chore: bump version to 3.2.5

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

backend/Cargo.lock

@@ -2581,6 +2581,7 @@ dependencies = [
  "cookie",
  "crypto_box",
  "ed25519-dalek",
+ "flate2",
  "futures",
  "gateway-admin",
  "gateway-core",
@@ -2608,6 +2609,7 @@ dependencies = [
  "tracing-subscriber",
  "tungstenite 0.27.0",
  "url",
+ "urlencoding",
  "uuid",
  "wiremock",
  "zstd",
@@ -3623,7 +3625,7 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "32497e9a4c7b38532efcdebeef879707aa9f794296a4f0244f6f69e9bc8574bd"
 dependencies = [
  "fastrand",
- "getrandom 0.3.4",
+ "getrandom 0.4.3",
  "once_cell",
  "rustix",
  "windows-sys 0.61.2",
@@ -4078,6 +4080,12 @@ dependencies = [
  "serde",
 ]
 
+[[package]]
+name = "urlencoding"
+version = "2.1.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "daf8dba3b7eb870caf1ddeed7bc9d2a049f3cfdfae7cb521b087cc33ae4c49da"
+
 [[package]]
 name = "utf-8"
 version = "0.7.6"

backend/apps/gateway/tests/bootstrap.rs

@@ -135,7 +135,10 @@ fn config_loader_should_reject_removed_fingerprint_section() {
 
 #[test]
 fn config_loader_should_reject_invalid_codex_cli_version() {
-    assert_rejected(valid_config().replace("codex_version: '0.146.0'", "codex_version: 'latest'"));
+    assert_rejected(valid_config().replace(
+        "codex_version: '0.147.0-alpha.6.6'",
+        "codex_version: 'latest'",
+    ));
 }
 
 #[test]
@@ -154,7 +157,7 @@ fn config_loader_should_reject_zero_server_port() {
 
 #[test]
 fn config_loader_should_reject_invalid_desktop_profile_fields() {
-    assert_rejected(valid_config().replace("desktop_build: '6119'", "desktop_build: 'build'"));
+    assert_rejected(valid_config().replace("desktop_build: '6415'", "desktop_build: 'build'"));
 }
 
 fn assert_rejected(config: String) {

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

@@ -608,8 +608,9 @@ pub struct ProviderQuota {
 impl ProviderQuota {
     /// 返回账号展开面板使用的当前额度窗口用量。
     ///
-    /// 只选择具备完整时间边界且可按账号归属的窗口；优先级与 Dashboard
-    /// 代表性额度窗口一致，同一优先级保持 Provider 投影顺序。
+    /// Provider 已提供账号级本地用量时，固定窗口和无固定重置点的滚动窗口
+    /// 都可作为代表用量；优先级与 Dashboard 代表性额度窗口一致，
+    /// 同一优先级保持 Provider 投影顺序。
     #[must_use]
     pub fn representative_window_usage(&self) -> Option<&AccountUsage> {
         self.windows
@@ -618,7 +619,6 @@ impl ProviderQuota {
             .filter(|(_, window)| {
                 window.local_usage_attribution == QuotaLocalUsageAttribution::AccountWide
                     && window.window_seconds.is_some()
-                    && window.reset_at.is_some()
                     && window.local_usage.is_some()
             })
             .min_by_key(|(index, window)| (quota_usage_priority(window), *index))
@@ -631,25 +631,74 @@ impl ProviderQuota {
     /// 这里只解释跨 Provider 共享的窗口语义，绝不读取 Provider 私有 JSON。
     #[must_use]
     pub fn representative_used_percent(&self) -> Option<f64> {
+        if self.limit_reached {
+            return Some(100.0);
+        }
+        self.representative_used_window()
+            .map(|(_, used_percent)| used_percent)
+    }
+
+    /// 将已确认的账号级额度耗尽事实投影到展示窗口。
+    ///
+    /// Provider 已指出具体触顶窗口时只归一化这些窗口；否则归一化 Dashboard
+    /// 同样会选择的代表窗口，避免把多个独立额度窗口全部伪造成已用尽。
+    pub fn apply_limit_reached_display(&mut self) {
+        if !self.limit_reached {
+            return;
+        }
+        let reached_windows = self
+            .windows
+            .iter()
+            .enumerate()
+            .filter_map(|(index, window)| window.limit_reached.then_some(index))
+            .collect::<Vec<_>>();
+        if !reached_windows.is_empty() {
+            for index in reached_windows {
+                self.windows[index].used_percent = Some(100.0);
+            }
+            return;
+        }
+        let representative = self
+            .representative_used_window()
+            .map(|(index, _)| index)
+            .or_else(|| {
+                self.windows
+                    .iter()
+                    .enumerate()
+                    .min_by_key(|(index, window)| (quota_usage_priority(window), *index))
+                    .map(|(index, _)| index)
+            });
+        if let Some(index) = representative {
+            self.windows[index].used_percent = Some(100.0);
+            self.windows[index].limit_reached = true;
+        }
+    }
+
+    fn representative_used_window(&self) -> Option<(usize, f64)> {
         self.windows
             .iter()
+            .enumerate()
             .filter_map(|window| {
+                let (index, window) = window;
                 let used_percent = window
                     .used_percent
                     .filter(|value| value.is_finite())
                     .map(|value| value.clamp(0.0, 100.0))?;
-                Some((quota_usage_priority(window), used_percent))
+                Some((index, quota_usage_priority(window), used_percent))
             })
-            .fold(None::<(u8, f64)>, |selected, candidate| match selected {
-                Some(current)
-                    if current.0 < candidate.0
-                        || (current.0 == candidate.0 && current.1 >= candidate.1) =>
-                {
+            .fold(
+                None::<(usize, u8, f64)>,
+                |selected, candidate| match selected {
                     Some(current)
-                }
-                _ => Some(candidate),
-            })
-            .map(|(_, used_percent)| used_percent)
+                        if current.1 < candidate.1
+                            || (current.1 == candidate.1 && current.2 >= candidate.2) =>
+                    {
+                        Some(current)
+                    }
+                    _ => Some(candidate),
+                },
+            )
+            .map(|(index, _, used_percent)| (index, used_percent))
     }
 }
 

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

@@ -6,6 +6,7 @@ use gateway_core::engine::credential::OpaqueProviderData;
 use gateway_admin::{
     model::{
         PageSize, Revision,
+        accounts::AccountUsage,
         auth::LoginCommand,
         client_keys::ClientKeyPageSize,
         observability::{DecimalAmount, RequestOutcome, TimeRange},
@@ -135,6 +136,95 @@ fn representative_quota_should_prefer_short_window_and_highest_usage() {
     assert_eq!(quota.representative_used_percent(), Some(45.0));
 }
 
+#[test]
+fn representative_window_usage_should_accept_account_wide_rolling_window_without_reset() {
+    let usage = AccountUsage {
+        account_id: "acct_xai".to_owned(),
+        request_count: 5,
+        success_count: 5,
+        input_tokens: Some(100),
+        output_tokens: Some(20),
+        cached_tokens: Some(80),
+        cache_write_tokens: Some(0),
+        reasoning_tokens: Some(0),
+        image_input_tokens: Some(0),
+        image_output_tokens: Some(0),
+        image_request_count: 0,
+        image_request_failed_count: 0,
+        total_tokens: Some(120),
+        cost_coverage: Default::default(),
+        costs: Vec::new(),
+        last_used_at: Some(Utc::now()),
+        request_buckets: Vec::new(),
+        models: Vec::new(),
+    };
+    let quota = ProviderQuota {
+        observed_at: None,
+        refresh_token_expires_at: None,
+        windows: vec![ProviderQuotaWindow {
+            key: "free-rolling-24h".to_owned(),
+            group: "shortTerm".to_owned(),
+            label: "日限额".to_owned(),
+            source: None,
+            local_usage_attribution: QuotaLocalUsageAttribution::AccountWide,
+            window_seconds: Some(86_400),
+            used_percent: None,
+            reset_at: None,
+            limit_reached: false,
+            local_usage: Some(usage.clone()),
+            provider_data: None,
+        }],
+        limit_reached: false,
+        provider_data: None,
+    };
+
+    assert_eq!(quota.representative_window_usage(), Some(&usage));
+}
+
+#[test]
+fn exhausted_quota_should_project_full_usage_to_only_the_representative_window() {
+    let mut quota = ProviderQuota {
+        observed_at: None,
+        refresh_token_expires_at: None,
+        windows: vec![
+            quota_window("monthly", Some(2_592_000), Some(99.0)),
+            quota_window("shortTerm", Some(604_800), Some(80.0)),
+            quota_window("shortTerm", Some(18_000), Some(30.0)),
+            quota_window("shortTerm", Some(14_400), Some(95.0)),
+        ],
+        limit_reached: true,
+        provider_data: None,
+    };
+
+    assert_eq!(quota.representative_used_percent(), Some(100.0));
+    quota.apply_limit_reached_display();
+
+    assert_eq!(quota.windows[0].used_percent, Some(99.0));
+    assert_eq!(quota.windows[1].used_percent, Some(80.0));
+    assert_eq!(quota.windows[2].used_percent, Some(30.0));
+    assert_eq!(quota.windows[3].used_percent, Some(100.0));
+    assert!(quota.windows[3].limit_reached);
+}
+
+#[test]
+fn exhausted_quota_should_preserve_the_provider_identified_reached_window() {
+    let mut reached = quota_window("monthly", Some(2_592_000), Some(98.0));
+    reached.limit_reached = true;
+    let mut quota = ProviderQuota {
+        observed_at: None,
+        refresh_token_expires_at: None,
+        windows: vec![reached, quota_window("shortTerm", Some(18_000), Some(95.0))],
+        limit_reached: true,
+        provider_data: None,
+    };
+
+    quota.apply_limit_reached_display();
+
+    assert_eq!(quota.windows[0].used_percent, Some(100.0));
+    assert_eq!(quota.windows[1].used_percent, Some(95.0));
+    assert!(!quota.windows[1].limit_reached);
+}
+
 fn quota_window(
     group: &str,
     window_seconds: Option<u64>,

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

@@ -1574,10 +1574,11 @@ fn account_view(item: AccountDirectoryItem, now: DateTime<Utc>) -> AccountView {
 }
 
 fn account_quota_view(
-    quota: ProviderQuota,
+    mut quota: ProviderQuota,
     rate_limited_until: Option<DateTime<Utc>>,
     now: DateTime<Utc>,
 ) -> (AccountQuotaView, Option<String>) {
+    quota.apply_limit_reached_display();
     let refresh_token_expires_at = quota
         .refresh_token_expires_at
         .map(|value| china_rfc3339(&value));

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

@@ -1588,14 +1588,26 @@ impl AccountSelector {
 
         let candidate = match context.policy.strategy() {
             RotationStrategy::QuotaResetPriority => {
-                eligible.sort_by_key(|candidate| {
+                let default_concurrency = context.policy.max_concurrent_per_account();
+                eligible.sort_by(|left, right| {
                     (
-                        candidate.signals.quota_reset_at.is_none(),
-                        candidate.signals.quota_reset_at,
-                        candidate.signals.in_flight,
-                        candidate.signals.last_started_at,
-                        candidate.account.id().clone(),
+                        left.signals.quota_reset_at.is_none(),
+                        left.signals.quota_reset_at,
                     )
+                        .cmp(&(
+                            right.signals.quota_reset_at.is_none(),
+                            right.signals.quota_reset_at,
+                        ))
+                        .then_with(|| {
+                            capacity_utilization(left, default_concurrency)
+                                .total_cmp(&capacity_utilization(right, default_concurrency))
+                        })
+                        .then_with(|| {
+                            left.signals
+                                .last_started_at
+                                .cmp(&right.signals.last_started_at)
+                        })
+                        .then_with(|| left.account.id().cmp(right.account.id()))
                 });
                 eligible.first().copied()?
             }
@@ -1604,9 +1616,11 @@ impl AccountSelector {
                 let index = context.round_robin_cursor as usize % eligible.len();
                 eligible.get(index).copied()?
             }
-            RotationStrategy::Smart => {
-                select_smart_candidate(&eligible, context.round_robin_cursor)?
-            }
+            RotationStrategy::Smart => select_smart_candidate(
+                &eligible,
+                context.policy.max_concurrent_per_account(),
+                context.round_robin_cursor,
+            )?,
             RotationStrategy::Sticky => {
                 eligible.sort_by_key(|candidate| {
                     (
@@ -1676,8 +1690,18 @@ const SMART_QUOTA_WEIGHT: f64 = 0.8;
 const SMART_FAILURE_WEIGHT: f64 = 1.0;
 const SMART_LATENCY_WEIGHT: f64 = 0.5;
 
+fn capacity_utilization(candidate: &AccountCandidate, default_concurrency: NonZeroU32) -> f64 {
+    f64::from(candidate.signals.in_flight)
+        / f64::from(
+            candidate
+                .account
+                .effective_concurrency(default_concurrency)
+                .get(),
+        )
+}
+
 struct SmartNormalization {
-    max_in_flight: u32,
+    max_capacity_utilization: f64,
     min_quota: Option<u64>,
     max_quota: Option<u64>,
     min_latency_ms: Option<u64>,
@@ -1685,11 +1709,11 @@ struct SmartNormalization {
 }
 
 impl SmartNormalization {
-    fn from_candidates(candidates: &[&AccountCandidate]) -> Self {
-        let max_in_flight = candidates
+    fn from_candidates(candidates: &[&AccountCandidate], default_concurrency: NonZeroU32) -> Self {
+        let max_capacity_utilization = candidates
             .iter()
-            .map(|candidate| candidate.signals.in_flight)
-            .max()
+            .map(|candidate| capacity_utilization(candidate, default_concurrency))
+            .reduce(f64::max)
             .unwrap_or_default();
         let min_quota = candidates
             .iter()
@@ -1710,7 +1734,7 @@ impl SmartNormalization {
             .filter(|latency| *latency > 0)
             .max();
         Self {
-            max_in_flight,
+            max_capacity_utilization,
             min_quota,
             max_quota,
             min_latency_ms,
@@ -1721,12 +1745,18 @@ impl SmartNormalization {
 
 fn select_smart_candidate<'a>(
     candidates: &[&'a AccountCandidate],
+    default_concurrency: NonZeroU32,
     cursor: u64,
 ) -> Option<&'a AccountCandidate> {
-    let normalization = SmartNormalization::from_candidates(candidates);
+    let normalization = SmartNormalization::from_candidates(candidates, default_concurrency);
     let mut ranked = candidates
         .iter()
-        .map(|candidate| (*candidate, smart_score(candidate, &normalization)))
+        .map(|candidate| {
+            (
+                *candidate,
+                smart_score(candidate, default_concurrency, &normalization),
+            )
+        })
         .collect::<Vec<_>>();
     ranked.sort_by(|(left, left_score), (right, right_score)| {
         right_score
@@ -1742,11 +1772,14 @@ fn select_smart_candidate<'a>(
     Some(ranked[index].0)
 }
 
-fn smart_score(candidate: &AccountCandidate, normalization: &SmartNormalization) -> f64 {
-    let load = lower_is_better(
-        u64::from(candidate.signals.in_flight),
-        0,
-        u64::from(normalization.max_in_flight),
+fn smart_score(
+    candidate: &AccountCandidate,
+    default_concurrency: NonZeroU32,
+    normalization: &SmartNormalization,
+) -> f64 {
+    let load = lower_ratio_is_better(
+        capacity_utilization(candidate, default_concurrency),
+        normalization.max_capacity_utilization,
     );
     let quota = candidate.signals.quota_remaining_rank.map_or(0.5, |quota| {
         higher_is_better(
@@ -1781,6 +1814,13 @@ fn smart_score(candidate: &AccountCandidate, normalization: &SmartNormalization)
         + SMART_LATENCY_WEIGHT * latency
 }
 
+fn lower_ratio_is_better(value: f64, maximum: f64) -> f64 {
+    if maximum <= 0.0 {
+        return 1.0;
+    }
+    1.0 - (value / maximum).clamp(0.0, 1.0)
+}
+
 fn lower_is_better(value: u64, minimum: u64, maximum: u64) -> f64 {
     if maximum <= minimum {
         return 1.0;

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

@@ -407,6 +407,84 @@ pub trait ProviderCatalogCachePort: Send + Sync {
     ) -> BoxFuture<'a, Result<Option<OpaqueProviderData>, ProviderStoreError>>;
 }
 
+/// Provider 从官方制品核验出的可重建请求画像。
+///
+/// Core 只用单调制品序号约束覆盖顺序；具体版本字段由对应 Provider 放在
+/// `profile` 中解释。每个 Provider 在 Store 中只保留一份最新画像。
+#[derive(Clone, PartialEq)]
+pub struct ProviderArtifactProfile {
+    provider_kind: ProviderKind,
+    artifact_sequence: u64,
+    verified_at: SystemTime,
+    profile: OpaqueProviderData,
+}
+
+impl ProviderArtifactProfile {
+    #[must_use]
+    pub const fn new(
+        provider_kind: ProviderKind,
+        artifact_sequence: u64,
+        verified_at: SystemTime,
+        profile: OpaqueProviderData,
+    ) -> Self {
+        Self {
+            provider_kind,
+            artifact_sequence,
+            verified_at,
+            profile,
+        }
+    }
+
+    #[must_use]
+    pub const fn provider_kind(&self) -> &ProviderKind {
+        &self.provider_kind
+    }
+
+    #[must_use]
+    pub const fn artifact_sequence(&self) -> u64 {
+        self.artifact_sequence
+    }
+
+    #[must_use]
+    pub const fn verified_at(&self) -> SystemTime {
+        self.verified_at
+    }
+
+    #[must_use]
+    pub const fn profile(&self) -> &OpaqueProviderData {
+        &self.profile
+    }
+}
+
+impl fmt::Debug for ProviderArtifactProfile {
+    fn fmt(&self, formatter: &mut fmt::Formatter<'_>) -> fmt::Result {
+        formatter
+            .debug_struct("ProviderArtifactProfile")
+            .field("provider_kind", &self.provider_kind)
+            .field("artifact_sequence", &self.artifact_sequence)
+            .field("verified_at", &self.verified_at)
+            .field("profile", &"[PROVIDER-OWNED]")
+            .finish()
+    }
+}
+
+pub trait ProviderArtifactProfileCachePort: Send + Sync {
+    /// 覆盖同一 Provider 的固定 cache key。
+    ///
+    /// 返回 `false` 表示 Store 已持有更高的制品序号；相同序号但内容不同必须返回
+    /// [`ProviderStoreErrorKind::Conflict`]，不能静默改写已核验画像。
+    fn replace_if_newer(
+        &self,
+        profile: ProviderArtifactProfile,
+        ttl: Duration,
+    ) -> BoxFuture<'_, Result<bool, ProviderStoreError>>;
+
+    fn read<'a>(
+        &'a self,
+        provider_kind: &'a ProviderKind,
+    ) -> BoxFuture<'a, Result<Option<ProviderArtifactProfile>, ProviderStoreError>>;
+}
+
 /// Redis 中可重建的账号状态投影。
 #[derive(Debug, Clone, PartialEq, Eq)]
 pub struct ProviderCredentialState {
@@ -946,6 +1024,7 @@ pub struct ProviderStorePorts {
     session_exclusions: Arc<dyn ProviderSessionExclusionPort>,
     account_feedback: Arc<AccountFeedbackStats>,
     catalog_cache: Arc<dyn ProviderCatalogCachePort>,
+    artifact_profiles: Arc<dyn ProviderArtifactProfileCachePort>,
     credential_state: Arc<dyn ProviderCredentialStatePort>,
     cooldowns: Arc<dyn ProviderCooldownPort>,
     runtime_policy: Arc<dyn ProviderRuntimePolicyPort>,
@@ -962,6 +1041,7 @@ impl ProviderStorePorts {
         session_affinity: Arc<dyn ProviderSessionAffinityPort>,
         session_exclusions: Arc<dyn ProviderSessionExclusionPort>,
         catalog_cache: Arc<dyn ProviderCatalogCachePort>,
+        artifact_profiles: Arc<dyn ProviderArtifactProfileCachePort>,
         credential_state: Arc<dyn ProviderCredentialStatePort>,
         cooldowns: Arc<dyn ProviderCooldownPort>,
         runtime_policy: Arc<dyn ProviderRuntimePolicyPort>,
@@ -974,6 +1054,7 @@ impl ProviderStorePorts {
             session_exclusions,
             account_feedback: Arc::new(AccountFeedbackStats::default()),
             catalog_cache,
+            artifact_profiles,
             credential_state,
             cooldowns,
             runtime_policy,
@@ -1011,6 +1092,11 @@ impl ProviderStorePorts {
         Arc::clone(&self.catalog_cache)
     }
 
+    #[must_use]
+    pub fn artifact_profiles(&self) -> Arc<dyn ProviderArtifactProfileCachePort> {
+        Arc::clone(&self.artifact_profiles)
+    }
+
     #[must_use]
     pub fn credential_state(&self) -> Arc<dyn ProviderCredentialStatePort> {
         Arc::clone(&self.credential_state)

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

@@ -51,6 +51,15 @@ fn candidate(id: &str, in_flight: u32, remaining: Option<u64>) -> AccountCandida
     }
 }
 
+fn candidate_with_concurrency(id: &str, in_flight: u32, concurrency: u32) -> AccountCandidate {
+    let mut candidate = candidate(id, in_flight, Some(100));
+    candidate.account = candidate.account.with_scheduling(
+        Some(AccountConcurrencyLimit::new(concurrency).expect("concurrency override")),
+        AccountWeight::DEFAULT,
+    );
+    candidate
+}
+
 fn weighted_candidate(id: &str, weight: u16, in_flight: u32) -> AccountCandidate {
     let mut candidate = candidate(id, in_flight, None);
     candidate.account = candidate.account.with_scheduling(
@@ -348,6 +357,20 @@ fn smart_selector_should_prefer_lower_inflight_count() {
     assert_eq!(selected.candidate().account.id().as_str(), "acct_idle");
 }
 
+#[test]
+fn smart_selector_should_prefer_lower_capacity_utilization_over_lower_inflight_count() {
+    let candidates = [
+        candidate_with_concurrency("acct_smaller", 1, 2),
+        candidate_with_concurrency("acct_larger", 2, 10),
+    ];
+
+    let selected = AccountSelector
+        .select(&candidates, &context(RotationStrategy::Smart))
+        .expect("candidate available");
+
+    assert_eq!(selected.candidate().account.id().as_str(), "acct_larger");
+}
+
 #[test]
 fn selector_should_prioritize_the_highest_weight_for_every_rotation_strategy() {
     let candidates = vec![
@@ -622,6 +645,22 @@ fn quota_reset_selector_should_prefer_known_earliest_window() {
     assert_eq!(selected.candidate().account.id().as_str(), "acct_earlier");
 }
 
+#[test]
+fn quota_reset_selector_should_use_capacity_utilization_as_load_tiebreaker() {
+    let reset_at = SystemTime::now() + Duration::from_secs(60);
+    let mut smaller = candidate_with_concurrency("acct_smaller", 1, 2);
+    smaller.signals.quota_reset_at = Some(reset_at);
+    let mut larger = candidate_with_concurrency("acct_larger", 2, 10);
+    larger.signals.quota_reset_at = Some(reset_at);
+    let candidates = [smaller, larger];
+
+    let selected = AccountSelector
+        .select(&candidates, &context(RotationStrategy::QuotaResetPriority))
+        .expect("candidate available");
+
+    assert_eq!(selected.candidate().account.id().as_str(), "acct_larger");
+}
+
 #[test]
 fn smart_selector_should_prefer_known_quota_over_unknown_after_load_is_equal() {
     let candidates = vec![

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

@@ -70,7 +70,7 @@ pub const MAX_SSE_EVENT_BUFFER_BYTES: usize = 64 * 1024 * 1024;
 pub const DONE_SSE_FRAME: &str = "data: [DONE]\n\n";
 
 /// 用于处理任意分块边界的增量 SSE 解码器。
-#[derive(Debug, Default)]
+#[derive(Debug)]
 pub struct SseEventDecoder {
     pending: Vec<u8>,
     /// `pending` 中已确认不含帧分隔符的前缀长度。新数据到达时从该边界
@@ -79,6 +79,19 @@ pub struct SseEventDecoder {
     /// 超大未完成帧已开始直接交付；在遇到帧分隔符前只保留跨 chunk 扫描尾部，
     /// 不再尝试解析该帧。
     opaque_frame: bool,
+    /// 只有整个 SSE 流的第一条物理行允许携带 UTF-8 BOM。
+    stream_start: bool,
+}
+
+impl Default for SseEventDecoder {
+    fn default() -> Self {
+        Self {
+            pending: Vec::new(),
+            scanned: 0,
+            opaque_frame: false,
+            stream_start: true,
+        }
+    }
 }
 
 impl SseEventDecoder {
@@ -115,7 +128,8 @@ impl SseEventDecoder {
         while let Some(end) = self.next_frame_end(consumed) {
             let frame = std::str::from_utf8(&self.pending[consumed..end])
                 .map_err(|error| SseError::ParseError(error.to_string()))?;
-            events.extend(parse_sse_events(frame)?);
+            events.extend(parse_sse_events_inner(frame, self.stream_start)?);
+            self.stream_start = false;
             consumed = end;
         }
 
@@ -149,9 +163,10 @@ impl SseEventDecoder {
             } else {
                 std::str::from_utf8(&raw)
                     .ok()
-                    .and_then(|frame| parse_sse_events(frame).ok())
+                    .and_then(|frame| parse_sse_events_inner(frame, self.stream_start).ok())
                     .unwrap_or_default()
             };
+            self.stream_start = false;
             frames.push(SseFrame::new(raw, events));
             consumed = end;
         }
@@ -173,7 +188,9 @@ impl SseEventDecoder {
         let pending = std::mem::take(&mut self.pending);
         let frame = std::str::from_utf8(&pending)
             .map_err(|error| SseError::ParseError(error.to_string()))?;
-        parse_sse_events(frame)
+        let events = parse_sse_events_inner(frame, self.stream_start)?;
+        self.stream_start = false;
+        Ok(events)
     }
 
     /// 在流结束时返回未带空行分隔符的最后一帧及其旁路解析结果。
@@ -191,9 +208,10 @@ impl SseEventDecoder {
         } else {
             std::str::from_utf8(&raw)
                 .ok()
-                .and_then(|frame| parse_sse_events(frame).ok())
+                .and_then(|frame| parse_sse_events_inner(frame, self.stream_start).ok())
                 .unwrap_or_default()
         };
+        self.stream_start = false;
         self.opaque_frame = false;
         vec![SseFrame::new(raw, events)]
     }
@@ -207,6 +225,7 @@ impl SseEventDecoder {
         let remaining = self.pending.split_off(end);
         let raw = std::mem::replace(&mut self.pending, remaining);
         frames.push(SseFrame::new(raw, Vec::new()));
+        self.stream_start = false;
         self.opaque_frame = false;
         self.scanned = 0;
         true
@@ -226,6 +245,7 @@ impl SseEventDecoder {
         let raw = std::mem::replace(&mut self.pending, tail);
         self.scanned = 0;
         frames.push(SseFrame::new(raw, Vec::new()));
+        self.stream_start = false;
     }
 }
 
@@ -351,12 +371,21 @@ impl EventBuilder {
 
 /// 解析 SSE 事件流。
 pub fn parse_sse_events(input: &str) -> Result<Vec<SseEvent>, SseError> {
+    parse_sse_events_inner(input, true)
+}
+
+fn parse_sse_events_inner(input: &str, strip_leading_bom: bool) -> Result<Vec<SseEvent>, SseError> {
     let mut events = Vec::new();
     let mut builder = EventBuilder::default();
     let mut saw_sse_syntax = false;
     let mut event_buffer_bytes = 0usize;
 
-    for raw_line in input.lines() {
+    for (index, raw_line) in input.lines().enumerate() {
+        let raw_line = if strip_leading_bom && index == 0 {
+            raw_line.strip_prefix('\u{feff}').unwrap_or(raw_line)
+        } else {
+            raw_line
+        };
         let line = raw_line.strip_suffix('\r').unwrap_or(raw_line);
         if line.is_empty() {
             event_buffer_bytes = 0;

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

@@ -31,6 +31,50 @@ fn parser_should_preserve_multiline_data_id_and_retry() {
     );
 }
 
+#[test]
+fn parser_should_ignore_a_leading_utf8_bom() {
+    let input = concat!(
+        "\u{feff}event: response.created\n",
+        "data: {\"type\":\"response.created\"}\n\n",
+    );
+
+    assert_eq!(
+        parse_sse_events(input).expect("SSE with BOM"),
+        vec![SseEvent {
+            event: Some("response.created".to_owned()),
+            data: "{\"type\":\"response.created\"}".to_owned(),
+            id: None,
+            retry: None,
+        }]
+    );
+}
+
+#[test]
+fn incremental_decoder_should_ignore_bom_only_at_the_stream_start() {
+    let mut decoder = SseEventDecoder::default();
+    let first = decoder
+        .push(
+            concat!(
+                "\u{feff}event: response.created\n",
+                "data: {\"type\":\"response.created\"}\n\n",
+            )
+            .as_bytes(),
+        )
+        .expect("first SSE frame");
+    let second = decoder
+        .push(
+            concat!(
+                "\u{feff}event: response.completed\n",
+                "data: {\"type\":\"response.completed\"}\n\n",
+            )
+            .as_bytes(),
+        )
+        .expect("second SSE frame");
+
+    assert_eq!(first[0].event.as_deref(), Some("response.created"));
+    assert_eq!(second[0].event, None);
+}
+
 #[test]
 fn parser_should_ignore_done_control_frame() {
     assert!(

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

@@ -79,6 +79,10 @@ pub async fn initialize(mut config: StoreConfig) -> StoreResult<StoreBundle> {
         redis_connection.clone(),
         REDIS_NAMESPACE,
     )?);
+    let artifact_profiles = Arc::new(redis::RedisProviderArtifactProfileRepository::new(
+        redis_connection.clone(),
+        REDIS_NAMESPACE,
+    )?);
     let runtime_policy = Arc::new(postgres::PgRuntimeSettingsRepository::new(pool.clone()));
     let oauth_pending = Arc::new(redis::RedisOAuthPendingFlowRepository::new(
         redis_connection.clone(),
@@ -167,6 +171,7 @@ pub async fn initialize(mut config: StoreConfig) -> StoreResult<StoreBundle> {
             REDIS_NAMESPACE,
         )?),
         credential_state.clone(),
+        artifact_profiles,
         credential_state,
         cooldowns,
         runtime_policy,

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

@@ -52,9 +52,12 @@ pub(crate) async fn provider_account_usage(
     }
     statement.push(
         " group by pa.id, pa.provider_kind, pa.authentication_kind, pa.name, pa.email,
-                   pa.plan_type
-          order by max(mr.started_at) desc nulls last, pa.name, pa.id limit ",
+                   pa.plan_type",
     );
+    if query.account_ids.is_none() {
+        statement.push(" having count(mr.id) > 0");
+    }
+    statement.push(" order by max(mr.started_at) desc nulls last, pa.name, pa.id limit ");
     statement.push_bind(i64::from(query.limit));
     let rows = statement
         .build()

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

@@ -0,0 +1,175 @@
+//! Provider 官方制品画像的单键 Redis cache。
+
+use std::time::{Duration, SystemTime};
+
+use gateway_core::engine::credential::OpaqueProviderData;
+use gateway_core::provider_ports::{
+    ProviderArtifactProfile, ProviderArtifactProfileCachePort, ProviderStoreError,
+    ProviderStoreErrorKind,
+};
+use gateway_core::routing::ProviderKind;
+use redis::{Script, aio::ConnectionManager};
+
+use crate::StoreResult;
+
+use super::{MAX_REDIS_EXACT_INTEGER, namespace};
+
+const MAX_PROFILE_BYTES: usize = 16 * 1024;
+
+const REPLACE_SCRIPT: &str = r#"
+local current_sequence = redis.call('HGET', KEYS[1], 'artifact_sequence')
+if current_sequence then
+  local current_number = tonumber(current_sequence)
+  local candidate_number = tonumber(ARGV[1])
+  if current_number > candidate_number then return 0 end
+  if current_number == candidate_number then
+    local current_profile = redis.call('HGET', KEYS[1], 'profile')
+    if current_profile ~= ARGV[3] then return -1 end
+  end
+end
+redis.call('HSET', KEYS[1],
+  'artifact_sequence', ARGV[1],
+  'verified_at_ms', ARGV[2],
+  'profile', ARGV[3])
+redis.call('PEXPIRE', KEYS[1], ARGV[4])
+return 1
+"#;
+
+/// 每个 Provider 只占一个覆盖写 key；TTL 负责进程长期停机后的最终清理。
+#[derive(Clone)]
+pub struct RedisProviderArtifactProfileRepository {
+    connection: ConnectionManager,
+    namespace: String,
+}
+
+impl RedisProviderArtifactProfileRepository {
+    pub fn new(connection: ConnectionManager, key_namespace: &str) -> StoreResult<Self> {
+        Ok(Self {
+            connection,
+            namespace: namespace(key_namespace)?,
+        })
+    }
+
+    fn key(&self, provider_kind: &ProviderKind) -> String {
+        format!(
+            "{}:provider:{}:artifact-profile:v1",
+            self.namespace,
+            provider_kind.as_str()
+        )
+    }
+}
+
+impl ProviderArtifactProfileCachePort for RedisProviderArtifactProfileRepository {
+    fn replace_if_newer(
+        &self,
+        profile: ProviderArtifactProfile,
+        ttl: Duration,
+    ) -> futures::future::BoxFuture<'_, Result<bool, ProviderStoreError>> {
+        Box::pin(async move {
+            let artifact_sequence = profile.artifact_sequence();
+            if artifact_sequence == 0 || artifact_sequence > MAX_REDIS_EXACT_INTEGER {
+                return Err(provider_invalid("validate artifact profile sequence"));
+            }
+            let ttl_ms = exact_positive_millis(ttl, "validate artifact profile TTL")?;
+            let verified_at_ms = profile
+                .verified_at()
+                .duration_since(SystemTime::UNIX_EPOCH)
+                .ok()
+                .and_then(|duration| u64::try_from(duration.as_millis()).ok())
+                .filter(|millis| *millis <= MAX_REDIS_EXACT_INTEGER)
+                .ok_or_else(|| provider_invalid("validate artifact profile timestamp"))?;
+            let payload = serde_json::to_vec(&serde_json::Value::Object(
+                profile.profile().expose_to_provider().clone(),
+            ))
+            .map_err(|_| provider_invalid("encode artifact profile"))?;
+            if payload.len() > MAX_PROFILE_BYTES {
+                return Err(provider_invalid("validate artifact profile size"));
+            }
+
+            let mut connection = self.connection.clone();
+            let outcome = Script::new(REPLACE_SCRIPT)
+                .key(self.key(profile.provider_kind()))
+                .arg(artifact_sequence)
+                .arg(verified_at_ms)
+                .arg(payload)
+                .arg(ttl_ms)
+                .invoke_async::<i64>(&mut connection)
+                .await
+                .map_err(|_| provider_unavailable("replace artifact profile"))?;
+            match outcome {
+                1 => Ok(true),
+                0 => Ok(false),
+                -1 => Err(ProviderStoreError::new(
+                    ProviderStoreErrorKind::Conflict,
+                    "replace artifact profile",
+                )),
+                _ => Err(provider_invalid("decode artifact profile write outcome")),
+            }
+        })
+    }
+
+    fn read<'a>(
+        &'a self,
+        provider_kind: &'a ProviderKind,
+    ) -> futures::future::BoxFuture<'a, Result<Option<ProviderArtifactProfile>, ProviderStoreError>>
+    {
+        Box::pin(async move {
+            let mut connection = self.connection.clone();
+            let values: (Option<u64>, Option<u64>, Option<Vec<u8>>) = redis::cmd("HMGET")
+                .arg(self.key(provider_kind))
+                .arg("artifact_sequence")
+                .arg("verified_at_ms")
+                .arg("profile")
+                .query_async(&mut connection)
+                .await
+                .map_err(|_| provider_unavailable("read artifact profile"))?;
+            let (artifact_sequence, verified_at_ms, payload) = values;
+            if artifact_sequence.is_none() && verified_at_ms.is_none() && payload.is_none() {
+                return Ok(None);
+            }
+            let (Some(artifact_sequence), Some(verified_at_ms), Some(payload)) =
+                (artifact_sequence, verified_at_ms, payload)
+            else {
+                return Err(provider_invalid("decode artifact profile"));
+            };
+            if artifact_sequence == 0 || artifact_sequence > MAX_REDIS_EXACT_INTEGER {
+                return Err(provider_invalid("decode artifact profile sequence"));
+            }
+            if payload.len() > MAX_PROFILE_BYTES {
+                return Err(provider_invalid("decode artifact profile size"));
+            }
+            let serde_json::Value::Object(fields) = serde_json::from_slice(&payload)
+                .map_err(|_| provider_invalid("decode artifact profile"))?
+            else {
+                return Err(provider_invalid("decode artifact profile"));
+            };
+            let verified_at = SystemTime::UNIX_EPOCH
+                .checked_add(Duration::from_millis(verified_at_ms))
+                .ok_or_else(|| provider_invalid("decode artifact profile timestamp"))?;
+            Ok(Some(ProviderArtifactProfile::new(
+                provider_kind.clone(),
+                artifact_sequence,
+                verified_at,
+                OpaqueProviderData::new(fields),
+            )))
+        })
+    }
+}
+
+fn exact_positive_millis(
+    duration: Duration,
+    operation: &'static str,
+) -> Result<u64, ProviderStoreError> {
+    u64::try_from(duration.as_millis())
+        .ok()
+        .filter(|millis| *millis > 0 && *millis <= MAX_REDIS_EXACT_INTEGER)
+        .ok_or_else(|| provider_invalid(operation))
+}
+
+fn provider_unavailable(operation: &'static str) -> ProviderStoreError {
+    ProviderStoreError::new(ProviderStoreErrorKind::Unavailable, operation)
+}
+
+fn provider_invalid(operation: &'static str) -> ProviderStoreError {
+    ProviderStoreError::new(ProviderStoreErrorKind::InvalidData, operation)
+}

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

@@ -7,6 +7,7 @@ use sha2::{Digest, Sha256};
 
 use async_trait::async_trait;
 
+mod artifact_profile;
 mod client_admission;
 mod coordination_buffer;
 mod credential_cooldown;
@@ -20,6 +21,7 @@ mod provider_session_exclusion;
 mod runtime_change;
 pub(crate) mod worker_lease;
 
+pub use artifact_profile::*;
 pub use client_admission::*;
 pub use coordination_buffer::*;
 pub use credential_cooldown::*;

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

@@ -730,8 +730,18 @@ async fn dashboard_summary_totals_include_history_outside_selected_range() {
         .await
         .expect("dashboard summary");
     assert_eq!(dashboard.requests.request_count, 2);
-    assert_eq!(dashboard.account_usage[0].request_count, 0);
-    assert_eq!(dashboard.account_usage[0].request_buckets.len(), 2);
+    assert!(dashboard.account_usage.is_empty());
+
+    let explicit_account_usage = repository
+        .provider_account_usage(
+            ProviderAccountUsageQuery::for_accounts(range, vec!["acct_observe".to_owned()])
+                .expect("account usage query"),
+        )
+        .await
+        .expect("provider account usage");
+    assert_eq!(explicit_account_usage.len(), 1);
+    assert_eq!(explicit_account_usage[0].request_count, 0);
+
     assert_eq!(
         (
             dashboard.totals.request_count,

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

@@ -0,0 +1,114 @@
+use std::time::{Duration, SystemTime};
+
+use gateway_core::engine::credential::OpaqueProviderData;
+use gateway_core::provider_ports::{
+    ProviderArtifactProfile, ProviderArtifactProfileCachePort, ProviderStoreErrorKind,
+};
+use gateway_core::routing::ProviderKind;
+use gateway_store::redis::RedisProviderArtifactProfileRepository;
+use redis::aio::ConnectionManager;
+use serde_json::{Map, Value};
+use uuid::Uuid;
+
+#[test]
+fn artifact_profile_adapter_implements_provider_port() {
+    fn assert_port<T: ProviderArtifactProfileCachePort>() {}
+    assert_port::<RedisProviderArtifactProfileRepository>();
+}
+
+#[tokio::test]
+async fn artifact_profile_uses_one_expiring_key_and_rejects_rollback_or_conflict() {
+    let Some((repository, mut connection, namespace)) = repository().await else {
+        return;
+    };
+    let provider = ProviderKind::new("openai").expect("provider kind");
+    let current = profile(provider.clone(), 6_570, "0.148.0-alpha.9");
+    assert!(
+        repository
+            .replace_if_newer(current.clone(), Duration::from_secs(60))
+            .await
+            .expect("store current profile")
+    );
+    assert_eq!(
+        repository.read(&provider).await.expect("read profile"),
+        Some(current.clone())
+    );
+
+    assert!(
+        !repository
+            .replace_if_newer(
+                profile(provider.clone(), 6_415, "0.147.0-alpha.6.6"),
+                Duration::from_secs(60),
+            )
+            .await
+            .expect("reject older profile")
+    );
+    let conflict = repository
+        .replace_if_newer(
+            profile(provider.clone(), 6_570, "0.148.0-alpha.10"),
+            Duration::from_secs(60),
+        )
+        .await
+        .expect_err("same build with different Core must conflict");
+    assert_eq!(conflict.kind(), ProviderStoreErrorKind::Conflict);
+    assert_eq!(
+        repository
+            .read(&provider)
+            .await
+            .expect("read retained profile"),
+        Some(current)
+    );
+
+    let keys = redis::cmd("KEYS")
+        .arg(format!("{namespace}:*"))
+        .query_async::<Vec<String>>(&mut connection)
+        .await
+        .expect("list isolated keys");
+    assert_eq!(keys.len(), 1);
+    assert_eq!(
+        keys[0],
+        format!("{namespace}:provider:openai:artifact-profile:v1")
+    );
+    let ttl = redis::cmd("PTTL")
+        .arg(&keys[0])
+        .query_async::<i64>(&mut connection)
+        .await
+        .expect("read profile TTL");
+    assert!((1..=60_000).contains(&ttl));
+}
+
+fn profile(
+    provider: ProviderKind,
+    artifact_sequence: u64,
+    codex_version: &str,
+) -> ProviderArtifactProfile {
+    let mut fields = Map::new();
+    fields.insert("schema_version".to_owned(), Value::from(1));
+    fields.insert(
+        "codex_version".to_owned(),
+        Value::String(codex_version.to_owned()),
+    );
+    ProviderArtifactProfile::new(
+        provider,
+        artifact_sequence,
+        SystemTime::UNIX_EPOCH + Duration::from_secs(1_800_000_000),
+        OpaqueProviderData::new(fields),
+    )
+}
+
+async fn repository() -> Option<(
+    RedisProviderArtifactProfileRepository,
+    ConnectionManager,
+    String,
+)> {
+    let redis_url = crate::support::test_env("CPR_TEST_REDIS_URL")?;
+    let client = redis::Client::open(redis_url).expect("valid CPR_TEST_REDIS_URL");
+    let connection = client
+        .get_connection_manager()
+        .await
+        .expect("connect test Redis");
+    let namespace = format!("gateway-store-test-{}", Uuid::new_v4());
+    let repository = RedisProviderArtifactProfileRepository::new(connection.clone(), &namespace)
+        .expect("valid test namespace");
+    Some((repository, connection, namespace))
+}

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

@@ -1,3 +1,4 @@
+mod artifact_profile;
 mod client_admission;
 mod coordination_buffer;
 mod credential_cooldown;

backend/crates/providers/openai/Cargo.toml

@@ -17,6 +17,7 @@ cookie = "0.18.1"
 crypto_box = { version = "0.9.1", features = ["seal"] }
 ed25519-dalek = { version = "3.0.0", features = ["alloc", "pkcs8"] }
 futures = "0.3.32"
+flate2 = "1.1.9"
 gateway-admin = { path = "../../gateway-admin" }
 gateway-core = { path = "../../gateway-core" }
 gateway-protocol = { path = "../../gateway-protocol" }
@@ -49,6 +50,7 @@ tokio-tungstenite = { git = "https://github.com/openai-oss-forks/tokio-tungsteni
 tracing = "0.1.44"
 tungstenite = { git = "https://github.com/openai-oss-forks/tungstenite-rs", rev = "4fffad30fe373adbdcffab9545e9e9bf4f2fc19f", features = ["deflate", "proxy"] }
 url = "2.5.8"
+urlencoding = "2.1.3"
 uuid = { version = "1.23.3", features = ["v4", "v7"] }
 # 官方 Codex app-server 对 /v1/responses 请求体默认做 zstd 压缩（level 3），
 # 并携带 Content-Encoding: zstd；保持同版本同级别以对齐上游指纹。

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

@@ -51,10 +51,10 @@ use crate::credential::{
 };
 use crate::credential::{CodexCredentialCodec, CodexOAuthSecret, oauth_owner_ref};
 use crate::transport::CodexWebSocketPool;
-use crate::transport::openai_billing_breakdown;
 use crate::transport::profile::{
     CodexDesktopReleaseSnapshot, CodexDesktopReleaseStatus, CodexWireProfile, CodexWireProfileState,
 };
+use crate::transport::{OpenAiBillingUsage, openai_billing_breakdown};
 
 const PROVIDER_NAME: &str = "openai";
 const PENDING_DOCUMENT_SCHEMA_VERSION: u64 = 3;
@@ -215,10 +215,12 @@ impl ProviderAdmin for OpenAiAdminProvider {
         };
         let Some(breakdown) = openai_billing_breakdown(
             &input.upstream_model_id,
-            input_tokens,
-            output_tokens,
-            input.cached_tokens.unwrap_or_default(),
-            input.cache_write_tokens.unwrap_or_default(),
+            OpenAiBillingUsage::new(
+                input_tokens,
+                output_tokens,
+                input.cached_tokens.unwrap_or_default(),
+                input.cache_write_tokens.unwrap_or_default(),
+            ),
             input.service_tier.as_deref(),
         ) else {
             return Ok(None);
@@ -1011,6 +1013,7 @@ struct PendingDocument {
     state: String,
     nonce: String,
     code_verifier: String,
+    installation_id: String,
     reauthorization_account_id: Option<String>,
     mutation: PendingMutationDocument,
 }
@@ -1071,6 +1074,10 @@ fn encode_pending(pending: &CodexPendingAuthorization) -> Map<String, Value> {
         "code_verifier".to_owned(),
         Value::String(pending.code_verifier().expose_secret().to_owned()),
     );
+    document.insert(
+        "installation_id".to_owned(),
+        Value::String(pending.installation_id().to_owned()),
+    );
     document.insert(
         "reauthorization_account_id".to_owned(),
         pending
@@ -1161,6 +1168,7 @@ fn decode_pending(
         state: SecretString::from(document.state),
         nonce: SecretString::from(document.nonce),
         code_verifier: SecretString::from(document.code_verifier),
+        installation_id: document.installation_id,
         reauthorization_account_id: document.reauthorization_account_id,
         mutation: decode_mutation(document.mutation)?,
     })

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

@@ -269,7 +269,7 @@ impl CodexAuthSettings {
 #[serde(deny_unknown_fields)]
 pub struct CodexWireProfileConfig {
     pub originator: String,
-    /// npm `@openai/codex` 启动版本；运行时发布任务会更新此画像字段。
+    /// 官方 Desktop ZIP 内嵌 Core 的启动基线；运行时按完整制品元组更新。
     pub codex_version: String,
     pub desktop_version: String,
     pub desktop_build: String,
@@ -284,9 +284,9 @@ impl Default for CodexWireProfileConfig {
     fn default() -> Self {
         Self {
             originator: "Codex Desktop".to_owned(),
-            codex_version: "0.146.0".to_owned(),
-            desktop_version: "26.727.51351".to_owned(),
-            desktop_build: "6119".to_owned(),
+            codex_version: "0.147.0-alpha.6.6".to_owned(),
+            desktop_version: "26.803.81509".to_owned(),
+            desktop_build: "6415".to_owned(),
             os_type: "Mac OS".to_owned(),
             os_version: "15.7.1".to_owned(),
             arch: "arm64".to_owned(),

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

@@ -55,6 +55,7 @@ pub struct ImportCodexOAuthCredential {
 pub(crate) struct UnresolvedCodexOAuthCredential {
     pub(crate) account_id: String,
     pub(crate) name: String,
+    pub(crate) installation_id: String,
     pub(crate) secret: CodexOAuthSecret,
     pub(crate) metadata: CodexOAuthMetadata,
     pub(crate) access_token_expires_at: Option<DateTime<Utc>>,
@@ -68,6 +69,7 @@ impl fmt::Debug for UnresolvedCodexOAuthCredential {
             .debug_struct("UnresolvedCodexOAuthCredential")
             .field("account_id", &self.account_id)
             .field("name", &self.name)
+            .field("installation_id", &"<pseudonymous>")
             .field("secret", &"<redacted>")
             .field("metadata", &self.metadata)
             .field("access_token_expires_at", &self.access_token_expires_at)
@@ -473,8 +475,12 @@ impl CodexCredentialAdmin {
             chatgpt_account_id: upstream_account_id,
         } = input.metadata;
         let has_upstream_user_id = upstream_user_id.is_some();
-        let credential = CodexCredentialCodec::encode_unresolved(&input.secret, Vec::new())
-            .map_err(|_| CodexCredentialAdminError::InvalidCredential)?;
+        let credential = CodexCredentialCodec::encode_unresolved(
+            &input.secret,
+            input.installation_id,
+            Vec::new(),
+        )
+        .map_err(|_| CodexCredentialAdminError::InvalidCredential)?;
         let account = ProviderAccount::new(
             account_id,
             provider,
@@ -917,6 +923,7 @@ impl CodexCredentialAdminService {
                                 .or_else(|| metadata.email.clone())
                                 .filter(|name| !name.trim().is_empty())
                                 .unwrap_or_else(|| "Codex OAuth".to_owned()),
+                            installation_id: uuid::Uuid::new_v4().to_string(),
                             secret,
                             metadata,
                             access_token_expires_at,

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

@@ -22,6 +22,8 @@ use url::Url;
 
 use super::repository::{CodexCredentialRepository, CredentialRepositoryError};
 use super::security::{CodexRuntimeAuthentication, CodexRuntimeCredential};
+use crate::transport::headers::build_codex_profile_headers;
+use crate::transport::profile::CodexWireProfileState;
 use crate::transport::websocket::CodexWebSocketExchangeError;
 use crate::transport::{CodexClientError, CodexWebSocketPool};
 
@@ -167,13 +169,21 @@ pub trait CodexAgentIdentityTaskRegistrar: Send + Sync + 'static {
 pub struct OfficialCodexAgentIdentityTaskRegistrar {
     client: reqwest::Client,
     base_url: Url,
+    profile: CodexWireProfileState,
 }
 
 impl OfficialCodexAgentIdentityTaskRegistrar {
-    pub fn new(client: reqwest::Client) -> Result<Self, CodexAgentIdentityError> {
+    pub fn new(
+        client: reqwest::Client,
+        profile: CodexWireProfileState,
+    ) -> Result<Self, CodexAgentIdentityError> {
         let base_url = Url::parse(OFFICIAL_AGENT_IDENTITY_AUTH_BASE_URL)
             .map_err(|_| CodexAgentIdentityError::InvalidConfiguration)?;
-        Ok(Self { client, base_url })
+        Ok(Self {
+            client,
+            base_url,
+            profile,
+        })
     }
 }
 
@@ -197,9 +207,13 @@ impl CodexAgentIdentityTaskRegistrar for OfficialCodexAgentIdentityTaskRegistrar
             .path_segments_mut()
             .map_err(|_| CodexAgentIdentityError::InvalidConfiguration)?
             .extend(["v1", "agent", credential.runtime_id(), "task", "register"]);
+        let profile = self.profile.snapshot();
+        let headers = build_codex_profile_headers(&profile)
+            .map_err(|_| CodexAgentIdentityError::InvalidConfiguration)?;
         let response = self
             .client
             .post(endpoint)
+            .headers(headers)
             .header(reqwest::header::ACCEPT, "application/json")
             .json(&credential.registration(Utc::now()))
             .send()

