feat: 实现统一模型 ID 机制

实现统一模型 ID 格式 (provider_id/model_name)，支持跨协议模型标识和 Smart Passthrough。

核心变更：
- 新增 pkg/modelid 包：解析、格式化、校验统一模型 ID
- 数据库迁移：models 表使用 UUID 主键 + UNIQUE(provider_id, model_name) 约束
- Repository 层：FindByProviderAndModelName、ListEnabled 方法
- Service 层：联合唯一校验、provider ID 字符集校验
- Conversion 层：ExtractModelName、RewriteRequestModelName/RewriteResponseModelName 方法
- Handler 层：统一模型 ID 路由、Smart Passthrough、Models API 本地聚合
- 新增 error-responses、unified-model-id 规范

测试覆盖：
- 单元测试：modelid、conversion、handler、service、repository
- 集成测试：统一模型 ID 路由、Smart Passthrough 保真性、跨协议转换
- 迁移测试：UUID 主键、UNIQUE 约束、级联删除

OpenSpec：
- 归档 unified-model-id 变更到 archive/2026-04-21-unified-model-id
- 同步 11 个 delta specs 到 main specs
- 新增 error-responses、unified-model-id 规范文件

backend/tests/integration/e2e_conversion_test.go

@@ -8,8 +8,6 @@ import (
 	"io"
 	"net/http"
 	"net/http/httptest"
-	"os"
-	"path/filepath"
 	"strings"
 	"testing"
 	"time"
@@ -17,10 +15,7 @@ import (
 	"github.com/gin-gonic/gin"
 	"github.com/stretchr/testify/assert"
 	"github.com/stretchr/testify/require"
-	"gorm.io/driver/sqlite"
-	"gorm.io/gorm"
 
-	"nex/backend/internal/config"
 	"nex/backend/internal/conversion"
 	"nex/backend/internal/conversion/anthropic"
 	openaiConv "nex/backend/internal/conversion/openai"
@@ -40,25 +35,20 @@ func setupE2ETest(t *testing.T) (*gin.Engine, *httptest.Server) {
 		w.Write([]byte(`{"error":"not mocked"}`))
 	}))
 
-	dir, _ := os.MkdirTemp("", "e2e-test-*")
-	db, err := gorm.Open(sqlite.Open(filepath.Join(dir, "test.db")), &gorm.Config{})
-	require.NoError(t, err)
-	err = db.AutoMigrate(&config.Provider{}, &config.Model{}, &config.UsageStats{})
-	require.NoError(t, err)
+	db := setupTestDB(t)
 	t.Cleanup(func() {
 		sqlDB, _ := db.DB()
 		if sqlDB != nil {
 			sqlDB.Close()
 		}
 		upstream.Close()
-		os.RemoveAll(dir)
 	})
 
 	providerRepo := repository.NewProviderRepository(db)
 	modelRepo := repository.NewModelRepository(db)
 	statsRepo := repository.NewStatsRepository(db)
 
-	providerService := service.NewProviderService(providerRepo)
+	providerService := service.NewProviderService(providerRepo, modelRepo)
 	modelService := service.NewModelService(modelRepo, providerRepo)
 	routingService := service.NewRoutingService(modelRepo, providerRepo)
 	statsService := service.NewStatsService(statsRepo)
@@ -105,7 +95,7 @@ func e2eCreateProviderAndModel(t *testing.T, r *gin.Engine, providerID, protocol
 	require.Equal(t, 201, w.Code)
 
 	modelBody, _ := json.Marshal(map[string]string{
-		"id": modelName, "provider_id": providerID, "model_name": modelName,
+		"provider_id": providerID, "model_name": modelName,
 	})
 	w = httptest.NewRecorder()
 	req = httptest.NewRequest("POST", "/api/models", bytes.NewReader(modelBody))
@@ -178,10 +168,10 @@ func TestE2E_OpenAI_NonStream_BasicText(t *testing.T) {
 			},
 		})
 	})
-	e2eCreateProviderAndModel(t, r, "openai-p", "openai", "gpt-4o", upstream.URL)
+	e2eCreateProviderAndModel(t, r, "openai_p", "openai", "gpt-4o", upstream.URL)
 
 	body, _ := json.Marshal(map[string]any{
-		"model": "gpt-4o",
+		"model": "openai_p/gpt-4o",
 		"messages": []map[string]any{
 			{"role": "user", "content": "你好"},
 		},
@@ -195,7 +185,7 @@ func TestE2E_OpenAI_NonStream_BasicText(t *testing.T) {
 	var resp map[string]any
 	require.NoError(t, json.Unmarshal(w.Body.Bytes(), &resp))
 	assert.Equal(t, "chat.completion", resp["object"])
-	assert.Equal(t, "gpt-4o", resp["model"])
+	assert.Equal(t, "openai_p/gpt-4o", resp["model"])
 
