diff --git a/crates/tw-gateway/src/glm.rs b/crates/tw-gateway/src/glm.rs
index d03cf387..6baa79c7 100644
--- a/crates/tw-gateway/src/glm.rs
+++ b/crates/tw-gateway/src/glm.rs
@@ -7,7 +7,8 @@
 //! 问的纪律：
 //! - **界面来要时问**，60 秒内的多次合成一次；**有请求去这一家时顺手问**，5 分钟最多一次。
 //! - 失败了退避（30 秒、60 秒、120 秒、300 秒），不跟着界面的每一次刷新重试。
-//! - 这把 key 没有开通套餐：记成「没有额度数据」，隔一小时才再问。
+//! - 这把 key 没有开通套餐：记成「没有额度数据」，隔一小时才再问。刚才还有额度的 key
+//!   要连着说两次才算（一次临时的 500 和「没有套餐」长得一样）。
 //!
 //! 额度用完时模型请求回 429，业务码在 body 里（[`exhausted`]）。**那一刻就记成用完**，
 //! 不等下一次问额度。
@@ -324,6 +325,8 @@ struct Slot {
     quiet_until: u64,
     /// 正在问。**同一时刻只问一次**
     running: bool,
+    /// 上一次问来的是额度
+    had_quota: bool,
 }
 
 impl Slot {
@@ -339,10 +342,20 @@ impl Slot {
             .is_none_or(|t| now_ms.saturating_sub(t) >= every)
     }
 
-    fn settle(&mut self, answer: &Answer, now_ms: u64) {
+    /// 记下这一次的结果，交回该按哪个结论办。
+    ///
+    /// **刚才还有额度的 key 头一次说「没有套餐」，当成一次失败**：「没有套餐」认的是
+    /// 业务码 500，而一次临时的 500 长得一模一样。照「没有套餐」办要清掉额度、一小时
+    /// 不再问；当成失败只是按退避过一会儿再问。连着第二次还这么说，才是真没有了
+    fn settle(&mut self, answer: Answer, now_ms: u64) -> Answer {
+        let answer = match answer {
+            Answer::NoPlan if self.had_quota => Answer::Failed,
+            a => a,
+        };
+        self.had_quota = matches!(answer, Answer::Quota(_));
         self.running = false;
         self.asked_at = Some(now_ms);
-        match answer {
+        match &answer {
             Answer::Quota(_) => {
                 self.failures = 0;
                 self.quiet_until = 0;
@@ -357,6 +370,7 @@ impl Slot {
                 self.quiet_until = now_ms + BACKOFF_MS[i];
             }
         }
+        answer
     }
 }
 
@@ -398,14 +412,18 @@ impl Tracker {
         due
     }
 
-    /// 问完了。凭据在这中间换了的话，这个结果说的是旧的 key，不记
-    pub fn settle(&self, provider: &str, ident: &str, answer: &Answer, now_ms: u64) {
-        if let Ok(mut g) = self.slots.lock()
-            && let Some(slot) = g.get_mut(provider)
-            && slot.ident == ident
-        {
-            slot.settle(answer, now_ms);
-        }
+    /// 问完了：记下节奏，交回该按哪个结论办（见 `Slot::settle`）。凭据在这中间换了
+    /// 的话，这个结果说的是旧的 key，不记也不办，是 None
+    pub fn settle(
+        &self,
+        provider: &str,
+        ident: &str,
+        answer: Answer,
+        now_ms: u64,
+    ) -> Option<Answer> {
+        let mut g = self.slots.lock().ok()?;
+        let slot = g.get_mut(provider).filter(|s| s.ident == ident)?;
+        Some(slot.settle(answer, now_ms))
     }
 }
 
