chore(release): v3.27.0

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"
@@ -1681,6 +1710,7 @@ dependencies = [
  "gateway-core",
  "heck",
  "hex",
+ "hkdf",
  "prettyplease",
  "quote",
  "rand_core 0.6.4",
@@ -3749,9 +3779,9 @@ 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",
@@ -4742,6 +4772,7 @@ 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/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/Cargo.toml

@@ -21,6 +21,7 @@ futures = "0.3.32"
 gateway-core = { path = "../gateway-core" }
 heck = "0.5.0"
 hex = "0.4.3"
+hkdf = "0.13.0"
 rand_core = { version = "0.6.4", features = ["getrandom"] }
 serde = { version = "1.0.228", features = ["derive"] }
 serde_json = { version = "1.0.150", features = ["arbitrary_precision", "preserve_order"] }

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

@@ -405,9 +405,7 @@ async fn initialize_inner(
     let registry = providers;
 
     let auth = Arc::new(DefaultAuthService::new(
-        config.default_username,
-        config.session_ttl_minutes,
-        config.session_absolute_ttl_minutes,
+        &config,
         client_config.session_ttl_minutes,
         store.auth(),
         client_key_verifier.clone(),

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/auth.rs

@@ -76,18 +76,23 @@ impl std::fmt::Debug for LoginCommand {
     }
 }
 
-/// 登录成功后返回的会话事实
+/// 登录或刷新后仅向当前浏览器交付的会话凭据
 #[derive(Clone, PartialEq, Eq)]
-pub struct LoginResult {
+pub struct SessionGrant {
     pub session_id: String,
+    pub refresh_token: Option<String>,
     pub session: AuthSession,
 }
 
-impl std::fmt::Debug for LoginResult {
+impl std::fmt::Debug for SessionGrant {
     fn fmt(&self, formatter: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
         formatter
-            .debug_struct("LoginResult")
+            .debug_struct("SessionGrant")
             .field("session_id", &"[REDACTED]")
+            .field(
+                "refresh_token",
+                &self.refresh_token.as_ref().map(|_| "[REDACTED]"),
+            )
             .field("session", &self.session)
             .finish()
     }
@@ -138,12 +143,41 @@ impl std::fmt::Debug for ChangePassword {
     }
 }
 
-/// 会话有效期与管理员续期上限；没有上限字段的旧会话保持固定有效期
+/// 访问有效期与逻辑会话；没有续期凭据的会话保持固定期限
 #[derive(Debug, Clone, PartialEq, Eq)]
 pub struct AuthSession {
     pub subject: SessionSubject,
     pub expires_at: DateTime<Utc>,
     pub absolute_expires_at: Option<DateTime<Utc>>,
+    pub renewal: Option<SessionRenewal>,
+}
+
+/// 长效登录的凭据摘要与续期窗口；访问到期不删除整段登录
+#[derive(Debug, Clone, PartialEq, Eq, serde::Serialize, serde::Deserialize)]
+#[serde(deny_unknown_fields)]
+pub struct SessionRenewal {
+    pub expires_at: DateTime<Utc>,
+    pub access_digest: [u8; 32],
+    pub refresh_digest: [u8; 32],
+    pub previous: Option<SessionRefreshReplay>,
+}
+
+/// 仅保存上一代摘要及派生盐；响应丢失时须持有上一代凭据才能重建同一响应
+#[derive(Debug, Clone, PartialEq, Eq, serde::Serialize, serde::Deserialize)]
+#[serde(deny_unknown_fields)]
+pub struct SessionRefreshReplay {
+    pub digest: [u8; 32],
+    pub salt: [u8; 32],
+    pub expires_at: DateTime<Utc>,
+}
+
+impl AuthSession {
+    #[must_use]
+    pub fn retained_until(&self) -> DateTime<Utc> {
+        self.renewal
+            .as_ref()
+            .map_or(self.expires_at, |renewal| renewal.expires_at)
+    }
 }
 
 /// 安全审计事件类型

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/ports/store.rs

@@ -294,12 +294,12 @@ pub trait AuthStore: Send + Sync {
 
     async fn store_session(&self, session_id: &str, session: &AuthSession) -> AdminStoreResult<()>;
 
-    /// 只延长仍存在且匹配的会话；并发续期返回当前值，已退出或过期时不重建
-    async fn renew_session(
+    /// 原子替换或撤销仍匹配的会话并返回当前值；过期或退出后不重新创建
+    async fn compare_exchange_session(
         &self,
         session_id: &str,
         expected: &AuthSession,
-        expires_at: chrono::DateTime<chrono::Utc>,
+        replacement: Option<&AuthSession>,
     ) -> AdminStoreResult<Option<AuthSession>>;
 
     async fn delete_session(&self, session_id: &str) -> AdminStoreResult<Option<AuthSession>>;

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

@@ -0,0 +1,139 @@
+//! 控制面认证端口、会话策略与访问和续期共用的身份校验
+
+use std::{net::IpAddr, sync::Arc};
+
+use async_trait::async_trait;
+use base64::{Engine as _, engine::general_purpose::URL_SAFE_NO_PAD};
+use chrono::{Duration, Utc};
+use gateway_core::engine::execution::ClientKeyVerifier;
+use sha2::{Digest as _, Sha256};
+
+use super::map_store_error;
+use crate::{
+    model::{
+        AdminError,
+        auth::{
+            AuthSession, ChangePassword, LoginCommand, LoginError, SessionGrant, SessionSubject,
+        },
+    },
+    ports::store::AuthStore,
+};
+
+mod service;
+mod session;
+
+/// 所有控制面接口消费同一个会话服务，权限由服务端身份决定
+#[async_trait]
+pub trait AuthService: Send + Sync {
+    async fn change_password(
+        &self,
+        session_id: Option<&str>,
+        command: ChangePassword,
+        source_ip: IpAddr,
+    ) -> Result<(), AdminError>;
+    async fn ensure_default_admin(&self, password: &str) -> Result<bool, AdminError>;
+    async fn session(&self, session_id: Option<&str>) -> Result<Option<AuthSession>, AdminError>;
+    async fn renew_session(
+        &self,
+        refresh_token: Option<&str>,
+        access_token: Option<&str>,
+    ) -> Result<Option<SessionGrant>, AdminError>;
+    async fn resolve_admin_user_id(
+        &self,
+        session_id: Option<&str>,
+    ) -> Result<Option<String>, AdminError>;
+    async fn verify_admin_api_key(&self, key: &str) -> Result<bool, AdminError>;
+    async fn login(
+        &self,
+        command: LoginCommand,
+        source_ip: IpAddr,
+        previous_session_id: Option<&str>,
+    ) -> Result<SessionGrant, LoginError>;
+    async fn logout(&self, session_id: &str) -> Result<(), AdminError>;
+}
+
+const MAX_SESSION_TTL_MINUTES: i64 = 366 * 24 * 60;
+
+pub(crate) struct DefaultAuthService {
+    default_admin_user_id: String,
+    admin_access_ttl: Duration,
+    admin_session_ttl: Duration,
+    admin_session_absolute_ttl: Duration,
+    key_session_ttl: Duration,
+    store: Arc<dyn AuthStore>,
+    verifier: Arc<dyn ClientKeyVerifier>,
+}
+
+impl DefaultAuthService {
+    #[must_use]
+    pub(crate) fn new(
+        config: &crate::AdminConfig,
+        key_session_ttl_minutes: u64,
+        store: Arc<dyn AuthStore>,
+        verifier: Arc<dyn ClientKeyVerifier>,
+    ) -> Self {
+        let admin_session_ttl = session_ttl(config.session_ttl_minutes);
+        Self {
+            default_admin_user_id: config.default_username.clone(),
+            // 访问凭据最多有效 24 小时，较短的不活动期限也要留出续期窗口
+            admin_access_ttl: Duration::hours(24).min(admin_session_ttl / 4),
+            admin_session_ttl,
+            admin_session_absolute_ttl: session_ttl(config.session_absolute_ttl_minutes),
+            key_session_ttl: session_ttl(key_session_ttl_minutes),
+            store,
+            verifier,
+        }
+    }
+
+    async fn load_live_session(&self, id: &str) -> Result<Option<AuthSession>, AdminError> {
+        let Some(session) = self
+            .store
+            .load_session(id)
+            .await
+            .map_err(|error| map_store_error(error, "authentication session"))?
+        else {
+            return Ok(None);
+        };
+        if session.retained_until() <= Utc::now()
+            || session
+                .absolute_expires_at
+                .is_some_and(|expiry| expiry <= Utc::now())
+        {
+            return Ok(None);
+        }
+        let valid = match &session.subject {
+            SessionSubject::Admin {
+                admin_user_id,
+                credential_fingerprint,
+            } => self
+                .store
+                .load_password_hash(admin_user_id)
+                .await
+                .map_err(|error| map_store_error(error, "administrator session"))?
+                .is_some_and(|hash| password_fingerprint(&hash) == *credential_fingerprint),
+            SessionSubject::Key { client_key_id } => self
+                .store
+                .client_key_enabled(client_key_id)
+                .await
+                .map_err(|error| map_store_error(error, "session key"))?,
+        };
+        if !valid {
+            let _ = self.store.delete_session(id).await;
+            return Ok(None);
+        }
+        Ok(Some(session))
+    }
+}
+
+fn session_ttl(minutes: u64) -> Duration {
+    Duration::minutes(
+        i64::try_from(minutes)
+            .unwrap_or(MAX_SESSION_TTL_MINUTES)
+            .clamp(1, MAX_SESSION_TTL_MINUTES),
+    )
+}
+
+// 指纹只绑定已加盐的密码哈希，不把密码或原始哈希复制到 Redis 会话
+fn password_fingerprint(password_hash: &str) -> String {
+    URL_SAFE_NO_PAD.encode(Sha256::digest(password_hash.as_bytes()))
+}

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

@@ -1,93 +1,35 @@
-//! 控制面登录、会话恢复与身份校验的唯一 owner
+//! 管理员和 Key 的登录、退出、密码策略与会话访问
 
-use std::{net::IpAddr, sync::Arc, time::Duration as StdDuration};
+use std::{net::IpAddr, time::Duration as StdDuration};
 
 use argon2::{Argon2, PasswordHash, PasswordHasher, PasswordVerifier};
 use async_trait::async_trait;
 use base64::{Engine as _, engine::general_purpose::URL_SAFE_NO_PAD};
-use chrono::{Duration, Utc};
-use gateway_core::engine::execution::ClientKeyVerifier;
+use chrono::Utc;
 use rand_core::{OsRng, RngCore as _};
-use sha2::{Digest, Sha256};
 use subtle::ConstantTimeEq as _;
 use uuid::Uuid;
 
+use super::{
+    AuthService, DefaultAuthService, password_fingerprint,
+    session::{self, Credential, storage_id},
+};
 use crate::{
     model::{
         AdminError, AdminErrorKind,
         auth::{
             AdminAuditEvent, AuditActorKind, AuthSession, ChangePassword, LoginCommand, LoginError,
-            LoginResult, SessionSubject,
+            SessionGrant, SessionSubject,
         },
     },
-    ports::store::AuthStore,
+    use_case::map_store_error,
 };
 
-use super::map_store_error;
-
-/// 所有控制面接口消费同一个会话服务，权限由服务端身份决定
-#[async_trait]
-pub trait AuthService: Send + Sync {
-    async fn change_password(
-        &self,
-        session_id: Option<&str>,
-        command: ChangePassword,
-        source_ip: IpAddr,
-    ) -> Result<(), AdminError>;
-    async fn ensure_default_admin(&self, password: &str) -> Result<bool, AdminError>;
-    async fn session(&self, session_id: Option<&str>) -> Result<Option<AuthSession>, AdminError>;
-    async fn renew_session(
-        &self,
-        session_id: Option<&str>,
-    ) -> Result<Option<AuthSession>, AdminError>;
-    async fn resolve_admin_user_id(
-        &self,
-        session_id: Option<&str>,
-    ) -> Result<Option<String>, AdminError>;
-    async fn verify_admin_api_key(&self, key: &str) -> Result<bool, AdminError>;
-    async fn login(
-        &self,
-        command: LoginCommand,
-        source_ip: IpAddr,
-        previous_session_id: Option<&str>,
-    ) -> Result<LoginResult, LoginError>;
-    async fn logout(&self, session_id: &str) -> Result<(), AdminError>;
-}
-
-const MAX_SESSION_TTL_MINUTES: i64 = 366 * 24 * 60;
 const LOGIN_WINDOW: StdDuration = StdDuration::from_secs(60);
 const LOGIN_ATTEMPTS_PER_SOURCE: u32 = 10;
 const LOGIN_ATTEMPTS_GLOBAL: u32 = 200;
 
-pub(crate) struct DefaultAuthService {
-    default_admin_user_id: String,
-    admin_session_ttl: Duration,
-    admin_session_absolute_ttl: Duration,
-    key_session_ttl: Duration,
-    store: Arc<dyn AuthStore>,
-    verifier: Arc<dyn ClientKeyVerifier>,
-}
-
 impl DefaultAuthService {
-    #[must_use]
-    pub(crate) fn new(
-        default_admin_user_id: impl Into<String>,
-        admin_session_ttl_minutes: u64,
-        admin_session_absolute_ttl_minutes: u64,
-        key_session_ttl_minutes: u64,
-        store: Arc<dyn AuthStore>,
-        verifier: Arc<dyn ClientKeyVerifier>,
-    ) -> Self {
-        Self {
-            default_admin_user_id: default_admin_user_id.into(),
-            admin_session_ttl: session_ttl(admin_session_ttl_minutes),
-            admin_session_absolute_ttl: session_ttl(admin_session_absolute_ttl_minutes),
-            key_session_ttl: session_ttl(key_session_ttl_minutes),
-            store,
-            verifier,
-        }
-    }
-
     fn auth_audit(&self, action: &str, admin_user_id: &str) -> AdminAuditEvent {
         AdminAuditEvent {
             id: format!("audit_{}", Uuid::now_v7().simple()),
@@ -228,50 +170,23 @@ impl AuthService for DefaultAuthService {
     }
 
     async fn session(&self, session_id: Option<&str>) -> Result<Option<AuthSession>, AdminError> {
-        let Some(session_id) = session_id.filter(|value| !value.is_empty()) else {
+        let Some(raw) = session_id.filter(|value| !value.is_empty()) else {
             return Ok(None);
         };
-        let Some(session) = self
-            .store
-            .load_session(session_id)
-            .await
-            .map_err(|error| map_store_error(error, "authentication session"))?
-        else {
+        let Some(session) = self.load_live_session(storage_id(raw)).await? else {
             return Ok(None);
         };
-        if session.expires_at <= Utc::now()
-            || session
-                .absolute_expires_at
-                .is_some_and(|expiry| expiry <= Utc::now())
-        {
-            // Redis 的 TTL 负责过期清理，避免旧读取删除另一个请求刚续期的会话
+        if session.expires_at <= Utc::now() {
             return Ok(None);
         }
-        if let SessionSubject::Admin {
-            admin_user_id,
-            credential_fingerprint,
-        } = &session.subject
-        {
-            let password_hash = self
-                .store
-                .load_password_hash(admin_user_id)
-                .await
-                .map_err(|error| map_store_error(error, "administrator session"))?;
-            if password_hash
-                .is_none_or(|hash| password_fingerprint(&hash) != *credential_fingerprint)
-            {
-                let _ = self.store.delete_session(session_id).await;
+        if let Some(renewal) = &session.renewal {
+            let Some(credential) = Credential::parse(raw).filter(|value| !value.refresh) else {
+                return Ok(None);
+            };
+            if !credential.owns(renewal) {
                 return Ok(None);
             }
-        }
-        if let SessionSubject::Key { client_key_id } = &session.subject
-            && !self
-                .store
-                .client_key_enabled(client_key_id)
-                .await
-                .map_err(|error| map_store_error(error, "session key"))?
-        {
-            let _ = self.store.delete_session(session_id).await;
+        } else if Credential::parse(raw).is_some() {
             return Ok(None);
         }
         Ok(Some(session))
@@ -279,33 +194,25 @@ impl AuthService for DefaultAuthService {
 
     async fn renew_session(
         &self,
-        session_id: Option<&str>,
-    ) -> Result<Option<AuthSession>, AdminError> {
-        let Some(session_id) = session_id else {
-            return Ok(None);
-        };
-        let Some(session) = self.session(Some(session_id)).await? else {
-            return Ok(None);
-        };
-        // Key 与升级前的会话保留原有期限，不能凭续期获得新的最长登录窗口
-        let Some(absolute) = session
-            .absolute_expires_at
-            .filter(|_| matches!(session.subject, SessionSubject::Admin { .. }))
-        else {
-            return Ok(Some(session));
-        };
-        let now = Utc::now();
-        let expires_at = (now + self.admin_session_ttl).min(absolute);
-        let interval = (self.admin_session_ttl / 4).min(Duration::minutes(1));
-        if expires_at <= session.expires_at
-            || session.expires_at - now > self.admin_session_ttl - interval
-        {
-            return Ok(Some(session));
+        refresh_token: Option<&str>,
+        access_token: Option<&str>,
+    ) -> Result<Option<SessionGrant>, AdminError> {
+        if let Some(refresh) = refresh_token {
+            return self.refresh(refresh).await;
         }
-        self.store
-            .renew_session(session_id, &session, expires_at)
-            .await
-            .map_err(|error| map_store_error(error, "authentication session renewal"))
+        // Key 与升级前的单凭据仅返回仍有效的原会话，不授予新的续期能力
+        let Some(session) = self
+            .session(access_token)
+            .await?
+            .filter(|session| session.renewal.is_none())
+        else {
+            return Ok(None);
+        };
+        Ok(Some(SessionGrant {
+            session_id: access_token.unwrap_or_default().to_owned(),
+            refresh_token: None,
+            session,
+        }))
     }
 
     async fn resolve_admin_user_id(
@@ -346,7 +253,7 @@ impl AuthService for DefaultAuthService {
         command: LoginCommand,
         source_ip: IpAddr,
         previous_session_id: Option<&str>,
-    ) -> Result<LoginResult, LoginError> {
+    ) -> Result<SessionGrant, LoginError> {
         if let Some(retry_after) = self
             .store
             .consume_login_attempt(
@@ -370,14 +277,31 @@ impl AuthService for DefaultAuthService {
         let now = Utc::now();
         let absolute_expires_at = matches!(subject, SessionSubject::Admin { .. })
             .then_some(now + self.admin_session_absolute_ttl);
-        let session = AuthSession {
-            subject,
-            expires_at: (now + ttl).min(absolute_expires_at.unwrap_or(now + ttl)),
-            absolute_expires_at,
+        let refresh_expiry = (now + ttl).min(absolute_expires_at.unwrap_or(now + ttl));
+        let grant = if let Some(absolute) = absolute_expires_at {
+            session::issue(
+                subject,
+                (now + self.admin_access_ttl).min(refresh_expiry),
+                refresh_expiry,
+                absolute,
+            )
+            .map_err(|_| LoginError::Unavailable)?
+        } else {
+            SessionGrant {
+                session_id: random_session_token(),
+                refresh_token: None,
+                session: AuthSession {
+                    subject,
+                    expires_at: refresh_expiry,
+                    absolute_expires_at: None,
+                    renewal: None,
+                },
+            }
         };
-        let session_id = random_session_token();
+        let session_id = storage_id(&grant.session_id);
+        let session = &grant.session;
         self.store
-            .store_session(&session_id, &session)
+            .store_session(session_id, session)
             .await
             .map_err(|_| LoginError::Unavailable)?;
         if let SessionSubject::Admin { admin_user_id, .. } = &session.subject
@@ -387,26 +311,39 @@ impl AuthService for DefaultAuthService {
                 .await
                 .is_err()
         {
-            let _ = self.store.delete_session(&session_id).await;
+            let _ = self.store.delete_session(session_id).await;
             return Err(LoginError::Unavailable);
         }
         // 新身份验证成功后才撤销旧会话；撤销失败时不向浏览器提交新会话
         if let Some(previous) = previous_session_id.filter(|value| !value.is_empty())
             && self.logout(previous).await.is_err()
         {
-            let _ = self.store.delete_session(&session_id).await;
+            let _ = self.store.delete_session(session_id).await;
             return Err(LoginError::Unavailable);
         }
-        Ok(LoginResult {
-            session_id,
-            session,
-        })
+        Ok(grant)
     }
 
     async fn logout(&self, session_id: &str) -> Result<(), AdminError> {
+        let id = storage_id(session_id);
+        let Some(stored) = self
+            .store
+            .load_session(id)
+            .await
+            .map_err(|error| map_store_error(error, "authentication session"))?
+        else {
+            return Ok(());
+        };
+        if let Some(renewal) = &stored.renewal {
+            if !Credential::parse(session_id).is_some_and(|credential| credential.owns(renewal)) {
+                return Ok(());
+            }
+        } else if Credential::parse(session_id).is_some() {
+            return Ok(());
+        }
         let session = self
             .store
-            .delete_session(session_id)
+            .delete_session(id)
             .await
             .map_err(|error| map_store_error(error, "authentication session"))?;
         if let Some(AuthSession {
@@ -423,14 +360,6 @@ impl AuthService for DefaultAuthService {
     }
 }
 
-fn session_ttl(minutes: u64) -> Duration {
-    Duration::minutes(
-        i64::try_from(minutes)
-            .unwrap_or(MAX_SESSION_TTL_MINUTES)
-            .clamp(1, MAX_SESSION_TTL_MINUTES),
-    )
-}
-
 fn validate_new_password(password: &str) -> Result<(), AdminError> {
     if password.trim().chars().count() < 12
         || password.len() > 1024
@@ -470,8 +399,3 @@ fn valid_admin_api_key_shape(value: &str) -> bool {
         && value.starts_with("admin-")
         && value[6..].bytes().all(|byte| byte.is_ascii_hexdigit())
 }
-
-// 指纹只绑定已加盐的密码哈希，不把密码或原始哈希复制到 Redis 会话
-fn password_fingerprint(password_hash: &str) -> String {
-    URL_SAFE_NO_PAD.encode(Sha256::digest(password_hash.as_bytes()))
-}

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

@@ -0,0 +1,214 @@
+//! 双凭据会话的签发、摘要校验与短窗口重试，不持久化可直接使用的凭据
+
+use base64::{Engine as _, engine::general_purpose::URL_SAFE_NO_PAD};
+use chrono::{DateTime, Duration, Utc};
+use rand_core::{OsRng, RngCore as _};
+use sha2::{Digest as _, Sha256};
+use subtle::ConstantTimeEq as _;
+
+use super::DefaultAuthService;
+use crate::{
+    model::{
+        AdminError, AdminErrorKind,
+        auth::{AuthSession, SessionGrant, SessionRefreshReplay, SessionRenewal, SessionSubject},
+    },
+    use_case::map_store_error,
+};
+
+const REPLAY_WINDOW: Duration = Duration::seconds(30);
+
+pub(super) struct Credential<'a> {
+    pub id: &'a str,
+    pub secret: &'a str,
+    pub refresh: bool,
+}
+
+impl<'a> Credential<'a> {
+    pub fn parse(value: &'a str) -> Option<Self> {
+        let mut parts = value.split('.');
+        let refresh = match parts.next()? {
+            "a2" => false,
+            "r2" => true,
+            _ => return None,
+        };
+        let id = parts.next()?;
+        let secret = parts.next()?;
+        if parts.next().is_some() || !valid_secret(id) || !valid_secret(secret) {
+            return None;
+        }
+        Some(Self {
+            id,
+            secret,
+            refresh,
+        })
+    }
+
+    pub fn digest(&self) -> [u8; 32] {
+        digest(self.secret)
+    }
+
+    pub fn owns(&self, renewal: &SessionRenewal) -> bool {
+        if self.refresh {
+            matches_digest(&renewal.refresh_digest, &self.digest())
+                || renewal
+                    .previous
+                    .as_ref()
+                    .is_some_and(|previous| matches_digest(&previous.digest, &self.digest()))
+        } else {
+            matches_digest(&renewal.access_digest, &self.digest())
+        }
+    }
+}
+
+fn valid_secret(value: &str) -> bool {
+    value.len() == 43
+        && URL_SAFE_NO_PAD
+            .decode(value)
+            .is_ok_and(|bytes| bytes.len() == 32)
+}
+
+pub(super) fn storage_id(value: &str) -> &str {
+    Credential::parse(value).map_or(value, |credential| credential.id)
+}
+
+fn matches_digest(left: &[u8; 32], right: &[u8; 32]) -> bool {
+    bool::from(left.ct_eq(right))
+}
+
+fn digest(secret: &str) -> [u8; 32] {
+    Sha256::digest(secret.as_bytes()).into()
+}
+
+fn random_bytes() -> [u8; 32] {
+    let mut bytes = [0; 32];
+    OsRng.fill_bytes(&mut bytes);
+    bytes
+}
+
+// 使用标准 HKDF 的独立用途标签；只有盐和摘要留在 Store，重试仍须持有原始凭据
+fn derive(secret: &str, salt: Option<&[u8]>, purpose: &[u8]) -> Result<String, AdminError> {
+    let mut output = [0; 32];
+    hkdf::Hkdf::<Sha256>::new(salt, secret.as_bytes())
+        .expand(purpose, &mut output)
+        .map_err(|_| AdminError::internal("会话凭据派生失败"))?;
+    Ok(URL_SAFE_NO_PAD.encode(output))
+}
+
+fn access_secret(refresh: &str) -> Result<String, AdminError> {
+    derive(refresh, None, b"cpr-session-v2/access")
+}
+
+pub(super) fn issue(
+    subject: SessionSubject,
+    access_expiry: DateTime<Utc>,
+    refresh_expiry: DateTime<Utc>,
+    absolute: DateTime<Utc>,
+) -> Result<SessionGrant, AdminError> {
+    let id = URL_SAFE_NO_PAD.encode(random_bytes());
+    let secret = URL_SAFE_NO_PAD.encode(random_bytes());
+    let session = AuthSession {
+        subject,
+        expires_at: access_expiry,
+        absolute_expires_at: Some(absolute),
+        renewal: Some(SessionRenewal {
+            expires_at: refresh_expiry,
+            access_digest: digest(&access_secret(&secret)?),
+            refresh_digest: digest(&secret),
+            previous: None,
+        }),
+    };
+    grant(&id, &secret, session)
+}
+
+fn grant(id: &str, refresh: &str, session: AuthSession) -> Result<SessionGrant, AdminError> {
+    Ok(SessionGrant {
+        session_id: format!("a2.{id}.{}", access_secret(refresh)?),
+        refresh_token: Some(format!("r2.{id}.{refresh}")),
+        session,
+    })
+}
+
+impl DefaultAuthService {
+    pub(super) async fn refresh(&self, raw: &str) -> Result<Option<SessionGrant>, AdminError> {
+        let Some(credential) = Credential::parse(raw).filter(|value| value.refresh) else {
+            return Ok(None);
+        };
+        let mut current = self.load_live_session(credential.id).await?;
+        // CAS 冲突只重新解释同一请求的凭据，不对并发退出进行补建
+        for _ in 0..3 {
+            let Some(session) = current else {
+                return Ok(None);
+            };
+            let Some(renewal) = &session.renewal else {
+                return Ok(None);
+            };
+            let now = Utc::now();
+            if session.retained_until() <= now
+                || session.absolute_expires_at.is_none_or(|limit| limit <= now)
+            {
+                return Ok(None);
+            }
+            if matches_digest(&renewal.refresh_digest, &credential.digest()) {
+                // 其他标签页已经刷新或客户端重试时复用当前代，避免 Cookie 响应相互覆盖
+                if session.expires_at > now {
+                    return grant(credential.id, credential.secret, session).map(Some);
+                }
+                let absolute = session
+                    .absolute_expires_at
+                    .ok_or_else(|| AdminError::internal("缺少会话最长有效期"))?;
+                let salt = random_bytes();
+                let next_secret =
+                    derive(credential.secret, Some(&salt), b"cpr-session-v2/refresh")?;
+                let refresh_expiry = (now + self.admin_session_ttl).min(absolute);
+                let access_expiry = (now + self.admin_access_ttl).min(refresh_expiry);
+                let replacement = AuthSession {
+                    expires_at: access_expiry,
+                    renewal: Some(SessionRenewal {
+                        expires_at: refresh_expiry,
+                        access_digest: digest(&access_secret(&next_secret)?),
+                        refresh_digest: digest(&next_secret),
+                        previous: Some(SessionRefreshReplay {
+                            digest: credential.digest(),
+                            salt,
+                            expires_at: (now + REPLAY_WINDOW).min(access_expiry),
+                        }),
+                    }),
+                    ..session.clone()
+                };
+                current = self
+                    .store
+                    .compare_exchange_session(credential.id, &session, Some(&replacement))
+                    .await
+                    .map_err(|error| map_store_error(error, "authentication session rotation"))?;
+                if current.as_ref() == Some(&replacement) {
+                    return grant(credential.id, &next_secret, replacement).map(Some);
+                }
+            } else if let Some(previous) = renewal
+                .previous
+                .as_ref()
+                .filter(|previous| matches_digest(&previous.digest, &credential.digest()))
+            {
+                if previous.expires_at > now {
+                    let secret = derive(
+                        credential.secret,
+                        Some(&previous.salt),
+                        b"cpr-session-v2/refresh",
+                    )?;
+                    return grant(credential.id, &secret, session).map(Some);
+                }
+                // 已知上一代在重试窗口外再次使用，原子撤销整个会话
+                current = self
+                    .store
+                    .compare_exchange_session(credential.id, &session, None)
+                    .await
+                    .map_err(|error| map_store_error(error, "authentication session revocation"))?;
+            } else {
+                return Ok(None);
+            }
+        }
+        Err(AdminError::new(
+            AdminErrorKind::Unavailable,
+            "会话正在变更，请重试",
+        ))
+    }
+}

