diff --git a/backend/crates/gateway-core/src/engine/coordinator.rs b/backend/crates/gateway-core/src/engine/coordinator.rs
index b5747c284..9dc2d2b5d 100644
--- a/backend/crates/gateway-core/src/engine/coordinator.rs
+++ b/backend/crates/gateway-core/src/engine/coordinator.rs
@@ -1582,6 +1582,12 @@ where
             .attempt(self.attempts)
             .record_provider_failure(&error);
         let mut atomic_client_events = error.take_atomic_client_events();
+        // 失败随附的观测也属于当前 attempt，换号记录不能等到最终交付才读取
+        for event in &atomic_client_events {
+            if let Some(observation) = event.response_observation() {
+                self.observe_response(observation.clone());
+            }
+        }
         let current = self.current.take().ok_or(EngineError::NoActiveAttempt)?;
         if self.request_observation.is_some() {
             let websocket_attempt = websocket_observation_attempt(&current);
diff --git a/backend/crates/gateway-core/src/engine/execution/contract.rs b/backend/crates/gateway-core/src/engine/execution/contract.rs
index 138995ea1..0a234884d 100644
--- a/backend/crates/gateway-core/src/engine/execution/contract.rs
+++ b/backend/crates/gateway-core/src/engine/execution/contract.rs
@@ -65,6 +65,22 @@ pub struct ExecutionRequestMetadata {
     pub previous_response_id: Option<PreviousResponseId>,
 }
 
+impl ExecutionRequestMetadata {
+    pub(super) fn inherit_request_origin(&mut self, scope: &ExtensionCallScope) {
+        let Some(request) = scope.request() else {
+            return;
+        };
+        // 执行入口可以是内部调用，客户端来源仍属于原始入口快照
+        self.client_ip = request.client_ip;
+        self.user_agent = request
+            .headers
+            .iter()
+            .find(|header| header.name().eq_ignore_ascii_case("user-agent"))
+            .and_then(|header| std::str::from_utf8(header.value()).ok())
+            .map(str::to_owned);
+    }
+}
+
 #[derive(Clone)]
 pub struct AuthenticatedClient {
     pub(super) settings: Option<RequestSettings>,
diff --git a/backend/crates/gateway-core/src/engine/execution/service.rs b/backend/crates/gateway-core/src/engine/execution/service.rs
index 8075fbf57..464e0687f 100644
--- a/backend/crates/gateway-core/src/engine/execution/service.rs
+++ b/backend/crates/gateway-core/src/engine/execution/service.rs
@@ -1001,8 +1001,7 @@ impl DefaultExecutionService {
         let execution_effects_baseline = parent.graph.effects.epoch();
         metadata.endpoint = "host.model".to_owned();
         metadata.transport = ClientTransport::InternalPlugin;
-        metadata.client_ip = None;
-        metadata.user_agent = None;
+        metadata.inherit_request_origin(&extension_scope);
         let authorization = AuthorizedExecution {
             account_scope,
             response_control: None,
@@ -1155,8 +1154,7 @@ impl DefaultExecutionService {
         let execution_effects_baseline = graph.effects.epoch();
         metadata.endpoint = "host.model".to_owned();
         metadata.transport = ClientTransport::InternalPlugin;
-        metadata.client_ip = None;
-        metadata.user_agent = None;
+        metadata.inherit_request_origin(&authority.extension_scope);
         self.client_api_key_usage
             .record_used(authority.client.policy.key_id());
         self.start_authorized(
diff --git a/backend/crates/gateway-core/src/event.rs b/backend/crates/gateway-core/src/event.rs
index 4ab98ccc1..fcf0fa920 100644
--- a/backend/crates/gateway-core/src/event.rs
+++ b/backend/crates/gateway-core/src/event.rs
@@ -547,7 +547,7 @@ pub struct ProviderResponseTimings {
 /// 这样 Provider 专有协议不会渗入路由
 /// 或管理领域
 #[derive(Clone, PartialEq, Eq)]
-pub struct ProviderResponseMetadata(String);
+pub struct ProviderResponseMetadata(Arc<str>);
 
 impl ProviderResponseMetadata {
     const MAX_BYTES: usize = 32 * 1024;
@@ -561,7 +561,7 @@ impl ProviderResponseMetadata {
         serde_json::from_str::<Value>(&json)
             .ok()
             .filter(Value::is_object)
-            .map(|_| Self(json))
+            .map(|_| Self(json.into()))
     }
 
     /// 返回不透明的 JSON 文本，供持久化 adapter 原样验证和写入
@@ -626,7 +626,8 @@ pub struct ProviderResponseObservation {
     service_tier: Option<String>,
     upstream_response_model: Option<String>,
     timings: ProviderResponseTimings,
-    client_headers: Vec<ProviderResponseHeader>,
+    // 流式事件分别持有快照，共享不可变头集合，避免逐事件复制头名和集合
+    client_headers: Arc<[ProviderResponseHeader]>,
     provider_metadata: Option<ProviderResponseMetadata>,
 }
 
@@ -642,7 +643,7 @@ impl ProviderResponseObservation {
             service_tier: None,
             upstream_response_model: None,
             timings: ProviderResponseTimings::default(),
-            client_headers: Vec::new(),
+            client_headers: Arc::default(),
             provider_metadata: None,
         }
     }
@@ -719,7 +720,7 @@ impl ProviderResponseObservation {
 
     #[must_use]
     pub fn with_client_headers(mut self, client_headers: Vec<ProviderResponseHeader>) -> Self {
-        self.client_headers = client_headers;
+        self.client_headers = client_headers.into();
         self
     }
 
diff --git a/backend/crates/gateway-core/tests/engine/coordinator.rs b/backend/crates/gateway-core/tests/engine/coordinator.rs
index 7d73b1464..950831c98 100644
--- a/backend/crates/gateway-core/tests/engine/coordinator.rs
+++ b/backend/crates/gateway-core/tests/engine/coordinator.rs
@@ -3467,13 +3467,16 @@ fn explicit_429_after_structural_event_should_retry_before_commit() {
 fn atomic_response_failed_should_rotate_account_before_stream_commit() {
     let operation = generate_operation();
     let route_plan = plan(&operation);
+    let mut failure = atomic_response_failed("response-first-failed", "first account failure");
+    let mut events = failure.take_atomic_client_events();
+    events[0].attach_observation(
+        ProviderResponseObservation::new(UpstreamTransport::new("http_sse").unwrap())
+            .with_request_id(OpaqueUpstreamValue::new("upstream-failed-attempt")),
+    );
     let (coordinator, store, provider) = coordinator(vec![
         Script::ObservedStream {
             account_id: "acct_first",
-            items: vec![Err(atomic_response_failed(
-                "response-first-failed",
-                "first account failure",
-            ))],
+            items: vec![Err(failure.with_atomic_client_events(events))],
         },
         Script::Stream {
             account_id: "acct_second",
@@ -3516,6 +3519,10 @@ fn atomic_response_failed_should_rotate_account_before_stream_commit() {
     let state = store.state.lock().expect("store lock");
     assert_eq!(state.commits, 1);
     assert_eq!(state.intermediate_failures, 1);
+    assert_eq!(
+        state.intermediate_request_ids,
+        [Some("upstream-failed-attempt".to_owned())]
+    );
     assert_eq!(state.finalizations[0].outcome, ExecutionOutcome::Succeeded);
 }
 
diff --git a/backend/crates/gateway-core/tests/engine/execution/mod.rs b/backend/crates/gateway-core/tests/engine/execution/mod.rs
index 3355865b4..c44c83a65 100644
--- a/backend/crates/gateway-core/tests/engine/execution/mod.rs
+++ b/backend/crates/gateway-core/tests/engine/execution/mod.rs
@@ -1195,6 +1195,114 @@ fn request_policy_snapshot(
     .with_extensions(Some(generation))
 }
 
+fn plugin_origin_scope() -> gateway_core::engine::extensions::ExtensionCallScope {
+    gateway_core::engine::extensions::ExtensionCallScope::default().with_request(Arc::new(
+        gateway_core::engine::extensions::RequestContext {
+            method: "POST".into(),
+            uri: "/plugin/proxy".into(),
+            http_version: "HTTP/1.1".into(),
+            headers: vec![gateway_core::engine::middleware::MiddlewareHeader::new(
+                "User-Agent",
+                Bytes::from_static(b"fixture-client/494"),
+            )],
+            peer_addr: None,
+            client_ip: Some("192.0.2.49".parse().unwrap()),
+        },
+    ))
+}
+
+#[test]
+fn explicit_key_model_execution_preserves_http_origin_and_leaves_background_origin_unknown() {
+    use gateway_core::engine::nested::{
+        BoundModelExecutionBinding, BoundModelExecutionRequest, NestedModelExecutionPort,
+    };
+    block_on(async {
+        for has_origin in [true, false] {
+            let store = Arc::new(TrackingExecutionStore::default());
+            let service = DefaultExecutionService::new(
+                RuntimeSnapshotHandle::new(request_policy_snapshot(
+                    super::extensions::reference("origin-fixture"),
+                    &[ProviderKind::new("openai").unwrap()],
+                )),
+                store.clone(),
+                ProviderRegistry::new([Arc::new(ProcessingProvider) as Arc<dyn Provider>]).unwrap(),
+                Arc::new(Admissions::default()),
+                Arc::new(UnusedContinuation),
+                Arc::new(RecordingClientApiKeyUsage::default()),
+                Arc::new(crate::RecordingDiagnostics::default()),
+            );
+            let input = request(&service, ClientTransport::InternalPlugin);
+            let context = NestedModelExecutionPort::bind(
+                &service,
+                BoundModelExecutionBinding {
+                    settings: None,
+                    client_key_id: ClientApiKeyId::new("key_start_test").unwrap(),
+                    initiating_plugin_instance_id: "origin-fixture".into(),
+                    cancellation: Default::default(),
+                    extension_scope: if has_origin {
+                        plugin_origin_scope()
+                    } else {
+                        Default::default()
+                    },
+                },
+            )
+            .await
+            .unwrap();
+            let mut started = NestedModelExecutionPort::start_bound(
+                &service,
+                BoundModelExecutionRequest {
+                    context: context.clone(),
+                    public_model: input.public_model,
+                    operation: input.operation,
+                    metadata: input.metadata,
+                    provider: None,
+                    account: None,
+                },
+            )
+            .await
+            .unwrap();
+            let mut child = NestedModelExecutionPort::start(
+                &service,
+                NestedModelExecutionRequest {
+                    parent_request_id: started.request_id.clone(),
+                    initiating_plugin_instance_id: "nested-origin-fixture".into(),
+                    public_model: PublicModelId::new("gpt-start").unwrap(),
+                    operation: start_operation(),
+                    metadata: request(&service, ClientTransport::InternalPlugin).metadata,
+                    provider: None,
+                    account: None,
+                    parent_account: None,
+                },
+            )
+            .await
+            .unwrap();
+            child.session.collect_uncommitted().await.unwrap();
+            child.session.commit_downstream(Some(200)).await.unwrap();
+            child.session.detach_finalize().await;
+            started.session.collect_uncommitted().await.unwrap();
+            {
+                let requests = store.requests.lock().unwrap();
+                assert_eq!(requests.len(), 2);
+                for request in requests.iter() {
+                    assert_eq!(
+                        request.client_ip,
+                        has_origin.then(|| "192.0.2.49".parse().unwrap())
+                    );
+                    assert_eq!(
+                        request.user_agent.as_deref(),
+                        has_origin.then_some("fixture-client/494")
+                    );
+                    assert_eq!(
+                        request.client_transport,
+                        ClientTransport::InternalPlugin.as_str()
+                    );
+                }
+            }
+            started.session.detach_finalize().await;
+        }
+    });
+}
+
 #[test]
 fn completed_parent_rejects_new_nested_execution_and_cancels_an_active_child() {
     block_on(async {
diff --git a/backend/crates/gateway-core/tests/event.rs b/backend/crates/gateway-core/tests/event.rs
index 01c0a0f8e..1733e42b1 100644
--- a/backend/crates/gateway-core/tests/event.rs
+++ b/backend/crates/gateway-core/tests/event.rs
@@ -3,8 +3,9 @@
 use bytes::Bytes;
 use gateway_core::event::{
     ContentItem, ContentKind, EventSequenceError, EventSequenceValidator, GatewayEvent,
-    ProtocolWireEvent, ProviderEvent, ProviderResponseHeader, ProviderResponseObservation,
-    ReasoningDelta, ResponseMeta, TextDelta, ToolCallDelta,
+    ProtocolWireEvent, ProviderEvent, ProviderResponseHeader, ProviderResponseMetadata,
+    ProviderResponseObservation, ProviderResponseTimings, ReasoningDelta, ResponseMeta, TextDelta,
+    ToolCallDelta,
 };
 use gateway_core::upstream::UpstreamTransport;
 use serde_json::json;
@@ -335,3 +336,61 @@ fn validator_should_accept_reasoning_and_tool_call_stream() {
 