 	choices := resp["choices"].([]any)
 	require.Len(t, choices, 1)
@@ -231,10 +221,10 @@ func TestE2E_OpenAI_NonStream_MultiTurn(t *testing.T) {
 			"usage": map[string]any{"prompt_tokens": 100, "completion_tokens": 20, "total_tokens": 120},
 		})
 	})
-	e2eCreateProviderAndModel(t, r, "openai-p", "openai", "gpt-4o", upstream.URL)
+	e2eCreateProviderAndModel(t, r, "openai_p", "openai", "gpt-4o", upstream.URL)
 
 	body, _ := json.Marshal(map[string]any{
-		"model": "gpt-4o",
+		"model": "openai_p/gpt-4o",
 		"messages": []map[string]any{
 			{"role": "system", "content": "你是编程助手"},
 			{"role": "user", "content": "什么是interface?"},
@@ -279,10 +269,10 @@ func TestE2E_OpenAI_NonStream_ToolCalls(t *testing.T) {
 			"usage": map[string]any{"prompt_tokens": 80, "completion_tokens": 18, "total_tokens": 98},
 		})
 	})
-	e2eCreateProviderAndModel(t, r, "openai-p", "openai", "gpt-4o", upstream.URL)
+	e2eCreateProviderAndModel(t, r, "openai_p", "openai", "gpt-4o", upstream.URL)
 
 	body, _ := json.Marshal(map[string]any{
-		"model": "gpt-4o",
+		"model": "openai_p/gpt-4o",
 		"messages": []map[string]any{
 			{"role": "user", "content": "北京天气"},
 		},
@@ -335,10 +325,10 @@ func TestE2E_OpenAI_NonStream_MaxTokens_Length(t *testing.T) {
 			"usage": map[string]any{"prompt_tokens": 20, "completion_tokens": 30, "total_tokens": 50},
 		})
 	})
-	e2eCreateProviderAndModel(t, r, "openai-p", "openai", "gpt-4o", upstream.URL)
+	e2eCreateProviderAndModel(t, r, "openai_p", "openai", "gpt-4o", upstream.URL)
 
 	body, _ := json.Marshal(map[string]any{
-		"model": "gpt-4o",
+		"model": "openai_p/gpt-4o",
 		"messages": []map[string]any{{"role": "user", "content": "介绍AI历史"}},
 		"max_tokens": 30,
 	})
@@ -372,10 +362,10 @@ func TestE2E_OpenAI_NonStream_UsageWithReasoning(t *testing.T) {
 			},
 		})
 	})
-	e2eCreateProviderAndModel(t, r, "openai-p", "openai", "o3", upstream.URL)
+	e2eCreateProviderAndModel(t, r, "openai_p", "openai", "o3", upstream.URL)
 
 	body, _ := json.Marshal(map[string]any{
-		"model": "o3",
+		"model": "openai_p/o3",
 		"messages": []map[string]any{{"role": "user", "content": "15+23*2=?"}},
 	})
 	w := httptest.NewRecorder()
@@ -413,10 +403,10 @@ func TestE2E_OpenAI_NonStream_Refusal(t *testing.T) {
 			"usage": map[string]any{"prompt_tokens": 12, "completion_tokens": 35, "total_tokens": 47},
 		})
 	})
-	e2eCreateProviderAndModel(t, r, "openai-p", "openai", "gpt-4o", upstream.URL)
+	e2eCreateProviderAndModel(t, r, "openai_p", "openai", "gpt-4o", upstream.URL)
 
 	body, _ := json.Marshal(map[string]any{
-		"model": "gpt-4o",
+		"model": "openai_p/gpt-4o",
 		"messages": []map[string]any{{"role": "user", "content": "做坏事"}},
 	})
 	w := httptest.NewRecorder()
@@ -455,10 +445,10 @@ func TestE2E_OpenAI_Stream_Text(t *testing.T) {
 			time.Sleep(10 * time.Millisecond)
 		}
 	})
-	e2eCreateProviderAndModel(t, r, "openai-p", "openai", "gpt-4o", upstream.URL)
+	e2eCreateProviderAndModel(t, r, "openai_p", "openai", "gpt-4o", upstream.URL)
 
 	body, _ := json.Marshal(map[string]any{
-		"model": "gpt-4o",
+		"model": "openai_p/gpt-4o",
 		"messages": []map[string]any{{"role": "user", "content": "你好"}},
 		"stream": true,
 	})
@@ -499,10 +489,10 @@ func TestE2E_OpenAI_Stream_ToolCalls(t *testing.T) {
 			time.Sleep(10 * time.Millisecond)
 		}
 	})
-	e2eCreateProviderAndModel(t, r, "openai-p", "openai", "gpt-4o", upstream.URL)
+	e2eCreateProviderAndModel(t, r, "openai_p", "openai", "gpt-4o", upstream.URL)
 