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

@@ -36,9 +36,16 @@ use super::token_client::{
     OFFICIAL_CODEX_OAUTH_CLIENT_ID, OFFICIAL_CODEX_REDIRECT_URI,
 };
 use super::types::parse_chatgpt_jwt_claims;
+use crate::transport::profile::CodexWireProfileState;
 
 const AUTHORIZATION_ENDPOINT: &str = "https://auth.openai.com/oauth/authorize";
-const AUTHORIZATION_SCOPE: &str = "openid profile email offline_access";
+const DESKTOP_AUTH_ENDPOINT: &str = "https://chatgpt.com/codex/desktop-auth";
+const AUTHORIZATION_SCOPE: &str =
+    "openid profile email offline_access api.connectors.read api.connectors.invoke";
+const AUTHORIZATION_ORIGINATOR: &str = "Codex Desktop";
+const STREAMLINED_LOGIN_QUERY: &str = "codex_streamlined_login";
+const NO_UNIVERSAL_LINKS_QUERY: &str = "no_universal_links";
+const SURFACE_STABLE_ID_DOMAIN: &[u8] = b"openai/codex-desktop/surface-stable-id/v1\0";
 const AUTHORIZATION_TTL: TimeDelta = TimeDelta::minutes(10);
 const AUTHORIZATION_CLAIM_TTL: Duration = Duration::from_secs(90);
 const MAX_CALLBACK_BYTES: usize = 16 * 1024;
@@ -137,6 +144,7 @@ pub struct CodexPendingAuthorization {
     state: SecretString,
     nonce: SecretString,
     code_verifier: SecretString,
+    installation_id: String,
     reauthorization: Option<ProviderAccountId>,
     mutation: PendingAuthorizationMutation,
 }
@@ -150,6 +158,7 @@ pub struct StoredCodexPendingAuthorization {
     pub state: SecretString,
     pub nonce: SecretString,
     pub code_verifier: SecretString,
+    pub installation_id: String,
     pub reauthorization_account_id: Option<String>,
     pub mutation: PendingAuthorizationMutation,
 }
@@ -172,6 +181,7 @@ impl CodexPendingAuthorization {
             state: input.state,
             nonce: input.nonce,
             code_verifier: input.code_verifier,
+            installation_id: input.installation_id,
             reauthorization,
             mutation: input.mutation,
         };
@@ -183,6 +193,7 @@ impl CodexPendingAuthorization {
             || !valid_secret(pending.state.expose_secret())
             || !valid_secret(pending.nonce.expose_secret())
             || !valid_secret(pending.code_verifier.expose_secret())
+            || !valid_installation_id(&pending.installation_id)
             || !pending_mutation_matches(&pending)
         {
             return Err(CodexOAuthPendingStoreError::InvalidValue);
@@ -230,6 +241,11 @@ impl CodexPendingAuthorization {
         &self.code_verifier
     }
 
+    #[must_use]
+    pub fn installation_id(&self) -> &str {
+        &self.installation_id
+    }
+
     #[must_use]
     pub const fn reauthorization(&self) -> Option<&ProviderAccountId> {
         self.reauthorization.as_ref()
@@ -253,6 +269,7 @@ impl fmt::Debug for CodexPendingAuthorization {
             .field("state", &"<redacted>")
             .field("nonce", &"<redacted>")
             .field("code_verifier", &"<redacted>")
+            .field("installation_id", &"<pseudonymous>")
             .field("reauthorization", &self.reauthorization)
             .field("mutation", &self.mutation)
             .finish()
@@ -348,6 +365,7 @@ pub struct CodexOAuthAdminService {
     store: Arc<dyn ProviderAccountStore>,
     credentials: CodexCredentialAdmin,
     oauth_client_id: String,
+    profile: CodexWireProfileState,
 }
 
 struct CodexOAuthAuthorizationCommitGuard {
@@ -397,6 +415,7 @@ impl CodexOAuthAdminService {
         exchanger: Arc<dyn AuthorizationCodeExchanger>,
         store: Arc<dyn ProviderAccountStore>,
         credentials: CodexCredentialAdmin,
+        profile: CodexWireProfileState,
     ) -> Self {
         Self {
             pending,
@@ -404,6 +423,7 @@ impl CodexOAuthAdminService {
             store,
             credentials,
             oauth_client_id: OFFICIAL_CODEX_OAUTH_CLIENT_ID.to_owned(),
+            profile,
         }
     }
 
@@ -420,8 +440,10 @@ impl CodexOAuthAdmin for CodexOAuthAdminService {
         &self,
         command: StartCodexOAuthAuthorization,
     ) -> Result<CodexOAuthAuthorizationStarted, CodexOAuthAdminError> {
-        let (name, reauthorization) = match command.mutation.target() {
-            AuthorizationMutationTarget::Create { name } => (name.clone(), None),
+        let (name, reauthorization, installation_id) = match command.mutation.target() {
+            AuthorizationMutationTarget::Create { name } => {
+                (name.clone(), None, Uuid::new_v4().to_string())
+            }
             AuthorizationMutationTarget::Reauthorize { account_id } => {
                 let current = self
                     .store
@@ -433,10 +455,15 @@ impl CodexOAuthAdmin for CodexOAuthAdminService {
                 {
                     return Err(CodexOAuthAdminError::NotFound);
                 }
-                (current.account.name().to_owned(), Some(account_id.clone()))
+                let installation_id = current_installation_id(&current)?;
+                (
+                    current.account.name().to_owned(),
+                    Some(account_id.clone()),
+                    installation_id,
+                )
             }
         };
-        self.start_pending(command.mutation, name, reauthorization)
+        self.start_pending(command.mutation, name, reauthorization, installation_id)
             .await
     }
 
@@ -488,6 +515,9 @@ impl CodexOAuthAdminService {
                 .load_current_credential(target)
                 .await
                 .map_err(map_store_error)?;
+            if current_installation_id(&current)? != pending.installation_id() {
+                return Err(CodexOAuthAdminError::Conflict);
+            }
             Some(current)
         } else {
             None
@@ -525,6 +555,7 @@ impl CodexOAuthAdminService {
                     .prepare_unresolved_oauth(UnresolvedCodexOAuthCredential {
                         account_id,
                         name: pending.name,
+                        installation_id: pending.installation_id,
                         secret,
                         metadata,
                         access_token_expires_at,
@@ -544,6 +575,7 @@ impl CodexOAuthAdminService {
         mutation: PendingAuthorizationMutation,
         name: String,
         reauthorization: Option<ProviderAccountId>,
+        installation_id: String,
     ) -> Result<CodexOAuthAuthorizationStarted, CodexOAuthAdminError> {
         if !valid_text(&name) {
             return Err(CodexOAuthAdminError::InvalidInput);
@@ -562,17 +594,21 @@ impl CodexOAuthAdminService {
             state: SecretString::from(random_secret()?),
             nonce: SecretString::from(random_secret()?),
             code_verifier: SecretString::from(random_secret()?),
+            installation_id,
             reauthorization_account_id: reauthorization.as_ref().map(ToString::to_string),
             mutation,
         })
         .map_err(map_pending_error)?;
+        let profile = self.profile.snapshot();
+        let authorization_url =
+            authorization_url(&pending, &self.oauth_client_id, &profile.desktop_version)?;
         self.pending
             .create(&pending)
             .await
             .map_err(map_pending_error)?;
         Ok(CodexOAuthAuthorizationStarted {
             flow_id: pending.flow_id.clone(),
-            authorization_url: authorization_url(&pending, &self.oauth_client_id)?,
+            authorization_url,
             expires_at,
         })
     }
@@ -688,27 +724,70 @@ fn current_oauth_refresh_token(
         .and_then(|oauth| oauth.refresh_token.clone()))
 }
 
+fn current_installation_id(current: &LoadedCredential) -> Result<String, CodexOAuthAdminError> {
+    let runtime = CodexCredentialCodec::decode(&current.credential)
+        .map_err(|_| CodexOAuthAdminError::Credential)?;
+    if runtime.authentication.oauth().is_none() {
+        return Err(CodexOAuthAdminError::Credential);
+    }
+    Ok(runtime.installation_id)
+}
+
+fn derive_surface_stable_id(installation_id: &str) -> Result<String, CodexOAuthAdminError> {
+    let installation_id =
+        Uuid::parse_str(installation_id).map_err(|_| CodexOAuthAdminError::Credential)?;
+    let mut digest = Sha256::new();
+    digest.update(SURFACE_STABLE_ID_DOMAIN);
+    digest.update(installation_id.as_bytes());
+    let digest = digest.finalize();
+    let mut bytes = [0_u8; 16];
+    bytes.copy_from_slice(&digest[..16]);
+    bytes[6] = (bytes[6] & 0x0f) | 0x40;
+    bytes[8] = (bytes[8] & 0x3f) | 0x80;
+    Ok(Uuid::from_bytes(bytes).hyphenated().to_string())
+}
+
 fn authorization_url(
     pending: &CodexPendingAuthorization,
     oauth_client_id: &str,
+    desktop_version: &str,
 ) -> Result<String, CodexOAuthAdminError> {
-    let mut url =
+    if !valid_text(desktop_version) {
+        return Err(CodexOAuthAdminError::InvalidInput);
+    }
+    let mut inner =
         Url::parse(AUTHORIZATION_ENDPOINT).map_err(|_| CodexOAuthAdminError::InvalidInput)?;
     let challenge = URL_SAFE_NO_PAD.encode(Sha256::digest(
         pending.code_verifier.expose_secret().as_bytes(),
     ));
-    url.query_pairs_mut()
-        .append_pair("response_type", "code")
-        .append_pair("client_id", oauth_client_id)
-        .append_pair("redirect_uri", OFFICIAL_CODEX_REDIRECT_URI)
-        .append_pair("scope", AUTHORIZATION_SCOPE)
-        .append_pair("state", pending.state.expose_secret())
-        .append_pair("nonce", pending.nonce.expose_secret())
-        .append_pair("code_challenge", &challenge)
-        .append_pair("code_challenge_method", "S256")
-        .append_pair("id_token_add_organizations", "true")
-        .append_pair("codex_cli_simplified_flow", "true");
-    Ok(url.into())
+    let stable_id = derive_surface_stable_id(pending.installation_id())?;
+    {
+        let mut query = inner.query_pairs_mut();
+        query
+            .append_pair("response_type", "code")
+            .append_pair("client_id", oauth_client_id)
+            .append_pair("redirect_uri", OFFICIAL_CODEX_REDIRECT_URI)
+            .append_pair("scope", AUTHORIZATION_SCOPE)
+            .append_pair("code_challenge", challenge.as_str())
+            .append_pair("code_challenge_method", "S256")
+            .append_pair("id_token_add_organizations", "true")
+            .append_pair("codex_cli_simplified_flow", "true")
+            .append_pair("state", pending.state.expose_secret())
+            .append_pair("originator", AUTHORIZATION_ORIGINATOR)
+            .append_pair("codex_app_version", desktop_version)
+            .append_pair("source_surface_stable_id", &stable_id)
+            .append_pair("codex_origin_stable_id", &stable_id)
+            .append_pair(STREAMLINED_LOGIN_QUERY, "true");
+    }
+
+    let mut outer =
+        Url::parse(DESKTOP_AUTH_ENDPOINT).map_err(|_| CodexOAuthAdminError::InvalidInput)?;
+    outer
+        .query_pairs_mut()
+        .append_pair("authorize_url", inner.as_str())
+        .append_pair(STREAMLINED_LOGIN_QUERY, "true")
+        .append_pair(NO_UNIVERSAL_LINKS_QUERY, "1");
+    Ok(outer.into())
 }
 
 fn callback_parts(value: &str) -> Result<(SecretString, SecretString), CodexOAuthAdminError> {
@@ -760,6 +839,12 @@ fn valid_secret(value: &str) -> bool {
     valid_text(value) && value.len() >= 16
 }
 
+fn valid_installation_id(value: &str) -> bool {
+    Uuid::parse_str(value)
+        .ok()
+        .is_some_and(|uuid| uuid.get_version_num() == 4)
+}
+
 pub(crate) fn oauth_owner_ref(owner: &AuthorizationOwner) -> String {
     let mut digest = Sha256::new();
     match owner {

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

@@ -108,6 +108,7 @@ impl CodexCredentialCodec {
                 oauth_subject: account.oauth_subject.clone(),
                 poid: account.poid.clone(),
             }),
+            uuid::Uuid::new_v4().to_string(),
             cookies,
         )
     }
@@ -115,20 +116,22 @@ impl CodexCredentialCodec {
     /// 为尚未解析资料的 OAuth 账号编码凭据。
     pub(crate) fn encode_unresolved(
         secret: &CodexOAuthSecret,
+        installation_id: String,
         cookies: Vec<CodexCookie>,
     ) -> Result<PlaintextCredential, CodexCredentialDataError> {
-        Self::encode_oauth(secret, None, cookies)
+        Self::encode_oauth(secret, None, installation_id, cookies)
     }
 
     fn encode_oauth(
         secret: &CodexOAuthSecret,
         principal: Option<CodexCredentialPrincipal>,
+        installation_id: String,
         cookies: Vec<CodexCookie>,
     ) -> Result<PlaintextCredential, CodexCredentialDataError> {
         Self::encode_complete(CodexCredentialData::OAuth(CodexOAuthCredentialData {
             schema_version: CODEX_CREDENTIAL_SCHEMA_VERSION,
             principal,
-            installation_id: uuid::Uuid::new_v4().to_string(),
+            installation_id,
             access_token: secret.access_token.expose_secret().to_owned(),
             refresh_token: secret
                 .refresh_token

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

@@ -22,8 +22,7 @@ use crate::credential::{
     CodexOAuthAdmin, CodexOAuthAdminService, OfficialCodexAgentIdentityTaskRegistrar,
 };
 use crate::transport::profile::{
-    CodexCliReleaseService, CodexDesktopReleaseService, OfficialCodexCliReleaseTransport,
-    OfficialCodexDesktopReleaseTransport,
+    CodexArtifactProfileCache, CodexDesktopReleaseService, OfficialCodexDesktopReleaseTransport,
 };
 use crate::transport::{CodexWebSocketPool, build_reqwest_client};
 
@@ -37,8 +36,8 @@ pub mod credential;
 pub mod transport;
 
 pub use transport::{
-    CodexCanonicalDecoder, CodexCanonicalError, CodexRequestEncodeError, encode_generate_request,
-    openai_billing_breakdown,
+    CodexCanonicalDecoder, CodexCanonicalError, CodexRequestEncodeError, OpenAiBillingUsage,
+    encode_generate_request, openai_billing_breakdown,
 };
 
 /// OpenAI 初始化后交给组装根的最小能力集。
@@ -63,6 +62,23 @@ pub async fn initialize(
     let runtime_policy = ports.runtime_policy();
     let credential_state = ports.credential_state();
     let profile = config.wire_profile_state();
+    let artifact_cache =
+        CodexArtifactProfileCache::new(provider_kind.clone(), ports.artifact_profiles());
+    let configured_build = profile.snapshot().desktop_build.parse::<u64>().ok();
+    let verified_profile = match artifact_cache.load().await {
+        Ok(cached) => cached.filter(|cached| {
+            cached
+                .desktop_build
+                .parse::<u64>()
+                .ok()
+                .zip(configured_build)
+                .is_some_and(|(cached, configured)| cached >= configured)
+        }),
+        Err(error) => {
+            tracing::warn!(error = %error, "OpenAI artifact profile cache could not be loaded");
+            None
+        }
+    };
     let session_identity = config
         .session_identity()
         .map_err(|_| OpenAiInitializeError::SessionIdentity)?;
@@ -73,15 +89,10 @@ pub async fn initialize(
             OfficialCodexDesktopReleaseTransport::new()
                 .map_err(|_| OpenAiInitializeError::DesktopRelease)?,
         ),
+        artifact_cache,
+        verified_profile,
     ));
     let desktop_release_status = desktop_release.status();
-    let cli_release = Arc::new(CodexCliReleaseService::new(
-        profile.clone(),
-        Arc::new(
-            OfficialCodexCliReleaseTransport::new()
-                .map_err(|_| OpenAiInitializeError::CliRelease)?,
-        ),
-    ));
     let repository = CodexCredentialRepository::new(Arc::clone(&accounts));
     let websocket_pool = Arc::new(CodexWebSocketPool::with_config(
         config.websocket_pool_config(),
@@ -89,7 +100,7 @@ pub async fn initialize(
     let agent_identity = Arc::new(CodexAgentIdentityTaskService::new(
         repository.clone(),
         Arc::new(
-            OfficialCodexAgentIdentityTaskRegistrar::new(http.clone())
+            OfficialCodexAgentIdentityTaskRegistrar::new(http.clone(), profile.clone())
                 .map_err(|_| OpenAiInitializeError::AgentIdentity)?,
         ),
         Arc::clone(&websocket_pool),
@@ -165,6 +176,7 @@ pub async fn initialize(
             exchanger,
             Arc::clone(&accounts),
             CodexCredentialAdmin,
+            profile.clone(),
         )
         .with_oauth_client_id(config.oauth_client_id()),
     );
@@ -187,7 +199,6 @@ pub async fn initialize(
         catalog,
         config.quota_refresh_policy(),
         config.oauth_refresh_enabled(),
-        cli_release,
         desktop_release,
     )
     .map_err(|_| OpenAiInitializeError::Worker)?;
@@ -243,8 +254,6 @@ pub enum OpenAiInitializeError {
     Refresh,
     #[error("OpenAI Desktop release service could not initialize")]
     DesktopRelease,
-    #[error("OpenAI CLI release service could not initialize")]
-    CliRelease,
     #[error("OpenAI worker plan is invalid")]
     Worker,
 }

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

@@ -57,8 +57,7 @@ use crate::transport::diagnostics::{
     CodexFailureCategory, CodexUpstreamFailure, CodexUpstreamSendPhase,
 };
 use crate::transport::profile::{
-    APPCAST_POLL_INTERVAL, CodexCliReleaseService, CodexDesktopReleaseService,
-    CodexWireProfileState,
+    APPCAST_POLL_INTERVAL, CodexDesktopReleaseService, CodexWireProfileState,
 };
 use crate::transport::protocol::responses::{
     CodexResponsesRequest, PreviousResponseScope, ResponseEventSignals,
@@ -737,6 +736,7 @@ fn cold_response_stream(response: ColdResponse) -> EventStream {
         // （transport/protocol/websocket.rs），push_frames 抽出的 data 即上游原文。
         let mut decoder = CodexCanonicalDecoder::new(upstream_model.as_str())
             .with_requested_service_tier(request.service_tier())
+            .with_request_tool_pricing(upstream_model.as_str(), request.tools())
             .with_raw_sse_passthrough();
         let mut pre_commit_events = PreCommitClientEvents::new();
         loop {
@@ -2339,7 +2339,6 @@ const WORKER_LEASE_TTL: Duration = Duration::from_secs(15 * 60);
 const WORKER_LEASE_RENEWAL: Duration = Duration::from_secs(5 * 60);
 const OAUTH_REFRESH_INTERVAL: Duration = Duration::from_secs(30);
 const DESKTOP_RELEASE_WORKER_OWNER: &str = "openai-desktop-release";
-const CLI_RELEASE_WORKER_OWNER: &str = "openai-cli-release";
 const MODEL_ETAG_WORKER_OWNER: &str = "openai-model-etag";
 
 pub(crate) fn worker_contributions(
@@ -2348,7 +2347,6 @@ pub(crate) fn worker_contributions(
     catalog: Arc<CodexCredentialCatalogService>,
     quota_refresh_policy: CodexQuotaRefreshPolicy,
     oauth_refresh_enabled: bool,
-    cli_release: Arc<CodexCliReleaseService>,
     desktop_release: Arc<CodexDesktopReleaseService>,
 ) -> Result<Vec<WorkerContribution>, WorkerDefinitionError> {
     let refresh_id = WorkerId::try_new(WorkerKind::OAuthRefresh, PROVIDER_NAME)?;
@@ -2356,8 +2354,6 @@ pub(crate) fn worker_contributions(
     let etag_id = WorkerId::try_new(WorkerKind::QuotaCatalogHealth, MODEL_ETAG_WORKER_OWNER)?;
     let desktop_release_id =
         WorkerId::try_new(WorkerKind::QuotaCatalogHealth, DESKTOP_RELEASE_WORKER_OWNER)?;
-    let cli_release_id =
-        WorkerId::try_new(WorkerKind::QuotaCatalogHealth, CLI_RELEASE_WORKER_OWNER)?;
     let mut contributions = Vec::new();
     if oauth_refresh_enabled {
         contributions.push(WorkerContribution::Registration(scheduled_registration(
@@ -2385,13 +2381,6 @@ pub(crate) fn worker_contributions(
                 task: Box::new(OpenAiCatalogEtagTask { catalog }),
             },
         )?),
-        WorkerContribution::Registration(scheduled_registration(
-            cli_release_id,
-            APPCAST_POLL_INTERVAL,
-            Box::new(OpenAiCliReleaseTask {
-                service: cli_release,
-            }),
-        )?),
         WorkerContribution::Registration(scheduled_registration(
             desktop_release_id,
             APPCAST_POLL_INTERVAL,
@@ -2512,27 +2501,6 @@ struct OpenAiDesktopReleaseTask {
     service: Arc<CodexDesktopReleaseService>,
 }
 
-struct OpenAiCliReleaseTask {
-    service: Arc<CodexCliReleaseService>,
-}
-
-impl ScheduledTask for OpenAiCliReleaseTask {
-    fn run_cycle(&self, context: WorkerCycleContext) -> BoxFuture<'_, Result<(), WorkerTaskError>> {
-        Box::pin(async move {
-            let refresh = self.service.refresh();
-            tokio::pin!(refresh);
-            let result = tokio::select! {
-                () = context.cancellation().cancelled() => return Ok(()),
-                result = &mut refresh => result,
-            };
-            if let Err(error) = result {
-                tracing::warn!(error = %error, "OpenAI CLI release check failed");
-            }
-            Ok(())
-        })
-    }
-}
-
 impl ScheduledTask for OpenAiDesktopReleaseTask {
     fn run_cycle(&self, context: WorkerCycleContext) -> BoxFuture<'_, Result<(), WorkerTaskError>> {
         Box::pin(async move {

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

@@ -21,7 +21,10 @@ use thiserror::Error;
 use super::protocol::responses::{
     ResponseEventSignals, ResponsesSseFailure, response_event_signals,
 };
-use super::usage::{normalize_service_tier, openai_billing_breakdown};
+use super::usage::{
+    OpenAiBillingUsage, WebSearchPricing, normalize_service_tier, openai_billing_breakdown,
+    web_search_pricing,
+};
 
 const CONTENTS_PER_OUTPUT: u32 = 1_024;
 
@@ -43,6 +46,7 @@ pub struct CodexCanonicalDecoder {
     semantic_output_seen: bool,
     requested_service_tier: Option<String>,
     response_service_tier: Option<String>,
+    web_search_pricing: Option<WebSearchPricing>,
     timing_signals: ResponseEventSignals,
     raw_sse_passthrough: bool,
 }
@@ -139,6 +143,7 @@ impl CodexCanonicalDecoder {
             semantic_output_seen: false,
             requested_service_tier: None,
             response_service_tier: None,
+            web_search_pricing: None,
             timing_signals: ResponseEventSignals::default(),
             raw_sse_passthrough: false,
         }
@@ -160,6 +165,13 @@ impl CodexCanonicalDecoder {
         self
     }
 
+    /// 记录请求工具类型，用于区分标准与预览 Web Search 的按次价格。
+    #[must_use]
+    pub fn with_request_tool_pricing(mut self, model: &str, tools: Option<&[Value]>) -> Self {
+        self.web_search_pricing = web_search_pricing(model, tools);
+        self
+    }
+
     pub fn push(&mut self, chunk: &[u8]) -> CodexCanonicalOutcome {
         self.timing_signals = ResponseEventSignals::default();
         if self.raw_sse_passthrough {
@@ -849,15 +861,15 @@ impl CodexCanonicalDecoder {
             .requested_service_tier
             .as_deref()
             .or(self.response_service_tier.as_deref());
+        let web_search_calls = web_search_call_count(response);
         if let Some(breakdown) = usage
             .filter(|usage| billable_usage_is_complete(response, *usage))
             .and_then(|usage| {
+                let web_search_calls = web_search_calls?;
                 openai_billing_breakdown(
                     &model,
-                    usage.input_tokens,
-                    usage.output_tokens,
-                    usage.cached_tokens,
-                    usage.cache_write_tokens,
+                    OpenAiBillingUsage::from(usage)
+                        .with_web_search_calls(web_search_calls, self.web_search_pricing),
                     service_tier,
                 )
             })
@@ -984,6 +996,28 @@ fn core_usage(usage: TokenUsage) -> Usage {
     normalized
 }
 
+fn web_search_call_count(response: &Value) -> Option<u64> {
+    let mut calls = 0_u64;
+    for item in response
+        .get("output")
+        .and_then(Value::as_array)
+        .into_iter()
+        .flatten()
+        .filter(|item| item.get("type").and_then(Value::as_str) == Some("web_search_call"))
+    {
+        match item
+            .get("action")
+            .and_then(|action| action.get("type"))
+            .and_then(Value::as_str)
+        {
+            Some("search") => calls = calls.checked_add(1)?,
+            Some("open_page" | "find_in_page") => {}
+            Some(_) | None => return None,
+        }
+    }
+    Some(calls)
+}
+
 fn billable_usage_is_complete(response: &Value, usage: TokenUsage) -> bool {
     let Some(raw) = response.get("usage").filter(|value| value.is_object()) else {
         return false;

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

@@ -677,7 +677,13 @@ impl CodexBackendClient {
             // API 透传黑名单也不能覆盖画像，避免下游客户端暴露不一致指纹。
             if matches!(
                 name.as_str(),
-                "originator" | "user-agent" | "x-oai-attestation" | "x-oai-is" | "x-oai-is-update"
+                "openai-beta"
+                    | "originator"
+                    | "user-agent"
+                    | "x-oai-attestation"
+                    | "x-oai-is"
+                    | "x-oai-is-update"
+                    | "x-openai-internal-codex-residency"
             ) {
                 continue;
             }
@@ -696,6 +702,10 @@ impl CodexBackendClient {
     ) -> CodexClientResult<HeaderMap> {
         let mut headers = self.request_headers_for_http_response(request, context)?;
         headers.remove(X_OPENAI_INTERNAL_CODEX_RESPONSES_LITE_HEADER);
+        headers.insert(
+            HeaderName::from_static("openai-beta"),
+            HeaderValue::from_static("responses_websockets=2026-02-06"),
+        );
         Ok(headers)
     }
 
@@ -709,14 +719,6 @@ impl CodexBackendClient {
         headers.insert(CONTENT_TYPE, HeaderValue::from_static("application/json"));
         insert_optional_header(&mut headers, "cookie", context.cookie_header)?;
         headers.insert(ACCEPT, HeaderValue::from_static("text/event-stream"));
-        headers.insert(
-            HeaderName::from_static("openai-beta"),
-            HeaderValue::from_static("responses_websockets=2026-02-06"),
-        );
-        headers.insert(
-            HeaderName::from_static("x-openai-internal-codex-residency"),
-            HeaderValue::from_static("us"),
-        );
         headers.insert(
             HeaderName::from_static("x-client-request-id"),
             HeaderValue::from_str(context.request_id)?,
@@ -788,10 +790,6 @@ impl CodexBackendClient {
         let mut headers = HeaderMap::new();
         headers.insert(USER_AGENT, HeaderValue::from_str(&profile.user_agent())?);
         headers.insert(AUTHORIZATION, HeaderValue::from_str(context.authorization)?);
-        headers.insert(
-            HeaderName::from_static("originator"),
-            HeaderValue::from_str(&profile.originator)?,
-        );
         insert_optional_header(&mut headers, "chatgpt-account-id", context.account_id)?;
         insert_optional_header(&mut headers, "cookie", context.cookie_header)?;
         headers.insert(ACCEPT, HeaderValue::from_static("application/json"));

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

@@ -1,7 +1,5 @@
 /// `/codex/responses`
 pub const CODEX_RESPONSES_PATH: &str = "/codex/responses";
-/// `/codex/usage`
-pub const CODEX_USAGE_PATH: &str = "/codex/usage";
 /// `/api/codex/usage`
 pub const CODEX_USAGE_API_PATH: &str = "/api/codex/usage";
 /// `/wham/usage`
@@ -16,20 +14,14 @@ pub fn endpoint_url(base_url: &str, endpoint_path: &str) -> String {
     )
 }
 
-/// 返回 usage 相关 endpoint 的完整 URL 列表。
-pub fn usage_endpoint_urls(base_url: &str) -> Vec<String> {
-    usage_endpoint_paths(base_url)
-        .into_iter()
-        .map(|path| endpoint_url(base_url, path))
-        .collect()
-}
-
-fn usage_endpoint_paths(base_url: &str) -> Vec<&'static str> {
-    if has_backend_api_base_path(base_url) {
-        vec![WHAM_USAGE_PATH, CODEX_USAGE_PATH]
+/// 返回与 base path 对应的唯一 usage endpoint。
+pub fn usage_endpoint_url(base_url: &str) -> String {
+    let path = if has_backend_api_base_path(base_url) {
+        WHAM_USAGE_PATH
     } else {
-        vec![CODEX_USAGE_API_PATH, CODEX_USAGE_PATH]
-    }
+        CODEX_USAGE_API_PATH
+    };
+    endpoint_url(base_url, path)
 }
 
 fn has_backend_api_base_path(base_url: &str) -> bool {

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

@@ -3,20 +3,33 @@ use reqwest::header::{AUTHORIZATION, HeaderMap, HeaderName, HeaderValue, USER_AG
 
 use super::client::CodexClientResult;
 
+const CODEX_RESIDENCY_HEADER: &str = "x-openai-internal-codex-residency";
+
 /// 构造 Codex Core 为模型请求设置的稳定身份请求头。
 pub fn build_codex_base_headers(
     profile: &CodexWireProfile,
     authorization: &str,
     account_id: Option<&str>,
 ) -> CodexClientResult<HeaderMap> {
-    let mut headers = HeaderMap::new();
+    let mut headers = build_codex_profile_headers(profile)?;
     headers.insert(AUTHORIZATION, HeaderValue::from_str(authorization)?);
     insert_optional_header(&mut headers, "chatgpt-account-id", account_id)?;
+    Ok(headers)
+}
+
+/// 构造不依赖访问凭据的 Codex Desktop 身份请求头。
+pub fn build_codex_profile_headers(profile: &CodexWireProfile) -> CodexClientResult<HeaderMap> {
+    let mut headers = HeaderMap::new();
     headers.insert(
         HeaderName::from_static("originator"),
         HeaderValue::from_str(&profile.originator)?,
     );
     headers.insert(USER_AGENT, HeaderValue::from_str(&profile.user_agent())?);
+    // 此部署固定采用官方支持的 US residency requirement。
+    headers.insert(
+        HeaderName::from_static(CODEX_RESIDENCY_HEADER),
+        HeaderValue::from_static("us"),
+    );
     Ok(headers)
 }
 

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

@@ -32,13 +32,13 @@ pub use self::{
     },
     diagnostics::{CodexUpstreamDiagnostics, CodexUpstreamSendPhase},
     endpoints::{
-        CODEX_RESPONSES_PATH, CODEX_USAGE_API_PATH, CODEX_USAGE_PATH, WHAM_USAGE_PATH,
-        endpoint_url, usage_endpoint_urls,
+        CODEX_RESPONSES_PATH, CODEX_USAGE_API_PATH, WHAM_USAGE_PATH, endpoint_url,
+        usage_endpoint_url,
     },
     headers::build_codex_base_headers,
     request::{CodexRequestEncodeError, encode_generate_request},
     response_meta::CodexResponseMetadata,
-    usage::{MAX_CODEX_USAGE_BODY_BYTES, openai_billing_breakdown},
+    usage::{MAX_CODEX_USAGE_BODY_BYTES, OpenAiBillingUsage, openai_billing_breakdown},
     websocket::{
         CodexWebSocketPool, CodexWebSocketPoolConfig, CodexWebSocketPoolKey, WebSocketPoolDecision,
     },

backend/crates/providers/openai/src/transport/profile/desktop_artifact.rs

@@ -0,0 +1,600 @@
+//! 有界读取 Codex Desktop ZIP 内嵌的 arm64 Core 版本。
+
+use std::cmp;
+
+use flate2::{Decompress, FlushDecompress, Status};
+use futures::StreamExt as _;
+use reqwest::header::{CONTENT_LENGTH, CONTENT_RANGE, RANGE};
+use reqwest::{Client, StatusCode};
+use url::Url;
+
+const OFFICIAL_ARTIFACT_HOST: &str = "persistent.oaistatic.com";
+const EOCD_SIGNATURE: &[u8; 4] = b"PK\x05\x06";
+const CENTRAL_SIGNATURE: &[u8; 4] = b"PK\x01\x02";
+const LOCAL_SIGNATURE: &[u8; 4] = b"PK\x03\x04";
+const CORE_PATH_SUFFIX: &[u8] = b"/Contents/Resources/codex";
+const CORE_VERSION_ANCHOR: &[u8] = b"codex-mcp-client/";
+const MAX_EOCD_SEARCH_BYTES: u64 = 65_535 + 22;
+const MAX_CENTRAL_DIRECTORY_BYTES: u64 = 16 * 1024 * 1024;
+const MAX_RANGE_BYTES: u64 = 16 * 1024 * 1024;
+const CORE_RANGE_CHUNK_BYTES: u64 = 4 * 1024 * 1024;
+const MAX_CORE_COMPRESSED_BYTES: u64 = 256 * 1024 * 1024;
+const MAX_CORE_UNCOMPRESSED_BYTES: u64 = 512 * 1024 * 1024;
+const MAX_VERSION_BYTES: usize = 64;
+const MACHO_HEADER_BYTES: usize = 8;
+const MACHO_64_LE_MAGIC: [u8; 4] = [0xcf, 0xfa, 0xed, 0xfe];
+const MACHO_ARM64_CPU_TYPE: u32 = 0x0100_000c;
+
+#[derive(Debug, thiserror::Error)]
+pub enum CodexDesktopArtifactError {
+    #[error("Codex Desktop artifact URL is invalid")]
+    InvalidUrl,
+    #[error("Codex Desktop artifact URL is not an approved official ZIP")]
+    UnapprovedUrl,
+    #[error("Codex Desktop artifact size is invalid")]
+    InvalidSize,
+    #[error("Codex Desktop artifact range request failed")]
+    Http(#[from] reqwest::Error),
+    #[error("Codex Desktop artifact did not honor an exact range request")]
+    RangeUnsupported,
+    #[error("Codex Desktop artifact returned an invalid range")]
+    InvalidRange,
+    #[error("Codex Desktop artifact ZIP metadata is invalid or unsupported")]
+    InvalidArchive,
+    #[error("Codex Desktop artifact does not contain exactly one bundled Core")]
+    MissingCore,
+    #[error("Codex Desktop bundled Core format is unsupported")]
+    UnsupportedCore,
+    #[error("Codex Desktop bundled Core exceeds the scan limit")]
+    CoreTooLarge,
+    #[error("Codex Desktop bundled Core version marker is missing or invalid")]
+    MissingCoreVersion,
+}
+
+#[derive(Debug)]
+struct CentralDirectory {
+    offset: u64,
+    size: u64,
+    entries: u16,
+}
+
+#[derive(Debug)]
+pub(super) struct CoreEntry {
+    pub(super) name: Vec<u8>,
+    flags: u16,
+    compression: u16,
+    pub(super) compressed_size: u64,
+    pub(super) uncompressed_size: u64,
+    pub(super) local_header_offset: u64,
+}
+
+pub(super) async fn fetch_codex_core_version(
+    client: &Client,
+    artifact_url: &str,
+    artifact_size: u64,
+) -> Result<String, CodexDesktopArtifactError> {
+    let url = validate_artifact(artifact_url, artifact_size)?;
+    let tail_start = artifact_size.saturating_sub(MAX_EOCD_SEARCH_BYTES);
+    let tail = read_range(client, &url, artifact_size, tail_start, artifact_size - 1).await?;
+    let directory = parse_eocd(&tail, tail_start, artifact_size)?;
+    let directory_end = directory
+        .offset
+        .checked_add(directory.size)
+        .and_then(|end| end.checked_sub(1))
+        .ok_or(CodexDesktopArtifactError::InvalidArchive)?;
+    let central = if directory.offset >= tail_start && directory_end < artifact_size {
+        let start = usize::try_from(directory.offset - tail_start)
+            .map_err(|_| CodexDesktopArtifactError::InvalidArchive)?;
+        let end = usize::try_from(directory_end - tail_start + 1)
+            .map_err(|_| CodexDesktopArtifactError::InvalidArchive)?;
+        tail.get(start..end)
+            .ok_or(CodexDesktopArtifactError::InvalidArchive)?
+            .to_vec()
+    } else {
+        read_range(client, &url, artifact_size, directory.offset, directory_end).await?
+    };
+    let entry = find_core_entry(&central, directory.entries)?;
+    let data_offset = core_data_offset(client, &url, artifact_size, &entry).await?;
+    scan_core_version(client, &url, artifact_size, data_offset, &entry).await
+}
+
+pub(super) fn validate_codex_artifact(
+    artifact_url: &str,
+    artifact_size: u64,
+) -> Result<(), CodexDesktopArtifactError> {
+    validate_artifact(artifact_url, artifact_size).map(drop)
+}
+
+fn validate_artifact(
+    artifact_url: &str,
+    artifact_size: u64,
+) -> Result<Url, CodexDesktopArtifactError> {
+    if artifact_size == 0 || artifact_size > u64::from(u32::MAX) {
+        return Err(CodexDesktopArtifactError::InvalidSize);
+    }
+    let url = Url::parse(artifact_url).map_err(|_| CodexDesktopArtifactError::InvalidUrl)?;
+    if url.scheme() != "https"
+        || url.host_str() != Some(OFFICIAL_ARTIFACT_HOST)
+        || !url.path().to_ascii_lowercase().ends_with(".zip")
+        || url.username() != ""
+        || url.password().is_some()
+        || url.fragment().is_some()
+    {
+        return Err(CodexDesktopArtifactError::UnapprovedUrl);
+    }
+    Ok(url)
+}
+
+async fn read_range(
+    client: &Client,
+    url: &Url,
+    artifact_size: u64,
+    start: u64,
+    end: u64,
+) -> Result<Vec<u8>, CodexDesktopArtifactError> {
+    let expected = end
+        .checked_sub(start)
+        .and_then(|length| length.checked_add(1))
+        .filter(|length| *length <= MAX_RANGE_BYTES)
+        .ok_or(CodexDesktopArtifactError::InvalidRange)?;
+    let response = client
+        .get(url.clone())
+        .header(RANGE, format!("bytes={start}-{end}"))
+        .send()
+        .await?;
+    if response.status() != StatusCode::PARTIAL_CONTENT {
+        return Err(CodexDesktopArtifactError::RangeUnsupported);
+    }
+    let content_range = response
+        .headers()
+        .get(CONTENT_RANGE)
+        .and_then(|value| value.to_str().ok())
+        .ok_or(CodexDesktopArtifactError::InvalidRange)?;
+    if parse_content_range(content_range) != Some((start, end, artifact_size)) {
+        return Err(CodexDesktopArtifactError::InvalidRange);
+    }
+    if response
+        .headers()
+        .get(CONTENT_LENGTH)
+        .and_then(|value| value.to_str().ok())
+        .and_then(|value| value.parse::<u64>().ok())
+        != Some(expected)
+    {
+        return Err(CodexDesktopArtifactError::InvalidRange);
+    }
+
+    let capacity =
+        usize::try_from(expected).map_err(|_| CodexDesktopArtifactError::InvalidRange)?;
+    let mut body = Vec::with_capacity(capacity);
+    let mut stream = response.bytes_stream();
+    while let Some(chunk) = stream.next().await {
+        let chunk = chunk?;
+        if body
+            .len()
+            .checked_add(chunk.len())
+            .is_none_or(|length| length > capacity)
+        {
+            return Err(CodexDesktopArtifactError::InvalidRange);
+        }
+        body.extend_from_slice(&chunk);
+    }
+    if body.len() != capacity {
+        return Err(CodexDesktopArtifactError::InvalidRange);
+    }
+    Ok(body)
+}
+
+pub(super) fn parse_content_range(value: &str) -> Option<(u64, u64, u64)> {
+    let value = value.strip_prefix("bytes ")?;
+    let (range, size) = value.split_once('/')?;
+    let (start, end) = range.split_once('-')?;
+    Some((start.parse().ok()?, end.parse().ok()?, size.parse().ok()?))
+}
+
+fn parse_eocd(
+    tail: &[u8],
+    tail_start: u64,
+    artifact_size: u64,
+) -> Result<CentralDirectory, CodexDesktopArtifactError> {
+    let position = tail
+        .windows(EOCD_SIGNATURE.len())
+        .rposition(|window| window == EOCD_SIGNATURE)
+        .ok_or(CodexDesktopArtifactError::InvalidArchive)?;
+    let eocd = tail
+        .get(position..)
+        .filter(|value| value.len() >= 22)
+        .ok_or(CodexDesktopArtifactError::InvalidArchive)?;
+    let comment_length = usize::from(le_u16(eocd, 20)?);
+    if eocd.len() != 22 + comment_length || le_u16(eocd, 4)? != 0 || le_u16(eocd, 6)? != 0 {
+        return Err(CodexDesktopArtifactError::InvalidArchive);
+    }
+    let entries_on_disk = le_u16(eocd, 8)?;
+    let entries = le_u16(eocd, 10)?;
+    let size = u64::from(le_u32(eocd, 12)?);
+    let offset = u64::from(le_u32(eocd, 16)?);
+    if entries == 0
+        || entries != entries_on_disk
+        || entries == u16::MAX
+        || size == u64::from(u32::MAX)
+        || offset == u64::from(u32::MAX)
+        || size == 0
+        || size > MAX_CENTRAL_DIRECTORY_BYTES
+    {
+        return Err(CodexDesktopArtifactError::InvalidArchive);
+    }
+    let eocd_offset = tail_start
+        .checked_add(
+            u64::try_from(position).map_err(|_| CodexDesktopArtifactError::InvalidArchive)?,
+        )
+        .ok_or(CodexDesktopArtifactError::InvalidArchive)?;
+    if offset
+        .checked_add(size)
+        .is_none_or(|end| end > eocd_offset || end > artifact_size)
+    {
+        return Err(CodexDesktopArtifactError::InvalidArchive);
+    }
+    Ok(CentralDirectory {
+        offset,
+        size,
+        entries,
+    })
+}
+
+pub(super) fn find_core_entry(
+    central: &[u8],
+    expected_entries: u16,
+) -> Result<CoreEntry, CodexDesktopArtifactError> {
+    let mut cursor = 0_usize;
+    let mut core = None;
+    for _ in 0..expected_entries {
+        let header = central
+            .get(cursor..cursor.saturating_add(46))
+            .filter(|header| header.get(..4) == Some(CENTRAL_SIGNATURE.as_slice()))
+            .ok_or(CodexDesktopArtifactError::InvalidArchive)?;
+        let name_length = usize::from(le_u16(header, 28)?);
+        let extra_length = usize::from(le_u16(header, 30)?);
+        let comment_length = usize::from(le_u16(header, 32)?);
+        let entry_length = 46_usize
+            .checked_add(name_length)
+            .and_then(|length| length.checked_add(extra_length))
+            .and_then(|length| length.checked_add(comment_length))
+            .ok_or(CodexDesktopArtifactError::InvalidArchive)?;
+        let entry = central
+            .get(cursor..cursor.saturating_add(entry_length))
+            .ok_or(CodexDesktopArtifactError::InvalidArchive)?;
+        let name = entry
+            .get(46..46 + name_length)
+            .ok_or(CodexDesktopArtifactError::InvalidArchive)?;
+        if name.ends_with(CORE_PATH_SUFFIX) {
+            if core.is_some() {
+                return Err(CodexDesktopArtifactError::MissingCore);
+            }
+            let flags = le_u16(header, 8)?;
+            let compression = le_u16(header, 10)?;
+            let compressed_size = u64::from(le_u32(header, 20)?);
+            let uncompressed_size = u64::from(le_u32(header, 24)?);
+            let local_header_offset = u64::from(le_u32(header, 42)?);
+            if flags & 1 != 0
+                || compression != 8
+                || compressed_size == 0
+                || compressed_size == u64::from(u32::MAX)
+                || compressed_size > MAX_CORE_COMPRESSED_BYTES
+                || uncompressed_size == 0
+                || uncompressed_size == u64::from(u32::MAX)
+                || uncompressed_size > MAX_CORE_UNCOMPRESSED_BYTES
+                || local_header_offset == u64::from(u32::MAX)
+            {
+                return Err(CodexDesktopArtifactError::UnsupportedCore);
+            }
+            core = Some(CoreEntry {
+                name: name.to_vec(),
+                flags,
+                compression,
+                compressed_size,
+                uncompressed_size,
+                local_header_offset,
+            });
+        }
+        cursor = cursor
+            .checked_add(entry_length)
+            .ok_or(CodexDesktopArtifactError::InvalidArchive)?;
+    }
+    core.ok_or(CodexDesktopArtifactError::MissingCore)
+}
+
+async fn core_data_offset(
+    client: &Client,
+    url: &Url,
+    artifact_size: u64,
+    entry: &CoreEntry,
+) -> Result<u64, CodexDesktopArtifactError> {
+    let fixed_end = entry
+        .local_header_offset
+        .checked_add(29)
+        .filter(|end| *end < artifact_size)
+        .ok_or(CodexDesktopArtifactError::InvalidArchive)?;
+    let fixed = read_range(
+        client,
+        url,
+        artifact_size,
+        entry.local_header_offset,
+        fixed_end,
+    )
+    .await?;
+    if fixed.get(..4) != Some(LOCAL_SIGNATURE.as_slice())
+        || le_u16(&fixed, 6)? != entry.flags
+        || le_u16(&fixed, 8)? != entry.compression
+    {
+        return Err(CodexDesktopArtifactError::InvalidArchive);
+    }
+    let name_length = u64::from(le_u16(&fixed, 26)?);
+    let extra_length = u64::from(le_u16(&fixed, 28)?);
+    let variable_length = name_length
+        .checked_add(extra_length)
+        .ok_or(CodexDesktopArtifactError::InvalidArchive)?;
+    let variable_start = entry
+        .local_header_offset
+        .checked_add(30)
+        .ok_or(CodexDesktopArtifactError::InvalidArchive)?;
+    let data_offset = variable_start
+        .checked_add(variable_length)
+        .ok_or(CodexDesktopArtifactError::InvalidArchive)?;
+    let data_end = data_offset
+        .checked_add(entry.compressed_size)
+        .ok_or(CodexDesktopArtifactError::InvalidArchive)?;
+    if data_end > artifact_size || name_length == 0 {
+        return Err(CodexDesktopArtifactError::InvalidArchive);
+    }
+    let variable_end = data_offset
+        .checked_sub(1)
+        .ok_or(CodexDesktopArtifactError::InvalidArchive)?;
+    let variable = read_range(client, url, artifact_size, variable_start, variable_end).await?;
+    let local_name_length =
+        usize::try_from(name_length).map_err(|_| CodexDesktopArtifactError::InvalidArchive)?;
+    if variable.get(..local_name_length) != Some(entry.name.as_slice()) {
+        return Err(CodexDesktopArtifactError::InvalidArchive);
+    }
+    Ok(data_offset)
+}
+
+async fn scan_core_version(
+    client: &Client,
+    url: &Url,
+    artifact_size: u64,
+    data_offset: u64,
+    entry: &CoreEntry,
+) -> Result<String, CodexDesktopArtifactError> {
+    let mut decompressor = Decompress::new(false);
+    let mut scanner = CoreVersionScanner::default();
+    let mut macho_header = Vec::with_capacity(MACHO_HEADER_BYTES);
+    let mut compressed_read = 0_u64;
+    let mut uncompressed_read = 0_u64;
+    while compressed_read < entry.compressed_size {
+        let chunk_length = cmp::min(
+            CORE_RANGE_CHUNK_BYTES,
+            entry.compressed_size - compressed_read,
+        );
+        let start = data_offset
+            .checked_add(compressed_read)
+            .ok_or(CodexDesktopArtifactError::InvalidArchive)?;
+        let end = start
+            .checked_add(chunk_length - 1)
+            .ok_or(CodexDesktopArtifactError::InvalidArchive)?;
+        let compressed = read_range(client, url, artifact_size, start, end).await?;
+        compressed_read = compressed_read
+            .checked_add(chunk_length)
+            .ok_or(CodexDesktopArtifactError::CoreTooLarge)?;
+        let final_chunk = compressed_read == entry.compressed_size;
+        let mut input_offset = 0_usize;
+        loop {
+            if input_offset == compressed.len() && !final_chunk {
+                break;
+            }
+            let mut output = [0_u8; 64 * 1024];
+            let input_before = decompressor.total_in();
+            let output_before = decompressor.total_out();
+            let status = decompressor
+                .decompress(
+                    &compressed[input_offset..],
+                    &mut output,
+                    if final_chunk {
+                        FlushDecompress::Finish
+                    } else {
+                        FlushDecompress::None
+                    },
+                )
+                .map_err(|_| CodexDesktopArtifactError::UnsupportedCore)?;
+            let consumed = usize::try_from(decompressor.total_in() - input_before)
+                .map_err(|_| CodexDesktopArtifactError::CoreTooLarge)?;
+            let produced = usize::try_from(decompressor.total_out() - output_before)
+                .map_err(|_| CodexDesktopArtifactError::CoreTooLarge)?;
+            if consumed == 0 && produced == 0 {
+                return Err(CodexDesktopArtifactError::UnsupportedCore);
+            }
+            input_offset = input_offset
+                .checked_add(consumed)
+                .ok_or(CodexDesktopArtifactError::CoreTooLarge)?;
+            uncompressed_read = uncompressed_read
+                .checked_add(
+                    u64::try_from(produced).map_err(|_| CodexDesktopArtifactError::CoreTooLarge)?,
+                )
+                .filter(|read| {
+                    *read <= entry.uncompressed_size && *read <= MAX_CORE_UNCOMPRESSED_BYTES
+                })
+                .ok_or(CodexDesktopArtifactError::CoreTooLarge)?;
+            let output = &output[..produced];
+            if macho_header.len() < MACHO_HEADER_BYTES {
+                let needed = MACHO_HEADER_BYTES - macho_header.len();
+                macho_header.extend_from_slice(&output[..cmp::min(needed, output.len())]);
+                if macho_header.len() == MACHO_HEADER_BYTES {
+                    validate_macho_header(&macho_header)?;
+                }
+            }
+            if let Some(version) = scanner.push(output)? {
+                if macho_header.len() != MACHO_HEADER_BYTES {
+                    return Err(CodexDesktopArtifactError::UnsupportedCore);
+                }
+                return Ok(version);
+            }
+            if status == Status::StreamEnd {
+                if !final_chunk
+                    || input_offset != compressed.len()
+                    || uncompressed_read != entry.uncompressed_size
+                    || macho_header.len() != MACHO_HEADER_BYTES
+                {
+                    return Err(CodexDesktopArtifactError::UnsupportedCore);
+                }
+                return scanner
+                    .finish()?
+                    .ok_or(CodexDesktopArtifactError::MissingCoreVersion);
+            }
+        }
+    }
+    Err(CodexDesktopArtifactError::UnsupportedCore)
+}
+
+fn validate_macho_header(header: &[u8]) -> Result<(), CodexDesktopArtifactError> {
+    if header.get(..4) != Some(MACHO_64_LE_MAGIC.as_slice())
+        || header
+            .get(4..8)
+            .and_then(|bytes| bytes.try_into().ok())
+            .map(u32::from_le_bytes)
+            != Some(MACHO_ARM64_CPU_TYPE)
+    {
+        return Err(CodexDesktopArtifactError::UnsupportedCore);
+    }
+    Ok(())
+}
+
+#[derive(Default)]
+pub(super) struct CoreVersionScanner {
+    carry: Vec<u8>,
+}
+
+impl CoreVersionScanner {
+    pub(super) fn push(
+        &mut self,
+        bytes: &[u8],
+    ) -> Result<Option<String>, CodexDesktopArtifactError> {
+        let mut window = Vec::with_capacity(self.carry.len() + bytes.len());
+        window.extend_from_slice(&self.carry);
+        window.extend_from_slice(bytes);
+        let mut search_start = 0_usize;
+        while let Some(relative_position) = find_bytes(&window[search_start..], CORE_VERSION_ANCHOR)
+        {
+            let position = search_start + relative_position;
+            if let Some(version) =
+                parse_codex_semver(&window[position + CORE_VERSION_ANCHOR.len()..], false)?
+            {
+                return Ok(Some(version));
+            }
+            search_start = position + CORE_VERSION_ANCHOR.len();
+        }
+        let carry_length = cmp::min(window.len(), CORE_VERSION_ANCHOR.len() + MAX_VERSION_BYTES);
+        self.carry.clear();
+        self.carry
+            .extend_from_slice(&window[window.len() - carry_length..]);
+        Ok(None)
+    }
+
+    fn finish(&self) -> Result<Option<String>, CodexDesktopArtifactError> {
+        let mut search_start = 0_usize;
+        while let Some(relative_position) =
+            find_bytes(&self.carry[search_start..], CORE_VERSION_ANCHOR)
+        {
+            let position = search_start + relative_position;
+            if let Some(version) =
+                parse_codex_semver(&self.carry[position + CORE_VERSION_ANCHOR.len()..], true)?
+            {
+                return Ok(Some(version));
+            }
+            search_start = position + CORE_VERSION_ANCHOR.len();
+        }
+        Ok(None)
+    }
+}
+
+fn parse_codex_semver(
+    bytes: &[u8],
+    allow_end: bool,
+) -> Result<Option<String>, CodexDesktopArtifactError> {
+    let mut cursor = 0_usize;
+    for component in 0..3 {
+        let start = cursor;
+        while bytes.get(cursor).is_some_and(u8::is_ascii_digit) {
+            cursor += 1;
+        }
+        if cursor == start || (cursor - start > 1 && bytes[start] == b'0') {
+            return Ok(None);
+        }
+        if component < 2 {
+            if bytes.get(cursor) != Some(&b'.') {
+                return Ok(None);
+            }
+            cursor += 1;
+        }
+    }
+    if bytes.get(cursor) == Some(&b'-') {
+        cursor += 1;
+        let channel = [b"alpha".as_slice(), b"beta".as_slice(), b"rc".as_slice()]
+            .into_iter()
+            .find(|channel| bytes.get(cursor..cursor + channel.len()) == Some(*channel));
+        let Some(channel) = channel else {
+            return Ok(None);
+        };
+        cursor += channel.len();
+        let mut numeric_parts = 0_usize;
+        while bytes.get(cursor) == Some(&b'.') {
+            cursor += 1;
+            let start = cursor;
+            while bytes.get(cursor).is_some_and(u8::is_ascii_digit) {
+                cursor += 1;
+            }
+            if cursor == start || (cursor - start > 1 && bytes[start] == b'0') {
+                return Ok(None);
+            }
+            numeric_parts += 1;
+        }
+        if numeric_parts == 0 {
+            return Ok(None);
+        }
+    }
+    if cursor == bytes.len() && !allow_end {
+        return Ok(None);
+    }
+    if bytes.get(cursor).is_some_and(u8::is_ascii_digit)
+        || bytes.get(cursor) == Some(&b'.')
+        || bytes
+            .get(cursor)
+            .is_some_and(|byte| matches!(*byte, b'+' | b'-'))
+        || cursor > MAX_VERSION_BYTES
+    {
+        return Ok(None);
+    }
+    let version = std::str::from_utf8(&bytes[..cursor])
+        .map_err(|_| CodexDesktopArtifactError::MissingCoreVersion)?;
+    semver::Version::parse(version).map_err(|_| CodexDesktopArtifactError::MissingCoreVersion)?;
+    Ok(Some(version.to_owned()))
+}
+
+fn find_bytes(haystack: &[u8], needle: &[u8]) -> Option<usize> {
+    haystack
+        .windows(needle.len())
+        .position(|window| window == needle)
+}
+
+fn le_u16(bytes: &[u8], offset: usize) -> Result<u16, CodexDesktopArtifactError> {
+    bytes
+        .get(offset..offset + 2)
+        .and_then(|value| value.try_into().ok())
+        .map(u16::from_le_bytes)
+        .ok_or(CodexDesktopArtifactError::InvalidArchive)
+}
+
+fn le_u32(bytes: &[u8], offset: usize) -> Result<u32, CodexDesktopArtifactError> {
+    bytes
+        .get(offset..offset + 4)
+        .and_then(|value| value.try_into().ok())
+        .map(u32::from_le_bytes)
+        .ok_or(CodexDesktopArtifactError::InvalidArchive)
+}