     assert_eq!(validator.finish(), Ok(()));
 }
+
+#[test]
+fn response_observation_snapshots_preserve_values_without_copying_large_payloads() {
+    let metadata = format!("{{\"raw\":\"{}\"}}", "a".repeat(32 * 1024 - 10));
+    let observation =
+        ProviderResponseObservation::new(UpstreamTransport::new("http_sse").expect("transport"))
+            .with_provider_metadata(
+                ProviderResponseMetadata::new(metadata.clone()).expect("metadata"),
+            )
+            .with_client_headers(
+                (0..32)
+                    .map(|_| {
+                        ProviderResponseHeader::new("set-cookie", Bytes::from_static(b"raw-\xff"))
+                    })
+                    .collect(),
+            )
+            .with_timings(ProviderResponseTimings {
+                first_event_ms: Some(17),
+                ..ProviderResponseTimings::default()
+            });
+
+    let allocations = allocation_counter::measure(|| {
+        for _ in 0..1_000 {
+            std::hint::black_box(observation.clone());
+        }
+    });
+    // 允许小标量字段的复制，但流式快照不能随大头集合和 metadata 重复分配
+    assert!(allocations.bytes_total < 256 * 1_000, "{allocations:?}");
+    assert_eq!(allocations.bytes_current, 0);
+
+    let updated = observation
+        .clone()
+        .with_client_headers(Vec::new())
+        .with_provider_metadata(
+            ProviderResponseMetadata::new("{\"updated\":true}".to_owned()).expect("metadata"),
+        )
+        .with_timings(ProviderResponseTimings {
+            first_event_ms: Some(23),
+            ..ProviderResponseTimings::default()
+        });
+    assert_eq!(
+        observation.provider_metadata().expect("metadata").as_json(),
+        metadata
+    );
+    assert_eq!(observation.client_headers().len(), 32);
+    assert!(
+        observation.client_headers().iter().all(|header| {
+            header.name() == "set-cookie" && header.value().as_ref() == b"raw-\xff"
+        })
+    );
+    assert_eq!(observation.timings().first_event_ms, Some(17));
+    assert_eq!(
+        updated.provider_metadata().expect("metadata").as_json(),
+        "{\"updated\":true}"
+    );
+    assert!(updated.client_headers().is_empty());
+    assert_eq!(updated.timings().first_event_ms, Some(23));
+}
diff --git a/backend/crates/gateway-host/src/outbound/http.rs b/backend/crates/gateway-host/src/outbound/http.rs
index 910e98c3e..5a6ebed48 100644
--- a/backend/crates/gateway-host/src/outbound/http.rs
+++ b/backend/crates/gateway-host/src/outbound/http.rs
@@ -46,6 +46,7 @@ pub struct HttpRequest {
 
 pub struct HttpResponse {
     pub status: u16,
+    pub version: http::Version,
     pub headers: Vec<(String, Vec<u8>)>,
     pub body: HttpBody,
 }
@@ -53,6 +54,7 @@ pub struct HttpResponse {
 pub struct HttpBody {
     response: hyper::body::Incoming,
     pending: Bytes,
+    trailers: Vec<(String, Vec<u8>)>,
     received: usize,
     deadline: Instant,
     _permit: OwnedSemaphorePermit,
@@ -66,7 +68,8 @@ pub struct HttpError {
     pub send_state: UpstreamSendState,
 }
 
-#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+#[derive(Debug, Clone, Copy, PartialEq, Eq, serde::Serialize)]
+#[serde(rename_all = "snake_case")]
 pub enum HttpErrorKind {
     InvalidRequest,
     AddressDenied,
@@ -366,10 +369,12 @@ impl PendingResponse {
     pub(super) fn into_http(self) -> HttpResponse {
         HttpResponse {
             status: self.status,
+            version: self.response.version(),
             headers: self.headers,
             body: HttpBody {
                 response: self.response.into_body(),
                 pending: Bytes::new(),
+                trailers: Vec::new(),
                 received: 0,
                 deadline: self.deadline,
                 _permit: self.permit,
@@ -379,6 +384,11 @@ impl PendingResponse {
 }
 
 impl HttpBody {
+    /// 取出截至当前读取位置收到的尾部头，保留多值与原始字节
+    pub fn take_trailers(&mut self) -> Vec<(String, Vec<u8>)> {
+        std::mem::take(&mut self.trailers)
+    }
+
     pub async fn read(&mut self, maximum: usize) -> Result<Option<Bytes>, HttpError> {
         self.read_until(maximum, self.deadline).await
     }
@@ -414,7 +424,15 @@ impl HttpBody {
             self.pending = match tokio::time::timeout_at(deadline, self.response.frame()).await {
                 Ok(Some(Ok(frame))) => match frame.into_data() {
                     Ok(chunk) => chunk,
-                    Err(_) => continue,
+                    Err(frame) => {
+                        let trailers = frame
+                            .into_trailers()
+                            .map_err(|_| HttpError::sent("response frame"))?;
+                        self.trailers.extend(trailers.iter().map(|(name, value)| {
+                            (name.as_str().to_owned(), value.as_bytes().to_vec())
+                        }));
+                        continue;
+                    }
                 },
                 Ok(None) => return Ok(None),
                 Ok(Some(Err(_))) => return Err(HttpError::sent("response body")),
diff --git a/backend/crates/gateway-host/src/outbound/websocket.rs b/backend/crates/gateway-host/src/outbound/websocket.rs
index c30c417c4..7265713b4 100644
--- a/backend/crates/gateway-host/src/outbound/websocket.rs
+++ b/backend/crates/gateway-host/src/outbound/websocket.rs
@@ -27,6 +27,8 @@ type Socket = WebSocketStream<TokioIo<hyper::upgrade::Upgraded>>;
 
 /// 读写分别串行，允许等待输出时发送控制消息；关闭会唤醒在途操作并释放容量
 pub struct ManagedWebSocket {
+    handshake_version: http::Version,
+    handshake_headers: Vec<(String, Vec<u8>)>,
     writer: Mutex<Option<(SplitSink<Socket, Message>, OwnedSemaphorePermit)>>,
     reader: Mutex<Option<SplitStream<Socket>>>,
     closed: watch::Sender<bool>,
@@ -39,6 +41,7 @@ pub enum WebSocketMessage {
 
 pub struct WebSocketResponse {
     pub status: u16,
+    pub version: http::Version,
     pub headers: Vec<(String, Vec<u8>)>,
     pub connection: Option<ManagedWebSocket>,
     pub body: Option<HttpBody>,
@@ -70,6 +73,7 @@ impl HttpClient {
             let response = response.into_http();
             return Ok(WebSocketResponse {
                 status: response.status,
+                version: response.version,
                 headers: response.headers,
                 connection: None,
                 body: Some(response.body),
@@ -112,8 +116,11 @@ impl HttpClient {
         let (writer, reader) = socket.split();
         Ok(WebSocketResponse {
             status: response.status,
-            headers: response.headers,
+            version: response.response.version(),
+            headers: response.headers.clone(),
             connection: Some(ManagedWebSocket {
+                handshake_version: response.response.version(),
+                handshake_headers: response.headers,
                 writer: Mutex::new(Some((writer, response.permit))),
                 reader: Mutex::new(Some(reader)),
                 closed: watch::channel(false).0,
@@ -124,6 +131,17 @@ impl HttpClient {
 }
 
 impl ManagedWebSocket {
+    /// 连接复用仍可读取最初握手的原始元数据，不伪造一次新握手
+    #[must_use]
+    pub const fn handshake_version(&self) -> http::Version {
+        self.handshake_version
+    }
+
+    #[must_use]
+    pub fn handshake_headers(&self) -> &[(String, Vec<u8>)] {
+        &self.handshake_headers
+    }
+
     pub async fn send(
         &self,
         message: WebSocketMessage,
diff --git a/backend/crates/gateway-plugin/runtime/plugin-host-compatibility.json b/backend/crates/gateway-plugin/runtime/plugin-host-compatibility.json
index aa469fbb2..4c26ac476 100644
--- a/backend/crates/gateway-plugin/runtime/plugin-host-compatibility.json
+++ b/backend/crates/gateway-plugin/runtime/plugin-host-compatibility.json
@@ -1,7 +1,7 @@
 {
   "schema_version": 2,
   "manifest_schema_versions": [2],
-  "protocol_versions": [4],
+  "protocol_versions": [5],
   "capabilities": [
     { "capability": "frontend_authentication", "versions": [2] },
     { "capability": "scheduler", "versions": [3] },
@@ -9,7 +9,7 @@
     { "capability": "model_catalog", "versions": [1] },
     { "capability": "retry_policy", "versions": [2] },
     { "capability": "middleware", "versions": [5] },
-    { "capability": "upstream_adapter", "versions": [3] },
+    { "capability": "upstream_adapter", "versions": [4] },
     { "capability": "observer", "versions": [2] },
     { "capability": "command_line", "versions": [1] },
     { "capability": "management", "versions": [2] },
diff --git a/backend/crates/gateway-plugin/runtime/src/adapter/upstream_adapter/event.rs b/backend/crates/gateway-plugin/runtime/src/adapter/upstream_adapter/event.rs
index a35b9d530..c5f2389e4 100644
--- a/backend/crates/gateway-plugin/runtime/src/adapter/upstream_adapter/event.rs
+++ b/backend/crates/gateway-plugin/runtime/src/adapter/upstream_adapter/event.rs
@@ -24,7 +24,7 @@ use gateway_protocol::openai::sse::SseEventDecoder;
 pub(super) struct EventDecoder {
     sequence: EventSequenceValidator,
     response_id: Option<String>,
-    service_tier: Option<String>,
+    observation: Option<ProviderResponseObservation>,
     usage: Usage,
     contents: usize,
     pub(super) completed: bool,
@@ -54,6 +54,7 @@ impl EventDecoder {
         transport: &str,
         account: &dyn UpstreamAccountConnection,
         sent: UpstreamSendState,
+        started_at: std::time::Instant,
     ) -> Result<DecodedEvent, EventDecodeError> {
         let message = UpstreamAdapterEvent::decode(payload)
             .map_err(|source| invalid(sent).with_source(source))?;
@@ -65,22 +66,31 @@ impl EventDecoder {
             .wire
             .map(|wire| decode_wire(wire, protocol, sent))
             .transpose()?;
+        let has_observation = self.observation.is_some() || message.observation.is_some();
+        let mut observation = match self.observation.take() {
+            Some(observation) => observation,
+            None => ProviderResponseObservation::new(
+                UpstreamTransport::new(transport)
+                    .map_err(|source| invalid(sent).with_source(source))?,
+            ),
+        };
+        if let Some(update) = message.observation {
+            observation = super::observation::merge(observation, update, sent)?;
+        }
         if let Some(failure) = message.failure {
             if !message.event.facts.is_empty() || message.continuation.is_some() {
                 return Err(invalid(sent).into());
             }
-            let mut error = failure_error(failure, sent)?;
+            let mut events = Vec::new();
+            if has_observation {
+                events.push(ProviderEvent::observation(observation));
+            }
             if let Some(wire) = wire {
-                error = error.with_atomic_client_events(vec![ProviderEvent::wire(wire)]);
+                events.push(ProviderEvent::wire(wire));
             }
-            return Err(EventDecodeError::Upstream(error));
-        }
-        if let Some(tier) = message.service_tier {
-            if tier.is_empty() || tier.len() > 64 || tier.chars().any(char::is_control) {
-                return Err(invalid(sent).into());
-            }
-            // 上游可能在完成时才把 auto 解析为实际档位；与原生观测一样保留最新值
-            self.service_tier = Some(tier);
+            return Err(EventDecodeError::Upstream(
+                failure_error(failure, sent)?.with_atomic_client_events(events),
+            ));
         }
         let mut facts = Vec::new();
         for fact in message.event.facts {
@@ -181,7 +191,7 @@ impl EventDecoder {
                         return Err(invalid(sent).into());
                     }
                     if let Some(cost) =
-                        account.calculate_cost(self.service_tier.as_deref(), &self.usage)
+                        account.calculate_cost(observation.service_tier(), &self.usage)
                     {
                         let cost = GatewayEvent::CalculatedCost(cost);
                         self.sequence
@@ -213,22 +223,26 @@ impl EventDecoder {
         if message.continuation.is_some() && !self.completed {
             return Err(invalid(sent).into());
         }
+        observation =
+            super::observation::observe_output(observation, &facts, wire.is_some(), started_at);
+        self.observation = Some(observation.clone());
         let mut event = if let Some(wire) = wire {
             ProviderEvent::canonical_with_wire(facts, wire)
         } else {
             let mut facts = facts.into_iter();
-            let first = facts.next().ok_or_else(|| invalid(sent))?;
-            facts.fold(ProviderEvent::canonical(first), ProviderEvent::with_fact)
+            match facts.next() {
+                Some(first) => {
+                    facts.fold(ProviderEvent::canonical(first), ProviderEvent::with_fact)
+                }
+                None => {
+                    return Ok(DecodedEvent {
+                        event: ProviderEvent::observation(observation),
+                        continuation: message.continuation,
+                    });
+                }
+            }
         };
-        if let Some(tier) = &self.service_tier {
-            let observation = ProviderResponseObservation::new(
-                UpstreamTransport::new(transport)
-                    .map_err(|source| invalid(sent).with_source(source))?,
-            )
-            .try_with_service_tier(tier.clone())
-            .map_err(|source| invalid(sent).with_source(source))?;
-            event.attach_observation(observation);
-        }
+        event.attach_observation(observation);
         Ok(DecodedEvent {
             event,
             continuation: message.continuation,
@@ -308,6 +322,20 @@ fn decode_wire(
                 ProtocolWireEvent::json_with_sse_metadata(protocol, event, data, id, retry)
             }
         }
+        WirePayload::WebSocket {
+            event,
+            data,
+            message,
+        } => {
+            let message =
+                String::from_utf8(message).map_err(|source| invalid(sent).with_source(source))?;
+            if let Some(data) = data {
+                ProtocolWireEvent::json(protocol, event, data)
+                    .map(|wire| wire.with_raw_websocket_message(message))
+            } else {
+                ProtocolWireEvent::raw_websocket(protocol, message)
+            }
+        }
         WirePayload::RawSse { frame } => {
             single_sse(&frame, sent)?;
             ProtocolWireEvent::raw_sse(protocol, Bytes::from(frame))
diff --git a/backend/crates/gateway-plugin/runtime/src/adapter/upstream_adapter/execution.rs b/backend/crates/gateway-plugin/runtime/src/adapter/upstream_adapter/execution.rs
index bfb60f44e..541c86e0a 100644
--- a/backend/crates/gateway-plugin/runtime/src/adapter/upstream_adapter/execution.rs
+++ b/backend/crates/gateway-plugin/runtime/src/adapter/upstream_adapter/execution.rs
@@ -106,6 +106,7 @@ impl Execution {
                 self.adapter.declaration.transport.as_str(),
                 self.invocation.account.as_ref(),
                 active.managed.send_state.get(),
+                self.invocation.context.timing_started_at(),
             )
             .map_err(|error| match error {
                 EventDecodeError::Invalid(error) => {
@@ -131,7 +132,7 @@ impl Execution {
             }.await;
             if let Err(error) = terminal {
                 // 尾部故障不得交付成功 wire，但已经解析的计量事实仍归 Core 结算
-                let metering = decoded
+                let mut metering: Vec<_> = decoded
                     .event
                     .canonical_facts()
                     .iter()
@@ -145,6 +146,11 @@ impl Execution {
                     .cloned()
                     .map(gateway_core::event::ProviderEvent::canonical)
                     .collect();
+                if let Some(observation) = decoded.event.response_observation() {
+                    metering.push(gateway_core::event::ProviderEvent::observation(
+                        observation.clone(),
+                    ));
+                }
                 return Err(error.with_atomic_client_events(metering));
             }
             self.ended = true;
@@ -243,6 +249,10 @@ impl Execution {
             protocol: invocation.operation.protocol().to_owned(),
             client_transport: invocation.context.client_transport().as_str().to_owned(),
             fast_mode: invocation.context.fast_mode().as_str().to_owned(),
+            request_elapsed_ms: u64::try_from(
+                invocation.context.timing_started_at().elapsed().as_millis(),
+            )
+            .unwrap_or(u64::MAX),
             headers: invocation
                 .headers
                 .iter()
diff --git a/backend/crates/gateway-plugin/runtime/src/adapter/upstream_adapter/mod.rs b/backend/crates/gateway-plugin/runtime/src/adapter/upstream_adapter/mod.rs
index 57c38833e..b95566b68 100644
--- a/backend/crates/gateway-plugin/runtime/src/adapter/upstream_adapter/mod.rs
+++ b/backend/crates/gateway-plugin/runtime/src/adapter/upstream_adapter/mod.rs
@@ -3,6 +3,7 @@
 mod event;
 mod execution;
 mod failure;
+mod observation;
 mod registration;
 pub(crate) mod target;
 
