self.observability.as_ref()
@@ -467,8 +474,10 @@ async fn initialize_inner(
         registry.clone(),
         pricing_source,
     );
+    let usage_data = initialize_usage_data(store.usage_data(), store.settings(), registry.clone());
     let mut public_services = public_service::Registry::new(service_middleware);
     public_services.register_settings(&settings)?;
+    public_services.register_usage_data(&usage_data)?;
     let services = AdminServices {
         timezone,
         public_services: Arc::new(public_services),
@@ -512,8 +521,10 @@ async fn initialize_inner(
             store.settings(),
             registry.clone(),
             timezone,
+            system.clone(),
         )),
         settings,
+        usage_data,
         system,
         credentials,
         _plugin_accounts: plugin_accounts,
@@ -644,3 +655,15 @@ pub fn initialize_settings(
         pricing_source,
     ))
 }
+
+/// HTTP、插件及 CLI 共用的用量领域用例，宿主提供已验证的调用主体
+#[must_use]
+pub fn initialize_usage_data(
+    store: Arc<dyn ports::usage_data::UsageDataStore>,
+    settings: Arc<dyn ports::store::SettingsStore>,
+    providers: ProviderAdminRegistry,
+) -> Arc<dyn UsageDataService> {
+    Arc::new(use_case::usage_data::DefaultUsageDataService::new(
+        store, settings, providers,
+    ))
+}

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

@@ -100,6 +100,7 @@ pub struct AccountGroupListQuery {
 /// 完整账号分组摘要
 #[derive(Debug, Clone, PartialEq, Eq)]
 pub struct AccountGroupRecord {
+    pub allow_image_generation: bool,
     pub fast_mode: FastMode,
     pub id: AccountGroupId,
     pub name: String,
@@ -139,6 +140,7 @@ pub struct AccountGroupOptionsPage {
 /// 创建账号分组
 #[derive(Debug, Clone, PartialEq, Eq)]
 pub struct CreateAccountGroup {
+    pub allow_image_generation: bool,
     pub fast_mode: FastMode,
     pub name: String,
     pub description: Option<String>,
@@ -148,6 +150,7 @@ pub struct CreateAccountGroup {
 /// 已生成稳定标识、可直接提交存储的创建命令
 #[derive(Debug, Clone, PartialEq, Eq)]
 pub struct NewAccountGroup {
+    pub allow_image_generation: bool,
     pub fast_mode: FastMode,
     pub id: AccountGroupId,
     pub name: String,
@@ -158,6 +161,7 @@ pub struct NewAccountGroup {
 /// 更新账号分组的描述字段
 #[derive(Debug, Clone, PartialEq, Eq)]
 pub struct UpdateAccountGroup {
+    pub allow_image_generation: Option<bool>,
     pub fast_mode: Option<FastMode>,
     pub id: AccountGroupId,
     pub name: String,

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

@@ -124,6 +124,7 @@ pub struct AccountRecord {
     pub weight: AccountWeight,
     pub model_access: gateway_core::account::AccountModelAccess,
     pub outbound_proxy: Option<gateway_core::account::OutboundProxy>,
+    pub outbound_proxy_enabled: bool,
     pub credential_state: CredentialState,
     pub credential_observed_at: DateTime<Utc>,
     pub quota: QuotaState,
@@ -133,6 +134,16 @@ pub struct AccountRecord {
     pub updated_at: DateTime<Utc>,
 }
 
+impl AccountRecord {
+    /// 配置保留绑定，执行请求时禁用代理按直连处理
+    #[must_use]
+    pub fn effective_outbound_proxy(&self) -> Option<&gateway_core::account::OutboundProxy> {
+        self.outbound_proxy
+            .as_ref()
+            .filter(|_| self.outbound_proxy_enabled)
+    }
+}
+
 pub use super::observability::{
     AccountModelUsage, AccountPoolMetrics as AccountSummary, AccountRequestBucket,
     CurrencyCost as AccountCost,

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

@@ -0,0 +1,44 @@
+//! 故障诊断导出的目标、已采集证据与缺口
+
+use chrono::{DateTime, Utc};
+
+use super::{
+    observability::{OpsError, UsageDetail},
+    system::SystemVersion,
+};
+
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub enum DiagnosticTarget {
+    Request(String),
+    Operation(String),
+}
+
+impl DiagnosticTarget {
+    #[must_use]
+    pub fn id(&self) -> &str {
+        match self {
+            Self::Request(id) | Self::Operation(id) => id,
+        }
+    }
+}
+
+pub struct DiagnosticGap {
+    pub section: String,
+    pub reason: &'static str,
+}
+
+pub struct DiagnosticRequest {
+    pub request_id: String,
+    pub detail: Option<UsageDetail>,
+    pub errors: Vec<OpsError>,
+}
+
+/// 原始受控详情只能经导出投影脱敏后交付，不参与普通日志或 Debug
+pub struct DiagnosticBundle {
+    pub target: DiagnosticTarget,
+    pub collected_at: DateTime<Utc>,
+    pub environment: Option<SystemVersion>,
+    pub error: Option<OpsError>,
+    pub requests: Vec<DiagnosticRequest>,
+    pub gaps: Vec<DiagnosticGap>,
+}

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

@@ -10,6 +10,7 @@ pub mod auth;
 pub mod backup;
 pub mod client_distribution;
 pub mod client_keys;
+pub mod diagnostics;
 pub mod import_tasks;
 pub mod key_usage;
 pub mod observability;
@@ -24,6 +25,7 @@ pub mod quota_forecast_sampling;
 pub mod retention;
 pub mod settings;
 pub mod system;
+pub mod usage_data;
 
 /// 管理用例对外返回的稳定错误分类
 #[derive(Debug, Clone, Copy, PartialEq, Eq)]

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

@@ -172,9 +172,19 @@ pub struct UsageFilter {
     pub search: Option<String>,
 }
 
+/// 请求列表的结果筛选，不改变汇总和计量口径
+#[derive(Debug, Clone, Copy, Default, PartialEq, Eq)]
+pub enum UsageRecordStatus {
+    #[default]
+    Success,
+    Error,
+    All,
+}
+
 /// 用量记录分页查询
 #[derive(Debug, Clone, PartialEq, Eq)]
 pub struct UsageQuery {
+    pub status: UsageRecordStatus,
     pub range: TimeRange,
     pub filter: UsageFilter,
     pub current_page: u32,
@@ -653,6 +663,7 @@ pub struct DashboardObservation {
 /// 使用记录表格的窄读模型
 #[derive(Debug, Clone, PartialEq)]
 pub struct UsageListRecord {
+    pub outcome: RequestOutcome,
     pub client_api_key_name: Option<String>,
     pub id: String,
     pub endpoint: String,
@@ -845,7 +856,7 @@ pub struct UsageAttempt {
     pub occurred_at: DateTime<Utc>,
 }
 
-/// 一条请求及其全部尝试
+/// 一条请求及其保留的尝试与诊断轨迹
 #[derive(Debug, Clone, PartialEq)]
 pub struct UsageDetail {
     pub trace: Option<serde_json::Value>,

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

@@ -69,6 +69,7 @@ pub struct ProxyTestResult {
 
 #[derive(Debug, Clone)]
 pub struct ProxyRecord {
+    pub enabled: bool,
     pub auto_location: bool,
     pub detected_location: Option<DetectedProxyLocation>,
     pub location: Option<gateway_core::account::RequestLocation>,
@@ -103,6 +104,7 @@ pub struct NewProxy {
 
 #[derive(Debug, Clone)]
 pub struct UpdateProxy {
+    pub enabled: Option<bool>,
     pub auto_location: Option<bool>,
     pub test: Option<ProxyTestResult>,
     /// 外层为空保留配置，内层为空恢复全局继承

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

@@ -24,7 +24,7 @@ impl RetentionPolicy {
         audit_days: u32,
     ) -> Result<(), &'static str> {
         for (valid, field) in [
-            (usage_days >= 31, "usage_retention_days"),
+            ((1..=31).contains(&usage_days), "usage_retention_days"),
             (ops_days > 0, "ops_event_retention_days"),
             (audit_days > 0, "audit_retention_days"),
         ] {
@@ -38,7 +38,7 @@ impl RetentionPolicy {
     pub fn try_new(usage_days: u32, ops_days: u32, audit_days: u32) -> Result<Self, AdminError> {
         if Self::validate_values(usage_days, ops_days, audit_days).is_err() {
             return Err(AdminError::invalid(
-                "请求保留期至少 31 天，事件与审计保留期必须为正数",
+                "请求保留期必须为 1～31 天，事件与审计保留期必须为正数",
             ));
         }
         Ok(Self {

backend/crates/gateway-admin/src/model/usage_data/actor.rs

@@ -0,0 +1,26 @@
+//! 由宿主传入的可信写入主体，不属于插件请求载荷
+
+use serde::{Deserialize, Serialize};
+
+#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
+#[serde(tag = "kind", rename_all = "snake_case", deny_unknown_fields)]
+pub enum UsageActor {
+    Admin {
+        subject: String,
+    },
+    Plugin {
+        publisher: String,
+        name: String,
+        instance_id: String,
+    },
+    System,
+}
+
+impl super::UsageRecordChange {
+    #[must_use]
+    pub fn id(&self) -> &str {
+        match self {
+            Self::Put { id, .. } | Self::Update { id, .. } | Self::Delete { id } => id,
+        }
+    }
+}

backend/crates/gateway-admin/src/model/usage_data/aggregate.rs

@@ -0,0 +1,87 @@
+//! 按归属时间和维度维护统计贡献，业务来源与调整历史由调用方保存
+
+use super::{UsageDelta, UsageRecordFilter};
+use serde::{Deserialize, Serialize};
+use std::collections::BTreeMap;
+
+#[derive(Debug, Clone, Copy, PartialEq, Eq, PartialOrd, Ord, Serialize, Deserialize)]
+#[serde(rename_all = "snake_case")]
+pub enum UsageMetric {
+    RequestCount,
+    SuccessCount,
+    FailureCount,
+    CancelledCount,
+    IncompleteCount,
+    CompletedRequestCount,
+    InputTokens,
+    OutputTokens,
+    CachedInputTokens,
+    CacheWriteInputTokens,
+    ReasoningOutputTokens,
+    ImageInputTokens,
+    ImageOutputTokens,
+    TotalTokens,
+}
+
+#[derive(Debug, Clone, Default, PartialEq, Eq, Serialize, Deserialize)]
+#[serde(deny_unknown_fields)]
+pub struct UsageDimensions {
+    pub provider: Option<String>,
+    pub model: Option<String>,
+    pub account_ref: Option<String>,
+    pub client_key_ref: Option<String>,
+    pub operation: Option<String>,
+    pub transport: Option<String>,
+}
+
+/// 整个键决定覆盖目标；调用方可将来源标识编码进 key，重试相同键不会累计
+#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
+#[serde(deny_unknown_fields)]
+pub struct UsageAggregateKey {
+    pub bucket_start_ms: i64,
+    pub dimensions: UsageDimensions,
+    pub key: String,
+}
+
+/// 值按 bucket_start_ms 归属，不自动摊分；负值用于相对原生统计的修正
+#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
+#[serde(deny_unknown_fields)]
+pub struct UsageAggregate {
+    pub identity: UsageAggregateKey,
+    pub counts: BTreeMap<UsageMetric, UsageDelta>,
+    pub costs: BTreeMap<String, String>,
+}
+
+#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
+#[serde(tag = "kind", rename_all = "snake_case", deny_unknown_fields)]
+pub enum UsageAggregateChange {
+    Put { aggregate: UsageAggregate },
+    Delete { identity: UsageAggregateKey },
+}
+
+#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
+#[serde(deny_unknown_fields)]
+pub struct WriteUsageAggregates {
+    pub changes: Vec<UsageAggregateChange>,
+}
+
+#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
+#[serde(deny_unknown_fields)]
+pub struct UsageAggregateWriteResult {
+    pub changed: Vec<UsageAggregateKey>,
+}
+
+#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
+#[serde(deny_unknown_fields)]
+pub struct ListUsageAggregates {
+    pub filter: UsageRecordFilter,
+    pub cursor: Option<String>,
+    pub limit: Option<u16>,
+}
+
+#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
+#[serde(deny_unknown_fields)]
+pub struct UsageAggregatePage {
+    pub items: Vec<UsageAggregate>,
+    pub next_cursor: Option<String>,
+}

backend/crates/gateway-admin/src/model/usage_data/error.rs

@@ -0,0 +1,44 @@
+//! 用量基础操作的错误分类，不保存失败或重试状态
+
+use serde::{Deserialize, Serialize};
+
+#[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize)]
+#[serde(rename_all = "snake_case")]
+pub enum UsageErrorKind {
+    InvalidInput,
+    NotFound,
+    Conflict,
+    CapacityExceeded,
+    Unavailable,
+}
+
+#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
+#[serde(deny_unknown_fields)]
+pub struct UsageDataError {
+    pub kind: UsageErrorKind,
+    pub message: String,
+}
+impl UsageDataError {
+    #[must_use]
+    pub fn new(kind: UsageErrorKind, message: impl Into<String>) -> Self {
+        Self {
+            kind,
+            message: message.into(),
+        }
+    }
+    #[must_use]
+    pub fn invalid(message: impl Into<String>) -> Self {
+        Self::new(UsageErrorKind::InvalidInput, message)
+    }
+    #[must_use]
+    pub fn unavailable(message: impl Into<String>) -> Self {
+        Self::new(UsageErrorKind::Unavailable, message)
+    }
+}
+impl std::fmt::Display for UsageDataError {
+    fn fmt(&self, formatter: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        formatter.write_str(&self.message)
+    }
+}
+impl std::error::Error for UsageDataError {}
+pub type UsageDataResult<T> = Result<T, UsageDataError>;

backend/crates/gateway-admin/src/model/usage_data/mod.rs

@@ -0,0 +1,22 @@
+//! 用量明细、聚合维护与纯计价的公开合同
+
+mod value;
+pub use value::*;
+mod record;
+pub use record::*;
+mod query;
+pub use query::*;
+mod pricing;
+pub use pricing::*;
+mod error;
+pub use error::*;
+
+pub const USAGE_DATA_MAX_TARGETS: usize = 500;
+pub const USAGE_DATA_MAX_REQUEST_BYTES: usize = 2 * 1024 * 1024;
+pub const USAGE_DATA_MAX_PAGE_SIZE: u16 = 200;
+
+mod actor;
+pub use actor::*;
+
+mod aggregate;
+pub use aggregate::*;

backend/crates/gateway-admin/src/model/usage_data/pricing.rs

@@ -0,0 +1,62 @@
+//! 纯 Token 计价合同，费用是否采用与额外费用合并由调用方决定
+
+use super::{UsageCost, UsageTokens};
+use gateway_core::metering::ModelPriceOverride;
+use serde::{Deserialize, Serialize};
+
+#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
+#[serde(tag = "kind", rename_all = "snake_case", deny_unknown_fields)]
+pub enum UsagePriceSelection {
+    Current,
+    Override { price: Box<ModelPriceOverride> },
+}
+
+#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
+#[serde(deny_unknown_fields)]
+pub struct UsagePricingInput {
+    pub id: String,
+    pub provider: String,
+    pub model: String,
+    pub tokens: UsageTokens,
+    pub service_tier: Option<String>,
+    pub price: UsagePriceSelection,
+}
+
+#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
+#[serde(deny_unknown_fields)]
+pub struct EvaluateUsagePricing {
+    pub items: Vec<UsagePricingInput>,
+}
+
+#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
+#[serde(deny_unknown_fields)]
+pub struct UsagePricingBreakdown {
+    pub input_amount: String,
+    pub output_amount: String,
+    pub cache_read_amount: String,
+    pub cache_write_amount: String,
+    pub standard_amount: String,
+    pub long_context_billing_applied: bool,
+    pub service_tier: Option<String>,
+    pub custom_multiplier_bps: u32,
+}
+
+#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
+#[serde(tag = "status", rename_all = "snake_case", deny_unknown_fields)]
+pub enum UsagePricingResult {
+    Estimated {
+        id: String,
+        cost: UsageCost,
+        breakdown: UsagePricingBreakdown,
+    },
+    NotEstimable {
+        id: String,
+        reason: String,
+    },
+}
+
+#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
+#[serde(deny_unknown_fields)]
+pub struct UsagePricingEvaluation {
+    pub items: Vec<UsagePricingResult>,
+}

backend/crates/gateway-admin/src/model/usage_data/query.rs

@@ -0,0 +1,36 @@
+//! 请求明细的有界筛选与稳定顺序分页
+
+use super::UsageRecord;
+use serde::{Deserialize, Serialize};
+
+#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
+#[serde(deny_unknown_fields)]
+pub struct GetUsageRecord {
+    pub id: String,
+}
+
+#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
+#[serde(deny_unknown_fields)]
+pub struct UsageRecordFilter {
+    pub start_ms: i64,
+    pub end_ms: i64,
+    pub provider: Option<String>,
+    pub model: Option<String>,
+    pub account_ref: Option<String>,
+    pub client_key_ref: Option<String>,
+}
+
+#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
+#[serde(deny_unknown_fields)]
+pub struct ListUsageRecords {
+    pub filter: UsageRecordFilter,
+    pub cursor: Option<String>,
+    pub limit: Option<u16>,
+}
+
+#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
+#[serde(deny_unknown_fields)]
+pub struct UsageRecordPage {
+    pub items: Vec<UsageRecord>,
+    pub next_cursor: Option<String>,
+}

backend/crates/gateway-admin/src/model/usage_data/record.rs

@@ -0,0 +1,106 @@
+//! 既有请求明细的用量读写合同，不承载插件任务或来源状态
+
+use super::UsageCount;
+use serde::{Deserialize, Serialize};
+
+/// Token 使用宿主口径：缓存包含在输入中，推理包含在输出中，未知值为 null
+#[derive(Debug, Clone, Default, PartialEq, Eq, Serialize, Deserialize)]
+#[serde(deny_unknown_fields)]
+pub struct UsageTokens {
+    pub input_tokens: Option<UsageCount>,
+    pub output_tokens: Option<UsageCount>,
+    pub cached_input_tokens: Option<UsageCount>,
+    pub cache_write_input_tokens: Option<UsageCount>,
+    pub reasoning_output_tokens: Option<UsageCount>,
+    pub image_input_tokens: Option<UsageCount>,
+    pub image_output_tokens: Option<UsageCount>,
+    pub total_tokens: Option<UsageCount>,
+}
+
+#[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize)]
+#[serde(rename_all = "snake_case")]
+pub enum UsageCostSource {
+    ProviderReported,
+    Calculated,
+}
+
+#[derive(Debug, Clone, Default, PartialEq, Eq, Serialize, Deserialize)]
+#[serde(tag = "kind", rename_all = "snake_case", deny_unknown_fields)]
+pub enum UsageCost {
+    #[default]
+    Unavailable,
+    Known {
+        source: UsageCostSource,
+        amount: String,
+        currency: String,
+    },
+}
+
+#[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize)]
+#[serde(rename_all = "snake_case")]
+pub enum UsageOutcome {
+    Running,
+    Succeeded,
+    Failed,
+    Cancelled,
+    Incomplete,
+}
+
+#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
+#[serde(deny_unknown_fields)]
+pub struct UsageRecordData {
+    pub client_key_ref: String,
+    pub provider: Option<String>,
+    pub account_ref: Option<String>,
+    pub requested_model: Option<String>,
+    pub model: Option<String>,
+    pub protocol: String,
+    pub operation: String,
+    pub endpoint: String,
+    pub client_transport: String,
+    pub service_tier: Option<String>,
+    pub started_at_ms: i64,
+    pub completed_at_ms: Option<i64>,
+    pub outcome: UsageOutcome,
+    pub tokens: UsageTokens,
+    pub cost: UsageCost,
+}
+
+#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
+#[serde(deny_unknown_fields)]
+pub struct UsageRecord {
+    pub id: String,
+    pub data: UsageRecordData,
+    pub billing_snapshot: Option<serde_json::Value>,
+    pub provider_observation: Option<serde_json::Value>,
+}
+
+#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
+#[serde(tag = "kind", rename_all = "snake_case", deny_unknown_fields)]
+pub enum UsageRecordChange {
+    Put {
+        id: String,
+        data: Box<UsageRecordData>,
+    },
+    /// tokens 替换全部 Token 字段；cost 的 unavailable 明确清空费用
+    Update {
+        id: String,
+        tokens: Option<UsageTokens>,
+        cost: Option<UsageCost>,
+    },
+    Delete {
+        id: String,
+    },
+}
+
+#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
+#[serde(deny_unknown_fields)]
+pub struct WriteUsageRecords {
+    pub changes: Vec<UsageRecordChange>,
+}
+
+#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
+#[serde(deny_unknown_fields)]
+pub struct UsageWriteResult {
+    pub changed_ids: Vec<String>,
+}

backend/crates/gateway-admin/src/model/usage_data/value.rs

@@ -0,0 +1,63 @@
+//! 用量计数使用十进制字符串，避免插件前端丢失整数精度
+
+use serde::{Deserialize, Deserializer, Serialize, Serializer, de::Error};
+use std::fmt;
+
+#[derive(Debug, Clone, Copy, Default, PartialEq, Eq, PartialOrd, Ord, Hash)]
+pub struct UsageCount(pub u64);
+impl UsageCount {
+    pub const ZERO: Self = Self(0);
+    #[must_use]
+    pub const fn get(self) -> u64 {
+        self.0
+    }
+}
+impl From<u64> for UsageCount {
+    fn from(value: u64) -> Self {
+        Self(value)
+    }
+}
+impl fmt::Display for UsageCount {
+    fn fmt(&self, formatter: &mut fmt::Formatter<'_>) -> fmt::Result {
+        self.0.fmt(formatter)
+    }
+}
+impl Serialize for UsageCount {
+    fn serialize<S: Serializer>(&self, serializer: S) -> Result<S::Ok, S::Error> {
+        serializer.serialize_str(&self.0.to_string())
+    }
+}
+impl<'de> Deserialize<'de> for UsageCount {
+    fn deserialize<D: Deserializer<'de>>(deserializer: D) -> Result<Self, D::Error> {
+        let value = String::deserialize(deserializer)?;
+        if value.is_empty() || !value.bytes().all(|byte| byte.is_ascii_digit()) {
+            return Err(D::Error::custom("count must be an unsigned decimal string"));
+        }
+        value.parse().map(Self).map_err(D::Error::custom)
+    }
+}
+
+/// 有符号统计贡献受 PostgreSQL bigint 范围限制，JSON 使用十进制字符串
+#[derive(Debug, Clone, Copy, Default, PartialEq, Eq, PartialOrd, Ord, Hash)]
+pub struct UsageDelta(pub i64);
+impl UsageDelta {
+    #[must_use]
+    pub const fn get(self) -> i64 {
+        self.0
+    }
+}
+impl Serialize for UsageDelta {
+    fn serialize<S: Serializer>(&self, serializer: S) -> Result<S::Ok, S::Error> {
+        serializer.serialize_str(&self.0.to_string())
+    }
+}
+impl<'de> Deserialize<'de> for UsageDelta {
+    fn deserialize<D: Deserializer<'de>>(deserializer: D) -> Result<Self, D::Error> {
+        let value = String::deserialize(deserializer)?;
+        let digits = value.strip_prefix('-').unwrap_or(&value);
+        if digits.is_empty() || !digits.bytes().all(|byte| byte.is_ascii_digit()) {
+            return Err(D::Error::custom("delta must be a signed decimal string"));
+        }
+        value.parse().map(Self).map_err(D::Error::custom)
+    }
+}

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

@@ -14,3 +14,4 @@ pub mod proxy;
 pub mod retention;
 pub mod store;
 pub mod system;
+pub mod usage_data;

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

@@ -26,6 +26,14 @@ use crate::model::{
     quota_forecast_sampling::QuotaForecastObservation,
 };
 
+/// 按原生 Token 口径计算的报价输入，额外费用由调用方处理
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub struct ProviderUsagePricingInput {
+    pub upstream_model_id: String,
+    pub service_tier: Option<String>,
+    pub usage: gateway_core::metering::Usage,
+}
+
 /// Provider 管理失败的稳定分类
 #[derive(Debug, Clone, Copy, PartialEq, Eq)]
 pub enum ProviderAdminErrorKind {
@@ -161,6 +169,14 @@ pub trait ProviderAdmin: Send + Sync {
         Default::default()
     }
 
+    fn evaluate_usage_pricing(
+        &self,
+        _input: &ProviderUsagePricingInput,
+        _pricing: Option<&gateway_core::metering::ModelPriceOverride>,
+    ) -> Result<Option<gateway_core::metering::CalculatedCostBreakdown>, ProviderAdminError> {
+        Err(ProviderAdminError::new(ProviderAdminErrorKind::Unsupported))
+    }
+
     fn provider_kind(&self) -> &ProviderKind;
 
     /// 只读取原生 Provider 能力及账号类型，不执行网络或数据库查询

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

@@ -485,6 +485,13 @@ pub trait ObservabilityStore: Send + Sync {
     ) -> AdminStoreResult<DiagnosticsObservation>;
 
     async fn list_ops_errors(&self, query: OpsErrorQuery) -> AdminStoreResult<OpsErrorPage>;
+
+    /// 按精确目标读取受控错误，不受列表时间窗口与当前页面影响
+    async fn diagnostic_errors(
+        &self,
+        target: &crate::model::diagnostics::DiagnosticTarget,
+        limit: u16,
+    ) -> AdminStoreResult<Vec<crate::model::observability::OpsError>>;
 }
 
 /// Runtime settings 与管理员 API Key 写入
@@ -565,9 +572,15 @@ pub struct AdminStorePorts {
     plugins: Arc<dyn super::plugins::PluginStore>,
     plugin_state: Arc<dyn super::plugins::PluginStateStore>,
     plugin_resources: Arc<dyn super::plugin_resources::PluginResourceStore>,
+    usage_data: Arc<dyn super::usage_data::UsageDataStore>,
 }
 
 impl AdminStorePorts {
+    #[must_use]
+    pub fn usage_data(&self) -> Arc<dyn super::usage_data::UsageDataStore> {
+        self.usage_data.clone()
+    }
+
     #[must_use]
     pub fn plugin_resources(&self) -> Arc<dyn super::plugin_resources::PluginResourceStore> {
         self.plugin_resources.clone()
@@ -598,6 +611,7 @@ impl AdminStorePorts {
         plugins: Arc<dyn super::plugins::PluginStore>,
         plugin_state: Arc<dyn super::plugins::PluginStateStore>,
         plugin_resources: Arc<dyn super::plugin_resources::PluginResourceStore>,
+        usage_data: Arc<dyn super::usage_data::UsageDataStore>,
     ) -> Self {
         Self {
             accounts,
@@ -609,6 +623,7 @@ impl AdminStorePorts {
             plugins,
             plugin_state,
             plugin_resources,
+            usage_data,
         }
     }
 