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

@@ -5,31 +5,42 @@ use std::time::Duration;
 
 use chrono::{DateTime, Utc};
 use futures::{StreamExt as _, future::BoxFuture};
+use gateway_core::engine::credential::OpaqueProviderData;
+use gateway_core::provider_ports::{
+    ProviderArtifactProfile, ProviderArtifactProfileCachePort, ProviderStoreError,
+};
+use gateway_core::routing::ProviderKind;
 use reqwest::Client;
 use reqwest::redirect::Policy;
 use roxmltree::{Document, Node};
+use serde::Deserialize;
+use serde_json::{Map, Value};
 use url::Url;
 
+use self::desktop_artifact::{CodexDesktopArtifactError, fetch_codex_core_version};
+
+mod desktop_artifact;
+
 /// Codex Desktop 官方 appcast 地址。
 pub const CODEX_DESKTOP_APPCAST_URL: &str =
     "https://persistent.oaistatic.com/codex-app-prod/appcast.xml";
-pub const CODEX_CLI_RELEASE_URL: &str = "https://registry.npmjs.org/@openai%2Fcodex/latest";
 /// 官方发布元数据检查周期。
 pub const APPCAST_POLL_INTERVAL: Duration = Duration::from_secs(24 * 60 * 60);
 
 const APPCAST_TIMEOUT: Duration = Duration::from_secs(30);
 const MAX_APPCAST_BYTES: usize = 1024 * 1024;
-const MAX_CLI_RELEASE_BYTES: usize = 64 * 1024;
+const ARTIFACT_PROFILE_CACHE_TTL: Duration = Duration::from_secs(30 * 24 * 60 * 60);
+const ARTIFACT_PROFILE_SCHEMA_VERSION: u64 = 1;
 
 /// Codex Desktop 上游请求身份。
 ///
-/// 启动配置提供经源码审计的 Core、运行环境和 Desktop 启动版本。npm CLI metadata
-/// 与官方 appcast 检查成功后，分别同步 Core 版本和 Desktop 版本及构建号。
+/// 启动配置提供经源码审计的 Core、运行环境和 Desktop 启动版本。运行时只会使用
+/// 同一个官方 Desktop ZIP 中核验出的 Core、Desktop 版本及构建号原子替换版本字段。
 #[derive(Debug, Clone, PartialEq, Eq)]
 pub struct CodexWireProfile {
     /// `originator` 请求头及 User-Agent 产品名。
     pub originator: String,
-    /// npm `@openai/codex` 启动版本；用于 `/codex/models?client_version=` 与 UA。
+    /// Desktop ZIP 内嵌 Core 版本；用于 `/codex/models?client_version=` 与 UA。
     pub codex_version: String,
     /// Desktop 应用版本，用于 app-server `clientInfo.version` 对应的 UA 后缀。
     pub desktop_version: String,
@@ -86,124 +97,148 @@ impl CodexWireProfileState {
             .clone()
     }
 
-    /// 使用官方 appcast 已验证的 Desktop 版本和构建号更新运行时画像。
-    pub fn update_desktop_release(&self, desktop_version: &str, desktop_build: &str) {
+    /// 原子发布同一 Desktop ZIP 中核验出的完整版本元组。
+    pub fn update_bundled_release(&self, release: &CodexBundledReleaseProfile) {
         let mut profile = self
             .profile
             .write()
             .unwrap_or_else(std::sync::PoisonError::into_inner);
-        if profile.desktop_version == desktop_version && profile.desktop_build == desktop_build {
-            return;
+        profile.codex_version.clone_from(&release.codex_version);
+        profile.desktop_version.clone_from(&release.desktop_version);
+        profile.desktop_build.clone_from(&release.desktop_build);
+        profile.verified_at = release.verified_at;
+    }
+}
+
+/// 同一个官方 Desktop ZIP 中核验出的原子版本元组。
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub struct CodexBundledReleaseProfile {
+    pub codex_version: String,
+    pub desktop_version: String,
+    pub desktop_build: String,
+    pub verified_at: DateTime<Utc>,
+}
+
+impl CodexBundledReleaseProfile {
+    fn artifact_sequence(&self) -> Result<u64, CodexDesktopReleaseError> {
+        parse_build_sequence(&self.desktop_build)
+    }
+
+    fn validate(&self) -> Result<(), CodexDesktopReleaseError> {
+        semver::Version::parse(&self.codex_version)
+            .map_err(|_| CodexDesktopReleaseError::InvalidCoreVersion)?;
+        if !numeric_dotted_version(&self.desktop_version) {
+            return Err(CodexDesktopReleaseError::InvalidVersion);
         }
-        profile.desktop_version = desktop_version.to_owned();
-        profile.desktop_build = desktop_build.to_owned();
-    }
-
-    /// 使用官方 npm metadata 发布的 Codex CLI 版本更新运行时请求画像。
-    pub fn update_cli_release(&self, codex_version: &str) {
-        let mut profile = self
-            .profile
-            .write()
-            .unwrap_or_else(std::sync::PoisonError::into_inner);
-        if profile.codex_version != codex_version {
-            profile.codex_version = codex_version.to_owned();
-        }
-    }
-}
-
-pub trait CodexCliReleaseTransport: Send + Sync {
-    fn fetch(&self) -> BoxFuture<'_, Result<String, CodexCliReleaseError>>;
-}
-
-#[derive(Clone)]
-pub struct OfficialCodexCliReleaseTransport {
-    client: Client,
-    endpoint: Url,
-}
-
-impl OfficialCodexCliReleaseTransport {
-    pub fn new() -> Result<Self, CodexCliReleaseError> {
-        let endpoint =
-            Url::parse(CODEX_CLI_RELEASE_URL).map_err(|_| CodexCliReleaseError::InvalidEndpoint)?;
-        let client = Client::builder()
-            .https_only(true)
-            .no_proxy()
-            .redirect(Policy::none())
-            .connect_timeout(Duration::from_secs(10))
-            .timeout(APPCAST_TIMEOUT)
-            .build()
-            .map_err(|_| CodexCliReleaseError::ClientInitialization)?;
-        Ok(Self { client, endpoint })
-    }
-}
-
-impl CodexCliReleaseTransport for OfficialCodexCliReleaseTransport {
-    fn fetch(&self) -> BoxFuture<'_, Result<String, CodexCliReleaseError>> {
-        Box::pin(async move {
-            let response = self.client.get(self.endpoint.clone()).send().await?;
-            if !response.status().is_success() {
-                return Err(CodexCliReleaseError::HttpStatus(response.status().as_u16()));
-            }
-            if response
-                .content_length()
-                .is_some_and(|size| size > MAX_CLI_RELEASE_BYTES as u64)
-            {
-                return Err(CodexCliReleaseError::ResponseTooLarge);
-            }
-            let bytes = response.bytes().await?;
-            if bytes.len() > MAX_CLI_RELEASE_BYTES {
-                return Err(CodexCliReleaseError::ResponseTooLarge);
-            }
-            let document: serde_json::Value = serde_json::from_slice(&bytes)
-                .map_err(|_| CodexCliReleaseError::InvalidDocument)?;
-            let version = document
-                .get("version")
-                .and_then(serde_json::Value::as_str)
-                .ok_or(CodexCliReleaseError::InvalidDocument)?;
-            semver::Version::parse(version).map_err(|_| CodexCliReleaseError::InvalidVersion)?;
-            Ok(version.to_owned())
-        })
+        self.artifact_sequence()?;
+        Ok(())
     }
 }
 
 #[derive(Clone)]
-pub struct CodexCliReleaseService {
-    transport: Arc<dyn CodexCliReleaseTransport>,
-    profile: CodexWireProfileState,
+pub struct CodexArtifactProfileCache {
+    provider_kind: ProviderKind,
+    store: Arc<dyn ProviderArtifactProfileCachePort>,
 }
 
-impl CodexCliReleaseService {
+impl CodexArtifactProfileCache {
     #[must_use]
     pub fn new(
-        profile: CodexWireProfileState,
-        transport: Arc<dyn CodexCliReleaseTransport>,
+        provider_kind: ProviderKind,
+        store: Arc<dyn ProviderArtifactProfileCachePort>,
     ) -> Self {
-        Self { transport, profile }
+        Self {
+            provider_kind,
+            store,
+        }
     }
 
-    pub async fn refresh(&self) -> Result<String, CodexCliReleaseError> {
-        let version = self.transport.fetch().await?;
-        self.profile.update_cli_release(&version);
-        Ok(version)
+    pub async fn load(
+        &self,
+    ) -> Result<Option<CodexBundledReleaseProfile>, CodexDesktopReleaseError> {
+        self.store
+            .read(&self.provider_kind)
+            .await
+            .map_err(CodexDesktopReleaseError::ArtifactCache)?
+            .map(decode_artifact_profile)
+            .transpose()
+    }
+
+    async fn replace_if_newer(
+        &self,
+        profile: &CodexBundledReleaseProfile,
+    ) -> Result<bool, CodexDesktopReleaseError> {
+        profile.validate()?;
+        self.store
+            .replace_if_newer(
+                encode_artifact_profile(self.provider_kind.clone(), profile)?,
+                ARTIFACT_PROFILE_CACHE_TTL,
+            )
+            .await
+            .map_err(CodexDesktopReleaseError::ArtifactCache)
     }
 }
 
-#[derive(Debug, thiserror::Error)]
-pub enum CodexCliReleaseError {
-    #[error("Codex CLI release client initialization failed")]
-    ClientInitialization,
-    #[error("Codex CLI release endpoint is invalid")]
-    InvalidEndpoint,
-    #[error("Codex CLI release request failed")]
-    Request(#[from] reqwest::Error),
-    #[error("Codex CLI release endpoint returned HTTP {0}")]
-    HttpStatus(u16),
-    #[error("Codex CLI release response exceeded the size limit")]
-    ResponseTooLarge,
-    #[error("Codex CLI release document is invalid")]
-    InvalidDocument,
-    #[error("Codex CLI release version is invalid")]
-    InvalidVersion,
+#[derive(Deserialize)]
+#[serde(deny_unknown_fields)]
+struct CachedArtifactProfile {
+    schema_version: u64,
+    codex_version: String,
+    desktop_version: String,
+    desktop_build: String,
+}
+
+fn encode_artifact_profile(
+    provider_kind: ProviderKind,
+    profile: &CodexBundledReleaseProfile,
+) -> Result<ProviderArtifactProfile, CodexDesktopReleaseError> {
+    let artifact_sequence = profile.artifact_sequence()?;
+    let mut fields = Map::new();
+    fields.insert(
+        "schema_version".to_owned(),
+        Value::from(ARTIFACT_PROFILE_SCHEMA_VERSION),
+    );
+    fields.insert(
+        "codex_version".to_owned(),
+        Value::String(profile.codex_version.clone()),
+    );
+    fields.insert(
+        "desktop_version".to_owned(),
+        Value::String(profile.desktop_version.clone()),
+    );
+    fields.insert(
+        "desktop_build".to_owned(),
+        Value::String(profile.desktop_build.clone()),
+    );
+    Ok(ProviderArtifactProfile::new(
+        provider_kind,
+        artifact_sequence,
+        profile.verified_at.into(),
+        OpaqueProviderData::new(fields),
+    ))
+}
+
+fn decode_artifact_profile(
+    profile: ProviderArtifactProfile,
+) -> Result<CodexBundledReleaseProfile, CodexDesktopReleaseError> {
+    let wire: CachedArtifactProfile = serde_json::from_value(Value::Object(
+        profile.profile().expose_to_provider().clone(),
+    ))
+    .map_err(|_| CodexDesktopReleaseError::InvalidArtifactCache)?;
+    if wire.schema_version != ARTIFACT_PROFILE_SCHEMA_VERSION {
+        return Err(CodexDesktopReleaseError::InvalidArtifactCache);
+    }
+    let decoded = CodexBundledReleaseProfile {
+        codex_version: wire.codex_version,
+        desktop_version: wire.desktop_version,
+        desktop_build: wire.desktop_build,
+        verified_at: DateTime::<Utc>::from(profile.verified_at()),
+    };
+    decoded.validate()?;
+    if decoded.artifact_sequence()? != profile.artifact_sequence() {
+        return Err(CodexDesktopReleaseError::InvalidArtifactCache);
+    }
+    Ok(decoded)
 }
 
 /// 官方 appcast 中按顺序出现的首个完整 Desktop 制品。
@@ -266,6 +301,11 @@ impl CodexDesktopReleaseStatus {
 /// Desktop appcast 获取边界。生产实现固定访问官方 HTTPS，测试实现只替换此边界。
 pub trait CodexDesktopReleaseTransport: Send + Sync {
     fn fetch(&self) -> BoxFuture<'_, Result<CodexDesktopRelease, CodexDesktopReleaseError>>;
+
+    fn fetch_bundled_core_version<'a>(
+        &'a self,
+        release: &'a CodexDesktopRelease,
+    ) -> BoxFuture<'a, Result<String, CodexDesktopReleaseError>>;
 }
 
 /// 固定访问官方 Desktop appcast 的生产 transport。
@@ -325,6 +365,24 @@ impl CodexDesktopReleaseTransport for OfficialCodexDesktopReleaseTransport {
             parse_desktop_release(xml)
         })
     }
+
+    fn fetch_bundled_core_version<'a>(
+        &'a self,
+        release: &'a CodexDesktopRelease,
+    ) -> BoxFuture<'a, Result<String, CodexDesktopReleaseError>> {
+        Box::pin(async move {
+            let artifact_url = release
+                .download_url
+                .as_deref()
+                .ok_or(CodexDesktopReleaseError::MissingArtifactIdentity)?;
+            let artifact_size = release
+                .download_size
+                .ok_or(CodexDesktopReleaseError::MissingArtifactIdentity)?;
+            fetch_codex_core_version(&self.client, artifact_url, artifact_size)
+                .await
+                .map_err(Into::into)
+        })
+    }
 }
 
 /// 拉取 appcast 并原子发布请求画像的 Provider 服务。
@@ -333,6 +391,8 @@ pub struct CodexDesktopReleaseService {
     transport: Arc<dyn CodexDesktopReleaseTransport>,
     status: CodexDesktopReleaseStatus,
     profile: CodexWireProfileState,
+    cache: CodexArtifactProfileCache,
+    verified: Arc<RwLock<Option<CodexBundledReleaseProfile>>>,
 }
 
 impl CodexDesktopReleaseService {
@@ -341,11 +401,18 @@ impl CodexDesktopReleaseService {
     pub fn new(
         profile: CodexWireProfileState,
         transport: Arc<dyn CodexDesktopReleaseTransport>,
+        cache: CodexArtifactProfileCache,
+        verified: Option<CodexBundledReleaseProfile>,
     ) -> Self {
+        if let Some(verified) = verified.as_ref() {
+            profile.update_bundled_release(verified);
+        }
         Self {
             transport,
             status: CodexDesktopReleaseStatus::default(),
             profile,
+            cache,
+            verified: Arc::new(RwLock::new(verified)),
         }
     }
 
@@ -357,11 +424,9 @@ impl CodexDesktopReleaseService {
     /// 执行一次有界检查；失败只更新观察状态，不修改上一份成功画像。
     pub async fn refresh(&self) -> Result<CodexDesktopRelease, CodexDesktopReleaseError> {
         let checked_at = Utc::now();
-        let result = self.transport.fetch().await;
+        let result = self.refresh_inner(checked_at).await;
         match result {
             Ok(release) => {
-                self.profile
-                    .update_desktop_release(&release.version, &release.build);
                 self.status.record_success(checked_at, release.clone());
                 Ok(release)
             }
@@ -371,6 +436,91 @@ impl CodexDesktopReleaseService {
             }
         }
     }
+
+    async fn refresh_inner(
+        &self,
+        checked_at: DateTime<Utc>,
+    ) -> Result<CodexDesktopRelease, CodexDesktopReleaseError> {
+        let release = self.transport.fetch().await?;
+        validate_release_artifact(&release)?;
+        let release_sequence = parse_build_sequence(&release.build)?;
+        let verified = self
+            .verified
+            .read()
+            .unwrap_or_else(std::sync::PoisonError::into_inner)
+            .clone();
+        if let Some(verified) = verified {
+            let verified_sequence = verified.artifact_sequence()?;
+            if release_sequence < verified_sequence {
+                return Err(CodexDesktopReleaseError::ReleaseRollback);
+            }
+            if release_sequence == verified_sequence {
+                if release.version != verified.desktop_version
+                    || release.build != verified.desktop_build
+                {
+                    return Err(CodexDesktopReleaseError::InconsistentRelease);
+                }
+                if !self.cache.replace_if_newer(&verified).await? {
+                    return Err(CodexDesktopReleaseError::ReleaseRollback);
+                }
+                self.profile.update_bundled_release(&verified);
+                return Ok(release);
+            }
+        }
+
+        let codex_version = self.transport.fetch_bundled_core_version(&release).await?;
+        let bundled = CodexBundledReleaseProfile {
+            codex_version,
+            desktop_version: release.version.clone(),
+            desktop_build: release.build.clone(),
+            verified_at: checked_at,
+        };
+        bundled.validate()?;
+        let published = if self.cache.replace_if_newer(&bundled).await? {
+            bundled
+        } else {
+            self.cache
+                .load()
+                .await?
+                .filter(|cached| {
+                    cached
+                        .artifact_sequence()
+                        .is_ok_and(|sequence| sequence >= release_sequence)
+                })
+                .ok_or(CodexDesktopReleaseError::ReleaseRollback)?
+        };
+        self.profile.update_bundled_release(&published);
+        *self
+            .verified
+            .write()
+            .unwrap_or_else(std::sync::PoisonError::into_inner) = Some(published);
+        Ok(release)
+    }
+}
+
+fn validate_release_artifact(
+    release: &CodexDesktopRelease,
+) -> Result<(), CodexDesktopReleaseError> {
+    if !release.signature_present {
+        return Err(CodexDesktopReleaseError::MissingArtifactSignature);
+    }
+    let artifact_url = release
+        .download_url
+        .as_deref()
+        .ok_or(CodexDesktopReleaseError::MissingArtifactIdentity)?;
+    let artifact_size = release
+        .download_size
+        .ok_or(CodexDesktopReleaseError::MissingArtifactIdentity)?;
+    desktop_artifact::validate_codex_artifact(artifact_url, artifact_size)?;
+    Ok(())
+}
+
+fn parse_build_sequence(value: &str) -> Result<u64, CodexDesktopReleaseError> {
+    value
+        .parse::<u64>()
+        .ok()
+        .filter(|sequence| *sequence > 0)
+        .ok_or(CodexDesktopReleaseError::InvalidBuild)
 }
 
 /// 解析 appcast 中按顺序出现的首个完整发布项。
@@ -489,4 +639,20 @@ pub enum CodexDesktopReleaseError {
     InvalidPublishedAt,
     #[error("Codex Desktop appcast download size is invalid")]
     InvalidDownloadSize,
+    #[error("Codex Desktop appcast artifact identity is incomplete")]
+    MissingArtifactIdentity,
+    #[error("Codex Desktop appcast artifact signature is missing")]
+    MissingArtifactSignature,
+    #[error("Codex Desktop bundled Core version is invalid")]
+    InvalidCoreVersion,
+    #[error(transparent)]
+    Artifact(#[from] CodexDesktopArtifactError),
+    #[error("Codex Desktop artifact profile cache operation failed")]
+    ArtifactCache(#[source] ProviderStoreError),
+    #[error("Codex Desktop artifact profile cache is invalid")]
+    InvalidArtifactCache,
+    #[error("Codex Desktop appcast attempted to roll back the verified artifact")]
+    ReleaseRollback,
+    #[error("Codex Desktop appcast release identity conflicts with the verified artifact")]
+    InconsistentRelease,
 }

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

@@ -2,19 +2,95 @@ use gateway_core::accounting::{
     CalculatedCostAmounts, CalculatedCostBreakdown, CalculatedCostRates, CurrencyCode, Decimal,
     Money,
 };
-use gateway_protocol::openai::events::retry_after_seconds_from_body;
+use gateway_protocol::openai::events::{TokenUsage, retry_after_seconds_from_body};
 use reqwest::StatusCode;
 use serde_json::Value;
 
 use super::{
     CodexBackendClient, CodexClientError, CodexClientResult, CodexRequestContext,
-    CodexUpstreamDiagnostics,
     client::{read_capped_response_body, retry_after_seconds, truncate_for_error},
-    endpoints::usage_endpoint_urls,
+    endpoints::usage_endpoint_url,
     response_meta,
 };
 
 const LONG_CONTEXT_THRESHOLD: u64 = 272_000;
+const WEB_SEARCH_CALL_TICKS: u128 = 100_000_000;
+const WEB_SEARCH_PREVIEW_NON_REASONING_CALL_TICKS: u128 = 250_000_000;
+
+/// OpenAI 公开 Token 价格计算所需的单次用量事实。
+#[derive(Debug, Clone, Copy, Default, PartialEq, Eq)]
+pub struct OpenAiBillingUsage {
+    input_tokens: u64,
+    output_tokens: u64,
+    cached_tokens: u64,
+    cache_write_tokens: u64,
+    image_input_tokens: u64,
+    image_output_tokens: u64,
+    web_search_calls: u64,
+    web_search_pricing: Option<WebSearchPricing>,
+}
+
+impl OpenAiBillingUsage {
+    /// 构造不含托管工具费用的 Token 用量。
+    #[must_use]
+    pub const fn new(
+        input_tokens: u64,
+        output_tokens: u64,
+        cached_tokens: u64,
+        cache_write_tokens: u64,
+    ) -> Self {
+        Self {
+            input_tokens,
+            output_tokens,
+            cached_tokens,
+            cache_write_tokens,
+            image_input_tokens: 0,
+            image_output_tokens: 0,
+            web_search_calls: 0,
+            web_search_pricing: None,
+        }
+    }
+
+    pub(crate) const fn with_web_search_calls(
+        mut self,
+        calls: u64,
+        pricing: Option<WebSearchPricing>,
+    ) -> Self {
+        self.web_search_calls = calls;
+        self.web_search_pricing = pricing;
+        self
+    }
+}
+
+impl From<TokenUsage> for OpenAiBillingUsage {
+    fn from(usage: TokenUsage) -> Self {
+        Self {
+            input_tokens: usage.input_tokens,
+            output_tokens: usage.output_tokens,
+            cached_tokens: usage.cached_tokens,
+            cache_write_tokens: usage.cache_write_tokens,
+            image_input_tokens: usage.image_input_tokens,
+            image_output_tokens: usage.image_output_tokens,
+            web_search_calls: 0,
+            web_search_pricing: None,
+        }
+    }
+}
+
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub(crate) enum WebSearchPricing {
+    Standard,
+    PreviewNonReasoning,
+}
+
+impl WebSearchPricing {
+    const fn price_per_call_ticks(self) -> u128 {
+        match self {
+            Self::Standard => WEB_SEARCH_CALL_TICKS,
+            Self::PreviewNonReasoning => WEB_SEARCH_PREVIEW_NON_REASONING_CALL_TICKS,
+        }
+    }
+}
 
 #[derive(Clone, Copy)]
 struct TokenRates {
@@ -43,9 +119,11 @@ impl TokenRates {
 #[derive(Clone, Copy)]
 struct ModelPricing {
     standard: TokenRates,
-    priority: TokenRates,
-    long: TokenRates,
-    long_priority: TokenRates,
+    flex: TokenRates,
+    fast: TokenRates,
+    long_standard: TokenRates,
+    long_flex: TokenRates,
+    long_fast: TokenRates,
     cache_write_percent: u32,
 }
 
@@ -53,29 +131,37 @@ impl ModelPricing {
     const fn new(input: u128, output: u128, cache_read: u128) -> Self {
         Self {
             standard: TokenRates::new(input, output, cache_read),
-            priority: TokenRates::ZERO,
-            long: TokenRates::ZERO,
-            long_priority: TokenRates::ZERO,
+            flex: TokenRates::ZERO,
+            fast: TokenRates::ZERO,
+            long_standard: TokenRates::ZERO,
+            long_flex: TokenRates::ZERO,
+            long_fast: TokenRates::ZERO,
             cache_write_percent: 0,
         }
     }
 
-    const fn with_priority(mut self, input: u128, output: u128, cache_read: u128) -> Self {
-        self.priority = TokenRates::new(input, output, cache_read);
+    const fn with_flex(mut self, input: u128, output: u128, cache_read: u128) -> Self {
+        self.flex = TokenRates::new(input, output, cache_read);
         self
     }
 
-    const fn with_long_context(
-        mut self,
-        input: u128,
-        output: u128,
-        cache_read: u128,
-        priority_input: u128,
-        priority_output: u128,
-        priority_cache_read: u128,
-    ) -> Self {
-        self.long = TokenRates::new(input, output, cache_read);
-        self.long_priority = TokenRates::new(priority_input, priority_output, priority_cache_read);
+    const fn with_fast(mut self, input: u128, output: u128, cache_read: u128) -> Self {
+        self.fast = TokenRates::new(input, output, cache_read);
+        self
+    }
+
+    const fn with_long(mut self, input: u128, output: u128, cache_read: u128) -> Self {
+        self.long_standard = TokenRates::new(input, output, cache_read);
+        self
+    }
+
+    const fn with_long_flex(mut self, input: u128, output: u128, cache_read: u128) -> Self {
+        self.long_flex = TokenRates::new(input, output, cache_read);
+        self
+    }
+
+    const fn with_long_fast(mut self, input: u128, output: u128, cache_read: u128) -> Self {
+        self.long_fast = TokenRates::new(input, output, cache_read);
         self
     }
 
@@ -83,6 +169,29 @@ impl ModelPricing {
         self.cache_write_percent = percent;
         self
     }
+
+    fn rates(self, tier: PricingTier, long_context: bool) -> Option<TokenRates> {
+        // Only models with a published long-context column switch at the threshold.
+        // A dash in that column is not a second, unavailable price tier for models
+        // whose entire supported context is covered by the short-context price.
+        let uses_long_rates = long_context && self.long_standard.is_configured();
+        let rates = match (tier, uses_long_rates) {
+            (PricingTier::Standard, false) => self.standard,
+            (PricingTier::Flex, false) => self.flex,
+            (PricingTier::Fast, false) => self.fast,
+            (PricingTier::Standard, true) => self.long_standard,
+            (PricingTier::Flex, true) => self.long_flex,
+            (PricingTier::Fast, true) => self.long_fast,
+        };
+        rates.is_configured().then_some(rates)
+    }
+}
+
+#[derive(Clone, Copy)]
+enum PricingTier {
+    Standard,
+    Flex,
+    Fast,
 }
 
 #[derive(Clone, Copy)]
@@ -91,86 +200,178 @@ struct PricingRule {
     pricing: ModelPricing,
 }
 
+// Source: https://developers.openai.com/api/docs/pricing (verified 2026-08-14).
 const PRICING_RULES: &[PricingRule] = &[
     PricingRule {
         model: "gpt-5.6-sol",
         pricing: ModelPricing::new(50_000, 300_000, 5_000)
             .with_cache_write(125)
-            .with_priority(100_000, 600_000, 10_000)
-            .with_long_context(100_000, 450_000, 10_000, 0, 0, 0),
+            .with_flex(25_000, 150_000, 2_500)
+            .with_fast(100_000, 600_000, 10_000)
+            .with_long(100_000, 450_000, 10_000)
+            .with_long_flex(50_000, 225_000, 5_000)
+            .with_long_fast(200_000, 900_000, 20_000),
     },
     PricingRule {
         model: "gpt-5.6-terra",
         pricing: ModelPricing::new(20_000, 120_000, 2_000)
             .with_cache_write(125)
-            .with_priority(40_000, 240_000, 4_000)
-            .with_long_context(40_000, 180_000, 4_000, 0, 0, 0),
+            .with_flex(10_000, 60_000, 1_000)
+            .with_fast(40_000, 240_000, 4_000)
+            .with_long(40_000, 180_000, 4_000)
+            .with_long_flex(20_000, 90_000, 2_000)
+            .with_long_fast(80_000, 360_000, 8_000),
     },
     PricingRule {
         model: "gpt-5.6-luna",
         pricing: ModelPricing::new(2_000, 12_000, 200)
             .with_cache_write(125)
-            .with_priority(4_000, 24_000, 400)
-            .with_long_context(4_000, 18_000, 400, 0, 0, 0),
+            .with_flex(1_000, 6_000, 100)
+            .with_fast(4_000, 24_000, 400)
+            .with_long(4_000, 18_000, 400)
+            .with_long_flex(2_000, 9_000, 200)
+            .with_long_fast(8_000, 36_000, 800),
     },
     PricingRule {
         model: "gpt-5.6",
         pricing: ModelPricing::new(50_000, 300_000, 5_000)
             .with_cache_write(125)
-            .with_priority(100_000, 600_000, 10_000)
-            .with_long_context(100_000, 450_000, 10_000, 0, 0, 0),
+            .with_flex(25_000, 150_000, 2_500)
+            .with_fast(100_000, 600_000, 10_000)
+            .with_long(100_000, 450_000, 10_000)
+            .with_long_flex(50_000, 225_000, 5_000)
+            .with_long_fast(200_000, 900_000, 20_000),
     },
     PricingRule {
         model: "gpt-5.5-pro",
         pricing: ModelPricing::new(300_000, 1_800_000, 0)
-            .with_priority(750_000, 4_500_000, 0)
-            .with_long_context(600_000, 2_700_000, 0, 1_500_000, 6_750_000, 0),
+            .with_flex(150_000, 900_000, 0)
+            .with_long(600_000, 2_700_000, 0),
     },
     PricingRule {
         model: "gpt-5.5",
         pricing: ModelPricing::new(50_000, 300_000, 5_000)
-            .with_priority(125_000, 750_000, 12_500)
-            .with_long_context(100_000, 450_000, 10_000, 250_000, 1_125_000, 25_000),
+            .with_flex(25_000, 150_000, 2_500)
+            .with_fast(125_000, 750_000, 12_500)
+            .with_long(100_000, 450_000, 10_000)
+            .with_long_flex(50_000, 225_000, 5_000),
     },
     PricingRule {
         model: "gpt-5.4-mini",
-        pricing: ModelPricing::new(7_500, 45_000, 750),
+        pricing: ModelPricing::new(7_500, 45_000, 750)
+            .with_flex(3_750, 22_500, 375)
+            .with_fast(15_000, 90_000, 1_500),
     },
     PricingRule {
         model: "gpt-5.4-nano",
-        pricing: ModelPricing::new(2_000, 12_500, 200),
+        pricing: ModelPricing::new(2_000, 12_500, 200).with_flex(1_000, 6_250, 100),
     },
     PricingRule {
         model: "gpt-5.4-pro",
         pricing: ModelPricing::new(300_000, 1_800_000, 0)
-            .with_priority(750_000, 4_500_000, 0)
-            .with_long_context(600_000, 2_700_000, 0, 1_500_000, 6_750_000, 0),
+            .with_flex(150_000, 900_000, 0)
+            .with_long(600_000, 2_700_000, 0)
+            .with_long_flex(300_000, 1_350_000, 0),
     },
     PricingRule {
         model: "gpt-5.4",
         pricing: ModelPricing::new(25_000, 150_000, 2_500)
-            .with_priority(50_000, 300_000, 5_000)
-            .with_long_context(50_000, 225_000, 5_000, 100_000, 450_000, 10_000),
-    },
-    PricingRule {
-        model: "gpt-5.3-codex-spark",
-        pricing: ModelPricing::new(12_500, 100_000, 1_250).with_priority(25_000, 200_000, 2_500),
+            .with_flex(12_500, 75_000, 1_300)
+            .with_fast(50_000, 300_000, 5_000)
+            .with_long(50_000, 225_000, 5_000)
+            .with_long_flex(25_000, 112_500, 2_500),
     },
     PricingRule {
         model: "gpt-5.3-codex",
-        pricing: ModelPricing::new(17_500, 140_000, 1_750).with_priority(35_000, 280_000, 3_500),
+        pricing: ModelPricing::new(17_500, 140_000, 1_750).with_fast(35_000, 280_000, 3_500),
+    },
+    PricingRule {
+        model: "gpt-5.2-pro",
+        pricing: ModelPricing::new(210_000, 1_680_000, 0),
     },
     PricingRule {
         model: "gpt-5.2",
-        pricing: ModelPricing::new(17_500, 140_000, 1_750).with_priority(35_000, 280_000, 3_500),
+        pricing: ModelPricing::new(17_500, 140_000, 1_750)
+            .with_flex(8_750, 70_000, 875)
+            .with_fast(35_000, 280_000, 3_500),
+    },
+    PricingRule {
+        model: "gpt-5.1",
+        pricing: ModelPricing::new(12_500, 100_000, 1_250)
+            .with_flex(6_250, 50_000, 625)
+            .with_fast(25_000, 200_000, 2_500),
+    },
+    PricingRule {
+        model: "gpt-5-mini",
+        pricing: ModelPricing::new(2_500, 20_000, 250)
+            .with_flex(1_250, 10_000, 125)
+            .with_fast(4_500, 36_000, 450),
+    },
+    PricingRule {
+        model: "gpt-5-nano",
+        pricing: ModelPricing::new(500, 4_000, 50).with_flex(250, 2_000, 25),
+    },
+    PricingRule {
+        model: "gpt-5-pro",
+        pricing: ModelPricing::new(150_000, 1_200_000, 0),
+    },
+    PricingRule {
+        model: "gpt-5",
+        pricing: ModelPricing::new(12_500, 100_000, 1_250)
+            .with_flex(6_250, 50_000, 625)
+            .with_fast(25_000, 200_000, 2_500),
+    },
+    PricingRule {
+        model: "gpt-4.1-mini",
+        pricing: ModelPricing::new(4_000, 16_000, 1_000).with_fast(7_000, 28_000, 1_750),
+    },
+    PricingRule {
+        model: "gpt-4.1-nano",
+        pricing: ModelPricing::new(1_000, 4_000, 250).with_fast(2_000, 8_000, 500),
+    },
+    PricingRule {
+        model: "gpt-4.1",
+        pricing: ModelPricing::new(20_000, 80_000, 5_000).with_fast(35_000, 140_000, 8_750),
+    },
+    PricingRule {
+        model: "gpt-4o-2024-05-13",
+        pricing: ModelPricing::new(50_000, 150_000, 0).with_fast(87_500, 262_500, 0),
     },
     PricingRule {
         model: "gpt-4o-mini",
-        pricing: ModelPricing::new(1_500, 6_000, 750),
+        pricing: ModelPricing::new(1_500, 6_000, 750).with_fast(2_500, 10_000, 1_250),
     },
     PricingRule {
         model: "gpt-4o",
-        pricing: ModelPricing::new(25_000, 100_000, 12_500),
+        pricing: ModelPricing::new(25_000, 100_000, 12_500).with_fast(42_500, 170_000, 21_250),
+    },
+    PricingRule {
+        model: "o1-pro",
+        pricing: ModelPricing::new(1_500_000, 6_000_000, 0),
+    },
+    PricingRule {
+        model: "o1",
+        pricing: ModelPricing::new(150_000, 600_000, 75_000),
+    },
+    PricingRule {
+        model: "o3-pro",
+        pricing: ModelPricing::new(200_000, 800_000, 0),
+    },
+    PricingRule {
+        model: "o3-mini",
+        pricing: ModelPricing::new(11_000, 44_000, 5_500),
+    },
+    PricingRule {
+        model: "o3",
+        pricing: ModelPricing::new(20_000, 80_000, 5_000)
+            .with_flex(10_000, 40_000, 2_500)
+            .with_fast(35_000, 140_000, 8_750),
+    },
+    PricingRule {
+        model: "o4-mini",
+        pricing: ModelPricing::new(11_000, 44_000, 2_750)
+            .with_flex(5_500, 22_000, 1_380)
+            .with_fast(20_000, 80_000, 5_000),
     },
     PricingRule {
         model: "gpt-4-turbo",
@@ -180,93 +381,86 @@ const PRICING_RULES: &[PricingRule] = &[
         model: "gpt-4",
         pricing: ModelPricing::new(300_000, 600_000, 0),
     },
+    PricingRule {
+        model: "gpt-3.5-turbo-instruct",
+        pricing: ModelPricing::new(15_000, 20_000, 0),
+    },
+    PricingRule {
+        model: "gpt-3.5-turbo-1106",
+        pricing: ModelPricing::new(10_000, 20_000, 0),
+    },
     PricingRule {
         model: "gpt-3.5-turbo",
         pricing: ModelPricing::new(5_000, 15_000, 0),
     },
 ];
 
+const UNPRICED_MODELS: &[&str] = &["gpt-5.3-codex-spark"];
+
+#[derive(Clone, Copy)]
+struct TokenAmounts {
+    input_ticks: u128,
+    output_ticks: u128,
+    cache_read_ticks: u128,
+    cache_write_ticks: u128,
+    total_ticks: u128,
+}
+
 /// 按 OpenAI Provider 当前受控价格规则计算费用明细。
 #[must_use]
 pub fn openai_billing_breakdown(
     model: &str,
-    input_tokens: u64,
-    output_tokens: u64,
-    cached_tokens: u64,
-    cache_write_tokens: u64,
+    usage: OpenAiBillingUsage,
     service_tier: Option<&str>,
 ) -> Option<CalculatedCostBreakdown> {
-    if cached_tokens > input_tokens {
+    if usage.cached_tokens > usage.input_tokens
+        || usage.image_input_tokens > 0
+        || usage.image_output_tokens > 0
+    {
         return None;
     }
+    let web_search_ticks = web_search_amount_ticks(usage)?;
     let pricing = model_pricing(model)?;
-    let long_context = input_tokens > LONG_CONTEXT_THRESHOLD;
+    let long_context = usage.input_tokens > LONG_CONTEXT_THRESHOLD;
     let normalized_tier = normalize_service_tier(service_tier);
-    let priority = matches!(normalized_tier.as_deref(), Some("priority" | "fast"));
-    let mut rates = if long_context && pricing.long.is_configured() {
-        pricing.long
-    } else {
-        pricing.standard
-    };
-    let multiplier_percent = if priority {
-        let priority_rates = if long_context && pricing.long_priority.is_configured() {
-            pricing.long_priority
-        } else {
-            pricing.priority
-        };
-        if priority_rates.is_configured() {
-            rates = priority_rates;
-            100
-        } else {
-            200
-        }
-    } else if normalized_tier.as_deref() == Some("flex") {
-        50
-    } else {
-        100
-    };
-
-    let billed_cache_read = if rates.cache_read_ticks > 0 {
-        cached_tokens.min(input_tokens)
-    } else {
-        0
-    };
-    let cache_write_rate = if pricing.cache_write_percent > 0 {
-        apply_percent(rates.input_ticks, pricing.cache_write_percent)?
-    } else {
-        0
-    };
-    let billed_cache_write = if cache_write_rate > 0 {
-        cache_write_tokens.min(input_tokens.saturating_sub(billed_cache_read))
-    } else {
-        0
-    };
-    let uncached_input = input_tokens
-        .saturating_sub(billed_cache_read)
-        .saturating_sub(billed_cache_write);
-    let input_amount = u128::from(uncached_input).checked_mul(rates.input_ticks)?;
-    let output_amount = u128::from(output_tokens).checked_mul(rates.output_ticks)?;
-    let cache_read_amount = u128::from(billed_cache_read).checked_mul(rates.cache_read_ticks)?;
-    let cache_write_amount = u128::from(billed_cache_write).checked_mul(cache_write_rate)?;
-    let standard_amount = input_amount
-        .checked_add(output_amount)?
-        .checked_add(cache_read_amount)?
-        .checked_add(cache_write_amount)?;
-    let total_amount = apply_percent(standard_amount, multiplier_percent)?;
+    let tier = pricing_tier(normalized_tier.as_deref())?;
+    let standard_rates = pricing.rates(PricingTier::Standard, long_context)?;
+    let selected_rates = pricing.rates(tier, long_context)?;
+    let mut standard = token_amounts(
+        standard_rates,
+        pricing.cache_write_percent,
+        usage.input_tokens,
+        usage.output_tokens,
+        usage.cached_tokens,
+        usage.cache_write_tokens,
+    )?;
+    let mut selected = token_amounts(
+        selected_rates,
+        pricing.cache_write_percent,
+        usage.input_tokens,
+        usage.output_tokens,
+        usage.cached_tokens,
+        usage.cache_write_tokens,
+    )?;
+    standard.total_ticks = standard.total_ticks.checked_add(web_search_ticks)?;
+    selected.total_ticks = selected.total_ticks.checked_add(web_search_ticks)?;
+    let multiplier_percent =
+        effective_multiplier_percent(selected.total_ticks, standard.total_ticks)?;
+    let cache_write_rate = cache_write_rate(selected_rates, pricing.cache_write_percent)?;
 
     Some(CalculatedCostBreakdown::new(
         CalculatedCostAmounts::new(
-            usd_money(input_amount)?,
-            usd_money(output_amount)?,
-            usd_money(cache_read_amount)?,
-            usd_money(cache_write_amount)?,
-            usd_money(standard_amount)?,
-            usd_money(total_amount)?,
+            usd_money(selected.input_ticks)?,
+            usd_money(selected.output_ticks)?,
+            usd_money(selected.cache_read_ticks)?,
+            usd_money(selected.cache_write_ticks)?,
+            usd_money(standard.total_ticks)?,
+            usd_money(selected.total_ticks)?,
         ),
         CalculatedCostRates::new(
-            usd_price_per_million(rates.input_ticks)?,
-            usd_price_per_million(rates.output_ticks)?,
-            usd_price_per_million(rates.cache_read_ticks)?,
+            usd_price_per_million(selected_rates.input_ticks)?,
+            usd_price_per_million(selected_rates.output_ticks)?,
+            usd_price_per_million(selected_rates.cache_read_ticks)?,
             usd_price_per_million(cache_write_rate)?,
         ),
         Some(normalized_tier.unwrap_or_else(|| "default".to_owned())),
@@ -276,6 +470,12 @@ pub fn openai_billing_breakdown(
 
 fn model_pricing(model: &str) -> Option<ModelPricing> {
     let normalized = normalize_model_name(model);
+    if UNPRICED_MODELS
+        .iter()
+        .any(|rule| model_matches_rule(&normalized, rule))
+    {
+        return None;
+    }
     PRICING_RULES
         .iter()
         .filter(|rule| model_matches_rule(&normalized, rule.model))
@@ -283,6 +483,120 @@ fn model_pricing(model: &str) -> Option<ModelPricing> {
         .map(|rule| rule.pricing)
 }
 
+pub(crate) fn web_search_pricing(model: &str, tools: Option<&[Value]>) -> Option<WebSearchPricing> {
+    let mut standard = false;
+    let mut preview = false;
+    for tool_type in tools
+        .into_iter()
+        .flatten()
+        .filter_map(|tool| tool.get("type").and_then(Value::as_str))
+    {
+        if tool_type == "web_search_preview" || tool_type.starts_with("web_search_preview_") {
+            preview = true;
+        } else if tool_type == "web_search" || tool_type.starts_with("web_search_") {
+            standard = true;
+        }
+    }
+    match (standard, preview) {
+        // These two models bill non-preview search content as a fixed 8K input block.
+        // The response does not expose enough detail to replace that block without
+        // double-counting ordinary input, so leave the request unpriced.
+        (true, false) if fixed_block_web_search_model(model) => None,
+        (true, false) => Some(WebSearchPricing::Standard),
+        (false, true) if reasoning_model(model) => Some(WebSearchPricing::Standard),
+        (false, true) => Some(WebSearchPricing::PreviewNonReasoning),
+        (false, false) | (true, true) => None,
+    }
+}
+
+fn fixed_block_web_search_model(model: &str) -> bool {
+    let normalized = normalize_model_name(model);
+    ["gpt-4o-mini", "gpt-4.1-mini"]
+        .iter()
+        .any(|rule| model_matches_rule(&normalized, rule))
+}
+
+fn reasoning_model(model: &str) -> bool {
+    let normalized = normalize_model_name(model);
+    normalized.starts_with("gpt-5")
+        || normalized.starts_with("o1")
+        || normalized.starts_with("o3")
+        || normalized.starts_with("o4")
+}
+
+fn pricing_tier(service_tier: Option<&str>) -> Option<PricingTier> {
+    match service_tier {
+        None | Some("auto" | "default" | "standard") => Some(PricingTier::Standard),
+        Some("flex") => Some(PricingTier::Flex),
+        Some("fast" | "priority") => Some(PricingTier::Fast),
+        Some(_) => None,
+    }
+}
+
+fn web_search_amount_ticks(usage: OpenAiBillingUsage) -> Option<u128> {
+    if usage.web_search_calls == 0 {
+        return Some(0);
+    }
+    u128::from(usage.web_search_calls).checked_mul(usage.web_search_pricing?.price_per_call_ticks())
+}
+
+fn token_amounts(
+    rates: TokenRates,
+    cache_write_percent: u32,
+    input_tokens: u64,
+    output_tokens: u64,
+    cached_tokens: u64,
+    cache_write_tokens: u64,
+) -> Option<TokenAmounts> {
+    let billed_cache_read = if rates.cache_read_ticks > 0 {
+        cached_tokens.min(input_tokens)
+    } else {
+        0
+    };
+    let cache_write_rate = cache_write_rate(rates, cache_write_percent)?;
+    let billed_cache_write = if cache_write_rate > 0 {
+        cache_write_tokens.min(input_tokens.saturating_sub(billed_cache_read))
+    } else {
+        0
+    };
+    let uncached_input = input_tokens
+        .saturating_sub(billed_cache_read)
+        .saturating_sub(billed_cache_write);
+    let input_ticks = u128::from(uncached_input).checked_mul(rates.input_ticks)?;
+    let output_ticks = u128::from(output_tokens).checked_mul(rates.output_ticks)?;
+    let cache_read_ticks = u128::from(billed_cache_read).checked_mul(rates.cache_read_ticks)?;
+    let cache_write_ticks = u128::from(billed_cache_write).checked_mul(cache_write_rate)?;
+    let total_ticks = input_ticks
+        .checked_add(output_ticks)?
+        .checked_add(cache_read_ticks)?
+        .checked_add(cache_write_ticks)?;
+    Some(TokenAmounts {
+        input_ticks,
+        output_ticks,
+        cache_read_ticks,
+        cache_write_ticks,
+        total_ticks,
+    })
+}
+
+fn cache_write_rate(rates: TokenRates, percent: u32) -> Option<u128> {
+    if percent == 0 {
+        return Some(0);
+    }
+    apply_percent(rates.input_ticks, percent)
+}
+
+fn effective_multiplier_percent(total: u128, standard: u128) -> Option<u32> {
+    if standard == 0 {
+        return Some(100);
+    }
+    let rounded = total
+        .checked_mul(100)?
+        .checked_add(standard / 2)?
+        .checked_div(standard)?;
+    u32::try_from(rounded).ok()
+}
+
 fn normalize_model_name(model: &str) -> String {
     model
         .trim()
@@ -336,78 +650,68 @@ pub const MAX_CODEX_USAGE_BODY_BYTES: usize = 1024 * 1024;
 impl CodexBackendClient {
     /// 获取 Codex usage JSON。
     pub async fn fetch_usage(&self, context: CodexRequestContext<'_>) -> CodexClientResult<Value> {
-        let mut last_invalid_body = None;
+        let headers = self.usage_request_headers(context)?;
+        let response = self
+            .client
+            .get(usage_endpoint_url(&self.base_url))
+            .headers(headers)
+            .send()
+            .await?;
+        let status = response.status();
+        let diagnostics = response_meta::diagnostics(Some(status.as_u16()), response.headers());
+        let retry_after_seconds = retry_after_seconds(response.headers(), None);
+        let body = read_capped_response_body(response, MAX_CODEX_USAGE_BODY_BYTES).await?;
+        if body.limit_exceeded() {
+            return Err(CodexClientError::Upstream {
+                status: if status.is_success() {
+                    StatusCode::BAD_GATEWAY
+                } else {
+                    status
+                },
+                retry_after_seconds,
+                body: "upstream usage response exceeded the body limit".to_owned(),
+                client_response: None,
+                diagnostics: Box::new(diagnostics),
+                set_cookie_headers: Vec::new(),
+                rate_limit_headers: Vec::new(),
+                transport: super::client::CodexBackendTransport::HttpSse,
+                transport_metrics: Box::default(),
+                send_phase: super::diagnostics::CodexUpstreamSendPhase::AfterPayload,
+            });
+        }
+        let body = body.into_string();
 