diff --git a/backend/crates/gateway-plugin/runtime/src/adapter/upstream_adapter/observation.rs b/backend/crates/gateway-plugin/runtime/src/adapter/upstream_adapter/observation.rs
new file mode 100644
index 000000000..9549a1096
--- /dev/null
+++ b/backend/crates/gateway-plugin/runtime/src/adapter/upstream_adapter/observation.rs
@@ -0,0 +1,129 @@
+//! 合并适配器的原始上游观测，并记录宿主实际收到规范化输出的时间
+
+use gateway_core::event::{
+    GatewayEvent, ProviderResponseHeader, ProviderResponseMetadata, ProviderResponseObservation,
+    ProviderResponseTimings, UpstreamHttpVersion, WebSocketPoolKind,
+};
+use gateway_core::upstream::{OpaqueUpstreamValue, UpstreamSendState};
+use gateway_plugin_sdk::call::model::facts::ResponseObservation;
+
+use super::failure::invalid;
+
+pub(super) fn merge(
+    mut current: ProviderResponseObservation,
+    update: ResponseObservation,
+    sent: UpstreamSendState,
+) -> Result<ProviderResponseObservation, gateway_core::error::ProviderError> {
+    // 传输类型由已选适配器固定，不能让不一致的报告绕过 Core 的观测归属
+    if update.transport != current.transport().as_str() {
+        return Err(invalid(sent));
+    }
+    if let Some(value) = update.http_version {
+        let version = if value == "unknown" {
+            UpstreamHttpVersion::Unknown
+        } else {
+            UpstreamHttpVersion::parse(&value).ok_or_else(|| invalid(sent))?
+        };
+        current = current.with_http_version(version);
+    }
+    if let Some(value) = update.websocket_pool {
+        current = current.with_websocket_pool(match value.as_str() {
+            "new" => WebSocketPoolKind::New,
+            "reuse" => WebSocketPoolKind::Reuse,
+            _ => return Err(invalid(sent)),
+        });
+    }
+    if let Some(value) = update.status_code {
+        if !(100..=599).contains(&value) {
+            return Err(invalid(sent));
+        }
+        current = current.with_status_code(value);
+    }
+    if let Some(value) = update.request_id {
+        current = current.with_request_id(OpaqueUpstreamValue::new(value));
+    }
+    if let Some(value) = update.service_tier {
+        current = current
+            .try_with_service_tier(value)
+            .map_err(|source| invalid(sent).with_source(source))?;
+    }
+    if let Some(value) = update.upstream_response_model {
+        current = current.with_upstream_response_model_if_valid(&value);
+        if current.upstream_response_model() != Some(value.trim()) {
+            return Err(invalid(sent));
+        }
+    }
+    if !update.client_headers.is_empty() {
+        current = current.with_client_headers(
+            update
+                .client_headers
+                .into_iter()
+                .map(|header| ProviderResponseHeader::new(header.name, header.value.into()))
+                .collect(),
+        );
+    }
+    if let Some(value) = update.provider_metadata {
+        current = current.with_provider_metadata(
+            ProviderResponseMetadata::new(value).ok_or_else(|| invalid(sent))?,
+        );
+    }
+    let previous = current.timings();
+    let update = update.timings;
+    let timings = ProviderResponseTimings {
+        transport_decision_wait_ms: update
+            .transport_decision_wait_ms
+            .or(previous.transport_decision_wait_ms),
+        connect_ms: update.connect_ms.or(previous.connect_ms),
+        headers_ms: update.headers_ms.or(previous.headers_ms),
+        first_event_ms: update.first_event_ms.or(previous.first_event_ms),
+        first_reasoning_ms: update.first_reasoning_ms.or(previous.first_reasoning_ms),
+        first_text_ms: update.first_text_ms.or(previous.first_text_ms),
+        first_token_ms: update.first_token_ms.or(previous.first_token_ms),
+        provider_processing_ms: update
+            .provider_processing_ms
+            .or(previous.provider_processing_ms),
+        upstream_response_ms: update
+            .upstream_response_ms
+            .or(previous.upstream_response_ms),
+        upstream_api_overhead_ms: update
+            .upstream_api_overhead_ms
+            .or(previous.upstream_api_overhead_ms),
+        upstream_engine_ms: update.upstream_engine_ms.or(previous.upstream_engine_ms),
+    };
+    Ok(current.with_timings(timings))
+}
+
+pub(super) fn observe_output(
+    current: ProviderResponseObservation,
+    facts: &[GatewayEvent],
+    has_wire: bool,
+    started_at: std::time::Instant,
+) -> ProviderResponseObservation {
+    let elapsed = u64::try_from(started_at.elapsed().as_millis()).unwrap_or(u64::MAX);
+    let mut timings = current.timings();
+    if has_wire || !facts.is_empty() {
+        timings.first_event_ms.get_or_insert(elapsed);
+    }
+    // 内容边界来自插件报告的 canonical 事实，不按上游私有事件名猜测
+    for fact in facts {
+        let has_output = match fact {
+            GatewayEvent::TextDelta(delta) if !delta.text.is_empty() => {
+                timings.first_text_ms.get_or_insert(elapsed);
+                true
+            }
+            GatewayEvent::ReasoningDelta(delta) if !delta.text.is_empty() => {
+                timings.first_reasoning_ms.get_or_insert(elapsed);
+                true
+            }
+            GatewayEvent::ToolCallDelta(delta) => {
+                !delta.arguments_delta.is_empty()
+                    || delta.name.as_ref().is_some_and(|name| !name.is_empty())
+            }
+            _ => false,
+        };
+        if has_output {
+            timings.first_token_ms.get_or_insert(elapsed);
+        }
+    }
+    current.with_timings(timings)
+}
diff --git a/backend/crates/gateway-plugin/runtime/src/callback/facts.rs b/backend/crates/gateway-plugin/runtime/src/callback/facts.rs
index cc12d9498..d10929a9a 100644
--- a/backend/crates/gateway-plugin/runtime/src/callback/facts.rs
+++ b/backend/crates/gateway-plugin/runtime/src/callback/facts.rs
@@ -18,7 +18,7 @@ use gateway_plugin_sdk::{
             FinishReason as WireFinishReason, Usage as WireUsage, WireEvent, WirePayload,
             facts::{self, ExecutionCost, ExecutionFacts},
         },
-        observation::{RequestMoney, RequestTimings},
+        observation::RequestMoney,
     },
 };
 