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/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/tests/backup/support.rs

@@ -725,11 +725,11 @@ impl AuthStore for FakeAuthStore {
     ) -> AdminStoreResult<()> {
         Ok(())
     }
-    async fn renew_session(
+    async fn compare_exchange_session(
         &self,
         _: &str,
         _: &gateway_admin::model::auth::AuthSession,
-        _: chrono::DateTime<chrono::Utc>,
+        _: Option<&gateway_admin::model::auth::AuthSession>,
     ) -> AdminStoreResult<Option<gateway_admin::model::auth::AuthSession>> {
         Ok(None)
     }

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/auth/mod.rs

@@ -1,4 +1,4 @@
-//! 验证管理员认证、密码变更、会话绑定与限流行为
+//! 验证管理员认证、密码变更与限流，提供会话轮换的内存存储辅助
 
 use std::{collections::BTreeMap, sync::Mutex};
 
@@ -69,6 +69,7 @@ async fn password_change_obeys_rate_limit_and_rejects_unbound_legacy_sessions()
                 admin_user_id: "admin".into(),
                 credential_fingerprint: String::new(),
             },
+            renewal: None,
             absolute_expires_at: None,
             expires_at: Utc::now() + TimeDelta::hours(1),
         },
@@ -138,19 +139,24 @@ impl AuthStore for MemoryAuthStore {
         Ok(())
     }
 
-    async fn renew_session(
+    async fn compare_exchange_session(
         &self,
         session_id: &str,
         expected: &AuthSession,
-        expires_at: chrono::DateTime<chrono::Utc>,
+        replacement: Option<&AuthSession>,
     ) -> AdminStoreResult<Option<AuthSession>> {
         let mut sessions = self.sessions.lock().unwrap();
         if let Some(session) = sessions.get_mut(session_id) {
-            if session.expires_at <= chrono::Utc::now() {
+            if session.retained_until() <= chrono::Utc::now() {
                 return Ok(None);
             }
             if session == expected {
-                session.expires_at = expires_at;
+                if let Some(replacement) = replacement {
+                    *session = replacement.clone();
+                } else {
+                    sessions.remove(session_id);
+                    return Ok(None);
+                }
             }
             return Ok(Some(session.clone()));
         }
@@ -190,8 +196,13 @@ impl AuthStore for MemoryAuthStore {
 #[tokio::test]
 async fn successful_login_should_create_expiring_session_and_audit() {
     let store = std::sync::Arc::new(MemoryAuthStore::default());
-    let services = super::AdminHarness::new().auth(store.clone()).build().await;
+    let services = super::AdminHarness::new()
+        .session_ttl_minutes(7 * 24 * 60)
+        .auth(store.clone())
+        .build()
+        .await;
 
+    let before = Utc::now();
     let result = services
         .auth()
         .login(
@@ -205,6 +216,17 @@ async fn successful_login_should_create_expiring_session_and_audit() {
         .await
         .expect("login");
 
+    let after = Utc::now();
+    for (expiry, ttl) in [
+        (result.session.expires_at, TimeDelta::hours(24)),
+        (result.session.retained_until(), TimeDelta::days(7)),
+        (
+            result.session.absolute_expires_at.unwrap(),
+            TimeDelta::days(30),
+        ),
+    ] {
+        assert!((before + ttl..=after + ttl).contains(&expiry));
+    }
     assert!(
         services
             .auth()
@@ -293,6 +315,7 @@ async fn expired_sessions_are_rejected_and_store_outages_are_not_treated_as_logo
                 credential_fingerprint: String::new(),
                 admin_user_id: "admin".to_owned(),
             },
+            renewal: None,
             absolute_expires_at: None,
             expires_at: Utc::now() - TimeDelta::seconds(1),
         },
@@ -362,7 +385,7 @@ async fn failed_rotation_discards_the_new_session_and_leaves_the_old_session_ret
         .await
         .unwrap();
     assert!(!format!("{original:?}").contains(&original.session_id));
-    *store.reject_delete.lock().unwrap() = Some(original.session_id.clone());
+    *store.reject_delete.lock().unwrap() = Some(stored_id(&original.session_id).to_owned());
     assert_eq!(
         services
             .auth()
@@ -391,109 +414,23 @@ async fn failed_rotation_discards_the_new_session_and_leaves_the_old_session_ret
     assert!(store.sessions.lock().unwrap().is_empty());
 }
 
-#[tokio::test]
-async fn renewal_is_throttled_and_capped_by_the_original_absolute_expiry() {
-    let store = std::sync::Arc::new(MemoryAuthStore::default());
-    let services = super::AdminHarness::new()
-        .auth(store.clone())
-        .session_ttl_minutes(60)
-        .build()
-        .await;
-    let auth = services.auth();
-    let login = auth
-        .login(
-            LoginCommand::Admin {
-                username: None,
-                password: "strong-test-password".into(),
-            },
-            std::net::Ipv4Addr::LOCALHOST.into(),
-            None,
-        )
-        .await
-        .unwrap();
-    assert_eq!(
-        auth.renew_session(Some(&login.session_id)).await.unwrap(),
-        Some(login.session.clone())
-    );
-    let limit = Utc::now() + TimeDelta::minutes(30);
-    {
-        let mut sessions = store.sessions.lock().unwrap();
-        let session = sessions.get_mut(&login.session_id).unwrap();
-        session.expires_at = Utc::now() + TimeDelta::minutes(1);
-        session.absolute_expires_at = Some(limit);
-    }
-    let renewed = auth
-        .renew_session(Some(&login.session_id))
-        .await
-        .unwrap()
-        .unwrap();
-    assert_eq!(renewed.expires_at, limit);
-    assert_eq!(renewed.absolute_expires_at, Some(limit));
-    assert_eq!(
-        auth.renew_session(Some(&login.session_id)).await.unwrap(),
-        Some(renewed)
-    );
+fn stored_id(token: &str) -> &str {
+    token.split('.').nth(1).unwrap_or(token)
 }
 
-#[tokio::test]
-async fn renewal_never_revives_expired_revoked_or_legacy_sessions() {
-    let store = std::sync::Arc::new(MemoryAuthStore::default());
-    let services = super::AdminHarness::new().auth(store.clone()).build().await;
-    let auth = services.auth();
-    let login = auth
-        .login(
-            LoginCommand::Admin {
-                username: None,
-                password: "strong-test-password".into(),
-            },
-            std::net::Ipv4Addr::LOCALHOST.into(),
-            None,
-        )
-        .await
-        .unwrap();
-    let mut session = login.session.clone();
-    session.absolute_expires_at = None;
-    session.expires_at = Utc::now() + TimeDelta::seconds(30);
-    store
-        .sessions
-        .lock()
-        .unwrap()
-        .insert(login.session_id.clone(), session.clone());
-    assert_eq!(
-        auth.renew_session(Some(&login.session_id)).await.unwrap(),
-        Some(session.clone())
-    );
-    session.expires_at = Utc::now() - TimeDelta::seconds(1);
-    store
-        .sessions
-        .lock()
-        .unwrap()
-        .insert(login.session_id.clone(), session);
-    assert!(
-        auth.renew_session(Some(&login.session_id))
-            .await
-            .unwrap()
-            .is_none()
-    );
-    let mut session = login.session;
-    session.absolute_expires_at = Some(Utc::now() - TimeDelta::seconds(1));
-    store
-        .sessions
-        .lock()
-        .unwrap()
-        .insert(login.session_id.clone(), session);
-    assert!(
-        auth.renew_session(Some(&login.session_id))
-            .await
-            .unwrap()
-            .is_none()
-    );
-    auth.logout(&login.session_id).await.unwrap();
-    assert!(
-        auth.renew_session(Some(&login.session_id))
-            .await
-            .unwrap()
-            .is_none()
-    );
-    assert!(store.sessions.lock().unwrap().is_empty());
+async fn admin_login(
+    auth: &dyn gateway_admin::AuthService,
+) -> gateway_admin::model::auth::SessionGrant {
+    auth.login(
+        LoginCommand::Admin {
+            username: None,
+            password: "strong-test-password".into(),
+        },
+        std::net::Ipv4Addr::LOCALHOST.into(),
+        None,
+    )
+    .await
+    .unwrap()
 }
+
+mod session;

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

@@ -0,0 +1,268 @@
+//! 验证双凭据续期、响应重试、并发撤销与固定期限边界
+
+use super::super::AdminHarness;
+use super::{MemoryAuthStore, admin_login, stored_id};
+use chrono::{TimeDelta, Utc};
+use gateway_admin::model::auth::AuthSession;
+
+#[tokio::test]
+async fn expired_access_is_restored_by_refresh_and_duplicate_refresh_replays_the_same_grant() {
+    let store = std::sync::Arc::new(MemoryAuthStore::default());
+    let services = AdminHarness::new().auth(store.clone()).build().await;
+    let auth = services.auth();
+    let login = admin_login(auth).await;
+    let refresh = login.refresh_token.as_deref().unwrap();
+    assert!(auth.session(Some(refresh)).await.unwrap().is_none());
+    assert!(
+        auth.renew_session(Some(&login.session_id), None)
+            .await
+            .unwrap()
+            .is_none()
+    );
+    assert!(
+        auth.renew_session(None, Some(&login.session_id))
+            .await
+            .unwrap()
+            .is_none()
+    );
+    assert_eq!(
+        auth.renew_session(Some(refresh), None).await.unwrap(),
+        Some(login.clone())
+    );
+    store
+        .sessions
+        .lock()
+        .unwrap()
+        .get_mut(stored_id(&login.session_id))
+        .unwrap()
+        .expires_at = Utc::now() - TimeDelta::seconds(1);
+    assert!(
+        auth.session(Some(&login.session_id))
+            .await
+            .unwrap()
+            .is_none()
+    );
+    let (first, concurrent) = tokio::join!(
+        auth.renew_session(Some(refresh), None),
+        auth.renew_session(Some(refresh), None)
+    );
+    let first = first.unwrap().unwrap();
+    assert_eq!(concurrent.unwrap().as_ref(), Some(&first));
+    assert_ne!(first.session_id, login.session_id);
+    assert_ne!(first.refresh_token, login.refresh_token);
+    assert_eq!(
+        first.session.absolute_expires_at,
+        login.session.absolute_expires_at
+    );
+    assert!(
+        auth.session(Some(&first.session_id))
+            .await
+            .unwrap()
+            .is_some()
+    );
+    assert_eq!(
+        auth.renew_session(first.refresh_token.as_deref(), None)
+            .await
+            .unwrap(),
+        Some(first.clone())
+    );
+    assert_eq!(
+        auth.renew_session(Some(refresh), None).await.unwrap(),
+        Some(first.clone())
+    );
+    let persisted = format!("{:?}", store.sessions.lock().unwrap());
+    for token in [
+        &first.session_id,
+        first.refresh_token.as_ref().unwrap(),
+        &login.session_id,
+        login.refresh_token.as_ref().unwrap(),
+    ] {
+        assert!(!persisted.contains(token));
+        assert!(!persisted.contains(token.rsplit('.').next().unwrap()));
+    }
+}
+
+#[tokio::test]
+async fn late_refresh_reuse_revokes_the_family_but_unknown_secrets_do_not() {
+    let store = std::sync::Arc::new(MemoryAuthStore::default());
+    let services = AdminHarness::new().auth(store.clone()).build().await;
+    let auth = services.auth();
+    let login = admin_login(auth).await;
+    store
+        .sessions
+        .lock()
+        .unwrap()
+        .get_mut(stored_id(&login.session_id))
+        .unwrap()
+        .expires_at = Utc::now() - TimeDelta::seconds(1);
+    let next = auth
+        .renew_session(login.refresh_token.as_deref(), None)
+        .await
+        .unwrap()
+        .unwrap();
+    let unknown = format!("r2.{}.{}", stored_id(&login.session_id), "A".repeat(43));
+    assert!(
+        auth.renew_session(Some(&unknown), None)
+            .await
+            .unwrap()
+            .is_none()
+    );
+    auth.logout(&unknown).await.unwrap();
+    auth.logout(stored_id(&login.session_id)).await.unwrap();
+    assert!(
+        auth.session(Some(&next.session_id))
+            .await
+            .unwrap()
+            .is_some()
+    );
+    store
+        .sessions
+        .lock()
+        .unwrap()
+        .get_mut(stored_id(&login.session_id))
+        .unwrap()
+        .renewal
+        .as_mut()
+        .unwrap()
+        .previous
+        .as_mut()
+        .unwrap()
+        .expires_at = Utc::now() - TimeDelta::seconds(1);
+    assert!(
+        auth.renew_session(login.refresh_token.as_deref(), None)
+            .await
+            .unwrap()
+            .is_none()
+    );
+    assert!(
+        auth.session(Some(&next.session_id))
+            .await
+            .unwrap()
+            .is_none()
+    );
+    assert!(
+        auth.renew_session(next.refresh_token.as_deref(), None)
+            .await
+            .unwrap()
+            .is_none()
+    );
+}
+
+#[tokio::test]
+async fn refresh_respects_idle_and_absolute_expiry_and_logout_after_access_expiry() {
+    let store = std::sync::Arc::new(MemoryAuthStore::default());
+    let services = AdminHarness::new().auth(store.clone()).build().await;
+    let auth = services.auth();
+    for absolute in [false, true] {
+        let login = admin_login(auth).await;
+        {
+            let mut sessions = store.sessions.lock().unwrap();
+            let session = sessions.get_mut(stored_id(&login.session_id)).unwrap();
+            if absolute {
+                session.absolute_expires_at = Some(Utc::now() - TimeDelta::seconds(1));
+            } else {
+                session.renewal.as_mut().unwrap().expires_at = Utc::now() - TimeDelta::seconds(1);
+            }
+        }
+        assert!(
+            auth.renew_session(login.refresh_token.as_deref(), None)
+                .await
+                .unwrap()
+                .is_none()
+        );
+        assert!(
+            auth.session(Some(&login.session_id))
+                .await
+                .unwrap()
+                .is_none()
+        );
+    }
+    let login = admin_login(auth).await;
+    store
+        .sessions
+        .lock()
+        .unwrap()
+        .get_mut(stored_id(&login.session_id))
+        .unwrap()
+        .expires_at = Utc::now() - TimeDelta::seconds(1);
+    auth.logout(login.refresh_token.as_deref().unwrap())
+        .await
+        .unwrap();
+    assert!(
+        auth.renew_session(login.refresh_token.as_deref(), None)
+            .await
+            .unwrap()
+            .is_none()
+    );
+}
+
+#[tokio::test]
+async fn refreshed_expiry_is_capped_and_store_outages_preserve_the_session() {
+    let store = std::sync::Arc::new(MemoryAuthStore::default());
+    let services = AdminHarness::new().auth(store.clone()).build().await;
+    let auth = services.auth();
+    let login = admin_login(auth).await;
+    let limit = Utc::now() + TimeDelta::minutes(1);
+    {
+        let mut sessions = store.sessions.lock().unwrap();
+        let session = sessions.get_mut(stored_id(&login.session_id)).unwrap();
+        session.absolute_expires_at = Some(limit);
+        session.expires_at = Utc::now() - TimeDelta::seconds(1);
+    }
+    store
+        .unavailable
+        .store(true, std::sync::atomic::Ordering::SeqCst);
+    assert!(
+        auth.renew_session(login.refresh_token.as_deref(), None)
+            .await
+            .is_err()
+    );
+    store
+        .unavailable
+        .store(false, std::sync::atomic::Ordering::SeqCst);
+    let refreshed = auth
+        .renew_session(login.refresh_token.as_deref(), None)
+        .await
+        .unwrap()
+        .unwrap();
+    assert_eq!(refreshed.session.expires_at, limit);
+    assert_eq!(refreshed.session.retained_until(), limit);
+    assert_eq!(refreshed.session.absolute_expires_at, Some(limit));
+}
+
+#[tokio::test]
+async fn fixed_sessions_cannot_acquire_refresh_credentials() {
+    let store = std::sync::Arc::new(MemoryAuthStore::default());
+    let services = AdminHarness::new().auth(store.clone()).build().await;
+    let auth = services.auth();
+    let login = admin_login(auth).await;
+    let fixed = AuthSession {
+        renewal: None,
+        ..login.session
+    };
+    store
+        .sessions
+        .lock()
+        .unwrap()
+        .insert("fixed".into(), fixed.clone());
+    let grant = auth
+        .renew_session(None, Some("fixed"))
+        .await
+        .unwrap()
+        .unwrap();
+    assert_eq!(grant.session, fixed);
+    assert!(grant.refresh_token.is_none());
+    store
+        .sessions
+        .lock()
+        .unwrap()
+        .get_mut("fixed")
+        .unwrap()
+        .expires_at = Utc::now() - TimeDelta::seconds(1);
+    assert!(
+        auth.renew_session(None, Some("fixed"))
+            .await
+            .unwrap()
+            .is_none()
+    );
+}

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

@@ -285,19 +285,24 @@ impl AuthStore for MemoryKeyAuthStore {
         Ok(())
     }
 
-    async fn renew_session(
+    async fn compare_exchange_session(
         &self,
         session_id: &str,
         expected: &AuthSession,
-        expires_at: chrono::DateTime<chrono::Utc>,
+        replacement: Option<&AuthSession>,
     ) -> AdminStoreResult<Option<AuthSession>> {
         let mut sessions = self.sessions.lock().unwrap();
         if let Some(session) = sessions.get_mut(session_id) {
-            if session.expires_at <= chrono::Utc::now() {
+            if session.retained_until() <= chrono::Utc::now() {
                 return Ok(None);
             }
             if session == expected {
-                session.expires_at = expires_at;
+                if let Some(replacement) = replacement {
+                    *session = replacement.clone();
+                } else {
+                    sessions.remove(session_id);
+                    return Ok(None);
+                }
             }
             return Ok(Some(session.clone()));
         }

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

@@ -369,11 +369,11 @@ impl AuthStore for BootstrapAuthStore {
         Err(unavailable("admin session"))
     }
 
-    async fn renew_session(
+    async fn compare_exchange_session(
         &self,
         _: &str,
         _: &gateway_admin::model::auth::AuthSession,
-        _: chrono::DateTime<chrono::Utc>,
+        _: Option<&gateway_admin::model::auth::AuthSession>,
     ) -> AdminStoreResult<Option<gateway_admin::model::auth::AuthSession>> {
         Ok(None)
     }

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/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-api/Cargo.toml

@@ -32,7 +32,7 @@ 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/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/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),
     }

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/auth.rs

@@ -138,7 +138,9 @@ where
         .login(
             payload.into(),
             peer.ip(),
-            session_cookie::value(&headers).as_deref(),
+            session_cookie::refresh_value(&headers)
+                .or_else(|| session_cookie::value(&headers))
+                .as_deref(),
         )
         .await
         .map_err(map_login_error)?;
@@ -149,8 +151,7 @@ where
         )
         .into_response(),
         &headers,
-        &result.session_id,
-        &result.session,
+        &result,
     )
 }
 
@@ -177,16 +178,18 @@ async fn refresh_session<S>(
 where
     S: SessionState + Send + Sync,
 {
-    let session_id = session_cookie::value(&headers);
-    let session = state
+    let grant = state
         .admin_services()
         .auth()
-        .renew_session(session_id.as_deref())
+        .renew_session(
+            session_cookie::refresh_value(&headers).as_deref(),
+            session_cookie::value(&headers).as_deref(),
+        )
         .await
         .map_err(map_admin_service_error)?;
-    let response = session_response(session.as_ref());
-    if let (Some(session_id), Some(session)) = (session_id, session) {
-        return set_session_cookie(response, &headers, &session_id, &session);
+    let response = session_response(grant.as_ref().map(|grant| &grant.session));
+    if let Some(grant) = grant {
+        return set_session_cookie(response, &headers, &grant);
     }
     // 失效响应不清 Cookie，避免晚到的请求覆盖另一个标签页刚建立的会话
     Ok(response)
@@ -206,32 +209,72 @@ fn session_response(session: Option<&AuthSession>) -> Response {
 fn set_session_cookie(
     mut response: Response,
     headers: &HeaderMap,
-    session_id: &str,
-    session: &AuthSession,
+    grant: &gateway_admin::model::auth::SessionGrant,
 ) -> Result<Response, AdminError> {
-    let max_age = session
-        .expires_at
-        .signed_duration_since(chrono::Utc::now())
-        .num_seconds()
-        .max(1);
-    let expires = session.expires_at.format("%a, %d %b %Y %H:%M:%S GMT");
-    let cookie = format!(
-        "{}={session_id}; {}; Max-Age={max_age}; Expires={expires}",
+    append_cookie(
+        &mut response,
+        headers,
         session_cookie::NAME,
-        session_cookie::attributes(headers)
-    );
-    response.headers_mut().insert(
-        header::SET_COOKIE,
-        HeaderValue::from_str(&cookie).map_err(|_| AdminError::internal())?,
-    );
+        &grant.session_id,
+        Some(grant.session.expires_at),
+    )?;
+    append_cookie(
+        &mut response,
+        headers,
+        session_cookie::REFRESH_NAME,
+        grant.refresh_token.as_deref().unwrap_or_default(),
+        grant
+            .session
+            .renewal
+            .as_ref()
+            .map(|renewal| renewal.expires_at),
+    )?;
     Ok(response)
 }
 
+fn append_cookie(
+    response: &mut Response,
+    headers: &HeaderMap,
+    name: &str,
+    value: &str,
+    expires_at: Option<chrono::DateTime<chrono::Utc>>,
+) -> Result<(), AdminError> {
+    let attributes = session_cookie::attributes(headers);
+    let path = if name == session_cookie::REFRESH_NAME {
+        "/api/auth"
+    } else {
+        "/"
+    };
+    let lifetime = expires_at.map_or_else(
+        || "Max-Age=0".to_owned(),
+        |expiry| {
+            format!(
+                "Max-Age={}; Expires={}",
+                expiry
+                    .signed_duration_since(chrono::Utc::now())
+                    .num_seconds()
+                    .max(1),
+                expiry.format("%a, %d %b %Y %H:%M:%S GMT")
+            )
+        },
+    );
+    response.headers_mut().append(
+        header::SET_COOKIE,
+        HeaderValue::from_str(&format!(
+            "{name}={value}; Path={path}; {attributes}; {lifetime}"
+        ))
+        .map_err(|_| AdminError::internal())?,
+    );
+    Ok(())
+}
+
 async fn logout<S>(State(state): State<S>, headers: HeaderMap) -> Result<Response, AdminError>
 where
     S: SessionState + Send + Sync,
 {
-    if let Some(session_id) = session_cookie::value(&headers) {
+    if let Some(session_id) =
+        session_cookie::refresh_value(&headers).or_else(|| session_cookie::value(&headers))
+    {
         state
             .admin_services()
             .auth()
@@ -285,15 +328,9 @@ fn clear_session_cookie(
     mut response: Response,
     headers: &HeaderMap,
 ) -> Result<Response, AdminError> {
-    let cookie = format!(
-        "{}=; {}; Max-Age=0",
-        session_cookie::NAME,
-        session_cookie::attributes(headers)
-    );
-    response.headers_mut().insert(
-        header::SET_COOKIE,
-        HeaderValue::from_str(&cookie).map_err(|_| AdminError::internal())?,
-    );
+    for name in [session_cookie::NAME, session_cookie::REFRESH_NAME] {
+        append_cookie(&mut response, headers, name, "", None)?;
+    }
     Ok(response)
 }
 

