AI 网关切换 OpenAI 兼容面并移除 chat 端点,新增模型黑白名单

2026-07-12 17:48:28 +08:00
parent 7706f59549
commit 489cb49cb3
34 changed files with 2602 additions and 2901 deletions
@@ -1,332 +0,0 @@
-package service
-
-import (
-	"encoding/json"
-	"fmt"
-	"strings"
-
-	"oci-portal/internal/aiwire"
-)
-
-// ErrAiUnsupportedBlock 表示请求含网关无法承接的内容块(仅支持文本与图片)。
-var ErrAiUnsupportedBlock = fmt.Errorf("暂不支持文本与图片以外的内容块")
-
-// AnthropicToIR 把 Anthropic Messages 请求转为 IR(OpenAI 线格式)。
-// tool_result 块拆为独立 tool 角色消息(一拆多),顶层 system 变首条 system 消息。
-func AnthropicToIR(req aiwire.MessagesRequest) (aiwire.ChatRequest, error) {
-	ir := aiwire.ChatRequest{
-		Model:       req.Model,
-		Temperature: req.Temperature,
-		TopP:        req.TopP,
-		TopK:        req.TopK,
-		Stop:        req.StopSequences,
-		Stream:      req.Stream,
-		Tools:       anthTools(req.Tools),
-		ToolChoice:  anthToolChoice(req.ToolChoice),
-	}
-	if req.MaxTokens > 0 {
-		mt := req.MaxTokens
-		ir.MaxTokens = &mt
-	}
-	if sys := req.SystemText(); sys != "" {
-		ir.Messages = append(ir.Messages, aiwire.ChatMessage{Role: "system", Content: aiwire.NewTextContent(sys)})
-	}
-	for _, m := range req.Messages {
-		msgs, err := anthMessageToIR(m)
-		if err != nil {
-			return ir, err
-		}
-		ir.Messages = append(ir.Messages, msgs...)
-	}
-	return ir, nil
-}
-
-// anthMessageToIR 拆解单条 Anthropic 消息;user 消息里的 tool_result 前置为独立 tool 消息,
-// image 块转为 IR image_url 部件。
-func anthMessageToIR(m aiwire.AnthMessage) ([]aiwire.ChatMessage, error) {
-	var out []aiwire.ChatMessage
-	var parts []aiwire.ContentPart
-	var toolCalls []aiwire.ToolCall
-	hasImage := false
-	for _, b := range m.Content.AllBlocks() {
-		switch b.Type {
-		case "text":
-			parts = append(parts, aiwire.ContentPart{Type: "text", Text: b.Text})
-		case "image":
-			p, err := anthImagePart(b.Source)
-			if err != nil {
-				return nil, err
-			}
-			parts, hasImage = append(parts, p), true
-		case "tool_use":
-			toolCalls = append(toolCalls, aiwire.ToolCall{ID: b.ID, Type: "function",
-				Function: aiwire.FunctionCall{Name: b.Name, Arguments: string(b.Input)}})
-		case "tool_result":
-			out = append(out, aiwire.ChatMessage{Role: "tool", ToolCallID: b.ToolUseID,
-				Content: aiwire.NewTextContent(b.ResultText())})
-		case "thinking", "redacted_thinking":
-			// 一期忽略 thinking 块
-		default:
-			return nil, ErrAiUnsupportedBlock
-		}
-	}
-	content := partsContent(parts, hasImage)
-	if hasImage || content.JoinText() != "" || len(toolCalls) > 0 {
-		out = append(out, aiwire.ChatMessage{Role: m.Role, Content: content, ToolCalls: toolCalls})
-	}
-	return out, nil
-}
-
-// anthImagePart 把 Anthropic image 块转为 IR image_url 部件(base64 → data URI)。
-func anthImagePart(source json.RawMessage) (aiwire.ContentPart, error) {
-	var src struct {
-		Type      string `json:"type"`
-		MediaType string `json:"media_type"`
-		Data      string `json:"data"`
-		URL       string `json:"url"`
-	}
-	if err := json.Unmarshal(source, &src); err != nil {
-		return aiwire.ContentPart{}, fmt.Errorf("image source 解析失败: %w", err)
-	}
-	switch src.Type {
-	case "base64":
-		if src.MediaType == "" || src.Data == "" {
-			return aiwire.ContentPart{}, fmt.Errorf("image source 缺少 media_type 或 data")
-		}
-		url := "data:" + src.MediaType + ";base64," + src.Data
-		return aiwire.ContentPart{Type: "image_url", ImageURL: &aiwire.ImageURL{URL: url}}, nil
-	case "url":
-		if src.URL == "" {
-			return aiwire.ContentPart{}, fmt.Errorf("image source 缺少 url")
-		}
-		return aiwire.ContentPart{Type: "image_url", ImageURL: &aiwire.ImageURL{URL: src.URL}}, nil
-	}
-	return aiwire.ContentPart{}, fmt.Errorf("不支持的 image source 类型 %q", src.Type)
-}
-
-// partsContent 无图片时退回字符串形态(与上游线格式习惯一致),含图片时保留块数组。
-func partsContent(parts []aiwire.ContentPart, hasImage bool) aiwire.Content {
-	if !hasImage {
-		var sb strings.Builder
-		for _, p := range parts {
-			sb.WriteString(p.Text)
-		}
-		return aiwire.NewTextContent(sb.String())
-	}
-	return aiwire.NewPartsContent(parts)
-}
-
-func anthTools(tools []aiwire.AnthTool) []aiwire.Tool {
-	if len(tools) == 0 {
-		return nil
-	}
-	out := make([]aiwire.Tool, 0, len(tools))
-	for _, t := range tools {
-		out = append(out, aiwire.Tool{Type: "function", Function: aiwire.FunctionDef{
-			Name: t.Name, Description: t.Description, Parameters: t.InputSchema}})
-	}
-	return out
-}
-
-// anthToolChoice 映射 {type:auto|any|tool,name} → OpenAI 形态。
-func anthToolChoice(raw json.RawMessage) json.RawMessage {
-	if len(raw) == 0 {
-		return nil
-	}
-	var tc struct {
-		Type string `json:"type"`
-		Name string `json:"name"`
-	}
-	if json.Unmarshal(raw, &tc) != nil {
-		return nil
-	}
-	switch tc.Type {
-	case "auto":
-		return json.RawMessage(`"auto"`)
-	case "any":
-		return json.RawMessage(`"required"`)
-	case "tool":
-		b, _ := json.Marshal(map[string]any{"type": "function", "function": map[string]string{"name": tc.Name}})
-		return b
-	case "none":
-		return json.RawMessage(`"none"`)
-	}
-	return nil
-}
-
-// IRRespToAnthropic 把 IR 非流式响应转为 Anthropic Messages 响应。
-func IRRespToAnthropic(resp *aiwire.ChatResponse, id string) aiwire.MessagesResponse {
-	out := aiwire.MessagesResponse{ID: id, Type: "message", Role: "assistant", Model: resp.Model, Content: []aiwire.AnthBlock{}}
-	if len(resp.Choices) > 0 {
-		choice := resp.Choices[0]
-		if text := choice.Message.Content.JoinText(); text != "" {
-			out.Content = append(out.Content, aiwire.AnthBlock{Type: "text", Text: text})
-		}
-		for _, tc := range choice.Message.ToolCalls {
-			out.Content = append(out.Content, aiwire.AnthBlock{Type: "tool_use", ID: tc.ID,
-				Name: tc.Function.Name, Input: argsToJSON(tc.Function.Arguments)})
-		}
-		out.StopReason = anthStopReason(choice.FinishReason)
-	}
-	if resp.Usage != nil {
-		out.Usage = aiwire.AnthUsage{InputTokens: resp.Usage.PromptTokens, OutputTokens: resp.Usage.CompletionTokens,
-			CacheReadInputTokens: resp.Usage.CachedTokens()}
-	}
-	return out
-}
-
-// argsToJSON 保证 tool_use.input 是合法 JSON 对象(模型可能产出非法片段)。
-func argsToJSON(args string) json.RawMessage {
-	trimmed := strings.TrimSpace(args)
-	if trimmed == "" {
-		return json.RawMessage(`{}`)
-	}
-	if json.Valid([]byte(trimmed)) {
-		return json.RawMessage(trimmed)
-	}
-	b, _ := json.Marshal(map[string]string{"_raw": args})
-	return b
-}
-
-func anthStopReason(finish string) string {
-	switch finish {
-	case "length":
-		return "max_tokens"
-	case "tool_calls":
-		return "tool_use"
-	default:
-		return "end_turn"
-	}
-}
-
-// ---- Anthropic 流式状态机 ----
-
-// AnthEvent 是一条待写出的 Anthropic SSE 事件。
-type AnthEvent struct {
-	Event string
-	Data  any
-}
-
-// AnthStream 把 IR chunk 流聚合为 Anthropic 事件序列:
-// message_start → content_block_start/delta/stop(text 与 tool_use 分块)→ message_delta → message_stop。
-type AnthStream struct {
-	id, model   string
-	started     bool
-	blockOpen   bool
-	blockIsTool bool
-	toolID      string
-	blockIndex  int
-	stopReason  string
-	usage       aiwire.AnthUsage
-}
-
-// NewAnthStream 构造状态机;id 为响应消息 ID。
-func NewAnthStream(id, model string) *AnthStream {
-	return &AnthStream{id: id, model: model, blockIndex: -1, stopReason: "end_turn"}
-}
-
-// Feed 消费一个 IR chunk,返回应立即写出的事件。
-func (st *AnthStream) Feed(chunk aiwire.ChatChunk) []AnthEvent {
-	var events []AnthEvent
-	if !st.started {
-		st.started = true
-		events = append(events, st.startEvent())
-	}
-	if chunk.Usage != nil {
-		st.usage = aiwire.AnthUsage{InputTokens: chunk.Usage.PromptTokens, OutputTokens: chunk.Usage.CompletionTokens,
-			CacheReadInputTokens: chunk.Usage.CachedTokens()}
-	}
-	for _, choice := range chunk.Choices {
-		events = append(events, st.feedDelta(choice.Delta)...)
-		if choice.FinishReason != nil && *choice.FinishReason != "" {
-			st.stopReason = anthStopReason(*choice.FinishReason)
-		}
-	}
-	return events
-}
-
-func (st *AnthStream) startEvent() AnthEvent {
-	return AnthEvent{Event: "message_start", Data: map[string]any{
-		"type": "message_start",
-		"message": map[string]any{
-			"id": st.id, "type": "message", "role": "assistant", "model": st.model,
-			"content": []any{}, "stop_reason": nil,
-			"usage": map[string]int{"input_tokens": 0, "output_tokens": 0},
-		},
-	}}
-}
-
-// feedDelta 处理文本与工具调用增量,必要时切块。
-func (st *AnthStream) feedDelta(d aiwire.Delta) []AnthEvent {
-	var events []AnthEvent
-	if d.Content != "" {
-		if !st.blockOpen || st.blockIsTool {
-			events = append(events, st.openBlock(false, "", "")...)
-		}
-		events = append(events, AnthEvent{Event: "content_block_delta", Data: map[string]any{
-			"type": "content_block_delta", "index": st.blockIndex,
-			"delta": map[string]string{"type": "text_delta", "text": d.Content},
-		}})
-	}
-	for _, tc := range d.ToolCalls {
-		if tc.ID != "" && (!st.blockOpen || !st.blockIsTool || st.toolID != tc.ID) {
-			events = append(events, st.openBlock(true, tc.ID, tc.Function.Name)...)
-		}
-		if tc.Function.Arguments != "" && st.blockOpen && st.blockIsTool {
-			events = append(events, AnthEvent{Event: "content_block_delta", Data: map[string]any{
-				"type": "content_block_delta", "index": st.blockIndex,
-				"delta": map[string]string{"type": "input_json_delta", "partial_json": tc.Function.Arguments},
-			}})
-		}
-	}
-	return events
-}
-
-// openBlock 关闭当前块并打开新块(text 或 tool_use)。
-func (st *AnthStream) openBlock(isTool bool, toolID, toolName string) []AnthEvent {
-	var events []AnthEvent
-	if st.blockOpen {
-		events = append(events, st.closeBlockEvent())
-	}
-	st.blockOpen, st.blockIsTool, st.toolID = true, isTool, toolID
-	st.blockIndex++
-	block := map[string]any{"type": "text", "text": ""}
-	if isTool {
-		block = map[string]any{"type": "tool_use", "id": toolID, "name": toolName, "input": map[string]any{}}
-	}
-	events = append(events, AnthEvent{Event: "content_block_start", Data: map[string]any{
-		"type": "content_block_start", "index": st.blockIndex, "content_block": block,
-	}})
-	return events
-}
-
-func (st *AnthStream) closeBlockEvent() AnthEvent {
-	return AnthEvent{Event: "content_block_stop", Data: map[string]any{
-		"type": "content_block_stop", "index": st.blockIndex,
-	}}
-}
-
-// Finish 在上游流结束后收尾:关块 → message_delta(stop_reason+usage)→ message_stop。
-func (st *AnthStream) Finish() []AnthEvent {
-	var events []AnthEvent
-	if !st.started {
-		events = append(events, st.startEvent())
-		st.started = true
-	}
-	if st.blockOpen {
-		events = append(events, st.closeBlockEvent())
-		st.blockOpen = false
-	}
-	events = append(events,
-		AnthEvent{Event: "message_delta", Data: map[string]any{
-			"type":  "message_delta",
-			"delta": map[string]any{"stop_reason": st.stopReason, "stop_sequence": nil},
-			"usage": map[string]int{"output_tokens": st.usage.OutputTokens},
-		}},
-		AnthEvent{Event: "message_stop", Data: map[string]any{"type": "message_stop"}},
-	)
-	return events
-}
-
-// Usage 返回聚合到的用量(供调用日志)。
-func (st *AnthStream) Usage() aiwire.AnthUsage { return st.usage }
@@ -1,136 +0,0 @@
-package service
-
-import (
-	"encoding/json"
-	"strings"
-	"testing"
-
-	"oci-portal/internal/aiwire"
-)
-
-func mustAnthReq(t *testing.T, body string) aiwire.MessagesRequest {
-	t.Helper()
-	var req aiwire.MessagesRequest
-	if err := json.Unmarshal([]byte(body), &req); err != nil {
-		t.Fatalf("unmarshal anthropic request: %v", err)
-	}
-	return req
-}
-
-func TestAnthropicToIR(t *testing.T) {
-	req := mustAnthReq(t, `{
-		"model": "meta.llama-3.3-70b-instruct", "max_tokens": 128, "system": "你是助手",
-		"messages": [
-			{"role": "user", "content": "东京天气?"},
-			{"role": "assistant", "content": [
-				{"type": "text", "text": "查询中"},
-				{"type": "tool_use", "id": "t1", "name": "get_weather", "input": {"city": "东京"}}
-			]},
-			{"role": "user", "content": [
-				{"type": "tool_result", "tool_use_id": "t1", "content": "晴 25 度"},
-				{"type": "text", "text": "继续"}
-			]}
-		],
-		"tools": [{"name": "get_weather", "input_schema": {"type": "object"}}],
-		"tool_choice": {"type": "any"}
-	}`)
-	ir, err := AnthropicToIR(req)
-	if err != nil {
-		t.Fatalf("AnthropicToIR: %v", err)
-	}
-	roles := make([]string, 0, len(ir.Messages))
-	for _, m := range ir.Messages {
-		roles = append(roles, m.Role)
-	}
-	want := []string{"system", "user", "assistant", "tool", "user"}
-	if strings.Join(roles, ",") != strings.Join(want, ",") {
-		t.Fatalf("roles = %v, want %v", roles, want)
-	}
-	if ir.Messages[2].ToolCalls[0].Function.Name != "get_weather" {
-		t.Errorf("tool_use 未映射: %+v", ir.Messages[2])
-	}
-	if ir.Messages[3].ToolCallID != "t1" || ir.Messages[3].Content.JoinText() != "晴 25 度" {
-		t.Errorf("tool_result 未拆为 tool 消息: %+v", ir.Messages[3])
-	}
-	if ir.MaxTokens == nil || *ir.MaxTokens != 128 || len(ir.Tools) != 1 {
-		t.Errorf("max_tokens/tools 未直通")
-	}
-	if string(ir.ToolChoice) != `"required"` {
-		t.Errorf("tool_choice any → %s, want required", ir.ToolChoice)
-	}
-	// image 块拒绝
-	bad := mustAnthReq(t, `{"model":"m","max_tokens":1,"messages":[{"role":"user","content":[{"type":"image","source":{}}]}]}`)
-	if _, err := AnthropicToIR(bad); err == nil {
-		t.Error("image 块应被拒绝")
-	}
-}
-
-func TestIRRespToAnthropic(t *testing.T) {
-	resp := &aiwire.ChatResponse{
-		Model: "m1",
-		Choices: []aiwire.Choice{{
-			Message: aiwire.ChatMessage{
-				Role:    "assistant",
-				Content: aiwire.NewTextContent("查到了"),
-				ToolCalls: []aiwire.ToolCall{{ID: "t1", Type: "function",
-					Function: aiwire.FunctionCall{Name: "get_weather", Arguments: `{"city":"东京"}`}}},
-			},
-			FinishReason: "tool_calls",
-		}},
-		Usage: &aiwire.Usage{PromptTokens: 10, CompletionTokens: 5},
-	}
-	out := IRRespToAnthropic(resp, "msg_1")
-	if len(out.Content) != 2 || out.Content[0].Type != "text" || out.Content[1].Type != "tool_use" {
-		t.Fatalf("content 块 = %+v", out.Content)
-	}
-	if out.StopReason != "tool_use" || out.Usage.InputTokens != 10 || out.Usage.OutputTokens != 5 {
-		t.Errorf("stop_reason/usage 未映射: %+v", out)
-	}
-	if string(out.Content[1].Input) != `{"city":"东京"}` {
-		t.Errorf("tool input = %s", out.Content[1].Input)
-	}
-}
-
-// eventTypes 提取事件类型序列便于断言。
-func eventTypes(events []AnthEvent) string {
-	types := make([]string, 0, len(events))
-	for _, e := range events {
-		types = append(types, e.Event)
-	}
-	return strings.Join(types, ",")
-}
-
-func TestAnthStreamTextAndTool(t *testing.T) {
-	st := NewAnthStream("msg_1", "m1")
-	fr := "tool_calls"
-	var events []AnthEvent
-	events = append(events, st.Feed(aiwire.ChatChunk{Choices: []aiwire.ChunkChoice{{Delta: aiwire.Delta{Content: "你"}}}})...)
-	events = append(events, st.Feed(aiwire.ChatChunk{Choices: []aiwire.ChunkChoice{{Delta: aiwire.Delta{Content: "好"}}}})...)
-	events = append(events, st.Feed(aiwire.ChatChunk{Choices: []aiwire.ChunkChoice{{Delta: aiwire.Delta{
-		ToolCalls: []aiwire.ToolCallDelta{{ID: "t1", Function: aiwire.FunctionCallDelta{Name: "get_weather", Arguments: `{"ci`}}},
-	}}}})...)
-	events = append(events, st.Feed(aiwire.ChatChunk{
-		Choices: []aiwire.ChunkChoice{{Delta: aiwire.Delta{ToolCalls: []aiwire.ToolCallDelta{{ID: "t1", Function: aiwire.FunctionCallDelta{Arguments: `ty":"东京"}`}}}}, FinishReason: &fr}},
-		Usage:   &aiwire.Usage{PromptTokens: 8, CompletionTokens: 4},
-	})...)
-	events = append(events, st.Finish()...)
-
-	got := eventTypes(events)
-	want := "message_start,content_block_start,content_block_delta,content_block_delta," +
-		"content_block_stop,content_block_start,content_block_delta,content_block_delta," +
-		"content_block_stop,message_delta,message_stop"
-	if got != want {
-		t.Fatalf("事件序列:\n got %s\nwant %s", got, want)
-	}
-	if st.Usage().OutputTokens != 4 {
-		t.Errorf("usage 未聚合: %+v", st.Usage())
-	}
-}
-
-func TestAnthStreamEmpty(t *testing.T) {
-	st := NewAnthStream("msg_1", "m1")
-	events := st.Finish()
-	if got := eventTypes(events); got != "message_start,message_delta,message_stop" {
-		t.Errorf("空流事件序列 = %s", got)
-	}
-}
@@ -5,6 +5,7 @@ import (
 	"crypto/rand"
 	"crypto/sha256"
 	"encoding/hex"
+	"encoding/json"
 	"errors"
 	"fmt"
 	"log"
@@ -30,10 +31,6 @@ const (
 	aiBackoffCap    = 30 * time.Minute
 	// aiKeyTouchGap 是 LastUsedAt 的最小写库间隔,避免高频调用刷库
 	aiKeyTouchGap = time.Minute
-	// modelRecheckGap 是已标记不可用模型的复检间隔(恢复供给自动解除标记);
-	// validateBatchCap 限制单渠道单轮验证的试调次数
-	modelRecheckGap  = 20 * time.Hour
-	validateBatchCap = 32
 	// 内容日志(红线例外)约束:开启必须限时(上限 7 天),正文截断,短保留
 	aiContentLogMaxHours  = 168
 	aiContentLogRetention = 7 * 24 * time.Hour
@@ -59,17 +56,13 @@ type AiGatewayService struct {
 	// touchMu 保护各密钥的最近触达时间(内存节流,不追求跨实例精确)
 	touchMu   sync.Mutex
 	lastTouch map[uint]time.Time
-	// validateMu 保护 validating:同一渠道的模型验证不并发
-	validateMu sync.Mutex
-	validating map[uint]bool
 	// onChannelsChanged 在渠道增删后触发,由 main 装配为探测任务同步钩子
 	onChannelsChanged func(context.Context)
 }