@@ -120,7 +120,7 @@ fn observation(value: &ProviderResponseObservation) -> facts::ResponseObservatio
         request_id: value.request_id().map(|value| value.as_str().to_owned()),
         service_tier: value.service_tier().map(str::to_owned),
         upstream_response_model: value.upstream_response_model().map(str::to_owned),
-        timings: RequestTimings {
+        timings: facts::ResponseTimings {
             transport_decision_wait_ms: timings.transport_decision_wait_ms,
             connect_ms: timings.connect_ms,
             headers_ms: timings.headers_ms,
@@ -132,7 +132,6 @@ fn observation(value: &ProviderResponseObservation) -> facts::ResponseObservatio
             upstream_response_ms: timings.upstream_response_ms,
             upstream_api_overhead_ms: timings.upstream_api_overhead_ms,
             upstream_engine_ms: timings.upstream_engine_ms,
-            latency_ms: None,
         },
         client_headers: value
             .client_headers()
@@ -271,6 +270,12 @@ fn project_wire(wire: &ProtocolWireEvent) -> Result<WireEvent, PluginFault> {
         WirePayload::RawJson {
             body: body.to_vec(),
         }
+    } else if let Some(message) = wire.raw_websocket_message() {
+        WirePayload::WebSocket {
+            event: wire.event_type().map(str::to_owned),
+            data: wire.has_json_data().then(|| wire.data().clone()),
+            message: message.as_bytes().to_vec(),
+        }
     } else if wire.has_json_data() {
         WirePayload::Json {
             event: wire.event_type().map(str::to_owned),
@@ -283,15 +288,6 @@ fn project_wire(wire: &ProtocolWireEvent) -> Result<WireEvent, PluginFault> {
         WirePayload::RawSse {
             frame: frame.to_vec(),
         }
-    } else if let Some(message) = wire.raw_websocket_message() {
-        // 子模型回调采用 SDK 的事件流表达，不把 WebSocket 文本冒充 HTTP body
-        WirePayload::RawSse {
-            frame: gateway_protocol::openai::sse::encode_sse_event(
-                wire.event_type().unwrap_or_default(),
-                message,
-            )
-            .into_bytes(),
-        }
     } else {
         return Err(PluginFault::new(
             ErrorCode::Fault,
diff --git a/backend/crates/gateway-plugin/runtime/src/callback/http.rs b/backend/crates/gateway-plugin/runtime/src/callback/http.rs
index 24a950353..206435f98 100644
--- a/backend/crates/gateway-plugin/runtime/src/callback/http.rs
+++ b/backend/crates/gateway-plugin/runtime/src/callback/http.rs
@@ -58,16 +58,12 @@ impl HttpCallbacks<'_> {
         );
         let response = response.map_err(http_error)?;
         let mut body = response.body;
-        let headers = response
-            .headers
-            .into_iter()
-            .map(|(name, value)| String::from_utf8(value).map(|value| (name, value)))
-            .collect::<Result<Vec<_>, _>>()
-            .map_err(|_| sent_fault(ErrorCode::Fault, "HTTP header cannot be represented"))?;
         let mut result = wire::HttpResponse {
             status: response.status,
-            headers,
+            http_version: format!("{:?}", response.version),
+            headers: response.headers,
             stream: None,
+            trailers: Vec::new(),
         };
         let mut payload = Vec::new();
         if streaming {
@@ -103,6 +99,7 @@ impl HttpCallbacks<'_> {
                 }
                 payload.extend_from_slice(&chunk);
             }
+            result.trailers = body.take_trailers();
         }
         Ok(RpcReply {
             result: serde_json::to_value(result).map_err(|_| invalid())?,
@@ -124,11 +121,7 @@ impl HttpCallbacks<'_> {
                     HttpRequest {
                         method: request.method,
                         url: request.url,
-                        headers: request
-                            .headers
-                            .into_iter()
-                            .map(|(name, value)| (name, value.into_bytes()))
-                            .collect(),
+                        headers: request.headers,
                         maximum_body_bytes: self.body_limit(&payload)?,
                         body: payload,
                     },
@@ -163,6 +156,9 @@ impl HttpCallbacks<'_> {
                     _ = closed.wait_for(|closed| *closed) => Err(denied()),
                     result = body.as_mut().ok_or_else(denied)?.read_with_timeout(request.maximum_bytes as usize, timeout) => result.map_err(http_error),
                 };
+                let trailers = body
+                    .as_mut()
+                    .map_or_else(Vec::new, |body| body.take_trailers());
                 if !matches!(&result, Ok(Some(_))) {
                     body.take();
                     stream.closed.send_replace(true);
@@ -174,7 +170,7 @@ impl HttpCallbacks<'_> {
                 }
                 let chunk = result?;
                 Ok(RpcReply {
-                    result: serde_json::json!({"eof":chunk.is_none()}),
+                    result: serde_json::json!({"eof":chunk.is_none(), "trailers": trailers}),
                     payload: chunk.map_or_else(Vec::new, |chunk| chunk.to_vec()),
                 })
             }
@@ -231,6 +227,10 @@ pub(super) fn http_error(error: HttpError) -> PluginFault {
         HttpErrorKind::Response => (ErrorCode::Upstream, "目标网站响应读取失败"),
     };
     let mut fault = PluginFault::new(code, message);
+    fault.details = Some(serde_json::json!({
+        "kind": error.kind(),
+        "reason": error.reason(),
+    }));
     fault.send_state = match error.send_state {
         UpstreamSendState::NotSent => SendState::NotSent,
         UpstreamSendState::Sent => SendState::Sent,
diff --git a/backend/crates/gateway-plugin/runtime/src/callback/upstream/mod.rs b/backend/crates/gateway-plugin/runtime/src/callback/upstream/mod.rs
index 66154e733..8b0308466 100644
--- a/backend/crates/gateway-plugin/runtime/src/callback/upstream/mod.rs
+++ b/backend/crates/gateway-plugin/runtime/src/callback/upstream/mod.rs
@@ -159,12 +159,7 @@ impl ManagedUpstream {
                 (name, value.to_vec())
             })
             .collect();
-        headers.extend(
-            request
-                .headers
-                .into_iter()
-                .map(|(name, value)| (name, value.into_bytes())),
-        );
+        headers.extend(request.headers);
         Ok((
             HttpRequest {
                 method: request.method,
@@ -237,7 +232,7 @@ pub(super) async fn dispatch(
                 return Err(denied());
             }
             let mut socket = managed.socket.try_lock().map_err(|_| denied())?;
-            if let Some((url, _)) = socket
+            if let Some((url, connection)) = socket
                 .as_ref()
                 .filter(|(_, connection)| !connection.is_closed())
             {
@@ -245,7 +240,13 @@ pub(super) async fn dispatch(
                     return Err(denied());
                 }
                 return Ok(RpcReply {
-                    result: serde_json::json!({"status":101,"headers":[],"stream":null}),
+                    result: serde_json::json!({
+                        "status": 101,
+                        "http_version": format!("{:?}", connection.handshake_version()),
+                        "headers": connection.handshake_headers(),
+                        "stream": null,
+                        "trailers": [],
+                    }),
                     payload: vec![],
                 });
             }
@@ -266,12 +267,8 @@ pub(super) async fn dispatch(
                     .map_or_else(|error| error.send_state, |_| UpstreamSendState::Sent),
             );
             let mut response = response.map_err(super::http::http_error)?;
-            let headers = response
-                .headers
-                .into_iter()
-                .map(|(name, value)| String::from_utf8(value).map(|value| (name, value)))
-                .collect::<Result<Vec<_>, _>>()
-                .map_err(|_| invalid())?;
+            let headers = response.headers;
+            let mut trailers = Vec::new();
             let mut payload = Vec::new();
             if let Some(body) = response.body.as_mut() {
                 while let Some(chunk) = body
@@ -284,6 +281,7 @@ pub(super) async fn dispatch(
                     }
                     payload.extend_from_slice(&chunk);
                 }
+                trailers = body.take_trailers();
             }
             if let Some(connection) = response.connection {
                 *socket = Some((url, Arc::new(connection)));
@@ -291,8 +289,10 @@ pub(super) async fn dispatch(
             Ok(RpcReply {
                 result: serde_json::to_value(wire::HttpResponse {
                     status: response.status,
+                    http_version: format!("{:?}", response.version),
                     headers,
                     stream: None,
+                    trailers,
                 })
                 .map_err(|_| invalid())?,
                 payload,
diff --git a/backend/crates/gateway-plugin/runtime/tests/generation/policy/facts/mod.rs b/backend/crates/gateway-plugin/runtime/tests/generation/policy/facts/mod.rs
index 22c595ffb..5c9366196 100644
--- a/backend/crates/gateway-plugin/runtime/tests/generation/policy/facts/mod.rs
+++ b/backend/crates/gateway-plugin/runtime/tests/generation/policy/facts/mod.rs
@@ -102,7 +102,21 @@ async fn reading_websocket_facts_preserves_original_text_and_opaque_messages() {
         );
     }
     assert_eq!(delivered, messages);
-    assert_eq!(marker_lines(&marker, 4).await.len(), 4);
+    let snapshots = marker_lines(&marker, 4).await;
+    for (snapshot, original) in snapshots[1..].iter().zip(&messages) {
+        let event: gateway_plugin_sdk::call::model::ExecutionEvent =
+            serde_json::from_value(snapshot.clone()).unwrap();
+        let gateway_plugin_sdk::call::model::WirePayload::WebSocket { data, message, .. } =
+            event.wire.unwrap().payload
+        else {
+            panic!("WebSocket facts must retain their original transport");
+        };
+        assert_eq!(message, original.as_bytes());
+        assert_eq!(
+            data.is_some(),
+            serde_json::from_str::<serde_json::Value>(original).is_ok()
+        );
+    }
     drop(stream);
     runtime.shutdown().await;
 }
diff --git a/backend/crates/gateway-plugin/runtime/tests/generation/upstream_adapter.rs b/backend/crates/gateway-plugin/runtime/tests/generation/upstream_adapter.rs
index 3c1ce5ba5..c546995e4 100644
--- a/backend/crates/gateway-plugin/runtime/tests/generation/upstream_adapter.rs
+++ b/backend/crates/gateway-plugin/runtime/tests/generation/upstream_adapter.rs
@@ -87,7 +87,7 @@ impl UpstreamAccountConnection for Account {
 fn configuration(base: &str, streaming: bool) -> Value {
     json!({
         "upstream_registration":{"adapters":[{"id":"example","provider":"openai","base_url":format!("{base}/adapter/"),"paths":[{"path":"responses","purpose":"inference"}],"authentication_kinds":["oauth"],"transport":if streaming {"http_sse"} else {"http_json"},"protocol":"openai"}]},
-        "upstream_callbacks":[{"method":if streaming {"host.upstream.http.do_stream"} else {"host.upstream.http.do"},"params":{"method":"POST","path":"responses","headers":[["content-type","application/json"]]},"body":"request"}],
+        "upstream_callbacks":[{"method":if streaming {"host.upstream.http.do_stream"} else {"host.upstream.http.do"},"params":{"method":"POST","path":"responses","headers":[["content-type",b"application/json".to_vec()]]},"body":"request"}],
         "upstream_events":[
             {"event":{"facts":[{"type":"started","id":"response-one","model":"native-model"}]}},
             {"event":{"facts":[{"type":"usage","usage":{"input_tokens":7,"output_tokens":2}},{"type":"completed","id":"response-one","model":"native-model","reason":"stop"}]},"wire":if streaming {"sse"} else {"http"},"continuation":{"scope":"persisted","upstream_response_id":"private-upstream-id","state":{"cursor":"opaque"}}}
@@ -457,6 +457,193 @@ async fn adapter_sse_outlives_rpc_budget_and_cancellation_closes_its_resources()
     }
 }
 
