diff --git a/backend/crates/providers/openai/src/transport/downstream/body.rs b/backend/crates/providers/openai/src/transport/downstream/body.rs
index 135a6aa84..14c82d174 100644
--- a/backend/crates/providers/openai/src/transport/downstream/body.rs
+++ b/backend/crates/providers/openai/src/transport/downstream/body.rs
@@ -2,7 +2,7 @@
 
 use serde_json::{Map, Value};
 
-/// 补齐 Codex 请求缺省字段并移除已确认不适用的顶层参数，不递归清洗业务正文。
+/// 补齐 Codex 请求缺省字段并适配已确认不兼容的请求形状，不递归清洗业务正文。
 ///
 /// 兼容基准是 Codex Core/Desktop 的模型请求，不是公开 OpenAI Responses API。
 /// 未知字段继续透传，不能因官方请求结构中没有某个字段就将其列入过滤规则。
@@ -10,6 +10,21 @@ pub(in crate::transport) fn normalize_codex_request_body(body: &mut Map<String,
     // 官方 Core/Desktop 显式发送 store=false；仅为缺字段的下游请求补齐，保留显式值。
     body.entry("store").or_insert(Value::Bool(false));
 
+    // Codex 上游拒绝显式 message 的 system role；沿用官方客户端的 developer
+    // role 承载指令，只转换已确认的消息形状，保留内容与其他字段。
+    if let Some(input) = body.get_mut("input").and_then(Value::as_array_mut) {
+        for item in input {
+            let Some(item) = item.as_object_mut() else {
+                continue;
+            };
+            if item.get("type").and_then(Value::as_str) == Some("message")
+                && item.get("role").and_then(Value::as_str) == Some("system")
+            {
+                item.insert("role".to_owned(), Value::String("developer".to_owned()));
+            }
+        }
+    }
+
     for field in [
         // Pi 普通 Responses 适配将 maxTokens 映射为 max_output_tokens，
         // temperature 则原样写入；Pi 的 Codex 适配也可能发送 temperature。
diff --git a/backend/crates/providers/openai/tests/transport/downstream/body.rs b/backend/crates/providers/openai/tests/transport/downstream/body.rs
index 656a708c8..6ed840867 100644
--- a/backend/crates/providers/openai/tests/transport/downstream/body.rs
+++ b/backend/crates/providers/openai/tests/transport/downstream/body.rs
@@ -87,8 +87,72 @@ fn encoder_should_preserve_explicit_store_values() {
     }
 }
 
+#[test]
+fn encoder_should_convert_system_messages_to_developer_without_changing_content() {
+    let body = json!({
+        "model": "client-model",
+        "instructions": "Keep the existing instructions.",
+        "input": [
+            {
+                "type": "message",
+                "role": "system",
+                "content": "You are a pirate.",
+                "future_item_field": {"keep": true}
+            },
+            {
+                "type": "message",
+                "role": "system",
+                "content": [{"type": "input_text", "text": "Be concise.", "role": "system"}]
+            },
+            {
+                "type": "message",
+                "role": "user",
+                "content": [{"type": "input_text", "text": "Say hello."}]
+            },
+            {"type": "message", "role": "developer", "content": "Existing guidance."},
+            {"role": "system", "content": "Shorthand message."},
+            {"type": "computer_screenshot", "role": "system"},
+            "opaque-item"
+        ],
+        "future_top_level_field": {"keep": true}
+    });
+    let encoded = encode_downstream_request(body);
+
+    assert_eq!(
+        Value::Object(encoded.body().clone()),
+        json!({
+            "model": "gpt-test",
+            "instructions": "Keep the existing instructions.",
+            "input": [
+                {
+                    "type": "message",
+                    "role": "developer",
+                    "content": "You are a pirate.",
+                    "future_item_field": {"keep": true}
+                },
+                {
+                    "type": "message",
+                    "role": "developer",
+                    "content": [{"type": "input_text", "text": "Be concise.", "role": "system"}]
+                },
+                {
+                    "type": "message",
+                    "role": "user",
+                    "content": [{"type": "input_text", "text": "Say hello."}]
+                },
+                {"type": "message", "role": "developer", "content": "Existing guidance."},
+                {"role": "system", "content": "Shorthand message."},
+                {"type": "computer_screenshot", "role": "system"},
+                "opaque-item"
+            ],
+            "store": false,
+            "future_top_level_field": {"keep": true}
+        })
+    );
+}
+
 #[tokio::test]
-async fn backend_http_should_send_default_store_for_downstream_request() {
+async fn backend_http_should_send_default_store_and_normalized_system_role() {
     let listener = TcpListener::bind("127.0.0.1:0")
         .await
         .expect("bind HTTP server");
@@ -105,7 +169,10 @@ async fn backend_http_should_send_default_store_for_downstream_request() {
         let body = zstd::stream::decode_all(&raw[head_end..]).expect("decode request body");
         serde_json::from_slice::<Value>(&body).expect("HTTP request JSON")
     });
-    let mut request = encode_downstream_request(json!({"model": "gpt-test", "input": "hello"}));
+    let mut request = encode_downstream_request(json!({
+        "model": "gpt-test",
+        "input": [{"type": "message", "role": "system", "content": "Be concise."}]
+    }));
     request.force_http_sse = true;
     let client = CodexBackendClient::new(
         reqwest::Client::builder()
@@ -127,13 +194,16 @@ async fn backend_http_should_send_default_store_for_downstream_request() {
     assert_eq!(
         server.await.expect("HTTP server task"),
         json!({
-            "model": "gpt-test", "input": "hello", "store": false, "stream": true
+            "model": "gpt-test",
+            "input": [{"type": "message", "role": "developer", "content": "Be concise."}],
+            "store": false,
+            "stream": true
         })
     );
 }
 
