Merge pull request #7432 from doublepi123/fix/antigravity-compat-bare-gemini-model

fix(antigravity): 裸 Gemini 模型名在所有转发入口解析到思考变体，修复 404/502

backend/internal/service/antigravity_gateway_claude.go

@@ -49,7 +49,11 @@ func (s *AntigravityGatewayService) Forward(ctx context.Context, c *gin.Context,
 	}
 
 	originalModel := claudeReq.Model
-	mappedModel := s.getMappedModel(account, claudeReq.Model)
+	mappedModel := s.getMappedModelForThinkingLevel(
+		account,
+		claudeReq.Model,
+		geminiThinkingLevelFromClaudeThinking(claudeReq.Thinking),
+	)
 	if mappedModel == "" {
 		MarkOpsClientBusinessLimited(c, OpsClientBusinessLimitedReasonLocalFeatureGate)
 		return nil, s.writeClaudeError(c, http.StatusForbidden, "permission_error", fmt.Sprintf("model %s not in whitelist", claudeReq.Model))

backend/internal/service/antigravity_gateway_compat.go

@@ -212,7 +212,11 @@ func (s *AntigravityGatewayService) prepareAntigravityCompatCall(
 		return nil, s.writeAntigravityCompatError(c, http.StatusBadRequest, "invalid_request_error", "Invalid request body")
 	}
 
-	mappedModel := s.getMappedModel(account, request.originalModel)
+	mappedModel := s.getMappedModelForThinkingLevel(
+		account,
+		request.originalModel,
+		geminiThinkingLevelFromClaudeThinking(claudeRequest.Thinking),
+	)
 	if mappedModel == "" {
 		MarkOpsClientBusinessLimited(c, OpsClientBusinessLimitedReasonLocalFeatureGate)
 		message := fmt.Sprintf("model %s not in whitelist", request.originalModel)

backend/internal/service/antigravity_gateway_gemini.go

@@ -86,13 +86,11 @@ func (s *AntigravityGatewayService) ForwardGemini(ctx context.Context, c *gin.Co
 		return nil, s.writeGoogleError(c, http.StatusNotFound, "Unsupported action: "+action)
 	}
 
-	// 裸模型名（gemini-3.8-flash）+ thinkingConfig → 账号映射表里对应的 -low/-medium/-high 变体；
+	// 裸模型名（gemini-3.8-flash）按 thinkingConfig 解析到 -low/-medium/-high 变体；
 	// 裸名有显式映射时保持原行为。
-	mappedModel, variantResolved := resolveGeminiThinkingVariant(account, originalModel, body)
-	if !variantResolved {
-		mappedModel = s.getMappedModel(account, originalModel)
-	} else {
-		logger.LegacyPrintf("service.antigravity_gateway", "%s resolved bare Gemini model %s to thinking variant %s", prefix, originalModel, mappedModel)
+	mappedModel := s.getMappedModelForThinkingLevel(account, originalModel, geminiThinkingLevelFromBody(body))
+	if mappedModel != "" && mappedModel != originalModel {
+		logger.LegacyPrintf("service.antigravity_gateway", "%s mapped Gemini model %s to %s", prefix, originalModel, mappedModel)
 	}
 	if mappedModel == "" {
 		MarkOpsClientBusinessLimited(c, OpsClientBusinessLimitedReasonLocalFeatureGate)

backend/internal/service/antigravity_gateway_service.go

@@ -294,6 +294,18 @@ func mapAntigravityModel(account *Account, requestedModel string) string {
 // getMappedModel 获取映射后的模型名
 // 完全依赖映射配置：账户映射（通配符）→ 默认映射兜底
 func (s *AntigravityGatewayService) getMappedModel(account *Account, requestedModel string) string {
+	return s.getMappedModelForThinkingLevel(account, requestedModel, "")
+}
+
+// getMappedModelForThinkingLevel 在常规映射之前先把裸 Gemini 模型名
+// （gemini-3.x-flash）解析到上游目录里真实存在的 -low/-medium/-high/-tiered 变体。
+// Antigravity 目录只登记带后缀的变体，裸名直接转发会被上游以
+// 404 "Requested entity was not found." 拒绝，因此所有转发入口都必须经过这一步。
+// thinkingLevel 为空时按 high 兜底；调用方可按自身协议传入推导出的档位。
+func (s *AntigravityGatewayService) getMappedModelForThinkingLevel(account *Account, requestedModel string, thinkingLevel string) string {
+	if mapped, ok := resolveGeminiThinkingVariantForLevel(account, requestedModel, thinkingLevel); ok {
+		return mapped
+	}
 	return mapAntigravityModel(account, requestedModel)
 }
 

backend/internal/service/antigravity_gemini_thinking_variant.go

@@ -3,6 +3,8 @@ package service
 import (
 	"encoding/json"
 	"strings"
+
+	"github.com/Wei-Shaw/sub2api/internal/pkg/antigravity"
 )
 
