fix(openai): 使用 prompt_cache_key 兜底粘性会话

opencode 请求不带 session_id/conversation_id，导致粘性会话失效。现在按 header 优先、prompt_cache_key 兜底生成 session hash，并补充单测验证优先级。

backend/internal/handler/openai_gateway_handler.go

@@ -186,8 +186,8 @@ func (h *OpenAIGatewayHandler) Responses(c *gin.Context) {
 		return
 	}
 
-	// Generate session hash (from header for OpenAI)
-	sessionHash := h.gatewayService.GenerateSessionHash(c)
+	// Generate session hash (header first; fallback to prompt_cache_key)
+	sessionHash := h.gatewayService.GenerateSessionHash(c, reqBody)
 
 	const maxAccountSwitches = 3
 	switchCount := 0

backend/internal/service/openai_gateway_service.go

@@ -133,12 +133,30 @@ func NewOpenAIGatewayService(
 	}
 }
 
-// GenerateSessionHash generates session hash from header (OpenAI uses session_id header)
-func (s *OpenAIGatewayService) GenerateSessionHash(c *gin.Context) string {
-	sessionID := c.GetHeader("session_id")
+// GenerateSessionHash generates a sticky-session hash for OpenAI requests.
+//
+// Priority:
+//  1. Header: session_id
+//  2. Header: conversation_id
+//  3. Body:   prompt_cache_key (opencode)
+func (s *OpenAIGatewayService) GenerateSessionHash(c *gin.Context, reqBody map[string]any) string {
+	if c == nil {
+		return ""
+	}
+
+	sessionID := strings.TrimSpace(c.GetHeader("session_id"))
+	if sessionID == "" {
+		sessionID = strings.TrimSpace(c.GetHeader("conversation_id"))
+	}
+	if sessionID == "" && reqBody != nil {
+		if v, ok := reqBody["prompt_cache_key"].(string); ok {
+			sessionID = strings.TrimSpace(v)
+		}
+	}
 	if sessionID == "" {
 		return ""
 	}
+
 	hash := sha256.Sum256([]byte(sessionID))
 	return hex.EncodeToString(hash[:])
 }

backend/internal/service/openai_gateway_service_test.go

@@ -49,6 +49,49 @@ func (c stubConcurrencyCache) GetAccountsLoadBatch(ctx context.Context, accounts
 	return out, nil
 }
 
+func TestOpenAIGatewayService_GenerateSessionHash_Priority(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+	rec := httptest.NewRecorder()
+	c, _ := gin.CreateTestContext(rec)
+	c.Request = httptest.NewRequest(http.MethodPost, "/openai/v1/responses", nil)
+
+	svc := &OpenAIGatewayService{}
+
+	// 1) session_id header wins
+	c.Request.Header.Set("session_id", "sess-123")
+	c.Request.Header.Set("conversation_id", "conv-456")
+	h1 := svc.GenerateSessionHash(c, map[string]any{"prompt_cache_key": "ses_aaa"})
+	if h1 == "" {
+		t.Fatalf("expected non-empty hash")
+	}
+
+	// 2) conversation_id used when session_id absent
+	c.Request.Header.Del("session_id")
+	h2 := svc.GenerateSessionHash(c, map[string]any{"prompt_cache_key": "ses_aaa"})
+	if h2 == "" {
+		t.Fatalf("expected non-empty hash")
+	}
+	if h1 == h2 {
+		t.Fatalf("expected different hashes for different keys")
+	}
+
+	// 3) prompt_cache_key used when both headers absent
+	c.Request.Header.Del("conversation_id")
+	h3 := svc.GenerateSessionHash(c, map[string]any{"prompt_cache_key": "ses_aaa"})
+	if h3 == "" {
+		t.Fatalf("expected non-empty hash")
+	}
+	if h2 == h3 {
+		t.Fatalf("expected different hashes for different keys")
+	}
+
+	// 4) empty when no signals
+	h4 := svc.GenerateSessionHash(c, map[string]any{})
+	if h4 != "" {
+		t.Fatalf("expected empty hash when no signals")
+	}
+}
+
 func TestOpenAISelectAccountWithLoadAwareness_FiltersUnschedulable(t *testing.T) {
 	now := time.Now()
 	resetAt := now.Add(10 * time.Minute)