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/src/session_cookie.rs

@@ -4,6 +4,7 @@ use axum::http::{HeaderMap, header::ORIGIN};
 use url::Url;
 
 pub(crate) const NAME: &str = "cpr_session";
+pub(crate) const REFRESH_NAME: &str = "cpr_refresh";
 
 pub(crate) fn attributes(headers: &HeaderMap) -> &'static str {
     // HTTPS 反代回源 HTTP 不改变浏览器 Origin
@@ -19,16 +20,24 @@ pub(crate) fn attributes(headers: &HeaderMap) -> &'static str {
                 })
             });
     if http_origin {
-        "Path=/; HttpOnly; SameSite=Lax"
+        "HttpOnly; SameSite=Lax"
     } else {
-        "Path=/; Secure; HttpOnly; SameSite=Lax"
+        "Secure; HttpOnly; SameSite=Lax"
     }
 }
 
 pub(crate) fn value(headers: &HeaderMap) -> Option<String> {
+    named_value(headers, NAME)
+}
+
+pub(crate) fn refresh_value(headers: &HeaderMap) -> Option<String> {
+    named_value(headers, REFRESH_NAME)
+}
+
+fn named_value(headers: &HeaderMap, name: &str) -> Option<String> {
     let cookie = headers.get("cookie")?.to_str().ok()?;
     cookie.split(';').find_map(|part| {
         let (candidate, value) = part.trim().split_once('=')?;
-        (candidate == NAME).then(|| value.to_owned())
+        (candidate == name && !value.is_empty()).then(|| value.to_owned())
     })
 }

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;
@@ -321,6 +322,7 @@ impl MemoryAuthStore {
                     },
                     admin_user_id: "admin_1".to_owned(),
                 },
+                renewal: None,
                 absolute_expires_at: None,
                 expires_at: Utc::now() + Duration::hours(1),
             },
@@ -339,7 +341,7 @@ impl MemoryAuthStore {
         self.sessions
             .lock()
             .unwrap()
-            .get_mut(session_id)
+            .get_mut(session_id.split('.').nth(1).unwrap_or(session_id))
             .unwrap()
             .expires_at = expires_at;
     }
@@ -415,19 +417,24 @@ impl AuthStore for MemoryAuthStore {
         Ok(())
     }
 
-    async fn renew_session(
+    async fn compare_exchange_session(
         &self,
         session_id: &str,
         expected: &AuthSession,
-        expires_at: chrono::DateTime<chrono::Utc>,
+        replacement: Option<&AuthSession>,
     ) -> AdminStoreResult<Option<AuthSession>> {
         let mut sessions = self.sessions.lock().unwrap();
         if let Some(session) = sessions.get_mut(session_id) {
-            if session.expires_at <= chrono::Utc::now() {
+            if session.retained_until() <= chrono::Utc::now() {
                 return Ok(None);
             }
             if session == expected {
-                session.expires_at = expires_at;
+                if let Some(replacement) = replacement {
+                    *session = replacement.clone();
+                } else {
+                    sessions.remove(session_id);
+                    return Ok(None);
+                }
             }
             return Ok(Some(session.clone()));
         }

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/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

@@ -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())

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/auth.rs

@@ -23,6 +23,27 @@ fn session_cookie(response: &axum::response::Response) -> String {
         .to_owned()
 }
 
+fn refresh_cookie(response: &axum::response::Response) -> String {
+    response
+        .headers()
+        .get_all(header::SET_COOKIE)
+        .iter()
+        .filter_map(|value| value.to_str().ok())
+        .find(|value| value.starts_with("cpr_refresh="))
+        .expect("refresh cookie")
+        .split(';')
+        .next()
+        .unwrap()
+        .to_owned()
+}
+
+async fn refresh(app: &axum::Router, cookie: &str) -> axum::response::Response {
+    app.clone()
+        .oneshot(cookie_request(Method::POST, "/api/auth/refresh", cookie))
+        .await
+        .unwrap()
+}
+
 async fn login(
     app: &axum::Router,
     body: Value,
@@ -537,64 +558,129 @@ async fn key_auth_should_issue_restore_and_clear_a_unified_cookie() {
 }
 
 #[tokio::test]
-async fn refresh_renews_admin_cookie_and_preserves_fixed_key_expiry() {
+async fn refresh_restores_expired_access_with_only_the_refresh_cookie() {
     let (app, fixture) = auth_app().await;
-    for body in [
+    let response = login(
+        &app,
         json!({"mode": "admin", "password": "strong-admin-password"}),
-        json!({"mode": "key", "apiKey": RAW_KEY}),
-    ] {
-        let response = login(&app, body.clone(), None).await;
-        let cookie = session_cookie(&response);
-        let original = response_json(response).await["data"]["expiresAt"]
-            .as_str()
-            .unwrap()
-            .to_owned();
-        if body["mode"] == "admin" {
-            fixture.set_session_expiry(
-                cookie.split_once('=').unwrap().1,
-                chrono::Utc::now() + chrono::Duration::minutes(1),
-            );
-        }
-        let response = app
-            .clone()
-            .oneshot(cookie_request(Method::POST, "/api/auth/refresh", &cookie))
+        None,
+    )
+    .await;
+    let access = session_cookie(&response);
+    let renewal = refresh_cookie(&response);
+    let cookies: Vec<_> = response
+        .headers()
+        .get_all(header::SET_COOKIE)
+        .iter()
+        .map(|value| value.to_str().unwrap())
+        .collect();
+    assert_eq!(cookies.len(), 2);
+    assert!(cookies[0].contains("Path=/;"));
+    assert!(cookies[1].contains("Path=/api/auth;"));
+    assert!(
+        cookies
+            .iter()
+            .all(|value| value.contains("HttpOnly") && value.contains("SameSite=Lax"))
+    );
+    assert_eq!(
+        response_json(refresh(&app, &access).await).await["data"]["authenticated"],
+        false
+    );
+    fixture.set_session_expiry(
+        access.split_once('=').unwrap().1,
+        chrono::Utc::now() - chrono::Duration::seconds(1),
+    );
+    assert_eq!(
+        get(&app, "/api/admin/system/version", &access)
             .await
-            .unwrap();
-        assert_eq!(response.status(), StatusCode::OK);
-        assert_eq!(response.headers()[header::CACHE_CONTROL], "no-store");
-        assert_eq!(session_cookie(&response), cookie);
-        assert!(
-            response.headers()[header::SET_COOKIE]
-                .to_str()
-                .unwrap()
-                .contains("HttpOnly")
-        );
-        let data = response_json(response).await;
-        assert_eq!(data["data"]["authenticated"], true);
-        let expiry = data["data"]["session"]["expiresAt"].as_str().unwrap();
-        if body["mode"] == "key" {
-            assert_eq!(expiry, original);
-        } else {
-            assert!(
-                chrono::DateTime::parse_from_rfc3339(expiry).unwrap()
-                    > chrono::Utc::now() + chrono::Duration::minutes(1)
-            );
-        }
-    }
+            .status(),
+        StatusCode::UNAUTHORIZED
+    );
+    assert_eq!(
+        response_json(get(&app, "/api/auth/status", &renewal).await).await["data"]["authenticated"],
+        false
+    );
+
+    let response = refresh(&app, &renewal).await;
+    assert_eq!(response.status(), StatusCode::OK);
+    assert_eq!(response.headers()[header::CACHE_CONTROL], "no-store");
+    let next_access = session_cookie(&response);
+    let next_renewal = refresh_cookie(&response);
+    assert_ne!(next_access, access);
+    assert_ne!(next_renewal, renewal);
+    assert_eq!(response_json(response).await["data"]["authenticated"], true);
+    assert_eq!(
+        get(&app, "/api/admin/system/version", &next_access)
+            .await
+            .status(),
+        StatusCode::OK
+    );
+    assert_eq!(
+        get(&app, "/api/admin/system/version", &access)
+            .await
+            .status(),
+        StatusCode::UNAUTHORIZED
+    );
+    // 模拟响应丢失后的重试，返回同一代 Cookie
+    let retry = refresh(&app, &renewal).await;
+    assert_eq!(session_cookie(&retry), next_access);
+    assert_eq!(refresh_cookie(&retry), next_renewal);
+}
+
+#[tokio::test]
+async fn key_refresh_preserves_fixed_expiry_and_clears_admin_refresh_cookie() {
+    let (app, _) = auth_app().await;
+    let admin = login(
+        &app,
+        json!({"mode": "admin", "password": "strong-admin-password"}),
+        None,
+    )
+    .await;
+    let renewal = refresh_cookie(&admin);
+    let response = login(
+        &app,
+        json!({"mode": "key", "apiKey": RAW_KEY}),
+        Some(&renewal),
+    )
+    .await;
+    assert_eq!(refresh_cookie(&response), "cpr_refresh=");
+    assert!(
+        response
+            .headers()
+            .get_all(header::SET_COOKIE)
+            .iter()
+            .any(|value| {
+                let value = value.to_str().unwrap();
+                value.starts_with("cpr_refresh=")
+                    && value.contains("Max-Age=0")
+                    && value.contains("Path=/api/auth;")
+            })
+    );
+    let cookie = session_cookie(&response);
+    let original = response_json(response).await["data"]["expiresAt"].clone();
+    let response = refresh(&app, &cookie).await;
+    assert_eq!(session_cookie(&response), cookie);
+    let data = response_json(response).await;
+    assert_eq!(data["data"]["authenticated"], true);
+    assert_eq!(data["data"]["session"]["expiresAt"], original);
+    assert_eq!(
+        response_json(refresh(&app, &renewal).await).await["data"]["authenticated"],
+        false
+    );
 }
 
 #[tokio::test]
 async fn refresh_after_logout_or_password_change_cannot_recreate_a_session() {
-    let (app, _) = auth_app().await;
+    let (app, fixture) = auth_app().await;
     for password_change in [false, true] {
-        let cookie = session_cookie(
-            &login(
-                &app,
-                json!({"mode": "admin", "password": "strong-admin-password"}),
-                None,
-            )
-            .await,
-        );
+        let login_response = login(
+            &app,
+            json!({"mode": "admin", "password": "strong-admin-password"}),
+            None,
+        )
+        .await;
+        let cookie = session_cookie(&login_response);
+        let renewal = refresh_cookie(&login_response);
         let response = if password_change {
             change_password(
                 &app,
@@ -604,15 +690,34 @@ async fn refresh_after_logout_or_password_change_cannot_recreate_a_session() {
             )
             .await
         } else {
+            fixture.set_session_expiry(
+                cookie.split_once('=').unwrap().1,
+                chrono::Utc::now() - chrono::Duration::seconds(1),
+            );
             app.clone()
-                .oneshot(cookie_request(Method::POST, "/api/auth/logout", &cookie))
+                .oneshot(cookie_request(Method::POST, "/api/auth/logout", &renewal))
                 .await
                 .unwrap()
         };
         assert_eq!(response.status(), StatusCode::OK);
+        assert_eq!(
+            response
+                .headers()
+                .get_all(header::SET_COOKIE)
+                .iter()
+                .count(),
+            2
+        );
+        assert!(
+            response
+                .headers()
+                .get_all(header::SET_COOKIE)
+                .iter()
+                .all(|value| value.to_str().unwrap().contains("Max-Age=0"))
+        );
         let response = app
             .clone()
-            .oneshot(cookie_request(Method::POST, "/api/auth/refresh", &cookie))
+            .oneshot(cookie_request(Method::POST, "/api/auth/refresh", &renewal))
             .await
             .unwrap();
         assert_eq!(response.status(), StatusCode::OK);

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/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/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,
@@ -358,6 +359,7 @@ 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 +370,26 @@ 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 +524,7 @@ 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 {
@@ -560,8 +583,6 @@ where
             match self.pull().await? {
                 PullOutcome::Events(events) => {
                     self.replay_barrier_reached |= self.middleware.is_some()
-                        || self.client_transport != super::execution::ClientTransport::WebSocket
-                        || self.response_control.is_none()
                         || events.iter().any(|event| !event.is_replayable_prelude());
                     let requirement = if self.delivery_committed {
                         CommitRequirement::AlreadyCommitted
@@ -577,6 +598,7 @@ where
                     error,
                     send_state,
                 } => {
+                    self.record_client_retry_failure(&error).await;
                     let requirement = if self.delivery_committed {
                         CommitRequirement::AlreadyCommitted
                     } else {
@@ -914,7 +936,13 @@ 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));
@@ -1584,22 +1612,32 @@ 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 {
@@ -1622,9 +1660,11 @@ 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
@@ -1670,6 +1710,7 @@ where
             "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,
@@ -1950,6 +1991,9 @@ 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
@@ -2024,6 +2068,7 @@ 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
@@ -2051,6 +2096,16 @@ 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();
@@ -2345,6 +2400,16 @@ 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);
@@ -829,8 +840,15 @@ 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/mod.rs

@@ -3,6 +3,7 @@
 pub mod admission;
 pub mod authentication;
 pub mod budget;
+pub mod client_retry;
 pub mod connection;
 pub mod continuation;
 pub mod coordinator;
@@ -559,12 +560,10 @@ impl AttemptContext {
         self.request.client_transport
     }
 
-    /// 原生双向通道可提前交付可替换前导事件；中间件可能改变其输出语义
+    /// 原生响应可交付可替换前导事件；中间件可能改变其输出语义
     #[must_use]
     pub fn allows_replayable_prelude(&self) -> bool {
-        self.client_transport() == execution::ClientTransport::WebSocket
-            && self.request.middleware.is_none()
-            && self.response_control().is_some()
+        self.request.middleware.is_none()
     }
 
     #[must_use]
@@ -1123,12 +1122,26 @@ pub enum EngineError {
 pub struct GatewayEngine<S: ?Sized> {
     store: Arc<S>,
     providers: provider::ProviderRegistry,
+    client_retry: Option<Arc<dyn client_retry::ClientRetryStore>>,
 }
 
 impl<S: ?Sized> GatewayEngine<S> {
     #[must_use]
     pub const fn new(store: Arc<S>, providers: provider::ProviderRegistry) -> Self {
-        Self { store, providers }
+        Self {
+            store,
+            providers,
+            client_retry: None,
+        }
+    }
+
+    #[must_use]
+    pub fn with_client_retry_store(
+        mut self,
+        store: Arc<dyn client_retry::ClientRetryStore>,
+    ) -> Self {
+        self.client_retry = Some(store);
+        self
     }
 
     #[must_use]

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

@@ -638,7 +638,8 @@ fn provider_event_to_middleware_frame(
 
 fn encode_sse_wire_event(wire: &ProtocolWireEvent) -> Result<Bytes, ProviderError> {
     let mut encoded = Vec::new();
-    if let Some(id) = wire.sse_id() {
+    // 源事件元数据保持不透明；仅编码可表示为单行 SSE 字段的值，避免生成额外帧
+    if let Some(id) = wire.sse_id().filter(|id| !id.contains(['\r', '\n', '\0'])) {
         encoded.extend_from_slice(b"id: ");
         encoded.extend_from_slice(id.as_bytes());
         encoded.push(b'\n');
@@ -646,7 +647,10 @@ fn encode_sse_wire_event(wire: &ProtocolWireEvent) -> Result<Bytes, ProviderErro
     if let Some(retry) = wire.sse_retry() {
         encoded.extend_from_slice(format!("retry: {retry}\n").as_bytes());
     }
-    if let Some(event_type) = wire.event_type() {
+    if let Some(event_type) = wire
+        .event_type()
+        .filter(|event_type| !event_type.is_empty() && !event_type.contains(['\r', '\n']))
+    {
         encoded.extend_from_slice(b"event: ");
         encoded.extend_from_slice(event_type.as_bytes());
         encoded.push(b'\n');

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

@@ -120,19 +120,34 @@ pub enum ContinuationRecoveryDisposition {
     ProviderReplayAllowed,
 }
 
+/// 同账号瞬态错误的恢复策略，明确重试耗尽后的去向
+///
+/// Provider 按上游错误场景选择，Core 共用同一套退避与安全重放检查
+#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash)]
+pub enum TransientRetryPolicy {
+    /// 同账号退避重试，耗尽后允许在安全边界与换号预算内选择其他账号
+    RetryThenRotateAccount,
+    /// 同账号退避重试，耗尽后返回最后一次上游错误，不因该错误换号
+    RetryThenReturnError,
+}
+
 /// Provider 在下游尚未提交事件时请求的隐藏恢复方式
 #[derive(Debug, Clone, Copy, PartialEq, Eq, Hash)]
 pub enum PreDeliveryRetry {
     /// 排除本次账号，按普通调度策略重新选号
     AccountRotation,
-    /// 上游明确拒绝后的同账号退避；耗尽后仍由普通安全重放规则决定是否换号
+    /// 上游明确拒绝后的同账号退避；耗尽动作由 Provider 显式选择
     SameAccountTransientRetry {
         /// 每个账号在本请求中的重试上限，同时受总路由预算约束
         max_retries: NonZeroU32,
         /// 指数退避的起始间隔
         initial_delay: Duration,
-        /// 单次等待上限，避免上游提示把请求拖入长时间等待
+        /// 本地指数退避的基准上限，不含附加抖动，也不缩短上游 Retry-After
         max_delay: Duration,
+        /// Provider 为本次失败采样的附加等待，仅在没有 Retry-After 时使用
+        jitter: Duration,
+        /// 同账号恢复策略，不提供额外安全重放资格
+        policy: TransientRetryPolicy,
     },
     /// 固定本次账号，并按 Provider 给出的序号重试当前传输
     /// 可用于发送结果不明确时的有界尽力恢复，不等于上游未执行的证明
@@ -682,18 +697,22 @@ impl ProviderError {
         self
     }
 
-    /// 请求有界同账号退避；该意图本身不提供安全重放证据
+    /// 请求有界同账号退避并显式选择耗尽去向；该意图本身不提供安全重放证据
     #[must_use]
     pub fn with_transient_retry(
         mut self,
         max_retries: NonZeroU32,
         initial_delay: Duration,
         max_delay: Duration,
+        jitter: Duration,
+        policy: TransientRetryPolicy,
     ) -> Self {
         self.pre_delivery_retry = Some(Box::new(PreDeliveryRetry::SameAccountTransientRetry {
             max_retries,
             initial_delay,
             max_delay,
+            jitter,
+            policy,
         }));
         self
     }
@@ -1048,6 +1067,8 @@ pub enum GatewayErrorKind {
     NoAvailableProvider,
     /// 符合条件的上游账号暂时没有调度容量
     AccountCapacityUnavailable,
+    /// 中间件在途资源或执行等待队列已满
+    MiddlewareCapacityUnavailable,
     ConcurrencyQueueFull,
     ConcurrencyQueueTimeout,
     /// Provider 的本地账号基础设施不可用
@@ -1079,6 +1100,7 @@ impl GatewayErrorKind {
             Self::ModelNotFound => "model_not_found",
             Self::NoAvailableProvider => "no_available_provider",
             Self::AccountCapacityUnavailable => "account_capacity_unavailable",
+            Self::MiddlewareCapacityUnavailable => "middleware_capacity_unavailable",
             Self::ConcurrencyQueueFull => "concurrency_queue_full",
             Self::ConcurrencyQueueTimeout => "concurrency_queue_timeout",
             Self::ProviderInfrastructureUnavailable => "provider_infrastructure_unavailable",

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

@@ -62,6 +62,7 @@ pub struct CoreStorePorts {
     client_api_key_usage: Arc<dyn ClientApiKeyUsageSink>,
     budget: Option<Arc<dyn engine::budget::ClientBudgetPort>>,
     session_affinity: Option<Arc<dyn ProviderSessionAffinityPort>>,
+    client_retry: Option<Arc<dyn engine::client_retry::ClientRetryStore>>,
 }
 
 impl CoreStorePorts {
@@ -91,6 +92,7 @@ impl CoreStorePorts {
             client_api_key_usage,
             budget: None,
             session_affinity: None,
+            client_retry: None,
         }
     }
 
@@ -100,6 +102,15 @@ impl CoreStorePorts {
         self
     }
 
+    #[must_use]
+    pub fn with_client_retry_store(
+        mut self,
+        store: Arc<dyn engine::client_retry::ClientRetryStore>,
+    ) -> Self {
+        self.client_retry = Some(store);
+        self
+    }
+
     #[must_use]
     pub fn with_session_affinity(
         mut self,
@@ -209,6 +220,9 @@ fn build_execution_service(
     if let Some(budget) = &ports.budget {
         service = service.with_budget(Arc::clone(budget));
     }
+    if let Some(store) = &ports.client_retry {
+        service = service.with_client_retry_store(Arc::clone(store));
+    }
     if let Some(request_observers) = request_observers {
         service = service.with_request_observers(request_observers);
     }

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

@@ -404,11 +404,38 @@ pub struct GenerateRequest {
 #[derive(Clone, PartialEq)]
 struct GeneratePayload {
     protocol_payload: ProtocolPayload,
+    client_turn: Option<ClientTurn>,
     provider_session_state: Option<ProviderSessionState>,
     source_requirements: Option<CapabilityRequirements>,
     inherited_capability_requirements: Option<CapabilityRequirements>,
 }
 
+/// 协议入口确认的客户端推理轮次，不使用连接路由键猜测身份
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub struct ClientTurn {
+    thread_id: String,
+    turn_id: String,
+}
+
+impl ClientTurn {
+    pub fn new(thread_id: String, turn_id: String) -> Option<Self> {
+        if [thread_id.as_str(), turn_id.as_str()].iter().any(|value| {
+            value.trim().is_empty() || value.len() > 256 || value.chars().any(char::is_control)
+        }) {
+            return None;
+        }
+        Some(Self { thread_id, turn_id })
+    }
+
+    pub fn thread_id(&self) -> &str {
+        &self.thread_id
+    }
+
+    pub fn turn_id(&self) -> &str {
+        &self.turn_id
+    }
+}
+
 impl GenerateRequest {
     /// 创建携带协议不透明正文的生成请求
     ///
@@ -419,6 +446,7 @@ impl GenerateRequest {
         Self {
             payload: Arc::new(GeneratePayload {
                 protocol_payload,
+                client_turn: None,
                 provider_session_state: None,
                 source_requirements: None,
                 inherited_capability_requirements: None,
@@ -426,27 +454,14 @@ impl GenerateRequest {
         }
     }
 
-    /// 写回原生 Provider 编码结果，保留已准入的能力、连接上下文和会话状态
-    ///
-    /// 原生编码不消耗中间件的一次跨协议改写额度：可以保留协议名，也可以在一次
-    /// 中间件改写后编码到 Provider wire
-    /// 调用方仍须验证保护字段没有被改动
-    ///
-    /// # Errors
-    ///
-    /// 目标协议名无效时返回错误
-    pub fn with_native_encoded_body(
-        mut self,
-        protocol: impl Into<String>,
-        body: Map<String, Value>,
-    ) -> Result<Self, OperationError> {
-        let encoded = ProtocolPayload::json_object(protocol, body)?
-            .with_context(self.protocol_payload().context().clone());
-        let requirements = self.requirements();
-        let payload = Arc::make_mut(&mut self.payload);
-        payload.protocol_payload = encoded;
-        payload.source_requirements = Some(requirements);
-        Ok(self)
+    #[must_use]
+    pub fn with_client_turn(mut self, client_turn: Option<ClientTurn>) -> Self {
+        Arc::make_mut(&mut self.payload).client_turn = client_turn;
+        self
+    }
+
+    pub fn client_turn(&self) -> Option<&ClientTurn> {
+        self.payload.client_turn.as_ref()
     }
 
     /// 附着同一客户端连接上一轮由 Provider 返回的不透明状态
@@ -951,41 +966,6 @@ impl Operation {
         self.replace_protocol_wire_inner(None, body, context)
     }
 
-    /// 替换 Provider HTTP 正文、非 wire 上下文和已解析 headers，同时保持 endpoint、
-    /// method 与 query 不变
-    /// headers 继续由 [`ProviderHttpRequest::new`] 统一验证；
-    /// Core 不解析插件传输使用的 base64 表示
-    ///
-    /// # Errors
-    ///
-    /// 当前 operation 不是 Provider HTTP，或替换后的 headers 超出请求边界时返回错误
-    pub fn replace_provider_http_wire(
-        self,
-        body: Option<Bytes>,
-        context: Map<String, Value>,
-        headers: Vec<ProviderHttpHeader>,
-    ) -> Result<Self, OperationError> {
-        let Self::ProviderHttp(request) = self else {
-            return Err(OperationError::EmptyField {
-                field: "provider_http_request",
-            });
-        };
-        let payload = request.payload;
-        ProviderHttpRequest::new(
-            request.endpoint,
-            request.method,
-            request.query,
-            headers,
-            RawHttpPayload {
-                protocol: payload.protocol,
-                body: body.unwrap_or(payload.body),
-                context,
-                translated: payload.translated,
-            },
-        )
-        .map(Self::ProviderHttp)
-    }
-
     /// 把一次已选择的协议转换结果写回 operation，同时保留非 wire 上下文和
     /// Provider 会话状态
     /// 该方法只验证公共载荷形状，不解释目标协议字段

backend/crates/gateway-core/src/settings/privacy.rs

@@ -56,6 +56,12 @@ pub struct PrivacyRule {
     pub id: String,
     pub name: String,
     pub enabled: bool,
+    pub steps: Vec<PrivacyStep>,
+}
+
+#[derive(Clone, PartialEq, Eq, Serialize, Deserialize)]
+#[serde(rename_all = "camelCase", deny_unknown_fields)]
+pub struct PrivacyStep {
     pub scope: PrivacyScope,
     pub selector: String,
     pub action: PrivacyAction,
@@ -71,9 +77,17 @@ pub struct PrivacyRule {
 impl fmt::Debug for PrivacyRule {
     fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
         f.debug_struct("PrivacyRule")
+            .field("enabled", &self.enabled)
+            .field("step_count", &self.steps.len())
+            .finish_non_exhaustive()
+    }
+}
+
+impl fmt::Debug for PrivacyStep {
+    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
+        f.debug_struct("PrivacyStep")
             .field("scope", &self.scope)
             .field("action", &self.action)
-            .field("enabled", &self.enabled)
             .finish_non_exhaustive()
     }
 }