@@ -749,7 +767,7 @@ mod tests {
         assert!(t.claim("glm", "k", Why::Demand, NOW));
         // 正在问：再来的不另问
         assert!(!t.claim("glm", "k", Why::Demand, NOW + 1));
-        t.settle("glm", "k", &Answer::Quota(Quota::default()), NOW + 100);
+        t.settle("glm", "k", Answer::Quota(Quota::default()), NOW + 100);
         assert!(!t.claim("glm", "k", Why::Demand, NOW + 30_000));
         assert!(t.claim("glm", "k", Why::Demand, NOW + 100 + MINUTE_MS));
     }
@@ -758,14 +776,14 @@ mod tests {
     fn traffic_asks_at_most_every_five_minutes() {
         let t = Tracker::default();
         assert!(t.claim("glm", "k", Why::Traffic, NOW));
-        t.settle("glm", "k", &Answer::Quota(Quota::default()), NOW);
+        t.settle("glm", "k", Answer::Quota(Quota::default()), NOW);
         assert!(!t.claim("glm", "k", Why::Traffic, NOW + 4 * MINUTE_MS));
         // 界面来要的照样按一分钟算
         assert!(t.claim("glm", "k", Why::Demand, NOW + 2 * MINUTE_MS));
         t.settle(
             "glm",
             "k",
-            &Answer::Quota(Quota::default()),
+            Answer::Quota(Quota::default()),
             NOW + 2 * MINUTE_MS,
         );
         assert!(!t.claim("glm", "k", Why::Traffic, NOW + 6 * MINUTE_MS));
@@ -778,16 +796,16 @@ mod tests {
         let mut now = NOW;
         for wait in [30_000, 60_000, 120_000, 300_000, 300_000] {
             assert!(t.claim("glm", "k", Why::Demand, now), "at {}", now - NOW);
-            t.settle("glm", "k", &Answer::Failed, now);
+            t.settle("glm", "k", Answer::Failed, now);
             // 界面每分钟来要一次也不提前
             assert!(!t.claim("glm", "k", Why::Demand, now + wait - 1));
             now += wait.max(DEMAND_EVERY_MS);
         }
         // 成功一次，退避清零
         assert!(t.claim("glm", "k", Why::Demand, now));
-        t.settle("glm", "k", &Answer::Quota(Quota::default()), now);
+        t.settle("glm", "k", Answer::Quota(Quota::default()), now);
         assert!(t.claim("glm", "k", Why::Demand, now + DEMAND_EVERY_MS));
-        t.settle("glm", "k", &Answer::Rejected, now + DEMAND_EVERY_MS);
+        t.settle("glm", "k", Answer::Rejected, now + DEMAND_EVERY_MS);
         assert!(t.claim("glm", "k", Why::Demand, now + 2 * DEMAND_EVERY_MS));
     }
 
@@ -795,23 +813,45 @@ mod tests {
     fn a_key_without_a_plan_is_asked_again_only_after_an_hour_or_a_new_key() {
         let t = Tracker::default();
         assert!(t.claim("glm", "k", Why::Demand, NOW));
-        t.settle("glm", "k", &Answer::NoPlan, NOW);
+        t.settle("glm", "k", Answer::NoPlan, NOW);
         assert!(!t.claim("glm", "k", Why::Demand, NOW + 30 * MINUTE_MS));
         assert!(!t.claim("glm", "k", Why::Traffic, NOW + 30 * MINUTE_MS));
         // 换了 key：之前的结论说的是旧的那把
         assert!(t.claim("glm", "k2", Why::Demand, NOW + 30 * MINUTE_MS));
-        t.settle("glm", "k2", &Answer::NoPlan, NOW + 30 * MINUTE_MS);
+        t.settle("glm", "k2", Answer::NoPlan, NOW + 30 * MINUTE_MS);
         assert!(t.claim("glm", "k2", Why::Demand, NOW + 90 * MINUTE_MS));
     }
 
