diff --git a/README.md b/README.md
index ec186f1de..040c1cafc 100644
--- a/README.md
+++ b/README.md
@@ -83,6 +83,9 @@ API Key 持有者可在同一登录页切换登录身份，进入 `/key-usage` 
 2. 按需建立账号分组，再创建客户端密钥并选择可用分组。**不选分组表示可使用全部账号**。
 3. 打开密钥的「使用密钥」，复制客户端配置。
 
+账号自动冻结默认关闭，可在「系统设置 → 运行参数」中启用。启用后，账号频繁命中上游容量错误时
+会暂停调度并显示为限流中；若开启恢复前探测，冷却到期后需探测成功才恢复。
+
 ## 客户端接入
 
 **Codex CLI / 桌面端**：在「使用密钥」中按操作系统复制配置，或通过 CCSwitch 导入。
diff --git a/backend/crates/gateway-admin/src/freeze_recovery.rs b/backend/crates/gateway-admin/src/freeze_recovery.rs
new file mode 100644
index 000000000..b3488b99d
--- /dev/null
+++ b/backend/crates/gateway-admin/src/freeze_recovery.rs
@@ -0,0 +1,304 @@
+//! 容量熔断冻结的恢复编排：到期探测、失败顺延与自适应并发下调。
+//!
+//! 冻结事实保存在可丢失的 Redis 冷却中，本服务只组合 Admin 端口：
+//! 观测（`AccountRuntimeStore`）、探测（`AccountsService`）与写回
+//! （原子降低并发 / 按冻结代次结束探测）。旧观测不会覆盖管理员操作。
+
+use std::collections::BTreeMap;
+use std::sync::Arc;
+use std::time::{Duration, SystemTime};
+
+use chrono::{DateTime, Utc};
+use futures::StreamExt as _;
+use gateway_core::account::{AccountConcurrencyLimit, ProviderAccountId};
+use gateway_core::routing::UpstreamModelId;
+use gateway_core::task::{ScheduledTask, WorkerCycleContext, WorkerTaskError};
+use tracing::warn;
+
+use crate::model::accounts::{
+    AccountConnectionTestEvent, AccountFreeze, AccountPageItem, AccountRuntimeSnapshot,
+};
+use crate::model::{MutationActor, MutationContext};
+use crate::ports::store::{AccountRuntimeStore, AccountStore, SettingsStore};
+use crate::use_case::accounts::AccountsService;
+
+/// 自适应并发下调保留系数：下调到观测在途峰值的 80%。
+const ADAPTIVE_CONCURRENCY_FACTOR: f64 = 0.8;
+/// 自适应并发下限：过低的并发让账号几乎不可用，宁可保持冻结。
+const ADAPTIVE_CONCURRENCY_FLOOR: u32 = 2;
+/// worker 周期；冷却观测与探测都依赖该节奏。
+pub const FREEZE_RECOVERY_INTERVAL: Duration = Duration::from_secs(30);
+pub const FREEZE_RECOVERY_WORKER_OWNER: &str = "account-freeze-recovery";
+pub const WORKER_INITIAL_BACKOFF: Duration = Duration::from_secs(1);
+pub const WORKER_MAXIMUM_BACKOFF: Duration = Duration::from_secs(60);
+pub const WORKER_LEASE_TTL: Duration = Duration::from_secs(15 * 60);
+pub const WORKER_LEASE_RENEWAL: Duration = Duration::from_secs(5 * 60);
+/// 系统审计上下文的固定请求 ID；账号并发更新审计由此溯源到自动冻结编排。
+const SYSTEM_REQUEST_ID: &str = "account-freeze-recovery";
+
+/// 冻结恢复所需的管理端口组合。
+#[derive(Clone)]
+pub struct FreezeRecoveryDeps {
+    pub accounts: Arc<dyn AccountsService>,
+    pub store: Arc<dyn AccountStore>,
+    pub runtime: Arc<dyn AccountRuntimeStore>,
+    pub settings: Arc<dyn SettingsStore>,
+}
+
+pub struct FreezeRecoveryTask {
+    deps: FreezeRecoveryDeps,
+}
+
+impl FreezeRecoveryTask {
+    pub fn new(deps: FreezeRecoveryDeps) -> Self {
+        Self { deps }
+    }
+
+    fn system_context() -> MutationContext {
+        MutationContext {
+            actor: MutationActor::System,
+            request_id: SYSTEM_REQUEST_ID.to_owned(),
+        }
+    }
+
+    /// 读取冻结策略；读取或校验失败跳过本轮，不能把未知配置当作允许解冻。
+    async fn freeze_policy(&self) -> Option<gateway_core::provider_ports::ProviderFreezePolicy> {
+        let settings = self.deps.settings.load_runtime_settings().await.ok()?;
+        gateway_core::provider_ports::ProviderFreezePolicy::try_new(
+            settings.account_auto_freeze_enabled,
+            settings.account_auto_freeze_threshold,
+            settings.account_auto_freeze_window_seconds,
+            settings.account_auto_freeze_duration_seconds,
+            settings.account_auto_freeze_probe_enabled,
+            settings.account_auto_freeze_probe_model,
+            settings.account_auto_freeze_adaptive_concurrency,
+        )
+        .ok()
+    }
+
+    async fn active_freezes(&self) -> BTreeMap<String, AccountFreeze> {
+        self.deps.runtime.active_freezes().await.unwrap_or_default()
+    }
+
+    async fn run_cycle_inner(&self) {
+        let Some(policy) = self.freeze_policy().await else {
+            return;
+        };
+        let freezes = self.active_freezes().await;
+        if freezes.is_empty() {
+            return;
+        }
+        if policy.enabled() && policy.adaptive_concurrency() {
+            self.adapt_concurrency_limits(&freezes, &policy).await;
+        }
+        self.recover_due_freezes(&freezes, &policy).await;
+    }
+
+    /// 对冻结中的账号执行自适应并发下调：目标为观测在途峰值的 80%（下限 2），
+    /// 只降不升；未观测到峰值证据的账号保持现状。
+    async fn adapt_concurrency_limits(
+        &self,
+        freezes: &BTreeMap<String, AccountFreeze>,
+        policy: &gateway_core::provider_ports::ProviderFreezePolicy,
+    ) {
+        let account_ids = freezes.keys().cloned().collect::<Vec<_>>();
+        let Ok(peaks) = self.deps.runtime.capacity_peaks(&account_ids).await else {
+            return;
+        };
+        for (account_id, peak) in peaks {
+            let Some(target) = adaptive_target(peak) else {
+                continue;
+            };
+            let Ok(account) = ProviderAccountId::new(account_id.clone()) else {
+                continue;
+            };
+            let Some(limit) = AccountConcurrencyLimit::new(target) else {
+                continue;
+            };
+            match self
+                .deps
+                .accounts
+                .lower_concurrency_limit(&Self::system_context(), account, limit)
+                .await
+            {
+                Ok(Some(_)) => {
+                    tracing::info!(
+                        account_id,
+                        adapted_limit = target,
+                        freeze_seconds = policy.freeze_duration().as_secs(),
+                        "容量熔断：按观测在途峰值下调账号并发上限",
+                    );
+                }
+                Ok(None) => {}
+                Err(error) => {
+                    warn!(account_id, error = %error, "自适应并发下调失败");
+                }
+            }
+        }
+    }
+
+    async fn load_account(&self, account_id: &str) -> Option<Option<AccountPageItem>> {
+        let account_id = ProviderAccountId::new(account_id.to_owned()).ok()?;
+        self.deps
+            .store
+            .load_account(account_id.as_str(), AccountRuntimeSnapshot::default())
+            .await
+            .ok()
+    }
+
+    /// 到达探测时间后才恢复；关闭自动冻结或探测时，已有冻结仍等待原冷却结束。
+    async fn recover_due_freezes(
+        &self,
+        freezes: &BTreeMap<String, AccountFreeze>,
+        policy: &gateway_core::provider_ports::ProviderFreezePolicy,
+    ) {
+        for (account_id, freeze) in freezes {
+            if freeze.until > Utc::now() {
+                continue;
+            }
+            if policy.enabled() && policy.probe_enabled() && freeze.requires_probe {
+                self.probe_and_recover(account_id, freeze, policy).await;
+            } else if let Err(error) = self
+                .deps
+                .runtime
+                .finish_freeze(account_id, freeze, None)
+                .await
+            {
+                warn!(account_id, error = %error, "解除到期冻结失败");
+            }
+        }
+    }
+
+    async fn probe_and_recover(
+        &self,
+        account_id: &str,
+        freeze: &AccountFreeze,
+        policy: &gateway_core::provider_ports::ProviderFreezePolicy,
+    ) {
+        let Some(Some(item)) = self.load_account(account_id).await else {
+            return;
+        };
+        if !item.account.enabled {
+            // 停用账号没有自动恢复意义；解冻交给管理员手动恢复。
+            return;
+        }
+        let Ok(account) = ProviderAccountId::new(account_id.to_owned()) else {
+            return;
+        };
+        let Some(model) = self.resolve_probe_model(&account, policy).await else {
+            // 没有可用探测模型时按失败处理，顺延冻结等待下一轮。
+            self.postpone(&account, freeze, policy).await;
+            return;
+        };
+        let probe_succeeded = match self
+            .deps
+            .accounts
+            .test_connection(account.clone(), model)
+            .await
+        {
+            Ok(events) => drain_probe(events).await,
+            Err(error) => {
+                tracing::debug!(account_id, error = %error, "冻结恢复探测发起失败");
+                false
+            }
+        };
+        if probe_succeeded {
+            match self
+                .deps
+                .runtime
+                .finish_freeze(account_id, freeze, None)
+                .await
+            {
+                Ok(true) => tracing::info!(account_id, "冻结恢复探测成功：账号已解冻"),
+                Ok(false) => {}
+                Err(error) => warn!(account_id, error = %error, "解除冻结失败"),
+            }
+        } else {
+            self.postpone(&account, freeze, policy).await;
+        }
+    }
+
+    /// 探测失败只顺延本次仍存在的冻结，旧探测不能重新冻结已手动恢复的账号。
+    async fn postpone(
+        &self,
+        account: &ProviderAccountId,
+        freeze: &AccountFreeze,
+        policy: &gateway_core::provider_ports::ProviderFreezePolicy,
+    ) {
+        let Some(until) = SystemTime::now().checked_add(policy.freeze_duration()) else {
+            return;
+        };
+        let until = DateTime::<Utc>::from(until);
+        match self
+            .deps
+            .runtime
+            .finish_freeze(account.as_str(), freeze, Some(until))
+            .await
+        {
+            Ok(true) => {
+                tracing::info!(
+                    account_id = account.as_str(),
+                    postpone_seconds = policy.freeze_duration().as_secs(),
+                    "冻结恢复探测失败：账号冻结顺延",
+                );
+            }
+            Ok(false) => {}
+            Err(error) => {
+                warn!(
+                    account_id = account.as_str(),
+                    error = %error,
+                    "冻结顺延失败"
+                );
+            }
+        }
+    }
+
+    async fn resolve_probe_model(
+        &self,
+        account: &ProviderAccountId,
+        policy: &gateway_core::provider_ports::ProviderFreezePolicy,
+    ) -> Option<UpstreamModelId> {
+        if let Some(model) = policy.probe_model() {
+            return UpstreamModelId::new(model.to_owned()).ok();
+        }
+        let models = self.deps.accounts.models(account, false).await.ok()?;
+        models.models.into_iter().next().map(|model| model.id)
+    }
+}
+
+impl ScheduledTask for FreezeRecoveryTask {
+    fn run_cycle(
+        &self,
+        context: WorkerCycleContext,
+    ) -> futures::future::BoxFuture<'_, Result<(), WorkerTaskError>> {
+        Box::pin(async move {
+            if context.cancellation().is_cancelled() {
+                return Ok(());
+            }
+            self.run_cycle_inner().await;
+            Ok(())
+        })
+    }
+}
+
+/// 自适应目标并发：观测峰值 × 0.8 向下取整，下限 2；无证据（峰值为 0）不调整。
+fn adaptive_target(peak_in_flight: u32) -> Option<u32> {
+    if peak_in_flight == 0 {
+        return None;
+    }
+    let scaled = (f64::from(peak_in_flight) * ADAPTIVE_CONCURRENCY_FACTOR).floor() as u32;
+    Some(scaled.max(ADAPTIVE_CONCURRENCY_FLOOR))
+}
+
+/// 消耗连接测试事件流并返回是否以 `Completed` 终止。
+async fn drain_probe(mut events: crate::model::accounts::AccountConnectionTestEventStream) -> bool {
+    let mut completed = false;
+    while let Some(event) = events.next().await {
+        match event {
+            AccountConnectionTestEvent::Completed => completed = true,
+            AccountConnectionTestEvent::Failed { .. } => return false,
+            _ => {}
+        }
+    }
+    completed
+}
diff --git a/backend/crates/gateway-admin/src/lib.rs b/backend/crates/gateway-admin/src/lib.rs
index e78d3052b..30a18f77d 100644
--- a/backend/crates/gateway-admin/src/lib.rs
+++ b/backend/crates/gateway-admin/src/lib.rs
@@ -10,14 +10,15 @@ use gateway_core::{
     routing::ProviderKind,
     runtime::SnapshotControl,
     task::{
-        DaemonRestartPolicy, WorkerContribution, WorkerId, WorkerKind, WorkerRegistration,
-        WorkerRunnable,
+        DaemonRestartPolicy, WorkerContribution, WorkerId, WorkerKind, WorkerLeaseRequest,
+        WorkerRegistration, WorkerRunnable, WorkerSchedule,
     },
 };
 use secrecy::{ExposeSecret as _, SecretString};
 use serde::Deserialize;
 
 pub mod backup;
+pub mod freeze_recovery;
 pub mod model;
 pub mod ports;
 mod use_case;
@@ -210,6 +211,12 @@ impl AdminServices {
         self.key_usage.as_ref()
     }
 
+    /// 取得账号服务的共享句柄；后台编排（冻结恢复 worker）需要持有 Arc。
+    #[must_use]
+    pub fn accounts_handle(&self) -> Arc<dyn AccountsService> {
+        Arc::clone(&self.accounts)
+    }
+
     #[must_use]
     pub fn proxies(&self) -> &dyn ProxiesService {
         self.proxies.as_ref()
@@ -392,7 +399,7 @@ pub async fn initialize(
             registry.clone(),
         )),
         auth,
-        accounts,
+        accounts: accounts.clone(),
         account_groups: Arc::new(DefaultAccountGroupService::new(
             store.account_groups(),
             store.account_runtime(),
@@ -419,6 +426,13 @@ pub async fn initialize(
         import_tasks,
         backups,
     };
+    let freeze_recovery =
+        freeze_recovery::FreezeRecoveryTask::new(freeze_recovery::FreezeRecoveryDeps {
+            accounts: Arc::clone(&accounts) as Arc<dyn AccountsService>,
+            store: store.accounts(),
+            runtime: store.account_runtime(),
+            settings: store.settings(),
+        });
     let mut worker_contributions = backup_worker_contribution(backup_task)?;
     let id = WorkerId::try_new(WorkerKind::AccountImport, "admin")
         .map_err(|_| AdminError::internal("导入 Worker ID 不合法"))?;
@@ -433,6 +447,7 @@ pub async fn initialize(
     )
     .map_err(|_| AdminError::internal("导入 Worker 注册信息不合法"))?;
     worker_contributions.push(WorkerContribution::Registration(registration));
+    worker_contributions.extend(freeze_recovery_worker_contribution(freeze_recovery)?);
     Ok(AdminBundle {
         services,
         worker_contributions,
@@ -458,6 +473,37 @@ fn backup_worker_contribution(
     Ok(vec![WorkerContribution::Registration(registration)])
 }
 
+/// 冻结恢复 Worker 注册：按固定周期扫描活跃冻结，owner 固定。
+fn freeze_recovery_worker_contribution(
+    task: freeze_recovery::FreezeRecoveryTask,
+) -> Result<Vec<WorkerContribution>, AdminError> {
+    let id = WorkerId::try_new(
+        WorkerKind::AccountFreezeRecovery,
+        freeze_recovery::FREEZE_RECOVERY_WORKER_OWNER,
+    )
+    .map_err(|_| AdminError::internal("冻结恢复 Worker ID 不合法"))?;
+    let schedule = WorkerSchedule::try_new(
+        freeze_recovery::FREEZE_RECOVERY_INTERVAL,
+        freeze_recovery::WORKER_INITIAL_BACKOFF,
+        freeze_recovery::WORKER_MAXIMUM_BACKOFF,
+        freeze_recovery::WORKER_LEASE_TTL,
+        freeze_recovery::WORKER_LEASE_RENEWAL,
+    )
+    .map_err(|_| AdminError::internal("冻结恢复 Worker 调度配置不合法"))?;
+    let lease = WorkerLeaseRequest::try_new(id.clone(), freeze_recovery::WORKER_LEASE_TTL)
+        .map_err(|_| AdminError::internal("冻结恢复 Worker 租约配置不合法"))?;
+    let registration = WorkerRegistration::try_new(
+        id,
+        WorkerRunnable::Scheduled {
+            schedule,
+            lease: Some(lease),
+            task: Box::new(task),
+        },
+    )
+    .map_err(|_| AdminError::internal("冻结恢复 Worker 注册信息不合法"))?;
+    Ok(vec![WorkerContribution::Registration(registration)])
+}
+
 fn provider_kind(value: &'static str) -> Result<ProviderKind, AdminError> {
     ProviderKind::new(value).map_err(|_| AdminError::internal("内置 Provider 类型不合法"))
 }
diff --git a/backend/crates/gateway-admin/src/model/accounts.rs b/backend/crates/gateway-admin/src/model/accounts.rs
index 751a05612..29abee64d 100644
--- a/backend/crates/gateway-admin/src/model/accounts.rs
+++ b/backend/crates/gateway-admin/src/model/accounts.rs
@@ -69,11 +69,20 @@ pub struct AccountListQuery {
 /// Admin query service 从运行态存储取得的当前账号冷却快照。
 #[derive(Debug, Clone, Default, PartialEq, Eq)]
 pub struct AccountRuntimeSnapshot {
-    pub rate_limited_until: BTreeMap<String, DateTime<Utc>>,
+    pub cooldown: BTreeMap<String, gateway_core::account::AccountCooldown>,
     /// `None` 表示实时 lease 存储不可用；`Some` 中未出现的账号当前使用量为零。
     pub in_flight: Option<BTreeMap<String, u64>>,
 }
 
+/// 恢复任务读取的冻结快照；generation 将异步结果绑定到本次冻结。
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub struct AccountFreeze {
+    pub credential_revision: Revision,
+    pub until: DateTime<Utc>,
+    pub generation: String,
+    pub requires_probe: bool,
+}
+
 /// Optional account membership filter.
 #[derive(Debug, Clone, PartialEq, Eq)]
 pub enum AccountGroupFilter {
diff --git a/backend/crates/gateway-admin/src/model/settings.rs b/backend/crates/gateway-admin/src/model/settings.rs
index da628b80e..031224dcc 100644
--- a/backend/crates/gateway-admin/src/model/settings.rs
+++ b/backend/crates/gateway-admin/src/model/settings.rs
@@ -32,6 +32,13 @@ pub struct RuntimeSettings {
     pub usage_retention_days: u32,
     pub ops_event_retention_days: u32,
     pub audit_retention_days: u32,
+    pub account_auto_freeze_enabled: bool,
+    pub account_auto_freeze_threshold: u32,
+    pub account_auto_freeze_window_seconds: u64,
+    pub account_auto_freeze_duration_seconds: u64,
+    pub account_auto_freeze_probe_enabled: bool,
+    pub account_auto_freeze_probe_model: Option<String>,
+    pub account_auto_freeze_adaptive_concurrency: bool,
     pub updated_at: DateTime<Utc>,
 }
 
@@ -52,6 +59,13 @@ pub struct ReplaceRuntimeSettings {
     pub usage_retention_days: u32,
     pub ops_event_retention_days: u32,
     pub audit_retention_days: u32,
+    pub account_auto_freeze_enabled: bool,
+    pub account_auto_freeze_threshold: u32,
+    pub account_auto_freeze_window_seconds: u64,
+    pub account_auto_freeze_duration_seconds: u64,
+    pub account_auto_freeze_probe_enabled: bool,
+    pub account_auto_freeze_probe_model: Option<String>,
+    pub account_auto_freeze_adaptive_concurrency: bool,
 }
 
 /// 明文管理员 API Key；按产品约束明文落库，但禁止 Debug 泄漏。
diff --git a/backend/crates/gateway-admin/src/ports/store.rs b/backend/crates/gateway-admin/src/ports/store.rs
index 394978dec..4c1147ec7 100644
--- a/backend/crates/gateway-admin/src/ports/store.rs
+++ b/backend/crates/gateway-admin/src/ports/store.rs
@@ -2,7 +2,7 @@
 //!
 //! 端口按业务资源拆分，方法使用领域模型，不暴露连接池、事务或 Redis client。
 
-use std::{net::IpAddr, sync::Arc, time::Duration};
+use std::{collections::BTreeMap, net::IpAddr, sync::Arc, time::Duration};
 
 use async_trait::async_trait;
 use chrono::{DateTime, Utc};
@@ -159,6 +159,14 @@ pub trait AccountStore: Send + Sync {
         context: &MutationContext,
     ) -> AdminStoreResult<AccountUpdateResult>;
 
+    /// 在事务内按最新启用状态、账号上限和全局默认值判断，只降低并发上限。
+    async fn lower_concurrency_limit(
+        &self,
+        account_id: &gateway_core::account::ProviderAccountId,
+        limit: gateway_core::account::AccountConcurrencyLimit,
+        context: &MutationContext,
+    ) -> AdminStoreResult<Option<AccountUpdateResult>>;
+
     async fn recover_account(
         &self,
         account_id: &gateway_core::account::ProviderAccountId,
@@ -193,6 +201,25 @@ pub trait AccountRuntimeStore: Send + Sync {
         &self,
         account_ids: &[String],
     ) -> AdminStoreResult<AccountRuntimeSnapshot>;
+
+    /// 容量熔断自动冻结中的账号与其冻结截止时间；429 临时限流不包含在内。
+    async fn active_freezes(
+        &self,
+    ) -> AdminStoreResult<BTreeMap<String, crate::model::accounts::AccountFreeze>>;
+
+    /// 读取容量失败窗口内观测到的在途并发峰值（自适应并发下调的证据）。
+    async fn capacity_peaks(
+        &self,
+        account_ids: &[String],
+    ) -> AdminStoreResult<BTreeMap<String, u32>>;
+
+    /// 仅当冻结快照仍匹配时解除或顺延；旧探测不得覆盖手动恢复或新一轮冻结。
+    async fn finish_freeze(
+        &self,
+        account_id: &str,
+        expected: &crate::model::accounts::AccountFreeze,
+        postpone_until: Option<DateTime<Utc>>,
+    ) -> AdminStoreResult<bool>;
 }
 
 /// 控制面凭据、统一会话、登录限流与管理员安全审计。
diff --git a/backend/crates/gateway-admin/src/use_case/account_groups.rs b/backend/crates/gateway-admin/src/use_case/account_groups.rs
index f6321877a..a278e0640 100644
--- a/backend/crates/gateway-admin/src/use_case/account_groups.rs
+++ b/backend/crates/gateway-admin/src/use_case/account_groups.rs
@@ -192,11 +192,7 @@ fn project_group_runtime(
     let mut accounts_by_group = BTreeMap::<&str, Vec<&str>>::new();
     for member in members {
         let mut status = member.status.clone();
-        status.rate_limited_until = runtime
-            .rate_limited_until
-            .get(&member.account_id)
-            .copied()
-            .map(Into::into);
+        status.cooldown = runtime.cooldown.get(&member.account_id).copied();
         status_by_account
             .entry(member.account_id.as_str())
             .or_insert_with(|| resolve_account_status(&status, now).status);
diff --git a/backend/crates/gateway-admin/src/use_case/accounts.rs b/backend/crates/gateway-admin/src/use_case/accounts.rs
index 785fc3ebd..cf7d9bfc6 100644
--- a/backend/crates/gateway-admin/src/use_case/accounts.rs
+++ b/backend/crates/gateway-admin/src/use_case/accounts.rs
@@ -73,6 +73,13 @@ pub trait AccountsService: Send + Sync {
         command: UpdateAccount,
     ) -> Result<AccountUpdateResult, AdminError>;
 
+    async fn lower_concurrency_limit(
+        &self,
+        context: &MutationContext,
+        account_id: ProviderAccountId,
+        limit: gateway_core::account::AccountConcurrencyLimit,
+    ) -> Result<Option<AccountUpdateResult>, AdminError>;
+
     async fn batch_update(
         &self,
         context: &MutationContext,
@@ -582,6 +589,27 @@ impl AccountsService for DefaultAccountsService {
         Ok(result)
     }
 
+    async fn lower_concurrency_limit(
+        &self,
+        context: &MutationContext,
+        account_id: ProviderAccountId,
+        limit: gateway_core::account::AccountConcurrencyLimit,
+    ) -> Result<Option<AccountUpdateResult>, AdminError> {
+        let (_, provider) = self.provider_for_account(&account_id).await?;
+        let result = self
+            .accounts
+            .lower_concurrency_limit(&account_id, limit, context)
+            .await
+            .map_err(|error| map_store_error(error, "provider account"))?;
+        if let Some(result) = &result {
+            provider
+                .account_facts_changed(std::slice::from_ref(&account_id))
+                .await;
+            publish_committed(self.snapshot.as_ref(), result.config_revision).await?;
+        }
+        Ok(result)
+    }
+
     async fn batch_update(
         &self,
         context: &MutationContext,
diff --git a/backend/crates/gateway-admin/src/use_case/settings.rs b/backend/crates/gateway-admin/src/use_case/settings.rs
index 330dc1038..60cfb4280 100644
--- a/backend/crates/gateway-admin/src/use_case/settings.rs
+++ b/backend/crates/gateway-admin/src/use_case/settings.rs
@@ -125,7 +125,11 @@ fn validate_settings(command: &ReplaceRuntimeSettings) -> Result<(), AdminError>
         && command.audit_retention_days > 0
         && valid_client_version(command.min_codex_desktop_version.as_deref())
         && valid_client_version(command.min_codex_cli_version.as_deref())
-        && i64::try_from(command.request_interval_ms).is_ok();
+        && valid_probe_model(command.account_auto_freeze_probe_model.as_deref())
+        && i64::try_from(command.request_interval_ms).is_ok()
+        && (2..=1_000).contains(&command.account_auto_freeze_threshold)
+        && (60..=3_600).contains(&command.account_auto_freeze_window_seconds)
+        && (300..=604_800).contains(&command.account_auto_freeze_duration_seconds);
     if valid {
         Ok(())
     } else {
@@ -136,3 +140,12 @@ fn validate_settings(command: &ReplaceRuntimeSettings) -> Result<(), AdminError>
 fn valid_client_version(value: Option<&str>) -> bool {
     value.is_none_or(|value| CodexClientVersion::parse(value).is_ok())
 }
+
+fn valid_probe_model(value: Option<&str>) -> bool {
+    value.is_none_or(|value| {
+        !value.is_empty()
+            && value.len() <= 128
+            && value == value.trim()
+            && !value.bytes().any(|byte| byte.is_ascii_control())
+    })
+}
diff --git a/backend/crates/gateway-admin/tests/use_case/account_groups.rs b/backend/crates/gateway-admin/tests/use_case/account_groups.rs
index 1bcecb1a2..8ba996ec8 100644
--- a/backend/crates/gateway-admin/tests/use_case/account_groups.rs
+++ b/backend/crates/gateway-admin/tests/use_case/account_groups.rs
@@ -5,7 +5,7 @@ use std::{
 };
 
 use async_trait::async_trait;
-use chrono::{Duration, Utc};
+use chrono::{DateTime, Duration, Utc};
 use gateway_admin::{
     model::{
         MutationContext, PageSize, Revision,
@@ -168,13 +168,35 @@ impl AccountRuntimeStore for FakeRuntimeStore {
     ) -> AdminStoreResult<AccountRuntimeSnapshot> {
         *self.requested_accounts.lock().expect("requested accounts") = account_ids.to_vec();
         Ok(AccountRuntimeSnapshot {
-            rate_limited_until: BTreeMap::from([(
+            cooldown: BTreeMap::from([(
                 "acct_limited".to_owned(),
-                Utc::now() + Duration::minutes(5),
+                std::time::SystemTime::from(Utc::now() + Duration::minutes(5)).into(),
             )]),
             in_flight: Some(BTreeMap::from([("acct_available".to_owned(), 2)])),
         })
     }
+
+    async fn active_freezes(
+        &self,
+    ) -> AdminStoreResult<BTreeMap<String, gateway_admin::model::accounts::AccountFreeze>> {
+        Ok(BTreeMap::new())
+    }
+
+    async fn capacity_peaks(
+        &self,
+        _account_ids: &[String],
+    ) -> AdminStoreResult<BTreeMap<String, u32>> {
+        Ok(BTreeMap::new())
+    }
+
+    async fn finish_freeze(
+        &self,
+        _account_id: &str,
+        _expected: &gateway_admin::model::accounts::AccountFreeze,
+        _postpone_until: Option<DateTime<Utc>>,
+    ) -> AdminStoreResult<bool> {
+        Ok(false)
+    }
 }
 
 fn group_record() -> AccountGroupRecord {
@@ -215,7 +237,7 @@ fn member(account_id: &str, total_slots: u64) -> AccountGroupMemberFact {
             credential_state: CredentialState::Ready,
             access_token_expires_at: None,
             quota: QuotaState::default(),
-            rate_limited_until: None,
+            cooldown: None,
             last_error_reason: None,
             last_error_message: None,
         },
diff --git a/backend/crates/gateway-admin/tests/use_case/accounts.rs b/backend/crates/gateway-admin/tests/use_case/accounts.rs
index f36c15d10..74b9fa897 100644
--- a/backend/crates/gateway-admin/tests/use_case/accounts.rs
+++ b/backend/crates/gateway-admin/tests/use_case/accounts.rs
@@ -544,6 +544,8 @@ pub(super) struct FakeAccountStore {
     quota_window_usage: Mutex<Vec<AccountUsageWindowResult>>,
     quota_window_queries: Mutex<Vec<AccountUsageWindowQuery>>,
     quota_forecast_history: Mutex<QuotaForecastHistory>,
+    update_commands: Mutex<Vec<UpdateAccount>>,
+    pub(super) lowered_limits: Mutex<Vec<(String, u32)>>,
 }
 
 impl FakeAccountStore {
@@ -551,7 +553,7 @@ impl FakeAccountStore {
         Self::with_account(account_record(kind), events)
     }
 
-    fn with_account(account: AccountRecord, events: EventLog) -> Arc<Self> {
+    pub(super) fn with_account(account: AccountRecord, events: EventLog) -> Arc<Self> {
         Arc::new(Self {
             events,
             accounts: Mutex::new(vec![account]),
@@ -562,9 +564,18 @@ impl FakeAccountStore {
             quota_window_usage: Mutex::new(Vec::new()),
             quota_window_queries: Mutex::new(Vec::new()),
             quota_forecast_history: Mutex::new(QuotaForecastHistory::default()),
+            update_commands: Mutex::new(Vec::new()),
+            lowered_limits: Mutex::new(Vec::new()),
         })
     }
 
+    pub(super) fn update_commands(&self) -> Vec<UpdateAccount> {
+        self.update_commands
+            .lock()
+            .expect("update commands")
+            .clone()
+    }
+
     pub(super) fn fail_next_commit(&self) {
         *self.fail_commit.lock().expect("store failure") = true;
     }
@@ -630,7 +641,7 @@ impl FakeAccountStore {
             credential_state: account.credential_state,
             access_token_expires_at: account.access_token_expires_at.map(Into::into),
             quota: account.quota,
-            rate_limited_until: None,
+            cooldown: None,
             last_error_reason: account.last_error_reason,
             last_error_message: account.last_error_message.clone(),
         };
@@ -852,12 +863,45 @@ impl AccountStore for FakeAccountStore {
         self.record("store.update_account");
         self.record_context(context);
         self.require_commit()?;
+        self.update_commands
+            .lock()
+            .expect("update commands")
+            .push(command.clone());
         Ok(AccountUpdateResult {
             config_revision: revision(2),
             account_id: ProviderAccountId::new(command.account_id).expect("account ID"),
         })
     }
 