-        for endpoint in usage_endpoint_urls(&self.base_url) {
-            let headers = self.usage_request_headers(context)?;
-            let response = self.client.get(endpoint).headers(headers).send().await?;
-            let status = response.status();
-            let diagnostics = response_meta::diagnostics(Some(status.as_u16()), response.headers());
-            let retry_after_seconds = retry_after_seconds(response.headers(), None);
-            let body = read_capped_response_body(response, MAX_CODEX_USAGE_BODY_BYTES).await?;
-            if body.limit_exceeded() {
-                return Err(CodexClientError::Upstream {
-                    status: if status.is_success() {
-                        StatusCode::BAD_GATEWAY
-                    } else {
-                        status
-                    },
-                    retry_after_seconds,
-                    body: "upstream usage response exceeded the body limit".to_owned(),
-                    client_response: None,
-                    diagnostics: Box::new(diagnostics),
-                    set_cookie_headers: Vec::new(),
-                    rate_limit_headers: Vec::new(),
-                    transport: super::client::CodexBackendTransport::HttpSse,
-                    transport_metrics: Box::default(),
-                    send_phase: super::diagnostics::CodexUpstreamSendPhase::AfterPayload,
-                });
-            }
-            let body = body.into_string();
-
-            if status == StatusCode::NOT_FOUND {
-                last_invalid_body = Some(body);
-                continue;
-            }
-            if !status.is_success() {
-                return Err(CodexClientError::Upstream {
-                    status,
-                    retry_after_seconds: retry_after_seconds
-                        .or_else(|| retry_after_seconds_from_body(&body)),
-                    body,
-                    client_response: None,
-                    diagnostics: Box::new(diagnostics),
-                    set_cookie_headers: Vec::new(),
-                    rate_limit_headers: Vec::new(),
-                    transport: super::client::CodexBackendTransport::HttpSse,
-                    transport_metrics: Box::default(),
-                    send_phase: super::diagnostics::CodexUpstreamSendPhase::AfterPayload,
-                });
-            }
-
-            match serde_json::from_str::<Value>(&body) {
-                Ok(parsed) if is_usage_response(&parsed) => return Ok(parsed),
-                _ => last_invalid_body = Some(body),
-            }
+        if !status.is_success() {
+            return Err(CodexClientError::Upstream {
+                status,
+                retry_after_seconds: retry_after_seconds
+                    .or_else(|| retry_after_seconds_from_body(&body)),
+                body,
+                client_response: None,
+                diagnostics: Box::new(diagnostics),
+                set_cookie_headers: Vec::new(),
+                rate_limit_headers: Vec::new(),
+                transport: super::client::CodexBackendTransport::HttpSse,
+                transport_metrics: Box::default(),
+                send_phase: super::diagnostics::CodexUpstreamSendPhase::AfterPayload,
+            });
         }
 
-        Err(CodexClientError::Upstream {
-            status: StatusCode::BAD_GATEWAY,
-            retry_after_seconds: None,
-            body: last_invalid_body.map_or_else(
-                || "usage endpoint is unavailable".to_string(),
-                |body| format!("invalid usage response: {}", truncate_for_error(&body)),
-            ),
-            client_response: None,
-            diagnostics: Box::new(CodexUpstreamDiagnostics::with_status(
-                StatusCode::BAD_GATEWAY.as_u16(),
-            )),
-            set_cookie_headers: Vec::new(),
-            rate_limit_headers: Vec::new(),
-            transport: super::client::CodexBackendTransport::HttpSse,
-            transport_metrics: Box::default(),
-            send_phase: super::diagnostics::CodexUpstreamSendPhase::AfterPayload,
-        })
+        match serde_json::from_str::<Value>(&body) {
+            Ok(parsed) if is_usage_response(&parsed) => Ok(parsed),
+            _ => Err(CodexClientError::Upstream {
+                status: StatusCode::BAD_GATEWAY,
+                retry_after_seconds: None,
+                body: format!("invalid usage response: {}", truncate_for_error(&body)),
+                client_response: None,
+                diagnostics: Box::new(diagnostics),
+                set_cookie_headers: Vec::new(),
+                rate_limit_headers: Vec::new(),
+                transport: super::client::CodexBackendTransport::HttpSse,
+                transport_metrics: Box::default(),
+                send_phase: super::diagnostics::CodexUpstreamSendPhase::AfterPayload,
+            }),
+        }
     }
 }
 

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

@@ -112,6 +112,7 @@ fn is_sensitive_opening_header(name: &str) -> bool {
             | "x-codex-turn-metadata"
             | "x-codex-turn-state"
             | "x-codex-parent-thread-id"
+            | "x-oai-attestation"
     )
 }
 

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

@@ -25,15 +25,15 @@ use gateway_core::operation::{GenerateRequest, Operation, ProtocolPayload};
 use gateway_core::provider_ports::{
     NewOAuthPendingFlow, OAuthPendingClaimOutcome, OAuthPendingConsumeOutcome,
     OAuthPendingFlowPort, OAuthPendingPutOutcome, OAuthPendingReleaseOutcome,
-    ProviderCatalogCacheKey, ProviderCatalogCachePort, ProviderCooldown, ProviderCooldownPort,
-    ProviderCooldownScope, ProviderCredentialState, ProviderCredentialStatePort,
-    ProviderRefreshPolicy, ProviderRuntimePolicyPort, ProviderScopedCooldown, ProviderStoreError,
-    ProviderStorePorts,
+    ProviderArtifactProfile, ProviderArtifactProfileCachePort, ProviderCatalogCacheKey,
+    ProviderCatalogCachePort, ProviderCooldown, ProviderCooldownPort, ProviderCooldownScope,
+    ProviderCredentialState, ProviderCredentialStatePort, ProviderRefreshPolicy,
+    ProviderRuntimePolicyPort, ProviderScopedCooldown, ProviderStoreError, ProviderStorePorts,
 };
 use gateway_core::routing::{ProviderKind, UpstreamModelId};
 use gateway_core::task::{WorkerContribution, WorkerKind, WorkerRunnable};
 use provider_openai::config::{CodexWireProfileConfig, OpenAiConfig};
-use provider_openai::credential::ImportCodexOAuthCredential;
+use provider_openai::credential::{CodexCredentialCodec, ImportCodexOAuthCredential};
 use provider_openai::transport::profile::APPCAST_POLL_INTERVAL;
 use secrecy::SecretString;
 use serde_json::{Map, Value, json};
@@ -54,7 +54,7 @@ async fn openai_bundle_exposes_one_core_provider_and_drains_worker_contributions
     assert_eq!(bundle.core_provider().name(), "openai");
     assert_eq!(bundle.admin_provider().provider_kind().as_str(), "openai");
     let contributions = bundle.take_worker_contributions();
-    assert_eq!(contributions.len(), 5);
+    assert_eq!(contributions.len(), 4);
     assert!(
         contributions
             .iter()
@@ -81,22 +81,6 @@ async fn openai_bundle_exposes_one_core_provider_and_drains_worker_contributions
         panic!("Desktop release worker must be scheduled");
     };
     assert_eq!(schedule.interval(), APPCAST_POLL_INTERVAL);
-    let cli_release_worker = contributions
-        .iter()
-        .find_map(|contribution| match contribution {
-            WorkerContribution::Registration(registration)
-                if registration.id.owner() == "openai-cli-release" =>
-            {
-                Some(registration)
-            }
-            WorkerContribution::Registration(_) | WorkerContribution::Disabled { .. } => None,
-        })
-        .expect("CLI release worker");
-    assert_eq!(cli_release_worker.id.kind(), WorkerKind::QuotaCatalogHealth);
-    let WorkerRunnable::Scheduled { schedule, .. } = &cli_release_worker.runnable else {
-        panic!("CLI release worker must be scheduled");
-    };
-    assert_eq!(schedule.interval(), APPCAST_POLL_INTERVAL);
     assert!(contributions.iter().any(|contribution| {
         matches!(
             contribution,
@@ -164,15 +148,15 @@ async fn openai_admin_provider_exposes_live_wire_profile_and_validated_billing()
             cache_write_tokens: Some(0),
             total: CurrencyCost {
                 currency: "USD".to_owned(),
-                amount: "5".parse().expect("fast amount"),
+                amount: "4.25".parse().expect("fast amount"),
             },
         })
         .expect("fast billing")
         .expect("known fast pricing");
     assert_eq!(fast_billing.service_tier.as_deref(), Some("priority"));
-    assert_eq!(fast_billing.multiplier_percent, 200);
+    assert_eq!(fast_billing.multiplier_percent, 170);
     assert_eq!(fast_billing.standard_amount.amount.as_str(), "2.5");
-    assert_eq!(fast_billing.total_amount.amount.as_str(), "5");
+    assert_eq!(fast_billing.total_amount.amount.as_str(), "4.25");
 }
 
 #[tokio::test]
@@ -249,6 +233,14 @@ async fn openai_admin_provider_persists_the_full_pending_envelope_and_binds_owne
                 .get("reauthorization_credential_revision")
                 .is_none()
         );
+        assert!(
+            payload
+                .expose_to_provider()
+                .get("installation_id")
+                .and_then(Value::as_str)
+                .and_then(|value| uuid::Uuid::parse_str(value).ok())
+                .is_some_and(|value| value.get_version_num() == 4)
+        );
         assert_eq!(
             mutation.get("schema_version").and_then(Value::as_u64),
             Some(3)
@@ -284,7 +276,7 @@ async fn openai_admin_provider_persists_the_full_pending_envelope_and_binds_owne
 }
 
 #[tokio::test]
-async fn openai_reauthorization_pending_payload_contains_only_stable_account_identity() {
+async fn openai_reauthorization_pending_payload_reuses_the_account_installation_id() {
     let accounts = Arc::new(MemoryAccountStore::default());
     accounts
         .seed_oauth_credential(ImportCodexOAuthCredential {
@@ -296,6 +288,14 @@ async fn openai_reauthorization_pending_payload_contains_only_stable_account_ide
             enabled: true,
         })
         .await;
+    let account_id = ProviderAccountId::new("acct_pending_reauth").expect("account id");
+    let existing = accounts
+        .load_current_credential(&account_id)
+        .await
+        .expect("seeded credential");
+    let expected_installation_id = CodexCredentialCodec::decode(&existing.credential)
+        .expect("decode seeded credential")
+        .installation_id;
     let pending = Arc::new(TestOAuthPending::default());
     let config = valid_config();
     let bundle = provider_openai::initialize(
@@ -313,9 +313,7 @@ async fn openai_reauthorization_pending_payload_contains_only_stable_account_ide
         .admin_provider()
         .start_authorization(PendingAuthorizationMutation::new(
             ProviderKind::new("openai").expect("provider"),
-            AuthorizationMutationTarget::Reauthorize {
-                account_id: ProviderAccountId::new("acct_pending_reauth").expect("account id"),
-            },
+            AuthorizationMutationTarget::Reauthorize { account_id },
             AuthorizationOwnerBinding::from_context(&context),
         ))
         .await
@@ -342,6 +340,10 @@ async fn openai_reauthorization_pending_payload_contains_only_stable_account_ide
             .and_then(Value::as_str),
         Some("acct_pending_reauth")
     );
+    assert_eq!(
+        document.get("installation_id").and_then(Value::as_str),
+        Some(expected_installation_id.as_str())
+    );
     assert!(
         document
             .get("reauthorization_credential_revision")
@@ -785,6 +787,7 @@ fn provider_ports_with_catalog(
         Arc::new(MemorySessionAffinity::default()),
         Arc::new(MemorySessionExclusions::default()),
         catalog_cache,
+        Arc::new(TestArtifactProfiles),
         Arc::new(TestCredentialState),
         Arc::new(TestCooldown),
         Arc::new(TestRuntimePolicy),
@@ -852,6 +855,25 @@ fn valid_config() -> TestOpenAiConfig {
     }
 }
 
+struct TestArtifactProfiles;
+
+impl ProviderArtifactProfileCachePort for TestArtifactProfiles {
+    fn replace_if_newer(
+        &self,
+        _profile: ProviderArtifactProfile,
+        _ttl: Duration,
+    ) -> BoxFuture<'_, Result<bool, ProviderStoreError>> {
+        Box::pin(async { Ok(true) })
+    }
+
+    fn read<'a>(
+        &'a self,
+        _provider_kind: &'a ProviderKind,
+    ) -> BoxFuture<'a, Result<Option<ProviderArtifactProfile>, ProviderStoreError>> {
+        Box::pin(async { Ok(None) })
+    }
+}
+
 #[derive(Default)]
 struct TestCatalogCache {
     values: Mutex<BTreeMap<String, OpaqueProviderData>>,

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

@@ -116,7 +116,7 @@ fn openai_config_defaults_to_the_provider_owned_operating_values() {
     );
     assert_eq!(
         config.wire_profile_state().snapshot().user_agent(),
-        "Codex Desktop/0.146.0 (Mac OS 15.7.1; arm64) unknown (Codex Desktop; 26.727.51351)"
+        "Codex Desktop/0.147.0-alpha.6.6 (Mac OS 15.7.1; arm64) unknown (Codex Desktop; 26.803.81509)"
     );
 }
 

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

@@ -13,6 +13,7 @@ use gateway_core::engine::credential::{
     ProviderAccountStore, QuotaState,
 };
 use gateway_core::routing::ProviderKind;
+use provider_openai::OpenAiConfig;
 use provider_openai::credential::{
     CODEX_AUTHENTICATION_KIND_AGENT_IDENTITY, CodexAgentIdentityAuthMode,
     CodexAgentIdentityCredentialData, CodexAgentIdentityError, CodexAgentIdentitySecret,
@@ -20,7 +21,9 @@ use provider_openai::credential::{
     CodexRuntimeAuthentication, is_agent_identity_task_invalid_response,
 };
 use provider_openai::transport::CodexWebSocketPool;
+use provider_openai::transport::headers::build_codex_profile_headers;
 use reqwest::StatusCode;
+use reqwest::header::USER_AGENT;
 use secrecy::ExposeSecret as _;
 use serde_json::Value;
 
@@ -144,6 +147,20 @@ fn agent_identity_task_error_classifier_requires_unauthorized_task_error() {
     ));
 }
 
+#[test]
+fn agent_identity_task_registration_should_use_the_shared_desktop_profile_headers() {
+    let profile = OpenAiConfig::default().wire_profile_state().snapshot();
+    let headers = build_codex_profile_headers(&profile).expect("Desktop profile headers");
+
+    assert_eq!(headers.len(), 3);
+    assert_eq!(headers["originator"], "Codex Desktop");
+    assert_eq!(headers["x-openai-internal-codex-residency"], "us");
+    assert_eq!(
+        headers[USER_AGENT],
+        "Codex Desktop/0.147.0-alpha.6.6 (Mac OS 15.7.1; arm64) unknown (Codex Desktop; 26.803.81509)"
+    );
+}
+
 struct SequenceTaskRegistrar {
     task_ids: Mutex<VecDeque<String>>,
     calls: AtomicUsize,

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

@@ -1,3 +1,4 @@
+use std::collections::BTreeMap;
 use std::sync::{Arc, Mutex};
 use std::time::{Duration, SystemTime};
 
@@ -7,22 +8,26 @@ use gateway_admin::model::provider_credentials::{
     AuthorizationMutationTarget, AuthorizationOwnerBinding, PendingAuthorizationMutation,
 };
 use gateway_admin::model::{MutationActor, MutationContext};
-use gateway_core::engine::credential::AccountStatus;
+use gateway_core::engine::credential::{AccountStatus, ProviderAccountId};
 use gateway_core::routing::ProviderKind;
+use provider_openai::OpenAiConfig;
 use provider_openai::credential::token_client::{
     AuthorizationCodeExchangeError, AuthorizationCodeExchanger, AuthorizationCodeGrant,
     AuthorizationTokenSet,
 };
 use provider_openai::credential::{
-    CodexCredentialAdmin, CodexOAuthAdmin, CodexOAuthAdminError, CodexOAuthAdminService,
-    CodexOAuthPendingClaimOutcome, CodexOAuthPendingStore, CodexOAuthPendingStoreError,
-    CodexOAuthSecret, CompleteCodexOAuthAuthorization, CompletedCodexOAuthCredential,
-    StartCodexOAuthAuthorization, StoredCodexPendingAuthorization,
+    CodexCredentialAdmin, CodexCredentialCodec, CodexOAuthAdmin, CodexOAuthAdminError,
+    CodexOAuthAdminService, CodexOAuthPendingClaimOutcome, CodexOAuthPendingStore,
+    CodexOAuthPendingStoreError, CodexOAuthSecret, CompleteCodexOAuthAuthorization,
+    CompletedCodexOAuthCredential, ImportCodexOAuthCredential, StartCodexOAuthAuthorization,
+    StoredCodexPendingAuthorization,
 };
+use provider_openai::transport::profile::CodexWireProfileState;
 use secrecy::SecretString;
 use url::Url;
+use uuid::Uuid;
 
-use crate::support::MemoryAccountStore;
+use crate::support::{MemoryAccountStore, profile as account_profile, secret};
 
 #[derive(Default)]
 struct PendingStore {
@@ -48,6 +53,7 @@ impl CodexOAuthPendingStore for PendingStore {
             state: pending.state().clone(),
             nonce: pending.nonce().clone(),
             code_verifier: pending.code_verifier().clone(),
+            installation_id: pending.installation_id().to_owned(),
             reauthorization_account_id: pending.reauthorization().map(ToString::to_string),
             mutation: pending.mutation().clone(),
         });
@@ -125,18 +131,37 @@ fn mutation() -> PendingAuthorizationMutation {
     )
 }
 
+fn reauthorization_mutation(
+    account_id: ProviderAccountId,
+    request_id: &str,
+) -> PendingAuthorizationMutation {
+    let context = MutationContext {
+        actor: MutationActor::System,
+        request_id: request_id.to_owned(),
+    };
+    PendingAuthorizationMutation::new(
+        ProviderKind::new("openai").expect("provider"),
+        AuthorizationMutationTarget::Reauthorize { account_id },
+        AuthorizationOwnerBinding::from_context(&context),
+    )
+}
+
 fn id_token(payload: serde_json::Value) -> String {
     let payload = URL_SAFE_NO_PAD.encode(serde_json::to_vec(&payload).expect("payload JSON"));
     // 官方逻辑只读取 payload；header/signature 不参与本地 metadata 解析。
     format!("unverified-header.{payload}.unverified-signature")
 }
 
-async fn complete(id_token: String) -> Result<CompletedCodexOAuthCredential, CodexOAuthAdminError> {
+async fn complete(
+    id_token: String,
+) -> Result<(CompletedCodexOAuthCredential, String, String), CodexOAuthAdminError> {
+    let pending_store = Arc::new(PendingStore::default());
     let service = CodexOAuthAdminService::new(
-        Arc::new(PendingStore::default()),
+        pending_store.clone(),
         Arc::new(Exchanger { id_token }),
         Arc::new(MemoryAccountStore::default()),
         CodexCredentialAdmin,
+        profile(),
     );
     let started = service
         .start_authorization(StartCodexOAuthAuthorization {
@@ -144,11 +169,26 @@ async fn complete(id_token: String) -> Result<CompletedCodexOAuthCredential, Cod
         })
         .await
         .expect("start OAuth authorization");
-    let state = Url::parse(&started.authorization_url)
-        .expect("authorization URL")
+    let outer = Url::parse(&started.authorization_url).expect("authorization URL");
+    let inner_url = outer
         .query_pairs()
-        .find_map(|(key, value)| (key == "state").then(|| value.into_owned()))
-        .expect("state parameter");
+        .find_map(|(key, value)| (key == "authorize_url").then(|| value.into_owned()))
+        .expect("inner authorization URL");
+    let inner = Url::parse(&inner_url).expect("authorization URL");
+    let parameters = inner.query_pairs().into_owned().collect::<BTreeMap<_, _>>();
+    let state = parameters.get("state").expect("state parameter");
+    let surface_stable_id = parameters
+        .get("source_surface_stable_id")
+        .expect("surface stable ID")
+        .to_owned();
+    let installation_id = pending_store
+        .pending
+        .lock()
+        .expect("pending lock")
+        .as_ref()
+        .expect("stored pending authorization")
+        .installation_id
+        .clone();
     let completed = service
         .complete_authorization(CompleteCodexOAuthAuthorization {
             owner_ref: "test-owner".to_owned(),
@@ -158,12 +198,279 @@ async fn complete(id_token: String) -> Result<CompletedCodexOAuthCredential, Cod
             )),
         })
         .await?;
-    Ok(completed.credential)
+    Ok((completed.credential, installation_id, surface_stable_id))
+}
+
+#[tokio::test]
+async fn authorize_url_matches_the_official_desktop_parameter_contract() {
+    let service = CodexOAuthAdminService::new(
+        Arc::new(PendingStore::default()),
+        Arc::new(Exchanger {
+            id_token: "unused".to_owned(),
+        }),
+        Arc::new(MemoryAccountStore::default()),
+        CodexCredentialAdmin,
+        profile(),
+    );
+    let started = service
+        .start_authorization(StartCodexOAuthAuthorization {
+            mutation: mutation(),
+        })
+        .await
+        .expect("start OAuth authorization");
+    let url = Url::parse(&started.authorization_url).expect("authorization URL");
+    let outer_parameters = url.query_pairs().into_owned().collect::<Vec<_>>();
+    assert_eq!(
+        outer_parameters
+            .iter()
+            .map(|(name, _)| name.as_str())
+            .collect::<Vec<_>>(),
+        vec![
+            "authorize_url",
+            "codex_streamlined_login",
+            "no_universal_links"
+        ]
+    );
+    let outer_parameters = outer_parameters.into_iter().collect::<BTreeMap<_, _>>();
+
+    assert_eq!(url.scheme(), "https");
+    assert_eq!(url.host_str(), Some("chatgpt.com"));
+    assert_eq!(url.path(), "/codex/desktop-auth");
+    assert_eq!(outer_parameters.len(), 3);
+    assert_eq!(
+        outer_parameters
+            .get("codex_streamlined_login")
+            .map(String::as_str),
+        Some("true")
+    );
+    assert_eq!(
+        outer_parameters
+            .get("no_universal_links")
+            .map(String::as_str),
+        Some("1")
+    );
+    let inner = Url::parse(
+        outer_parameters
+            .get("authorize_url")
+            .expect("authorize_url parameter"),
+    )
+    .expect("inner authorization URL");
+    let ordered_parameters = inner.query_pairs().into_owned().collect::<Vec<_>>();
+    assert_eq!(
+        ordered_parameters
+            .iter()
+            .map(|(name, _)| name.as_str())
+            .collect::<Vec<_>>(),
+        vec![
+            "response_type",
+            "client_id",
+            "redirect_uri",
+            "scope",
+            "code_challenge",
+            "code_challenge_method",
+            "id_token_add_organizations",
+            "codex_cli_simplified_flow",
+            "state",
+            "originator",
+            "codex_app_version",
+            "source_surface_stable_id",
+            "codex_origin_stable_id",
+            "codex_streamlined_login",
+        ]
+    );
+    let parameters = ordered_parameters.into_iter().collect::<BTreeMap<_, _>>();
+
+    assert_eq!(inner.host_str(), Some("auth.openai.com"));
+    assert_eq!(inner.path(), "/oauth/authorize");
+    assert_eq!(parameters.len(), 14);
+    assert!(
+        inner
+            .query()
+            .is_some_and(|query| query.contains("scope=openid+profile+email+offline_access"))
+    );
+    assert!(
+        !inner
+            .query()
+            .is_some_and(|query| query.contains("scope=openid%20profile"))
+    );
+    assert_eq!(
+        parameters.get("response_type").map(String::as_str),
+        Some("code")
+    );
+    assert_eq!(
+        parameters.get("redirect_uri").map(String::as_str),
+        Some("http://localhost:1455/auth/callback")
+    );
+    assert_eq!(
+        parameters.get("scope").map(String::as_str),
+        Some("openid profile email offline_access api.connectors.read api.connectors.invoke")
+    );
+    assert_eq!(
+        parameters.get("originator").map(String::as_str),
+        Some("Codex Desktop")
+    );
+    assert_eq!(
+        parameters.get("codex_app_version").map(String::as_str),
+        Some("26.803.81509")
+    );
+    let surface_stable_id = parameters
+        .get("source_surface_stable_id")
+        .expect("surface stable ID");
+    let parsed_surface_id = Uuid::parse_str(surface_stable_id).expect("UUID surface stable ID");
+    assert_eq!(parsed_surface_id.get_version_num(), 4);
+    assert_eq!(
+        parameters.get("codex_origin_stable_id"),
+        Some(surface_stable_id)
+    );
+    assert_eq!(
+        parameters
+            .get("codex_streamlined_login")
+            .map(String::as_str),
+        Some("true")
+    );
+    assert_eq!(
+        parameters.get("code_challenge_method").map(String::as_str),
+        Some("S256")
+    );
+    assert_eq!(
+        parameters
+            .get("id_token_add_organizations")
+            .map(String::as_str),
+        Some("true")
+    );
+    assert_eq!(
+        parameters
+            .get("codex_cli_simplified_flow")
+            .map(String::as_str),
+        Some("true")
+    );
+    assert!(
+        parameters
+            .get("client_id")
+            .is_some_and(|value| !value.is_empty())
+    );
+    assert!(
+        parameters
+            .get("state")
+            .is_some_and(|value| !value.is_empty())
+    );
+    assert!(
+        parameters
+            .get("code_challenge")
+            .is_some_and(|value| !value.is_empty())
+    );
+    assert!(!parameters.contains_key("nonce"));
+}
+
+fn profile() -> CodexWireProfileState {
+    OpenAiConfig::default().wire_profile_state()
+}
+
+#[tokio::test]
+async fn separate_new_authorizations_use_distinct_surface_stable_ids() {
+    let mut ids = Vec::new();
+    for _ in 0..2 {
+        let service = CodexOAuthAdminService::new(
+            Arc::new(PendingStore::default()),
+            Arc::new(Exchanger {
+                id_token: "unused".to_owned(),
+            }),
+            Arc::new(MemoryAccountStore::default()),
+            CodexCredentialAdmin,
+            profile(),
+        );
+        let started = service
+            .start_authorization(StartCodexOAuthAuthorization {
+                mutation: mutation(),
+            })
+            .await
+            .expect("start OAuth authorization");
+        let outer = Url::parse(&started.authorization_url).expect("outer authorization URL");
+        let inner = outer
+            .query_pairs()
+            .find_map(|(key, value)| (key == "authorize_url").then(|| value.into_owned()))
+            .and_then(|value| Url::parse(&value).ok())
+            .expect("inner authorization URL");
+        let surface_stable_id = inner
+            .query_pairs()
+            .find_map(|(key, value)| {
+                (key == "source_surface_stable_id").then(|| value.into_owned())
+            })
+            .expect("surface stable ID");
+        ids.push(surface_stable_id);
+    }
+    assert_ne!(ids[0], ids[1]);
+}
+
+#[tokio::test]
+async fn repeated_reauthorization_derives_the_same_surface_id_for_one_account() {
+    let store = Arc::new(MemoryAccountStore::default());
+    store
+        .seed_oauth_credential(ImportCodexOAuthCredential {
+            account_id: "acct_surface_reauth".to_owned(),
+            name: "surface reauthorization".to_owned(),
+            secret: secret("surface-reauth-access"),
+            verified_account: account_profile("chatgpt-surface-reauth"),
+            next_refresh_at: None,
+            enabled: true,
+        })
+        .await;
+    let account_id = ProviderAccountId::new("acct_surface_reauth").expect("account id");
+    let mut surface_ids = Vec::new();
+    for request_id in ["surface-reauth-first", "surface-reauth-second"] {
+        let service = CodexOAuthAdminService::new(
+            Arc::new(PendingStore::default()),
+            Arc::new(Exchanger {
+                id_token: "unused".to_owned(),
+            }),
+            store.clone(),
+            CodexCredentialAdmin,
+            profile(),
+        );
+        let started = service
+            .start_authorization(StartCodexOAuthAuthorization {
+                mutation: reauthorization_mutation(account_id.clone(), request_id),
+            })
+            .await
+            .expect("start OAuth reauthorization");
+        let outer = Url::parse(&started.authorization_url).expect("outer authorization URL");
+        let inner = outer
+            .query_pairs()
+            .find_map(|(key, value)| (key == "authorize_url").then(|| value.into_owned()))
+            .and_then(|value| Url::parse(&value).ok())
+            .expect("inner authorization URL");
+        surface_ids.push(
+            inner
+                .query_pairs()
+                .find_map(|(key, value)| {
+                    (key == "source_surface_stable_id").then(|| value.into_owned())
+                })
+                .expect("surface stable ID"),
+        );
+    }
+
+    assert_eq!(surface_ids[0], surface_ids[1]);
+}
+
+#[tokio::test]
+async fn first_exchange_persists_the_installation_id_without_reusing_it_as_the_surface_id() {
+    let (credential, pending_installation_id, surface_stable_id) = complete(id_token(
+        serde_json::json!({ "email": "identity@example.com" }),
+    ))
+    .await
+    .expect("completed OAuth exchange");
+    let CompletedCodexOAuthCredential::Create(account) = credential else {
+        panic!("expected account creation");
+    };
+    let runtime = CodexCredentialCodec::decode(&account.credential).expect("stored credential");
+
+    assert_eq!(runtime.installation_id, pending_installation_id);
+    assert_ne!(runtime.installation_id, surface_stable_id);
 }
 
 #[tokio::test]
 async fn first_exchange_uses_official_id_token_claim_mapping_without_signature_validation() {
-    let credential = complete(id_token(serde_json::json!({
+    let (credential, _, _) = complete(id_token(serde_json::json!({
         "email": "top@example.com",
         "https://api.openai.com/profile": { "email": "fallback@example.com" },
         "https://api.openai.com/auth": {
@@ -198,7 +505,7 @@ async fn first_exchange_uses_official_id_token_claim_mapping_without_signature_v
 
 #[tokio::test]
 async fn first_exchange_permits_missing_identity_claims_but_requires_a_parseable_id_token() {
-    let credential = complete(id_token(serde_json::json!({
+    let (credential, _, _) = complete(id_token(serde_json::json!({
         "https://api.openai.com/profile": { "email": "profile@example.com" }
     })))
     .await

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

@@ -485,8 +485,11 @@ pub(crate) fn agent_identity_service_with_pool(
         .no_proxy()
         .build()
         .expect("agent task client");
-    let registrar =
-        OfficialCodexAgentIdentityTaskRegistrar::new(client).expect("agent task registrar");
+    let registrar = OfficialCodexAgentIdentityTaskRegistrar::new(
+        client,
+        provider_openai::OpenAiConfig::default().wire_profile_state(),
+    )
+    .expect("agent task registrar");
     Arc::new(CodexAgentIdentityTaskService::new(
         store.repository(),
         Arc::new(registrar),

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

@@ -4,6 +4,7 @@ use provider_openai::transport::canonical::{
     CodexCanonicalDecoder, CodexCanonicalError, CodexCanonicalFailure, CodexCanonicalOutcome,
 };
 use provider_openai::transport::protocol::websocket::websocket_event_to_sse_frame;
+use serde_json::json;
 
 const METADATA_PREFIX_FIXTURE: &str = include_str!(
     "../../../../gateway-api/tests/openai/responses/fixtures/http_sse/metadata_only_prefix.sse"
@@ -262,6 +263,69 @@ fn decoder_should_use_response_service_tier_when_request_omits_it() {
     )));
 }
 
+#[test]
+fn decoder_should_add_standard_web_search_call_cost() {
+    let body = concat!(
+        "event: response.created\n",
+        "data: {\"type\":\"response.created\",\"response\":{\"id\":\"resp_search_cost\",\"model\":\"gpt-5.4\"}}\n\n",
+        "event: response.completed\n",
+        "data: {\"type\":\"response.completed\",\"response\":{\"id\":\"resp_search_cost\",\"model\":\"gpt-5.4\",\"status\":\"completed\",\"output\":[{\"type\":\"web_search_call\",\"id\":\"ws_1\",\"status\":\"completed\",\"action\":{\"type\":\"search\"}}],\"usage\":{\"input_tokens\":0,\"output_tokens\":0,\"total_tokens\":0}}}\n\n",
+    );
+    let tools = vec![json!({ "type": "web_search" })];
+    let events = CodexCanonicalDecoder::new("fallback")
+        .with_request_tool_pricing("gpt-5.4", Some(&tools))
+        .push(body.as_bytes())
+        .expect("canonical web search response");
+
+    assert!(canonical_facts(&events).into_iter().any(|event| matches!(
+        event,
+        GatewayEvent::CalculatedCost(cost)
+            if cost.total().amount().scaled() == 100_000_000
+    )));
+}
+
+#[test]
+fn decoder_should_use_non_reasoning_preview_web_search_price() {
+    let body = concat!(
+        "event: response.created\n",
+        "data: {\"type\":\"response.created\",\"response\":{\"id\":\"resp_preview_search_cost\",\"model\":\"gpt-4o\"}}\n\n",
+        "event: response.completed\n",
+        "data: {\"type\":\"response.completed\",\"response\":{\"id\":\"resp_preview_search_cost\",\"model\":\"gpt-4o\",\"status\":\"completed\",\"output\":[{\"type\":\"web_search_call\",\"id\":\"ws_1\",\"status\":\"completed\",\"action\":{\"type\":\"search\"}}],\"usage\":{\"input_tokens\":0,\"output_tokens\":0,\"total_tokens\":0}}}\n\n",
+    );
+    let tools = vec![json!({ "type": "web_search_preview" })];
+    let events = CodexCanonicalDecoder::new("fallback")
+        .with_request_tool_pricing("gpt-4o", Some(&tools))
+        .push(body.as_bytes())
+        .expect("canonical preview web search response");
+
+    assert!(canonical_facts(&events).into_iter().any(|event| matches!(
+        event,
+        GatewayEvent::CalculatedCost(cost)
+            if cost.total().amount().scaled() == 250_000_000
+    )));
+}
+
+#[test]
+fn decoder_should_fail_closed_for_fixed_block_web_search_content() {
+    let body = concat!(
+        "event: response.created\n",
+        "data: {\"type\":\"response.created\",\"response\":{\"id\":\"resp_fixed_search_cost\",\"model\":\"gpt-4o-mini\"}}\n\n",
+        "event: response.completed\n",
+        "data: {\"type\":\"response.completed\",\"response\":{\"id\":\"resp_fixed_search_cost\",\"model\":\"gpt-4o-mini\",\"status\":\"completed\",\"output\":[{\"type\":\"web_search_call\",\"id\":\"ws_1\",\"status\":\"completed\",\"action\":{\"type\":\"search\"}}],\"usage\":{\"input_tokens\":1,\"output_tokens\":1,\"total_tokens\":2}}}\n\n",
+    );
+    let tools = vec![json!({ "type": "web_search" })];
+    let events = CodexCanonicalDecoder::new("fallback")
+        .with_request_tool_pricing("gpt-4o-mini", Some(&tools))
+        .push(body.as_bytes())
+        .expect("canonical fixed-block search response");
+
+    assert!(
+        canonical_facts(&events)
+            .into_iter()
+            .all(|event| !matches!(event, GatewayEvent::CalculatedCost(_)))
+    );
+}
+
 #[test]
 fn websocket_decoder_should_prefer_requested_service_tier_over_response_tier_for_billing() {
     let created = websocket_event_to_sse_frame(
@@ -423,6 +487,25 @@ fn decoder_should_preserve_image_tool_tokens_in_canonical_usage() {
     assert_eq!(image_usage, Some((Some(31), Some(9))));
 }
 
+#[test]
+fn decoder_should_not_understate_cost_when_image_tool_price_is_ambiguous() {
+    let body = concat!(
+        "event: response.created\n",
+        "data: {\"type\":\"response.created\",\"response\":{\"id\":\"resp_image_cost\",\"model\":\"gpt-5.4\"}}\n\n",
+        "event: response.completed\n",
+        "data: {\"type\":\"response.completed\",\"response\":{\"id\":\"resp_image_cost\",\"model\":\"gpt-5.4\",\"status\":\"completed\",\"usage\":{\"input_tokens\":12,\"output_tokens\":5,\"total_tokens\":17},\"tool_usage\":{\"image_gen\":{\"input_tokens\":31,\"output_tokens\":9}}}}\n\n",
+    );
+    let events = CodexCanonicalDecoder::new("fallback")
+        .push(body.as_bytes())
+        .expect("canonical image cost response");
+
+    assert!(
+        canonical_facts(&events)
+            .into_iter()
+            .all(|event| !matches!(event, GatewayEvent::CalculatedCost(_)))
+    );
+}
+
 #[test]
 fn decoder_should_drop_rate_limit_control_and_preserve_metadata_lifecycle_events() {
     let body = concat!(

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

@@ -223,6 +223,7 @@ async fn fetch_should_send_official_catalog_headers_and_query() {
             "user-agent",
             "codex_cli_rs/0.144.0 (linux 6.8; x86_64) xterm (codex_cli_rs; 1.0.0)",
         ))
+        .and(header("x-openai-internal-codex-residency", "us"))
         .and(header("accept", "application/json"))
         .and(header("x-codex-installation-id", "installation-123"))
         .respond_with(
@@ -254,7 +255,6 @@ async fn fetch_should_send_official_catalog_headers_and_query() {
     for forbidden in [
         "content-type",
         "openai-beta",
-        "x-openai-internal-codex-residency",
         "x-client-request-id",
         "session_id",
         "session-id",

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

@@ -1,4 +1,4 @@
-use provider_openai::transport::{endpoint_url, usage_endpoint_urls};
+use provider_openai::transport::{endpoint_url, usage_endpoint_url};
 
 #[test]
 fn endpoints_should_join_origin_and_backend_paths_without_double_slashes() {
@@ -9,12 +9,9 @@ fn endpoints_should_join_origin_and_backend_paths_without_double_slashes() {
 }
 
 #[test]
-fn usage_endpoints_should_preserve_the_official_backend_api_fallback_order() {
+fn usage_endpoint_should_use_the_official_wham_path_for_backend_api() {
     assert_eq!(
-        usage_endpoint_urls("https://chatgpt.com/backend-api"),
-        vec![
-            "https://chatgpt.com/backend-api/wham/usage".to_owned(),
-            "https://chatgpt.com/backend-api/codex/usage".to_owned(),
-        ]
+        usage_endpoint_url("https://chatgpt.com/backend-api"),
+        "https://chatgpt.com/backend-api/wham/usage"
     );
 }

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

@@ -178,6 +178,7 @@ fn websocket_opening_audit_should_redact_sensitive_headers() {
                 "x-codex-turn-metadata".to_owned(),
                 "{\"secret\":true}".to_owned(),
             ),
+            ("x-oai-attestation".to_owned(), "device-token".to_owned()),
         ],
     );
     let audit = connection.opening_audit_snapshot();
@@ -190,6 +191,7 @@ fn websocket_opening_audit_should_redact_sensitive_headers() {
             "user-agent",
             "x-client-request-id",
             "x-codex-turn-metadata",
+            "x-oai-attestation",
         ]
     );
     assert!(
@@ -313,6 +315,7 @@ async fn backend_websocket_should_forward_context_headers_and_preserve_payload_f
         ("content-type", "application/json"),
         ("accept", "text/event-stream"),
         ("x-codex-installation-id", "install-123"),
+        ("openai-beta", "responses_websockets=2026-02-06"),
         ("x-openai-internal-codex-residency", "us"),
         ("x-codex-turn-state", "turn-state"),
         ("x-codex-turn-metadata", "{\"thread_source\":\"subagent\"}"),
@@ -382,10 +385,7 @@ async fn backend_http_should_restore_opaque_multivalue_header_bytes_and_lease_id
         raw_header_values(&raw, "x-openai-future-mode"),
         vec![b"future-ascii".to_vec(), b"\x80\xff".to_vec()]
     );
-    assert_eq!(
-        raw_header_values(&raw, "openai-beta"),
-        vec![b"future_responses=v2".to_vec(), b"future_tools=v3".to_vec()]
-    );
+    assert!(raw_header_values(&raw, "openai-beta").is_empty());
     assert_eq!(
         raw_header_values(&raw, "x-codex-turn-state"),
         vec![b"turn-ascii".to_vec(), b"turn-\x80".to_vec()]
@@ -396,10 +396,7 @@ async fn backend_http_should_restore_opaque_multivalue_header_bytes_and_lease_id
             "content-type",
             b"application/vnd.openai.request+json".as_slice(),
         ),
-        (
-            "x-openai-internal-codex-residency",
-            b"future-region".as_slice(),
-        ),
+        ("x-openai-internal-codex-residency", b"us".as_slice()),
         ("x-still-valid", b"after-invalid".as_slice()),
         ("authorization", b"Bearer lease-token".as_slice()),
         ("chatgpt-account-id", b"lease-account".as_slice()),
@@ -416,7 +413,12 @@ async fn backend_http_should_restore_opaque_multivalue_header_bytes_and_lease_id
         raw_header_values(&raw, "originator"),
         vec![b"codex_cli_rs".to_vec()]
     );
-    for dropped in ["x-oai-attestation", "x-oai-is", "x-oai-is-update"] {
+    for dropped in [
+        "openai-beta",
+        "x-oai-attestation",
+        "x-oai-is",
+        "x-oai-is-update",
+    ] {
         assert!(
             raw_header_values(&raw, dropped).is_empty(),
             "unexpected {dropped}"
@@ -509,7 +511,7 @@ async fn backend_websocket_should_drop_only_unrepresentable_opaque_header_values
     );
     assert_eq!(
         values("openai-beta"),
-        vec![b"future_responses=v2".to_vec(), b"future_tools=v3".to_vec()]
+        vec![b"responses_websockets=2026-02-06".to_vec()]
     );
     assert_eq!(
         values("accept"),
@@ -524,6 +526,10 @@ async fn backend_websocket_should_drop_only_unrepresentable_opaque_header_values
         vec![test_wire_profile().snapshot().user_agent().into_bytes()]
     );
     assert_eq!(values("originator"), vec![b"codex_cli_rs".to_vec()]);
+    assert_eq!(
+        values("x-openai-internal-codex-residency"),
+        vec![b"us".to_vec()]
+    );
     for dropped in ["x-oai-attestation", "x-oai-is", "x-oai-is-update"] {
         assert!(values(dropped).is_empty(), "unexpected {dropped}");
     }
@@ -631,7 +637,6 @@ async fn backend_http_should_send_codex_context_without_browser_headers() {
         "content-type",
         "cookie",
         "accept",
-        "openai-beta",
         "x-openai-internal-codex-residency",
         "x-client-request-id",
         "x-codex-installation-id",
@@ -654,6 +659,7 @@ async fn backend_http_should_send_codex_context_without_browser_headers() {
         "sec-fetch-site",
         "sec-fetch-mode",
         "sec-fetch-dest",
+        "openai-beta",
     ] {
         assert!(header_names.iter().all(|name| name != forbidden));
     }
@@ -947,7 +953,12 @@ async fn websocket_should_keep_an_exact_chain_while_new_connections_adopt_the_la
         )
         .await
         .expect("first response");
-    profile.update_cli_release("1.2.4");
+    profile.update_bundled_release(&CodexBundledReleaseProfile {
+        codex_version: "1.2.4".to_owned(),
+        desktop_version: "1.2.4".to_owned(),
+        desktop_build: "124".to_owned(),
+        verified_at: Utc::now(),
+    });
 
     let mut continuation = request.clone();
     continuation.set_previous_response_id(Some("resp_profile_first".to_owned()));

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

@@ -10,7 +10,9 @@ use std::{
 use chrono::Utc;
 use futures::{SinkExt, StreamExt};
 use gateway_protocol::openai::events::{TokenUsage, extract_sse_usage};
-use provider_openai::transport::profile::{CodexWireProfile, CodexWireProfileState};
+use provider_openai::transport::profile::{
+    CodexBundledReleaseProfile, CodexWireProfile, CodexWireProfileState,
+};
 use provider_openai::transport::protocol::responses::{
     CodexResponsesRequest, PreviousResponseScope,
 };

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

@@ -1,60 +1,131 @@
 use std::collections::VecDeque;
 use std::sync::{Arc, Mutex};
+use std::time::Duration;
 
 use chrono::{TimeZone, Utc};
 use futures::future::BoxFuture;
+use gateway_core::provider_ports::{
+    ProviderArtifactProfile, ProviderArtifactProfileCachePort, ProviderStoreError,
+    ProviderStoreErrorKind,
+};
+use gateway_core::routing::ProviderKind;
 use provider_openai::transport::profile::{
-    CodexCliReleaseError, CodexCliReleaseService, CodexCliReleaseTransport, CodexDesktopRelease,
+    CodexArtifactProfileCache, CodexBundledReleaseProfile, CodexDesktopRelease,
     CodexDesktopReleaseError, CodexDesktopReleaseService, CodexDesktopReleaseTransport,
-    CodexWireProfile, CodexWireProfileState, parse_desktop_release,
+    CodexWireProfile, CodexWireProfileState, OfficialCodexDesktopReleaseTransport,
+    parse_desktop_release,
 };
 
+mod desktop_artifact;
+
 struct ReleaseTransport {
-    outcomes: Mutex<VecDeque<Result<CodexDesktopRelease, CodexDesktopReleaseError>>>,
+    releases: Mutex<VecDeque<Result<CodexDesktopRelease, CodexDesktopReleaseError>>>,
+    core_versions: Mutex<VecDeque<Result<String, CodexDesktopReleaseError>>>,
+    core_probes: Mutex<usize>,
 }
 
 impl ReleaseTransport {
     fn new(
-        outcomes: impl IntoIterator<Item = Result<CodexDesktopRelease, CodexDesktopReleaseError>>,
+        releases: impl IntoIterator<Item = Result<CodexDesktopRelease, CodexDesktopReleaseError>>,
+        core_versions: impl IntoIterator<Item = Result<String, CodexDesktopReleaseError>>,
     ) -> Self {
         Self {
-            outcomes: Mutex::new(outcomes.into_iter().collect()),
+            releases: Mutex::new(releases.into_iter().collect()),
+            core_versions: Mutex::new(core_versions.into_iter().collect()),
+            core_probes: Mutex::new(0),
         }
     }
+
+    fn core_probes(&self) -> usize {
+        *self.core_probes.lock().expect("core probe count")
+    }
 }
 
 impl CodexDesktopReleaseTransport for ReleaseTransport {
     fn fetch(&self) -> BoxFuture<'_, Result<CodexDesktopRelease, CodexDesktopReleaseError>> {
         Box::pin(async move {
-            self.outcomes
+            self.releases
                 .lock()
                 .expect("release outcomes")
                 .pop_front()
                 .expect("release outcome")
         })
     }
+
+    fn fetch_bundled_core_version<'a>(
+        &'a self,
+        _release: &'a CodexDesktopRelease,
+    ) -> BoxFuture<'a, Result<String, CodexDesktopReleaseError>> {
+        Box::pin(async move {
+            *self.core_probes.lock().expect("core probe count") += 1;
+            self.core_versions
+                .lock()
+                .expect("Core version outcomes")
+                .pop_front()
+                .expect("Core version outcome")
+        })
+    }
 }
 
-struct CliReleaseTransport {
-    outcomes: Mutex<VecDeque<Result<String, CodexCliReleaseError>>>,
+#[derive(Default)]
+struct ArtifactProfiles {
+    profile: Mutex<Option<ProviderArtifactProfile>>,
+    fail_writes: bool,
 }
 
-impl CliReleaseTransport {
-    fn new(outcomes: impl IntoIterator<Item = Result<String, CodexCliReleaseError>>) -> Self {
+impl ArtifactProfiles {
+    fn failing() -> Self {
         Self {
-            outcomes: Mutex::new(outcomes.into_iter().collect()),
+            profile: Mutex::new(None),
+            fail_writes: true,
         }
     }
 }
 
-impl CodexCliReleaseTransport for CliReleaseTransport {
-    fn fetch(&self) -> BoxFuture<'_, Result<String, CodexCliReleaseError>> {
+impl ProviderArtifactProfileCachePort for ArtifactProfiles {
+    fn replace_if_newer(
+        &self,
+        profile: ProviderArtifactProfile,
+        _ttl: Duration,
+    ) -> BoxFuture<'_, Result<bool, ProviderStoreError>> {
         Box::pin(async move {
-            self.outcomes
+            if self.fail_writes {
+                return Err(ProviderStoreError::new(
+                    ProviderStoreErrorKind::Unavailable,
+                    "replace test artifact profile",
+                ));
+            }
+            let mut current = self.profile.lock().expect("artifact profile");
+            if let Some(current) = current.as_ref() {
+                if current.artifact_sequence() > profile.artifact_sequence() {
+                    return Ok(false);
+                }
+                if current.artifact_sequence() == profile.artifact_sequence()
+                    && current.profile() != profile.profile()
+                {
+                    return Err(ProviderStoreError::new(
+                        ProviderStoreErrorKind::Conflict,
+                        "replace test artifact profile",
+                    ));
+                }
+            }
+            *current = Some(profile);
+            Ok(true)
+        })
+    }
+
+    fn read<'a>(
+        &'a self,
+        provider_kind: &'a ProviderKind,
+    ) -> BoxFuture<'a, Result<Option<ProviderArtifactProfile>, ProviderStoreError>> {
+        Box::pin(async move {
+            Ok(self
+                .profile
                 .lock()
-                .expect("CLI release outcomes")
-                .pop_front()
-                .expect("CLI release outcome")
+                .expect("artifact profile")
+                .as_ref()
+                .filter(|profile| profile.provider_kind() == provider_kind)
+                .cloned())
         })
     }
 }
@@ -63,22 +134,22 @@ impl CodexCliReleaseTransport for CliReleaseTransport {
 fn wire_profile_should_generate_bundled_core_app_server_user_agent() {
     let profile = CodexWireProfile {
         originator: "Codex Desktop".to_owned(),
-        codex_version: "0.144.2".to_owned(),
-        desktop_version: "26.707.72221".to_owned(),
-        desktop_build: "72221".to_owned(),
+        codex_version: "0.147.0-alpha.6.6".to_owned(),
+        desktop_version: "26.803.81509".to_owned(),
+        desktop_build: "6415".to_owned(),
         os_type: "Mac OS".to_owned(),
         os_version: "15.7.1".to_owned(),
         arch: "arm64".to_owned(),
         terminal: "unknown".to_owned(),
         verified_at: Utc
-            .with_ymd_and_hms(2026, 7, 18, 0, 0, 0)
+            .with_ymd_and_hms(2026, 8, 3, 0, 0, 0)
             .single()
             .expect("valid fixture time"),
     };
 
     assert_eq!(
         profile.user_agent(),
-        "Codex Desktop/0.144.2 (Mac OS 15.7.1; arm64) unknown (Codex Desktop; 26.707.72221)"
+        "Codex Desktop/0.147.0-alpha.6.6 (Mac OS 15.7.1; arm64) unknown (Codex Desktop; 26.803.81509)"
     );
 }
 