+    /// 刚才还有额度的 key 头一次说「没有套餐」：多半是一次临时的 500，当成失败，
+    /// 过一会儿再问；连着第二次才信
+    #[test]
+    fn a_key_that_had_a_plan_must_say_no_plan_twice() {
+        let t = Tracker::default();
+        let q = Answer::Quota(Quota::default());
+        assert!(t.claim("glm", "k", Why::Demand, NOW));
+        assert_eq!(t.settle("glm", "k", q.clone(), NOW), Some(q));
+        assert!(t.claim("glm", "k", Why::Demand, NOW + MINUTE_MS));
+        assert_eq!(
+            t.settle("glm", "k", Answer::NoPlan, NOW + MINUTE_MS),
+            Some(Answer::Failed)
+        );
+        let later = NOW + 2 * MINUTE_MS;
+        assert!(t.claim("glm", "k", Why::Demand, later));
+        assert_eq!(
+            t.settle("glm", "k", Answer::NoPlan, later),
+            Some(Answer::NoPlan)
+        );
+        assert!(!t.claim("glm", "k", Why::Demand, later + 30 * MINUTE_MS));
+    }
+
     #[test]
     fn an_answer_for_a_replaced_key_is_not_kept() {
         let t = Tracker::default();
         assert!(t.claim("glm", "old", Why::Demand, NOW));
         assert!(t.claim("glm", "new", Why::Demand, NOW));
-        t.settle("glm", "old", &Answer::NoPlan, NOW);
+        t.settle("glm", "old", Answer::NoPlan, NOW);
         // 新 key 还在问，旧的结论没把它记成「没有套餐」
-        t.settle("glm", "new", &Answer::Quota(Quota::default()), NOW);
+        t.settle("glm", "new", Answer::Quota(Quota::default()), NOW);
         assert!(t.claim("glm", "new", Why::Demand, NOW + DEMAND_EVERY_MS));
     }
 }
diff --git a/crates/tw-gateway/src/lib.rs b/crates/tw-gateway/src/lib.rs
index f032d52d..4457cd9a 100644
--- a/crates/tw-gateway/src/lib.rs
+++ b/crates/tw-gateway/src/lib.rs
@@ -64,6 +64,13 @@ pub use state::{AppState, Runtime};
 /// 那个上限，又不至于让一条正常的流塞满心跳
 pub const PING_EVERY: std::time::Duration = std::time::Duration::from_secs(15);
 
+/// 上游一个字节都不发（连 SSE 注释都没有）超过这么久，就不再替它补 `ping`。
+///
+/// 心跳让客户端相信连接还活着；上游的连接半开了（NAT、代理把它丢了）时，网关自己
+/// 永远等不到下一个字节，**一直补下去这个请求就永远挂着**。十分钟之后停下，客户端
+/// 自己的静默计时会断开它。正常在排队的上游会发 `: keep-alive` 这样的注释，不受影响
+pub const PING_FOR: std::time::Duration = std::time::Duration::from_secs(600);
+
 /// 请求来自谁。**如实写 ThinkWatch** —— 我们从不把自己报成别的客户端。
 pub const ORIGINATOR: &str = "thinkwatch";
 