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

@@ -0,0 +1,27 @@
+//! 原生请求用量的有界读取和事务写入端口
+
+use crate::model::{auth::AdminAuditEvent, usage_data::*};
+use async_trait::async_trait;
+
+pub struct StoredUsageWrite {
+    pub changes: Vec<(UsageRecordChange, AdminAuditEvent)>,
+}
+
+pub struct StoredAggregateWrite {
+    pub changes: Vec<(UsageAggregateChange, AdminAuditEvent)>,
+}
+
+#[async_trait]
+pub trait UsageDataStore: Send + Sync {
+    async fn list_aggregates(
+        &self,
+        request: ListUsageAggregates,
+    ) -> UsageDataResult<UsageAggregatePage>;
+    async fn write_aggregates(
+        &self,
+        command: StoredAggregateWrite,
+    ) -> UsageDataResult<UsageAggregateWriteResult>;
+    async fn get_record(&self, request: GetUsageRecord) -> UsageDataResult<UsageRecord>;
+    async fn list_records(&self, request: ListUsageRecords) -> UsageDataResult<UsageRecordPage>;
+    async fn write_records(&self, command: StoredUsageWrite) -> UsageDataResult<UsageWriteResult>;
+}

backend/crates/gateway-admin/src/public_service/actor.rs

@@ -0,0 +1,21 @@
+//! 公开服务写入主体由受信宿主边界设置，插件输入不能覆盖
+
+use crate::model::usage_data::UsageActor;
+use std::future::Future;
+
+tokio::task_local! {
+    static ACTOR: UsageActor;
+}
+
+/// 与 middleware Origin 独立传递，嵌套服务不会丢失真实调用主体
+pub async fn scope_actor<F: Future>(actor: UsageActor, future: F) -> F::Output {
+    ACTOR.scope(actor, future).await
+}
+
+pub(super) fn current_actor() -> Result<UsageActor, crate::model::usage_data::UsageDataError> {
+    ACTOR.try_with(Clone::clone).map_err(|_| {
+        crate::model::usage_data::UsageDataError::invalid(
+            "usage mutation requires a trusted caller identity",
+        )
+    })
+}

backend/crates/gateway-admin/src/public_service/mod.rs

@@ -1,7 +1,10 @@
 //! 主动服务调用的统一组合入口；普通业务方法由所属 HTTP 或执行边界包裹
 
+mod actor;
 mod registry;
 mod settings;
+mod usage_data;
+pub use actor::scope_actor;
 pub use registry::Registry;
 
 use std::{future::Future, sync::Arc};

backend/crates/gateway-admin/src/public_service/registry.rs

@@ -54,6 +54,57 @@ impl Registry {
         Ok(())
     }
 
+    pub fn register_usage_data(
+        &mut self,
+        service: &Arc<dyn crate::UsageDataService>,
+    ) -> Result<(), AdminError> {
+        super::usage_data::register(self, service)
+    }
+
+    pub(crate) fn register_usage<I, O>(
+        &mut self,
+        name: &'static str,
+        handler: impl Fn(I) -> BoxFuture<'static, crate::model::usage_data::UsageDataResult<O>>
+        + Send
+        + Sync
+        + 'static,
+    ) -> Result<(), AdminError>
+    where
+        I: DeserializeOwned + Send + 'static,
+        O: Serialize + Send + 'static,
+    {
+        let std::collections::btree_map::Entry::Vacant(entry) = self.operations.entry(name) else {
+            return Err(AdminError::internal(format!("公开服务重复登记：{name}")));
+        };
+        let handler = Arc::new(handler);
+        entry.insert(Arc::new(move |input| {
+            let handler = handler.clone();
+            Box::pin(async move {
+                // 在反序列化之前限制编码后的完整请求，避免单条记录绕过批量行数约束
+                let bytes =
+                    serde_json::to_vec(&input).map_err(|_| Error::invalid("服务输入编码无效"))?;
+                if bytes.len() > crate::model::usage_data::USAGE_DATA_MAX_REQUEST_BYTES {
+                    return Err(super::usage_data::encode_error(
+                        crate::model::usage_data::UsageDataError::new(
+                            crate::model::usage_data::UsageErrorKind::CapacityExceeded,
+                            "usage input exceeds 2 MiB",
+                        ),
+                    ));
+                }
+                let input = serde_json::from_value(input).map_err(|_| {
+                    super::usage_data::encode_error(
+                        crate::model::usage_data::UsageDataError::invalid("服务输入类型无效"),
+                    )
+                })?;
+                let output = handler(input)
+                    .await
+                    .map_err(super::usage_data::encode_error)?;
+                serde_json::to_value(output).map_err(|_| Error::unavailable("服务结果编码失败"))
+            })
+        }));
+        Ok(())
+    }
+
     pub async fn call(&self, operation: &str, input: Value) -> Result<Value, Error> {
         let (operation, handler) =
             self.operations

backend/crates/gateway-admin/src/public_service/usage_data.rs

@@ -0,0 +1,63 @@
+//! 原生用量能力的公开服务登记
+
+use crate::{
+    UsageDataService,
+    model::{AdminError, usage_data::*},
+};
+use std::sync::Arc;
+
+pub(super) fn register(
+    registry: &mut super::Registry,
+    service: &Arc<dyn UsageDataService>,
+) -> Result<(), AdminError> {
+    macro_rules! read {
+        ($name:literal, $method:ident, $input:ty) => {{
+            let service = service.clone();
+            registry.register_usage($name, move |request: $input| {
+                let service = service.clone();
+                Box::pin(async move { service.$method(request).await })
+            })?;
+        }};
+    }
+    read!("usage.records.get", get_record, GetUsageRecord);
+    read!("usage.records.list", list_records, ListUsageRecords);
+    read!(
+        "usage.aggregates.list",
+        list_aggregates,
+        ListUsageAggregates
+    );
+    read!("pricing.evaluate", evaluate_pricing, EvaluateUsagePricing);
+    let aggregates = service.clone();
+    registry.register_usage(
+        "usage.aggregates.write",
+        move |request: WriteUsageAggregates| {
+            let service = aggregates.clone();
+            Box::pin(async move {
+                let actor = super::actor::current_actor()?;
+                service.write_aggregates(&actor, request).await
+            })
+        },
+    )?;
+    let service = service.clone();
+    registry.register_usage("usage.records.write", move |request: WriteUsageRecords| {
+        let service = service.clone();
+        Box::pin(async move {
+            let actor = super::actor::current_actor()?;
+            service.write_records(&actor, request).await
+        })
+    })?;
+    Ok(())
+}
+
+pub(super) fn encode_error(
+    error: UsageDataError,
+) -> gateway_core::engine::middleware::service::Error {
+    gateway_core::engine::middleware::service::Error {
+        kind: serde_json::to_value(error.kind)
+            .ok()
+            .and_then(|v| v.as_str().map(str::to_owned))
+            .unwrap_or_else(|| "unavailable".into()),
+        message: error.message.clone(),
+        details: serde_json::to_value(error).ok(),
+    }
+}

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

@@ -160,6 +160,7 @@ impl AccountGroupService for DefaultAccountGroupService {
             self.store
                 .create_account_group(
                     NewAccountGroup {
+                        allow_image_generation: command.allow_image_generation,
                         id,
                         name: command.name,
                         description: command.description,

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

@@ -106,7 +106,8 @@ pub(super) async fn pending_authorization(
             proxy = required_credential(accounts, provider_kind, account_id, resource)
                 .await?
                 .credential
-                .outbound_proxy;
+                .effective_outbound_proxy()
+                .cloned();
             AuthorizationMutationTarget::Reauthorize {
                 account_id: account_id.clone(),
             }
@@ -121,6 +122,9 @@ pub(super) async fn pending_authorization(
                         .await
                         .map_err(|error| map_store_error(error, "proxy"))?;
                     proxy_id = Some(record.id);
+                    if !record.enabled {
+                        return Err(AdminError::invalid("代理已禁用，请启用后再选择"));
+                    }
                     Some(record.proxy)
                 }
                 Some(AccountProxySelection::Direct) | None => None,

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

@@ -225,6 +225,7 @@ impl KeyUsageService for DefaultKeyUsageService {
             KeyUsageRecordKind::Success => KeyUsageRecords::Success(
                 self.observations
                     .list_usage_records(UsageQuery {
+                        status: Default::default(),
                         range: query.usage.range,
                         filter: usage_filter(&id, query.usage.model),
                         current_page: query.current_page,

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

@@ -18,6 +18,7 @@ pub mod plugins;
 pub mod proxies;
 pub mod settings;
 pub mod system;
+pub mod usage_data;
 
 mod credential_mutation;
 mod error;

backend/crates/gateway-admin/src/use_case/observability/export.rs

@@ -0,0 +1,227 @@
+//! 按故障目标收集诊断证据，补充读取失败不丢弃已取得的记录
+
+use std::{
+    collections::{HashSet, VecDeque},
+    time::Duration,
+};
+
+use tokio::time::{Instant, timeout_at};
+
+use super::{DefaultObservabilityService, map_store_error};
+use crate::model::{
+    AdminError, AdminErrorKind,
+    diagnostics::{DiagnosticBundle, DiagnosticGap, DiagnosticRequest, DiagnosticTarget},
+    observability::OpsError,
+};
+
+const MAX_REQUESTS: usize = 8;
+const MAX_ERRORS: u16 = 128;
+
+impl DefaultObservabilityService {
+    pub(super) async fn collect_diagnostics(
+        &self,
+        target: DiagnosticTarget,
+    ) -> Result<DiagnosticBundle, AdminError> {
+        if target.id().trim().is_empty() || target.id().len() > 256 {
+            return Err(AdminError::invalid("诊断目标 ID 不合法"));
+        }
+        let deadline = Instant::now() + Duration::from_secs(20);
+        let (collected, environment) = tokio::join!(
+            self.collect_records(target, deadline),
+            timeout_at(deadline, self.system.version()),
+        );
+        let mut bundle = collected?;
+        match environment {
+            Ok(Ok(version)) => bundle.environment = Some(version),
+            _ => gap(&mut bundle, "environment", "collection_failed"),
+        }
+        Ok(bundle)
+    }
+
+    async fn collect_records(
+        &self,
+        target: DiagnosticTarget,
+        deadline: Instant,
+    ) -> Result<DiagnosticBundle, AdminError> {
+        let mut bundle = DiagnosticBundle {
+            target,
+            collected_at: chrono::Utc::now(),
+            environment: None,
+            error: None,
+            requests: Vec::new(),
+            gaps: Vec::new(),
+        };
+        let mut remaining_errors = MAX_ERRORS;
+        let first_id = match &bundle.target {
+            DiagnosticTarget::Request(id) => Some(id.clone()),
+            DiagnosticTarget::Operation(_) => {
+                let mut errors = self.export_errors(&bundle.target, 1, deadline).await?;
+                let error = errors
+                    .pop()
+                    .ok_or_else(|| AdminError::not_found("错误记录不存在或已过期"))?;
+                let id = error.request_id.clone();
+                bundle.error = Some(error);
+                remaining_errors -= 1;
+                id
+            }
+        };
+        let mut pending = VecDeque::from_iter(first_id);
+        let mut seen = HashSet::new();
+        while let Some(id) = pending.pop_front() {
+            if !seen.insert(id.clone()) {
+                continue;
+            }
+            if bundle.requests.len() >= MAX_REQUESTS {
+                gap(&mut bundle, "requests", "related_request_limit");
+                break;
+            }
+            let section = format!("requests.{id}");
+            let target = DiagnosticTarget::Request(id.clone());
+            let selected = bundle
+                .error
+                .as_ref()
+                .filter(|error| error.request_id.as_deref() == Some(id.as_str()));
+            // 事件导出已经持有选中记录，额外读取一个槽位用于去重后判断是否超限
+            let read_limit = remaining_errors + 1 + u16::from(selected.is_some());
+            let (detail, errors) = tokio::join!(
+                timeout_at(deadline, self.store.usage_record_detail(&id)),
+                self.export_errors(&target, read_limit, deadline),
+            );
+            let (detail, detail_error) = match detail {
+                Ok(Ok(detail)) => (Some(detail), None),
+                Ok(Err(error)) => (None, Some(map_store_error(error, "diagnostic request"))),
+                Err(_) => (
+                    None,
+                    Some(AdminError::new(AdminErrorKind::Unavailable, "诊断采集超时")),
+                ),
+            };
+            let mut errors = match errors {
+                Ok(mut errors) => {
+                    if let Some(selected) = selected {
+                        errors.retain(|error| {
+                            error.source != selected.source || error.event_id != selected.event_id
+                        });
+                    }
+                    errors
+                }
+                Err(error) => {
+                    if detail.is_none() && bundle.requests.is_empty() && bundle.error.is_none() {
+                        return Err(error);
+                    }
+                    gap(
+                        &mut bundle,
+                        &format!("{section}.errors"),
+                        "collection_failed",
+                    );
+                    Vec::new()
+                }
+            };
+            if detail.is_none()
+                && errors.is_empty()
+                && bundle.requests.is_empty()
+                && bundle.error.is_none()
+            {
+                return Err(
+                    detail_error.unwrap_or_else(|| AdminError::not_found("请求记录不存在或已过期"))
+                );
+            }
+            if errors.len() > usize::from(remaining_errors) {
+                gap(
+                    &mut bundle,
+                    &format!("{section}.errors"),
+                    "error_record_limit",
+                );
+                errors.truncate(usize::from(remaining_errors));
+            }
+            remaining_errors =
+                remaining_errors.saturating_sub(u16::try_from(errors.len()).unwrap_or(MAX_ERRORS));
+            if let Some(record) = &detail {
+                for related in &record.related_requests {
+                    if let Some(id) = related.get("requestId").and_then(serde_json::Value::as_str) {
+                        pending.push_back(id.to_owned());
+                    }
+                }
+                if record.request.outcome == crate::model::observability::RequestOutcome::Running {
+                    gap(&mut bundle, &section, "request_not_finalized");
+                }
+                match &record.trace {
+                    None => gap(
+                        &mut bundle,
+                        &format!("{section}.trace"),
+                        "not_recorded_or_retained",
+                    ),
+                    Some(trace) => {
+                        if trace["droppedEvents"].as_u64().unwrap_or(0) > 0 {
+                            gap(&mut bundle, &format!("{section}.trace"), "events_evicted");
+                        }
+                        let events = trace["events"].as_array();
+                        if !events.is_some_and(|events| {
+                            events.iter().any(|event| {
+                                event["stage"] == "request.routed"
+                                    && event["data"]["accountPolicy"].is_object()
+                            })
+                        }) {
+                            gap(&mut bundle, &section, "incident_policy_not_recorded");
+                        }
+                        for event in events.into_iter().flatten() {
+                            if event["stage"] == "capture.gap" || event["data"]["truncated"] == true
+                            {
+                                gap(
+                                    &mut bundle,
+                                    &format!("{section}.trace"),
+                                    "capture_gap_or_truncation",
+                                );
+                            }
+                            if event["stage"] == "retry.decided" && event["data"]["phase"].is_null()
+                            {
+                                gap(
+                                    &mut bundle,
+                                    &format!("{section}.trace"),
+                                    "legacy_retry_facts_not_recorded",
+                                );
+                            }
+                        }
+                    }
+                }
+                gap(
+                    &mut bundle,
+                    &format!("{section}.attempts"),
+                    "best_effort_persistence",
+                );
+            } else {
+                gap(&mut bundle, &section, "request_detail_unavailable");
+            }
+            bundle.requests.push(DiagnosticRequest {
+                request_id: id,
+                detail,
+                errors,
+            });
+        }
+        Ok(bundle)
+    }
+
+    async fn export_errors(
+        &self,
+        target: &DiagnosticTarget,
+        limit: u16,
+        deadline: Instant,
+    ) -> Result<Vec<OpsError>, AdminError> {
+        timeout_at(deadline, self.store.diagnostic_errors(target, limit))
+            .await
+            .map_err(|_| AdminError::new(AdminErrorKind::Unavailable, "诊断采集超时"))?
+            .map_err(|error| map_store_error(error, "diagnostic errors"))
+    }
+}
+
+fn gap(bundle: &mut DiagnosticBundle, section: &str, reason: &'static str) {
+    if !bundle
+        .gaps
+        .iter()
+        .any(|gap| gap.section == section && gap.reason == reason)
+    {
+        bundle.gaps.push(DiagnosticGap {
+            section: section.to_owned(),
+            reason,
+        });
+    }
+}

backend/crates/gateway-admin/src/use_case/observability/mod.rs

@@ -38,6 +38,7 @@ const HEALTH_TIMELINE_SLOT_MINUTES: i64 = 15;
 const HEALTH_TIMELINE_MIN_SAMPLE_SIZE: u64 = 10;
 const HEALTH_TIMELINE_UNAVAILABLE_FAILURE_THRESHOLD: u64 = 3;
 const HEALTH_TIMELINE_STABLE_RELIABILITY: f64 = 95.0;
+mod export;
 mod store_cache;
 pub(crate) use store_cache::CachedObservabilityStore;
 
@@ -70,6 +71,10 @@ pub trait ObservabilityService: Send + Sync {
         dimension: DiagnosticDimension,
     ) -> Result<DiagnosticsResult, AdminError>;
     async fn ops_errors(&self, query: OpsErrorQuery) -> Result<OpsErrorPage, AdminError>;
+    async fn export_diagnostics(
+        &self,
+        target: crate::model::diagnostics::DiagnosticTarget,
+    ) -> Result<crate::model::diagnostics::DiagnosticBundle, AdminError>;
 }
 
 pub(crate) struct DefaultObservabilityService {
@@ -78,6 +83,7 @@ pub(crate) struct DefaultObservabilityService {
     accounts: Arc<dyn AccountStore>,
     settings: Arc<dyn SettingsStore>,
     providers: ProviderAdminRegistry,
+    system: Arc<dyn super::system::SystemService>,
 }
 
 impl DefaultObservabilityService {
@@ -88,6 +94,7 @@ impl DefaultObservabilityService {
         settings: Arc<dyn SettingsStore>,
         providers: ProviderAdminRegistry,
         timezone: gateway_core::time::DeploymentTimeZone,
+        system: Arc<dyn super::system::SystemService>,
     ) -> Self {
         Self {
             timezone,
@@ -95,6 +102,7 @@ impl DefaultObservabilityService {
             accounts,
             settings,
             providers,
+            system,
         }
     }
 
@@ -192,6 +200,13 @@ impl DefaultObservabilityService {
 
 #[async_trait]
 impl ObservabilityService for DefaultObservabilityService {
+    async fn export_diagnostics(
+        &self,
+        target: crate::model::diagnostics::DiagnosticTarget,
+    ) -> Result<crate::model::diagnostics::DiagnosticBundle, AdminError> {
+        self.collect_diagnostics(target).await
+    }
+
     async fn dashboard_summary(
         &self,
         range: TimeRange,

backend/crates/gateway-admin/src/use_case/observability/store_cache.rs

@@ -1,4 +1,4 @@
-//! 有界统计缓存：身份和完整筛选参与键，删除版本阻止旧数据复活
+//! 有界统计缓存：身份和完整筛选参与键，统计发布代次使旧快照失效
 
 use crate::model::observability::*;
 use crate::ports::store::{
@@ -121,7 +121,7 @@ impl<K: Eq + Hash, T: AnalyticsValue> Cache<K, T> {
         };
         // 同键刷新只执行一次；失败不会缓存成成功，也不延长旧结果寿命
         let mut cached = slot.lock().await;
-        // 等锁期间代次可能推进；取得同键锁后再核对，失败时也不能回用已删除事实
+        // 等锁期间代次可能推进；取得同键锁后再核对，失败时也不能回用已被新统计替代的快照
         let generation = store.analytics_cache_generation().await?;
         if let Some(entry) = cached.as_mut()
             && entry.value.metadata().generation == generation
@@ -289,4 +289,12 @@ impl ObservabilityStore for CachedObservabilityStore {
     async fn list_ops_errors(&self, query: OpsErrorQuery) -> AdminStoreResult<OpsErrorPage> {
         self.inner.list_ops_errors(query).await
     }
+
+    async fn diagnostic_errors(
+        &self,
+        target: &crate::model::diagnostics::DiagnosticTarget,
+        limit: u16,
+    ) -> AdminStoreResult<Vec<crate::model::observability::OpsError>> {
+        self.inner.diagnostic_errors(target, limit).await
+    }
 }

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

@@ -52,6 +52,7 @@ impl PluginResourceAccess for DefaultPluginResourceAccess {
                     owner,
                     resource_key,
                     NewAccountGroup {
+                        allow_image_generation: command.allow_image_generation,
                         id,
                         name: command.name,
                         description: command.description,

backend/crates/gateway-admin/src/use_case/plugins/distribution.rs

@@ -305,6 +305,9 @@ impl PluginsService {
             .get(id)
             .await
             .map_err(|error| map_store_error(error, "proxy"))?;
+        if !record.enabled {
+            return Err(AdminError::invalid("下载代理已禁用，请启用或改为直连"));
+        }
         Ok(Some(PluginDistributionEgress::new(
             record.id,
             record.revision,

backend/crates/gateway-admin/src/use_case/usage_data/mod.rs

@@ -0,0 +1,184 @@
+//! 公开原生用量读写与纯计价，插件拥有跨调用流程和恢复策略
+
+mod pricing;
+mod validation;
+
+use crate::{
+    model::{
+        auth::{AdminAuditEvent, AuditActorKind},
+        usage_data::*,
+    },
+    ports::{provider::ProviderAdminRegistry, store::SettingsStore, usage_data::*},
+};
+use async_trait::async_trait;
+use std::sync::Arc;
+
+#[async_trait]
+pub trait UsageDataService: Send + Sync {
+    async fn list_aggregates(
+        &self,
+        request: ListUsageAggregates,
+    ) -> UsageDataResult<UsageAggregatePage>;
+    async fn write_aggregates(
+        &self,
+        actor: &UsageActor,
+        request: WriteUsageAggregates,
+    ) -> UsageDataResult<UsageAggregateWriteResult>;
+    async fn get_record(&self, request: GetUsageRecord) -> UsageDataResult<UsageRecord>;
+    async fn list_records(&self, request: ListUsageRecords) -> UsageDataResult<UsageRecordPage>;
+    async fn write_records(
+        &self,
+        actor: &UsageActor,
+        request: WriteUsageRecords,
+    ) -> UsageDataResult<UsageWriteResult>;
+    async fn evaluate_pricing(
+        &self,
+        request: EvaluateUsagePricing,
+    ) -> UsageDataResult<UsagePricingEvaluation>;
+}
+
+pub struct DefaultUsageDataService {
+    store: Arc<dyn UsageDataStore>,
+    settings: Arc<dyn SettingsStore>,
+    providers: ProviderAdminRegistry,
+}
+impl DefaultUsageDataService {
+    pub fn new(
+        store: Arc<dyn UsageDataStore>,
+        settings: Arc<dyn SettingsStore>,
+        providers: ProviderAdminRegistry,
+    ) -> Self {
+        Self {
+            store,
+            settings,
+            providers,
+        }
+    }
+}
+
+#[async_trait]
+impl UsageDataService for DefaultUsageDataService {
+    async fn list_aggregates(
+        &self,
+        request: ListUsageAggregates,
+    ) -> UsageDataResult<UsageAggregatePage> {
+        validation::query(&request.filter, request.cursor.as_deref(), request.limit)?;
+        self.store.list_aggregates(request).await
+    }
+    async fn write_aggregates(
+        &self,
+        actor: &UsageActor,
+        request: WriteUsageAggregates,
+    ) -> UsageDataResult<UsageAggregateWriteResult> {
+        validation::bounded(&request, request.changes.len())?;
+        let mut changes = Vec::with_capacity(request.changes.len());
+        for change in request.changes {
+            validation::aggregate(&change)?;
+            let identity = match &change {
+                UsageAggregateChange::Put { aggregate } => &aggregate.identity,
+                UsageAggregateChange::Delete { identity } => identity,
+            };
+            let action = match &change {
+                UsageAggregateChange::Put { .. } => "usage.aggregate.put",
+                UsageAggregateChange::Delete { .. } => "usage.aggregate.delete",
+            };
+            let audit = audit_event(
+                actor,
+                action,
+                "usage_aggregate",
+                &identity.key,
+                vec!["aggregate".into()],
+            );
+            changes.push((change, audit));
+        }
+        self.store
+            .write_aggregates(StoredAggregateWrite { changes })
+            .await
+    }
+    async fn get_record(&self, request: GetUsageRecord) -> UsageDataResult<UsageRecord> {
+        validation::text(&request.id, 128)?;
+        self.store.get_record(request).await
+    }
+    async fn list_records(&self, request: ListUsageRecords) -> UsageDataResult<UsageRecordPage> {
+        validation::query(&request.filter, request.cursor.as_deref(), request.limit)?;
+        self.store.list_records(request).await
+    }
+    async fn write_records(
+        &self,
+        actor: &UsageActor,
+        request: WriteUsageRecords,
+    ) -> UsageDataResult<UsageWriteResult> {
+        validation::bounded(&request, request.changes.len())?;
+        let mut changes = Vec::with_capacity(request.changes.len());
+        for change in request.changes {
+            validation::change(&change)?;
+            let action = match &change {
+                UsageRecordChange::Put { .. } => "usage.put",
+                UsageRecordChange::Update { .. } => "usage.update",
+                UsageRecordChange::Delete { .. } => "usage.delete",
+            };
+            let fields = match &change {
+                UsageRecordChange::Update { tokens, cost, .. } => tokens
+                    .as_ref()
+                    .map(|_| "tokens".to_owned())
+                    .into_iter()
+                    .chain(cost.as_ref().map(|_| "cost".to_owned()))
+                    .collect(),
+                _ => vec!["record".into()],
+            };
+            let audit = audit_event(actor, action, "model_request", change.id(), fields);
+            changes.push((change, audit));
+        }
+        self.store.write_records(StoredUsageWrite { changes }).await
+    }
+    async fn evaluate_pricing(
+        &self,
+        request: EvaluateUsagePricing,
+    ) -> UsageDataResult<UsagePricingEvaluation> {
+        validation::bounded(&request, request.items.len())?;
+        let stored = self
+            .settings
+            .load_pricing()
+            .await
+            .map_err(|_| UsageDataError::unavailable("读取价目失败"))?;
+        let prices = gateway_core::metering::merge_pricing(stored.synced, &stored.overrides);
+        let mut items = Vec::with_capacity(request.items.len());
+        for input in request.items {
+            validation::text(&input.id, 128)?;
+            items.push(self.evaluate_one(input, &prices)?);
+        }
+        Ok(UsagePricingEvaluation { items })
+    }
+}
+
+fn audit_event(
+    actor: &UsageActor,
+    action: &str,
+    entity_kind: &str,
+    id: &str,
+    changed_fields: Vec<String>,
+) -> AdminAuditEvent {
+    let actor_ref = match actor {
+        UsageActor::Plugin {
+            instance_id,
+            publisher,
+            name,
+            ..
+        } => format!("plugin:{publisher}/{name}:{instance_id}"),
+        UsageActor::Admin { subject } => format!("admin:{subject}"),
+        UsageActor::System => "system".into(),
+    };
+    AdminAuditEvent {
+        id: format!("audit_{}", uuid::Uuid::now_v7().simple()),
+        actor_kind: AuditActorKind::System,
+        actor_admin_user_id: None,
+        actor_ref,
+        request_id: None,
+        action: action.into(),
+        entity_kind: entity_kind.into(),
+        entity_ref: id.to_owned(),
+        config_revision: None,
+        changed_fields,
+        occurred_at: chrono::Utc::now(),
+    }
+}