@@ -95,7 +166,7 @@ fn desktop_appcast_should_publish_the_first_complete_release() {
               <pubDate>Sun, 19 Jul 2026 08:00:00 +0000</pubDate>
               <sparkle:minimumSystemVersion>14.0</sparkle:minimumSystemVersion>
               <sparkle:hardwareRequirements>arm64</sparkle:hardwareRequirements>
-              <enclosure url="https://persistent.oaistatic.com/codex.dmg"
+              <enclosure url="https://persistent.oaistatic.com/codex.zip"
                 length="123456" sparkle:shortVersionString="26.720.1"
                 sparkle:version="72001" sparkle:edSignature="signature" />
             </item>
@@ -135,77 +206,75 @@ fn desktop_appcast_should_reject_invalid_version_and_build() {
 }
 
 #[tokio::test]
-async fn wire_profile_release_update_should_change_only_desktop_identity() {
+async fn bundled_release_update_should_change_core_and_desktop_identity_atomically() {
     let original = wire_profile();
     let original_user_agent = original.user_agent();
     let state = CodexWireProfileState::new(original.clone());
-    let service = CodexDesktopReleaseService::new(
-        state.clone(),
-        Arc::new(ReleaseTransport::new([Ok(release("26.720.1", "72001"))])),
-    );
+    let transport = Arc::new(ReleaseTransport::new(
+        [Ok(release("26.810.41047", "6570"))],
+        [Ok("0.148.0-alpha.9".to_owned())],
+    ));
+    let cache = Arc::new(ArtifactProfiles::default());
+    let service = service(state.clone(), transport.clone(), cache.clone(), None);
 
     service.refresh().await.expect("release refresh");
 
     let updated = state.snapshot();
-    assert_eq!(updated.desktop_version, "26.720.1");
-    assert_eq!(updated.desktop_build, "72001");
-    assert_eq!(updated.codex_version, original.codex_version);
+    assert_eq!(updated.desktop_version, "26.810.41047");
+    assert_eq!(updated.desktop_build, "6570");
+    assert_eq!(updated.codex_version, "0.148.0-alpha.9");
     assert_eq!(updated.originator, original.originator);
     assert_eq!(updated.os_type, original.os_type);
     assert_eq!(updated.os_version, original.os_version);
     assert_eq!(updated.arch, original.arch);
     assert_eq!(updated.terminal, original.terminal);
-    assert_eq!(updated.verified_at, original.verified_at);
+    assert!(updated.verified_at > original.verified_at);
     assert_ne!(updated.user_agent(), original_user_agent);
     assert_eq!(
         updated.user_agent(),
-        "Codex Desktop/0.144.2 (Mac OS 15.7.1; arm64) unknown (Codex Desktop; 26.720.1)"
+        "Codex Desktop/0.148.0-alpha.9 (Mac OS 15.7.1; arm64) unknown (Codex Desktop; 26.810.41047)"
     );
+    assert_eq!(transport.core_probes(), 1);
+    assert!(cache.profile.lock().expect("artifact profile").is_some());
     let status = service.status().snapshot();
-    assert_eq!(status.latest, Some(release("26.720.1", "72001")));
+    assert_eq!(status.latest, Some(release("26.810.41047", "6570")));
     assert!(status.checked_at.is_some());
     assert!(status.last_error.is_none());
 }
 
 #[tokio::test]
-async fn cli_release_update_should_change_only_codex_core_identity() {
-    let original = wire_profile();
-    let original_user_agent = original.user_agent();
-    let state = CodexWireProfileState::new(original.clone());
-    let service = CodexCliReleaseService::new(
-        state.clone(),
-        Arc::new(CliReleaseTransport::new([Ok("0.145.0".to_owned())])),
-    );
+async fn unchanged_verified_release_should_refresh_cache_without_downloading_core_again() {
+    let state = CodexWireProfileState::new(wire_profile());
+    let latest = release("26.810.41047", "6570");
+    let transport = Arc::new(ReleaseTransport::new(
+        [Ok(latest.clone()), Ok(latest)],
+        [Ok("0.148.0-alpha.9".to_owned())],
+    ));
+    let cache = Arc::new(ArtifactProfiles::default());
+    let service = service(state, transport.clone(), cache, None);
 
-    service.refresh().await.expect("CLI release refresh");
+    service.refresh().await.expect("first release refresh");
+    service.refresh().await.expect("cached release refresh");
 
-    let updated = state.snapshot();
-    assert_eq!(updated.codex_version, "0.145.0");
-    assert_eq!(updated.desktop_version, original.desktop_version);
-    assert_eq!(updated.desktop_build, original.desktop_build);
-    assert_eq!(updated.originator, original.originator);
-    assert_eq!(updated.os_type, original.os_type);
-    assert_eq!(updated.os_version, original.os_version);
-    assert_eq!(updated.arch, original.arch);
-    assert_eq!(updated.terminal, original.terminal);
-    assert_eq!(updated.verified_at, original.verified_at);
-    assert_ne!(updated.user_agent(), original_user_agent);
-    assert_eq!(
-        updated.user_agent(),
-        "Codex Desktop/0.145.0 (Mac OS 15.7.1; arm64) unknown (Codex Desktop; 26.707.72221)"
-    );
+    assert_eq!(transport.core_probes(), 1);
 }
 
 #[tokio::test]
 async fn failed_release_refresh_should_preserve_the_last_successful_profile_and_release() {
     let state = CodexWireProfileState::new(wire_profile());
-    let latest = release("26.720.1", "72001");
-    let service = CodexDesktopReleaseService::new(
-        state.clone(),
-        Arc::new(ReleaseTransport::new([
+    let latest = release("26.810.41047", "6570");
+    let transport = Arc::new(ReleaseTransport::new(
+        [
             Ok(latest.clone()),
             Err(CodexDesktopReleaseError::InvalidDocument),
-        ])),
+        ],
+        [Ok("0.148.0-alpha.9".to_owned())],
+    ));
+    let service = service(
+        state.clone(),
+        transport,
+        Arc::new(ArtifactProfiles::default()),
+        None,
     );
     service.refresh().await.expect("first release refresh");
     let successful_profile = state.snapshot();
@@ -225,18 +294,74 @@ async fn failed_release_refresh_should_preserve_the_last_successful_profile_and_
     assert!(status.checked_at.is_some());
 }
 
+#[tokio::test]
+async fn cache_failure_should_not_publish_a_partially_verified_profile() {
+    let original = wire_profile();
+    let state = CodexWireProfileState::new(original.clone());
+    let transport = Arc::new(ReleaseTransport::new(
+        [Ok(release("26.810.41047", "6570"))],
+        [Ok("0.148.0-alpha.9".to_owned())],
+    ));
+    let service = service(
+        state.clone(),
+        transport,
+        Arc::new(ArtifactProfiles::failing()),
+        None,
+    );
+
+    service
+        .refresh()
+        .await
+        .expect_err("cache failure must fail the refresh");
+
+    assert_eq!(state.snapshot(), original);
+    assert!(service.status().snapshot().latest.is_none());
+}
+
+#[tokio::test]
+#[ignore = "downloads bounded ranges from the current official macOS artifact"]
+async fn official_desktop_artifact_should_expose_its_bundled_core_version() {
+    let transport = OfficialCodexDesktopReleaseTransport::new().expect("official transport");
+    let release = transport.fetch().await.expect("official appcast release");
+    let core_version = transport
+        .fetch_bundled_core_version(&release)
+        .await
+        .expect("bundled Core version");
+
+    semver::Version::parse(&core_version).expect("semantic bundled Core version");
+    eprintln!(
+        "official Desktop {} (build {}) bundles Core {core_version}",
+        release.version, release.build
+    );
+}
+
+fn service(
+    state: CodexWireProfileState,
+    transport: Arc<ReleaseTransport>,
+    cache: Arc<ArtifactProfiles>,
+    verified: Option<CodexBundledReleaseProfile>,
+) -> CodexDesktopReleaseService {
+    let provider = ProviderKind::new("openai").expect("provider kind");
+    CodexDesktopReleaseService::new(
+        state,
+        transport,
+        CodexArtifactProfileCache::new(provider, cache),
+        verified,
+    )
+}
+
 fn wire_profile() -> CodexWireProfile {
     CodexWireProfile {
         originator: "Codex Desktop".to_owned(),
-        codex_version: "0.144.2".to_owned(),
-        desktop_version: "26.707.72221".to_owned(),
-        desktop_build: "72221".to_owned(),
+        codex_version: "0.147.0-alpha.6.6".to_owned(),
+        desktop_version: "26.803.81509".to_owned(),
+        desktop_build: "6415".to_owned(),
         os_type: "Mac OS".to_owned(),
         os_version: "15.7.1".to_owned(),
         arch: "arm64".to_owned(),
         terminal: "unknown".to_owned(),
         verified_at: Utc
-            .with_ymd_and_hms(2026, 7, 18, 0, 0, 0)
+            .with_ymd_and_hms(2026, 8, 3, 0, 0, 0)
             .single()
             .expect("valid fixture time"),
     }
@@ -249,8 +374,8 @@ fn release(version: &str, build: &str) -> CodexDesktopRelease {
         published_at: None,
         minimum_system_version: Some("14.0".to_owned()),
         hardware_requirements: Some("arm64".to_owned()),
-        download_url: Some("https://persistent.oaistatic.com/codex.dmg".to_owned()),
-        download_size: Some(123_456),
+        download_url: Some("https://persistent.oaistatic.com/codex.zip".to_owned()),
+        download_size: Some(589_926_710),
         signature_present: true,
     }
 }

backend/crates/providers/openai/tests/transport/profile/desktop_artifact.rs

@@ -0,0 +1,83 @@
+#[allow(dead_code)]
+#[path = "../../../src/transport/profile/desktop_artifact.rs"]
+mod subject;
+
+use subject::{CoreVersionScanner, find_core_entry, parse_content_range};
+
+#[test]
+fn scanner_accepts_stable_and_multi_part_alpha_versions_across_chunks() {
+    for expected in ["0.148.0", "0.147.0-alpha.6.6"] {
+        let bytes = format!("prefixcodex-mcp-client/{expected}OAuth suffix");
+        let split = bytes.find("client").expect("split point") + 3;
+        let mut scanner = CoreVersionScanner::default();
+        assert_eq!(
+            scanner.push(&bytes.as_bytes()[..split]).expect("scan"),
+            None
+        );
+        assert_eq!(
+            scanner.push(&bytes.as_bytes()[split..]).expect("scan"),
+            Some(expected.to_owned())
+        );
+    }
+}
+
+#[test]
+fn scanner_rejects_unrecognized_prerelease_shapes() {
+    let mut scanner = CoreVersionScanner::default();
+    assert!(
+        scanner
+            .push(b"codex-mcp-client/0.149.0-nightly.1 next")
+            .expect("scan")
+            .is_none()
+    );
+}
+
+#[test]
+fn scanner_rejects_build_metadata_instead_of_truncating_the_version() {
+    for version in ["0.149.0+desktop.1", "0.149.0-alpha.1+desktop.1"] {
+        let mut scanner = CoreVersionScanner::default();
+        let bytes = format!("codex-mcp-client/{version}OAuth");
+        assert!(scanner.push(bytes.as_bytes()).expect("scan").is_none());
+    }
+}
+
+#[test]
+fn scanner_skips_an_invalid_marker_before_the_bundled_version() {
+    let mut scanner = CoreVersionScanner::default();
+    assert_eq!(
+        scanner
+            .push(b"codex-mcp-client/not-a-version codex-mcp-client/0.147.0-alpha.6.6OAuth",)
+            .expect("scan"),
+        Some("0.147.0-alpha.6.6".to_owned())
+    );
+}
+
+#[test]
+fn central_directory_requires_one_deflated_core_entry() {
+    let name = b"ChatGPT.app/Contents/Resources/codex";
+    let mut central = vec![0_u8; 46 + name.len()];
+    central[..4].copy_from_slice(b"PK\x01\x02");
+    central[10..12].copy_from_slice(&8_u16.to_le_bytes());
+    central[20..24].copy_from_slice(&100_u32.to_le_bytes());
+    central[24..28].copy_from_slice(&200_u32.to_le_bytes());
+    central[28..30].copy_from_slice(
+        &u16::try_from(name.len())
+            .expect("name length")
+            .to_le_bytes(),
+    );
+    central[42..46].copy_from_slice(&50_u32.to_le_bytes());
+    central[46..].copy_from_slice(name);
+
+    let entry = find_core_entry(&central, 1).expect("core entry");
+    assert_eq!(entry.name, name);
+    assert_eq!(entry.compressed_size, 100);
+    assert_eq!(entry.uncompressed_size, 200);
+    assert_eq!(entry.local_header_offset, 50);
+}
+
+#[test]
+fn content_range_parser_is_strict() {
+    assert_eq!(parse_content_range("bytes 10-19/100"), Some((10, 19, 100)));
+    assert_eq!(parse_content_range("bytes */100"), None);
+    assert_eq!(parse_content_range("10-19/100"), None);
+}

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

@@ -8,7 +8,7 @@ use chrono::{TimeZone, Utc};
 use provider_openai::transport::profile::{CodexWireProfile, CodexWireProfileState};
 use provider_openai::transport::{
     CodexBackendClient, CodexClientError, CodexRequestContext, MAX_CODEX_USAGE_BODY_BYTES,
-    build_reqwest_client, openai_billing_breakdown,
+    OpenAiBillingUsage, build_reqwest_client, openai_billing_breakdown,
 };
 use reqwest::StatusCode;
 use wiremock::matchers::{method, path};
@@ -16,10 +16,24 @@ use wiremock::{Mock, MockServer, ResponseTemplate};
 
 const OVERSIZED_BODY_ERROR: &str = "upstream usage response exceeded the body limit";
 
+fn billing_usage(
+    input_tokens: u64,
+    output_tokens: u64,
+    cached_tokens: u64,
+    cache_write_tokens: u64,
+) -> OpenAiBillingUsage {
+    OpenAiBillingUsage::new(
+        input_tokens,
+        output_tokens,
+        cached_tokens,
+        cache_write_tokens,
+    )
+}
+
 #[test]
 fn billing_breakdown_should_preserve_input_output_and_cache_components() {
-    let breakdown =
-        openai_billing_breakdown("gpt-5.6-sol", 100, 5, 20, 10, None).expect("known model pricing");
+    let breakdown = openai_billing_breakdown("gpt-5.6-sol", billing_usage(100, 5, 20, 10), None)
+        .expect("known model pricing");
 
     assert_eq!(breakdown.input_amount().amount().scaled(), 3_500_000);
     assert_eq!(breakdown.output_amount().amount().scaled(), 1_500_000);
@@ -32,50 +46,120 @@ fn billing_breakdown_should_preserve_input_output_and_cache_components() {
 
 #[test]
 fn billing_breakdown_should_use_latest_gpt_5_6_and_cached_input_prices() {
-    let terra = openai_billing_breakdown("gpt-5.6-terra", 1, 1, 0, 0, None)
+    let terra = openai_billing_breakdown("gpt-5.6-terra", billing_usage(1, 1, 0, 0), None)
         .expect("gpt-5.6-terra standard pricing");
     assert_eq!(terra.total_amount().amount().scaled(), 140_000);
 
-    let terra_fast = openai_billing_breakdown("gpt-5.6-terra", 1, 1, 0, 0, Some("fast"))
-        .expect("gpt-5.6-terra fast pricing");
+    let terra_fast =
+        openai_billing_breakdown("gpt-5.6-terra", billing_usage(1, 1, 0, 0), Some("fast"))
+            .expect("gpt-5.6-terra fast pricing");
     assert_eq!(terra_fast.total_amount().amount().scaled(), 280_000);
 
-    let terra_long = openai_billing_breakdown("gpt-5.6-terra", 272_001, 0, 0, 0, None)
-        .expect("gpt-5.6-terra long-context pricing");
+    let terra_long =
+        openai_billing_breakdown("gpt-5.6-terra", billing_usage(272_001, 0, 0, 0), None)
+            .expect("gpt-5.6-terra long-context pricing");
     assert_eq!(terra_long.total_amount().amount().scaled(), 10_880_040_000);
 
-    let luna = openai_billing_breakdown("gpt-5.6-luna", 1, 1, 0, 0, None)
+    let luna = openai_billing_breakdown("gpt-5.6-luna", billing_usage(1, 1, 0, 0), None)
         .expect("gpt-5.6-luna standard pricing");
     assert_eq!(luna.total_amount().amount().scaled(), 14_000);
 
-    let gpt_4o =
-        openai_billing_breakdown("gpt-4o", 1, 1, 1, 0, None).expect("gpt-4o cached-input pricing");
+    let gpt_4o = openai_billing_breakdown("gpt-4o", billing_usage(1, 1, 1, 0), None)
+        .expect("gpt-4o cached-input pricing");
     assert_eq!(gpt_4o.total_amount().amount().scaled(), 112_500);
 
-    let gpt_4o_mini = openai_billing_breakdown("gpt-4o-mini", 1, 1, 1, 0, None)
+    let gpt_4o_mini = openai_billing_breakdown("gpt-4o-mini", billing_usage(1, 1, 1, 0), None)
         .expect("gpt-4o-mini cached-input pricing");
     assert_eq!(gpt_4o_mini.total_amount().amount().scaled(), 6_750);
 }
 
 #[test]
 fn billing_breakdown_should_apply_fast_and_flex_tiers_without_guessing_unknown_models() {
-    let fast = openai_billing_breakdown("gpt-5.4", 1, 1, 0, 0, Some("fast")).expect("fast pricing");
-    let flex = openai_billing_breakdown("gpt-5.4", 1, 1, 0, 0, Some("flex")).expect("flex pricing");
+    let fast = openai_billing_breakdown("gpt-5.4", billing_usage(1, 1, 0, 0), Some("fast"))
+        .expect("fast pricing");
+    let flex = openai_billing_breakdown("gpt-5.4", billing_usage(1, 1, 0, 0), Some("flex"))
+        .expect("flex pricing");
 
     assert_eq!(fast.total_amount().amount().scaled(), 350_000);
     assert_eq!(fast.service_tier(), Some("fast"));
-    assert_eq!(fast.multiplier_percent(), 100);
+    assert_eq!(fast.multiplier_percent(), 200);
     assert_eq!(flex.total_amount().amount().scaled(), 87_500);
     assert_eq!(flex.multiplier_percent(), 50);
-    assert!(openai_billing_breakdown("unknown-model", 1, 1, 0, 0, None).is_none());
+    assert!(openai_billing_breakdown("unknown-model", billing_usage(1, 1, 0, 0), None).is_none());
+}
+
+#[test]
+fn billing_breakdown_should_use_official_fast_long_context_prices() {
+    let breakdown =
+        openai_billing_breakdown("gpt-5.6-sol", billing_usage(272_001, 1, 1, 1), Some("fast"))
+            .expect("gpt-5.6-sol fast long-context pricing");
+
+    assert_eq!(
+        (
+            breakdown.input_price_per_million().amount().scaled(),
+            breakdown.cache_read_price_per_million().amount().scaled(),
+            breakdown.cache_write_price_per_million().amount().scaled(),
+            breakdown.output_price_per_million().amount().scaled(),
+        ),
+        (
+            200_000_000_000,
+            20_000_000_000,
+            250_000_000_000,
+            900_000_000_000,
+        )
+    );
+}
+
+#[test]
+fn billing_breakdown_should_fail_closed_for_unpublished_prices() {
+    assert_eq!(
+        (
+            openai_billing_breakdown("gpt-5.5", billing_usage(272_001, 1, 0, 0), Some("fast"))
+                .is_none(),
+            openai_billing_breakdown("gpt-5.5-pro", billing_usage(1, 1, 0, 0), Some("fast"))
+                .is_none(),
+            openai_billing_breakdown("gpt-5.3-codex-spark", billing_usage(1, 1, 0, 0), None)
+                .is_none(),
+            openai_billing_breakdown("gpt-4o", billing_usage(1, 1, 0, 0), Some("flex")).is_none(),
+        ),
+        (true, true, true, true)
+    );
+}
+
+#[test]
+fn billing_breakdown_should_use_official_gpt_4o_fast_prices() {
+    let breakdown = openai_billing_breakdown("gpt-4o", billing_usage(1, 1, 1, 0), Some("priority"))
+        .expect("gpt-4o fast pricing");
+
+    assert_eq!(
+        (
+            breakdown.cache_read_price_per_million().amount().scaled(),
+            breakdown.output_price_per_million().amount().scaled(),
+            breakdown.standard_amount().amount().scaled(),
+            breakdown.total_amount().amount().scaled(),
+            breakdown.multiplier_percent(),
+        ),
+        (21_250_000_000, 170_000_000_000, 112_500, 191_250, 170,)
+    );
+}
+
+#[test]
+fn billing_breakdown_should_use_published_flex_cache_price() {
+    let breakdown = openai_billing_breakdown("gpt-5.4", billing_usage(1, 0, 1, 0), Some("flex"))
+        .expect("gpt-5.4 flex pricing");
+
+    assert_eq!(
+        breakdown.cache_read_price_per_million().amount().scaled(),
+        1_300_000_000
+    );
 }
 
 #[test]
 fn billing_breakdown_should_switch_only_after_the_long_context_threshold() {
-    let boundary = openai_billing_breakdown("gpt-5.4", 272_000, 0, 0, 0, None)
+    let boundary = openai_billing_breakdown("gpt-5.4", billing_usage(272_000, 0, 0, 0), None)
         .expect("short-context boundary");
-    let long =
-        openai_billing_breakdown("gpt-5.4", 272_001, 0, 0, 0, None).expect("long-context pricing");
+    let long = openai_billing_breakdown("gpt-5.4", billing_usage(272_001, 0, 0, 0), None)
+        .expect("long-context pricing");
 
     assert_eq!(
         boundary.input_price_per_million().amount().scaled(),
@@ -85,7 +169,7 @@ fn billing_breakdown_should_switch_only_after_the_long_context_threshold() {
         long.input_price_per_million().amount().scaled(),
         50_000_000_000
     );
-    assert!(openai_billing_breakdown("gpt-5.4", 1, 0, 2, 0, None).is_none());
+    assert!(openai_billing_breakdown("gpt-5.4", billing_usage(1, 0, 2, 0), None).is_none());
 }
 
 #[tokio::test]
@@ -153,7 +237,6 @@ async fn fetch_should_use_wham_usage_headers_only() {
     for (name, expected) in [
         ("authorization", "Bearer oauth-access"),
         ("chatgpt-account-id", "acct_123"),
-        ("originator", "codex_cli_rs"),
         ("accept", "application/json"),
         ("cookie", "session=old"),
     ] {
@@ -181,12 +264,12 @@ async fn fetch_should_use_wham_usage_headers_only() {
             "authorization",
             "chatgpt-account-id",
             "cookie",
-            "originator",
             "user-agent",
         ])
     );
     for forbidden in [
         "content-type",
+        "originator",
         "sec-ch-ua",
         "x-openai-internal-codex-residency",
         "x-client-request-id",
@@ -210,6 +293,30 @@ async fn fetch_should_use_wham_usage_headers_only() {
     }
 }
 
+#[tokio::test]
+async fn usage_not_found_should_preserve_status_without_a_nonofficial_fallback() {
+    let server = MockServer::start().await;
+    Mock::given(method("GET"))
+        .and(path("/api/codex/usage"))
+        .respond_with(ResponseTemplate::new(404).set_body_string("missing"))
+        .expect(1)
+        .mount(&server)
+        .await;
+
+    let error = client(&server.uri())
+        .fetch_usage(context())
+        .await
+        .expect_err("usage 404");
+    let CodexClientError::Upstream { status, .. } = error else {
+        panic!("expected upstream status");
+    };
+    assert_eq!(status, StatusCode::NOT_FOUND);
+
+    let requests = server.received_requests().await.expect("usage requests");
+    assert_eq!(requests.len(), 1);
+    assert_eq!(requests[0].url.path(), "/api/codex/usage");
+}
+
 #[tokio::test]
 async fn content_length_over_limit_success_body_should_be_rejected() {
     let server = MockServer::start().await;

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

@@ -109,7 +109,12 @@ async fn exact_continuation_reuses_owning_socket_after_connection_profile_update
         )
         .await
         .expect("connection-local response");
-    profile.update_desktop_release("2.0.0", "200");
+    profile.update_bundled_release(&CodexBundledReleaseProfile {
+        codex_version: "2.0.0".to_owned(),
+        desktop_version: "2.0.0".to_owned(),
+        desktop_build: "200".to_owned(),
+        verified_at: Utc::now(),
+    });
     let mut continuation = first_request;
     continuation.set_previous_response_id(Some("resp_connection_local".to_owned()));
     continuation.previous_response_scope = Some(PreviousResponseScope::ConnectionLocal);

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

@@ -41,17 +41,16 @@ use gateway_core::routing::{ProviderKind, UpstreamModelId};
 use serde::Deserialize;
 use serde_json::{Map, Number, Value};
 use sha2::{Digest as _, Sha256};
-use url::Url;
 use uuid::Uuid;
 
 use crate::XaiWireProfileState;
 use crate::credential::{
-    AuthorizationCallback, FailureClass, GrokAccountProfile, GrokCredentialAdmin,
-    GrokCredentialCatalogError, GrokCredentialCatalogService, GrokCredentialQuotaService,
-    GrokCredentialRefreshError, GrokCredentialRefreshService, GrokCredentialRepository,
-    GrokCredentialRepositoryError, GrokOAuthClient, GrokOAuthConfig, GrokOAuthImportCandidate,
-    GrokOAuthImportDocument, GrokOAuthImportMetadata, GrokOAuthImportTokens, GrokOAuthSecret,
-    GrokQuotaError, GrokQuotaPeriodKind, GrokQuotaSnapshot, OAuthError, PendingAuthorization,
+    FailureClass, GrokAccountProfile, GrokCredentialAdmin, GrokCredentialCatalogError,
+    GrokCredentialCatalogService, GrokCredentialQuotaService, GrokCredentialRefreshError,
+    GrokCredentialRefreshService, GrokCredentialRepository, GrokCredentialRepositoryError,
+    GrokOAuthClient, GrokOAuthConfig, GrokOAuthImportCandidate, GrokOAuthImportDocument,
+    GrokOAuthImportMetadata, GrokOAuthImportTokens, GrokOAuthSecret, GrokQuotaError,
+    GrokQuotaPeriodKind, GrokQuotaSnapshot, OAuthError, PendingAuthorization,
     PreparedGrokCredentialRotation, PreparedGrokCredentialRotationGuard, RedirectUriAllowlist,
     RotateManagedGrokCredential, SecretValue, VerifiedGrokAccount, VerifiedTokenSet,
 };
@@ -771,9 +770,8 @@ impl XaiAdminProvider {
             &SecretValue::new(stored.server_state),
         )
         .map_err(map_oauth_error)?;
-        let callback = callback(callback_url)?;
         let grant = authorization
-            .accept_callback(callback)
+            .accept_authorization_input(callback_url)
             .map_err(map_oauth_error)?;
         let discovery = self.oauth.discover().await.map_err(map_oauth_error)?;
         let tokens = self
@@ -1343,23 +1341,6 @@ fn decode_mutation(
     ))
 }
 
-fn callback(value: &str) -> Result<AuthorizationCallback, ProviderAdminError> {
-    let mut callback =
-        Url::parse(value).map_err(|_| provider_error(ProviderAdminErrorKind::Invalid))?;
-    if callback.fragment().is_some() {
-        return Err(provider_error(ProviderAdminErrorKind::Invalid));
-    }
-    let query = callback.query().unwrap_or_default().to_owned();
-    callback.set_query(None);
-    let expected = Url::parse(crate::OFFICIAL_REDIRECT_URI)
-        .map_err(|_| provider_error(ProviderAdminErrorKind::Internal))?;
-    if callback != expected {
-        return Err(provider_error(ProviderAdminErrorKind::Invalid));
-    }
-    AuthorizationCallback::parse(&query)
-        .map_err(|_| provider_error(ProviderAdminErrorKind::Invalid))
-}
-
 fn random_flow_id() -> Result<String, ProviderAdminError> {
     let mut random = [0_u8; 32];
     getrandom::fill(&mut random)

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

@@ -197,6 +197,51 @@ impl PendingAuthorization {
             return Err(rejection.into());
         }
         let code = callback.code.ok_or(CallbackRejection::MissingCode)?;
+        self.accept_code(code)
+    }
+
+    /// 消费一次性流程，并接受管理员粘贴的完整回调 URL、回调 query 或裸授权码。
+    ///
+    /// 完整 URL 仍须精确匹配本流程的 redirect URI；query 输入仍须回传正确 state。
+    /// 裸授权码依赖已经绑定 owner 的一次性 flow 与 PKCE verifier，不降低 token 交换约束。
+    ///
+    /// # Errors
+    ///
+    /// URL、state 或授权码不满足当前一次性流程约束时返回回调拒绝错误。
+    pub fn accept_authorization_input(
+        self,
+        input: &str,
+    ) -> Result<AuthorizationCodeGrant, OAuthError> {
+        let input = input.trim();
+        if let Ok(mut callback_url) = Url::parse(input)
+            && matches!(callback_url.scheme(), "http" | "https")
+            && callback_url.has_host()
+        {
+            if callback_url.fragment().is_some() {
+                return Err(CallbackRejection::ProviderRejected.into());
+            }
+            let query = callback_url.query().unwrap_or_default().to_owned();
+            callback_url.set_query(None);
+            if &callback_url != self.redirect_uri.as_url() {
+                return Err(CallbackRejection::ProviderRejected.into());
+            }
+            let callback = AuthorizationCallback::parse(&query)?;
+            return self.accept_callback(callback);
+        }
+
+        if looks_like_callback_url(input) {
+            return Err(CallbackRejection::ProviderRejected.into());
+        }
+
+        if looks_like_callback_query(input) {
+            let callback = AuthorizationCallback::parse(input)?;
+            return self.accept_callback(callback);
+        }
+
+        self.accept_code(SecretValue::new(input.to_owned()))
+    }
+
+    fn accept_code(self, code: SecretValue) -> Result<AuthorizationCodeGrant, OAuthError> {
         if code.is_empty()
             || code.len() > MAX_CALLBACK_VALUE_BYTES
             || code.expose().chars().any(char::is_control)
@@ -213,6 +258,15 @@ impl PendingAuthorization {
     }
 }
 
+fn looks_like_callback_url(value: &str) -> bool {
+    let value = value.to_ascii_lowercase();
+    value.starts_with("http:") || value.starts_with("https:") || value.starts_with("//")
+}
+
+fn looks_like_callback_query(value: &str) -> bool {
+    value.starts_with('?') || value.starts_with("code=")
+}
+
 #[derive(Serialize)]
 struct PendingAuthorizationWire<'a> {
     schema_version: u32,

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

@@ -26,12 +26,17 @@ use crate::{
     SelectedGrokSession,
 };
 
-const DEFAULT_RATE_LIMIT_COOLDOWN: Duration = Duration::from_secs(60);
+const DEFAULT_RATE_LIMIT_COOLDOWN: Duration = Duration::from_secs(2 * 60);
 const MAX_RATE_LIMIT_COOLDOWN: Duration = Duration::from_secs(24 * 60 * 60);
 const STREAM_INTERRUPTION_COOLDOWN: Duration = Duration::from_secs(30);
-const UNAUTHORIZED_COOLDOWN: Duration = Duration::from_secs(60);
+const UNAUTHORIZED_COOLDOWN: Duration = Duration::from_secs(10 * 60);
+const ACCESS_DENIED_COOLDOWN: Duration = Duration::from_secs(30 * 60);
+const PAYMENT_REQUIRED_COOLDOWN: Duration = Duration::from_secs(30 * 60);
+const MODEL_CAPACITY_COOLDOWN: Duration = Duration::from_secs(3 * 60);
+const EMPTY_UPSTREAM_COOLDOWN: Duration = Duration::from_secs(4 * 60);
+const UPSTREAM_UNAVAILABLE_COOLDOWN: Duration = Duration::from_secs(2 * 60);
 const MODEL_ACCESS_DENIED_COOLDOWN: Duration = Duration::from_secs(5 * 60);
-const MODEL_QUOTA_COOLDOWN: Duration = Duration::from_secs(24 * 60 * 60);
+const MODEL_QUOTA_COOLDOWN: Duration = Duration::from_secs(10 * 60);
 const MAX_MODEL_COOLDOWN: Duration = Duration::from_secs(24 * 60 * 60);
 
 #[derive(Debug, Clone, Copy, PartialEq, Eq)]
@@ -479,7 +484,7 @@ impl GrokSessionSelector for GrokAccountSessionSelector {
                 // 不持久化 QuotaExhausted（结构化 QuotaExhausted/FreeQuotaExhausted 走上面）。
                 GrokCredentialFailure::PaymentRequired { retry_after } => {
                     let retry_after = retry_after
-                        .unwrap_or(DEFAULT_RATE_LIMIT_COOLDOWN)
+                        .unwrap_or(PAYMENT_REQUIRED_COOLDOWN)
                         .min(MAX_RATE_LIMIT_COOLDOWN);
                     self.record_runtime_cooldown(session, retry_after).await;
                 }
@@ -507,6 +512,31 @@ impl GrokSessionSelector for GrokAccountSessionSelector {
                     self.record_model_runtime_cooldown(session, upstream_model, retry_after)
                         .await;
                 }
+                GrokCredentialFailure::AccessDenied => {
+                    self.record_runtime_cooldown(session, ACCESS_DENIED_COOLDOWN)
+                        .await;
+                }
+                GrokCredentialFailure::ModelCapacity { upstream_model } => {
+                    if upstream_model.as_str().contains("multi-agent") {
+                        self.record_model_runtime_cooldown(
+                            session,
+                            upstream_model,
+                            MODEL_CAPACITY_COOLDOWN,
+                        )
+                        .await;
+                    } else {
+                        self.record_runtime_cooldown(session, MODEL_CAPACITY_COOLDOWN)
+                            .await;
+                    }
+                }
+                GrokCredentialFailure::EmptyUpstream => {
+                    self.record_runtime_cooldown(session, EMPTY_UPSTREAM_COOLDOWN)
+                        .await;
+                }
+                GrokCredentialFailure::UpstreamUnavailable => {
+                    self.record_runtime_cooldown(session, UPSTREAM_UNAVAILABLE_COOLDOWN)
+                        .await;
+                }
                 GrokCredentialFailure::RateLimited { retry_after } => {
                     let retry_after = retry_after
                         .unwrap_or(DEFAULT_RATE_LIMIT_COOLDOWN)

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

@@ -75,7 +75,8 @@ const RESPONSE_NOT_FOUND_CODE: &str = "not_found";
 
 /// 官方 Grok Build Provider；会话选择与 HTTP SSE transport 均由外部注入。
 ///
-/// 每次调用只选择一个 OAuth 会话，只准备一次可见的上游 POST。重试、凭据轮换、
+/// 每次调用只选择一个 OAuth 会话。仅当 xAI 明确拒绝历史 reasoning 密文时，
+/// 允许在同账号、同凭据、同会话绑定上剥离密文后有界重试一次。凭据轮换、
 /// endpoint fallback 以及公开 xAI API key 推理都不在该 adapter 内。
 pub struct GrokBuildProvider {
     selector: Arc<dyn GrokSessionSelector>,
@@ -202,6 +203,13 @@ impl GrokBuildProvider {
             context.client_api_key_ref(),
         )
         .map_err(map_request_error)?;
+        let wire_upstream_model = UpstreamModelId::new(
+            upstream_request
+                .upstream_model()
+                .ok_or_else(protocol_not_sent)?
+                .to_owned(),
+        )
+        .map_err(|_| protocol_not_sent())?;
         let request_input = upstream_request.input_items();
         if let Some(previous) = previous_session.as_ref() {
             upstream_request.inherit_session(previous.session_id.as_deref());
@@ -209,6 +217,7 @@ impl GrokBuildProvider {
         let selected = select_grok_session(
             self.selector.as_ref(),
             candidate,
+            &wire_upstream_model,
             &context,
             continuation_account,
             upstream_request.affinity().cloned(),
@@ -237,7 +246,7 @@ impl GrokBuildProvider {
             .or(inherited_replay_session_id)
             .and_then(|session_id| {
                 self.reasoning_replay.key(
-                    candidate.upstream_model().as_str(),
+                    wire_upstream_model.as_str(),
                     session_id,
                     selected.account_id().as_str(),
                 )
@@ -278,7 +287,7 @@ impl GrokBuildProvider {
                 credential_recovery: Arc::clone(&self.credential_recovery),
                 responses_url: self.responses_url.clone(),
                 request: upstream_request,
-                upstream_model: candidate.upstream_model().clone(),
+                upstream_model: wire_upstream_model,
                 context,
                 session: Arc::clone(&selected),
                 output_started_at,
@@ -323,6 +332,13 @@ impl GrokBuildProvider {
             context.client_api_key_ref(),
         )
         .map_err(map_request_error)?;
+        let wire_upstream_model = UpstreamModelId::new(
+            upstream_request
+                .upstream_model()
+                .ok_or_else(protocol_not_sent)?
+                .to_owned(),
+        )
+        .map_err(|_| protocol_not_sent())?;
         let explicit_replay_session_id = upstream_request
             .reasoning_replay_session_id()
             .map(str::to_owned);
@@ -333,6 +349,7 @@ impl GrokBuildProvider {
             select_grok_session(
                 self.selector.as_ref(),
                 candidate,
+                &wire_upstream_model,
                 &context,
                 operation_account,
                 upstream_request.affinity().cloned(),
@@ -344,7 +361,7 @@ impl GrokBuildProvider {
             .or(inherited_session_id.as_deref())
             .and_then(|session_id| {
                 self.reasoning_replay.key(
-                    candidate.upstream_model().as_str(),
+                    wire_upstream_model.as_str(),
                     session_id,
                     selected.account_id().as_str(),
                 )
@@ -360,7 +377,7 @@ impl GrokBuildProvider {
                 credential_recovery: Arc::clone(&self.credential_recovery),
                 responses_url: self.responses_url.clone(),
                 request: upstream_request,
-                upstream_model: candidate.upstream_model().clone(),
+                upstream_model: wire_upstream_model,
                 upstream_session_id,
                 context,
                 session: Arc::clone(&selected),
@@ -380,13 +397,14 @@ impl GrokBuildProvider {
 async fn select_grok_session(
     selector: &dyn GrokSessionSelector,
     candidate: &ProviderCandidate,
+    wire_upstream_model: &UpstreamModelId,
     context: &AttemptContext,
     operation_account: Option<ProviderAccountId>,
     affinity: Option<GrokSessionAffinityKey>,
 ) -> Result<SelectedGrokSession, ProviderError> {
     let required_account = context.required_account().cloned().or(operation_account);
     let selection = GrokSessionSelection::new(
-        candidate.upstream_model().clone(),
+        wire_upstream_model.clone(),
         context.excluded_accounts().clone(),
         required_account.clone(),
         context.account_selection_policy(),
@@ -941,81 +959,6 @@ struct AcceptedGrokInference {
     observation: ProviderResponseObservation,
 }
 
-struct GrokInferenceStartFailure {
-    observation: Option<ProviderResponseObservation>,
-    error: ProviderError,
-}
-
-async fn start_grok_inference(
-    selector: &dyn GrokSessionSelector,
-    transport: &dyn GrokInferenceTransport,
-    credential_recovery: &dyn GrokCredentialRecovery,
-    request: GrokInferenceRequest,
-    upstream_model: &UpstreamModelId,
-    context: &AttemptContext,
-    session: &SelectedGrokSession,
-) -> Result<AcceptedGrokInference, GrokInferenceStartFailure> {
-    if context.cancellation().is_cancelled() {
-        return Err(GrokInferenceStartFailure {
-            observation: None,
-            error: provider_error(ProviderErrorKind::Cancelled, UpstreamSendState::NotSent),
-        });
-    }
-    let Some(handshake_deadline) = remaining(context.deadline()) else {
-        return Err(GrokInferenceStartFailure {
-            observation: None,
-            error: provider_error(ProviderErrorKind::Timeout, UpstreamSendState::NotSent),
-        });
-    };
-    let cancellation = context.cancellation().clone();
-    let boundary = tokio::select! {
-        biased;
-        _ = cancellation.cancelled() => InferenceBoundary::Cancelled,
-        _ = tokio::time::sleep(handshake_deadline) => InferenceBoundary::Deadline,
-        response = transport.execute(request) => InferenceBoundary::Response(response),
-    };
-    let response = match boundary {
-        InferenceBoundary::Cancelled => {
-            return Err(GrokInferenceStartFailure {
-                observation: None,
-                error: provider_error(ProviderErrorKind::Cancelled, UpstreamSendState::Ambiguous),
-            });
-        }
-        InferenceBoundary::Deadline => {
-            return Err(GrokInferenceStartFailure {
-                observation: None,
-                error: provider_error(ProviderErrorKind::Timeout, UpstreamSendState::Ambiguous),
-            });
-        }
-        InferenceBoundary::Response(Ok(response)) => response,
-        InferenceBoundary::Response(Err(error)) => {
-            let observation = xai_error_observation(&error).ok();
-            let credential_failure = transport_credential_failure(&error, upstream_model);
-            let error =
-                map_continuation_failure(context, map_transport_error_for_context(error, context));
-            let error = recover_or_record_failure(
-                selector,
-                credential_recovery,
-                session,
-                error,
-                credential_failure,
-                context.credential_recovery_attempted(),
-            )
-            .await;
-            return Err(GrokInferenceStartFailure { observation, error });
-        }
-    };
-    let observation =
-        xai_response_observation(&response).map_err(|error| GrokInferenceStartFailure {
-            observation: None,
-            error,
-        })?;
-    Ok(AcceptedGrokInference {
-        response,
-        observation,
-    })
-}
-
 async fn next_grok_chunk(
     body: &mut GrokInferenceChunkStream,
     selector: &dyn GrokSessionSelector,
@@ -1063,7 +1006,7 @@ fn cold_compaction_http_sse_stream(
         wire_profile,
         credential_recovery,
         responses_url,
-        request,
+        mut request,
         upstream_model,
         upstream_session_id,
         context,
@@ -1081,32 +1024,90 @@ fn cold_compaction_http_sse_stream(
             None,
             &upstream_model,
         );
-        let body = request.to_json_bytes().map_err(map_request_error)?;
-        let inference_request = GrokInferenceRequest::new(
-            responses_url,
-            headers,
-            body,
-            session.binding().clone(),
-        );
-        let accepted = match start_grok_inference(
-            selector.as_ref(),
-            transport.as_ref(),
-            credential_recovery.as_ref(),
-            inference_request,
-            &upstream_model,
-            &context,
-            &session,
-        )
-        .await
-        {
-            Ok(accepted) => accepted,
-            Err(failure) => {
-                if let Some(observation) = failure.observation {
-                    yield ProviderEvent::observation(observation);
-                }
-                Err(mark_transient_compaction_failure(failure.error))?;
+        let mut invalid_encrypted_content_retried = false;
+        let accepted = loop {
+            if context.cancellation().is_cancelled() {
+                Err(provider_error(
+                    ProviderErrorKind::Cancelled,
+                    UpstreamSendState::NotSent,
+                ))?;
                 return;
             }
+            let body = request.to_json_bytes().map_err(map_request_error)?;
+            let inference_request = GrokInferenceRequest::new(
+                responses_url.clone(),
+                headers.clone(),
+                body,
+                session.binding().clone(),
+            );
+            let Some(handshake_deadline) = remaining(context.deadline()) else {
+                Err(mark_transient_compaction_failure(provider_error(
+                    ProviderErrorKind::Timeout,
+                    UpstreamSendState::NotSent,
+                )))?;
+                return;
+            };
+            let cancellation = context.cancellation().clone();
+            let boundary = tokio::select! {
+                biased;
+                _ = cancellation.cancelled() => InferenceBoundary::Cancelled,
+                _ = tokio::time::sleep(handshake_deadline) => InferenceBoundary::Deadline,
+                response = transport.execute(inference_request) => InferenceBoundary::Response(response),
+            };
+            match boundary {
+                InferenceBoundary::Cancelled => {
+                    Err(provider_error(
+                        ProviderErrorKind::Cancelled,
+                        UpstreamSendState::Ambiguous,
+                    ))?;
+                    return;
+                }
+                InferenceBoundary::Deadline => {
+                    Err(mark_transient_compaction_failure(provider_error(
+                        ProviderErrorKind::Timeout,
+                        UpstreamSendState::Ambiguous,
+                    )))?;
+                    return;
+                }
+                InferenceBoundary::Response(Ok(response)) => {
+                    let observation = xai_response_observation(&response)
+                        .map_err(mark_transient_compaction_failure)?;
+                    break AcceptedGrokInference {
+                        response,
+                        observation,
+                    };
+                }
+                InferenceBoundary::Response(Err(error)) => {
+                    if !invalid_encrypted_content_retried
+                        && is_invalid_encrypted_content_failure(&error)
+                        && request.strip_invalid_encrypted_reasoning()
+                    {
+                        invalid_encrypted_content_retried = true;
+                        continue;
+                    }
+                    let observation = xai_error_observation(&error).ok();
+                    let credential_failure =
+                        transport_credential_failure(&error, &upstream_model);
+                    let error = map_continuation_failure(
+                        &context,
+                        map_transport_error_for_context(error, &context),
+                    );
+                    let error = recover_or_record_failure(
+                        selector.as_ref(),
+                        credential_recovery.as_ref(),
+                        &session,
+                        error,
+                        credential_failure,
+                        context.credential_recovery_attempted(),
+                    )
+                    .await;
+                    if let Some(observation) = observation {
+                        yield ProviderEvent::observation(observation);
+                    }
+                    Err(mark_transient_compaction_failure(error))?;
+                    return;
+                }
+            }
         };
         yield ProviderEvent::observation(accepted.observation);
 
@@ -1174,22 +1175,27 @@ fn cold_compaction_http_sse_stream(
             }
         }
 
-        let summary = summary.finish().map_err(map_compaction_decode_error)?;
-        let started = facts.started.ok_or_else(protocol_sent)?;
-        let upstream_completed = facts.completed.is_some();
-        let completed = facts.completed.unwrap_or_else(|| started.clone());
+        let started = facts
+            .started
+            .ok_or_else(|| mark_transient_compaction_failure(protocol_sent()))?;
+        let completed = facts
+            .completed
+            .ok_or_else(|| mark_transient_compaction_failure(protocol_sent()))?;
+        let (summary, encrypted_content) = summary
+            .finish_with_encrypted_content()
+            .map_err(map_compaction_decode_error)?;
         let (created, output_done, terminal) = crate::transport::compaction::compaction_wire_events(
             &started,
             &completed,
-            &summary,
+            summary.as_deref(),
+            &encrypted_content,
             facts.created_response.as_ref(),
             facts.terminal_response.as_ref(),
-            facts.terminal_was_incomplete,
         )
-        .map_err(|_| protocol_sent())?
+        .map_err(|_| mark_transient_compaction_failure(protocol_sent()))?
         .into_parts();
         ensure_sent_context(&context)?;
-        if upstream_completed && session.allows_account_state_mutation() {
+        if session.allows_account_state_mutation() {
             selector.record_success(&session).await;
         }
         if let Some(key) = reasoning_replay_key.as_ref() {
@@ -1210,7 +1216,6 @@ struct CompactionFacts {
     accounting: Vec<GatewayEvent>,
     created_response: Option<Value>,
     terminal_response: Option<Value>,
-    terminal_was_incomplete: bool,
 }
 
 impl CompactionFacts {
@@ -1249,7 +1254,6 @@ impl CompactionFacts {
             Some("response.completed" | "response.incomplete")
                 if self.terminal_response.is_none() =>
             {
-                self.terminal_was_incomplete = event_type == Some("response.incomplete");
                 self.terminal_response = response;
             }
             _ => {}
@@ -1259,7 +1263,9 @@ impl CompactionFacts {
 
 fn map_compaction_decode_error(error: GrokCompactionDecodeError) -> ProviderError {
     match error {
-        GrokCompactionDecodeError::Degenerate => mark_transient_compaction_failure(protocol_sent()),
+        GrokCompactionDecodeError::MissingEncryptedContent => {
+            mark_transient_compaction_failure(protocol_sent())
+        }
     }
 }
 
@@ -1288,7 +1294,7 @@ fn cold_http_sse_stream(
         wire_profile,
         credential_recovery,
         responses_url,
-        request,
+        mut request,
         upstream_model,
         context,
         session,
@@ -1312,54 +1318,69 @@ fn cold_http_sse_stream(
             None,
             &upstream_model,
         );
-        let body = request.to_json_bytes().map_err(map_request_error)?;
-        let inference_request = GrokInferenceRequest::new(
-            responses_url,
-            headers,
-            body,
-            session.binding().clone(),
-        );
-        let Some(handshake_deadline) = remaining(context.deadline()) else {
-            Err(provider_error(
-                ProviderErrorKind::Timeout,
-                UpstreamSendState::NotSent,
-            ))?;
-            return;
-        };
         let cancellation = context.cancellation().clone();
-        let boundary = tokio::select! {
-            biased;
-            _ = cancellation.cancelled() => InferenceBoundary::Cancelled,
-            _ = tokio::time::sleep(handshake_deadline) => InferenceBoundary::Deadline,
-            response = transport.execute(inference_request) => InferenceBoundary::Response(response),
-        };
-        let response = match boundary {
-            InferenceBoundary::Cancelled => {
-                Err(provider_error(ProviderErrorKind::Cancelled, UpstreamSendState::Ambiguous))?;
-                return;
-            }
-            InferenceBoundary::Deadline => {
-                Err(provider_error(ProviderErrorKind::Timeout, UpstreamSendState::Ambiguous))?;
-                return;
-            }
-            InferenceBoundary::Response(Ok(response)) => response,
-            InferenceBoundary::Response(Err(error)) => {
-                let observation = xai_error_observation(&error)?;
-                let credential_failure = transport_credential_failure(&error, &upstream_model);
-                let error =
-                    map_continuation_failure(&context, map_transport_error_for_context(error, &context));
-                let error = recover_or_record_failure(
-                    selector.as_ref(),
-                    credential_recovery.as_ref(),
-                    &session,
-                    error,
-                    credential_failure,
-                    context.credential_recovery_attempted(),
-                )
-                .await;
-                yield ProviderEvent::observation(observation);
-                Err(error)?;
+        let mut invalid_encrypted_content_retried = false;
+        let response = loop {
+            let body = request.to_json_bytes().map_err(map_request_error)?;
+            let inference_request = GrokInferenceRequest::new(
+                responses_url.clone(),
+                headers.clone(),
+                body,
+                session.binding().clone(),
+            );
+            let Some(handshake_deadline) = remaining(context.deadline()) else {
+                Err(provider_error(
+                    ProviderErrorKind::Timeout,
+                    UpstreamSendState::NotSent,
+                ))?;
                 return;
+            };
+            let boundary = tokio::select! {
+                biased;
+                _ = cancellation.cancelled() => InferenceBoundary::Cancelled,
+                _ = tokio::time::sleep(handshake_deadline) => InferenceBoundary::Deadline,
+                response = transport.execute(inference_request) => InferenceBoundary::Response(response),
+            };
+            match boundary {
+                InferenceBoundary::Cancelled => {
+                    Err(provider_error(ProviderErrorKind::Cancelled, UpstreamSendState::Ambiguous))?;
+                    return;
+                }
+                InferenceBoundary::Deadline => {
+                    Err(provider_error(ProviderErrorKind::Timeout, UpstreamSendState::Ambiguous))?;
+                    return;
+                }
+                InferenceBoundary::Response(Ok(response)) => break response,
+                InferenceBoundary::Response(Err(error)) => {
+                    if !invalid_encrypted_content_retried
+                        && is_invalid_encrypted_content_failure(&error)
+                        && request.strip_invalid_encrypted_reasoning()
+                    {
+                        invalid_encrypted_content_retried = true;
+                        if let Some(capture) = session_capture.as_mut() {
+                            capture.request_input = request.input_items();
+                        }
+                        continue;
+                    }
+                    let observation = xai_error_observation(&error)?;
+                    let credential_failure = transport_credential_failure(&error, &upstream_model);
+                    let error = map_continuation_failure(
+                        &context,
+                        map_transport_error_for_context(error, &context),
+                    );
+                    let error = recover_or_record_failure(
+                        selector.as_ref(),
+                        credential_recovery.as_ref(),
+                        &session,
+                        error,
+                        credential_failure,
+                        context.credential_recovery_attempted(),
+                    )
+                    .await;
+                    yield ProviderEvent::observation(observation);
+                    Err(error)?;
+                    return;
+                }
             }
         };
 