 // NewAiGatewayService 组装依赖;调用 StartCleanup 后开始调用日志周期清理。
 func NewAiGatewayService(db *gorm.DB, configs *OciConfigService, client oci.Client) *AiGatewayService {
-	return &AiGatewayService{db: db, configs: configs, client: client,
-		lastTouch: map[uint]time.Time{}, validating: map[uint]bool{}}
+	return &AiGatewayService{db: db, configs: configs, client: client, lastTouch: map[uint]time.Time{}}
 }

 // SetOnChannelsChanged 注册渠道数量变化钩子(渠道创建/删除成功后调用)。
@@ -87,7 +80,7 @@ func (s *AiGatewayService) fireChannelsChanged(ctx context.Context) {

 // CreateKey 生成网关密钥并返回明文(仅此一次);customValue 非空时使用给定值,
 // group 非空时该密钥只在同分组渠道内路由。
-func (s *AiGatewayService) CreateKey(ctx context.Context, name, customValue, group string) (string, *model.AiKey, error) {
+func (s *AiGatewayService) CreateKey(ctx context.Context, name, customValue, group string, models []string) (string, *model.AiKey, error) {
 	name = strings.TrimSpace(name)
 	if name == "" {
 		return "", nil, fmt.Errorf("密钥名称不能为空")
@@ -103,13 +96,28 @@ func (s *AiGatewayService) CreateKey(ctx context.Context, name, customValue, gro
 	if len(raw) < 8 {
 		return "", nil, fmt.Errorf("自定义密钥至少 8 个字符")
 	}
-	key := &model.AiKey{Name: name, KeyHash: hashKey(raw), Tail: raw[len(raw)-4:], Group: strings.TrimSpace(group), Enabled: true}
+	key := &model.AiKey{Name: name, KeyHash: hashKey(raw), Tail: raw[len(raw)-4:], Group: strings.TrimSpace(group), Models: normalizeKeyModels(models), Enabled: true}
 	if err := s.db.WithContext(ctx).Create(key).Error; err != nil {
 		return "", nil, fmt.Errorf("密钥名称或取值与现有密钥重复")
 	}
 	return raw, key, nil
 }

+// normalizeKeyModels 规范化模型白名单:trim、剔空串、保序去重;结果为空返回 nil(= 不限)。
+func normalizeKeyModels(models []string) []string {
+	var out []string
+	seen := map[string]bool{}
+	for _, m := range models {
+		m = strings.TrimSpace(m)
+		if m == "" || seen[m] {
+			continue
+		}
+		seen[m] = true
+		out = append(out, m)
+	}
+	return out
+}
+
 func hashKey(raw string) string {
 	sum := sha256.Sum256([]byte(raw))
 	return hex.EncodeToString(sum[:])
@@ -122,8 +130,8 @@ func (s *AiGatewayService) Keys(ctx context.Context) ([]model.AiKey, error) {
 	return keys, err
 }

-// UpdateKey 修改密钥名称 / 启用状态 / 分组(group 指针非空即覆盖,可置空)。
-func (s *AiGatewayService) UpdateKey(ctx context.Context, id uint, name string, enabled *bool, group *string) error {
+// UpdateKey 修改密钥名称 / 启用状态 / 分组 / 模型白名单(指针非空即覆盖,可置空)。
+func (s *AiGatewayService) UpdateKey(ctx context.Context, id uint, name string, enabled *bool, group *string, models *[]string) error {
 	updates := map[string]any{}
 	if name = strings.TrimSpace(name); name != "" {
 		updates["name"] = name
@@ -134,6 +142,14 @@ func (s *AiGatewayService) UpdateKey(ctx context.Context, id uint, name string,
 	if group != nil {
 		updates["key_group"] = strings.TrimSpace(*group)
 	}
+	if models != nil {
+		// 手动序列化走 map 更新,不依赖 GORM map 路径对 serializer 的支持
+		b, err := json.Marshal(normalizeKeyModels(*models))
+		if err != nil {
+			return fmt.Errorf("serialize models: %w", err)
+		}
+		updates["models"] = string(b)
+	}
 	if len(updates) == 0 {
 		return nil
 	}
@@ -298,12 +314,17 @@ func (s *AiGatewayService) ProbeChannel(ctx context.Context, id uint) (*model.Ai
 	return &fresh, nil
 }

-// probe 执行探测并返回 (状态, 错误摘要);同时完成模型缓存同步(保留不可用标记)。
+// probe 执行探测并返回 (状态, 错误摘要);同时完成模型缓存同步(黑名单模型不入库)。
 func (s *AiGatewayService) probe(ctx context.Context, cred oci.Credentials, ch *model.AiChannel) (string, string) {
 	models, err := s.client.ListGenAiModels(ctx, cred, ch.Region)
 	if err != nil {
 		return classifyProbeErr(err), truncateErr(oci.CompactError(err))
 	}
+	models = supportedGatewayModels(models)
+	models, err = s.withoutBlacklisted(ctx, models)
+	if err != nil {
+		return "error", truncateErr(err.Error())
+	}
 	if len(models) == 0 {
 		_ = s.replaceModels(ctx, ch.ID, nil)
 		return "no_service", "区域无可用模型(GenAI 服务不可用或未开放)"
@@ -311,31 +332,12 @@ func (s *AiGatewayService) probe(ctx context.Context, cred oci.Credentials, ch *
 	if err := s.replaceModels(ctx, ch.ID, models); err != nil {
 		return "error", truncateErr(err.Error())
 	}
-	usable, err := s.usableModels(ctx, ch.ID, models)
-	if err != nil {
-		return "error", truncateErr(err.Error())
-	}
-	return s.probeChat(ctx, cred, ch, usable)
+	return s.probeChat(ctx, cred, ch, models)
 }

-// usableModels 过滤掉缓存中已标记不可按需调用的模型,探测候选不再反复踩坑。
-func (s *AiGatewayService) usableModels(ctx context.Context, channelID uint, models []oci.GenAiModel) ([]oci.GenAiModel, error) {
-	marks, err := s.loadModelMarks(ctx, channelID)
-	if err != nil {
-		return nil, err
-	}
-	out := make([]oci.GenAiModel, 0, len(models))
-	for _, m := range models {
-		if !marks[m.Ocid].Unusable {
-			out = append(out, m)
-		}
-	}
-	return out, nil
-}
-
-// probeChat 按候选顺序试调(上限 8):遇「模型不可按需调用」标记剔除并换下一个;
-// 401/403 与鉴权类 404 属租户级直接定论 no_quota;其余错误(部分模型元数据标 CHAT
-// 但实际不可对话,如 voice agent)累计 3 次止损。
+// probeChat 按候选顺序试调(上限 8):遇「模型不可按需调用」(微调基座 400 / 实体
+// 不存在 404)换下一个候选,错误信息带模型名供用户加入黑名单;401/403 与鉴权类 404
+// 属租户级直接定论 no_quota;其余错误(元数据标 CHAT 但实际不可对话等)累计 3 次止损。
 func (s *AiGatewayService) probeChat(ctx context.Context, cred oci.Credentials, ch *model.AiChannel, models []oci.GenAiModel) (string, string) {
 	status, detail := "error", "无可试调对话模型"
 	errBudget := 3
@@ -345,8 +347,7 @@ func (s *AiGatewayService) probeChat(ctx context.Context, cred oci.Credentials,
 		case code == 200 || code == 429:
 			return "ok", ""
 		case oci.IsModelUnavailable(err):
-			s.markModelUnusable(ctx, ch.ID, m.Ocid, oci.CompactError(err))
-			status, detail = "error", truncateErr(fmt.Sprintf("%s: 不可按需调用,已从模型池剔除", m.Name))
+			status, detail = "error", truncateErr(fmt.Sprintf("%s: 不可按需调用,建议加入模型黑名单", m.Name))
 		case code == 401 || code == 403 || code == 404:
 			return "no_quota", truncateErr(oci.CompactError(err))
 		default:
@@ -359,106 +360,8 @@ func (s *AiGatewayService) probeChat(ctx context.Context, cred oci.Credentials,
 	return status, detail
 }

-// markModelUnusable 把 (渠道, 模型OCID) 标记为不可按需调用;失败仅记日志不阻断主流程。
-func (s *AiGatewayService) markModelUnusable(ctx context.Context, channelID uint, ocid, reason string) {
-	err := s.db.WithContext(ctx).Model(&model.AiModelCache{}).
-		Where("channel_id = ? AND model_ocid = ?", channelID, ocid).
-		Updates(map[string]any{"unusable": true, "unusable_reason": shortReason(reason), "checked_at": time.Now()}).Error
-	if err != nil {
-		log.Printf("mark model unusable: %v", err)
-	}
-}
-
-func shortReason(reason string) string {
-	if len(reason) > 200 {
-		return reason[:200]
-	}
-	return reason
-}
-
-// beginValidate 抢占渠道的验证执行权,同渠道同时只跑一轮。
-func (s *AiGatewayService) beginValidate(id uint) bool {
-	s.validateMu.Lock()
-	defer s.validateMu.Unlock()
-	if s.validating[id] {
-		return false
-	}
-	s.validating[id] = true
-	return true
-}
-
-func (s *AiGatewayService) endValidate(id uint) {
-	s.validateMu.Lock()
-	delete(s.validating, id)
-	s.validateMu.Unlock()
-}
-
-// validateTargets 取待验证行:从未验证的新模型,以及标记超过复检间隔的模型(仅对话能力,
-// 兼容存量空串);单轮上限 validateBatchCap 控制调用量。
-func (s *AiGatewayService) validateTargets(ctx context.Context, channelID uint) ([]model.AiModelCache, error) {
-	stale := time.Now().Add(-modelRecheckGap)
-	var rows []model.AiModelCache
-	err := s.db.WithContext(ctx).
-		Where("channel_id = ? AND capability IN ?", channelID, []string{"CHAT", ""}).
-		Where("checked_at IS NULL OR (unusable = ? AND checked_at < ?)", true, stale).
-		Order("id ASC").Limit(validateBatchCap).Find(&rows).Error
-	return rows, err
-}
-
-// validateChannelModels 逐个试调渠道内待验证模型并落结论,把不可按需调用的模型
-// 从池中剔除、把恢复供给的解除标记;ListModels 无字段可事先判别,只能试调习得。
-func (s *AiGatewayService) validateChannelModels(ctx context.Context, channelID uint) {
-	if !s.beginValidate(channelID) {
-		return
-	}
-	defer s.endValidate(channelID)
-	var ch model.AiChannel
-	if err := s.db.WithContext(ctx).First(&ch, channelID).Error; err != nil {
-		return
-	}
-	cred, err := s.configs.credentialsByID(ctx, ch.OciConfigID)
-	if err != nil {
-		return
-	}
-	rows, err := s.validateTargets(ctx, channelID)
-	if err != nil {
-		log.Printf("validate channel %d models: %v", channelID, err)
-		return
-	}
-	for _, r := range rows {
-		s.validateOne(ctx, cred, &ch, r)
-	}
-}
-
-// validateOne 试调单个模型并落库:可用(200/429)清除标记;模型级不可用标记剔除;
-// 其余 4xx 记录已检不改状态;5xx/网络类瞬态不落 checked_at,下轮再验。
-func (s *AiGatewayService) validateOne(ctx context.Context, cred oci.Credentials, ch *model.AiChannel, r model.AiModelCache) {
-	code, err := s.client.GenAiProbeChat(ctx, cred, ch.Region, r.ModelOcid, r.Name)
-	updates := map[string]any{"checked_at": time.Now()}
-	switch {
-	case code == 200 || code == 429:
-		updates["unusable"], updates["unusable_reason"] = false, ""
-	case oci.IsModelUnavailable(err):
-		updates["unusable"], updates["unusable_reason"] = true, shortReason(oci.CompactError(err))
-	case code == 0 || code >= 500:
-		return
-	}
-	s.db.WithContext(ctx).Model(&model.AiModelCache{}).Where("id = ?", r.ID).Updates(updates)
-}
-
-// validateModelsAsync 后台验证渠道模型池(手动同步后触发,不阻塞接口响应)。
-func (s *AiGatewayService) validateModelsAsync(ctx context.Context, channelID uint) {
-	s.wg.Add(1)
-	go func() {
-		defer s.wg.Done()
-		vctx, cancel := context.WithTimeout(context.WithoutCancel(ctx), 3*time.Minute)
-		defer cancel()
-		s.validateChannelModels(vctx, channelID)
-	}()
-}
-
-// probeCandidateCap 是单次探测的候选上限:模型级不可用会被标记剔除不再重试,
-// 放宽到 8 让一次探测有机会越过整批坏模型找到可用者;其他错误另有 3 次止损预算。
+// probeCandidateCap 是单次探测的候选上限:放宽到 8 让一次探测有机会越过整批
+// 不可按需调用的坏模型找到可用者;其他错误另有 3 次止损预算。
 const probeCandidateCap = 8