diff --git a/crates/tw-gateway/src/server/pipeline/hop.rs b/crates/tw-gateway/src/server/pipeline/hop.rs
index aa352cc0..58bc7975 100644
--- a/crates/tw-gateway/src/server/pipeline/hop.rs
+++ b/crates/tw-gateway/src/server/pipeline/hop.rs
@@ -347,8 +347,9 @@ fn prepare(
     let client_dialect = req.api.map(|a| a.dialect());
     let target = crate::translate::plan(req.api, generates, provider.effective_protocol());
     let chatgpt = generates && provider.effective_protocol() == Some(tw_config::Protocol::Chatgpt);
-    // DeepSeek Harness 的扩展只有 DeepSeek 官方认。**发给它时一个字节都不改**
-    let harness = generates && reading.harness.is_some();
+    // DeepSeek Harness 的扩展只有 DeepSeek 官方认。**发给它时一个字节都不改**。
+    // 不止生成请求：数 token 这样的请求一样带着它的头，也可能带着会话日志
+    let harness = reading.harness.is_some();
     let to_deepseek = tw_dialect::official::is_deepseek_host(&provider.base_url);
     let mut path = req.uri.path().to_string();
     let mut query = req.query.clone();
@@ -525,11 +526,17 @@ async fn send(
     // DeepSeek Harness 发给别家：它自己的头不转发；直通时 `anthropic-beta` 去掉对话中途
     // 增删工具那一项（那些块已经去掉了），剩下的照发
     let harness = out.harness_elsewhere;
-    let beta = headers
-        .get("anthropic-beta")
-        .filter(|_| harness && target.is_none())
-        .and_then(|v| v.to_str().ok())
-        .and_then(tw_dialect::harness::anthropic_beta);
+    // 分几行发的也要都看：下面一律不转发原来的，只转发这一份
+    let beta = (harness && target.is_none())
+        .then(|| {
+            headers
+                .get_all("anthropic-beta")
+                .iter()
+                .filter_map(|v| v.to_str().ok())
+                .collect::<Vec<_>>()
+                .join(",")
+        })
+        .and_then(|v| tw_dialect::harness::anthropic_beta(&v));
     let build = |upstream_headers: &[(String, String)]| {
         let mut req = http.request(method.clone(), &url);
         req = forward::forward_headers_filtered(req, headers, |n| {
diff --git a/crates/tw-gateway/src/server/pipeline/relay.rs b/crates/tw-gateway/src/server/pipeline/relay.rs
index a28fa18f..6bddbb09 100644
--- a/crates/tw-gateway/src/server/pipeline/relay.rs
+++ b/crates/tw-gateway/src/server/pipeline/relay.rs
@@ -97,6 +97,7 @@ pub(super) fn respond(
         && plan.client_sse
         && status.is_success())
     .then_some(state.ping_every);
+    let ping_for = state.ping_for;
     let stream = async_stream::stream! {
         // **通行证跟着响应体走。**这个流被丢掉的时候它才还回去：正常
         // 发完是一种，客户端中途断开、hyper 丢掉响应体是另一种 —— 两种
@@ -106,17 +107,27 @@ pub(super) fn respond(
         let mut ending = ending;
         let mut chunks = std::pin::pin!(chunks);
         let mut broke: Option<GatewayError> = None;
+        // 客户端上一次收到字节的时刻。**心跳看的是客户端那一边的静默**，不是上游的：
+        // 上游在排队时发的 SSE 注释（DeepSeek 的 `: keep-alive`、OpenRouter 的
+        // `: OPENROUTER PROCESSING`）转换时被丢掉，上游不算静默，客户端却什么都没收到
+        let mut quiet_since = tokio::time::Instant::now();
+        // 上游上一次发来任何字节（注释也算）的时刻。**上游整个没了声音太久就不再补**：
+        // 半开的连接永远等不到下一个字节，一直补心跳的话客户端也永远不会放弃，这个
+        // 请求就挂在那儿了。停下之后由客户端自己的静默计时来断
+        let mut upstream_since = tokio::time::Instant::now();
         loop {
             let next = match ping_every {
                 None => chunks.next().await,
                 // `next()` 被超时丢掉不丢数据：它只是去问一次流，没拿走任何东西
-                Some(every) => match tokio::time::timeout(every, chunks.next()).await {
+                Some(every) => match tokio::time::timeout_at(quiet_since + every, chunks.next()).await {
                     Ok(next) => next,
                     Err(_) => {
+                        // 停在一帧中间时这一轮不补，也要重新计时，否则会原地空转
+                        quiet_since = tokio::time::Instant::now();
                         // **不经过留档、计量和审查**：心跳不是上游说的话，不进请求记录，
                         // 也不算输出。**只在帧的边界上插**，上游停在一帧中间时插进去
                         // 会把那一帧拆坏 —— 那时宁可不补
-                        if relay.between_frames() {
+                        if relay.between_frames() && upstream_since.elapsed() < ping_for {
                             yield Ok::<Bytes, std::io::Error>(Bytes::from_static(PING));
                         }
                         continue;
@@ -126,6 +137,7 @@ pub(super) fn respond(
             let Some(item) = next else { break };
             match item {
                 Ok(chunk) => {
+                    upstream_since = tokio::time::Instant::now();
                     // **旁路嗅探和留档，不缓冲**：字节照常流向客户端，同时
                     // 喂它一份。上游返回的 usage 是真相，而拿不到它就只能估。
                     //
@@ -136,6 +148,7 @@ pub(super) fn respond(
                     let (out, cut) = relay.chunk(&chunk);
                     relay.sent(&out);
                     if !out.is_empty() {
+                        quiet_since = tokio::time::Instant::now();
                         yield Ok::<Bytes, std::io::Error>(Bytes::from(out));
                     }
                     if let Some(err) = cut {
diff --git a/crates/tw-gateway/src/state.rs b/crates/tw-gateway/src/state.rs
index 8202749a..84eb9e11 100644
--- a/crates/tw-gateway/src/state.rs
+++ b/crates/tw-gateway/src/state.rs
@@ -210,6 +210,8 @@ pub struct AppState {
     /// Anthropic 流里上游静默多久就补一个 `ping`（见 `relay`）。**测试会把它调短**，
     /// 否则一条心跳的测试要干等十五秒
     pub ping_every: std::time::Duration,
+    /// 上游整个静默（连注释都没有）超过这么久，就不再补 `ping`（见 `relay`）。**测试会把它调短**
+    pub ping_for: std::time::Duration,
 }
 