@@ -1498,6 +1519,17 @@ fn cold_http_sse_stream(
     })
 }
 
+fn is_invalid_encrypted_content_failure(error: &GrokInferenceTransportError) -> bool {
+    error.kind() == GrokInferenceTransportErrorKind::InvalidRequest
+        && error.status() == Some(400)
+        && error.upstream_code().is_some_and(|code| {
+            matches!(
+                code.as_str(),
+                REASONING_DECODE_FAILED_CODE | "invalid_encrypted_content"
+            )
+        })
+}
+
 enum InferenceBoundary {
     Response(Result<crate::transport::GrokInferenceResponse, GrokInferenceTransportError>),
     Cancelled,
@@ -1689,6 +1721,14 @@ fn transport_credential_failure(
 ) -> Option<GrokCredentialFailure> {
     match error.kind() {
         GrokInferenceTransportErrorKind::Unauthorized => Some(GrokCredentialFailure::Unauthorized),
+        GrokInferenceTransportErrorKind::PermissionDenied => {
+            Some(GrokCredentialFailure::AccessDenied)
+        }
+        GrokInferenceTransportErrorKind::RateLimited if model_capacity_failure(error) => {
+            Some(GrokCredentialFailure::ModelCapacity {
+                upstream_model: upstream_model.clone(),
+            })
+        }
         GrokInferenceTransportErrorKind::RateLimited => Some(GrokCredentialFailure::RateLimited {
             retry_after: error.retry_after(),
         }),
@@ -1715,10 +1755,58 @@ fn transport_credential_failure(
                 retry_after: error.retry_after(),
             })
         }
+        GrokInferenceTransportErrorKind::Unavailable if empty_upstream_failure(error) => {
+            Some(GrokCredentialFailure::EmptyUpstream)
+        }
+        GrokInferenceTransportErrorKind::Unavailable => {
+            Some(GrokCredentialFailure::UpstreamUnavailable)
+        }
         _ => None,
     }
 }
 
+fn model_capacity_failure(error: &GrokInferenceTransportError) -> bool {
+    transport_error_contains_any(
+        error,
+        &[
+            "model_capacity",
+            "capacity",
+            "overloaded",
+            "server_busy",
+            "too many concurrent",
+            "engine_overloaded",
+        ],
+    )
+}
+
+fn empty_upstream_failure(error: &GrokInferenceTransportError) -> bool {
+    transport_error_contains_any(
+        error,
+        &[
+            "empty_upstream",
+            "empty model output",
+            "no content/tool_calls",
+            "no client-visible content",
+            "empty upstream",
+        ],
+    )
+}
+
+fn transport_error_contains_any(error: &GrokInferenceTransportError, signals: &[&str]) -> bool {
+    let code = error
+        .upstream_code()
+        .map(|code| code.as_str())
+        .unwrap_or_default();
+    let message = error
+        .client_visible_upstream_error()
+        .map(ClientVisibleUpstreamError::message)
+        .unwrap_or_default()
+        .to_ascii_lowercase();
+    signals
+        .iter()
+        .any(|signal| code.contains(signal) || message.contains(signal))
+}
+
 async fn map_and_record_stream_transport_failure(
     selector: &dyn GrokSessionSelector,
     session: &SelectedGrokSession,
@@ -1916,7 +2004,7 @@ fn map_transport_error_with_state(
         mapped = mapped.with_status(status);
         if allow_explicit_replay
             && forced_send_state.is_none()
-            && explicit_rejection_is_replay_safe(transport_kind, status)
+            && explicit_rejection_is_replay_safe(&error, status)
         {
             mapped = mapped.with_replay_safe();
         }
@@ -1942,26 +2030,24 @@ fn map_transport_error_with_state(
     mapped
 }
 
-fn explicit_rejection_is_replay_safe(kind: GrokInferenceTransportErrorKind, status: u16) -> bool {
-    matches!(
-        (kind, status),
-        (GrokInferenceTransportErrorKind::Unauthorized, 401)
-            | (
-                GrokInferenceTransportErrorKind::QuotaExhausted,
-                402 | 403 | 429
-            )
-            | (
-                GrokInferenceTransportErrorKind::FreeQuotaExhausted,
-                402 | 403 | 429
-            )
-            | (
-                GrokInferenceTransportErrorKind::ModelQuotaExhausted,
-                402 | 403 | 429
-            )
-            | (GrokInferenceTransportErrorKind::PaymentRequired, 402)
-            | (GrokInferenceTransportErrorKind::ModelAccessDenied, 403)
-            | (GrokInferenceTransportErrorKind::RateLimited, 429)
-    )
+fn explicit_rejection_is_replay_safe(error: &GrokInferenceTransportError, status: u16) -> bool {
+    let kind = error.kind();
+    if kind == GrokInferenceTransportErrorKind::SafetyRejected {
+        return false;
+    }
+    if matches!(status, 401 | 402 | 403 | 405 | 429 | 529) || (500..=599).contains(&status) {
+        return true;
+    }
+    status == 400
+        && (matches!(
+            kind,
+            GrokInferenceTransportErrorKind::QuotaExhausted
+                | GrokInferenceTransportErrorKind::FreeQuotaExhausted
+                | GrokInferenceTransportErrorKind::ModelQuotaExhausted
+                | GrokInferenceTransportErrorKind::PaymentRequired
+                | GrokInferenceTransportErrorKind::Unavailable
+        ) || (kind == GrokInferenceTransportErrorKind::RateLimited
+            && model_capacity_failure(error)))
 }
 
 fn provider_error(kind: ProviderErrorKind, send_state: UpstreamSendState) -> ProviderError {
@@ -1981,9 +2067,9 @@ const WORKER_LEASE_TTL: Duration = Duration::from_secs(15 * 60);
 const WORKER_LEASE_RENEWAL: Duration = Duration::from_secs(5 * 60);
 const OAUTH_REFRESH_INTERVAL: Duration = Duration::from_secs(30);
 const QUOTA_CATALOG_INTERVAL: Duration = Duration::from_secs(5 * 60);
-// grok2api 对 Free/未知额度耗尽采用 24 小时恢复探测；当前公共账号状态没有
-// 独立的 paid period-end 字段，因此 Build 账号统一使用同一保守下限。
-const EXHAUSTED_QUOTA_FALLBACK_RECHECK_INTERVAL: Duration = Duration::from_secs(24 * 60 * 60);
+// rolling 24h 描述的是上游用量窗口，不代表从本次观测起封禁 24 小时；
+// 缺少可信 reset 时间时按 sub2api 的短周期探测策略恢复检查。
+const EXHAUSTED_QUOTA_FALLBACK_RECHECK_INTERVAL: Duration = Duration::from_secs(10 * 60);
 const EXHAUSTED_QUOTA_REFRESH_RETRY_INTERVAL: Duration = QUOTA_CATALOG_INTERVAL;
 const CLI_RELEASE_WORKER_OWNER: &str = "xai-cli-release";
 

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

@@ -23,6 +23,7 @@ use super::{classify_grok_quota_failure, scrub_account_fingerprints};
 
 const CONTENTS_PER_OUTPUT: u32 = 1_024;
 const LONG_CONTEXT_THRESHOLD: u64 = 200_000;
+const GROK_PING_SSE_COMMENT: &[u8] = b": ping\n\n";
 
 #[derive(Clone, Copy)]
 struct TokenRates {
@@ -198,27 +199,64 @@ impl GrokCanonicalDecoder {
     fn decode(&mut self, events: Vec<SseEvent>) -> Result<Vec<ProviderEvent>, ProviderError> {
         let mut output = Vec::new();
         for event in events {
+            if event
+                .event
+                .as_deref()
+                .is_some_and(|event_type| event_type.trim() == "response.doom_loop_check")
+            {
+                continue;
+            }
             if event.data.trim() == "[DONE]" {
                 if !self.completed {
                     return Err(protocol_error_marker());
                 }
                 continue;
             }
-            let Ok(value) = serde_json::from_str::<Value>(&event.data) else {
-                continue;
-            };
-            let body_type = value.get("type").and_then(Value::as_str);
-            let Some(event_type) = body_type.or(event.event.as_deref()) else {
-                let wire = ProtocolWireEvent::json_with_sse_metadata(
+            let parsed = serde_json::from_str::<Value>(&event.data);
+            if event
+                .event
+                .as_deref()
+                .is_some_and(|event_type| event_type.trim() == "ping")
+                && parsed
+                    .as_ref()
+                    .ok()
+                    .and_then(|value| value.get("type").and_then(Value::as_str))
+                    .is_none_or(|event_type| event_type == "ping")
+            {
+                let wire = ProtocolWireEvent::raw_sse(
                     "openai",
-                    event.event,
-                    value,
-                    event.id,
-                    event.retry,
+                    bytes::Bytes::from_static(GROK_PING_SSE_COMMENT),
                 )
                 .map_err(|_| protocol_error_marker())?;
                 output.push(ProviderEvent::wire(wire));
                 continue;
+            }
+            let Ok(value) = parsed else {
+                continue;
+            };
+            let body_type = value.get("type").and_then(Value::as_str);
+            if body_type == Some("response.doom_loop_check") {
+                continue;
+            }
+            let Some(event_type) = body_type.or(event.event.as_deref()) else {
+                let Ok(transformed) = self.response_transform.rewrite_stream_event("", value)
+                else {
+                    continue;
+                };
+                for (index, transformed) in transformed.into_iter().enumerate() {
+                    let mut value = transformed.into_value();
+                    self.response_transform.resequence_stream_value(&mut value);
+                    let wire = ProtocolWireEvent::json_with_sse_metadata(
+                        "openai",
+                        None,
+                        value,
+                        (index == 0).then(|| event.id.clone()).flatten(),
+                        (index == 0).then_some(event.retry).flatten(),
+                    )
+                    .map_err(|_| protocol_error_marker())?;
+                    output.push(ProviderEvent::wire(wire));
+                }
+                continue;
             };
             let event_type = event_type.to_owned();
             // 无法转换的上游帧（多为畸形帧）直接丢弃，保持已开始的客户端流不中断。
@@ -238,7 +276,8 @@ impl GrokCanonicalDecoder {
                     transformed_type.as_str(),
                     "response.created" | "response.in_progress" | "response.failed" | "error"
                 );
-                let value = transformed.into_value();
+                let mut value = transformed.into_value();
+                self.response_transform.resequence_stream_value(&mut value);
                 let mut canonical = Vec::new();
                 // 终态事件（completed/incomplete）fail-closed：用量/计费校验失败即断流。
                 // 其余内容事件容忍字段校验失败——正常上游变体（空 delta、重复 index、
@@ -611,7 +650,8 @@ impl GrokCanonicalDecoder {
 }
 
 fn client_visible_event(event_type: &str) -> bool {
-    event_type == "error" || event_type.starts_with("response.")
+    event_type == "error"
+        || (event_type.starts_with("response.") && event_type != "response.doom_loop_check")
 }
 
 fn response_object(value: &Value) -> Option<&Value> {

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

@@ -2,16 +2,15 @@ use std::fmt;
 use std::time::{SystemTime, UNIX_EPOCH};
 
 use gateway_core::error::IdentifierError;
-use gateway_core::event::{
-    FinishReason, GatewayEvent, ProtocolWireEvent, ProviderEvent, ResponseMeta,
-};
+use gateway_core::event::{GatewayEvent, ProtocolWireEvent, ProviderEvent, ResponseMeta};
 use gateway_core::operation::GenerateRequest;
 use gateway_core::policy::ClientApiKeyId;
 use serde_json::{Map, Value, json};
+use uuid::Uuid;
 
+use super::request::strip_invalid_encrypted_reasoning_from_body;
 use super::{GrokRequestEncodeError, GrokResponsesRequest, GrokSessionAffinityKey};
 
-const MIN_GROK_COMPACTION_SUMMARY_CHARS: usize = 500;
 // 来自 xAI grok-build 的 full-replace compaction 模板；当前协议没有额外的
 // `/compact <text>` 用户上下文，因此不保留上游模板中的占位符。
 const GROK_COMPACTION_PROMPT: &str = r#"Your task is to produce a faithful, concise summary of the conversation so far so that a successor assistant can continue the work seamlessly after the earlier turns are discarded. The successor will see the user's original query plus this summary. Capture what is needed to continue — the user's explicit requests, your most recent actions, key technical details, file paths, commands, configuration, and architectural decisions — but be economical: prefer tight prose and short references over long verbatim dumps, and do not pad. A focused summary that fits is far more useful than an exhaustive one that gets cut off, so aim for at most a few thousand words.
@@ -68,27 +67,16 @@ impl GrokCompactionRequest {
         input.push(summary_prompt_item());
         body.insert("input".to_owned(), Value::Array(input));
 
-        for field in [
-            "background",
-            "include",
-            "max_output_tokens",
-            "parallel_tool_calls",
-            "previous_response_id",
-            "prompt_cache_key",
-            "response_format",
-            "service_tier",
-            "text",
-            "truncation",
-        ] {
-            body.remove(field);
-        }
-        body.insert("temperature".to_owned(), json!(1.0));
+        // 压缩请求已经携带规范化后的完整历史，不能让账号派生的缓存身份
+        // 把它重新变成依赖原账号状态的原生续接请求。
+        body.remove("prompt_cache_key");
+        body.insert("include".to_owned(), json!(["reasoning.encrypted_content"]));
         if body
             .get("tools")
             .and_then(Value::as_array)
             .is_some_and(|tools| !tools.is_empty())
         {
-            body.insert("tool_choice".to_owned(), Value::String("auto".to_owned()));
+            body.insert("tool_choice".to_owned(), Value::String("none".to_owned()));
         } else {
             body.remove("tool_choice");
         }
@@ -120,6 +108,11 @@ impl GrokCompactionRequest {
         self.reasoning_replay_session_id.as_deref()
     }
 
+    /// 返回经 sub2api 别名表归一化后的 xAI wire 模型。
+    pub(crate) fn upstream_model(&self) -> Option<&str> {
+        self.body.get("model").and_then(Value::as_str)
+    }
+
     /// 序列化上游请求正文。
     ///
     /// # Errors
@@ -128,6 +121,11 @@ impl GrokCompactionRequest {
     pub fn to_json_bytes(&self) -> Result<Vec<u8>, GrokRequestEncodeError> {
         serde_json::to_vec(&self.body).map_err(|_| GrokRequestEncodeError::Serialization)
     }
+
+    /// 为同账号的一次 `invalid_encrypted_content` 恢复请求剥离被拒绝密文。
+    pub(crate) fn strip_invalid_encrypted_reasoning(&mut self) -> bool {
+        strip_invalid_encrypted_reasoning_from_body(&mut self.body)
+    }
 }
 
 /// 返回 xAI 专用全历史压缩是否由客户端请求。
@@ -166,6 +164,8 @@ impl fmt::Debug for GrokCompactionRequest {
 #[derive(Default)]
 pub struct GrokCompactionSummaryDecoder {
     text: String,
+    terminal_text: Option<String>,
+    encrypted_content: Option<String>,
 }
 
 impl GrokCompactionSummaryDecoder {
@@ -174,6 +174,8 @@ impl GrokCompactionSummaryDecoder {
     pub const fn new() -> Self {
         Self {
             text: String::new(),
+            terminal_text: None,
+            encrypted_content: None,
         }
     }
 
@@ -188,20 +190,48 @@ impl GrokCompactionSummaryDecoder {
                 self.text.push_str(&delta.text);
             }
         }
+        if let Some(wire) = event.wire_event().filter(|wire| wire.has_json_data()) {
+            capture_encrypted_reasoning(wire.data(), &mut self.encrypted_content);
+            capture_completed_summary(wire.data(), &mut self.terminal_text);
+        }
         Ok(())
     }
 
-    /// 完成摘要解码并返回清理后的摘要文本。
+    /// 完成摘要解码并返回上游摘要文本。
     ///
     /// # Errors
     ///
-    /// 摘要为空或过短时返回错误。
+    /// 保留此 `Result` 签名以稳定 decoder contract。
     pub fn finish(self) -> Result<String, GrokCompactionDecodeError> {
-        let summary = clean_summary(&self.text);
-        if summary.is_empty() || summary.chars().count() < MIN_GROK_COMPACTION_SUMMARY_CHARS {
-            return Err(GrokCompactionDecodeError::Degenerate);
+        Ok(self.summary_text().unwrap_or_default())
+    }
+
+    /// 完成摘要解码，并返回可见摘要与 xAI 真实 reasoning 密文。
+    pub(crate) fn finish_with_encrypted_content(
+        self,
+    ) -> Result<(Option<String>, String), GrokCompactionDecodeError> {
+        let summary = self.summary_text();
+        let encrypted_content = self
+            .encrypted_content
+            .filter(|value| !value.trim().is_empty())
+            .ok_or(GrokCompactionDecodeError::MissingEncryptedContent)?;
+        Ok((summary, encrypted_content))
+    }
+
+    fn summary_text(&self) -> Option<String> {
+        if let Some(text) = self
+            .terminal_text
+            .as_deref()
+            .map(str::trim)
+            .filter(|text| !text.is_empty())
+        {
+            return Some(text.to_owned());
         }
-        Ok(summary)
+        let text = self.text.trim();
+        if !text.is_empty() {
+            return Some(text.to_owned());
+        }
+        None
     }
 }
 
@@ -210,6 +240,8 @@ impl fmt::Debug for GrokCompactionSummaryDecoder {
         formatter
             .debug_struct("GrokCompactionSummaryDecoder")
             .field("summary_bytes", &self.text.len())
+            .field("has_terminal_summary", &self.terminal_text.is_some())
+            .field("has_encrypted_content", &self.encrypted_content.is_some())
             .finish()
     }
 }
@@ -217,8 +249,8 @@ impl fmt::Debug for GrokCompactionSummaryDecoder {
 /// Grok 摘要响应不满足压缩 contract。
 #[derive(Debug, thiserror::Error, Clone, Copy, PartialEq, Eq)]
 pub enum GrokCompactionDecodeError {
-    #[error("Grok compaction response summary was too short")]
-    Degenerate,
+    #[error("Grok compaction response omitted reasoning.encrypted_content")]
+    MissingEncryptedContent,
 }
 
 pub(crate) struct GrokCompactionWireEvents {
@@ -240,10 +272,10 @@ impl GrokCompactionWireEvents {
 pub(crate) fn compaction_wire_events(
     started: &ResponseMeta,
     completed: &ResponseMeta,
-    summary: &str,
+    summary: Option<&str>,
+    encrypted_content: &str,
     created_source: Option<&Value>,
     terminal_source: Option<&Value>,
-    terminal_was_incomplete: bool,
 ) -> Result<GrokCompactionWireEvents, IdentifierError> {
     let created_response = compaction_response(
         started,
@@ -253,17 +285,14 @@ pub(crate) fn compaction_wire_events(
         Vec::new(),
         Value::Null,
     );
-    let item = compaction_item(summary);
-    let (status, terminal_event, incomplete_details) = terminal_shape(
-        completed.finish_reason(),
-        terminal_was_incomplete,
-        terminal_source,
-    );
+    let item = compaction_item(summary, encrypted_content);
+    // sub2api 的 compact contract 以真实密文为成功依据，不向客户端延续
+    // 摘要生成阶段的 incomplete 状态。
     let terminal_response = compaction_response(
         completed,
         terminal_source.or(created_source),
-        status,
-        incomplete_details,
+        "completed",
+        Value::Null,
         vec![item.clone()],
         terminal_source
             .and_then(|source| source.get("usage"))
@@ -294,9 +323,9 @@ pub(crate) fn compaction_wire_events(
         )?,
         terminal: ProtocolWireEvent::json_with_sse_metadata(
             "openai",
-            Some(terminal_event.to_owned()),
+            Some("response.completed".to_owned()),
             json!({
-                "type": terminal_event,
+                "type": "response.completed",
                 "response": terminal_response,
             }),
             None,
@@ -333,38 +362,96 @@ fn compaction_response(
     response.insert("model".to_owned(), Value::String(meta.model().to_owned()));
     response.insert("output".to_owned(), Value::Array(output));
     response.insert("usage".to_owned(), usage);
+    response.remove("output_text");
     Value::Object(response)
 }
 
-fn compaction_item(summary: &str) -> Value {
-    json!({
-        "type": "compaction",
-        "encrypted_content": summary,
-    })
+fn compaction_item(summary: Option<&str>, encrypted_content: &str) -> Value {
+    let mut item = Map::from_iter([
+        (
+            "id".to_owned(),
+            Value::String(format!("cmp_{}", Uuid::new_v4().simple())),
+        ),
+        ("type".to_owned(), Value::String("compaction".to_owned())),
+        ("status".to_owned(), Value::String("completed".to_owned())),
+        (
+            "encrypted_content".to_owned(),
+            Value::String(encrypted_content.to_owned()),
+        ),
+    ]);
+    if let Some(summary) = summary.map(str::trim).filter(|value| !value.is_empty()) {
+        item.insert(
+            "summary".to_owned(),
+            json!([{
+                "type": "summary_text",
+                "text": summary,
+            }]),
+        );
+    }
+    Value::Object(item)
 }
 
-fn terminal_shape(
-    finish_reason: Option<FinishReason>,
-    terminal_was_incomplete: bool,
-    terminal_source: Option<&Value>,
-) -> (&'static str, &'static str, Value) {
-    let incomplete = terminal_was_incomplete
-        || matches!(
-            finish_reason,
-            Some(FinishReason::Length | FinishReason::ContentFilter)
-        );
-    if !incomplete {
-        return ("completed", "response.completed", Value::Null);
+fn capture_encrypted_reasoning(value: &Value, captured: &mut Option<String>) {
+    if let Some(item) = value.get("item") {
+        capture_encrypted_reasoning_item(item, captured);
+    }
+    for output in [value.get("output"), value.pointer("/response/output")]
+        .into_iter()
+        .flatten()
+        .filter_map(Value::as_array)
+    {
+        for item in output {
+            capture_encrypted_reasoning_item(item, captured);
+        }
+    }
+}
+
+fn capture_encrypted_reasoning_item(item: &Value, captured: &mut Option<String>) {
+    let Some(item) = item.as_object() else {
+        return;
+    };
+    if item.get("type").and_then(Value::as_str) != Some("reasoning") {
+        return;
+    }
+    if let Some(encrypted_content) = item
+        .get("encrypted_content")
+        .and_then(Value::as_str)
+        .map(str::trim)
+        .filter(|value| !value.is_empty())
+    {
+        *captured = Some(encrypted_content.to_owned());
+    }
+}
+
+fn capture_completed_summary(value: &Value, captured: &mut Option<String>) {
+    let mut parts = Vec::new();
+    for output in [value.get("output"), value.pointer("/response/output")]
+        .into_iter()
+        .flatten()
+        .filter_map(Value::as_array)
+    {
+        for item in output {
+            let Some(item) = item.as_object() else {
+                continue;
+            };
+            if item.get("type").and_then(Value::as_str) != Some("message") {
+                continue;
+            }
+            let Some(content) = item.get("content").and_then(Value::as_array) else {
+                continue;
+            };
+            parts.extend(content.iter().filter_map(|part| {
+                part.get("text")
+                    .and_then(Value::as_str)
+                    .map(str::trim)
+                    .filter(|text| !text.is_empty())
+                    .map(str::to_owned)
+            }));
+        }
+    }
+    if !parts.is_empty() {
+        *captured = Some(parts.join("\n"));
     }
-    let details = terminal_source
-        .and_then(|source| source.get("incomplete_details"))
-        .cloned()
-        .unwrap_or_else(|| match finish_reason {
-            Some(FinishReason::Length) => json!({ "reason": "max_output_tokens" }),
-            Some(FinishReason::ContentFilter) => json!({ "reason": "content_filter" }),
-            _ => json!({ "reason": "other" }),
-        });
-    ("incomplete", "response.incomplete", details)
 }
 
 fn unix_seconds() -> u64 {
@@ -384,71 +471,3 @@ fn summary_prompt_item() -> Value {
         }],
     })
 }
-
-fn clean_summary(raw: &str) -> String {
-    let mut summary = raw.to_owned();
-    while let Some(start) = summary.find("<analysis>") {
-        let leading = match summary.find("<summary>") {
-            Some(summary_start) => {
-                start < summary_start
-                    || summary[summary_start + "<summary>".len()..start]
-                        .trim()
-                        .is_empty()
-            }
-            None => summary[..start].trim().is_empty(),
-        };
-        if !leading {
-            break;
-        }
-        if let Some(relative_end) = summary[start..].find("</analysis>") {
-            let end = start + relative_end + "</analysis>".len();
-            summary.replace_range(start..end, "");
-        } else {
-            let drop_end = summary[start..]
-                .find("<summary>")
-                .map_or(summary.len(), |relative| start + relative);
-            summary.replace_range(start..drop_end, "");
-            break;
-        }
-    }
-
-    if let Some(start) = summary.find("<summary>")
-        && let Some(end) = summary.rfind("</summary>")
-        && end > start
-    {
-        let before = summary[..start].to_owned();
-        let after = summary[end + "</summary>".len()..].to_owned();
-        let inner = strip_leading_scratchpad(summary[start + "<summary>".len()..end].trim());
-        summary = format!("{before}Summary:\n{inner}{after}");
-    }
-
-    summary = neutralize_compaction_control_tokens(&summary);
-    while summary.contains("\n\n\n") {
-        summary = summary.replace("\n\n\n", "\n\n");
-    }
-    summary.trim().to_owned()
-}
-
-fn strip_leading_scratchpad(inner: &str) -> String {
-    let mut summary = inner.trim();
-    let lead = summary.trim_start_matches(['#', '*', '-', '>', ' ', '\t']);
-    if !lead.starts_with(|character: char| character.is_ascii_digit())
-        && let Some(end) = summary.rfind("</analysis>")
-    {
-        summary = summary[end + "</analysis>".len()..].trim_start();
-    }
-    if let Some(inner) = summary.strip_prefix("<summary>") {
-        summary = inner.trim_start();
-    }
-    summary.to_owned()
-}
-
-fn neutralize_compaction_control_tokens(summary: &str) -> String {
-    summary
-        .replace("</summary>", "<\u{200b}/summary>")
-        .replace("<summary>", "<\u{200b}summary>")
-        .replace("</analysis>", "<\u{200b}/analysis>")
-        .replace("<analysis>", "<\u{200b}analysis>")
-        .replace("</summary_request>", "<\u{200b}/summary_request>")
-        .replace("<summary_request>", "<\u{200b}summary_request>")
-}

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

@@ -61,8 +61,26 @@ const FREE_USAGE_SIGNALS: &[&str] = &[
     "subscription_free_usage_exhausted",
     "free-usage-exhausted",
     "free_usage_exhausted",
+    "usage-limit-exceeded",
+    "usage_limit_exceeded",
+    "free-tier-limit",
+    "free_tier_limit",
     "used all the included free usage",
     "used all your free usage",
+    "free quota",
+    "no remaining free",
+    "out of free",
+    "usage resets over a rolling",
+    "额度耗尽",
+    "额度用完",
+    "额度不足",
+    "额度已用尽",
+    "免费额度",
+    "免费用量",
+    "用量耗尽",
+    "配额耗尽",
+    "配额不足",
+    "模型额度",
 ];
 const ACCOUNT_QUOTA_SIGNALS: &[&str] = &[
     "personal-team-blocked:spending-limit",

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

@@ -50,7 +50,7 @@ const HTTP2_KEEP_ALIVE_INTERVAL: Duration = Duration::from_secs(15);
 const HTTP2_KEEP_ALIVE_TIMEOUT: Duration = Duration::from_secs(5);
 const MAX_ERROR_BODY_BYTES: usize = 64 * 1024;
 const MAX_INFERENCE_BODY_BYTES: usize = 256 * 1024 * 1024;
-const MAX_RETRY_AFTER_SECONDS: u64 = 120;
+const MAX_RETRY_AFTER_SECONDS: u64 = 24 * 60 * 60;
 const TRUSTED_DOH_HOST: &str = "dns.google";
 const TRUSTED_DOH_URL: &str = "https://dns.google/resolve";
 const MAX_DOH_RESPONSE_BYTES: usize = 64 * 1024;
@@ -992,19 +992,32 @@ async fn classify_inference_status(response: Response) -> GrokInferenceTransport
         Ok(BoundedBody::TooLarge) | Err(_) => Vec::new(),
     };
     let metadata = inference_error_metadata(&body);
+    let body_failure = classify_grok_body_failure(&metadata, &body);
     let kind = match status {
-        StatusCode::BAD_REQUEST
-        | StatusCode::NOT_FOUND
-        | StatusCode::CONFLICT
-        | StatusCode::UNPROCESSABLE_ENTITY => GrokInferenceTransportErrorKind::InvalidRequest,
-        StatusCode::UNAUTHORIZED => GrokInferenceTransportErrorKind::Unauthorized,
-        StatusCode::PAYMENT_REQUIRED => classify_payment_required(&metadata),
-        StatusCode::FORBIDDEN => classify_forbidden(&metadata, &body),
-        StatusCode::REQUEST_TIMEOUT | StatusCode::GATEWAY_TIMEOUT => {
-            GrokInferenceTransportErrorKind::Timeout
+        StatusCode::BAD_REQUEST => match body_failure {
+            Some(GrokBodyFailure::RateLimit) | None => {
+                GrokInferenceTransportErrorKind::InvalidRequest
+            }
+            Some(failure) => failure.transport_kind(),
+        },
+        StatusCode::NOT_FOUND | StatusCode::CONFLICT | StatusCode::UNPROCESSABLE_ENTITY => {
+            GrokInferenceTransportErrorKind::InvalidRequest
         }
-        StatusCode::TOO_MANY_REQUESTS => classify_too_many_requests(&metadata),
-        status if status.is_server_error() => GrokInferenceTransportErrorKind::Unavailable,
+        StatusCode::UNAUTHORIZED => GrokInferenceTransportErrorKind::Unauthorized,
+        StatusCode::PAYMENT_REQUIRED => body_failure
+            .map(GrokBodyFailure::transport_kind)
+            .unwrap_or_else(|| classify_payment_required(&metadata)),
+        StatusCode::FORBIDDEN => classify_forbidden(&metadata, &body, body_failure),
+        StatusCode::REQUEST_TIMEOUT => GrokInferenceTransportErrorKind::Timeout,
+        StatusCode::TOO_MANY_REQUESTS => body_failure
+            .map(GrokBodyFailure::transport_kind)
+            .unwrap_or_else(|| classify_too_many_requests(&metadata)),
+        _ if status_code == 529 => body_failure
+            .map(GrokBodyFailure::transport_kind)
+            .unwrap_or(GrokInferenceTransportErrorKind::Unavailable),
+        status if status.is_server_error() => body_failure
+            .map(GrokBodyFailure::transport_kind)
+            .unwrap_or(GrokInferenceTransportErrorKind::Unavailable),
         _ => GrokInferenceTransportErrorKind::Protocol,
     };
     let credential_recovery_required = kind == GrokInferenceTransportErrorKind::Unauthorized;
@@ -1015,7 +1028,11 @@ async fn classify_inference_status(response: Response) -> GrokInferenceTransport
     let upstream_code = if status == StatusCode::BAD_REQUEST && reasoning_decode_failed(&metadata) {
         Some("reasoning_decode_failed".to_owned())
     } else {
-        metadata.code.as_deref().and_then(normalize_failure_code)
+        metadata
+            .code
+            .as_deref()
+            .and_then(normalize_failure_code)
+            .or_else(|| body_failure.map(GrokBodyFailure::marker).map(str::to_owned))
     };
     if let Some(message) = metadata.client_message.as_deref() {
         error = error.with_client_visible_upstream_error(ClientVisibleUpstreamError::new(
@@ -1045,6 +1062,115 @@ struct InferenceErrorMetadata {
     client_message: Option<String>,
 }
 
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+enum GrokBodyFailure {
+    Quota(GrokQuotaFailureKind),
+    Billing,
+    EmptyUpstream,
+    ModelCapacity,
+    RateLimit,
+}
+
+impl GrokBodyFailure {
+    fn transport_kind(self) -> GrokInferenceTransportErrorKind {
+        match self {
+            Self::Quota(failure) => quota_transport_error_kind(failure),
+            Self::Billing => GrokInferenceTransportErrorKind::PaymentRequired,
+            Self::EmptyUpstream => GrokInferenceTransportErrorKind::Unavailable,
+            Self::ModelCapacity | Self::RateLimit => GrokInferenceTransportErrorKind::RateLimited,
+        }
+    }
+
+    const fn marker(self) -> &'static str {
+        match self {
+            Self::Quota(GrokQuotaFailureKind::Account) => "quota_exhausted",
+            Self::Quota(GrokQuotaFailureKind::FreeAccount) => "free_usage_exhausted",
+            Self::Quota(GrokQuotaFailureKind::FreeModelUsage) => "model_quota_exhausted",
+            Self::Billing => "billing_quota",
+            Self::EmptyUpstream => "empty_upstream",
+            Self::ModelCapacity => "model_capacity",
+            Self::RateLimit => "rate_limit",
+        }
+    }
+}
+
+fn classify_grok_body_failure(
+    metadata: &InferenceErrorMetadata,
+    body: &[u8],
+) -> Option<GrokBodyFailure> {
+    let text = format!(
+        "{} {}",
+        inference_metadata_text(metadata),
+        String::from_utf8_lossy(body).to_ascii_lowercase()
+    );
+    let quota_failure = classify_grok_quota_failure(
+        metadata.code.as_deref(),
+        metadata.error_type.as_deref(),
+        Some(&text),
+    );
+    if let Some(
+        failure @ (GrokQuotaFailureKind::FreeAccount | GrokQuotaFailureKind::FreeModelUsage),
+    ) = quota_failure
+    {
+        return Some(GrokBodyFailure::Quota(failure));
+    }
+    if contains_any(
+        &text,
+        &[
+            "billing quota",
+            "insufficient_quota",
+            "payment required",
+            "payment failed",
+            "spending limit",
+            "run out of credits",
+            "out of credits",
+            "余额不足",
+            "欠费",
+            "需要付费",
+        ],
+    ) {
+        return Some(GrokBodyFailure::Billing);
+    }
+    if let Some(failure) = quota_failure {
+        return Some(GrokBodyFailure::Quota(failure));
+    }
+    if contains_any(
+        &text,
+        &[
+            "empty model output",
+            "no content/tool_calls",
+            "no client-visible content",
+            "empty_upstream",
+            "empty upstream",
+        ],
+    ) {
+        return Some(GrokBodyFailure::EmptyUpstream);
+    }
+    if contains_any(
+        &text,
+        &[
+            "capacity",
+            "overloaded",
+            "server_busy",
+            "too many concurrent",
+            "engine_overloaded",
+        ],
+    ) {
+        return Some(GrokBodyFailure::ModelCapacity);
+    }
+    contains_any(
+        &text,
+        &[
+            "rate limit",
+            "rate_limit",
+            "too many requests",
+            "请求过于频繁",
+            "速率限制",
+        ],
+    )
+    .then_some(GrokBodyFailure::RateLimit)
+}
+
 fn inference_error_metadata(body: &[u8]) -> InferenceErrorMetadata {
     let Ok(value) = serde_json::from_slice::<Value>(body) else {
         let message = String::from_utf8_lossy(body).trim().to_owned();
@@ -1091,13 +1217,18 @@ fn classify_payment_required(metadata: &InferenceErrorMetadata) -> GrokInference
 fn classify_forbidden(
     metadata: &InferenceErrorMetadata,
     body: &[u8],
+    body_failure: Option<GrokBodyFailure>,
 ) -> GrokInferenceTransportErrorKind {
     let text = inference_metadata_text(metadata);
     let body_text = String::from_utf8_lossy(body).to_ascii_lowercase();
-    if safety_rejected(&text) || safety_rejected(&body_text) {
+    if account_access_rejected(&text) || account_access_rejected(&body_text) {
+        GrokInferenceTransportErrorKind::PermissionDenied
+    } else if safety_rejected(&text) || safety_rejected(&body_text) {
         GrokInferenceTransportErrorKind::SafetyRejected
     } else if credential_rejected(&text) {
         GrokInferenceTransportErrorKind::Unauthorized
+    } else if let Some(failure) = body_failure {
+        failure.transport_kind()
     } else if let Some(failure) = quota_failure(metadata) {
         quota_transport_error_kind(failure)
     } else if model_access_denied(metadata) {
@@ -1117,10 +1248,27 @@ fn classify_too_many_requests(
 }
 
 fn reasoning_decode_failed(metadata: &InferenceErrorMetadata) -> bool {
-    metadata.message.as_deref().is_some_and(|message| {
-        message.contains("could not decode the compaction blob")
-            || message.contains("could not decrypt the provided encrypted_content")
-    })
+    let code = metadata.code.as_deref().map(str::trim).unwrap_or_default();
+    let message = metadata
+        .message
+        .as_deref()
+        .map(str::trim)
+        .unwrap_or_default()
+        .to_ascii_lowercase();
+    if message.contains("could not decode the compaction blob") {
+        return true;
+    }
+    if code.eq_ignore_ascii_case("invalid_encrypted_content") {
+        return !message.is_empty();
+    }
+    if !code.is_empty() && !code.eq_ignore_ascii_case("invalid-argument") {
+        return false;
+    }
+    if code.is_empty() && !message.contains("decrypt") {
+        return false;
+    }
+    message.contains("encrypted_content")
+        && (message.contains("decrypt") || message.contains("unmodified"))
 }
 
 fn quota_failure(metadata: &InferenceErrorMetadata) -> Option<GrokQuotaFailureKind> {
@@ -1152,7 +1300,48 @@ fn model_access_denied(metadata: &InferenceErrorMetadata) -> bool {
 fn safety_rejected(text: &str) -> bool {
     contains_any(
         text,
-        &["content violates usage guidelines", "safety_check_type_"],
+        &[
+            "content violates usage guidelines",
+            "safety_check_type_",
+            "content_filter",
+            "content_policy",
+            "content moderation",
+            "cyber_policy",
+            "new_sensitive",
+            "moderation feature is not available",
+            "image is sensitive",
+            "text is sensitive",
+            "prohibited content",
+            "forbidden content",
+            "request blocked by policy",
+            "request rejected by policy",
+            "request violates policy",
+            "prompt violates policy",
+            "input violates policy",
+        ],
+    )
+}
+
+fn account_access_rejected(text: &str) -> bool {
+    contains_any(
+        text,
+        &[
+            "account_suspended",
+            "account_disabled",
+            "user_suspended",
+            "user_disabled",
+            "subscription_required",
+            "entitlement_required",
+            "not_entitled",
+            "plan_required",
+            "permission_denied",
+            "account suspended",
+            "account disabled",
+            "user suspended",
+            "subscription required",
+            "entitlement required",
+            "not entitled",
+        ],
     )
 }
 

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

@@ -347,6 +347,14 @@ pub enum GrokCredentialFailure {
         upstream_model: UpstreamModelId,
         retry_after: Option<Duration>,
     },
+    /// 非内容策略的 403：当前账号的访问资格暂不可用。
+    AccessDenied,
+    /// 上游明确报告模型容量不足；重模型按模型隔离，其余账号短暂冷却。
+    ModelCapacity { upstream_model: UpstreamModelId },
+    /// 上游返回空模型输出。
+    EmptyUpstream,
+    /// 上游 5xx/529 暂时不可用。
+    UpstreamUnavailable,
     /// 成功的 SSE 响应在终止 Responses 事件前结束。
     StreamInterrupted,
 }

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

@@ -27,10 +27,11 @@ use gateway_core::policy::ClientApiKeyId;
 use gateway_core::provider_ports::{
     NewOAuthPendingFlow, OAuthPendingBinding, OAuthPendingClaimOutcome, OAuthPendingConsumeOutcome,
     OAuthPendingFlowPort, OAuthPendingPutOutcome, OAuthPendingReleaseOutcome,
-    ProviderCatalogCacheKey, ProviderCatalogCachePort, ProviderCooldown, ProviderCooldownPort,
-    ProviderCooldownScope, ProviderCredentialState, ProviderCredentialStatePort,
-    ProviderLeaseAcquisition, ProviderLeasePort, ProviderLeaseRequest, ProviderRefreshPolicy,
-    ProviderRuntimePolicyPort, ProviderScopedCooldown, ProviderStoreError, ProviderStorePorts,
+    ProviderArtifactProfile, ProviderArtifactProfileCachePort, ProviderCatalogCacheKey,
+    ProviderCatalogCachePort, ProviderCooldown, ProviderCooldownPort, ProviderCooldownScope,
+    ProviderCredentialState, ProviderCredentialStatePort, ProviderLeaseAcquisition,
+    ProviderLeasePort, ProviderLeaseRequest, ProviderRefreshPolicy, ProviderRuntimePolicyPort,
+    ProviderScopedCooldown, ProviderStoreError, ProviderStorePorts,
 };
 use gateway_core::routing::{ProviderKind, UpstreamModelId};
 use gateway_core::task::{
@@ -448,6 +449,7 @@ fn provider_ports_with_catalog(
         Arc::new(TestSessionAffinity),
         Arc::new(TestSessionExclusions),
         catalog_cache,
+        Arc::new(TestArtifactProfiles),
         Arc::new(TestCredentialState),
         Arc::new(TestCooldown),
         Arc::new(TestRuntimePolicy),
@@ -529,6 +531,25 @@ impl ProviderLeasePort for TestLeases {
     }
 }
 
+struct TestArtifactProfiles;
+
+impl ProviderArtifactProfileCachePort for TestArtifactProfiles {
+    fn replace_if_newer(
+        &self,
+        _profile: ProviderArtifactProfile,
+        _ttl: Duration,
+    ) -> BoxFuture<'_, Result<bool, ProviderStoreError>> {
+        Box::pin(async { Ok(true) })
+    }
+
+    fn read<'a>(
+        &'a self,
+        _provider_kind: &'a ProviderKind,
+    ) -> BoxFuture<'a, Result<Option<ProviderArtifactProfile>, ProviderStoreError>> {
+        Box::pin(async { Ok(None) })
+    }
+}
+
 #[derive(Default)]
 struct TestCatalogCache {
     values: Mutex<BTreeMap<String, OpaqueProviderData>>,

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

@@ -3,6 +3,8 @@ use provider_xai::{
     PendingAuthorization, RedirectUriAllowlist,
 };
 
+const REDIRECT_URI: &str = "https://gateway.example/admin/xai/callback";
+
 #[test]
 fn callback_should_reject_duplicate_state() {
     let result = AuthorizationCallback::parse("code=fake&state=one&state=two");
@@ -28,10 +30,9 @@ fn pending_authorization_should_round_trip_only_through_server_state() {
     let config = GrokOAuthConfig::official().expect("fixture config");
     let discovery = DiscoveryDocument::parse(&config, include_bytes!("fixtures/discovery.json"))
         .expect("fixture discovery");
-    let redirect = "https://gateway.example/admin/xai/callback";
-    let allowed = RedirectUriAllowlist::new([redirect])
+    let allowed = RedirectUriAllowlist::new([REDIRECT_URI])
         .expect("redirect allowlist")
-        .authorize(redirect)
+        .authorize(REDIRECT_URI)
         .expect("allowlisted redirect");
     let pending = PendingAuthorization::start(&config, &discovery, allowed, None)
         .expect("start pending flow");
@@ -61,3 +62,121 @@ fn pending_authorization_should_round_trip_only_through_server_state() {
         .accept_callback(callback)
         .expect("restored state should validate callback");
 }
+
+#[test]
+fn authorization_input_should_accept_callback_query_and_bare_code() {
+    let (pending, state) = pending_authorization();
+    pending
+        .accept_authorization_input(&format!("{REDIRECT_URI}?code=full-url-code&state={state}"))
+        .expect("official callback URL");
+
+    let (pending, state) = pending_authorization();
+    pending
+        .accept_authorization_input(&format!("?code=query-code&state={state}"))
+        .expect("callback query");
+
+    let (pending, state) = pending_authorization();
+    pending
+        .accept_authorization_input(&format!("code=query-code&state={state}"))
+        .expect("callback query without question mark");
+
+    let (pending, _) = pending_authorization();
+    pending
+        .accept_authorization_input("  bare-code  ")
+        .expect("bare authorization code");
+
+    for code in [
+        "opaque&state=still-code",
+        "opaque&code=still-code",
+        "opaque-error=value",
+        "opaque://still-code",
+    ] {
+        let (pending, _) = pending_authorization();
+        pending
+            .accept_authorization_input(code)
+            .expect("opaque bare authorization code");
+    }
+}
+
+#[test]
+fn authorization_input_should_keep_url_state_and_code_validation_strict() {
+    let (pending, state) = pending_authorization();
+    pending
+        .accept_authorization_input(&format!(
+            "https://attacker.invalid/callback?code=code&state={state}"
+        ))
+        .expect_err("foreign callback host");
+
+    let (pending, _) = pending_authorization();
+    pending
+        .accept_authorization_input(&format!("{REDIRECT_URI}?code=code&state=wrong-state"))
+        .expect_err("callback state mismatch");
+
+    let (pending, state) = pending_authorization();
+    pending
+        .accept_authorization_input(&format!("{REDIRECT_URI}?code=code&state={state}#fragment"))
+        .expect_err("callback fragment");
+
+    let (pending, _) = pending_authorization();
+    pending
+        .accept_authorization_input("https:// malformed callback")
+        .expect_err("malformed callback URL");
+
+    let (pending, _) = pending_authorization();
+    pending
+        .accept_authorization_input("//attacker.invalid/callback?code=code")
+        .expect_err("scheme-relative callback URL");
+
+    let (pending, _) = pending_authorization();
+    pending
+        .accept_authorization_input("?unrelated=value")
+        .expect_err("query input without code or state");
+
+    let (pending, _) = pending_authorization();
+    pending
+        .accept_authorization_input("code=query-code")
+        .expect_err("query input without state");
+
+    let (pending, state) = pending_authorization();
+    pending
+        .accept_authorization_input(&format!("code=one&code=two&state={state}"))
+        .expect_err("duplicate code");
+
+    let (pending, _) = pending_authorization();
+    pending
+        .accept_authorization_input("code-with\n-control")
+        .expect_err("control character");
+
+    let (pending, state) = pending_authorization();
+    pending
+        .accept_authorization_input(&format!("?code=line%0Abreak&state={state}"))
+        .expect_err("percent-decoded control character");
+
+    let (pending, _) = pending_authorization();
+    pending
+        .accept_authorization_input(&"x".repeat(64 * 1024))
+        .expect("maximum-size code");
+
+    let (pending, _) = pending_authorization();
+    pending
+        .accept_authorization_input(&"x".repeat(64 * 1024 + 1))
+        .expect_err("oversized code");
+}
+
+fn pending_authorization() -> (PendingAuthorization, String) {
+    let config = GrokOAuthConfig::official().expect("fixture config");
+    let discovery = DiscoveryDocument::parse(&config, include_bytes!("fixtures/discovery.json"))
+        .expect("fixture discovery");
+    let redirect = RedirectUriAllowlist::new([REDIRECT_URI])
+        .expect("redirect allowlist")
+        .authorize(REDIRECT_URI)
+        .expect("allowlisted redirect");
+    let pending = PendingAuthorization::start(&config, &discovery, redirect, None)
+        .expect("start pending flow");
+    let state = pending
+        .authorization_url()
+        .query_pairs()
+        .find_map(|(name, value)| (name == "state").then(|| value.into_owned()))
+        .expect("authorization state");
+    (pending, state)
+}

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

@@ -560,7 +560,7 @@ async fn model_quota_feedback_writes_model_scoped_cooldown_without_blocking_othe
 
     let account = fixture.store.account(&selected).expect("selected account");
     assert_eq!(account.credential_state(), CredentialState::Ready);
-    let minimum_until = SystemTime::now() + Duration::from_secs(23 * 60 * 60);
+    let minimum_until = SystemTime::now() + Duration::from_secs(9 * 60);
     let scope = ProviderCooldownScope::upstream_model(failed_model);
     // model-scoped cooldown：目标模型被排除，账号级无 cooldown。
     assert!(

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

@@ -42,7 +42,7 @@ use provider_xai::{
     GrokSessionSelection, GrokSessionSelector, GrokSessionSelectorError, GrokSessionSelectorFuture,
     SecretValue, SelectedGrokSession,
 };
-use serde_json::{Map, json};
+use serde_json::{Map, Value, json};
 
 use crate::support::{
     MemoryGrokCatalogCache, MemoryProviderAccountStore, account_id, create_input, seed_input,
@@ -208,6 +208,8 @@ fn compaction_sse(summary: &str, reasoning: Option<&str>) -> Vec<u8> {
             "event: response.created\n",
             "data: {{\"type\":\"response.created\",\"response\":{{\"id\":\"resp_compaction\",\"model\":\"grok-4.5\"}}}}\n\n",
             "{reasoning}",
+            "event: response.output_item.done\n",
+            "data: {{\"type\":\"response.output_item.done\",\"output_index\":0,\"item\":{{\"id\":\"reason_compaction\",\"type\":\"reasoning\",\"status\":\"completed\",\"summary\":[],\"encrypted_content\":\"grok-compaction-ciphertext\"}}}}\n\n",
             "event: response.content_part.added\n",
             "data: {{\"type\":\"response.content_part.added\",\"output_index\":1,\"content_index\":0,\"part\":{{\"type\":\"output_text\"}}}}\n\n",
             "event: response.output_text.delta\n",
@@ -221,7 +223,53 @@ fn compaction_sse(summary: &str, reasoning: Option<&str>) -> Vec<u8> {
     .into_bytes()
 }
 
