diff --git a/backend/crates/providers/openai/src/transport/headers.rs b/backend/crates/providers/openai/src/transport/headers.rs
index cdf78df30..e9d338342 100644
--- a/backend/crates/providers/openai/src/transport/headers.rs
+++ b/backend/crates/providers/openai/src/transport/headers.rs
@@ -254,5 +254,32 @@ fn header_pairs(headers: &HeaderMap) -> Vec<(String, String)> {
 }
 
 pub(super) fn websocket_header_pairs(headers: &HeaderMap) -> Vec<(String, String)> {
-    header_pairs(headers)
+    let mut pairs = header_pairs(headers);
+    // WebSocket opening 的 HeaderMap 先由业务头构造，再被 tungstenite 插入协议头；
+    // 这里复现官方 HeaderMap 交给 tungstenite 时的迭代顺序，最终序列化后的
+    // 线级顺序由 fingerprint 测试锁定。未知扩展头仍保持相对顺序。
+    pairs.sort_by_key(|(name, _)| websocket_header_order(name));
+    pairs
+}
+
+fn websocket_header_order(name: &str) -> usize {
+    const OFFICIAL_INSERTION_ORDER: &[&str] = &[
+        "version",
+        "x-codex-beta-features",
+        "x-client-request-id",
+        "session-id",
+        "thread-id",
+        "x-codex-window-id",
+        "x-codex-turn-metadata",
+        "x-codex-routing-hint",
+        "openai-beta",
+        "originator",
+        "user-agent",
+        "authorization",
+        "chatgpt-account-id",
+    ];
+    OFFICIAL_INSERTION_ORDER
+        .iter()
+        .position(|candidate| name.eq_ignore_ascii_case(candidate))
+        .unwrap_or(OFFICIAL_INSERTION_ORDER.len())
 }
diff --git a/backend/crates/providers/openai/tests/transport/websocket/fingerprint.rs b/backend/crates/providers/openai/tests/transport/websocket/fingerprint.rs
new file mode 100644
index 000000000..01f38a17f
--- /dev/null
+++ b/backend/crates/providers/openai/tests/transport/websocket/fingerprint.rs
@@ -0,0 +1,276 @@
+use base64::{Engine as _, engine::general_purpose::STANDARD};
+use flate2::{Decompress, FlushDecompress};
+use tokio_tungstenite::tungstenite::handshake::derive_accept_key;
+
+use super::*;
+
+#[derive(Debug)]
+struct ClientFrame {
+    first_byte: u8,
+    masked: bool,
+    mask_len: usize,
+    payload: Vec<u8>,
+}
+
+#[tokio::test]
+async fn official_websocket_wire_fingerprint_should_remain_aligned() {
+    let listener = TcpListener::bind("127.0.0.1:0")
+        .await
+        .expect("bind fingerprint server");
+    let address = listener.local_addr().expect("fingerprint server address");
+    let server = tokio::spawn(async move {
+        let (mut stream, _) = listener.accept().await.expect("accept fingerprint client");
+        let opening = read_http_request(&mut stream).await;
+        let websocket_key = read_header_value(&opening, "sec-websocket-key")
+            .expect("WebSocket key")
+            .to_owned();
+        let response = format!(
+            concat!(
+                "HTTP/1.1 101 Switching Protocols\r\n",
+                "Connection: Upgrade\r\n",
+                "Upgrade: websocket\r\n",
+                "Sec-WebSocket-Accept: {}\r\n",
+                "Sec-WebSocket-Extensions: permessage-deflate\r\n",
+                "\r\n"
+            ),
+            derive_accept_key(websocket_key.as_bytes())
+        );
+        stream
+            .write_all(response.as_bytes())
+            .await
+            .expect("write upgrade response");
+
+        let frame = read_client_frame(&mut stream).await;
+        write_server_text_frame(
+            &mut stream,
+            &completed_websocket_response("resp_wire_fingerprint", 1, 1),
+        )
+        .await;
+        (opening, websocket_key, frame)
+    });
+
+    let mut request = CodexResponsesRequest::from_body(
+        json!({
+            "model": "gpt-5.5",
+            "instructions": "be brief",
+            "input": [],
+            "tools": [],
+            "tool_choice": "auto",
+            "parallel_tool_calls": true,
+            "reasoning": {"effort": "medium", "summary": "auto"},
+            "store": false,
+            "stream": true,
+            "include": ["reasoning.encrypted_content"],