 // probeCandidates 只取对话模型,按可靠度排序后跨厂商取候选:主流文本模型优先;
@@ -551,8 +454,7 @@ func truncateErr(msg string) string {
 	return msg
 }

-// SyncModels 重新拉取渠道区域的模型列表并覆盖缓存(标记按 OCID 结转),
-// 随后触发后台验证:新模型逐个试调,不可按需调用的数十秒内从池中剔除。
+// SyncModels 重新拉取渠道区域的模型列表并覆盖缓存,黑名单中的模型不入库。
 func (s *AiGatewayService) SyncModels(ctx context.Context, id uint) ([]model.AiModelCache, error) {
 	var ch model.AiChannel
 	if err := s.db.WithContext(ctx).First(&ch, id).Error; err != nil {
@@ -566,29 +468,23 @@ func (s *AiGatewayService) SyncModels(ctx context.Context, id uint) ([]model.AiM
 	if err != nil {
 		return nil, fmt.Errorf("同步模型失败:%s", oci.CompactError(err))
 	}
+	models = supportedGatewayModels(models)
+	if models, err = s.withoutBlacklisted(ctx, models); err != nil {
+		return nil, err
+	}
 	if err := s.replaceModels(ctx, id, models); err != nil {
 		return nil, err
 	}
-	s.validateModelsAsync(ctx, id)
 	return s.channelModels(ctx, id)
 }

-// replaceModels 以事务整组覆盖渠道模型缓存,按 OCID 结转不可用标记与验证时间
-// (OCID 变化视为新条目,自然回到待验证状态)。
+// replaceModels 以事务整组覆盖渠道模型缓存。
 func (s *AiGatewayService) replaceModels(ctx context.Context, channelID uint, models []oci.GenAiModel) error {
-	marks, err := s.loadModelMarks(ctx, channelID)
-	if err != nil {
-		return err
-	}
 	rows := make([]model.AiModelCache, 0, len(models))
 	now := time.Now()
 	for _, m := range models {
-		row := model.AiModelCache{ChannelID: channelID, ModelOcid: m.Ocid, Name: m.Name, Vendor: m.Vendor,
-			Capability: m.Capability, SyncedAt: now, DeprecatedAt: m.Deprecated, RetiredAt: m.Retired}
-		if prev, ok := marks[m.Ocid]; ok {
-			row.Unusable, row.UnusableReason, row.CheckedAt = prev.Unusable, prev.UnusableReason, prev.CheckedAt
-		}
-		rows = append(rows, row)
+		rows = append(rows, model.AiModelCache{ChannelID: channelID, ModelOcid: m.Ocid, Name: m.Name, Vendor: m.Vendor,
+			Capability: m.Capability, SyncedAt: now, DeprecatedAt: m.Deprecated, RetiredAt: m.Retired})
 	}
 	return s.db.WithContext(ctx).Transaction(func(tx *gorm.DB) error {
 		if err := tx.Where("channel_id = ?", channelID).Delete(&model.AiModelCache{}).Error; err != nil {
@@ -601,33 +497,17 @@ func (s *AiGatewayService) replaceModels(ctx context.Context, channelID uint, mo
 	})
 }

-// loadModelMarks 取渠道内带标记或已验证过的行(OCID → 行),供同步结转与候选过滤。
-func (s *AiGatewayService) loadModelMarks(ctx context.Context, channelID uint) (map[string]model.AiModelCache, error) {
-	var rows []model.AiModelCache
-	err := s.db.WithContext(ctx).Select("model_ocid", "unusable", "unusable_reason", "checked_at").
-		Where("channel_id = ? AND (unusable = ? OR checked_at IS NOT NULL)", channelID, true).Find(&rows).Error
-	if err != nil {
-		return nil, fmt.Errorf("load model marks: %w", err)
-	}
-	marks := make(map[string]model.AiModelCache, len(rows))
-	for _, r := range rows {
-		marks[r.ModelOcid] = r
-	}
-	return marks, nil
-}
-
 func (s *AiGatewayService) channelModels(ctx context.Context, channelID uint) ([]model.AiModelCache, error) {
 	var rows []model.AiModelCache
 	err := s.db.WithContext(ctx).Where("channel_id = ?", channelID).Order("name ASC").Find(&rows).Error
 	return rows, err
 }

-// GatewayModels 聚合启用渠道的可用模型(按名称去重,剔除不可按需调用标记),
-// 供 /ai/v1/models;group 非空时仅聚合该分组渠道(与密钥分组路由口径一致)。
+// GatewayModels 聚合启用渠道的可用模型(按名称去重),供 /ai/v1/models;
+// group 非空时仅聚合该分组渠道(与密钥分组路由口径一致)。
 func (s *AiGatewayService) GatewayModels(ctx context.Context, group string) (aiwire.ModelList, error) {
 	q := s.db.WithContext(ctx).
-		Joins("JOIN ai_channels ON ai_channels.id = ai_model_caches.channel_id AND ai_channels.enabled = ?", true).
-		Where("(ai_model_caches.unusable = ? OR ai_model_caches.unusable IS NULL)", false)
+		Joins("JOIN ai_channels ON ai_channels.id = ai_model_caches.channel_id AND ai_channels.enabled = ?", true)
 	if group != "" {
 		q = q.Where("ai_channels.channel_group = ?", group)
 	}
@@ -658,7 +538,6 @@ func (s *AiGatewayService) DeprecatingModels(ctx context.Context, within time.Du
 	err := s.db.WithContext(ctx).
 		Where("(retired_at IS NOT NULL AND retired_at > ? AND retired_at <= ?) OR (deprecated_at IS NOT NULL AND deprecated_at >= ? AND deprecated_at <= ?)",
 			now, deadline, now, deadline).
-		Where("(unusable = ? OR unusable IS NULL)", false).
 		Order("name ASC").Find(&rows).Error
 	if err != nil {
 		return nil, err
@@ -679,7 +558,7 @@ func (s *AiGatewayService) DeprecatingModels(ctx context.Context, within time.Du
 	return out, nil
 }

-// ProbeAll 逐个探测全部渠道并顺带验证模型池,返回状态汇总;供 AI 探测后台任务调用。
+// ProbeAll 逐个探测全部渠道,返回状态汇总;供 AI 探测后台任务调用。
 func (s *AiGatewayService) ProbeAll(ctx context.Context) (string, error) {
 	chs, err := s.Channels(ctx)
 	if err != nil {
@@ -695,8 +574,6 @@ func (s *AiGatewayService) ProbeAll(ctx context.Context) (string, error) {
 			continue
 		}
 		counts[fresh.ProbeStatus]++
-		// 后台任务里同步执行:新模型验证入池、坏模型定期复检,零人工收敛
-		s.validateChannelModels(ctx, ch.ID)
 	}
 	msg := fmt.Sprintf("probed %d: %d ok, %d no_service, %d no_quota, %d error",
 		len(chs), counts["ok"], counts["no_service"], counts["no_quota"], counts["error"])
@@ -706,6 +583,99 @@ func (s *AiGatewayService) ProbeAll(ctx context.Context) (string, error) {
 	return msg, nil
 }

+// ---- 模型黑名单 ----
+
+// Blacklist 列出全部黑名单模型(按名称排序)。
+func (s *AiGatewayService) Blacklist(ctx context.Context) ([]model.AiModelBlacklist, error) {
+	var rows []model.AiModelBlacklist
+	err := s.db.WithContext(ctx).Order("name ASC").Find(&rows).Error
+	return rows, err
+}
+
+// AddBlacklist 把模型名加入黑名单并删除全部渠道缓存中的同名条目;
+// 该模型此后同步 / 探测均被过滤,直到移出黑名单后重新同步。
+func (s *AiGatewayService) AddBlacklist(ctx context.Context, name string) (*model.AiModelBlacklist, error) {
+	name = strings.TrimSpace(name)
+	if name == "" {
+		return nil, fmt.Errorf("模型名不能为空")
+	}
+	row := model.AiModelBlacklist{Name: name}
+	err := s.db.WithContext(ctx).Transaction(func(tx *gorm.DB) error {
+		var n int64
+		if err := tx.Model(&model.AiModelBlacklist{}).Where("name = ?", name).Count(&n).Error; err != nil {
+			return err
+		}
+		if n > 0 {
+			return fmt.Errorf("模型已在黑名单中")
+		}
+		if err := tx.Create(&row).Error; err != nil {
+			return err
+		}
+		return tx.Where("name = ?", name).Delete(&model.AiModelCache{}).Error
+	})
+	if err != nil {
+		return nil, err
+	}
+	return &row, nil
+}
+
+// RemoveBlacklist 把模型移出黑名单;缓存不回填,下次同步 / 探测自然恢复入池。
+func (s *AiGatewayService) RemoveBlacklist(ctx context.Context, id uint) error {
+	res := s.db.WithContext(ctx).Delete(&model.AiModelBlacklist{}, id)
+	if res.Error != nil {
+		return res.Error
+	}
+	if res.RowsAffected == 0 {
+		return fmt.Errorf("黑名单条目不存在")
+	}
+	return nil
+}
+
+// withoutBlacklisted 过滤掉黑名单中的模型,同步与探测入库前统一经此收口。
+// supportedGatewayModels 过滤模型目录:对话模型仅保留实测支持 OpenAI 兼容面的
+// 厂商(xai / meta / openai)——typed chat 面已剔除,google / cohere 对话模型无
+// 上游通路,不入目录(不出现在列表、路由与探测候选);EMBEDDING 模型不受影响。
+func supportedGatewayModels(models []oci.GenAiModel) []oci.GenAiModel {
+	out := make([]oci.GenAiModel, 0, len(models))
+	for _, m := range models {
+		if m.Capability == "CHAT" && !compatChatVendor(m.Name) {
+			continue
+		}
+		out = append(out, m)
+	}
+	return out
+}
+
+func compatChatVendor(model string) bool {
+	for _, prefix := range []string{"xai.", "meta.", "openai."} {
+		if strings.HasPrefix(model, prefix) {
+			return true
+		}
+	}
+	return false
+}
+
+func (s *AiGatewayService) withoutBlacklisted(ctx context.Context, models []oci.GenAiModel) ([]oci.GenAiModel, error) {
+	var names []string
+	if err := s.db.WithContext(ctx).Model(&model.AiModelBlacklist{}).Pluck("name", &names).Error; err != nil {
+		return nil, fmt.Errorf("读取模型黑名单: %w", err)
+	}
+	if len(names) == 0 {
+		return models, nil
+	}
+	banned := make(map[string]bool, len(names))
+	for _, n := range names {
+		banned[n] = true
+	}
+	out := make([]oci.GenAiModel, 0, len(models))
+	for _, m := range models {
+		if !banned[m.Name] {
+			out = append(out, m)
+		}
+	}
+	return out, nil
+}
+
 // ---- 调用日志 ----

 // LogCall 落一条调用日志(仅元数据与用量,永不含请求 / 响应正文),返回落库 ID 供内容日志关联(失败为 0)。
@@ -3,6 +3,7 @@ package service
 import (
 	"context"
 	"errors"
+	"io"
 	"math/rand"
 	"time"

@@ -26,24 +27,25 @@ type aiCandidate struct {
 	modelOcid string
 }

-// Chat 编排非流式调用:选渠道 → 调用 → 可重试错误换渠道(整请求上限 3 次)。
-// group 非空时只在同分组渠道内路由(取自调用密钥)。
-func (s *AiGatewayService) Chat(ctx context.Context, ir aiwire.ChatRequest, group string) (*aiwire.ChatResponse, ChatMeta, error) {
+// RespPassthrough 编排一次非流式直通调用:选渠道(priority→加权随机)→ 调用 →
+// 可重试错误换渠道(整请求上限 3 次)并维护熔断;group 非空时只在同分组渠道内路由。
+// 上游为 OpenAI-compatible /actions/v1/responses(实测可用,无 Oracle 文档合同)。
+func (s *AiGatewayService) RespPassthrough(ctx context.Context, raw []byte, modelName, group string) ([]byte, ChatMeta, error) {
 	meta := ChatMeta{}
 	excluded := map[uint]bool{}
 	var lastErr error
 	for attempt := 0; attempt < 3; attempt++ {
-		cand, err := s.pick(ctx, ir.Model, group, "CHAT", excluded)
+		cand, err := s.pick(ctx, modelName, group, "CHAT", excluded)
 		if err != nil {
 			return nil, meta, firstErr(lastErr, err)
 		}
 		meta.ChannelID, meta.ChannelName = cand.ch.ID, cand.ch.Name
-		resp, err := s.callOnce(ctx, cand, ir)
+		payload, err := s.passthroughOnce(ctx, cand, raw)
 		if err == nil {
 			s.markSuccess(ctx, cand.ch.ID)
-			return resp, meta, nil
+			return payload, meta, nil
 		}
-		retry, penalize := s.noteCallErr(ctx, cand, err)
+		retry, penalize := switchable(err)
 		if !retry {
 			return nil, meta, err
 		}
@@ -57,22 +59,22 @@ func (s *AiGatewayService) Chat(ctx context.Context, ir aiwire.ChatRequest, grou
 	return nil, meta, lastErr
 }

-func (s *AiGatewayService) callOnce(ctx context.Context, cand *aiCandidate, ir aiwire.ChatRequest) (*aiwire.ChatResponse, error) {
+func (s *AiGatewayService) passthroughOnce(ctx context.Context, cand *aiCandidate, raw []byte) ([]byte, error) {
 	cred, err := s.configs.credentialsByID(ctx, cand.ch.OciConfigID)
 	if err != nil {
 		return nil, err
 	}
-	return s.client.GenAiChat(ctx, cred, cand.ch.Region, cand.modelOcid, ir)
+	return s.client.GenAiCompatResponses(ctx, cred, cand.ch.Region, raw)
 }

-// OpenStream 编排流式调用:流建立成功后即绑定渠道,建立失败可换渠道重试。
-// group 语义与 Chat 相同。
-func (s *AiGatewayService) OpenStream(ctx context.Context, ir aiwire.ChatRequest, group string) (oci.GenAiStream, ChatMeta, error) {
+// RespPassthroughStream 编排流式直通:流建立成功即绑定渠道,建立失败按 switchable
+// 换渠道重试;建立后的中断不重试、不计熔断(与 OpenStream 语义一致)。
+func (s *AiGatewayService) RespPassthroughStream(ctx context.Context, raw []byte, modelName, group string) (io.ReadCloser, ChatMeta, error) {
 	meta := ChatMeta{}
 	excluded := map[uint]bool{}
 	var lastErr error
 	for attempt := 0; attempt < 3; attempt++ {
-		cand, err := s.pick(ctx, ir.Model, group, "CHAT", excluded)
+		cand, err := s.pick(ctx, modelName, group, "CHAT", excluded)
 		if err != nil {
 			return nil, meta, firstErr(lastErr, err)
 		}
@@ -81,12 +83,12 @@ func (s *AiGatewayService) OpenStream(ctx context.Context, ir aiwire.ChatRequest
 		if err != nil {
 			return nil, meta, err
 		}
-		stream, err := s.client.GenAiChatStream(ctx, cred, cand.ch.Region, cand.modelOcid, ir)
+		stream, err := s.client.GenAiCompatResponsesStream(ctx, cred, cand.ch.Region, raw)
 		if err == nil {
 			s.markSuccess(ctx, cand.ch.ID)
 			return stream, meta, nil
 		}
-		retry, penalize := s.noteCallErr(ctx, cand, err)
+		retry, penalize := switchable(err)
 		if !retry {
 			return nil, meta, err
 		}
@@ -108,12 +110,11 @@ func firstErr(lastErr, pickErr error) error {
 	return pickErr
 }