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

@@ -16,6 +16,7 @@ use futures::executor::block_on;
 
 use gateway_core::account::FastMode;
 use gateway_core::account::{AccountSelectionPolicy, ProviderAccountId, RotationStrategy};
+use gateway_core::engine::client_retry::{ClientRetryKey, ClientRetryStore};
 use gateway_core::engine::continuation::{
     ContinuationBinding, NativeContinuationPin, PreviousResponseId,
 };
@@ -31,7 +32,7 @@ use gateway_core::engine::{
 use gateway_core::error::{
     ClientVisibleUpstreamError, ClientVisibleUpstreamResponse, ContinuationFailure,
     ContinuationRecoveryDisposition, GatewayErrorKind, OpaqueUpstreamValue, ProviderError,
-    ProviderErrorKind, StoreError, StoreErrorKind,
+    ProviderErrorKind, StoreError, StoreErrorKind, TransientRetryPolicy,
 };
 use gateway_core::event::{
     ContentItem, ContentKind, GatewayEvent, ProtocolWireEvent, ProviderEvent,
@@ -40,6 +41,7 @@ use gateway_core::event::{
 };
 use gateway_core::lifecycle::CancellationToken;
 use gateway_core::metering::{CalculatedCost, CostSource, ProviderReportedCost, Usage};
+use gateway_core::operation::ClientTurn;
 use gateway_core::operation::{GenerateRequest, Operation, ProtocolPayload, ProviderSessionState};
 use gateway_core::policy::ClientApiKeyId;
 use gateway_core::routing::{
@@ -51,6 +53,407 @@ use gateway_core::routing::{
 use gateway_core::upstream::{UpstreamSendState, UpstreamTransport};
 use serde_json::{Map, Value, json};
 
+#[derive(Default)]
+struct RetryMarkers {
+    entries: Mutex<BTreeMap<String, ModelRequestId>>,
+    unavailable: bool,
+    stalled: bool,
+}
+
+fn retry_marker_key(key: &ClientRetryKey) -> String {
+    format!(
+        "{}/{}/{}",
+        key.client_key_id.as_str(),
+        key.turn.thread_id(),
+        key.turn.turn_id()
+    )
+}
+
+impl ClientRetryStore for RetryMarkers {
+    fn load<'a>(
+        &'a self,
+        key: &'a ClientRetryKey,
+    ) -> futures::future::BoxFuture<'a, Result<Option<ModelRequestId>, StoreError>> {
+        Box::pin(async move {
+            if self.stalled {
+                return futures::future::pending().await;
+            }
+            if self.unavailable {
+                return Err(StoreError::new(StoreErrorKind::Unavailable));
+            }
+            Ok(self
+                .entries
+                .lock()
+                .unwrap()
+                .get(&retry_marker_key(key))
+                .cloned())
+        })
+    }
+
+    fn record_failure<'a>(
+        &'a self,
+        key: &'a ClientRetryKey,
+        version: &'a ModelRequestId,
+        ttl: Duration,
+    ) -> futures::future::BoxFuture<'a, Result<(), StoreError>> {
+        Box::pin(async move {
+            assert_eq!(ttl, Duration::from_secs(300));
+            if self.stalled {
+                return futures::future::pending().await;
+            }
+            if self.unavailable {
+                return Err(StoreError::new(StoreErrorKind::Unavailable));
+            }
+            self.entries
+                .lock()
+                .unwrap()
+                .insert(retry_marker_key(key), version.clone());
+            Ok(())
+        })
+    }
+
+    fn clear_failure<'a>(
+        &'a self,
+        key: &'a ClientRetryKey,
+        expected: &'a ModelRequestId,
+    ) -> futures::future::BoxFuture<'a, Result<(), StoreError>> {
+        Box::pin(async move {
+            if self.stalled {
+                return futures::future::pending().await;
+            }
+            if self.unavailable {
+                return Err(StoreError::new(StoreErrorKind::Unavailable));
+            }
+            let mut entries = self.entries.lock().unwrap();
+            if entries.get(&retry_marker_key(key)) == Some(expected) {
+                entries.remove(&retry_marker_key(key));
+            }
+            Ok(())
+        })
+    }
+}
+
+fn retry_turn_operation(thread: &str, turn: &str) -> Operation {
+    let Operation::Generate(request) = generate_operation() else {
+        unreachable!()
+    };
+    Operation::Generate(
+        request.with_client_turn(ClientTurn::new(thread.to_owned(), turn.to_owned())),
+    )
+}
+
+fn retry_failure() -> Script {
+    Script::Stream {
+        account_id: "acct_first",
+        items: vec![Err(ProviderError::new(
+            ProviderErrorKind::Unavailable,
+            UpstreamSendState::Sent,
+        )
+        .with_status(503)
+        .with_replay_safe()
+        .with_transient_retry(
+            NonZeroU32::new(5).unwrap(),
+            Duration::ZERO,
+            Duration::ZERO,
+            Duration::ZERO,
+            TransientRetryPolicy::RetryThenReturnError,
+        ))],
+    }
+}
+
+fn retry_coordinator(
+    scripts: Vec<Script>,
+    markers: Arc<RetryMarkers>,
+) -> (AttemptCoordinator<FakeStore>, Arc<ScriptedProvider>) {
+    let provider = Arc::new(ScriptedProvider::new(scripts));
+    let mut registry = ProviderRegistry::builder();
+    registry.register(provider.clone()).unwrap();
+    let engine = GatewayEngine::new(Arc::new(FakeStore::default()), registry.build())
+        .with_client_retry_store(markers);
+    (AttemptCoordinator::new(engine), provider)
+}
+
+fn retry_session(
+    coordinator: &AttemptCoordinator<FakeStore>,
+    id: &str,
+    operation: Operation,
+    client: &str,
+) -> gateway_core::engine::ResponseExecutionSession<FakeStore> {
+    let mut request = model_request(&operation, SystemTime::now() + Duration::from_secs(30));
+    request.id = ModelRequestId::new(id).unwrap();
+    request.client_api_key_ref = ClientApiKeyId::new(client).unwrap();
+    block_on(coordinator.start(
+        request,
+        operation.clone(),
+        plan(&operation),
+        None,
+        None,
+        CancellationToken::new(),
+    ))
+    .unwrap()
+}
+
+#[test]
+fn client_retry_should_skip_repeated_budget_and_clear_before_completed_delivery() {
+    let markers = Arc::new(RetryMarkers::default());
+    let mut scripts = (0..7).map(|_| retry_failure()).collect::<Vec<_>>();
+    scripts.push(Script::HangingStream {
+        account_id: "acct_first",
+        items: complete_stream(None),
+    });
+    scripts.extend([
+        retry_failure(),
+        retry_failure(),
+        Script::Stream {
+            account_id: "acct_first",
+            items: complete_stream(None),
+        },
+    ]);
+    let (coordinator, provider) = retry_coordinator(scripts, markers.clone());
+    let operation = retry_turn_operation("thread", "turn");
+    for (id, expected) in [("req_initial", 6), ("req_retry", 7)] {
+        let mut session = retry_session(&coordinator, id, operation.clone(), "key_client_1");
+        assert!(block_on(session.collect_uncommitted()).is_err());
+        assert_eq!(provider.contexts.lock().unwrap().len(), expected);
+    }
+    let mut recovered = retry_session(
+        &coordinator,
+        "req_recovered",
+        operation.clone(),
+        "key_client_1",
+    );
+    block_on(recovered.next_event()).unwrap().unwrap();
+    block_on(recovered.commit_downstream(Some(200))).unwrap();
+    let completed = block_on(recovered.next_event()).unwrap().unwrap();
+    assert!(completed.into_provider_events().iter().any(|event| {
+        event
+            .canonical_facts()
+            .iter()
+            .any(|fact| matches!(fact, GatewayEvent::Completed(_)))
+    }));
+    assert!(markers.entries.lock().unwrap().is_empty());
+    assert!(!recovered.is_finalized());
+    // 输入可以随工具结果改变，正常续请求沿用同轮次但重新获得网关预算
+    let mut followup = retry_session(&coordinator, "req_followup", operation, "key_client_1");
+    block_on(followup.collect_uncommitted()).unwrap();
+    block_on(followup.commit_downstream(Some(200))).unwrap();
+    assert_eq!(provider.contexts.lock().unwrap().len(), 11);
+    block_on(recovered.cancel_and_finalize()).unwrap();
+}
+
+#[test]
+fn client_retry_should_record_atomic_failure_before_delivering_its_events() {
+    let markers = Arc::new(RetryMarkers::default());
+    let (coordinator, provider) = retry_coordinator(
+        vec![
+            Script::Stream {
+                account_id: "acct_first",
+                items: vec![
+                    Ok(GatewayEvent::Started(ResponseMeta::new("resp", "gpt-5"))),
+                    Err(
+                        atomic_response_failed("resp", "failed").with_transient_retry(
+                            NonZeroU32::new(5).unwrap(),
+                            Duration::ZERO,
+                            Duration::ZERO,
+                            Duration::ZERO,
+                            TransientRetryPolicy::RetryThenReturnError,
+                        ),
+                    ),
+                ],
+            },
+            retry_failure(),
+        ],
+        markers.clone(),
+    );
+    let operation = retry_turn_operation("thread", "turn");
+    let mut first = retry_session(
+        &coordinator,
+        "req_atomic",
+        operation.clone(),
+        "key_client_1",
+    );
+    block_on(first.next_event()).unwrap().unwrap();
+    block_on(first.commit_downstream(Some(200))).unwrap();
+    block_on(first.next_event()).unwrap().unwrap();
+    assert_eq!(
+        markers
+            .entries
+            .lock()
+            .unwrap()
+            .values()
+            .next()
+            .unwrap()
+            .as_str(),
+        "req_atomic"
+    );
+    // 不驱动旧请求 EOF，新请求此时已经必须看到失败标记
+    let mut retry = retry_session(&coordinator, "req_atomic_retry", operation, "key_client_1");
+    assert!(block_on(retry.collect_uncommitted()).is_err());
+    assert_eq!(provider.contexts.lock().unwrap().len(), 2);
+    block_on(first.cancel_and_finalize()).unwrap();
+    assert_eq!(
+        markers
+            .entries
+            .lock()
+            .unwrap()
+            .values()
+            .next()
+            .unwrap()
+            .as_str(),
+        "req_atomic_retry"
+    );
+}
+
+#[test]
+fn client_retry_late_success_should_keep_a_newer_failure() {
+    let markers = Arc::new(RetryMarkers::default());
+    markers.entries.lock().unwrap().insert(
+        "key_client_1/thread/turn".to_owned(),
+        ModelRequestId::new("req_old").unwrap(),
+    );
+    let (coordinator, _) = retry_coordinator(
+        vec![
+            Script::HangingStream {
+                account_id: "acct_first",
+                items: complete_stream(None),
+            },
+            retry_failure(),
+        ],
+        markers.clone(),
+    );
+    let operation = retry_turn_operation("thread", "turn");
+    let mut slow = retry_session(&coordinator, "req_slow", operation.clone(), "key_client_1");
+    block_on(slow.next_event()).unwrap().unwrap();
+    block_on(slow.commit_downstream(Some(200))).unwrap();
+    let mut newer = retry_session(&coordinator, "req_newer", operation, "key_client_1");
+    assert!(block_on(newer.collect_uncommitted()).is_err());
+    block_on(slow.next_event()).unwrap().unwrap();
+    assert_eq!(
+        markers
+            .entries
+            .lock()
+            .unwrap()
+            .values()
+            .next()
+            .unwrap()
+            .as_str(),
+        "req_newer"
+    );
+    block_on(slow.cancel_and_finalize()).unwrap();
+}
+
+#[test]
+fn client_retry_should_isolate_client_thread_and_turn_and_ignore_missing_identity() {
+    for (client, operation) in [
+        ("key_other", retry_turn_operation("thread", "turn")),
+        ("key_client_1", retry_turn_operation("other", "turn")),
+        ("key_client_1", retry_turn_operation("thread", "other")),
+        ("key_client_1", generate_operation()),
+    ] {
+        let markers = Arc::new(RetryMarkers::default());
+        markers.entries.lock().unwrap().insert(
+            "key_client_1/thread/turn".to_owned(),
+            ModelRequestId::new("req_old").unwrap(),
+        );
+        let (coordinator, provider) = retry_coordinator(
+            vec![
+                retry_failure(),
+                Script::Stream {
+                    account_id: "acct_first",
+                    items: complete_stream(None),
+                },
+            ],
+            markers.clone(),
+        );
+        let mut session = retry_session(&coordinator, "req_distinct", operation, client);
+        block_on(session.collect_uncommitted()).unwrap();
+        block_on(session.commit_downstream(Some(200))).unwrap();
+        assert_eq!(provider.contexts.lock().unwrap().len(), 2);
+        assert_eq!(markers.entries.lock().unwrap().len(), 1);
+    }
+}
+
+#[test]
+fn client_retry_store_failure_or_timeout_should_preserve_original_retries_and_error() {
+    for stalled in [false, true] {
+        let markers = Arc::new(RetryMarkers {
+            unavailable: !stalled,
+            stalled,
+            ..Default::default()
+        });
+        let (coordinator, provider) =
+            retry_coordinator((0..6).map(|_| retry_failure()).collect(), markers);
+        let started = std::time::Instant::now();
+        let mut session = retry_session(
+            &coordinator,
+            "req_unavailable",
+            retry_turn_operation("thread", "turn"),
+            "key_client_1",
+        );
+        let Err(EngineError::Provider(error)) = block_on(session.collect_uncommitted()) else {
+            panic!("original upstream failure");
+        };
+        assert_eq!(error.upstream_status(), Some(503));
+        assert_eq!(provider.contexts.lock().unwrap().len(), 6);
+        assert!(started.elapsed() < Duration::from_secs(2));
+    }
+}
+
+#[test]
+fn client_retry_should_preserve_other_retry_policies_and_not_mark_their_failures() {
+    for seeded in [false, true] {
+        let markers = Arc::new(RetryMarkers::default());
+        if seeded {
+            markers.entries.lock().unwrap().insert(
+                "key_client_1/thread/turn".to_owned(),
+                ModelRequestId::new("req_old").unwrap(),
+            );
+        }
+        let failure = || {
+            ProviderError::new(ProviderErrorKind::RateLimited, UpstreamSendState::Sent)
+                .with_status(429)
+                .with_replay_safe()
+                .with_transient_retry(
+                    NonZeroU32::new(3).unwrap(),
+                    Duration::ZERO,
+                    Duration::ZERO,
+                    Duration::ZERO,
+                    TransientRetryPolicy::RetryThenRotateAccount,
+                )
+        };
+        let scripts = vec![
+            Script::Stream {
+                account_id: "acct_first",
+                items: vec![Err(failure())],
+            },
+            // 已提交输出后的失败直接交回客户端，不应创建或刷新本策略标记
+            Script::Stream {
+                account_id: "acct_first",
+                items: vec![
+                    Ok(GatewayEvent::Started(ResponseMeta::new("resp", "gpt-5"))),
+                    Err(failure()),
+                ],
+            },
+        ];
+        let (coordinator, provider) = retry_coordinator(scripts, markers.clone());
+        let mut session = retry_session(
+            &coordinator,
+            "req_other_policy",
+            retry_turn_operation("thread", "turn"),
+            "key_client_1",
+        );
+        block_on(session.next_event()).unwrap().unwrap();
+        block_on(session.commit_downstream(Some(200))).unwrap();
+        assert!(block_on(session.next_event()).is_err());
+        assert_eq!(provider.contexts.lock().unwrap().len(), 2);
+        let entries = markers.entries.lock().unwrap();
+        assert_eq!(entries.len(), usize::from(seeded));
+        if seeded {
+            assert_eq!(entries.values().next().unwrap().as_str(), "req_old");
+        }
+    }
+}
+
 #[derive(Debug, Clone, PartialEq)]
 struct FinalState {
     error_message: Option<String>,
@@ -3668,6 +4071,8 @@ fn transient_rejection() -> ProviderError {
         NonZeroU32::new(3).expect("retry budget"),
         Duration::ZERO,
         Duration::ZERO,
+        Duration::ZERO,
+        TransientRetryPolicy::RetryThenRotateAccount,
     )
 }
 
@@ -3679,6 +4084,15 @@ fn server_retry_advice_should_override_local_backoff_and_transport_fallback() {
             NonZeroU32::MIN,
             Duration::from_millis(1),
             Duration::from_millis(1),
+            Duration::from_secs(1),
+            TransientRetryPolicy::RetryThenRotateAccount,
+        ),
+        transient_rejection().with_transient_retry(
+            NonZeroU32::MIN,
+            Duration::from_millis(1),
+            Duration::from_millis(1),
+            Duration::from_secs(1),
+            TransientRetryPolicy::RetryThenReturnError,
         ),
         ProviderError::new(ProviderErrorKind::Transport, UpstreamSendState::NotSent)
             .with_pre_delivery_transport_fallback(),
@@ -3710,6 +4124,14 @@ fn server_retry_advice_should_override_local_backoff_and_transport_fallback() {
         assert_eq!(provider.contexts.lock().unwrap().len(), 1);
         block_on(collection).unwrap();
         assert!(started.elapsed() >= advised_delay);
+        let trace = session.trace().snapshot().unwrap();
+        let retry = trace["events"]
+            .as_array()
+            .unwrap()
+            .iter()
+            .find(|event| event["stage"] == "retry.decided")
+            .unwrap();
+        assert_eq!(retry["data"]["delayMs"], 120);
         let contexts = provider.contexts.lock().unwrap();
         assert_eq!(contexts.len(), 2);
         assert_eq!(
@@ -3721,38 +4143,45 @@ fn server_retry_advice_should_override_local_backoff_and_transport_fallback() {
 
 #[test]
 fn zero_server_retry_advice_should_override_nonzero_local_backoff() {
-    let operation = generate_operation();
-    let route_plan = plan(&operation);
-    let (coordinator, _, _) = coordinator(vec![
-        Script::Stream {
-            account_id: "acct_first",
-            items: vec![Err(transient_rejection()
-                .with_transient_retry(
-                    NonZeroU32::MIN,
-                    Duration::from_secs(60),
-                    Duration::from_secs(60),
-                )
-                .with_retry_after(Duration::ZERO))],
-        },
-        Script::Stream {
-            account_id: "acct_first",
-            items: complete_stream(None),
-        },
-    ]);
-    let mut session = block_on(coordinator.start(
-        model_request(&operation, SystemTime::now() + Duration::from_secs(30)),
-        operation,
-        route_plan,
-        None,
-        None,
-        CancellationToken::new(),
-    ))
-    .unwrap();
-    let mut collection = Box::pin(session.collect_uncommitted());
-    assert!(matches!(
-        block_on(async { futures::poll!(&mut collection) }),
-        std::task::Poll::Ready(Ok(_))
-    ));
+    for policy in [
+        TransientRetryPolicy::RetryThenRotateAccount,
+        TransientRetryPolicy::RetryThenReturnError,
+    ] {
+        let operation = generate_operation();
+        let route_plan = plan(&operation);
+        let (coordinator, _, _) = coordinator(vec![
+            Script::Stream {
+                account_id: "acct_first",
+                items: vec![Err(transient_rejection()
+                    .with_transient_retry(
+                        NonZeroU32::MIN,
+                        Duration::from_secs(60),
+                        Duration::from_secs(60),
+                        Duration::from_secs(60),
+                        policy,
+                    )
+                    .with_retry_after(Duration::ZERO))],
+            },
+            Script::Stream {
+                account_id: "acct_first",
+                items: complete_stream(None),
+            },
+        ]);
+        let mut session = block_on(coordinator.start(
+            model_request(&operation, SystemTime::now() + Duration::from_secs(30)),
+            operation,
+            route_plan,
+            None,
+            None,
+            CancellationToken::new(),
+        ))
+        .unwrap();
+        let mut collection = Box::pin(session.collect_uncommitted());
+        assert!(matches!(
+            block_on(async { futures::poll!(&mut collection) }),
+            std::task::Poll::Ready(Ok(_))
+        ));
+    }
 }
 
 #[test]
@@ -3842,6 +4271,113 @@ fn transient_retries_back_off_on_the_same_account_then_rotate_with_a_fresh_budge
     );
 }
 
+#[test]
+fn transient_return_to_client_policy_keeps_the_account_and_preserves_the_final_error() {
+    for succeeds in [false, true] {
+        let operation = generate_operation();
+        let route_plan = plan(&operation);
+        let jitter_millis = [3, 7, 2, 6, 4, 0];
+        let failure = |attempt: usize| {
+            ProviderError::new(ProviderErrorKind::Unavailable, UpstreamSendState::Sent)
+                .with_status(503)
+                .with_replay_safe()
+                .with_client_visible_upstream_response(ClientVisibleUpstreamResponse::new(
+                    503,
+                    Some(b"application/json".to_vec()),
+                    format!(
+                        r#"{{"error":{{"code":"server_error","message":"attempt {attempt}"}}}}"#
+                    )
+                    .into(),
+                ))
+                .with_transient_retry(
+                    NonZeroU32::new(5).expect("retry budget"),
+                    Duration::from_millis(5),
+                    Duration::from_millis(20),
+                    Duration::from_millis(jitter_millis[attempt - 1]),
+                    TransientRetryPolicy::RetryThenReturnError,
+                )
+        };
+        let mut scripts = (1..=5)
+            .map(|attempt| Script::Stream {
+                account_id: "acct_first",
+                items: vec![Err(failure(attempt))],
+            })
+            .collect::<Vec<_>>();
+        scripts.push(Script::Stream {
+            account_id: "acct_first",
+            items: if succeeds {
+                complete_stream(None)
+            } else {
+                vec![Err(failure(6))]
+            },
+        });
+        scripts.push(Script::Stream {
+            account_id: "acct_second",
+            items: complete_stream(None),
+        });
+        let (coordinator, store, provider) = coordinator(scripts);
+        let mut session = block_on(coordinator.start(
+            model_request(&operation, SystemTime::now() + Duration::from_secs(30)),
+            operation,
+            route_plan,
+            None,
+            None,
+            CancellationToken::new(),
+        ))
+        .expect("start execution");
+        let result = block_on(session.collect_uncommitted());
+        if succeeds {
+            result.expect("same account recovers");
+            block_on(session.commit_downstream(Some(200))).expect("commit");
+        } else {
+            let Err(EngineError::Provider(error)) = result else {
+                panic!("retry exhaustion must return the last failure without rotating");
+            };
+            assert_eq!(error.upstream_status(), Some(503));
+            assert_eq!(
+                error
+                    .client_visible_upstream_response()
+                    .unwrap()
+                    .body()
+                    .as_ref(),
+                failure(6)
+                    .client_visible_upstream_response()
+                    .unwrap()
+                    .body()
+                    .as_ref(),
+            );
+        }
+        let original = ProviderAccountId::new("acct_first").unwrap();
+        let trace = session.trace().snapshot().unwrap();
+        let delays: Vec<_> = trace["events"]
+            .as_array()
+            .unwrap()
+            .iter()
+            .filter(|event| event["stage"] == "retry.decided")
+            .filter_map(|event| event["data"]["delayMs"].as_u64())
+            .collect();
+        // 每次使用该次失败的抖动，指数基准封顶后再相加，不放大或截断抖动
+        assert_eq!(delays, [8, 17, 22, 26, 24]);
+        let contexts = provider.contexts.lock().unwrap();
+        assert_eq!(contexts.len(), 6);
+        for context in &contexts[1..] {
+            assert_eq!(context.required_account(), Some(&original));
+            assert!(!context.excluded_accounts().contains(&original));
+        }
+        let state = store.state.lock().unwrap();
+        assert_eq!(state.intermediate_failures, 5);
+        assert_eq!(state.finalizations[0].attempt_count, 6);
+        assert_eq!(
+            state.finalizations[0].outcome,
+            if succeeds {
+                ExecutionOutcome::Succeeded
+            } else {
+                ExecutionOutcome::Failed
+            }
+        );
+    }
+}
+
 #[test]
 fn transient_retries_consume_the_total_routing_budget() {
     let operation = generate_operation();
@@ -3854,6 +4390,8 @@ fn transient_retries_consume_the_total_routing_budget() {
                 max_attempts,
                 Duration::ZERO,
                 Duration::ZERO,
+                Duration::ZERO,
+                TransientRetryPolicy::RetryThenRotateAccount,
             ))],
         })
         .collect();
@@ -3890,7 +4428,13 @@ fn transient_retry_intent_does_not_supply_replay_proof_or_override_ambiguous_sen
         let operation = generate_operation();
         let route_plan = plan(&operation);
         let mut error = ProviderError::new(ProviderErrorKind::Unavailable, send_state)
-            .with_transient_retry(NonZeroU32::MIN, Duration::ZERO, Duration::ZERO);
+            .with_transient_retry(
+                NonZeroU32::MIN,
+                Duration::ZERO,
+                Duration::ZERO,
+                Duration::ZERO,
+                TransientRetryPolicy::RetryThenRotateAccount,
+            );
         if prove_safe {
             error = error.with_replay_safe();
         }
@@ -3917,30 +4461,37 @@ fn transient_retry_intent_does_not_supply_replay_proof_or_override_ambiguous_sen
 
 #[test]
 fn transient_backoff_obeys_request_deadline() {
-    let operation = generate_operation();
-    let route_plan = plan(&operation);
-    let (coordinator, _, provider) = coordinator(vec![Script::Stream {
-        account_id: "acct_first",
-        items: vec![Err(transient_rejection().with_transient_retry(
-            NonZeroU32::MIN,
-            Duration::from_secs(60),
-            Duration::from_secs(60),
-        ))],
-    }]);
-    let mut session = block_on(coordinator.start(
-        model_request(&operation, SystemTime::now() + Duration::from_millis(100)),
-        operation,
-        route_plan,
-        None,
-        None,
-        CancellationToken::new(),
-    ))
-    .expect("start execution");
-    assert!(matches!(
-        block_on(session.collect_uncommitted()),
-        Err(EngineError::Deadline)
-    ));
-    assert_eq!(provider.contexts.lock().expect("contexts").len(), 1);
+    for policy in [
+        TransientRetryPolicy::RetryThenRotateAccount,
+        TransientRetryPolicy::RetryThenReturnError,
+    ] {
+        let operation = generate_operation();
+        let route_plan = plan(&operation);
+        let (coordinator, _, provider) = coordinator(vec![Script::Stream {
+            account_id: "acct_first",
+            items: vec![Err(transient_rejection().with_transient_retry(
+                NonZeroU32::MIN,
+                Duration::from_secs(60),
+                Duration::from_secs(60),
+                Duration::ZERO,
+                policy,
+            ))],
+        }]);
+        let mut session = block_on(coordinator.start(
+            model_request(&operation, SystemTime::now() + Duration::from_millis(100)),
+            operation,
+            route_plan,
+            None,
+            None,
+            CancellationToken::new(),
+        ))
+        .expect("start execution");
+        assert!(matches!(
+            block_on(session.collect_uncommitted()),
+            Err(EngineError::Deadline)
+        ));
+        assert_eq!(provider.contexts.lock().expect("contexts").len(), 1);
+    }
 }
 
 #[test]
@@ -4022,36 +4573,43 @@ fn transient_retry_is_suppressed_for_required_accounts_and_committed_output() {
 