 	body, _ := json.Marshal(map[string]any{
-		"model": "gpt-4o",
+		"model": "openai_p/gpt-4o",
 		"messages": []map[string]any{{"role": "user", "content": "北京天气"}},
 		"tools": []map[string]any{{
 			"type": "function",
@@ -548,10 +538,10 @@ func TestE2E_OpenAI_Stream_WithUsage(t *testing.T) {
 			time.Sleep(10 * time.Millisecond)
 		}
 	})
-	e2eCreateProviderAndModel(t, r, "openai-p", "openai", "gpt-4o", upstream.URL)
+	e2eCreateProviderAndModel(t, r, "openai_p", "openai", "gpt-4o", upstream.URL)
 
 	body, _ := json.Marshal(map[string]any{
-		"model": "gpt-4o",
+		"model": "openai_p/gpt-4o",
 		"messages": []map[string]any{{"role": "user", "content": "hi"}},
 		"stream": true,
 	})
@@ -583,10 +573,10 @@ func TestE2E_Anthropic_NonStream_BasicText(t *testing.T) {
 			"usage": map[string]any{"input_tokens": 15, "output_tokens": 25},
 		})
 	})
-	e2eCreateProviderAndModel(t, r, "anthropic-p", "anthropic", "claude-opus-4-7", upstream.URL)
+	e2eCreateProviderAndModel(t, r, "anthropic_p", "anthropic", "claude-opus-4-7", upstream.URL)
 
 	body, _ := json.Marshal(map[string]any{
-		"model": "claude-opus-4-7", "max_tokens": 1024,
+		"model": "anthropic_p/claude-opus-4-7", "max_tokens": 1024,
 		"messages": []map[string]any{{"role": "user", "content": "你好"}},
 	})
 	w := httptest.NewRecorder()
@@ -599,7 +589,7 @@ func TestE2E_Anthropic_NonStream_BasicText(t *testing.T) {
 	require.NoError(t, json.Unmarshal(w.Body.Bytes(), &resp))
 	assert.Equal(t, "message", resp["type"])
 	assert.Equal(t, "assistant", resp["role"])
-	assert.Equal(t, "claude-opus-4-7", resp["model"])
+	assert.Equal(t, "anthropic_p/claude-opus-4-7", resp["model"])
 	assert.Equal(t, "end_turn", resp["stop_reason"])
 
 	content := resp["content"].([]any)
@@ -629,10 +619,10 @@ func TestE2E_Anthropic_NonStream_WithSystem(t *testing.T) {
 			"usage": map[string]any{"input_tokens": 30, "output_tokens": 15},
 		})
 	})
-	e2eCreateProviderAndModel(t, r, "anthropic-p", "anthropic", "claude-opus-4-7", upstream.URL)
+	e2eCreateProviderAndModel(t, r, "anthropic_p", "anthropic", "claude-opus-4-7", upstream.URL)
 
 	body, _ := json.Marshal(map[string]any{
-		"model": "claude-opus-4-7", "max_tokens": 1024,
+		"model": "anthropic_p/claude-opus-4-7", "max_tokens": 1024,
 		"system": "你是编程助手",
 		"messages": []map[string]any{{"role": "user", "content": "什么是递归?"}},
 	})
@@ -658,10 +648,10 @@ func TestE2E_Anthropic_NonStream_ToolUse(t *testing.T) {
 			"usage": map[string]any{"input_tokens": 180, "output_tokens": 42},
 		})
 	})
-	e2eCreateProviderAndModel(t, r, "anthropic-p", "anthropic", "claude-opus-4-7", upstream.URL)
+	e2eCreateProviderAndModel(t, r, "anthropic_p", "anthropic", "claude-opus-4-7", upstream.URL)
 
 	body, _ := json.Marshal(map[string]any{
-		"model": "claude-opus-4-7", "max_tokens": 1024,
+		"model": "anthropic_p/claude-opus-4-7", "max_tokens": 1024,
 		"messages": []map[string]any{{"role": "user", "content": "北京天气"}},
 		"tools": []map[string]any{{
 			"name": "get_weather", "description": "获取天气",
@@ -704,10 +694,10 @@ func TestE2E_Anthropic_NonStream_Thinking(t *testing.T) {
 			"usage": map[string]any{"input_tokens": 95, "output_tokens": 280},
 		})
 	})
-	e2eCreateProviderAndModel(t, r, "anthropic-p", "anthropic", "claude-opus-4-7", upstream.URL)
+	e2eCreateProviderAndModel(t, r, "anthropic_p", "anthropic", "claude-opus-4-7", upstream.URL)
 
