fix: preserve system cache breakpoint in OAuth mimic

backend/internal/service/gateway_anthropic_apikey_passthrough_test.go

@@ -798,11 +798,12 @@ func TestGatewayService_AnthropicOAuthMimic_RewritesSystemWithBillingBlock(t *te
 	gin.SetMode(gin.TestMode)
 
 	tests := []struct {
-		name               string
-		body               string
-		wantModel          string
-		wantOriginalSystem string
-		wantMetadataUserID string
+		name                       string
+		body                       string
+		wantModel                  string
+		wantOriginalSystem         string
+		wantOriginalSystemCacheTTL string
+		wantMetadataUserID         string
 	}{
 		{
 			name:               "sonnet system array",
@@ -817,11 +818,12 @@ func TestGatewayService_AnthropicOAuthMimic_RewritesSystemWithBillingBlock(t *te
 			wantOriginalSystem: "x-anthropic-billing-header keep",
 		},
 		{
-			name:               "haiku full mimicry",
-			body:               `{"model":"claude-haiku-4-5","metadata":{"user_id":"pi-session-metadata"},"system":[{"type":"text","text":"Pi project instructions","cache_control":{"type":"ephemeral"}}],"thinking":{"type":"enabled","budget_tokens":1024},"messages":[{"role":"user","content":[{"type":"text","text":"hello"}]}]}`,
-			wantModel:          "claude-haiku-4-5-20251001",
-			wantOriginalSystem: "Pi project instructions",
-			wantMetadataUserID: "pi-session-metadata",
+			name:                       "haiku full mimicry",
+			body:                       `{"model":"claude-haiku-4-5","metadata":{"user_id":"pi-session-metadata"},"system":[{"type":"text","text":"Pi project instructions","cache_control":{"type":"ephemeral","ttl":"1h"}}],"thinking":{"type":"enabled","budget_tokens":1024},"messages":[{"role":"user","content":[{"type":"text","text":"hello"}]}]}`,
+			wantModel:                  "claude-haiku-4-5-20251001",
+			wantOriginalSystem:         "Pi project instructions",
+			wantOriginalSystemCacheTTL: "1h",
+			wantMetadataUserID:         "pi-session-metadata",
 		},
 	}
 
@@ -912,6 +914,12 @@ func TestGatewayService_AnthropicOAuthMimic_RewritesSystemWithBillingBlock(t *te
 			firstMsg := messages.Array()[0]
 			require.Equal(t, "user", firstMsg.Get("role").String())
 			require.Contains(t, firstMsg.Get("content.0.text").String(), tt.wantOriginalSystem)
+			if tt.wantOriginalSystemCacheTTL != "" {
+				require.Equal(t, "ephemeral", firstMsg.Get("content.0.cache_control.type").String())
+				require.Equal(t, tt.wantOriginalSystemCacheTTL, firstMsg.Get("content.0.cache_control.ttl").String())
+			} else {
+				require.False(t, firstMsg.Get("content.0.cache_control").Exists())
+			}
 
 			if tt.wantMetadataUserID != "" {
 				require.Equal(t, tt.wantMetadataUserID, gjson.GetBytes(upstream.lastBody, "metadata.user_id").String())

backend/internal/service/gateway_claude_oauth_body.go

@@ -854,26 +854,42 @@ func ValidateClaudeOAuthSystemPromptBlocksConfig(raw string) error {
 	return nil
 }
 
+func extractSystemTextAndCacheControl(system any) (string, any) {
+	switch v := system.(type) {
+	case string:
+		return strings.TrimSpace(v), nil
+	case []any:
+		var parts []string
+		var cacheControl any
+		for _, item := range v {
+			m, ok := item.(map[string]any)
+			if !ok {
+				continue
+			}
+			text, ok := m["text"].(string)
+			if !ok || strings.TrimSpace(text) == "" {
+				continue
+			}
+			parts = append(parts, text)
+			// system blocks are collapsed into one messages text block below.
+			// Preserve the last original breakpoint as the closest equivalent
+			// boundary, including its client-selected TTL.
+			if cc, exists := m["cache_control"]; exists && cc != nil {
+				cacheControl = cc
+			}
+		}
+		return strings.Join(parts, "\n\n"), cacheControl
+	default:
+		return "", nil
+	}
+}
+
 func rewriteSystemForNonClaudeCodeWithPromptBlocks(body []byte, system any, expansionPrompt string, blocksConfig string) []byte {
 	system = normalizeSystemParam(system)
 	expansionPrompt = defaultClaudeOAuthExpansionPrompt(expansionPrompt)
 