 #[test]
 fn transient_backoff_can_be_cancelled_before_the_next_attempt() {
-    let operation = generate_operation();
-    let route_plan = plan(&operation);
-    let (coordinator, _, provider) = coordinator(vec![Script::Stream {
-        account_id: "acct_first",
-        items: vec![Err(transient_rejection().with_transient_retry(
-            NonZeroU32::MIN,
-            Duration::from_secs(60),
-            Duration::from_secs(60),
-        ))],
-    }]);
-    let cancellation = CancellationToken::new();
-    let mut session = block_on(coordinator.start(
-        model_request(&operation, SystemTime::now() + Duration::from_secs(30)),
-        operation,
-        route_plan,
-        None,
-        None,
-        cancellation.clone(),
-    ))
-    .expect("start execution");
-    let cancel = std::thread::spawn(move || {
-        std::thread::sleep(Duration::from_millis(50));
-        cancellation.cancel();
-    });
-    assert!(matches!(
-        block_on(session.collect_uncommitted()),
-        Err(EngineError::Cancelled)
-    ));
-    cancel.join().expect("cancel task");
-    assert_eq!(provider.contexts.lock().expect("contexts").len(), 1);
+    for policy in [
+        TransientRetryPolicy::RetryThenRotateAccount,
+        TransientRetryPolicy::RetryThenReturnError,
+    ] {
+        let operation = generate_operation();
+        let route_plan = plan(&operation);
+        let (coordinator, _, provider) = coordinator(vec![Script::Stream {
+            account_id: "acct_first",
+            items: vec![Err(transient_rejection().with_transient_retry(
+                NonZeroU32::MIN,
+                Duration::from_secs(60),
+                Duration::from_secs(60),
+                Duration::ZERO,
+                policy,
+            ))],
+        }]);
+        let cancellation = CancellationToken::new();
+        let mut session = block_on(coordinator.start(
+            model_request(&operation, SystemTime::now() + Duration::from_secs(30)),
+            operation,
+            route_plan,
+            None,
+            None,
+            cancellation.clone(),
+        ))
+        .expect("start execution");
+        let cancel = std::thread::spawn(move || {
+            std::thread::sleep(Duration::from_millis(50));
+            cancellation.cancel();
+        });
+        assert!(matches!(
+            block_on(session.collect_uncommitted()),
+            Err(EngineError::Cancelled)
+        ));
+        cancel.join().expect("cancel task");
+        assert_eq!(provider.contexts.lock().expect("contexts").len(), 1);
+    }
 }
 
 #[test]
@@ -4286,6 +4844,85 @@ fn structural_event_before_replay_safe_failure_should_switch_account_before_comm
     assert_eq!(state.finalizations[0].client_status_code, Some(200));
 }
 
+#[test]
+fn delivered_prelude_allows_retry_without_control_until_non_replayable_delivery() {
+    for close_window in [false, true] {
+        let operation = generate_operation();
+        let route_plan = plan(&operation);
+        let prelude = || {
+            ProviderEvent::wire(
+                ProtocolWireEvent::json(
+                    "openai",
+                    Some("response.created".to_owned()),
+                    json!({"type":"response.created","response":{"id":"response-1","output":[]}}),
+                )
+                .unwrap(),
+            )
+            .with_replayable_prelude()
+        };
+        let mut items = vec![Ok(prelude())];
+        if close_window {
+            items.push(Ok(ProviderEvent::wire(
+                ProtocolWireEvent::json(
+                    "openai",
+                    Some("response.future".to_owned()),
+                    json!({"type":"response.future"}),
+                )
+                .unwrap(),
+            )));
+            // 后续前导不能撤销已经交付未知事件的事实
+            items.push(Ok(prelude()));
+        }
+        items.push(Err(ProviderError::new(
+            ProviderErrorKind::Transport,
+            UpstreamSendState::Sent,
+        )
+        .with_replay_safe()));
+        let (coordinator, store, provider) = coordinator(vec![
+            Script::ObservedStream {
+                account_id: "acct_first",
+                items,
+            },
+            Script::Stream {
+                account_id: "acct_second",
+                items: complete_stream(None),
+            },
+        ]);
+        let mut session = block_on(coordinator.start(
+            model_request(&operation, SystemTime::now() + Duration::from_secs(30)),
+            operation,
+            route_plan,
+            None,
+            None,
+            CancellationToken::new(),
+        ))
+        .unwrap();
+        let first = block_on(session.next_event()).unwrap().unwrap();
+        assert!(first.into_provider_events()[0].is_replayable_prelude());
+        block_on(session.commit_downstream(Some(200))).unwrap();
+        let result = loop {
+            match block_on(session.next_event()) {
+                Ok(Some(_)) => {}
+                terminal => break terminal,
+            }
+        };
+        assert_eq!(result.is_err(), close_window);
+        let attempts = if close_window { 1 } else { 2 };
+        assert_eq!(provider.contexts.lock().unwrap().len(), attempts);
+        let state = store.state.lock().unwrap();
+        assert_eq!(state.commits, 1);
+        assert_eq!(state.intermediate_failures, attempts - 1);
+        assert_eq!(
+            state.finalizations[0].outcome,
+            if close_window {
+                ExecutionOutcome::Incomplete
+            } else {
+                ExecutionOutcome::Succeeded
+            }
+        );
+    }
+}
+
 #[test]
 fn cancellation_before_pending_delivery_commit_reaches_terminal_state() {
     let operation = generate_operation();
@@ -5632,6 +6269,8 @@ fn provider_candidate_advance_stops_at_the_account_rotation_budget() {
                 NonZeroU32::new(3).expect("retry budget"),
                 Duration::ZERO,
                 Duration::ZERO,
+                Duration::ZERO,
+                TransientRetryPolicy::RetryThenRotateAccount,
             ))],
         },
         Script::Error(ProviderError::new(

backend/crates/gateway-core/tests/engine/middleware/mod.rs

@@ -509,6 +509,84 @@ fn attempt_passthrough_preserves_parsed_wire_for_every_client_transport() {
     });
 }
 
+#[test]
+fn attempt_sse_encoding_keeps_opaque_metadata_from_creating_extra_frames() {
+    #[derive(Debug)]
+    struct InspectingPlan {
+        expected: &'static str,
+    }
+
+    impl MiddlewarePlan for InspectingPlan {
+        fn handle(
+            &self,
+            _: MiddlewareContext,
+            request: MiddlewareRequest,
+            next: MiddlewareNext,
+        ) -> BoxFuture<'static, Result<MiddlewareResponse, MiddlewareError>> {
+            let expected = self.expected;
+            Box::pin(async move {
+                let (protocol, status, headers, mut body, envelope) =
+                    next.run(request).await?.into_parts();
+                let frame = body.next_frame().await?.expect("provider event");
+                assert_eq!(frame.bytes().as_ref(), expected.as_bytes());
+                Ok(MiddlewareResponse::new(protocol, status, headers, body)
+                    .with_envelope(envelope.expect("provider metadata")))
+            })
+        }
+    }
+
+    futures::executor::block_on(async {
+        for (event_type, id, expected) in [
+            (
+                "future.event",
+                "event-1",
+                "id: event-1\nretry: 10\nevent: future.event\ndata: {\"value\":1}\n\n",
+            ),
+            (
+                "future.event\ndata: injected\n\n",
+                "event-1\r\ndata: injected\r\n\r\n",
+                "retry: 10\ndata: {\"value\":1}\n\n",
+            ),
+            ("", "event\0id", "retry: 10\ndata: {\"value\":1}\n\n"),
+        ] {
+            let plan = FrozenMiddlewarePlan::new(
+                Arc::new(InspectingPlan { expected }),
+                ExtensionSetReference::new(
+                    ExtensionSetId::new("sse-metadata".to_owned()).unwrap(),
+                    Arc::new(ReadyGeneration),
+                ),
+            );
+            let wire = ProtocolWireEvent::json_with_sse_metadata(
+                "openai",
+                Some(event_type.to_owned()),
+                json!({"value":1}),
+                Some(id.to_owned()),
+                Some(10),
+            )
+            .unwrap();
+            let terminal: ProviderMiddlewareTerminal = Box::new(move |_, _| {
+                Box::pin(async move {
+                    Ok(ProviderStream::new(
+                        empty_provider_stream().metadata().clone(),
+                        stream::iter([Ok(ProviderEvent::wire(wire))]),
+                        (),
+                    ))
+                })
+            });
+
+            execute_attempt_middleware(
+                Some(&plan),
+                context(),
+                operation(),
+                ClientTransport::HttpSse,
+                terminal,
+            )
+            .await
+            .unwrap();
+        }
+    });
+}
+
 #[test]
 fn attempt_middleware_preserves_multi_value_and_authentication_headers() {
     futures::executor::block_on(async {

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

@@ -8,7 +8,6 @@ mod execution;
 mod extensions;
 mod middleware;
 mod policy;
-mod probe;
 mod provider;
 mod response_control;
 

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

@@ -1,9 +0,0 @@
-//! 验证账号探测可通过对象安全的执行链端口调用
-
-use gateway_core::engine::probe::AccountProbe;
-
-#[test]
-fn account_probe_is_an_object_safe_execution_chain_port() {
-    fn accepts(_: Option<&dyn AccountProbe>) {}
-    accepts(None);
-}

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

@@ -1,24 +1,10 @@
-//! 验证健康探针的对象安全接口与 Worker 身份投影
+//! 验证 Worker 健康状态的身份投影
 
-use futures::future::BoxFuture;
 use gateway_core::health::{
-    HealthProbe, HealthState, WorkerHealthKey, WorkerHealthSnapshot, WorkerHealthSource,
-    WorkerRuntimeState,
+    WorkerHealthKey, WorkerHealthSnapshot, WorkerHealthSource, WorkerRuntimeState,
 };
 use gateway_core::task::{WorkerId, WorkerKind};
 
-struct HealthyProbe;
-
-impl HealthProbe for HealthyProbe {
-    fn name(&self) -> &'static str {
-        "test"
-    }
-
-    fn check(&self) -> BoxFuture<'_, HealthState> {
-        Box::pin(async { HealthState::Healthy })
-    }
-}
-
 struct SingleSnapshot;
 
 impl WorkerHealthSource for SingleSnapshot {
@@ -38,13 +24,6 @@ impl WorkerHealthSource for SingleSnapshot {
     }
 }
 
-#[test]
-fn health_probe_contract_is_object_safe() {
-    let probe: &dyn HealthProbe = &HealthyProbe;
-
-    assert_eq!(probe.name(), "test");
-}
-
 #[test]
 fn worker_health_source_preserves_worker_identity() {
     let source: &dyn WorkerHealthSource = &SingleSnapshot;

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

@@ -315,14 +315,6 @@ fn connection_registration_rejects_after_drain_linearization() {
     assert!(matches!(lifecycle.try_register(), Err(ConnectionDraining)));
 }
 
-#[test]
-fn connection_lifecycle_contract_is_object_safe() {
-    let lifecycle = TestLifecycle::new();
-    let object: &dyn ConnectionLifecycle = &lifecycle;
-
-    assert!(!object.is_draining());
-}
-
 #[test]
 fn cancellation_reason_survives_cleanup_and_propagates_only_to_descendants() {
     use gateway_core::lifecycle::{CancellationReason, LeaseFailure, RequestLeaseKind};

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

@@ -109,68 +109,6 @@ fn middleware_protocol_change_keeps_source_capability_facts_without_interpreting
     assert_eq!(replaced.capability_requirements(), source_requirements);
 }
 
-#[test]
-fn native_encoding_keeps_admission_context_and_session_after_optional_translation() {
-    let context = Map::from_iter([("conversation_id".to_owned(), json!("conversation"))]);
-    let session = ProviderSessionState::new(
-        "xai",
-        Map::from_iter([("session_id".to_owned(), json!("session"))]),
-    )
-    .unwrap();
-    let request = GenerateRequest::from_protocol_payload(
-        ProtocolPayload::json_object(
-            "openai",
-            json!({"input":"hello", "max_output_tokens":64, "reasoning":{"effort":"high"}})
-                .as_object()
-                .unwrap()
-                .clone(),
-        )
-        .unwrap()
-        .with_context(context.clone()),
-    )
-    .with_provider_session_state(session.clone());
-    let original = Operation::Generate(request.clone());
-    let translated = original
-        .clone()
-        .replace_middleware_wire("native-input", Bytes::from_static(br#"{"prompt":"hello"}"#))
-        .unwrap();
-    let Operation::Generate(translated) = translated else {
-        panic!("generate request");
-    };
-
-    for (request, target) in [(request, "openai"), (translated, "xai")] {
-        let body = Map::from_iter([("native_prompt".to_owned(), json!("encoded"))]);
-        let encoded = request
-            .with_native_encoded_body(target, body.clone())
-            .unwrap();
-        assert_eq!(encoded.protocol_payload().protocol(), target);
-        assert_eq!(encoded.protocol_payload().body(), &body);
-        assert_eq!(encoded.protocol_payload().context(), &context);
-        assert_eq!(encoded.provider_session_state("xai"), Some(&session));
-        let operation = Operation::Generate(encoded);
-        assert_eq!(
-            operation.capability_requirements(),
-            original.capability_requirements()
-        );
-        assert!(
-            operation
-                .replace_middleware_wire("another", Bytes::from_static(b"{}"))
-                .is_err()
-        );
-    }
-}
-
-#[test]
-fn native_encoding_rejects_invalid_protocol_names() {
-    for name in ["", "invalid\nprotocol"] {
-        assert!(
-            generate(json!({"input":"hello"}))
-                .with_native_encoded_body(name, Map::new())
-                .is_err()
-        );
-    }
-}
-
 #[test]
 fn middleware_processing_recomputes_requirements_but_protocol_change_is_direct_once() {
     let operation = Operation::Generate(generate(json!({
@@ -336,68 +274,6 @@ fn provider_http_should_reject_urls_controls_and_unbounded_headers() {
     );
 }
 
-#[test]
-fn provider_http_wire_replacement_keeps_frozen_target_and_validates_typed_headers() {
-    let operation = Operation::ProviderHttp(
-        ProviderHttpRequest::new(
-            "inspect",
-            ProviderHttpMethod::Post,
-            Some("mode=stable".to_owned()),
-            vec![ProviderHttpHeader::new(
-                "x-original",
-                Bytes::from_static(b"old"),
-            )],
-            RawHttpPayload::new("provider-http", Bytes::from_static(b"old-body"))
-                .expect("provider HTTP payload"),
-        )
-        .expect("provider HTTP request"),
-    );
-    assert!(
-        operation
-            .clone()
-            .replace_provider_http_wire(
-                None,
-                Map::new(),
-                vec![ProviderHttpHeader::new(
-                    "x-invalid",
-                    Bytes::from_static(b"line\r\nbreak"),
-                )],
-            )
-            .is_err()
-    );
-    let replacement = operation
-        .replace_provider_http_wire(
-            Some(Bytes::from_static(&[0x00, 0xff, b'n', b'e', b'w'])),
-            Map::from_iter([("connection_fact".to_owned(), json!("kept"))]),
-            vec![
-                ProviderHttpHeader::new("x-repeat", Bytes::from_static(&[0xff, 0xfe])),
-                ProviderHttpHeader::new("x-repeat", Bytes::from_static(b"two")),
-            ],
-        )
-        .expect("typed Provider HTTP replacement");
-    let Operation::ProviderHttp(request) = replacement else {
-        panic!("operation kind changed")
-    };
-    assert_eq!(request.endpoint(), "inspect");
-    assert_eq!(request.method(), ProviderHttpMethod::Post);
-    assert_eq!(request.query(), Some("mode=stable"));
-    assert_eq!(
-        request.payload().body().as_ref(),
-        &[0x00, 0xff, b'n', b'e', b'w']
-    );
-    assert_eq!(request.payload().context()["connection_fact"], "kept");
-    assert_eq!(request.headers().len(), 2);
-    assert_eq!(request.headers()[0].value().as_ref(), &[0xff, 0xfe]);
-    assert_eq!(request.headers()[1].value().as_ref(), b"two");
-
-    let generate = Operation::Generate(generate(json!({"model":"gpt-test"})));
-    assert!(
-        generate
-            .replace_provider_http_wire(None, Map::new(), Vec::new())
-            .is_err()
-    );
-}
-
 #[test]
 fn standalone_search_should_keep_raw_body_and_context_opaque() {
     let body = Bytes::from_static(

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

@@ -1,4 +1,4 @@
-//! 隐私策略在快照与请求计划中的冻结、复用及关闭行为
+//! 隐私策略的数据合同，以及快照与请求计划中的冻结、复用及关闭行为
 
 use super::*;
 use gateway_core::settings::privacy::*;
@@ -16,7 +16,7 @@ impl PrivacyPolicyCompiler for Compiler {
         policy: &CodexPrivacyPolicy,
     ) -> Result<Arc<dyn CompiledPrivacyPolicy>, PrivacyError> {
         self.0.fetch_add(1, Ordering::SeqCst);
-        Ok(Arc::new(Compiled(policy.rules[0].value.clone())))
+        Ok(Arc::new(Compiled(policy.rules[0].steps[0].value.clone())))
     }
 }
 
@@ -41,19 +41,41 @@ fn policy(value: &str) -> CodexPrivacyPolicy {
             id: "rule".into(),
             name: "规则".into(),
             enabled: true,
-            scope: PrivacyScope::RequestBody,
-            selector: "$.metadata".into(),
-            action: PrivacyAction::SetValue,
-            pattern: None,
-            replacement: String::new(),
-            value: serde_json::json!(value),
-            replace_all: false,
-            case_insensitive: false,
-            multi_line: false,
+            steps: vec![PrivacyStep {
+                scope: PrivacyScope::RequestBody,
+                selector: "$.metadata".into(),
+                action: PrivacyAction::SetValue,
+                pattern: None,
+                replacement: String::new(),
+                value: serde_json::json!(value),
+                replace_all: false,
+                case_insensitive: false,
+                multi_line: false,
+            }],
         }],
     }
 }
 
+#[test]
+fn privacy_rule_requires_steps_and_rejects_unknown_fields() {
+    let expected = policy("user").rules.remove(0);
+    let value = serde_json::to_value(&expected).unwrap();
+    assert_eq!(
+        serde_json::from_value::<PrivacyRule>(value.clone()).unwrap(),
+        expected
+    );
+
+    let mut missing_steps = value.clone();
+    missing_steps.as_object_mut().unwrap().remove("steps");
+    assert!(serde_json::from_value::<PrivacyRule>(missing_steps).is_err());
+    let mut unknown_rule_field = value.clone();
+    unknown_rule_field["scope"] = serde_json::json!("request_body");
+    assert!(serde_json::from_value::<PrivacyRule>(unknown_rule_field).is_err());
+    let mut unknown_step_field = value;
+    unknown_step_field["steps"][0]["unexpected"] = serde_json::json!(true);
+    assert!(serde_json::from_value::<PrivacyRule>(unknown_step_field).is_err());
+}
+
 fn compiled(snapshot: &RuntimeSnapshot) -> Option<Arc<dyn CompiledPrivacyPolicy>> {
     snapshot
         .plan_provider_endpoint(

backend/crates/gateway-core/tests/runtime/mod.rs

@@ -324,25 +324,6 @@ fn publisher_should_suspend_but_still_notify_after_committed_refresh_failure() {
     );
 }
 
-#[test]
-fn snapshot_ports_and_control_should_remain_object_safe() {
-    fn accept_store(_: &dyn SnapshotStorePort) {}
-    fn accept_subscriptions(_: &dyn SnapshotSubscriptionPort) {}
-    fn accept_control(_: &dyn SnapshotControl) {}
-
-    let store = Arc::new(TestSnapshotStore::new(Ok(facts(1, 1))));
-    let subscriptions = Arc::new(TestSnapshotSubscriptions::default());
-    let publisher = RuntimeSnapshotPublisher::new(
-        Arc::new(compiler(store.clone())),
-        RuntimeSnapshotHandle::new(empty_snapshot(1)),
-        subscriptions.clone(),
-    );
-
-    accept_store(store.as_ref());
-    accept_subscriptions(subscriptions.as_ref());
-    accept_control(&publisher);
-}
-
 #[test]
 fn publisher_should_contribute_reconciliation_and_subscription_workers() {
     let store = Arc::new(TestSnapshotStore::new(Ok(facts(1, 1))));

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

@@ -6,7 +6,7 @@ use std::time::Duration;
 use futures::future::BoxFuture;
 use gateway_core::lifecycle::CancellationToken;
 use gateway_core::task::{
-    DaemonRestartPolicy, DaemonTask, ScheduledTask, WorkerCycleContext, WorkerDefinitionError,
+    DaemonRestartPolicy, ScheduledTask, WorkerCycleContext, WorkerDefinitionError,
     WorkerFencingToken, WorkerId, WorkerKind, WorkerLeaseRequest, WorkerRegistration,
     WorkerRunnable, WorkerSchedule, WorkerTaskError,
 };
@@ -22,14 +22,6 @@ impl ScheduledTask for NoopScheduledTask {
     }
 }
 
-struct NoopDaemonTask;
-
-impl DaemonTask for NoopDaemonTask {
-    fn run(&self, _cancellation: CancellationToken) -> BoxFuture<'_, Result<(), WorkerTaskError>> {
-        Box::pin(async { Ok(()) })
-    }
-}
-
 #[test]
 fn schedule_rejects_every_invalid_duration_boundary() {
     let one = Duration::from_nanos(1);
@@ -135,15 +127,6 @@ fn worker_cycle_context_preserves_optional_fencing_token() {
     assert_eq!(context.fencing_token(), Some(token));
 }
 
-#[test]
-fn scheduled_and_daemon_tasks_are_object_safe() {
-    fn accept_scheduled(_task: &dyn ScheduledTask) {}
-    fn accept_daemon(_task: &dyn DaemonTask) {}
-
-    accept_scheduled(&NoopScheduledTask);
-    accept_daemon(&NoopDaemonTask);
-}
-
 fn worker_id(owner: &str) -> WorkerId {
     WorkerId::try_new(WorkerKind::OAuthRefresh, owner).expect("valid worker identity")
 }

backend/crates/gateway-host/Cargo.toml

@@ -27,7 +27,7 @@ hyper-util = { version = "0.1.20", features = ["client-legacy", "client-proxy",
 ipnet = "2.11.0"
 percent-encoding = "2.3.2"
 reqwest = { version = "=0.12.28", default-features = false, features = ["json", "stream", "rustls-tls-native-roots", "socks"] }
-rustls = { version = "=0.23.45", default-features = false, features = ["aws_lc_rs", "std", "tls12"] }
+rustls = { version = "=0.23.46", default-features = false, features = ["aws_lc_rs", "std", "tls12"] }
 rustls-native-certs = "=0.8.3"
 scraper = "0.27.0"
 semver = "1.0.27"

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

@@ -24,6 +24,7 @@ const MAXIMUM_RESPONSE_BYTES: usize = 512 * 1024;
 const DYNAMIC_CACHE_SECONDS: i64 = 300;
 const FALLBACK_CACHE_SECONDS: u64 = 60;
 const MINIMUM_REMAINING_LINK_SECONDS: i64 = 600;
+const OFFICIAL_METADATA_TIMEOUT: Duration = Duration::from_secs(3);
 
 const OFFICIAL_X64_URL: &str = "https://persistent.oaistatic.com/codex-app-prod/ChatGPT-x64.msix";
 const OFFICIAL_ARM64_URL: &str =
@@ -53,7 +54,7 @@ impl RgAdguardClientDistribution {
     pub(crate) fn new() -> Self {
         let client = Client::builder()
             .connect_timeout(Duration::from_secs(5))
-            .timeout(Duration::from_secs(12))
+            .timeout(Duration::from_secs(30))
             .redirect(Policy::none())
             .build()
             .inspect_err(|_| {
@@ -69,10 +70,11 @@ impl RgAdguardClientDistribution {
     }
 
     async fn resolve_uncached(&self) -> (CodexDesktopWindowsDownloads, Duration) {
-        let now = Utc::now();
-        match self.fetch_store_packages(now).await {
+        match self.fetch_store_packages().await {
             Ok(packages) => {
-                let (packages, warning) = complete_with_official_fallback(packages);
+                let (mut packages, warning) = complete_with_official_fallback(packages);
+                self.fill_official_package_sizes(&mut packages).await;
+                let now = Utc::now();
                 let valid_seconds = packages
                     .iter()
                     .filter_map(|package| package.expires_at)
@@ -97,17 +99,33 @@ impl RgAdguardClientDistribution {
                     error_kind = ?error_kind,
                     "RG-Adguard package resolution failed; using official downloads"
                 );
-                (
-                    official_fallback(now),
-                    Duration::from_secs(FALLBACK_CACHE_SECONDS),
-                )
+                let mut downloads = official_fallback(Utc::now());
+                self.fill_official_package_sizes(&mut downloads.packages)
+                    .await;
+                downloads.resolved_at = Utc::now();
+                (downloads, Duration::from_secs(FALLBACK_CACHE_SECONDS))
             }
         }
     }
 
+    async fn fill_official_package_sizes(&self, packages: &mut [ClientDownloadPackage]) {
+        let Some(client) = self.client.as_ref() else {
+            return;
+        };
+        // 只补充回退包的展示大小，超时或缺失长度不能撤销官方稳定下载地址
+        futures::future::join_all(
+            packages
+                .iter_mut()
+                .filter(|package| package.source == ClientDownloadSource::OfficialOpenAi)
+                .map(|package| async {
+                    package.size_bytes = official_package_size(client, &package.download_url).await;
+                }),
+        )
+        .await;
+    }
+
     async fn fetch_store_packages(
         &self,
-        now: DateTime<Utc>,
     ) -> Result<Vec<ClientDownloadPackage>, ClientDistributionResolutionErrorKind> {
         let client = self
             .client
@@ -145,7 +163,8 @@ impl RgAdguardClientDistribution {
         }
         let html = std::str::from_utf8(&bytes)
             .map_err(|_| ClientDistributionResolutionErrorKind::InvalidResponse)?;
-        parse_store_packages(html, now)
+        // 上游解析可能耗时几十秒，按收到结果时的时间检查剩余有效期
+        parse_store_packages(html, Utc::now())
     }
 }
 
