From d8065cc628fcd373d00c4363d718095f19e78c9e Mon Sep 17 00:00:00 2001
From: htyvista <htyvista@gmail.com>
Date: Sat, 29 Aug 2026 22:55:03 +0800
Subject: [PATCH] fix(proxy): preserve mid-conversation system messages for
 prefix cache (#6941)

* fix(proxy): preserve mid-conversation system messages for prefix cache

Remove normalize_openai_system_messages which merged all system messages
to the head, breaking radix prefix cache on every turn when Claude Code
injects <total_tokens> metadata as a mid-conversation system message.

Now system messages from the top-level  field are placed first,
and all messages (including mid-conversation system) are extended in
their original order without any merging or reordering.

Also add custom-protocol feature for standalone release builds.

* fix(proxy): restore top-level system merge, add regression test

Restore merging of top-level system array blocks into a single system
message (cross-turn stable, does not affect prefix cache). Mid-conversation
system messages from the messages array remain in place without merging
or reordering.

Add regression test to lock the new behavior: mid-conversation system
messages stay at their original position.
---
 src-tauri/src/proxy/providers/transform.rs | 93 ++++++++++------------
 1 file changed, 40 insertions(+), 53 deletions(-)

diff --git a/src-tauri/src/proxy/providers/transform.rs b/src-tauri/src/proxy/providers/transform.rs
index 6bd6c511d..7dbb9d63d 100644
--- a/src-tauri/src/proxy/providers/transform.rs
+++ b/src-tauri/src/proxy/providers/transform.rs
@@ -158,15 +158,20 @@ pub fn anthropic_to_openai_with_reasoning_content(
                 messages.push(json!({"role": "system", "content": text}));
             }
         } else if let Some(arr) = system.as_array() {
+            // 顶层 system 数组合并为一条 system 消息（跨轮字节稳定，不影响前缀缓存）
+            let mut parts = Vec::new();
             for msg in arr {
                 if let Some(text) = msg.get("text").and_then(|t| t.as_str()) {
                     let text = strip_leading_anthropic_billing_header(text);
                     if text.is_empty() {
                         continue;
                     }
-                    messages.push(json!({"role": "system", "content": text}));
+                    parts.push(text.to_string());
                 }
             }
+            if !parts.is_empty() {
+                messages.push(json!({"role": "system", "content": parts.join("\n")}));
+            }
         }
     }
 
@@ -180,7 +185,6 @@ pub fn anthropic_to_openai_with_reasoning_content(
         }
     }
 
-    normalize_openai_system_messages(&mut messages);
     result["messages"] = json!(messages);
 
     // 转换参数 — o-series 模型需要 max_completion_tokens
@@ -305,57 +309,6 @@ fn map_tool_choice_to_chat(tool_choice: &Value) -> Value {
     }
 }
 
-fn normalize_openai_system_messages(messages: &mut Vec<Value>) {
-    let system_count = messages
-        .iter()
-        .filter(|message| message.get("role").and_then(|value| value.as_str()) == Some("system"))
-        .count();
-
-    if system_count == 0 {
-        return;
-    }
-
-    if system_count == 1 {
-        if let Some(index) = messages.iter().position(|message| {
-            message.get("role").and_then(|value| value.as_str()) == Some("system")
-        }) {
-            if index > 0 {
-                let message = messages.remove(index);
-                messages.insert(0, message);
-            }
-        }
-        return;
-    }
-
-    let mut parts = Vec::new();
-    messages.retain(|message| {
-        if message.get("role").and_then(|value| value.as_str()) != Some("system") {
-            return true;
-        }
-
-        match message.get("content") {
-            Some(Value::String(text)) if !text.is_empty() => parts.push(text.clone()),
-            Some(Value::Array(content_parts)) => {
-                let text = content_parts
-                    .iter()
-                    .filter_map(|part| part.get("text").and_then(|value| value.as_str()))
-                    .collect::<Vec<_>>()
-                    .join("\n");
-                if !text.is_empty() {
-                    parts.push(text);
-                }
-            }
-            _ => {}
-        }
-
-        false
-    });
-
-    if !parts.is_empty() {
-        messages.insert(0, json!({"role": "system", "content": parts.join("\n")}));
-    }
-}
-
 /// 转换单条消息到 OpenAI 格式（可能产生多条消息）
 fn convert_message_to_openai(
     role: &str,
@@ -988,6 +941,40 @@ mod tests {
         assert!(result["messages"][0].get("cache_control").is_none());
     }
 
+    #[test]
+    fn test_anthropic_to_openai_preserves_mid_conversation_system_in_place() {
+        // Claude Code 会在对话中间注入 system 消息（如 <total_tokens>），
+        // 必须保持原位，不合并不上提，否则破坏前缀缓存。
+        let input = json!({
+            "model": "claude-3-sonnet",
+            "max_tokens": 1024,
+            "system": "You are Claude Code.",
+            "messages": [
+                {"role": "user", "content": "Hello"},
+                {"role": "assistant", "content": "Hi there!"},
+                {"role": "system", "content": "<total_tokens>14963538 tokens left</total_tokens>"},
+                {"role": "user", "content": "Continue"}
+            ]
+        });
+
+        let result = anthropic_to_openai(input).unwrap();
+        let messages = result["messages"].as_array().unwrap();
+
+        // 顶层 system 在最前面
+        assert_eq!(messages[0]["role"], "system");
+        assert_eq!(messages[0]["content"], "You are Claude Code.");
+
+        // 中途 system 保持原位（第 3 条，index=3），不被合并或上提
+        assert_eq!(messages[3]["role"], "system");
+        assert_eq!(
+            messages[3]["content"],
+            "<total_tokens>14963538 tokens left</total_tokens>"
+        );
+
+        // 总共 5 条消息，没有合并
+        assert_eq!(messages.len(), 5);
+    }
+
     #[test]
     fn test_anthropic_to_openai_strips_cache_control_from_conflicting_system() {
         let input = json!({