-// noteCallErr 汇总一次调用失败:模型级不可用(微调基座 400 / 实体不存在 404)先标记
-// 剔除该 (渠道, 模型),换渠道重试且不计熔断——这是模型×区域供给问题而非渠道健康问题;
-// 429 / 5xx / 网络错误换渠道并计熔断;其余 4xx 直接透传。
-func (s *AiGatewayService) noteCallErr(ctx context.Context, cand *aiCandidate, err error) (retry, penalize bool) {
+// switchable 判断调用失败是否换渠道重试、是否计入熔断:模型级不可用(微调基座
+// 400 / 实体不存在 404)换渠道但不计熔断——这是模型×区域供给问题而非渠道健康
+// 问题,持久解决靠加入模型黑名单;429 / 5xx / 网络错误换渠道并计熔断;其余 4xx 直接透传。
+func switchable(err error) (retry, penalize bool) {
 	if oci.IsModelUnavailable(err) {
-		s.markModelUnusable(ctx, cand.ch.ID, cand.modelOcid, oci.CompactError(err))
 		return true, false
 	}
 	if status, ok := oci.ServiceStatus(err); ok {
@@ -154,11 +155,10 @@ func (s *AiGatewayService) pick(ctx context.Context, modelName, group, capabilit
 	return &aiCandidate{ch: chosen, modelOcid: ocids[chosen.ID]}, nil
 }

-// modelChannels 查出提供该模型的渠道 ID 及各自的模型 OCID(剔除不可按需调用标记);
-// capability=CHAT 时兼容存量空串(加列前只同步对话模型);unusable 需兼容 NULL
-// (AutoMigrate 加列后、首次重同步前的存量行)。
+// modelChannels 查出提供该模型的渠道 ID 及各自的模型 OCID;
+// capability=CHAT 时兼容存量空串(加列前只同步对话模型)。
 func (s *AiGatewayService) modelChannels(ctx context.Context, modelName, capability string) (map[uint]string, []uint, error) {
-	q := s.db.WithContext(ctx).Where("name = ? AND (unusable = ? OR unusable IS NULL)", modelName, false)
+	q := s.db.WithContext(ctx).Where("name = ?", modelName)
 	if capability == "CHAT" {
 		q = q.Where("capability IN ?", []string{"CHAT", ""})
 	} else {
@@ -234,7 +234,7 @@ func (s *AiGatewayService) Embeddings(ctx context.Context, req aiwire.Embeddings
 			s.markSuccess(ctx, cand.ch.ID)
 			return resp, meta, nil
 		}
-		retry, penalize := s.noteCallErr(ctx, cand, err)
+		retry, penalize := switchable(err)
 		if !retry {
 			return nil, meta, err
 		}
@@ -0,0 +1,29 @@
+package service
+
+import (
+	"strings"
+	"testing"
+
+	"oci-portal/internal/oci"
+)
+
+// TestSupportedGatewayModels 断言 CHAT 目录仅保留兼容面厂商,EMBEDDING 不受影响。
+func TestSupportedGatewayModels(t *testing.T) {
+	in := []oci.GenAiModel{
+		{Name: "xai.grok-4.3", Capability: "CHAT"},
+		{Name: "meta.llama-4-maverick-17b-128e", Capability: "CHAT"},
+		{Name: "openai.gpt-oss-120b", Capability: "CHAT"},
+		{Name: "google.gemini-2.5-flash", Capability: "CHAT"},
+		{Name: "cohere.command-a-03-2025", Capability: "CHAT"},
+		{Name: "cohere.embed-v4.0", Capability: "EMBEDDING"},
+	}
+	out := supportedGatewayModels(in)
+	names := make([]string, 0, len(out))
+	for _, m := range out {
+		names = append(names, m.Name)
+	}
+	want := "xai.grok-4.3,meta.llama-4-maverick-17b-128e,openai.gpt-oss-120b,cohere.embed-v4.0"
+	if got := strings.Join(names, ","); got != want {
+		t.Fatalf("过滤结果 = %s, want %s", got, want)
+	}
+}
@@ -1,9 +1,11 @@
 package service

 import (
+	"bytes"
 	"context"
 	"errors"
 	"fmt"
+	"io"
 	"strings"
 	"testing"
 	"time"
@@ -45,15 +47,46 @@ type gatewayStubClient struct {
 	probeCode int
 	probeErr  error
 	// probeSeq 非空时逐次弹出,模拟按候选依次试调;弹尽后回落 probeCode/probeErr
-	probeSeq  []probeResult
-	chatResp  *aiwire.ChatResponse
-	chatErrs  []error
-	chatCalls int
-	regions   []string
+	probeSeq []probeResult
+	// probedNames 记录试调过的模型名,供断言候选过滤
+	probedNames []string
+	chatCalls   int
+	regions     []string

 	embedVecs  [][]float32
 	embedUsage *aiwire.Usage
 	embedErr   error
+
+	passPayload []byte
+	passErrs    []error
+	passCalls   int
+	passRegions []string
+}
+
+func (f *gatewayStubClient) GenAiCompatResponses(ctx context.Context, cred oci.Credentials, region string, body []byte) ([]byte, error) {
+	f.passCalls++
+	f.passRegions = append(f.passRegions, region)
+	if len(f.passErrs) > 0 {
+		err := f.passErrs[0]
+		f.passErrs = f.passErrs[1:]
+		if err != nil {
+			return nil, err
+		}
+	}
+	return f.passPayload, nil
+}
+
+func (f *gatewayStubClient) GenAiCompatResponsesStream(ctx context.Context, cred oci.Credentials, region string, body []byte) (io.ReadCloser, error) {
+	f.passCalls++
+	f.passRegions = append(f.passRegions, region)
+	if len(f.passErrs) > 0 {
+		err := f.passErrs[0]
+		f.passErrs = f.passErrs[1:]
+		if err != nil {
+			return nil, err
+		}
+	}
+	return io.NopCloser(bytes.NewReader(f.passPayload)), nil
 }

 func (f *gatewayStubClient) GenAiEmbed(ctx context.Context, cred oci.Credentials, region, modelOcid string, inputs []string, dimensions *int) ([][]float32, *aiwire.Usage, error) {
@@ -71,6 +104,7 @@ type probeResult struct {
 }

 func (f *gatewayStubClient) GenAiProbeChat(ctx context.Context, cred oci.Credentials, region, modelOcid, modelName string) (int, error) {
+	f.probedNames = append(f.probedNames, modelName)
 	if len(f.probeSeq) > 0 {
 		r := f.probeSeq[0]
 		f.probeSeq = f.probeSeq[1:]
@@ -79,23 +113,10 @@ func (f *gatewayStubClient) GenAiProbeChat(ctx context.Context, cred oci.Credent
 	return f.probeCode, f.probeErr
 }

-func (f *gatewayStubClient) GenAiChat(ctx context.Context, cred oci.Credentials, region, modelOcid string, ir aiwire.ChatRequest) (*aiwire.ChatResponse, error) {
-	f.chatCalls++
-	f.regions = append(f.regions, region)
-	if len(f.chatErrs) > 0 {
-		err := f.chatErrs[0]
-		f.chatErrs = f.chatErrs[1:]
-		if err != nil {
-			return nil, err
-		}
-	}
-	return f.chatResp, nil
-}
-
 func newTestGateway(t *testing.T, client oci.Client) (*AiGatewayService, *OciConfigService) {
 	t.Helper()
 	svc := newTestService(t, client)
-	if err := svc.db.AutoMigrate(&model.AiKey{}, &model.AiChannel{}, &model.AiModelCache{}, &model.AiCallLog{}); err != nil {
+	if err := svc.db.AutoMigrate(&model.AiKey{}, &model.AiChannel{}, &model.AiModelCache{}, &model.AiModelBlacklist{}, &model.AiCallLog{}); err != nil {
 		t.Fatalf("auto migrate ai tables: %v", err)
 	}
 	return NewAiGatewayService(svc.db, svc, client), svc
@@ -105,14 +126,14 @@ func TestAiKeyLifecycle(t *testing.T) {
 	gw, _ := newTestGateway(t, &gatewayStubClient{fakeClient: &fakeClient{tenancy: oci.TenancyInfo{Name: "t"}}})
 	ctx := context.Background()

-	raw, key, err := gw.CreateKey(ctx, "免费-ai-api-key", "test-api-key", "")
+	raw, key, err := gw.CreateKey(ctx, "免费-ai-api-key", "test-api-key", "", nil)
 	if err != nil || raw != "test-api-key" || key.Tail != "-key" {
 		t.Fatalf("CreateKey 自定义值 = %q %+v, %v", raw, key, err)
 	}
-	if _, _, err := gw.CreateKey(ctx, "short", "abc", ""); err == nil {
+	if _, _, err := gw.CreateKey(ctx, "short", "abc", "", nil); err == nil {
 		t.Error("过短自定义密钥应被拒绝")
 	}
-	auto, _, err := gw.CreateKey(ctx, "auto", "", "")
+	auto, _, err := gw.CreateKey(ctx, "auto", "", "", nil)
 	if err != nil || len(auto) < 40 || auto[:3] != "sk-" {
 		t.Fatalf("CreateKey 随机值 = %q, %v", auto, err)
 	}
@@ -124,7 +145,7 @@ func TestAiKeyLifecycle(t *testing.T) {
 		t.Errorf("错误密钥 err = %v", err)
 	}
 	off := false
-	_ = gw.UpdateKey(ctx, key.ID, "", &off, nil)
+	_ = gw.UpdateKey(ctx, key.ID, "", &off, nil, nil)
 	if _, err := gw.VerifyKey(ctx, "test-api-key"); !errors.Is(err, ErrAiKeyInvalid) {
 		t.Errorf("禁用后 VerifyKey err = %v", err)
 	}
@@ -187,51 +208,6 @@ func seedChannel(t *testing.T, gw *AiGatewayService, cfgID uint, region string,
 	return ch
 }

-func TestAiChatRetrySwitchesChannel(t *testing.T) {
-	client := &gatewayStubClient{
-		fakeClient: &fakeClient{tenancy: oci.TenancyInfo{Name: "t"}},
-		chatResp:   &aiwire.ChatResponse{Model: "meta.llama-3.3-70b-instruct", Choices: []aiwire.Choice{{Message: aiwire.ChatMessage{Role: "assistant", Content: aiwire.NewTextContent("hi")}, FinishReason: "stop"}}},
-		chatErrs:   []error{stubServiceError{status: 429}},
-	}
-	gw, svc := newTestGateway(t, client)
-	cfg := importAliveConfig(t, svc)
-	// 两个同优先级渠道
-	seedChannel(t, gw, cfg.ID, "eu-frankfurt-1", 1, 1)
-	seedChannel(t, gw, cfg.ID, "us-chicago-1", 1, 1)
-	ctx := context.Background()
-
-	resp, meta, err := gw.Chat(ctx, aiwire.ChatRequest{Model: "meta.llama-3.3-70b-instruct", Messages: []aiwire.ChatMessage{{Role: "user", Content: aiwire.NewTextContent("你好")}}}, "")
-	if err != nil || resp == nil {
-		t.Fatalf("Chat = %v, %v", resp, err)
-	}
-	if meta.Retries != 1 || client.chatCalls != 2 {
-		t.Errorf("应换渠道重试一次: retries=%d calls=%d", meta.Retries, client.chatCalls)
-	}
-	if len(client.regions) != 2 && client.regions[0] == client.regions[1] {
-		t.Errorf("重试未换渠道: %v", client.regions)
-	}
-	// 未知模型
-	if _, _, err := gw.Chat(ctx, aiwire.ChatRequest{Model: "no-such-model"}, ""); !errors.Is(err, ErrAiUnknownModel) {
-		t.Errorf("未知模型 err = %v", err)
-	}
-}
-
-func TestAiChatNonRetryablePassThrough(t *testing.T) {
-	client := &gatewayStubClient{
-		fakeClient: &fakeClient{tenancy: oci.TenancyInfo{Name: "t"}},
-		chatErrs:   []error{stubServiceError{status: 400}},
-	}
-	gw, svc := newTestGateway(t, client)
-	cfg := importAliveConfig(t, svc)
-	seedChannel(t, gw, cfg.ID, "eu-frankfurt-1", 1, 1)
-	seedChannel(t, gw, cfg.ID, "us-chicago-1", 1, 1)
-
-	_, meta, err := gw.Chat(context.Background(), aiwire.ChatRequest{Model: "meta.llama-3.3-70b-instruct"}, "")
-	if err == nil || meta.Retries != 0 || client.chatCalls != 1 {
-		t.Errorf("400 应直接透传不重试: err=%v retries=%d calls=%d", err, meta.Retries, client.chatCalls)
-	}
-}
-
 func TestPickPriorityAndBreaker(t *testing.T) {
 	gw, svc := newTestGateway(t, &gatewayStubClient{fakeClient: &fakeClient{tenancy: oci.TenancyInfo{Name: "t"}}})
 	cfg := importAliveConfig(t, svc)
@@ -282,8 +258,8 @@ func TestMarkFailureBackoff(t *testing.T) {

 func TestAiGroupRouting(t *testing.T) {
 	client := &gatewayStubClient{
-		fakeClient: &fakeClient{tenancy: oci.TenancyInfo{Name: "t"}},
-		chatResp:   &aiwire.ChatResponse{Model: "meta.llama-3.3-70b-instruct", Choices: []aiwire.Choice{{Message: aiwire.ChatMessage{Role: "assistant", Content: aiwire.NewTextContent("hi")}, FinishReason: "stop"}}},
+		fakeClient:  &fakeClient{tenancy: oci.TenancyInfo{Name: "t"}},
+		passPayload: []byte(`{"id":"resp_1","usage":{"input_tokens":3,"output_tokens":1}}`),
 	}
 	gw, svc := newTestGateway(t, client)
 	cfg := importAliveConfig(t, svc)
@@ -292,17 +268,17 @@ func TestAiGroupRouting(t *testing.T) {
 	gw.db.Model(vip).Update("channel_group", "vip")
 	gw.db.Create(&model.AiModelCache{ChannelID: other.ID, ModelOcid: "ocid1..cr", Name: "cohere.command-r", Vendor: "cohere", SyncedAt: time.Now()})
 	ctx := context.Background()
-	req := aiwire.ChatRequest{Model: "meta.llama-3.3-70b-instruct", Messages: []aiwire.ChatMessage{{Role: "user", Content: aiwire.NewTextContent("你好")}}}
+	body := []byte(`{"model":"meta.llama-3.3-70b-instruct","input":"你好"}`)