@@ -318,11 +337,12 @@ fn parse_size_bytes(value: &str) -> Option<u64> {
     if parts.next().is_some() {
         return None;
     }
+    // RG 页面以 KB/MB/GB 标记二进制大小，换算后才能与 CDN Content-Length 对齐
     let multiplier = match unit {
         "B" => 1_u128,
-        "KB" => 1_000,
-        "MB" => 1_000_000,
-        "GB" => 1_000_000_000,
+        "KB" => 1_024,
+        "MB" => 1_048_576,
+        "GB" => 1_073_741_824,
         _ => return None,
     };
     let (whole, fraction) = amount.split_once('.').unwrap_or((amount.as_str(), ""));
@@ -351,7 +371,7 @@ fn official_fallback(now: DateTime<Utc>) -> CodexDesktopWindowsDownloads {
         resolved_at: now,
         cached: false,
         warning: Some(
-            "Microsoft Store 离线直链暂不可用，已切换为 OpenAI 官方稳定安装包。".to_owned(),
+            "Microsoft Store 离线直链暂不可用，已切换为 OpenAI 官方稳定安装包".to_owned(),
         ),
         packages,
     }
@@ -375,7 +395,7 @@ pub fn complete_with_official_fallback(
     }
     let warning = (!missing.is_empty()).then(|| {
         format!(
-            "{} 的 Microsoft Store 临时直链不可用，已切换为 OpenAI 官方稳定安装包。",
+            "{} 的 Microsoft Store 临时直链不可用，已切换为 OpenAI 官方稳定安装包",
             missing.join("、")
         )
     });
@@ -397,3 +417,25 @@ fn official_package(architecture: ClientArchitecture) -> ClientDownloadPackage {
         expires_at: None,
     }
 }
+
+/// 读取官方安装包的响应头大小，不下载包字节
+pub async fn official_package_size(client: &Client, download_url: &str) -> Option<u64> {
+    let response = client
+        .head(download_url)
+        .timeout(OFFICIAL_METADATA_TIMEOUT)
+        .send()
+        .await
+        .ok()?;
+    if !response.status().is_success() {
+        return None;
+    }
+    // HEAD 没有正文，不能用正文长度推断实际安装包大小
+    response
+        .headers()
+        .get(reqwest::header::CONTENT_LENGTH)?
+        .to_str()
+        .ok()?
+        .parse::<u64>()
+        .ok()
+        .filter(|size| *size > 0)
+}

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

@@ -72,7 +72,7 @@ fn missing_config_field(error: &config::ConfigError) -> Option<String> {
 }
 
 /// Host 唯一拥有的进程配置
-#[derive(Debug, Clone, Deserialize, PartialEq, Eq)]
+#[derive(Debug, Clone, Deserialize, PartialEq)]
 pub struct HostConfig {
     #[serde(default)]
     pub timezone: gateway_core::time::DeploymentTimeZone,
@@ -145,7 +145,7 @@ pub struct ListenConfig {
 }
 
 /// Host 结构化日志配置
-#[derive(Debug, Clone, Deserialize, PartialEq, Eq)]
+#[derive(Debug, Clone, Deserialize, PartialEq)]
 pub struct LoggingConfig {
     pub level: String,
     pub stdout: bool,
@@ -157,17 +157,17 @@ pub struct LoggingConfig {
     /// 请求报文的留存窗口由 request_dump_retention_days 独立控制
     #[serde(default)]
     pub request_dump: bool,
-    /// 完整报文的最少保留天数；按 UTC 日期整组清理，默认 1 天
+    /// 完整报文的最少保留天数；支持小数，每天按 24 小时换算，默认 1 天
     #[serde(default = "default_request_dump_retention_days")]
-    pub request_dump_retention_days: usize,
+    pub request_dump_retention_days: f64,
 }
 
 fn default_file_retention_days() -> usize {
     7
 }
 
-fn default_request_dump_retention_days() -> usize {
-    1
+fn default_request_dump_retention_days() -> f64 {
+    1.0
 }
 
 impl LoggingConfig {
@@ -190,14 +190,19 @@ impl LoggingConfig {
                 "host.logging.file.max_file_size_mb",
             ));
         }
-        if self.request_dump_retention_days == 0 {
-            return Err(ConfigError::InvalidField(
-                "host.logging.request_dump_retention_days",
-            ));
-        }
+        self.request_dump_retention()?;
         resolve_relative_path(source_dir, &mut self.file.directory);
         Ok(())
     }
+
+    pub(crate) fn request_dump_retention(&self) -> Result<Duration, ConfigError> {
+        Duration::try_from_secs_f64(self.request_dump_retention_days * 86_400.0)
+            .ok()
+            .filter(|duration| !duration.is_zero())
+            .ok_or(ConfigError::InvalidField(
+                "host.logging.request_dump_retention_days",
+            ))
+    }
 }
 
 /// Host 文件日志配置

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

@@ -15,7 +15,7 @@ mod sink;
 mod writer;
 
 use sink::{FileLogGuard, FileLogSink, LogHealth};
-use writer::RotatingLogWriter;
+use writer::{LogRetention, RotatingLogWriter};
 
 // 文件前缀统一为 codex-proxy-rs-<kebab-case 类别>；专用 tracing target 使用 snake_case
 const APPLICATION_LOG_FILE_PREFIX: &str = "codex-proxy-rs-application";
@@ -45,9 +45,11 @@ pub enum LogError {
     AlreadyInitialized,
     #[error("logging size limit is too large")]
     SizeOverflow,
+    #[error(transparent)]
+    InvalidConfig(#[from] crate::config::ConfigError),
 }
 
-/// 初始化日志；文件按完整日期留存，不以分片数量淘汰窗口内记录
+/// 初始化日志；按各通道的保留窗口清理，不以分片数量淘汰记录
 pub fn initialize_logging(
     config: &LoggingConfig,
     timezone: gateway_core::time::DeploymentTimeZone,
@@ -74,7 +76,7 @@ pub fn initialize_logging(
         let (writer, guard) = create_file_writer(
             config,
             timezone,
-            config.file.retention_days,
+            LogRetention::CalendarDays(config.file.retention_days),
             APPLICATION_LOG_FILE_PREFIX,
             "gateway-log-application-file",
             Arc::clone(&health),
@@ -88,7 +90,7 @@ pub fn initialize_logging(
         let (writer, guard) = create_file_writer(
             config,
             timezone,
-            config.file.retention_days,
+            LogRetention::CalendarDays(config.file.retention_days),
             OAUTH_RECOVERY_LOG_FILE_PREFIX,
             "gateway-log-oauth-recovery-file",
             Arc::clone(&health),
@@ -102,7 +104,7 @@ pub fn initialize_logging(
         let (writer, guard) = create_file_writer(
             config,
             timezone,
-            config.request_dump_retention_days,
+            LogRetention::Elapsed(config.request_dump_retention()?),
             REQUEST_DUMP_LOG_FILE_PREFIX,
             "gateway-log-request-dump-file",
             Arc::clone(&health),
@@ -220,7 +222,7 @@ impl LogGuard {
 fn create_file_writer(
     config: &LoggingConfig,
     timezone: gateway_core::time::DeploymentTimeZone,
-    retention_days: usize,
+    retention: LogRetention,
     prefix: &'static str,
     thread_name: &'static str,
     health: Arc<LogHealth>,
@@ -234,7 +236,7 @@ fn create_file_writer(
         config.file.directory.clone(),
         prefix,
         maximum_bytes,
-        retention_days,
+        retention,
         Arc::clone(&health),
         timezone,
     )?;

backend/crates/gateway-host/src/logging/writer.rs

@@ -1,22 +1,30 @@
-//! 按大小轮转、压缩已关闭分片、按部署时区自然日期组清理
+//! 按大小与时间轮转、压缩已关闭分片，按自然日期或实际时长清理
 
 use std::collections::BTreeMap;
 use std::fs::{self, File, OpenOptions};
 use std::io::{self, BufWriter, Write};
 use std::path::{Path, PathBuf};
 use std::sync::Arc;
+use std::time::{Duration, Instant, SystemTime};
 
 use chrono::{NaiveDate, Utc};
 use flate2::{Compression, write::GzEncoder};
 
 use super::sink::LogHealth;
 
+#[derive(Clone, Copy)]
+pub(super) enum LogRetention {
+    CalendarDays(usize),
+    Elapsed(Duration),
+}
+
 pub(super) struct RotatingLogWriter {
     timezone: gateway_core::time::DeploymentTimeZone,
     directory: PathBuf,
     prefix: &'static str,
     maximum_bytes: u64,
-    retention_days: usize,
+    retention: LogRetention,
+    opened_at: Instant,
     date: NaiveDate,
     segment: usize,
     bytes_written: u64,
@@ -29,19 +37,23 @@ impl RotatingLogWriter {
         directory: PathBuf,
         prefix: &'static str,
         maximum_bytes: u64,
-        retention_days: usize,
+        retention: LogRetention,
         health: Arc<LogHealth>,
         timezone: gateway_core::time::DeploymentTimeZone,
     ) -> io::Result<Self> {
         fs::create_dir_all(&directory)?;
         let date = timezone.local(Utc::now()).date_naive();
-        cleanup_log_files(&directory, prefix, date, retention_days, timezone)?;
+        cleanup_log_files(&directory, prefix, retention, timezone, None)?;
         let latest = managed_log_files(&directory, prefix)?
             .into_iter()
             .filter(|entry| entry.date == date)
             .max_by_key(|entry| (entry.segment, entry.compressed));
         let (segment, bytes_written) = match latest {
-            Some(entry) if !entry.compressed && entry.path.metadata()?.len() < maximum_bytes => {
+            Some(entry)
+                if matches!(retention, LogRetention::CalendarDays(_))
+                    && !entry.compressed
+                    && entry.path.metadata()?.len() < maximum_bytes =>
+            {
                 (entry.segment, entry.path.metadata()?.len())
             }
             Some(entry) => (next_segment(entry.segment)?, 0),
@@ -54,7 +66,8 @@ impl RotatingLogWriter {
             directory,
             prefix,
             maximum_bytes,
-            retention_days,
+            retention,
+            opened_at: Instant::now(),
             date,
             segment,
             bytes_written,
@@ -78,7 +91,14 @@ impl RotatingLogWriter {
         let day_changed = date != self.date;
         let size_exceeded = self.bytes_written > 0
             && self.bytes_written.saturating_add(incoming_bytes as u64) > self.maximum_bytes;
-        if !day_changed && !size_exceeded {
+        // 小流量转储也需及时关闭分片，避免小数保留期仍要等自然日切换才能生效
+        let age_exceeded = match self.retention {
+            LogRetention::CalendarDays(_) => false,
+            LogRetention::Elapsed(duration) => {
+                self.opened_at.elapsed() >= duration.min(Duration::from_secs(3600))
+            }
+        };
+        if !day_changed && !size_exceeded && !age_exceeded {
             return Ok(());
         }
         let previous = self
@@ -104,6 +124,7 @@ impl RotatingLogWriter {
         self.date = date;
         self.segment = segment;
         self.bytes_written = self.file.get_ref().metadata()?.len();
+        self.opened_at = Instant::now();
         if let Err(error) = compress_log_file(&previous) {
             self.health.maintenance_failed(error.kind());
         }
@@ -111,9 +132,9 @@ impl RotatingLogWriter {
         if let Err(error) = cleanup_log_files(
             &self.directory,
             self.prefix,
-            date,
-            self.retention_days,
+            self.retention,
             self.timezone,
+            Some(&path),
         ) {
             self.health.maintenance_failed(error.kind());
         }
@@ -203,17 +224,38 @@ fn managed_log_files(directory: &Path, prefix: &str) -> io::Result<Vec<ManagedLo
 fn cleanup_log_files(
     directory: &Path,
     prefix: &str,
-    today: NaiveDate,
-    retention_days: usize,
+    retention: LogRetention,
     timezone: gateway_core::time::DeploymentTimeZone,
+    active_path: Option<&Path>,
 ) -> io::Result<()> {
+    let files = managed_log_files(directory, prefix)?;
+    let retention_days = match retention {
+        LogRetention::CalendarDays(days) => days,
+        LogRetention::Elapsed(duration) => {
+            let now = SystemTime::now();
+            for entry in files {
+                if active_path == Some(entry.path.as_path()) {
+                    continue;
+                }
+                let modified = entry.path.metadata()?.modified()?;
+                if now
+                    .duration_since(modified)
+                    .is_ok_and(|age| age >= duration)
+                {
+                    fs::remove_file(entry.path)?;
+                }
+            }
+            return Ok(());
+        }
+    };
+    let today = timezone.local(Utc::now()).date_naive();
     let days =
         u64::try_from(retention_days).map_err(|_| io::Error::other("log retention overflow"))?;
     let cutoff = today
         .checked_sub_days(chrono::Days::new(days))
         .unwrap_or(NaiveDate::MIN);
     let mut dates = BTreeMap::<NaiveDate, Vec<PathBuf>>::new();
-    for entry in managed_log_files(directory, prefix)? {
+    for entry in files {
         dates.entry(entry.date).or_default().push(entry.path);
     }
     for (date, paths) in dates {

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

@@ -39,7 +39,11 @@ pub struct HttpPluginDistribution {
 
 impl HttpPluginDistribution {
     pub fn new(http: Arc<HttpClient>) -> Result<Self, AdminError> {
-        Self::with_transport("https://api.github.com/", http, source_network()?)
+        Self::with_transport(
+            "https://api.github.com/",
+            http,
+            NetworkPolicy::unrestricted(),
+        )
     }
 
     /// 测试可注入受限网络；生产来源由管理员显式配置，因此保持既有内外网可达范围
@@ -173,10 +177,6 @@ impl PluginDistribution for HttpPluginDistribution {
     }
 }
 
-fn source_network() -> Result<NetworkPolicy, AdminError> {
-    NetworkPolicy::new(&["0.0.0.0/0".into(), "::/0".into()]).map_err(|_| source_error())
-}
-
 fn credential_ids(credentials: &[SourceCredential]) -> Vec<String> {
     credentials
         .iter()

backend/crates/gateway-host/tests/client_distribution.rs

@@ -30,7 +30,7 @@ fn parser_selects_latest_valid_package_for_each_architecture() {
     assert_eq!(packages[0].architecture, ClientArchitecture::X64);
     assert_eq!(packages[0].version.as_deref(), Some("26.825.6671.0"));
     assert!(packages[0].download_url.starts_with("http://"));
-    assert_eq!(packages[0].size_bytes, Some(744_250_000));
+    assert_eq!(packages[0].size_bytes, Some(780_402_688));
     assert_eq!(packages[1].architecture, ClientArchitecture::Arm64);
     assert!(packages[1].download_url.starts_with("https://"));
 }
@@ -72,3 +72,74 @@ fn missing_architecture_uses_only_its_official_fallback() {
     assert_eq!(packages[1].source, ClientDownloadSource::OfficialOpenAi);
     assert!(warning.is_some_and(|value| value.contains("arm64")));
 }
+
+#[tokio::test]
+async fn official_package_size_reads_head_length_without_downloading_the_body() {
+    use tokio::io::{AsyncReadExt as _, AsyncWriteExt as _};
+
+    let listener = tokio::net::TcpListener::bind("127.0.0.1:0").await.unwrap();
+    let url = format!("http://{}/ChatGPT-x64.msix", listener.local_addr().unwrap());
+    let server = tokio::spawn(async move {
+        let (mut stream, _) = listener.accept().await.unwrap();
+        let mut request = Vec::new();
+        let mut buffer = [0; 1024];
+        while !request.windows(4).any(|bytes| bytes == b"\r\n\r\n") {
+            let read = stream.read(&mut buffer).await.unwrap();
+            assert!(read > 0);
+            request.extend_from_slice(&buffer[..read]);
+        }
+        assert!(request.starts_with(b"HEAD /ChatGPT-x64.msix HTTP/1.1\r\n"));
+        // 只发送响应头，不发送大包正文，证明解析使用文件长度而不是 HEAD 正文长度
+        stream
+            .write_all(b"HTTP/1.1 200 OK\r\ncontent-length: 944929441\r\nconnection: close\r\n\r\n")
+            .await
+            .unwrap();
+    });
+    let client = reqwest::Client::builder().no_proxy().build().unwrap();
+    assert_eq!(
+        gateway_host::client_distribution::official_package_size(&client, &url).await,
+        Some(944_929_441)
+    );
+    server.await.unwrap();
+}
+
+#[tokio::test]
+async fn official_package_size_failure_keeps_metadata_optional() {
+    use wiremock::{
+        Mock, MockServer, ResponseTemplate,
+        matchers::{method, path},
+    };
+
+    let server = MockServer::start().await;
+    for (route, response) in [
+        ("/missing.msix", ResponseTemplate::new(404)),
+        ("/empty.msix", ResponseTemplate::new(200)),
+        (
+            "/slow.msix",
+            ResponseTemplate::new(200).set_delay(std::time::Duration::from_secs(4)),
+        ),
+    ] {
+        Mock::given(method("HEAD"))
+            .and(path(route))
+            .respond_with(response)
+            .expect(1)
+            .mount(&server)
+            .await;
+    }
+    let client = reqwest::Client::builder().no_proxy().build().unwrap();
+    for route in ["/missing.msix", "/empty.msix", "/slow.msix"] {
+        assert_eq!(
+            tokio::time::timeout(
+                std::time::Duration::from_millis(3500),
+                gateway_host::client_distribution::official_package_size(
+                    &client,
+                    &format!("{}{route}", server.uri()),
+                ),
+            )
+            .await
+            .expect("metadata timeout must not wait for the slow response"),
+            None
+        );
+    }
+    server.verify().await;
+}

backend/crates/gateway-host/tests/config.rs

@@ -146,7 +146,7 @@ fn logging_rejects_zero_retention_windows() {
     for request_dump in [false, true] {
         let mut config = valid_config();
         if request_dump {
-            config.logging.request_dump_retention_days = 0;
+            config.logging.request_dump_retention_days = 0.0;
         } else {
             config.logging.file.retention_days = 0;
         }
@@ -168,7 +168,7 @@ fn logging_defaults_to_time_retention_and_ignores_removed_count_retention() {
         "file": { "enabled": true, "directory": "logs", "max_file_size_mb": 20 }
     });
     let config: LoggingConfig = serde_json::from_value(value.clone()).unwrap();
-    assert_eq!(config.request_dump_retention_days, 1);
+    assert_eq!(config.request_dump_retention_days, 1.0);
     assert_eq!(config.file.retention_days, 7);
     assert!(!config.oauth_recovery);
     value["file"]["max_files"] = serde_json::json!(20);
@@ -178,6 +178,55 @@ fn logging_defaults_to_time_retention_and_ignores_removed_count_retention() {
     );
 }
 
+#[test]
+fn request_dump_retention_accepts_integer_and_fractional_yaml_days() {
+    for (input, expected) in [("1", 1.0), ("0.5", 0.5), ("0.25", 0.25), ("1.5", 1.5)] {
+        let yaml = format!(
+            "level: info\nstdout: true\nrequest_dump_retention_days: {input}\nfile:\n  enabled: true\n  directory: logs\n  max_file_size_mb: 20\n"
+        );
+        let mut config = valid_config();
+        config.logging = config::Config::builder()
+            .add_source(config::File::from_str(&yaml, config::FileFormat::Yaml))
+            .build()
+            .unwrap()
+            .try_deserialize()
+            .unwrap();
+        config
+            .resolve_and_validate(
+                std::path::Path::new("/srv/gateway"),
+                std::path::Path::new("/srv/gateway/web/dist"),
+            )
+            .unwrap();
+        assert_eq!(config.logging.request_dump_retention_days, expected);
+    }
+}
+
+#[test]
+fn request_dump_retention_rejects_invalid_durations() {
+    for days in [
+        0.0,
+        -0.5,
+        f64::NAN,
+        f64::INFINITY,
+        f64::NEG_INFINITY,
+        f64::MAX,
+        1e-20,
+    ] {
+        let mut config = valid_config();
+        config.logging.request_dump_retention_days = days;
+        let error = config
+            .resolve_and_validate(
+                std::path::Path::new("/srv/gateway"),
+                std::path::Path::new("/srv/gateway/web/dist"),
+            )
+            .unwrap_err();
+        assert!(matches!(
+            error,
+            gateway_host::ConfigError::InvalidField("host.logging.request_dump_retention_days")
+        ));
+    }
+}
+
 #[test]
 fn dedicated_file_channels_can_be_enabled_without_application_sinks() {
     for oauth_recovery in [false, true] {
@@ -220,7 +269,7 @@ fn valid_config() -> HostConfig {
             },
             oauth_recovery: false,
             request_dump: false,
-            request_dump_retention_days: 1,
+            request_dump_retention_days: 1.0,
         },
         system_update,
         worker_shutdown_timeout_seconds: 30,

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

@@ -48,7 +48,7 @@ fn logging_requires_at_least_one_sink() {
             },
             oauth_recovery: false,
             request_dump: false,
-            request_dump_retention_days: 1,
+            request_dump_retention_days: 1.0,
         },
         system_update: SystemUpdateConfig::default(),
         worker_shutdown_timeout_seconds: 30,
@@ -325,7 +325,7 @@ fn logging_config(directory: PathBuf, request_dump: bool) -> HostConfig {
             },
             oauth_recovery: true,
             request_dump,
-            request_dump_retention_days: 1,
+            request_dump_retention_days: 1.0,
         },
         system_update: SystemUpdateConfig::default(),
         worker_shutdown_timeout_seconds: 30,

backend/crates/gateway-host/tests/logging/writer.rs

@@ -1,4 +1,4 @@
-//! 验证日志压缩、轮转与重启后按日期保留完整记录
+//! 验证日志压缩、轮转，以及重启后的自然日与小数天数保留窗口
 
 use super::*;
 use std::time::SystemTime;
@@ -10,12 +10,11 @@ fn retention_preserves_complete_dates_across_compression_rotation_and_restart()
     if env::var_os(CHILD_PROCESS_ENV).is_some() {
         with_file_logging(
             PathBuf::from(env::var_os(LOG_DIRECTORY_ENV).unwrap()),
-            true,
+            false,
             || {
                 let payload = "x".repeat(1024 * 1024);
                 // 三条记录已跨过 1 MiB 轮转边界；保留日期的 25 个分段由下方单独构造
                 for sequence in 0..ROTATION_RECORDS {
-                    tracing::info!(target: REQUEST_DUMP_LOG_TARGET, sequence, payload, "retention record");
                     tracing::info!(target: APPLICATION_LOG_TARGET, sequence, payload, "retention record");
                     tracing::info!(target: OAUTH_RECOVERY_LOG_TARGET, sequence, payload, "retention record");
                 }
@@ -32,7 +31,6 @@ fn retention_preserves_complete_dates_across_compression_rotation_and_restart()
     for (prefix, days) in [
         (APPLICATION_LOG_FILE_PREFIX, 7),
         (OAUTH_RECOVERY_LOG_FILE_PREFIX, 7),
-        (REQUEST_DUMP_LOG_FILE_PREFIX, 1),
     ] {
         let boundary = today - chrono::Days::new(days);
         for segment in 1..=25 {
@@ -90,11 +88,7 @@ fn retention_preserves_complete_dates_across_compression_rotation_and_restart()
         );
         assert!(unrelated.exists());
 
-        for prefix in [
-            APPLICATION_LOG_FILE_PREFIX,
-            OAUTH_RECOVERY_LOG_FILE_PREFIX,
-            REQUEST_DUMP_LOG_FILE_PREFIX,
-        ] {
+        for prefix in [APPLICATION_LOG_FILE_PREFIX, OAUTH_RECOVERY_LOG_FILE_PREFIX] {
             let body = read_log_file_set(directory.path(), prefix);
             let mut sequences = [0; ROTATION_RECORDS];
             let mut boundaries = std::collections::BTreeSet::new();
@@ -141,6 +135,151 @@ fn retention_preserves_complete_dates_across_compression_rotation_and_restart()
     }
 }
 
+#[test]
+fn request_dump_retention_uses_elapsed_time_across_compression_and_restart() {
+    const RETENTION_ENV: &str = "CPR_LOGGING_TEST_RETENTION_DAYS";
+    if env::var_os(CHILD_PROCESS_ENV).is_some() {
+        let mut config =
+            logging_config(PathBuf::from(env::var_os(LOG_DIRECTORY_ENV).unwrap()), true);
+        config.logging.request_dump_retention_days =
+            env::var(RETENTION_ENV).unwrap().parse().unwrap();
+        with_logging(config, || {
+            let payload = "x".repeat(1024 * 1024);
+            for sequence in 0..ROTATION_RECORDS {
+                tracing::info!(target: REQUEST_DUMP_LOG_TARGET, sequence, payload, "elapsed retention record");
+            }
+        });
+        return;
+    }
+
+    for days in [0.25, 0.5, 1.0, 1.5] {
+        let directory = tempfile::tempdir().unwrap();
+        let now = chrono::Utc::now();
+        let cutoff = now - chrono::Duration::seconds((days * 86_400.0) as i64);
+        let mut retained = Vec::new();
+        let mut expired = Vec::new();
+        // 同一日期内既有过期文件也有有效文件，文件名日期不覆盖实际修改时间
+        for (segment, modified, keep) in [
+            (0, cutoff - chrono::Duration::minutes(1), false),
+            (1, cutoff + chrono::Duration::minutes(1), true),
+            (2, now + chrono::Duration::hours(1), true),
+        ] {
+            let path = directory.path().join(format!(
+                "{REQUEST_DUMP_LOG_FILE_PREFIX}{}.{segment}.log",
+                now.date_naive()
+            ));
+            fs::write(&path, format!("seed-{segment}\n")).unwrap();
+            fs::File::open(&path)
+                .unwrap()
+                .set_modified(modified.into())
+                .unwrap();
+            if keep {
+                let modified = path.metadata().unwrap().modified().unwrap();
+                retained.push((path, modified));
+            } else {
+                expired.push(path);
+            }
+        }
+        // 归档的修改时间也参与清理，压缩不能重置保留窗口
+        let archive = directory.path().join(format!(
+            "{REQUEST_DUMP_LOG_FILE_PREFIX}{}.3.log.gz",
+            now.date_naive() - chrono::Days::new(3)
+        ));
+        let mut encoder = flate2::write::GzEncoder::new(
+            fs::File::create(&archive).unwrap(),
+            flate2::Compression::fast(),
+        );
+        std::io::Write::write_all(&mut encoder, b"expired archive\n").unwrap();
+        encoder
+            .finish()
+            .unwrap()
+            .set_modified((cutoff - chrono::Duration::hours(1)).into())
+            .unwrap();
+        expired.push(archive);
+        let unrelated = directory.path().join("unmanaged.log");
+        fs::write(&unrelated, "unmanaged\n").unwrap();
+
+        for run in 1..=2 {
+            let output = Command::new(env::current_exe().unwrap())
+                .args(["--exact", "logging::writer::request_dump_retention_uses_elapsed_time_across_compression_and_restart"])
+                .env(CHILD_PROCESS_ENV, "1")
+                .env(LOG_DIRECTORY_ENV, directory.path())
+                .env(LOG_TIMEZONE_ENV, "America/New_York")
+                .env(RETENTION_ENV, days.to_string())
+                .env("RUST_LOG", "off")
+                .output().unwrap();
+            assert!(
+                output.status.success(),
+                "{}",
+                String::from_utf8_lossy(&output.stderr)
+            );
+            for (path, modified) in &retained {
+                let archive = path.with_extension("log.gz");
+                assert!(archive.exists());
+                assert_eq!(archive.metadata().unwrap().modified().unwrap(), *modified);
+            }
+            assert!(
+                expired
+                    .iter()
+                    .all(|path| !path.exists() && !path.with_extension("log.gz").exists())
+            );
+            assert!(unrelated.exists());
+            let body = read_log_file_set(directory.path(), REQUEST_DUMP_LOG_FILE_PREFIX);
+            assert!(body.contains("seed-1\n") && body.contains("seed-2\n"));
+            assert_eq!(
+                body.matches("elapsed retention record").count(),
+                run * ROTATION_RECORDS
+            );
+        }
+    }
+}
+
+#[test]
+fn request_dump_rotates_small_files_when_the_time_window_elapses() {
+    if env::var_os(CHILD_PROCESS_ENV).is_some() {
+        let directory = PathBuf::from(env::var_os(LOG_DIRECTORY_ENV).unwrap());
+        let mut config = logging_config(directory.clone(), true);
+        config.logging.request_dump_retention_days = 1.0 / 86_400.0;
+        with_logging(config, || {
+            tracing::info!(target: REQUEST_DUMP_LOG_TARGET, "before retention window");
+            let deadline = std::time::Instant::now() + std::time::Duration::from_secs(5);
+            loop {
+                let body = read_log_file_set(&directory, REQUEST_DUMP_LOG_FILE_PREFIX);
+                if body.contains("before retention window") {
+                    break;
+                }
+                assert!(
+                    std::time::Instant::now() < deadline,
+                    "first record was not flushed"
+                );
+                std::thread::sleep(std::time::Duration::from_millis(10));
+            }
+            std::thread::sleep(std::time::Duration::from_millis(1100));
+            tracing::info!(target: REQUEST_DUMP_LOG_TARGET, "after retention window");
+        });
+        return;
+    }
+    let directory = tempfile::tempdir().unwrap();
+    let output = Command::new(env::current_exe().unwrap())
+        .args([
+            "--exact",
+            "logging::writer::request_dump_rotates_small_files_when_the_time_window_elapses",
+        ])
+        .env(CHILD_PROCESS_ENV, "1")
+        .env(LOG_DIRECTORY_ENV, directory.path())
+        .env("RUST_LOG", "off")
+        .output()
+        .unwrap();
+    assert!(
+        output.status.success(),
+        "{}",
+        String::from_utf8_lossy(&output.stderr)
+    );
+    let body = read_log_file_set(directory.path(), REQUEST_DUMP_LOG_FILE_PREFIX);
+    assert!(!body.contains("before retention window"));
+    assert!(body.contains("after retention window"));
+}
+
 fn seed_log(path: &Path, date: chrono::NaiveDate, body: &str) {
     fs::write(path, body).unwrap();
     let modified = gateway_core::time::DeploymentTimeZone::default()