-	// 1. 提取原始 system prompt 文本
-	var originalSystemText string
-	switch v := system.(type) {
-	case string:
-		originalSystemText = strings.TrimSpace(v)
-	case []any:
-		var parts []string
-		for _, item := range v {
-			if m, ok := item.(map[string]any); ok {
-				if text, ok := m["text"].(string); ok && strings.TrimSpace(text) != "" {
-					parts = append(parts, text)
-				}
-			}
-		}
-		originalSystemText = strings.Join(parts, "\n\n")
-	}
+	// 1. 提取原始 system prompt 文本及其缓存断点
+	originalSystemText, originalSystemCacheControl := extractSystemTextAndCacheControl(system)
 
 	// 2. 构造 system 数组，对齐真实 Claude Code CLI 的 3-block 形态：
 	//    [0] billing attribution block（cc_version={cliVer}.{fp}; cc_entrypoint=cli;）
@@ -906,10 +922,17 @@ func rewriteSystemForNonClaudeCodeWithPromptBlocks(body []byte, system any, expa
 	//    模型仍通过 messages 接收完整指令，保留客户端功能
 	ccPromptTrimmed := strings.TrimSpace(claudeCodeSystemPrompt)
 	if originalSystemText != "" && originalSystemText != ccPromptTrimmed && !hasClaudeCodePrefix(originalSystemText) {
+		instructionBlock := map[string]any{
+			"type": "text",
+			"text": "[System Instructions]\n" + originalSystemText,
+		}
+		if originalSystemCacheControl != nil {
+			instructionBlock["cache_control"] = originalSystemCacheControl
+		}
 		instrMsg, err1 := json.Marshal(map[string]any{
 			"role": "user",
 			"content": []map[string]any{
-				{"type": "text", "text": "[System Instructions]\n" + originalSystemText},
+				instructionBlock,
 			},
 		})
 		ackMsg, err2 := json.Marshal(map[string]any{

backend/internal/service/gateway_prompt_test.go

@@ -482,6 +482,34 @@ func TestRewriteSystemForNonClaudeCodeWithPrompt_UsesCustomExpansionPrompt(t *te
 	require.Equal(t, "ephemeral", system.Array()[2].Get("cache_control.type").String())
 }
 
+func TestRewriteSystemForNonClaudeCode_PreservesSystemCacheControlOnMigratedMessage(t *testing.T) {
+	body := []byte(`{"model":"claude-3","system":[{"type":"text","text":"Stable project instructions","cache_control":{"type":"ephemeral","ttl":"1h"}}],"messages":[{"role":"user","content":"hello"}]}`)
+	system := []any{
+		map[string]any{
+			"type":          "text",
+			"text":          "Stable project instructions",
+			"cache_control": map[string]any{"type": "ephemeral", "ttl": "1h"},
+		},
+	}
+
+	result := rewriteSystemForNonClaudeCode(body, system)
+
+	require.Equal(t, "[System Instructions]\nStable project instructions", gjson.GetBytes(result, "messages.0.content.0.text").String())
+	require.Equal(t, "ephemeral", gjson.GetBytes(result, "messages.0.content.0.cache_control.type").String())
+	require.Equal(t, "1h", gjson.GetBytes(result, "messages.0.content.0.cache_control.ttl").String())
+}
+
+func TestRewriteSystemForNonClaudeCode_LeavesMigratedMessageUncachedWithoutSystemBreakpoint(t *testing.T) {
+	body := []byte(`{"model":"claude-3","system":[{"type":"text","text":"Project instructions"}],"messages":[{"role":"user","content":"hello"}]}`)
+	system := []any{
+		map[string]any{"type": "text", "text": "Project instructions"},
+	}
+
+	result := rewriteSystemForNonClaudeCode(body, system)
+
+	require.False(t, gjson.GetBytes(result, "messages.0.content.0.cache_control").Exists())
+}
+
 func TestRewriteSystemForNonClaudeCodeWithPromptBlocks_UsesConfiguredBlocks(t *testing.T) {
 	body := []byte(`{"model":"claude-3","system":"Project instructions","messages":[{"role":"user","content":"hello"}]}`)
 	blocks := `{