From c807c6e3b0c74b77ad4c7a8213a1d5690d34e462 Mon Sep 17 00:00:00 2001
From: kvc0 <3454741+kvc0@users.noreply.github.com>
Date: Thu, 23 Apr 2026 08:45:02 -0700
Subject: [PATCH] server: (anthropic API) fix prefix caching (#21793)

When testing claude code against llama.cpp, I noticed that only
n_past 18577 was used even when context was 60k or more. The log
in llama-server says:
```
slot update_slots: id  3 | task 10342 | old: ... ; cch= | defa0;You are
slot update_slots: id  3 | task 10342 | new: ... ; cch= | 1c8b4;
```
I observed that the cch value changed every time. Reading about that,
the x-anthropic-billing-header system message seems to be specially
handled inside of the anthropic api. I could remove it, but there
is a meaningful string sometimes included at the end. So instead,
I just replace the changing cch checksum with fffff.

I'm treating this as an anthropic message body API detail - I think this
is the right way to do this, but by all means please correct me!

It's always 5 hexadecimal characters, but I've written the replacement
defensively in case they change the protocol.
---
 tools/server/server-chat.cpp | 41 +++++++++++++++++++++++++++++++++++-
 1 file changed, 40 insertions(+), 1 deletion(-)

diff --git a/tools/server/server-chat.cpp b/tools/server/server-chat.cpp
index ef586d1e1..34ec7982b 100644
--- a/tools/server/server-chat.cpp
+++ b/tools/server/server-chat.cpp
@@ -281,6 +281,42 @@ json server_chat_convert_responses_to_chatcmpl(const json & response_body) {
     return chatcmpl_body;
 }
 
+// Edits the cch section of an "x-anthropic-billing-header" system prompt.
+// Does nothing to any other prompt.
+//
+// This is a claude message with a "cch=ef01a" attribute that breaks prefix caching.
+// The cch stamp is a whitebox end-to-end integrity hint. It's not meaningful as a
+// system prompt data, particularly to llama.cpp, but its presence means the prefix
+// cache will not get past it: It changes on each request.
+//
+// Reference: https://github.com/ggml-org/llama.cpp/pull/21793
+// Example header:
+// ```
+// x-anthropic-billing-header: cc_version=2.1.101.e51; cc_entrypoint=cli; cch=a5145;You are Claude Code, Anthropic's official CLI for Claude.
+//                                                                            ^^^^^
+// ```
+static void normalize_anthropic_billing_header(std::string & system_text) {
+    if (system_text.rfind("x-anthropic-billing-header:", 0) != 0) {
+        return;
+    }
+
+    const size_t header_prefix_length = strlen("x-anthropic-billing-header:");
+    const size_t cch_length = 5;
+    const size_t index_cch = system_text.find("cch=", header_prefix_length);
+    if (index_cch == std::string::npos) {
+        return;
+    }
+
+    const size_t index_replace = index_cch + 4;
+    if (index_replace + cch_length < system_text.length() && system_text[index_replace + cch_length] == ';') {
+        for (size_t i = 0; i < cch_length; ++i) {
+            system_text[index_replace + i] = 'f';
+        }
+    } else {
+        LOG_ERR("anthropic string not as expected: %s", system_text.c_str());
+    }
+}
+
 json server_chat_convert_anthropic_to_oai(const json & body) {
     json oai_body;
 
@@ -292,10 +328,13 @@ json server_chat_convert_anthropic_to_oai(const json & body) {
 
         if (system_param.is_string()) {
             system_content = system_param.get<std::string>();
+            normalize_anthropic_billing_header(system_content);
         } else if (system_param.is_array()) {
             for (const auto & block : system_param) {
                 if (json_value(block, "type", std::string()) == "text") {
-                    system_content += json_value(block, "text", std::string());
+                    auto system_text = json_value(block, "text", std::string());
+                    normalize_anthropic_billing_header(system_text);
+                    system_content += system_text;
                 }
             }
         }