backend/crates/gateway-admin/src/use_case/usage_data/pricing.rs

@@ -0,0 +1,87 @@
+//! 复用 Provider 的单请求 Token 计价规则，不写入历史或账本
+
+use super::{DefaultUsageDataService, validation};
+use crate::{
+    model::usage_data::*,
+    ports::provider::{ProviderAdminErrorKind, ProviderUsagePricingInput},
+};
+use gateway_core::{
+    metering::{PricingOverrides, Usage},
+    routing::ProviderKind,
+};
+
+impl DefaultUsageDataService {
+    pub(super) fn evaluate_one(
+        &self,
+        input: UsagePricingInput,
+        prices: &PricingOverrides,
+    ) -> UsageDataResult<UsagePricingResult> {
+        validation::text(&input.provider, 128)?;
+        validation::text(&input.model, 128)?;
+        validation::tokens(&input.tokens)?;
+        let unavailable = |reason: &str| UsagePricingResult::NotEstimable {
+            id: input.id.clone(),
+            reason: reason.into(),
+        };
+        let kind = ProviderKind::new(input.provider.clone())
+            .map_err(|_| UsageDataError::invalid("Provider 无效"))?;
+        let Ok(provider) = self.providers.require(&kind) else {
+            return Ok(unavailable("Provider 没有可用的计价规则"));
+        };
+        let tokens = &input.tokens;
+        let price = match &input.price {
+            UsagePriceSelection::Current => prices
+                .get(&input.provider)
+                .and_then(|models| models.get(&input.model)),
+            UsagePriceSelection::Override { price } => {
+                price.validate().map_err(UsageDataError::invalid)?;
+                Some(price.as_ref())
+            }
+        };
+        let request = ProviderUsagePricingInput {
+            upstream_model_id: input.model.clone(),
+            service_tier: input.service_tier.clone(),
+            usage: Usage {
+                input_tokens: tokens.input_tokens.map(UsageCount::get),
+                output_tokens: tokens.output_tokens.map(UsageCount::get),
+                cached_input_tokens: tokens.cached_input_tokens.map(UsageCount::get),
+                cache_write_input_tokens: tokens.cache_write_input_tokens.map(UsageCount::get),
+                reasoning_output_tokens: tokens.reasoning_output_tokens.map(UsageCount::get),
+                image_input_tokens: tokens.image_input_tokens.map(UsageCount::get),
+                image_output_tokens: tokens.image_output_tokens.map(UsageCount::get),
+                total_tokens: tokens.total_tokens.map(UsageCount::get),
+            },
+        };
+        let calculated = match provider.evaluate_usage_pricing(&request, price) {
+            Ok(Some(value)) => value,
+            Ok(None) => return Ok(unavailable("用量不完整、包含图像或价格不可用")),
+            Err(error)
+                if matches!(
+                    error.kind(),
+                    ProviderAdminErrorKind::Unsupported | ProviderAdminErrorKind::Invalid
+                ) =>
+            {
+                return Ok(unavailable("Provider 不支持该计价输入"));
+            }
+            Err(_) => return Err(UsageDataError::unavailable("Provider 计价失败")),
+        };
+        Ok(UsagePricingResult::Estimated {
+            id: input.id,
+            cost: UsageCost::Known {
+                source: UsageCostSource::Calculated,
+                amount: calculated.total_amount().amount().canonical(),
+                currency: calculated.total_amount().currency().as_str().into(),
+            },
+            breakdown: UsagePricingBreakdown {
+                input_amount: calculated.input_amount().amount().canonical(),
+                output_amount: calculated.output_amount().amount().canonical(),
+                cache_read_amount: calculated.cache_read_amount().amount().canonical(),
+                cache_write_amount: calculated.cache_write_amount().amount().canonical(),
+                standard_amount: calculated.standard_amount().amount().canonical(),
+                long_context_billing_applied: calculated.long_context_billing_applied(),
+                service_tier: calculated.service_tier().map(str::to_owned),
+                custom_multiplier_bps: calculated.custom_multiplier_bps(),
+            },
+        })
+    }
+}

backend/crates/gateway-admin/src/use_case/usage_data/validation.rs

@@ -0,0 +1,194 @@
+//! 用量输入沿用原生记录的数值、时间和资源边界
+
+use crate::model::usage_data::*;
+use gateway_core::metering::Decimal;
+use serde::Serialize;
+
+pub(super) fn text(value: &str, max: usize) -> UsageDataResult<()> {
+    if value.trim().is_empty() || value.len() > max || value.chars().any(char::is_control) {
+        return Err(UsageDataError::invalid("文本为空、过长或包含控制字符"));
+    }
+    Ok(())
+}
+pub(super) fn bounded(value: &impl Serialize, count: usize) -> UsageDataResult<()> {
+    if count == 0
+        || count > USAGE_DATA_MAX_TARGETS
+        || serde_json::to_vec(value)
+            .map_err(|_| UsageDataError::invalid("请求编码失败"))?
+            .len()
+            > USAGE_DATA_MAX_REQUEST_BYTES
+    {
+        return Err(UsageDataError::new(
+            UsageErrorKind::CapacityExceeded,
+            "每次操作接受 1～500 项且不超过 2 MiB",
+        ));
+    }
+    Ok(())
+}
+pub(super) fn tokens(value: &UsageTokens) -> UsageDataResult<()> {
+    if [
+        value.input_tokens,
+        value.output_tokens,
+        value.cached_input_tokens,
+        value.cache_write_input_tokens,
+        value.reasoning_output_tokens,
+        value.image_input_tokens,
+        value.image_output_tokens,
+        value.total_tokens,
+    ]
+    .into_iter()
+    .flatten()
+    .any(|v| v.get() > i64::MAX as u64)
+    {
+        return Err(UsageDataError::invalid("Token 超出原生存储范围"));
+    }
+    Ok(())
+}
+fn cost(value: &UsageCost) -> UsageDataResult<()> {
+    if let UsageCost::Known {
+        amount, currency, ..
+    } = value
+    {
+        let _ = amount
+            .parse::<Decimal>()
+            .map_err(|_| UsageDataError::invalid("费用不是有效十进制数"))?;
+        if currency.len() != 3 || !currency.bytes().all(|b| b.is_ascii_uppercase()) {
+            return Err(UsageDataError::invalid("费用或币种不符合原生存储范围"));
+        }
+    }
+    Ok(())
+}
+pub(super) fn change(value: &UsageRecordChange) -> UsageDataResult<()> {
+    text(value.id(), 128)?;
+    match value {
+        UsageRecordChange::Put { data, .. } => {
+            text(&data.client_key_ref, 128)?;
+            for v in [
+                &data.provider,
+                &data.account_ref,
+                &data.requested_model,
+                &data.model,
+                &data.service_tier,
+            ]
+            .into_iter()
+            .flatten()
+            {
+                text(v, 128)?;
+            }
+            for v in [
+                &data.protocol,
+                &data.operation,
+                &data.endpoint,
+                &data.client_transport,
+            ] {
+                text(v, 256)?;
+            }
+            if data.outcome == UsageOutcome::Running
+                || data
+                    .completed_at_ms
+                    .is_none_or(|end| end < data.started_at_ms)
+            {
+                return Err(UsageDataError::invalid("只允许写入已结束的请求明细"));
+            }
+            for at in [Some(data.started_at_ms), data.completed_at_ms]
+                .into_iter()
+                .flatten()
+            {
+                if chrono::DateTime::from_timestamp_millis(at).is_none() {
+                    return Err(UsageDataError::invalid("请求时间无效"));
+                }
+            }
+            tokens(&data.tokens)?;
+            cost(&data.cost)?;
+        }
+        UsageRecordChange::Update {
+            tokens: new_tokens,
+            cost: new_cost,
+            ..
+        } => {
+            if new_tokens.is_none() && new_cost.is_none() {
+                return Err(UsageDataError::invalid("修改必须指定 Token 或费用"));
+            }
+            if let Some(value) = new_tokens {
+                tokens(value)?;
+            }
+            if let Some(value) = new_cost {
+                cost(value)?;
+            }
+        }
+        UsageRecordChange::Delete { .. } => {}
+    }
+    Ok(())
+}
+pub(super) fn query(
+    filter: &UsageRecordFilter,
+    cursor: Option<&str>,
+    limit: Option<u16>,
+) -> UsageDataResult<()> {
+    if limit.is_some_and(|v| v == 0 || v > USAGE_DATA_MAX_PAGE_SIZE)
+        || filter.start_ms >= filter.end_ms
+        || chrono::DateTime::from_timestamp_millis(filter.start_ms).is_none()
+        || chrono::DateTime::from_timestamp_millis(filter.end_ms).is_none()
+    {
+        return Err(UsageDataError::invalid("查询时间范围或分页大小无效"));
+    }
+    for v in [
+        &filter.provider,
+        &filter.model,
+        &filter.account_ref,
+        &filter.client_key_ref,
+    ]
+    .into_iter()
+    .flatten()
+    {
+        text(v, 128)?;
+    }
+    if cursor.is_some_and(|v| v.len() > 4096) {
+        return Err(UsageDataError::invalid("游标过长"));
+    }
+    Ok(())
+}
+
+pub(super) fn aggregate(value: &UsageAggregateChange) -> UsageDataResult<()> {
+    let identity = match value {
+        UsageAggregateChange::Put { aggregate } => &aggregate.identity,
+        UsageAggregateChange::Delete { identity } => identity,
+    };
+    text(&identity.key, 128)?;
+    if chrono::DateTime::from_timestamp_millis(identity.bucket_start_ms).is_none() {
+        return Err(UsageDataError::invalid("聚合归属时间无效"));
+    }
+    let d = &identity.dimensions;
+    for value in [
+        &d.provider,
+        &d.model,
+        &d.account_ref,
+        &d.client_key_ref,
+        &d.operation,
+        &d.transport,
+    ]
+    .into_iter()
+    .flatten()
+    {
+        text(value, 128)?;
+    }
+    if let UsageAggregateChange::Put { aggregate } = value {
+        if aggregate.costs.len() > 32 {
+            return Err(UsageDataError::invalid("费用币种数量超过限制"));
+        }
+        for (currency, amount) in &aggregate.costs {
+            if currency.len() != 3
+                || !currency.bytes().all(|b| b.is_ascii_uppercase())
+                || amount.len() > 128
+            {
+                return Err(UsageDataError::invalid("聚合费用格式无效"));
+            }
+            amount
+                .strip_prefix('-')
+                .unwrap_or(amount)
+                .parse::<Decimal>()
+                .map_err(|_| UsageDataError::invalid("聚合费用不是有效十进制数"))?;
+        }
+    }
+    Ok(())
+}

backend/crates/gateway-admin/tests/model/retention.rs

@@ -4,7 +4,13 @@ use gateway_admin::model::retention::{RetentionPolicy, RetentionTarget};
 
 #[test]
 fn retention_policy_rejects_invalid_windows_and_preserves_independent_windows() {
-    for (usage, ops, audit) in [(30, 1, 1), (31, 0, 1), (31, 1, 0)] {
+    for (usage, ops, audit) in [
+        (0, 1, 1),
+        (32, 1, 1),
+        (u32::MAX, 1, 1),
+        (1, 0, 1),
+        (1, 1, 0),
+    ] {
         assert!(RetentionPolicy::try_new(usage, ops, audit).is_err());
     }
     let policy = RetentionPolicy::try_new(31, 7, 90).unwrap();
@@ -12,3 +18,13 @@ fn retention_policy_rejects_invalid_windows_and_preserves_independent_windows()
     assert_eq!(policy.days(RetentionTarget::OpsEvents), 7);
     assert_eq!(policy.days(RetentionTarget::AdminAuditEvents), 90);
 }
+
+#[test]
+fn request_retention_accepts_short_windows_without_changing_other_targets() {
+    for days in [1, 3, 7, 15, 30, 31] {
+        let policy = RetentionPolicy::try_new(days, 30, 90).unwrap();
+        assert_eq!(policy.days(RetentionTarget::ModelRequests), days);
+        assert_eq!(policy.days(RetentionTarget::OpsEvents), 30);
+        assert_eq!(policy.days(RetentionTarget::AdminAuditEvents), 90);
+    }
+}

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

@@ -317,6 +317,7 @@ impl AccountRuntimeStore for FakeRuntimeStore {
 fn group_record() -> AccountGroupRecord {
     let now = Utc::now();
     AccountGroupRecord {
+        allow_image_generation: true,
         fast_mode: FastMode::Default,
         id: group_id(),
         name: "Primary".to_owned(),
@@ -390,6 +391,7 @@ async fn invalid_group_fields_are_rejected_before_store_mutations() {
         .create(
             &context,
             CreateAccountGroup {
+                allow_image_generation: true,
                 fast_mode: FastMode::Default,
                 name: " invalid".to_owned(),
                 description: None,
@@ -403,6 +405,7 @@ async fn invalid_group_fields_are_rejected_before_store_mutations() {
         .update(
             &context,
             UpdateAccountGroup {
+                allow_image_generation: None,
                 fast_mode: None,
                 id: record.id,
                 name: "valid".to_owned(),

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

@@ -3137,6 +3137,7 @@ fn quota_local_usage(account_id: &str, total_tokens: u64) -> AccountUsage {
 pub(super) fn account_record(kind: &str) -> AccountRecord {
     let now = Utc::now();
     AccountRecord {
+        outbound_proxy_enabled: true,
         notes: None,
         model_access: Default::default(),
         outbound_proxy: None,

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

@@ -15,6 +15,7 @@ mod plugins;
 mod proxies;
 mod settings;
 mod system;
+mod usage_data;
 
 use std::{
     str::FromStr,
@@ -273,6 +274,7 @@ impl AdminHarness {
                 self.plugin_store,
                 Arc::new(plugins::TestPluginPorts),
                 Arc::new(plugins::TestPluginPorts),
+                Arc::new(plugins::TestPluginPorts),
             ),
             gateway_admin::AdminRuntimePorts {
                 timezone: Default::default(),
@@ -747,6 +749,14 @@ impl ClientKeyStore for UnavailableStore {
 
 #[async_trait]
 impl ObservabilityStore for UnavailableStore {
+    async fn diagnostic_errors(
+        &self,
+        _target: &gateway_admin::model::diagnostics::DiagnosticTarget,
+        _limit: u16,
+    ) -> AdminStoreResult<Vec<gateway_admin::model::observability::OpsError>> {
+        Err(unavailable("diagnostic errors"))
+    }
+
     fn query_timeout(&self) -> std::time::Duration {
         std::time::Duration::from_secs(10)
     }

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

@@ -945,6 +945,14 @@ impl FixtureObservabilityStore {
 
 #[async_trait]
 impl ObservabilityStore for FixtureObservabilityStore {
+    async fn diagnostic_errors(
+        &self,
+        _target: &gateway_admin::model::diagnostics::DiagnosticTarget,
+        _limit: u16,
+    ) -> AdminStoreResult<Vec<gateway_admin::model::observability::OpsError>> {
+        Err(super::unavailable("diagnostic errors"))
+    }
+
     fn query_timeout(&self) -> std::time::Duration {
         std::time::Duration::from_secs(10)
     }
@@ -1194,6 +1202,7 @@ fn observation_range(end: DateTime<Utc>) -> TimeRange {
 
 fn usage_query(now: DateTime<Utc>) -> UsageQuery {
     UsageQuery {
+        status: Default::default(),
         range: observation_range(now),
         filter: UsageFilter::default(),
         current_page: 1,
@@ -1208,6 +1217,7 @@ fn total_record(
     now: DateTime<Utc>,
 ) -> UsageListRecord {
     UsageListRecord {
+        outcome: gateway_admin::model::observability::RequestOutcome::Succeeded,
         client_api_key_name: Some("Production".to_owned()),
         id: id.to_owned(),
         endpoint: "/v1/responses".to_owned(),

backend/crates/gateway-admin/tests/use_case/plugins/distribution.rs

@@ -534,6 +534,7 @@ fn preview_fixture() -> Fixture {
 fn source_proxy() -> ProxyRecord {
     let now = chrono::Utc::now();
     ProxyRecord {
+        enabled: true,
         auto_location: false,
         detected_location: None,
         location: None,
@@ -608,6 +609,38 @@ async fn saved_source_proxy_is_resolved_for_queries_and_frozen_into_verified_sou
     );
 }
 
+#[tokio::test]
+async fn disabled_source_proxy_rejects_queries_and_downloads_before_network_access() {
+    let mut fixture = preview_fixture();
+    fixture.binding.get_mut().unwrap().outbound_proxy_id = Some("source-proxy".into());
+    let mut proxy = source_proxy();
+    proxy.enabled = false;
+    *fixture.proxy.get_mut().unwrap() = Some(proxy);
+    let fixture = Arc::new(fixture);
+    assert_eq!(
+        fixture
+            .service()
+            .check_update("test.example", &[])
+            .await
+            .unwrap_err()
+            .kind(),
+        AdminErrorKind::Invalid
+    );
+    let mut request = preview_request();
+    request.outbound_proxy_id = Some("source-proxy".into());
+    assert_eq!(
+        fixture
+            .service()
+            .verify_remote(request)
+            .await
+            .unwrap_err()
+            .kind(),
+        AdminErrorKind::Invalid
+    );
+    assert!(fixture.egresses.lock().unwrap().is_empty());
+    assert_eq!(*fixture.downloads.lock().unwrap(), 0);
+}
+
 #[tokio::test]
 async fn proxy_revision_change_during_download_rejects_the_verified_candidate() {
     let mut fixture = preview_fixture();

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

@@ -1,5 +1,7 @@
 //! 插件管理用例的测试入口与共用端口替身
 
+mod usage_data;
+
 use async_trait::async_trait;
 use gateway_admin::model::plugins::distribution::{
     DownloadedPlugin, GithubReleaseQuery, PluginRelease, RemotePluginLocation, SourceCredential,

backend/crates/gateway-admin/tests/use_case/plugins/usage_data.rs

@@ -0,0 +1,24 @@
+//! 通用用量端口的插件管理测试夹具，非本测试调用路径禁止访问
+use super::TestPluginPorts;
+use gateway_admin::{model::usage_data::*, ports::usage_data::*};
+#[async_trait::async_trait]
+impl UsageDataStore for TestPluginPorts {
+    async fn get_record(&self, _: GetUsageRecord) -> UsageDataResult<UsageRecord> {
+        unreachable!("unexpected usage call")
+    }
+    async fn list_records(&self, _: ListUsageRecords) -> UsageDataResult<UsageRecordPage> {
+        unreachable!("unexpected usage call")
+    }
+    async fn write_records(&self, _: StoredUsageWrite) -> UsageDataResult<UsageWriteResult> {
+        unreachable!("unexpected usage call")
+    }
+    async fn list_aggregates(&self, _: ListUsageAggregates) -> UsageDataResult<UsageAggregatePage> {
+        unreachable!("unexpected usage call")
+    }
+    async fn write_aggregates(
+        &self,
+        _: StoredAggregateWrite,
+    ) -> UsageDataResult<UsageAggregateWriteResult> {
+        unreachable!("unexpected usage call")
+    }
+}

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

@@ -126,6 +126,7 @@ async fn authorization_uses_selected_proxy_regardless_of_probe_status() {
                 .accounts(FakeAccountStore::new(kind, events.clone()))
                 .proxies(Arc::new(TestProxies {
                     record: Some(ProxyRecord {
+                        enabled: true,
                         auto_location: false,
                         detected_location: None,
                         location: None,

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

@@ -82,6 +82,7 @@ fn contract_type(mut item: Item, name: &str) -> Item {
                     | "RegeneratedAdminApiKey"
                     | "CodexPrivacyPolicy"
                     | "PrivacyRule"
+                    | "PrivacyStep"
                     | "PrivacyPreviewRequest"
                     | "PrivacyPreviewResult"
             );
@@ -137,6 +138,7 @@ fn sdk_settings_contract_matches_host_declarations() {
                 "PrivacyScope",
                 "PrivacyAction",
                 "PrivacyRule",
+                "PrivacyStep",
                 "PrivacyRuleOutcome",
                 "PrivacyPreviewRequest",
                 "PrivacyPreviewResult",

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

@@ -0,0 +1,119 @@
+//! 用量服务的 wire 精度、可信主体和独立 SDK 合同校验
+use gateway_admin::model::usage_data::*;
+use serde_json::json;
+#[test]
+fn counters_preserve_integer_precision_and_reject_json_numbers() {
+    assert_eq!(
+        serde_json::from_value::<UsageCount>(json!(u64::MAX.to_string()))
+            .unwrap()
+            .get(),
+        u64::MAX
+    );
+    assert_eq!(
+        serde_json::from_value::<UsageDelta>(json!(i64::MIN.to_string()))
+            .unwrap()
+            .get(),
+        i64::MIN
+    );
+    for value in [json!(1), json!("1.5"), json!("+1")] {
+        assert!(serde_json::from_value::<UsageCount>(value.clone()).is_err());
+        assert!(serde_json::from_value::<UsageDelta>(value).is_err());
+    }
+}
+#[tokio::test]
+async fn usage_mutations_require_trusted_actor_and_old_workflow_operations_are_absent() {
+    let admin = super::AdminHarness::new().build().await;
+    let registry = admin.public_services();
+    let denied = registry
+        .call("usage.records.write", json!({"changes":[]}))
+        .await
+        .unwrap_err();
+    assert!(denied.message.contains("trusted caller"));
+    assert!(
+        registry
+            .call(
+                "usage.records.write",
+                json!({"changes":[],"actor":{"kind":"system"}})
+            )
+            .await
+            .is_err()
+    );
+    let ops = registry
+        .operations()
+        .filter(|op| op.starts_with("usage.") || *op == "pricing.evaluate")
+        .collect::<std::collections::BTreeSet<_>>();
+    assert_eq!(
+        ops,
+        std::collections::BTreeSet::from([
+            "usage.records.get",
+            "usage.records.list",
+            "usage.records.write",
+            "usage.aggregates.list",
+            "usage.aggregates.write",
+            "pricing.evaluate"
+        ])
+    );
+}
+#[test]
+fn sdk_wire_types_match_host_contracts() {
+    use quote::ToTokens as _;
+    let root = std::path::Path::new(env!("CARGO_MANIFEST_DIR"));
+    let items = |path: std::path::PathBuf| {
+        syn::parse_file(&std::fs::read_to_string(path).unwrap())
+            .unwrap()
+            .items
+            .into_iter()
+            .filter(|item| {
+                matches!(
+                    item,
+                    syn::Item::Struct(_)
+                        | syn::Item::Enum(_)
+                        | syn::Item::Impl(_)
+                        | syn::Item::Type(_)
+                )
+            })
+            .map(|item| item.into_token_stream().to_string())
+            .collect::<Vec<_>>()
+    };
+    for name in [
+        "value.rs",
+        "record.rs",
+        "query.rs",
+        "aggregate.rs",
+        "pricing.rs",
+        "error.rs",
+    ] {
+        assert_eq!(
+            items(root.join("src/model/usage_data").join(name)),
+            items(
+                root.join("../gateway-plugin/sdk/src/call/services/usage")
+                    .join(name)
+            ),
+            "{name}"
+        );
+    }
+}
+
+#[tokio::test]
+async fn batch_and_wire_limits_fail_before_reaching_storage() {
+    let admin = super::AdminHarness::new().build().await;
+    let registry = admin.public_services();
+    let changes = (0..501)
+        .map(|i| json!({"kind":"delete","id":format!("item_{i}")}))
+        .collect::<Vec<_>>();
+    let too_many = gateway_admin::public_service::scope_actor(
+        UsageActor::System,
+        registry.call("usage.records.write", json!({"changes":changes})),
+    )
+    .await
+    .unwrap_err();
+    assert_eq!(too_many.kind, "capacity_exceeded");
+    let oversized = registry
+        .call(
+            "usage.records.get",
+            json!({"id":"x".repeat(USAGE_DATA_MAX_REQUEST_BYTES)}),
+        )
+        .await
+        .unwrap_err();
+    assert_eq!(oversized.kind, "capacity_exceeded");
+}

backend/crates/gateway-api/Cargo.toml

@@ -25,6 +25,7 @@ gateway-admin = { path = "../gateway-admin" }
 gateway-core = { path = "../gateway-core" }
 gateway-protocol = { path = "../gateway-protocol" }
 headers = "0.4.1"
+regex = "1.13.1"
 secrecy = "0.10"
 serde = { version = "1.0.228", features = ["derive"] }
 serde_json = { version = "1.0.150", features = ["arbitrary_precision", "preserve_order"] }
@@ -32,7 +33,7 @@ sha2 = "0.11.0"
 thiserror = "2.0.18"
 tokio = { version = "1.52.3", default-features = false, features = ["macros", "rt", "sync", "time"] }
 tower = { version = "0.5.3", features = ["util"] }
-tower-http = { version = "0.7.0", features = ["cors", "fs", "request-id", "timeout", "trace"] }
+tower-http = { version = "0.7.0", features = ["compression-gzip", "cors", "fs", "request-id", "timeout", "trace"] }
 tracing = "0.1.44"
 url = "2.5.8"
 uuid = { version = "1.23.3", features = ["v7"] }

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

@@ -91,9 +91,15 @@ impl ListAccountGroupsQuery {
     }
 }
 
+const fn default_allow_image_generation() -> bool {
+    true
+}
+
 #[derive(Debug, Clone, Deserialize)]
 #[serde(rename_all = "camelCase", deny_unknown_fields)]
 struct CreateAccountGroupRequest {
+    #[serde(default = "default_allow_image_generation")]
+    allow_image_generation: bool,
     #[serde(default)]
     fast_mode: FastMode,
     name: String,
@@ -104,6 +110,7 @@ struct CreateAccountGroupRequest {
 #[derive(Debug, Clone, Deserialize)]
 #[serde(rename_all = "camelCase", deny_unknown_fields)]
 struct UpdateAccountGroupRequest {
+    allow_image_generation: Option<bool>,
     fast_mode: Option<FastMode>,
     id: String,
     name: String,
@@ -120,6 +127,7 @@ struct AccountGroupIdRequest {
 #[derive(Debug, Clone, Serialize)]
 #[serde(rename_all = "camelCase")]
 struct AccountGroupView {
+    allow_image_generation: bool,
     fast_mode: FastMode,
     id: String,
     name: String,
@@ -167,6 +175,7 @@ impl From<(AccountGroupRecord, crate::time::TimePresenter)> for AccountGroupView
             name: record.name,
             description: record.description,
             fast_mode: record.fast_mode,
+            allow_image_generation: record.allow_image_generation,
             color: record.color.as_str().to_owned(),
             enabled: record.enabled,
             member_count: record.member_count,
@@ -362,6 +371,7 @@ where
             .create(
                 &auth.context().mutation_context(),
                 CreateAccountGroup {
+                    allow_image_generation: request.allow_image_generation,
                     name: request.name,
                     description: request.description,
                     fast_mode: request.fast_mode,
@@ -390,6 +400,7 @@ where
             .update(
                 &auth.context().mutation_context(),
                 UpdateAccountGroup {
+                    allow_image_generation: request.allow_image_generation,
                     id: group_id(request.id)?,
                     name: request.name,
                     description: request.description,

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

@@ -0,0 +1,25 @@
+//! 管理端普通 JSON 响应的压缩范围与大小门槛
+
+use http::{Response, header};
+use http_body::Body;
+use tower_http::compression::predicate::Predicate;
+
+#[derive(Clone, Copy)]
+pub(super) struct AdminJson;
+
+impl Predicate for AdminJson {
+    fn should_compress<B: Body>(&self, response: &Response<B>) -> bool {
+        // 只处理长度已知的 JSON，不为判断门槛收集流式正文
+        response
+            .body()
+            .size_hint()
+            .exact()
+            .is_some_and(|size| size >= 1024)
+            && response
+                .headers()
+                .get(header::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("application/json"))
+    }
+}