 	// 分组密钥只落同分组渠道
 	for i := 0; i < 5; i++ {
-		_, meta, err := gw.Chat(ctx, req, "vip")
+		_, meta, err := gw.RespPassthrough(ctx, body, "meta.llama-3.3-70b-instruct", "vip")
 		if err != nil || meta.ChannelID != vip.ID {
 			t.Fatalf("vip 组第 %d 次落点 = %d, err %v, want %d", i, meta.ChannelID, err, vip.ID)
 		}
 	}
 	// 分组内无渠道 → 无可用渠道
-	if _, _, err := gw.Chat(ctx, req, "nope"); !errors.Is(err, ErrAiNoChannel) {
+	if _, _, err := gw.RespPassthrough(ctx, body, "meta.llama-3.3-70b-instruct", "nope"); !errors.Is(err, ErrAiNoChannel) {
 		t.Errorf("空分组 err = %v, want ErrAiNoChannel", err)
 	}
 	// 模型列表按分组过滤
@@ -315,12 +291,12 @@ func TestAiGroupRouting(t *testing.T) {
 		t.Errorf("不限组模型数 = %d, want 2", len(all.Data))
 	}
 	// 密钥分组落库
-	_, key, err := gw.CreateKey(ctx, "vip-key", "vip-secret-1234", "vip")
+	_, key, err := gw.CreateKey(ctx, "vip-key", "vip-secret-1234", "vip", nil)
 	if err != nil || key.Group != "vip" {
 		t.Fatalf("CreateKey group = %+v, %v", key, err)
 	}
 	empty := ""
-	_ = gw.UpdateKey(ctx, key.ID, "", nil, &empty)
+	_ = gw.UpdateKey(ctx, key.ID, "", nil, &empty, nil)
 	var fresh model.AiKey
 	gw.db.First(&fresh, key.ID)
 	if fresh.Group != "" {
@@ -465,7 +441,8 @@ func entityNotFoundErr() stubServiceError {
 }

 func TestProbeSkipsUnavailableModels(t *testing.T) {
-	// 首选 llama 不可按需调用(基座 400 / 实体 404)→ 标记剔除 → 换候选成功 → 渠道判可用
+	// 首选 llama 不可按需调用(基座 400 / 实体 404)→ 换候选成功 → 渠道判可用;
+	// 模型全量入库不再自动标记,持久剔除交由用户手动拉黑
 	tests := []struct {
 		name string
 		bad  probeResult
@@ -496,112 +473,16 @@ func TestProbeSkipsUnavailableModels(t *testing.T) {
 			if err != nil || probed.ProbeStatus != "ok" {
 				t.Fatalf("坏候选后应换候选并判可用: %+v, %v", probed, err)
 			}
-			var row model.AiModelCache
-			if err := gw.db.Where("channel_id = ? AND model_ocid = ?", ch.ID, "m1").First(&row).Error; err != nil || !row.Unusable {
-				t.Fatalf("首个坏候选应被标记不可用: %+v, %v", row, err)
-			}
-			// 再次探测:同步保留标记,m1 不再进候选(probeSeq 只需一次 200)
-			client.probeSeq = []probeResult{{200, nil}}
-			probed, err = gw.ProbeChannel(ctx, ch.ID)
-			if err != nil || probed.ProbeStatus != "ok" {
-				t.Fatalf("复测应跳过已标记模型: %+v, %v", probed, err)
-			}
-			var again model.AiModelCache
-			gw.db.Where("channel_id = ? AND model_ocid = ?", ch.ID, "m1").First(&again)
-			if !again.Unusable {
-				t.Error("探测触发的同步应保留不可用标记")
-			}
-			// 手动同步同样保留标记(坏模型不随重新同步复活);近期已检的标记不被后台验证翻转
-			client.probeCode = 200
-			if _, err := gw.SyncModels(ctx, ch.ID); err != nil {
-				t.Fatalf("SyncModels: %v", err)
-			}
-			gw.Wait()
-			var kept model.AiModelCache
-			gw.db.Where("channel_id = ? AND model_ocid = ?", ch.ID, "m1").First(&kept)
-			if !kept.Unusable {
-				t.Error("手动同步不应清除不可用标记")
+			rows, _ := gw.channelModels(ctx, ch.ID)
+			if len(rows) != 3 {
+				t.Errorf("模型应全量入库, got %d", len(rows))
 			}
 		})
 	}
 }

-func TestValidateModelsAfterSync(t *testing.T) {
-	// 同步后后台验证:坏模型标记剔除、好模型记录已检、其他 4xx 不改状态、非对话模型不试调
-	client := &gatewayStubClient{
-		fakeClient: &fakeClient{tenancy: oci.TenancyInfo{Name: "t"}},
-		models: []oci.GenAiModel{
-			{Ocid: "v1", Name: "meta.llama-4-maverick", Vendor: "meta", Capability: "CHAT"},
-			{Ocid: "v2", Name: "xai.grok-4", Vendor: "xai", Capability: "CHAT"},
-			{Ocid: "v3", Name: "xai.grok-voice-agent", Vendor: "xai", Capability: "CHAT"},
-			{Ocid: "v4", Name: "cohere.embed-v4.0", Vendor: "cohere", Capability: "EMBEDDING"},
-		},
-		probeSeq: []probeResult{{404, entityNotFoundErr()}, {200, nil}, {400, stubServiceError{status: 400}}},
-	}
-	gw, svc := newTestGateway(t, client)
-	cfg := importAliveConfig(t, svc)
-	ctx := context.Background()
-	ch, err := gw.CreateChannel(ctx, ChannelInput{OciConfigID: cfg.ID, Region: "us-ashburn-1"})
-	if err != nil {
-		t.Fatalf("CreateChannel: %v", err)
-	}
-	if _, err := gw.SyncModels(ctx, ch.ID); err != nil {
-		t.Fatalf("SyncModels: %v", err)
-	}
-	gw.Wait()
-
-	want := map[string]struct {
-		unusable bool
-		checked  bool
-	}{
-		"v1": {true, true},   // 实体 404 → 标记剔除
-		"v2": {false, true},  // 200 → 可用已检
-		"v3": {false, true},  // 普通 400 → 已检不标记
-		"v4": {false, false}, // EMBEDDING 不试调
-	}
-	rows, _ := gw.channelModels(ctx, ch.ID)
-	for _, r := range rows {
-		w := want[r.ModelOcid]
-		if r.Unusable != w.unusable || (r.CheckedAt != nil) != w.checked {
-			t.Errorf("%s: unusable=%v checked=%v, want %+v", r.ModelOcid, r.Unusable, r.CheckedAt != nil, w)
-		}
-	}
-	list, err := gw.GatewayModels(ctx, "")
-	if err != nil || len(list.Data) != 3 {
-		t.Errorf("网关列表应只剔除被标记的坏模型(余 v2/v3/v4), got %+v, %v", list.Data, err)
-	}
-}
-
-func TestValidateRecheckUnmarksRecovered(t *testing.T) {
-	// 已标记模型超过复检间隔后重验:恢复供给(200)自动解除标记
-	client := &gatewayStubClient{fakeClient: &fakeClient{tenancy: oci.TenancyInfo{Name: "t"}}, probeCode: 200}
-	gw, svc := newTestGateway(t, client)
-	cfg := importAliveConfig(t, svc)
-	ch := seedChannel(t, gw, cfg.ID, "eu-frankfurt-1", 1, 1)
-	ctx := context.Background()
-	old := time.Now().Add(-25 * time.Hour)
-	gw.db.Model(&model.AiModelCache{}).Where("channel_id = ?", ch.ID).
-		Updates(map[string]any{"unusable": true, "unusable_reason": "x", "checked_at": old})
-
-	gw.validateChannelModels(ctx, ch.ID)
-	var row model.AiModelCache
-	gw.db.Where("channel_id = ?", ch.ID).First(&row)
-	if row.Unusable || row.UnusableReason != "" || row.CheckedAt == nil || !row.CheckedAt.After(old) {
-		t.Errorf("超期复检应解除标记并刷新已检时间: %+v", row)
-	}
-	// 未超期的标记不复检(probeSeq 为空、fallback 200 也不会被消费)
-	fresh := time.Now()
-	gw.db.Model(&model.AiModelCache{}).Where("channel_id = ?", ch.ID).
-		Updates(map[string]any{"unusable": true, "checked_at": fresh})
-	gw.validateChannelModels(ctx, ch.ID)
-	gw.db.Where("channel_id = ?", ch.ID).First(&row)
-	if !row.Unusable {
-		t.Error("未超期的标记不应被复检翻转")
-	}
-}
-
 func TestProbeAuth404StillNoQuota(t *testing.T) {
-	// 鉴权类 404(NotAuthorizedOrNotFound)仍属租户级,直接定论 no_quota 且不标记模型
+	// 鉴权类 404(NotAuthorizedOrNotFound)仍属租户级,直接定论 no_quota
 	client := &gatewayStubClient{
 		fakeClient: &fakeClient{tenancy: oci.TenancyInfo{Name: "t"}},
 		models:     []oci.GenAiModel{{Ocid: "m1", Name: "meta.llama-3.3-70b-instruct", Vendor: "meta"}},
@@ -617,31 +498,26 @@ func TestProbeAuth404StillNoQuota(t *testing.T) {
 	if probed.ProbeStatus != "no_quota" {
 		t.Errorf("鉴权 404 status = %q, want no_quota", probed.ProbeStatus)
 	}
-	var marked int64
-	gw.db.Model(&model.AiModelCache{}).Where("unusable = ?", true).Count(&marked)
-	if marked != 0 {
-		t.Errorf("鉴权 404 不应标记模型, marked=%d", marked)
-	}
 }

 func TestAiChatFinetuneSwitchesChannelWithoutPenalty(t *testing.T) {
 	// 微调基座 400 换渠道重试成功,且不计入熔断失败
 	client := &gatewayStubClient{
-		fakeClient: &fakeClient{tenancy: oci.TenancyInfo{Name: "t"}},
-		chatResp:   &aiwire.ChatResponse{Model: "meta.llama-3.3-70b-instruct", Choices: []aiwire.Choice{{Message: aiwire.ChatMessage{Role: "assistant", Content: aiwire.NewTextContent("hi")}, FinishReason: "stop"}}},
-		chatErrs:   []error{finetuneBaseErr()},
+		fakeClient:  &fakeClient{tenancy: oci.TenancyInfo{Name: "t"}},
+		passPayload: []byte(`{"id":"resp_1","usage":{"input_tokens":3,"output_tokens":1}}`),
+		passErrs:    []error{finetuneBaseErr()},
 	}
 	gw, svc := newTestGateway(t, client)
 	cfg := importAliveConfig(t, svc)
 	seedChannel(t, gw, cfg.ID, "eu-frankfurt-1", 1, 1)
 	seedChannel(t, gw, cfg.ID, "us-chicago-1", 1, 1)

-	resp, meta, err := gw.Chat(context.Background(), aiwire.ChatRequest{Model: "meta.llama-3.3-70b-instruct", Messages: []aiwire.ChatMessage{{Role: "user", Content: aiwire.NewTextContent("你好")}}}, "")
+	resp, meta, err := gw.RespPassthrough(context.Background(), []byte(`{}`), "meta.llama-3.3-70b-instruct", "")
 	if err != nil || resp == nil {
-		t.Fatalf("Chat = %v, %v", resp, err)
+		t.Fatalf("RespPassthrough = %v, %v", resp, err)
 	}
-	if meta.Retries != 1 || client.chatCalls != 2 {
-		t.Errorf("应换渠道重试一次: retries=%d calls=%d", meta.Retries, client.chatCalls)
+	if meta.Retries != 1 || client.passCalls != 2 {
+		t.Errorf("应换渠道重试一次: retries=%d calls=%d", meta.Retries, client.passCalls)
 	}
 	var chs []model.AiChannel
 	gw.db.Find(&chs)
@@ -650,38 +526,93 @@ func TestAiChatFinetuneSwitchesChannelWithoutPenalty(t *testing.T) {
 			t.Errorf("微调基座 400 不应计入熔断: 渠道 %s failCount=%d", ch.Name, ch.FailCount)
 		}
 	}
-	// 失败渠道的该模型被标记,不再参与路由;成功渠道不受影响
-	region := client.regions[0]
-	var row model.AiModelCache
-	gw.db.Where("model_ocid = ?", "ocid1..m-"+region).First(&row)
-	if !row.Unusable {
-		t.Errorf("失败渠道的模型应被标记不可用: %+v", row)
+}
+
+func TestBlacklistLifecycle(t *testing.T) {
+	// 拉黑:全渠道同名缓存删除、列表与路由立即不可见;重复/空名被拒;
+	// 移出黑名单后重新同步恢复入池
+	client := &gatewayStubClient{
+		fakeClient: &fakeClient{tenancy: oci.TenancyInfo{Name: "t"}},
+		models:     []oci.GenAiModel{{Ocid: "ocid1..m-eu-frankfurt-1", Name: "meta.llama-3.3-70b-instruct", Vendor: "meta"}},
 	}
-	var usable int64
-	gw.db.Model(&model.AiModelCache{}).Where("unusable = ?", false).Count(&usable)
-	if usable != 1 {
-		t.Errorf("成功渠道模型应保持可用, usable=%d", usable)
+	gw, svc := newTestGateway(t, client)
+	cfg := importAliveConfig(t, svc)
+	ch := seedChannel(t, gw, cfg.ID, "eu-frankfurt-1", 1, 1)
+	seedChannel(t, gw, cfg.ID, "us-chicago-1", 1, 1)
+	ctx := context.Background()
+
+	item, err := gw.AddBlacklist(ctx, " meta.llama-3.3-70b-instruct ")
+	if err != nil || item.Name != "meta.llama-3.3-70b-instruct" {
+		t.Fatalf("AddBlacklist = %+v, %v", item, err)
+	}
+	var left int64
+	gw.db.Model(&model.AiModelCache{}).Count(&left)
+	if left != 0 {
+		t.Errorf("拉黑应删除全部渠道的同名缓存, 剩 %d", left)
+	}
+	if _, _, err := gw.RespPassthrough(ctx, []byte(`{}`), "meta.llama-3.3-70b-instruct", ""); !errors.Is(err, ErrAiUnknownModel) {
+		t.Errorf("拉黑后路由应按未知模型拒绝: %v", err)
+	}
+	if _, err := gw.AddBlacklist(ctx, "meta.llama-3.3-70b-instruct"); err == nil {
+		t.Error("重复拉黑应被拒绝")
+	}
+	if _, err := gw.AddBlacklist(ctx, "  "); err == nil {
+		t.Error("空模型名应被拒绝")
+	}
+	rows, err := gw.Blacklist(ctx)
+	if err != nil || len(rows) != 1 {
+		t.Fatalf("Blacklist = %+v, %v", rows, err)
+	}
+	if err := gw.RemoveBlacklist(ctx, rows[0].ID); err != nil {
+		t.Fatalf("RemoveBlacklist: %v", err)
+	}
+	if err := gw.RemoveBlacklist(ctx, rows[0].ID); err == nil {
+		t.Error("移除不存在的条目应报错")
+	}
+	// 移出后重新同步:模型恢复入池
+	if _, err := gw.SyncModels(ctx, ch.ID); err != nil {
+		t.Fatalf("SyncModels: %v", err)
+	}
+	list, _ := gw.GatewayModels(ctx, "")
+	if len(list.Data) != 1 {
+		t.Errorf("移出黑名单并同步后应恢复入池: %+v", list.Data)
 	}
 }

-func TestUnusableModelExcludedFromPoolAndRouting(t *testing.T) {
-	// 唯一渠道的模型被标记后:网关列表不再展示,路由按未知模型拒绝,渠道详情仍可见标记
-	gw, svc := newTestGateway(t, &gatewayStubClient{fakeClient: &fakeClient{tenancy: oci.TenancyInfo{Name: "t"}}})
+func TestSyncAndProbeFilterBlacklisted(t *testing.T) {
+	// 同步与探测都过滤黑名单模型:不入库、不进试调候选
+	client := &gatewayStubClient{
+		fakeClient: &fakeClient{tenancy: oci.TenancyInfo{Name: "t"}},
+		models: []oci.GenAiModel{
+			{Ocid: "m1", Name: "meta.llama-3-70b-instruct", Vendor: "meta"},
+			{Ocid: "m2", Name: "cohere.command-a-03-2025", Vendor: "cohere"},
+		},
+		probeCode: 200,
+	}
+	gw, svc := newTestGateway(t, client)
 	cfg := importAliveConfig(t, svc)
-	ch := seedChannel(t, gw, cfg.ID, "eu-frankfurt-1", 1, 1)
 	ctx := context.Background()
-
-	gw.markModelUnusable(ctx, ch.ID, "ocid1..m-eu-frankfurt-1", "Entity with key … not found")
-	list, err := gw.GatewayModels(ctx, "")
-	if err != nil || len(list.Data) != 0 {
-		t.Errorf("已标记模型不应出现在网关列表: %+v, %v", list.Data, err)
+	if _, err := gw.AddBlacklist(ctx, "meta.llama-3-70b-instruct"); err != nil {
+		t.Fatalf("AddBlacklist: %v", err)
 	}
-	if _, _, err := gw.Chat(ctx, aiwire.ChatRequest{Model: "meta.llama-3.3-70b-instruct"}, ""); !errors.Is(err, ErrAiUnknownModel) {
-		t.Errorf("已标记模型路由应拒绝: %v", err)
+	ch, err := gw.CreateChannel(ctx, ChannelInput{OciConfigID: cfg.ID, Region: "eu-frankfurt-1"})
+	if err != nil {
+		t.Fatalf("CreateChannel: %v", err)
 	}
-	rows, err := gw.channelModels(ctx, ch.ID)
-	if err != nil || len(rows) != 1 || !rows[0].Unusable || rows[0].UnusableReason == "" {
-		t.Errorf("渠道详情应保留标记行便于排查: %+v, %v", rows, err)
+	probed, err := gw.ProbeChannel(ctx, ch.ID)
+	if err != nil || probed.ProbeStatus != "ok" {
+		t.Fatalf("ProbeChannel = %+v, %v", probed, err)
+	}
+	rows, _ := gw.channelModels(ctx, ch.ID)
+	if len(rows) != 1 || rows[0].Name != "cohere.command-a-03-2025" {
+		t.Errorf("探测同步应过滤黑名单模型: %+v", rows)
+	}
+	if len(client.probedNames) != 1 || client.probedNames[0] != "cohere.command-a-03-2025" {
+		t.Errorf("试调候选不应包含黑名单模型: %v", client.probedNames)
+	}
+	models, err := gw.SyncModels(ctx, ch.ID)
+	if err != nil || len(models) != 1 || models[0].Name != "cohere.command-a-03-2025" {
+		t.Errorf("SyncModels 应过滤黑名单模型: %+v, %v", models, err)
 	}
 }