 // Gemini 原生请求（/v1beta/models/{model}:generateContent 等）经 Antigravity 账号转发时，
@@ -83,6 +85,29 @@ func geminiThinkingLevelFromBody(body []byte) string {
 	}
 }
 
+// geminiThinkingLevelFromClaudeThinking 用 Claude Messages 协议的 thinking 配置推导档位，
+// 阈值与 geminiThinkingLevelFromBody 保持一致，使同一请求无论走 Gemini 原生还是
+// Chat Completions / Messages 兼容层都落到同一个上游变体。
+func geminiThinkingLevelFromClaudeThinking(thinking *antigravity.ThinkingConfig) string {
+	if thinking == nil {
+		return "high"
+	}
+	if strings.EqualFold(strings.TrimSpace(thinking.Type), "disabled") {
+		return "low"
+	}
+	budget := thinking.BudgetTokens
+	switch {
+	case budget <= 0:
+		return "high" // 动态思考 / 未指定预算
+	case budget <= geminiThinkingBudgetLowMax:
+		return "low"
+	case budget <= geminiThinkingBudgetMediumMax:
+		return "medium"
+	default:
+		return "high"
+	}
+}
+
 // accountRawModelMappingHasKey 判断 key 是否由用户写在账号 credentials.model_mapping 里
 // （而不是 resolveModelMapping 运行时补进来的默认透传条目）。
 func accountRawModelMappingHasKey(account *Account, key string) bool {
@@ -97,6 +122,13 @@ func accountRawModelMappingHasKey(account *Account, key string) bool {
 // resolveGeminiThinkingVariant 为裸 Gemini 模型名挑选账号映射表里存在的思考深度变体。
 // 返回 (映射后的上游模型名, 是否命中)。未命中时调用方应回退到常规 getMappedModel 流程。
 func resolveGeminiThinkingVariant(account *Account, requestedModel string, body []byte) (string, bool) {
+	return resolveGeminiThinkingVariantForLevel(account, requestedModel, geminiThinkingLevelFromBody(body))
+}
+
+// resolveGeminiThinkingVariantForLevel 是 resolveGeminiThinkingVariant 的协议无关内核：
+// 调用方负责按自身协议推导 preferred 档位（Gemini 原生读 generationConfig.thinkingConfig，
+// Claude/OpenAI 兼容层读 thinking.budget_tokens），此处只做映射查找与降级。
+func resolveGeminiThinkingVariantForLevel(account *Account, requestedModel string, preferred string) (string, bool) {
 	if account == nil {
 		return "", false
 	}
@@ -120,7 +152,9 @@ func resolveGeminiThinkingVariant(account *Account, requestedModel string, body
 		}
 	}
 
-	preferred := geminiThinkingLevelFromBody(body)
+	if preferred == "" {
+		preferred = "high"
+	}
 	order := []string{preferred}
 	for _, level := range []string{"high", "medium", "low", "tiered"} {
 		if level != preferred {

backend/internal/service/antigravity_gemini_thinking_variant_test.go

@@ -3,6 +3,7 @@ package service
 import (
 	"testing"
 
+	"github.com/Wei-Shaw/sub2api/internal/pkg/antigravity"
 	"github.com/stretchr/testify/require"
 )
 