+            "prompt_cache_key": "conversation-fingerprint",
+            "generate": false,
+            "client_metadata": {"request_kind": "prewarm"}
+        })
+        .as_object()
+        .expect("request object")
+        .clone(),
+    );
+    request.use_websocket = true;
+    request.local_conversation_id = Some("conversation-fingerprint".to_owned());
+    request.turn_metadata = Some(r#"{"kind":"fingerprint"}"#.to_owned());
+    request.beta_features = Some("feature-a".to_owned());
+    request.version = Some("26.908.40834".to_owned());
+    request.codex_window_id = Some("window-fingerprint".to_owned());
+    let client = CodexBackendClient::new(
+        reqwest::Client::builder()
+            .no_proxy()
+            .build()
+            .expect("HTTP client"),
+        format!("http://{address}/backend-api"),
+        test_wire_profile(),
+    )
+    .with_websocket_pool(Arc::new(CodexWebSocketPool::new(Duration::from_mins(1))));
+
+    let result = client
+        .create_response(
+            &request,
+            CodexRequestContext {
+                trace: None,
+                turn_metadata: request.turn_metadata.as_deref(),
+                beta_features: request.beta_features.as_deref(),
+                version: request.version.as_deref(),
+                codex_window_id: request.codex_window_id.as_deref(),
+                session_id: Some("session-fingerprint"),
+                thread_id: Some("thread-fingerprint"),
+                client_request_id: Some("request-fingerprint"),
+                ..request_context("req_wire_fingerprint", Some("account-fingerprint"))
+            },
+        )
+        .await;
+
+    result.unwrap_or_else(|error| panic!("fingerprint WebSocket response: {error}"));
+    let (opening, websocket_key, frame) = server.await.expect("fingerprint server task");
+    assert_eq!(
+        opening.lines().next(),
+        Some("GET /backend-api/codex/responses HTTP/1.1")
+    );
+    // 官方 Codex Core 0.154.0（Desktop 26.908.40834）的原始 opening 抓包顺序。
+    assert_eq!(
+        read_header_names(&opening),
+        vec![
+            "host",
+            "connection",
+            "upgrade",
+            "sec-websocket-version",
+            "sec-websocket-key",
+            "chatgpt-account-id",
+            "authorization",
+            "user-agent",
+            "originator",
+            "openai-beta",
+            "version",
+            "x-codex-beta-features",
+            "x-client-request-id",
+            "session-id",
+            "thread-id",
+            "x-codex-window-id",
+            "x-codex-turn-metadata",
+            "x-codex-routing-hint",
+            "sec-websocket-extensions",
+        ]
+    );
+    assert_eq!(
+        read_header_value(&opening, "sec-websocket-extensions"),
+        Some("permessage-deflate; client_max_window_bits")
+    );
+    assert_eq!(
+        (
+            read_header_value(&opening, "chatgpt-account-id"),
+            read_header_value(&opening, "authorization"),
+            read_header_value(&opening, "originator"),
+            read_header_value(&opening, "version"),
+            read_header_value(&opening, "openai-beta"),
+            read_header_value(&opening, "x-client-request-id"),
+            read_header_value(&opening, "x-codex-routing-hint"),
+        ),
+        (
+            Some("account-fingerprint"),
+            Some("Bearer access-token"),
+            Some("codex_cli_rs"),
+            Some("1.2.3"),
+            Some("responses_websockets=2026-02-06"),
+            Some("request-fingerprint"),
+            Some("model=gpt-5.5"),
+        )
+    );
+    assert_eq!(
+        STANDARD
+            .decode(websocket_key)
+            .expect("base64 WebSocket key")
+            .len(),
+        16
+    );
+
+    assert_eq!(frame.first_byte, 0b1100_0001, "FIN + RSV1 + text opcode");
+    assert!(frame.masked, "official client frames are masked");
+    assert_eq!(frame.mask_len, 4);
+    let payload = decompress_permessage_deflate(&frame.payload);
