diff --git a/openless-all/app/src-tauri/src/polish.rs b/openless-all/app/src-tauri/src/polish.rs
index a6adbdf..38a8ca7 100644
--- a/openless-all/app/src-tauri/src/polish.rs
+++ b/openless-all/app/src-tauri/src/polish.rs
@@ -514,15 +514,7 @@ impl OpenAICompatibleLLMProvider {
         }
         let request = request.json(body);
 
-        let response = match request.send().await {
-            Ok(r) => r,
-            Err(e) => {
-                if e.is_timeout() {
-                    return Err(LLMError::Timeout);
-                }
-                return Err(LLMError::Network(e.to_string()));
-            }
-        };
+        let response = send_with_transient_retry(request).await?;
 
         let status = response.status();
         let body_text = response
@@ -588,15 +580,7 @@ impl OpenAICompatibleLLMProvider {
         }
         let request = request.json(&body);
 
-        let response = match request.send().await {
-            Ok(r) => r,
-            Err(e) => {
-                if e.is_timeout() {
-                    return Err(LLMError::Timeout);
-                }
-                return Err(LLMError::Network(e.to_string()));
-            }
-        };
+        let response = send_with_transient_retry(request).await?;
 
         let status = response.status();
         if !status.is_success() {
@@ -717,15 +701,7 @@ impl OpenAICompatibleLLMProvider {
         }
         let request = request.json(&body);
 
-        let response = match request.send().await {
-            Ok(r) => r,
-            Err(e) => {
-                if e.is_timeout() {
-                    return Err(LLMError::Timeout);
-                }
-                return Err(LLMError::Network(e.to_string()));
-            }
-        };
+        let response = send_with_transient_retry(request).await?;
 
         let status = response.status();
         if !status.is_success() {
@@ -1260,6 +1236,54 @@ pub(crate) fn http_client_builder(base_url: &str, timeout_secs: u64) -> reqwest:
     }
 }
 
+/// 发请求 + 网络抖动 retry：**只**对 `is_connect()` / `is_request()` 这两类「服务端
+/// 必然没收到」的失败重试一次。`is_timeout()` 故意**不**重试——超时时服务端可能已经
+/// 在处理请求并扣计费（LLM completion 是非幂等动作），重试会导致重复 billing + 重复
+/// completion。HTTP 4xx/5xx 不在这里触发——那些走 response.status() 分支单独处理。
+///
+/// 调用前提：传入的 RequestBuilder body 必须是内存型（json / form），不能是 stream
+/// reader——retry 用 `try_clone()` 复制 RequestBuilder，stream body 不支持。
+///
+/// 对流式 SSE 路径 retry 是安全的：connect / request 类失败发生在 TCP 握手 / HTTP
+/// 请求写出阶段，response 还没回 → on_delta 必然未被调用 → 不会有「已流式输出的字
+/// 被重复」的问题。
+async fn send_with_transient_retry(
+    request: reqwest::RequestBuilder,
+) -> Result<reqwest::Response, LLMError> {
+    const RETRY_DELAY_MS: u64 = 500;
+    let initial = request
+        .try_clone()
+        .expect("memory-backed body (json/form) must be clonable for retry");
+    match initial.send().await {
+        Ok(r) => Ok(r),
+        Err(e) if e.is_connect() || e.is_request() => {
+            log::warn!(
+                "[llm] send transient failure, retry in {}ms: {}",
+                RETRY_DELAY_MS,
+                e
+            );
+            tokio::time::sleep(Duration::from_millis(RETRY_DELAY_MS)).await;
+            match request.send().await {
+                Ok(r) => Ok(r),
+                Err(e2) => {
+                    if e2.is_timeout() {
+                        Err(LLMError::Timeout)
+                    } else {
+                        Err(LLMError::Network(e2.to_string()))
+                    }
+                }
+            }
+        }
+        Err(e) => {
+            if e.is_timeout() {
+                Err(LLMError::Timeout)
+            } else {
+                Err(LLMError::Network(e.to_string()))
+            }
+        }
+    }
+}
+
 fn should_bypass_proxy_for_base_url(base_url: &str) -> bool {
     let Ok(url) = reqwest::Url::parse(base_url.trim()) else {
         return false;