 	body, _ := json.Marshal(map[string]any{
-		"model": "claude-opus-4-7", "max_tokens": 4096,
+		"model": "anthropic_p/claude-opus-4-7", "max_tokens": 4096,
 		"messages": []map[string]any{{"role": "user", "content": "15+23*2=?"}},
 		"thinking": map[string]any{"type": "enabled", "budget_tokens": 2048},
 	})
@@ -736,10 +726,10 @@ func TestE2E_Anthropic_NonStream_MaxTokens(t *testing.T) {
 			"usage": map[string]any{"input_tokens": 22, "output_tokens": 20},
 		})
 	})
-	e2eCreateProviderAndModel(t, r, "anthropic-p", "anthropic", "claude-opus-4-7", upstream.URL)
+	e2eCreateProviderAndModel(t, r, "anthropic_p", "anthropic", "claude-opus-4-7", upstream.URL)
 
 	body, _ := json.Marshal(map[string]any{
-		"model": "claude-opus-4-7", "max_tokens": 20,
+		"model": "anthropic_p/claude-opus-4-7", "max_tokens": 20,
 		"messages": []map[string]any{{"role": "user", "content": "介绍AI历史"}},
 	})
 	w := httptest.NewRecorder()
@@ -764,10 +754,10 @@ func TestE2E_Anthropic_NonStream_StopSequence(t *testing.T) {
 			"usage": map[string]any{"input_tokens": 22, "output_tokens": 10},
 		})
 	})
-	e2eCreateProviderAndModel(t, r, "anthropic-p", "anthropic", "claude-opus-4-7", upstream.URL)
+	e2eCreateProviderAndModel(t, r, "anthropic_p", "anthropic", "claude-opus-4-7", upstream.URL)
 
 	body, _ := json.Marshal(map[string]any{
-		"model": "claude-opus-4-7", "max_tokens": 1024,
+		"model": "anthropic_p/claude-opus-4-7", "max_tokens": 1024,
 		"messages": []map[string]any{{"role": "user", "content": "从1数到10"}},
 		"stop_sequences": []string{"5"},
 	})
@@ -800,10 +790,10 @@ func TestE2E_Anthropic_NonStream_MetadataUserID(t *testing.T) {
 			"usage": map[string]any{"input_tokens": 12, "output_tokens": 5},
 		})
 	})
-	e2eCreateProviderAndModel(t, r, "anthropic-p", "anthropic", "claude-opus-4-7", upstream.URL)
+	e2eCreateProviderAndModel(t, r, "anthropic_p", "anthropic", "claude-opus-4-7", upstream.URL)
 
 	body, _ := json.Marshal(map[string]any{
-		"model": "claude-opus-4-7", "max_tokens": 1024,
+		"model": "anthropic_p/claude-opus-4-7", "max_tokens": 1024,
 		"messages": []map[string]any{{"role": "user", "content": "你好"}},
 		"metadata": map[string]any{"user_id": "user_12345"},
 	})
@@ -829,10 +819,10 @@ func TestE2E_Anthropic_NonStream_UsageWithCache(t *testing.T) {
 			},
 		})
 	})
-	e2eCreateProviderAndModel(t, r, "anthropic-p", "anthropic", "claude-opus-4-7", upstream.URL)
+	e2eCreateProviderAndModel(t, r, "anthropic_p", "anthropic", "claude-opus-4-7", upstream.URL)
 
 	body, _ := json.Marshal(map[string]any{
-		"model": "claude-opus-4-7", "max_tokens": 1024,
+		"model": "anthropic_p/claude-opus-4-7", "max_tokens": 1024,
 		"system": []map[string]any{{"type": "text", "text": "你是编程助手。"}},
 		"messages": []map[string]any{{"role": "user", "content": "你好"}},
 	})
@@ -874,10 +864,10 @@ func TestE2E_Anthropic_Stream_Text(t *testing.T) {
 			time.Sleep(10 * time.Millisecond)
 		}
 	})
-	e2eCreateProviderAndModel(t, r, "anthropic-p", "anthropic", "claude-opus-4-7", upstream.URL)
+	e2eCreateProviderAndModel(t, r, "anthropic_p", "anthropic", "claude-opus-4-7", upstream.URL)
 
 	body, _ := json.Marshal(map[string]any{
-		"model": "claude-opus-4-7", "max_tokens": 1024,
+		"model": "anthropic_p/claude-opus-4-7", "max_tokens": 1024,
 		"messages": []map[string]any{{"role": "user", "content": "你好"}},
 		"stream": true,
 	})
@@ -921,10 +911,10 @@ func TestE2E_Anthropic_Stream_Thinking(t *testing.T) {
 			time.Sleep(10 * time.Millisecond)
 		}
 	})
-	e2eCreateProviderAndModel(t, r, "anthropic-p", "anthropic", "claude-opus-4-7", upstream.URL)
+	e2eCreateProviderAndModel(t, r, "anthropic_p", "anthropic", "claude-opus-4-7", upstream.URL)
 