 #[tokio::test]
-async fn backend_websocket_should_send_default_store_for_downstream_request() {
+async fn backend_websocket_should_send_default_store_and_normalized_system_role() {
     let listener = TcpListener::bind("127.0.0.1:0")
         .await
         .expect("bind WebSocket server");
@@ -156,7 +226,10 @@ async fn backend_websocket_should_send_default_store_for_downstream_request() {
             .expect("send terminal event");
         body
     });
-    let mut request = encode_downstream_request(json!({"model": "gpt-test", "input": "hello"}));
+    let mut request = encode_downstream_request(json!({
+        "model": "gpt-test",
+        "input": [{"type": "message", "role": "system", "content": "Be concise."}]
+    }));
     request.use_websocket = true;
     let client = CodexBackendClient::new(
         reqwest::Client::builder()
@@ -178,6 +251,10 @@ async fn backend_websocket_should_send_default_store_for_downstream_request() {
 
     let body = server.await.expect("WebSocket server task");
     assert_eq!(body.get("store"), Some(&json!(false)));
+    assert_eq!(
+        body["input"],
+        json!([{"type": "message", "role": "developer", "content": "Be concise."}])
+    );
 }
 
 fn encode_downstream_request(body: Value) -> CodexResponsesRequest {
diff --git a/docs/api.md b/docs/api.md
index 08faf50ec..4f56f72ea 100644
--- a/docs/api.md
+++ b/docs/api.md
@@ -190,7 +190,9 @@ Responses 也不透传 `x-stainless-*`、`Origin`、`Referer`、`sec-ch-ua*` 和
 
 Responses 上游编码会移除 Codex 不接受的顶层 `temperature`、`max_output_tokens` 和
 `prompt_cache_retention`。缺少顶层 `store` 时补齐 `false`，与官方 Codex 客户端一致；
-显式提供的值保持原样。HTTP/SSE 与 WebSocket 共用这条正文兼容规则。
+显式提供的值保持原样。`input` 数组中显式指定 `type: "message"` 且 `role: "system"` 的消息，
+其角色转换为 Codex 接受的 `developer`；消息内容、顺序、其他字段和顶层 `instructions` 保持不变。
+HTTP/SSE 与 WebSocket 共用这条正文兼容规则。
 `prompt_cache_key`、`reasoning`、`include` 等 Codex 参数继续保留。过滤只作用于顶层，
 不删除工具参数 schema、输入内容或 `client_metadata` 内的同名业务字段；其他未知字段继续透传。