+    let payload: Value = serde_json::from_slice(&payload).expect("response.create JSON");
+    assert_eq!(payload["type"], "response.create");
+    assert_eq!(payload["generate"], false);
+    assert_eq!(
+        payload
+            .as_object()
+            .expect("response.create object")
+            .keys()
+            .map(String::as_str)
+            .collect::<Vec<_>>(),
+        vec![
+            "type",
+            "model",
+            "instructions",
+            "input",
+            "tools",
+            "tool_choice",
+            "parallel_tool_calls",
+            "reasoning",
+            "store",
+            "stream",
+            "include",
+            "prompt_cache_key",
+            "generate",
+            "client_metadata",
+        ]
+    );
+}
+
+async fn read_client_frame(stream: &mut TcpStream) -> ClientFrame {
+    let first_byte = stream.read_u8().await.expect("frame first byte");
+    let second_byte = stream.read_u8().await.expect("frame second byte");
+    let payload_len = match second_byte & 0x7f {
+        len @ 0..=125 => u64::from(len),
+        126 => u64::from(stream.read_u16().await.expect("16-bit frame length")),
+        127 => stream.read_u64().await.expect("64-bit frame length"),
+        _ => unreachable!(),
+    };
+    let masked = second_byte & 0x80 != 0;
+    let mut mask = [0_u8; 4];
+    if masked {
+        stream.read_exact(&mut mask).await.expect("frame mask");
+    }
+    let mut payload = vec![0_u8; usize::try_from(payload_len).expect("frame length fits usize")];
+    stream
+        .read_exact(&mut payload)
+        .await
+        .expect("frame payload");
+    if masked {
+        for (index, byte) in payload.iter_mut().enumerate() {
+            *byte ^= mask[index % mask.len()];
+        }
+    }
+    ClientFrame {
+        first_byte,
+        masked,
+        mask_len: if masked { mask.len() } else { 0 },
+        payload,
+    }
+}
+
+fn decompress_permessage_deflate(payload: &[u8]) -> Vec<u8> {
+    let mut compressed = Vec::with_capacity(payload.len() + 4);
+    compressed.extend_from_slice(payload);
+    compressed.extend_from_slice(&[0x00, 0x00, 0xff, 0xff]);
+    let mut output = vec![0_u8; 64 * 1024];
+    let mut decompressor = Decompress::new(false);
+    decompressor
+        .decompress(&compressed, &mut output, FlushDecompress::Sync)
+        .expect("decompress response.create frame");
+    output.truncate(
+        usize::try_from(decompressor.total_out()).expect("decompressed length fits usize"),
+    );
+    output
+}
+
+async fn write_server_text_frame(stream: &mut TcpStream, payload: &str) {
+    let payload = payload.as_bytes();
+    let mut frame = Vec::with_capacity(payload.len() + 10);
+    frame.push(0x81);
+    match payload.len() {
+        len @ 0..=125 => frame.push(u8::try_from(len).expect("short frame length")),
+        len @ 126..=65_535 => {
+            frame.push(126);
+            frame.extend_from_slice(
+                &u16::try_from(len)
+                    .expect("16-bit frame length")
+                    .to_be_bytes(),
+            );
+        }
+        len => {
+            frame.push(127);
+            frame.extend_from_slice(
+                &u64::try_from(len)
+                    .expect("64-bit frame length")
+                    .to_be_bytes(),
+            );
+        }
+    }
+    frame.extend_from_slice(payload);
+    stream
+        .write_all(&frame)
+        .await
+        .expect("write completed frame");
+}
diff --git a/backend/crates/providers/openai/tests/transport/websocket/mod.rs b/backend/crates/providers/openai/tests/transport/websocket/mod.rs
index 9e2577527..8eb40bd10 100644
--- a/backend/crates/providers/openai/tests/transport/websocket/mod.rs
+++ b/backend/crates/providers/openai/tests/transport/websocket/mod.rs
@@ -5,6 +5,7 @@ use provider_openai::transport::{
 };
 
 mod diagnostics;
+mod fingerprint;
 mod metadata;
 
 fn rate_limit_event(used_percent: u64) -> String {