 	body, _ := json.Marshal(map[string]any{
-		"model": "claude-opus-4-7", "max_tokens": 4096,
+		"model": "anthropic_p/claude-opus-4-7", "max_tokens": 4096,
 		"messages": []map[string]any{{"role": "user", "content": "1+1=?"}},
 		"thinking": map[string]any{"type": "enabled", "budget_tokens": 1024},
 		"stream": true,
@@ -970,10 +960,10 @@ func TestE2E_CrossProtocol_OpenAIToAnthropic_RequestFormat(t *testing.T) {
 			"usage": map[string]any{"input_tokens": 10, "output_tokens": 5},
 		})
 	})
-	e2eCreateProviderAndModel(t, r, "anthropic-p", "anthropic", "claude-model", upstream.URL)
+	e2eCreateProviderAndModel(t, r, "anthropic_p", "anthropic", "claude-model", upstream.URL)
 
 	body, _ := json.Marshal(map[string]any{
-		"model": "claude-model",
+		"model": "anthropic_p/claude-model",
 		"messages": []map[string]any{{"role": "user", "content": "Hello"}},
 	})
 	w := httptest.NewRecorder()
@@ -1011,10 +1001,10 @@ func TestE2E_CrossProtocol_AnthropicToOpenAI_RequestFormat(t *testing.T) {
 			"usage": map[string]any{"prompt_tokens": 10, "completion_tokens": 8, "total_tokens": 18},
 		})
 	})
-	e2eCreateProviderAndModel(t, r, "openai-p", "openai", "gpt-4", upstream.URL)
+	e2eCreateProviderAndModel(t, r, "openai_p", "openai", "gpt-4", upstream.URL)
 
 	body, _ := json.Marshal(map[string]any{
-		"model": "gpt-4", "max_tokens": 1024,
+		"model": "openai_p/gpt-4", "max_tokens": 1024,
 		"messages": []map[string]any{{"role": "user", "content": "Hello"}},
 	})
 	w := httptest.NewRecorder()
@@ -1052,10 +1042,10 @@ func TestE2E_CrossProtocol_OpenAIToAnthropic_Stream(t *testing.T) {
 			time.Sleep(10 * time.Millisecond)
 		}
 	})
-	e2eCreateProviderAndModel(t, r, "anthropic-p", "anthropic", "claude-model", upstream.URL)
+	e2eCreateProviderAndModel(t, r, "anthropic_p", "anthropic", "claude-model", upstream.URL)
 
 	body, _ := json.Marshal(map[string]any{
-		"model": "claude-model",
+		"model": "anthropic_p/claude-model",
 		"messages": []map[string]any{{"role": "user", "content": "Hello"}},
 		"stream": true,
 	})
@@ -1092,10 +1082,10 @@ func TestE2E_CrossProtocol_AnthropicToOpenAI_Stream(t *testing.T) {
 			time.Sleep(10 * time.Millisecond)
 		}
 	})
-	e2eCreateProviderAndModel(t, r, "openai-p", "openai", "gpt-4", upstream.URL)
+	e2eCreateProviderAndModel(t, r, "openai_p", "openai", "gpt-4", upstream.URL)
 
 	body, _ := json.Marshal(map[string]any{
-		"model": "gpt-4", "max_tokens": 1024,
+		"model": "openai_p/gpt-4", "max_tokens": 1024,
 		"messages": []map[string]any{{"role": "user", "content": "Hello"}},
 		"stream": true,
 	})
@@ -1130,10 +1120,10 @@ func TestE2E_OpenAI_ErrorResponse(t *testing.T) {
 			},
 		})
 	})
-	e2eCreateProviderAndModel(t, r, "openai-p", "openai", "nonexistent", upstream.URL)
+	e2eCreateProviderAndModel(t, r, "openai_p", "openai", "nonexistent", upstream.URL)
 
 	body, _ := json.Marshal(map[string]any{
-		"model": "nonexistent",
+		"model": "openai_p/nonexistent",
 		"messages": []map[string]any{{"role": "user", "content": "test"}},
 	})
 	w := httptest.NewRecorder()
@@ -1157,10 +1147,10 @@ func TestE2E_Anthropic_ErrorResponse(t *testing.T) {
 			},
 		})
 	})
-	e2eCreateProviderAndModel(t, r, "anthropic-p", "anthropic", "claude-opus-4-7", upstream.URL)
+	e2eCreateProviderAndModel(t, r, "anthropic_p", "anthropic", "claude-opus-4-7", upstream.URL)
 