+    async fn lower_concurrency_limit(
+        &self,
+        account_id: &ProviderAccountId,
+        limit: gateway_core::account::AccountConcurrencyLimit,
+        _: &MutationContext,
+    ) -> AdminStoreResult<Option<AccountUpdateResult>> {
+        let mut accounts = self.accounts.lock().expect("accounts");
+        let Some(account) = accounts
+            .iter_mut()
+            .find(|account| account.id == account_id.as_str())
+        else {
+            return Ok(None);
+        };
+        if !account.enabled
+            || account.concurrency_limit.map_or(5, |value| value.get()) <= limit.get()
+        {
+            return Ok(None);
+        }
+        account.concurrency_limit = Some(limit);
+        self.lowered_limits
+            .lock()
+            .expect("lowered limits")
+            .push((account.id.clone(), limit.get()));
+        Ok(Some(AccountUpdateResult {
+            config_revision: revision(2),
+            account_id: account_id.clone(),
+        }))
+    }
+
     async fn recover_account(
         &self,
         account_id: &ProviderAccountId,
@@ -944,6 +988,13 @@ impl SettingsStore for StaticSettingsStore {
             usage_retention_days: 30,
             ops_event_retention_days: 30,
             audit_retention_days: 30,
+            account_auto_freeze_enabled: true,
+            account_auto_freeze_threshold: 12,
+            account_auto_freeze_window_seconds: 600,
+            account_auto_freeze_duration_seconds: 7_200,
+            account_auto_freeze_probe_enabled: true,
+            account_auto_freeze_probe_model: None,
+            account_auto_freeze_adaptive_concurrency: true,
             updated_at: Utc::now(),
         })
     }
diff --git a/backend/crates/gateway-admin/tests/use_case/freeze_recovery.rs b/backend/crates/gateway-admin/tests/use_case/freeze_recovery.rs
new file mode 100644
index 000000000..9dd2ce6a8
--- /dev/null
+++ b/backend/crates/gateway-admin/tests/use_case/freeze_recovery.rs
@@ -0,0 +1,425 @@
+//! 容量熔断冻结恢复编排回归：到期探测、失败顺延与自适应并发下调。
+
+use std::collections::BTreeMap;
+use std::sync::{Arc, Mutex};
+
+use async_trait::async_trait;
+use chrono::{DateTime, Duration as TimeDelta, Utc};
+use gateway_admin::freeze_recovery::{FreezeRecoveryDeps, FreezeRecoveryTask};
+use gateway_admin::model::MutationContext;
+use gateway_admin::model::accounts::{AccountFreeze, AccountRuntimeSnapshot};
+use gateway_admin::model::settings::{
+    AdminApiKey, AdminApiKeyMutation, ReplaceRuntimeSettings, RuntimeSettings,
+};
+use gateway_admin::ports::store::{
+    AccountRuntimeStore, AccountStore, AdminStoreError, AdminStoreResult, SettingsStore,
+};
+use gateway_core::engine::probe::{
+    AccountProbe, AccountProbeError, AccountProbeRequest, AccountProbeResult,
+};
+use gateway_core::lifecycle::CancellationToken;
+use gateway_core::task::{ScheduledTask as _, WorkerCycleContext, WorkerId, WorkerKind};
+
+use super::AdminHarness;
+use super::accounts::{FakeAccountStore, FakeProviderAdmin, account_record, events, revision};
+
+fn runtime_settings(enabled: bool, probe_enabled: bool, adaptive: bool) -> RuntimeSettings {
+    RuntimeSettings {
+        config_revision: revision(1),
+        model_mappings: Default::default(),
+        refresh_margin_seconds: 300,
+        refresh_concurrency: 2,
+        max_concurrent_per_account: 5,
+        request_interval_ms: 0,
+        max_waiting_per_key: 0,
+        max_waiting_per_account: 0,
+        concurrency_wait_timeout_seconds: 30,
+        rotation_strategy: gateway_admin::model::settings::RotationStrategy::Smart,
+        min_codex_desktop_version: None,
+        min_codex_cli_version: None,
+        usage_retention_days: 31,
+        ops_event_retention_days: 30,
+        audit_retention_days: 90,
+        account_auto_freeze_enabled: enabled,
+        account_auto_freeze_threshold: 12,
+        account_auto_freeze_window_seconds: 600,
+        account_auto_freeze_duration_seconds: 7_200,
+        account_auto_freeze_probe_enabled: probe_enabled,
+        account_auto_freeze_probe_model: Some("gpt-5.5".to_owned()),
+        account_auto_freeze_adaptive_concurrency: adaptive,
+        updated_at: Utc::now(),
+    }
+}
+
+struct FreezeSettingsStore {
+    settings: RuntimeSettings,
+}
+
+#[async_trait]
+impl SettingsStore for FreezeSettingsStore {
+    async fn load_runtime_settings(&self) -> AdminStoreResult<RuntimeSettings> {
+        Ok(self.settings.clone())
+    }
+
+    async fn admin_api_key_exists(&self) -> AdminStoreResult<bool> {
+        Ok(false)
+    }
+
+    async fn replace_runtime_settings(
+        &self,
+        _: ReplaceRuntimeSettings,
+        _: &MutationContext,
+    ) -> AdminStoreResult<RuntimeSettings> {
+        Err(store_unavailable())
+    }
+
+    async fn replace_admin_api_key(
+        &self,
+        _: AdminApiKey,
+        _: &MutationContext,
+    ) -> AdminStoreResult<AdminApiKeyMutation> {
+        Err(store_unavailable())
+    }
+
+    async fn delete_admin_api_key(
+        &self,
+        _: &MutationContext,
+    ) -> AdminStoreResult<AdminApiKeyMutation> {
+        Err(store_unavailable())
+    }
+}
+
+fn store_unavailable() -> AdminStoreError {
+    AdminStoreError::new(
+        gateway_admin::ports::store::AdminStoreErrorKind::Unavailable,
+        "freeze recovery test",
+        "unused",
+    )
+}
+
+/// 记录解冻/顺延调用的运行态 fake；冻结与峰值证据由测试预置。
+struct FreezeRuntimeStore {
+    freezes: BTreeMap<String, AccountFreeze>,
+    peaks: BTreeMap<String, u32>,
+    cleared: Mutex<Vec<String>>,
+    extended: Mutex<Vec<(String, DateTime<Utc>)>>,
+}
+
+impl FreezeRuntimeStore {
+    fn new(freezes: BTreeMap<String, AccountFreeze>, peaks: BTreeMap<String, u32>) -> Arc<Self> {
+        Arc::new(Self {
+            freezes,
+            peaks,
+            cleared: Mutex::new(Vec::new()),
+            extended: Mutex::new(Vec::new()),
+        })
+    }
+
+    fn cleared(&self) -> Vec<String> {
+        self.cleared.lock().expect("cleared lock").clone()
+    }
+
+    fn extended(&self) -> Vec<(String, DateTime<Utc>)> {
+        self.extended.lock().expect("extended lock").clone()
+    }
+}
+
+#[async_trait]
+impl AccountRuntimeStore for FreezeRuntimeStore {
+    async fn active_rate_limits(&self) -> AdminStoreResult<AccountRuntimeSnapshot> {
+        Ok(AccountRuntimeSnapshot::default())
+    }
+
+    async fn account_runtime(
+        &self,
+        _account_ids: &[String],
+    ) -> AdminStoreResult<AccountRuntimeSnapshot> {
+        Ok(AccountRuntimeSnapshot::default())
+    }
+
+    async fn active_freezes(&self) -> AdminStoreResult<BTreeMap<String, AccountFreeze>> {
+        Ok(self.freezes.clone())
+    }
+
+    async fn capacity_peaks(
+        &self,
+        account_ids: &[String],
+    ) -> AdminStoreResult<BTreeMap<String, u32>> {
+        Ok(account_ids
+            .iter()
+            .filter_map(|id| self.peaks.get(id).map(|peak| (id.clone(), *peak)))
+            .collect())
+    }
+
+    async fn finish_freeze(
+        &self,
+        account_id: &str,
+        expected: &AccountFreeze,
+        postpone_until: Option<DateTime<Utc>>,
+    ) -> AdminStoreResult<bool> {
+        assert_eq!(self.freezes.get(account_id), Some(expected));
+        if let Some(until) = postpone_until {
+            self.extended
+                .lock()
+                .expect("extended")
+                .push((account_id.to_owned(), until));
+        } else {
+            self.cleared
+                .lock()
+                .expect("cleared")
+                .push(account_id.to_owned());
+        }
+        Ok(true)
+    }
+}
+
+struct SuccessfulProbe;
+
+impl AccountProbe for SuccessfulProbe {
+    fn probe(
+        &self,
+        _: AccountProbeRequest,
+    ) -> futures::future::BoxFuture<'_, Result<AccountProbeResult, AccountProbeError>> {
+        Box::pin(async {
+            Ok(AccountProbeResult {
+                text: vec!["OK".to_owned()],
+            })
+        })
+    }
+}
+
+struct FailingProbe;
+
+impl AccountProbe for FailingProbe {
+    fn probe(
+        &self,
+        _: AccountProbeRequest,
+    ) -> futures::future::BoxFuture<'_, Result<AccountProbeResult, AccountProbeError>> {
+        Box::pin(async {
+            Err(AccountProbeError::new(
+                gateway_core::error::GatewayError::new(
+                    gateway_core::error::GatewayErrorKind::RateLimited,
+                    "still overloaded",
+                ),
+                gateway_core::engine::probe::AccountProbeErrorSource::Upstream,
+                Some(gateway_core::upstream::UpstreamSendState::NotSent),
+                None,
+            ))
+        })
+    }
+}
+
+async fn recovery_task(
+    settings: RuntimeSettings,
+    runtime: Arc<FreezeRuntimeStore>,
+    probe: Arc<dyn AccountProbe>,
+) -> (FreezeRecoveryTask, Arc<FakeAccountStore>) {
+    recovery_task_with_store(settings, runtime, probe).await
+}
+
+async fn recovery_task_with_store(
+    settings: RuntimeSettings,
+    runtime: Arc<FreezeRuntimeStore>,
+    probe: Arc<dyn AccountProbe>,
+) -> (FreezeRecoveryTask, Arc<FakeAccountStore>) {
+    let store = FakeAccountStore::new("openai", events());
+    let services = AdminHarness::new()
+        .accounts(Arc::clone(&store) as Arc<dyn AccountStore>)
+        .account_runtime(Arc::clone(&runtime) as Arc<dyn AccountRuntimeStore>)
+        .settings(Arc::new(FreezeSettingsStore {
+            settings: settings.clone(),
+        }))
+        .provider(FakeProviderAdmin::new("openai", events()))
+        .probe(probe)
+        .build()
+        .await;
+    let task = FreezeRecoveryTask::new(FreezeRecoveryDeps {
+        accounts: services.accounts_handle(),
+        store: store.clone(),
+        runtime,
+        settings: Arc::new(FreezeSettingsStore {
+            settings: settings.clone(),
+        }),
+    });
+    (task, store)
+}
+
+async fn run_cycle(task: &FreezeRecoveryTask) {
+    let worker = WorkerId::try_new(WorkerKind::AccountFreezeRecovery, "test").expect("worker id");
+    let context = WorkerCycleContext::new(worker, None, CancellationToken::new());
+    task.run_cycle(context)
+        .await
+        .expect("freeze recovery cycle");
+}
+
+fn freeze_until(until: DateTime<Utc>) -> BTreeMap<String, AccountFreeze> {
+    BTreeMap::from([(
+        "acct_test".to_owned(),
+        AccountFreeze {
+            credential_revision: revision(1),
+            until,
+            generation: "freeze-generation".to_owned(),
+            requires_probe: true,
+        },
+    )])
+}
+
+fn freeze_due() -> BTreeMap<String, AccountFreeze> {
+    freeze_until(Utc::now() - TimeDelta::seconds(1))
+}
+
+#[tokio::test]
+async fn probe_success_clears_freeze() {
+    let runtime = FreezeRuntimeStore::new(freeze_due(), BTreeMap::new());
+    let (task, _store) = recovery_task(
+        runtime_settings(true, true, false),
+        Arc::clone(&runtime),
+        Arc::new(SuccessfulProbe),
+    )
+    .await;
+
+    run_cycle(&task).await;
+
+    assert_eq!(runtime.cleared(), vec!["acct_test".to_owned()]);
+    assert!(runtime.extended().is_empty());
+}
+
+#[tokio::test]
+async fn probe_failure_postpones_freeze() {
+    let runtime = FreezeRuntimeStore::new(freeze_due(), BTreeMap::new());
+    let (task, _store) = recovery_task(
+        runtime_settings(true, true, false),
+        Arc::clone(&runtime),
+        Arc::new(FailingProbe),
+    )
+    .await;
+
+    run_cycle(&task).await;
+
+    assert!(runtime.cleared().is_empty());
+    let extended = runtime.extended();
+    assert_eq!(extended.len(), 1);
+    let (account_id, until) = &extended[0];
+    assert_eq!(account_id, "acct_test");
+    let remaining = until.timestamp_millis() - Utc::now().timestamp_millis();
+    assert!(
+        remaining > 7_000_000 && remaining <= 7_200_000,
+        "postponed freeze should be about 2 hours, got {remaining}ms"
+    );
+}
+
+#[tokio::test]
+async fn disabled_policy_releases_due_freeze_without_probing() {
+    let runtime = FreezeRuntimeStore::new(freeze_due(), BTreeMap::new());
+    let (task, _store) = recovery_task(
+        runtime_settings(false, true, false),
+        Arc::clone(&runtime),
+        Arc::new(SuccessfulProbe),
+    )
+    .await;
+
+    run_cycle(&task).await;
+
+    assert_eq!(runtime.cleared(), vec!["acct_test".to_owned()]);
+    assert!(runtime.extended().is_empty());
+}
+
+#[tokio::test]
+async fn adaptive_concurrency_lowers_limit_to_observed_peak() {
+    let runtime = FreezeRuntimeStore::new(
+        freeze_until(Utc::now() + TimeDelta::hours(1)),
+        BTreeMap::from([("acct_test".to_owned(), 4_u32)]),
+    );
+    let (task, store) = recovery_task(
+        runtime_settings(true, false, true),
+        Arc::clone(&runtime),
+        Arc::new(SuccessfulProbe),
+    )
+    .await;
+
+    run_cycle(&task).await;
+
+    // 峰值 4 × 0.8 = 3（不低于下限 2），低于全局默认 5，应下调到 3。
+    assert_eq!(
+        *store.lowered_limits.lock().expect("lowered limits"),
+        vec![("acct_test".to_owned(), 3)]
+    );
+    assert!(store.update_commands().is_empty());
+    assert!(runtime.extended().is_empty());
+    assert!(runtime.cleared().is_empty());
+}
+
+#[tokio::test]
+async fn adaptive_concurrency_never_raises_limit() {
+    let runtime = FreezeRuntimeStore::new(
+        freeze_until(Utc::now() + TimeDelta::hours(1)),
+        BTreeMap::from([("acct_test".to_owned(), 40_u32)]),
+    );
+    let (task, store) = recovery_task(
+        runtime_settings(true, false, true),
+        Arc::clone(&runtime),
+        Arc::new(SuccessfulProbe),
+    )
+    .await;
+
+    run_cycle(&task).await;
+
+    // 峰值 40 的目标 32 高于全局默认 5；只降不升，不下发任何更新。
+    assert!(store.update_commands().is_empty());
+}
+
+/// 到期时间还早（超过探测提前量）的冻结不发起探测。
+#[tokio::test]
+async fn probe_skips_freezes_far_from_expiry() {
+    let runtime = FreezeRuntimeStore::new(
+        freeze_until(Utc::now() + TimeDelta::hours(1)),
+        BTreeMap::new(),
+    );
+    let (task, _store) = recovery_task(
+        runtime_settings(true, true, false),
+        Arc::clone(&runtime),
+        Arc::new(SuccessfulProbe),
+    )
+    .await;
+
+    run_cycle(&task).await;
+
+    assert!(runtime.cleared().is_empty());
+    assert!(runtime.extended().is_empty());
+}
+
+/// 停用账号不参与探测恢复；并发下调同样跳过。
+#[tokio::test]
+async fn disabled_accounts_are_skipped() {
+    let runtime = FreezeRuntimeStore::new(
+        freeze_due(),
+        BTreeMap::from([("acct_test".to_owned(), 4_u32)]),
+    );
+    let mut record = account_record("openai");
+    record.enabled = false;
+    let store = FakeAccountStore::with_account(record, events());
+    let services = AdminHarness::new()
+        .accounts(Arc::clone(&store) as Arc<dyn AccountStore>)
+        .account_runtime(Arc::clone(&runtime) as Arc<dyn AccountRuntimeStore>)
+        .settings(Arc::new(FreezeSettingsStore {
+            settings: runtime_settings(true, true, true),
+        }))
+        .provider(FakeProviderAdmin::new("openai", events()))
+        .probe(Arc::new(SuccessfulProbe))
+        .build()
+        .await;
+    let task = FreezeRecoveryTask::new(FreezeRecoveryDeps {
+        accounts: services.accounts_handle(),
+        store: store.clone(),
+        runtime: Arc::clone(&runtime) as Arc<dyn AccountRuntimeStore>,
+        settings: Arc::new(FreezeSettingsStore {
+            settings: runtime_settings(true, true, true),
+        }),
+    });
+
+    run_cycle(&task).await;
+
+    assert!(runtime.cleared().is_empty());
+    assert!(runtime.extended().is_empty());
+    assert!(store.update_commands().is_empty());
+}
diff --git a/backend/crates/gateway-admin/tests/use_case/mod.rs b/backend/crates/gateway-admin/tests/use_case/mod.rs
index 4c726f9db..f23e98188 100644
--- a/backend/crates/gateway-admin/tests/use_case/mod.rs
+++ b/backend/crates/gateway-admin/tests/use_case/mod.rs
@@ -4,6 +4,7 @@ mod auth;
 mod auth_key;
 mod backup;
 mod client_keys;
+mod freeze_recovery;
 mod import_tasks;
 mod observability;
 mod openai;
@@ -82,6 +83,7 @@ use gateway_core::{
     routing::{ConfigRevision, ProviderKind},
     runtime::SnapshotControl,
 };
+use std::collections::BTreeMap;
 
 pub(super) struct AdminHarness {
     default_password: String,
@@ -488,6 +490,15 @@ impl AccountStore for UnavailableStore {
         Err(unavailable("account enabled"))
     }
 
+    async fn lower_concurrency_limit(
+        &self,
+        _: &gateway_core::account::ProviderAccountId,
+        _: gateway_core::account::AccountConcurrencyLimit,
+        _: &MutationContext,
+    ) -> AdminStoreResult<Option<gateway_admin::model::accounts::AccountUpdateResult>> {
+        Ok(None)
+    }
+
     async fn recover_account(
         &self,
         _: &ProviderAccountId,
@@ -530,6 +541,28 @@ impl AccountRuntimeStore for UnavailableStore {
     async fn account_runtime(&self, _: &[String]) -> AdminStoreResult<AccountRuntimeSnapshot> {
         Ok(AccountRuntimeSnapshot::default())
     }
+
+    async fn active_freezes(
+        &self,
+    ) -> AdminStoreResult<BTreeMap<String, gateway_admin::model::accounts::AccountFreeze>> {
+        Ok(BTreeMap::new())
+    }
+
+    async fn capacity_peaks(
+        &self,
+        _account_ids: &[String],
+    ) -> AdminStoreResult<BTreeMap<String, u32>> {
+        Ok(BTreeMap::new())
+    }
+
+    async fn finish_freeze(
+        &self,
+        _account_id: &str,
+        _expected: &gateway_admin::model::accounts::AccountFreeze,
+        _postpone_until: Option<DateTime<Utc>>,
+    ) -> AdminStoreResult<bool> {
+        Ok(false)
+    }
 }
 
 #[async_trait]
diff --git a/backend/crates/gateway-admin/tests/use_case/observability.rs b/backend/crates/gateway-admin/tests/use_case/observability.rs
index f460d7c28..87dd2bd6f 100644
--- a/backend/crates/gateway-admin/tests/use_case/observability.rs
+++ b/backend/crates/gateway-admin/tests/use_case/observability.rs
@@ -769,6 +769,13 @@ impl SettingsStore for FixtureSettingsStore {
             ops_event_retention_days: 30,
             audit_retention_days: 30,
             updated_at: Utc::now(),
+            account_auto_freeze_enabled: true,
+            account_auto_freeze_threshold: 12,
+            account_auto_freeze_window_seconds: 600,
+            account_auto_freeze_duration_seconds: 7_200,
+            account_auto_freeze_probe_enabled: true,
+            account_auto_freeze_probe_model: None,
+            account_auto_freeze_adaptive_concurrency: true,
         })
     }
 
diff --git a/backend/crates/gateway-admin/tests/use_case/settings.rs b/backend/crates/gateway-admin/tests/use_case/settings.rs
index 31a1b0b69..7f67c2abb 100644
--- a/backend/crates/gateway-admin/tests/use_case/settings.rs
+++ b/backend/crates/gateway-admin/tests/use_case/settings.rs
@@ -75,6 +75,13 @@ async fn settings_should_reject_zero_refresh_margin_before_store_call() {
                 usage_retention_days: 31,
                 ops_event_retention_days: 30,
                 audit_retention_days: 30,
+                account_auto_freeze_enabled: true,
+                account_auto_freeze_threshold: 12,
+                account_auto_freeze_window_seconds: 600,
+                account_auto_freeze_duration_seconds: 7_200,
+                account_auto_freeze_probe_enabled: true,
+                account_auto_freeze_probe_model: None,
+                account_auto_freeze_adaptive_concurrency: true,
             },
         )
         .await
diff --git a/backend/crates/gateway-api/src/admin/accounts/presenter.rs b/backend/crates/gateway-api/src/admin/accounts/presenter.rs
index 9bc6bb252..2f2980f1b 100644
--- a/backend/crates/gateway-api/src/admin/accounts/presenter.rs
+++ b/backend/crates/gateway-api/src/admin/accounts/presenter.rs
@@ -66,7 +66,7 @@ pub(super) fn account_view(item: AccountDirectoryItem, now: DateTime<Utc>) -> Ac
         quota,
     } = item;
     let status = projection.status.as_str().to_owned();
-    let rate_limited_until = projection.rate_limited_until.map(DateTime::<Utc>::from);
+    let cooldown = projection.cooldown;
     let expires_at = account.access_token_expires_at.as_ref().map(china_rfc3339);
     let added_at = china_rfc3339(&account.created_at);
     let updated_at = china_rfc3339(&account.updated_at);
@@ -75,7 +75,7 @@ pub(super) fn account_view(item: AccountDirectoryItem, now: DateTime<Utc>) -> Ac
     if account.authentication_kind == "api_key" {
         usage.window_label_display = "通用额度".to_owned();
     }
-    let (quota, refresh_token_expires_at) = account_quota_view(quota, rate_limited_until, now);
+    let (quota, refresh_token_expires_at) = account_quota_view(quota, cooldown, now);
     AccountView {
         id: account.id.clone(),
         name: account.name,
@@ -189,7 +189,7 @@ fn forecast_usd_display(value: Option<f64>) -> String {
 
 pub(super) fn account_quota_view(
     mut quota: ProviderQuota,
-    rate_limited_until: Option<DateTime<Utc>>,
+    cooldown: Option<gateway_core::account::AccountCooldown>,
     now: DateTime<Utc>,
 ) -> (AccountQuotaView, Option<String>) {
     quota.apply_limit_reached_display();
@@ -200,12 +200,22 @@ pub(super) fn account_quota_view(
         .observed_at
         .map_or_else(|| "—".to_owned(), |value| relative_time(value, now));
     let windows = quota.windows.into_iter().map(quota_window_view).collect();
-    let rate_limited_until = rate_limited_until.map(|until| china_datetime(&until));
+    let rate_limited_until = cooldown.map(|value| china_datetime(&value.until.into()));
+    let rate_limit_reason = cooldown.map(|value| {
+        if value.kind.is_capacity_freeze() {
+            "capacity_freeze".to_owned()
+        } else {
+            "upstream_rate_limit".to_owned()
+        }
+    });
+    let recovery_probe_required = cooldown.is_some_and(|value| value.kind.requires_probe());
     (
         AccountQuotaView {
             refreshed_at_display,
             limit_reached: quota.limit_reached,
             rate_limited_until,
+            rate_limit_reason,
+            recovery_probe_required,
             windows,
         },
         refresh_token_expires_at,
diff --git a/backend/crates/gateway-api/src/admin/accounts/wire.rs b/backend/crates/gateway-api/src/admin/accounts/wire.rs
index e90ca705b..cee3b5eab 100644
--- a/backend/crates/gateway-api/src/admin/accounts/wire.rs
+++ b/backend/crates/gateway-api/src/admin/accounts/wire.rs
@@ -361,8 +361,10 @@ pub struct AccountGroupRefView {
 pub struct AccountQuotaView {
     pub refreshed_at_display: String,
     pub limit_reached: bool,
-    /// 429 临时限流（Redis 冷却）到期时间展示；非限流中为 `null`。
+    /// 冷却到期或可开始恢复探测的时间；非限流中为 `null`。
     pub rate_limited_until: Option<String>,
+    pub rate_limit_reason: Option<String>,
+    pub recovery_probe_required: bool,
     pub windows: Vec<AccountQuotaWindowView>,
 }
 
diff --git a/backend/crates/gateway-api/src/admin/settings.rs b/backend/crates/gateway-api/src/admin/settings.rs
index 34046b692..7547a8774 100644
--- a/backend/crates/gateway-api/src/admin/settings.rs
+++ b/backend/crates/gateway-api/src/admin/settings.rs
@@ -48,6 +48,13 @@ pub struct RuntimeSettingsView {
     pub usage_retention_days: u64,
     pub ops_event_retention_days: u64,
     pub audit_retention_days: u64,
+    pub account_auto_freeze_enabled: bool,
+    pub account_auto_freeze_threshold: u64,
+    pub account_auto_freeze_window_seconds: u64,
+    pub account_auto_freeze_duration_seconds: u64,
+    pub account_auto_freeze_probe_enabled: bool,
+    pub account_auto_freeze_probe_model: Option<String>,
+    pub account_auto_freeze_adaptive_concurrency: bool,
     pub updated_at: DateTime<Utc>,
 }
 
@@ -69,6 +76,13 @@ pub struct UpdateRuntimeSettingsRequest {
     pub usage_retention_days: u64,
     pub ops_event_retention_days: u64,
     pub audit_retention_days: u64,
+    pub account_auto_freeze_enabled: bool,
+    pub account_auto_freeze_threshold: u64,
+    pub account_auto_freeze_window_seconds: u64,
+    pub account_auto_freeze_duration_seconds: u64,
+    pub account_auto_freeze_probe_enabled: bool,
+    pub account_auto_freeze_probe_model: Option<String>,
+    pub account_auto_freeze_adaptive_concurrency: bool,
 }
 
 impl UpdateRuntimeSettingsRequest {
@@ -110,6 +124,35 @@ impl UpdateRuntimeSettingsRequest {
             self.min_codex_cli_version.as_deref(),
             "minCodexCliVersion",
         )?;
+        for (value, field) in [
+            (
+                self.account_auto_freeze_threshold,
+                "accountAutoFreezeThreshold",
+            ),
+            (
+                self.account_auto_freeze_window_seconds,
+                "accountAutoFreezeWindowSeconds",
+            ),
+            (
+                self.account_auto_freeze_duration_seconds,
+                "accountAutoFreezeDurationSeconds",
+            ),
+        ] {
+            require_positive_i64(value, field)?;
+        }
+        if !(2..=1_000).contains(&self.account_auto_freeze_threshold) {
+            return Err(WireValidationError::new("accountAutoFreezeThreshold"));
+        }
+        if !(60..=3_600).contains(&self.account_auto_freeze_window_seconds) {
+            return Err(WireValidationError::new("accountAutoFreezeWindowSeconds"));
+        }
+        if !(300..=604_800).contains(&self.account_auto_freeze_duration_seconds) {
+            return Err(WireValidationError::new("accountAutoFreezeDurationSeconds"));
+        }
+        validate_optional_probe_model(
+            self.account_auto_freeze_probe_model.as_deref(),
+            "accountAutoFreezeProbeModel",
+        )?;
         Ok(())
     }
 
@@ -136,6 +179,14 @@ impl UpdateRuntimeSettingsRequest {
                 .map_err(|_| WireValidationError::new("settingsOpsRetentionOverflow"))?,
             audit_retention_days: u32::try_from(self.audit_retention_days)
                 .map_err(|_| WireValidationError::new("settingsAuditRetentionOverflow"))?,
+            account_auto_freeze_enabled: self.account_auto_freeze_enabled,
+            account_auto_freeze_threshold: u32::try_from(self.account_auto_freeze_threshold)
+                .map_err(|_| WireValidationError::new("settingsFreezeThresholdOverflow"))?,
+            account_auto_freeze_window_seconds: self.account_auto_freeze_window_seconds,
+            account_auto_freeze_duration_seconds: self.account_auto_freeze_duration_seconds,
+            account_auto_freeze_probe_enabled: self.account_auto_freeze_probe_enabled,
+            account_auto_freeze_probe_model: self.account_auto_freeze_probe_model,
+            account_auto_freeze_adaptive_concurrency: self.account_auto_freeze_adaptive_concurrency,
         })
     }
 }
@@ -157,6 +208,14 @@ impl From<RuntimeSettings> for RuntimeSettingsView {
             usage_retention_days: u64::from(settings.usage_retention_days),
             ops_event_retention_days: u64::from(settings.ops_event_retention_days),
             audit_retention_days: u64::from(settings.audit_retention_days),
+            account_auto_freeze_enabled: settings.account_auto_freeze_enabled,
+            account_auto_freeze_threshold: u64::from(settings.account_auto_freeze_threshold),
+            account_auto_freeze_window_seconds: settings.account_auto_freeze_window_seconds,
+            account_auto_freeze_duration_seconds: settings.account_auto_freeze_duration_seconds,
+            account_auto_freeze_probe_enabled: settings.account_auto_freeze_probe_enabled,
+            account_auto_freeze_probe_model: settings.account_auto_freeze_probe_model,
+            account_auto_freeze_adaptive_concurrency: settings
+                .account_auto_freeze_adaptive_concurrency,
             updated_at: settings.updated_at,
         }
     }
@@ -452,6 +511,22 @@ fn validate_optional_client_version(
     Ok(())
 }
 