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

@@ -9,12 +9,14 @@ use axum::{
     response::Response,
     routing::any,
 };
+use tower_http::compression::{CompressionLayer, CompressionLevel};
 
 pub mod account_groups;
 pub mod accounts;
 pub mod auth;
 pub mod backups;
 pub mod client_keys;
+mod compression;
 pub(crate) mod diagnostics;
 mod extract;
 pub mod observability;
@@ -54,6 +56,12 @@ where
         .method_not_allowed_fallback(method_not_allowed)
         .route("/api/admin", any(admin_not_found))
         .route("/api/admin/{*path}", any(admin_not_found))
+        .layer(
+            CompressionLayer::new()
+                .quality(CompressionLevel::Fastest)
+                .compress_when(compression::AdminJson),
+        )
+        .merge(plugins::public_router().method_not_allowed_fallback(method_not_allowed))
         .layer(middleware::map_response(no_store))
 }
 

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

@@ -0,0 +1,136 @@
+//! 诊断文件的 HTTP 下载与受控数据投影，普通详情响应不携带时间线
+
+use axum::{extract::State, http::header, response::IntoResponse};
+use gateway_admin::model::{
+    diagnostics::{DiagnosticBundle, DiagnosticTarget},
+    observability::OpsError,
+};
+use serde::Deserialize;
+use serde_json::{Value, json};
+
+use super::{
+    AdminAuth, AdminError, AdminQuery, map_service_error, ops_error_view, usage_detail_view,
+};
+use crate::{auth::SessionState, time::TimePresenter};
+
+mod privacy;
+
+#[derive(Deserialize)]
+#[serde(rename_all = "camelCase", deny_unknown_fields)]
+pub(super) struct ExportQuery {
+    request_id: Option<String>,
+    error_id: Option<String>,
+}
+
+pub(super) async fn download<S>(
+    _auth: AdminAuth,
+    State(state): State<S>,
+    AdminQuery(query): AdminQuery<ExportQuery>,
+) -> Result<impl IntoResponse, AdminError>
+where
+    S: SessionState + Send + Sync,
+{
+    let target = match (query.request_id, query.error_id) {
+        (Some(id), None) => DiagnosticTarget::Request(id),
+        (None, Some(id)) => DiagnosticTarget::Operation(id),
+        _ => {
+            return Err(map_service_error(
+                gateway_admin::model::AdminError::invalid("请指定一个请求 ID 或错误记录 ID"),
+            ));
+        }
+    };
+    let bundle = state
+        .admin_services()
+        .observability()
+        .export_diagnostics(target)
+        .await
+        .map_err(map_service_error)?;
+    let time = TimePresenter::new(state.admin_services().timezone());
+    let document = privacy::redact(document(bundle, time));
+    let body = serde_json::to_vec_pretty(&document).map_err(|_| {
+        map_service_error(gateway_admin::model::AdminError::internal(
+            "诊断文件生成失败",
+        ))
+    })?;
+    Ok((
+        [
+            (header::CONTENT_TYPE, "application/json; charset=utf-8"),
+            (
+                header::CONTENT_DISPOSITION,
+                "attachment; filename=diagnostics.json",
+            ),
+            (header::CACHE_CONTROL, "private, no-store"),
+            (header::X_CONTENT_TYPE_OPTIONS, "nosniff"),
+        ],
+        body,
+    ))
+}
+
+fn document(bundle: DiagnosticBundle, time: TimePresenter) -> Value {
+    let target = match &bundle.target {
+        DiagnosticTarget::Request(id) => json!({"requestId": id}),
+        DiagnosticTarget::Operation(id) => json!({"errorId": id}),
+    };
+    let requests = bundle.requests.into_iter().map(|record| {
+        let evidence = record.detail.map(|mut detail| {
+            let trace = detail.trace.take();
+            let related = std::mem::take(&mut detail.related_requests);
+            let configuration = json!({
+                "revision": detail.request.config_revision,
+                "routingScope": detail.request.routing_scope,
+                "routingGroupIds": detail.request.routing_group_refs,
+                "deadlineAt": detail.request.deadline_at,
+            });
+            // 页面 DTO 没有展示的尝试原因与等待事实仍属于导出证据
+            let attempt_details = detail.attempts.iter().map(|attempt| json!({
+                "id": attempt.id, "attemptIndex": attempt.attempt_index,
+                "component": attempt.component, "operation": attempt.operation,
+                "message": attempt.message, "retryAfterMs": attempt.retry_after_ms,
+            })).collect::<Vec<_>>();
+            json!({
+                "detail": usage_detail_view(detail, time), "trace": trace,
+                "relatedRequests": related, "incidentConfiguration": configuration,
+                "attemptDetails": attempt_details,
+            })
+        });
+        json!({
+            "requestId": record.request_id,
+            "evidence": evidence,
+            "errors": record.errors.into_iter().map(|error| export_error(error, time)).collect::<Vec<_>>(),
+        })
+    }).collect::<Vec<_>>();
+    json!({
+        "schemaVersion": 5,
+        "source": "gateway.diagnostics",
+        "exportPolicy": "support_credentials_redacted",
+        "target": target,
+        "collectedAt": bundle.collected_at,
+        "environment": {
+            "scope": "export_time", "observedAt": bundle.collected_at,
+            "gateway": bundle.environment.map(|version| json!({
+                "version": version.version, "gitSha": version.git_sha,
+                "buildTime": version.build_time, "deploymentMode": version.deployment_mode,
+            })),
+        },
+        "error": bundle.error.map(|error| export_error(error, time)),
+        "requests": requests,
+        "availability": {"gaps": bundle.gaps.into_iter().map(|gap| json!({"section": gap.section, "reason": gap.reason})).collect::<Vec<_>>()},
+        "omitted": ["credentials_and_authentication", "account_contact_details", "uncollected_wire_payloads", "unrelated_requests_and_system_secrets"],
+        "notes": [
+            "诊断包包含错误原因链、上游错误与业务上下文，用于故障分析",
+            "认证字段、账号联系信息与可识别的内嵌凭据已脱敏，任意业务内容不属于自动识别范围",
+            "environment 为导出时环境，incidentConfiguration 与 request.routed 为事发配置事实，当前版本不能代替事发版本",
+            "缺失、截断、淘汰与读取失败见 availability.gaps，没有记录不表示没有发生",
+            "attempts 为异步 best-effort 投影，firstTokenMs 为空不证明没有输出",
+            "未采集的请求正文、过期记录、上游内部状态与退出前未落盘信息无法由导出补回"
+        ],
+    })
+}
+
+// 复用错误字段投影，原始详情直接进入脱敏遍历，避免先为页面缩进再解析一遍
+fn export_error(mut error: OpsError, time: TimePresenter) -> Value {
+    let details = error.error_details.take();
+    let mut value = json!(ops_error_view(error, time));
+    value["errorDetails"] = json!(details);
+    value
+}

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

@@ -0,0 +1,220 @@
+//! 诊断文件按认证与观测合同移除敏感字段，保留错误事实并明确导出截断
+
+use regex::Regex;
+use serde::Serialize;
+use serde_json::{Value, json};
+use std::sync::LazyLock;
+
+const REDACTED: &str = "<redacted>";
+const MAX_TEXT: usize = 256 * 1024;
+static INLINE: LazyLock<Vec<(Regex, &str)>> = LazyLock::new(|| {
+    [
+        (r"(?i)\b((?:set-cookie|cookie2?|(?:proxy-)?authorization|x-api-key|x-oai-is(?:-update)?|x-oai-attestation)\s*:\s*)[^\r\n]+", "${1}<redacted>"),
+        (r#"(?i)\b(Bearer|Basic)\s+[^\s"'<>;,]+"#, "${1} <redacted>"),
+        // 官方 TokenData 的 access_token / id_token 为 JWT，不按猜测的令牌前缀过滤普通文本
+        (r"\beyJ[\w-]+\.[\w-]+\.[\w-]+\b", "<redacted>"),
+        (r"(?i)([a-z][a-z0-9+.-]*://)[^\s/@]+@", "${1}<redacted>@"),
+        (r#"(?i)\b((?:(?:personal[_-]?)?access[_-]?token|refresh[_-]?token|id[_-]?token|(?:(?:openai|x)[_-]?)?api[_-]?key|(?:current[_-]?|new[_-]?)?password|authorization|proxy[_-]?authorization|x[_-]?oai[_-]?(?:is(?:[_-]?update)?|attestation))["']?\s*[=:]\s*)("[^"\r\n]*"|'[^'\r\n]*'|[^\s&,;]+)"#, "${1}<redacted>"),
+    ].into_iter().map(|(pattern, replacement)| (Regex::new(pattern).expect("static credential pattern"), replacement)).collect()
+});
+
+#[derive(Default, Serialize)]
+#[serde(rename_all = "camelCase")]
+struct Report {
+    credential_fields: usize,
+    personal_fields: usize,
+    inline_credentials: usize,
+    truncated_values: usize,
+}
+
+struct Redactor {
+    report: Report,
+    bytes: usize,
+    nodes: usize,
+}
+
+pub(super) fn redact(mut value: Value) -> Value {
+    let mut redactor = Redactor {
+        report: Report::default(),
+        bytes: 8 * 1024 * 1024,
+        nodes: 100_000,
+    };
+    // 先处理有界文件信息，正文耗尽预算也不能删除 schema、目标或既有缺口
+    let mut error = value["error"].take();
+    let mut requests = value["requests"].take();
+    redactor.visit(&mut value, 0);
+    redactor.visit(&mut error, 0);
+    redactor.visit(&mut requests, 0);
+    value["error"] = error;
+    value["requests"] = requests;
+    if redactor.report.truncated_values > 0 {
+        // 缺口清单可能也触及预算，顶层统计始终保留以免截断变成静默丢失
+        value["exportTruncated"] = json!(true);
+        if !value["availability"]["gaps"].is_array() {
+            value["availability"] = json!({"gaps": []});
+        }
+        value["availability"]["gaps"]
+            .as_array_mut()
+            .expect("gap array")
+            .push(json!({
+                "section": "export", "reason": "export_size_limit",
+            }));
+    }
+    value["redaction"] = json!(redactor.report);
+    value
+}
+
+fn normalized(key: &str) -> String {
+    key.chars()
+        .filter(|character| !matches!(character, '_' | '-'))
+        .flat_map(char::to_lowercase)
+        .collect()
+}
+
+// OAuth 字段来自 Provider credential schema 与官方 Codex TokenData；API Key 来自
+// CodexCprApiKeyExportAccount / AuthDotJson；密码来自管理登录合同，头部来自传输身份合同
+// 不用 token、secret、credential 等泛称推断敏感性，避免删除业务错误上下文
+fn secret(key: &str) -> bool {
+    matches!(
+        key,
+        "accesstoken"
+            | "personalaccesstoken"
+            | "refreshtoken"
+            | "idtoken"
+            | "apikey"
+            | "openaiapikey"
+            | "password"
+            | "currentpassword"
+            | "newpassword"
+            | "authorization"
+            | "proxyauthorization"
+            | "xapikey"
+            | "cookie"
+            | "cookie2"
+            | "setcookie"
+            | "xoaiis"
+            | "xoaiisupdate"
+            | "xoaiattestation"
+    )
+}
+
+impl Redactor {
+    fn omitted(&mut self) -> Value {
+        self.report.truncated_values += 1;
+        json!({"omitted": "export_size_limit"})
+    }
+
+    fn text(&mut self, text: &str) -> String {
+        let mut result = text.to_owned();
+        for (pattern, replacement) in INLINE.iter() {
+            self.report.inline_credentials += pattern.find_iter(&result).count();
+            result = pattern.replace_all(&result, *replacement).into_owned();
+        }
+        result
+    }
+
+    fn visit(&mut self, value: &mut Value, depth: usize) {
+        if self.nodes == 0 || self.bytes == 0 || depth > 32 {
+            *value = self.omitted();
+            return;
+        }
+        self.nodes -= 1;
+        // 预算同时计入缩进和容器分隔符，避免深层 JSON 的下载体积显著超过正文预算
+        self.bytes = self.bytes.saturating_sub(24 + depth * 4);
+        match value {
+            Value::String(text) => {
+                if text.trim_start().starts_with('{') || text.trim_start().starts_with('[') {
+                    if text.len() > MAX_TEXT {
+                        *value = self.omitted();
+                        return;
+                    }
+                    if let Ok(parsed) = serde_json::from_str::<Value>(text) {
+                        *value = parsed;
+                        self.visit(value, depth + 1);
+                        return;
+                    }
+                }
+                let cleaned = self.text(text);
+                let limit = MAX_TEXT.min(self.bytes / 2);
+                if cleaned.len() > limit {
+                    let boundary = cleaned.floor_char_boundary(limit);
+                    *value = json!({"text": &cleaned[..boundary], "omittedBytes": cleaned.len() - boundary});
+                    self.report.truncated_values += 1;
+                } else {
+                    *text = cleaned;
+                }
+                self.bytes = self.bytes.saturating_sub(value.to_string().len());
+            }
+            Value::Array(items) => {
+                // 头部也可能以二元组承载，不能只识别对象字段名
+                if items.len() == 2
+                    && items[0]
+                        .as_str()
+                        .is_some_and(|name| secret(&normalized(name)))
+                {
+                    items[1] = json!(REDACTED);
+                    self.report.credential_fields += 1;
+                }
+                let original_len = items.len();
+                let mut kept = 0;
+                for item in items.iter_mut().take(4096) {
+                    if self.nodes == 0 || self.bytes == 0 {
+                        break;
+                    }
+                    self.visit(item, depth + 1);
+                    kept += 1;
+                }
+                if kept < original_len {
+                    items.truncate(kept);
+                    items.push(json!({"omittedItems": original_len - kept}));
+                    self.report.truncated_values += 1;
+                }
+            }
+            Value::Object(object) => {
+                if object
+                    .get("name")
+                    .and_then(Value::as_str)
+                    .is_some_and(|name| secret(&normalized(name)))
+                    && object.contains_key("value")
+                {
+                    object.insert("value".to_owned(), json!(REDACTED));
+                    self.report.credential_fields += 1;
+                }
+                let fields = std::mem::take(object);
+                let total = fields.len();
+                for (key, mut item) in fields {
+                    if object.len() >= 4096 || self.nodes == 0 || self.bytes == 0 {
+                        break;
+                    }
+                    let normalized = normalized(&key);
+                    if secret(&normalized) {
+                        item = json!(REDACTED);
+                        self.report.credential_fields += 1;
+                    } else if matches!(
+                        normalized.as_str(),
+                        "email"
+                            | "accountemail"
+                            | "clientip"
+                            | "accountname"
+                            | "accountnotes"
+                            | "credentialname"
+                            | "clientapikeyname"
+                            | "accountlabel"
+                    ) {
+                        item = json!(REDACTED);
+                        self.report.personal_fields += 1;
+                    } else {
+                        self.visit(&mut item, depth + 1);
+                    }
+                    self.bytes = self.bytes.saturating_sub(key.len());
+                    object.insert(self.text(&key), item);
+                }
+                if object.len() < total {
+                    object.insert("_omittedFields".to_owned(), json!(total - object.len()));
+                    self.report.truncated_values += 1;
+                }
+            }
+            _ => {}
+        }
+    }
+}

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

@@ -16,6 +16,7 @@ use super::{
     wire::map_admin_service_error,
 };
 
+mod export;
 mod presenter;
 mod query;
 mod routes;

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

@@ -1,6 +1,7 @@
 //! 观测领域事实到管理页面的展示投影
 
 use super::*;
+use gateway_core::metering::Usage;
 
 pub(crate) fn display_duration(value: Option<u64>) -> String {
     let Some(value) = value.and_then(|value| i64::try_from(value).ok()) else {
@@ -100,7 +101,7 @@ pub(crate) fn attempt_metrics_view(metrics: &domain::AttemptMetrics) -> AttemptM
 }
 
 pub(crate) fn token_details(record: &domain::UsageRecord) -> TokenDetailsView {
-    TokenDetailsView {
+    token_details_view(&Usage {
         input_tokens: record.input_tokens,
         output_tokens: record.output_tokens,
         cached_input_tokens: record.cached_input_tokens,
@@ -109,35 +110,11 @@ pub(crate) fn token_details(record: &domain::UsageRecord) -> TokenDetailsView {
         image_input_tokens: record.image_input_tokens,
         image_output_tokens: record.image_output_tokens,
         total_tokens: record.total_tokens,
-        input_tokens_display: record
-            .input_tokens
-            .map_or_else(|| "-".to_owned(), format_number),
-        output_tokens_display: record
-            .output_tokens
-            .map_or_else(|| "-".to_owned(), format_number),
-        cached_input_tokens_display: record
-            .cached_input_tokens
-            .map_or_else(|| "-".to_owned(), format_compact_number),
-        cache_write_input_tokens_display: record
-            .cache_write_input_tokens
-            .map_or_else(|| "-".to_owned(), format_compact_number),
-        reasoning_output_tokens_display: record
-            .reasoning_output_tokens
-            .map_or_else(|| "-".to_owned(), format_number),
-        image_input_tokens_display: record
-            .image_input_tokens
-            .map_or_else(|| "-".to_owned(), format_number),
-        image_output_tokens_display: record
-            .image_output_tokens
-            .map_or_else(|| "-".to_owned(), format_number),
-        total_tokens_display: record
-            .total_tokens
-            .map_or_else(|| "-".to_owned(), format_number),
-    }
+    })
 }
 
 pub(crate) fn usage_list_token_details(record: &domain::UsageListRecord) -> TokenDetailsView {
-    TokenDetailsView {
+    token_details_view(&Usage {
         input_tokens: record.input_tokens,
         output_tokens: record.output_tokens,
         cached_input_tokens: record.cached_input_tokens,
@@ -146,28 +123,41 @@ pub(crate) fn usage_list_token_details(record: &domain::UsageListRecord) -> Toke
         image_input_tokens: record.image_input_tokens,
         image_output_tokens: record.image_output_tokens,
         total_tokens: record.total_tokens,
-        input_tokens_display: record
+    })
+}
+
+fn token_details_view(usage: &Usage) -> TokenDetailsView {
+    TokenDetailsView {
+        input_tokens: usage.input_tokens,
+        output_tokens: usage.output_tokens,
+        cached_input_tokens: usage.cached_input_tokens,
+        cache_write_input_tokens: usage.cache_write_input_tokens,
+        reasoning_output_tokens: usage.reasoning_output_tokens,
+        image_input_tokens: usage.image_input_tokens,
+        image_output_tokens: usage.image_output_tokens,
+        total_tokens: usage.total_tokens,
+        input_tokens_display: usage
             .input_tokens
             .map_or_else(|| "-".to_owned(), format_number),
-        output_tokens_display: record
+        output_tokens_display: usage
             .output_tokens
             .map_or_else(|| "-".to_owned(), format_number),
-        cached_input_tokens_display: record
+        cached_input_tokens_display: usage
             .cached_input_tokens
             .map_or_else(|| "-".to_owned(), format_compact_number),
-        cache_write_input_tokens_display: record
+        cache_write_input_tokens_display: usage
             .cache_write_input_tokens
             .map_or_else(|| "-".to_owned(), format_compact_number),
-        reasoning_output_tokens_display: record
+        reasoning_output_tokens_display: usage
             .reasoning_output_tokens
             .map_or_else(|| "-".to_owned(), format_number),
-        image_input_tokens_display: record
+        image_input_tokens_display: usage
             .image_input_tokens
             .map_or_else(|| "-".to_owned(), format_number),
-        image_output_tokens_display: record
+        image_output_tokens_display: usage
             .image_output_tokens
             .map_or_else(|| "-".to_owned(), format_number),
-        total_tokens_display: record
+        total_tokens_display: usage
             .total_tokens
             .map_or_else(|| "-".to_owned(), format_number),
     }
@@ -272,6 +262,7 @@ pub(crate) fn usage_list_record_view(
         .clone()
         .or_else(|| record.requested_model_id.clone());
     UsageListRecordView {
+        logical_outcome: record.outcome.as_str().to_owned(),
         client_api_key_name: record.client_api_key_name,
         id: record.id,
         provider: record.provider_kind,
@@ -545,8 +536,6 @@ pub(crate) fn usage_detail_view(
     time: crate::time::TimePresenter,
 ) -> UsageRecordDetailView {
     UsageRecordDetailView {
-        trace: detail.trace,
-        related_requests: detail.related_requests,
         request: usage_record_view(detail.request, time),
         attempts: detail
             .attempts
@@ -626,7 +615,7 @@ pub(crate) fn ops_error_view(
         subagent_kind: error.subagent_kind,
         compact: error.compact,
         message: error.message,
-        error_details: error.error_details,
+        error_details: error.error_details.map(error_details_view),
         metadata: OpsErrorMetadataView {
             source: error.source,
             component: error.component,
@@ -650,6 +639,20 @@ pub(crate) fn ops_error_view(
     }
 }
 
+// 受控错误正文保持原文语义，嵌套 JSON 只在展示投影中展开缩进
+fn error_details_view(raw: String) -> String {
+    let Ok(mut value) = serde_json::from_str::<Value>(&raw) else {
+        return raw;
+    };
+    if let Some(upstream) = value.get_mut("upstream")
+        && let Some(text) = upstream.as_str()
+        && let Ok(parsed) = serde_json::from_str::<Value>(text)
+    {
+        *upstream = parsed;
+    }
+    serde_json::to_string_pretty(&value).unwrap_or(raw)
+}
+
 pub(crate) fn ops_page_view(
     page: domain::OpsErrorPage,
     time: crate::time::TimePresenter,

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

@@ -29,6 +29,7 @@ impl DashboardQuery {
 #[derive(Clone, Default, Deserialize)]
 #[serde(rename_all = "camelCase", deny_unknown_fields)]
 pub struct UsageQuery {
+    pub status: Option<String>,
     pub current_page: Option<u32>,
     pub page_size: Option<u16>,
     pub kind: Option<String>,
@@ -330,6 +331,12 @@ pub(crate) fn usage_command(
     let page_size_value =
         DomainPageSize::new(page_size).map_err(|_| WireValidationError::new("pageSize"))?;
     Ok(domain::UsageQuery {
+        status: match query.status.as_deref() {
+            None | Some("success") => domain::UsageRecordStatus::Success,
+            Some("error") => domain::UsageRecordStatus::Error,
+            Some("all") => domain::UsageRecordStatus::All,
+            Some(_) => return Err(WireValidationError::new("status")),
+        },
         range: usage_range(
             query.start_time.as_deref(),
             query.end_time.as_deref(),

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

@@ -29,6 +29,10 @@ where
             get(usage_insights_diagnostics::<S>),
         )
         .route("/api/admin/operations/errors", get(ops_errors::<S>))
+        .route(
+            "/api/admin/diagnostics/export",
+            get(super::export::download::<S>),
+        )
 }
 
 pub(crate) async fn dashboard_summary<S>(

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

@@ -86,6 +86,7 @@ pub struct ImageBillingView {
 #[derive(Serialize)]
 #[serde(rename_all = "camelCase")]
 pub struct UsageListRecordView {
+    pub logical_outcome: String,
     pub client_api_key_name: Option<String>,
     pub id: String,
     pub provider: Option<String>,
@@ -275,8 +276,6 @@ pub struct UsageAttemptView {
 #[derive(Serialize)]
 #[serde(rename_all = "camelCase")]
 pub struct UsageRecordDetailView {
-    pub trace: Option<serde_json::Value>,
-    pub related_requests: Vec<serde_json::Value>,
     #[serde(flatten)]
     pub request: UsageRecordView,
     pub attempts: Vec<UsageAttemptView>,

backend/crates/gateway-api/src/admin/plugins/management.rs

@@ -53,8 +53,19 @@ pub(super) fn router<S: SessionState + Clone + Send + Sync + 'static>() -> Route
         .route("/api/admin/plugins/extensions/{instance_id}/{artifact_sha256}/{revision}/callback-tickets", post(start_callback::<S>))
         .route("/api/admin/plugins/extensions/{instance_id}/{artifact_sha256}/{revision}/api/{*path}", any(handle::<S>))
         .route("/api/admin/plugins/extensions/{instance_id}/{artifact_sha256}/{revision}/resources/{*path}", get(resource::<S>))
-        .route("/plugins/resources/{instance_id}/{artifact_sha256}/{revision}/{*path}", get(public_resource::<S>))
-        .route("/plugins/callbacks/{instance_id}/{artifact_sha256}/{revision}/{*path}", get(callback::<S>))
+}
+
+pub(in crate::admin) fn public_router<S: SessionState + Clone + Send + Sync + 'static>() -> Router<S>
+{
+    Router::new()
+        .route(
+            "/plugins/resources/{instance_id}/{artifact_sha256}/{revision}/{*path}",
+            get(public_resource::<S>),
+        )
+        .route(
+            "/plugins/callbacks/{instance_id}/{artifact_sha256}/{revision}/{*path}",
+            get(callback::<S>),
+        )
 }
 
 pub(in crate::admin) fn model_router() -> Router<crate::ApiState> {

backend/crates/gateway-api/src/admin/plugins/mod.rs

@@ -5,7 +5,7 @@ mod distribution;
 mod instances;
 mod management;
 
-pub(super) use management::model_router;
+pub(super) use management::{model_router, public_router};
 
 #[derive(serde::Deserialize)]
 #[serde(deny_unknown_fields)]

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

@@ -60,6 +60,7 @@ struct CreateRequest {
 #[derive(Debug, Deserialize)]
 #[serde(rename_all = "camelCase", deny_unknown_fields)]
 struct UpdateRequest {
+    enabled: Option<bool>,
     auto_location: Option<bool>,
     #[serde(default, deserialize_with = "deserialize_location_update")]
     location: Option<Option<gateway_core::account::RequestLocation>>,
@@ -145,6 +146,7 @@ struct ProxyAccountView {
 #[derive(Serialize)]
 #[serde(rename_all = "camelCase")]
 struct ProxyView {
+    enabled: bool,
     auto_location: bool,
     detected_location: Option<gateway_admin::model::proxies::DetectedProxyLocation>,
     location: Option<gateway_core::account::RequestLocation>,
@@ -168,6 +170,7 @@ impl From<(ProxyRecord, crate::time::TimePresenter)> for ProxyView {
         let endpoint = record.proxy.endpoint();
         Self {
             auto_location: record.auto_location,
+            enabled: record.enabled,
             detected_location: record.detected_location,
             location: record.location,
             id: record.id,
@@ -419,6 +422,7 @@ where
         .proxies()
         .update(
             UpdateProxy {
+                enabled: request.enabled,
                 auto_location: request.auto_location,
                 test: None,
                 location: request.location,

backend/crates/gateway-api/src/middleware/mod.rs

@@ -17,9 +17,10 @@ use axum::{
     http::{HeaderName, HeaderValue, Request, StatusCode},
     response::{IntoResponse as _, Response},
 };
-use gateway_core::engine::middleware::FrozenMiddlewarePlan;
 use gateway_core::engine::middleware::http as contract;
 use gateway_core::engine::middleware::http::Settings;
+use gateway_core::engine::middleware::{FrozenMiddlewarePlan, MiddlewareError};
+use gateway_core::error::GatewayErrorKind;
 use gateway_core::lifecycle::CancellationToken;
 use gateway_core::middleware::compose;
 use http_body::{Body as HttpBody, Frame, SizeHint};
@@ -270,6 +271,11 @@ async fn handle(
             }
             response.map(|body| Body::new(ResponseBody::new(body, guard)))
         }
+        Err(MiddlewareError::Gateway(error))
+            if error.kind() == GatewayErrorKind::MiddlewareCapacityUnavailable =>
+        {
+            crate::openai::error::gateway_error_response(&error)
+        }
         Err(error) => {
             let status = if error.is_rejected() {
                 StatusCode::FORBIDDEN