+fn terminal_only_compaction_sse(summary: &str) -> Vec<u8> {
+    let summary = serde_json::to_string(summary).expect("summary JSON string");
+    format!(
+        concat!(
+            "event: response.created\n",
+            "data: {{\"type\":\"response.created\",\"response\":{{\"id\":\"resp_compaction\",\"model\":\"grok-4.5\"}}}}\n\n",
+            "event: response.completed\n",
+            "data: {{\"type\":\"response.completed\",\"response\":{{\"id\":\"resp_compaction\",\"model\":\"grok-4.5\",\"status\":\"completed\",\"output\":[{{\"id\":\"reason_compaction\",\"type\":\"reasoning\",\"status\":\"completed\",\"encrypted_content\":\"grok-compaction-ciphertext\"}},{{\"id\":\"msg_compaction\",\"type\":\"message\",\"status\":\"completed\",\"role\":\"assistant\",\"content\":[{{\"type\":\"output_text\",\"text\":{summary}}}]}}],\"usage\":{{\"input_tokens\":100,\"output_tokens\":20,\"total_tokens\":120}}}}}}\n\n",
+        ),
+        summary = summary,
+    )
+    .into_bytes()
+}
+
+fn incomplete_compaction_sse(summary: &str) -> Vec<u8> {
+    let summary = serde_json::to_string(summary).expect("summary JSON string");
+    format!(
+        concat!(
+            "event: response.created\n",
+            "data: {{\"type\":\"response.created\",\"response\":{{\"id\":\"resp_compaction\",\"model\":\"grok-4.5\"}}}}\n\n",
+            "event: response.incomplete\n",
+            "data: {{\"type\":\"response.incomplete\",\"response\":{{\"id\":\"resp_compaction\",\"model\":\"grok-4.5\",\"status\":\"incomplete\",\"incomplete_details\":{{\"reason\":\"max_output_tokens\"}},\"output_text\":{summary},\"output\":[{{\"id\":\"reason_compaction\",\"type\":\"reasoning\",\"status\":\"completed\",\"encrypted_content\":\"grok-compaction-ciphertext\"}},{{\"id\":\"msg_compaction\",\"type\":\"message\",\"status\":\"incomplete\",\"role\":\"assistant\",\"content\":[{{\"type\":\"output_text\",\"text\":{summary}}}]}}],\"usage\":{{\"input_tokens\":100,\"output_tokens\":20,\"total_tokens\":120}}}}}}\n\n",
+        ),
+        summary = summary,
+    )
+    .into_bytes()
+}
+
 fn compaction_sse_without_terminal(summary: &str) -> Vec<u8> {
+    let summary = serde_json::to_string(summary).expect("summary JSON string");
+    format!(
+        concat!(
+            "event: response.created\n",
+            "data: {{\"type\":\"response.created\",\"response\":{{\"id\":\"resp_compaction\",\"model\":\"grok-4.5\"}}}}\n\n",
+            "event: response.output_item.done\n",
+            "data: {{\"type\":\"response.output_item.done\",\"output_index\":0,\"item\":{{\"id\":\"reason_compaction\",\"type\":\"reasoning\",\"status\":\"completed\",\"summary\":[],\"encrypted_content\":\"grok-compaction-ciphertext\"}}}}\n\n",
+            "event: response.content_part.added\n",
+            "data: {{\"type\":\"response.content_part.added\",\"output_index\":0,\"content_index\":0,\"part\":{{\"type\":\"output_text\"}}}}\n\n",
+            "event: response.output_text.delta\n",
+            "data: {{\"type\":\"response.output_text.delta\",\"output_index\":0,\"content_index\":0,\"delta\":{summary}}}\n\n",
+        ),
+        summary = summary,
+    )
+    .into_bytes()
+}
+
+fn compaction_sse_without_encrypted_content(summary: &str) -> Vec<u8> {
     let summary = serde_json::to_string(summary).expect("summary JSON string");
     format!(
         concat!(
@@ -231,6 +279,8 @@ fn compaction_sse_without_terminal(summary: &str) -> Vec<u8> {
             "data: {{\"type\":\"response.content_part.added\",\"output_index\":0,\"content_index\":0,\"part\":{{\"type\":\"output_text\"}}}}\n\n",
             "event: response.output_text.delta\n",
             "data: {{\"type\":\"response.output_text.delta\",\"output_index\":0,\"content_index\":0,\"delta\":{summary}}}\n\n",
+            "event: response.completed\n",
+            "data: {{\"type\":\"response.completed\",\"response\":{{\"id\":\"resp_compaction\",\"model\":\"grok-4.5\",\"status\":\"completed\"}}}}\n\n",
         ),
         summary = summary,
     )
@@ -260,6 +310,7 @@ struct StubSelector {
     feedback: Mutex<Vec<GrokCredentialFailure>>,
     error: Mutex<Option<GrokSessionSelectorError>>,
     required_accounts: Mutex<Vec<Option<gateway_core::engine::credential::ProviderAccountId>>>,
+    upstream_models: Mutex<Vec<String>>,
 }
 
 impl StubSelector {
@@ -269,6 +320,7 @@ impl StubSelector {
             feedback: Mutex::new(Vec::new()),
             error: Mutex::new(None),
             required_accounts: Mutex::new(Vec::new()),
+            upstream_models: Mutex::new(Vec::new()),
         })
     }
 
@@ -278,6 +330,7 @@ impl StubSelector {
             feedback: Mutex::new(Vec::new()),
             error: Mutex::new(Some(error)),
             required_accounts: Mutex::new(Vec::new()),
+            upstream_models: Mutex::new(Vec::new()),
         })
     }
 }
@@ -285,6 +338,10 @@ impl StubSelector {
 impl GrokSessionSelector for StubSelector {
     fn select(&self, request: GrokSessionSelection) -> GrokSessionSelectorFuture<'_> {
         self.calls.fetch_add(1, Ordering::SeqCst);
+        self.upstream_models
+            .lock()
+            .expect("upstream models")
+            .push(request.upstream_model().as_str().to_owned());
         let error = self.error.lock().expect("selector error").take();
         let required_account = request.required_account().cloned();
         self.required_accounts
@@ -630,6 +687,70 @@ async fn mapped_transport_error(
     next_provider_error(&mut stream).await
 }
 
+#[tokio::test]
+async fn invalid_encrypted_reasoning_retries_once_on_the_same_account_without_ciphertext() {
+    let transport = StubInferenceTransport::sequence([
+        InferenceMode::Error(
+            GrokInferenceTransportError::new(
+                GrokInferenceTransportErrorKind::InvalidRequest,
+                UpstreamSendState::Sent,
+            )
+            .with_status(400)
+            .with_upstream_code(OpaqueUpstreamValue::new("reasoning_decode_failed")),
+        ),
+        InferenceMode::Success,
+    ]);
+    let provider = provider(StubSelector::success(), transport.clone()).await;
+    let operation = Operation::Generate(GenerateRequest::from_protocol_payload(
+        ProtocolPayload::json_object(
+            "openai",
+            Map::from_iter([
+                ("model".to_owned(), json!("client-model")),
+                (
+                    "input".to_owned(),
+                    json!([
+                        {
+                            "type": "reasoning",
+                            "id": "reason_stale",
+                            "summary": [{"type": "summary_text", "text": "keep summary"}],
+                            "content": null,
+                            "encrypted_content": "stale-ciphertext"
+                        },
+                        {"type": "message", "role": "user", "content": "continue"}
+                    ]),
+                ),
+            ]),
+        )
+        .expect("OpenAI payload"),
+    ));
+    let mut stream = provider
+        .execute(
+            provider_request_with_operation("xai", operation),
+            context(CancellationToken::new(), None),
+        )
+        .await
+        .expect("retrying stream");
+    while let Some(event) = stream.next().await {
+        event.expect("recovered response");
+    }
+
+    let requests = transport.requests.lock().expect("requests");
+    assert_eq!(requests.len(), 2);
+    let first: serde_json::Value = serde_json::from_slice(requests[0].body()).expect("first body");
+    let second: serde_json::Value = serde_json::from_slice(requests[1].body()).expect("retry body");
+    assert_eq!(
+        first.pointer("/input/0/encrypted_content"),
+        Some(&json!("stale-ciphertext"))
+    );
+    assert_eq!(second.pointer("/input/0/encrypted_content"), None);
+    assert_eq!(second.pointer("/input/0/content"), None);
+    assert_eq!(
+        second.pointer("/input/0/summary/0/text"),
+        Some(&json!("keep summary"))
+    );
+    assert_eq!(requests[0].binding(), requests[1].binding());
+}
+
 async fn next_provider_error(
     stream: &mut gateway_core::engine::provider::ProviderStream,
 ) -> ProviderError {
@@ -803,10 +924,18 @@ fn provider_request(provider_kind: &str) -> ProviderRequest {
 }
 
 fn provider_request_with_operation(provider_kind: &str, operation: Operation) -> ProviderRequest {
+    provider_request_with_upstream_model(provider_kind, MODEL, operation)
+}
+
+fn provider_request_with_upstream_model(
+    provider_kind: &str,
+    upstream_model: &str,
+    operation: Operation,
+) -> ProviderRequest {
     let provider = ProviderKind::new(provider_kind).expect("provider");
     let provider_model = ProviderModel::new(
         provider.clone(),
-        UpstreamModelId::new(MODEL).expect("model"),
+        UpstreamModelId::new(upstream_model).expect("model"),
         ModelCapabilities::new(BTreeSet::from([OperationKind::Generate]), Some(131_072))
             .with_feature(Feature::Tools, SupportLevel::Native)
             .with_feature(Feature::NativeContinuation, SupportLevel::Native),
@@ -828,7 +957,7 @@ fn provider_request_with_operation(provider_kind: &str, operation: Operation) ->
     .expect("snapshot");
     let plan = snapshot
         .plan(
-            &PublicModelId::new(MODEL).expect("model"),
+            &PublicModelId::new(upstream_model).expect("model"),
             &operation,
             account_scope,
             &RoutingContext::default(),
@@ -904,6 +1033,31 @@ async fn execute_successfully(provider: &GrokBuildProvider, operation: Operation
     assert!(events.iter().all(Result::is_ok));
 }
 
+#[tokio::test]
+async fn model_alias_should_be_canonical_before_account_selection() {
+    let selector = StubSelector::success();
+    let provider = provider(selector.clone(), StubInferenceTransport::success()).await;
+    let mut stream = provider
+        .execute(
+            provider_request_with_upstream_model("xai", "grok-4.6-latest", operation()),
+            context(CancellationToken::new(), None),
+        )
+        .await
+        .expect("provider stream");
+    while let Some(event) = stream.next().await {
+        event.expect("provider event");
+    }
+
+    assert_eq!(
+        selector
+            .upstream_models
+            .lock()
+            .expect("upstream models")
+            .as_slice(),
+        ["grok-4.6"]
+    );
+}
+
 #[tokio::test]
 async fn request_observation_preserves_raw_xai_reasoning_effort() {
     let provider = provider(StubSelector::success(), StubInferenceTransport::success()).await;
@@ -1125,6 +1279,145 @@ async fn compaction_stream_should_emit_openai_wire_and_keep_only_accounting_cano
     );
 }
 
+#[tokio::test]
+async fn compaction_should_use_terminal_output_when_upstream_omits_text_deltas() {
+    let summary = valid_compaction_summary("terminal-only summary");
+    let transport = StubInferenceTransport::sequence([InferenceMode::SuccessBody(
+        terminal_only_compaction_sse(&summary),
+    )]);
+    let provider = provider(StubSelector::success(), transport).await;
+    let mut stream = provider
+        .execute(
+            provider_request_with_operation("xai", compaction_operation()),
+            context(CancellationToken::new(), None),
+        )
+        .await
+        .expect("compaction stream");
+    let events = stream.by_ref().collect::<Vec<_>>().await;
+    let compact_item = events
+        .iter()
+        .map(|event| event.as_ref().expect("successful compaction event"))
+        .filter_map(|event| event.wire_event())
+        .find(|wire| wire.event_type() == Some("response.output_item.done"))
+        .and_then(|wire| wire.data().get("item"))
+        .expect("compaction output item");
+
+    assert_eq!(
+        compact_item.pointer("/summary/0/text"),
+        Some(&Value::String(summary))
+    );
+}
+
+#[tokio::test]
+async fn compaction_should_complete_incomplete_upstream_when_ciphertext_is_present() {
+    let summary = valid_compaction_summary("incomplete upstream summary");
+    let transport = StubInferenceTransport::sequence([InferenceMode::SuccessBody(
+        incomplete_compaction_sse(&summary),
+    )]);
+    let provider = provider(StubSelector::success(), transport).await;
+    let mut stream = provider
+        .execute(
+            provider_request_with_operation("xai", compaction_operation()),
+            context(CancellationToken::new(), None),
+        )
+        .await
+        .expect("compaction stream");
+    let events = stream.by_ref().collect::<Vec<_>>().await;
+    let terminal = events
+        .iter()
+        .map(|event| event.as_ref().expect("successful compaction event"))
+        .filter_map(|event| event.wire_event())
+        .find(|wire| wire.event_type() == Some("response.completed"))
+        .expect("completed compaction response");
+    let response = terminal
+        .data()
+        .get("response")
+        .expect("terminal response object");
+
+    assert_eq!(
+        terminal.data().get("type"),
+        Some(&json!("response.completed"))
+    );
+    assert_eq!(response.get("status"), Some(&json!("completed")));
+    assert_eq!(response.get("incomplete_details"), Some(&Value::Null));
+    assert!(
+        !response
+            .as_object()
+            .expect("response object")
+            .contains_key("output_text")
+    );
+    assert_eq!(
+        response.pointer("/output/0/encrypted_content"),
+        Some(&json!("grok-compaction-ciphertext"))
+    );
+}
+
+#[tokio::test]
+async fn compaction_invalid_encrypted_reasoning_should_retry_once_on_the_same_account() {
+    let transport = StubInferenceTransport::sequence([
+        InferenceMode::Error(
+            GrokInferenceTransportError::new(
+                GrokInferenceTransportErrorKind::InvalidRequest,
+                UpstreamSendState::Sent,
+            )
+            .with_status(400)
+            .with_upstream_code(OpaqueUpstreamValue::new("reasoning_decode_failed")),
+        ),
+        InferenceMode::SuccessBody(compaction_sse(
+            &valid_compaction_summary("recovered compaction"),
+            None,
+        )),
+    ]);
+    let provider = provider(StubSelector::success(), transport.clone()).await;
+    let payload = ProtocolPayload::json_object(
+        "openai",
+        object(json!({
+            "model": "client-model",
+            "input": [
+                {
+                    "type": "reasoning",
+                    "summary": [{"type": "summary_text", "text": "keep summary"}],
+                    "content": null,
+                    "encrypted_content": "stale-compaction-ciphertext"
+                },
+                {"type": "message", "role": "user", "content": "history"},
+                {"type": "compaction_trigger"}
+            ],
+            "stream": true
+        })),
+    )
+    .expect("OpenAI payload");
+    let mut stream = provider
+        .execute(
+            provider_request_with_operation(
+                "xai",
+                Operation::Generate(GenerateRequest::from_protocol_payload(payload)),
+            ),
+            context(CancellationToken::new(), None),
+        )
+        .await
+        .expect("retrying compaction stream");
+    while let Some(event) = stream.next().await {
+        event.expect("recovered compaction response");
+    }
+
+    let requests = transport.requests.lock().expect("requests");
+    assert_eq!(requests.len(), 2);
+    let first: serde_json::Value = serde_json::from_slice(requests[0].body()).expect("first body");
+    let second: serde_json::Value = serde_json::from_slice(requests[1].body()).expect("retry body");
+    assert_eq!(
+        first.pointer("/input/0/encrypted_content"),
+        Some(&json!("stale-compaction-ciphertext"))
+    );
+    assert_eq!(second.pointer("/input/0/encrypted_content"), None);
+    assert_eq!(second.pointer("/input/0/content"), None);
+    assert_eq!(
+        second.pointer("/input/0/summary/0/text"),
+        Some(&json!("keep summary"))
+    );
+    assert_eq!(requests[0].binding(), requests[1].binding());
+}
+
 #[tokio::test]
 async fn compaction_should_pin_recorded_account_and_forward_session_headers() {
     let state = ProviderSessionState::new(
@@ -1180,7 +1473,7 @@ async fn compaction_should_pin_recorded_account_and_forward_session_headers() {
 }
 
 #[tokio::test]
-async fn compaction_should_accept_summary_when_upstream_stream_ends_without_terminal() {
+async fn compaction_should_reject_upstream_stream_that_ends_without_terminal() {
     let transport = StubInferenceTransport::sequence([InferenceMode::SuccessBody(
         compaction_sse_without_terminal(&valid_compaction_summary("usable eof summary")),
     )]);
@@ -1192,32 +1485,10 @@ async fn compaction_should_accept_summary_when_upstream_stream_ends_without_term
         )
         .await
         .expect("compaction stream");
-    let events = stream
-        .by_ref()
-        .map(|event| event.expect("successful compaction event"))
-        .collect::<Vec<_>>()
-        .await;
-    let facts = events
-        .iter()
-        .flat_map(|event| event.canonical_facts())
-        .collect::<Vec<_>>();
-    let wire_types = events
-        .iter()
-        .filter_map(|event| event.wire_event().and_then(|wire| wire.event_type()))
-        .collect::<Vec<_>>();
+    let error = next_provider_error(&mut stream).await;
 
-    assert!(matches!(
-        facts.as_slice(),
-        [GatewayEvent::Started(_), GatewayEvent::Completed(_)]
-    ));
-    assert_eq!(
-        wire_types,
-        [
-            "response.created",
-            "response.output_item.done",
-            "response.completed"
-        ]
-    );
+    assert_eq!(error.kind(), ProviderErrorKind::Protocol);
+    assert!(error.allows_pre_delivery_retry());
 }
 
 #[tokio::test]
@@ -1236,21 +1507,34 @@ async fn compaction_wire_should_exclude_reasoning_from_summary_content() {
         .await
         .expect("compaction stream");
     let events = stream.by_ref().collect::<Vec<_>>().await;
-    let summary = events
+    let item = events
         .iter()
         .map(|event| event.as_ref().expect("successful compaction event"))
         .filter_map(|event| event.wire_event())
         .find(|wire| wire.event_type() == Some("response.output_item.done"))
-        .and_then(|wire| wire.data().pointer("/item/encrypted_content"))
-        .and_then(serde_json::Value::as_str)
+        .and_then(|wire| wire.data().get("item"))
         .expect("compaction output wire");
+    let summary = item
+        .pointer("/summary/0/text")
+        .and_then(serde_json::Value::as_str)
+        .expect("visible compaction summary");
 
     assert!(summary.contains("continuation marker"));
     assert!(!summary.contains("private reasoning"));
+    assert_eq!(
+        item.get("encrypted_content"),
+        Some(&json!("grok-compaction-ciphertext"))
+    );
+    assert_eq!(item.get("status"), Some(&json!("completed")));
+    assert!(
+        item.get("id")
+            .and_then(serde_json::Value::as_str)
+            .is_some_and(|id| id.starts_with("cmp_"))
+    );
 }
 
 #[tokio::test]
-async fn compaction_stream_should_reject_degenerate_summary_before_commit() {
+async fn compaction_stream_should_accept_short_summary_like_sub2api() {
     let transport = StubInferenceTransport::sequence([InferenceMode::SuccessBody(compaction_sse(
         "<summary>too short</summary>",
         None,
@@ -1264,12 +1548,77 @@ async fn compaction_stream_should_reject_degenerate_summary_before_commit() {
         .await
         .expect("compaction stream");
 
+    let events = stream
+        .by_ref()
+        .map(|event| event.expect("successful compaction event"))
+        .collect::<Vec<_>>()
+        .await;
+    let item = events
+        .iter()
+        .filter_map(|event| event.wire_event())
+        .find(|wire| wire.event_type() == Some("response.output_item.done"))
+        .and_then(|wire| wire.data().get("item"))
+        .expect("compaction item");
+
+    assert_eq!(
+        item.pointer("/summary/0/text"),
+        Some(&json!("<summary>too short</summary>"))
+    );
+    assert_eq!(
+        item.get("encrypted_content"),
+        Some(&json!("grok-compaction-ciphertext"))
+    );
+}
+
+#[tokio::test]
+async fn compaction_stream_should_allow_ciphertext_without_visible_summary() {
+    let transport =
+        StubInferenceTransport::sequence([InferenceMode::SuccessBody(compaction_sse("   ", None))]);
+    let provider = provider(StubSelector::success(), transport).await;
+    let mut stream = provider
+        .execute(
+            provider_request_with_operation("xai", compaction_operation()),
+            context(CancellationToken::new(), None),
+        )
+        .await
+        .expect("compaction stream");
+    let events = stream
+        .by_ref()
+        .map(|event| event.expect("successful compaction event"))
+        .collect::<Vec<_>>()
+        .await;
+    let item = events
+        .iter()
+        .filter_map(|event| event.wire_event())
+        .find(|wire| wire.event_type() == Some("response.output_item.done"))
+        .and_then(|wire| wire.data().get("item"))
+        .expect("compaction item");
+
+    assert_eq!(item.get("summary"), None);
+    assert_eq!(
+        item.get("encrypted_content"),
+        Some(&json!("grok-compaction-ciphertext"))
+    );
+}
+
+#[tokio::test]
+async fn compaction_stream_should_require_real_reasoning_ciphertext_before_commit() {
+    let transport = StubInferenceTransport::sequence([InferenceMode::SuccessBody(
+        compaction_sse_without_encrypted_content(&valid_compaction_summary("visible summary")),
+    )]);
+    let provider = provider(StubSelector::success(), transport).await;
+    let mut stream = provider
+        .execute(
+            provider_request_with_operation("xai", compaction_operation()),
+            context(CancellationToken::new(), None),
+        )
+        .await
+        .expect("compaction stream");
+
     let error = next_provider_error(&mut stream).await;
 
     assert_eq!(error.kind(), ProviderErrorKind::Protocol);
-    assert!(!error.replay_is_safe());
     assert!(error.allows_pre_delivery_retry());
-    assert!(!error.retries_same_account());
 }
 
 #[tokio::test]
@@ -1701,7 +2050,7 @@ async fn explicit_http_408_does_not_mark_provider_error_replay_safe() {
 }
 
 #[tokio::test]
-async fn generic_http_403_does_not_mutate_or_switch_credential() {
+async fn generic_http_403_cools_account_and_allows_next_account_replay() {
     let selector = StubSelector::success();
     let transport = StubInferenceTransport::error(
         GrokInferenceTransportError::new(
@@ -1721,14 +2070,17 @@ async fn generic_http_403_does_not_mutate_or_switch_credential() {
     let error = next_provider_error(&mut stream).await;
 
     assert_eq!(error.kind(), ProviderErrorKind::PermissionDenied);
-    assert!(!error.replay_is_safe());
+    assert!(error.replay_is_safe());
     assert_eq!(selector.calls.load(Ordering::SeqCst), 1);
     assert_eq!(transport.calls.load(Ordering::SeqCst), 1);
-    assert!(selector.feedback.lock().expect("feedback").is_empty());
+    assert_eq!(
+        selector.feedback.lock().expect("feedback").as_slice(),
+        &[GrokCredentialFailure::AccessDenied]
+    );
 }
 
 #[tokio::test]
-async fn generic_http_500_does_not_mark_provider_error_replay_safe() {
+async fn generic_http_500_allows_next_account_replay() {
     let error = mapped_transport_error(
         GrokInferenceTransportError::new(
             GrokInferenceTransportErrorKind::Unavailable,
@@ -1740,7 +2092,64 @@ async fn generic_http_500_does_not_mark_provider_error_replay_safe() {
     .await;
 
     assert_eq!(error.kind(), ProviderErrorKind::Unavailable);
+    assert!(error.replay_is_safe());
+}
+
+#[tokio::test]
+async fn rate_limit_wording_on_http_400_stays_request_scoped() {
+    let selector = StubSelector::success();
+    let transport = StubInferenceTransport::error(
+        GrokInferenceTransportError::new(
+            GrokInferenceTransportErrorKind::InvalidRequest,
+            UpstreamSendState::Sent,
+        )
+        .with_status(400)
+        .with_upstream_code(OpaqueUpstreamValue::new("rate_limit")),
+    );
+    let provider = provider(selector.clone(), transport).await;
+    let mut stream = provider
+        .execute(
+            provider_request("xai"),
+            context(CancellationToken::new(), None),
+        )
+        .await
+        .expect("stream");
+    let error = next_provider_error(&mut stream).await;
+
+    assert_eq!(error.kind(), ProviderErrorKind::InvalidRequest);
     assert!(!error.replay_is_safe());
+    assert!(selector.feedback.lock().expect("feedback").is_empty());
+}
+
+#[tokio::test]
+async fn model_capacity_on_http_400_cools_scope_and_allows_replay() {
+    let selector = StubSelector::success();
+    let transport = StubInferenceTransport::error(
+        GrokInferenceTransportError::new(
+            GrokInferenceTransportErrorKind::RateLimited,
+            UpstreamSendState::Sent,
+        )
+        .with_status(400)
+        .with_upstream_code(OpaqueUpstreamValue::new("model_capacity")),
+    );
+    let provider = provider(selector.clone(), transport).await;
+    let mut stream = provider
+        .execute(
+            provider_request("xai"),
+            context(CancellationToken::new(), None),
+        )
+        .await
+        .expect("stream");
+    let error = next_provider_error(&mut stream).await;
+
+    assert_eq!(error.kind(), ProviderErrorKind::RateLimited);
+    assert!(error.replay_is_safe());
+    assert_eq!(
+        selector.feedback.lock().expect("feedback").as_slice(),
+        &[GrokCredentialFailure::ModelCapacity {
+            upstream_model: UpstreamModelId::new("grok-4.5").expect("model"),
+        }]
+    );
 }
 
 #[tokio::test]

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

@@ -2,6 +2,7 @@ use gateway_core::error::ProviderError;
 use gateway_core::event::{ContentKind, FinishReason, GatewayEvent, ProviderEvent};
 use gateway_core::operation::{GenerateRequest, ProtocolPayload};
 use gateway_core::policy::ClientApiKeyId;
+use gateway_protocol::openai::sse::{encode_sse_event, encode_sse_event_with_metadata};
 use serde_json::Value;
 
 use provider_xai::{GrokCanonicalDecoder, GrokResponsesRequest, grok_billing_breakdown};
@@ -579,6 +580,274 @@ fn decoder_should_buffer_only_custom_arguments_until_the_done_event() {
     }));
 }
 
+#[test]
+fn decoder_should_normalize_schema_integer_arguments_and_resequence_split_events() {
+    let request = tool_request(serde_json::json!({
+        "model": "client",
+        "input": "wait",
+        "tools": [{
+            "type": "function",
+            "name": "wait",
+            "parameters": {
+                "$ref": "#/$defs/Args",
+                "$defs": {
+                    "Args": {
+                        "type": "object",
+                        "properties": {
+                            "timeout_ms": {"allOf": [{"type": "integer"}]},
+                            "fractional": {"type": "integer"},
+                            "unsafe": {"type": "integer"},
+                            "list": {
+                                "type": "array",
+                                "prefixItems": [{"type": "integer"}],
+                                "items": {"type": "number"}
+                            },
+                            "extra": {
+                                "type": "object",
+                                "additionalProperties": {"type": "integer"}
+                            }
+                        }
+                    },
+                    "Unused": {"type": "integer"}
+                }
+            }
+        }]
+    }));
+    let arguments = concat!(
+        "{\"timeout_ms\":60000.0,\"fractional\":1.5,",
+        "\"unsafe\":9007199254740992.0,\"list\":[2.0,3.0],",
+        "\"extra\":{\"retries\":4.0},\"untyped\":7.0}"
+    );
+    let body = [
+        encode_sse_event(
+            "response.created",
+            &serde_json::json!({
+                "type": "response.created",
+                "response": {"id": "resp_integer"}
+            })
+            .to_string(),
+        ),
+        encode_sse_event(
+            "response.output_item.added",
+            &serde_json::json!({
+                "type": "response.output_item.added",
+                "sequence_number": 40.0,
+                "output_index": 0,
+                "item": {
+                    "id": "item_wait",
+                    "type": "function_call",
+                    "call_id": "call_wait",
+                    "name": "wait"
+                }
+            })
+            .to_string(),
+        ),
+        encode_sse_event(
+            "",
+            &serde_json::json!({
+                "sequence_number": "invalid",
+                "future": true
+            })
+            .to_string(),
+        ),
+        encode_sse_event(
+            "response.function_call_arguments.delta",
+            &serde_json::json!({
+                "type": "response.function_call_arguments.delta",
+                "sequence_number": 91,
+                "item_id": "item_wait",
+                "call_id": "call_wait",
+                "output_index": 0,
+                "delta": arguments
+            })
+            .to_string(),
+        ),
+        encode_sse_event_with_metadata(
+            "response.function_call_arguments.done",
+            &serde_json::json!({
+                "type": "response.function_call_arguments.done",
+                "sequence_number": 120,
+                "item_id": "item_wait",
+                "call_id": "call_wait",
+                "output_index": 0,
+                "arguments": arguments
+            })
+            .to_string(),
+            Some("evt_arguments_done"),
+            Some(250),
+        ),
+        encode_sse_event(
+            "response.output_item.done",
+            &serde_json::json!({
+                "type": "response.output_item.done",
+                "sequence_number": 200,
+                "output_index": 0,
+                "item": {
+                    "id": "item_wait",
+                    "type": "function_call",
+                    "call_id": "call_wait",
+                    "name": "wait",
+                    "arguments": arguments
+                }
+            })
+            .to_string(),
+        ),
+    ]
+    .concat();
+
+    let events = GrokCanonicalDecoder::for_request("grok-4.5", &request)
+        .push(body.as_bytes())
+        .expect("integer argument stream");
+    let wire = wire_events(&events);
+    let sequences = wire
+        .iter()
+        .filter_map(|event| event.data().get("sequence_number")?.as_u64())
+        .collect::<Vec<_>>();
+
+    assert_eq!(sequences, vec![40, 41, 42, 43, 44]);
+    assert!(wire.iter().any(|event| {
+        event.event_type().is_none()
+            && event.data().get("future") == Some(&serde_json::json!(true))
+            && event.data().get("sequence_number") == Some(&serde_json::json!(41))
+    }));
+
+    let argument_events = wire
+        .iter()
+        .filter(|event| {
+            matches!(
+                event.event_type(),
+                Some(
+                    "response.function_call_arguments.delta"
+                        | "response.function_call_arguments.done"
+                        | "response.output_item.done"
+                )
+            )
+        })
+        .collect::<Vec<_>>();
+    assert_eq!(argument_events.len(), 3);
+    for event in &argument_events {
+        let raw = event
+            .data()
+            .pointer(if event.event_type() == Some("response.output_item.done") {
+                "/item/arguments"
+            } else if event.event_type() == Some("response.function_call_arguments.delta") {
+                "/delta"
+            } else {
+                "/arguments"
+            })
+            .and_then(Value::as_str)
+            .expect("normalized arguments");
+        let normalized = serde_json::from_str::<Value>(raw).expect("argument JSON");
+        assert_eq!(
+            normalized.pointer("/timeout_ms"),
+            Some(&serde_json::json!(60000))
+        );
+        assert_eq!(
+            normalized.pointer("/fractional"),
+            Some(&serde_json::json!(1.5))
+        );
+        assert_eq!(
+            normalized.pointer("/unsafe"),
+            Some(&serde_json::json!(9007199254740992.0))
+        );
+        assert_eq!(normalized.pointer("/list/0"), Some(&serde_json::json!(2)));
+        assert_eq!(normalized.pointer("/list/1"), Some(&serde_json::json!(3.0)));
+        assert_eq!(
+            normalized.pointer("/extra/retries"),
+            Some(&serde_json::json!(4))
+        );
+        assert_eq!(
+            normalized.pointer("/untyped"),
+            Some(&serde_json::json!(7.0))
+        );
+    }
+    assert_eq!(argument_events[0].sse_id(), Some("evt_arguments_done"));
+    assert_eq!(argument_events[0].sse_retry(), Some(250));
+    assert_eq!(argument_events[1].sse_id(), None);
+    assert_eq!(argument_events[1].sse_retry(), None);
+}
+
+#[test]
+fn decoder_should_fail_open_after_the_function_argument_buffer_limit() {
+    let request = tool_request(serde_json::json!({
+        "model": "client",
+        "input": "wait",
+        "tools": [{
+            "type": "function",
+            "name": "wait",
+            "parameters": {
+                "type": "object",
+                "properties": {"timeout_ms": {"type": "integer"}}
+            }
+        }]
+    }));
+    let arguments = format!(
+        "{{\"timeout_ms\":1.0,\"padding\":\"{}\"}}",
+        "x".repeat((1 << 20) + 1)
+    );
+    let mut decoder = GrokCanonicalDecoder::for_request("grok-4.5", &request);
+    let added = encode_sse_event(
+        "response.output_item.added",
+        &serde_json::json!({
+            "type": "response.output_item.added",
+            "output_index": 0,
+            "item": {
+                "id": "item_large",
+                "type": "function_call",
+                "call_id": "call_large",
+                "name": "wait"
+            }
+        })
+        .to_string(),
+    );
+    decoder
+        .push(added.as_bytes())
+        .expect("large function item added");
+
+    let delta = encode_sse_event(
+        "response.function_call_arguments.delta",
+        &serde_json::json!({
+            "type": "response.function_call_arguments.delta",
+            "item_id": "item_large",
+            "call_id": "call_large",
+            "output_index": 0,
+            "delta": arguments
+        })
+        .to_string(),
+    );
+    let delta = decoder
+        .push(delta.as_bytes())
+        .expect("oversized arguments should pass through");
+    let delta_wire = wire_events(&delta);
+    assert_eq!(delta_wire.len(), 1);
+    assert_eq!(
+        delta_wire[0].data().get("delta").and_then(Value::as_str),
+        Some(arguments.as_str())
+    );
+
+    let done = encode_sse_event(
+        "response.function_call_arguments.done",
+        &serde_json::json!({
+            "type": "response.function_call_arguments.done",
+            "item_id": "item_large",
+            "call_id": "call_large",
+            "output_index": 0,
+            "arguments": arguments
+        })
+        .to_string(),
+    );
+    let done = decoder
+        .push(done.as_bytes())
+        .expect("oversized arguments done");
+    let normalized = wire_events(&done)[0]
+        .data()
+        .get("arguments")
+        .and_then(Value::as_str)
+        .and_then(|arguments| serde_json::from_str::<Value>(arguments).ok())
+        .expect("normalized final arguments");
+    assert_eq!(normalized.get("timeout_ms"), Some(&serde_json::json!(1)));
+}
+
 #[test]
 fn decoder_should_coalesce_reasoning_item_part_and_summary_index() {
     let body = concat!(
@@ -612,6 +881,10 @@ fn decoder_should_preserve_unknown_events_as_openai_wire() {
         "data: {\"type\":\"response.created\",\"response\":{\"id\":\"resp_1\"}}\n\n",
         "event: xai.internal.trace\n",
         "data: {\"type\":\"xai.internal.trace\",\"secret\":\"must-not-reach-client\"}\n\n",
+        "event: response.doom_loop_check \n",
+        "data: {\"type\":\"response.backend_tool_call.started\",\"secret\":\"drop-header\"}\n\n",
+        "event: response.backend_tool_call.started\n",
+        "data: {\"type\":\"response.doom_loop_check\",\"secret\":\"drop-body\"}\n\n",
         "event: response.backend_tool_call.started\n",
         "id: evt_future\n",
         "retry: 1250\n",
@@ -639,6 +912,32 @@ fn decoder_should_preserve_unknown_events_as_openai_wire() {
     assert!(events[1].canonical_facts().is_empty());
 }
 
+#[test]
+fn decoder_should_rewrite_grok_ping_frames_as_sse_comments() {
+    let body = concat!(
+        "event: response.created\n",
+        "data: {\"type\":\"response.created\",\"response\":{\"id\":\"resp_ping\"}}\n\n",
+        "event: ping\n",
+        "data: {\"type\":\"ping\",\"x-opencode-type\":\"inference-cost\",\"cost\":2.75}\n\n",
+        "event: ping\n",
+        "data: {not-json}\n\n",
+        "event: response.completed\n",
+        "data: {\"type\":\"response.completed\",\"response\":{\"id\":\"resp_ping\",\"status\":\"completed\"}}\n\n",
+    );
+
+    let events = GrokCanonicalDecoder::new("fallback")
+        .push(body.as_bytes())
+        .expect("ping-filtered response");
+    let wire = wire_events(&events);
+    let ping_comments = wire
+        .iter()
+        .filter(|event| event.raw_sse_frame().map(AsRef::as_ref) == Some(b": ping\n\n".as_slice()))
+        .count();
+
+    assert_eq!(ping_comments, 2);
+    assert!(wire.iter().all(|event| event.event_type() != Some("ping")));
+}
+
 #[test]
 fn decoder_should_preserve_image_and_hosted_tool_items_without_inventing_facts() {
     let body = concat!(

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

@@ -1,4 +1,6 @@
-use gateway_core::event::{GatewayEvent, ProviderEvent, ReasoningDelta, TextDelta, ToolCallDelta};
+use gateway_core::event::{
+    GatewayEvent, ProtocolWireEvent, ProviderEvent, ReasoningDelta, TextDelta, ToolCallDelta,
+};
 use gateway_core::operation::{GenerateRequest, ProtocolPayload};
 use gateway_core::policy::ClientApiKeyId;
 use provider_xai::{
@@ -74,7 +76,7 @@ fn encoder_should_preserve_history_order_and_append_summary_prompt() {
 }
 
 #[test]
-fn encoder_should_preserve_tools_and_remove_non_prefix_constraints() {
+fn encoder_should_preserve_sub2api_compaction_constraints() {
     let request = encode(json!({
         "model": "client-model",
         "input": [
@@ -89,21 +91,24 @@ fn encoder_should_preserve_tools_and_remove_non_prefix_constraints() {
         "prompt_cache_key": "cache_private",
         "service_tier": "priority",
         "max_output_tokens": 16,
+        "background": true,
+        "truncation": "auto",
         "stream": false,
         "store": true
     }))
     .expect("compaction request");
 
     let body = request.body();
-    let forbidden = [
-        "parallel_tool_calls",
-        "text",
-        "previous_response_id",
-        "prompt_cache_key",
-        "service_tier",
-        "max_output_tokens",
-    ];
+    let forbidden = ["previous_response_id", "prompt_cache_key", "service_tier"];
     assert!(forbidden.into_iter().all(|field| !body.contains_key(field)));
+    assert_eq!(body.get("parallel_tool_calls"), Some(&Value::Bool(true)));
+    assert_eq!(
+        body.get("text"),
+        Some(&json!({"format": {"type": "json_object"}}))
+    );
+    assert_eq!(body.get("max_output_tokens"), Some(&json!(16)));
+    assert_eq!(body.get("background"), Some(&Value::Bool(true)));
+    assert_eq!(body.get("truncation"), Some(&json!("auto")));
     assert_eq!(
         body.get("tools"),
         Some(&json!([{
@@ -112,8 +117,12 @@ fn encoder_should_preserve_tools_and_remove_non_prefix_constraints() {
             "parameters": {"type": "object"}
         }]))
     );
-    assert_eq!(body.get("tool_choice"), Some(&json!("auto")));
-    assert_eq!(body.get("temperature"), Some(&json!(1.0)));
+    assert_eq!(body.get("tool_choice"), Some(&json!("none")));
+    assert_eq!(
+        body.get("include"),
+        Some(&json!(["reasoning.encrypted_content"]))
+    );
+    assert!(!body.contains_key("temperature"));
     assert_eq!(body.get("stream"), Some(&Value::Bool(true)));
     assert_eq!(body.get("store"), Some(&Value::Bool(false)));
 }
@@ -181,54 +190,56 @@ fn encoder_should_consume_only_the_terminal_compaction_trigger() {
 }
 
 #[test]
-fn decoder_should_return_clean_typed_summary_after_normal_terminal() {
-    let summary = decode_summary(&valid_summary("checkpoint-only-secret")).expect("summary");
+fn decoder_should_preserve_sub2api_summary_text_after_trimming_edges() {
+    let raw = valid_summary("checkpoint-only-secret");
+    let summary = decode_summary(&format!("  {raw}  \n")).expect("summary");
 
-    assert!(summary.as_str().starts_with("Summary:\nPrivate state:"));
-    assert!(!summary.as_str().contains("<summary>"));
+    assert_eq!(summary, raw);
 }
 
 #[test]
-fn decoder_should_remove_nested_leading_scratchpad() {
+fn decoder_should_not_interpret_nested_summary_or_analysis_tokens() {
     let raw = format!(
         "<summary>\n<analysis>private scratchpad</analysis>\nActual state\n{}\n</summary>",
         "preserved implementation context ".repeat(20)
     );
-    let summary = decode_summary(&raw).expect("clean summary");
+    let summary = decode_summary(&raw).expect("summary");
 
     assert!(summary.as_str().contains("Actual state"));
-    assert!(!summary.as_str().contains("private scratchpad"));
+    assert!(summary.as_str().contains("private scratchpad"));
+    assert!(summary.as_str().starts_with("<summary>"));
 }
 
 #[test]
-fn decoder_should_remove_top_level_leading_scratchpad() {
+fn decoder_should_preserve_top_level_analysis_as_upstream_text() {
     let raw = format!(
         "<analysis>private scratchpad</analysis>\n{}",
         valid_summary("actual state")
     );
-    let summary = decode_summary(&raw).expect("clean summary");
+    let summary = decode_summary(&raw).expect("summary");
 
-    assert!(!summary.as_str().contains("private scratchpad"));
+    assert!(summary.as_str().contains("private scratchpad"));
 }
 
 #[test]
-fn decoder_should_keep_summary_after_unclosed_leading_scratchpad() {
+fn decoder_should_preserve_unclosed_leading_analysis() {
     let raw = format!(
         "<analysis>truncated private scratchpad\n{}",
         valid_summary("actual state")
     );
-    let summary = decode_summary(&raw).expect("clean summary");
+    let summary = decode_summary(&raw).expect("summary");
 
+    assert!(summary.as_str().contains("truncated private scratchpad"));
     assert!(summary.as_str().contains("actual state"));
 }
 
 #[test]
-fn decoder_should_neutralize_live_compaction_control_tokens_in_body() {
+fn decoder_should_preserve_compaction_control_tokens_like_sub2api() {
     let raw = valid_summary(
         "Quoted <summary> and </summary>, <analysis> and </analysis>, \
          plus <summary_request> and </summary_request>.",
     );
-    let summary = decode_summary(&raw).expect("clean summary");
+    let summary = decode_summary(&raw).expect("summary");
 
     for token in [
         "<summary>",
@@ -238,20 +249,20 @@ fn decoder_should_neutralize_live_compaction_control_tokens_in_body() {
         "<summary_request>",
         "</summary_request>",
     ] {
-        assert!(!summary.as_str().contains(token), "live token {token}");
+        assert!(summary.as_str().contains(token), "token {token}");
     }
 }
 
 #[test]
-fn decoder_should_preserve_unclosed_summary_body_without_live_token() {
+fn decoder_should_preserve_unclosed_summary_token() {
     let raw = format!(
         "<summary>\nActual state\n{}",
         "preserved implementation context ".repeat(20)
     );
-    let summary = decode_summary(&raw).expect("clean summary");
+    let summary = decode_summary(&raw).expect("summary");
 
     assert!(summary.as_str().contains("Actual state"));
-    assert!(!summary.as_str().contains("<summary>"));
+    assert!(summary.as_str().contains("<summary>"));
 }
 
 #[test]
@@ -267,17 +278,95 @@ fn decoder_should_preserve_numbered_sections_that_quote_analysis_close() {
 }
 
 #[test]
-fn decoder_should_reject_degenerate_summary() {
-    let error = decode_summary("<summary>too short</summary>").expect_err("short summary");
+fn decoder_should_accept_short_summary_like_sub2api() {
+    let summary = decode_summary("<summary>too short</summary>").expect("short summary");
 
-    assert_eq!(error, GrokCompactionDecodeError::Degenerate);
+    assert_eq!(summary, "<summary>too short</summary>");
 }
 
 #[test]
-fn decoder_should_reject_empty_cleaned_summary_as_degenerate() {
-    let error = decode_summary("<analysis>draft only</analysis>").expect_err("empty summary");
+fn decoder_should_accept_empty_summary_text() {
+    let summary = decode_summary("  \n\t ").expect("empty summary");
 
-    assert_eq!(error, GrokCompactionDecodeError::Degenerate);
+    assert!(summary.is_empty());
+}
+
+#[test]
+fn decoder_should_fall_back_to_completed_output_without_text_deltas() {
+    let mut decoder = GrokCompactionSummaryDecoder::new();
+    let wire = ProtocolWireEvent::json_with_sse_metadata(
+        "openai",
+        Some("response.completed".to_owned()),
+        json!({
+            "type": "response.completed",
+            "response": {
+                "output": [
+                    {
+                        "type": "reasoning",
+                        "encrypted_content": "real-ciphertext"
+                    },
+                    {
+                        "type": "message",
+                        "content": [
+                            {"type": "output_text", "text": "first part"},
+                            {"type": "output_text", "text": "second part"}
+                        ]
+                    }
+                ]
+            }
+        }),
+        None,
+        None,
+    )
+    .expect("terminal wire event");
+    decoder
+        .observe(&ProviderEvent::wire(wire))
+        .expect("terminal response");
+
+    assert_eq!(
+        decoder.finish().expect("summary"),
+        "first part\nsecond part"
+    );
+}
+
+#[test]
+fn decoder_should_prefer_complete_terminal_output_over_partial_deltas() {
+    let mut decoder = GrokCompactionSummaryDecoder::new();
+    decoder
+        .observe(&ProviderEvent::canonical(GatewayEvent::TextDelta(
+            TextDelta {
+                content_index: 0,
+                text: "partial summary".to_owned(),
+            },
+        )))
+        .expect("partial summary delta");
+    let wire = ProtocolWireEvent::json_with_sse_metadata(
+        "openai",
+        Some("response.completed".to_owned()),
+        json!({
+            "type": "response.completed",
+            "response": {
+                "output": [{
+                    "type": "message",
+                    "content": [{
+                        "type": "output_text",
+                        "text": "complete terminal summary"
+                    }]
+                }]
+            }
+        }),
+        None,
+        None,
+    )
+    .expect("terminal wire event");
+    decoder
+        .observe(&ProviderEvent::wire(wire))
+        .expect("terminal response");
+
+    assert_eq!(
+        decoder.finish().expect("summary"),
+        "complete terminal summary"
+    );
 }
 
 #[test]

backend/crates/providers/xai/tests/transport/fixtures/grok_compaction_request.json

@@ -26,8 +26,10 @@
       }
     }
   ],
-  "tool_choice": "auto",
-  "temperature": 1.0,
+  "tool_choice": "none",
+  "include": [
+    "reasoning.encrypted_content"
+  ],
   "store": false,
   "stream": true
 }

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

@@ -262,6 +262,8 @@ async fn inference_transport_should_classify_http_failures_without_retaining_bod
         (408, GrokInferenceTransportErrorKind::Timeout),
         (429, GrokInferenceTransportErrorKind::RateLimited),
         (500, GrokInferenceTransportErrorKind::Unavailable),
+        (504, GrokInferenceTransportErrorKind::Unavailable),
+        (529, GrokInferenceTransportErrorKind::Unavailable),
         (418, GrokInferenceTransportErrorKind::Protocol),
     ];
 
@@ -332,7 +334,7 @@ async fn inference_transport_should_expose_safe_flat_json_error_details() {
 
     assert_eq!(
         error.kind(),
-        GrokInferenceTransportErrorKind::QuotaExhausted
+        GrokInferenceTransportErrorKind::PaymentRequired
     );
     assert_eq!(
         (detail.message(), detail.code(), detail.error_type()),
@@ -395,16 +397,22 @@ async fn inference_transport_should_scope_forbidden_failures_from_safe_metadata(
         ),
         (
             json!({"error": {"code": "permission_denied", "message": "access to the chat endpoint is denied"}}),
-            GrokInferenceTransportErrorKind::ModelAccessDenied,
+            GrokInferenceTransportErrorKind::PermissionDenied,
             false,
             Some("permission_denied"),
         ),
         (
             json!({"error": {"code": "permission_denied", "message": "request rejected", "details": {"check": "SAFETY_CHECK_TYPE_INPUT"}}}),
-            GrokInferenceTransportErrorKind::SafetyRejected,
+            GrokInferenceTransportErrorKind::PermissionDenied,
             false,
             Some("permission_denied"),
         ),
+        (
+            json!({"error": {"code": "invalid_token", "message": "request blocked by policy"}}),
+            GrokInferenceTransportErrorKind::SafetyRejected,
+            false,
+            Some("invalid_token"),
+        ),
         (
             json!({"error": {"code": "policy_denied", "message": "request rejected"}}),
             GrokInferenceTransportErrorKind::PermissionDenied,
@@ -558,11 +566,70 @@ async fn inference_transport_should_classify_free_usage_429_as_account_free_quot
     );
 }
 
+#[tokio::test]
+async fn inference_transport_should_apply_sub2api_body_aware_400_failures() {
+    let cases = [
+        (
+            json!({"error": {"code": "subscription:free-usage-exhausted", "message": "free usage exhausted"}}),
+            GrokInferenceTransportErrorKind::FreeQuotaExhausted,
+            Some("subscription_free_usage_exhausted"),
+        ),
+        (
+            json!({"error": {"message": "selected model is at capacity"}}),
+            GrokInferenceTransportErrorKind::RateLimited,
+            Some("model_capacity"),
+        ),
+        (
+            json!({"error": {"message": "spending limit reached"}}),
+            GrokInferenceTransportErrorKind::PaymentRequired,
+            Some("billing_quota"),
+        ),
+        (
+            json!({"error": {"message": "empty model output: no content/tool_calls"}}),
+            GrokInferenceTransportErrorKind::Unavailable,
+            Some("empty_upstream"),
+        ),
+        (
+            json!({"error": {"message": "rate limit exceeded"}}),
+            GrokInferenceTransportErrorKind::InvalidRequest,
+            Some("rate_limit"),
+        ),
+        (
+            json!({"error": {"message": "invalid tool schema"}}),
+            GrokInferenceTransportErrorKind::InvalidRequest,
+            None,
+        ),
+    ];
+
+    for (body, expected_kind, expected_code) in cases {
+        let server = MockServer::start().await;
+        Mock::given(method("POST"))
+            .and(path("/v1/responses"))
+            .respond_with(ResponseTemplate::new(400).set_body_json(body))
+            .expect(1)
+            .mount(&server)
+            .await;
+        let origin = Url::parse(&server.uri()).expect("wiremock origin");
+        let error = inference_transport(&origin)
+            .execute(inference_request(&origin))
+            .await
+            .expect_err("400 response must be classified");
+
+        assert_eq!(error.kind(), expected_kind);
+        assert_eq!(
+            error.upstream_code().map(|code| code.as_str()),
+            expected_code
+        );
+    }
+}
+
 #[tokio::test]
 async fn inference_transport_should_bound_retry_after_to_the_safe_window() {
     for (header, expected) in [
         ("120", Some(Duration::from_secs(120))),
-        ("121", None),
+        ("121", Some(Duration::from_secs(121))),
+        ("86400", Some(Duration::from_secs(86_400))),
+        ("86401", None),
         ("0", None),
     ] {
         let server = MockServer::start().await;

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

@@ -51,6 +51,10 @@ fn encoder_should_preserve_raw_images_hosted_tools_and_unknown_fields() {
     assert_eq!(body.pointer("/model"), Some(&json!("grok-routed")));
     assert_eq!(body.pointer("/stream"), Some(&json!(true)));
     assert_eq!(body.pointer("/store"), Some(&json!(true)));
+    assert_eq!(
+        body.pointer("/include"),
+        Some(&json!(["reasoning.encrypted_content"]))
+    );
     assert_eq!(
         body.pointer("/input/0/content/1/detail"),
         Some(&json!("high"))
@@ -71,25 +75,230 @@ fn encoder_should_preserve_raw_images_hosted_tools_and_unknown_fields() {
 }
 
 #[test]
-fn encoder_should_strip_openai_only_client_metadata_before_grok_build() {
+fn encoder_should_strip_all_client_metadata_before_grok_build() {
     let request = raw_request(json!({
         "model": "client-model",
         "input": "hello",
         "client_metadata": {
             "x-openai-subagent": "review",
             "application_tag": "preserve-me"
-        }
+        },
+        "metadata": {"application_tag": "keep-this"}
     }));
 
     let encoded = GrokResponsesRequest::encode(&request, "grok-routed", &client_key())
         .expect("sanitized request");
     let body = Value::Object(encoded.body().clone());
 