+#[tokio::test]
+async fn adapter_preserves_parsed_and_opaque_websocket_messages() {
+    use gateway_plugin_sdk::call::model::{WireEvent, WirePayload};
+    let messages = [
+        "{ \"type\": \"future.event\", \"value\": 1e3 }".to_owned(),
+        "opaque future message".to_owned(),
+        format!("{{\"value\":{}0{}}}", "[".repeat(140), "]".repeat(140)),
+    ];
+    let mut config = configuration("http://127.0.0.1:1", false);
+    config["upstream_callbacks"] = json!([]);
+    config["upstream_registration"]["adapters"][0]["transport"] = json!("websocket");
+    let mut events = vec![config["upstream_events"][0].clone()];
+    for message in &messages {
+        events.push(json!({"event":{"facts":[],"wire":WireEvent {
+            protocol: "openai".into(),
+            payload: WirePayload::WebSocket {
+                event: None, data: serde_json::from_str(message).ok(), message: message.as_bytes().to_vec(),
+            },
+        }}}));
+    }
+    events.push(json!({"event":config["upstream_events"][1]["event"].clone()}));
+    config["upstream_events"] = json!(events);
+    let (cache, _, runtime) = setup(config).await;
+    let generation = ExtensionPreparationPort::prepare(&runtime, ConfigRevision::new(1).unwrap())
+        .await
+        .unwrap();
+    let mut stream = execute(&runtime, &generation, Account::new(), "key-one", None);
+    let mut received = Vec::new();
+    while let Some(event) = stream.next().await {
+        if let Some(wire) = event.unwrap().wire_event() {
+            assert!(wire.raw_sse_frame().is_none());
+            received.push(wire.raw_websocket_message().unwrap().to_owned());
+        }
+    }
+    assert_eq!(received, messages);
+    drop(stream);
+    drop(generation);
+    runtime.shutdown().await;
+    super::wait_until_empty(cache.path()).await;
+}
+
+#[tokio::test]
+async fn managed_http_failure_keeps_host_reason_and_error_kind() {
+    let mut config = configuration("http://127.0.0.1:1", false);
+    config["upstream_callbacks"][0]["params"]["method"] = json!("invalid method");
+    let (cache, _, runtime) = setup(config).await;
+    let generation = ExtensionPreparationPort::prepare(&runtime, ConfigRevision::new(1).unwrap())
+        .await
+        .unwrap();
+    let mut stream = execute(&runtime, &generation, Account::new(), "key-one", None);
+    let error = stream.next().await.unwrap().unwrap_err();
+    let raw: Value = serde_json::from_str(error.raw_upstream_error().unwrap().as_str()).unwrap();
+    assert_eq!(raw["fault"]["details"]["kind"], "invalid_request");
+    assert_eq!(raw["fault"]["details"]["reason"], "method");
+    assert_eq!(error.send_state(), UpstreamSendState::NotSent);
+    drop(stream);
+    drop(generation);
+    runtime.shutdown().await;
+    super::wait_until_empty(cache.path()).await;
+}
+
+#[tokio::test]
+async fn adapter_preserves_observation_updates_and_measures_only_real_output() {
+    let mut config = configuration("http://127.0.0.1:1", true);
+    config["upstream_callbacks"] = json!([]);
+    config["upstream_events"] = json!([
+        {"event":{"facts":[{"type":"started","id":"response-one","model":"native-model"},
+            {"type":"content_added","index":0,"kind":"reasoning"}]},
+         "observation":{"transport":"http_sse","http_version":"HTTP/2","status_code":200,
+            "request_id":"upstream-494","upstream_response_model":"real-model",
+            "client_headers":[{"name":"x-bytes","value":[255]},{"name":"x-bytes","value":[254]}],
+            "provider_metadata":"{ \"future\": 1e3 }","timings":{"connect_ms":5,"headers_ms":8}}},
+        {"event":{"facts":[{"type":"reasoning_delta","index":0,"text":""}]}},
+        {"event":{"facts":[{"type":"reasoning_delta","index":0,"text":"thinking"}]}},
+        {"event":{"facts":[{"type":"content_added","index":1,"kind":"text"},{"type":"text_delta","index":1,"text":"answer"}]}},
+        {"event":{"facts":[]},"observation":{"transport":"http_sse","service_tier":"priority",
+            "timings":{"upstream_response_ms":1000,"upstream_api_overhead_ms":1.25,"upstream_engine_ms":998.75}}},
+        {"event":{"facts":[{"type":"usage","usage":{"input_tokens":7,"output_tokens":2}},
+            {"type":"completed","id":"response-one","model":"native-model","reason":"stop"}]},
+         "observation":{"transport":"http_sse","timings":{"first_token_ms":41}}}
+    ]);
+    let (cache, _, runtime) = setup(config).await;
+    let generation = ExtensionPreparationPort::prepare(&runtime, ConfigRevision::new(1).unwrap())
+        .await
+        .unwrap();
+    let context = context(&runtime, &generation, "key-one");
+    tokio::time::sleep(Duration::from_millis(50)).await;
+    let mut stream = execute_with_context(Account::new(), context, None);
+    let mut observations = Vec::new();
+    while let Some(event) = stream.next().await {
+        observations.push(event.unwrap().response_observation().unwrap().clone());
+    }
+    assert_eq!(observations.len(), 6);
+    assert_eq!(observations[0].timings().first_token_ms, None);
+    assert_eq!(observations[1].timings().first_token_ms, None);
+    let first_output = observations[2].timings().first_token_ms.unwrap();
+    assert!(first_output >= 50, "首字计时必须包括调用前的等待");
+    assert_eq!(
+        observations[2].timings().first_reasoning_ms,
+        Some(first_output)
+    );
+    assert!(observations[3].timings().first_text_ms.unwrap() >= first_output);
+    assert_eq!(observations[3].timings().upstream_response_ms, None);
+    let last = observations.last().unwrap();
+    assert_eq!(last.timings().first_token_ms, Some(41));
+    assert_eq!(last.timings().connect_ms, Some(5));
+    assert_eq!(last.timings().headers_ms, Some(8));
+    assert_eq!(last.timings().upstream_response_ms, Some(1000));
+    assert_eq!(last.timings().upstream_api_overhead_ms, Some(1.25));
+    assert_eq!(last.timings().upstream_engine_ms, Some(998.75));
+    assert_eq!(last.status_code(), Some(200));
+    assert_eq!(last.http_version().unwrap().as_str(), "HTTP/2");
+    assert_eq!(last.request_id().unwrap().as_str(), "upstream-494");
+    assert_eq!(last.upstream_response_model(), Some("real-model"));
+    assert_eq!(last.service_tier(), Some("priority"));
+    assert_eq!(
+        last.provider_metadata().unwrap().as_json(),
+        "{ \"future\": 1e3 }"
+    );
+    assert_eq!(last.client_headers().len(), 2);
+    assert_eq!(last.client_headers()[0].value().as_ref(), [255]);
+    assert_eq!(last.client_headers()[1].value().as_ref(), [254]);
+    drop(stream);
+    drop(generation);
+    runtime.shutdown().await;
+    super::wait_until_empty(cache.path()).await;
+}
+
+#[tokio::test]
+async fn managed_http_preserves_byte_headers_trailers_and_version_for_plugins() {
+    use tokio::io::{AsyncReadExt, AsyncWriteExt};
+    for streaming in [false, true] {
+        let listener = tokio::net::TcpListener::bind("127.0.0.1:0").await.unwrap();
+        let base = format!("http://{}", listener.local_addr().unwrap());
+        let upstream = tokio::spawn(async move {
+            let (mut socket, _) = listener.accept().await.unwrap();
+            let mut request = Vec::new();
+            while !request.windows(4).any(|part| part == b"\r\n\r\n") {
+                assert!(socket.read_buf(&mut request).await.unwrap() > 0);
+            }
+            socket.write_all(b"HTTP/1.1 200 OK\r\nX-Bytes: \xff\r\nX-Bytes: other\r\nTransfer-Encoding: chunked\r\nTrailer: x-tail\r\n\r\n").await.unwrap();
+            let body = if streaming { "data: {}\n\n" } else { "{}" };
+            socket
+                .write_all(format!("{:x}\r\n{body}\r\n", body.len()).as_bytes())
+                .await
+                .unwrap();
+            socket
+                .write_all(b"0\r\nx-tail: \xfe\r\nx-tail: final\r\n\r\n")
+                .await
+                .unwrap();
+        });
+        let records = tempfile::tempdir().unwrap();
+        let marker = records.path().join("http.jsonl");
+        let mut config = configuration(&base, streaming);
+        config["upstream_callbacks"][0]["body"] = json!("");
+        config["upstream_http_marker"] = json!(marker);
+        let (cache, _, runtime) = setup(config).await;
+        let generation =
+            ExtensionPreparationPort::prepare(&runtime, ConfigRevision::new(1).unwrap())
+                .await
+                .unwrap();
+        let mut stream = execute(&runtime, &generation, Account::new(), "key-one", None);
+        while let Some(event) = stream.next().await {
+            event.unwrap();
+        }
+        upstream.await.unwrap();
+        let observations: Vec<Value> = std::fs::read_to_string(&marker)
+            .unwrap()
+            .lines()
+            .map(|line| serde_json::from_str(line).unwrap())
+            .collect();
+        assert_eq!(observations[0]["http_version"], "HTTP/1.1");
+        let headers = observations[0]["headers"].as_array().unwrap();
+        assert!(headers.contains(&json!(["x-bytes", [255]])));
+        assert!(headers.contains(&json!(["x-bytes", b"other".to_vec()])));
+        let trailers = &observations.last().unwrap()["trailers"];
+        assert_eq!(
+            trailers,
+            &json!([["x-tail", [254]], ["x-tail", b"final".to_vec()]])
+        );
+        drop(stream);
+        drop(generation);
+        runtime.shutdown().await;
+        super::wait_until_empty(cache.path()).await;
+    }
+}
+
 #[tokio::test]
 async fn adapter_accepts_service_tier_resolved_by_the_terminal_response() {
     let server = MockServer::start().await;
@@ -466,8 +653,10 @@ async fn adapter_accepts_service_tier_resolved_by_the_terminal_response() {
         .mount(&server)
         .await;
     let mut config = configuration(&server.uri(), false);
-    config["upstream_events"][0]["service_tier"] = json!("auto");
-    config["upstream_events"][1]["service_tier"] = json!("default");
+    config["upstream_events"][0]["observation"] =
+        json!({"transport":"http_json","service_tier":"auto"});
+    config["upstream_events"][1]["observation"] =
+        json!({"transport":"http_json","service_tier":"default"});
     let (cache, _, runtime) = setup(config).await;
     let generation = ExtensionPreparationPort::prepare(&runtime, ConfigRevision::new(1).unwrap())
         .await
@@ -509,6 +698,7 @@ async fn adapter_failure_preserves_upstream_diagnostics_and_native_feedback() {
     let mut config = configuration(&server.uri(), false);
     config["upstream_events"] = json!([{
         "event": {"facts": []},
+        "observation":{"transport":"http_json","status_code":429,"timings":{"upstream_response_ms":230}},
         "failure": {
             "kind": "rate_limited", "status": 429, "retry_after_ms": 1500,
             "message": "fixture rate limit", "code": "rate_limit_exceeded"
@@ -520,7 +710,7 @@ async fn adapter_failure_preserves_upstream_diagnostics_and_native_feedback() {
         .unwrap();
     let account = Account::new();
     let mut stream = execute(&runtime, &generation, Arc::clone(&account), "key-one", None);
-    let error = stream.next().await.unwrap().unwrap_err();
+    let mut error = stream.next().await.unwrap().unwrap_err();
     assert_eq!(error.kind(), ProviderErrorKind::RateLimited);
     assert_eq!(error.send_state(), UpstreamSendState::Sent);
     assert_eq!(error.upstream_status(), Some(429));
@@ -529,6 +719,10 @@ async fn adapter_failure_preserves_upstream_diagnostics_and_native_feedback() {
         "rate_limit_exceeded"
     );
     assert_eq!(error.retry_after(), Some(Duration::from_millis(1500)));
+    let observations = error.take_atomic_client_events();
+    let observation = observations[0].response_observation().unwrap();
+    assert_eq!(observation.status_code(), Some(429));
+    assert_eq!(observation.timings().upstream_response_ms, Some(230));
     let snapshot = error.stable_snapshot();
     let raw: Value = serde_json::from_str(snapshot.raw_upstream_error().unwrap().as_str()).unwrap();
     assert_eq!(raw["message"], "fixture rate limit");
@@ -584,7 +778,8 @@ async fn adapter_overrides_auth_headers_and_rejects_stale_continuation() {
     super::wait_until_empty(cache.path()).await;
 
     let mut config = configuration(&server.uri(), false);
-    config["upstream_callbacks"][0]["params"]["headers"] = json!([["Authorization", "injected"]]);
+    config["upstream_callbacks"][0]["params"]["headers"] =
+        json!([["Authorization", b"injected".to_vec()]]);
     let (cache, _, runtime) = setup(config).await;
     let generation = ExtensionPreparationPort::prepare(&runtime, ConfigRevision::new(1).unwrap())
         .await
@@ -636,8 +831,13 @@ async fn adapter_does_not_publish_completed_when_rpc_fails_after_terminal_frame(
             .contains("fixture terminal fault")
     );
     assert!(!format!("{error:?}").contains("fixture terminal fault"));
-    let facts = error
-        .take_atomic_client_events()
+    let atomic = error.take_atomic_client_events();
+    assert!(
+        atomic
+            .iter()
+            .any(|event| event.response_observation().is_some())
+    );
+    let facts = atomic
         .into_iter()
         .flat_map(|event| event.into_parts().0)
         .collect::<Vec<_>>();
@@ -726,11 +926,12 @@ async fn websocket_continuation_reuses_exact_connection_and_consumes_each_handle
         let (socket, _) = listener.accept().await.unwrap();
         let mut ws = tokio_tungstenite::accept_hdr_async(
             socket,
-            |request: &tokio_tungstenite::tungstenite::handshake::server::Request, response| {
+            |request: &tokio_tungstenite::tungstenite::handshake::server::Request, mut response: tokio_tungstenite::tungstenite::handshake::server::Response| {
                 assert_eq!(
                     request.headers()["authorization"],
                     "Bearer fixture-native-token"
                 );
+                response.headers_mut().append("x-bytes", http::HeaderValue::from_static("handshake-value"));
                 Ok(response)
             },
         )
@@ -745,7 +946,10 @@ async fn websocket_continuation_reuses_exact_connection_and_consumes_each_handle
         }
         assert!(ws.next().await.is_none_or(|result| result.is_err()));
     });
+    let records = tempfile::tempdir().unwrap();
+    let marker = records.path().join("handshake.jsonl");
     let mut config = configuration(&base, false);
+    config["upstream_http_marker"] = json!(marker);
     config["upstream_registration"]["adapters"][0]["transport"] = json!("websocket");
     config["upstream_events"][1]["continuation"]["scope"] = json!("connection_local");
     config["upstream_callbacks"] = json!([
@@ -789,6 +993,22 @@ async fn websocket_continuation_reuses_exact_connection_and_consumes_each_handle
         event.unwrap();
     }
     drop(second);
+    let handshakes: Vec<Value> = std::fs::read_to_string(&marker)
+        .unwrap()
+        .lines()
+        .map(|line| serde_json::from_str::<Value>(line).unwrap())
+        .filter(|value| value["status"] == 101)
+        .collect();
+    assert_eq!(handshakes.len(), 2);
+    assert_eq!(handshakes[0], handshakes[1]);
+    assert_eq!(handshakes[0]["http_version"], "HTTP/1.1");
+    assert!(
+        handshakes[0]["headers"]
+            .as_array()
+            .unwrap()
+            .contains(&json!(["x-bytes", b"handshake-value".to_vec()]))
+    );
+
     let error = execute(
         &runtime,
         &generation,
diff --git a/backend/crates/gateway-plugin/runtime/tests/package/compatibility.rs b/backend/crates/gateway-plugin/runtime/tests/package/compatibility.rs
index bd892c43b..5e59fa96f 100644
--- a/backend/crates/gateway-plugin/runtime/tests/package/compatibility.rs
+++ b/backend/crates/gateway-plugin/runtime/tests/package/compatibility.rs
@@ -48,7 +48,8 @@ fn host_compatibility_requires_the_trusted_middleware_contract() {
     assert!(compatibility.supports_capability("middleware", 5));
     assert!(!compatibility.supports_capability("upstream_adapter", 1));
     assert!(!compatibility.supports_capability("upstream_adapter", 2));
-    assert!(compatibility.supports_capability("upstream_adapter", 3));
+    assert!(!compatibility.supports_capability("upstream_adapter", 3));
+    assert!(compatibility.supports_capability("upstream_adapter", 4));
     assert!(!compatibility.supports_capability("openai", 1));
 }
 
diff --git a/backend/crates/gateway-plugin/runtime/tests/support/worker/mod.rs b/backend/crates/gateway-plugin/runtime/tests/support/worker/mod.rs
index 4a4245edf..a1607c334 100644
--- a/backend/crates/gateway-plugin/runtime/tests/support/worker/mod.rs
+++ b/backend/crates/gateway-plugin/runtime/tests/support/worker/mod.rs
@@ -1683,6 +1683,7 @@ impl Peer {
             }))
             .await?;
             for (callback, (result, payload)) in callbacks.iter().zip(results) {
+                self.append_observation_marker("upstream_http_marker", &result);
                 if let Some(stream) = result["stream"].as_str() {
                     response_body.clear();
                     let read_method = if callback["method"] == "host.http.do_stream" {
@@ -1699,6 +1700,7 @@ impl Peer {
                                 vec![],
                             )
                             .await?;
+                        self.append_observation_marker("upstream_http_marker", &result);
                         response_body.extend(chunk);
                         if result["eof"] == true {
                             break;
@@ -1715,7 +1717,9 @@ impl Peer {
             let mut message = UpstreamAdapterEvent::new(
                 serde_json::from_value::<ExecutionEvent>(event["event"].clone()).unwrap(),
             );
-            message.service_tier = event["service_tier"].as_str().map(str::to_owned);
+            message.observation = event
+                .get("observation")
+                .map(|value| serde_json::from_value(value.clone()).unwrap());
             message.continuation = event
                 .get("continuation")
                 .map(|value| serde_json::from_value(value.clone()).unwrap());
diff --git a/backend/crates/gateway-plugin/sdk/README.md b/backend/crates/gateway-plugin/sdk/README.md
index 89a218565..7f1c03289 100644
--- a/backend/crates/gateway-plugin/sdk/README.md
+++ b/backend/crates/gateway-plugin/sdk/README.md
@@ -8,8 +8,8 @@ SDK 处于实验阶段；能力是否可用取决于宿主支持、清单声明
 | --- | --- | --- |
 | SDK 包 | `0.1.0` | Rust 开发依赖，独立于网关版本 |
 | `manifestVersion` | `2` | 安装清单格式 |
-| `package.protocolVersion` | `4` | 宿主与插件的进程通信 |
-| `contributes.*.version` | `middleware` 为 `5`，`scheduler` / `upstream_adapter` 为 `3`，其余见声明 | 单项能力合同，见[扩展项声明](docs/manifest.md#扩展项简写) |
+| `package.protocolVersion` | `5` | 宿主与插件的进程通信 |
+| `contributes.*.version` | `middleware` 为 `5`，`scheduler` 为 `3`，`upstream_adapter` 为 `4`，其余见声明 | 单项能力合同，见[扩展项声明](docs/manifest.md#扩展项简写) |
 