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

@@ -13,6 +13,8 @@ use gateway_core::{
 };
 use serde_json::{Value, json};
 
+pub use gateway_core::engine::execution::gateway_error_from_engine;
+
 use super::responses::ProtocolErrorBody;
 
 /// OpenAI 风格错误响应
@@ -148,39 +150,6 @@ pub fn model_not_found_response() -> (StatusCode, Json<Value>) {
     )
 }
 
-/// 将 core Engine 错误收敛为稳定 Gateway 分类
-#[must_use]
-pub fn gateway_error_from_engine(error: &EngineError) -> GatewayError {
-    match error {
-        EngineError::Provider(error) => GatewayError::from_provider(error),
-        EngineError::Cancelled => {
-            GatewayError::new(GatewayErrorKind::Cancelled, "request was cancelled")
-        }
-        EngineError::Deadline => {
-            GatewayError::new(GatewayErrorKind::Timeout, "gateway request timed out")
-        }
-        EngineError::ProviderNotRegistered { .. } | EngineError::EmptyRoutingPlan => {
-            GatewayError::new(
-                GatewayErrorKind::NoAvailableProvider,
-                "no upstream provider is currently available for this request",
-            )
-        }
-        EngineError::Store(source) => {
-            GatewayError::new(GatewayErrorKind::Internal, "gateway execution failed")
-                .with_source(source.clone())
-        }
-        EngineError::ProviderMetadataMismatch
-        | EngineError::ContinuationPinMismatch
-        | EngineError::RequiredAccountMismatch
-        | EngineError::AccountOutsideClientScope
-        | EngineError::DownstreamCommitRequired
-        | EngineError::NoActiveAttempt
-        | EngineError::InvalidDeliveryState => {
-            GatewayError::new(GatewayErrorKind::Internal, "gateway execution failed")
-        }
-    }
-}
-
 /// Gateway 错误的 OpenAI HTTP 表达
 pub fn gateway_error_response(error: &GatewayError) -> Response {
     let (status, default_type, default_code) = gateway_error_contract(error.kind());
@@ -310,6 +279,11 @@ pub const fn gateway_error_contract(
             "server_error",
             "provider_infrastructure_unavailable",
         ),
+        GatewayErrorKind::MiddlewareCapacityUnavailable => (
+            StatusCode::SERVICE_UNAVAILABLE,
+            "server_error",
+            "middleware_capacity_unavailable",
+        ),
         GatewayErrorKind::ConcurrencyQueueFull => (
             StatusCode::TOO_MANY_REQUESTS,
             "rate_limit_error",

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

@@ -5,7 +5,9 @@ use std::{borrow::Cow, fmt, net::IpAddr};
 use axum::http::HeaderMap;
 use base64::{Engine as _, engine::general_purpose::STANDARD};
 use bytes::Bytes;
-use gateway_core::operation::{GenerateRequest, Operation, ProtocolPayload, ProviderSessionState};
+use gateway_core::operation::{
+    ClientTurn, GenerateRequest, Operation, ProtocolPayload, ProviderSessionState,
+};
 use gateway_protocol::openai::{
     X_OPENAI_INTERNAL_CODEX_RESPONSES_LITE_HEADER, X_OPENAI_MEMGEN_REQUEST_HEADER,
     is_transport_managed_request_header,
@@ -497,12 +499,30 @@ pub(super) fn decode_request_object(
     };
     let protocol_context = request_headers.protocol_context(use_websocket, frame_turn_metadata);
 
+    let semantics =
+        gateway_protocol::openai::codex_responses_request_semantics(&object, &protocol_context);
+    let client_turn = if object.get("generate").and_then(Value::as_bool) != Some(false)
+        && semantics.request_kind.as_deref() == Some("turn")
+        && !semantics.compact
+    {
+        gateway_protocol::openai::codex_account_thread_id(&object, &protocol_context)
+            .zip(gateway_protocol::openai::codex_turn_id(
+                &object,
+                &protocol_context,
+            ))
+            .and_then(|(thread, turn)| ClientTurn::new(thread, turn))
+    } else {
+        None
+    };
+
     let payload = ProtocolPayload::json_object(OPENAI_PROTOCOL, object)
         .map_err(|_| RequestDecodeError::CanonicalContract {
             field: "request".to_owned(),
         })?
         .with_context(protocol_context);
-    let operation = Operation::Generate(GenerateRequest::from_protocol_payload(payload));
+    let operation = Operation::Generate(
+        GenerateRequest::from_protocol_payload(payload).with_client_turn(client_turn),
+    );
     Ok(DecodedResponsesRequest {
         operation,
         metadata: ResponsesRequestMetadata {

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

@@ -106,6 +106,7 @@ async fn create_route_should_create_an_empty_group() {
     assert_eq!(value["data"]["record"]["memberCount"], 0);
     assert_eq!(value["data"]["record"]["color"], "#A855F780");
     assert_eq!(value["data"]["record"]["fastMode"], "default");
+    assert_eq!(value["data"]["record"]["allowImageGeneration"], true);
     assert_eq!(value["data"]["configRevision"], 8);
 }
 
@@ -393,3 +394,40 @@ async fn fast_mode_rejects_invalid_values_and_creates_groups_in_each_mode() {
         }
     }
 }
+
+#[tokio::test]
+async fn image_generation_flag_round_trips_and_omitted_updates_preserve_it() {
+    let fixture = authenticated_fixture().await;
+    let created = request(router(&fixture), Method::POST, "/api/admin/account-groups/create", Some(json!({
+        "name":"No images", "description":null, "color":"#2563EBFF", "allowImageGeneration":false
+    })), true).await;
+    assert_eq!(created.status(), StatusCode::CREATED);
+    let created = response_json(created).await;
+    assert_eq!(created["data"]["record"]["allowImageGeneration"], false);
+    let id = created["data"]["id"].as_str().unwrap();
+    for (value, expected) in [
+        (None, false),
+        (Some(true), true),
+        (Some(false), false),
+        (None, false),
+    ] {
+        let mut body =
+            json!({"id":id,"name":"Image policy","description":null,"color":"#2563EBFF"});
+        if let Some(value) = value {
+            body["allowImageGeneration"] = json!(value);
+        }
+        let response = request(
+            router(&fixture),
+            Method::POST,
+            "/api/admin/account-groups/update",
+            Some(body),
+            true,
+        )
+        .await;
+        assert_eq!(response.status(), StatusCode::OK);
+        assert_eq!(
+            response_json(response).await["data"]["record"]["allowImageGeneration"],
+            expected
+        );
+    }
+}

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

@@ -112,6 +112,7 @@ fn account_page_item(
 
     AccountPageItem {
         account: AccountRecord {
+            outbound_proxy_enabled: true,
             id: "acct_cooldown".to_owned(),
             provider_kind: gateway_core::routing::ProviderKind::new("openai").unwrap(),
             groups: Vec::new(),

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

@@ -0,0 +1,111 @@
+//! 管理端 JSON 压缩的协商、响应合同与重复编码回归测试
+
+use std::io::Read as _;
+
+use axum::{
+    Router,
+    body::{Body, to_bytes},
+    http::{Request, StatusCode, header},
+    response::Response,
+};
+use flate2::read::GzDecoder;
+use tower::ServiceExt as _;
+use tower_http::compression::CompressionLayer;
+
+use super::AdminTestFixture;
+
+async fn app() -> Router {
+    let fixture = AdminTestFixture::new().await;
+    fixture.auth.insert_session("valid-session");
+    crate::openai::api_router_with_admin(fixture.services)
+}
+
+async fn get(router: Router, uri: &str, encoding: Option<&str>) -> Response {
+    let mut request = Request::builder()
+        .uri(uri)
+        .header(header::COOKIE, "cpr_session=valid-session")
+        .header("x-request-id", "req_admin_compression");
+    if let Some(encoding) = encoding {
+        request = request.header(header::ACCEPT_ENCODING, encoding);
+    }
+    router
+        .oneshot(request.body(Body::empty()).expect("request"))
+        .await
+        .expect("response")
+}
+
+async fn bytes(response: Response) -> Vec<u8> {
+    to_bytes(response.into_body(), 1024 * 1024)
+        .await
+        .expect("response body")
+        .to_vec()
+}
+
+fn decode_gzip(body: &[u8]) -> Vec<u8> {
+    let mut decoded = Vec::new();
+    GzDecoder::new(body)
+        .read_to_end(&mut decoded)
+        .expect("valid gzip body");
+    decoded
+}
+
+#[tokio::test]
+async fn gzip_should_preserve_admin_json_and_response_headers() {
+    let router = app().await;
+    let expected = bytes(get(router.clone(), "/api/admin/settings", None).await).await;
+    assert!(expected.len() >= 1024);
+
+    for encoding in ["gzip", "br, gzip;q=0.8, deflate;q=0.5", "*;q=0.5"] {
+        let response = get(router.clone(), "/api/admin/settings", Some(encoding)).await;
+        assert_eq!(response.status(), StatusCode::OK);
+        assert_eq!(response.headers()[header::CONTENT_ENCODING], "gzip");
+        assert_eq!(response.headers()[header::CONTENT_TYPE], "application/json");
+        assert_eq!(response.headers()[header::CACHE_CONTROL], "no-store");
+        assert_eq!(response.headers()["x-request-id"], "req_admin_compression");
+        assert_eq!(response.headers()[header::VARY], "accept-encoding");
+        assert!(!response.headers().contains_key(header::CONTENT_LENGTH));
+        let compressed = bytes(response).await;
+        assert!(compressed.len() < expected.len());
+        assert_eq!(decode_gzip(&compressed), expected, "{encoding}");
+    }
+}
+
+#[tokio::test]
+async fn clients_without_gzip_support_should_receive_original_json() {
+    let router = app().await;
+    let expected = bytes(get(router.clone(), "/api/admin/settings", None).await).await;
+    for encoding in [None, Some("identity"), Some("br"), Some("gzip;q=0")] {
+        let response = get(router.clone(), "/api/admin/settings", encoding).await;
+        assert_eq!(response.status(), StatusCode::OK);
+        assert!(!response.headers().contains_key(header::CONTENT_ENCODING));
+        assert_eq!(response.headers()[header::VARY], "accept-encoding");
+        assert_eq!(bytes(response).await, expected, "{encoding:?}");
+    }
+}
+
+#[tokio::test]
+async fn small_admin_json_should_remain_uncompressed() {
+    let router = app().await;
+    let uri = "/api/admin/settings/admin-api-key";
+    let expected = bytes(get(router.clone(), uri, None).await).await;
+    assert!(expected.len() < 1024);
+    let response = get(router, uri, Some("gzip")).await;
+    assert_eq!(response.status(), StatusCode::OK);
+    assert!(!response.headers().contains_key(header::CONTENT_ENCODING));
+    assert_eq!(bytes(response).await, expected);
+}
+
+#[tokio::test]
+async fn a_second_compression_layer_should_not_encode_gzip_again() {
+    let router = app().await;
+    let expected = bytes(get(router.clone(), "/api/admin/settings", None).await).await;
+    let response = get(
+        router.layer(CompressionLayer::new()),
+        "/api/admin/settings",
+        Some("gzip"),
+    )
+    .await;
+    assert_eq!(response.headers()[header::CONTENT_ENCODING], "gzip");
+    assert_eq!(response.headers()[header::VARY], "accept-encoding");
+    assert_eq!(decode_gzip(&bytes(response).await), expected);
+}

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

@@ -511,6 +511,7 @@ mod provider {
                 total_slots: None,
             },
             account: AccountRecord {
+                outbound_proxy_enabled: true,
                 notes: None,
                 model_access: Default::default(),
                 id: "acct_error_test".to_owned(),

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

@@ -88,6 +88,7 @@ mod account_groups;
 mod accounts;
 mod auth;
 mod client_keys;
+mod compression;
 mod error_diagnostics;
 mod errors;
 mod observability;
@@ -184,6 +185,7 @@ impl AdminTestFixture {
             plugin_ports.clone(),
             plugin_ports.clone(),
             plugin_ports.clone(),
+            plugin_ports.clone(),
         );
         let providers: Vec<Arc<dyn ProviderAdmin>> = vec![
             Arc::new(UnusedProvider::new("openai", Arc::clone(&provider_error))),
@@ -709,6 +711,7 @@ impl MemoryAccountGroupStore {
             (
                 primary_id.clone(),
                 AccountGroupRecord {
+                    allow_image_generation: true,
                     fast_mode: FastMode::Default,
                     id: primary_id,
                     name: "Alpha routing".to_owned(),
@@ -731,6 +734,7 @@ impl MemoryAccountGroupStore {
             (
                 secondary_id.clone(),
                 AccountGroupRecord {
+                    allow_image_generation: true,
                     fast_mode: FastMode::Default,
                     id: secondary_id,
                     name: "Beta routing".to_owned(),
@@ -889,6 +893,7 @@ impl AccountGroupStore for MemoryAccountGroupStore {
         let mut state = self.state.lock().expect("account groups");
         let now = Utc::now();
         let record = AccountGroupRecord {
+            allow_image_generation: command.allow_image_generation,
             fast_mode: command.fast_mode,
             id: command.id.clone(),
             name: command.name,
@@ -924,6 +929,9 @@ impl AccountGroupStore for MemoryAccountGroupStore {
         if let Some(fast_mode) = command.fast_mode {
             record.fast_mode = fast_mode;
         }
+        if let Some(allowed) = command.allow_image_generation {
+            record.allow_image_generation = allowed;
+        }
         record.updated_at = Utc::now();
         mutation(&mut state, command.id, true)
     }
@@ -1379,6 +1387,27 @@ impl AccountRuntimeStore for UnusedStore {
 
 #[async_trait]
 impl ObservabilityStore for UnusedStore {
+    async fn diagnostic_errors(
+        &self,
+        target: &gateway_admin::model::diagnostics::DiagnosticTarget,
+        limit: u16,
+    ) -> AdminStoreResult<Vec<gateway_admin::model::observability::OpsError>> {
+        let errors = self.ops_errors.lock().expect("ops errors");
+        Ok(errors
+            .iter()
+            .filter(|error| match target {
+                gateway_admin::model::diagnostics::DiagnosticTarget::Request(id) => {
+                    error.request_id.as_ref() == Some(id)
+                }
+                gateway_admin::model::diagnostics::DiagnosticTarget::Operation(id) => {
+                    &error.event_id == id && error.source == "ops_event"
+                }
+            })
+            .take(usize::from(limit))
+            .cloned()
+            .collect())
+    }
+
     fn query_timeout(&self) -> std::time::Duration {
         std::time::Duration::from_secs(10)
     }
@@ -1468,11 +1497,12 @@ impl ObservabilityStore for UnusedStore {
         })
     }
 
-    async fn usage_record_detail(&self, _: &str) -> AdminStoreResult<UsageDetail> {
+    async fn usage_record_detail(&self, request_id: &str) -> AdminStoreResult<UsageDetail> {
         self.usage_detail
             .lock()
             .expect("usage detail")
             .clone()
+            .filter(|detail| detail.request.id == request_id)
             .ok_or_else(|| unavailable("usage detail"))
     }
 

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

@@ -0,0 +1,402 @@
+//! 诊断下载的权限、证据完整性、脱敏与部分失败合同
+
+use axum::{
+    body::{Body, to_bytes},
+    http::{Request, StatusCode, header},
+};
+use chrono::Utc;
+use gateway_admin::model::observability::UsageDetail;
+use gateway_api::admin::observability;
+use serde_json::{Value, json};
+use tower::ServiceExt as _;
+
+use super::response::{ops_error_record, usage_record_with_account};
+use crate::admin::{AdminTestFixture, AdminTestState};
+
+async fn export(
+    fixture: &AdminTestFixture,
+    query: &str,
+    authenticated: bool,
+) -> (StatusCode, axum::http::HeaderMap, Value) {
+    let mut request = Request::builder()
+        .uri(format!("/api/admin/diagnostics/export{query}"))
+        .header("x-request-id", "diagnostics-export-test");
+    if authenticated {
+        request = request.header(header::COOKIE, "cpr_session=valid-session");
+    }
+    let response = observability::router::<AdminTestState>()
+        .with_state(fixture.state())
+        .oneshot(request.body(Body::empty()).unwrap())
+        .await
+        .unwrap();
+    let status = response.status();
+    let headers = response.headers().clone();
+    let body = to_bytes(response.into_body(), 20 * 1024 * 1024)
+        .await
+        .unwrap();
+    (status, headers, serde_json::from_slice(&body).unwrap())
+}
+
+#[tokio::test]
+async fn diagnostic_export_requires_admin_and_exactly_one_target() {
+    let fixture = AdminTestFixture::new().await;
+    fixture.auth.insert_session("valid-session");
+    assert_eq!(
+        export(&fixture, "?errorId=missing", false).await.0,
+        StatusCode::UNAUTHORIZED
+    );
+    for query in [
+        "",
+        "?requestId=a&errorId=b",
+        "?requestId=",
+        "?errorId=a&payload=untrusted",
+    ] {
+        assert_eq!(
+            export(&fixture, query, true).await.0,
+            StatusCode::BAD_REQUEST,
+            "{query}"
+        );
+    }
+    assert_eq!(
+        export(&fixture, "?errorId=missing", true).await.0,
+        StatusCode::NOT_FOUND
+    );
+}
+
+#[tokio::test]
+async fn diagnostic_export_keeps_retry_evidence_and_original_errors_but_not_credentials() {
+    let fixture = AdminTestFixture::new().await;
+    fixture.auth.insert_session("valid-session");
+    let mut error = ops_error_record();
+    error.error_details = Some(json!({
+        "causes": ["TLS certificate verification failed", "pool timeout after 5 seconds"],
+        "upstream": json!({"error": {"code": "unknown_vendor_failure", "message": "quota exhausted"},
+            "headers": {"Authorization": "fixture-authorization", "set-cookie": "fixture-cookie"},
+            "access_token": "fixture-access", "refreshToken": "fixture-refresh",
+        }).to_string(),
+        "endpoint": "postgres://fixture-user:fixture-password@database:5432/main",
+        "message": "connect failed: Bearer fixture-bearer, retry after 30 seconds",
+    }).to_string());
+    let mut unrelated = error.clone();
+    unrelated.request_id = Some("another-request".into());
+    unrelated.message = "unrelated evidence must not leak".into();
+    fixture
+        .ops_errors
+        .lock()
+        .unwrap()
+        .extend([error, unrelated]);
+    let trace = json!({"droppedEvents": 0, "events": [
+        {"stage": "request.routed", "data": {"accountPolicy": {"maxAccountRotations": 3}}},
+        {"stage": "account.selection", "data": {"candidates": [{"id": "acct_high", "weight": 2}, {"id": "acct_low", "weight": 1}]}},
+        {"stage": "attempt.failed", "data": {"kind": "unavailable", "recovery": {"replaySafe": true}}},
+        {"stage": "retry.decided", "data": {"phase": "execute", "retryable": false, "clientRetrySuppressed": true, "replayBarrierReached": false}},
+    ]});
+    fixture.usage_detail.lock().unwrap().replace(UsageDetail {
+        request: usage_record_with_account(
+            "req_err",
+            "acct_err",
+            "private name",
+            "fixture@example.invalid",
+            "api_key",
+            Utc::now(),
+        ),
+        attempts: Vec::new(),
+        trace: Some(trace.clone()),
+        related_requests: Vec::new(),
+    });
+    let (status, headers, file) = export(&fixture, "?requestId=req_err", true).await;
+    assert_eq!(status, StatusCode::OK, "{file}");
+    assert_eq!(
+        headers[header::CONTENT_DISPOSITION],
+        "attachment; filename=diagnostics.json"
+    );
+    assert_eq!(headers[header::CACHE_CONTROL], "private, no-store");
+    assert_eq!(file["schemaVersion"], 5);
+    assert_eq!(file["requests"][0]["evidence"]["trace"], trace);
+    assert_eq!(
+        file["requests"][0]["evidence"]["incidentConfiguration"]["revision"],
+        1
+    );
+    let details = &file["requests"][0]["errors"][0]["errorDetails"];
+    assert_eq!(
+        details["upstream"]["error"]["code"],
+        "unknown_vendor_failure"
+    );
+    assert_eq!(details["causes"][0], "TLS certificate verification failed");
+    assert_eq!(file["requests"][0]["errors"][0]["accountId"], "acct_err");
+    assert!(
+        file["availability"]["gaps"]
+            .as_array()
+            .unwrap()
+            .iter()
+            .any(|gap| gap["section"] == "environment")
+    );
+    let text = file.to_string();
+    for marker in [
+        "fixture-authorization",
+        "fixture-cookie",
+        "fixture-access",
+        "fixture-refresh",
+        "fixture-password",
+        "fixture-user",
+        "fixture-bearer",
+        "fixture@example.invalid",
+        "private name",
+        "unrelated evidence must not leak",
+        "203.0.113.8",
+    ] {
+        assert!(!text.contains(marker), "leaked {marker}");
+    }
+    assert!(text.contains("retry after 30 seconds"));
+}
+
+#[tokio::test]
+async fn background_error_exports_without_fabricating_a_model_request() {
+    let fixture = AdminTestFixture::new().await;
+    fixture.auth.insert_session("valid-session");
+    let mut error = ops_error_record();
+    error.source = "ops_event".into();
+    error.request_id = None;
+    error.component = "worker".into();
+    error.error_details =
+        Some(json!({"causes": ["database connection refused"], "upstream": null}).to_string());
+    fixture.ops_errors.lock().unwrap().push(error);
+    let (status, _, file) = export(&fixture, "?errorId=err_snapshot", true).await;
+    assert_eq!(status, StatusCode::OK);
+    assert_eq!(file["requests"], json!([]));
+    assert_eq!(file["error"]["requestId"], Value::Null);
+    assert_eq!(
+        file["error"]["errorDetails"]["causes"][0],
+        "database connection refused"
+    );
+}
+
+#[tokio::test]
+async fn error_survives_missing_request_and_nested_credentials_are_redacted() {
+    let fixture = AdminTestFixture::new().await;
+    fixture.auth.insert_session("valid-session");
+    let mut error = ops_error_record();
+    error.error_details = Some(json!({
+        "cookies": "Cookie: a=fixture-cookie-a; b=fixture-cookie-b\nupstream timeout",
+        "uri": "https://provider.invalid/path?api_key=fixture-query&model=gpt-test",
+        "headers": [["Authorization", "fixture-tuple"], {"name": "Cookie", "value": "fixture-header"}],
+    }).to_string());
+    fixture.ops_errors.lock().unwrap().push(error);
+    let (status, _, file) = export(&fixture, "?requestId=req_err", true).await;
+    assert_eq!(status, StatusCode::OK);
+    assert!(file["requests"][0]["evidence"].is_null());
+    assert!(
+        file["availability"]["gaps"]
+            .as_array()
+            .unwrap()
+            .iter()
+            .any(|gap| gap["reason"] == "request_detail_unavailable")
+    );
+    let text = file.to_string();
+    for marker in [
+        "fixture-cookie-a",
+        "fixture-cookie-b",
+        "fixture-query",
+        "fixture-tuple",
+        "fixture-header",
+    ] {
+        assert!(!text.contains(marker), "leaked {marker}");
+    }
+    assert!(text.contains("upstream timeout"));
+    assert!(text.contains("model=gpt-test"));
+}
+
+#[tokio::test]
+async fn oversized_error_export_reports_truncation_instead_of_silent_loss() {
+    let fixture = AdminTestFixture::new().await;
+    fixture.auth.insert_session("valid-session");
+    let mut error = ops_error_record();
+    error.source = "ops_event".into();
+    error.request_id = None;
+    error.message = "诊断文本".repeat(100_000);
+    fixture.ops_errors.lock().unwrap().push(error);
+    let (status, _, file) = export(&fixture, "?errorId=err_snapshot", true).await;
+    assert_eq!(status, StatusCode::OK);
+    assert_eq!(file["exportTruncated"], true);
+    assert!(file["redaction"]["truncatedValues"].as_u64().unwrap() > 0);
+    assert!(file["error"]["message"]["omittedBytes"].as_u64().unwrap() > 0);
+}
+
+#[tokio::test]
+async fn related_read_failures_and_limits_keep_primary_evidence() {
+    let fixture = AdminTestFixture::new().await;
+    fixture.auth.insert_session("valid-session");
+    let mut related = vec![json!({"requestId": "req_err", "relation": "recovers"})];
+    related.extend(
+        (0..10).map(
+            |index| json!({"requestId": format!("missing_{index}"), "relation": "recovered_by"}),
+        ),
+    );
+    fixture.usage_detail.lock().unwrap().replace(UsageDetail {
+        request: usage_record_with_account(
+            "req_err",
+            "acct_err",
+            "fixture",
+            "fixture@example.invalid",
+            "api_key",
+            Utc::now(),
+        ),
+        attempts: Vec::new(),
+        trace: None,
+        related_requests: related,
+    });
+    fixture
+        .ops_errors
+        .lock()
+        .unwrap()
+        .extend((0..130).map(|index| {
+            let mut error = ops_error_record();
+            error.event_id = format!("error_{index}");
+            error
+        }));
+    let (status, _, file) = export(&fixture, "?requestId=req_err", true).await;
+    assert_eq!(status, StatusCode::OK);
+    assert_eq!(file["requests"].as_array().unwrap().len(), 8);
+    assert_eq!(file["requests"][0]["errors"].as_array().unwrap().len(), 128);
+    assert_eq!(
+        file["requests"][0]["evidence"]["detail"]["requestId"],
+        "req_err"
+    );
+    assert!(file["requests"][1]["evidence"].is_null());
+    let gaps = file["availability"]["gaps"].as_array().unwrap();
+    for reason in [
+        "related_request_limit",
+        "error_record_limit",
+        "request_detail_unavailable",
+        "not_recorded_or_retained",
+    ] {
+        assert!(
+            gaps.iter().any(|gap| gap["reason"] == reason),
+            "missing {reason}"
+        );
+    }
+}
+
+#[tokio::test]
+async fn credential_schemas_are_redacted_without_guessing_business_field_names() {
+    let fixture = AdminTestFixture::new().await;
+    fixture.auth.insert_session("valid-session");
+    let mut error = ops_error_record();
+    let business = json!({
+        "token": {"input": 128, "output": 64},
+        "credentials": {"status": "expired", "provider": "openai"},
+        "secret": "tool argument name", "clientSecret": "business field description",
+        "accountId": "account-correlation", "userId": "user-correlation",
+        "planType": "plus", "status": "active", "accessTokenExpiresAt": "2026-10-16T00:00:00Z",
+        "error": {"code": "unknown_code", "message": "upstream capacity exceeded"},
+    });
+    error.error_details = Some(json!({
+        "business": business,
+        "account": {"accessToken": "fixture-access", "refreshToken": "fixture-refresh", "idToken": "fixture-id"},
+        "auth": {"OPENAI_API_KEY": "fixture-key", "personal_access_token": "fixture-pat", "tokens": {
+            "access_token": "fixture-official-access", "refresh_token": "fixture-official-refresh",
+            "id_token": "fixture-official-id", "account_id": "account-correlation",
+        }},
+    }).to_string());
+    fixture.ops_errors.lock().unwrap().push(error);
+    let (status, _, file) = export(&fixture, "?requestId=req_err", true).await;
+    assert_eq!(status, StatusCode::OK);
+    let details = &file["requests"][0]["errors"][0]["errorDetails"];
+    assert_eq!(details["business"], business);
+    assert_eq!(
+        details["auth"]["tokens"]["account_id"],
+        "account-correlation"
+    );
+    assert!(!file.to_string().contains("fixture-"));
+}
+
+#[tokio::test]
+async fn export_budget_keeps_format_target_and_existing_gap_records() {
+    let fixture = AdminTestFixture::new().await;
+    fixture.auth.insert_session("valid-session");
+    fixture.usage_detail.lock().unwrap().replace(UsageDetail {
+        request: usage_record_with_account(
+            "req_err",
+            "acct_err",
+            "QA",
+            "qa@example.invalid",
+            "oauth",
+            Utc::now(),
+        ),
+        attempts: Vec::new(),
+        trace: Some(json!({"events": vec![vec![true; 32]; 3500]})),
+        related_requests: Vec::new(),
+    });
+    let (status, _, file) = export(&fixture, "?requestId=req_err", true).await;
+    assert_eq!(status, StatusCode::OK);
+    assert_eq!(file["schemaVersion"], 5);
+    assert_eq!(file["source"], "gateway.diagnostics");
+    assert_eq!(file["target"]["requestId"], "req_err");
+    assert_eq!(file["exportTruncated"], true);
+    let gaps = file["availability"]["gaps"].as_array().unwrap();
+    assert!(gaps.iter().any(|gap| gap["reason"] == "export_size_limit"));
+    assert!(
+        gaps.iter()
+            .any(|gap| gap["reason"] == "incident_policy_not_recorded")
+    );
+    assert!(serde_json::to_vec_pretty(&file).unwrap().len() < 16 * 1024 * 1024);
+}
+
+#[tokio::test]
+async fn selected_event_is_exported_once_and_shares_the_total_error_budget() {
+    let fixture = AdminTestFixture::new().await;
+    fixture.auth.insert_session("valid-session");
+    let mut selected = ops_error_record();
+    selected.source = "ops_event".into();
+    let mut errors = vec![selected];
+    errors.extend((0..129).map(|index| {
+        let mut error = ops_error_record();
+        error.source = "ops_event".into();
+        error.event_id = format!("related_{index}");
+        error
+    }));
+    fixture.ops_errors.lock().unwrap().extend(errors);
+    let (status, _, file) = export(&fixture, "?errorId=err_snapshot", true).await;
+    assert_eq!(status, StatusCode::OK);
+    assert_eq!(file["error"]["id"], "err_snapshot");
+    let errors = file["requests"][0]["errors"].as_array().unwrap();
+    assert_eq!(errors.len(), 127);
+    assert!(errors.iter().all(|error| error["id"] != "err_snapshot"));
+    assert!(
+        file["availability"]["gaps"]
+            .as_array()
+            .unwrap()
+            .iter()
+            .any(|gap| gap["reason"] == "error_record_limit")
+    );
+}
+
+#[tokio::test]
+async fn inline_authentication_headers_are_redacted_without_losing_error_context() {
+    let fixture = AdminTestFixture::new().await;
+    fixture.auth.insert_session("valid-session");
+    let mut error = ops_error_record();
+    error.error_details = Some(json!({
+        "headers": "x-api-key: fixture-api-key\nx-oai-is: fixture-session\nx-oai-is-update: fixture-update\nx-oai-attestation: fixture-attestation fixture-attestation-tail\nauthorization: Custom fixture-auth\nproxy-authorization: Custom fixture-proxy\nupstream timeout",
+        "message": "access-token=fixture-access; refresh-token=fixture-refresh; model=gpt-test",
+    }).to_string());
+    fixture.ops_errors.lock().unwrap().push(error);
+    let (status, _, file) = export(&fixture, "?requestId=req_err", true).await;
+    assert_eq!(status, StatusCode::OK);
+    let text = file.to_string();
+    for marker in [
+        "fixture-api-key",
+        "fixture-session",
+        "fixture-update",
+        "fixture-attestation",
+        "fixture-attestation-tail",
+        "fixture-auth",
+        "fixture-proxy",
+        "fixture-access",
+        "fixture-refresh",
+    ] {
+        assert!(!text.contains(marker), "leaked {marker}");
+    }
+    assert!(text.contains("upstream timeout"));
+    assert!(text.contains("model=gpt-test"));
+}