+/// 探测模型为可选自由文本：非空、去首尾空白后不变、无控制字符且不超过 128 字节。
+fn validate_optional_probe_model(
+    value: Option<&str>,
+    field: &'static str,
+) -> Result<(), WireValidationError> {
+    if value.is_some_and(|value| {
+        value.is_empty()
+            || value.len() > 128
+            || value != value.trim()
+            || value.bytes().any(|byte| byte.is_ascii_control())
+    }) {
+        return Err(WireValidationError::new(field));
+    }
+    Ok(())
+}
+
 fn map_wire_error(error: WireValidationError) -> AdminError {
     let message = match error.field() {
         "settingsRefreshConcurrencyOverflow" => "refreshConcurrency 不合法".to_owned(),
@@ -459,6 +534,11 @@ fn map_wire_error(error: WireValidationError) -> AdminError {
         "settingsUsageRetentionOverflow" => "usageRetentionDays 不合法".to_owned(),
         "settingsOpsRetentionOverflow" => "opsEventRetentionDays 不合法".to_owned(),
         "settingsAuditRetentionOverflow" => "auditRetentionDays 不合法".to_owned(),
+        "settingsFreezeThresholdOverflow" => "accountAutoFreezeThreshold 不合法".to_owned(),
+        "accountAutoFreezeThreshold" => "账号自动冻结阈值应为 2～1000 的整数".to_owned(),
+        "accountAutoFreezeWindowSeconds" => "账号自动冻结统计窗口应为 60～3600 秒".to_owned(),
+        "accountAutoFreezeDurationSeconds" => "账号自动冻结时长应为 300～604800 秒".to_owned(),
+        "accountAutoFreezeProbeModel" => "探测模型格式不合法".to_owned(),
         "minCodexDesktopVersion" => "Codex Desktop 最低版本格式不合法".to_owned(),
         "minCodexCliVersion" => "Codex CLI 最低版本格式不合法".to_owned(),
         field => format!("{field} 字段不合法"),
diff --git a/backend/crates/gateway-api/tests/admin/errors.rs b/backend/crates/gateway-api/tests/admin/errors.rs
index b84c653e3..b4ecc9932 100644
--- a/backend/crates/gateway-api/tests/admin/errors.rs
+++ b/backend/crates/gateway-api/tests/admin/errors.rs
@@ -463,7 +463,7 @@ mod provider {
             credential_state: CredentialState::Ready,
             access_token_expires_at: None,
             quota: QuotaState::unknown(),
-            rate_limited_until: None,
+            cooldown: None,
             last_error_reason: None,
             last_error_message: None,
         };
diff --git a/backend/crates/gateway-api/tests/admin/mod.rs b/backend/crates/gateway-api/tests/admin/mod.rs
index cd8079eb6..0f2e9b0cf 100644
--- a/backend/crates/gateway-api/tests/admin/mod.rs
+++ b/backend/crates/gateway-api/tests/admin/mod.rs
@@ -417,6 +417,13 @@ impl SettingsStore for MemorySettingsStore {
             usage_retention_days: command.usage_retention_days,
             ops_event_retention_days: command.ops_event_retention_days,
             audit_retention_days: command.audit_retention_days,
+            account_auto_freeze_enabled: true,
+            account_auto_freeze_threshold: 12,
+            account_auto_freeze_window_seconds: 600,
+            account_auto_freeze_duration_seconds: 7_200,
+            account_auto_freeze_probe_enabled: true,
+            account_auto_freeze_probe_model: None,
+            account_auto_freeze_adaptive_concurrency: true,
             updated_at: Utc::now(),
         };
         *settings = updated.clone();
@@ -587,7 +594,7 @@ impl AccountGroupStore for MemoryAccountGroupStore {
                     credential_state: CredentialState::Ready,
                     access_token_expires_at: None,
                     quota: QuotaState::default(),
-                    rate_limited_until: None,
+                    cooldown: None,
                     last_error_reason: None,
                     last_error_message: None,
                 },
@@ -601,7 +608,7 @@ impl AccountGroupStore for MemoryAccountGroupStore {
                     credential_state: CredentialState::Ready,
                     access_token_expires_at: None,
                     quota: QuotaState::default(),
-                    rate_limited_until: None,
+                    cooldown: None,
                     last_error_reason: None,
                     last_error_message: None,
                 },
@@ -896,6 +903,15 @@ impl AccountStore for UnusedStore {
         Err(unavailable("account enabled"))
     }
 
+    async fn lower_concurrency_limit(
+        &self,
+        _: &gateway_core::account::ProviderAccountId,
+        _: gateway_core::account::AccountConcurrencyLimit,
+        _: &MutationContext,
+    ) -> AdminStoreResult<Option<gateway_admin::model::accounts::AccountUpdateResult>> {
+        Ok(None)
+    }
+
     async fn recover_account(
         &self,
         _: &ProviderAccountId,
@@ -933,17 +949,39 @@ impl AccountStore for UnusedStore {
 impl AccountRuntimeStore for UnusedStore {
     async fn active_rate_limits(&self) -> AdminStoreResult<AccountRuntimeSnapshot> {
         Ok(AccountRuntimeSnapshot {
-            rate_limited_until: BTreeMap::new(),
+            cooldown: BTreeMap::new(),
             in_flight: Some(BTreeMap::new()),
         })
     }
 
     async fn account_runtime(&self, _: &[String]) -> AdminStoreResult<AccountRuntimeSnapshot> {
         Ok(AccountRuntimeSnapshot {
-            rate_limited_until: BTreeMap::new(),
+            cooldown: BTreeMap::new(),
             in_flight: Some(BTreeMap::new()),
         })
     }
+
+    async fn active_freezes(
+        &self,
+    ) -> AdminStoreResult<BTreeMap<String, gateway_admin::model::accounts::AccountFreeze>> {
+        Ok(BTreeMap::new())
+    }
+
+    async fn capacity_peaks(
+        &self,
+        _account_ids: &[String],
+    ) -> AdminStoreResult<BTreeMap<String, u32>> {
+        Ok(BTreeMap::new())
+    }
+
+    async fn finish_freeze(
+        &self,
+        _account_id: &str,
+        _expected: &gateway_admin::model::accounts::AccountFreeze,
+        _postpone_until: Option<DateTime<Utc>>,
+    ) -> AdminStoreResult<bool> {
+        Ok(false)
+    }
 }
 
 #[async_trait]
@@ -1245,6 +1283,13 @@ fn test_runtime_settings() -> RuntimeSettings {
         usage_retention_days: 31,
         ops_event_retention_days: 30,
         audit_retention_days: 90,
+        account_auto_freeze_enabled: true,
+        account_auto_freeze_threshold: 12,
+        account_auto_freeze_window_seconds: 600,
+        account_auto_freeze_duration_seconds: 7_200,
+        account_auto_freeze_probe_enabled: true,
+        account_auto_freeze_probe_model: None,
+        account_auto_freeze_adaptive_concurrency: true,
         updated_at: Utc::now(),
     }
 }
diff --git a/backend/crates/gateway-api/tests/admin/settings.rs b/backend/crates/gateway-api/tests/admin/settings.rs
index 222d38a14..139342e75 100644
--- a/backend/crates/gateway-api/tests/admin/settings.rs
+++ b/backend/crates/gateway-api/tests/admin/settings.rs
@@ -53,7 +53,14 @@ fn update_body() -> Value {
         "minCodexCliVersion": "0.40.0",
         "usageRetentionDays": 32,
         "opsEventRetentionDays": 31,
-        "auditRetentionDays": 91
+        "auditRetentionDays": 91,
+        "accountAutoFreezeEnabled": true,
+        "accountAutoFreezeThreshold": 12,
+        "accountAutoFreezeWindowSeconds": 600,
+        "accountAutoFreezeDurationSeconds": 7200,
+        "accountAutoFreezeProbeEnabled": true,
+        "accountAutoFreezeProbeModel": null,
+        "accountAutoFreezeAdaptiveConcurrency": true
     })
 }
 
@@ -116,6 +123,13 @@ fn settings_response_should_cover_the_full_runtime_settings_contract() {
         usage_retention_days: 32,
         ops_event_retention_days: 31,
         audit_retention_days: 91,
+        account_auto_freeze_enabled: true,
+        account_auto_freeze_threshold: 12,
+        account_auto_freeze_window_seconds: 600,
+        account_auto_freeze_duration_seconds: 7_200,
+        account_auto_freeze_probe_enabled: true,
+        account_auto_freeze_probe_model: None,
+        account_auto_freeze_adaptive_concurrency: true,
         updated_at: Utc
             .with_ymd_and_hms(2026, 8, 2, 10, 30, 0)
             .single()
@@ -143,6 +157,13 @@ fn settings_response_should_cover_the_full_runtime_settings_contract() {
             "usageRetentionDays": 32,
             "opsEventRetentionDays": 31,
             "auditRetentionDays": 91,
+            "accountAutoFreezeEnabled": true,
+            "accountAutoFreezeThreshold": 12,
+            "accountAutoFreezeWindowSeconds": 600,
+            "accountAutoFreezeDurationSeconds": 7200,
+            "accountAutoFreezeProbeEnabled": true,
+            "accountAutoFreezeProbeModel": null,
+            "accountAutoFreezeAdaptiveConcurrency": true,
             "updatedAt": "2026-08-02T10:30:00Z"
         })
     );
@@ -195,6 +216,13 @@ fn settings_request_and_response_fields_should_stay_in_lockstep() {
         usage_retention_days: u32::try_from(request.usage_retention_days).expect("u32"),
         ops_event_retention_days: u32::try_from(request.ops_event_retention_days).expect("u32"),
         audit_retention_days: u32::try_from(request.audit_retention_days).expect("u32"),
+        account_auto_freeze_enabled: true,
+        account_auto_freeze_threshold: 12,
+        account_auto_freeze_window_seconds: 600,
+        account_auto_freeze_duration_seconds: 7_200,
+        account_auto_freeze_probe_enabled: true,
+        account_auto_freeze_probe_model: None,
+        account_auto_freeze_adaptive_concurrency: true,
         updated_at: chrono::Utc::now(),
     };
 
diff --git a/backend/crates/gateway-core/src/account/model.rs b/backend/crates/gateway-core/src/account/model.rs
index dcedeceb3..aa4c60d96 100644
--- a/backend/crates/gateway-core/src/account/model.rs
+++ b/backend/crates/gateway-core/src/account/model.rs
@@ -583,6 +583,72 @@ impl AccountErrorReason {
     }
 }
 
+/// 账号冷却的原因与恢复方式。
+#[derive(Debug, Clone, Copy, PartialEq, Eq, Default)]
+pub enum AccountCooldownKind {
+    /// 上游 429 临时限流；默认类别，兼容升级前未标记的存量 key。
+    #[default]
+    RateLimit,
+    /// 容量类错误高频触发后由熔断策略写入的自动冻结。
+    CapacityFreeze,
+    /// 到期后仍阻止调度，必须由恢复探测成功或管理员恢复解除。
+    CapacityFreezeProbe,
+}
+
+impl AccountCooldownKind {
+    #[must_use]
+    pub const fn as_str(self) -> &'static str {
+        match self {
+            Self::RateLimit => "rate_limit",
+            Self::CapacityFreeze => "capacity_freeze",
+            Self::CapacityFreezeProbe => "capacity_freeze_probe",
+        }
+    }
+
+    #[must_use]
+    pub const fn is_capacity_freeze(self) -> bool {
+        matches!(self, Self::CapacityFreeze | Self::CapacityFreezeProbe)
+    }
+
+    #[must_use]
+    pub const fn requires_probe(self) -> bool {
+        matches!(self, Self::CapacityFreezeProbe)
+    }
+
+    #[must_use]
+    pub fn parse(value: &str) -> Option<Self> {
+        match value {
+            "rate_limit" => Some(Self::RateLimit),
+            "capacity_freeze" => Some(Self::CapacityFreeze),
+            "capacity_freeze_probe" => Some(Self::CapacityFreezeProbe),
+            _ => None,
+        }
+    }
+}
+
+/// 账号级冷却投影；探测冻结的 until 是下次探测时间，不是自动放行时间。
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub struct AccountCooldown {
+    pub until: SystemTime,
+    pub kind: AccountCooldownKind,
+}
+
+impl AccountCooldown {
+    #[must_use]
+    pub fn is_active(self, now: SystemTime) -> bool {
+        self.kind.requires_probe() || self.until > now
+    }
+}
+
+impl From<SystemTime> for AccountCooldown {
+    fn from(until: SystemTime) -> Self {
+        Self {
+            until,
+            kind: AccountCooldownKind::RateLimit,
+        }
+    }
+}
+
 /// 唯一状态解析器的完整输入事实。
 #[derive(Debug, Clone, PartialEq, Eq)]
 pub struct AccountStatusFacts {
@@ -590,7 +656,7 @@ pub struct AccountStatusFacts {
     pub credential_state: CredentialState,
     pub access_token_expires_at: Option<SystemTime>,
     pub quota: QuotaState,
-    pub rate_limited_until: Option<SystemTime>,
+    pub cooldown: Option<AccountCooldown>,
     pub last_error_reason: Option<AccountErrorReason>,
     pub last_error_message: Option<String>,
 }
@@ -601,8 +667,8 @@ pub struct AccountStatusProjection {
     pub status: AccountStatus,
     pub error_reason: Option<AccountErrorReason>,
     pub error_message: Option<String>,
-    /// 仅 `rate_limited` 状态携带仍有效的运行时冷却截止时间。
-    pub rate_limited_until: Option<SystemTime>,
+    /// 仅 `rate_limited` 状态携带仍有效的运行时冷却事实。
+    pub cooldown: Option<AccountCooldown>,
 }
 
 /// 从独立事实派生唯一、互斥的对外状态。
@@ -625,21 +691,21 @@ pub fn resolve_account_status(
             status: AccountStatus::Error,
             error_reason: Some(facts.last_error_reason.unwrap_or(default_reason)),
             error_message: facts.last_error_message.clone(),
-            rate_limited_until: None,
+            cooldown: None,
         };
     }
     if facts.quota.is_exhausted() {
         return status_projection(AccountStatus::QuotaExhausted);
     }
     if facts
-        .rate_limited_until
-        .is_some_and(|rate_limited_until| rate_limited_until > now)
+        .cooldown
+        .is_some_and(|cooldown| cooldown.is_active(now))
     {
         return AccountStatusProjection {
             status: AccountStatus::RateLimited,
             error_reason: None,
             error_message: None,
-            rate_limited_until: facts.rate_limited_until,
+            cooldown: facts.cooldown,
         };
     }
     status_projection(AccountStatus::Normal)
@@ -650,7 +716,7 @@ const fn status_projection(status: AccountStatus) -> AccountStatusProjection {
         status,
         error_reason: None,
         error_message: None,
-        rate_limited_until: None,
+        cooldown: None,
     }
 }
 
@@ -905,7 +971,7 @@ impl ProviderAccount {
     pub fn status_projection(
         &self,
         now: SystemTime,
-        rate_limited_until: Option<SystemTime>,
+        cooldown: Option<AccountCooldown>,
     ) -> AccountStatusProjection {
         resolve_account_status(
             &AccountStatusFacts {
@@ -913,7 +979,7 @@ impl ProviderAccount {
                 credential_state: self.credential_state,
                 access_token_expires_at: self.access_token_expires_at,
                 quota: self.quota,
-                rate_limited_until,
+                cooldown,
                 last_error_reason: self.last_error_reason,
                 last_error_message: self.last_error_message.clone(),
             },
diff --git a/backend/crates/gateway-core/src/account/selection.rs b/backend/crates/gateway-core/src/account/selection.rs
index 12c926801..77a13a09e 100644
--- a/backend/crates/gateway-core/src/account/selection.rs
+++ b/backend/crates/gateway-core/src/account/selection.rs
@@ -106,7 +106,7 @@ pub struct AccountRuntimeSignals {
     pub quota_reset_at: Option<SystemTime>,
     /// Provider 归一化的剩余额度基点：0 耗尽，10_000 全部可用。
     pub quota_remaining_rank: Option<u64>,
-    pub rate_limited_until: Option<SystemTime>,
+    pub cooldown: Option<super::AccountCooldown>,
     pub failure_rate_basis_points: Option<u16>,
     pub first_output_latency_ms: Option<u64>,
 }
@@ -316,8 +316,8 @@ impl AccountRuntimeSignals {
     }
 
     #[must_use]
-    pub const fn with_rate_limit(mut self, rate_limited_until: Option<SystemTime>) -> Self {
-        self.rate_limited_until = rate_limited_until;
+    pub const fn with_rate_limit(mut self, cooldown: Option<super::AccountCooldown>) -> Self {
+        self.cooldown = cooldown;
         self
     }
 
@@ -654,7 +654,7 @@ impl AccountSelector {
         if !context.eligibility.bypasses_local_eligibility() {
             let status = candidate
                 .account
-                .status_projection(context.now, candidate.signals.rate_limited_until)
+                .status_projection(context.now, candidate.signals.cooldown)
                 .status;
             if status != AccountStatus::Normal {
                 return Some(AccountSchedulingBlocker::LocalAvailability);
diff --git a/backend/crates/gateway-core/src/provider_ports/mod.rs b/backend/crates/gateway-core/src/provider_ports/mod.rs
index ed621a7f7..56e6c1482 100644
--- a/backend/crates/gateway-core/src/provider_ports/mod.rs
+++ b/backend/crates/gateway-core/src/provider_ports/mod.rs
@@ -181,6 +181,16 @@ pub trait ProviderLeasePort: Send + Sync {
         &self,
         request: ProviderLeaseRequest,
     ) -> BoxFuture<'_, Result<ProviderLeaseAcquisition, ProviderStoreError>>;
+
+    /// 读取指定账号当前的在途请求数；只用于容量熔断的峰值证据，
+    /// 支持租约信号的存储实现覆盖，否则视为不可观测（空映射）。
+    fn account_in_flight<'a>(
+        &'a self,
+        account_ids: &'a [ProviderAccountId],
+    ) -> BoxFuture<'a, Result<BTreeMap<ProviderAccountId, u32>, ProviderStoreError>> {
+        let _ = account_ids;
+        Box::pin(async move { Ok(BTreeMap::new()) })
+    }
 }
 
 /// Provider 从原始会话锚点派生的不可逆亲和键。
@@ -596,12 +606,16 @@ pub trait ProviderCredentialStatePort: Send + Sync {
     ) -> BoxFuture<'a, Result<(), ProviderStoreError>>;
 }
 
-/// 临时 cooldown 只保存可丢失的调度截止时间，不进入账号持久状态。
+/// 账号级 cooldown 的来源类别；调度状态统一按 `rate_limited` 处理。
+pub use crate::account::AccountCooldownKind as ProviderCooldownKind;
+
+/// 可丢失的账号冷却事实，不进入持久状态；探测冻结到期后仍需确认恢复。
 #[derive(Debug, Clone, PartialEq, Eq)]
 pub struct ProviderCooldown {
     account_id: ProviderAccountId,
     credential_revision: CredentialRevision,
     until: SystemTime,
+    kind: ProviderCooldownKind,
 }
 
 impl ProviderCooldown {
@@ -610,11 +624,27 @@ impl ProviderCooldown {
         account_id: ProviderAccountId,
         credential_revision: CredentialRevision,
         until: SystemTime,
+    ) -> Self {
+        Self::new_with_kind(
+            account_id,
+            credential_revision,
+            until,
+            ProviderCooldownKind::RateLimit,
+        )
+    }
+
+    #[must_use]
+    pub const fn new_with_kind(
+        account_id: ProviderAccountId,
+        credential_revision: CredentialRevision,
+        until: SystemTime,
+        kind: ProviderCooldownKind,
     ) -> Self {
         Self {
             account_id,
             credential_revision,
             until,
+            kind,
         }
     }
 
@@ -632,6 +662,19 @@ impl ProviderCooldown {
     pub const fn until(&self) -> SystemTime {
         self.until
     }
+
+    #[must_use]
+    pub const fn scheduling_state(&self) -> crate::account::AccountCooldown {
+        crate::account::AccountCooldown {
+            until: self.until,
+            kind: self.kind,
+        }
+    }
+
+    #[must_use]
+    pub const fn kind(&self) -> ProviderCooldownKind {
+        self.kind
+    }
 }
 
 /// 可丢失 cooldown 的细粒度作用域。
@@ -749,6 +792,29 @@ pub trait ProviderCooldownPort: Send + Sync {
         &'a self,
         account_id: &'a ProviderAccountId,
     ) -> BoxFuture<'a, Result<bool, ProviderStoreError>>;
+
+    /// 记录一次容量类失败并返回滑动窗口内的累计次数，同时把观测到的账号
+    /// 在途并发并入窗口峰值（`in_flight` 为 0 表示本次未观测，跳过峰值更新）。
+    /// 只服务容量熔断触发器；调用频率受失败频率约束，不需要批量接口。
+    fn record_capacity_failure<'a>(
+        &'a self,
+        account_id: &'a ProviderAccountId,
+        window: Duration,
+        in_flight: u32,
+    ) -> BoxFuture<'a, Result<u32, ProviderStoreError>>;
+
+    /// 普通请求成功后原子清除临时限流及失败证据；必须保留任何容量冻结。
+    fn clear_after_success<'a>(
+        &'a self,
+        account_id: &'a ProviderAccountId,
+        through_revision: CredentialRevision,
+    ) -> BoxFuture<'a, Result<(), ProviderStoreError>>;
+
+    /// 读取窗口内观测到的在途并发峰值；无证据时返回 `None`。
+    fn capacity_peak_in_flight<'a>(
+        &'a self,
+        account_id: &'a ProviderAccountId,
+    ) -> BoxFuture<'a, Result<Option<u32>, ProviderStoreError>>;
 }
 
 #[derive(Debug, Clone, Copy, PartialEq, Eq)]
@@ -864,6 +930,118 @@ pub trait ProviderRuntimePolicyPort: Send + Sync {
     fn load_refresh_policy(
         &self,
     ) -> BoxFuture<'_, Result<ProviderRefreshPolicy, ProviderStoreError>>;
+
+    /// 读取账号容量熔断策略；默认关闭，只有实现运行时设置的存储需要覆盖。
+    fn load_freeze_policy(
+        &self,
+    ) -> BoxFuture<'_, Result<ProviderFreezePolicy, ProviderStoreError>> {
+        Box::pin(async move { Ok(ProviderFreezePolicy::disabled()) })
+    }
+}
+
+/// 账号容量熔断（自动冻结）策略；来源于 `runtime_settings`，
+/// 由 Provider 触发路径与恢复 worker 共享同一份配置事实。
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub struct ProviderFreezePolicy {
+    enabled: bool,
+    threshold: u32,
+    window: Duration,
+    freeze_duration: Duration,
+    probe_enabled: bool,
+    probe_model: Option<String>,
+    adaptive_concurrency: bool,
+}
+
+impl ProviderFreezePolicy {
+    /// 边界与迁移 `0010_account_auto_freeze.sql` 的 check 约束一致；
+    /// store 层写入前已校验，这里兜底防御越界配置。
+    pub fn try_new(
+        enabled: bool,
+        threshold: u32,
+        window_seconds: u64,
+        freeze_duration_seconds: u64,
+        probe_enabled: bool,
+        probe_model: Option<String>,
+        adaptive_concurrency: bool,
+    ) -> Result<Self, ProviderStoreError> {
+        if !(2..=1_000).contains(&threshold)
+            || !(60..=3_600).contains(&window_seconds)
+            || !(300..=604_800).contains(&freeze_duration_seconds)
+            || probe_model.as_deref().is_some_and(|model| {
+                model.is_empty()
+                    || model.len() > 128
+                    || model != model.trim()
+                    || model.bytes().any(|byte| byte.is_ascii_control())
+            })
+        {
+            return Err(ProviderStoreError::new(
+                ProviderStoreErrorKind::InvalidData,
+                "validate freeze policy",
+            ));
+        }
+        Ok(Self {
+            enabled,
+            threshold,
+            window: Duration::from_secs(window_seconds),
+            freeze_duration: Duration::from_secs(freeze_duration_seconds),
+            probe_enabled,
+            probe_model,
+            adaptive_concurrency,
+        })
+    }
+
+    /// 功能关闭时的全零策略；触发路径与 worker 都以此短路。
+    #[must_use]
+    pub const fn disabled() -> Self {
+        Self {
+            enabled: false,
+            threshold: 2,
+            window: Duration::from_secs(60),
+            freeze_duration: Duration::from_secs(300),
+            probe_enabled: false,
+            probe_model: None,
+            adaptive_concurrency: false,
+        }
+    }
+
+    #[must_use]
+    pub const fn enabled(&self) -> bool {
+        self.enabled
+    }
+
+    /// 窗口内触发冻结的请求级失败次数阈值。
+    #[must_use]
+    pub const fn threshold(&self) -> u32 {
+        self.threshold
+    }
+
+    /// 失败计数滑动窗口；每次失败都会顺延窗口。
+    #[must_use]
+    pub const fn window(&self) -> Duration {
+        self.window
+    }
+
+    /// 冻结时长；探测失败后的顺延也使用该值。
+    #[must_use]
+    pub const fn freeze_duration(&self) -> Duration {
+        self.freeze_duration
+    }
+
+    #[must_use]
+    pub const fn probe_enabled(&self) -> bool {
+        self.probe_enabled
+    }
+
+    /// 探测模型；`None` 表示由 worker 选择账号可用的第一个模型。
+    #[must_use]
+    pub fn probe_model(&self) -> Option<&str> {
+        self.probe_model.as_deref()
+    }
+
+    #[must_use]
+    pub const fn adaptive_concurrency(&self) -> bool {
+        self.adaptive_concurrency
+    }
 }
 
 /// OAuth pending flow 的原始绑定只在 Provider 与 Store 边界内短暂存在。
diff --git a/backend/crates/gateway-core/src/task/mod.rs b/backend/crates/gateway-core/src/task/mod.rs
index 166fee18c..613bfbbae 100644
--- a/backend/crates/gateway-core/src/task/mod.rs
+++ b/backend/crates/gateway-core/src/task/mod.rs
@@ -23,10 +23,11 @@ pub enum WorkerKind {
     OpsFlush,
     Backup,
     AccountImport,
+    AccountFreezeRecovery,
 }
 
 impl WorkerKind {
-    pub const ALL: [Self; 9] = [
+    pub const ALL: [Self; 10] = [
         Self::OAuthRefresh,
         Self::QuotaCatalogHealth,
         Self::RuntimeSnapshotReconciliation,
@@ -36,6 +37,7 @@ impl WorkerKind {
         Self::OpsFlush,
         Self::Backup,
         Self::AccountImport,
+        Self::AccountFreezeRecovery,
     ];
 
     #[must_use]
@@ -50,6 +52,7 @@ impl WorkerKind {
             Self::OpsFlush => "ops_flush",
             Self::Backup => "backup",
             Self::AccountImport => "account_import",
+            Self::AccountFreezeRecovery => "account_freeze_recovery",
         }
     }
 }
diff --git a/backend/crates/gateway-core/tests/account/mod.rs b/backend/crates/gateway-core/tests/account/mod.rs
index c92bece3c..0c4bb707d 100644
--- a/backend/crates/gateway-core/tests/account/mod.rs
+++ b/backend/crates/gateway-core/tests/account/mod.rs
@@ -49,7 +49,7 @@ fn candidate(id: &str, in_flight: u32, remaining: Option<u64>) -> AccountCandida
             last_started_at: None,
             quota_reset_at: None,
             quota_remaining_rank: remaining,
-            rate_limited_until: None,
+            cooldown: None,
             failure_rate_basis_points: None,
             first_output_latency_ms: None,
         },
@@ -222,7 +222,7 @@ fn diagnostic_selection_bypasses_all_local_account_eligibility() {
             last_started_at: None,
             quota_reset_at: None,
             quota_remaining_rank: None,
-            rate_limited_until: None,
+            cooldown: None,
             failure_rate_basis_points: None,
             first_output_latency_ms: None,
         },
@@ -240,7 +240,7 @@ fn diagnostic_selection_bypasses_all_local_account_eligibility() {
             last_started_at: None,
             quota_reset_at: None,
             quota_remaining_rank: None,
-            rate_limited_until: None,
+            cooldown: None,
             failure_rate_basis_points: None,
             first_output_latency_ms: None,
         },
@@ -313,13 +313,13 @@ fn rate_limited_projection_carries_only_the_active_cooldown_deadline() {
     let active_until = now + Duration::from_secs(60);
     let account = account("acct_rate_limited");
 
-    let active = account.status_projection(now, Some(active_until));
+    let active = account.status_projection(now, Some(active_until.into()));
     assert_eq!(active.status, AccountStatus::RateLimited);
-    assert_eq!(active.rate_limited_until, Some(active_until));
+    assert_eq!(active.cooldown, Some(active_until.into()));
 
-    let elapsed = account.status_projection(now, Some(now - Duration::from_secs(1)));
+    let elapsed = account.status_projection(now, Some((now - Duration::from_secs(1)).into()));
     assert_eq!(elapsed.status, AccountStatus::Normal);
-    assert_eq!(elapsed.rate_limited_until, None);
+    assert_eq!(elapsed.cooldown, None);
 }
 
 #[test]
@@ -492,8 +492,8 @@ fn highest_priority_affinity_should_still_obey_existing_scheduling_constraints()
                 );
             }
             AccountSchedulingBlocker::LocalAvailability => {
-                candidates[0].signals.rate_limited_until =
-                    Some(selection.now + Duration::from_secs(60))
+                candidates[0].signals.cooldown =
+                    Some((selection.now + Duration::from_secs(60)).into())
             }
             _ => unreachable!("test cases"),
         }
@@ -643,7 +643,7 @@ fn provider_quota_overlay_should_preserve_store_concurrency_facts() {
         last_started_at: Some(last_started_at),
         quota_reset_at: None,
         quota_remaining_rank: None,
-        rate_limited_until: None,
+        cooldown: None,
         failure_rate_basis_points: None,
         first_output_latency_ms: None,
     }
@@ -1062,3 +1062,32 @@ fn elapsed_quota_reset_does_not_fabricate_recovery() {
         .expect("healthy candidate available");
     assert_eq!(selected.candidate().account.id().as_str(), "acct_healthy");
 }
+
+#[test]
+fn due_probe_freeze_blocks_scheduling_and_remains_visible_as_rate_limited() {
+    use gateway_core::account::{AccountCooldown, AccountCooldownKind};
+    let mut candidates = [
+        weighted_candidate("acct_frozen", 100, 0),
+        weighted_candidate("acct_ready", 1, 0),
+    ];
+    let selection = context(RotationStrategy::Smart);
+    candidates[0].signals.cooldown = Some(AccountCooldown {
+        until: selection.now - Duration::from_secs(120),
+        kind: AccountCooldownKind::CapacityFreezeProbe,
+    });
+    let projection = candidates[0]
+        .account
+        .status_projection(selection.now, candidates[0].signals.cooldown);
+    assert_eq!(projection.status, AccountStatus::RateLimited);
+    assert!(
+        projection
+            .cooldown
+            .expect("freeze details")
+            .kind
+            .requires_probe()
+    );
+    let selected = AccountSelector
+        .select(&candidates, &selection)
+        .expect("available account");
+    assert_eq!(selected.candidate().account.id().as_str(), "acct_ready");
+}
diff --git a/backend/crates/gateway-core/tests/provider_ports/mod.rs b/backend/crates/gateway-core/tests/provider_ports/mod.rs
index 30950a01b..511cf357d 100644
--- a/backend/crates/gateway-core/tests/provider_ports/mod.rs
+++ b/backend/crates/gateway-core/tests/provider_ports/mod.rs
@@ -92,7 +92,7 @@ fn scheduling_state_preserves_provider_neutral_signals() {
             last_started_at: None,
             quota_reset_at: None,
             quota_remaining_rank: Some(7),
-            rate_limited_until: None,
+            cooldown: None,
             failure_rate_basis_points: Some(125),
             first_output_latency_ms: Some(250),
         },
diff --git a/backend/crates/gateway-host/tests/workers/mod.rs b/backend/crates/gateway-host/tests/workers/mod.rs
index 1a9f87ddd..5af943534 100644
--- a/backend/crates/gateway-host/tests/workers/mod.rs
+++ b/backend/crates/gateway-host/tests/workers/mod.rs
@@ -17,7 +17,7 @@ use gateway_core::task::{
 use gateway_host::workers::{WorkerStartError, WorkerSupervisor};
 use tokio::sync::Notify;
 
-const ACTIVE_KINDS: [WorkerKind; 8] = [
+const ACTIVE_KINDS: [WorkerKind; 9] = [
     WorkerKind::OAuthRefresh,
     WorkerKind::QuotaCatalogHealth,
     WorkerKind::RuntimeSnapshotReconciliation,
@@ -26,6 +26,7 @@ const ACTIVE_KINDS: [WorkerKind; 8] = [
     WorkerKind::Retention,
     WorkerKind::Backup,
     WorkerKind::AccountImport,
+    WorkerKind::AccountFreezeRecovery,
 ];
 