 impl AppState {
@@ -260,6 +262,7 @@ impl AppState {
             live: crate::live::Live::default(),
             sessions: Default::default(),
             ping_every: crate::PING_EVERY,
+            ping_for: crate::PING_FOR,
         };
         // 手写的清单马上可用；向上游问是后台的事，不挡启动
         state.publish_catalog();
diff --git a/crates/tw-gateway/src/state/glm.rs b/crates/tw-gateway/src/state/glm.rs
index dd480fd6..0d670ed1 100644
--- a/crates/tw-gateway/src/state/glm.rs
+++ b/crates/tw-gateway/src/state/glm.rs
@@ -22,8 +22,12 @@ impl AppState {
 
     /// 这一家是 GLM Coding Plan 的话，额度接口的地址和凭据的指纹。
     ///
-    /// 没有 key 的不算：额度跟着 key 走，没有 key 就没有可问的
+    /// 没有 key 的不算：额度跟着 key 走，没有 key 就没有可问的。**停用的也不算**：
+    /// 用户不想用它了，不该再拿它的 key 去问
     fn glm_target(&self, p: &tw_config::Provider) -> Option<(String, String)> {
+        if p.disabled {
+            return None;
+        }
         let key = p.key.as_ref()?;
         let url = self.glm.sites().quota_url(&p.base_url)?;
         // 指纹，不是 key 本身：它只用来认出「换了 key」
@@ -67,6 +71,11 @@ impl AppState {
         let p = p.clone();
         Some(tokio::spawn(async move {
             let answer = state.ask_glm(&p, &url).await;
+            // 节奏先记上，**按记下的那个结论办**：刚才还有套餐的 key 头一次说没有，
+            // 当成一次失败（见 `Tracker::settle`）
+            let Some(answer) = state.glm.settle(&p.name, &ident, answer, now_ms()) else {
+                return;
+            };
             match &answer {
                 Answer::Quota(q) => state.record_quota(state.bus.next_id(), &p.name, q.clone()),
                 // 没有套餐就没有额度数据：之前记着的也不再作数
@@ -78,7 +87,6 @@ impl AppState {
                     tracing::debug!(provider = %p.name, "the GLM quota could not be read")
                 }
             }
-            state.glm.settle(&p.name, &ident, &answer, now_ms());
         }))
     }
 