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

@@ -1,4 +1,5 @@
 //! Admin 观测 wire/路由的固定合同测试
 
+mod export;
 mod query;
 mod response;

backend/crates/gateway-api/tests/admin/observability/response.rs

@@ -398,7 +398,7 @@ async fn usage_detail_should_keep_attempt_snapshot_contract() {
 }
 
 #[tokio::test]
-async fn zero_attempt_failure_detail_keeps_missing_upstream_facts_and_preparation_trace() {
+async fn zero_attempt_failure_detail_omits_trace_and_keeps_missing_upstream_facts() {
     use axum::{
         body::{Body, to_bytes},
         http::{Request, StatusCode, header},
@@ -470,7 +470,8 @@ async fn zero_attempt_failure_detail_keeps_missing_upstream_facts_and_preparatio
     assert_eq!(data["attempts"], json!([]));
     // 明细仍沿用全局 best-effort 合同，不从预备 trace 推断完整尝试列表
     assert_eq!(data["attemptsComplete"], false);
-    assert_eq!(data["trace"], trace);
+    assert!(data.get("trace").is_none());
+    assert!(data.get("relatedRequests").is_none());
     assert_eq!(data["logicalOutcome"], "failed");
     assert_eq!(data["clientStatusCode"], 503);
     assert_eq!(data["message"], "no upstream account is available");
@@ -495,8 +496,6 @@ async fn ops_errors_should_keep_account_label_and_authentication_contract() {
         body::{Body, to_bytes},
         http::{Request, StatusCode, header},
     };
-    use chrono::Utc;
-    use gateway_admin::model::observability::OpsError;
     use gateway_api::admin::observability;
     use tower::ServiceExt as _;
 
@@ -508,63 +507,7 @@ async fn ops_errors_should_keep_account_label_and_authentication_contract() {
         .ops_errors
         .lock()
         .expect("ops errors")
-        .push(OpsError {
-            client_api_key_name: Some("Production".to_owned()),
-            source: "model_request".to_owned(),
-            event_id: "err_snapshot".to_owned(),
-            request_id: Some("req_err".to_owned()),
-            attempt_index: Some(1),
-            client_api_key_ref: Some("key_err".to_owned()),
-            component: "model_request".to_owned(),
-            operation: "responses".to_owned(),
-            protocol: Some("openai".to_owned()),
-            client_transport: Some("http_sse".to_owned()),
-            requested_model_id: Some("gpt-5.4".to_owned()),
-            service_tier: Some("priority".to_owned()),
-            endpoint: Some("/v1/responses".to_owned()),
-            provider_kind: Some("openai".to_owned()),
-            provider_account_ref: Some("acct_err".to_owned()),
-            provider_account_name: None,
-            provider_account_email: Some("err@example.invalid".to_owned()),
-            provider_account_plan_type: Some("pro".to_owned()),
-            provider_account_plan_type_display: Some("Pro".to_owned()),
-            provider_account_authentication_kind: Some("api_key".to_owned()),
-            upstream_model_id: Some("upstream-err".to_owned()),
-            upstream_transport: Some("http_sse".to_owned()),
-            failure_kind: "upstream_error".to_owned(),
-            upstream_send_state: Some("sent".to_owned()),
-            client_status_code: Some(502),
-            upstream_status_code: Some(502),
-            provider_error_code: Some("upstream".to_owned()),
-            client_response_id: None,
-            upstream_request_id: None,
-            latency_ms: Some(90),
-            message: "snapshot error".to_owned(),
-            error_details: Some(
-                r#"{"error":{"code":"upstream","message":"raw upstream marker"}}"#.to_owned(),
-            ),
-            client_ip: Some("203.0.113.8".to_owned()),
-            user_agent: Some("codex-cli/0.144.0".to_owned()),
-            reasoning_effort: Some("medium".to_owned()),
-            reasoning_preset: None,
-            request_kind: Some("root".to_owned()),
-            subagent_kind: None,
-            compact: Some(false),
-            continuation_affinity_hash: Some("affinity-hash".to_owned()),
-            continuation_previous_response_id_hash: Some("response-hash".to_owned()),
-            continuation_unavailable_reason: Some("reused_connection_lost".to_owned()),
-            upstream_connection_id: Some("connection-id".to_owned()),
-            upstream_connection_exit_reason: Some("tcp_reset".to_owned()),
-            upstream_connection_age_ms: Some(12_000),
-            upstream_connection_idle_ms: Some(4_000),
-            recovery_request_id: None,
-            recovered_at: None,
-            recovery_attempt_count: 0,
-            recovery_retry_delay_ms: None,
-            recovery_total_latency_ms: None,
-            occurred_at: Utc::now(),
-            stable_sort_id: "model_request:req_err".to_owned(),
-        });
+        .push(ops_error_record());
     let response = observability::router::<AdminTestState>()
         .with_state(fixture.state())
         .oneshot(
@@ -621,7 +564,7 @@ async fn ops_errors_should_keep_account_label_and_authentication_contract() {
             "userAgent": "codex-cli/0.144.0",
             "providerErrorCode": "upstream",
             "upstreamSendState": "sent",
-            "errorDetails": "{\"error\":{\"code\":\"upstream\",\"message\":\"raw upstream marker\"}}",
+            "errorDetails": serde_json::to_string_pretty(&json!({"error": {"code": "upstream", "message": "raw upstream marker"}})).unwrap(),
             "continuationUnavailableReason": "reused_connection_lost",
             "upstreamConnectionExitReason": "tcp_reset",
             "upstreamConnectionAgeMs": 12000,
@@ -707,7 +650,7 @@ async fn diagnostics_should_keep_stable_key_and_display_name_contract() {
     );
 }
 
-fn usage_record_with_account(
+pub(super) fn usage_record_with_account(
     id: &str,
     account_id: &str,
     name: &str,
@@ -858,6 +801,7 @@ async fn usage_route_should_expose_table_facts_without_detail_payload() {
         .lock()
         .expect("usage records")
         .push(UsageListRecord {
+            outcome: gateway_admin::model::observability::RequestOutcome::Succeeded,
             client_api_key_name: Some("Production".to_owned()),
             id: "request_endpoint".to_owned(),
             endpoint: "/v1/responses".to_owned(),
@@ -1303,3 +1247,123 @@ async fn observability_queries_reject_group_filters_and_analytics_search() {
     assert!(observations.records.is_empty());
     assert!(observations.errors.is_empty());
 }
+
+pub(super) fn ops_error_record() -> gateway_admin::model::observability::OpsError {
+    use gateway_admin::model::observability::OpsError;
+    OpsError {
+        client_api_key_name: Some("Production".to_owned()),
+        source: "model_request".to_owned(),
+        event_id: "err_snapshot".to_owned(),
+        request_id: Some("req_err".to_owned()),
+        attempt_index: Some(1),
+        client_api_key_ref: Some("key_err".to_owned()),
+        component: "model_request".to_owned(),
+        operation: "responses".to_owned(),
+        protocol: Some("openai".to_owned()),
+        client_transport: Some("http_sse".to_owned()),
+        requested_model_id: Some("gpt-5.4".to_owned()),
+        service_tier: Some("priority".to_owned()),
+        endpoint: Some("/v1/responses".to_owned()),
+        provider_kind: Some("openai".to_owned()),
+        provider_account_ref: Some("acct_err".to_owned()),
+        provider_account_name: None,
+        provider_account_email: Some("err@example.invalid".to_owned()),
+        provider_account_plan_type: Some("pro".to_owned()),
+        provider_account_plan_type_display: Some("Pro".to_owned()),
+        provider_account_authentication_kind: Some("api_key".to_owned()),
+        upstream_model_id: Some("upstream-err".to_owned()),
+        upstream_transport: Some("http_sse".to_owned()),
+        failure_kind: "upstream_error".to_owned(),
+        upstream_send_state: Some("sent".to_owned()),
+        client_status_code: Some(502),
+        upstream_status_code: Some(502),
+        provider_error_code: Some("upstream".to_owned()),
+        client_response_id: None,
+        upstream_request_id: None,
+        latency_ms: Some(90),
+        message: "snapshot error".to_owned(),
+        error_details: Some(
+            r#"{"error":{"code":"upstream","message":"raw upstream marker"}}"#.to_owned(),
+        ),
+        client_ip: Some("203.0.113.8".to_owned()),
+        user_agent: Some("codex-cli/0.144.0".to_owned()),
+        reasoning_effort: Some("medium".to_owned()),
+        reasoning_preset: None,
+        request_kind: Some("root".to_owned()),
+        subagent_kind: None,
+        compact: Some(false),
+        continuation_affinity_hash: Some("affinity-hash".to_owned()),
+        continuation_previous_response_id_hash: Some("response-hash".to_owned()),
+        continuation_unavailable_reason: Some("reused_connection_lost".to_owned()),
+        upstream_connection_id: Some("connection-id".to_owned()),
+        upstream_connection_exit_reason: Some("tcp_reset".to_owned()),
+        upstream_connection_age_ms: Some(12_000),
+        upstream_connection_idle_ms: Some(4_000),
+        recovery_request_id: None,
+        recovered_at: None,
+        recovery_attempt_count: 0,
+        recovery_retry_delay_ms: None,
+        recovery_total_latency_ms: None,
+        occurred_at: Utc::now(),
+        stable_sort_id: "model_request:req_err".to_owned(),
+    }
+}
+
+#[tokio::test]
+async fn usage_status_query_accepts_three_modes_and_rejects_unknown_values() {
+    use crate::admin::{AdminTestFixture, AdminTestState};
+    use axum::{
+        body::{Body, to_bytes},
+        http::{Request, StatusCode, header},
+    };
+    use gateway_api::admin::observability;
+    use tower::ServiceExt as _;
+    let fixture = AdminTestFixture::new().await;
+    fixture.auth.insert_session("valid-session");
+    for (status, expected) in [
+        ("success", StatusCode::OK),
+        ("error", StatusCode::OK),
+        ("all", StatusCode::OK),
+        ("failed", StatusCode::BAD_REQUEST),
+    ] {
+        let response = observability::router::<AdminTestState>()
+            .with_state(fixture.state())
+            .oneshot(
+                Request::builder()
+                    .uri(format!("/api/admin/usage/records?status={status}"))
+                    .header(header::COOKIE, "cpr_session=valid-session")
+                    .header("x-request-id", "usage_status_fixture")
+                    .body(Body::empty())
+                    .unwrap(),
+            )
+            .await
+            .unwrap();
+        let actual = response.status();
+        let body = to_bytes(response.into_body(), 1024 * 1024).await.unwrap();
+        assert_eq!(
+            actual,
+            expected,
+            "{status}: {}",
+            String::from_utf8_lossy(&body)
+        );
+        if expected == StatusCode::OK {
+            use gateway_admin::model::observability::UsageRecordStatus;
+            let status_value = match status {
+                "success" => UsageRecordStatus::Success,
+                "error" => UsageRecordStatus::Error,
+                _ => UsageRecordStatus::All,
+            };
+            assert_eq!(
+                fixture
+                    .observations
+                    .lock()
+                    .unwrap()
+                    .records
+                    .last()
+                    .unwrap()
+                    .status,
+                status_value
+            );
+        }
+    }
+}

backend/crates/gateway-api/tests/admin/plugins/management.rs

@@ -435,6 +435,35 @@ async fn resource_request(
     .unwrap()
 }
 
+#[tokio::test]
+async fn large_html_resource_should_remain_uncompressed() {
+    let fixture = published_fixture().await;
+    let path = format!("ui/{}.html", "a".repeat(2048));
+    let response = crate::openai::api_router_with_admin(fixture.services)
+        .oneshot(
+            Request::builder()
+                .uri(format!(
+                    "/api/admin/plugins/extensions/plugin-instance-diagnostic/{}/7/resources/{path}",
+                    "a".repeat(64)
+                ))
+                .header("cookie", "cpr_session=management-session")
+                .header("accept-encoding", "gzip")
+                .body(Body::empty())
+                .unwrap(),
+        )
+        .await
+        .unwrap();
+    assert_eq!(response.status(), StatusCode::OK);
+    assert_eq!(response.headers()["content-type"], "text/html");
+    assert!(!response.headers().contains_key("content-encoding"));
+    assert_eq!(
+        axum::body::to_bytes(response.into_body(), 4096)
+            .await
+            .unwrap(),
+        path
+    );
+}
+
 #[tokio::test]
 async fn static_resources_revalidate_privately_after_authorization() {
     use axum::body::to_bytes;

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

@@ -1,5 +1,7 @@
 //! 插件管理 HTTP 测试入口，以及制品、实例与权限合同测试
 
+mod usage_data;
+
 use async_trait::async_trait;
 use gateway_admin::model::plugins::distribution::{
     DownloadedPlugin, GithubReleaseQuery, PluginRelease, RemotePluginLocation, SourceCredential,

backend/crates/gateway-api/tests/admin/plugins/usage_data.rs

@@ -0,0 +1,24 @@
+//! 通用用量端口的插件管理测试夹具，非本测试调用路径禁止访问
+use super::TestPluginPorts;
+use gateway_admin::{model::usage_data::*, ports::usage_data::*};
+#[async_trait::async_trait]
+impl UsageDataStore for TestPluginPorts {
+    async fn get_record(&self, _: GetUsageRecord) -> UsageDataResult<UsageRecord> {
+        unreachable!("unexpected usage call")
+    }
+    async fn list_records(&self, _: ListUsageRecords) -> UsageDataResult<UsageRecordPage> {
+        unreachable!("unexpected usage call")
+    }
+    async fn write_records(&self, _: StoredUsageWrite) -> UsageDataResult<UsageWriteResult> {
+        unreachable!("unexpected usage call")
+    }
+    async fn list_aggregates(&self, _: ListUsageAggregates) -> UsageDataResult<UsageAggregatePage> {
+        unreachable!("unexpected usage call")
+    }
+    async fn write_aggregates(
+        &self,
+        _: StoredAggregateWrite,
+    ) -> UsageDataResult<UsageAggregateWriteResult> {
+        unreachable!("unexpected usage call")
+    }
+}

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

@@ -89,6 +89,43 @@ async fn proxy_location_rejects_invalid_and_incomplete_input() {
 #[derive(Default)]
 pub(super) struct MemoryProxies(Mutex<Option<ProxyRecord>>);
 
+#[tokio::test]
+async fn proxy_enabled_updates_preserve_connection_and_omitted_state() {
+    let fixture = AdminTestFixture::new().await;
+    fixture.auth.insert_session("valid-session");
+    let (status, created) = request(
+        &fixture,
+        "/api/admin/proxies/create",
+        Some(json!({
+            "name":"Office", "proxyUrl":"http://proxy.example:8080"
+        })),
+        true,
+    )
+    .await;
+    assert_eq!(status, StatusCode::CREATED);
+    assert_eq!(created["data"]["record"]["enabled"], true);
+    for (revision, enabled, expected) in [
+        (1, Some(false), false),
+        (2, None, false),
+        (3, Some(true), true),
+    ] {
+        let mut body = json!({"id":"proxy_test", "revision":revision, "name":"Office"});
+        if let Some(enabled) = enabled {
+            body["enabled"] = json!(enabled);
+        }
+        let (status, changed) =
+            request(&fixture, "/api/admin/proxies/update", Some(body), true).await;
+        assert_eq!(status, StatusCode::OK);
+        assert_eq!(changed["data"]["record"]["enabled"], expected);
+        assert_eq!(
+            changed["data"]["record"]["endpoint"],
+            created["data"]["record"]["endpoint"]
+        );
+        let (_, listed) = request(&fixture, "/api/admin/proxies", None, true).await;
+        assert_eq!(listed["data"]["items"][0]["enabled"], expected);
+    }
+}
+
 fn missing() -> AdminStoreError {
     AdminStoreError::new(AdminStoreErrorKind::NotFound, "proxy", "missing proxy")
 }
@@ -165,6 +202,7 @@ impl ProxyStore for MemoryProxies {
         _: &MutationContext,
     ) -> AdminStoreResult<ProxyMutation> {
         let mut record = ProxyRecord {
+            enabled: true,
             auto_location: command.auto_location,
             detected_location: None,
             location: command.location,
@@ -197,6 +235,9 @@ impl ProxyStore for MemoryProxies {
         let mut stored = self.0.lock().unwrap();
         let record = stored.as_mut().ok_or_else(missing)?;
         record.name = command.name;
+        if let Some(enabled) = command.enabled {
+            record.enabled = enabled;
+        }
         if let Some(auto_location) = command.auto_location {
             if auto_location != record.auto_location {
                 record.detected_location = None;

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

@@ -62,7 +62,7 @@ fn update_body() -> Value {
         "smartScheduling": gateway_core::account::SmartSchedulingConfig::default(),
         "minCodexDesktopVersion": "26.825.6671",
         "minCodexCliVersion": "0.40.0",
-        "usageRetentionDays": 32,
+        "usageRetentionDays": 31,
         "opsEventRetentionDays": 31,
         "auditRetentionDays": 91,
         "accountAutoFreezeEnabled": true,
@@ -84,9 +84,10 @@ async fn privacy_save_requires_provider_validation_and_preview_requires_admin_au
     fixture.auth.insert_session("valid-session");
     let mut body = update_body();
     let policy = json!({"enabled":true,"onError":"reject_request","rules":[{
-        "id":"auth","name":"自定义认证头","enabled":true,"scope":"request_header",
-        "selector":"authorization","action":"remove_field","pattern":null,
-        "replacement":"","value":null,"replaceAll":true,"caseInsensitive":false,"multiLine":false
+        "id":"auth","name":"自定义认证头","enabled":true,"steps":[{
+            "scope":"request_header","selector":"authorization","action":"remove_field","pattern":null,
+            "replacement":"","value":null,"replaceAll":true,"caseInsensitive":false,"multiLine":false
+        }]
     }]});
     body["codexPrivacyPolicy"] = policy.clone();
     let response = app(fixture.state())
@@ -140,7 +141,7 @@ async fn smart_settings_round_trip_and_invalid_updates_leave_the_saved_value_int
     let fixture = AdminTestFixture::new().await;
     fixture.auth.insert_session("valid-session");
     let mut body = update_body();
-    let custom = json!({"loadWeight": 2.0, "quotaWeight": 0.0, "healthWeight": 1.0, "latencyWeight": 0.5, "resetWeight": 1.2, "queueWeight": 2.3, "preferHigherWeight": true});
+    let custom = json!({"loadWeight": 2.0, "quotaWeight": 0.0, "healthWeight": 1.0, "latencyWeight": 0.5, "resetWeight": 1.2, "queueWeight": 2.3});
     body["smartScheduling"] = custom.clone();
     let response = app(fixture.state())
         .oneshot(request(
@@ -160,8 +161,8 @@ async fn smart_settings_round_trip_and_invalid_updates_leave_the_saved_value_int
     let mut invalid_values = vec![
         json!(null),
         json!({}),
-        json!({"loadWeight":0,"quotaWeight":0,"healthWeight":0,"latencyWeight":0,"resetWeight":0,"queueWeight":0,"preferHigherWeight":true}),
-        json!({"loadWeight":0.01,"quotaWeight":1,"healthWeight":1,"latencyWeight":1,"resetWeight":0,"queueWeight":0,"preferHigherWeight":false}),
+        json!({"loadWeight":0,"quotaWeight":0,"healthWeight":0,"latencyWeight":0,"resetWeight":0,"queueWeight":0}),
+        json!({"loadWeight":0.01,"quotaWeight":1,"healthWeight":1,"latencyWeight":1,"resetWeight":0,"queueWeight":0}),
     ];
     for field in ["resetWeight", "queueWeight"] {
         for value in [json!(-0.1), json!(10.1), json!(0.01), json!(null)] {
@@ -215,6 +216,56 @@ fn settings_request_should_reject_unknown_rotation_strategy() {
     assert_eq!(request.validate().unwrap_err().field(), "rotationStrategy");
 }
 
+#[tokio::test]
+async fn short_request_retention_round_trips_and_invalid_days_do_not_replace_it() {
+    let fixture = AdminTestFixture::new().await;
+    fixture.auth.insert_session("valid-session");
+    let mut body = update_body();
+    for days in [1, 3, 7, 15, 30, 31] {
+        body["usageRetentionDays"] = json!(days);
+        let response = app(fixture.state())
+            .oneshot(request(
+                Method::POST,
+                "/api/admin/settings/update",
+                Some(body.clone()),
+            ))
+            .await
+            .unwrap();
+        assert_eq!(response.status(), StatusCode::OK);
+        let saved = response_json(response).await;
+        assert_eq!(saved["data"]["usageRetentionDays"], days);
+        body["configRevision"] = saved["data"]["configRevision"].clone();
+    }
+    for (days, expected_status) in [
+        (json!(0), StatusCode::BAD_REQUEST),
+        (json!(32), StatusCode::BAD_REQUEST),
+        (json!(u32::MAX), StatusCode::BAD_REQUEST),
+        (json!(-1), StatusCode::UNPROCESSABLE_ENTITY),
+        (json!(1.5), StatusCode::UNPROCESSABLE_ENTITY),
+        (json!(null), StatusCode::UNPROCESSABLE_ENTITY),
+    ] {
+        body["usageRetentionDays"] = days;
+        let response = app(fixture.state())
+            .oneshot(request(
+                Method::POST,
+                "/api/admin/settings/update",
+                Some(body.clone()),
+            ))
+            .await
+            .unwrap();
+        assert_eq!(response.status(), expected_status);
+    }
+    let response = app(fixture.state())
+        .oneshot(request(Method::GET, "/api/admin/settings", None))
+        .await
+        .unwrap();
+    assert_eq!(response.status(), StatusCode::OK);
+    assert_eq!(
+        response_json(response).await["data"]["usageRetentionDays"],
+        31
+    );
+}
+
 #[test]
 fn settings_request_accepts_unlimited_default_account_concurrency() {
     let mut body = update_body();
@@ -297,7 +348,7 @@ fn settings_response_should_cover_the_full_runtime_settings_contract() {
             smart_scheduling: gateway_core::account::SmartSchedulingConfig::default(),
             min_codex_desktop_version: Some("26.825.6671".to_owned()),
             min_codex_cli_version: Some("0.40.0".to_owned()),
-            usage_retention_days: 32,
+            usage_retention_days: 31,
             ops_event_retention_days: 31,
             audit_retention_days: 91,
             account_auto_freeze_enabled: true,
@@ -348,7 +399,7 @@ fn settings_response_should_cover_the_full_runtime_settings_contract() {
             "smartSchedulingDefaults": gateway_core::account::SmartSchedulingConfig::default(),
             "minCodexDesktopVersion": "26.825.6671",
             "minCodexCliVersion": "0.40.0",
-            "usageRetentionDays": 32,
+            "usageRetentionDays": 31,
             "opsEventRetentionDays": 31,
             "auditRetentionDays": 91,
             "accountAutoFreezeEnabled": true,