 #[derive(Clone)]
diff --git a/backend/crates/gateway-store/src/admin_adapter.rs b/backend/crates/gateway-store/src/admin_adapter.rs
index f08a3000c..82a0f6de6 100644
--- a/backend/crates/gateway-store/src/admin_adapter.rs
+++ b/backend/crates/gateway-store/src/admin_adapter.rs
@@ -54,6 +54,14 @@ impl SettingsStore for AdminSettingsStoreAdapter {
                 usage_retention_days: command.usage_retention_days,
                 ops_event_retention_days: command.ops_event_retention_days,
                 audit_retention_days: command.audit_retention_days,
+                account_auto_freeze_enabled: command.account_auto_freeze_enabled,
+                account_auto_freeze_threshold: command.account_auto_freeze_threshold,
+                account_auto_freeze_window_seconds: command.account_auto_freeze_window_seconds,
+                account_auto_freeze_duration_seconds: command.account_auto_freeze_duration_seconds,
+                account_auto_freeze_probe_enabled: command.account_auto_freeze_probe_enabled,
+                account_auto_freeze_probe_model: command.account_auto_freeze_probe_model,
+                account_auto_freeze_adaptive_concurrency: command
+                    .account_auto_freeze_adaptive_concurrency,
             },
             audit: mutation_audit(
                 context,
@@ -73,6 +81,7 @@ impl SettingsStore for AdminSettingsStoreAdapter {
                     "min_codex_desktop_version".to_owned(),
                     "min_codex_cli_version".to_owned(),
                     "retention".to_owned(),
+                    "account_auto_freeze".to_owned(),
                 ],
             ),
         };
@@ -181,6 +190,13 @@ pub(crate) fn admin_runtime_settings(
         usage_retention_days: settings.usage_retention_days,
         ops_event_retention_days: settings.ops_event_retention_days,
         audit_retention_days: settings.audit_retention_days,
+        account_auto_freeze_enabled: settings.account_auto_freeze_enabled,
+        account_auto_freeze_threshold: settings.account_auto_freeze_threshold,
+        account_auto_freeze_window_seconds: settings.account_auto_freeze_window_seconds,
+        account_auto_freeze_duration_seconds: settings.account_auto_freeze_duration_seconds,
+        account_auto_freeze_probe_enabled: settings.account_auto_freeze_probe_enabled,
+        account_auto_freeze_probe_model: settings.account_auto_freeze_probe_model,
+        account_auto_freeze_adaptive_concurrency: settings.account_auto_freeze_adaptive_concurrency,
         updated_at: settings.updated_at,
     })
 }
diff --git a/backend/crates/gateway-store/src/postgres/account_groups.rs b/backend/crates/gateway-store/src/postgres/account_groups.rs
index b27896aff..e31a3df4f 100644
--- a/backend/crates/gateway-store/src/postgres/account_groups.rs
+++ b/backend/crates/gateway-store/src/postgres/account_groups.rs
@@ -208,7 +208,7 @@ impl AccountGroupStore for PgAccountGroupRepository {
                         credential_state: account.credential_state,
                         access_token_expires_at: account.access_token_expires_at.map(Into::into),
                         quota: account.quota,
-                        rate_limited_until: None,
+                        cooldown: None,
                         last_error_reason: account.last_error_reason,
                         last_error_message: account.last_error_message,
                     },
diff --git a/backend/crates/gateway-store/src/postgres/observability/admin_adapter.rs b/backend/crates/gateway-store/src/postgres/observability/admin_adapter.rs
index fd532616c..21f7c1421 100644
--- a/backend/crates/gateway-store/src/postgres/observability/admin_adapter.rs
+++ b/backend/crates/gateway-store/src/postgres/observability/admin_adapter.rs
@@ -9,7 +9,7 @@ use super::*;
 
 use crate::postgres::{
     PgProviderAccountRepository, ProviderAccountRepository, ProviderAccountSummary,
-    account_status_projection, load_rate_limited_until,
+    account_status_projection, load_cooldown,
 };
 
 use crate::redis::{CredentialLeaseRepository as _, RedisCredentialLeaseRepository};
@@ -49,19 +49,15 @@ impl PgObservabilityRepository {
             )
             .await?;
         let now = observed_at.into();
-        let rate_limited_until =
-            load_rate_limited_until(self.cooldowns.as_deref(), &accounts, now).await;
+        let cooldown = load_cooldown(self.cooldowns.as_deref(), &accounts, now).await;
         let mut metrics = ProviderAccountMetrics {
             total: u64::try_from(accounts.len()).unwrap_or(u64::MAX),
             ..ProviderAccountMetrics::default()
         };
         let mut normal_accounts = Vec::new();
         for account in &accounts {
-            let projection = account_status_projection(
-                account,
-                now,
-                rate_limited_until.get(&account.id).copied(),
-            );
+            let projection =
+                account_status_projection(account, now, cooldown.get(&account.id).copied());
             match projection.status {
                 gateway_core::account::AccountStatus::Normal => {
                     metrics.normal = metrics.normal.saturating_add(1);
diff --git a/backend/crates/gateway-store/src/postgres/provider_accounts/admin_adapter.rs b/backend/crates/gateway-store/src/postgres/provider_accounts/admin_adapter.rs
index ce46b26e7..c20fdd77c 100644
--- a/backend/crates/gateway-store/src/postgres/provider_accounts/admin_adapter.rs
+++ b/backend/crates/gateway-store/src/postgres/provider_accounts/admin_adapter.rs
@@ -407,15 +407,11 @@ impl AccountStore for PgAdminAccountStore {
             ));
         }
         let now = Utc::now();
-        let rate_limited_until = runtime
-            .rate_limited_until
-            .into_iter()
-            .map(|(account_id, until)| (account_id, until.into()))
-            .collect::<BTreeMap<_, _>>();
+        let cooldown = runtime.cooldown;
         let now_system_time = std::time::SystemTime::from(now);
-        let active_rate_limited_ids = rate_limited_until
+        let active_rate_limited_ids = cooldown
             .iter()
-            .filter(|(_, until)| **until > now_system_time)
+            .filter(|(_, cooldown)| cooldown.is_active(now_system_time))
             .map(|(account_id, _)| account_id.clone())
             .collect::<Vec<_>>();
         let page =
@@ -434,7 +430,7 @@ impl AccountStore for PgAdminAccountStore {
                 let projection = account_status_projection(
                     &summary,
                     now.into(),
-                    rate_limited_until.get(&account_id).copied(),
+                    cooldown.get(&account_id).copied(),
                 );
                 let mut account = admin_account_record(summary)?;
                 account.groups = groups_by_account.remove(&account_id).unwrap_or_default();
@@ -466,12 +462,8 @@ impl AccountStore for PgAdminAccountStore {
             return Ok(None);
         };
         let now = Utc::now();
-        let rate_limited_until = runtime
-            .rate_limited_until
-            .get(account_id)
-            .copied()
-            .map(Into::into);
-        let projection = account_status_projection(&record.summary, now.into(), rate_limited_until);
+        let cooldown = runtime.cooldown.get(account_id).copied();
+        let projection = account_status_projection(&record.summary, now.into(), cooldown);
         let account_id = record.summary.id.clone();
         let mut groups = self
             .account_groups_by_account(std::slice::from_ref(&account_id))
@@ -745,6 +737,69 @@ impl AccountStore for PgAdminAccountStore {
         })
     }
 
+    async fn lower_concurrency_limit(
+        &self,
+        account_id: &CoreProviderAccountId,
+        limit: AccountConcurrencyLimit,
+        context: &MutationContext,
+    ) -> AdminStoreResult<Option<AccountUpdateResult>> {
+        let mut transaction = self.pool.begin().await.map_err(|_| {
+            admin_store_error(ENTITY, postgres_unavailable("begin concurrency reduction"))
+        })?;
+        let result =
+            async {
+                // 与管理写入采用相同锁顺序；锁住默认值后再检查账号最新设置，避免把旧快照写回。
+                let default_limit: i64 = sqlx::query_scalar(
+                "select max_concurrent_per_account from runtime_settings where id = 1 for update"
+            ).fetch_one(&mut *transaction).await
+                .map_err(|_| postgres_unavailable("lock default concurrency"))?;
+                let changed = sqlx::query_scalar::<_, String>(
+                    "update provider_accounts set concurrency_limit = $2, updated_at = now()
+                 where id = $1 and enabled = true and coalesce(concurrency_limit, $3) > $2
+                 returning id",
+                )
+                .bind(account_id.as_str())
+                .bind(i64::from(limit.get()))
+                .bind(default_limit)
+                .fetch_optional(&mut *transaction)
+                .await
+                .map_err(|_| postgres_unavailable("lower account concurrency"))?;
+                if changed.is_none() {
+                    return Ok(None);
+                }
+                let revision = bump_config_revision_in_transaction(&mut transaction).await?;
+                append_admin_audit_event_in_transaction(
+                    &mut transaction,
+                    mutation_audit(
+                        context,
+                        "adapt_concurrency",
+                        "provider_account",
+                        account_id.as_str(),
+                        vec!["concurrency_limit".to_owned()],
+                    ),
+                    revision,
+                )
+                .await?;
+                Ok(Some(revision))
+            }
+            .await;
+        let revision = super::repository::finish_admin_transaction(
+            transaction,
+            result,
+            "lower account concurrency",
+        )
+        .await
+        .map_err(|error| admin_store_error(ENTITY, error))?;
+        revision
+            .map(|revision| {
+                Ok(AccountUpdateResult {
+                    config_revision: admin_revision(revision)?,
+                    account_id: account_id.clone(),
+                })
+            })
+            .transpose()
+    }
+
     async fn recover_account(
         &self,
         account_id: &CoreProviderAccountId,
diff --git a/backend/crates/gateway-store/src/postgres/provider_accounts/runtime.rs b/backend/crates/gateway-store/src/postgres/provider_accounts/runtime.rs
index c328a246b..318b64f3a 100644
--- a/backend/crates/gateway-store/src/postgres/provider_accounts/runtime.rs
+++ b/backend/crates/gateway-store/src/postgres/provider_accounts/runtime.rs
@@ -16,7 +16,7 @@ use super::ProviderAccountSummary;
 pub(crate) fn account_status_projection(
     account: &ProviderAccountSummary,
     now: SystemTime,
-    rate_limited_until: Option<SystemTime>,
+    cooldown: Option<gateway_core::account::AccountCooldown>,
 ) -> AccountStatusProjection {
     resolve_account_status(
         &AccountStatusFacts {
@@ -24,7 +24,7 @@ pub(crate) fn account_status_projection(
             credential_state: account.credential_state,
             access_token_expires_at: account.access_token_expires_at.map(Into::into),
             quota: account.quota,
-            rate_limited_until,
+            cooldown,
             last_error_reason: account.last_error_reason,
             last_error_message: account.last_error_message.clone(),
         },
@@ -36,18 +36,21 @@ pub(crate) fn account_status_projection(
 ///
 /// 冷却属于可丢失的账号级运行时事实：端口不可用、单条读取失败或脏账号 ID
 /// 都视为没有冷却，不污染持久账号状态。凭据轮换不会解除上游对同一账号的限流。
-pub(crate) async fn load_rate_limited_until(
+pub(crate) async fn load_cooldown(
     cooldowns: Option<&dyn ProviderCooldownPort>,
     accounts: &[ProviderAccountSummary],
     now: SystemTime,
-) -> BTreeMap<String, SystemTime> {
+) -> BTreeMap<String, gateway_core::account::AccountCooldown> {
     let Some(cooldowns) = cooldowns else {
         return BTreeMap::new();
     };
     let reads = accounts.iter().map(|account| async move {
         let account_id = ProviderAccountId::new(account.id.clone()).ok()?;
         let cooldown = cooldowns.read(&account_id).await.ok().flatten()?;
-        (cooldown.until() > now).then(|| (account.id.clone(), cooldown.until()))
+        cooldown
+            .scheduling_state()
+            .is_active(now)
+            .then(|| (account.id.clone(), cooldown.scheduling_state()))
     });
     join_all(reads).await.into_iter().flatten().collect()
 }
diff --git a/backend/crates/gateway-store/src/postgres/runtime_settings.rs b/backend/crates/gateway-store/src/postgres/runtime_settings.rs
index ddb821652..df0f9a40a 100644
--- a/backend/crates/gateway-store/src/postgres/runtime_settings.rs
+++ b/backend/crates/gateway-store/src/postgres/runtime_settings.rs
@@ -11,7 +11,8 @@ use sqlx::{PgPool, Postgres, Transaction};
 use gateway_core::account::RotationStrategy;
 use gateway_core::policy::CodexClientVersion;
 use gateway_core::provider_ports::{
-    ProviderRefreshPolicy, ProviderRuntimePolicyPort, ProviderStoreError, ProviderStoreErrorKind,
+    ProviderFreezePolicy, ProviderRefreshPolicy, ProviderRuntimePolicyPort, ProviderStoreError,
+    ProviderStoreErrorKind,
 };
 
 use crate::{Revision, StoreError, StoreResult, postgres_unavailable};
@@ -34,6 +35,13 @@ pub struct RuntimeSettings {
     pub usage_retention_days: u32,
     pub ops_event_retention_days: u32,
     pub audit_retention_days: u32,
+    pub account_auto_freeze_enabled: bool,
+    pub account_auto_freeze_threshold: u32,
+    pub account_auto_freeze_window_seconds: u64,
+    pub account_auto_freeze_duration_seconds: u64,
+    pub account_auto_freeze_probe_enabled: bool,
+    pub account_auto_freeze_probe_model: Option<String>,
+    pub account_auto_freeze_adaptive_concurrency: bool,
     pub updated_at: DateTime<Utc>,
 }
 
@@ -60,6 +68,34 @@ impl fmt::Debug for RuntimeSettings {
             .field("usage_retention_days", &self.usage_retention_days)
             .field("ops_event_retention_days", &self.ops_event_retention_days)
             .field("audit_retention_days", &self.audit_retention_days)
+            .field(
+                "account_auto_freeze_enabled",
+                &self.account_auto_freeze_enabled,
+            )
+            .field(
+                "account_auto_freeze_threshold",
+                &self.account_auto_freeze_threshold,
+            )
+            .field(
+                "account_auto_freeze_window_seconds",
+                &self.account_auto_freeze_window_seconds,
+            )
+            .field(
+                "account_auto_freeze_duration_seconds",
+                &self.account_auto_freeze_duration_seconds,
+            )
+            .field(
+                "account_auto_freeze_probe_enabled",
+                &self.account_auto_freeze_probe_enabled,
+            )
+            .field(
+                "account_auto_freeze_probe_model",
+                &self.account_auto_freeze_probe_model,
+            )
+            .field(
+                "account_auto_freeze_adaptive_concurrency",
+                &self.account_auto_freeze_adaptive_concurrency,
+            )
             .field("updated_at", &self.updated_at)
             .finish()
     }
@@ -82,6 +118,13 @@ pub struct RuntimeSettingsUpdate {
     pub usage_retention_days: u32,
     pub ops_event_retention_days: u32,
     pub audit_retention_days: u32,
+    pub account_auto_freeze_enabled: bool,
+    pub account_auto_freeze_threshold: u32,
+    pub account_auto_freeze_window_seconds: u64,
+    pub account_auto_freeze_duration_seconds: u64,
+    pub account_auto_freeze_probe_enabled: bool,
+    pub account_auto_freeze_probe_model: Option<String>,
+    pub account_auto_freeze_adaptive_concurrency: bool,
 }
 
 impl fmt::Debug for RuntimeSettingsUpdate {
@@ -109,9 +152,13 @@ impl RuntimeSettingsUpdate {
             || self.usage_retention_days < 31
             || self.ops_event_retention_days == 0
             || self.audit_retention_days == 0
+            || !(2..=1_000).contains(&self.account_auto_freeze_threshold)
+            || !(60..=3_600).contains(&self.account_auto_freeze_window_seconds)
+            || !(300..=604_800).contains(&self.account_auto_freeze_duration_seconds)
             || !valid_model_mappings(&self.model_mappings)
             || !valid_client_version(self.min_codex_desktop_version.as_deref())
             || !valid_client_version(self.min_codex_cli_version.as_deref())
+            || !valid_probe_model(self.account_auto_freeze_probe_model.as_deref())
             || RotationStrategy::parse(&self.rotation_strategy).is_none()
         {
             return Err(StoreError::InvalidData {
@@ -173,7 +220,11 @@ pub(crate) async fn load_runtime_settings_from_pool(pool: &PgPool) -> StoreResul
                     refresh_concurrency, max_concurrent_per_account, request_interval_ms,
                     rotation_strategy, model_mappings_json, usage_retention_days, ops_event_retention_days,
                     audit_retention_days, min_codex_desktop_version,
-                    min_codex_cli_version, updated_at, max_waiting_per_key, max_waiting_per_account, concurrency_wait_timeout_seconds
+                    min_codex_cli_version, updated_at, max_waiting_per_key, max_waiting_per_account, concurrency_wait_timeout_seconds,
+                    account_auto_freeze_enabled, account_auto_freeze_threshold,
+                    account_auto_freeze_window_seconds, account_auto_freeze_duration_seconds,
+                    account_auto_freeze_probe_enabled, account_auto_freeze_probe_model,
+                    account_auto_freeze_adaptive_concurrency
              from runtime_settings where id = 1",
         )
     .fetch_optional(pool)
@@ -202,6 +253,25 @@ impl ProviderRuntimePolicyPort for PgRuntimeSettingsRepository {
             )
         })
     }
+
+    fn load_freeze_policy(
+        &self,
+    ) -> futures::future::BoxFuture<'_, Result<ProviderFreezePolicy, ProviderStoreError>> {
+        Box::pin(async move {
+            let settings = RuntimeSettingsRepository::load_runtime_settings(self)
+                .await
+                .map_err(|_| provider_unavailable("load freeze policy"))?;
+            ProviderFreezePolicy::try_new(
+                settings.account_auto_freeze_enabled,
+                settings.account_auto_freeze_threshold,
+                settings.account_auto_freeze_window_seconds,
+                settings.account_auto_freeze_duration_seconds,
+                settings.account_auto_freeze_probe_enabled,
+                settings.account_auto_freeze_probe_model,
+                settings.account_auto_freeze_adaptive_concurrency,
+            )
+        })
+    }
 }
 
 pub(crate) async fn load_runtime_settings_in_transaction(
@@ -212,7 +282,11 @@ pub(crate) async fn load_runtime_settings_in_transaction(
                 refresh_concurrency, max_concurrent_per_account, request_interval_ms,
                 rotation_strategy, model_mappings_json, usage_retention_days, ops_event_retention_days,
                 audit_retention_days, min_codex_desktop_version,
-                min_codex_cli_version, updated_at, max_waiting_per_key, max_waiting_per_account, concurrency_wait_timeout_seconds
+                min_codex_cli_version, updated_at, max_waiting_per_key, max_waiting_per_account, concurrency_wait_timeout_seconds,
+                account_auto_freeze_enabled, account_auto_freeze_threshold,
+                account_auto_freeze_window_seconds, account_auto_freeze_duration_seconds,
+                account_auto_freeze_probe_enabled, account_auto_freeze_probe_model,
+                account_auto_freeze_adaptive_concurrency
          from runtime_settings where id = 1",
     )
     .fetch_optional(&mut **transaction)
@@ -250,6 +324,13 @@ pub(crate) async fn update_runtime_settings_in_transaction(
                      max_waiting_per_key = $13,
                      max_waiting_per_account = $14,
                      concurrency_wait_timeout_seconds = $15,
+                     account_auto_freeze_enabled = $16,
+                     account_auto_freeze_threshold = $17,
+                     account_auto_freeze_window_seconds = $18,
+                     account_auto_freeze_duration_seconds = $19,
+                     account_auto_freeze_probe_enabled = $20,
+                     account_auto_freeze_probe_model = $21,
+                     account_auto_freeze_adaptive_concurrency = $22,
 	                 updated_at = now()
 	             where id = 1
 	             returning config_revision",
@@ -269,6 +350,16 @@ pub(crate) async fn update_runtime_settings_in_transaction(
     .bind(i64::from(update.max_waiting_per_key))
     .bind(i64::from(update.max_waiting_per_account))
     .bind(i64::from(update.concurrency_wait_timeout_seconds))
+    .bind(update.account_auto_freeze_enabled)
+    .bind(i64::from(update.account_auto_freeze_threshold))
+    .bind(i64::try_from(update.account_auto_freeze_window_seconds).map_err(|_| invalid_numeric())?)
+    .bind(
+        i64::try_from(update.account_auto_freeze_duration_seconds)
+            .map_err(|_| invalid_numeric())?,
+    )
+    .bind(update.account_auto_freeze_probe_enabled)
+    .bind(update.account_auto_freeze_probe_model.as_deref())
+    .bind(update.account_auto_freeze_adaptive_concurrency)
     .fetch_optional(&mut **transaction)
     .await
     .map_err(|_| postgres_unavailable("update runtime settings in transaction"))?
@@ -335,6 +426,13 @@ struct RuntimeSettingsRow {
     max_waiting_per_key: i64,
     max_waiting_per_account: i64,
     concurrency_wait_timeout_seconds: i64,
+    account_auto_freeze_enabled: bool,
+    account_auto_freeze_threshold: i64,
+    account_auto_freeze_window_seconds: i64,
+    account_auto_freeze_duration_seconds: i64,
+    account_auto_freeze_probe_enabled: bool,
+    account_auto_freeze_probe_model: Option<String>,
+    account_auto_freeze_adaptive_concurrency: bool,
 }
 
 fn runtime_settings_from_row(row: RuntimeSettingsRow) -> StoreResult<RuntimeSettings> {
@@ -356,6 +454,13 @@ fn runtime_settings_from_row(row: RuntimeSettingsRow) -> StoreResult<RuntimeSett
         max_waiting_per_key: to_u32(row.max_waiting_per_key)?,
         max_waiting_per_account: to_u32(row.max_waiting_per_account)?,
         concurrency_wait_timeout_seconds: to_u32(row.concurrency_wait_timeout_seconds)?,
+        account_auto_freeze_enabled: row.account_auto_freeze_enabled,
+        account_auto_freeze_threshold: to_u32(row.account_auto_freeze_threshold)?,
+        account_auto_freeze_window_seconds: to_u64(row.account_auto_freeze_window_seconds)?,
+        account_auto_freeze_duration_seconds: to_u64(row.account_auto_freeze_duration_seconds)?,
+        account_auto_freeze_probe_enabled: row.account_auto_freeze_probe_enabled,
+        account_auto_freeze_probe_model: row.account_auto_freeze_probe_model,
+        account_auto_freeze_adaptive_concurrency: row.account_auto_freeze_adaptive_concurrency,
     })
 }
 
@@ -398,3 +503,12 @@ fn valid_model_name(value: &str, max_len: usize) -> bool {
 fn valid_client_version(value: Option<&str>) -> bool {
     value.is_none_or(|value| CodexClientVersion::parse(value).is_ok())
 }
+
+fn valid_probe_model(value: Option<&str>) -> bool {
+    value.is_none_or(|value| {
+        !value.is_empty()
+            && value.len() <= 128
+            && value == value.trim()
+            && !value.bytes().any(|byte| byte.is_ascii_control())
+    })
+}
diff --git a/backend/crates/gateway-store/src/redis/admin_account_runtime.rs b/backend/crates/gateway-store/src/redis/admin_account_runtime.rs
index 1d236259f..5014698db 100644
--- a/backend/crates/gateway-store/src/redis/admin_account_runtime.rs
+++ b/backend/crates/gateway-store/src/redis/admin_account_runtime.rs
@@ -3,6 +3,7 @@
 use std::collections::BTreeMap;
 
 use async_trait::async_trait;
+use chrono::{DateTime, Utc};
 use futures::future::join_all;
 use gateway_admin::{
     model::accounts::AccountRuntimeSnapshot,
@@ -49,15 +50,23 @@ impl AccountRuntimeStore for RedisAdminAccountRuntimeStore {
                 .read_credential_cooldown(account_id)
                 .await
                 .map(|cooldown| {
-                    cooldown.map(|cooldown| (account_id.clone(), cooldown.cooldown_until))
+                    cooldown.map(|cooldown| {
+                        (
+                            account_id.clone(),
+                            gateway_core::account::AccountCooldown {
+                                until: cooldown.cooldown_until.into(),
+                                kind: cooldown.kind,
+                            },
+                        )
+                    })
                 })
         });
-        let mut rate_limited_until = BTreeMap::new();
+        let mut cooldown = BTreeMap::new();
         for result in join_all(reads).await {
             if let Some((account_id, until)) =
                 result.map_err(|error| crate::admin_store_error("account runtime", error))?
             {
-                rate_limited_until.insert(account_id, until);
+                cooldown.insert(account_id, until);
             }
         }
         let in_flight = self
@@ -72,8 +81,50 @@ impl AccountRuntimeStore for RedisAdminAccountRuntimeStore {
                     .collect()
             });
         Ok(AccountRuntimeSnapshot {
-            rate_limited_until,
+            cooldown,
             in_flight,
         })
     }
+
+    async fn active_freezes(
+        &self,
+    ) -> AdminStoreResult<BTreeMap<String, gateway_admin::model::accounts::AccountFreeze>> {
+        self.cooldowns
+            .active_freezes()
+            .await
+            .map_err(|error| crate::admin_store_error("account runtime", error))
+    }
+
+    async fn capacity_peaks(
+        &self,
+        account_ids: &[String],
+    ) -> AdminStoreResult<BTreeMap<String, u32>> {
+        let reads = account_ids.iter().map(|account_id| async move {
+            self.cooldowns
+                .read_capacity_peak(account_id)
+                .await
+                .map(|peak| peak.map(|value| (account_id.clone(), value)))
+        });
+        let mut peaks = BTreeMap::new();
+        for result in join_all(reads).await {
+            if let Some((account_id, peak)) =
+                result.map_err(|error| crate::admin_store_error("account runtime", error))?
+            {
+                peaks.insert(account_id, peak);
+            }
+        }
+        Ok(peaks)
+    }
+
+    async fn finish_freeze(
+        &self,
+        account_id: &str,
+        expected: &gateway_admin::model::accounts::AccountFreeze,
+        postpone_until: Option<DateTime<Utc>>,
+    ) -> AdminStoreResult<bool> {
+        self.cooldowns
+            .finish_freeze(account_id, expected, postpone_until)
+            .await
+            .map_err(|error| crate::admin_store_error("account runtime", error))
+    }
 }
diff --git a/backend/crates/gateway-store/src/redis/credential_cooldown.rs b/backend/crates/gateway-store/src/redis/credential_cooldown.rs
index 78e80ab43..354d0a307 100644
--- a/backend/crates/gateway-store/src/redis/credential_cooldown.rs
+++ b/backend/crates/gateway-store/src/redis/credential_cooldown.rs
@@ -6,11 +6,13 @@ use gateway_admin::model::accounts::AccountRuntimeSnapshot;
 use gateway_core::{
     account::{CredentialRevision, ProviderAccountId},
     provider_ports::{
-        ProviderCooldown, ProviderCooldownPort, ProviderCooldownScope, ProviderScopedCooldown,
-        ProviderStoreError, ProviderStoreErrorKind,
+        ProviderCooldown, ProviderCooldownKind, ProviderCooldownPort, ProviderCooldownScope,
+        ProviderScopedCooldown, ProviderStoreError, ProviderStoreErrorKind,
     },
 };
 use redis::{Script, aio::ConnectionManager};
+use std::collections::BTreeMap;
+use std::time::Duration;
 
 use crate::{Revision, StoreError, StoreResult, redis_unavailable, require_nonempty};
 
@@ -22,17 +24,21 @@ local incoming = tonumber(ARGV[1])
 local incoming_until = tonumber(ARGV[2])
 if current > incoming then return 0 end
 local current_until = tonumber(redis.call('HGET', KEYS[1], 'until_ms') or '0')
-if current == incoming and current_until >= incoming_until then return 0 end
+local current_kind = redis.call('HGET', KEYS[1], 'kind') or 'rate_limit'
+if current == incoming then
+  if current_kind ~= 'rate_limit' and ARGV[4] == 'rate_limit' then return 0 end
+  if current_kind == ARGV[4] and current_until >= incoming_until then return 0 end
+  incoming_until = math.max(current_until, incoming_until)
+end
 local clock = redis.call('TIME')
 local now_ms = (tonumber(clock[1]) * 1000) + math.floor(tonumber(clock[2]) / 1000)
-if incoming_until <= now_ms then
-  redis.call('DEL', KEYS[1])
-  if #KEYS > 1 then redis.call('ZREM', KEYS[2], ARGV[3]) end
-  return 0
+if incoming_until <= now_ms and ARGV[4] ~= 'capacity_freeze_probe' then return 0 end
+redis.call('HSET', KEYS[1], 'revision', ARGV[1], 'until_ms', incoming_until, 'kind', ARGV[4], 'generation', ARGV[5])
+if ARGV[4] == 'capacity_freeze_probe' then
+  redis.call('PERSIST', KEYS[1])
+else
+  redis.call('PEXPIRE', KEYS[1], incoming_until - now_ms + 60000)
 end
-redis.call('HSET', KEYS[1], 'revision', ARGV[1], 'until_ms', ARGV[2])
-local ttl = incoming_until - now_ms + 60000
-redis.call('PEXPIRE', KEYS[1], ttl)
 if #KEYS > 1 then redis.call('ZADD', KEYS[2], incoming_until, ARGV[3]) end
 return 1
 "#;
@@ -40,19 +46,15 @@ return 1
 const READ_SCRIPT: &str = r#"
 local revision = redis.call('HGET', KEYS[1], 'revision')
 local until_ms = redis.call('HGET', KEYS[1], 'until_ms')
-if revision == false or until_ms == false then
-  redis.call('DEL', KEYS[1])
-  if #KEYS > 1 then redis.call('ZREM', KEYS[2], ARGV[1]) end
-  return {0, '0', '0'}
-end
+local kind = redis.call('HGET', KEYS[1], 'kind') or 'rate_limit'
 local clock = redis.call('TIME')
 local now_ms = (tonumber(clock[1]) * 1000) + math.floor(tonumber(clock[2]) / 1000)
-if tonumber(until_ms) <= now_ms then
+if revision == false or until_ms == false or (kind ~= 'capacity_freeze_probe' and tonumber(until_ms) <= now_ms) then
   redis.call('DEL', KEYS[1])
   if #KEYS > 1 then redis.call('ZREM', KEYS[2], ARGV[1]) end
-  return {0, '0', '0'}
+  return {0, '0', '0', 'rate_limit', ''}
 end
-return {1, revision, until_ms}
+return {1, revision, until_ms, kind, redis.call('HGET', KEYS[1], 'generation') or ''}
 "#;
 
 const INVALIDATE_SCRIPT: &str = r#"
@@ -63,11 +65,56 @@ if #KEYS > 1 then redis.call('ZREM', KEYS[2], ARGV[2]) end
 return 1
 "#;
 
-const ACTIVE_COOLDOWNS_SCRIPT: &str = r#"
-local clock = redis.call('TIME')
-local now_ms = (tonumber(clock[1]) * 1000) + math.floor(tonumber(clock[2]) / 1000)
-redis.call('ZREMRANGEBYSCORE', KEYS[1], '-inf', now_ms)
-return redis.call('ZRANGEBYSCORE', KEYS[1], '(' .. now_ms, '+inf', 'WITHSCORES')
+// 普通推理成功只能清除临时限流和未形成冻结的证据，判断与删除必须原子执行。
+const SUCCESS_SCRIPT: &str = r#"
+local current = tonumber(redis.call('HGET', KEYS[1], 'revision') or '0')
+local kind = redis.call('HGET', KEYS[1], 'kind') or 'rate_limit'
+if current > tonumber(ARGV[1]) or kind ~= 'rate_limit' then return 0 end
+redis.call('DEL', KEYS[1], KEYS[3], KEYS[4])
+redis.call('ZREM', KEYS[2], ARGV[2])
+return 1
+"#;
+
+// 探测结果只能修改读到的这一代冻结；删除后重建同 revision 的冻结也不匹配。
+const FINISH_FREEZE_SCRIPT: &str = r#"
+if redis.call('HGET', KEYS[1], 'generation') ~= ARGV[2]
+  or redis.call('HGET', KEYS[1], 'revision') ~= ARGV[1] then return 0 end
+local kind = redis.call('HGET', KEYS[1], 'kind')
+if kind ~= 'capacity_freeze' and kind ~= 'capacity_freeze_probe' then return 0 end
+if ARGV[4] == '' then
+  redis.call('DEL', KEYS[1], KEYS[3], KEYS[4])
+  redis.call('ZREM', KEYS[2], ARGV[3])
+else
+  local until_ms = math.max(tonumber(redis.call('HGET', KEYS[1], 'until_ms')), tonumber(ARGV[4]))
+  redis.call('HSET', KEYS[1], 'until_ms', until_ms, 'generation', ARGV[5])
+  redis.call('ZADD', KEYS[2], until_ms, ARGV[3])
+  if kind == 'capacity_freeze_probe' then
+    redis.call('PERSIST', KEYS[1])
+  else
+    local clock = redis.call('TIME')
+    local now_ms = (tonumber(clock[1]) * 1000) + math.floor(tonumber(clock[2]) / 1000)
+    redis.call('PEXPIRE', KEYS[1], math.max(1, until_ms - now_ms + 60000))
+  end
+end
+return 1
+"#;
+
+// 每次容量失败都顺延窗口 TTL（与刷新退避计数同语义），并把本次观测到的
+// 在途并发并入峰值证据；峰值与计数共享同一窗口生命周期。
+const RECORD_CAPACITY_FAILURE_SCRIPT: &str = r#"
+local count = redis.call('INCR', KEYS[1])
+local ttl_ms = tonumber(ARGV[1])
+redis.call('PEXPIRE', KEYS[1], ttl_ms)
+local in_flight = tonumber(ARGV[2])
+if in_flight > 0 then
+  local peak = tonumber(redis.call('GET', KEYS[2]) or '0')
+  if in_flight > peak then
+    redis.call('SET', KEYS[2], in_flight, 'PX', ttl_ms)
+  else
+    redis.call('PEXPIRE', KEYS[2], ttl_ms)
+  end
+end
+return count
 "#;
 