 	body, _ := json.Marshal(map[string]any{
-		"model": "claude-opus-4-7", "max_tokens": 1024,
+		"model": "anthropic_p/claude-opus-4-7", "max_tokens": 1024,
 		"messages": []map[string]any{{"role": "user", "content": "test"}},
 	})
 	w := httptest.NewRecorder()
@@ -1203,10 +1193,10 @@ func TestE2E_OpenAI_NonStream_ParallelToolCalls(t *testing.T) {
 			"usage": map[string]any{"prompt_tokens": 100, "completion_tokens": 36, "total_tokens": 136},
 		})
 	})
-	e2eCreateProviderAndModel(t, r, "openai-p", "openai", "gpt-4o", upstream.URL)
+	e2eCreateProviderAndModel(t, r, "openai_p", "openai", "gpt-4o", upstream.URL)
 
 	body, _ := json.Marshal(map[string]any{
-		"model": "gpt-4o",
+		"model": "openai_p/gpt-4o",
 		"messages": []map[string]any{{"role": "user", "content": "北京和上海的天气"}},
 		"tools": []map[string]any{{
 			"type": "function",
@@ -1255,10 +1245,10 @@ func TestE2E_OpenAI_NonStream_StopSequence(t *testing.T) {
 			"usage": map[string]any{"prompt_tokens": 10, "completion_tokens": 8, "total_tokens": 18},
 		})
 	})
-	e2eCreateProviderAndModel(t, r, "openai-p", "openai", "gpt-4o", upstream.URL)
+	e2eCreateProviderAndModel(t, r, "openai_p", "openai", "gpt-4o", upstream.URL)
 
 	body, _ := json.Marshal(map[string]any{
-		"model": "gpt-4o",
+		"model": "openai_p/gpt-4o",
 		"messages": []map[string]any{{"role": "user", "content": "从1数到10"}},
 		"stop":   []string{"5"},
 	})
@@ -1293,10 +1283,10 @@ func TestE2E_OpenAI_NonStream_ContentFilter(t *testing.T) {
 			"usage": map[string]any{"prompt_tokens": 8, "completion_tokens": 0, "total_tokens": 8},
 		})
 	})
-	e2eCreateProviderAndModel(t, r, "openai-p", "openai", "gpt-4o", upstream.URL)
+	e2eCreateProviderAndModel(t, r, "openai_p", "openai", "gpt-4o", upstream.URL)
 
 	body, _ := json.Marshal(map[string]any{
-		"model": "gpt-4o",
+		"model": "openai_p/gpt-4o",
 		"messages": []map[string]any{{"role": "user", "content": "危险内容"}},
 	})
 	w := httptest.NewRecorder()
@@ -1325,10 +1315,10 @@ func TestE2E_Anthropic_NonStream_MultiToolUse(t *testing.T) {
 			"usage": map[string]any{"input_tokens": 200, "output_tokens": 84},
 		})
 	})
-	e2eCreateProviderAndModel(t, r, "anthropic-p", "anthropic", "claude-opus-4-7", upstream.URL)
+	e2eCreateProviderAndModel(t, r, "anthropic_p", "anthropic", "claude-opus-4-7", upstream.URL)
 
 	body, _ := json.Marshal(map[string]any{
-		"model": "claude-opus-4-7", "max_tokens": 1024,
+		"model": "anthropic_p/claude-opus-4-7", "max_tokens": 1024,
 		"messages": []map[string]any{{"role": "user", "content": "北京和上海的天气"}},
 		"tools": []map[string]any{{
 			"name": "get_weather", "description": "获取天气",
@@ -1374,10 +1364,10 @@ func TestE2E_Anthropic_NonStream_ToolChoiceAny(t *testing.T) {
 			"usage": map[string]any{"input_tokens": 100, "output_tokens": 30},
 		})
 	})
-	e2eCreateProviderAndModel(t, r, "anthropic-p", "anthropic", "claude-opus-4-7", upstream.URL)
+	e2eCreateProviderAndModel(t, r, "anthropic_p", "anthropic", "claude-opus-4-7", upstream.URL)
 
 	body, _ := json.Marshal(map[string]any{
-		"model": "claude-opus-4-7", "max_tokens": 1024,
+		"model": "anthropic_p/claude-opus-4-7", "max_tokens": 1024,
 		"messages": []map[string]any{{"role": "user", "content": "现在几点了?"}},
 		"tools": []map[string]any{{
 			"name": "get_time", "description": "获取当前时间",
@@ -1417,10 +1407,10 @@ func TestE2E_Anthropic_NonStream_ArraySystemPrompt(t *testing.T) {
 			"usage": map[string]any{"input_tokens": 50, "output_tokens": 10},
 		})
 	})
-	e2eCreateProviderAndModel(t, r, "anthropic-p", "anthropic", "claude-opus-4-7", upstream.URL)
+	e2eCreateProviderAndModel(t, r, "anthropic_p", "anthropic", "claude-opus-4-7", upstream.URL)
 