 开发顺序：编写[清单](docs/manifest.md) → 实现[能力处理器](docs/capabilities.md) → 用 [CLI](../../../apps/plugin-cli/README.md)
 打包 → [安装、配置与启用](../../../../docs/plugins.md)。完整可运行示例位于独立仓库
@@ -92,7 +92,7 @@ sequenceDiagram
 首个流式结果成功返回后，正文、会话及关联回调由资源取消与终态回收，不再沿用初始期限。
 普通调用继续使用原期限；资源模式仍保留并发容量、Credit、父调用关联及取消约束
 
-协议 v4 的帧头由大端 `u32` 元数据长度和大端 `u64` 载荷长度组成，随后依次传输 JSON 元数据和二进制载荷。
+协议 v5 的帧头由大端 `u32` 元数据长度和大端 `u64` 载荷长度组成，随后依次传输 JSON 元数据和二进制载荷。
 元数据保留 64 KiB 上限；正文总量不由插件传输层额外限制。传输内部以 256 KiB 为单次 I/O 块，
 不逐块发起 RPC，也不要求插件配置与宿主同步。读取按进度扩展缓冲区；截断正文必须报错，不能作为完整消息交付
 
diff --git a/backend/crates/gateway-plugin/sdk/docs/capabilities.md b/backend/crates/gateway-plugin/sdk/docs/capabilities.md
index 43f9aeeae..fb92115f1 100644
--- a/backend/crates/gateway-plugin/sdk/docs/capabilities.md
+++ b/backend/crates/gateway-plugin/sdk/docs/capabilities.md
@@ -221,8 +221,9 @@ HTTP 和 request 中间件中的模型回调沿用有效请求设置；attempt
 同一父调用内复用 Key 时不受后续配置发布影响；父调用结束后重新绑定才读取新配置，无需重新启用插件
 
 `ModelExecuteRequest.client_key_id` 可显式选择执行 Key，不与触发插件的父请求 Key 求交。省略时继承现有模型父请求身份；没有模型父请求的调用必须提供 Key。所选 Key 的启用状态、模型规则、账号组、并发和预算继续由 Core 执行
+嵌套调用与显式 Key 调用继承入口快照中的客户端 IP 和 User-Agent；没有 HTTP 入口的后台调用保持来源未知
 
-模型事件使用 `call::model::ExecutionEvent` 的 `GPE2` 封套，将 canonical 事实、宿主事实和原始 wire 分段编码。`host` 包含费用及定价明细、完整上游观测、会话状态和改写来源；当前 wire 与原始 wire 都保留二进制字节。读取或回传快照不触发第二次计费、续接提交或发送状态更新。
+模型事件使用 `call::model::ExecutionEvent` 的 `GPE2` 封套，将 canonical 事实、宿主事实和原始 wire 分段编码。`host` 包含费用及定价明细、完整上游观测、会话状态和改写来源；当前 wire 与原始 wire 都保留二进制字节；`WirePayload::WebSocket` 独立保留完整文本消息和可选 JSON 视图，不把未知文本改造成 SSE。读取或回传快照不触发第二次计费、续接提交或发送状态更新。
 非流式模型结果通过 `ModelEventBatch` 的 `HME1` 有界二进制批次返回；流式句柄只在创建它的父调用内有效。
 发起回调的插件实例在子调用中跳过，Core 在单次执行的子调用图内限制递归深度、总数和并发，避免插件调用自身形成环。
 显式 Key 的身份快照随父调用保留，每个模型执行独立计时并建立调用图；父调用结束会取消尚未完成的执行，连接空闲不消耗后续执行的期限
@@ -237,6 +238,10 @@ HTTP 和 request 中间件中的模型回调沿用有效请求设置；attempt
 父调用取消或结束时宿主回收未关闭的流，SDK 不后台预读。取消一次 `read()` 等待后，再次读取会继续同一次操作，不丢弃已经返回的分块。
 账号上游使用同一正文对象，使用方式见[受管上游](upstream-adapters.md)
 
+请求和响应 `headers` 使用 `(String, Vec<u8>)`，保留重复头与非 UTF-8 值；响应另有 `http_version`。
+读取至 EOF 后通过 `body.trailers()` 取得完整尾部头，读取中只包含已收到的部分。
+`collect` 只返回正文，需要尾部头时使用 `read`；底层非流式回复携带 `trailers`，流式回复在 `stream_read` 中交付尾部头
+
 `host.http.do/do_stream` 和流读取／关闭方法提供受管 HTTP。正文使用独立二进制载荷，Host 统一执行代理、超时和背压，不按插件身份限制目标地址段。DNS 由宿主解析，连接保留目标 Host 与 TLS 域名校验；重定向不会自动跟随。
 受管 HTTP（包括 `host.upstream.http.do[_stream]`）的请求正文按有效设置
 `responses_max_decompressed_body_bytes` 检查。HTTP/request 中间件使用有效请求设置；
@@ -247,7 +252,7 @@ attempt、路由、调度、重试和上游适配器使用执行冻结预算，
 这里检查实际出站正文，内部模型请求的转义余量不增加出站预算。响应缓冲及流窗口不随请求配置扩大
 普通调用使用父调用剩余期限；已建立的受管 HTTP 流或 WebSocket 会话按每次网络操作计时，仍受父资源取消约束
 
-期限耗尽返回 `timeout`，解析、连接或响应读取失败返回 `upstream`。错误消息提供分类提示，响应正文仍由插件读取；协议故障不应直接展示内部消息
+期限耗尽返回 `timeout`，解析、连接或响应读取失败返回 `upstream`。错误 `details.kind` 和 `details.reason` 保留 Host 的故障分类与原因标记，响应正文仍由插件读取；协议故障不应直接展示内部消息
 
 ## 请求扩展
 
diff --git a/backend/crates/gateway-plugin/sdk/docs/manifest.md b/backend/crates/gateway-plugin/sdk/docs/manifest.md
index e9ed902e7..96dda0378 100644
--- a/backend/crates/gateway-plugin/sdk/docs/manifest.md
+++ b/backend/crates/gateway-plugin/sdk/docs/manifest.md
@@ -34,7 +34,7 @@
 
 清单与宿主兼容声明的格式版本为 `2`，进程协议版本为 `4`；清单不接受 `permissions` 字段
 
-RPC v4 的帧布局、编码与会话生命周期见[通信边界](../README.md#通信边界)。插件使用 SDK 构建，并由 CLI 生成包含协议版本的安装包
+RPC v5 的帧布局、编码与会话生命周期见[通信边界](../README.md#通信边界)。插件使用 SDK 构建，并由 CLI 生成包含协议版本的安装包
 
 安装包清单校验与已安装元数据读取是两个边界：清单拒绝未知字段，并检查必需字段、类型和版本；
 数据库元数据的读取规则见[制品与发布](../../../../../docs/architecture.md#制品配置与发布)，不改变清单或 RPC 合同。
@@ -49,7 +49,7 @@ RPC v4 的帧布局、编码与会话生命周期见[通信边界](../README.md#
 
 ## 扩展项简写
 
-作者声明可以省略 `id` 和固定阶段。版本字段省略时为 `1`，适用于 `model_catalog`、`command_line` 和 `maintenance`；其余能力显式声明当前版本：`middleware` 为 `5`，`scheduler` 和 `upstream_adapter` 为 `3`，认证、路由、重试、观察和管理能力为 `2`：
+作者声明可以省略 `id` 和固定阶段。版本字段省略时为 `1`，适用于 `model_catalog`、`command_line` 和 `maintenance`；其余能力显式声明当前版本：`middleware` 为 `5`，`scheduler` 为 `3`，`upstream_adapter` 为 `4`，认证、路由、重试、观察和管理能力为 `2`：
 
 ```json
 {
diff --git a/backend/crates/gateway-plugin/sdk/docs/upstream-adapters.md b/backend/crates/gateway-plugin/sdk/docs/upstream-adapters.md
index c6fa994bd..bfebde7a2 100644
--- a/backend/crates/gateway-plugin/sdk/docs/upstream-adapters.md
+++ b/backend/crates/gateway-plugin/sdk/docs/upstream-adapters.md
@@ -1,12 +1,12 @@
 # 受管上游适配器
 
-`upstream_adapter` v3 是现有生成请求洋葱链的终端，只挂到内置 `openai` / `xai`，不替换独立图像或压缩端点。
+`upstream_adapter` v4 是现有生成请求洋葱链的终端，只挂到内置 `openai` / `xai`，不替换独立图像或压缩端点。
 插件定义 Base URL、业务路径、请求编码和响应解析；宿主继续持有身份、ACL、选号、租约、凭据、代理、重试和账本。
 普通请求或响应加工使用 [middleware](capabilities.md#洋葱中间件)
 
 ## 声明与绑定
 
-清单声明 `upstream_adapter`、版本 `3`，以及匹配的 `inputFormats` / `outputFormats`；固定阶段 `upstream` 由作者清单规范化入口生成。
+清单声明 `upstream_adapter`、版本 `4`，以及匹配的 `inputFormats` / `outputFormats`；固定阶段 `upstream` 由作者清单规范化入口生成。
 使用 `PluginBuilder::on` 注册
 `methods::UPSTREAM_ADAPTER_REGISTER` 与 `methods::UPSTREAM_ADAPTER_EXECUTE`
 
@@ -55,15 +55,15 @@ while let Some(bytes) = body.read().await? {
 
 正文与普通 `HostClient::http` 共用 `read`、`collect(maximum_bytes)` 和 `close`；每次读取至多 64 KiB，
 不后台预读，不向作者暴露流 ID。提前关闭及收集超限会释放正文，父调用结束或取消时宿主兜底回收。
-JSON 可有界收集，SSE 应增量解析，不能把读取块当成完整事件
+JSON 可有界收集，SSE 应增量解析，不能把读取块当成完整事件。HTTP 版本、字节头及 trailers 的读取见[网络](capabilities.md#网络)
 
 WebSocket 使用 `call.host.upstream_websocket(request).await?`，返回
-`UpstreamWebSocketUpgrade::Connected { headers, connection }` 或 `Rejected { status, headers, body }`。
+`UpstreamWebSocketUpgrade::Connected { http_version, headers, connection }` 或 `Rejected { status, http_version, headers, body, trailers }`。
 连接提供 `send(kind, body)`、`read()`、`close()`，支持完整文本与二进制消息，Ping/Pong 由宿主处理。
 读写可同时进行，等待上游输出时可以发送控制消息；有先后依赖的消息须顺序等待 `send`。
 同一时刻只允许一个读取者；取消一次 `read()` 等待后，再次读取继续同一次接收，不跳过消息。
 `close` 会唤醒在途操作并释放连接，并发关闭等待同一次完成；读取超时或连接错误也会关闭连接。
-握手拒绝保留 HTTP 状态和正文。底层 `host.upstream.*` 是 SDK 使用的资源协议
+握手拒绝保留 HTTP 状态、正文和 trailers；连接复用返回该连接最初的握手版本和响应头。底层 `host.upstream.*` 是 SDK 使用的资源协议
 
 ```rust,ignore
 let (message, ()) = tokio::try_join!(
@@ -80,16 +80,28 @@ let (message, ()) = tokio::try_join!(
 
 执行返回 `TypedReply<Empty>` 的有界 `ResponseStream`。每个 chunk 使用 `UpstreamAdapterEvent::encode()`，
 包含标准事实与可选原始 wire；事件需符合开始、内容、工具、用量和唯一结束的顺序。
-原始 JSON / SSE 字节与标准事实分开传递，错误使用 `UpstreamFailure`；宿主网络观测决定发送状态，插件不能将其降为未发送
+原始 JSON / SSE / WebSocket 字节与标准事实分开传递，错误使用 `UpstreamFailure`；宿主网络观测决定发送状态，插件不能将其降为未发送
 