 #[derive(Debug, Clone, PartialEq, Eq)]
@@ -75,6 +122,7 @@ pub struct CredentialCooldown {
     pub provider_account_id: String,
     pub credential_revision: Revision,
     pub cooldown_until: DateTime<Utc>,
+    pub kind: ProviderCooldownKind,
 }
 
 #[async_trait]
@@ -116,6 +164,22 @@ impl RedisCredentialCooldownRepository {
         format!("{}:account:active-cooldowns", self.namespace)
     }
 
+    fn capacity_failures_key(&self, provider_account_id: &str) -> StoreResult<String> {
+        let fingerprint = resource_fingerprint("credential cooldown", provider_account_id)?;
+        Ok(format!(
+            "{}:account:{fingerprint}:capacity-failures",
+            self.namespace
+        ))
+    }
+
+    fn capacity_peak_key(&self, provider_account_id: &str) -> StoreResult<String> {
+        let fingerprint = resource_fingerprint("credential cooldown", provider_account_id)?;
+        Ok(format!(
+            "{}:account:{fingerprint}:capacity-peak-inflight",
+            self.namespace
+        ))
+    }
+
     fn scoped_key(
         &self,
         provider_account_id: &str,
@@ -135,6 +199,7 @@ impl RedisCredentialCooldownRepository {
         key: String,
         credential_revision: Revision,
         cooldown_until: DateTime<Utc>,
+        kind: ProviderCooldownKind,
         index_member: Option<&str>,
     ) -> StoreResult<bool> {
         let until_ms = cooldown_until.timestamp_millis();
@@ -149,6 +214,8 @@ impl RedisCredentialCooldownRepository {
                 .arg(credential_revision.get())
                 .arg(until_ms)
                 .arg(index_member)
+                .arg(kind.as_str())
+                .arg(uuid::Uuid::new_v4().to_string())
                 .invoke_async::<i64>(&mut connection)
                 .await
         } else {
@@ -157,6 +224,8 @@ impl RedisCredentialCooldownRepository {
                 .arg(credential_revision.get())
                 .arg(until_ms)
                 .arg("")
+                .arg(kind.as_str())
+                .arg(uuid::Uuid::new_v4().to_string())
                 .invoke_async::<i64>(&mut connection)
                 .await
         }
@@ -168,7 +237,7 @@ impl RedisCredentialCooldownRepository {
         &self,
         key: String,
         index_member: Option<&str>,
-    ) -> StoreResult<Option<(Revision, DateTime<Utc>)>> {
+    ) -> StoreResult<Option<(Revision, DateTime<Utc>, ProviderCooldownKind, String)>> {
         let mut connection = self.connection.clone();
         let result = if let Some(index_member) = index_member {
             Script::new(READ_SCRIPT)
@@ -184,7 +253,7 @@ impl RedisCredentialCooldownRepository {
                 .invoke_async(&mut connection)
                 .await
         };
-        let (present, revision, until_ms): (i64, String, String) =
+        let (present, revision, until_ms, kind, generation): (i64, String, String, String, String) =
             result.map_err(|_| redis_unavailable("read credential cooldown"))?;
         if present == 0 {
             return Ok(None);
@@ -197,7 +266,13 @@ impl RedisCredentialCooldownRepository {
             .map_err(|_| invalid("cached cooldown expiry is invalid"))?;
         let cooldown_until = DateTime::from_timestamp_millis(until_ms)
             .ok_or_else(|| invalid("cached cooldown expiry is invalid"))?;
-        Ok(Some((Revision::new(revision)?, cooldown_until)))
+        let kind = ProviderCooldownKind::parse(&kind).unwrap_or(ProviderCooldownKind::RateLimit);
+        Ok(Some((
+            Revision::new(revision)?,
+            cooldown_until,
+            kind,
+            generation,
+        )))
     }
 
     async fn invalidate_at_key(
@@ -227,30 +302,107 @@ impl RedisCredentialCooldownRepository {
         Ok(removed == 1)
     }
 
-    pub(crate) async fn active_cooldowns(&self) -> StoreResult<AccountRuntimeSnapshot> {
+    async fn indexed_accounts(&self) -> StoreResult<Vec<String>> {
         let mut connection = self.connection.clone();
-        let values = Script::new(ACTIVE_COOLDOWNS_SCRIPT)
-            .key(self.active_index_key())
-            .invoke_async::<Vec<String>>(&mut connection)
+        redis::cmd("ZRANGE")
+            .arg(self.active_index_key())
+            .arg(0)
+            .arg(-1)
+            .query_async(&mut connection)
             .await
-            .map_err(|_| redis_unavailable("list active credential cooldowns"))?;
-        if values.len() % 2 != 0 {
-            return Err(invalid("active cooldown index is invalid"));
-        }
-        let mut rate_limited_until = std::collections::BTreeMap::new();
-        for pair in values.chunks_exact(2) {
-            let until_ms = pair[1]
-                .parse::<i64>()
-                .map_err(|_| invalid("active cooldown expiry is invalid"))?;
-            let until = DateTime::from_timestamp_millis(until_ms)
-                .ok_or_else(|| invalid("active cooldown expiry is invalid"))?;
-            rate_limited_until.insert(pair[0].clone(), until);
+            .map_err(|_| redis_unavailable("list indexed cooldowns"))
+    }
+
+    pub(crate) async fn active_cooldowns(&self) -> StoreResult<AccountRuntimeSnapshot> {
+        let mut cooldown = BTreeMap::new();
+        for account_id in self.indexed_accounts().await? {
+            if let Some((_, until, kind, _)) = self
+                .read_at_key(self.key(&account_id)?, Some(&account_id))
+                .await?
+            {
+                cooldown.insert(
+                    account_id,
+                    gateway_core::account::AccountCooldown {
+                        until: until.into(),
+                        kind,
+                    },
+                );
+            }
         }
         Ok(AccountRuntimeSnapshot {
-            rate_limited_until,
+            cooldown,
             in_flight: None,
         })
     }
+
+    /// 包含已到探测时间但尚未确认恢复的冻结；到期不能从 worker 工作集中移除。
+    pub(crate) async fn active_freezes(
+        &self,
+    ) -> StoreResult<BTreeMap<String, gateway_admin::model::accounts::AccountFreeze>> {
+        let mut freezes = BTreeMap::new();
+        for account_id in self.indexed_accounts().await? {
+            if let Some((revision, until, kind, generation)) = self
+                .read_at_key(self.key(&account_id)?, Some(&account_id))
+                .await?
+                && kind.is_capacity_freeze()
+            {
+                freezes.insert(
+                    account_id,
+                    gateway_admin::model::accounts::AccountFreeze {
+                        credential_revision: gateway_admin::model::Revision::new(revision.get())
+                            .map_err(|_| invalid("freeze revision"))?,
+                        until,
+                        generation,
+                        requires_probe: kind.requires_probe(),
+                    },
+                );
+            }
+        }
+        Ok(freezes)
+    }
+
+    pub(crate) async fn finish_freeze(
+        &self,
+        account_id: &str,
+        expected: &gateway_admin::model::accounts::AccountFreeze,
+        postpone_until: Option<DateTime<Utc>>,
+    ) -> StoreResult<bool> {
+        let mut connection = self.connection.clone();
+        let changed: i64 = Script::new(FINISH_FREEZE_SCRIPT)
+            .key(self.key(account_id)?)
+            .key(self.active_index_key())
+            .key(self.capacity_failures_key(account_id)?)
+            .key(self.capacity_peak_key(account_id)?)
+            .arg(expected.credential_revision.get())
+            .arg(&expected.generation)
+            .arg(account_id)
+            .arg(
+                postpone_until
+                    .map(|until| until.timestamp_millis().to_string())
+                    .unwrap_or_default(),
+            )
+            .arg(uuid::Uuid::new_v4().to_string())
+            .invoke_async(&mut connection)
+            .await
+            .map_err(|_| redis_unavailable("finish capacity freeze"))?;
+        Ok(changed == 1)
+    }
+
+    /// 读取窗口内观测到的在途并发峰值；key 随窗口 TTL 过期，无需额外清理。
+    pub(crate) async fn read_capacity_peak(
+        &self,
+        provider_account_id: &str,
+    ) -> StoreResult<Option<u32>> {
+        let mut connection = self.connection.clone();
+        let peak: Option<i64> = redis::cmd("GET")
+            .arg(self.capacity_peak_key(provider_account_id)?)
+            .query_async(&mut connection)
+            .await
+            .map_err(|_| redis_unavailable("read capacity peak in-flight"))?;
+        peak.map(u32::try_from)
+            .transpose()
+            .map_err(|_| invalid("capacity peak in-flight is invalid"))
+    }
 }
 
 #[async_trait]
@@ -265,6 +417,7 @@ impl CredentialCooldownRepository for RedisCredentialCooldownRepository {
             self.key(&cooldown.provider_account_id)?,
             cooldown.credential_revision,
             cooldown.cooldown_until,
+            cooldown.kind,
             Some(&cooldown.provider_account_id),
         )
         .await
@@ -282,11 +435,14 @@ impl CredentialCooldownRepository for RedisCredentialCooldownRepository {
         self.read_at_key(self.key(provider_account_id)?, Some(provider_account_id))
             .await
             .map(|value| {
-                value.map(|(credential_revision, cooldown_until)| CredentialCooldown {
-                    provider_account_id: provider_account_id.to_owned(),
-                    credential_revision,
-                    cooldown_until,
-                })
+                value.map(
+                    |(credential_revision, cooldown_until, kind, _)| CredentialCooldown {
+                        provider_account_id: provider_account_id.to_owned(),
+                        credential_revision,
+                        cooldown_until,
+                        kind,
+                    },
+                )
             })
     }
 
@@ -317,8 +473,11 @@ impl CredentialCooldownRepository for RedisCredentialCooldownRepository {
         // 账号删除：清除该账号的 account key 与全部 model-scoped key。
         // 用 SCAN 精确匹配命名空间内该账号前缀，避免 KEYS 阻塞。
         let mut connection = self.connection.clone();
-        let account_key = self.key(provider_account_id)?;
-        let mut keys = vec![account_key.clone()];
+        let mut keys = vec![
+            self.key(provider_account_id)?,
+            self.capacity_failures_key(provider_account_id)?,
+            self.capacity_peak_key(provider_account_id)?,
+        ];
         let pattern = format!(
             "{}:account:{}:cooldown:*",
             self.namespace,
@@ -341,21 +500,21 @@ impl CredentialCooldownRepository for RedisCredentialCooldownRepository {
                 break;
             }
         }
-        if keys.is_empty() {
-            return Ok(false);
-        }
-        let removed: i64 = redis::cmd("DEL")
-            .arg(keys)
-            .query_async(&mut connection)
-            .await
-            .map_err(|_| redis_unavailable("delete account cooldown keys"))?;
-        let index_removed: i64 = redis::cmd("ZREM")
-            .arg(self.active_index_key())
-            .arg(provider_account_id)
-            .query_async(&mut connection)
-            .await
-            .map_err(|_| redis_unavailable("remove active account cooldown"))?;
-        Ok(removed > 0 || index_removed > 0)
+        // 删除与索引移除同属一个原子边界；否则新冻结可能在两步之间写入后丢失索引。
+        let removed: i64 = Script::new(
+            r#"
+            local removed = 0
+            for i = 2, #KEYS do removed = removed + redis.call('DEL', KEYS[i]) end
+            return removed + redis.call('ZREM', KEYS[1], ARGV[1])
+        "#,
+        )
+        .key(self.active_index_key())
+        .key(keys)
+        .arg(provider_account_id)
+        .invoke_async(&mut connection)
+        .await
+        .map_err(|_| redis_unavailable("delete account cooldowns and index"))?;
+        Ok(removed > 0)
     }
 }
 
@@ -370,6 +529,7 @@ impl ProviderCooldownPort for RedisCredentialCooldownRepository {
                 credential_revision: Revision::new(cooldown.credential_revision().get())
                     .map_err(|_| provider_invalid("encode credential cooldown"))?,
                 cooldown_until: cooldown.until().into(),
+                kind: cooldown.kind(),
             };
             CredentialCooldownRepository::cache_credential_cooldown(self, &record)
                 .await
@@ -390,10 +550,11 @@ impl ProviderCooldownPort for RedisCredentialCooldownRepository {
                         .map_err(|_| provider_invalid("decode credential cooldown"))?;
                     let revision = CredentialRevision::new(record.credential_revision.get())
                         .map_err(|_| provider_invalid("decode credential cooldown"))?;
-                    Ok(ProviderCooldown::new(
+                    Ok(ProviderCooldown::new_with_kind(
                         account_id,
                         revision,
                         record.cooldown_until.into(),
+                        record.kind,
                     ))
                 })
                 .transpose()
@@ -430,6 +591,7 @@ impl ProviderCooldownPort for RedisCredentialCooldownRepository {
                     .map_err(|_| provider_invalid("encode scoped credential cooldown"))?,
                 revision,
                 cooldown.until().into(),
+                ProviderCooldownKind::RateLimit,
                 None,
             )
             .await
@@ -451,7 +613,7 @@ impl ProviderCooldownPort for RedisCredentialCooldownRepository {
             )
             .await
             .map_err(|_| provider_unavailable("read scoped credential cooldown"))?
-            .map(|(revision, until)| {
+            .map(|(revision, until, _, _)| {
                 Ok(ProviderScopedCooldown::new(
                     account_id.clone(),
                     CredentialRevision::new(revision.get())
@@ -494,6 +656,75 @@ impl ProviderCooldownPort for RedisCredentialCooldownRepository {
                 .map_err(|_| provider_unavailable("clear all credential cooldowns"))
         })
     }
+
+    fn record_capacity_failure<'a>(
+        &'a self,
+        account_id: &'a ProviderAccountId,
+        window: Duration,
+        in_flight: u32,
+    ) -> futures::future::BoxFuture<'a, Result<u32, ProviderStoreError>> {
+        Box::pin(async move {
+            let mut connection = self.connection.clone();
+            let window_ms = u64::try_from(window.as_millis())
+                .map_err(|_| provider_invalid("encode capacity failure window"))?;
+            let count: i64 = Script::new(RECORD_CAPACITY_FAILURE_SCRIPT)
+                .key(
+                    self.capacity_failures_key(account_id.as_str())
+                        .map_err(|_| provider_invalid("encode capacity failure key"))?,
+                )
+                .key(
+                    self.capacity_peak_key(account_id.as_str())
+                        .map_err(|_| provider_invalid("encode capacity peak key"))?,
+                )
+                .arg(i64::try_from(window_ms).unwrap_or(i64::MAX))
+                .arg(i64::from(in_flight))
+                .invoke_async(&mut connection)
+                .await
+                .map_err(|_| provider_unavailable("record capacity failure"))?;
+            u32::try_from(count).map_err(|_| provider_invalid("decode capacity failure count"))
+        })
+    }
+
+    fn clear_after_success<'a>(
+        &'a self,
+        account_id: &'a ProviderAccountId,
+        through_revision: CredentialRevision,
+    ) -> futures::future::BoxFuture<'a, Result<(), ProviderStoreError>> {
+        Box::pin(async move {
+            let mut connection = self.connection.clone();
+            let _: i64 = Script::new(SUCCESS_SCRIPT)
+                .key(
+                    self.key(account_id.as_str())
+                        .map_err(|_| provider_invalid("cooldown key"))?,
+                )
+                .key(self.active_index_key())
+                .key(
+                    self.capacity_failures_key(account_id.as_str())
+                        .map_err(|_| provider_invalid("capacity count key"))?,
+                )
+                .key(
+                    self.capacity_peak_key(account_id.as_str())
+                        .map_err(|_| provider_invalid("capacity peak key"))?,
+                )
+                .arg(through_revision.get())
+                .arg(account_id.as_str())
+                .invoke_async(&mut connection)
+                .await
+                .map_err(|_| provider_unavailable("clear cooldown after success"))?;
+            Ok(())
+        })
+    }
+
+    fn capacity_peak_in_flight<'a>(
+        &'a self,
+        account_id: &'a ProviderAccountId,
+    ) -> futures::future::BoxFuture<'a, Result<Option<u32>, ProviderStoreError>> {
+        Box::pin(async move {
+            self.read_capacity_peak(account_id.as_str())
+                .await
+                .map_err(|_| provider_unavailable("read capacity peak in-flight"))
+        })
+    }
 }
 
 fn invalid(message: &str) -> StoreError {
diff --git a/backend/crates/gateway-store/src/redis/credential_leases.rs b/backend/crates/gateway-store/src/redis/credential_leases.rs
index da192142b..ed9fa8e0f 100644
--- a/backend/crates/gateway-store/src/redis/credential_leases.rs
+++ b/backend/crates/gateway-store/src/redis/credential_leases.rs
@@ -524,7 +524,7 @@ impl RedisProviderLeaseCoordinator {
                         last_started_at: signal.last_started_at.map(Into::into),
                         quota_reset_at: None,
                         quota_remaining_rank: None,
-                        rate_limited_until: None,
+                        cooldown: None,
                         failure_rate_basis_points: None,
                         first_output_latency_ms: None,
                     },
@@ -645,6 +645,36 @@ impl ProviderLeasePort for RedisProviderLeaseCoordinator {
             }
         })
     }
+
+    fn account_in_flight<'a>(
+        &'a self,
+        account_ids: &'a [ProviderAccountId],
+    ) -> futures::future::BoxFuture<
+        'a,
+        Result<std::collections::BTreeMap<ProviderAccountId, u32>, ProviderStoreError>,
+    > {
+        Box::pin(async move {
+            if account_ids.is_empty() {
+                return Ok(std::collections::BTreeMap::new());
+            }
+            let ids = account_ids
+                .iter()
+                .map(|account_id| account_id.as_str().to_owned())
+                .collect::<Vec<_>>();
+            let signals = self
+                .repository
+                .credential_runtime_signals(&ids)
+                .await
+                .map_err(|_| provider_unavailable("load account in-flight signals"))?;
+            Ok(signals
+                .into_iter()
+                .filter_map(|signal| {
+                    let account_id = ProviderAccountId::new(signal.resource_id).ok()?;
+                    Some((account_id, signal.in_flight))
+                })
+                .collect())
+        })
+    }
 }
 
 fn provider_unavailable(operation: &'static str) -> ProviderStoreError {
diff --git a/backend/crates/gateway-store/tests/postgres/observability/mod.rs b/backend/crates/gateway-store/tests/postgres/observability/mod.rs
index 53a1d6e6d..c0e9c158a 100644
--- a/backend/crates/gateway-store/tests/postgres/observability/mod.rs
+++ b/backend/crates/gateway-store/tests/postgres/observability/mod.rs
@@ -855,6 +855,30 @@ impl ProviderCooldownPort for StaticCooldowns {
     ) -> BoxFuture<'a, Result<bool, ProviderStoreError>> {
         Box::pin(async { Ok(false) })
     }
+
+    fn record_capacity_failure<'a>(
+        &'a self,
+        _account_id: &'a ProviderAccountId,
+        _window: Duration,
+        _in_flight: u32,
+    ) -> BoxFuture<'a, Result<u32, ProviderStoreError>> {
+        Box::pin(async { Ok(0) })
+    }
+
+    fn clear_after_success<'a>(
+        &'a self,
+        _account_id: &'a ProviderAccountId,
+        _through_revision: gateway_core::account::CredentialRevision,
+    ) -> BoxFuture<'a, Result<(), ProviderStoreError>> {
+        Box::pin(async { Ok(()) })
+    }
+
+    fn capacity_peak_in_flight<'a>(
+        &'a self,
+        _account_id: &'a ProviderAccountId,
+    ) -> BoxFuture<'a, Result<Option<u32>, ProviderStoreError>> {
+        Box::pin(async { Ok(None) })
+    }
 }
 
 fn test_cooldown(account_id: &str, revision: u64) -> ProviderCooldown {
diff --git a/backend/crates/gateway-store/tests/postgres/provider_accounts/mod.rs b/backend/crates/gateway-store/tests/postgres/provider_accounts/mod.rs
index 96b0e9e92..bfb3f12f0 100644
--- a/backend/crates/gateway-store/tests/postgres/provider_accounts/mod.rs
+++ b/backend/crates/gateway-store/tests/postgres/provider_accounts/mod.rs
@@ -852,9 +852,9 @@ async fn terminal_admin_list_filters_and_sorts_before_pagination_with_retained_u
                 sort: None,
             },
             AccountRuntimeSnapshot {
-                rate_limited_until: BTreeMap::from([(
+                cooldown: BTreeMap::from([(
                     "acct_alpha".to_owned(),
-                    now + TimeDelta::minutes(5),
+                    SystemTime::from(now + TimeDelta::minutes(5)).into(),
                 )]),
                 in_flight: None,
             },
@@ -3387,3 +3387,83 @@ async fn model_access_only_batch_update_preserves_other_settings_and_survives_re
     assert_eq!(loaded.summary.model_access, AccountModelAccess::all());
     database.close().await;
 }
+
+#[tokio::test]
+async fn adaptive_concurrency_uses_latest_locked_settings_without_overwriting_admin_fields() {
+    let Some(database) = TestDatabase::create("adaptive_concurrency").await else {
+        return;
+    };
+    let repository = PgProviderAccountRepository::new(database.pool.clone());
+    let store = admin_account_store(&database.pool);
+    let id = ProviderAccountId::new("acct_adaptive").expect("id");
+    repository
+        .insert_provider_account(account(id.as_str(), "adaptive-user"))
+        .await
+        .expect("account");
+    let group = "grp_aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa";
+    sqlx::query("insert into account_groups (id, name, color, created_at, updated_at) values ($1, 'kept group', '#112233FF', now(), now())")
+        .bind(group).execute(&database.pool).await.expect("group");
+    sqlx::query("insert into account_group_accounts (account_group_id, provider_account_id, created_at) values ($1, $2, now())")
+        .bind(group).bind(id.as_str()).execute(&database.pool).await.expect("membership");
+    for (enabled, account_limit, default_limit, expected_limit, changes) in [
+        (false, Some(8_i64), 10_i64, Some(8_i64), false),
+        (true, Some(2), 10, Some(2), false),
+        (true, None, 2, None, false),
+        (true, None, 10, Some(3), true),
+    ] {
+        let mut admin = database.pool.begin().await.expect("admin transaction");
+        sqlx::query("update runtime_settings set max_concurrent_per_account = $1 where id = 1")
+            .bind(default_limit)
+            .execute(&mut *admin)
+            .await
+            .expect("latest default");
+        let worker_store = store.clone();
+        let worker_id = id.clone();
+        let worker = tokio::spawn(async move {
+            worker_store
+                .lower_concurrency_limit(
+                    &worker_id,
+                    gateway_core::account::AccountConcurrencyLimit::new(3).expect("limit"),
+                    &MutationContext {
+                        actor: MutationActor::System,
+                        request_id: "adaptive-worker".to_owned(),
+                    },
+                )
+                .await
+                .expect("atomic adaptation")
+        });
+        // 管理员持有设置锁时提交新的账号事实，worker 必须依据提交后的值判断。
+        sqlx::query("update provider_accounts set enabled = $2, concurrency_limit = $3, weight = 7, notes = 'administrator edit' where id = $1")
+            .bind(id.as_str()).bind(enabled).bind(account_limit).execute(&mut *admin).await.expect("concurrent admin edit");
+        admin.commit().await.expect("commit admin edit");
+        let result = worker.await.expect("worker");
+        assert_eq!(result.is_some(), changes);
+        let actual = repository
+            .load_provider_account(id.as_str())
+            .await
+            .expect("read account")
+            .expect("account");
+        assert_eq!(actual.summary.enabled, enabled);
+        assert_eq!(
+            actual
+                .summary
+                .concurrency_limit
+                .map(|value| i64::from(value.get())),
+            expected_limit
+        );
+        assert_eq!(actual.summary.weight.get(), 7);
+        assert_eq!(actual.summary.notes.as_deref(), Some("administrator edit"));
+        assert_eq!(
+            account_group_ids(&database.pool, id.as_str()).await,
+            vec![group.to_owned()]
+        );
+    }
+    let audited_fields: Vec<Vec<String>> = sqlx::query_scalar(
+        "select changed_fields from admin_audit_events where action = 'adapt_concurrency'",
+    )
+    .fetch_all(&database.pool)
+    .await
+    .expect("audit");
+    assert_eq!(audited_fields, vec![vec!["concurrency_limit".to_owned()]]);
+    database.close().await;
+}
diff --git a/backend/crates/gateway-store/tests/postgres/runtime_settings.rs b/backend/crates/gateway-store/tests/postgres/runtime_settings.rs
index 10d22bc53..988c163cd 100644
--- a/backend/crates/gateway-store/tests/postgres/runtime_settings.rs
+++ b/backend/crates/gateway-store/tests/postgres/runtime_settings.rs
@@ -27,6 +27,13 @@ fn settings_with_margin(refresh_margin_seconds: u64) -> RuntimeSettingsUpdate {
         usage_retention_days: 31,
         ops_event_retention_days: 30,
         audit_retention_days: 90,
+        account_auto_freeze_enabled: true,
+        account_auto_freeze_threshold: 12,
+        account_auto_freeze_window_seconds: 600,
+        account_auto_freeze_duration_seconds: 7_200,
+        account_auto_freeze_probe_enabled: true,
+        account_auto_freeze_probe_model: None,
+        account_auto_freeze_adaptive_concurrency: true,
     }
 }
 
@@ -46,6 +53,30 @@ fn runtime_settings_reject_invalid_model_mapping() {
     assert!(settings.validate().is_err());
 }
 
+#[test]
+fn runtime_settings_reject_out_of_range_auto_freeze() {
+    for update in [
+        RuntimeSettingsUpdate {
+            account_auto_freeze_threshold: 1,
+            ..settings_with_margin(3_600)
+        },
+        RuntimeSettingsUpdate {
+            account_auto_freeze_window_seconds: 59,
+            ..settings_with_margin(3_600)
+        },
+        RuntimeSettingsUpdate {
+            account_auto_freeze_duration_seconds: 299,
+            ..settings_with_margin(3_600)
+        },
+        RuntimeSettingsUpdate {
+            account_auto_freeze_probe_model: Some(" pad ".to_owned()),
+            ..settings_with_margin(3_600)
+        },
+    ] {
+        assert!(update.validate().is_err());
+    }
+}
+
 #[test]
 fn runtime_settings_reject_non_semver_client_min() {
     let settings = RuntimeSettingsUpdate {
@@ -238,3 +269,52 @@ async fn concurrency_queue_settings_round_trip_into_the_runtime_snapshot() {
     assert!(snapshot.config_revision > before.config_revision);
     database.close().await;
 }
+
+#[tokio::test]
+async fn auto_freeze_defaults_off_and_migration_preserves_explicit_opt_in() {
+    let Some(database) = TestDatabase::create("freeze_opt_in").await else {
+        return;
+    };
+    let repository = PgRuntimeSettingsRepository::new(database.pool.clone());
+    assert!(
+        !repository
+            .load_runtime_settings()
+            .await
+            .expect("default settings")
+            .account_auto_freeze_enabled
+    );
+    sqlx::query("update runtime_settings set account_auto_freeze_enabled = true where id = 1")
+        .execute(&database.pool)
+        .await
+        .expect("old default");
+    let migration = include_str!("../../../../migrations/0011_account_auto_freeze_opt_in.sql");
+    sqlx::raw_sql(migration)
+        .execute(&database.pool)
+        .await
+        .expect("migrate untouched default");
+    assert!(
+        !repository
+            .load_runtime_settings()
+            .await
+            .expect("settings")
+            .account_auto_freeze_enabled
+    );
+    sqlx::query("update runtime_settings set account_auto_freeze_enabled = true where id = 1")
+        .execute(&database.pool)
+        .await
+        .expect("explicit opt-in");
+    sqlx::query("insert into admin_audit_events (id, actor_kind, actor_ref, action, entity_kind, entity_ref, changed_fields, created_at) values ('freeze-opt-in', 'system', 'test', 'settings.replace', 'runtime_settings', '1', array['account_auto_freeze'], now())")
+        .execute(&database.pool).await.expect("saved settings audit");
+    sqlx::raw_sql(migration)
+        .execute(&database.pool)
+        .await
+        .expect("preserve opt-in");
+    assert!(
+        repository
+            .load_runtime_settings()
+            .await
+            .expect("settings")
+            .account_auto_freeze_enabled
+    );
+    database.close().await;
+}
diff --git a/backend/crates/gateway-store/tests/redis/credential_cooldown.rs b/backend/crates/gateway-store/tests/redis/credential_cooldown.rs
index 0593dddfb..c5a1b2317 100644
--- a/backend/crates/gateway-store/tests/redis/credential_cooldown.rs
+++ b/backend/crates/gateway-store/tests/redis/credential_cooldown.rs
@@ -3,7 +3,9 @@ use std::time::{Duration as StdDuration, SystemTime};
 use chrono::{DateTime, Duration, Utc};
 use gateway_core::{
     account::{CredentialRevision, ProviderAccountId},
-    provider_ports::{ProviderCooldownPort, ProviderCooldownScope, ProviderScopedCooldown},
+    provider_ports::{
+        ProviderCooldownKind, ProviderCooldownPort, ProviderCooldownScope, ProviderScopedCooldown,
+    },
 };
 use gateway_store::{
     Revision,
@@ -18,6 +20,7 @@ fn credential_cooldown_is_revision_fenced() {
         provider_account_id: "account-1".to_owned(),
         credential_revision: Revision::new(2).expect("positive revision"),
         cooldown_until: Utc::now() + Duration::seconds(30),
+        kind: ProviderCooldownKind::RateLimit,
     };
     assert_eq!(cooldown.credential_revision.get(), 2);
 }
@@ -159,6 +162,7 @@ async fn credential_cooldown_read_removes_expired_grace_key() {
         provider_account_id: "acct_cooldown_expiry".to_owned(),
         credential_revision: Revision::new(1).expect("positive revision"),
         cooldown_until: millisecond_precision(cooldown_until),
+        kind: ProviderCooldownKind::RateLimit,
     };
     repository
         .cache_credential_cooldown(&cooldown)
@@ -284,6 +288,7 @@ fn cooldown(account_id: &str, revision: u64, seconds: i64) -> CredentialCooldown
         provider_account_id: account_id.to_owned(),
         credential_revision: Revision::new(revision).expect("positive revision"),
         cooldown_until: millisecond_precision(Utc::now() + Duration::seconds(seconds)),
+        kind: ProviderCooldownKind::RateLimit,
     }
 }
 