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

@@ -9,7 +9,7 @@ use axum::{
     http::{Request, StatusCode, header},
 };
 use chrono::Utc;
-use futures::stream;
+use futures::{StreamExt as _, stream};
 use gateway_admin::{
     model::system::{
         SystemOperationAccepted, SystemOperationKind, SystemOperationState, SystemOperationStatus,
@@ -19,6 +19,7 @@ use gateway_admin::{
     ports::system::{SystemOperationError, SystemOperations, SystemUpdateEventStream},
 };
 use gateway_api::admin::system::{self, UpdateDetailQuery, UpdateRequest};
+use http_body_util::BodyExt as _;
 use tower::ServiceExt as _;
 
 use super::{AdminTestFixture, AdminTestState, unavailable_system};
@@ -40,25 +41,39 @@ fn update_request_should_preserve_target_version_for_domain_validation() {
     );
 }
 
-#[tokio::test]
+#[tokio::test(start_paused = true)]
 async fn update_event_stream_should_preserve_event_id_and_download_progress_percent() {
-    let fixture = AdminTestFixture::with_system(Arc::new(ProgressSystem)).await;
+    let fixture = AdminTestFixture::with_system(Arc::new(ProgressSystem::default())).await;
     fixture.auth.insert_session("valid-session");
-    let response = app(fixture.state())
+    let response = gateway_api::admin::router::<AdminTestState>()
+        .with_state(fixture.state())
         .oneshot(
             Request::builder()
                 .uri("/api/admin/system/update/events")
                 .header(header::COOKIE, "cpr_session=valid-session")
+                .header(header::ACCEPT_ENCODING, "gzip")
                 .header("x-request-id", "req_system_update_events")
                 .body(Body::empty())
                 .expect("update event request"),
-        )
+        );
+    let response = tokio::time::timeout(std::time::Duration::from_secs(1), response)
         .await
+        .expect("SSE response must not wait for stream completion")
         .expect("update event response");
-    let body = to_bytes(response.into_body(), 1024 * 1024)
+    assert_eq!(response.status(), StatusCode::OK);
+    assert_eq!(
+        response.headers()[header::CONTENT_TYPE],
+        "text/event-stream"
+    );
+    assert!(!response.headers().contains_key(header::CONTENT_ENCODING));
+    let mut stream = response.into_body();
+    let frame = tokio::time::timeout(std::time::Duration::from_secs(1), stream.frame())
         .await
-        .expect("update event body");
-    let body = String::from_utf8(body.to_vec()).expect("UTF-8 event stream");
+        .expect("first SSE event must arrive before stream completion")
+        .expect("first event")
+        .expect("event body");
+    let body = String::from_utf8(frame.into_data().expect("event data").to_vec())
+        .expect("UTF-8 event stream");
 
     assert!(
         body.contains(r#""id":"update-progress-1""#),
@@ -72,7 +87,7 @@ async fn update_event_stream_should_preserve_event_id_and_download_progress_perc
 
 #[tokio::test]
 async fn update_should_return_accepted_and_report_restart_only_through_status() {
-    let fixture = AdminTestFixture::with_system(Arc::new(ProgressSystem)).await;
+    let fixture = AdminTestFixture::with_system(Arc::new(ProgressSystem::default())).await;
     fixture.auth.insert_session("valid-session");
     let router = app(fixture.state());
     let response = router
@@ -119,7 +134,55 @@ fn app(state: AdminTestState) -> Router {
     system::router::<AdminTestState>().with_state(state)
 }
 
-struct ProgressSystem;
+#[tokio::test]
+async fn update_detail_should_compress_json_starting_at_one_kib() {
+    async fn detail(notes: String, encoding: &str) -> axum::response::Response {
+        let fixture =
+            AdminTestFixture::with_system(Arc::new(ProgressSystem { notes: Some(notes) })).await;
+        fixture.auth.insert_session("valid-session");
+        crate::openai::api_router_with_admin(fixture.services)
+            .oneshot(
+                Request::builder()
+                    .uri("/api/admin/system/update/detail")
+                    .header(header::COOKIE, "cpr_session=valid-session")
+                    .header(header::ACCEPT_ENCODING, encoding)
+                    .body(Body::empty())
+                    .unwrap(),
+            )
+            .await
+            .unwrap()
+    }
+
+    let response = detail(String::new(), "identity").await;
+    assert_eq!(response.status(), StatusCode::OK);
+    let base = to_bytes(response.into_body(), 4096).await.unwrap();
+    assert!(base.len() < 1023);
+    for size in [1023, 1024] {
+        let response = detail("x".repeat(size - base.len()), "gzip").await;
+        assert_eq!(response.status(), StatusCode::OK);
+        let compressed = response.headers().get(header::CONTENT_ENCODING).cloned();
+        let bytes = to_bytes(response.into_body(), 4096).await.unwrap();
+        if size == 1023 {
+            assert!(compressed.is_none());
+            assert_eq!(bytes.len(), size);
+        } else {
+            use std::io::Read as _;
+            assert_eq!(compressed.unwrap(), "gzip");
+            let mut decoded = Vec::new();
+            flate2::read::GzDecoder::new(bytes.as_ref())
+                .read_to_end(&mut decoded)
+                .unwrap();
+            assert_eq!(decoded.len(), size);
+            let body: serde_json::Value = serde_json::from_slice(&decoded).unwrap();
+            assert_eq!(body["data"]["notes"], "x".repeat(size - base.len()));
+        }
+    }
+}
+
+#[derive(Default)]
+struct ProgressSystem {
+    notes: Option<String>,
+}
 
 #[async_trait]
 impl SystemOperations for ProgressSystem {
@@ -144,7 +207,7 @@ impl SystemOperations for ProgressSystem {
             deployment_mode: "docker".into(),
             build_type: "release".into(),
             release_url: None,
-            notes: None,
+            notes: self.notes.clone(),
             cached: false,
             update_supported: true,
             unsupported_reason: None,
@@ -153,18 +216,21 @@ impl SystemOperations for ProgressSystem {
     }
 
     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(),
-            }
-        }))
+        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: false,
+                    progress_percent: Some(40),
+                    occurred_at: Utc::now(),
+                }
+            })
+            .chain(stream::pending()),
+        )
     }
 
     async fn perform_update(
@@ -216,7 +282,7 @@ impl SystemOperations for ProgressSystem {
 
 #[tokio::test]
 async fn update_detail_should_validate_and_forward_the_temporary_channel() {
-    let fixture = AdminTestFixture::with_system(Arc::new(ProgressSystem)).await;
+    let fixture = AdminTestFixture::with_system(Arc::new(ProgressSystem::default())).await;
     fixture.auth.insert_session("valid-session");
     let router = app(fixture.state());
     for (authenticated, query, expected) in [

backend/crates/gateway-api/tests/architecture.rs

@@ -22,13 +22,6 @@ fn manifest_should_depend_only_on_core_protocol_and_http_adapter_layers() {
     );
 }
 
-#[test]
-fn workspace_should_include_gateway_api() {
-    let workspace = include_str!("../../../Cargo.toml");
-
-    assert!(workspace.contains("\"crates/gateway-api\""));
-}
-
 #[test]
 fn admin_routes_should_use_only_get_post_and_static_paths() {
     let root = Path::new(env!("CARGO_MANIFEST_DIR")).join("src/admin");

backend/crates/gateway-api/tests/key_usage/fixtures.rs

@@ -93,6 +93,7 @@ pub(super) async fn fixture() -> AdminTestFixture {
 
 fn usage_record() -> UsageListRecord {
     UsageListRecord {
+        outcome: gateway_admin::model::observability::RequestOutcome::Succeeded,
         client_api_key_name: Some("Production".to_owned()),
         id: "req-visible".to_owned(),
         endpoint: "/v1/responses".to_owned(),

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

@@ -35,6 +35,7 @@ use tower::ServiceExt as _;
 
 #[derive(Clone, Copy, Debug)]
 enum Mode {
+    Overloaded,
     Rewrite,
     Upgrade {
         reject: bool,
@@ -142,6 +143,12 @@ impl MiddlewarePlan for Plan {
         let mode = self.mode;
         Box::pin(async move {
             match mode {
+                Mode::Overloaded => Err(MiddlewareError::Gateway(
+                    gateway_core::error::GatewayError::new(
+                        gateway_core::error::GatewayErrorKind::MiddlewareCapacityUnavailable,
+                        "Request middleware is temporarily busy",
+                    ),
+                )),
                 Mode::Service => unreachable!("service-only plan bypasses HTTP layers"),
                 Mode::Upgrade { reject, idle, .. } => {
                     let upgrade = request
@@ -659,3 +666,25 @@ async fn plugin_http_dispatch_uses_host_identity_without_client_credentials() {
         .unwrap();
     assert_eq!(response.status(), StatusCode::UNAUTHORIZED);
 }
+
+#[tokio::test]
+async fn http_middleware_overload_is_an_explicit_503_before_native_routing() {
+    let (router, plan) = router(Mode::Overloaded, None).await;
+    for path in [
+        "/healthz",
+        "/assets/index.js",
+        "/api/settings",
+        "/v1/responses",
+    ] {
+        let response = router
+            .clone()
+            .oneshot(Request::builder().uri(path).body(Body::empty()).unwrap())
+            .await
+            .unwrap();
+        assert_eq!(response.status(), StatusCode::SERVICE_UNAVAILABLE);
+        let body = response.into_body().collect().await.unwrap().to_bytes();
+        let body: serde_json::Value = serde_json::from_slice(&body).unwrap();
+        assert_eq!(body["error"]["code"], "middleware_capacity_unavailable");
+    }
+    assert_eq!(plan.calls.load(Ordering::Acquire), 4);
+}

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

@@ -263,6 +263,18 @@ fn rate_limited_error_should_map_to_openai_retryable_status() {
     );
 }
 
+#[test]
+fn middleware_capacity_is_service_unavailable() {
+    assert_eq!(
+        gateway_error_contract(GatewayErrorKind::MiddlewareCapacityUnavailable),
+        (
+            StatusCode::SERVICE_UNAVAILABLE,
+            "server_error",
+            "middleware_capacity_unavailable"
+        ),
+    );
+}
+
 #[test]
 fn queue_rejections_keep_distinct_retryable_http_codes() {
     for (kind, code) in [
@@ -548,7 +560,7 @@ fn engine_store_error_should_collapse_to_safe_internal_error() {
 
     let projected = gateway_error_from_engine(&error);
     assert_eq!(projected.kind(), GatewayErrorKind::Internal);
-    assert_eq!(projected.safe_message(), "gateway execution failed");
+    assert_eq!(projected.safe_message(), "request execution failed");
 }
 #[test]
 fn openai_error_response_should_preserve_only_safe_contract_fields() {

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

@@ -0,0 +1,112 @@
+//! 真实 HTTP / WebSocket 路由与 Core 执行入口共同拒绝生图，不触发准入或上游
+
+use super::*;
+use axum::{
+    body::{Body, to_bytes},
+    http::{Request, StatusCode, header::AUTHORIZATION},
+};
+use futures::{SinkExt, StreamExt};
+use serde_json::{Value, json};
+use std::{sync::Arc, time::Duration};
+use tokio_tungstenite::tungstenite::{Message, client::IntoClientRequest};
+use tower::ServiceExt;
+
+fn execution() -> Arc<DefaultExecutionService> {
+    Arc::new(DefaultExecutionService::new(
+        RuntimeSnapshotHandle::new(snapshot_with_image_policy(
+            "sk_image_policy",
+            "openai",
+            "key_image_policy",
+            &["acct_api_test"],
+            false,
+        )),
+        Arc::new(UnusedExecutionStore),
+        ProviderRegistry::default(),
+        Arc::new(UnusedAdmissions),
+        Arc::new(UnusedContinuation),
+        Arc::new(IgnoredClientApiKeyUsage),
+        Arc::new(crate::support::RecordingDiagnostics::default()),
+    ))
+}
+
+#[tokio::test]
+async fn denied_image_requests_return_403_before_admission_or_upstream() {
+    let app = api_router(execution()).await;
+    for (path, body) in [
+        (
+            "/v1/images/generations",
+            json!({"model":"gpt-image-2","prompt":"test"}),
+        ),
+        (
+            "/v1/images/edits",
+            json!({"model":"gpt-image-2","prompt":"test","images":[{"image_url":"data:image/png;base64,AA=="}]}),
+        ),
+        (
+            "/v1/responses",
+            json!({"model":"model-a","input":"test","tools":[{"type":"image_generation"}]}),
+        ),
+        (
+            "/v1/responses",
+            json!({"model":"model-a","input":"test","stream":true,"tool_choice":{"type":"image_generation"}}),
+        ),
+    ] {
+        let response = app
+            .clone()
+            .oneshot(
+                Request::post(path)
+                    .header(AUTHORIZATION, "Bearer sk_image_policy")
+                    .header("content-type", "application/json")
+                    .body(Body::from(body.to_string()))
+                    .unwrap(),
+            )
+            .await
+            .unwrap();
+        let status = response.status();
+        let body: Value =
+            serde_json::from_slice(&to_bytes(response.into_body(), usize::MAX).await.unwrap())
+                .unwrap();
+        assert_eq!(status, StatusCode::FORBIDDEN, "{path}: {body}");
+        assert_eq!(body["error"]["code"], "image_generation_disabled");
+    }
+}
+
+#[tokio::test]
+async fn each_websocket_create_enforces_image_policy() {
+    let app = api_router(execution()).await;
+    let listener = tokio::net::TcpListener::bind("127.0.0.1:0").await.unwrap();
+    let address = listener.local_addr().unwrap();
+    let server = tokio::spawn(async move {
+        axum::serve(listener, app).await.unwrap();
+    });
+    let mut request = format!("ws://{address}/v1/responses")
+        .into_client_request()
+        .unwrap();
+    request
+        .headers_mut()
+        .insert(AUTHORIZATION, "Bearer sk_image_policy".parse().unwrap());
+    let (mut socket, _) = tokio_tungstenite::connect_async(request).await.unwrap();
+    for image_fields in [
+        json!({"tools":[{"type":"image_generation"}]}),
+        json!({"tool_choice":{"type":"image_generation"}}),
+    ] {
+        let mut body = json!({"type":"response.create","model":"model-a","input":"test"});
+        body.as_object_mut()
+            .unwrap()
+            .extend(image_fields.as_object().unwrap().clone());
+        socket
+            .send(Message::Text(body.to_string().into()))
+            .await
+            .unwrap();
+        let frame = tokio::time::timeout(Duration::from_secs(2), socket.next())
+            .await
+            .unwrap()
+            .unwrap()
+            .unwrap();
+        let body: Value = serde_json::from_str(frame.to_text().unwrap()).unwrap();
+        assert_eq!(body["type"], "error", "{body}");
+        assert_eq!(body["status"], 403);
+        assert_eq!(body["error"]["code"], "image_generation_disabled");
+    }
+    socket.close(None).await.unwrap();
+    server.abort();
+}

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

@@ -3,6 +3,7 @@
 mod auth;
 mod endpoint;
 mod error;
+mod image_policy;
 mod images;
 mod live;
 mod middleware;
@@ -292,6 +293,16 @@ fn snapshot_with_accounts(
     provider_name: &str,
     client_key_id: &str,
     accounts: &[&str],
+) -> RuntimeSnapshot {
+    snapshot_with_image_policy(plaintext, provider_name, client_key_id, accounts, true)
+}
+
+fn snapshot_with_image_policy(
+    plaintext: &str,
+    provider_name: &str,
+    client_key_id: &str,
+    accounts: &[&str],
+    allow_image_generation: bool,
 ) -> RuntimeSnapshot {
     let provider = ProviderKind::new(provider_name).expect("provider");
     let account_directory = Arc::new(RuntimeAccountDirectory::new(BTreeMap::from_iter(
@@ -323,10 +334,13 @@ fn snapshot_with_accounts(
         vec![ClientPolicy::new(
             ClientApiKeyId::new(client_key_id).expect("key ID"),
             PlaintextClientApiKey::new(plaintext).expect("plaintext key"),
-            Arc::new(FrozenAccountScope::new(
-                Arc::clone(&account_directory),
-                ClientRoutingScope::all_accounts(),
-            )),
+            Arc::new(
+                FrozenAccountScope::new(
+                    Arc::clone(&account_directory),
+                    ClientRoutingScope::all_accounts(),
+                )
+                .with_image_generation(allow_image_generation),
+            ),
             true,
             RateLimits::unlimited(),
         )],

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

@@ -34,6 +34,49 @@ fn generate_request(body: Value) -> DecodedResponsesRequest {
     decode_request(body.to_string().as_bytes()).expect("test request should decode")
 }
 
+#[test]
+fn decoder_should_attach_only_identified_codex_inference_turns() {
+    let body = json!({"model":"smart-code", "input":"hello", "client_metadata": {
+        "thread_id":"thread", "turn_id":"turn", "x-codex-turn-metadata": r#"{"request_kind":"turn"}"#
+    }});
+    let decoded = generate_request(body.clone());
+    let identity = generate_operation(&decoded).client_turn().unwrap();
+    assert_eq!(identity.thread_id(), "thread");
+    assert_eq!(identity.turn_id(), "turn");
+    assert_eq!(Value::Object(openai_wire_body(&decoded).clone()), body);
+    for variant in [
+        "warmup",
+        "compaction",
+        "missing_thread",
+        "missing_turn",
+        "unknown_kind",
+        "oversized_id",
+    ] {
+        let mut request = body.clone();
+        match variant {
+            "warmup" => request["generate"] = json!(false),
+            "compaction" => {
+                request["client_metadata"]["x-codex-turn-metadata"] =
+                    json!(r#"{"request_kind":"compaction"}"#)
+            }
+            "missing_thread" => request["client_metadata"]["thread_id"] = json!(""),
+            "missing_turn" => request["client_metadata"]["turn_id"] = json!(""),
+            "unknown_kind" => {
+                request["client_metadata"]["x-codex-turn-metadata"] =
+                    json!(r#"{"request_kind":"other"}"#)
+            }
+            "oversized_id" => request["client_metadata"]["turn_id"] = json!("x".repeat(257)),
+            _ => unreachable!(),
+        }
+        assert!(
+            generate_operation(&generate_request(request))
+                .client_turn()
+                .is_none(),
+            "{variant}"
+        );
+    }
+}
+
 fn generate_operation(
     decoded: &DecodedResponsesRequest,
 ) -> &gateway_core::operation::GenerateRequest {

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

@@ -1,10 +1,192 @@
 //! 验证未知控制消息、迟到中断、结算期间收发和创建请求的串行准入
 
 use super::*;
+use gateway_core::engine::client_retry::{ClientRetryKey, ClientRetryStore};
 use gateway_core::engine::response_control::{
     ResponseControlTransport, ResponseControlUnavailable,
 };
 
+#[derive(Default)]
+struct RetryMarker(Mutex<Option<ModelRequestId>>);
+
+impl ClientRetryStore for RetryMarker {
+    fn load<'a>(
+        &'a self,
+        _: &'a ClientRetryKey,
+    ) -> BoxFuture<'a, Result<Option<ModelRequestId>, StoreError>> {
+        Box::pin(async { Ok(self.0.lock().unwrap().clone()) })
+    }
+    fn record_failure<'a>(
+        &'a self,
+        _: &'a ClientRetryKey,
+        version: &'a ModelRequestId,
+        _: Duration,
+    ) -> BoxFuture<'a, Result<(), StoreError>> {
+        Box::pin(async move {
+            *self.0.lock().unwrap() = Some(version.clone());
+            Ok(())
+        })
+    }
+    fn clear_failure<'a>(
+        &'a self,
+        _: &'a ClientRetryKey,
+        expected: &'a ModelRequestId,
+    ) -> BoxFuture<'a, Result<(), StoreError>> {
+        Box::pin(async move {
+            let mut marker = self.0.lock().unwrap();
+            if marker.as_ref() == Some(expected) {
+                *marker = None;
+            }
+            Ok(())
+        })
+    }
+}
+
+#[derive(Default)]
+struct RetryBudgetProvider(AtomicUsize);
+
+#[async_trait]
+impl Provider for RetryBudgetProvider {
+    fn name(&self) -> &'static str {
+        "openai"
+    }
+    fn catalog_generation(&self) -> ProviderCatalogGeneration {
+        ProviderCatalogGeneration::default()
+    }
+    async fn query_model_capabilities(
+        &self,
+    ) -> Result<Vec<ProviderModelCapabilities>, ProviderError> {
+        Ok(Vec::new())
+    }
+    async fn execute(
+        self: Arc<Self>,
+        request: ProviderRequest,
+        context: AttemptContext,
+    ) -> Result<ProviderStream, ProviderError> {
+        let number = self.0.fetch_add(1, Ordering::SeqCst) + 1;
+        if number <= 7 || matches!(number, 9 | 10) {
+            let error = ProviderError::new(ProviderErrorKind::Unavailable, UpstreamSendState::Sent)
+                .with_status(503)
+                .with_replay_safe()
+                .with_client_visible_upstream_response(gateway_core::error::ClientVisibleUpstreamResponse::new(
+                    503, Some(b"application/json".to_vec()),
+                    Bytes::from_static(br#"{"error":{"code":"server_error","message":"synthetic unavailable"}}"#),
+                ))
+                .with_transient_retry(
+                    std::num::NonZeroU32::new(5).unwrap(),
+                    Duration::ZERO,
+                    Duration::ZERO,
+                    Duration::ZERO,
+                    gateway_core::error::TransientRetryPolicy::RetryThenReturnError,
+                );
+            return Ok(ProviderStream::new(
+                ProviderCallMetadata::new(
+                    request.candidate().provider().clone(),
+                    request.candidate().upstream_model().unwrap().clone(),
+                    ProviderAccountId::new("acct_api_test").unwrap(),
+                    UpstreamTransport::new("websocket").unwrap(),
+                ),
+                futures::stream::iter([Err(error)]),
+                (),
+            ));
+        }
+        Arc::new(ChargedWebSocketProvider)
+            .execute(request, context)
+            .await
+    }
+}
+
+#[tokio::test]
+async fn client_retry_budget_should_flow_through_http_and_live_websocket() {
+    for websocket in [false, true] {
+        let provider = Arc::new(RetryBudgetProvider::default());
+        let marker = Arc::new(RetryMarker::default());
+        let execution = Arc::new(
+            DefaultExecutionService::new(
+                RuntimeSnapshotHandle::new(snapshot("sk_retry_probe", "openai")),
+                Arc::new(SettlementPorts::default()),
+                ProviderRegistry::new([provider.clone() as Arc<dyn Provider>]).unwrap(),
+                Arc::new(InterruptAdmissions::default()),
+                Arc::new(UnusedContinuation),
+                Arc::new(IgnoredClientApiKeyUsage),
+                Arc::new(crate::support::RecordingDiagnostics::default()),
+            )
+            .with_client_retry_store(marker.clone()),
+        );
+        let app = api_router(execution).await;
+        let mut body = json!({"model":"model-a", "input":"hello", "stream":true, "client_metadata": {
+            "thread_id":"thread", "turn_id":"turn", "x-codex-turn-metadata":r#"{"request_kind":"turn"}"#
+        }});
+        let mut socket = None;
+        let mut server = None;
+        if websocket {
+            let listener = tokio::net::TcpListener::bind("127.0.0.1:0").await.unwrap();
+            let address = listener.local_addr().unwrap();
+            let app = app.clone();
+            server = Some(Server(tokio::spawn(async move {
+                axum::serve(listener, app).await.unwrap()
+            })));
+            let mut request = format!("ws://{address}/v1/responses")
+                .into_client_request()
+                .unwrap();
+            request.headers_mut().insert(
+                AUTHORIZATION,
+                HeaderValue::from_static("Bearer sk_retry_probe"),
+            );
+            socket = Some(connect_async(request).await.unwrap().0);
+            body["type"] = json!("response.create");
+        }
+        for (expected_attempts, success) in [(6, false), (7, false), (8, true), (11, true)] {
+            if let Some(socket) = &mut socket {
+                send(socket, body.clone()).await;
+                loop {
+                    let event = next(socket).await;
+                    match event["type"].as_str() {
+                        Some("response.completed") => {
+                            assert!(success);
+                            break;
+                        }
+                        Some("response.failed" | "error") => {
+                            assert!(!success);
+                            break;
+                        }
+                        _ => {}
+                    }
+                }
+            } else {
+                let request = Request::builder()
+                    .method("POST")
+                    .uri("/v1/responses")
+                    .header(AUTHORIZATION, "Bearer sk_retry_probe")
+                    .header("content-type", "application/json")
+                    .body(Body::from(body.to_string()))
+                    .unwrap();
+                let response = app.clone().oneshot(request).await.unwrap();
+                assert_eq!(
+                    response.status(),
+                    if success {
+                        StatusCode::OK
+                    } else {
+                        StatusCode::SERVICE_UNAVAILABLE
+                    }
+                );
+                let payload = axum::body::to_bytes(response.into_body(), usize::MAX)
+                    .await
+                    .unwrap();
+                if success {
+                    assert!(String::from_utf8_lossy(&payload).contains("response.completed"));
+                }
+            }
+            assert_eq!(provider.0.load(Ordering::SeqCst), expected_attempts);
+            assert_eq!(marker.0.lock().unwrap().is_none(), success);
+        }
+        if let Some(mut socket) = socket {
+            socket.close(None).await.unwrap();
+        }
+        drop(server);
+    }
+}
+
 struct ControlTransport {
     id: String,
     active: std::sync::atomic::AtomicBool,

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

@@ -8,6 +8,30 @@ use serde_json::json;
 
 use super::decode_response_create;
 
+#[test]
+fn client_retry_identity_should_follow_each_frame_instead_of_handshake_metadata() {
+    let mut headers = HeaderMap::new();
+    headers.insert(
+        "x-codex-turn-metadata",
+        HeaderValue::from_static(
+            r#"{"thread_id":"old-thread","turn_id":"old-turn","request_kind":"compaction"}"#,
+        ),
+    );
+    let headers = OpenAiRequestHeaders::from_headers(&headers);
+    for turn in ["turn-one", "turn-two"] {
+        let decoded = decode_response_create_with_context(&json!({
+            "type":"response.create", "model":"smart-code", "input":"hello",
+            "client_metadata": {"thread_id":"thread", "turn_id":turn, "x-codex-turn-metadata":r#"{"request_kind":"turn"}"#}
+        }).to_string(), &headers).unwrap();
+        let Operation::Generate(request) = decoded.operation() else {
+            panic!("generate");
+        };
+        let identity = request.client_turn().unwrap();
+        assert_eq!(identity.thread_id(), "thread");
+        assert_eq!(identity.turn_id(), turn);
+    }
+}
+
 #[test]
 fn response_create_should_carry_source_headers_to_provider_on_every_frame() {
     let mut headers = HeaderMap::new();

backend/crates/gateway-core/src/account/scope.rs

@@ -247,6 +247,7 @@ impl ClientRoutingScope {
 /// 一次认证随 RuntimeSnapshot 冻结的账号目录与 Key 权限
 #[derive(Debug, Clone, PartialEq, Eq)]
 pub struct FrozenAccountScope {
+    allow_image_generation: bool,
     fast_mode: FastMode,
     request_profiles: BTreeMap<ProviderKind, super::OpaqueProviderData>,
     directory: Arc<RuntimeAccountDirectory>,
@@ -257,6 +258,18 @@ pub struct FrozenAccountScope {
 }
 
 impl FrozenAccountScope {
+    /// Key 绑定分组合并后的生图权限，与账号成员资格一同冻结
+    #[must_use]
+    pub const fn with_image_generation(mut self, allowed: bool) -> Self {
+        self.allow_image_generation = allowed;
+        self
+    }
+
+    #[must_use]
+    pub const fn allows_image_generation(&self) -> bool {
+        self.allow_image_generation
+    }
+
     /// 与 Key 授权范围一同冻结；具体字段只由对应 Provider 解释
     #[must_use]
     pub fn with_request_profiles(
@@ -296,6 +309,7 @@ impl FrozenAccountScope {
             ClientRoutingScope::Restricted { provider_kinds, .. } => Arc::clone(provider_kinds),
         };
         Self {
+            allow_image_generation: true,
             fast_mode: FastMode::Default,
             request_profiles: BTreeMap::new(),
             directory,

backend/crates/gateway-core/src/account/selection.rs

@@ -510,6 +510,8 @@ pub struct AccountSelectionContext {
     pub policy: AccountSelectionPolicy,
     pub now: SystemTime,
     pub excluded_accounts: BTreeSet<ProviderAccountId>,
+    /// 租约争用或队列让位只阻止立即选号，不排除等待资格
+    pub busy_accounts: BTreeSet<ProviderAccountId>,
     pub preferred_account: Option<ProviderAccountId>,
     /// 优先账号可调度时，是否优先于其它账号的持久权重
     pub preferred_account_overrides_weight: bool,
@@ -602,7 +604,7 @@ impl<'a> AccountSelection<'a> {
 pub struct AccountSelector;
 
 impl AccountSelector {
-    /// 返回调度策略可见的全部合格候选；权重层授权由调用策略的宿主适配器裁剪
+    /// 策略同时约束立即选号与等待，候选保留仅因本地容量受阻的账号
     #[must_use]
     pub(crate) fn policy_candidates<'a>(
         &self,
@@ -611,7 +613,12 @@ impl AccountSelector {
     ) -> Vec<&'a AccountCandidate> {
         candidates
             .iter()
-            .filter(|candidate| self.scheduling_blocker(candidate, context).is_none())
+            .filter(|candidate| {
+                matches!(
+                    self.scheduling_blocker(candidate, context),
+                    None | Some(AccountSchedulingBlocker::ConcurrencyLimit)
+                )
+            })
             .collect()
     }
 