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

@@ -56,7 +56,7 @@ pub fn configuration(directory: &Path) -> HostConfig {
             },
             oauth_recovery: false,
             request_dump: false,
-            request_dump_retention_days: 1,
+            request_dump_retention_days: 1.0,
         },
         system_update: Default::default(),
         worker_shutdown_timeout_seconds: 1,

backend/crates/gateway-plugin/runtime/src/adapter/command_line/execution.rs

@@ -139,7 +139,7 @@ fn rpc_failure(error: RpcError) -> &'static str {
         RpcError::Handshake(_) | RpcError::Protocol(_) | RpcError::InvalidResponse(..) => {
             "插件协议错误"
         }
-        RpcError::Capacity => "插件调用容量不足",
+        RpcError::Capacity | RpcError::Overloaded => "插件调用容量不足",
         RpcError::Context(_) => "调用上下文无效",
         RpcError::Remote(_) => "插件拒绝命令",
     }

backend/crates/gateway-plugin/runtime/src/adapter/middleware/request.rs

@@ -221,6 +221,7 @@ async fn recover_rpc(
         RpcError::Remote(fault) if fault.code == ErrorCode::Rejected
     );
     let delegate_safe = !rejected
+        && !error.is_overloaded()
         && !matches!(
             &error,
             RpcError::Remote(fault) if fault.send_state != SendState::NotSent

backend/crates/gateway-plugin/runtime/src/callback/error.rs

@@ -4,7 +4,7 @@ use gateway_core::{
     engine::{EngineError, execution::gateway_error_from_engine, middleware::MiddlewareError},
     error::{
         ContinuationFailure, ContinuationRecoveryDisposition, GatewayError, GatewayErrorKind,
-        PreDeliveryRetry, ProviderError,
+        PreDeliveryRetry, ProviderError, TransientRetryPolicy,
     },
     upstream::UpstreamSendState,
 };
@@ -20,6 +20,7 @@ pub(super) fn gateway(error: GatewayError) -> PluginFault {
         GatewayErrorKind::Unauthorized => ErrorCode::PermissionDenied,
         GatewayErrorKind::PolicyDenied => ErrorCode::Rejected,
         GatewayErrorKind::AccountCapacityUnavailable
+        | GatewayErrorKind::MiddlewareCapacityUnavailable
         | GatewayErrorKind::ConcurrencyQueueFull
         | GatewayErrorKind::ConcurrencyQueueTimeout
         | GatewayErrorKind::RateLimited
@@ -85,7 +86,15 @@ fn provider(error: &ProviderError) -> PluginFault {
                 gateway_core::engine::AttemptTransport::Fallback => json!({"kind":"fallback"}),
             }}),
             PreDeliveryRetry::SameAccountTransportRetry {retry_index,delay} => json!({"kind":"same_account_transport_retry","retry_index":retry_index.get(),"delay_ms":delay.as_millis()}),
-            PreDeliveryRetry::SameAccountTransientRetry {max_retries,initial_delay,max_delay} => json!({"kind":"same_account_transient_retry","max_retries":max_retries.get(),"initial_delay_ms":initial_delay.as_millis(),"max_delay_ms":max_delay.as_millis()}),
+            PreDeliveryRetry::SameAccountTransientRetry {max_retries,initial_delay,max_delay,jitter,policy} => json!({
+                "kind":"same_account_transient_retry","max_retries":max_retries.get(),
+                "initial_delay_ms":initial_delay.as_millis(),"max_delay_ms":max_delay.as_millis(),
+                "jitter_ms":jitter.as_millis(),
+                "policy": match policy {
+                    TransientRetryPolicy::RetryThenRotateAccount => "retry_then_rotate_account",
+                    TransientRetryPolicy::RetryThenReturnError => "retry_then_return_error",
+                },
+            }),
         }),
         "replay_safe": error.replay_is_safe(),
         "retry_same_account": error.retries_same_account(),
@@ -134,13 +143,22 @@ pub(crate) fn rpc(error: &crate::RpcError) -> PluginFault {
         RpcError::Remote(fault) => return fault.clone(),
         RpcError::Timeout => ErrorCode::Timeout,
         RpcError::Cancelled => ErrorCode::Cancelled,
-        RpcError::Capacity => ErrorCode::Capacity,
+        RpcError::Capacity | RpcError::Overloaded => ErrorCode::Capacity,
         _ => ErrorCode::Fault,
     };
     PluginFault::new(code, error.to_string())
 }
 
 pub(crate) fn rpc_middleware(error: crate::RpcError) -> MiddlewareError {
+    if error.is_overloaded() {
+        return MiddlewareError::Gateway(
+            GatewayError::new(
+                GatewayErrorKind::MiddlewareCapacityUnavailable,
+                "Request middleware is temporarily busy",
+            )
+            .with_source(gateway_core::error::ErrorSource::new(error)),
+        );
+    }
     let rejected =
         matches!(&error, crate::RpcError::Remote(fault) if fault.code == ErrorCode::Rejected);
     MiddlewareError::Remote {

backend/crates/gateway-plugin/runtime/src/generation/prepare/snapshot.rs

@@ -194,7 +194,10 @@ impl PluginRuntime {
             .map_err(|_| AdminError::internal("插件集合 ID 无效"))?;
         let observers = crate::adapter::observer::PluginObserverPlan::compile(
             observer_entries,
-            self.config.rpc_limits.maximum_calls,
+            self.config
+                .rpc_limits
+                .maximum_calls
+                .min(self.config.rpc_limits.maximum_executing_calls),
             self.config.rpc_limits.maximum_call_timeout,
             self.config.rpc_limits.maximum_buffered_body_bytes,
         )

backend/crates/gateway-plugin/runtime/src/rpc/admission.rs

@@ -0,0 +1,50 @@
+//! 区分插件执行名额与初次准入队列，恢复中的调用沿用在途资源预算
+
+use std::{sync::Arc, time::Duration};
+
+use tokio::sync::{OwnedSemaphorePermit, Semaphore};
+
+use super::RpcError;
+
+pub(super) struct ExecutionCapacity {
+    active: Arc<Semaphore>,
+    queued: Arc<Semaphore>,
+    wait_timeout: Duration,
+}
+
+impl ExecutionCapacity {
+    pub(super) fn new(active: usize, queued: usize, wait_timeout: Duration) -> Self {
+        Self {
+            active: Arc::new(Semaphore::new(active)),
+            queued: Arc::new(Semaphore::new(queued)),
+            wait_timeout,
+        }
+    }
+
+    pub(super) async fn enter(&self) -> Result<OwnedSemaphorePermit, RpcError> {
+        if let Ok(permit) = self.active.clone().try_acquire_owned() {
+            return Ok(permit);
+        }
+        let _queued = self.queued.try_acquire().map_err(|error| {
+            if matches!(error, tokio::sync::TryAcquireError::Closed) {
+                RpcError::Closed(None)
+            } else {
+                RpcError::Overloaded
+            }
+        })?;
+        self.resume().await
+    }
+
+    /// 恢复的等待者已持有在途名额，不再竞争初次准入队列
+    pub(super) async fn resume(&self) -> Result<OwnedSemaphorePermit, RpcError> {
+        tokio::time::timeout(self.wait_timeout, self.active.clone().acquire_owned())
+            .await
+            .map_err(|_| RpcError::Overloaded)?
+            .map_err(|_| RpcError::Closed(None))
+    }
+
+    pub(super) fn close(&self) {
+        self.active.close();
+        self.queued.close();
+    }
+}

backend/crates/gateway-plugin/runtime/src/rpc/dispatch.rs

@@ -125,7 +125,7 @@ async fn read_loop<R: AsyncRead + Unpin>(
                     Err(RpcError::Protocol(None))
                 } else if let Some(context) = shared.context(parent_id) {
                     last_callback = id;
-                    if let Some(permit) = shared.try_callback_slot() {
+                    if let Some(permit) = shared.try_callback_slot(&method) {
                         let handler = Arc::clone(&callbacks);
                         let response = Arc::clone(&shared);
                         let (ready, started) = oneshot::channel();
@@ -138,6 +138,12 @@ async fn read_loop<R: AsyncRead + Unpin>(
                                 return;
                             }
                             let result = handler.call(context, method, params, frame.payload).await;
+                            if pause.is_some()
+                                && let Err(error) = response.resume_next(parent_id).await
+                            {
+                                response.cancel(parent_id, error);
+                                return;
+                            }
                             drop(pause);
                             let mut reply = match result {
                                 Ok(reply) => Frame {

backend/crates/gateway-plugin/runtime/src/rpc/mod.rs

@@ -1,5 +1,6 @@
 //! 插件双向 RPC 会话、调用结果与生命周期的公共入口
 
+mod admission;
 mod dispatch;
 mod session;
 

backend/crates/gateway-plugin/runtime/src/rpc/session.rs

@@ -15,17 +15,21 @@ use gateway_plugin_sdk::{
     CallContext, Frame, Handshake, Message, PROTOCOL_VERSION, PluginFault, Stage,
     client::{CallDeadline, CallDeadlinePause, read_frame, validate_frame, write_frame},
 };
-use tokio::sync::{Mutex as AsyncMutex, Semaphore, mpsc, oneshot, watch};
+use tokio::sync::{Mutex as AsyncMutex, OwnedSemaphorePermit, Semaphore, mpsc, oneshot, watch};
 
 use crate::{PreparedPackage, RpcStream, stream::StreamIngress};
 
-use super::dispatch;
+use super::{admission::ExecutionCapacity, dispatch};
 
 #[derive(Debug, Clone, Copy)]
 pub struct RpcLimits {
     /// 受管回调与观察队列的缓冲预算，不是请求正文或 IPC 消息总量上限
     pub maximum_buffered_body_bytes: usize,
+    /// 包含初次排队、下游等待、取消确认和未消费流的总在途上限
     pub maximum_calls: usize,
+    pub maximum_executing_calls: usize,
+    pub maximum_queued_calls: usize,
+    pub maximum_queue_timeout: Duration,
     pub maximum_callbacks: usize,
     pub handshake_timeout: Duration,
     pub maximum_call_timeout: Duration,
@@ -36,8 +40,11 @@ impl Default for RpcLimits {
     fn default() -> Self {
         Self {
             maximum_buffered_body_bytes: 1024 * 1024,
-            maximum_calls: 16,
-            maximum_callbacks: 16,
+            maximum_calls: 256,
+            maximum_executing_calls: 64,
+            maximum_queued_calls: 64,
+            maximum_queue_timeout: Duration::from_secs(30),
+            maximum_callbacks: 64,
             handshake_timeout: Duration::from_secs(5),
             maximum_call_timeout: Duration::from_secs(120),
             maximum_stderr_bytes: 64 * 1024,
@@ -63,6 +70,8 @@ pub enum RpcError {
     Cancelled,
     #[error("plugin call capacity is exhausted")]
     Capacity,
+    #[error("plugin admission capacity is exhausted")]
+    Overloaded,
     #[error("plugin call context is invalid")]
     Context(#[source] Option<gateway_core::error::ErrorSource>),
     #[error("plugin returned an error")]
@@ -70,6 +79,11 @@ pub enum RpcError {
 }
 
 impl RpcError {
+    pub(crate) fn is_overloaded(&self) -> bool {
+        matches!(self, Self::Overloaded)
+            || matches!(self, Self::Remote(fault) if fault.code == gateway_plugin_sdk::ErrorCode::Capacity)
+    }
+
     /// 会话状态与请求错误共用安全分类，不展开插件返回正文
     pub(crate) fn diagnostic(&self) -> (&'static str, &'static str) {
         match self {
@@ -101,6 +115,7 @@ impl RpcError {
             RpcError::Timeout => ("timeout", "插件调用超时"),
             RpcError::Cancelled => ("cancelled", "插件会话已取消"),
             RpcError::Capacity => ("capacity", "插件调用容量已耗尽"),
+            RpcError::Overloaded => ("overloaded", "插件暂时繁忙"),
             RpcError::Context(_) => ("context", "插件调用上下文无效"),
             RpcError::Remote(_) => ("remote", "插件返回错误"),
         }
@@ -216,6 +231,8 @@ pub(crate) struct Shared {
     pub process: ProcessControl,
     pub stopped: watch::Sender<bool>,
     callback_slots: Arc<Semaphore>,
+    next_slots: Arc<Semaphore>,
+    execution: ExecutionCapacity,
 }
 
 struct State {
@@ -224,6 +241,8 @@ struct State {
 }
 
 struct PendingCall {
+    _slot: Arc<OwnedSemaphorePermit>,
+    execution: Option<OwnedSemaphorePermit>,
     deadline: CallDeadline,
     handler: Arc<dyn CallbackHandler>,
     transmitted: bool,
@@ -274,6 +293,7 @@ impl Shared {
                 .flatten();
             (std::mem::take(&mut state.calls), observer)
         };
+        self.execution.close();
         for call in calls.into_values() {
             call.finish(Err(error.clone()));
         }
@@ -303,6 +323,7 @@ impl Shared {
                 call.deadline.clear();
             }
             let initial = call.result.take().ok_or(RpcError::Protocol(None))?;
+            call.execution.take();
             call.handler.streaming(&call.context);
             let _ = initial.send(result);
         } else {
@@ -384,11 +405,40 @@ impl Shared {
     }
 
     pub fn pause_next(&self, id: u64) -> Option<CallDeadlinePause> {
-        self.state()
+        let mut state = self.state();
+        let call = state
             .calls
-            .get(&id)
+            .get_mut(&id)
+            .filter(|call| call.cancellation.is_none())?;
+        let pause = call.deadline.pause()?;
+        call.execution.take();
+        Some(pause)
+    }
+
+    pub async fn resume_next(&self, id: u64) -> Result<(), RpcError> {
+        {
+            let state = self.state();
+            let call = state
+                .calls
+                .get(&id)
+                .filter(|call| call.cancellation.is_none())
+                .ok_or(RpcError::Cancelled)?;
+            // 已交付响应头的流只占在途资源，不重新占用首结果执行名额
+            if call.result.is_none() {
+                return Ok(());
+            }
+        }
+        let permit = self.execution.resume().await?;
+        let mut state = self.state();
+        let call = state
+            .calls
+            .get_mut(&id)
             .filter(|call| call.cancellation.is_none())
-            .and_then(|call| call.deadline.pause())
+            .ok_or(RpcError::Cancelled)?;
+        if call.result.is_some() {
+            call.execution = Some(permit);
+        }
+        Ok(())
     }
 
     pub fn is_cancelling(&self, id: u64) -> bool {
@@ -495,21 +545,23 @@ impl Shared {
         }
     }
 
-    pub(super) fn try_callback_slot(&self) -> Option<tokio::sync::OwnedSemaphorePermit> {
-        self.callback_slots.clone().try_acquire_owned().ok()
+    pub(super) fn try_callback_slot(&self, method: &str) -> Option<OwnedSemaphorePermit> {
+        let slots = if method == gateway_plugin_sdk::call::middleware::NEXT_METHOD {
+            &self.next_slots
+        } else {
+            &self.callback_slots
+        };
+        slots.clone().try_acquire_owned().ok()
     }
 
-    async fn wait_for_callbacks(&self, maximum_callbacks: usize) {
-        let Ok(maximum_callbacks) = u32::try_from(maximum_callbacks) else {
-            return;
-        };
-        if let Ok(permits) = self
-            .callback_slots
-            .clone()
-            .acquire_many_owned(maximum_callbacks)
-            .await
-        {
-            drop(permits);
+    async fn wait_for_callbacks(&self, limits: RpcLimits) {
+        for (slots, maximum) in [
+            (&self.callback_slots, limits.maximum_callbacks),
+            (&self.next_slots, limits.maximum_calls),
+        ] {
+            if let Ok(permits) = slots.clone().acquire_many_owned(maximum as u32).await {
+                drop(permits);
+            }
         }
     }
 }
@@ -558,6 +610,10 @@ impl RpcSession {
     ) -> Result<Self, RpcError> {
         if limits.maximum_calls == 0
             || limits.maximum_calls > 256
+            || limits.maximum_executing_calls == 0
+            || limits.maximum_executing_calls > 256
+            || limits.maximum_queued_calls > 256
+            || limits.maximum_queue_timeout.is_zero()
             || limits.maximum_callbacks == 0
             || limits.maximum_callbacks > 256
             || limits.maximum_buffered_body_bytes < 1024
@@ -657,7 +713,7 @@ impl RpcSession {
         }
         let (data, data_receiver) = mpsc::channel(limits.maximum_calls);
         let (control, control_receiver) =
-            mpsc::channel(limits.maximum_callbacks + limits.maximum_calls + 8);
+            mpsc::channel(limits.maximum_callbacks + 2 * limits.maximum_calls + 8);
         let (stopped, _) = watch::channel(false);
         let callback_slots = Arc::new(Semaphore::new(limits.maximum_callbacks));
         let shared = Arc::new(Shared {
@@ -674,6 +730,12 @@ impl RpcSession {
             process,
             stopped,
             callback_slots,
+            next_slots: Arc::new(Semaphore::new(limits.maximum_calls)),
+            execution: ExecutionCapacity::new(
+                limits.maximum_executing_calls,
+                limits.maximum_queued_calls,
+                limits.maximum_queue_timeout,
+            ),
         });
         dispatch::start(
             input,
@@ -754,11 +816,7 @@ impl RpcSession {
         }
     }
 
-    fn call_deadline(
-        &self,
-        method: &str,
-        context: &CallContext,
-    ) -> Result<tokio::time::Instant, RpcError> {
+    fn call_timeout(&self, method: &str, context: &CallContext) -> Result<Duration, RpcError> {
         if context.instance_id != self.handshake.instance_id
             || context.generation != self.handshake.generation
             || context.incarnation != self.handshake.incarnation
@@ -769,7 +827,7 @@ impl RpcSession {
         {
             return Err(RpcError::Context(None));
         }
-        Ok(tokio::time::Instant::now() + Duration::from_millis(context.timeout_ms))
+        Ok(Duration::from_millis(context.timeout_ms))
     }
 
     fn begin_call(
@@ -777,6 +835,8 @@ impl RpcSession {
         context: CallContext,
         deadline: tokio::time::Instant,
         stream: Option<StreamIngress>,
+        slot: Arc<OwnedSemaphorePermit>,
+        execution: OwnedSemaphorePermit,
     ) -> Result<StartedCall, RpcError> {
         let id = context.call_id;
         let (result, received) = oneshot::channel();
@@ -784,14 +844,13 @@ impl RpcSession {
         if let Some(error) = &state.failure {
             return Err(error.clone());
         }
-        if state.calls.len() >= self.limits.maximum_calls {
-            return Err(RpcError::Capacity);
-        }
         let deadline = CallDeadline::new(deadline);
         self.callbacks.begin(&context, &deadline);
         state.calls.insert(
             id,
             PendingCall {
+                _slot: slot,
+                execution: Some(execution),
                 deadline: deadline.clone(),
                 handler: self.callbacks.clone(),
                 transmitted: false,
@@ -818,7 +877,16 @@ impl RpcSession {
         stream: Option<StreamIngress>,
         credit: Option<(u32, u32)>,
     ) -> Result<CompletedCall, RpcError> {
-        let deadline = self.call_deadline(method, &context)?;
+        let timeout = self.call_timeout(method, &context)?;
+        let slot = Arc::new(self.slots.clone().try_acquire_owned().map_err(|error| {
+            if matches!(error, tokio::sync::TryAcquireError::Closed) {
+                RpcError::Closed(None)
+            } else {
+                RpcError::Overloaded
+            }
+        })?);
+        let execution = self.shared.execution.enter().await?;
+        let deadline = tokio::time::Instant::now() + timeout;
         // FIFO 锁同时拥有 ID 与 data 入队顺序；Call/初始 Credit 入队后必须立即释放，
         // 不能把插件回复或重入回调纳入串行区，也不能为等待锁重置调用期限
         let mut next_call = tokio::time::timeout_at(deadline, self.next_call.lock())
@@ -841,7 +909,7 @@ impl RpcSession {
         validate_frame(&frame).map_err(|source| {
             RpcError::Context(Some(gateway_core::error::ErrorSource::new(source)))
         })?;
-        let started = self.begin_call(context, deadline, stream)?;
+        let started = self.begin_call(context, deadline, stream, slot.clone(), execution)?;
         let mut guard = CallGuard {
             shared: Arc::clone(&self.shared),
             id: started.id,
@@ -906,6 +974,7 @@ impl RpcSession {
             }
         };
         Ok(CompletedCall {
+            slot,
             id: started.id,
             deadline: started.deadline,
             reply,
@@ -919,7 +988,6 @@ impl RpcSession {
         params: serde_json::Value,
         payload: Vec<u8>,
     ) -> Result<RpcReply, RpcError> {
-        let _slot = self.slots.try_acquire().map_err(|_| RpcError::Capacity)?;
         self.send_call(method, context, params, payload, None, None)
             .await
             .map(|completed| completed.reply)
@@ -932,9 +1000,6 @@ impl RpcSession {
         params: serde_json::Value,
         payload: Vec<u8>,
     ) -> Result<RpcStream, RpcError> {
-        let slot = Arc::clone(&self.slots)
-            .try_acquire_owned()
-            .map_err(|_| RpcError::Capacity)?;
         let bytes = self.limits.maximum_buffered_body_bytes.min(256 * 1024) as u32;
         let frames = 32;
         let (stream, chunks, terminal) = StreamIngress::new(bytes, frames);
@@ -955,7 +1020,7 @@ impl RpcSession {
             deadline: completed.deadline,
             id: completed.id,
             shared: Arc::clone(&self.shared),
-            _slot: slot,
+            _slot: completed.slot,
             _session: Arc::clone(self),
         })
     }
@@ -988,9 +1053,7 @@ impl RpcSession {
         self.shared.process.exited().await;
         // 不等待监控任务调度；先撤销父调用并中止其回调，再确认回调 future 已析构
         self.shared.fail(RpcError::Closed(None));
-        self.shared
-            .wait_for_callbacks(self.limits.maximum_callbacks)
-            .await;
+        self.shared.wait_for_callbacks(self.limits).await;
     }
 }
 
