diff --git a/backend/Cargo.lock b/backend/Cargo.lock
index 729b48f15..b0a057344 100644
--- a/backend/Cargo.lock
+++ b/backend/Cargo.lock
@@ -3238,6 +3238,7 @@ dependencies = [
 name = "provider-openai"
 version = "0.0.0"
 dependencies = [
+ "allocation-counter",
  "async-stream",
  "async-trait",
  "aws-lc-rs",
diff --git a/backend/crates/gateway-api/src/openai/responses/http.rs b/backend/crates/gateway-api/src/openai/responses/http.rs
index 72470bc37..afed0b4f6 100644
--- a/backend/crates/gateway-api/src/openai/responses/http.rs
+++ b/backend/crates/gateway-api/src/openai/responses/http.rs
@@ -134,6 +134,8 @@ pub(crate) async fn execute_prepared_responses(
         ClientTransport::HttpJson
     };
     let model_hint = Some(decoded.metadata().requested_model().to_owned());
+    // 终端会重新校验中间件改写后的正文，初步解码不必保留到响应交付
+    drop(decoded);
     let execution = service.execution();
     let request_id = prepared.request_id().clone();
     let cancellation = prepared.cancellation();
diff --git a/backend/crates/gateway-api/src/openai/responses/websocket/mod.rs b/backend/crates/gateway-api/src/openai/responses/websocket/mod.rs
index 20696965f..342479b25 100644
--- a/backend/crates/gateway-api/src/openai/responses/websocket/mod.rs
+++ b/backend/crates/gateway-api/src/openai/responses/websocket/mod.rs
@@ -214,8 +214,9 @@ async fn serve_responses_websocket(socket: WebSocket, session: ResponsesWebSocke
         }
         request_count = request_count.saturating_add(1);
         let correlation_id = Arc::<str>::from(service.next_request_id());
-        let decoded = match decode_response_create_with_context(&payload, &request_headers) {
-            Ok(decoded) => decoded,
+        // 初步解码只保留路由事实，避免整份正文跨越准入等待和响应交付
+        let model_hint = match decode_response_create_with_context(&payload, &request_headers) {
+            Ok(decoded) => Some(decoded.metadata().requested_model().to_owned()),
             Err(error) => {
                 trace_rejected_request(
                     &correlation_id,
@@ -298,9 +299,9 @@ async fn serve_responses_websocket(socket: WebSocket, session: ResponsesWebSocke
             protocol: "openai".to_owned(),
             operation: Some(OperationKind::Generate),
             transport: ClientTransport::WebSocket,
-            model_hint: Some(decoded.metadata().requested_model().to_owned()),
+            model_hint,
             headers: encode_headers(&raw_headers),
-            body: Bytes::from(payload.clone()),
+            body: Bytes::from(payload),
         };
         let service_for_terminal = service.clone();
         let replay_for_terminal = replay.clone();
diff --git a/backend/crates/providers/openai/Cargo.toml b/backend/crates/providers/openai/Cargo.toml
index 86efa7666..5ae2d6ac6 100644
--- a/backend/crates/providers/openai/Cargo.toml
+++ b/backend/crates/providers/openai/Cargo.toml
@@ -72,6 +72,7 @@ openssl = "=0.10.75"
 openssl-sys = { version = "=0.9.111", features = ["vendored"] }
 
 [dev-dependencies]
+allocation-counter = "0.8.1"
 tempfile = "3.27.0"
 tokio = { version = "1.52.3", default-features = false, features = ["rt-multi-thread", "test-util"] }
 tracing-subscriber = { version = "0.3.23", features = ["json"] }
diff --git a/backend/crates/providers/openai/src/transport/profile/platform_release/xz.rs b/backend/crates/providers/openai/src/transport/profile/platform_release/xz.rs
index 0edb2ea8b..4651fc6d9 100644
--- a/backend/crates/providers/openai/src/transport/profile/platform_release/xz.rs
+++ b/backend/crates/providers/openai/src/transport/profile/platform_release/xz.rs
@@ -140,6 +140,8 @@ impl<R: Read + Seek> Read for IndexedXz<R> {
                 .map_err(|_| invalid())?;
             let decoder = xz2::read::XzDecoder::new_stream(bytes.as_slice(), stream);
             self.buffer.clear();
+            // 索引长度已受 MAX_BLOCK 限制，多留一字节检查越界，避免 read_to_end 倍增容量
+            self.buffer.reserve_exact(block.unpacked as usize + 1);
             decoder
                 .take(block.unpacked + 1)
                 .read_to_end(&mut self.buffer)?;
diff --git a/backend/crates/providers/openai/tests/transport/profile/platform_release/xz.rs b/backend/crates/providers/openai/tests/transport/profile/platform_release/xz.rs
index 7280ece44..55e48e871 100644
--- a/backend/crates/providers/openai/tests/transport/profile/platform_release/xz.rs
+++ b/backend/crates/providers/openai/tests/transport/profile/platform_release/xz.rs
@@ -1,4 +1,4 @@
-//! 验证 XZ 索引读取的定位、跨块访问与无关块跳过
+//! 验证 XZ 索引读取的定位、跨块访问、无关块跳过与大块缓冲保留量
 
 use provider_openai::transport::profile::platform_release::xz::IndexedXz;
 use std::io::{Cursor, Read, Seek, SeekFrom, Write};
@@ -8,6 +8,30 @@ pub(super) fn compress(bytes: &[u8]) -> Vec<u8> {
     encoder.write_all(bytes).unwrap();
     encoder.finish().unwrap()
 }
+
+#[test]
+fn indexed_xz_keeps_large_block_buffer_close_to_unpacked_size() {
+    // 同时覆盖实际制品的 24 MiB 块和会因 EOF 探测再次扩容的二次幂边界
+    for size in [24 * 1024 * 1024, 32 * 1024 * 1024] {
+        let compressed = compress(&vec![0x5a; size]);
+        let mut reader =
+            IndexedXz::new(Cursor::new(&compressed), 0, compressed.len() as u64).unwrap();
+        let mut byte = [0];
+        let allocations = allocation_counter::measure(|| reader.read_exact(&mut byte).unwrap());
+
+        assert_eq!(byte, [0x5a]);
+        // 测量读取器仍存活时的 Rust 分配；为块元数据留余量，不等同于进程 RSS
+        assert!(
+            allocations.bytes_current <= size as i64 + 64 * 1024,
+            "size={size}, {allocations:?}"
+        );
+        reader.seek(SeekFrom::End(-1)).unwrap();
+        reader.read_exact(&mut byte).unwrap();
+        assert_eq!(byte, [0x5a]);
+        assert_eq!(reader.read(&mut byte).unwrap(), 0);
+    }
+}
+
 #[test]
 fn indexed_xz_supports_backward_and_end_relative_seeks() {
     let bytes: Vec<u8> = (0..100_000).map(|i| (i % 251) as u8).collect();