@@ -745,7 +676,7 @@ func TestAiEmbeddings(t *testing.T) {
 		t.Errorf("chat 模型走 embeddings err = %v, want ErrAiUnknownModel", err)
 	}
 	// embedding 模型名打 chat:同样未知模型
-	if _, _, err := gw.Chat(ctx, aiwire.ChatRequest{Model: "cohere.embed-v4.0", Messages: []aiwire.ChatMessage{{Role: "user", Content: aiwire.NewTextContent("x")}}}, ""); !errors.Is(err, ErrAiUnknownModel) {
+	if _, _, err := gw.RespPassthrough(ctx, []byte(`{}`), "cohere.embed-v4.0", ""); !errors.Is(err, ErrAiUnknownModel) {
 		t.Errorf("embedding 模型走 chat err = %v, want ErrAiUnknownModel", err)
 	}
 }
@@ -756,7 +687,7 @@ func TestAiContentLogSwitch(t *testing.T) {
 		t.Fatalf("migrate content log: %v", err)
 	}
 	ctx := context.Background()
-	_, key, err := gw.CreateKey(ctx, "k1", "content-key-1234", "")
+	_, key, err := gw.CreateKey(ctx, "k1", "content-key-1234", "", nil)
 	if err != nil {
 		t.Fatalf("CreateKey: %v", err)
 	}
@@ -793,23 +724,130 @@ func TestAiContentLogSwitch(t *testing.T) {
 	}
 }

-func TestLegacyNullUnusableRowsStayVisible(t *testing.T) {
-	// 升级路径回归:AutoMigrate 加列后、首次重同步前,存量行 unusable 为 NULL,
-	// 网关列表与路由必须照常包含这些行,不能因 unusable = false 过滤而整池消失
-	gw, svc := newTestGateway(t, &gatewayStubClient{fakeClient: &fakeClient{tenancy: oci.TenancyInfo{Name: "t"}}})
+func TestRespPassthroughSwitchesChannel(t *testing.T) {
+	client := &gatewayStubClient{
+		fakeClient:  &fakeClient{tenancy: oci.TenancyInfo{Name: "t"}},
+		passPayload: []byte(`{"id":"resp_1","usage":{"input_tokens":9,"output_tokens":3}}`),
+		passErrs:    []error{stubServiceError{status: 429}},
+	}
+	gw, svc := newTestGateway(t, client)
 	cfg := importAliveConfig(t, svc)
-	ch := seedChannel(t, gw, cfg.ID, "eu-frankfurt-1", 1, 1)
-	ctx := context.Background()
-	if err := gw.db.Exec("UPDATE ai_model_caches SET unusable = NULL WHERE channel_id = ?", ch.ID).Error; err != nil {
-		t.Fatalf("set legacy null: %v", err)
-	}
+	seedChannel(t, gw, cfg.ID, "eu-frankfurt-1", 1, 1)
+	seedChannel(t, gw, cfg.ID, "us-chicago-1", 1, 1)

-	list, err := gw.GatewayModels(ctx, "")
-	if err != nil || len(list.Data) != 1 {
-		t.Errorf("存量 NULL 行应仍在网关列表: %+v, %v", list.Data, err)
+	payload, meta, err := gw.RespPassthrough(context.Background(), []byte(`{}`), "meta.llama-3.3-70b-instruct", "")
+	if err != nil || len(payload) == 0 {
+		t.Fatalf("RespPassthrough = %v, %v", payload, err)
 	}
-	_, ids, err := gw.modelChannels(ctx, "meta.llama-3.3-70b-instruct", "CHAT")
-	if err != nil || len(ids) != 1 {
-		t.Errorf("存量 NULL 行应仍参与路由: %v, %v", ids, err)
+	if meta.Retries != 1 || client.passCalls != 2 {
+		t.Errorf("应换渠道重试一次: retries=%d calls=%d", meta.Retries, client.passCalls)
+	}
+	if len(client.passRegions) == 2 && client.passRegions[0] == client.passRegions[1] {
+		t.Errorf("重试未换渠道: %v", client.passRegions)
+	}
+}
+
+func TestRespPassthroughNonRetryable(t *testing.T) {
+	client := &gatewayStubClient{
+		fakeClient: &fakeClient{tenancy: oci.TenancyInfo{Name: "t"}},
+		passErrs:   []error{stubServiceError{status: 400}},
+	}
+	gw, svc := newTestGateway(t, client)
+	cfg := importAliveConfig(t, svc)
+	seedChannel(t, gw, cfg.ID, "eu-frankfurt-1", 1, 1)
+	seedChannel(t, gw, cfg.ID, "us-chicago-1", 1, 1)
+
+	_, meta, err := gw.RespPassthrough(context.Background(), []byte(`{}`), "meta.llama-3.3-70b-instruct", "")
+	if err == nil || meta.Retries != 0 || client.passCalls != 1 {
+		t.Errorf("400 应直接透传不重试: err=%v retries=%d calls=%d", err, meta.Retries, client.passCalls)
+	}
+}
+
+func TestNormalizeKeyModels(t *testing.T) {
+	tests := []struct {
+		name string
+		in   []string
+		want []string
+	}{
+		{"nil 输入", nil, nil},
+		{"全空串", []string{"", "  "}, nil},
+		{"trim 与保序去重", []string{" a ", "b", "a", "", "b "}, []string{"a", "b"}},
+	}
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			got := normalizeKeyModels(tt.in)
+			if len(got) != len(tt.want) {
+				t.Fatalf("normalizeKeyModels(%v) = %v, want %v", tt.in, got, tt.want)
+			}
+			for i := range tt.want {
+				if got[i] != tt.want[i] {
+					t.Fatalf("normalizeKeyModels(%v) = %v, want %v", tt.in, got, tt.want)
+				}
+			}
+		})
+	}
+}
+
+// TestAiKeyModelsPersistence 钉住 serializer:json 字段经 Create 与 Updates(map) 两条路径的往返。
+func TestAiKeyModelsPersistence(t *testing.T) {
+	gw, _ := newTestGateway(t, &gatewayStubClient{fakeClient: &fakeClient{tenancy: oci.TenancyInfo{Name: "t"}}})
+	ctx := context.Background()
+
+	_, key, err := gw.CreateKey(ctx, "m-key", "model-key-1234", "", []string{" x-model ", "x-model", "y-model", ""})
+	if err != nil {
+		t.Fatalf("CreateKey: %v", err)
+	}
+	fresh, err := gw.VerifyKey(ctx, "model-key-1234")
+	if err != nil || len(fresh.Models) != 2 || fresh.Models[0] != "x-model" || fresh.Models[1] != "y-model" {
+		t.Fatalf("创建后读回 models = %v, %v", fresh.Models, err)
+	}
+	set := []string{"z-model"}
+	if err := gw.UpdateKey(ctx, key.ID, "", nil, nil, &set); err != nil {
+		t.Fatalf("UpdateKey 覆盖: %v", err)
+	}
+	fresh, _ = gw.VerifyKey(ctx, "model-key-1234")
+	if len(fresh.Models) != 1 || fresh.Models[0] != "z-model" {
+		t.Fatalf("覆盖后 models = %v", fresh.Models)
+	}
+	if err := gw.UpdateKey(ctx, key.ID, "renamed", nil, nil, nil); err != nil {
+		t.Fatalf("UpdateKey 不传 models: %v", err)
+	}
+	fresh, _ = gw.VerifyKey(ctx, "model-key-1234")
+	if len(fresh.Models) != 1 {
+		t.Fatalf("未传 models 却被改动: %v", fresh.Models)
+	}
+	empty := []string{}
+	if err := gw.UpdateKey(ctx, key.ID, "", nil, nil, &empty); err != nil {
+		t.Fatalf("UpdateKey 清空: %v", err)
+	}
+	fresh, _ = gw.VerifyKey(ctx, "model-key-1234")
+	if len(fresh.Models) != 0 {
+		t.Fatalf("清空后 models = %v, want 空", fresh.Models)
+	}
+}
+
+// TestRespPassthroughStreamSwitchesChannel 断言流式直通建立失败按 switchable 换渠道。
+func TestRespPassthroughStreamSwitchesChannel(t *testing.T) {
+	client := &gatewayStubClient{
+		fakeClient:  &fakeClient{tenancy: oci.TenancyInfo{Name: "t"}},
+		passPayload: []byte("data: {\"type\":\"response.completed\"}\n\n"),
+		passErrs:    []error{stubServiceError{status: 503}},
+	}
+	gw, svc := newTestGateway(t, client)
+	cfg := importAliveConfig(t, svc)
+	seedChannel(t, gw, cfg.ID, "eu-frankfurt-1", 1, 1)
+	seedChannel(t, gw, cfg.ID, "us-chicago-1", 1, 1)
+
+	stream, meta, err := gw.RespPassthroughStream(context.Background(), []byte(`{}`), "meta.llama-3.3-70b-instruct", "")
+	if err != nil || stream == nil {
+		t.Fatalf("RespPassthroughStream = %v", err)
+	}
+	defer stream.Close()
+	if meta.Retries != 1 || client.passCalls != 2 {
+		t.Errorf("应换渠道重试一次: retries=%d calls=%d", meta.Retries, client.passCalls)
+	}
+	payload, _ := io.ReadAll(stream)
+	if !bytes.Contains(payload, []byte("response.completed")) {
+		t.Errorf("流内容未透传: %s", payload)
 	}
 }
@@ -8,38 +8,8 @@ import (
 	"oci-portal/internal/aiwire"
 )

-// ResponsesToIR 把 Responses API 请求转换为网关 IR(OpenAI Chat 形态)。
-// 有状态特性与内置工具不支持,直接报错(API 层 400)。
-func ResponsesToIR(req aiwire.RespRequest) (aiwire.ChatRequest, error) {
-	if err := respRejectUnsupported(req); err != nil {
-		return aiwire.ChatRequest{}, err
-	}
-	ir := aiwire.ChatRequest{
-		Model:       req.Model,
-		MaxTokens:   req.MaxOutputTokens,
-		Temperature: req.Temperature,
-		TopP:        req.TopP,
-		Stream:      req.Stream,
-		ToolChoice:  respToolChoice(req.ToolChoice),
-	}
-	if req.Reasoning != nil {
-		ir.ReasoningEffort = req.Reasoning.Effort
-	}
-	if req.Instructions != "" {
-		ir.Messages = append(ir.Messages, aiwire.ChatMessage{Role: "system", Content: aiwire.NewTextContent(req.Instructions)})
-	}
-	msgs, err := respInputToMessages(req.Input)
-	if err != nil {
-		return aiwire.ChatRequest{}, err
-	}
-	ir.Messages = append(ir.Messages, msgs...)
-	ir.Tools = respTools(req.Tools)
-	ir.ResponseFormat = respFormat(req.Text)
-	return ir, nil
-}
-
-// respRejectUnsupported 拒绝无状态网关无法承接的请求特性。
-func respRejectUnsupported(req aiwire.RespRequest) error {
+// respRejectStateful 拒绝有状态特性(网关无状态)。
+func respRejectStateful(req aiwire.RespRequest) error {
 	if req.PreviousResponseID != "" {
 		return fmt.Errorf("previous_response_id 不支持:网关不保存历史响应,请在 input 中自带完整上下文(store:false 模式)")
 	}
@@ -49,195 +19,80 @@ func respRejectUnsupported(req aiwire.RespRequest) error {
 	if req.Background != nil && *req.Background {
 		return fmt.Errorf("background 模式不支持")
 	}
+	return nil
+}
+
+// RespServerTools 报告工具列表是否含 xAI 服务端工具(web_search / x_search)。
+func RespServerTools(tools []aiwire.RespTool) bool {
+	for _, t := range tools {
+		if t.Type == "web_search" || t.Type == "x_search" {
+			return true
+		}
+	}
+	return false
+}
+
+// RespPassthroughValidate 校验直通请求:模型必填,有状态特性不支持,工具类型
+// 只放行 function 与已实测的 web_search / x_search;流式仅在含服务端工具时拒绝
+// (工具流式事件形态未实测,不放开)。
+func RespPassthroughValidate(req aiwire.RespRequest) error {
+	if strings.TrimSpace(req.Model) == "" {
+		return fmt.Errorf("model 不能为空")
+	}
+	if req.Stream && RespServerTools(req.Tools) {
+		return fmt.Errorf("服务端工具暂不支持流式:请去掉 stream 或改用 function 工具")
+	}
+	if err := respRejectStateful(req); err != nil {
+		return err
+	}
 	for _, t := range req.Tools {
-		if t.Type != "function" {
-			return fmt.Errorf("不支持的工具类型 %q:仅支持 function 工具", t.Type)
+		switch t.Type {
+		case "function", "web_search", "x_search":
+		default:
+			return fmt.Errorf("不支持的工具类型 %q:服务端工具仅支持 web_search / x_search", t.Type)
 		}
 	}
 	return nil
 }

-// respInputToMessages 把 input(string 或 item 数组)展开为 IR 消息序列。
-func respInputToMessages(in aiwire.RespInput) ([]aiwire.ChatMessage, error) {
-	if !in.IsArray {
-		if strings.TrimSpace(in.Text) == "" {
-			return nil, fmt.Errorf("input 不能为空")
-		}
-		return []aiwire.ChatMessage{{Role: "user", Content: aiwire.NewTextContent(in.Text)}}, nil
+// RespPassthroughBody 以原始请求体为基构造上游 body:强制 store:false(禁上游
+// 存态),stream 原样保留(流式直通);用 json.Number 保真未知字段与数值。
+func RespPassthroughBody(raw []byte) ([]byte, error) {
+	dec := json.NewDecoder(strings.NewReader(string(raw)))
+	dec.UseNumber()
+	var body map[string]any
+	if err := dec.Decode(&body); err != nil {
+		return nil, fmt.Errorf("解析请求体: %w", err)
 	}
-	var msgs []aiwire.ChatMessage
-	for _, it := range in.Items {
-		out, err := respItemToMessages(it, msgs)
-		if err != nil {
-			return nil, err
-		}
-		msgs = out
-	}
-	return msgs, nil
+	body["store"] = false
+	return json.Marshal(body)
 }