 初始 RPC 回复受插件处理预算约束，默认上限为 120 秒；持续接收与解析应放在返回的流中。
 初始流式回复后，流的寿命遵循模型请求期限与取消，受管 HTTP 正文每次读取单独计时。
 持续有数据的流不会因初始 RPC 预算耗尽而结束，单次读取超时、客户端取消和流结束仍会回收资源
 
 用量提交标准 `Usage`，费用由内置 Provider 按已冻结的上游模型与价格计算，响应回显模型不切换计价。
-`service_tier` 记录上游回显的最新档位，允许上游在完成时确定实际值；OpenAI 计价沿用请求策略，回显档位仅作观测。
+`observation.service_tier` 记录上游回显的最新档位，允许上游在完成时确定实际值；OpenAI 计价沿用请求策略，回显档位仅作观测。
 未知用量或价格保留未知。宿主确认 RPC 正常结束后才发布完成事实，插件不能独立结算或组织换号重试
 
+`UpstreamAdapterEvent.observation` 接收 `ResponseObservation` 原始观测：HTTP 版本、状态、请求 ID、回显模型、
+服务档位、字节响应头、Provider JSON 元数据及 `ResponseTimings`。`transport` 与注册传输类型一致；
+已提供字段覆盖同次执行的旧值，缺失字段或空 headers 保留已有观测，允许独立的观测事件及失败事件携带观测。
+`ExecutionEvent.host` 是宿主结果的只读快照，不能用于回写费用、发送状态或续接
+
+`first_*_ms` 相对 Core 请求计时起点，包含本次请求的排队和重试等待。
+`request_elapsed_ms` 提供宿主发送本次适配器调用前已耗的时间，插件自行计时须对齐该基线；
+没有报告时 Runtime 在收到首事件、非空正文／推理／工具增量时记录相应时间。
+开始、心跳和空增量不构成首个输出，显式报告的计时优先于宿主缺省观测。
+`upstream_response_ms`、`upstream_api_overhead_ms`、`upstream_engine_ms` 只能来自上游原始计时，
+没有来源时保持缺失。观测合同不包含总耗时、速率或计算费用，这些由宿主业务层消费原始值计算
+
 成功终态可附带 `UpstreamContinuation`，包含上游响应 ID、不超过 32 KiB 的私有续接材料和作用域：
 
 - `persisted`：续接材料随宿主会话保存，仍受实例、代次、Key、账号和凭据版本限制
diff --git a/backend/crates/gateway-plugin/sdk/src/call/host.rs b/backend/crates/gateway-plugin/sdk/src/call/host.rs
index 0ca79fa39..8b9316035 100644
--- a/backend/crates/gateway-plugin/sdk/src/call/host.rs
+++ b/backend/crates/gateway-plugin/sdk/src/call/host.rs
@@ -63,15 +63,17 @@ pub struct HttpRequest {
     pub method: String,
     pub url: String,
     #[serde(default)]
-    pub headers: Vec<(String, String)>,
+    pub headers: Vec<(String, Vec<u8>)>,
 }
 
 #[derive(Clone, Serialize, Deserialize)]
 #[serde(deny_unknown_fields)]
 pub struct HttpResponse {
     pub status: u16,
-    pub headers: Vec<(String, String)>,
+    pub http_version: String,
+    pub headers: Vec<(String, Vec<u8>)>,
     pub stream: Option<String>,
+    pub trailers: Vec<(String, Vec<u8>)>,
 }
 
 /// 宿主账号列表查询；cursor 是上次结果返回的稳定账号 ID
diff --git a/backend/crates/gateway-plugin/sdk/src/call/model/codec.rs b/backend/crates/gateway-plugin/sdk/src/call/model/codec.rs
index 4c1228d7f..66c2f65a8 100644
--- a/backend/crates/gateway-plugin/sdk/src/call/model/codec.rs
+++ b/backend/crates/gateway-plugin/sdk/src/call/model/codec.rs
@@ -142,6 +142,7 @@ fn attach(target: Option<&mut Vec<u8>>, bytes: &[u8]) -> Result<(), ExecutionEnc
 fn raw_wire(wire: Option<&mut super::WireEvent>) -> Option<&mut Vec<u8>> {
     match &mut wire?.payload {
         WirePayload::Json { raw_sse, .. } => raw_sse.as_mut(),
+        WirePayload::WebSocket { message, .. } => Some(message),
         WirePayload::RawSse { frame } => Some(frame),
         WirePayload::RawJson { body } | WirePayload::RawBody { body } => Some(body),
     }
diff --git a/backend/crates/gateway-plugin/sdk/src/call/model/facts.rs b/backend/crates/gateway-plugin/sdk/src/call/model/facts.rs
index 8957f8b9b..63e327672 100644
--- a/backend/crates/gateway-plugin/sdk/src/call/model/facts.rs
+++ b/backend/crates/gateway-plugin/sdk/src/call/model/facts.rs
@@ -4,10 +4,7 @@
 use serde::{Deserialize, Serialize};
 
 use super::WireEvent;
-use crate::call::{
-    middleware::MiddlewareHeader,
-    observation::{RequestMoney, RequestTimings},
-};
+use crate::call::{middleware::MiddlewareHeader, observation::RequestMoney};
 
 #[derive(Clone, Default, PartialEq, Serialize, Deserialize)]
 #[serde(deny_unknown_fields)]
@@ -76,12 +73,60 @@ pub struct ResponseObservation {
     pub request_id: Option<String>,
     pub service_tier: Option<String>,
     pub upstream_response_model: Option<String>,
-    pub timings: RequestTimings,
+    #[serde(default)]
+    pub timings: ResponseTimings,
+    #[serde(default)]
     pub client_headers: Vec<MiddlewareHeader>,
     /// Provider 已产生的完整观测 JSON 文本，保留原始表示
     pub provider_metadata: Option<String>,
 }
 
+/// 原始上游计时；不包含最终总耗时、速率或费用等业务结果
+#[derive(Debug, Clone, Default, PartialEq, Serialize, Deserialize)]
+#[serde(deny_unknown_fields)]
+pub struct ResponseTimings {
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub transport_decision_wait_ms: Option<u64>,
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub connect_ms: Option<u64>,
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub headers_ms: Option<u64>,
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub first_event_ms: Option<u64>,
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub first_reasoning_ms: Option<u64>,
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub first_text_ms: Option<u64>,
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub first_token_ms: Option<u64>,
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub provider_processing_ms: Option<u64>,
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub upstream_response_ms: Option<u64>,
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub upstream_api_overhead_ms: Option<f64>,
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub upstream_engine_ms: Option<f64>,
+}
+
+impl ResponseObservation {
+    #[must_use]
+    pub fn new(transport: impl Into<String>) -> Self {
+        Self {
+            transport: transport.into(),
+            http_version: None,
+            websocket_pool: None,
+            status_code: None,
+            request_id: None,
+            service_tier: None,
+            upstream_response_model: None,
+            timings: ResponseTimings::default(),
+            client_headers: Vec::new(),
+            provider_metadata: None,
+        }
+    }
+}
+
 #[derive(Clone, PartialEq, Serialize, Deserialize)]
 #[serde(deny_unknown_fields)]
 pub struct SessionState {
diff --git a/backend/crates/gateway-plugin/sdk/src/call/model/mod.rs b/backend/crates/gateway-plugin/sdk/src/call/model/mod.rs
index 0013e6e97..d258b612f 100644
--- a/backend/crates/gateway-plugin/sdk/src/call/model/mod.rs
+++ b/backend/crates/gateway-plugin/sdk/src/call/model/mod.rs
@@ -158,6 +158,12 @@ pub enum WirePayload {
         #[serde(default, skip_serializing_if = "Option::is_none")]
         raw_sse: Option<Vec<u8>>,
     },
+    /// 完整 WebSocket 文本消息及可用的解析视图，原文独立放入二进制段
+    WebSocket {
+        event: Option<String>,
+        data: Option<serde_json::Value>,
+        message: Vec<u8>,
+    },
     /// 注释、非 JSON data 等原生 SSE 表达，不能夹带第二个帧
     RawSse { frame: Vec<u8> },
     /// 一个完整的非流式 JSON 响应正文，保留空白与字段顺序
diff --git a/backend/crates/gateway-plugin/sdk/src/call/upstream_adapter/codec.rs b/backend/crates/gateway-plugin/sdk/src/call/upstream_adapter/codec.rs
index 973e5425b..202a53dd1 100644
--- a/backend/crates/gateway-plugin/sdk/src/call/upstream_adapter/codec.rs
+++ b/backend/crates/gateway-plugin/sdk/src/call/upstream_adapter/codec.rs
@@ -10,14 +10,14 @@ use crate::call::model::{
 
 #[derive(Serialize, Deserialize)]
 struct Metadata {
-    service_tier: Option<String>,
+    observation: Option<super::super::model::facts::ResponseObservation>,
     continuation: Option<UpstreamContinuation>,
     failure: Option<UpstreamFailure>,
 }
 
 pub(super) fn encode(value: UpstreamAdapterEvent) -> Result<Vec<u8>, ExecutionEncodingError> {
     let metadata = Metadata {
-        service_tier: value.service_tier,
+        observation: value.observation,
         continuation: value.continuation,
         failure: value.failure,
     };
@@ -28,7 +28,7 @@ pub(super) fn decode(bytes: &[u8]) -> Result<UpstreamAdapterEvent, ExecutionEnco
     let (metadata, [event]): (Metadata, _) = unpack(*b"GPA1", bytes)?;
     Ok(UpstreamAdapterEvent {
         event: ExecutionEvent::decode(event)?,
-        service_tier: metadata.service_tier,
+        observation: metadata.observation,
         continuation: metadata.continuation,
         failure: metadata.failure,
     })
diff --git a/backend/crates/gateway-plugin/sdk/src/call/upstream_adapter/mod.rs b/backend/crates/gateway-plugin/sdk/src/call/upstream_adapter/mod.rs
index 62cde6323..1d13612a3 100644
--- a/backend/crates/gateway-plugin/sdk/src/call/upstream_adapter/mod.rs
+++ b/backend/crates/gateway-plugin/sdk/src/call/upstream_adapter/mod.rs
@@ -95,6 +95,8 @@ pub struct UpstreamAdapterRequest {
     pub client_transport: String,
     /// 宿主冻结的 Fast 策略：default、enabled 或 disabled
     pub fast_mode: String,
+    /// 宿主发送调用前相对 Core 请求计时起点已经过的毫秒数
+    pub request_elapsed_ms: u64,
     pub headers: Vec<(String, Vec<u8>)>,
     /// 只能来自同账号、同插件代次的宿主续接记录，不从客户端正文信任此值
     pub continuation: Option<UpstreamContinuation>,
@@ -129,8 +131,8 @@ pub struct UpstreamContinuation {
 /// 一个完整 GPE2 事件与可选的结算／续接事实；见 [`Self::encode`]
 pub struct UpstreamAdapterEvent {
     pub event: ExecutionEvent,
-    /// 上游回显的最新服务档位，可在终态确定；是否参与计价由内置 Provider 决定
-    pub service_tier: Option<String>,
+    /// 本次执行的上游观测；不使用 ExecutionEvent.host 回写宿主账本
+    pub observation: Option<super::model::facts::ResponseObservation>,
     pub continuation: Option<UpstreamContinuation>,
     pub failure: Option<UpstreamFailure>,
 }
@@ -140,7 +142,7 @@ impl UpstreamAdapterEvent {
     pub const fn new(event: ExecutionEvent) -> Self {
         Self {
             event,
-            service_tier: None,
+            observation: None,
             continuation: None,
             failure: None,
         }
@@ -186,7 +188,7 @@ pub struct UpstreamFailure {
 pub struct UpstreamHttpRequest {
     pub method: String,
     pub path: String,
-    pub headers: Vec<(String, String)>,
+    pub headers: Vec<(String, Vec<u8>)>,
     #[serde(default)]
     pub query: Vec<(String, String)>,
 }
@@ -195,7 +197,7 @@ pub struct UpstreamHttpRequest {
 #[serde(deny_unknown_fields)]
 pub struct UpstreamWebSocketRequest {
     pub path: String,
-    pub headers: Vec<(String, String)>,
+    pub headers: Vec<(String, Vec<u8>)>,
     #[serde(default)]
     pub query: Vec<(String, String)>,
 }
diff --git a/backend/crates/gateway-plugin/sdk/src/capability.rs b/backend/crates/gateway-plugin/sdk/src/capability.rs
index aa36593d5..601779c6f 100644
--- a/backend/crates/gateway-plugin/sdk/src/capability.rs
+++ b/backend/crates/gateway-plugin/sdk/src/capability.rs
@@ -32,7 +32,8 @@ impl Capability {
             | Self::RetryPolicy
             | Self::Observer
             | Self::Management => &[2],
-            Self::Scheduler | Self::UpstreamAdapter => &[3],
+            Self::Scheduler => &[3],
+            Self::UpstreamAdapter => &[4],
             Self::ModelCatalog | Self::CommandLine | Self::Maintenance => &[1],
         }
     }
diff --git a/backend/crates/gateway-plugin/sdk/src/client/http.rs b/backend/crates/gateway-plugin/sdk/src/client/http.rs
index 10d78dfb1..972ac81ed 100644
--- a/backend/crates/gateway-plugin/sdk/src/client/http.rs
+++ b/backend/crates/gateway-plugin/sdk/src/client/http.rs
@@ -11,7 +11,8 @@ use crate::{
 /// 响应头与惰性正文；不包含宿主的内部流句柄
 pub struct HostHttpResponse {
     pub status: u16,
-    pub headers: Vec<(String, String)>,
+    pub http_version: String,
+    pub headers: Vec<(String, Vec<u8>)>,
     pub body: HostHttpBody,
 }
 