@@ -1022,6 +1085,7 @@ struct StartedCall {
 }
 
 struct CompletedCall {
+    slot: Arc<OwnedSemaphorePermit>,
     id: u64,
     deadline: CallDeadline,
     reply: RpcReply,

backend/crates/gateway-plugin/runtime/src/stream/bridge.rs

@@ -62,7 +62,7 @@ pub struct RpcStream {
     pub(crate) deadline: CallDeadline,
     pub(crate) id: u64,
     pub(crate) shared: Arc<Shared>,
-    pub(crate) _slot: OwnedSemaphorePermit,
+    pub(crate) _slot: Arc<OwnedSemaphorePermit>,
     pub(crate) _session: Arc<RpcSession>,
 }
 

backend/crates/gateway-plugin/runtime/tests/callback/log.rs

@@ -99,6 +99,8 @@ impl tracing::Subscriber for Capture {
 
 #[tokio::test]
 async fn real_worker_logs_are_correlated_sanitized_validated_and_bounded() {
+    // 低于每秒日志上限，单独验证阻塞写入的槽位及跨代保留行为
+    const MAXIMUM_CALLBACKS: usize = 16;
     let capture = Capture::default();
     tracing::subscriber::set_global_default(capture.clone()).unwrap();
     let directory = tempfile::tempdir().unwrap();
@@ -155,7 +157,10 @@ async fn real_worker_logs_are_correlated_sanitized_validated_and_bounded() {
             cache_directory: directory.path().join("cache"),
             host_version: "1.0.0".parse().unwrap(),
             package_limits: PackageLimits::default(),
-            rpc_limits: RpcLimits::default(),
+            rpc_limits: RpcLimits {
+                maximum_callbacks: MAXIMUM_CALLBACKS,
+                ..RpcLimits::default()
+            },
             restart_circuit: Default::default(),
         },
         Arc::new(gateway_host::outbound::HttpClient::new().unwrap()),
@@ -270,7 +275,7 @@ async fn real_worker_logs_are_correlated_sanitized_validated_and_bounded() {
     blocked.config_revision = Revision::new(9).unwrap();
     blocked.instances[0].id = "log-blocked".into();
     blocked.instances[0].configuration = json!({
-        "log_method":"plugin.register", "log_entries":[{"params":{"event":"fixture.blocked"},"repeat":32}],
+        "log_method":"plugin.register", "log_entries":[{"params":{"event":"fixture.blocked"},"repeat":MAXIMUM_CALLBACKS * 2}],
         "log_marker":blocked_marker,
     });
     // 日志后端停住时，注册的业务结果和父调用结束仍须及时完成
@@ -288,7 +293,7 @@ async fn real_worker_logs_are_correlated_sanitized_validated_and_bounded() {
             .iter()
             .filter(|result| result["recorded"] == true)
             .count(),
-        RpcLimits::default().maximum_callbacks
+        MAXIMUM_CALLBACKS
     );
     assert!(
         blocked_results

backend/crates/gateway-plugin/runtime/tests/generation/data.rs

@@ -27,6 +27,7 @@ struct Facts(AtomicUsize, AtomicUsize, Option<String>);
 fn account() -> AccountRecord {
     let at = "2026-01-01T00:00:00Z".parse().unwrap();
     AccountRecord {
+        outbound_proxy_enabled: true,
         id: "acct_facts".into(),
         provider_kind: ProviderKind::new("openai").unwrap(),
         groups: vec![gateway_admin::model::account_groups::AccountGroupRef {

backend/crates/gateway-plugin/runtime/tests/generation/policy/http/capacity.rs

@@ -0,0 +1,162 @@
+//! 验证 next 等待不占执行名额，硬上限拒绝请求且不跳过插件
+
+use super::*;
+use gateway_core::error::GatewayErrorKind;
+
+fn context() -> core::Context {
+    core::Context {
+        plugin_instance_id: None,
+        call_id: uuid::Uuid::new_v4().to_string(),
+        parent_call_id: None,
+        extensions: Default::default(),
+        plan: None,
+        request_id: uuid::Uuid::new_v4().to_string(),
+        cancellation: CancellationToken::new(),
+    }
+}
+
+fn plugin_request() -> core::Request {
+    ::http::Request::builder()
+        .uri("/v1/responses")
+        .header("authorization", "Bearer fixture-secret")
+        .body(core::empty_body())
+        .unwrap()
+}
+
+#[tokio::test]
+async fn http_waiting_next_releases_execution_for_delegate_plugin() {
+    assert_capacity(PluginFailurePolicy::Delegate, RpcLimits::default(), false).await;
+}
+
+#[tokio::test]
+async fn http_waiting_next_releases_execution_for_reject_plugin() {
+    assert_capacity(PluginFailurePolicy::Reject, RpcLimits::default(), false).await;
+}
+
+#[tokio::test]
+async fn http_in_flight_limit_never_delegates_and_recovers_after_cancellation() {
+    for policy in [PluginFailurePolicy::Delegate, PluginFailurePolicy::Reject] {
+        assert_capacity(
+            policy,
+            RpcLimits {
+                maximum_calls: 64,
+                ..Default::default()
+            },
+            true,
+        )
+        .await;
+    }
+}
+
+async fn assert_capacity(policy: PluginFailurePolicy, limits: RpcLimits, saturated: bool) {
+    const CONCURRENT_CALLS: usize = 64;
+    let (cache, runtime, generation, plan) =
+        http_plugin_with_policy("passthrough", limits, policy).await;
+    let release = Arc::new(tokio::sync::Semaphore::new(0));
+    let (started, mut entered) = tokio::sync::mpsc::unbounded_channel();
+    let mut slow_calls = tokio::task::JoinSet::new();
+    for index in 0..CONCURRENT_CALLS {
+        let plan = plan.clone();
+        let release = release.clone();
+        let started = started.clone();
+        slow_calls.spawn(async move {
+            plan.handle_http(
+                context(),
+                plugin_request(),
+                compose(Vec::new(), move |request: core::Request| {
+                    assert_eq!(request.uri(), "/rewritten?encoded=%2F");
+                    let release = release.clone();
+                    Box::pin(async move {
+                        started.send(()).unwrap();
+                        release.acquire().await.unwrap().forget();
+                        Ok(core::Response::new(core::empty_body()))
+                    })
+                }),
+            )
+            .await
+        });
+        tokio::time::timeout(Duration::from_secs(5), entered.recv())
+            .await
+            .unwrap_or_else(|_| panic!("第 {} 个调用未能进入下游", index + 1))
+            .unwrap();
+    }
+
+    for uri in [
+        "/healthz",
+        "/assets/index.js",
+        "/api/settings",
+        "/v1/responses",
+    ] {
+        let request = ::http::Request::builder()
+            .uri(uri)
+            .header("authorization", "Bearer fixture-secret")
+            .body(core::empty_body())
+            .unwrap();
+        let result = tokio::time::timeout(
+            Duration::from_secs(3),
+            plan.handle_http(
+                context(),
+                request,
+                compose(Vec::new(), move |request: core::Request| {
+                    assert!(!saturated, "容量耗尽时不得执行下游");
+                    assert_eq!(request.uri(), "/rewritten?encoded=%2F");
+                    Box::pin(async { Ok(core::Response::new(core::empty_body())) })
+                }),
+            ),
+        )
+        .await
+        .expect("下游等待不应阻塞其它请求");
+        if saturated {
+            assert!(matches!(result, Err(MiddlewareError::Gateway(error))
+                if error.kind() == GatewayErrorKind::MiddlewareCapacityUnavailable));
+        } else {
+            assert_eq!(result.unwrap().headers()["x-http-plugin"], "active");
+        }
+    }
+
+    if saturated {
+        slow_calls.abort_all();
+        while let Some(result) = slow_calls.join_next().await {
+            assert!(matches!(result, Err(error) if error.is_cancelled()));
+        }
+    }
+    release.add_permits(CONCURRENT_CALLS);
+    while let Some(result) = slow_calls.join_next().await {
+        let response = result.unwrap().unwrap();
+        assert_eq!(response.headers()["x-http-plugin"], "active");
+    }
+    // 取消确认到达前仍持有在途名额，探测只允许遇到明确的暂时过载
+    let response = tokio::time::timeout(Duration::from_secs(2), async {
+        loop {
+            let result = plan
+                .handle_http(
+                    context(),
+                    plugin_request(),
+                    compose(Vec::new(), |request: core::Request| {
+                        Box::pin(async move {
+                            assert_eq!(request.uri(), "/rewritten?encoded=%2F");
+                            Ok(core::Response::new(core::empty_body()))
+                        })
+                    }),
+                )
+                .await;
+            match result {
+                Ok(response) => break response,
+                Err(MiddlewareError::Gateway(error))
+                    if error.kind() == GatewayErrorKind::MiddlewareCapacityUnavailable =>
+                {
+                    tokio::task::yield_now().await
+                }
+                Err(error) => panic!("unexpected error: {error}"),
+            }
+        }
+    })
+    .await
+    .unwrap();
+    assert_eq!(response.headers()["x-http-plugin"], "active");
+    drop(response);
+    drop(plan);
+    drop(generation);
+    runtime.shutdown().await;
+    crate::generation::wait_until_empty(cache.path()).await;
+}

backend/crates/gateway-plugin/runtime/tests/generation/policy/http/mod.rs

@@ -1,5 +1,6 @@
 //! 验证真实插件 HTTP 中间件的请求改写、流式转换与短路响应
 
+mod capacity;
 mod dispatch;
 mod websocket;
 
@@ -27,6 +28,19 @@ async fn http_plugin_with_limits(
     PluginRuntime,
     ExtensionSetReference,
     gateway_core::engine::middleware::FrozenMiddlewarePlan,
+) {
+    http_plugin_with_policy(mode, limits, PluginFailurePolicy::Reject).await
+}
+
+async fn http_plugin_with_policy(
+    mode: &str,
+    limits: RpcLimits,
+    failure_policy: PluginFailurePolicy,
+) -> (
+    tempfile::TempDir,
+    PluginRuntime,
+    ExtensionSetReference,
+    gateway_core::engine::middleware::FrozenMiddlewarePlan,
 ) {
     let worker = std::fs::read(env!("CARGO_BIN_EXE_gateway-plugin-test-middleware")).unwrap();
     let package = crate::support::package_with_contributions(
@@ -42,12 +56,7 @@ async fn http_plugin_with_limits(
         vec![InstanceFixture {
             id: "http-plugin",
             configuration: serde_json::json!({"http":true,"mode":mode}),
-            bindings: vec![binding(
-                MIDDLEWARE_CONTRIBUTION,
-                "http",
-                0,
-                PluginFailurePolicy::Reject,
-            )],
+            bindings: vec![binding(MIDDLEWARE_CONTRIBUTION, "http", 0, failure_policy)],
         }],
         package,
         limits,

backend/crates/gateway-plugin/runtime/tests/generation/policy/lifecycle.rs

@@ -327,3 +327,79 @@ async fn sdk_waiting_next_obeys_request_deadline_and_client_cancellation() {
         super::super::wait_until_empty(cache.path()).await;
     }
 }
+
+#[tokio::test]
+async fn request_middleware_never_delegates_when_total_in_flight_capacity_is_full() {
+    for policy in [PluginFailurePolicy::Delegate, PluginFailurePolicy::Reject] {
+        let package = crate::support::package_with_contributions(
+            crate::support::worker(),
+            Contributions::from([crate::support::contribution(
+                Capability::Middleware,
+                vec![Stage::Request],
+                vec!["openai".into()],
+                vec!["openai".into()],
+            )]),
+        );
+        let (cache, runtime) = setup_package_with_limits(
+            vec![InstanceFixture {
+                id: "busy-middleware",
+                configuration: serde_json::json!({}),
+                bindings: vec![binding(MIDDLEWARE_CONTRIBUTION, "request", 0, policy)],
+            }],
+            package,
+            RpcLimits {
+                maximum_calls: 1,
+                ..Default::default()
+            },
+        )
+        .await;
+        let generation = prepare(&runtime).await;
+        let plan = runtime
+            .execution_registry()
+            .middleware(&generation)
+            .unwrap();
+        let (entered, waiting) = tokio::sync::oneshot::channel();
+        let first = tokio::spawn({
+            let plan = plan.clone();
+            async move {
+                plan.handle(
+                    middleware_context(ClientTransport::HttpJson),
+                    MiddlewareRequest::new(
+                        "openai",
+                        middleware_headers(),
+                        Bytes::from_static(b"{}"),
+                    ),
+                    gateway_core::middleware::compose(Vec::new(), move |_| {
+                        Box::pin(async move {
+                            entered.send(()).unwrap();
+                            std::future::pending().await
+                        })
+                    }),
+                )
+                .await
+            }
+        });
+        tokio::time::timeout(Duration::from_secs(3), waiting)
+            .await
+            .unwrap()
+            .unwrap();
+        let result = plan
+            .handle(
+                middleware_context(ClientTransport::HttpJson),
+                MiddlewareRequest::new("openai", middleware_headers(), Bytes::from_static(b"{}")),
+                gateway_core::middleware::compose(Vec::new(), |_| {
+                    Box::pin(async { panic!("RPC 过载不能跳过请求中间件") })
+                }),
+            )
+            .await;
+        assert!(
+            matches!(result, Err(MiddlewareError::Gateway(error)) if error.kind() == gateway_core::error::GatewayErrorKind::MiddlewareCapacityUnavailable)
+        );
+        first.abort();
+        assert!(matches!(first.await, Err(error) if error.is_cancelled()));
+        drop(plan);
+        drop(generation);
+        runtime.shutdown().await;
+        super::super::wait_until_empty(cache.path()).await;
+    }
+}

backend/crates/gateway-plugin/runtime/tests/rpc/admission.rs

@@ -0,0 +1,328 @@
+//! 验证执行排队、next 恢复、流式在途预算与取消回收
+
+use std::{
+    sync::{Arc, atomic::Ordering},
+    time::Duration,
+};
+
+use super::session::{Callbacks, LifecycleCallbacks, prepared_session};
+use futures::future::BoxFuture;
+use gateway_plugin_runtime::{CallbackHandler, RpcError, RpcLimits, RpcReply, RpcSession};
+use gateway_plugin_sdk::{CallContext, PluginFault, Stage, call::middleware::NEXT_METHOD};
+use serde_json::json;
+use tokio::sync::{Semaphore, mpsc};
+
+async fn limited<C: CallbackHandler + 'static>(
+    callbacks: Arc<C>,
+    limits: RpcLimits,
+) -> (tempfile::TempDir, Arc<RpcSession>) {
+    let (cache, prepared, handshake, processes) = prepared_session();
+    let session = RpcSession::start(prepared, handshake, limits, &processes, callbacks)
+        .await
+        .unwrap();
+    (cache, Arc::new(session))
+}
+
+fn limits() -> RpcLimits {
+    RpcLimits {
+        maximum_calls: 4,
+        maximum_executing_calls: 1,
+        maximum_queued_calls: 1,
+        maximum_queue_timeout: Duration::from_millis(200),
+        maximum_callbacks: 1,
+        ..Default::default()
+    }
+}
+
+async fn wait_begun(callbacks: &LifecycleCallbacks, count: usize) {
+    tokio::time::timeout(Duration::from_secs(2), async {
+        while callbacks.begun.load(Ordering::Acquire) < count {
+            tokio::task::yield_now().await;
+        }
+    })
+    .await
+    .unwrap();
+}
+
+fn call(session: &RpcSession, method: &str) -> impl Future<Output = Result<RpcReply, RpcError>> {
+    session.call(
+        method,
+        session.context(Stage::Request, Duration::from_secs(2)),
+        json!({}),
+        Vec::new(),
+    )
+}
+
+#[tokio::test]
+async fn bounded_initial_queue_waits_and_releases_cancelled_or_expired_waiters() {
+    let callbacks = Arc::new(LifecycleCallbacks::default());
+    let (_cache, session) = limited(callbacks.clone(), limits()).await;
+    let active = tokio::spawn({
+        let session = session.clone();
+        async move { call(&session, "hang").await }
+    });
+    wait_begun(&callbacks, 1).await;
+
+    let mut queued = Box::pin(call(&session, "echo"));
+    assert!(futures::poll!(&mut queued).is_pending());
+    assert!(matches!(
+        call(&session, "echo").await,
+        Err(RpcError::Overloaded)
+    ));
+    assert_eq!(callbacks.begun.load(Ordering::Acquire), 1);
+    drop(queued);
+
+    let expired = session
+        .call(
+            "echo",
+            session.context(Stage::Request, Duration::from_millis(30)),
+            json!({}),
+            Vec::new(),
+        )
+        .await;
+    assert!(matches!(expired, Err(RpcError::Overloaded)));
+    assert_eq!(callbacks.begun.load(Ordering::Acquire), 1);
+
+    let mut queued = Box::pin(call(&session, "echo"));
+    assert!(futures::poll!(&mut queued).is_pending());
+    active.abort();
+    assert!(matches!(active.await, Err(error) if error.is_cancelled()));
+    assert!(queued.await.is_ok());
+    assert_eq!(callbacks.begun.load(Ordering::Acquire), 2);
+    assert!(session.is_ready());
+    session.shutdown(Duration::from_secs(1)).await;
+}
+
+struct NextCallbacks {
+    entered: mpsc::UnboundedSender<()>,
+    release: Arc<Semaphore>,
+}
+
+impl CallbackHandler for NextCallbacks {
+    fn call(
+        &self,
+        _context: CallContext,
+        method: String,
+        params: serde_json::Value,
+        payload: Vec<u8>,
+    ) -> BoxFuture<'static, Result<RpcReply, PluginFault>> {
+        let release = self.release.clone();
+        self.entered.send(()).unwrap();
+        Box::pin(async move {
+            if method == NEXT_METHOD {
+                release.acquire().await.unwrap().forget();
+            }
+            Ok(RpcReply {
+                result: params,
+                payload,
+            })
+        })
+    }
+}
+
+#[tokio::test]
+async fn next_resumption_does_not_need_a_vacant_initial_queue_slot() {
+    let (entered, mut started) = mpsc::unbounded_channel();
+    let release = Arc::new(Semaphore::new(0));
+    let (_cache, session) = limited(
+        Arc::new(NextCallbacks {
+            entered,
+            release: release.clone(),
+        }),
+        limits(),
+    )
+    .await;
+    let waiting = tokio::spawn({
+        let session = session.clone();
+        async move {
+            session
+                .call(
+                    "callback_method",
+                    session.context(Stage::Request, Duration::from_secs(2)),
+                    json!({"method": NEXT_METHOD}),
+                    Vec::new(),
+                )
+                .await
+        }
+    });
+    started.recv().await.unwrap();
+    // 等待 next 不挤占唯一的普通回调额度
+    let mut occupying = Box::pin(call(&session, "callback"));
+    assert!(futures::poll!(&mut occupying).is_pending());
+    started.recv().await.unwrap();
+    occupying.await.unwrap();
+    let mut active = Box::pin(call(&session, "hang"));
+    assert!(futures::poll!(&mut active).is_pending());
+    let mut queued = Box::pin(call(&session, "echo"));
+    assert!(futures::poll!(&mut queued).is_pending());
+    assert!(matches!(
+        call(&session, "echo").await,
+        Err(RpcError::Overloaded)
+    ));
+    release.add_permits(1);
+    // 恢复等待不能立即因初次队列已满而拒绝，也不能绕过执行名额完成
+    tokio::time::sleep(Duration::from_millis(30)).await;
+    assert!(!waiting.is_finished());
+    drop(active);
+    assert!(queued.await.is_ok());
+    assert!(waiting.await.unwrap().is_ok());
+    assert!(session.is_ready());
+    session.shutdown(Duration::from_secs(1)).await;
+}
+
+#[tokio::test]
+async fn response_stream_releases_execution_but_retains_total_capacity_until_drop() {
+    let callbacks = Arc::new(Callbacks::default());
+    let (_cache, session) = limited(
+        callbacks,
+        RpcLimits {
+            maximum_calls: 2,
+            ..limits()
+        },
+    )
+    .await;
+    let mut streams = Vec::new();
+    for _ in 0..2 {
+        streams.push(
+            session
+                .call_stream(
+                    "stream",
+                    session.context(Stage::Request, Duration::from_secs(5)),
+                    json!({}),
+                    Vec::new(),
+                )
+                .await
+                .unwrap(),
+        );
+    }
+    assert!(matches!(
+        call(&session, "echo").await,
+        Err(RpcError::Overloaded)
+    ));
+    drop(streams.pop());
+    tokio::time::timeout(Duration::from_secs(1), async {
+        loop {
+            match call(&session, "echo").await {
+                Ok(_) => break,
+                Err(RpcError::Overloaded) => tokio::task::yield_now().await,
+                Err(error) => panic!("unexpected error: {error}"),
+            }
+        }
+    })
+    .await
+    .unwrap();
+    assert!(session.is_ready());
+    drop(streams);
+    session.shutdown(Duration::from_secs(1)).await;
+}
+
+#[tokio::test]
+async fn queued_resumption_is_bounded_and_cancellable_without_failing_session() {
+    for cancelled in [false, true] {
+        let (entered, mut started) = mpsc::unbounded_channel();
+        let release = Arc::new(Semaphore::new(0));
+        let (_cache, session) = limited(
+            Arc::new(NextCallbacks {
+                entered,
+                release: release.clone(),
+            }),
+            limits(),
+        )
+        .await;
+        let waiting = tokio::spawn({
+            let session = session.clone();
+            async move {
+                session
+                    .call(
+                        "callback_method",
+                        session.context(Stage::Request, Duration::from_millis(200)),
+                        json!({"method": NEXT_METHOD}),
+                        Vec::new(),
+                    )
+                    .await
+            }
+        });
+        started.recv().await.unwrap();
+        let mut active = Box::pin(call(&session, "hang"));
+        assert!(futures::poll!(&mut active).is_pending());
+        release.add_permits(1);
+        if cancelled {
+            tokio::time::sleep(Duration::from_millis(30)).await;
+            assert!(!waiting.is_finished());
+            waiting.abort();
+            assert!(matches!(waiting.await, Err(error) if error.is_cancelled()));
+        } else {
+            assert!(matches!(waiting.await.unwrap(), Err(RpcError::Overloaded)));
+        }
+        drop(active);
+        assert!(call(&session, "echo").await.is_ok());
+        assert!(session.is_ready());
+        session.shutdown(Duration::from_secs(1)).await;
+    }
+}
+
+#[tokio::test]
+async fn initial_queue_wait_does_not_consume_the_plugin_processing_budget() {
+    let (_cache, session) = limited(
+        Arc::new(Callbacks::default()),
+        RpcLimits {
+            maximum_queue_timeout: Duration::from_secs(2),
+            ..limits()
+        },
+    )
+    .await;
+    let mut active = Box::pin(call(&session, "hang"));
+    assert!(futures::poll!(&mut active).is_pending());
+    let mut queued = Box::pin(session.call(
+        "echo",
+        session.context(Stage::Request, Duration::from_millis(200)),
+        json!({}),
+        Vec::new(),
+    ));
+    assert!(futures::poll!(&mut queued).is_pending());
+    tokio::time::sleep(Duration::from_millis(350)).await;
+    drop(active);
+    assert!(queued.await.is_ok());
+    assert!(session.is_ready());
+    session.shutdown(Duration::from_secs(1)).await;
+}
+
+#[tokio::test]
+async fn resumption_queue_wait_preserves_the_remaining_plugin_processing_budget() {
+    let (entered, mut started) = mpsc::unbounded_channel();
+    let release = Arc::new(Semaphore::new(0));
+    let (_cache, session) = limited(
+        Arc::new(NextCallbacks {
+            entered,
+            release: release.clone(),
+        }),
+        RpcLimits {
+            maximum_queue_timeout: Duration::from_secs(2),
+            ..limits()
+        },
+    )
+    .await;
+    let waiting = tokio::spawn({
+        let session = session.clone();
+        async move {
+            session
+                .call(
+                    "callback_method",
+                    session.context(Stage::Request, Duration::from_millis(200)),
+                    json!({"method": NEXT_METHOD}),
+                    Vec::new(),
+                )
+                .await
+        }
+    });
+    started.recv().await.unwrap();
+    let mut active = Box::pin(call(&session, "hang"));
+    assert!(futures::poll!(&mut active).is_pending());
+    release.add_permits(1);
+    tokio::time::sleep(Duration::from_millis(350)).await;
+    assert!(!waiting.is_finished());
+    drop(active);
+    assert!(waiting.await.unwrap().is_ok());
+    assert!(session.is_ready());
+    session.shutdown(Duration::from_secs(1)).await;
+}

backend/crates/gateway-plugin/runtime/tests/rpc/mod.rs

@@ -1,4 +1,5 @@
 //! 插件双向 RPC 分派与会话生命周期的测试入口
 
+mod admission;
 mod dispatch;
 mod session;

backend/crates/gateway-plugin/runtime/tests/rpc/session.rs

@@ -17,7 +17,7 @@ use serde_json::json;
 use tokio::sync::{Barrier, Notify};
 
 #[derive(Default)]
-struct Callbacks {
+pub(super) struct Callbacks {
     called: AtomicUsize,
     nested: Mutex<Option<Weak<RpcSession>>>,
 }
@@ -29,8 +29,8 @@ struct BlockingCallbacks {
 }
 
 #[derive(Default)]
-struct LifecycleCallbacks {
-    begun: AtomicUsize,
+pub(super) struct LifecycleCallbacks {
+    pub(super) begun: AtomicUsize,
     finished: AtomicUsize,
 }
 
@@ -193,7 +193,7 @@ where
     (cache, Arc::new(session))
 }
 
-fn prepared_session() -> (
+pub(super) fn prepared_session() -> (
     tempfile::TempDir,
     Arc<gateway_plugin_runtime::PreparedPackage>,
     Handshake,

backend/crates/gateway-plugin/sdk/README.md

@@ -52,6 +52,12 @@ use gateway_plugin_sdk::client::{read_frame, write_frame};
 `PluginSession::accept/run` 和 `PluginHandler`。组合插件从作者清单建立类型化方法，SDK 自动生成
 `plugin.register`；作者不需要复制贡献项或手写方法名、阶段和敏感载荷编解码。
 会话管理握手、有界并发、父调用关联、回调、流信用、取消、排空与关闭。
+
+`SessionConfig.maximum_calls` 默认 256，限制包含下游等待和响应流在内的在途调用；`maximum_callbacks` 默认 64，
+限制普通宿主回调。`next` 使用独立回调预算，上限为 `maximum_calls`，不挤占普通回调容量。
+宿主负责执行与准入调度，等待与处理独立计时；默认预算与过载行为见[插件使用](../../../../docs/plugins.md#故障处理)。
+插件显式覆盖配置时需与宿主协调在途预算；较小的插件侧上限会提前拒绝新调用
+
 `ResponseStream::channel` 提供有界流生产入口，`ResponseStream::pull` 按消费进度驱动生产者；`PluginCall` 的宿主客户端只代表当前父调用，
 不能把一个调用的资源误用到另一个调用。插件程序自行选择运行时和启动入口，不会启动网关。
 `SessionConfig.maximum_stream_chunk_bytes` 默认 1 MiB，只约束单个业务流分块；

backend/crates/gateway-plugin/sdk/docs/capabilities.md

@@ -243,7 +243,9 @@ SDK 不自动重试。超时或断连不能证明写入未提交；重试上限
 `next.run(input)` 消费自身，最多执行一次；结果按相反顺序经过外层。重试或独立子请求通过宿主执行服务发起
 
 调用上下文的 `timeout_ms` 限制插件自身处理时间，默认上限为 120 秒。等待宿主 `next` 时暂停这份预算，
-收到结果或错误后恢复剩余时间；入站处理与返回处理合计使用同一预算，普通宿主回调不会暂停计时。
+宿主等待期间归还执行名额，返回后先重新取得名额再交付结果或错误，插件继续同一次调用。
+初次准入和恢复排队各有独立等待期限，默认 30 秒，不扣插件处理预算；恢复等待受总在途上限约束，不占初次准入队列。
+入站处理与返回处理合计使用同一预算，普通宿主回调不会暂停计时。
 请求总期限、下游操作超时和取消信号独立生效，客户端取消或插件退出仍会结束等待并回收资源。
 宿主与插件 SDK 必须同时支持此语义；仍使用固定绝对期限的 SDK 构建的插件需重新构建，单独升级宿主不能解除插件侧期限