-// respItemToMessages 追加一个 input item;连续 function_call 合并进同一 assistant 消息。
-func respItemToMessages(it aiwire.RespItem, msgs []aiwire.ChatMessage) ([]aiwire.ChatMessage, error) {
-	switch it.Type {
-	case "", "message":
-		if it.Content.HasUnsupported() {
-			return nil, ErrAiUnsupportedBlock
-		}
-		role := respRole(it.Role)
-		return append(msgs, aiwire.ChatMessage{Role: role, Content: respContentToIR(it.Content)}), nil
-	case "function_call":
-		call := aiwire.ToolCall{ID: it.CallID, Type: "function",
-			Function: aiwire.FunctionCall{Name: it.Name, Arguments: it.Arguments}}
-		if n := len(msgs); n > 0 && msgs[n-1].Role == "assistant" && len(msgs[n-1].ToolCalls) > 0 {
-			msgs[n-1].ToolCalls = append(msgs[n-1].ToolCalls, call)
-			return msgs, nil
-		}
-		return append(msgs, aiwire.ChatMessage{Role: "assistant", ToolCalls: []aiwire.ToolCall{call}}), nil
-	case "function_call_output":
-		return append(msgs, aiwire.ChatMessage{Role: "tool", ToolCallID: it.CallID,
-			Content: aiwire.NewTextContent(it.OutputText())}), nil
-	case "reasoning":
-		return msgs, nil // 推理块不回灌上游
-	default:
-		return nil, fmt.Errorf("不支持的 input item 类型 %q", it.Type)
+// RespPassthroughUsage 从直通响应提取用量;缺失时返回 nil(日志记零)。
+func RespPassthroughUsage(payload []byte) *aiwire.Usage {
+	var root struct {
+		Usage *aiwire.RespUsage `json:"usage"`
 	}
-}
-
-// respRole 归一 item 角色;developer 视为 system,缺省 user。
-func respRole(role string) string {
-	switch role {
-	case "system", "developer":
-		return "system"
-	case "assistant":
-		return "assistant"
-	default:
-		return "user"
-	}
-}
-
-// respContentToIR 把 Responses 内容转 IR;纯文本拍平,含图片时保留部件顺序。
-func respContentToIR(c aiwire.RespContent) aiwire.Content {
-	var parts []aiwire.ContentPart
-	hasImage := false
-	for _, p := range c.Parts {
-		switch p.Type {
-		case "input_image":
-			parts = append(parts, aiwire.ContentPart{Type: "image_url",
-				ImageURL: &aiwire.ImageURL{URL: p.ImageURL, Detail: p.Detail}})
-			hasImage = true
-		default:
-			if p.Text != "" {
-				parts = append(parts, aiwire.ContentPart{Type: "text", Text: p.Text})
-			}
-		}
-	}
-	if !c.IsArray || !hasImage {
-		return aiwire.NewTextContent(c.JoinText())
-	}
-	return aiwire.NewPartsContent(parts)
-}
-
-// respTools 把扁平工具定义还原为嵌套 Chat 形态;strict 无 OCI 对应,忽略。
-func respTools(tools []aiwire.RespTool) []aiwire.Tool {
-	if len(tools) == 0 {
+	if json.Unmarshal(payload, &root) != nil || root.Usage == nil {
 		return nil
 	}
-	out := make([]aiwire.Tool, 0, len(tools))
-	for _, t := range tools {
-		out = append(out, aiwire.Tool{Type: "function",
-			Function: aiwire.FunctionDef{Name: t.Name, Description: t.Description, Parameters: t.Parameters}})
+	usage := &aiwire.Usage{PromptTokens: root.Usage.InputTokens,
+		CompletionTokens: root.Usage.OutputTokens, TotalTokens: root.Usage.TotalTokens}
+	if cached := root.Usage.InputTokensDetails.CachedTokens; cached > 0 {
+		usage.PromptTokensDetails = &aiwire.PromptTokensDetails{CachedTokens: cached}
 	}
-	return out
+	return usage
 }

-// respToolChoice 转换 tool_choice:字符串直通;{type:function,name} 转嵌套;
-// allowed_tools 等对象形态降级 auto。
-func respToolChoice(raw json.RawMessage) json.RawMessage {
-	if len(raw) == 0 {
+// RespStreamCompletedUsage 从一行 SSE data JSON 中提取 response.completed 事件的
+// usage;非 completed 事件或解析失败返回 nil。流式直通逐行喂入,最后一次非 nil 生效。
+func RespStreamCompletedUsage(data []byte) *aiwire.Usage {
+	var ev struct {
+		Type     string          `json:"type"`
+		Response json.RawMessage `json:"response"`
+	}
+	if json.Unmarshal(data, &ev) != nil || ev.Type != "response.completed" || len(ev.Response) == 0 {
 		return nil
 	}
-	t := strings.TrimSpace(string(raw))
-	if strings.HasPrefix(t, "\"") {
-		return raw
-	}
-	var obj struct {
-		Type string `json:"type"`
-		Name string `json:"name"`
-	}
-	if err := json.Unmarshal(raw, &obj); err != nil || obj.Type != "function" || obj.Name == "" {
-		return json.RawMessage(`"auto"`)
-	}
-	out, _ := json.Marshal(map[string]any{"type": "function", "function": map[string]string{"name": obj.Name}})
-	return out
-}
-
-// respFormat 把 text.format 转成 response_format;text 形态无需显式指定。
-func respFormat(text *aiwire.RespText) *aiwire.ResponseFormat {
-	if text == nil || text.Format == nil {
-		return nil
-	}
-	switch text.Format.Type {
-	case "json_object":
-		return &aiwire.ResponseFormat{Type: "json_object"}
-	case "json_schema":
-		spec, _ := json.Marshal(map[string]any{
-			"name": text.Format.Name, "schema": json.RawMessage(text.Format.Schema), "strict": text.Format.Strict,
-		})
-		return &aiwire.ResponseFormat{Type: "json_schema", JSONSchema: spec}
-	default:
-		return nil
-	}
-}
-
-// IRRespToResponses 把 IR 非流式响应装配为 Response 对象。
-func IRRespToResponses(resp *aiwire.ChatResponse, id string, created int64) aiwire.Response {
-	out := aiwire.Response{ID: id, Object: "response", CreatedAt: created,
-		Status: "completed", Model: resp.Model, Output: []aiwire.RespOutItem{}, Store: false}
-	if len(resp.Choices) == 0 {
-		return out
-	}
-	choice := resp.Choices[0]
-	if text := choice.Message.Content.JoinText(); text != "" {
-		out.Output = append(out.Output, respMessageItem(id, 0, text))
-	}
-	for i, tc := range choice.Message.ToolCalls {
-		out.Output = append(out.Output, aiwire.RespOutItem{Type: "function_call",
-			ID: respItemID(id, "fc", len(out.Output)+i), Status: "completed",
-			CallID: tc.ID, Name: tc.Function.Name, Arguments: respArgs(tc.Function.Arguments)})
-	}
-	if choice.FinishReason == "length" {
-		out.Status = "incomplete"
-		out.IncompleteDetails = &aiwire.RespIncomplete{Reason: "max_output_tokens"}
-	}
-	out.Usage = respUsage(resp.Usage)
-	return out
-}
-
-func respMessageItem(respID string, idx int, text string) aiwire.RespOutItem {
-	return aiwire.RespOutItem{Type: "message", ID: respItemID(respID, "msg", idx),
-		Status: "completed", Role: "assistant",
-		Content: []aiwire.RespOutPart{{Type: "output_text", Text: text, Annotations: []any{}}}}
-}
-
-// respItemID 从响应 ID 派生确定性 item ID。
-func respItemID(respID, kind string, idx int) string {
-	return fmt.Sprintf("%s_%s_%d", kind, strings.TrimPrefix(respID, "resp_"), idx)
-}
-
-// respArgs 保证 arguments 是合法 JSON 字符串(空实参回退 {})。
-func respArgs(args string) string {
-	if strings.TrimSpace(args) == "" {
-		return "{}"
-	}
-	return args
-}
-
-// respUsage 把 IR 用量改写为 Responses 命名口径。
-func respUsage(u *aiwire.Usage) *aiwire.RespUsage {
-	if u == nil {
-		return nil
-	}
-	return &aiwire.RespUsage{InputTokens: u.PromptTokens, OutputTokens: u.CompletionTokens,
-		TotalTokens:        u.TotalTokens,
-		InputTokensDetails: aiwire.RespInDetails{CachedTokens: u.CachedTokens()}}
+	return RespPassthroughUsage(ev.Response)
 }
@@ -1,199 +0,0 @@
-package service
-
-import (
-	"strings"
-
-	"oci-portal/internal/aiwire"
-)
-
-// RespEvent 是一条 Responses SSE 语义事件(event 名 + data 载荷)。
-type RespEvent struct {
-	Event string
-	Data  map[string]any
-}
-
-// RespStream 把 IR chunk 流聚合为 Responses 语义事件序列。
-// 与 AnthStream 同构,但终态事件须携带完整 Response 快照,故全程缓冲文本与实参。
-type RespStream struct {
-	id      string
-	model   string
-	created int64
-	seq     int
-	started bool
-	items   []aiwire.RespOutItem
-	msgOpen bool
-	text    strings.Builder
-	tool    aiwire.RespOutItem
-	toolIdx int
-	tOpen   bool
-	args    strings.Builder
-	finish  string
-	usage   *aiwire.Usage
-}
-
-// NewRespStream 构造状态机;id 形如 resp_*,created 为响应时间戳。
-func NewRespStream(id, model string, created int64) *RespStream {
-	return &RespStream{id: id, model: model, created: created}
-}
-
-// Feed 消化一个上游 chunk,返回应立即下发的事件。
-func (s *RespStream) Feed(chunk aiwire.ChatChunk) []RespEvent {
-	var evs []RespEvent
-	if !s.started {
-		s.started = true
-		evs = append(evs, s.respEvent("response.created", "in_progress"),
-			s.respEvent("response.in_progress", "in_progress"))
-	}
-	if chunk.Usage != nil {
-		s.usage = chunk.Usage
-	}
-	if len(chunk.Choices) == 0 {
-		return evs
-	}
-	choice := chunk.Choices[0]
-	if choice.Delta.Content != "" {
-		evs = append(evs, s.feedText(choice.Delta.Content)...)
-	}
-	for _, tc := range choice.Delta.ToolCalls {
-		evs = append(evs, s.feedTool(tc)...)
-	}
-	if choice.FinishReason != nil {
-		s.finish = *choice.FinishReason
-	}
-	return evs
-}
-
-// Finish 关闭未闭合的块并产出终态事件(带完整 Response 快照)。
-func (s *RespStream) Finish() []RespEvent {
-	var evs []RespEvent
-	if !s.started {
-		s.started = true
-		evs = append(evs, s.respEvent("response.created", "in_progress"))
-	}
-	evs = append(evs, s.closeMsg()...)
-	evs = append(evs, s.closeTool()...)
-	status := "completed"
-	if s.finish == "length" {
-		status = "incomplete"
-	}
-	return append(evs, s.respEvent("response."+status, status))
-}
-
-// Usage 返回聚合到的用量(可能为 nil),供调用日志。
-func (s *RespStream) Usage() *aiwire.Usage { return s.usage }
-
-// ev 构造带自增 sequence_number 的事件。
-func (s *RespStream) ev(typ string, kv map[string]any) RespEvent {
-	s.seq++
-	data := map[string]any{"type": typ, "sequence_number": s.seq}
-	for k, v := range kv {
-		data[k] = v
-	}
-	return RespEvent{Event: typ, Data: data}
-}
-
-// respEvent 构造携带 Response 快照的生命周期事件。
-func (s *RespStream) respEvent(typ, status string) RespEvent {
-	return s.ev(typ, map[string]any{"response": s.snapshot(status)})
-}
-
-// snapshot 组装当前累计状态的 Response 对象。
-func (s *RespStream) snapshot(status string) aiwire.Response {
-	resp := aiwire.Response{ID: s.id, Object: "response", CreatedAt: s.created,
-		Status: status, Model: s.model, Store: false,
-		Output: append([]aiwire.RespOutItem{}, s.items...)}
-	resp.Usage = respUsage(s.usage)
-	if status == "incomplete" {
-		resp.IncompleteDetails = &aiwire.RespIncomplete{Reason: "max_output_tokens"}
-	}
-	return resp
-}
-
-// feedText 处理文本增量:必要时开 message item 与 content part。
-func (s *RespStream) feedText(delta string) []RespEvent {
-	evs := s.closeTool()
-	if !s.msgOpen {
-		s.msgOpen = true
-		s.text.Reset()
-		item := aiwire.RespOutItem{Type: "message", ID: s.itemID("msg"), Status: "in_progress",
-			Role: "assistant", Content: []aiwire.RespOutPart{}}
-		evs = append(evs, s.ev("response.output_item.added", map[string]any{
-			"output_index": len(s.items), "item": item}))
-		evs = append(evs, s.ev("response.content_part.added", map[string]any{
-			"item_id": item.ID, "output_index": len(s.items), "content_index": 0,
-			"part": aiwire.RespOutPart{Type: "output_text", Text: "", Annotations: []any{}}}))
-	}
-	s.text.WriteString(delta)
-	return append(evs, s.ev("response.output_text.delta", map[string]any{
-		"item_id": s.itemID("msg"), "output_index": len(s.items), "content_index": 0, "delta": delta}))
-}
-
-// closeMsg 闭合当前 message item(text done → part done → item done)。
-func (s *RespStream) closeMsg() []RespEvent {
-	if !s.msgOpen {
-		return nil
-	}
-	s.msgOpen = false
-	id, idx, full := s.itemID("msg"), len(s.items), s.text.String()
-	part := aiwire.RespOutPart{Type: "output_text", Text: full, Annotations: []any{}}
-	item := aiwire.RespOutItem{Type: "message", ID: id, Status: "completed",
-		Role: "assistant", Content: []aiwire.RespOutPart{part}}
-	evs := []RespEvent{
-		s.ev("response.output_text.done", map[string]any{"item_id": id, "output_index": idx, "content_index": 0, "text": full}),
-		s.ev("response.content_part.done", map[string]any{"item_id": id, "output_index": idx, "content_index": 0, "part": part}),
-		s.ev("response.output_item.done", map[string]any{"output_index": idx, "item": item}),
-	}
-	s.items = append(s.items, item)
-	return evs
-}
-
-// feedTool 处理工具调用增量:新 Index 先闭合旧调用再开新 item。
-func (s *RespStream) feedTool(tc aiwire.ToolCallDelta) []RespEvent {
-	evs := s.closeMsg()
-	if s.tOpen && tc.Index != s.toolIdx {
-		evs = append(evs, s.closeTool()...)
-	}
-	if !s.tOpen {
-		s.tOpen, s.toolIdx = true, tc.Index
-		s.args.Reset()
-		s.tool = aiwire.RespOutItem{Type: "function_call", ID: s.itemID("fc"), Status: "in_progress",
-			CallID: tc.ID, Name: tc.Function.Name}
-		evs = append(evs, s.ev("response.output_item.added", map[string]any{
-			"output_index": len(s.items), "item": s.tool}))
-	}
-	if tc.ID != "" && s.tool.CallID == "" {
-		s.tool.CallID = tc.ID
-	}
-	if tc.Function.Name != "" && s.tool.Name == "" {
-		s.tool.Name = tc.Function.Name
-	}
-	if tc.Function.Arguments == "" {
-		return evs
-	}
-	s.args.WriteString(tc.Function.Arguments)
-	return append(evs, s.ev("response.function_call_arguments.delta", map[string]any{
-		"item_id": s.tool.ID, "output_index": len(s.items), "delta": tc.Function.Arguments}))
-}
-
-// closeTool 闭合当前 function_call item(arguments done → item done)。
-func (s *RespStream) closeTool() []RespEvent {
-	if !s.tOpen {
-		return nil
-	}
-	s.tOpen = false
-	s.tool.Arguments = respArgs(s.args.String())
-	s.tool.Status = "completed"
-	idx := len(s.items)
-	evs := []RespEvent{
-		s.ev("response.function_call_arguments.done", map[string]any{
-			"item_id": s.tool.ID, "output_index": idx, "arguments": s.tool.Arguments}),
-		s.ev("response.output_item.done", map[string]any{"output_index": idx, "item": s.tool}),
-	}
-	s.items = append(s.items, s.tool)
-	return evs
-}
-
-// itemID 按当前 output 序号派生确定性 item ID。
-func (s *RespStream) itemID(kind string) string {
-	return respItemID(s.id, kind, len(s.items))
-}
@@ -17,47 +17,6 @@ func respReq(t *testing.T, raw string) aiwire.RespRequest {
 	return req
 }