@@ -20,6 +21,7 @@ pub struct HostHttpBody {
     host: HostClient,
     stream: Option<String>,
     upstream: bool,
+    trailers: Vec<(String, Vec<u8>)>,
     read: PendingRead<Result<HostReply, SessionError>>,
 }
 
@@ -64,11 +66,13 @@ impl HostClient {
         }
         Ok(HostHttpResponse {
             status: response.status,
+            http_version: response.http_version,
             headers: response.headers,
             body: HostHttpBody {
                 host: self.clone(),
                 stream: response.stream,
                 upstream,
+                trailers: response.trailers,
                 read: PendingRead::default(),
             },
         })
@@ -76,6 +80,12 @@ impl HostClient {
 }
 
 impl HostHttpBody {
+    /// 已收到的响应尾部头；读取至 EOF 后完整，名称、多值和字节原样保留
+    #[must_use]
+    pub fn trailers(&self) -> &[(String, Vec<u8>)] {
+        &self.trailers
+    }
+
     /// 按需读取至多 64 KiB，EOF 后再次读取不再调用宿主
     /// 取消一次等待后，再次调用会继续同一次读取，不丢弃分块
     ///
@@ -89,6 +99,8 @@ impl HostHttpBody {
         #[serde(deny_unknown_fields)]
         struct Read {
             eof: bool,
+            #[serde(default)]
+            trailers: Vec<(String, Vec<u8>)>,
         }
         let maximum_bytes = u32::try_from(self.host.maximum_stream_chunk_bytes().min(64 * 1024))
             .map_err(|_| invalid())?;
@@ -112,6 +124,7 @@ impl HostHttpBody {
             let _ = self.close().await;
             return Err(invalid());
         }
+        self.trailers.extend(result.trailers);
         if result.eof {
             self.stream = None;
             Ok(None)
diff --git a/backend/crates/gateway-plugin/sdk/src/client/upstream_adapter.rs b/backend/crates/gateway-plugin/sdk/src/client/upstream_adapter.rs
index 7290e0759..6ac15c48a 100644
--- a/backend/crates/gateway-plugin/sdk/src/client/upstream_adapter.rs
+++ b/backend/crates/gateway-plugin/sdk/src/client/upstream_adapter.rs
@@ -21,13 +21,16 @@ use crate::{
 /// 握手拒绝仍保留 HTTP 状态及正文，供适配器解析上游错误
 pub enum UpstreamWebSocketUpgrade {
     Connected {
-        headers: Vec<(String, String)>,
+        http_version: String,
+        headers: Vec<(String, Vec<u8>)>,
         connection: UpstreamWebSocket,
     },
     Rejected {
         status: u16,
-        headers: Vec<(String, String)>,
+        http_version: String,
+        headers: Vec<(String, Vec<u8>)>,
         body: Vec<u8>,
+        trailers: Vec<(String, Vec<u8>)>,
     },
 }
 
@@ -57,6 +60,7 @@ impl HostClient {
         }
         Ok(if response.status == 101 {
             UpstreamWebSocketUpgrade::Connected {
+                http_version: response.http_version,
                 headers: response.headers,
                 connection: UpstreamWebSocket {
                     host: self.clone(),
@@ -67,8 +71,10 @@ impl HostClient {
         } else {
             UpstreamWebSocketUpgrade::Rejected {
                 status: response.status,
+                http_version: response.http_version,
                 headers: response.headers,
                 body,
+                trailers: response.trailers,
             }
         })
     }
diff --git a/backend/crates/gateway-plugin/sdk/src/message.rs b/backend/crates/gateway-plugin/sdk/src/message.rs
index 55b390c5e..223847c54 100644
--- a/backend/crates/gateway-plugin/sdk/src/message.rs
+++ b/backend/crates/gateway-plugin/sdk/src/message.rs
@@ -4,7 +4,7 @@ use serde::{Deserialize, Serialize};
 
 use crate::{CallContext, Handshake, PluginFault};
 
-pub const PROTOCOL_VERSION: u32 = 4;
+pub const PROTOCOL_VERSION: u32 = 5;
 
 /// 元数据与二进制载荷分开，流分块不经过 JSON/base64
 #[derive(Clone, PartialEq, Serialize, Deserialize)]
diff --git a/backend/crates/gateway-plugin/sdk/tests/call/upstream_adapter/codec.rs b/backend/crates/gateway-plugin/sdk/tests/call/upstream_adapter/codec.rs
index 8e012a396..8d0566def 100644
--- a/backend/crates/gateway-plugin/sdk/tests/call/upstream_adapter/codec.rs
+++ b/backend/crates/gateway-plugin/sdk/tests/call/upstream_adapter/codec.rs
@@ -11,7 +11,8 @@ fn request() -> UpstreamAdapterRequest {
         "adapter_id":"example", "provider":"openai", "upstream_model":"fixture",
         "client_key_id":"key", "account_id":"account", "credential_revision":1,
         "protocol":"openai", "client_transport":"http_json", "fast_mode":"default",
-        "headers":[["x-fixture",[97,98]]], "continuation":null
+        "headers":[["x-fixture",[97,98]]], "continuation":null,
+        "request_elapsed_ms":0
     }))
     .unwrap()
 }
diff --git a/backend/crates/gateway-plugin/sdk/tests/client/http.rs b/backend/crates/gateway-plugin/sdk/tests/client/http.rs
index 5c8e2fee8..7afd32dfc 100644
--- a/backend/crates/gateway-plugin/sdk/tests/client/http.rs
+++ b/backend/crates/gateway-plugin/sdk/tests/client/http.rs
@@ -108,10 +108,23 @@ impl PluginHandler for NetworkHandler {
             if call.params["collect"] == true {
                 return Ok(CallReply::unary(json!({}), response.body.collect(3).await?));
             }
+            if call.params["metadata"] == true {
+                assert_eq!(response.http_version, "HTTP/1.1");
+                assert_eq!(
+                    response.headers,
+                    vec![
+                        ("x-bytes".to_owned(), vec![0xff]),
+                        ("x-bytes".to_owned(), b"other".to_vec())
+                    ]
+                );
+            }
             let mut body = response.body;
             let bytes = body.read().await?.unwrap();
             assert!(body.read().await?.is_none());
             assert!(body.read().await?.is_none());
+            if call.params["metadata"] == true {
+                assert_eq!(body.trailers(), &[("x-tail".to_owned(), vec![0xfe])]);
+            }
             body.close().await?;
             body.close().await?;
             Ok(CallReply::unary(json!({}), bytes))
@@ -146,7 +159,14 @@ async fn reply_callback(
 async fn http_response_objects_share_pull_semantics_and_retire_eof_handles() {
     for method in ["http", "upstream"] {
         let (mut host, task) = start_session(NetworkHandler).await;
-        send_call(&mut host, 1, method, json!({}), b"request".to_vec()).await;
+        send_call(
+            &mut host,
+            1,
+            method,
+            json!({"metadata":true}),
+            b"request".to_vec(),
+        )
+        .await;
         let prefix = if method == "http" {
             "host.http"
         } else {
@@ -155,7 +175,7 @@ async fn http_response_objects_share_pull_semantics_and_retire_eof_handles() {
         let frame = reply_callback(
             &mut host,
             &format!("{prefix}.do_stream"),
-            json!({"status":200,"headers":[],"stream":"private-handle"}),
+            json!({"status":200,"http_version":"HTTP/1.1","trailers":[],"headers":[["x-bytes",[255]],["x-bytes",b"other".to_vec()]],"stream":"private-handle"}),
             vec![],
         )
         .await;
@@ -173,7 +193,7 @@ async fn http_response_objects_share_pull_semantics_and_retire_eof_handles() {
         reply_callback(
             &mut host,
             &format!("{prefix}.stream_read"),
-            json!({"eof":true}),
+            json!({"eof":true,"trailers":[["x-tail",[254]]]}),
             vec![],
         )
         .await;
@@ -191,7 +211,7 @@ async fn bounded_collection_closes_the_host_stream_before_reporting_overflow() {
     reply_callback(
         &mut host,
         "host.upstream.http.do_stream",
-        json!({"status":200,"headers":[],"stream":"private-handle"}),
+        json!({"status":200,"http_version":"HTTP/1.1","trailers":[],"headers":[],"stream":"private-handle"}),
         vec![],
     )
     .await;
@@ -222,7 +242,7 @@ async fn websocket_session_handles_messages_eof_and_http_rejection() {
     reply_callback(
         &mut host,
         "host.upstream.websocket.open",
-        json!({"status":101,"headers":[],"stream":null}),
+        json!({"status":101,"http_version":"HTTP/1.1","trailers":[],"headers":[],"stream":null}),
         vec![],
     )
     .await;
@@ -249,7 +269,7 @@ async fn websocket_session_handles_messages_eof_and_http_rejection() {
     reply_callback(
         &mut host,
         "host.upstream.websocket.open",
-        json!({"status":401,"headers":[],"stream":null}),
+        json!({"status":401,"http_version":"HTTP/1.1","trailers":[],"headers":[],"stream":null}),
         b"unauthorized".to_vec(),
     )
     .await;
@@ -268,7 +288,7 @@ async fn websocket_send_and_read_callbacks_can_be_in_flight_together() {
     reply_callback(
         &mut host,
         "host.upstream.websocket.open",
-        json!({"status":101,"headers":[],"stream":null}),
+        json!({"status":101,"http_version":"HTTP/1.1","trailers":[],"headers":[],"stream":null}),
         vec![],
     )
     .await;
@@ -329,7 +349,7 @@ async fn websocket_close_can_be_retried_after_cancelling_its_future() {
     reply_callback(
         &mut host,
         "host.upstream.websocket.open",
-        json!({"status":101,"headers":[],"stream":null}),
+        json!({"status":101,"http_version":"HTTP/1.1","trailers":[],"headers":[],"stream":null}),
         vec![],
     )
     .await;
@@ -362,7 +382,7 @@ async fn websocket_close_does_not_wait_for_a_pending_reader() {
     reply_callback(
         &mut host,
         "host.upstream.websocket.open",
-        json!({"status":101,"headers":[],"stream":null}),
+        json!({"status":101,"http_version":"HTTP/1.1","trailers":[],"headers":[],"stream":null}),
         vec![],
     )
     .await;
diff --git a/backend/crates/gateway-plugin/sdk/tests/client/read.rs b/backend/crates/gateway-plugin/sdk/tests/client/read.rs
index 6ea609be8..bcbf3136e 100644
--- a/backend/crates/gateway-plugin/sdk/tests/client/read.rs
+++ b/backend/crates/gateway-plugin/sdk/tests/client/read.rs
@@ -105,7 +105,7 @@ async fn verify_cancelled_read(method: &str) {
     reply(
         &mut host,
         id,
-        json!({"status":status,"headers":[],"stream":stream}),
+        json!({"status":status,"http_version":"HTTP/1.1","trailers":[],"headers":[],"stream":stream}),
         b"",
     )
     .await;
diff --git a/backend/crates/gateway-plugin/sdk/tests/manifest.rs b/backend/crates/gateway-plugin/sdk/tests/manifest.rs
index 45bca3dcc..42d460311 100644
--- a/backend/crates/gateway-plugin/sdk/tests/manifest.rs
+++ b/backend/crates/gateway-plugin/sdk/tests/manifest.rs
@@ -135,7 +135,7 @@ fn capability_contract_versions_are_checked_independently_of_rpc_version() {
     for (capability, stage, supported, needs_formats) in [
         (Capability::Middleware, Stage::Request, 5, true),
         (Capability::Management, Stage::Management, 2, false),
-        (Capability::UpstreamAdapter, Stage::Upstream, 3, true),
+        (Capability::UpstreamAdapter, Stage::Upstream, 4, true),
         (Capability::Scheduler, Stage::Scheduling, 3, false),
     ] {
         for mut manifest in [source_manifest(), packaged_manifest()] {
@@ -396,7 +396,7 @@ fn contributes_rejects_duplicate_capability_keys_in_direct_json_parsing() {
 
     let handshake = format!(
         r#"{{
-            "protocol_version":4,"artifact_sha256":"digest","plugin_id":"test.example",
+            "protocol_version":5,"artifact_sha256":"digest","plugin_id":"test.example",
             "instance_id":"instance","generation":1,"incarnation":"incarnation",
             "contributes":{{"middleware":{declaration},"middleware":{declaration}}}
         }}"#
diff --git a/modules/plugins b/modules/plugins
index 750073216..28199b811 160000
--- a/modules/plugins
+++ b/modules/plugins
@@ -1 +1 @@
-Subproject commit 7500732163f6a6ca0702235390d130963cd001ca
+Subproject commit 28199b8114d23875ad92278d3857a2903badcd48