@@ -311,3 +316,233 @@ async fn namespace_keys(connection: &mut ConnectionManager, namespace: &str) ->
         .await
         .expect("list isolated cooldown keys")
 }
+
+fn runtime_store(
+    repository: &RedisCredentialCooldownRepository,
+    connection: ConnectionManager,
+    namespace: &str,
+) -> gateway_store::redis::RedisAdminAccountRuntimeStore {
+    gateway_store::redis::RedisAdminAccountRuntimeStore::new(
+        repository.clone(),
+        gateway_store::redis::RedisCredentialLeaseRepository::new(connection, namespace)
+            .expect("lease repository"),
+    )
+}
+
+#[tokio::test]
+async fn capacity_freeze_survives_due_time_and_runtime_restart_until_probe_succeeds() {
+    use gateway_admin::ports::store::AccountRuntimeStore as _;
+    let Some((repository, mut connection, namespace)) = repository().await else {
+        return;
+    };
+    let frozen = CredentialCooldown {
+        kind: ProviderCooldownKind::CapacityFreezeProbe,
+        cooldown_until: millisecond_precision(Utc::now() - Duration::seconds(90)),
+        ..cooldown("acct_probe_due", 1, 30)
+    };
+    repository
+        .cache_credential_cooldown(&frozen)
+        .await
+        .expect("write due freeze");
+    let runtime = runtime_store(&repository, connection.clone(), &namespace);
+    let snapshot = runtime.active_rate_limits().await.expect("status snapshot");
+    assert!(snapshot.cooldown[&frozen.provider_account_id].is_active(SystemTime::now()));
+    assert!(
+        snapshot.cooldown[&frozen.provider_account_id]
+            .kind
+            .requires_probe()
+    );
+    let key = namespace_keys(&mut connection, &namespace)
+        .await
+        .into_iter()
+        .find(|key| key.ends_with(":cooldown"))
+        .expect("freeze key");
+    let ttl: i64 = redis::cmd("PTTL")
+        .arg(key)
+        .query_async(&mut connection)
+        .await
+        .expect("TTL");
+    assert_eq!(
+        ttl, -1,
+        "awaiting probe must not expire while worker is unavailable"
+    );
+    drop(runtime);
+    let restarted = runtime_store(&repository, connection, &namespace);
+    let freezes = restarted
+        .active_freezes()
+        .await
+        .expect("due freezes after restart");
+    assert_eq!(freezes.len(), 1);
+    assert!(
+        restarted
+            .finish_freeze(
+                &frozen.provider_account_id,
+                &freezes[&frozen.provider_account_id],
+                None
+            )
+            .await
+            .expect("probe completion")
+    );
+    assert!(
+        repository
+            .read_credential_cooldown(&frozen.provider_account_id)
+            .await
+            .expect("read")
+            .is_none()
+    );
+}
+
+#[tokio::test]
+async fn stale_probe_cannot_recreate_manual_recovery_or_clear_or_extend_new_freeze() {
+    use gateway_admin::ports::store::AccountRuntimeStore as _;
+    let Some((repository, connection, namespace)) = repository().await else {
+        return;
+    };
+    let frozen = CredentialCooldown {
+        kind: ProviderCooldownKind::CapacityFreezeProbe,
+        ..cooldown("acct_probe_generation", 1, 30)
+    };
+    let runtime = runtime_store(&repository, connection, &namespace);
+    repository
+        .cache_credential_cooldown(&frozen)
+        .await
+        .expect("freeze");
+    let original =
+        runtime.active_freezes().await.expect("snapshot")[&frozen.provider_account_id].clone();
+    repository
+        .delete_account_cooldowns(&frozen.provider_account_id)
+        .await
+        .expect("manual recovery");
+    let postponed = Utc::now() + Duration::hours(2);
+    assert!(
+        !runtime
+            .finish_freeze(&frozen.provider_account_id, &original, Some(postponed))
+            .await
+            .expect("stale failure")
+    );
+    assert!(
+        repository
+            .read_credential_cooldown(&frozen.provider_account_id)
+            .await
+            .expect("read")
+            .is_none()
+    );
+    // 同凭据版本、相同截止时间的新冻结仍必须视为不同代次。
+    repository
+        .cache_credential_cooldown(&frozen)
+        .await
+        .expect("new freeze");
+    assert!(
+        !runtime
+            .finish_freeze(&frozen.provider_account_id, &original, None)
+            .await
+            .expect("stale success")
+    );
+    assert!(
+        !runtime
+            .finish_freeze(&frozen.provider_account_id, &original, Some(postponed))
+            .await
+            .expect("stale failure")
+    );
+    let current =
+        runtime.active_freezes().await.expect("new snapshot")[&frozen.provider_account_id].clone();
+    assert_ne!(current.generation, original.generation);
+    assert!(
+        runtime
+            .finish_freeze(&frozen.provider_account_id, &current, Some(postponed))
+            .await
+            .expect("postpone current freeze")
+    );
+    assert!(
+        !runtime
+            .finish_freeze(&frozen.provider_account_id, &current, None)
+            .await
+            .expect("result from previous probe cycle")
+    );
+    let extended = runtime.active_freezes().await.expect("extended snapshot")
+        [&frozen.provider_account_id]
+        .clone();
+    assert!(extended.until >= current.until);
+    assert!(
+        runtime
+            .finish_freeze(&frozen.provider_account_id, &extended, None)
+            .await
+            .expect("recover current cycle")
+    );
+}
+
+#[tokio::test]
+async fn ordinary_success_and_later_429_preserve_capacity_freeze_and_peak_evidence() {
+    let Some((repository, _connection, _namespace)) = repository().await else {
+        return;
+    };
+    let id = ProviderAccountId::new("acct_late_success").expect("account id");
+    let revision = CredentialRevision::new(1).expect("revision");
+    repository
+        .record_capacity_failure(&id, StdDuration::from_secs(600), 8)
+        .await
+        .expect("failure evidence");
+    let frozen = CredentialCooldown {
+        kind: ProviderCooldownKind::CapacityFreezeProbe,
+        ..cooldown(id.as_str(), 1, 30)
+    };
+    repository
+        .cache_credential_cooldown(&frozen)
+        .await
+        .expect("freeze");
+    repository
+        .clear_after_success(&id, revision)
+        .await
+        .expect("late success");
+    assert_eq!(
+        repository.capacity_peak_in_flight(&id).await.expect("peak"),
+        Some(8)
+    );
+    assert!(
+        !repository
+            .cache_credential_cooldown(&cooldown(id.as_str(), 1, 60))
+            .await
+            .expect("429 must not downgrade freeze")
+    );
+    assert_eq!(
+        repository
+            .read_credential_cooldown(id.as_str())
+            .await
+            .expect("read"),
+        Some(frozen)
+    );
+}
+
+#[tokio::test]
+async fn ordinary_success_still_clears_rate_limit_and_unfrozen_failure_evidence() {
+    let Some((repository, _connection, _namespace)) = repository().await else {
+        return;
+    };
+    let id = ProviderAccountId::new("acct_success_429").expect("id");
+    repository
+        .record_capacity_failure(&id, StdDuration::from_secs(600), 3)
+        .await
+        .expect("failure evidence");
+    repository
+        .cache_credential_cooldown(&cooldown(id.as_str(), 1, 30))
+        .await
+        .expect("429");
+    repository
+        .clear_after_success(&id, CredentialRevision::new(1).expect("revision"))
+        .await
+        .expect("success");
+    assert!(
+        repository
+            .read_credential_cooldown(id.as_str())
+            .await
+            .expect("read")
+            .is_none()
+    );
+    assert!(
+        repository
+            .capacity_peak_in_flight(&id)
+            .await
+            .expect("peak")
+            .is_none()
+    );
+}
diff --git a/backend/crates/providers/openai/src/credential/quota/mod.rs b/backend/crates/providers/openai/src/credential/quota/mod.rs
index 43b4ab9d1..b31b4d0a2 100644
--- a/backend/crates/providers/openai/src/credential/quota/mod.rs
+++ b/backend/crates/providers/openai/src/credential/quota/mod.rs
@@ -24,7 +24,10 @@ use gateway_core::account::{
     QuotaAccessChange, QuotaAccessState, QuotaEvidence, QuotaObservation, QuotaObservationTouch,
     QuotaState, QuotaWriteOutcome,
 };
-use gateway_core::provider_ports::{ProviderCooldown, ProviderCooldownPort};
+use gateway_core::provider_ports::{
+    ProviderCooldown, ProviderCooldownKind, ProviderCooldownPort, ProviderFreezePolicy,
+    ProviderLeasePort, ProviderRuntimePolicyPort,
+};
 use gateway_protocol::openai::events::{
     ParsedRateLimits, RateLimitDetails, RateLimitWindow, parse_rate_limit_headers,
 };
@@ -184,10 +187,17 @@ pub struct CodexCredentialQuotaService {
     http: Client,
     base_url: String,
     cooldowns: Arc<dyn ProviderCooldownPort>,
+    leases: Arc<dyn ProviderLeasePort>,
+    runtime_policy: Arc<dyn ProviderRuntimePolicyPort>,
+    /// 冻结策略的短 TTL 缓存：失败路径热读，避免每个容量错误都查询设置。
+    freeze_policy_cache: Mutex<Option<(ProviderFreezePolicy, Instant)>>,
     scheduling: CodexQuotaSchedulingProjection,
     reset_consume_locks: Mutex<HashMap<ProviderAccountId, Arc<Mutex<()>>>>,
 }
 
+/// 冻结策略缓存活跃期；过期后下一次容量错误重新读取运行时设置。
+const FREEZE_POLICY_CACHE_TTL: Duration = Duration::from_secs(30);
+
 #[derive(Debug, Clone, Copy, PartialEq, Eq)]
 enum QuotaRefreshAuthority {
     ObserveAccess,
@@ -481,6 +491,8 @@ impl CodexCredentialQuotaService {
         http: Client,
         base_url: String,
         cooldowns: Arc<dyn ProviderCooldownPort>,
+        leases: Arc<dyn ProviderLeasePort>,
+        runtime_policy: Arc<dyn ProviderRuntimePolicyPort>,
     ) -> Self {
         Self {
             store: Arc::clone(repository.store()),
@@ -489,11 +501,89 @@ impl CodexCredentialQuotaService {
             http,
             base_url,
             cooldowns,
+            leases,
+            runtime_policy,
+            freeze_policy_cache: Mutex::new(None),
             scheduling: CodexQuotaSchedulingProjection::default(),
             reset_consume_locks: Mutex::new(HashMap::new()),
         }
     }
 
+    /// 读取容量熔断策略（带短 TTL 缓存）；读取失败退化为关闭，熔断不得放大失败。
+    async fn freeze_policy(&self) -> ProviderFreezePolicy {
+        {
+            let cache = self.freeze_policy_cache.lock().await;
+            if let Some((policy, loaded_at)) = cache.as_ref()
+                && loaded_at.elapsed() < FREEZE_POLICY_CACHE_TTL
+            {
+                return policy.clone();
+            }
+        }
+        let loaded = self.runtime_policy.load_freeze_policy().await.ok();
+        let mut cache = self.freeze_policy_cache.lock().await;
+        if let Some(policy) = loaded {
+            *cache = Some((policy.clone(), Instant::now()));
+            return policy;
+        }
+        cache
+            .as_ref()
+            .filter(|(_, loaded_at)| loaded_at.elapsed() < FREEZE_POLICY_CACHE_TTL)
+            .map_or_else(ProviderFreezePolicy::disabled, |(policy, _)| policy.clone())
+    }
+
+    /// 容量熔断入口：滑动窗口内累计容量类失败，达到阈值即写入带
+    /// `CapacityFreeze` 类别的账号级冷却。调度侧立即屏蔽该账号，恢复由
+    /// freeze-recovery worker 处理；本路径只依赖 Redis 可丢失事实。
+    pub async fn apply_capacity_failure(&self, account: &ProviderAccount, observed_at: SystemTime) {
+        let policy = self.freeze_policy().await;
+        if !policy.enabled() {
+            return;
+        }
+        let in_flight = self.current_in_flight(account.id()).await;
+        let Ok(count) = self
+            .cooldowns
+            .record_capacity_failure(account.id(), policy.window(), in_flight)
+            .await
+        else {
+            return;
+        };
+        if count < policy.threshold() {
+            return;
+        }
+        let Some(until) = observed_at.checked_add(policy.freeze_duration()) else {
+            return;
+        };
+        let cooldown = ProviderCooldown::new_with_kind(
+            account.id().clone(),
+            account.revision(),
+            until,
+            if policy.probe_enabled() {
+                ProviderCooldownKind::CapacityFreezeProbe
+            } else {
+                ProviderCooldownKind::CapacityFreeze
+            },
+        );
+        if self.cooldowns.put_if_later(cooldown).await.is_ok() {
+            tracing::warn!(
+                account_id = account.id().as_str(),
+                threshold = policy.threshold(),
+                window_seconds = policy.window().as_secs(),
+                freeze_seconds = policy.freeze_duration().as_secs(),
+                peak_in_flight = in_flight,
+                "账号容量熔断触发：冻结该账号一段时间",
+            );
+        }
+    }
+
+    async fn current_in_flight(&self, account_id: &ProviderAccountId) -> u32 {
+        self.leases
+            .account_in_flight(std::slice::from_ref(account_id))
+            .await
+            .ok()
+            .and_then(|signals| signals.get(account_id).copied())
+            .unwrap_or(0)
+    }
+
     /// 查询当前账号由 Codex Desktop 暴露的主动额度重置卡。
     pub async fn list_reset_credits(
         &self,
@@ -607,7 +697,7 @@ impl CodexCredentialQuotaService {
     }
 
     /// 成功推理是额度可访问的权威证据，同时解除账号级 429 冷却。
-    pub(crate) async fn record_successful_inference(
+    pub async fn record_successful_inference(
         &self,
         account: &ProviderAccount,
         observed_at: SystemTime,
@@ -628,7 +718,7 @@ impl CodexCredentialQuotaService {
             }
         }
         self.cooldowns
-            .clear(account.id(), account.revision())
+            .clear_after_success(account.id(), account.revision())
             .await
             .map_err(|error| CodexCredentialQuotaError::Store {
                 detail: error.to_string(),
@@ -975,11 +1065,11 @@ impl CodexCredentialQuotaService {
         Ok(())
     }
 
-    /// 读取账号当前是否处于临时限流（429）冷却，及到期时间。
-    pub async fn rate_limited_until(
+    /// 读取有效的账号冷却事实；等待恢复探测的冻结到期后仍有效。
+    pub async fn cooldown(
         &self,
         account_id: &ProviderAccountId,
-    ) -> Result<Option<SystemTime>, CodexCredentialQuotaError> {
+    ) -> Result<Option<gateway_core::account::AccountCooldown>, CodexCredentialQuotaError> {
         let Some(cooldown) = self.cooldowns.read(account_id).await.map_err(|error| {
             CodexCredentialQuotaError::Store {
                 detail: error.to_string(),
@@ -988,11 +1078,8 @@ impl CodexCredentialQuotaService {
         else {
             return Ok(None);
         };
-        let until = cooldown.until();
-        if until <= SystemTime::now() {
-            return Ok(None);
-        }
-        Ok(Some(until))
+        let state = cooldown.scheduling_state();
+        Ok(state.is_active(SystemTime::now()).then_some(state))
     }
 
     /// 读取单账号最后一次落库的 Provider quota，并由 Codex 域解析展示窗口。
diff --git a/backend/crates/providers/openai/src/credential/selector.rs b/backend/crates/providers/openai/src/credential/selector.rs
index e08945f23..db683bbc6 100644
--- a/backend/crates/providers/openai/src/credential/selector.rs
+++ b/backend/crates/providers/openai/src/credential/selector.rs
@@ -451,11 +451,7 @@ impl CodexCredentialSelector {
             let mut rate_limits = HashMap::with_capacity(accounts.len());
             if !diagnostic {
                 for account in &accounts {
-                    let until = self
-                        .quota
-                        .rate_limited_until(account.id())
-                        .await
-                        .unwrap_or(None);
+                    let until = self.quota.cooldown(account.id()).await.unwrap_or(None);
                     rate_limits.insert(account.id().clone(), until);
                 }
             }
@@ -487,7 +483,7 @@ impl CodexCredentialSelector {
                             last_started_at: None,
                             quota_reset_at: None,
                             quota_remaining_rank: None,
-                            rate_limited_until: None,
+                            cooldown: None,
                             failure_rate_basis_points: None,
                             first_output_latency_ms: None,
                         })
@@ -1330,7 +1326,7 @@ fn affinity_selection_for_bound_account(
     };
     match candidate
         .account
-        .status_projection(now, candidate.signals.rate_limited_until)
+        .status_projection(now, candidate.signals.cooldown)
         .status
     {
         AccountStatus::Normal => AffinitySelection::preferred(account_id),
@@ -1349,7 +1345,7 @@ fn affinity_unavailable_reason(
 ) -> AffinityEscapeReason {
     match candidate
         .account
-        .status_projection(now, candidate.signals.rate_limited_until)
+        .status_projection(now, candidate.signals.cooldown)
         .status
     {
         AccountStatus::QuotaExhausted => AffinityEscapeReason::QuotaExhausted,
diff --git a/backend/crates/providers/openai/src/lib.rs b/backend/crates/providers/openai/src/lib.rs
index dbc98ec8d..49f030037 100644
--- a/backend/crates/providers/openai/src/lib.rs
+++ b/backend/crates/providers/openai/src/lib.rs
@@ -112,6 +112,8 @@ pub async fn initialize(
         http.clone(),
         config.base_url().to_owned(),
         ports.cooldowns(),
+        Arc::clone(&leases),
+        Arc::clone(&runtime_policy),
     ));
     let profile_statistics = Arc::new(CodexCredentialProfileService::new(
         repository.clone(),
diff --git a/backend/crates/providers/openai/src/provider/execution.rs b/backend/crates/providers/openai/src/provider/execution.rs
index 162fec634..0f6953e66 100644
--- a/backend/crates/providers/openai/src/provider/execution.rs
+++ b/backend/crates/providers/openai/src/provider/execution.rs
@@ -403,6 +403,7 @@ pub(super) fn cold_json_response_stream(request: ColdJsonResponse) -> EventStrea
             response_origin: &request.response_origin,
             cyber_policy_scope: None,
             allows_account_state_mutation,
+            allows_capacity_feedback: !request.context.is_diagnostic_required_account(),
         };
         let active_account = request.lease.account().clone();
         let cookie_header = build_cookie_header(request.lease.cookies())?;
@@ -577,6 +578,7 @@ pub(super) fn cold_response_stream(response: ColdResponse) -> EventStream {
             response_origin: &response_origin,
             cyber_policy_scope: cyber_policy_scope.as_ref(),
             allows_account_state_mutation,
+            allows_capacity_feedback: !context.is_diagnostic_required_account(),
         };
         let mut active_account = lease.account().clone();
         let cookie_header = build_cookie_header(lease.cookies())?;
diff --git a/backend/crates/providers/openai/src/provider/failure.rs b/backend/crates/providers/openai/src/provider/failure.rs
index 93024dd4f..75de8acab 100644
--- a/backend/crates/providers/openai/src/provider/failure.rs
+++ b/backend/crates/providers/openai/src/provider/failure.rs
@@ -1,13 +1,15 @@
 //! OpenAI 上游失败分类、恢复决策与稳定错误投影。
 
 use super::*;
-use crate::transport::diagnostics::is_capacity_error;
 
 /// OpenAI 失败对 Smart 账号分数的结构化 reason 闭集。
 ///
 /// 已归一的上游 code 优先；code 缺失时才读取结构化客户端错误的 code/type。
 /// 新增错误、裸 HTTP 状态和内部错误 kind 默认都不会进入该闭集。
+/// 容量拒绝影响短期调度健康度，但不证明账号凭据或额度失效。
 const OPENAI_ACCOUNT_SCORE_FAILURE_REASONS: &[&str] = &[
+    "server_is_overloaded",
+    "slow_down",
     "rate_limit_exceeded",
     "rate_limit_error",
     "server_error",
@@ -32,17 +34,6 @@ fn openai_account_score_failure_reason(error: &ProviderError) -> Option<&str> {
 /// 返回该 OpenAI 失败是否属于 Smart 账号计分闭集。
 #[doc(hidden)]
 pub fn openai_failure_affects_account_score(error: &ProviderError) -> bool {
-    let client_error = error.client_visible_upstream_error();
-    if is_capacity_error(
-        error
-            .upstream_code()
-            .map(OpaqueUpstreamValue::as_str)
-            .or_else(|| client_error.and_then(ClientVisibleUpstreamError::code)),
-        client_error.and_then(ClientVisibleUpstreamError::error_type),
-        client_error.map(ClientVisibleUpstreamError::message),
-    ) {
-        return false;
-    }
     openai_account_score_failure_reason(error).is_some_and(is_openai_account_score_failure_reason)
 }
 
@@ -53,6 +44,7 @@ pub(super) struct MappedProviderFailure {
     /// 原始上游错误描述，仅在凭据错误状态下持久化。
     pub(super) error_message: Option<String>,
     pub(super) cyber_policy_failure: bool,
+    pub(super) upstream_capacity_failure: bool,
     pub(super) set_cookie_headers: Vec<String>,
     pub(super) rate_limit_headers: Vec<(String, String)>,
     pub(super) observation: Option<ProviderResponseObservation>,
@@ -67,6 +59,7 @@ impl MappedProviderFailure {
             account_failure: None,
             error_message: None,
             cyber_policy_failure: false,
+            upstream_capacity_failure: false,
             set_cookie_headers: Vec::new(),
             rate_limit_headers: Vec::new(),
             observation: None,
@@ -211,6 +204,7 @@ pub(super) struct OpenAiFailureContext<'a> {
     pub(super) response_origin: &'a Url,
     pub(super) cyber_policy_scope: Option<&'a CodexCyberPolicyScope>,
     pub(super) allows_account_state_mutation: bool,
+    pub(super) allows_capacity_feedback: bool,
 }
 
 #[derive(Clone, Copy)]
@@ -410,6 +404,13 @@ pub(super) async fn apply_failure(
             "Failed to persist OpenAI response cookies"
         );
     }
+    // 探测只验证恢复，不能把探测自身的单请求并发当作业务负载证据。
+    if context.allows_capacity_feedback && failure.upstream_capacity_failure {
+        context
+            .quota
+            .apply_capacity_failure(account, SystemTime::now())
+            .await;
+    }
 }
 
 pub(super) fn schedule_authoritative_quota_refresh_after_failure(
@@ -1269,6 +1270,12 @@ pub(super) fn map_upstream_failure(
         ),
         error_message: failure.client_message,
         cyber_policy_failure,
+        // 仅上游明确返回的容量/服务不可用计数；本地连接保护也会映射为
+        // Unavailable，不能用通用 ProviderErrorKind 推断容量证据。
+        upstream_capacity_failure: matches!(
+            category,
+            CodexFailureCategory::CapacityUnavailable | CodexFailureCategory::Unavailable
+        ),
         set_cookie_headers: failure.set_cookie_headers,
         rate_limit_headers: failure.rate_limit_headers,
         observation,
diff --git a/backend/crates/providers/openai/tests/admin.rs b/backend/crates/providers/openai/tests/admin.rs
index b494bf6ce..0b9e41d3f 100644
--- a/backend/crates/providers/openai/tests/admin.rs
+++ b/backend/crates/providers/openai/tests/admin.rs
@@ -1591,6 +1591,30 @@ impl ProviderCooldownPort for TestCooldown {
     ) -> BoxFuture<'a, Result<bool, ProviderStoreError>> {
         Box::pin(async { Ok(false) })
     }
+
+    fn record_capacity_failure<'a>(
+        &'a self,
+        _account_id: &'a ProviderAccountId,
+        _window: Duration,
+        _in_flight: u32,
+    ) -> BoxFuture<'a, Result<u32, ProviderStoreError>> {
+        Box::pin(async { Ok(0) })
+    }
+
+    fn clear_after_success<'a>(
+        &'a self,
+        _account_id: &'a ProviderAccountId,
+        _through_revision: gateway_core::account::CredentialRevision,
+    ) -> BoxFuture<'a, Result<(), ProviderStoreError>> {
+        Box::pin(async { Ok(()) })
+    }
+
+    fn capacity_peak_in_flight<'a>(
+        &'a self,
+        _account_id: &'a ProviderAccountId,
+    ) -> BoxFuture<'a, Result<Option<u32>, ProviderStoreError>> {
+        Box::pin(async { Ok(None) })
+    }
 }
 
 struct TestRuntimePolicy;
diff --git a/backend/crates/providers/openai/tests/credential/contract.rs b/backend/crates/providers/openai/tests/credential/contract.rs
index 373b8b06b..50bc14fba 100644
--- a/backend/crates/providers/openai/tests/credential/contract.rs
+++ b/backend/crates/providers/openai/tests/credential/contract.rs
@@ -20,7 +20,8 @@ use gateway_core::engine::{
 use gateway_core::lifecycle::CancellationToken;
 use gateway_core::policy::ClientApiKeyId;
 use gateway_core::provider_ports::{
-    ProviderCooldownPort, ProviderSessionAffinityKey, ProviderSessionAffinityPort,
+    ProviderCooldownPort, ProviderLeasePort, ProviderSessionAffinityKey,
+    ProviderSessionAffinityPort,
 };
 use gateway_core::routing::{
     ClientRoutingScope, FrozenAccountScope, ProviderKind, RuntimeAccount, RuntimeAccountDirectory,
@@ -178,6 +179,8 @@ fn selector_with_runtime(
         http,
         OFFICIAL_CODEX_BASE_URL.to_owned(),
         cooldowns,
+        Arc::clone(&leases) as Arc<dyn ProviderLeasePort>,
+        crate::support::runtime_policy(),
     ));
     CodexCredentialSelector::new(
         ProviderKind::new("openai").expect("provider"),
diff --git a/backend/crates/providers/openai/tests/credential/quota/capacity_freeze.rs b/backend/crates/providers/openai/tests/credential/quota/capacity_freeze.rs
new file mode 100644
index 000000000..546f7f537
--- /dev/null
+++ b/backend/crates/providers/openai/tests/credential/quota/capacity_freeze.rs
@@ -0,0 +1,195 @@
+//! 容量熔断触发：高频容量类失败按滑动窗口计数冻结账号，成功调用清空证据。
+
+use std::sync::Arc;
+use std::time::{Duration, SystemTime};
+
+use gateway_core::provider_ports::{
+    ProviderCooldownKind, ProviderCooldownPort as _, ProviderFreezePolicy,
+};
+use provider_openai::credential::CodexCredentialQuotaService;
+
+use super::{MemoryAccountStore, create_account, wire_profile};
+use crate::support::{MemoryCooldownPort, StaticFreezePolicy, TestLeaseCoordinator};
+
+fn freeze_policy(threshold: u32) -> ProviderFreezePolicy {
+    ProviderFreezePolicy::try_new(true, threshold, 600, 7_200, true, None, true)
+        .expect("valid freeze policy")
+}
+
+async fn freeze_service(
+    store: &Arc<MemoryAccountStore>,
+    policy: ProviderFreezePolicy,
+) -> (Arc<CodexCredentialQuotaService>, Arc<MemoryCooldownPort>) {
+    let cooldowns = Arc::new(MemoryCooldownPort::new());
+    let quota = Arc::new(CodexCredentialQuotaService::new(
+        store.repository(),
+        wire_profile(),
+        reqwest::Client::builder()
+            .no_proxy()
+            .build()
+            .expect("client"),
+        provider_openai::OFFICIAL_CODEX_BASE_URL.to_owned(),
+        Arc::clone(&cooldowns) as Arc<dyn gateway_core::provider_ports::ProviderCooldownPort>,
+        Arc::new(TestLeaseCoordinator::default()),
+        StaticFreezePolicy::policy_port(policy),
+    ));
+    (quota, cooldowns)
+}
+
+#[tokio::test]
+async fn capacity_failures_below_threshold_do_not_freeze() {
+    let store = Arc::new(MemoryAccountStore::default());
+    create_account(&store, "acct_capacity_below").await;
+    let account = store
+        .account("acct_capacity_below")
+        .expect("created account");
+    let (quota, cooldowns) = freeze_service(&store, freeze_policy(3)).await;
+
+    for _ in 0..2 {
+        quota
+            .apply_capacity_failure(&account, SystemTime::now())
+            .await;
+    }
+
+    assert_eq!(
+        cooldowns
+            .capacity_evidence(account.id())
+            .map(|(count, _)| count),
+        Some(2)
+    );
+    assert!(
+        quota
+            .cooldown(account.id())
+            .await
+            .expect("cooldown read")
+            .is_none()
+    );
+}
+
+#[tokio::test]
+async fn capacity_failures_reaching_threshold_write_capacity_freeze_cooldown() {
+    let store = Arc::new(MemoryAccountStore::default());
+    create_account(&store, "acct_capacity_freeze").await;
+    let account = store
+        .account("acct_capacity_freeze")
+        .expect("created account");
+    let (quota, cooldowns) = freeze_service(&store, freeze_policy(3)).await;
+
+    for _ in 0..3 {
+        quota
+            .apply_capacity_failure(&account, SystemTime::now())
+            .await;
+    }
+
+    let until = quota
+        .cooldown(account.id())
+        .await
+        .expect("cooldown read")
+        .expect("freeze cooldown written");
+    let remaining = until
+        .until
+        .duration_since(SystemTime::now())
+        .expect("freeze in the future");
+    assert!(
+        remaining > Duration::from_secs(7_000) && remaining <= Duration::from_secs(7_200),
+        "freeze should last about 2 hours, got {remaining:?}"
+    );
+    let cooldown = cooldowns.read(account.id()).await.expect("read cooldown");
+    assert_eq!(
+        cooldown.expect("cooldown").kind(),
+        ProviderCooldownKind::CapacityFreezeProbe
+    );
+}
+
+#[tokio::test]
+async fn disabled_freeze_policy_records_nothing() {
+    let store = Arc::new(MemoryAccountStore::default());
+    create_account(&store, "acct_capacity_disabled").await;
+    let account = store
+        .account("acct_capacity_disabled")
+        .expect("created account");
+    let (quota, cooldowns) = freeze_service(&store, ProviderFreezePolicy::disabled()).await;
+
+    for _ in 0..5 {
+        quota
+            .apply_capacity_failure(&account, SystemTime::now())
+            .await;
+    }
+
+    assert!(cooldowns.capacity_evidence(account.id()).is_none());
+    assert!(
+        quota
+            .cooldown(account.id())
+            .await
+            .expect("cooldown read")
+            .is_none()
+    );
+}
+
+#[tokio::test]
+async fn successful_inference_clears_capacity_evidence() {
+    let store = Arc::new(MemoryAccountStore::default());
+    create_account(&store, "acct_capacity_success").await;
+    let account = store
+        .account("acct_capacity_success")
+        .expect("created account");
+    let (quota, cooldowns) = freeze_service(&store, freeze_policy(100)).await;
+
+    for _ in 0..3 {
+        quota
+            .apply_capacity_failure(&account, SystemTime::now())
+            .await;
+    }
+    assert!(cooldowns.capacity_evidence(account.id()).is_some());
+
+    quota
+        .record_successful_inference(&account, SystemTime::now())
+        .await
+        .expect("record success");
+
+    assert!(cooldowns.capacity_evidence(account.id()).is_none());
+}
+
+/// 阈值与窗口边界使用与迁移 check 约束一致的构造校验。
+#[test]
+fn freeze_policy_rejects_out_of_range_configuration() {
+    for (threshold, window, duration) in [(1, 600, 7_200), (12, 59, 7_200), (12, 600, 299)] {
+        assert!(
+            ProviderFreezePolicy::try_new(true, threshold, window, duration, true, None, true)
+                .is_err()
+        );
+    }
+    assert!(
+        ProviderFreezePolicy::try_new(true, 12, 600, 7_200, true, Some(" ".to_owned()), true)
+            .is_err()
+    );
+    assert!(ProviderFreezePolicy::try_new(true, 12, 600, 7_200, true, None, true).is_ok());
+}
+
+#[tokio::test]
+async fn in_flight_success_does_not_release_a_new_capacity_freeze() {
+    let store = Arc::new(MemoryAccountStore::default());
+    create_account(&store, "acct_freeze_late_success").await;
+    let account = store.account("acct_freeze_late_success").expect("account");
+    let (quota, cooldowns) = freeze_service(&store, freeze_policy(2)).await;
+    for _ in 0..2 {
+        quota
+            .apply_capacity_failure(&account, SystemTime::now())
+            .await;
+    }
+    let before = cooldowns.read(account.id()).await.expect("freeze");
+    quota
+        .record_successful_inference(&account, SystemTime::now())
+        .await
+        .expect("late success");
+    assert_eq!(
+        cooldowns.read(account.id()).await.expect("freeze remains"),
+        before
+    );
+    assert_eq!(
+        cooldowns
+            .capacity_evidence(account.id())
+            .map(|(count, _)| count),
+        Some(2)
+    );
+}
diff --git a/backend/crates/providers/openai/tests/credential/quota/mod.rs b/backend/crates/providers/openai/tests/credential/quota/mod.rs
index 3bee3a98b..ae25554e5 100644
--- a/backend/crates/providers/openai/tests/credential/quota/mod.rs
+++ b/backend/crates/providers/openai/tests/credential/quota/mod.rs
@@ -1,5 +1,6 @@
 //! OpenAI 额度事实边界与展示快照回归。
 
