Preserve cache hints when collapsing system prompts

Strict OpenAI-compatible chat backends still need fragmented Claude\nsystem prompts collapsed into one leading system message, but that\nnormalization should not silently drop stable cache hints. Preserve\nmessage-level cache_control when the merged system fragments agree,\nand fall back to omitting it when the fragments conflict.\n\nConstraint: Must keep single-system normalization for Nvidia/Qwen-style chat backends\nRejected: Always copy the first cache_control | could misrepresent conflicting cache boundaries\nConfidence: high\nScope-risk: narrow\nReversibility: clean\nDirective: If system prompt merging changes again, preserve cache_control whenever the merged metadata is unambiguous\nTested: cargo test proxy::providers::transform --manifest-path src-tauri/Cargo.toml\nNot-tested: End-to-end prompt caching behavior against cache-aware OpenAI-compatible upstreams\nRelated: #1881

src-tauri/src/proxy/providers/transform.rs

@@ -113,6 +113,7 @@ pub fn anthropic_to_openai(body: Value, cache_key: Option<&str>) -> Result<Value
         }
     }
 
+    normalize_openai_system_messages(&mut messages);
     result["messages"] = json!(messages);
 
     // 转换参数 — o-series 模型需要 max_completion_tokens
@@ -182,6 +183,73 @@ pub fn anthropic_to_openai(body: Value, cache_key: Option<&str>) -> Result<Value
     Ok(result)
 }
 
+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();
+    let mut inherited_cache_control: Option<Value> = None;
+    let mut cache_control_conflict = false;
+    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);
+                }
+            }
+            _ => {}
+        }
+
+        if let Some(cache_control) = message.get("cache_control") {
+            match &inherited_cache_control {
+                None => inherited_cache_control = Some(cache_control.clone()),
+                Some(existing) if existing == cache_control => {}
+                Some(_) => cache_control_conflict = true,
+            }
+        }
+
+        false
+    });
+
+    if !parts.is_empty() {
+        let mut merged = json!({"role": "system", "content": parts.join("\n")});
+        if !cache_control_conflict {
+            if let Some(cache_control) = inherited_cache_control {
+                merged["cache_control"] = cache_control;
+            }
+        }
+        messages.insert(0, merged);
+    }
+}
+
 /// 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,
@@ -560,6 +628,50 @@ mod tests {
         assert_eq!(result["tools"][0]["function"]["name"], "get_weather");
     }
 
+    #[test]
+    fn test_anthropic_to_openai_normalizes_fragmented_system_messages() {
+        let input = json!({
+            "model": "claude-3-sonnet",
+            "max_tokens": 1024,
+            "system": [
+                {"type": "text", "text": "You are Claude Code.", "cache_control": {"type": "ephemeral"}},
+                {"type": "text", "text": "Be concise."}
+            ],
+            "messages": [
+                {"role": "system", "content": "Follow repo conventions."},
+                {"role": "user", "content": "Hello"}
+            ]
+        });
+
+        let result = anthropic_to_openai(input, None).unwrap();
+        assert_eq!(result["messages"].as_array().unwrap().len(), 2);
+        assert_eq!(result["messages"][0]["role"], "system");
+        assert_eq!(
+            result["messages"][0]["content"],
+            "You are Claude Code.\nBe concise.\nFollow repo conventions."
+        );
+        assert_eq!(result["messages"][0]["cache_control"]["type"], "ephemeral");
+        assert_eq!(result["messages"][1]["role"], "user");
+    }
+
+    #[test]
+    fn test_anthropic_to_openai_drops_conflicting_system_cache_control_when_merging() {
+        let input = json!({
+            "model": "claude-3-sonnet",
+            "max_tokens": 1024,
+            "system": [
+                {"type": "text", "text": "You are Claude Code.", "cache_control": {"type": "ephemeral"}},
+                {"type": "text", "text": "Be concise.", "cache_control": {"type": "ephemeral", "ttl": "5m"}}
+            ],
+            "messages": [{"role": "user", "content": "Hello"}]
+        });
+
+        let result = anthropic_to_openai(input, None).unwrap();
+        assert_eq!(result["messages"][0]["role"], "system");
+        assert_eq!(result["messages"][0]["content"], "You are Claude Code.\nBe concise.");
+        assert!(result["messages"][0].get("cache_control").is_none());
+    }
+
     #[test]
     fn test_anthropic_to_openai_tool_use() {
         let input = json!({