From b11e36963f680c9754713a64c98b9130c30ef758 Mon Sep 17 00:00:00 2001
From: YoVinchen <gzh298255@gmail.com>
Date: Wed, 8 Apr 2026 09:57:49 +0800
Subject: [PATCH] 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 | 39 ++++++++++++++++++++--
 1 file changed, 37 insertions(+), 2 deletions(-)

diff --git a/src-tauri/src/proxy/providers/transform.rs b/src-tauri/src/proxy/providers/transform.rs
index 2b660f33..4d3aae4b 100644
--- a/src-tauri/src/proxy/providers/transform.rs
+++ b/src-tauri/src/proxy/providers/transform.rs
@@ -206,6 +206,8 @@ fn normalize_openai_system_messages(messages: &mut Vec<Value>) {
     }
 
     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;
@@ -226,11 +228,25 @@ fn normalize_openai_system_messages(messages: &mut Vec<Value>) {
             _ => {}
         }
 
+        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() {
-        messages.insert(0, json!({"role": "system", "content": parts.join("\n")}));
+        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);
     }
 }
 
@@ -618,7 +634,7 @@ mod tests {
             "model": "claude-3-sonnet",
             "max_tokens": 1024,
             "system": [
-                {"type": "text", "text": "You are Claude Code."},
+                {"type": "text", "text": "You are Claude Code.", "cache_control": {"type": "ephemeral"}},
                 {"type": "text", "text": "Be concise."}
             ],
             "messages": [
@@ -634,9 +650,28 @@ mod tests {
             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!({