 	body, _ := json.Marshal(map[string]any{
-		"model": "claude-opus-4-7", "max_tokens": 1024,
+		"model": "anthropic_p/claude-opus-4-7", "max_tokens": 1024,
 		"system": []map[string]any{
 			{"type": "text", "text": "你是编程助手。"},
 			{"type": "text", "text": "请用中文回答。"},
@@ -1454,10 +1444,10 @@ func TestE2E_Anthropic_NonStream_ToolResultMessage(t *testing.T) {
 			"usage": map[string]any{"input_tokens": 150, "output_tokens": 20},
 		})
 	})
-	e2eCreateProviderAndModel(t, r, "anthropic-p", "anthropic", "claude-opus-4-7", upstream.URL)
+	e2eCreateProviderAndModel(t, r, "anthropic_p", "anthropic", "claude-opus-4-7", upstream.URL)
 
 	body, _ := json.Marshal(map[string]any{
-		"model": "claude-opus-4-7", "max_tokens": 1024,
+		"model": "anthropic_p/claude-opus-4-7", "max_tokens": 1024,
 		"messages": []map[string]any{
 			{"role": "user", "content": "北京天气"},
 			{"role": "assistant", "content": []map[string]any{
@@ -1507,10 +1497,10 @@ func TestE2E_Anthropic_Stream_ToolCalls(t *testing.T) {
 			time.Sleep(10 * time.Millisecond)
 		}
 	})
-	e2eCreateProviderAndModel(t, r, "anthropic-p", "anthropic", "claude-opus-4-7", upstream.URL)
+	e2eCreateProviderAndModel(t, r, "anthropic_p", "anthropic", "claude-opus-4-7", upstream.URL)
 
 	body, _ := json.Marshal(map[string]any{
-		"model": "claude-opus-4-7", "max_tokens": 1024,
+		"model": "anthropic_p/claude-opus-4-7", "max_tokens": 1024,
 		"messages": []map[string]any{{"role": "user", "content": "北京天气"}},
 		"tools": []map[string]any{{
 			"name": "get_weather", "description": "获取天气",
@@ -1561,10 +1551,10 @@ func TestE2E_CrossProtocol_OpenAIToAnthropic_NonStream_ToolCalls(t *testing.T) {
 			"usage": map[string]any{"input_tokens": 100, "output_tokens": 30},
 		})
 	})
-	e2eCreateProviderAndModel(t, r, "anthropic-p", "anthropic", "claude-model", upstream.URL)
+	e2eCreateProviderAndModel(t, r, "anthropic_p", "anthropic", "claude-model", upstream.URL)
 
 	body, _ := json.Marshal(map[string]any{
-		"model": "claude-model",
+		"model": "anthropic_p/claude-model",
 		"messages": []map[string]any{{"role": "user", "content": "北京天气"}},
 		"tools": []map[string]any{{
 			"type": "function",
@@ -1613,10 +1603,10 @@ func TestE2E_CrossProtocol_AnthropicToOpenAI_NonStream_Thinking(t *testing.T) {
 			"usage": map[string]any{"prompt_tokens": 50, "completion_tokens": 100, "total_tokens": 150},
 		})
 	})
-	e2eCreateProviderAndModel(t, r, "openai-p", "openai", "gpt-4", upstream.URL)
+	e2eCreateProviderAndModel(t, r, "openai_p", "openai", "gpt-4", upstream.URL)
 
 	body, _ := json.Marshal(map[string]any{
-		"model": "gpt-4", "max_tokens": 4096,
+		"model": "openai_p/gpt-4", "max_tokens": 4096,
 		"messages": []map[string]any{{"role": "user", "content": "宇宙的答案"}},
 	})
 	w := httptest.NewRecorder()
@@ -1643,10 +1633,10 @@ func TestE2E_CrossProtocol_StopReasonMapping(t *testing.T) {
 			"usage": map[string]any{"input_tokens": 10, "output_tokens": 20},
 		})
 	})
-	e2eCreateProviderAndModel(t, r, "anthropic-p", "anthropic", "claude-model", upstream.URL)
+	e2eCreateProviderAndModel(t, r, "anthropic_p", "anthropic", "claude-model", upstream.URL)
 