@@ -116,3 +117,96 @@ func TestResolveGeminiThinkingVariant(t *testing.T) {
 		require.Empty(t, got)
 	})
 }
+
+func TestGeminiThinkingLevelFromClaudeThinking(t *testing.T) {
+	tests := []struct {
+		name     string
+		thinking *antigravity.ThinkingConfig
+		want     string
+	}{
+		{"nil thinking", nil, "high"},
+		{"disabled", &antigravity.ThinkingConfig{Type: "disabled"}, "low"},
+		{"enabled without budget", &antigravity.ThinkingConfig{Type: "enabled"}, "high"},
+		{"budget 1024 -> low", &antigravity.ThinkingConfig{Type: "enabled", BudgetTokens: 1024}, "low"},
+		{"budget 4000 -> medium", &antigravity.ThinkingConfig{Type: "enabled", BudgetTokens: 4000}, "medium"},
+		{"budget 20000 -> high", &antigravity.ThinkingConfig{Type: "enabled", BudgetTokens: 20000}, "high"},
+	}
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			require.Equal(t, tt.want, geminiThinkingLevelFromClaudeThinking(tt.thinking))
+		})
+	}
+}
+
+// 锁定线上 404 回归：Chat Completions / Messages 兼容层收到裸 Gemini 模型名时，
+// 必须和 Gemini 原生路径一样解析到带后缀的上游变体，而不是把裸名透传给上游。
+func TestResolveGeminiThinkingVariantForLevel_CompatPathResolvesBareModel(t *testing.T) {
+	catalog := map[string]string{
+		"gemini-3.8-flash-low":    "gemini-3.8-flash-low",
+		"gemini-3.8-flash-medium": "gemini-3.8-flash-medium",
+		"gemini-3.8-flash-high":   "gemini-3.8-flash-high",
+	}
+	account := newAntigravityAccountWithMapping(catalog)
+
+	for _, tc := range []struct{ level, want string }{
+		{"low", "gemini-3.8-flash-low"},
+		{"medium", "gemini-3.8-flash-medium"},
+		{"high", "gemini-3.8-flash-high"},
+		{"", "gemini-3.8-flash-high"},
+	} {
+		t.Run("level="+tc.level, func(t *testing.T) {
+			got, matched := resolveGeminiThinkingVariantForLevel(account, "gemini-3.8-flash", tc.level)
+			require.True(t, matched)
+			require.Equal(t, tc.want, got)
+		})
+	}
+
+	t.Run("already suffixed model is untouched", func(t *testing.T) {
+		got, matched := resolveGeminiThinkingVariantForLevel(account, "gemini-3.8-flash-high", "low")
+		require.False(t, matched)
+		require.Empty(t, got)
+	})
+}
+
+// 同一请求无论走 Gemini 原生还是兼容层，都必须落到同一个上游变体。
+func TestGeminiThinkingLevel_NativeAndCompatAgree(t *testing.T) {
+	native := geminiThinkingLevelFromBody([]byte(`{"generationConfig":{"thinkingConfig":{"thinkingBudget":4000}}}`))
+	compat := geminiThinkingLevelFromClaudeThinking(&antigravity.ThinkingConfig{Type: "enabled", BudgetTokens: 4000})
+	require.Equal(t, native, compat)
+}
+
+// 回归锁：裸名解析必须发生在 getMappedModel 这一层，从而覆盖全部 Antigravity
+// 转发入口（/v1beta 原生、/v1/chat/completions、/v1/messages）以及账号测试。
+// 早前只在单个入口打补丁，导致 /v1/messages 仍把裸名透传给上游并触发
+// 404 "Requested entity was not found."
+func TestGetMappedModelResolvesBareGeminiModelForAllEntrypoints(t *testing.T) {
+	svc := &AntigravityGatewayService{}
+	account := newAntigravityAccountWithMapping(map[string]string{
+		"gemini-3.8-flash-low":    "gemini-3.8-flash-low",
+		"gemini-3.8-flash-medium": "gemini-3.8-flash-medium",
+		"gemini-3.8-flash-high":   "gemini-3.8-flash-high",
+	})
+
+	t.Run("default entrypoint falls back to high", func(t *testing.T) {
+		require.Equal(t, "gemini-3.8-flash-high", svc.getMappedModel(account, "gemini-3.8-flash"))
+	})
+
+	t.Run("caller supplied level wins", func(t *testing.T) {
+		require.Equal(t, "gemini-3.8-flash-low",
+			svc.getMappedModelForThinkingLevel(account, "gemini-3.8-flash", "low"))
+		require.Equal(t, "gemini-3.8-flash-medium",
+			svc.getMappedModelForThinkingLevel(account, "gemini-3.8-flash", "medium"))
+	})
+
+	t.Run("models/ prefix is handled", func(t *testing.T) {
+		require.Equal(t, "gemini-3.8-flash-high", svc.getMappedModel(account, "models/gemini-3.8-flash"))
+	})
+
+	t.Run("suffixed model keeps normal mapping", func(t *testing.T) {
+		require.Equal(t, "gemini-3.8-flash-low", svc.getMappedModel(account, "gemini-3.8-flash-low"))
+	})
+
+	t.Run("unsupported model still reports unsupported", func(t *testing.T) {
+		require.Empty(t, svc.getMappedModel(account, "gemini-9.9-flash"))
+	})
+}