From 7ff3936efe988034200918cd3ededb0189f8e5bf Mon Sep 17 00:00:00 2001
From: Luis Pater <webmaster@idotorg.org>
Date: Sat, 31 Jan 2026 01:42:58 +0800
Subject: [PATCH] fix(caching): ensure prompt-caching beta is always appended
 and add multi-turn cache control tests

---
 .../runtime/executor/caching_verify_test.go   | 48 +++++++++++++++++++
 internal/runtime/executor/claude_executor.go  |  6 ++-
 2 files changed, 53 insertions(+), 1 deletion(-)

diff --git a/internal/runtime/executor/caching_verify_test.go b/internal/runtime/executor/caching_verify_test.go
index 599c1aec..6088d304 100644
--- a/internal/runtime/executor/caching_verify_test.go
+++ b/internal/runtime/executor/caching_verify_test.go
@@ -165,6 +165,54 @@ func TestEnsureCacheControl(t *testing.T) {
 			t.Errorf("system should have cache_control even with empty tools array")
 		}
 	})
+
+	// Test case 8: Messages caching for multi-turn (second-to-last user)
+	t.Run("Messages Caching Second-To-Last User", func(t *testing.T) {
+		input := []byte(`{
+			"model": "claude-3-5-sonnet",
+			"messages": [
+				{"role": "user", "content": "First user"},
+				{"role": "assistant", "content": "Assistant reply"},
+				{"role": "user", "content": "Second user"},
+				{"role": "assistant", "content": "Assistant reply 2"},
+				{"role": "user", "content": "Third user"}
+			]
+		}`)
+		output := ensureCacheControl(input)
+
+		cacheType := gjson.GetBytes(output, "messages.2.content.0.cache_control.type")
+		if cacheType.String() != "ephemeral" {
+			t.Errorf("cache_control not found on second-to-last user turn. Output: %s", string(output))
+		}
+
+		lastUserCache := gjson.GetBytes(output, "messages.4.content.0.cache_control")
+		if lastUserCache.Exists() {
+			t.Errorf("last user turn should NOT have cache_control")
+		}
+	})
+
+	// Test case 9: Existing message cache_control should skip injection
+	t.Run("Messages Skip When Cache Control Exists", func(t *testing.T) {
+		input := []byte(`{
+			"model": "claude-3-5-sonnet",
+			"messages": [
+				{"role": "user", "content": [{"type": "text", "text": "First user"}]},
+				{"role": "assistant", "content": [{"type": "text", "text": "Assistant reply", "cache_control": {"type": "ephemeral"}}]},
+				{"role": "user", "content": [{"type": "text", "text": "Second user"}]}
+			]
+		}`)
+		output := ensureCacheControl(input)
+
+		userCache := gjson.GetBytes(output, "messages.0.content.0.cache_control")
+		if userCache.Exists() {
+			t.Errorf("cache_control should NOT be injected when a message already has cache_control")
+		}
+
+		existingCache := gjson.GetBytes(output, "messages.1.content.0.cache_control.type")
+		if existingCache.String() != "ephemeral" {
+			t.Errorf("existing cache_control should be preserved. Output: %s", string(output))
+		}
+	})
 }
 
 // TestCacheControlOrder verifies the correct order: tools -> system -> messages
diff --git a/internal/runtime/executor/claude_executor.go b/internal/runtime/executor/claude_executor.go
index 3edf5080..83c231bd 100644
--- a/internal/runtime/executor/claude_executor.go
+++ b/internal/runtime/executor/claude_executor.go
@@ -642,13 +642,17 @@ func applyClaudeHeaders(r *http.Request, auth *cliproxyauth.Auth, apiKey string,
 		ginHeaders = ginCtx.Request.Header
 	}
 
-	baseBetas := "claude-code-20250219,oauth-2025-04-20,interleaved-thinking-2025-05-14,fine-grained-tool-streaming-2025-05-14,prompt-caching-2024-07-31"
+	promptCachingBeta := "prompt-caching-2024-07-31"
+	baseBetas := "claude-code-20250219,oauth-2025-04-20,interleaved-thinking-2025-05-14,fine-grained-tool-streaming-2025-05-14," + promptCachingBeta
 	if val := strings.TrimSpace(ginHeaders.Get("Anthropic-Beta")); val != "" {
 		baseBetas = val
 		if !strings.Contains(val, "oauth") {
 			baseBetas += ",oauth-2025-04-20"
 		}
 	}
+	if !strings.Contains(baseBetas, promptCachingBeta) {
+		baseBetas += "," + promptCachingBeta
+	}
 
 	// Merge extra betas from request body
 	if len(extraBetas) > 0 {