@@ -279,5 +287,8 @@ mod tests {
         let mut keyless = cfg.providers[0].clone();
         keyless.key = None;
         assert!(s.glm_target(&keyless).is_none());
+        let mut disabled = cfg.providers[0].clone();
+        disabled.disabled = true;
+        assert!(s.glm_target(&disabled).is_none());
     }
 }
diff --git a/crates/tw-gateway/tests/claude_desktop.rs b/crates/tw-gateway/tests/claude_desktop.rs
index a566ccfd..2d4dfa34 100644
--- a/crates/tw-gateway/tests/claude_desktop.rs
+++ b/crates/tw-gateway/tests/claude_desktop.rs
@@ -97,6 +97,10 @@ struct Gateway {
 }
 
 async fn gateway(p: Provider) -> Gateway {
+    gateway_pinging_for(p, tw_gateway::PING_FOR).await
+}
+
+async fn gateway_pinging_for(p: Provider, ping_for: Duration) -> Gateway {
     let cfg = Config {
         version: 1,
         listen: Listen::default(),
@@ -111,6 +115,7 @@ async fn gateway(p: Provider) -> Gateway {
     let mut state = tw_gateway::AppState::new(cfg).unwrap();
     // 心跳的间隔调短，一条测试不必干等十五秒
     state.ping_every = Duration::from_millis(100);
+    state.ping_for = ping_for;
     let events = state.bus.subscribe();
     let (tx, bodies) = tokio::sync::mpsc::channel(16);
     state.set_body_sink(tx);
@@ -343,6 +348,68 @@ async fn a_silent_converted_upstream_is_covered_with_pings() {
     assert!(!recorded.contains("ping"), "心跳进了请求记录：{recorded}");
 }
 
+/// 上游在排队时只发 SSE 注释（DeepSeek 的 `: keep-alive`）。转换时注释被丢掉，
+/// 客户端什么都收不到 —— **心跳要看客户端那一边的静默**，不能被上游的注释推迟
+#[tokio::test]
+async fn upstream_comments_that_never_reach_the_client_do_not_hold_pings_back() {
+    let mut parts = vec![Some(CHAT_FIRST)];
+    for _ in 0..14 {
+        parts.push(Some(": keep-alive\n\n"));
+        parts.push(None);
+    }
+    parts.push(Some(CHAT_REST));
+    let up = stalling_upstream(parts, Duration::from_millis(50)).await;
+    let gw = gateway(provider(up, Protocol::OpenaiChat, &[])).await;
+    let text = stream_from(
+        &gw,
+        "/v1/messages",
+        json!({"model": "gpt-x", "max_tokens": 16, "stream": true,
+               "messages": [{"role": "user", "content": "hi"}]}),
+    )
+    .await;
+    let ping = "event: ping\ndata: {\"type\": \"ping\"}\n\n";
+    assert!(text.matches(ping).count() >= 2, "{text}");
+    let rest = text.replace(ping, "");
+    assert!(!rest.contains("keep-alive"), "{rest}");
+    assert!(
+        rest.trim_end().ends_with("\"type\":\"message_stop\"}"),
+        "{rest}"
+    );
+}
+
+/// 上游一个字节都不发太久（连接半开了）：**不再补心跳**，让客户端自己的静默计时断开它。
+/// 一直补的话这个请求永远挂着
+#[tokio::test]
+async fn pings_stop_once_the_upstream_has_been_silent_too_long() {
+    let up = stalling_upstream(
+        vec![Some(CHAT_FIRST), None, Some(CHAT_REST)],
+        Duration::from_millis(1200),
+    )
+    .await;
+    let gw = gateway_pinging_for(
+        provider(up, Protocol::OpenaiChat, &[]),
+        Duration::from_millis(350),
+    )
+    .await;
+    let text = stream_from(
+        &gw,
+        "/v1/messages",
+        json!({"model": "gpt-x", "max_tokens": 16, "stream": true,
+               "messages": [{"role": "user", "content": "hi"}]}),
+    )
+    .await;
+    let ping = "event: ping\ndata: {\"type\": \"ping\"}\n\n";
+    // 静默 1.2 秒、每 0.1 秒一次：一直补的话有十一个。只补前 0.35 秒的
+    let n = text.matches(ping).count();
+    assert!((1..=4).contains(&n), "{n}: {text}");
+    assert!(
+        text.replace(ping, "")
+            .trim_end()
+            .ends_with("\"type\":\"message_stop\"}"),
+        "{text}"
+    );
+}
+
 #[tokio::test]
 async fn a_ping_never_splits_a_frame_the_upstream_left_half_sent() {
     const START: &str = "event: message_start\ndata: {\"type\":\"message_start\",\"message\":{\"id\":\"m1\",\"type\":\"message\",\"role\":\"assistant\",\"model\":\"claude-sonnet-4-5\",\"content\":[],\"usage\":{\"input_tokens\":5,\"output_tokens\":1}}}\n\n";
diff --git a/crates/tw-gateway/tests/harness.rs b/crates/tw-gateway/tests/harness.rs
index 8c6f83b6..c29bf815 100644
--- a/crates/tw-gateway/tests/harness.rs
+++ b/crates/tw-gateway/tests/harness.rs
@@ -345,6 +345,57 @@ async fn messages_to_another_anthropic_upstream_are_cleaned_in_place() {
     );
 }
 