 	body, _ := json.Marshal(map[string]any{
-		"model": "claude-model",
+		"model": "anthropic_p/claude-model",
 		"messages": []map[string]any{{"role": "user", "content": "长文"}},
 	})
 	w := httptest.NewRecorder()
@@ -1685,10 +1675,10 @@ func TestE2E_OpenAI_NonStream_AssistantWithToolResult(t *testing.T) {
 			"usage": map[string]any{"prompt_tokens": 100, "completion_tokens": 20, "total_tokens": 120},
 		})
 	})
-	e2eCreateProviderAndModel(t, r, "openai-p", "openai", "gpt-4o", upstream.URL)
+	e2eCreateProviderAndModel(t, r, "openai_p", "openai", "gpt-4o", upstream.URL)
 
 	body, _ := json.Marshal(map[string]any{
-		"model": "gpt-4o",
+		"model": "openai_p/gpt-4o",
 		"messages": []map[string]any{
 			{"role": "user", "content": "北京天气"},
 			{"role": "assistant", "content": nil, "tool_calls": []map[string]any{{
@@ -1732,10 +1722,10 @@ func TestE2E_CrossProtocol_AnthropicToOpenAI_Stream_ToolCalls(t *testing.T) {
 			time.Sleep(10 * time.Millisecond)
 		}
 	})
-	e2eCreateProviderAndModel(t, r, "anthropic-p", "anthropic", "claude-model", upstream.URL)
+	e2eCreateProviderAndModel(t, r, "anthropic_p", "anthropic", "claude-model", upstream.URL)
 
 	body, _ := json.Marshal(map[string]any{
-		"model": "claude-model",
+		"model": "anthropic_p/claude-model",
 		"messages": []map[string]any{{"role": "user", "content": "北京天气"}},
 		"tools": []map[string]any{{
 			"type": "function",
@@ -1781,10 +1771,10 @@ func TestE2E_CrossProtocol_OpenAIToAnthropic_Stream_ToolCalls(t *testing.T) {
 			time.Sleep(10 * time.Millisecond)
 		}
 	})
-	e2eCreateProviderAndModel(t, r, "openai-p", "openai", "gpt-4", upstream.URL)
+	e2eCreateProviderAndModel(t, r, "openai_p", "openai", "gpt-4", upstream.URL)
 
 	body, _ := json.Marshal(map[string]any{
-		"model": "gpt-4", "max_tokens": 1024,
+		"model": "openai_p/gpt-4", "max_tokens": 1024,
 		"messages": []map[string]any{{"role": "user", "content": "北京天气"}},
 		"tools": []map[string]any{{
 			"name": "get_weather", "description": "获取天气",
@@ -1819,10 +1809,10 @@ func TestE2E_OpenAI_Upstream5xx_ErrorPassthrough(t *testing.T) {
 			},
 		})
 	})
-	e2eCreateProviderAndModel(t, r, "openai-p", "openai", "gpt-4o", upstream.URL)
+	e2eCreateProviderAndModel(t, r, "openai_p", "openai", "gpt-4o", upstream.URL)
 
 	body, _ := json.Marshal(map[string]any{
-		"model": "gpt-4o",
+		"model": "openai_p/gpt-4o",
 		"messages": []map[string]any{{"role": "user", "content": "test"}},
 	})
 	w := httptest.NewRecorder()
@@ -1851,10 +1841,10 @@ func TestE2E_Anthropic_Upstream5xx_ErrorPassthrough(t *testing.T) {
 			},
 		})
 	})
-	e2eCreateProviderAndModel(t, r, "anthropic-p", "anthropic", "claude-opus-4-7", upstream.URL)
+	e2eCreateProviderAndModel(t, r, "anthropic_p", "anthropic", "claude-opus-4-7", upstream.URL)
 
 	body, _ := json.Marshal(map[string]any{
-		"model": "claude-opus-4-7", "max_tokens": 1024,
+		"model": "anthropic_p/claude-opus-4-7", "max_tokens": 1024,
 		"messages": []map[string]any{{"role": "user", "content": "test"}},
 	})
 	w := httptest.NewRecorder()
@@ -1889,10 +1879,10 @@ func TestE2E_Anthropic_Stream_TruncatedSSE(t *testing.T) {
 			time.Sleep(10 * time.Millisecond)
 		}
 	})
-	e2eCreateProviderAndModel(t, r, "anthropic-p", "anthropic", "claude-opus-4-7", upstream.URL)
+	e2eCreateProviderAndModel(t, r, "anthropic_p", "anthropic", "claude-opus-4-7", upstream.URL)
 
 	body, _ := json.Marshal(map[string]any{
-		"model": "claude-opus-4-7", "max_tokens": 1024,
+		"model": "anthropic_p/claude-opus-4-7", "max_tokens": 1024,
 		"messages": []map[string]any{{"role": "user", "content": "test"}},
 		"stream": true,
 	})