-    assert_eq!(body.pointer("/client_metadata/x-openai-subagent"), None);
+    assert_eq!(body.pointer("/client_metadata"), None);
     assert_eq!(
-        body.pointer("/client_metadata/application_tag"),
-        Some(&json!("preserve-me"))
+        body.pointer("/metadata/application_tag"),
+        Some(&json!("keep-this"))
     );
+
+    let request = raw_request(json!({
+        "model": "client-model",
+        "input": "hello",
+        "client_metadata": "opaque-local-envelope"
+    }));
+    let encoded = GrokResponsesRequest::encode(&request, "grok-routed", &client_key())
+        .expect("scalar client metadata");
+    assert_eq!(encoded.body().get("client_metadata"), None);
+}
+
+#[test]
+fn encoder_should_apply_build_defaults_and_normalize_reasoning_effort() {
+    let request = raw_request(json!({
+        "model": "client-model",
+        "input": "hello",
+        "include": ["web_search_call.action.sources"],
+        "reasoning": {"effort": "XHIGH", "summary": "auto"}
+    }));
+
+    let encoded = GrokResponsesRequest::encode(&request, "grok-4.5", &client_key())
+        .expect("normalized request");
+    let body = Value::Object(encoded.body().clone());
+
+    assert_eq!(body.pointer("/store"), Some(&json!(false)));
+    assert_eq!(
+        body.pointer("/include"),
+        Some(&json!([
+            "web_search_call.action.sources",
+            "reasoning.encrypted_content"
+        ]))
+    );
+    assert_eq!(body.pointer("/reasoning/effort"), Some(&json!("high")));
+    assert_eq!(body.pointer("/reasoning/summary"), Some(&json!("auto")));
+
+    let request = raw_request(json!({
+        "model": "client-model",
+        "input": "hello",
+        "reasoning": {"effort": "max"},
+        "reasoning_effort": "extra_high",
+        "reasoningEffort": "minimal"
+    }));
+    let encoded = GrokResponsesRequest::encode(&request, "xai/grok-4.6-latest", &client_key())
+        .expect("aliased request");
+    let body = Value::Object(encoded.body().clone());
+    assert_eq!(body.pointer("/model"), Some(&json!("grok-4.6")));
+    assert_eq!(body.pointer("/reasoning/effort"), Some(&json!("high")));
+    assert_eq!(body.pointer("/reasoning_effort"), Some(&json!("high")));
+    assert_eq!(body.pointer("/reasoningEffort"), None);
+
+    let request = raw_request(json!({
+        "model": "client-model",
+        "input": "hello",
+        "reasoning": {"effort": "high", "summary": "auto"}
+    }));
+    let encoded = GrokResponsesRequest::encode(&request, "grok-composer-2.5-fast", &client_key())
+        .expect("Composer request");
+    let body = Value::Object(encoded.body().clone());
+    assert_eq!(body.pointer("/reasoning"), None);
+
+    let request = raw_request(json!({
+        "model": "client-model",
+        "input": "hello",
+        "store": null,
+        "include": null,
+        "reasoning": {"effort": "max"}
+    }));
+    let encoded = GrokResponsesRequest::encode(&request, "grok-composer-2.5-fast", &client_key())
+        .expect("empty Composer reasoning");
+    let body = Value::Object(encoded.body().clone());
+    assert_eq!(body.pointer("/store"), Some(&json!(false)));
+    assert_eq!(
+        body.pointer("/include"),
+        Some(&json!(["reasoning.encrypted_content"]))
+    );
+    assert_eq!(body.pointer("/reasoning"), None);
+
+    let request = raw_request(json!({
+        "model": "client-model",
+        "input": "hello",
+        "reasoning": {}
+    }));
+    let encoded = GrokResponsesRequest::encode(&request, "grok-composer-2.5-fast", &client_key())
+        .expect("Composer reasoning without effort");
+    assert_eq!(
+        Value::Object(encoded.body().clone()).pointer("/reasoning"),
+        None
+    );
+
+    let request = raw_request(json!({
+        "model": "client-model",
+        "input": "hello",
+        "reasoning": {"effort": "max"}
+    }));
+    let encoded = GrokResponsesRequest::encode(&request, "tenant/foo/grok-4.6", &client_key())
+        .expect("unknown provider prefix");
+    assert_eq!(
+        Value::Object(encoded.body().clone()).pointer("/reasoning/effort"),
+        None
+    );
+}
+
+#[test]
+fn encoder_should_strip_sub2api_grok_unsupported_fields() {
+    let request = raw_request(json!({
+        "model": "client-model",
+        "input": "hello",
+        "prompt_cache_retention": "24h",
+        "safety_identifier": "user-1",
+        "presence_penalty": 0.1,
+        "presencePenalty": 0.2,
+        "frequency_penalty": 0.3,
+        "frequencyPenalty": 0.4,
+        "stop": ["done"],
+        "external_web_access": true,
+        "metadata": {"external_web_access": false},
+        "tools": [{
+            "type": "function",
+            "name": "lookup",
+            "external_web_access": true,
+            "parameters": {
+                "type": "object",
+                "properties": {"q": {"type": "string", "external_web_access": true}}
+            }
+        }]
+    }));
+
+    let encoded = GrokResponsesRequest::encode(&request, "grok-latest", &client_key())
+        .expect("sanitized 4.5 request");
+    let body = Value::Object(encoded.body().clone());
+    for pointer in [
+        "/prompt_cache_retention",
+        "/safety_identifier",
+        "/presence_penalty",
+        "/presencePenalty",
+        "/frequency_penalty",
+        "/frequencyPenalty",
+        "/stop",
+        "/external_web_access",
+        "/metadata/external_web_access",
+        "/tools/0/external_web_access",
+        "/tools/0/parameters/properties/q/external_web_access",
+    ] {
+        assert_eq!(body.pointer(pointer), None, "field survived at {pointer}");
+    }
+
+    let request = raw_request(json!({
+        "model": "client-model",
+        "input": "hello",
+        "logprobs": true,
+        "top_logprobs": 5
+    }));
+    let encoded = GrokResponsesRequest::encode(&request, "grok-4.20-reasoning", &client_key())
+        .expect("sanitized 4.20 request");
+    let body = Value::Object(encoded.body().clone());
+    assert_eq!(
+        body.pointer("/model"),
+        Some(&json!("grok-4.20-0309-reasoning"))
+    );
+    assert_eq!(body.pointer("/logprobs"), None);
+    assert_eq!(body.pointer("/top_logprobs"), None);
+}
+
+#[test]
+fn external_web_access_false_should_strip_only_the_unsupported_field() {
+    let request = raw_request(json!({
+        "model": "client-model",
+        "input": "hello",
+        "tools": [{
+            "type": "web_search",
+            "external_web_access": false
+        }],
+        "tool_choice": {"type": "web_search"}
+    }));
+
+    let encoded = GrokResponsesRequest::encode(&request, "grok-4.5", &client_key())
+        .expect("web search survives field sanitization");
+    let body = Value::Object(encoded.body().clone());
+
+    assert_eq!(
+        body.pointer("/tools/0"),
+        Some(&json!({"type": "web_search"}))
+    );
+    assert_eq!(
+        body.pointer("/tool_choice"),
+        Some(&json!({"type": "web_search"}))
+    );
+}
+
+#[test]
+fn encoder_should_reject_non_string_build_include_values() {
+    for include in [json!("reasoning.encrypted_content"), json!([1])] {
+        let request = raw_request(json!({
+            "model": "client-model",
+            "input": "hello",
+            "include": include
+        }));
+        let error = GrokResponsesRequest::encode(&request, "grok-routed", &client_key())
+            .expect_err("invalid include");
+        assert!(matches!(
+            error,
+            GrokRequestEncodeError::InvalidRequestField { field: "include" }
+        ));
+    }
 }
 
 #[test]
@@ -326,6 +535,28 @@ fn function_parameters_should_remove_only_nullable_object_roots() {
     );
 }
 
+#[test]
+fn function_parameters_should_default_missing_or_null_schemas() {
+    let request = raw_request(json!({
+        "model": "client",
+        "input": "hello",
+        "tools": [
+            {"type": "function", "name": "lookup"},
+            {"type": "function", "name": "wait", "parameters": null}
+        ]
+    }));
+
+    let encoded = GrokResponsesRequest::encode(&request, "grok-4.5", &client_key())
+        .expect("default function schemas");
+    let body = Value::Object(encoded.body().clone());
+    for index in 0..2 {
+        assert_eq!(
+            body.pointer(&format!("/tools/{index}/parameters")),
+            Some(&json!({"type": "object", "properties": {}}))
+        );
+    }
+}
+
 #[test]
 fn function_parameters_should_keep_local_object_refs_after_removing_null() {
     let request = raw_request(json!({
@@ -411,6 +642,8 @@ fn explicit_session_should_add_x_search_after_codex_additional_tools_normalizati
     let body = Value::Object(encoded.body().clone());
 
     assert_eq!(body.pointer("/input/0/type"), Some(&json!("message")));
+    assert_eq!(body.pointer("/input/0/role"), Some(&json!("user")));
+    assert_eq!(body.pointer("/input/1"), None);
     assert_eq!(body.pointer("/tools/0/name"), Some(&json!("apply_patch")));
     assert_eq!(
         body.pointer("/tools/0/parameters/required"),
@@ -422,7 +655,55 @@ fn explicit_session_should_add_x_search_after_codex_additional_tools_normalizati
 }
 
 #[test]
-fn explicit_session_should_add_x_search_to_allowed_tools_choice() {
+fn additional_tools_should_keep_top_level_definition_and_drop_unsupported_types() {
+    let request = raw_request(json!({
+        "model": "client",
+        "tools": [
+            {
+                "type": "function",
+                "name": "existing",
+                "description": "top-level wins",
+                "parameters": {"type": "object"}
+            },
+            {"type": "image_generation", "model": "gpt-image-2"}
+        ],
+        "tool_choice": "auto",
+        "input": [
+            {
+                "type": "additional_tools",
+                "tools": [
+                    {
+                        "type": "function",
+                        "name": "existing",
+                        "description": "carrier duplicate",
+                        "parameters": {"type": "object"}
+                    },
+                    {"type": "function", "name": "wait"},
+                    {"type": "future_client_tool", "name": "drop-me"}
+                ]
+            },
+            {"type": "message", "role": "user", "content": "hello"}
+        ]
+    }));
+
+    let encoded = GrokResponsesRequest::encode(&request, "grok-4.5", &client_key())
+        .expect("promoted additional tools");
+    let body = Value::Object(encoded.body().clone());
+
+    assert_eq!(body.pointer("/input/0/role"), Some(&json!("user")));
+    assert_eq!(body.pointer("/input/1"), None);
+    assert_eq!(body.pointer("/tools/0/name"), Some(&json!("existing")));
+    assert_eq!(
+        body.pointer("/tools/0/description"),
+        Some(&json!("top-level wins"))
+    );
+    assert_eq!(body.pointer("/tools/1/name"), Some(&json!("wait")));
+    assert_eq!(body.pointer("/tools/2"), None);
+    assert_eq!(body.pointer("/tool_choice"), Some(&json!("auto")));
+}
+
+#[test]
+fn explicit_session_should_drop_xai_unsupported_allowed_tools_choice() {
     let request = raw_request(json!({
         "model": "client",
         "prompt_cache_key": "conversation-42",
@@ -443,10 +724,8 @@ fn explicit_session_should_add_x_search_to_allowed_tools_choice() {
         GrokResponsesRequest::encode(&request, "grok-4.5", &client_key()).expect("request");
     let body = Value::Object(encoded.body().clone());
 
-    assert_eq!(
-        body.pointer("/tool_choice/tools/1"),
-        Some(&json!({"type": "x_search"}))
-    );
+    assert_eq!(body.pointer("/tool_choice"), None);
+    assert_eq!(body.pointer("/tools/1"), Some(&json!({"type": "x_search"})));
 }
 
 #[test]
@@ -664,7 +943,7 @@ fn custom_apply_patch_declaration_should_use_patch_parameter() {
 }
 
 #[test]
-fn hosted_tool_choice_should_narrow_visible_tools_and_require_the_selected_kind() {
+fn hosted_tool_choice_should_preserve_the_sub2api_specific_choice() {
     let request = raw_request(json!({
         "model": "client",
         "input": "search",
@@ -679,13 +958,19 @@ fn hosted_tool_choice_should_narrow_visible_tools_and_require_the_selected_kind(
         GrokResponsesRequest::encode(&request, "grok-4.5", &client_key()).expect("hosted choice");
     let body = Value::Object(encoded.body().clone());
 
-    assert_eq!(body.pointer("/tool_choice"), Some(&json!("required")));
+    assert_eq!(
+        body.pointer("/tool_choice"),
+        Some(&json!({"type": "web_search"}))
+    );
     assert_eq!(
         body.pointer("/tools"),
-        Some(&json!([{
-            "type": "web_search",
-            "filters": {"allowed_domains": ["example.com"]}
-        }]))
+        Some(&json!([
+            {
+                "type": "web_search",
+                "filters": {"allowed_domains": ["example.com"]}
+            },
+            {"type": "x_search"}
+        ]))
     );
 }
 
@@ -927,7 +1212,6 @@ fn compaction_history_should_become_plaintext_user_continuation_in_place() {
             {"type": "message", "role": "user", "content": "before compaction"},
             {
                 "type": "compaction",
-                "id": "cmp_1",
                 "encrypted_content": "Repository state and pending work."
             },
             {"type": "message", "role": "user", "content": "continue"}
@@ -960,7 +1244,53 @@ fn compaction_history_should_become_plaintext_user_continuation_in_place() {
 }
 
 #[test]
-fn compaction_history_should_reject_missing_empty_or_non_string_plaintext() {
+fn sub2api_compaction_history_should_restore_reasoning_ciphertext_and_visible_summary() {
+    for item_type in ["compaction", "compaction_summary"] {
+        let request = raw_request(json!({
+            "model": "client",
+            "input": [
+                {
+                    "type": item_type,
+                    "id": "cmp_1",
+                    "status": "completed",
+                    "encrypted_content": "grok-encrypted-state",
+                    "summary": [
+                        {"type": "summary_text", "text": "first part"},
+                        {"type": "summary_text", "text": "second part"}
+                    ]
+                },
+                {"type": "message", "role": "user", "content": "continue"}
+            ]
+        }));
+
+        let encoded = GrokResponsesRequest::encode(&request, "grok-4.5", &client_key())
+            .expect("sub2api compaction continuation");
+        let body = Value::Object(encoded.body().clone());
+
+        assert_eq!(
+            body.pointer("/input"),
+            Some(&json!([
+                {
+                    "type": "reasoning",
+                    "summary": [],
+                    "encrypted_content": "grok-encrypted-state"
+                },
+                {
+                    "type": "message",
+                    "role": "user",
+                    "content": [{
+                        "type": "input_text",
+                        "text": "<conversation_summary>\nfirst part\nsecond part\n</conversation_summary>"
+                    }]
+                },
+                {"type": "message", "role": "user", "content": "continue"}
+            ]))
+        );
+    }
+}
+
+#[test]
+fn malformed_compaction_history_should_be_dropped_like_sub2api() {
     for encrypted_content in [
         None,
         Some(json!(null)),
@@ -978,12 +1308,9 @@ fn compaction_history_should_reject_missing_empty_or_non_string_plaintext() {
             "model": "client",
             "input": [Value::Object(compaction)]
         }));
-
-        assert_eq!(
-            GrokResponsesRequest::encode(&request, "grok-4.5", &client_key())
-                .expect_err("invalid compaction plaintext"),
-            GrokRequestEncodeError::InvalidRequestField { field: "input" }
-        );
+        let encoded = GrokResponsesRequest::encode(&request, "grok-4.5", &client_key())
+            .expect("malformed compaction item is omitted");
+        assert_eq!(encoded.body().get("input"), Some(&json!([])));
     }
 }
 
@@ -1025,12 +1352,34 @@ fn tool_search_history_should_load_returned_tools_at_the_original_turn() {
 }
 
 #[test]
-fn unsupported_or_ambiguous_tool_contracts_should_fail_before_upstream_io() {
+fn unsupported_tools_should_be_filtered_with_their_orphaned_choice() {
+    let request = raw_request(json!({
+        "model": "client",
+        "input": "x",
+        "tools": [{"type": "future_tool"}],
+        "tool_choice": {"type": "future_tool"}
+    }));
+    let encoded = GrokResponsesRequest::encode(&request, "grok-4.5", &client_key())
+        .expect("unsupported tool filter");
+    assert_eq!(encoded.body().get("tools"), None);
+    assert_eq!(encoded.body().get("tool_choice"), None);
+
+    let request = raw_request(json!({
+        "model": "client",
+        "input": "x",
+        "tools": [{"type": "function", "name": "kept"}],
+        "tool_choice": {"type": "function", "name": "missing"}
+    }));
+    let encoded = GrokResponsesRequest::encode(&request, "grok-4.5", &client_key())
+        .expect("orphaned function choice filter");
+    let body = Value::Object(encoded.body().clone());
+    assert_eq!(body.pointer("/tools/0/name"), Some(&json!("kept")));
+    assert_eq!(encoded.body().get("tool_choice"), None);
+}
+
+#[test]
+fn ambiguous_tool_contracts_should_fail_before_upstream_io() {
     for (body, field) in [
-        (
-            json!({"model": "client", "input": "x", "tools": [{"type": "future_tool"}]}),
-            "tools",
-        ),
         (
             json!({"model": "client", "input": "x", "tools": [
                 {"type": "tool_search", "execution": "client"},

deploy/README.md

@@ -47,7 +47,7 @@ UID/GID 和 mode，因此配置由当前用户持有，并只向容器组 `10001
 
 模板中的 `openai` / `xai` 只保留请求画像启动基线。OpenAI 的上游地址、WebSocket 池、额度刷新
 与 OAuth 设置，以及 xAI 的 OAuth、额度和模型目录策略，均由各自 Provider 使用代码内默认值管理；
-模板不重复列出这些默认项。运行后，OpenAI CLI/Desktop 与 xAI CLI 的官方发布检查只更新进程内
+模板不重复列出这些默认项。运行后，OpenAI Desktop 与 xAI CLI 的官方发布检查只更新进程内
 版本字段，不回写 `config.yaml`；检查失败时继续使用上一份有效画像。
 
 ## 启动

deploy/config.example.yaml

@@ -57,17 +57,17 @@ api:
 openai:
   # 上游地址、WebSocket 池、额度刷新和 OAuth 使用 Provider 内置默认值；模板只保留画像启动基线。
   wire_profile:
-    # codex_version 是 npm CLI 的启动基线，并用于模型目录与请求 UA；Provider 每 24 小时更新它。
-    # Desktop appcast 同期只更新 Desktop 版本与构建号。
+    # 版本元组来自同一个官方 macOS Desktop ZIP；Provider 每 24 小时检查 appcast，
+    # 仅在包内 Core 核验和 Redis 条件覆盖都成功后原子更新。
     originator: 'Codex Desktop'
-    codex_version: '0.146.0'
-    desktop_version: '26.727.51351'
-    desktop_build: '6119'
+    codex_version: '0.147.0-alpha.6.6'
+    desktop_version: '26.803.81509'
+    desktop_build: '6415'
     os_type: 'Mac OS'
     os_version: '15.7.1'
     arch: 'arm64'
     terminal: 'unknown'
-    verified_at: '2026-08-04T00:00:00+08:00'
+    verified_at: '2026-08-03T00:00:00+08:00'
 
 xai:
   # OAuth 端点、额度与模型目录策略由 xAI Provider 固定管理；这里只声明请求画像启动基线。

docs/architecture.md

@@ -10,8 +10,8 @@ Codex Proxy RS 是单进程、单副本部署的多 Provider AI 网关：
 - 管理面提供 `/api/admin/*` 和 Vue 静态管理端。
 - 当前 Provider 为 OpenAI 与 xAI；账号 credential 均由所属 Provider 独占解释。
 - OpenAI 行为以重构前正式实现为语义基准；xAI/Grok 行为以 `grok2api` 验证结果为基准。
-- 两个 Provider 的请求画像都由启动配置提供基线，并通过进程级共享状态向每次新请求发布一致快照。OpenAI CLI、OpenAI Desktop 与 xAI CLI 的官方版本检查会分别原子更新自己负责的版本字段。
-- OpenAI CLI 使用官方 npm 包 `@openai/codex`，OpenAI Desktop 使用官方 appcast，xAI CLI 使用官方 npm 包 `@xai-official/grok`。检查失败保留上一份成功画像；检查成功后 OAuth、catalog、quota/billing、inference 与 Dashboard 自动采用最新版本。
+- 两个 Provider 的请求画像都由启动配置提供基线，并通过进程级共享状态向每次新请求发布一致快照。OpenAI Desktop 与 xAI CLI 的官方版本检查会分别原子更新自己负责的版本字段。
+- OpenAI Desktop 使用官方 appcast，并从同一 ZIP 的内嵌 Core 做有界 Range 探测；xAI CLI 使用官方 npm 包 `@xai-official/grok`。检查失败保留上一份成功画像；检查成功后 OAuth、catalog、quota/billing、inference 与 Dashboard 自动采用最新版本。OpenAI 制品画像写入 Redis 单键并使用 TTL，避免按发布次数堆积。
 - PostgreSQL 是业务与配置事实的唯一权威存储。
 - Redis 只保存可丢失、可重建或有自然过期时间的协调状态。
 
@@ -148,7 +148,16 @@ OpenAI OAuth 导入先取得 access token（仅有 refresh token 时先交换）
 由 `accessToken` 补齐。解析器只读取 `email`、`https://api.openai.com/profile` 与
 `https://api.openai.com/auth` 中的官方 claims，不调用 `whoami`，也不使用导入文档顶层的
 `userId/accountId` 补身份。两个令牌都没有用户身份 claims 时，账号以未验证错误状态入库。首次 OAuth
-在服务端 `state`、PKCE 与官方 token exchange 成功后解析并保存同一 token set。重新授权的目标绑定
+返回与当前 Codex Desktop 一致的两层登录地址：内层使用 `auth.openai.com/oauth/authorize` 的
+官方参数顺序，外层使用 `https://chatgpt.com/codex/desktop-auth` 并开启
+`codex_streamlined_login`。`codex_app_version` 每次从共享 Desktop 画像快照读取；
+首次授权在 10 分钟 pending flow 中生成账号级 `installation_id`，换码成功后把同一个值写入 credential。
+`source_surface_stable_id` 与 `codex_origin_stable_id` 由该 ID 经过带版本域的 SHA-256 确定性派生，并映射为
+UUIDv4；重新授权读取既有 `installation_id`，因此得到相同的 surface ID。系统不额外持久化 surface ID，
+也不再从全局 HMAC 身份锚点派生；它不等同于 DeviceCheck attestation。为了对齐 macOS Desktop
+交给外部浏览器的最终 `chatgpt.com` URL，Provider 同样携带 `no_universal_links=1` 浏览器路由参数。
+服务端随后仍以 `state`、PKCE
+与官方 token exchange 成功后解析并保存同一 token set。重新授权的目标绑定
 只保存目标账号 ID，不接受或冻结客户端提供的 credential revision；complete 阶段重新读取目标账号及其当前 revision，并以当前 revision 做最终 CAS。
 重新授权和手工或后台 RT refresh 只轮换目标账号的 token，保留既有账号资料与 OAuth principal。
 
@@ -262,7 +271,7 @@ HTTP serve/drain 与 worker 运行时都在 `gateway-host`：Host 负责注册
   Cron 游标、恢复中间状态、完成删除收敛、领取并执行一个 queued 任务、执行一小批保留清理；长时间
   `pg_dump` 与 multipart 上传由 daemon 自身持有，Host 只负责 panic 后重启、健康与关闭。它不使用
   Redis leader lease（单副本无竞争），只依赖 PostgreSQL 部分唯一约束防御并发 API/调度创建。
-- Provider：credential refresh、quota/catalog 健康，以及模型 etag 与官方版本 release 检查（OpenAI CLI、OpenAI Desktop、xAI CLI）。
+- Provider：credential refresh、quota/catalog 健康，以及模型 etag 与官方版本 release 检查（OpenAI Desktop、xAI CLI）。
 
 带 lease 的周期任务在每个周期开始向 Redis 申请一次 leader lease，周期内续租，周期结束即释放。这只是单周期互斥，不是跨周期 leader 选举：各副本独立计时，N 副本部署下同一任务的实际执行频率最坏可达单副本的 N 倍。
 

frontend/src/components/base/BaseButton.vue

@@ -10,6 +10,7 @@ const props = withDefaults(
     variant?: ButtonVariant
     size?: ButtonSize
     loading?: boolean
+    spinIconOnLoading?: boolean
     disabled?: boolean
     active?: boolean
     iconOnly?: boolean
@@ -21,6 +22,7 @@ const props = withDefaults(
     variant: 'primary',
     size: 'default',
     loading: false,
+    spinIconOnLoading: false,
     disabled: false,
     active: false,
     iconOnly: false,
@@ -80,6 +82,7 @@ const loadingIndicatorStyle = computed(() => {
 const classes = computed(() => [
   'inline-flex items-center justify-center border-0 font-bold leading-[1.15] transition-[background-color,box-shadow,color,opacity,transform] duration-150 cursor-pointer outline-none motion-safe:active:translate-y-px motion-safe:active:scale-[0.985]',
   props.iconOnly ? iconOnlySizeClasses[props.size] : sizeClasses[props.size],
+  props.iconOnly ? 'shrink-0' : undefined,
   props.iconOnly ? '' : 'rounded-(--cp-button-radius-base)',
   'focus-visible:ring-2 focus-visible:ring-(--cp-info-border) focus-visible:ring-offset-2 focus-visible:ring-offset-(--cp-bg-surface)',
   'disabled:cursor-not-allowed disabled:transform-none disabled:bg-(--cp-disabled-bg) disabled:text-(--cp-disabled-text) disabled:shadow-none',
@@ -124,7 +127,7 @@ const loadingIconClasses = computed(() => [
     :aria-busy="loading"
   >
     <span
-      v-if="loading"
+      v-if="loading && !spinIconOnLoading"
       :class="loadingIconClasses"
       :style="loadingIndicatorStyle"
       aria-hidden="true"
@@ -140,7 +143,7 @@ const loadingIconClasses = computed(() => [
       <slot />
     </span>
     <span
-      v-if="iconOnly && !$slots.icon && !loading"
+      v-if="iconOnly && !$slots.icon && (!loading || spinIconOnLoading)"
       :class="loadingIconClasses"
       aria-hidden="true"
     >

frontend/src/composables/usePagedQuery.ts

@@ -27,10 +27,12 @@ export function usePagedQuery<Result extends PageResult>(options: {
   const error = shallowRef('')
   let requestSequence = 0
 
-  async function execute() {
+  async function execute(execution: { silent?: boolean } = {}) {
     const requestId = ++requestSequence
-    loading.value = true
-    error.value = ''
+    if (!execution.silent) {
+      loading.value = true
+      error.value = ''
+    }
 
     try {
       const result = await options.load({
@@ -42,7 +44,7 @@ export function usePagedQuery<Result extends PageResult>(options: {
 
       if (result.items.length === 0 && result.page.total > 0 && result.page.page > 1) {
         page.value = clamp(result.page.totalPages, 1, Number.POSITIVE_INFINITY)
-        return execute()
+        return execute(execution)
       }
 
       items.value = result.items
@@ -55,8 +57,10 @@ export function usePagedQuery<Result extends PageResult>(options: {
     catch (cause: unknown) {
       if (requestId !== requestSequence)
         return false
-      error.value = errorMessage(cause, '加载失败')
-      options.onError?.(cause)
+      if (!execution.silent) {
+        error.value = errorMessage(cause, '加载失败')
+        options.onError?.(cause)
+      }
       return false
     }
     finally {

frontend/src/layout/components/SystemUpdateModal.vue

@@ -91,8 +91,16 @@ const statusView = computed(() => {
       iconClass: 'text-(--cp-success)',
     }
   }
+  if (updateInfo.value) {
+    return {
+      label: '已是最新',
+      icon: CheckCircle2,
+      badge: 'bg-(--cp-success-bg) text-(--cp-success-text)',
+      iconClass: 'text-(--cp-success)',
+    }
+  }
   return {
-    label: updateInfo.value ? '已是最新' : '未检查',
+    label: '未检查',
     icon: CheckCircle2,
     badge: 'bg-(--cp-bg-muted) text-(--cp-text-secondary)',
     iconClass: 'text-(--cp-text-muted)',
@@ -318,6 +326,7 @@ watch(
   <BaseModal
     v-model="open"
     title="系统更新"
+    description="检查版本、查看发布说明并执行在线更新"
     variant="success"
     width="820px"
     body-max-height="72dvh"

frontend/src/views/account-groups/constants.ts

@@ -66,8 +66,8 @@ export const accountGroupColumns = [
   {
     key: 'actions',
     label: '操作',
-    width: '100px',
-    minWidth: '100px',
+    width: '136px',
+    minWidth: '136px',
     cellClass,
   },
 ] satisfies BaseTableColumn<AccountGroup>[]

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

@@ -116,13 +116,13 @@ const modalTitle = computed(() => {
 const oauthPanelTitle = computed(() => {
   if (props.reauthorizing)
     return accountName.value
-  return isXai.value ? '使用 xAI OAuth 完成账号接入' : '使用 OpenAI OAuth 完成账号接入'
+  return isXai.value ? 'xAI OAuth 授权' : 'OpenAI OAuth 授权'
 })
 
 const oauthPanelDescription = computed(() => {
-  if (props.reauthorizing)
-    return '生成新的授权链接，完成后粘贴回调地址更新账号凭据'
-  return '复制授权链接到浏览器打开，完成后把回调地址粘贴到下方即可导入'
+  return isXai.value
+    ? '生成并打开授权链接 → 完成浏览器授权 → 粘贴回调地址、查询字符串或授权码'
+    : '生成并打开授权链接 → 完成浏览器授权 → 粘贴回调地址'
 })
 
 const canGenerateOauth = computed(() =>
@@ -145,20 +145,23 @@ const canSubmit = computed(() => {
 const description = computed<string | undefined>(() => {
   if (isChoosingProvider.value)
     return undefined
-  if (props.reauthorizing)
-    return '完成授权后粘贴回调地址，系统会更新账号凭据'
+  if (props.reauthorizing) {
+    return isXai.value
+      ? '完成新的 xAI 授权并替换此账号凭据'
+      : '完成新的 OpenAI 授权并替换此账号凭据'
+  }
   if (isBatch.value)
-    return '导入 CPR 多平台账号包，系统会按文档中的平台自动分流'
+    return '粘贴或上传 CPR 账号包，一次导入多个平台账号'
   if (isXai.value) {
     return mode.value === 'oauth'
-      ? '复制 xAI 授权链接，完成后粘贴回调地址，不使用 xAI API Key'
-      : '导入 xAI 账号文件，已存在账号会更新'
+      ? '通过浏览器授权导入 xAI 账号'
+      : '粘贴或上传 xAI 账号文件，匹配已有账号时更新凭据'
   }
   if (mode.value === 'oauth')
-    return '复制 OpenAI 授权链接，完成后粘贴回调地址，系统会自动写入或更新账号'
+    return '通过浏览器授权导入 OpenAI 账号'
   if (mode.value === 'agent_identity')
-    return '导入 Agent 身份文件，系统会按身份信息写入或更新账号'
-  return '导入 CPR、Sub2API 或 CPA 账号文件，已存在账号会更新'
+    return '粘贴或上传 Agent 身份文件，匹配已有账号时更新凭据'
+  return '粘贴或上传 CPR、Sub2API 或 CPA 账号文件，匹配已有账号时更新凭据'
 })
 
 function selectProvider(value: 'openai' | 'xai' | 'batch') {
@@ -275,12 +278,12 @@ async function copyText(value: string, successText: string) {
         </BaseForm>
 
         <BaseForm>
-          <BaseFormItem label="回调地址" required>
+          <BaseFormItem :label="isXai ? '回调地址或授权码' : '回调地址'" required>
             <BaseTextarea
               v-model="oauthCallback"
-              aria-label="回调地址"
+              :aria-label="isXai ? '回调地址或授权码' : '回调地址'"
               size="sm"
-              :placeholder="isXai ? 'http://127.0.0.1:56121/callback?code=...&state=...' : 'http://localhost:1455/auth/callback?code=...&state=...'"
+              :placeholder="isXai ? '回调地址、?code=...&state=... 或授权码' : 'http://localhost:1455/auth/callback?code=...&state=...'"
               :disabled="saving"
             />
           </BaseFormItem>

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

@@ -45,6 +45,7 @@ const quotaWindows = computed(() => orderedPanelQuotaWindows(props.account.quota
         variant="ghost"
         size="sm"
         title="刷新额度"
+        spin-icon-on-loading
         :loading="refreshing"
         :disabled="refreshing"
         @click="emit('refreshQuota', account.id)"

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

@@ -1,5 +1,6 @@
 <script setup lang="ts">
 import type { AccountRow } from '../constants'
+import { modelSuccessRateTextClass } from '../constants'
 
 defineProps<{
   account: AccountRow
@@ -62,7 +63,7 @@ defineProps<{
       </div>
 
       <div
-        class="grid grid-cols-[1.5fr_0.7fr_0.8fr_1fr_1fr_1fr_1fr_1fr_1.4fr] gap-3 pb-2 text-[11px] font-heavy text-(--cp-text-secondary) shadow-[inset_0_-1px_0_var(--cp-divider-subtle)]"
+        class="grid grid-cols-[1.5fr_0.7fr_0.8fr_1fr_1fr_1fr_1fr_1fr_1.2fr] gap-3 pb-2 text-[11px] font-heavy text-(--cp-text-secondary) shadow-[inset_0_-1px_0_var(--cp-divider-subtle)]"
       >
         <span>模型</span>
         <span>调用</span>
@@ -84,11 +85,13 @@ defineProps<{
         <div
           v-for="model in account.usage.models"
           :key="model.model"
-          class="grid grid-cols-[1.5fr_0.7fr_0.8fr_1fr_1fr_1fr_1fr_1fr_1.4fr] gap-3 pt-3 text-[12px] font-emphasis text-(--cp-text-primary)"
+          class="grid grid-cols-[1.5fr_0.7fr_0.8fr_1fr_1fr_1fr_1fr_1fr_1.2fr] gap-3 pt-3 text-[12px] font-emphasis text-(--cp-text-primary)"
         >
           <span class="truncate">{{ model.model }}</span>
           <span>{{ model.requestCountDisplay }}</span>
-          <span class="text-(--cp-warning-text)">{{ model.successRateDisplay }}</span>
+          <span :class="modelSuccessRateTextClass(model.successRate)">
+            {{ model.successRateDisplay }}
+          </span>
           <span>{{ model.inputTokensDisplay }}</span>
           <span>{{ model.outputTokensDisplay }}</span>
           <span>{{ model.cachedTokensDisplay }}</span>

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

@@ -103,7 +103,7 @@ export function useAccountConnectionTest(options: { reload: () => Promise<unknow
     if (connectionTestStatus.value === 'running') {
       return {
         label: '正在测试',
-        description: '正在发送一条真实 Responses 流式请求',
+        description: '正在向所选模型发送请求并接收流式响应',
         icon: Clock3,
         badge: 'bg-(--cp-info-bg) text-(--cp-info-text)',
         iconClass: 'text-(--cp-info)',
@@ -112,7 +112,7 @@ export function useAccountConnectionTest(options: { reload: () => Promise<unknow
     if (connectionTestStatus.value === 'success') {
       return {
         label: '连接正常',
-        description: '账号凭据可用，已完成 Responses 流式验证',
+        description: '请求已完成，可在下方查看模型、耗时和事件轨迹',
         icon: CheckCircle2,
         badge: 'bg-(--cp-success-bg) text-(--cp-success-text)',
         iconClass: 'text-(--cp-success)',
@@ -121,7 +121,7 @@ export function useAccountConnectionTest(options: { reload: () => Promise<unknow
     if (connectionTestStatus.value === 'error') {
       return {
         label: '测试失败',
-        description: '真实请求未完成，优先检查令牌状态、账号权限或上游网络',
+        description: '请求未完成，请在下方查看上游错误与事件轨迹',
         icon: XCircle,
         badge: 'bg-(--cp-danger-bg) text-(--cp-danger-text)',
         iconClass: 'text-(--cp-danger)',
@@ -129,7 +129,7 @@ export function useAccountConnectionTest(options: { reload: () => Promise<unknow
     }
     return {
       label: '准备测试',
-      description: '点击开始后发送一条轻量 Responses 流式请求',
+      description: '选择模型后，点击“开始测试”发送真实请求',
       icon: Wifi,
       badge: 'bg-(--cp-bg-subtle) text-(--cp-text-secondary)',
       iconClass: 'text-(--cp-text-muted)',

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

@@ -98,8 +98,11 @@ export function useAccountOnboarding(options: {
           throw new Error('请先生成授权链接')
 
         const callbackUrl = createForm.value.oauthCallback.trim()
-        if (!callbackUrl)
-          throw new Error('请粘贴 OAuth 回调地址')
+        if (!callbackUrl) {
+          throw new Error(createForm.value.provider === 'xai'
+            ? '请粘贴 OAuth 回调地址、含 code 和 state 的查询字符串或授权码'
+            : '请粘贴 OAuth 回调地址')
+        }
         await completeAccountOAuth({
           provider: createForm.value.provider,
           flowId: createForm.value.oauthFlowId,
@@ -186,11 +189,11 @@ export function useAccountOnboarding(options: {
     }
 
     if (importedCount === 0) {
-      throw new Error(failures.length > 0 ? `批量导入失败：${failures.join('；')}` : '批量文件没有可导入的账号')
+      throw new Error(failures.length > 0 ? `批量导入失败：${failures.join('、')}` : '批量文件没有可导入的账号')
     }
 
     if (failures.length > 0) {
-      return `已导入 ${importedCount} 个账号，${failures.length} 个文档失败：${failures.join('；')}`
+      return `已导入 ${importedCount} 个账号，${failures.length} 个文档失败：${failures.join('、')}`
     }
     return `已导入 ${importedCount} 个账号`
   }

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

@@ -141,6 +141,7 @@ export function useAccountsQuery() {
     loading: query.loading,
     accounts: query.items,
     loadAccounts: query.execute,
+    refreshAccountsSilently: () => query.execute({ silent: true }),
     searchQuery,
     providerQuery,
     statusQuery,

frontend/src/views/accounts/constants.ts

@@ -121,8 +121,8 @@ export const accountColumns = [
   {
     key: 'actions',
     label: '操作',
-    width: '120px',
-    minWidth: '120px',
+    width: '136px',
+    minWidth: '136px',
   },
 ] satisfies BaseTableColumn<AccountRow>[]
 
@@ -211,6 +211,18 @@ export function quotaWindowPercentTextClass(window: AccountQuotaWindow) {
   return 'text-(--cp-success-text)'
 }
 
+export function modelSuccessRateTextClass(successRate: number | null) {
+  if (successRate === null)
+    return 'text-(--cp-text-muted)'
+  if (successRate >= 99.5)
+    return 'text-(--cp-success-text)'
+  if (successRate >= 98)
+    return 'text-(--cp-normal-text)'
+  if (successRate >= 95)
+    return 'text-(--cp-warning-text)'
+  return 'text-(--cp-danger-text)'
+}
+
 function compareQuotaWindows(left: AccountQuotaWindow, right: AccountQuotaWindow) {
   const groupDifference = groupOrder(left) - groupOrder(right)
   if (groupDifference !== 0)

frontend/src/views/accounts/index.vue

@@ -37,6 +37,7 @@ const {
   loading,
   accounts,
   loadAccounts,
+  refreshAccountsSilently,
   searchQuery,
   providerQuery,
   statusQuery,
@@ -108,7 +109,7 @@ const {
   openConnectionTest,
   handleRefreshConnectionTestModels,
   handleTestConnection,
-} = useAccountConnectionTest({ reload: loadAccounts })
+} = useAccountConnectionTest({ reload: refreshAccountsSilently })
 
 const {
   expandedAccountIds,
@@ -262,9 +263,9 @@ const {
           </template>
 
           <template #groups="{ row }">
-            <AccountGroupMarks
-              :groups="row.groups"
-            />
+            <div class="flex w-full justify-center">
+              <AccountGroupMarks :groups="row.groups" />
+            </div>
           </template>
 
           <template #lastUsedAt="{ row }">

frontend/src/views/accounts/utils/schedulingForm.ts

@@ -21,7 +21,7 @@ export function parseAccountSchedulingForm(
       || parsedLimit < 1
       || parsedLimit > MAX_ACCOUNT_CONCURRENCY)
   ) {
-    return { valid: false, message: `并发限制必须留空或填写 1 到 ${MAX_ACCOUNT_CONCURRENCY} 的整数` }
+    return { valid: false, message: `并发限制必须留空，或为 1 到 ${MAX_ACCOUNT_CONCURRENCY} 的整数` }
   }
 
   const parsedWeight = Number(weight.trim())

frontend/src/views/api-keys/components/ApiKeyCreateModal.vue

@@ -27,16 +27,13 @@ const open = defineModel<boolean>({ default: false })
 const createdOpen = defineModel<boolean>('createdOpen', { default: false })
 const form = defineModel<ApiKeyFormValue>('form', { required: true })
 const title = computed(() => props.editing ? '编辑 API Key' : '创建 API Key')
-const description = computed(() => props.editing
-  ? '调整访问范围与调度限制，分组替换原子生效。'
-  : '为调用方选择一个或多个账号分组。')
 </script>
 
 <template>
   <BaseModal
     v-model="open"
     :title="title"
-    :description="description"
+    description="设置调用方可用的账号分组、并发和每分钟请求上限"
     variant="info"
     width="680px"
     body-max-height="min(680px, calc(100dvh - 190px))"
@@ -47,7 +44,7 @@ const description = computed(() => props.editing
         <BaseInput
           v-model="form.name"
           aria-label="名称"
-          placeholder="例如：生产环境、测试调用方..."
+          placeholder="例如：生产环境"
           :disabled="saving"
         />
       </BaseFormItem>
@@ -56,7 +53,7 @@ const description = computed(() => props.editing
         <BaseInput
           v-model="form.label"
           aria-label="标签（可选）"
-          placeholder="备注信息..."
+          placeholder="例如：后端服务"
           :disabled="saving"
         />
       </BaseFormItem>
@@ -108,7 +105,7 @@ const description = computed(() => props.editing
   <BaseModal
     v-model="createdOpen"
     title="API Key 已创建"
-    description="可以立即复制或导入 CCSwitch，之后也可从密钥列表再次操作"
+    description="复制密钥，或直接导入 CCSwitch"
     variant="success"
     width="540px"
   >

frontend/src/views/api-keys/components/ApiKeyScopeCell.vue

@@ -9,8 +9,8 @@ defineProps<{
 </script>
 
 <template>
-  <div class="grid gap-1.5">
-    <div v-if="apiKey.routingScope === 'all'" class="flex items-center gap-2">
+  <div class="grid w-full justify-items-center gap-1.5">
+    <div v-if="apiKey.routingScope === 'all'" class="flex items-center justify-center gap-2">
       <span class="inline-flex h-6 items-center rounded-lg bg-(--cp-warning-bg) px-2 text-[11px] font-bold text-(--cp-warning-text)">
         全部账号权限
       </span>

frontend/src/views/api-keys/components/ApiKeyUseModal.vue

@@ -38,13 +38,13 @@ const authPath = computed(() =>
   activePlatform.value === 'windows' ? '%userprofile%\\.codex\\auth.json' : '~/.codex/auth.json',
 )
 const codexAuthJson = computed(() => JSON.stringify({ OPENAI_API_KEY: keyValue.value }, null, 2))
-const defaultModel = 'gpt-5.5'
+const defaultModel = 'gpt-5.6-terra'
 
 const codexConfigToml = computed(
   () => `model_provider = "OpenAI"
 model = "${defaultModel}"
 review_model = "${defaultModel}"
-model_reasoning_effort = "xhigh"
+model_reasoning_effort = "max"
 service_tier = "default"
 disable_response_storage = true
 network_access = "enabled"
@@ -70,7 +70,7 @@ const visibleFiles = computed(() => [
   <BaseModal
     v-model="open"
     title="使用密钥"
-    description="将以下配置文件添加到 Codex CLI 配置目录中"
+    description="将下方内容分别保存到显示的 Codex CLI 配置文件"
     width="760px"
   >
     <div class="flex flex-col gap-5">

frontend/src/views/api-keys/constants.ts

@@ -43,8 +43,8 @@ export const apiKeyColumns = [
   {
     key: 'actions',
     label: '操作',
-    width: '172px',
-    minWidth: '172px',
+    width: '208px',
+    minWidth: '208px',
     headerClass: relaxedActionCellClass,
     cellClass: relaxedActionCellClass,
   },

frontend/src/views/dashboard/components/AccountOverviewCard.vue

@@ -1,6 +1,6 @@
 <script setup lang="ts">
 import type { dashboardSnapshotView, MetricTone } from '../composables/presenter'
-import { CircleCheck, CircleOff, RefreshCw, ShieldAlert, TriangleAlert } from '@lucide/vue'
+import { CircleCheck, RefreshCw, ShieldAlert, TriangleAlert } from '@lucide/vue'
 import { clamp } from 'es-toolkit'
 
 import { computed } from 'vue'
@@ -106,8 +106,7 @@ const statusRows = computed(() => {
       { label: '正常', description: '暂无账号池观测', value: '—', tone: 'success', icon: CircleCheck },
       { label: '配额耗尽', description: '暂无账号池观测', value: '—', tone: 'warning', icon: TriangleAlert },
       { label: '限流中', description: '暂无账号池观测', value: '—', tone: 'normal', icon: RefreshCw },
-      { label: '已停用', description: '暂无账号池观测', value: '—', tone: 'normal', icon: CircleOff },
-      { label: '错误', description: '暂无账号池观测', value: '—', tone: 'danger', icon: ShieldAlert },
+      { label: '待处理', description: '停用 — · 错误 —', value: '—', tone: 'danger', icon: ShieldAlert },
     ]
   }
 
@@ -134,16 +133,9 @@ const statusRows = computed(() => {
       icon: RefreshCw,
     },
     {
-      label: '已停用',
-      description: '账号已停用',
-      value: String(counts.disabled),
-      tone: 'normal',
-      icon: CircleOff,
-    },
-    {
-      label: '错误',
-      description: '账号状态异常',
-      value: String(counts.error),
+      label: '待处理',
+      description: `停用 ${counts.disabled} · 错误 ${counts.error}`,
+      value: String(counts.disabled + counts.error),
       tone: 'danger',
       icon: ShieldAlert,
     },
@@ -165,8 +157,7 @@ const statusBars = computed(() => {
     { pct: (counts.normal / counts.total) * 100, cls: 'bg-(--cp-success)' },
     { pct: (counts.quotaExhausted / counts.total) * 100, cls: 'bg-(--cp-warning)' },
     { pct: (counts.rateLimited / counts.total) * 100, cls: 'bg-(--cp-normal)' },
-    { pct: (counts.disabled / counts.total) * 100, cls: 'bg-(--cp-text-secondary)' },
-    { pct: (counts.error / counts.total) * 100, cls: 'bg-(--cp-danger)' },
+    { pct: ((counts.disabled + counts.error) / counts.total) * 100, cls: 'bg-(--cp-danger)' },
   ].filter(b => b.pct > 0)
 })
 </script>
@@ -333,11 +324,11 @@ const statusBars = computed(() => {
           </div>
         </div>
 
-        <div class="mt-5 grid h-67 w-full grid-rows-5 gap-2">
+        <div class="mt-6.5 grid h-65.5 w-full gap-2.5">
           <div
             v-for="row in statusRows"
             :key="row.label"
-            class="grid min-h-0 grid-cols-[28px_14px_minmax(0,1fr)_76px] items-center rounded-[14px] bg-(--cp-bg-subtle) px-3.5"
+            class="grid h-14.5 grid-cols-[28px_14px_minmax(0,1fr)_76px] items-center rounded-[14px] bg-(--cp-bg-subtle) px-3.5"
           >
             <span
               class="inline-flex size-7 items-center justify-center rounded-[9px]"

frontend/src/views/dashboard/components/WireProfileCard.vue

@@ -50,6 +50,18 @@ const profile = computed(() =>
   props.profiles.find(item => item.provider === activeProvider.value) ?? props.profiles[0] ?? null,
 )
 
+const versionParts = computed(() => {
+  const version = profile.value?.version ?? ''
+  const separator = version.indexOf('-')
+  if (separator <= 0 || separator === version.length - 1)
+    return { release: version, prerelease: '' }
+
+  return {
+    release: version.slice(0, separator),
+    prerelease: version.slice(separator + 1),
+  }
+})
+
 const releaseLabel = computed(() => {
   const release = profile.value?.release
   if (!release?.latestVersion)
@@ -221,10 +233,20 @@ function providerLabel(provider: string) {
           <div class="grid min-h-0 content-center">
             <div class="flex min-w-0 flex-wrap items-baseline gap-x-2.5 gap-y-1.5">
               <strong
-                class="block max-w-full wrap-break-word font-mono text-[27px] leading-[1.05] font-heavy tabular-nums text-(--cp-text-primary)"
+                :aria-label="profile.version"
+                class="inline-flex max-w-full min-w-0 items-baseline gap-2 font-mono leading-none tabular-nums"
                 :title="profile.version"
               >
-                {{ profile.version }}
+                <span class="wrap-break-word text-[27px] leading-[1.05] font-heavy text-(--cp-text-primary)">
+                  {{ versionParts.release }}
+                </span>
+                <span
+                  v-if="versionParts.prerelease"
+                  aria-hidden="true"
+                  class="truncate text-[12px] font-bold text-(--cp-text-secondary)"
+                >
+                  {{ versionParts.prerelease }}
+                </span>
               </strong>
               <span
                 v-if="profile.build"

frontend/src/views/dashboard/index.vue

@@ -46,6 +46,7 @@ const {
           class="text-(--cp-normal)"
           size="md"
           label="刷新概览"
+          spin-icon-on-loading
           :loading="loading || refreshing"
           :disabled="loading || refreshing"
           @click="refresh"

frontend/src/views/settings/components/backup/BackupRecordsCard.vue

@@ -45,7 +45,7 @@ const columns: BaseTableColumn<BackupRecord>[] = [
   { key: 'expiresAt', label: '过期时间', width: 170 },
   { key: 'trigger', label: '触发方式', width: 100 },
   { key: 'startedAt', label: '开始时间', width: 170 },
-  { key: 'actions', label: '操作', width: 96, fixed: 'right' },
+  { key: 'actions', label: '操作', width: 112, minWidth: 112, fixed: 'right' },
 ]
 
 function triggerLabel(value: string): string {

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

@@ -69,6 +69,7 @@ function showDetail(record: Awaited<ReturnType<typeof getOpsErrors>>['items'][nu
           variant="ghost"
           size="md"
           label="刷新错误明细"
+          spin-icon-on-loading
           :loading="refreshing"
           :disabled="loading || refreshing"
           @click="refresh"

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

@@ -139,7 +139,6 @@ function diagnosticNameDisplay(name: string) {
         compact
         row-key="key"
         empty-text="暂无诊断数据"
-        max-height="230px"
         min-width="640px"
       >
         <template #header-nameDisplay>

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

@@ -32,6 +32,7 @@ const search = defineModel<string>('search', { required: true })
         variant="ghost"
         size="md"
         label="刷新使用记录"
+        spin-icon-on-loading
         :loading="refreshing"
         :disabled="loading || refreshing"
         @click="emit('refresh')"

release/version.yaml

@@ -1 +1 @@
-version: "3.2.1"
+version: "3.2.5"

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

@@ -35,7 +35,7 @@
 - 换号只由 Core 在下游 commit 前按重放安全性决定。
 - fallback 只允许同一 Provider kind 内的账号；不跨 Provider，也不存在 Provider Instance 层。
 - continuation 升级链由 Core 冻结：native（原生 handle 绑定原账号）→ replay owner（原账号完整 transcript 重放）→ replay any（transcript 已可携带，允许换账号）；两个 Provider 都实现全链，xAI 的 transcript 由不透明 session state 承载，跨账号重放前做脱敏投影。
-- Provider wire profile 以配置为启动基线，并由共享运行时状态统一发布。OpenAI CLI 读取 `@openai/codex`、Desktop 读取官方 appcast、xAI CLI 读取 `@xai-official/grok`；发现新版本后自动更新对应版本字段，所有消费边界不得维护独立常量。
+- Provider wire profile 以配置为启动基线，并由共享运行时状态统一发布。OpenAI Desktop 读取官方 appcast，并从同一 ZIP 的内嵌 Core 做有界 Range 探测；xAI CLI 读取 `@xai-official/grok`。发现新版本后自动更新对应版本字段，所有消费边界不得维护独立常量。
 - `downstream_committed_at` 是不可撤回交付承诺，不是首字节已经写达的证明。
 
 ## 存储