+mod capacity_freeze;
 mod recovery;
 mod refresh_timing;
 mod scheduling;
@@ -68,6 +69,8 @@ fn quota_service_with_base_url(
         http,
         base_url,
         Arc::new(crate::support::MemoryCooldownPort::new()),
+        Arc::new(crate::support::TestLeaseCoordinator::default()),
+        crate::support::runtime_policy(),
     )
 }
 
diff --git a/backend/crates/providers/openai/tests/credential/quota/scheduling.rs b/backend/crates/providers/openai/tests/credential/quota/scheduling.rs
index c59da5043..0ac7f7693 100644
--- a/backend/crates/providers/openai/tests/credential/quota/scheduling.rs
+++ b/backend/crates/providers/openai/tests/credential/quota/scheduling.rs
@@ -84,7 +84,7 @@ async fn old_weekly_quota_should_outweigh_one_in_flight_title_request() {
                 last_started_at: None,
                 quota_reset_at: None,
                 quota_remaining_rank: None,
-                rate_limited_until: None,
+                cooldown: None,
                 failure_rate_basis_points: None,
                 first_output_latency_ms: None,
             }
diff --git a/backend/crates/providers/openai/tests/provider/contract.rs b/backend/crates/providers/openai/tests/provider/contract.rs
index 2d4b12636..2d4641099 100644
--- a/backend/crates/providers/openai/tests/provider/contract.rs
+++ b/backend/crates/providers/openai/tests/provider/contract.rs
@@ -32,6 +32,7 @@ use gateway_core::operation::{
     OperationKind, ProtocolPayload, ProviderSessionState, RawJsonPayload, StandaloneSearchRequest,
 };
 use gateway_core::policy::ClientApiKeyId;
+use gateway_core::provider_ports::ProviderLeasePort;
 use gateway_core::routing::{
     ClientRoutingScope, ConfigRevision, FrozenAccountScope, ModelCapabilities, ModelServiceTier,
     ProviderKind, ProviderModel, PublicModelId, RoutingContext, RuntimeAccount,
@@ -227,6 +228,31 @@ fn provider_and_quota_with_affinity_and_base_url_and_leases(
     leases: Arc<TestLeaseCoordinator>,
     stream_max_retries: u32,
 ) -> (CodexProvider, Arc<CodexCredentialQuotaService>) {
+    let (provider, quota, _) = provider_and_quota_with_runtime_ports(
+        store,
+        session_affinity,
+        base_url,
+        leases,
+        stream_max_retries,
+        Arc::new(MemoryCooldownPort::new()),
+        crate::support::runtime_policy(),
+    );
+    (provider, quota)
+}
+
+fn provider_and_quota_with_runtime_ports(
+    store: &Arc<MemoryAccountStore>,
+    session_affinity: Arc<MemorySessionAffinity>,
+    base_url: String,
+    leases: Arc<TestLeaseCoordinator>,
+    stream_max_retries: u32,
+    cooldowns: Arc<MemoryCooldownPort>,
+    policy: Arc<dyn gateway_core::provider_ports::ProviderRuntimePolicyPort>,
+) -> (
+    CodexProvider,
+    Arc<CodexCredentialQuotaService>,
+    Arc<CodexWebSocketPool>,
+) {
     let profile = wire_profile();
     let http = reqwest::Client::builder()
         .no_proxy()
@@ -245,7 +271,9 @@ fn provider_and_quota_with_affinity_and_base_url_and_leases(
         profile.clone(),
         http.clone(),
         base_url.clone(),
-        Arc::new(MemoryCooldownPort::new()),
+        cooldowns,
+        Arc::clone(&leases) as Arc<dyn ProviderLeasePort>,
+        policy,
     ));
     let account_feedback = Arc::new(AccountFeedbackStats::default());
     let selector = Arc::new(CodexCredentialSelector::new(
@@ -268,11 +296,11 @@ fn provider_and_quota_with_affinity_and_base_url_and_leases(
         http,
         profile,
         base_url,
-        websocket_pool,
+        Arc::clone(&websocket_pool),
         stream_max_retries,
     )
     .expect("official OpenAI provider");
-    (provider, quota)
+    (provider, quota, websocket_pool)
 }
 
 async fn create_account(store: &Arc<MemoryAccountStore>, id: &str) {
@@ -5476,7 +5504,11 @@ async fn capacity_http_and_websocket_opening_rejections_use_bounded_business_ret
             create_account(&store, account_id).await;
             let server = MockServer::start().await;
             let body = json!({"error": {
-                "code": if status == 503 { Some("server_is_overloaded") } else { None },
+                "code": match status {
+                    429 => Some("slow_down"),
+                    503 => Some("server_is_overloaded"),
+                    _ => None,
+                },
                 "type": "server_error",
                 "message": "Selected model is at capacity. Please try a different model.",
                 "extension": {"preserved": true}
@@ -5518,7 +5550,7 @@ async fn capacity_http_and_websocket_opening_rejections_use_bounded_business_ret
                 max_retries, initial_delay, max_delay,
             }) if max_retries.get() == 3 && initial_delay == Duration::from_secs(8) && max_delay == Duration::from_secs(8))
             );
-            assert!(!provider_openai::openai_failure_affects_account_score(
+            assert!(provider_openai::openai_failure_affects_account_score(
                 &error
             ));
             assert_eq!(
@@ -5540,6 +5572,196 @@ async fn capacity_http_and_websocket_opening_rejections_use_bounded_business_ret
     }
 }
 
+fn capacity_websocket_operation() -> Operation {
+    let payload = ProtocolPayload::json_object(
+        "openai",
+        Map::from_iter([
+            ("model".to_owned(), json!("gpt-5.4")),
+            ("input".to_owned(), json!("hello")),
+            ("store".to_owned(), json!(false)),
+            ("session_id".to_owned(), json!("capacity-feedback-session")),
+        ]),
+    )
+    .expect("WebSocket payload")
+    .with_context(Map::from_iter([(
+        "downstream_websocket_connection_id".to_owned(),
+        json!("ws_capacity_feedback"),
+    )]));
+    Operation::Generate(
+        GenerateRequest::from_protocol_payload(payload).with_provider_session_state(
+            ProviderSessionState::new(
+                "openai",
+                Map::from_iter([
+                    ("account_id".to_owned(), json!("acct_provider_contract")),
+                    (
+                        "conversation_id".to_owned(),
+                        json!("capacity-feedback-session"),
+                    ),
+                    ("continuation_scope".to_owned(), json!("persisted")),
+                ]),
+            )
+            .expect("WebSocket session scope"),
+        ),
+    )
+}
+
+fn provider_with_capacity_tracking(
+    store: &Arc<MemoryAccountStore>,
+    base_url: String,
+    cooldowns: Arc<MemoryCooldownPort>,
+) -> (CodexProvider, Arc<CodexWebSocketPool>) {
+    let (provider, _, pool) = provider_and_quota_with_runtime_ports(
+        store,
+        Arc::new(MemorySessionAffinity::default()),
+        base_url,
+        Arc::new(TestLeaseCoordinator::default()),
+        0,
+        cooldowns,
+        crate::support::StaticFreezePolicy::policy_port(
+            gateway_core::provider_ports::ProviderFreezePolicy::try_new(
+                true, 12, 600, 7_200, true, None, true,
+            )
+            .expect("freeze policy"),
+        ),
+    );
+    (provider, pool)
+}
+
+#[tokio::test]
+async fn capacity_feedback_counts_business_rejections_but_excludes_diagnostic_probes() {
+    use gateway_core::provider_ports::ProviderCooldownPort as _;
+
+    for websocket in [false, true] {
+        for (status, code) in [
+            (429, "slow_down"),
+            (503, "server_is_overloaded"),
+            (500, "server_error"),
+        ] {
+            for diagnostic in [false, true] {
+                let store = Arc::new(MemoryAccountStore::default());
+                let account_id = "acct_provider_contract";
+                create_account(&store, account_id).await;
+                let account = store.account(account_id).expect("account");
+                let cooldowns = Arc::new(MemoryCooldownPort::new());
+                let server = MockServer::start().await;
+                Mock::given(method(if websocket { "GET" } else { "POST" }))
+                    .and(path("/codex/responses"))
+                    .respond_with(
+                        ResponseTemplate::new(status).set_body_json(json!({"error": {
+                            "code": code,
+                            "message": "Upstream temporarily unavailable"
+                        }})),
+                    )
+                    .expect(2)
+                    .mount(&server)
+                    .await;
+                let (provider, _) =
+                    provider_with_capacity_tracking(&store, server.uri(), Arc::clone(&cooldowns));
+                // 同时验证窗口证据已过期与尚未过期：探测不能重建峰值，也不能改写原计数。
+                for existing_evidence in [None, Some((4, 20))] {
+                    if let Some((count, peak)) = existing_evidence {
+                        for _ in 0..count {
+                            cooldowns
+                                .record_capacity_failure(
+                                    account.id(),
+                                    Duration::from_secs(600),
+                                    peak,
+                                )
+                                .await
+                                .expect("seed normal-request evidence");
+                        }
+                    }
+                    let before = cooldowns.capacity_evidence(account.id());
+                    let attempt = if diagnostic {
+                        diagnostic_context("req_capacity_feedback_probe", account_id)
+                    } else {
+                        context("req_capacity_feedback_business", CancellationToken::new())
+                    };
+                    let operation = if websocket {
+                        capacity_websocket_operation()
+                    } else {
+                        http_generate_operation()
+                    };
+                    let mut stream = provider
+                        .execute(planned_request("openai", operation), attempt)
+                        .await
+                        .expect("prepare upstream attempt");
+                    let error = loop {
+                        match stream.next().await {
+                            Some(Ok(_)) => {}
+                            Some(Err(error)) => break error,
+                            None => panic!("expected upstream rejection"),
+                        }
+                    };
+                    assert_eq!(
+                        error.upstream_status(),
+                        Some(status),
+                        "unexpected upstream error: {error:?}"
+                    );
+                    let after = cooldowns.capacity_evidence(account.id());
+                    if diagnostic {
+                        assert_eq!(after, before, "probe must preserve capacity count and peak");
+                    } else {
+                        assert_eq!(
+                            after.map(|(count, _)| count),
+                            Some(before.map_or(1, |(count, _)| count + 1))
+                        );
+                    }
+                }
+            }
+        }
+    }
+}
+
+#[tokio::test]
+async fn local_websocket_connection_cancellation_does_not_supply_capacity_evidence() {
+    let store = Arc::new(MemoryAccountStore::default());
+    let account_id = "acct_provider_contract";
+    create_account(&store, account_id).await;
+    let account = store.account(account_id).expect("account");
+    let cooldowns = Arc::new(MemoryCooldownPort::new());
+    let listener = TcpListener::bind("127.0.0.1:0").await.expect("listener");
+    let base_url = format!("http://{}", listener.local_addr().expect("address"));
+    let (provider, pool) =
+        provider_with_capacity_tracking(&store, base_url, Arc::clone(&cooldowns));
+    let operation = capacity_websocket_operation();
+    let mut stream = provider
+        .execute(
+            planned_request("openai", operation),
+            context("req_capacity_local", CancellationToken::new()),
+        )
+        .await
+        .expect("prepare WebSocket attempt");
+    let attempt = tokio::spawn(async move {
+        loop {
+            match stream.next().await {
+                Some(Ok(_)) => {}
+                Some(Err(error)) => break error,
+                None => panic!("cancelled opening must fail"),
+            }
+        }
+    });
+    let (mut opening, _) = timeout(Duration::from_secs(5), listener.accept())
+        .await
+        .expect("opening deadline")
+        .expect("opening");
+    read_http_request(&mut opening).await;
+    // 复现账号更新驱逐正在建连的连接，未收到任何上游容量拒绝。
+    pool.evict_account(account_id).await;
+    let error = timeout(Duration::from_secs(5), attempt)
+        .await
+        .expect("cancelled attempt deadline")
+        .expect("attempt task");
+    assert_eq!(error.kind(), ProviderErrorKind::Unavailable);
+    assert_eq!(error.send_state(), UpstreamSendState::NotSent);
+    assert_eq!(
+        error.diagnostic().and_then(|diagnostic| diagnostic.code()),
+        Some("shared_connect_failed")
+    );
+    assert_eq!(cooldowns.capacity_evidence(account.id()), None);
+    pool.shutdown().await;
+}
+
 #[tokio::test]
 async fn websocket_usage_limit_rejection_preserves_quota_state_for_account_rotation() {
     let store = Arc::new(MemoryAccountStore::default());
@@ -5578,7 +5800,12 @@ async fn websocket_usage_limit_rejection_preserves_quota_state_for_account_rotat
 
 #[tokio::test]
 async fn capacity_stream_rejection_only_retries_before_semantic_output_on_both_transports() {
-    for use_websocket in [false, true] {
+    for (use_websocket, code) in [
+        (false, "server_is_overloaded"),
+        (true, "server_is_overloaded"),
+        (false, "slow_down"),
+        (true, "slow_down"),
+    ] {
         for semantic_output in [false, true] {
             let store = Arc::new(MemoryAccountStore::default());
             create_account(&store, "acct_provider_contract").await;
@@ -5588,7 +5815,7 @@ async fn capacity_stream_rejection_only_retries_before_semantic_output_on_both_t
             if semantic_output {
                 events.push(json!({"type": "response.output_text.delta", "output_index": 0, "content_index": 0, "delta": "hello"}));
             }
-            let original = json!({"type": "response.failed", "response": {"id": "resp_capacity", "error": {"code": "server_is_overloaded", "message": "Selected model is at capacity. Please try a different model."}}});
+            let original = json!({"type": "response.failed", "response": {"id": "resp_capacity", "error": {"code": code, "message": "Selected model is at capacity. Please try a different model."}}});
             events.push(original.clone());
             let (base_url, _http_server, websocket_server) = if use_websocket {
                 let listener = TcpListener::bind("127.0.0.1:0").await.expect("listener");
@@ -5653,7 +5880,7 @@ async fn capacity_stream_rejection_only_retries_before_semantic_output_on_both_t
             assert_eq!(error.replay_is_safe(), !semantic_output);
             assert_eq!(error.kind(), ProviderErrorKind::UpstreamCapacityUnavailable);
             assert_eq!(error.pre_delivery_retry().is_some(), !semantic_output);
-            assert!(!provider_openai::openai_failure_affects_account_score(
+            assert!(provider_openai::openai_failure_affects_account_score(
                 &error
             ));
             assert_eq!(
diff --git a/backend/crates/providers/openai/tests/provider/failure.rs b/backend/crates/providers/openai/tests/provider/failure.rs
index e0ac0aa3a..cd8aea2c6 100644
--- a/backend/crates/providers/openai/tests/provider/failure.rs
+++ b/backend/crates/providers/openai/tests/provider/failure.rs
@@ -1,4 +1,6 @@
+use std::num::NonZeroU32;
 use std::sync::Arc;
+use std::time::Duration;
 
 use futures::{StreamExt, stream};
 use gateway_core::account::{AccountFeedbackStats, ProviderAccountId};
@@ -45,7 +47,7 @@ fn deliver_error_with_openai_feedback(
 }
 
 #[test]
-fn account_score_failure_filter_should_exclude_request_scoped_capacity_errors() {
+fn account_score_failure_filter_should_accept_capacity_errors_with_scored_reasons() {
     for code in ["server_is_overloaded", "slow_down", "server_error"] {
         let error = sent_error(ProviderErrorKind::Unavailable, Some(code))
             .with_client_visible_upstream_error(ClientVisibleUpstreamError::new(
@@ -53,13 +55,15 @@ fn account_score_failure_filter_should_exclude_request_scoped_capacity_errors()
                 Some("server_error".to_owned()),
                 Some("server_error".to_owned()),
             ));
-        assert!(!openai_failure_affects_account_score(&error));
+        assert!(openai_failure_affects_account_score(&error));
     }
 }
 
 #[test]
 fn account_score_failure_filter_should_accept_the_closed_reason_list() {
     for code in [
+        "server_is_overloaded",
+        "slow_down",
         "rate_limit_exceeded",
         "rate_limit_error",
         "server_error",
@@ -82,14 +86,20 @@ fn account_score_failure_filter_should_normalize_case_and_whitespace() {
 
 #[test]
 fn account_score_failure_filter_should_use_a_structured_type_when_code_is_absent() {
-    let error = ProviderError::new(ProviderErrorKind::Unavailable, UpstreamSendState::Sent)
-        .with_client_visible_upstream_error(ClientVisibleUpstreamError::new(
-            "upstream unavailable",
-            None,
-            Some("SERVICE_UNAVAILABLE_ERROR".to_owned()),
-        ));
+    for error_type in [
+        "SERVICE_UNAVAILABLE_ERROR",
+        "SERVER_IS_OVERLOADED",
+        "SLOW_DOWN",
+    ] {
+        let error = ProviderError::new(ProviderErrorKind::Unavailable, UpstreamSendState::Sent)
+            .with_client_visible_upstream_error(ClientVisibleUpstreamError::new(
+                "upstream unavailable",
+                None,
+                Some(error_type.to_owned()),
+            ));
 
-    assert!(openai_failure_affects_account_score(&error));
+        assert!(openai_failure_affects_account_score(&error));
+    }
 }
 
 #[test]
@@ -185,17 +195,27 @@ fn openai_feedback_should_not_amplify_repeated_client_configuration_errors() {
 }
 
 #[test]
-fn openai_feedback_should_not_penalize_an_account_for_server_overload() {
-    let feedback = Arc::new(AccountFeedbackStats::default());
-    let provider = ProviderKind::new("openai").expect("provider");
-    let account = ProviderAccountId::new("acct_overloaded").expect("account");
+fn openai_feedback_should_penalize_capacity_rejections_even_with_transient_retry() {
+    for code in ["server_is_overloaded", "slow_down"] {
+        let feedback = Arc::new(AccountFeedbackStats::default());
+        let provider = ProviderKind::new("openai").expect("provider");
+        let account = ProviderAccountId::new("acct_overloaded").expect("account");
+        let error = sent_error(ProviderErrorKind::UpstreamCapacityUnavailable, Some(code))
+            .with_replay_safe()
+            .with_transient_retry(
+                NonZeroU32::new(3).expect("retry count"),
+                Duration::from_millis(500),
+                Duration::from_secs(8),
+            );
 
-    deliver_error_with_openai_feedback(
-        &feedback,
-        &provider,
-        &account,
-        sent_error(ProviderErrorKind::Unavailable, Some("server_is_overloaded")),
-    );
+        deliver_error_with_openai_feedback(&feedback, &provider, &account, error);
 
-    assert_eq!(feedback.scheduling_signals(&provider, &account).0, None);
+        assert!(
+            feedback
+                .scheduling_signals(&provider, &account)
+                .0
+                .is_some_and(|failure_rate| failure_rate > 0),
+            "capacity rejection did not affect the account score: {code}"
+        );
+    }
 }
diff --git a/backend/crates/providers/openai/tests/support.rs b/backend/crates/providers/openai/tests/support.rs
index cb51a8bc6..18f8c9b98 100644
--- a/backend/crates/providers/openai/tests/support.rs
+++ b/backend/crates/providers/openai/tests/support.rs
@@ -20,11 +20,11 @@ use gateway_core::error::{StoreError, StoreErrorKind};
 use gateway_core::policy::ClientApiKeyId;
 use gateway_core::provider_ports::{
     ProviderCatalogCacheKey, ProviderCatalogCachePort, ProviderCooldown, ProviderCooldownPort,
-    ProviderCooldownScope, ProviderLeaseAcquisition, ProviderLeasePort, ProviderLeaseRequest,
-    ProviderRefreshPolicy, ProviderRuntimePolicyPort, ProviderSchedulingLeaseRequest,
-    ProviderSchedulingState, ProviderScopedCooldown, ProviderSessionAffinityKey,
-    ProviderSessionAffinityPort, ProviderSessionExclusionPort, ProviderSessionExclusions,
-    ProviderStoreError,
+    ProviderCooldownScope, ProviderFreezePolicy, ProviderLeaseAcquisition, ProviderLeasePort,
+    ProviderLeaseRequest, ProviderRefreshPolicy, ProviderRuntimePolicyPort,
+    ProviderSchedulingLeaseRequest, ProviderSchedulingState, ProviderScopedCooldown,
+    ProviderSessionAffinityKey, ProviderSessionAffinityPort, ProviderSessionExclusionPort,
+    ProviderSessionExclusions, ProviderStoreError,
 };
 use gateway_core::routing::ProviderKind;
 use provider_openai::credential::{
@@ -615,7 +615,7 @@ impl ProviderLeasePort for TestLeaseCoordinator {
                             last_started_at: None,
                             quota_reset_at: None,
                             quota_remaining_rank: None,
-                            rate_limited_until: None,
+                            cooldown: None,
                             failure_rate_basis_points: None,
                             first_output_latency_ms: None,
                         },
@@ -956,6 +956,36 @@ pub(crate) fn runtime_policy() -> Arc<dyn ProviderRuntimePolicyPort> {
     Arc::new(StaticRuntimePolicy)
 }
 
+/// 冻结策略可配置的运行时策略 fake；容量熔断触发测试用。
+pub(crate) struct StaticFreezePolicy(ProviderFreezePolicy);
+
+impl StaticFreezePolicy {
+    /// 构造带固定冻结策略的端口句柄；命名沿用测试构造器语义。
+    #[must_use]
+    pub(crate) fn policy_port(policy: ProviderFreezePolicy) -> Arc<dyn ProviderRuntimePolicyPort> {
+        Arc::new(Self(policy))
+    }
+}
+
+impl ProviderRuntimePolicyPort for StaticFreezePolicy {
+    fn load_refresh_policy(
+        &self,
+    ) -> BoxFuture<'_, Result<ProviderRefreshPolicy, ProviderStoreError>> {
+        Box::pin(async {
+            ProviderRefreshPolicy::try_new(
+                Duration::from_secs(60 * 60),
+                NonZeroU32::new(2).expect("positive concurrency"),
+            )
+        })
+    }
+
+    fn load_freeze_policy(
+        &self,
+    ) -> BoxFuture<'_, Result<ProviderFreezePolicy, ProviderStoreError>> {
+        Box::pin(async { Ok(self.0.clone()) })
+    }
+}
+
 pub(crate) fn secret(access_token: &str) -> CodexOAuthSecret {
     CodexOAuthSecret {
         access_token: SecretString::from(access_token.to_owned()),
@@ -972,11 +1002,13 @@ pub(crate) fn account_policy() -> gateway_core::account::AccountSelectionPolicy
     )
 }
 
-/// 内存 `ProviderCooldownPort`：只实现 `read`/`put_if_later`（openai selector
-/// 429 冷却路径用），其余 scope 变体测试不涉及，返回占位。
+/// 内存 `ProviderCooldownPort`：实现 `read`/`put_if_later` 与容量失败计数
+/// 证据（openai selector 429 冷却与容量熔断路径用），scope 变体测试不涉及，
+/// 返回占位。
 #[derive(Clone, Default)]
 pub(crate) struct MemoryCooldownPort {
     pub(crate) cooldowns: Arc<Mutex<BTreeMap<ProviderAccountId, ProviderCooldown>>>,
+    capacity_failures: Arc<Mutex<BTreeMap<ProviderAccountId, (u32, u32)>>>,
 }
 
 impl MemoryCooldownPort {
@@ -984,6 +1016,16 @@ impl MemoryCooldownPort {
     pub(crate) fn new() -> Self {
         Self::default()
     }
+
+    /// 测试断言用：窗口内累计失败次数与在途峰值。
+    #[must_use]
+    pub(crate) fn capacity_evidence(&self, account_id: &ProviderAccountId) -> Option<(u32, u32)> {
+        self.capacity_failures
+            .lock()
+            .expect("capacity failures lock")
+            .get(account_id)
+            .copied()
+    }
 }
 
 impl ProviderCooldownPort for MemoryCooldownPort {
@@ -1057,4 +1099,57 @@ impl ProviderCooldownPort for MemoryCooldownPort {
     ) -> BoxFuture<'a, Result<bool, ProviderStoreError>> {
         Box::pin(async { Ok(false) })
     }
+
+    fn record_capacity_failure<'a>(
+        &'a self,
+        account_id: &'a ProviderAccountId,
+        _window: Duration,
+        in_flight: u32,
+    ) -> BoxFuture<'a, Result<u32, ProviderStoreError>> {
+        Box::pin(async move {
+            let mut failures = self
+                .capacity_failures
+                .lock()
+                .expect("capacity failures lock");
+            let entry = failures.entry(account_id.clone()).or_insert((0, 0));
+            entry.0 = entry.0.saturating_add(1);
+            entry.1 = entry.1.max(in_flight);
+            Ok(entry.0)
+        })
+    }
+
+    fn clear_after_success<'a>(
+        &'a self,
+        account_id: &'a ProviderAccountId,
+        _through_revision: gateway_core::account::CredentialRevision,
+    ) -> BoxFuture<'a, Result<(), ProviderStoreError>> {
+        Box::pin(async move {
+            let mut cooldowns = self.cooldowns.lock().expect("cooldown lock");
+            if cooldowns.get(account_id).is_some_and(|value| {
+                value.kind().is_capacity_freeze() || value.credential_revision() > _through_revision
+            }) {
+                return Ok(());
+            }
+            cooldowns.remove(account_id);
+            self.capacity_failures
+                .lock()
+                .expect("capacity failures lock")
+                .remove(account_id);
+            Ok(())
+        })
+    }
+
+    fn capacity_peak_in_flight<'a>(
+        &'a self,
+        account_id: &'a ProviderAccountId,
+    ) -> BoxFuture<'a, Result<Option<u32>, ProviderStoreError>> {
+        Box::pin(async move {
+            Ok(self
+                .capacity_failures
+                .lock()
+                .expect("capacity failures lock")
+                .get(account_id)
+                .map(|(_, peak)| *peak))
+        })
+    }
 }
diff --git a/backend/crates/providers/xai/tests/admin.rs b/backend/crates/providers/xai/tests/admin.rs
index f5fad18ff..d15610aa1 100644
--- a/backend/crates/providers/xai/tests/admin.rs
+++ b/backend/crates/providers/xai/tests/admin.rs
@@ -666,7 +666,7 @@ impl ProviderLeasePort for TestLeases {
                             last_started_at: None,
                             quota_reset_at: None,
                             quota_remaining_rank: None,
-                            rate_limited_until: None,
+                            cooldown: None,
                             failure_rate_basis_points: None,
                             first_output_latency_ms: None,
                         },
@@ -862,6 +862,30 @@ impl ProviderCooldownPort for TestCooldown {
     ) -> BoxFuture<'a, Result<bool, ProviderStoreError>> {
         Box::pin(async { Ok(false) })
     }
+
+    fn record_capacity_failure<'a>(
+        &'a self,
+        _account_id: &'a ProviderAccountId,
+        _window: Duration,
+        _in_flight: u32,
+    ) -> BoxFuture<'a, Result<u32, ProviderStoreError>> {
+        Box::pin(async { Ok(0) })
+    }
+
+    fn clear_after_success<'a>(
+        &'a self,
+        _account_id: &'a ProviderAccountId,
+        _through_revision: gateway_core::account::CredentialRevision,
+    ) -> BoxFuture<'a, Result<(), ProviderStoreError>> {
+        Box::pin(async { Ok(()) })
+    }
+
+    fn capacity_peak_in_flight<'a>(
+        &'a self,
+        _account_id: &'a ProviderAccountId,
+    ) -> BoxFuture<'a, Result<Option<u32>, ProviderStoreError>> {
+        Box::pin(async { Ok(None) })
+    }
 }
 
 struct TestRuntimePolicy;
diff --git a/backend/crates/providers/xai/tests/credential/selector.rs b/backend/crates/providers/xai/tests/credential/selector.rs
index a673b2ced..e693b74f6 100644
--- a/backend/crates/providers/xai/tests/credential/selector.rs
+++ b/backend/crates/providers/xai/tests/credential/selector.rs
@@ -138,7 +138,7 @@ impl SelectorFixture {
                     last_started_at: None,
                     quota_reset_at: None,
                     quota_remaining_rank: None,
-                    rate_limited_until: None,
+                    cooldown: None,
                     failure_rate_basis_points: None,
                     first_output_latency_ms: None,
                 },
diff --git a/backend/crates/providers/xai/tests/provider/contract.rs b/backend/crates/providers/xai/tests/provider/contract.rs
index 44063d0f4..633b1d0b7 100644
--- a/backend/crates/providers/xai/tests/provider/contract.rs
+++ b/backend/crates/providers/xai/tests/provider/contract.rs
@@ -1065,7 +1065,7 @@ impl ProviderLeasePort for DiagnosticLeasePort {
                                 last_started_at: None,
                                 quota_reset_at: None,
                                 quota_remaining_rank: None,
-                                rate_limited_until: None,
+                                cooldown: None,
                                 failure_rate_basis_points: None,
                                 first_output_latency_ms: None,
                             },
diff --git a/backend/crates/providers/xai/tests/support.rs b/backend/crates/providers/xai/tests/support.rs
index 4c3c84ea4..2b7ff0c7a 100644
--- a/backend/crates/providers/xai/tests/support.rs
+++ b/backend/crates/providers/xai/tests/support.rs
@@ -972,6 +972,30 @@ impl ProviderCooldownPort for MemoryCooldownPort {
             Ok(account_removed || scoped_removed)
         })
     }
+
+    fn record_capacity_failure<'a>(
+        &'a self,
+        _account_id: &'a ProviderAccountId,
+        _window: Duration,
+        _in_flight: u32,
+    ) -> futures::future::BoxFuture<'a, Result<u32, ProviderStoreError>> {
+        Box::pin(async { Ok(0) })
+    }
+
+    fn clear_after_success<'a>(
+        &'a self,
+        _account_id: &'a ProviderAccountId,
+        _through_revision: gateway_core::account::CredentialRevision,
+    ) -> futures::future::BoxFuture<'a, Result<(), ProviderStoreError>> {
+        Box::pin(async { Ok(()) })
+    }
+
+    fn capacity_peak_in_flight<'a>(
+        &'a self,
+        _account_id: &'a ProviderAccountId,
+    ) -> futures::future::BoxFuture<'a, Result<Option<u32>, ProviderStoreError>> {
+        Box::pin(async { Ok(None) })
+    }
 }
 
 impl ProviderRuntimePolicyPort for StaticRuntimePolicy {
diff --git a/backend/migrations/.frozen-sha256 b/backend/migrations/.frozen-sha256
index 1af694aee..e59e5c8d0 100644
--- a/backend/migrations/.frozen-sha256
+++ b/backend/migrations/.frozen-sha256
@@ -7,3 +7,5 @@ b7d3df7b48d835a00a3b74cbefc958a039d38fea5529cdad04bfbc703dca72aa  0001_initial.s
 18e8c1b71c896a9d576d9e4ef9701506f654f55383af908898195b9eb2980fde  0007_account_notes.sql
 36b720c9d3e6bf0bdb799b9b6a89b81638c642ac866992a880bd82357012a208  0008_concurrency_queues.sql
 161df9ae43da34d51b340bbf95d71c61a8c1a2aa142382f693a9d127c834630d  0009_account_model_access.sql
+dd4aa34e9bcf1698bb9688421964bbf410314d8e905448b8229460b972cff2f6  0010_account_auto_freeze.sql
+9bd261b7e94872126754eb7f749c56187264a47c567b0bef5c394d6776b4f463  0011_account_auto_freeze_opt_in.sql
diff --git a/backend/migrations/0010_account_auto_freeze.sql b/backend/migrations/0010_account_auto_freeze.sql
new file mode 100644
index 000000000..9abaa3c9d
--- /dev/null
+++ b/backend/migrations/0010_account_auto_freeze.sql
@@ -0,0 +1,22 @@
+-- 账号容量熔断：统计窗口内高频出现容量类上游错误时自动冻结账号一段时间；
+-- 恢复前可选执行一次探测调用验证上游已恢复；冻结期间按观测到的在途并发峰值
+-- 下调账号并发上限，让账号带着更安全的并发配置自愈。默认启用，已有实例可关闭。
+alter table runtime_settings
+    add column account_auto_freeze_enabled boolean not null default true,
+    add column account_auto_freeze_threshold bigint not null default 12
+        check (account_auto_freeze_threshold between 2 and 1000),
+    add column account_auto_freeze_window_seconds bigint not null default 600
+        check (account_auto_freeze_window_seconds between 60 and 3600),
+    add column account_auto_freeze_duration_seconds bigint not null default 7200
+        check (account_auto_freeze_duration_seconds between 300 and 604800),
+    add column account_auto_freeze_probe_enabled boolean not null default true,
+    add column account_auto_freeze_probe_model text
+        check (
+          account_auto_freeze_probe_model is null
+          or (
+            octet_length(account_auto_freeze_probe_model) between 1 and 128
+            and account_auto_freeze_probe_model = btrim(account_auto_freeze_probe_model)
+            and account_auto_freeze_probe_model !~ '[[:cntrl:]]'
+          )
+        ),
+    add column account_auto_freeze_adaptive_concurrency boolean not null default true;
diff --git a/backend/migrations/0011_account_auto_freeze_opt_in.sql b/backend/migrations/0011_account_auto_freeze_opt_in.sql
new file mode 100644
index 000000000..70f0dcb1b
--- /dev/null
+++ b/backend/migrations/0011_account_auto_freeze_opt_in.sql
@@ -0,0 +1,12 @@
+-- 自动冻结默认关闭；保留已经通过设置页显式保存的冻结配置。
+alter table runtime_settings
+    alter column account_auto_freeze_enabled set default false;
+
+update runtime_settings
+set account_auto_freeze_enabled = false
+where account_auto_freeze_enabled = true
+  and not exists (
+    select 1 from admin_audit_events
+    where action = 'settings.replace'
+      and changed_fields @> array['account_auto_freeze']::text[]
+  );
diff --git a/docs/api.md b/docs/api.md
index b885c942d..131812788 100644
--- a/docs/api.md
+++ b/docs/api.md
@@ -253,7 +253,10 @@ Responses wire 之间的协议转换层，转换只在 xAI Provider 内完成。
 OpenAI 明确返回 `server_is_overloaded`、`slow_down` 或模型容量不足错误时，代理在允许安全重放且
 尚未交付输出的前提下，先做最多 3 次同账号指数退避，再通过现有调度换号。默认间隔从 500ms 开始，
 上游 `Retry-After` 参与退避计算，单次等待不超过 8 秒；重试同时受请求总尝试次数和截止时间约束。
-容量不足不扣 Smart 账号健康分，不触发 Provider 全局熔断，也不作为账号额度耗尽写入冷却状态。
+`server_is_overloaded`、`slow_down` 与既有可计分的结构化错误按已发送的失败尝试计入 Smart 账号
+健康分；容量分类不再豁免计分。失败率沿用账号级平滑与时间衰减，影响后续普通选路，已有可用账号的
+会话亲和仍优先。容量不足不触发 Provider 全局熔断，也不作为账号额度耗尽；启用账号自动冻结时，
+达到容量失败阈值会另外写入临时冷却。
 最终交付的上游错误仍按上述透明边界保留原始状态码、错误码和正文。
 明确额度耗尽继续走现有账号隔离与安全换号流程，
 包括 WebSocket 握手返回的 429；不会因其长 `Retry-After` 而转入同账号传输恢复等待。
@@ -367,6 +370,11 @@ Token 明细、费用明细、用时/首字与状态。Token 和费用复用现
 - `sortBy`: `email`、`status`、`planType`、`usage`、`lastUsedAt`、`expiresAt`；
 - `sortDirection`: `asc`、`desc`。
 