@@ -697,10 +704,7 @@ impl AccountSelector {
             .max()?;
         let (mut preferred, preferred_candidate) = self.preferred_decision(candidates, context);
         if let Some(candidate) = preferred_candidate {
-            // 权重回切只裁决内置策略的软亲和，插件显式选号按实际选择记录结果
-            let prefer_higher_weight = context.policy.strategy() == RotationStrategy::Smart
-                && context.policy.smart_scheduling().prefer_higher_weight();
-            if (!context.preferred_account_overrides_weight || prefer_higher_weight)
+            if !context.preferred_account_overrides_weight
                 && candidate.account.weight() < highest_weight
             {
                 preferred =
@@ -860,6 +864,9 @@ impl AccountSelector {
         if context.excluded_accounts.contains(candidate.account.id()) {
             return Some(AccountSchedulingBlocker::Excluded);
         }
+        if context.busy_accounts.contains(candidate.account.id()) {
+            return Some(AccountSchedulingBlocker::ConcurrencyLimit);
+        }
         if context
             .concurrency_limit(&candidate.account)
             .limit()

backend/crates/gateway-core/src/account/smart_scheduling.rs

@@ -6,7 +6,6 @@ use serde::{Deserialize, Serialize};
 #[serde(into = "SmartSchedulingValues")]
 pub struct SmartSchedulingConfig {
     weights: [u8; 6],
-    prefer_higher_weight: bool,
 }
 
 impl<'de> Deserialize<'de> for SmartSchedulingConfig {
@@ -24,7 +23,6 @@ impl Default for SmartSchedulingConfig {
     fn default() -> Self {
         Self {
             weights: [10, 8, 10, 5, 0, 0],
-            prefer_higher_weight: false,
         }
     }
 }
@@ -39,10 +37,7 @@ pub enum SmartSchedulingConfigError {
 
 impl SmartSchedulingConfig {
     /// 顺序为负载、剩余额度、健康、首输出延迟、额度重置和排队压力
-    pub fn new(
-        weights: [f64; 6],
-        prefer_higher_weight: bool,
-    ) -> Result<Self, SmartSchedulingConfigError> {
+    pub fn new(weights: [f64; 6]) -> Result<Self, SmartSchedulingConfigError> {
         let mut tenths = [0; 6];
         for (index, field) in [
             "loadWeight",
@@ -67,10 +62,7 @@ impl SmartSchedulingConfig {
         if tenths == [0; 6] {
             return Err(SmartSchedulingConfigError::EmptyWeights);
         }
-        Ok(Self {
-            weights: tenths,
-            prefer_higher_weight,
-        })
+        Ok(Self { weights: tenths })
     }
 
     #[must_use]
@@ -78,11 +70,6 @@ impl SmartSchedulingConfig {
         self.weights.map(|weight| f64::from(weight) / 10.0)
     }
 
-    #[must_use]
-    pub const fn prefer_higher_weight(self) -> bool {
-        self.prefer_higher_weight
-    }
-
     pub(crate) fn score_tolerance(self) -> f64 {
         // 与系数同步缩放，避免仅放大相同比例的系数就改变近似最优候选集合
         // 排队系数只参与入队选择，不能放大立即选号的候选容差
@@ -105,24 +92,20 @@ struct SmartSchedulingValues {
     latency_weight: f64,
     reset_weight: f64,
     queue_weight: f64,
-    prefer_higher_weight: bool,
 }
 
 impl TryFrom<SmartSchedulingValues> for SmartSchedulingConfig {
     type Error = SmartSchedulingConfigError;
 
     fn try_from(value: SmartSchedulingValues) -> Result<Self, Self::Error> {
-        Self::new(
-            [
-                value.load_weight,
-                value.quota_weight,
-                value.health_weight,
-                value.latency_weight,
-                value.reset_weight,
-                value.queue_weight,
-            ],
-            value.prefer_higher_weight,
-        )
+        Self::new([
+            value.load_weight,
+            value.quota_weight,
+            value.health_weight,
+            value.latency_weight,
+            value.reset_weight,
+            value.queue_weight,
+        ])
     }
 }
 
@@ -143,7 +126,6 @@ impl From<SmartSchedulingConfig> for SmartSchedulingValues {
             latency_weight,
             reset_weight,
             queue_weight,
-            prefer_higher_weight: value.prefer_higher_weight,
         }
     }
 }

backend/crates/gateway-core/src/diagnostics/trace.rs

@@ -132,6 +132,15 @@ impl TraceContext {
         }
         let data = json!({
             "kind": error.kind().as_str(), "sendState": error.send_state().as_str(),
+            "recovery": {
+                "replaySafe": error.replay_is_safe(),
+                "retryProhibited": error.retry_is_prohibited(),
+                "sameAccountCredentialRetry": error.retries_same_account(),
+                "intent": error.pre_delivery_retry().map(|intent| format!("{intent:?}")),
+                "continuationFailure": error.continuation_failure().map(|failure| format!("{failure:?}")),
+                "continuationDisposition": error.continuation_recovery_disposition().map(|disposition| format!("{disposition:?}")),
+                "retryAfterMs": error.retry_after().map(|delay| delay.as_millis()),
+            },
             "diagnostic": error.diagnostic().map(|diagnostic| json!({
                 "stage": diagnostic.stage(), "code": diagnostic.code(), "message": diagnostic.as_str(),
             })),
@@ -291,6 +300,27 @@ impl TraceContext {
                 }
                 summary["upstreamStatus"] =
                     json!(data.get("upstreamStatus").and_then(Value::as_u64));
+                if let Some(recovery) = data.get("recovery") {
+                    let mut kept = json!({});
+                    for key in [
+                        "replaySafe",
+                        "retryProhibited",
+                        "sameAccountCredentialRetry",
+                    ] {
+                        kept[key] = json!(recovery.get(key).and_then(Value::as_bool));
+                    }
+                    for key in ["intent", "continuationFailure", "continuationDisposition"] {
+                        kept[key] = json!(
+                            recovery
+                                .get(key)
+                                .and_then(Value::as_str)
+                                .map(|value| bounded(value, 256))
+                        );
+                    }
+                    kept["retryAfterMs"] =
+                        json!(recovery.get("retryAfterMs").and_then(Value::as_u64));
+                    summary["recovery"] = kept;
+                }
                 if let Some(diagnostic) = data.get("diagnostic").and_then(Value::as_object) {
                     let mut kept = json!({});
                     for (key, limit) in [("stage", 96), ("code", 96), ("message", 256)] {
@@ -359,6 +389,7 @@ impl TraceContext {
                             | "account.selected"
                             | "attempt.failed"
                             | "retry.decided"
+                            | "retry.policy"
                             | "upstream.close"
                             | "upstream.read.failed"
                             | "capture.gap"

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

@@ -0,0 +1,124 @@
+//! 客户端失败轮次的短期重试抑制与终态前状态更新
+
+use std::{sync::Arc, time::Duration};
+
+use futures::{FutureExt, future::BoxFuture, pin_mut, select_biased};
+use futures_timer::Delay;
+
+use super::ModelRequestId;
+use crate::{error::StoreError, operation::ClientTurn, policy::ClientApiKeyId};
+
+const FAILURE_TTL: Duration = Duration::from_secs(300);
+const STORE_TIMEOUT: Duration = Duration::from_millis(100);
+
+/// 不同客户端不能共享轮次失败状态
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub struct ClientRetryKey {
+    pub client_key_id: ClientApiKeyId,
+    pub turn: ClientTurn,
+}
+
+/// 可丢失的失败标记，版本使用创建该标记的模型请求 ID
+pub trait ClientRetryStore: Send + Sync {
+    fn load<'a>(
+        &'a self,
+        key: &'a ClientRetryKey,
+    ) -> BoxFuture<'a, Result<Option<ModelRequestId>, StoreError>>;
+
+    fn record_failure<'a>(
+        &'a self,
+        key: &'a ClientRetryKey,
+        version: &'a ModelRequestId,
+        ttl: Duration,
+    ) -> BoxFuture<'a, Result<(), StoreError>>;
+
+    /// 比较与删除必须原子执行，迟到成功不能删除较新的失败
+    fn clear_failure<'a>(
+        &'a self,
+        key: &'a ClientRetryKey,
+        expected: &'a ModelRequestId,
+    ) -> BoxFuture<'a, Result<(), StoreError>>;
+}
+
+pub(super) struct ClientRetryGuard {
+    store: Arc<dyn ClientRetryStore>,
+    key: ClientRetryKey,
+    request_id: ModelRequestId,
+    observed: Option<ModelRequestId>,
+    settled: bool,
+}
+
+impl ClientRetryGuard {
+    pub(super) async fn load(
+        store: Arc<dyn ClientRetryStore>,
+        key: ClientRetryKey,
+        request_id: ModelRequestId,
+    ) -> Self {
+        let observed = bounded("load", &request_id, store.load(&key))
+            .await
+            .flatten();
+        Self {
+            store,
+            key,
+            request_id,
+            observed,
+            settled: false,
+        }
+    }
+
+    pub(super) fn has_failure(&self) -> bool {
+        self.observed.is_some()
+    }
+
+    pub(super) async fn failure(&mut self) {
+        if self.settled {
+            return;
+        }
+        bounded(
+            "record",
+            &self.request_id,
+            self.store
+                .record_failure(&self.key, &self.request_id, FAILURE_TTL),
+        )
+        .await;
+        self.settled = true;
+    }
+
+    pub(super) async fn success(&mut self) {
+        if self.settled {
+            return;
+        }
+        if let Some(observed) = &self.observed {
+            bounded(
+                "clear",
+                &self.request_id,
+                self.store.clear_failure(&self.key, observed),
+            )
+            .await;
+        }
+        self.settled = true;
+    }
+}
+
+async fn bounded<T>(
+    action: &'static str,
+    request_id: &ModelRequestId,
+    operation: BoxFuture<'_, Result<T, StoreError>>,
+) -> Option<T> {
+    let operation = operation.fuse();
+    let timeout = Delay::new(STORE_TIMEOUT).fuse();
+    pin_mut!(operation, timeout);
+    select_biased! {
+        result = operation => match result {
+            Ok(value) => Some(value),
+            Err(error) => {
+                tracing::warn!(action, request_id = request_id.as_str(), error_kind = ?error.kind(), "客户端重试标记不可用，保留原请求结果与重试策略");
+                None
+            }
+        },
+        () = timeout => {
+            tracing::warn!(action, request_id = request_id.as_str(), "客户端重试标记操作超时，保留原请求结果与重试策略");
+            None
+        },
+    }
+}

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

@@ -9,6 +9,7 @@ use std::num::NonZeroU32;
 use std::sync::Arc;
 use std::time::{Duration, Instant, SystemTime};
 
+use super::client_retry::{ClientRetryGuard, ClientRetryKey};
 use super::nested::ExecutionEffects;
 use super::observation::{
     RequestObservationDispatch, ResponseObservation, WebSocketResponseAttempt,
@@ -293,6 +294,20 @@ where
                 "admissionMs": request.admission_decision_ms,
                 "continuation": format!("{continuation_attempt:?}"),
                 "candidateCount": plan.candidates().len(),
+                "accountPolicy": {
+                    "strategy": plan.account_selection_policy().strategy().as_str(),
+                    "affinity": plan.account_selection_policy().openai_account_affinity().as_str(),
+                    "maxAccountRotations": plan.account_selection_policy().max_account_rotations(),
+                    "maxRoutingAttempts": plan.max_attempts().get(),
+                    "maxConcurrentPerAccount": plan.account_selection_policy().max_concurrent_per_account().get(),
+                    "queueMaxWaiting": plan.account_selection_policy().queue_policy().max_waiting,
+                    "queueTimeoutMs": plan.account_selection_policy().queue_policy().timeout.as_millis(),
+                    "sessionAffinityTtlMs": plan.account_selection_policy().openai_session_affinity_ttl().as_millis(),
+                    "smartWeights": plan.account_selection_policy().smart_scheduling().weights(),
+                },
+                "requiredAccount": account_selection.required_account().map(|account| account.as_str()),
+                "hasMiddleware": middleware.is_some(),
+                "hasRetryPolicy": request_policy.is_some(),
             }),
         );
         let image_generation_requested = operation.image_generation_requested();
@@ -358,6 +373,7 @@ where
             last_retryable_failure_events: Vec::new(),
             last_observed_provider: None,
             pending_terminal_failure: None,
+            client_retry: None,
         };
 
         if session.cancellation.is_cancelled() {
@@ -368,6 +384,26 @@ where
             session.finish_interruption(&EngineError::Deadline).await?;
             return Err(EngineError::Deadline);
         }
+        if !session.account_selection.is_diagnostic()
+            && let Some(store) = &self.engine.client_retry
+            && let Operation::Generate(generate) = &session.operation
+            && let Some(turn) = generate.client_turn()
+        {
+            let guard = ClientRetryGuard::load(
+                Arc::clone(store),
+                ClientRetryKey {
+                    client_key_id: session.client_api_key_ref.clone(),
+                    turn: turn.clone(),
+                },
+                session.request_id.clone(),
+            )
+            .await;
+            session.trace.record(
+                "client_retry.loaded",
+                json!({"previousFailure": guard.has_failure()}),
+            );
+            session.client_retry = Some(guard);
+        }
         Ok(session)
     }
 }
@@ -502,6 +538,7 @@ pub struct ResponseExecutionSession<S: ?Sized> {
     last_observed_provider: Option<crate::identity::ProviderKind>,
     /// 原子失败批次已交给协议层、但尚待下游提交后收敛的原 Provider 错误
     pending_terminal_failure: Option<PendingTerminalFailure>,
+    client_retry: Option<ClientRetryGuard>,
 }
 
 struct PendingTerminalFailure {
@@ -575,6 +612,7 @@ where
                     error,
                     send_state,
                 } => {
+                    self.record_client_retry_failure(&error).await;
                     let requirement = if self.delivery_committed {
                         CommitRequirement::AlreadyCommitted
                     } else {
@@ -912,7 +950,13 @@ where
                         .translate_native_response(event, terminal);
                     match translated {
                         Ok(events) if events.is_empty() => continue,
-                        Ok(events) => return Ok(PullOutcome::Events(events)),
+                        Ok(events) => {
+                            // 成功终态交付后客户端即可续发工具结果，不能等流 EOF 或异步观测再清理
+                            if terminal && let Some(guard) = &mut self.client_retry {
+                                guard.success().await;
+                            }
+                            return Ok(PullOutcome::Events(events));
+                        }
                         Err(error) => {
                             self.finish_provider_error(&error).await?;
                             return Err(provider_engine_error(error));
@@ -1170,6 +1214,17 @@ where
                             retry_allowed: continuation_retry || candidate_retry,
                         })
                         .await?;
+                    attempt_trace.record("retry.decided", json!({
+                        "phase": "prepare", "retryable": retryable,
+                        "continuationRetry": continuation_retry, "candidateRetry": candidate_retry,
+                        "retryProhibited": error.retry_is_prohibited(),
+                        "sendState": error.send_state().as_str(),
+                        "requestSendState": self.current_send_state().as_str(),
+                        "continuation": format!("{:?}", self.continuation_attempt),
+                        "accountRotations": self.account_rotations,
+                        "maxAccountRotations": self.plan.account_selection_policy().max_account_rotations(),
+                        "remainingRoutingAttempts": self.plan.max_attempts().get().saturating_sub(self.routing_attempts),
+                    }));
                     if retryable {
                         return Ok(Some(PullOutcome::AttemptDiscarded));
                     }
@@ -1582,22 +1637,32 @@ where
             && attempt_send_state != UpstreamSendState::Ambiguous
             && provider_proved_replay_safe
             && self.routing_attempts < self.plan.max_attempts().get();
+        let suppress_transient_retry = is_return_error_retry(&error)
+            && self
+                .client_retry
+                .as_ref()
+                .is_some_and(ClientRetryGuard::has_failure);
         let transient_retry = match error.pre_delivery_retry() {
             Some(crate::error::PreDeliveryRetry::SameAccountTransientRetry {
                 max_retries,
                 initial_delay,
                 max_delay,
-            }) if ordinary_retry => {
+                jitter,
+                ..
+            }) if ordinary_retry && !suppress_transient_retry => {
                 let retries = self
                     .transient_retry_counts
                     .entry(current.metadata.provider_account_id().clone())
                     .or_default();
                 if *retries < max_retries.get() {
                     let multiplier = 1_u32.checked_shl(*retries).unwrap_or(u32::MAX);
-                    // 服务器建议优先于本地退避，不能被本地上限缩短或再次指数放大
-                    let delay = error
-                        .retry_after()
-                        .unwrap_or_else(|| initial_delay.saturating_mul(multiplier).min(max_delay));
+                    // 服务器建议优先于本地退避，不截断、不放大，也不叠加随机等待
+                    let delay = error.retry_after().unwrap_or_else(|| {
+                        initial_delay
+                            .saturating_mul(multiplier)
+                            .min(max_delay)
+                            .saturating_add(jitter)
+                    });
                     *retries = retries.saturating_add(1);
                     Some(delay)
                 } else {
@@ -1620,9 +1685,11 @@ where
         // 同一排除分支。换号预算只在这里消耗；同账号分支（瞬态退避、传输恢复、
         // 凭据恢复重放、continuation 精确重连）不消耗。不能直接把预算门写进
         // ordinary_retry：它是同账号瞬态退避的前置条件，会被连带误伤
+        // RetryThenReturnError 耗尽后直接交付错误，不能落入 ordinary_retry 的换号兜底
         let rotation_retry = !continuation_retry
             && !same_account_retry
             && transient_retry.is_none()
+            && !is_return_error_retry(&error)
             && transport_recovery.is_none()
             && (ordinary_retry || account_rotation_retry)
             && self.account_rotations
@@ -1664,10 +1731,12 @@ where
         };
 
         self.trace.attempt(current.index.get()).record("retry.decided", json!({
+            "phase": "execute",
             "retryable": retryable, "continuationRetry": continuation_retry,
             "sameAccountRetry": same_account_retry, "accountRotationRetry": account_rotation_retry,
             "ordinaryRetry": ordinary_retry, "transportRecovery": transport_recovery.is_some(),
             "transientRetry": transient_retry.is_some(),
+            "clientRetrySuppressed": suppress_transient_retry,
             "rotationRetry": rotation_retry, "accountRotations": self.account_rotations,
             "connectionRetry": connection_retry_requested,
             "connectionRetries": self.connection_retries,
@@ -1676,6 +1745,17 @@ where
             "delayMs": transient_retry.or(transport_recovery.map(|(_, delay)| delay)).map(duration_ms),
             "downstreamCommitted": self.response_commit.committed_at().is_some(),
             "replayBarrierReached": self.replay_barrier_reached,
+            "deliveryPending": self.delivery_pending,
+            "providerProvedReplaySafe": provider_proved_replay_safe,
+            "retryProhibited": error.retry_is_prohibited(),
+            "requiredAccount": self.account_selection.required_account().map(|account| account.as_str()),
+            "continuation": format!("{:?}", self.continuation_attempt),
+            "connectionBudgetExhausted": self.connection_budget.exhausted(),
+            "remainingRoutingAttempts": self.plan.max_attempts().get().saturating_sub(self.routing_attempts),
+            "maxAccountRotations": self.plan.account_selection_policy().max_account_rotations(),
+            "transientRetriesUsed": self.transient_retry_counts.get(current.metadata.provider_account_id()).copied().unwrap_or_default(),
+            "credentialRecoveryAttempted": self.credential_recovery_attempted_accounts.contains(current.metadata.provider_account_id()),
+            "hasRetryPolicy": self.request_policy.is_some(),
             "sendState": format!("{attempt_send_state:?}"),
         }));
         if retryable {
@@ -1775,6 +1855,7 @@ where
             && !self.cancellation.is_cancelled()
             && !self.execution_effect_observed();
         let allowed = facts.retry_allowed;
+        let attempt_index = facts.attempt_index;
         let decision = {
             let cancellation = self.cancellation.clone();
             let cancelled = cancellation.cancelled().fuse();
@@ -1787,6 +1868,16 @@ where
                 decision = decision => Ok(decision),
             }
         }?;
+        self.trace.attempt(attempt_index.get()).record(
+            "retry.policy",
+            json!({
+                "hostAllowed": allowed,
+                "policyStopped": decision == super::policy::RetryDecision::Stop,
+                "executionEffectObserved": self.execution_effect_observed(),
+                "cancelled": self.cancellation.is_cancelled(),
+                "deadlineElapsed": self.deadline.is_elapsed(),
+            }),
+        );
         // 策略只收窄宿主已允许的恢复，返回后仍复核期限和外部副作用
         Ok(allowed
             && decision != super::policy::RetryDecision::Stop
@@ -1948,6 +2039,9 @@ where
         if self.is_finalized() {
             return Ok(());
         }
+        if let Some(guard) = &mut self.client_retry {
+            guard.success().await;
+        }
         let completed_at = SystemTime::now();
         self.observation.finish();
         let upstream_request_id = self
@@ -2022,6 +2116,7 @@ where
         error: &ProviderError,
         send_state: UpstreamSendState,
     ) -> Result<(), EngineError> {
+        self.record_client_retry_failure(error).await;
         let send_state = self.raise_send_watermark(send_state);
         let outcome = if error.kind() == ProviderErrorKind::Cancelled {
             ExecutionOutcome::Cancelled
@@ -2049,6 +2144,16 @@ where
         .await
     }
 
+    async fn record_client_retry_failure(&mut self, error: &ProviderError) {
+        // 只记录本策略的终态失败，不让其他错误改变客户端轮次的退避预算
+        if self.attempts > 0
+            && is_return_error_retry(error)
+            && let Some(guard) = &mut self.client_retry
+        {
+            guard.failure().await;
+        }
+    }
+
     async fn finish_interruption(&mut self, error: &EngineError) -> Result<(), EngineError> {
         // 结算 I/O 可能继续等待，先冻结交付事实以保持终态与实际 HTTP 状态一致
         self.response_commit.finish();
@@ -2343,6 +2448,16 @@ where
     }
 }
 
+fn is_return_error_retry(error: &ProviderError) -> bool {
+    matches!(
+        error.pre_delivery_retry(),
+        Some(crate::error::PreDeliveryRetry::SameAccountTransientRetry {
+            policy: crate::error::TransientRetryPolicy::RetryThenReturnError,
+            ..
+        })
+    )
+}
+
 async fn best_effort_store_write<T>(
     operation: &'static str,
     request_id: &ModelRequestId,

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

@@ -92,6 +92,7 @@ pub struct DefaultExecutionService {
     execution_extensions: Option<ExecutionExtensionIndex>,
     frontend_authentication: Option<FrontendAuthenticationExtensionIndex>,
     session_affinity: Option<Arc<dyn ProviderSessionAffinityPort>>,
+    client_retry: Option<Arc<dyn crate::engine::client_retry::ClientRetryStore>>,
     snapshot_refresh: Option<Arc<RuntimeSnapshotPublisher>>,
     active_requests: ActiveRequestRegistry,
 }
@@ -122,6 +123,7 @@ impl DefaultExecutionService {
             execution_extensions: None,
             frontend_authentication: None,
             session_affinity: None,
+            client_retry: None,
             snapshot_refresh: None,
             active_requests: Arc::default(),
         }
@@ -133,6 +135,15 @@ impl DefaultExecutionService {
         self
     }
 
+    #[must_use]
+    pub fn with_client_retry_store(
+        mut self,
+        store: Arc<dyn crate::engine::client_retry::ClientRetryStore>,
+    ) -> Self {
+        self.client_retry = Some(store);
+        self
+    }
+
     #[must_use]
     pub fn with_request_observers(mut self, observers: RequestObserverExtensionIndex) -> Self {
         self.request_observers = Some(observers);
@@ -389,6 +400,10 @@ impl DefaultExecutionService {
         started_at: SystemTime,
         request_id: Option<ModelRequestId>,
     ) -> Result<StartedExecution, GatewayError> {
+        crate::engine::image_generation::check_image_generation(
+            authorization.account_scope.allows_image_generation(),
+            &request.operation,
+        )?;
         let request_id = request_id.map_or_else(new_request_id, Ok)?;
         if authorization.cancellation.is_cancelled() {
             return Err(GatewayError::new(
@@ -829,8 +844,15 @@ impl DefaultExecutionService {
             started_at,
             deadline_at: authorization.deadline_at,
         };
-        let coordinator =
-            AttemptCoordinator::new(GatewayEngine::new(self.observations.clone(), providers));
+        let mut engine = GatewayEngine::new(self.observations.clone(), providers);
+        // 插件内部调用不共享外部客户端的失败轮次，探测另走 transient engine
+        if authorization.nested.is_none()
+            && authorization.bound.is_none()
+            && let Some(store) = &self.client_retry
+        {
+            engine = engine.with_client_retry_store(Arc::clone(store));
+        }
+        let coordinator = AttemptCoordinator::new(engine);
         let core = match coordinator
             .start_observed(
                 new_request,

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

@@ -0,0 +1,20 @@
+//! 冻结分组生图权限的请求与发送前门禁
+
+use crate::{
+    error::{GatewayError, GatewayErrorKind},
+    operation::Operation,
+};
+
+pub(super) fn check_image_generation(
+    allowed: bool,
+    operation: &Operation,
+) -> Result<(), GatewayError> {
+    if !allowed && operation.image_generation_requested() {
+        return Err(GatewayError::new(
+            GatewayErrorKind::PolicyDenied,
+            "Image generation is not allowed by the API key's account groups",
+        )
+        .with_client_code("image_generation_disabled"));
+    }
+    Ok(())
+}