+/// `anthropic-beta` 分两行发：清理的那一行去掉，**另一行照发**
+#[tokio::test]
+async fn betas_sent_on_separate_lines_all_survive_the_cleaning() {
+    let (gw, _rx, log) = gateway(elsewhere(Protocol::Anthropic).await).await;
+    let (status, body) = send(
+        gw,
+        "/v1/messages",
+        &[
+            ("anthropic-version", "2023-06-01"),
+            ("anthropic-beta", TOOL_BETA),
+            ("anthropic-beta", "files-api-2025-04-14"),
+        ],
+        &messages_request().to_string(),
+    )
+    .await;
+    assert_eq!(status, 200, "{body}");
+    let s = generation(&log);
+    assert_clean(&s);
+    let betas: Vec<_> = s.headers.get_all("anthropic-beta").iter().collect();
+    assert_eq!(betas, ["files-api-2025-04-14"]);
+}
+
+/// 不生成回答的请求（数 token）也是 dsh 发的：发给别家时一样不带它的头和会话日志
+#[tokio::test]
+async fn counting_tokens_elsewhere_is_cleaned_too() {
+    let (gw, _rx, log) = gateway(elsewhere(Protocol::Anthropic).await).await;
+    let mut req = messages_request();
+    req.as_object_mut().unwrap().remove("max_tokens");
+    let (status, body) = send(
+        gw,
+        "/v1/messages/count_tokens",
+        &[
+            ("anthropic-version", "2023-06-01"),
+            ("anthropic-beta", TOOL_BETA),
+        ],
+        &req.to_string(),
+    )
+    .await;
+    assert_eq!(status, 200, "{body}");
+    let s = log
+        .lock()
+        .unwrap()
+        .iter()
+        .rev()
+        .find(|s| s.uri.ends_with("/count_tokens"))
+        .cloned()
+        .expect("the upstream got no count_tokens request");
+    assert_clean(&s);
+    assert!(s.headers.get("anthropic-beta").is_none());
+}
+
 #[tokio::test]
 async fn messages_to_an_openai_upstream_are_cleaned_by_the_conversion() {
     let (gw, mut rx, log) = gateway(elsewhere(Protocol::OpenaiChat).await).await;