+账号限流详情在 `quota` 中返回：`rateLimitReason` 为 `upstream_rate_limit`（上游临时限流）、
+`capacity_freeze`（容量错误触发自动冻结）或 `null`。`recoveryProbeRequired` 表示解除冻结是否需要成功探测；
+此时 `rateLimitedUntil` 是最早探测时间，到期后仍保持 `rate_limited`，直到探测成功或手动恢复。
+未要求探测时，该字段表示冷却结束时间。所有此类情况统一显示“限流中”，仅详情原因和恢复条件不同。
+
 账号列表和详情返回 `notes`（无备注时为 `null`）。编辑时省略或 `null` 保留原备注；字符串最多 500 个 Unicode
 字符，允许换行和制表符，保存时去除首尾空白，空字符串清空备注。备注独立于上游身份，导入时未显式提供备注、
 重新授权、凭据刷新及批量调度更新均保留已有备注。
@@ -910,6 +918,13 @@ minCodexCliVersion
 usageRetentionDays
 opsEventRetentionDays
 auditRetentionDays
+accountAutoFreezeEnabled
+accountAutoFreezeThreshold
+accountAutoFreezeWindowSeconds
+accountAutoFreezeDurationSeconds
+accountAutoFreezeProbeEnabled
+accountAutoFreezeProbeModel
+accountAutoFreezeAdaptiveConcurrency
 ```
 
 `maxWaitingPerKey` 与 `maxWaitingPerAccount` 是全局统一的排队容量，取值 0～1,000，默认 0（关闭）；
@@ -922,6 +937,17 @@ auditRetentionDays
 `rotationStrategy` 可取 `smart`、`quota_reset_priority`、`round_robin`、`sticky`。
 两个 `minCodex*Version` 字段为 `string | null`，只设置最低版本，不存在最大版本字段。
 
+账号自动冻结（`accountAutoFreezeEnabled`）默认关闭。启用后，在统计窗口内按尝试累计容量类上游错误（`server_is_overloaded`
+等与 5xx 不可用），达到阈值后把该账号冻结为带恢复倒计时的 `rate_limited` 状态。`accountAutoFreezeThreshold`
+取值 2～1,000（默认 12，按普通请求的 attempt 计数，含请求内同账号重试，不含诊断探测与本地连接保护错误）；`accountAutoFreezeWindowSeconds`
+取值 60～3,600（默认 600，随每次失败滑动顺延）；`accountAutoFreezeDurationSeconds` 取值 300～604,800
+（默认 7,200，即 2 小时，探测失败后按该时长顺延）。`accountAutoFreezeProbeEnabled` 开启时恢复 worker
+在到期后执行真实探测调用，成功才解冻；探测过程中和服务重启后继续阻止普通请求。关闭自动冻结或探测后，
+已有冻结等待当前冷却结束再恢复调度。`accountAutoFreezeProbeModel` 为 `string | null`，留空时自动
+选择账号可用的第一个模型。`accountAutoFreezeAdaptiveConcurrency` 开启时冻结期间把账号并发上限下调到
+观测在途峰值的 80%（下限 2，只降不升）。这会持久修改账号并发设置；跟随全局默认的账号也会设为独立上限，
+解冻后不自动恢复，管理员可手动改回。
+
 Windows 离线包接口固定解析 Microsoft Store Product ID `9PLM9XGG6VKS` 的 Retail 包，不接受调用方提供
 产品 ID、上游地址、ring 或文件名。后端只返回通过包名、架构、Microsoft CDN host/path、scheme 和失效
 时间校验的 `x64` / `arm64` MSIX 直链，不代理安装包字节。Store 内容通道返回 HTTP/80 临时地址时保留
diff --git a/docs/architecture.md b/docs/architecture.md
index 455823a67..0d7428006 100644
--- a/docs/architecture.md
+++ b/docs/architecture.md
@@ -422,6 +422,10 @@ PostgreSQL 周期对账才是正确性基础。
 `normal`、`quota_exhausted`、`rate_limited`、`disabled`、`error`。只有明确上游证据才能恢复或终态化账号，
 本地时钟和不确定响应不能伪造事实。
 
+容量冻结与普通上游限流共享账号冷却投影，但恢复条件不同：需要探测的冻结在截止时间后仍阻止调度，
+普通推理成功不能解除。Admin 恢复任务按冻结代次原子提交探测结果，避免旧结果覆盖手动恢复或新冻结；
+自动降低并发通过现有管理事务与快照发布链路，仅按数据库最新设置更新并发上限。
+
 PostgreSQL schema 由迁移目录按编号管理。已应用迁移按字节冻结，后续 schema 变化只能新增编号迁移，
 详见 [迁移规则](../backend/migrations/README.md)。
 
@@ -497,8 +501,19 @@ Worker 由各 Bundle 贡献、由 Host 统一监督：
 - Store：过期请求恢复、历史保留和 PostgreSQL/Redis 观测队列；
 - Core：`runtime` owner 的 RuntimeSnapshot 周期对账和 Redis change 订阅；
 - Admin：S3/R2 备份 daemon，负责调度、执行、删除收敛与保留清理；
+  以及账号冻结恢复 worker（容量熔断的自适应并发下调与到期探测解冻）；
 - Provider：credential refresh、quota/catalog 健康和官方版本/etag 检查。
 
+账号容量熔断默认关闭。启用后，仅普通请求收到的容量类上游错误（`server_is_overloaded` 等与
+5xx 不可用）按滑动窗口计数，并把当时观测到的在途并发并入峰值证据；本地连接保护与诊断探测
+不参与容量计数或峰值采样。这些错误不证明凭据或配额失效，不进入账号失败状态。达到阈值后写入
+账号级 Redis 容量冷却，调度立即跳过该账号；管理端沿用限流状态，通过原因区分容量冻结与上游限流。
+冻结与计数由 Redis 保存，不改变 PostgreSQL 账号状态。需要探测的冻结在到期后仍阻止调度，恢复
+worker 复用连接测试探针执行真实上游调用，并按冻结代次处理结果：成功清除，失败按配置时长顺延，
+过期探测结果不得覆盖更新的冻结或管理员操作。关闭自动冻结或探测后，已有冻结在冷却到期后恢复。
+自适应并发下调按观测峰值的 80%（下限 2）原子更新最新有效账号上限，只降不升，不覆盖其他账号
+设置，审计标注为系统变更；该上限持久保存，恢复调度后不自动调高。
+
 周期任务的 Redis lease 只保证单周期互斥，不构成多副本 leader 选举。备份 daemon 依赖单副本部署边界，
 自更新也只替换处理请求的当前进程，因此整个应用必须保持单副本。
 
diff --git a/frontend/src/api/modules/accounts.ts b/frontend/src/api/modules/accounts.ts
index 4f2a0c40b..7dafccf77 100644
--- a/frontend/src/api/modules/accounts.ts
+++ b/frontend/src/api/modules/accounts.ts
@@ -33,6 +33,8 @@ export interface AccountQuota {
   limitReached: boolean
   // 429 临时限流（Redis 冷却）到期时间；非限流中为 null。
   rateLimitedUntil: string | null
+  rateLimitReason: 'upstream_rate_limit' | 'capacity_freeze' | null
+  recoveryProbeRequired: boolean
   windows: AccountQuotaWindow[]
 }
 
diff --git a/frontend/src/api/modules/settings.ts b/frontend/src/api/modules/settings.ts
index 437409e1e..34ff6a17e 100644
--- a/frontend/src/api/modules/settings.ts
+++ b/frontend/src/api/modules/settings.ts
@@ -18,6 +18,13 @@ export interface RuntimeSettings {
   usageRetentionDays: number
   opsEventRetentionDays: number
   auditRetentionDays: number
+  accountAutoFreezeEnabled: boolean
+  accountAutoFreezeThreshold: number
+  accountAutoFreezeWindowSeconds: number
+  accountAutoFreezeDurationSeconds: number
+  accountAutoFreezeProbeEnabled: boolean
+  accountAutoFreezeProbeModel: string | null
+  accountAutoFreezeAdaptiveConcurrency: boolean
   updatedAt: string
 }
 
diff --git a/frontend/src/views/accounts/components/AccountConnectionTestModal.vue b/frontend/src/views/accounts/components/AccountConnectionTestModal.vue
index 7bd3bf163..5519aba47 100644
--- a/frontend/src/views/accounts/components/AccountConnectionTestModal.vue
+++ b/frontend/src/views/accounts/components/AccountConnectionTestModal.vue
@@ -62,6 +62,9 @@ function connectionLogClass(tone: string) {
           :status="account.status"
           :error-reason="account.errorReason"
           :error-message="account.errorMessage"
+          :rate-limited-until="account.quota.rateLimitedUntil"
+          :rate-limit-reason="account.quota.rateLimitReason"
+          :recovery-probe-required="account.quota.recoveryProbeRequired"
           :next-refresh-at="account.nextRefreshAt"
           variant="pill"
         />
diff --git a/frontend/src/views/accounts/components/AccountStatusBadge/index.vue b/frontend/src/views/accounts/components/AccountStatusBadge/index.vue
index e47d4f96c..ea5e94f40 100644
--- a/frontend/src/views/accounts/components/AccountStatusBadge/index.vue
+++ b/frontend/src/views/accounts/components/AccountStatusBadge/index.vue
@@ -13,6 +13,8 @@ const props = withDefaults(
     errorReason?: AccountErrorReason | null
     errorMessage?: string | null
     rateLimitedUntil?: string | null
+    rateLimitReason?: 'upstream_rate_limit' | 'capacity_freeze' | null
+    recoveryProbeRequired?: boolean
     nextRefreshAt?: string | null
     variant?: 'inline' | 'pill'
   }>(),
@@ -20,6 +22,8 @@ const props = withDefaults(
     errorReason: null,
     errorMessage: null,
     rateLimitedUntil: null,
+    rateLimitReason: null,
+    recoveryProbeRequired: false,
     nextRefreshAt: null,
     variant: 'inline',
   },
@@ -32,6 +36,8 @@ const presentation = computed(() =>
     errorReason: props.errorReason,
     errorMessage: props.errorMessage,
     rateLimitedUntil: props.rateLimitedUntil,
+    rateLimitReason: props.rateLimitReason,
+    recoveryProbeRequired: props.recoveryProbeRequired,
     nextRefreshAt: props.nextRefreshAt,
     now: now.value.getTime(),
   }),
@@ -102,7 +108,7 @@ const presentation = computed(() =>
           v-if="presentation.rateLimitRecovery"
           class="flex items-center justify-between gap-3 rounded-cp bg-cp-fill-quaternary px-3 py-2"
         >
-          <span class="font-heavy text-cp-text-tertiary">预计恢复</span>
+          <span class="font-heavy text-cp-text-tertiary">{{ presentation.recoveryTimeLabel }}</span>
           <span class="font-mono font-emphasis tabular-nums text-cp-text">
             {{ presentation.rateLimitRecovery }}
           </span>
diff --git a/frontend/src/views/accounts/components/AccountStatusBadge/presenter.ts b/frontend/src/views/accounts/components/AccountStatusBadge/presenter.ts
index 37fd95a92..c95ae5f31 100644
--- a/frontend/src/views/accounts/components/AccountStatusBadge/presenter.ts
+++ b/frontend/src/views/accounts/components/AccountStatusBadge/presenter.ts
@@ -37,6 +37,7 @@ export interface AccountStatusPresentation {
   errorText: string | null
   nextRefreshDisplay: string | null
   rateLimitRecovery: string | null
+  recoveryTimeLabel: string
   triggerLabel: string
 }
 
@@ -45,6 +46,8 @@ export interface AccountStatusPresentationInput {
   errorReason: AccountErrorReason | null
   errorMessage: string | null
   rateLimitedUntil: string | null
+  rateLimitReason: 'upstream_rate_limit' | 'capacity_freeze' | null
+  recoveryProbeRequired: boolean
   nextRefreshAt: string | null
   now: number
 }
@@ -146,19 +149,32 @@ export function resolveAccountStatusPresentation(
     && nextRefreshTimestamp !== null
     && nextRefreshTimestamp > input.now
   const mode: AccountStatusDisplayMode = isBackoff ? 'refresh_backoff' : input.status
-  const definition = displayDefinitions[mode]
+  const isFreeze = mode === 'rate_limited' && input.rateLimitReason === 'capacity_freeze'
+  const waitsForProbe = isFreeze && input.recoveryProbeRequired
+  const definition = isFreeze
+    ? {
+        ...displayDefinitions.rate_limited,
+        description: '容量类请求失败累计达到阈值，系统已暂停该账号的调度。',
+        recoveryHint: waitsForProbe
+          ? '冷却结束后进行恢复探测，成功后恢复调度；也可手动恢复账号。'
+          : '冷却结束后自动恢复调度，也可手动恢复账号。',
+      }
+    : displayDefinitions[mode]
   const nextRefreshDisplay = isBackoff ? formatDateTime(nextRefreshTimestamp) : null
   const reasonLabel = input.errorReason ? errorReasonLabels[input.errorReason] : null
   const title = definition.title ?? reasonLabel ?? definition.label
-  const rateLimitRecovery = input.status === 'rate_limited'
-    ? remainingTime(input.rateLimitedUntil, input.now)
+  const until = input.rateLimitedUntil ? parseTimestamp(input.rateLimitedUntil) : null
+  const rateLimitRecovery = mode === 'rate_limited'
+    ? waitsForProbe && until !== null && until <= input.now
+      ? '等待探测成功'
+      : remainingTime(input.rateLimitedUntil, input.now)
     : null
   const recoveryHint = mode !== 'refresh_backoff'
     && mode !== 'rate_limited'
     && input.errorReason
     ? errorRecoveryHints[input.errorReason]
     : definition.recoveryHint
-  const hasDetail = input.status === 'error' || rateLimitRecovery !== null || isBackoff
+  const hasDetail = input.status === 'error' || mode === 'rate_limited' || isBackoff
   const retry = nextRefreshDisplay ? ` 下次尝试：${nextRefreshDisplay}。` : ''
   const triggerLabel = `${title}。${definition.description}${retry} 点击或聚焦查看详情。`
 
@@ -174,6 +190,7 @@ export function resolveAccountStatusPresentation(
     errorText: input.errorMessage || null,
     nextRefreshDisplay,
     rateLimitRecovery,
+    recoveryTimeLabel: waitsForProbe ? '恢复探测' : '预计恢复',
     triggerLabel,
   }
 }
diff --git a/frontend/src/views/accounts/index.vue b/frontend/src/views/accounts/index.vue
index 0e7f6d1b1..daa70104d 100644
--- a/frontend/src/views/accounts/index.vue
+++ b/frontend/src/views/accounts/index.vue
@@ -291,6 +291,8 @@ const {
                 :error-reason="row.errorReason"
                 :error-message="row.errorMessage"
                 :rate-limited-until="row.quota.rateLimitedUntil"
+                :rate-limit-reason="row.quota.rateLimitReason"
+                :recovery-probe-required="row.quota.recoveryProbeRequired"
                 :next-refresh-at="row.nextRefreshAt"
               />
             </template>
diff --git a/frontend/src/views/settings/components/AccountAutoFreezeCard.vue b/frontend/src/views/settings/components/AccountAutoFreezeCard.vue
new file mode 100644
index 000000000..bc838043a
--- /dev/null
+++ b/frontend/src/views/settings/components/AccountAutoFreezeCard.vue
@@ -0,0 +1,135 @@
+<script setup lang="ts">
+import { Activity, Gauge, Snowflake, Timer } from '@lucide/vue'
+
+import BaseCard from '@/components/base/BaseCard.vue'
+import BaseCheckbox from '@/components/base/BaseCheckbox.vue'
+import BaseFormItem from '@/components/base/BaseForm/FormItem.vue'
+import BaseForm from '@/components/base/BaseForm/index.vue'
+import BaseInput from '@/components/base/BaseInput.vue'
+import BaseSwitch from '@/components/base/BaseSwitch.vue'
+
+const enabled = defineModel<boolean>('enabled', { required: true })
+const threshold = defineModel<string>('threshold', { required: true })
+const windowSeconds = defineModel<string>('windowSeconds', { required: true })
+const durationSeconds = defineModel<string>('durationSeconds', { required: true })
+const probeEnabled = defineModel<boolean>('probeEnabled', { required: true })
+const probeModel = defineModel<string>('probeModel', { required: true })
+const adaptiveConcurrency = defineModel<boolean>('adaptiveConcurrency', { required: true })
+</script>
+
+<template>
+  <BaseCard
+    title="账号自动冻结"
+    description="容量类错误高频出现时暂停账号调度，在账号管理中显示为限流中"
+  >
+    <BaseForm class="max-w-6xl sm:grid-cols-2">
+      <BaseSwitch
+        v-model="enabled"
+        class="col-span-full justify-self-start"
+        label="启用自动冻结"
+        show-label
+      />
+
+      <BaseFormItem
+        label="触发阈值"
+        description="窗口内累计失败尝试达到此次数后触发"
+      >
+        <BaseInput
+          v-model="threshold"
+          :disabled="!enabled"
+          aria-label="触发阈值"
+          type="number"
+          min="2"
+          max="1000"
+          step="1"
+        >
+          <template #prefix>
+            <Activity class="size-4" />
+          </template>
+          <template #suffix>
+            <span class="text-cp-sm">次</span>
+          </template>
+        </BaseInput>
+      </BaseFormItem>
+
+      <BaseFormItem
+        label="统计窗口"
+        description="每次失败后，计数有效期向后顺延"
+      >
+        <BaseInput
+          v-model="windowSeconds"
+          :disabled="!enabled"
+          aria-label="统计窗口秒数"
+          type="number"
+          min="60"
+          max="3600"
+          step="1"
+        >
+          <template #prefix>
+            <Timer class="size-4" />
+          </template>
+          <template #suffix>
+            <span class="text-cp-sm">秒</span>
+          </template>
+        </BaseInput>
+      </BaseFormItem>
+
+      <BaseFormItem
+        label="冷却时长"
+        description="探测失败时，按此时长延后恢复"
+      >
+        <BaseInput
+          v-model="durationSeconds"
+          :disabled="!enabled"
+          aria-label="冷却时长秒数"
+          type="number"
+          min="300"
+          max="604800"
+          step="1"
+        >
+          <template #prefix>
+            <Snowflake class="size-4" />
+          </template>
+          <template #suffix>
+            <span class="text-cp-sm">秒</span>
+          </template>
+        </BaseInput>
+      </BaseFormItem>
+      <BaseFormItem
+        label="恢复探测模型"
+        description="启用后需探测成功才恢复，模型留空时选择首个可用模型"
+      >
+        <template #extra>
+          <BaseCheckbox
+            v-model="probeEnabled"
+            :disabled="!enabled"
+            label="恢复前探测"
+            show-label
+          />
+        </template>
+        <BaseInput
+          v-model="probeModel"
+          :disabled="!enabled || !probeEnabled"
+          aria-label="恢复探测模型"
+          placeholder="留空自动选择"
+        >
+          <template #prefix>
+            <Gauge class="size-4" />
+          </template>
+        </BaseInput>
+      </BaseFormItem>
+
+      <div class="col-span-full flex flex-wrap items-center gap-x-3 gap-y-2">
+        <BaseCheckbox
+          v-model="adaptiveConcurrency"
+          :disabled="!enabled"
+          label="自适应并发下调"
+          show-label
+        />
+        <p class="m-0 text-cp-sm leading-5 text-cp-text-secondary">
+          修改账号并发上限，恢复后不自动调高
+        </p>
+      </div>
+    </BaseForm>
+  </BaseCard>
+</template>
diff --git a/frontend/src/views/settings/composables/useSettingsForm.ts b/frontend/src/views/settings/composables/useSettingsForm.ts
index d78dd1bfe..89982f04f 100644
--- a/frontend/src/views/settings/composables/useSettingsForm.ts
+++ b/frontend/src/views/settings/composables/useSettingsForm.ts
@@ -30,9 +30,17 @@ export function useSettingsForm() {
     usageRetentionDays: 31,
     opsEventRetentionDays: 30,
     auditRetentionDays: 90,
+
+    accountAutoFreezeEnabled: false,
+    accountAutoFreezeThreshold: null as number | null,
+    accountAutoFreezeWindowSeconds: null as number | null,
+    accountAutoFreezeDurationSeconds: null as number | null,
+    accountAutoFreezeProbeEnabled: true,
+    accountAutoFreezeProbeModel: '',
+    accountAutoFreezeAdaptiveConcurrency: true,
   })
 
-  function numericModel(key: 'refreshMarginSeconds' | 'refreshConcurrency' | 'maxConcurrentPerAccount' | 'requestIntervalMs' | 'maxWaitingPerKey' | 'maxWaitingPerAccount' | 'concurrencyWaitTimeoutSeconds') {
+  function numericModel(key: 'refreshMarginSeconds' | 'refreshConcurrency' | 'maxConcurrentPerAccount' | 'requestIntervalMs' | 'maxWaitingPerKey' | 'maxWaitingPerAccount' | 'concurrencyWaitTimeoutSeconds' | 'accountAutoFreezeThreshold' | 'accountAutoFreezeWindowSeconds' | 'accountAutoFreezeDurationSeconds') {
     return computed({
       get: () => (form[key] === null ? '' : String(form[key])),
       set: (value: string) => {
@@ -53,6 +61,9 @@ export function useSettingsForm() {
   const maxWaitingPerKeyValue = numericModel('maxWaitingPerKey')
   const maxWaitingPerAccountValue = numericModel('maxWaitingPerAccount')
   const concurrencyWaitTimeoutSecondsValue = numericModel('concurrencyWaitTimeoutSeconds')
+  const accountAutoFreezeThresholdValue = numericModel('accountAutoFreezeThreshold')
+  const accountAutoFreezeWindowSecondsValue = numericModel('accountAutoFreezeWindowSeconds')
+  const accountAutoFreezeDurationSecondsValue = numericModel('accountAutoFreezeDurationSeconds')
 
   const minCodexDesktopVersionError = computed(() => versionError(form.minCodexDesktopVersion))
   const minCodexCliVersionError = computed(() => versionError(form.minCodexCliVersion))
@@ -77,6 +88,13 @@ export function useSettingsForm() {
     form.usageRetentionDays = data.usageRetentionDays
     form.opsEventRetentionDays = data.opsEventRetentionDays
     form.auditRetentionDays = data.auditRetentionDays
+    form.accountAutoFreezeEnabled = data.accountAutoFreezeEnabled
+    form.accountAutoFreezeThreshold = data.accountAutoFreezeThreshold
+    form.accountAutoFreezeWindowSeconds = data.accountAutoFreezeWindowSeconds
+    form.accountAutoFreezeDurationSeconds = data.accountAutoFreezeDurationSeconds
+    form.accountAutoFreezeProbeEnabled = data.accountAutoFreezeProbeEnabled
+    form.accountAutoFreezeProbeModel = data.accountAutoFreezeProbeModel ?? ''
+    form.accountAutoFreezeAdaptiveConcurrency = data.accountAutoFreezeAdaptiveConcurrency
     mappings.value = Object.entries(data.modelMappings || {}).map(([requestedModel, upstreamModel]) => ({
       requestedModel,
       upstreamModel: String(upstreamModel),
@@ -132,7 +150,7 @@ export function useSettingsForm() {
   async function saveSettings() {
     if (saving.value || loading.value)
       return
-    const { refreshMarginSeconds, refreshConcurrency, maxConcurrentPerAccount, requestIntervalMs, rotationStrategy, maxWaitingPerKey, maxWaitingPerAccount, concurrencyWaitTimeoutSeconds } = form
+    const { refreshMarginSeconds, refreshConcurrency, maxConcurrentPerAccount, requestIntervalMs, rotationStrategy, maxWaitingPerKey, maxWaitingPerAccount, concurrencyWaitTimeoutSeconds, accountAutoFreezeThreshold, accountAutoFreezeWindowSeconds, accountAutoFreezeDurationSeconds } = form
     if (refreshMarginSeconds === null || refreshConcurrency === null || maxConcurrentPerAccount === null || requestIntervalMs === null || !rotationStrategy || maxWaitingPerKey === null || maxWaitingPerAccount === null || concurrencyWaitTimeoutSeconds === null) {
       toast.warning('请完整填写运行参数和调度策略')
       return
@@ -146,6 +164,21 @@ export function useSettingsForm() {
       toast.warning('请修正客户端最低版本格式')
       return
     }
+    if (accountAutoFreezeThreshold === null || accountAutoFreezeWindowSeconds === null || accountAutoFreezeDurationSeconds === null) {
+      toast.warning('请完整填写账号自动冻结参数')
+      return
+    }
+    if (!Number.isInteger(accountAutoFreezeThreshold) || accountAutoFreezeThreshold < 2 || accountAutoFreezeThreshold > 1000
+      || !Number.isInteger(accountAutoFreezeWindowSeconds) || accountAutoFreezeWindowSeconds < 60 || accountAutoFreezeWindowSeconds > 3600
+      || !Number.isInteger(accountAutoFreezeDurationSeconds) || accountAutoFreezeDurationSeconds < 300 || accountAutoFreezeDurationSeconds > 604800) {
+      toast.warning('自动冻结阈值应为 2～1000，统计窗口为 60～3600 秒，冻结时长为 300～604800 秒')
+      return
+    }
+    const probeModel = form.accountAutoFreezeProbeModel.trim()
+    if (probeModel && (probeModel.length > 128 || probeModel !== probeModel.trim())) {
+      toast.warning('探测模型名称不能超过 128 个字符')
+      return
+    }
     await saveAction.run(async () => {
       const result = await updateSettings({
         modelMappings: mappingPayload(),
@@ -162,6 +195,13 @@ export function useSettingsForm() {
         usageRetentionDays: form.usageRetentionDays,
         opsEventRetentionDays: form.opsEventRetentionDays,
         auditRetentionDays: form.auditRetentionDays,
+        accountAutoFreezeEnabled: form.accountAutoFreezeEnabled,
+        accountAutoFreezeThreshold,
+        accountAutoFreezeWindowSeconds,
+        accountAutoFreezeDurationSeconds,
+        accountAutoFreezeProbeEnabled: form.accountAutoFreezeProbeEnabled,
+        accountAutoFreezeProbeModel: probeModel || null,
+        accountAutoFreezeAdaptiveConcurrency: form.accountAutoFreezeAdaptiveConcurrency,
       })
       applySettings(result)
       toast.success('设置已保存')
@@ -189,6 +229,9 @@ export function useSettingsForm() {
     maxWaitingPerKeyValue,
     maxWaitingPerAccountValue,
     concurrencyWaitTimeoutSecondsValue,
+    accountAutoFreezeThresholdValue,
+    accountAutoFreezeWindowSecondsValue,
+    accountAutoFreezeDurationSecondsValue,
     minCodexDesktopVersionError,
     minCodexCliVersionError,
     saveSettings,
diff --git a/frontend/src/views/settings/index.vue b/frontend/src/views/settings/index.vue
index 5a7197990..d17488ef0 100644
--- a/frontend/src/views/settings/index.vue
+++ b/frontend/src/views/settings/index.vue
@@ -8,6 +8,7 @@ import BaseConfirmModal from '@/components/base/BaseConfirmModal.vue'
 import BasePageHeader from '@/components/base/BasePageHeader.vue'
 import BaseSegmented from '@/components/base/BaseSegmented.vue'
 
+import AccountAutoFreezeCard from './components/AccountAutoFreezeCard.vue'
 import AdminApiKeyCard from './components/AdminApiKeyCard.vue'
 import SettingsBackupSection from './components/backup/SettingsBackupSection.vue'
 import ClientVersionSettings from './components/client-version/index.vue'
@@ -52,6 +53,9 @@ const {
   maxWaitingPerKeyValue,
   maxWaitingPerAccountValue,
   concurrencyWaitTimeoutSecondsValue,
+  accountAutoFreezeThresholdValue,
+  accountAutoFreezeWindowSecondsValue,
+  accountAutoFreezeDurationSecondsValue,
 
   minCodexDesktopVersionError,
   minCodexCliVersionError,
@@ -136,6 +140,16 @@ watch(
         v-model:concurrency-wait-timeout-seconds="concurrencyWaitTimeoutSecondsValue"
       />
 
+      <AccountAutoFreezeCard
+        v-model:enabled="form.accountAutoFreezeEnabled"
+        v-model:threshold="accountAutoFreezeThresholdValue"
+        v-model:window-seconds="accountAutoFreezeWindowSecondsValue"
+        v-model:duration-seconds="accountAutoFreezeDurationSecondsValue"
+        v-model:probe-enabled="form.accountAutoFreezeProbeEnabled"
+        v-model:probe-model="form.accountAutoFreezeProbeModel"
+        v-model:adaptive-concurrency="form.accountAutoFreezeAdaptiveConcurrency"
+      />
+
       <ClientVersionSettings
         v-model:min-codex-desktop-version="form.minCodexDesktopVersion"
         v-model:min-codex-cli-version="form.minCodexCliVersion"