-func TestResponsesToIR(t *testing.T) {
-	req := respReq(t, `{
-		"model": "meta.llama-3.3-70b-instruct",
-		"instructions": "你是助手",
-		"max_output_tokens": 128,
-		"input": [
-			{"role": "user", "content": "查天气"},
-			{"type": "function_call", "call_id": "call_1", "name": "get_weather", "arguments": "{\"city\":\"上海\"}"},
-			{"type": "function_call_output", "call_id": "call_1", "output": "{\"temp\":31}"},
-			{"type": "message", "role": "assistant", "content": [{"type": "output_text", "text": "31 度"}]}
-		],
-		"tools": [{"type": "function", "name": "get_weather", "description": "查天气", "parameters": {"type": "object"}, "strict": true}],
-		"tool_choice": {"type": "function", "name": "get_weather"},
-		"text": {"format": {"type": "json_schema", "name": "out", "schema": {"type": "object"}}}
-	}`)
-	ir, err := ResponsesToIR(req)
-	if err != nil {
-		t.Fatalf("ResponsesToIR: %v", err)
-	}
-	roles := make([]string, 0, len(ir.Messages))
-	for _, m := range ir.Messages {
-		roles = append(roles, m.Role)
-	}
-	want := []string{"system", "user", "assistant", "tool", "assistant"}
-	if strings.Join(roles, ",") != strings.Join(want, ",") {
-		t.Errorf("roles = %v, want %v", roles, want)
-	}
-	if ir.Messages[2].ToolCalls[0].ID != "call_1" || ir.Messages[3].ToolCallID != "call_1" {
-		t.Errorf("tool call 链接错误: %+v", ir.Messages)
-	}
-	if deref(ir.MaxTokens) != 128 || len(ir.Tools) != 1 || ir.Tools[0].Function.Name != "get_weather" {
-		t.Errorf("参数映射错误: max=%v tools=%+v", ir.MaxTokens, ir.Tools)
-	}
-	if !strings.Contains(string(ir.ToolChoice), `"function"`) || !strings.Contains(string(ir.ToolChoice), "get_weather") {
-		t.Errorf("tool_choice = %s", ir.ToolChoice)
-	}
-	if ir.ResponseFormat == nil || ir.ResponseFormat.Type != "json_schema" {
-		t.Errorf("response_format = %+v", ir.ResponseFormat)
-	}
-}
-
 func deref(p *int) int {
 	if p == nil {
 		return 0
@@ -65,112 +24,124 @@ func deref(p *int) int {
 	return *p
 }

-func TestResponsesToIRRejects(t *testing.T) {
-	cases := []struct{ name, raw string }{
-		{"previous_response_id", `{"model":"m","input":"hi","previous_response_id":"resp_x"}`},
-		{"conversation", `{"model":"m","input":"hi","conversation":{"id":"conv_1"}}`},
-		{"background", `{"model":"m","input":"hi","background":true}`},
-		{"builtin tool", `{"model":"m","input":"hi","tools":[{"type":"web_search"}]}`},
-		{"file_id image", `{"model":"m","input":[{"role":"user","content":[{"type":"input_image","file_id":"file_1"}]}]}`},
-		{"unknown item", `{"model":"m","input":[{"type":"item_reference","id":"x"}]}`},
+func TestRespServerTools(t *testing.T) {
+	tests := []struct {
+		name  string
+		tools []aiwire.RespTool
+		want  bool
+	}{
+		{"web_search", []aiwire.RespTool{{Type: "web_search"}}, true},
+		{"x_search混用", []aiwire.RespTool{{Type: "function", Name: "f"}, {Type: "x_search"}}, true},
+		{"仅function", []aiwire.RespTool{{Type: "function", Name: "f"}}, false},
+		{"空", nil, false},
+		{"其他类型", []aiwire.RespTool{{Type: "code_interpreter"}}, false},
 	}
-	for _, c := range cases {
-		if _, err := ResponsesToIR(respReq(t, c.raw)); err == nil {
-			t.Errorf("%s 应被拒绝", c.name)
-		}
+	for _, test := range tests {
+		t.Run(test.name, func(t *testing.T) {
+			if got := RespServerTools(test.tools); got != test.want {
+				t.Fatalf("got %v, want %v", got, test.want)
+			}
+		})
 	}
-	// input 字符串形态 + store/reasoning 忽略项不报错
-	req := respReq(t, `{"model":"m","input":"你好","store":false,"reasoning":{"effort":"low"},"parallel_tool_calls":true}`)
-	ir, err := ResponsesToIR(req)
-	if err != nil || len(ir.Messages) != 1 || ir.Messages[0].Role != "user" || ir.ReasoningEffort != "low" {
-		t.Errorf("字符串 input = %+v, %v", ir.Messages, err)
+}
+
+func TestRespPassthroughValidate(t *testing.T) {
+	prev := "resp_1"
+	bg := true
+	tests := []struct {
+		name    string
+		req     aiwire.RespRequest
+		wantErr bool
+	}{
+		{"合法", aiwire.RespRequest{Model: "m", Tools: []aiwire.RespTool{{Type: "web_search"}}}, false},
+		{"混用function", aiwire.RespRequest{Model: "m", Tools: []aiwire.RespTool{{Type: "web_search"}, {Type: "function", Name: "f"}}}, false},
+		{"缺model", aiwire.RespRequest{Tools: []aiwire.RespTool{{Type: "web_search"}}}, true},
+		{"工具加流式拒绝", aiwire.RespRequest{Model: "m", Stream: true, Tools: []aiwire.RespTool{{Type: "web_search"}}}, true},
+		{"无工具流式放行", aiwire.RespRequest{Model: "m", Stream: true}, false},
+		{"function工具流式放行", aiwire.RespRequest{Model: "m", Stream: true, Tools: []aiwire.RespTool{{Type: "function", Name: "f"}}}, false},
+		{"有状态拒绝", aiwire.RespRequest{Model: "m", PreviousResponseID: prev}, true},
+		{"background拒绝", aiwire.RespRequest{Model: "m", Background: &bg}, true},
+		{"未知工具拒绝", aiwire.RespRequest{Model: "m", Tools: []aiwire.RespTool{{Type: "web_search"}, {Type: "mcp"}}}, true},
 	}
-	// input_image(url 形态)放行并保留图文顺序
-	req2 := respReq(t, `{"model":"m","input":[{"role":"user","content":[{"type":"input_text","text":"看图"},{"type":"input_image","image_url":"https://x/1.png","detail":"low"}]}]}`)
-	ir2, err := ResponsesToIR(req2)
+	for _, test := range tests {
+		t.Run(test.name, func(t *testing.T) {
+			err := RespPassthroughValidate(test.req)
+			if (err != nil) != test.wantErr {
+				t.Fatalf("err = %v, wantErr %v", err, test.wantErr)
+			}
+		})
+	}
+}
+
+func TestRespPassthroughBody(t *testing.T) {
+	raw := []byte(`{"model":"m","input":"hi","stream":true,"store":true,"max_output_tokens":128,"custom_field":{"a":1.5}}`)
+	out, err := RespPassthroughBody(raw)
 	if err != nil {
-		t.Fatalf("input_image 应放行: %v", err)
+		t.Fatalf("RespPassthroughBody: %v", err)
 	}
-	parts := ir2.Messages[0].Content.Parts
-	if len(parts) != 2 || parts[1].Type != "image_url" || parts[1].ImageURL == nil || parts[1].ImageURL.URL != "https://x/1.png" {
-		t.Errorf("parts = %+v", parts)
+	var body map[string]any
+	if err := json.Unmarshal(out, &body); err != nil {
+		t.Fatalf("unmarshal out: %v", err)
+	}
+	if body["store"] != false {
+		t.Errorf("store 应强制 false, got %v", body["store"])
+	}
+	if body["stream"] != true {
+		t.Errorf("stream 应原样保留, got %v", body["stream"])
+	}
+	if string(out) == "" || !strings.Contains(string(out), `"max_output_tokens":128`) {
+		t.Errorf("数值字段应保真: %s", out)
+	}
+	if !strings.Contains(string(out), `"custom_field"`) {
+		t.Errorf("未知字段应保留: %s", out)
+	}
+	if _, err := RespPassthroughBody([]byte("not-json")); err == nil {
+		t.Error("非法 JSON 应报错")
 	}
 }

-func TestIRRespToResponses(t *testing.T) {
-	resp := &aiwire.ChatResponse{Model: "m", Choices: []aiwire.Choice{{
-		Message: aiwire.ChatMessage{Role: "assistant", Content: aiwire.NewTextContent("你好"),
-			ToolCalls: []aiwire.ToolCall{{ID: "call_9", Type: "function",
-				Function: aiwire.FunctionCall{Name: "f", Arguments: ""}}}},
-		FinishReason: "length",
-	}}, Usage: &aiwire.Usage{PromptTokens: 10, CompletionTokens: 5, TotalTokens: 15}}
-	out := IRRespToResponses(resp, "resp_abc", 1751966400)
-	if out.Status != "incomplete" || out.IncompleteDetails == nil || out.IncompleteDetails.Reason != "max_output_tokens" {
-		t.Errorf("length 应映射 incomplete: %+v", out)
+func TestRespPassthroughUsage(t *testing.T) {
+	tests := []struct {
+		name           string
+		payload        string
+		wantNil        bool
+		prompt, cached int
+	}{
+		{"完整", `{"usage":{"input_tokens":9,"output_tokens":3,"total_tokens":12,"input_tokens_details":{"cached_tokens":5}}}`, false, 9, 5},
+		{"无细分", `{"usage":{"input_tokens":9,"output_tokens":3,"total_tokens":12}}`, false, 9, 0},
+		{"无usage", `{"id":"resp_1"}`, true, 0, 0},
+		{"非法JSON", `xx`, true, 0, 0},
 	}
-	if len(out.Output) != 2 || out.Output[0].Type != "message" || out.Output[1].Type != "function_call" {
-		t.Fatalf("output = %+v", out.Output)
-	}
-	if out.Output[0].Content[0].Text != "你好" || out.Output[1].CallID != "call_9" || out.Output[1].Arguments != "{}" {
-		t.Errorf("item 装配错误: %+v", out.Output)
-	}
-	if out.Usage.InputTokens != 10 || out.Usage.OutputTokens != 5 || out.Store {
-		t.Errorf("usage/store 错误: %+v", out)
+	for _, test := range tests {
+		t.Run(test.name, func(t *testing.T) {
+			usage := RespPassthroughUsage([]byte(test.payload))
+			if (usage == nil) != test.wantNil {
+				t.Fatalf("usage = %v, wantNil %v", usage, test.wantNil)
+			}
+			if usage == nil {
+				return
+			}
+			if usage.PromptTokens != test.prompt || usage.CachedTokens() != test.cached {
+				t.Fatalf("prompt=%d cached=%d, want %d/%d", usage.PromptTokens, usage.CachedTokens(), test.prompt, test.cached)
+			}
+		})
 	}
 }

-func chunkText(text string) aiwire.ChatChunk {
-	return aiwire.ChatChunk{Choices: []aiwire.ChunkChoice{{Delta: aiwire.Delta{Content: text}}}}
-}
-
-func TestRespStreamTextAndTool(t *testing.T) {
-	st := NewRespStream("resp_x", "m", 1751966400)
-	var types []string
-	collect := func(evs []RespEvent) {
-		for _, ev := range evs {
-			types = append(types, ev.Event)
+// TestRespStreamCompletedUsage 断言流式 usage 只从 completed 事件提取。
+func TestRespStreamCompletedUsage(t *testing.T) {
+	completed := []byte(`{"type":"response.completed","response":{"usage":{"input_tokens":10,"output_tokens":5,"total_tokens":15,"input_tokens_details":{"cached_tokens":4}}}}`)
+	if u := RespStreamCompletedUsage(completed); u == nil || u.PromptTokens != 10 || u.CompletionTokens != 5 ||
+		u.PromptTokensDetails == nil || u.PromptTokensDetails.CachedTokens != 4 {
+		t.Fatalf("completed 事件 usage 解析失败: %+v", u)
+	}
+	for _, data := range []string{
+		`{"type":"response.output_text.delta","delta":"hi"}`,
+		`{"type":"response.completed"}`,
+		`not-json`,
+	} {
+		if u := RespStreamCompletedUsage([]byte(data)); u != nil {
+			t.Fatalf("非 completed 或缺 response 应返回 nil: %s", data)
 		}
 	}
-	collect(st.Feed(chunkText("你")))
-	collect(st.Feed(chunkText("好")))
-	collect(st.Feed(aiwire.ChatChunk{Choices: []aiwire.ChunkChoice{{Delta: aiwire.Delta{
-		ToolCalls: []aiwire.ToolCallDelta{{Index: 0, ID: "call_1", Function: aiwire.FunctionCallDelta{Name: "f", Arguments: "{\"a\""}}}}}}}))
-	collect(st.Feed(aiwire.ChatChunk{Choices: []aiwire.ChunkChoice{{Delta: aiwire.Delta{
-		ToolCalls: []aiwire.ToolCallDelta{{Index: 0, Function: aiwire.FunctionCallDelta{Arguments: ":1}"}}}}}},
-		Usage: &aiwire.Usage{PromptTokens: 3, CompletionTokens: 7, TotalTokens: 10}}))
-	finish := st.Finish()
-	collect(finish)
-	want := []string{
-		"response.created", "response.in_progress",
-		"response.output_item.added", "response.content_part.added", "response.output_text.delta",
-		"response.output_text.delta",
-		"response.output_text.done", "response.content_part.done", "response.output_item.done",
-		"response.output_item.added", "response.function_call_arguments.delta",
-		"response.function_call_arguments.delta",
-		"response.function_call_arguments.done", "response.output_item.done",
-		"response.completed",
-	}
-	if strings.Join(types, "\n") != strings.Join(want, "\n") {
-		t.Errorf("事件序列 =\n%s\nwant\n%s", strings.Join(types, "\n"), strings.Join(want, "\n"))
-	}
-	last := finish[len(finish)-1].Data
-	resp := last["response"].(aiwire.Response)
-	if len(resp.Output) != 2 || resp.Output[0].Content[0].Text != "你好" || resp.Output[1].Arguments != "{\"a\":1}" {
-		t.Errorf("终态快照 = %+v", resp.Output)
-	}
-	if resp.Usage == nil || resp.Usage.TotalTokens != 10 {
-		t.Errorf("终态 usage = %+v", resp.Usage)
-	}
-	// sequence_number 严格递增
-	if seq, ok := last["sequence_number"].(int); !ok || seq != len(want) {
-		t.Errorf("最终 sequence_number = %v, want %d", last["sequence_number"], len(want))
-	}
-}
-
-func TestRespStreamEmpty(t *testing.T) {
-	st := NewRespStream("resp_e", "m", 1)
-	evs := st.Finish()
-	if len(evs) != 2 || evs[0].Event != "response.created" || evs[1].Event != "response.completed" {
-		t.Errorf("空流事件 = %+v", evs)
-	}
 }
@@ -0,0 +1,411 @@
+package service
+
+import (
+	"encoding/json"
+	"fmt"
+	"strings"
+
+	"oci-portal/internal/aiwire"
+)
+
+// Anthropic Messages ↔ OCI OpenAI 兼容面(/actions/v1/responses)直通转换。
+// typed chat 面剔除后 Messages 入口的唯一上游通路。语义损失(README 已披露):
+// stop_sequences / top_k / metadata / thinking 无对应字段,忽略;上游 reasoning
+// 输出项与增量事件丢弃(Anthropic thinking 块含签名语义,不伪造)。
+
+// ErrAiUnsupportedBlock 表示请求含网关无法承接的内容块(仅支持文本/图片/工具块)。
+var ErrAiUnsupportedBlock = fmt.Errorf("暂不支持文本与图片以外的内容块")
+
+// AnthropicToResponsesBody 把 Messages 请求转为直通 body(强制 store:false)。
+func AnthropicToResponsesBody(req aiwire.MessagesRequest) ([]byte, error) {
+	input, err := anthInputItems(req.Messages)
+	if err != nil {
+		return nil, err
+	}
+	body := map[string]any{"model": req.Model, "max_output_tokens": req.MaxTokens,
+		"input": input, "store": false}
+	if sys := req.SystemText(); sys != "" {
+		body["instructions"] = sys
+	}
+	if req.Temperature != nil {
+		body["temperature"] = *req.Temperature
+	}
+	if req.TopP != nil {
+		body["top_p"] = *req.TopP
+	}
+	if tools := anthRespTools(req.Tools); tools != nil {
+		body["tools"] = tools
+	}
+	if tc := anthRespToolChoice(req.ToolChoice); tc != nil {
+		body["tool_choice"] = tc
+	}
+	if req.OutputConfig != nil && req.OutputConfig.Effort != "" {
+		body["reasoning"] = map[string]string{"effort": strings.ToLower(req.OutputConfig.Effort)}
+	}
+	if req.Stream {
+		body["stream"] = true
+	}
+	return json.Marshal(body)
+}
+
+// anthInputItems 把消息序列展开为 Responses input 项:tool_use / tool_result 为
+// 独立 function_call / function_call_output 项,其余聚合为 message 项;遇独立项时
+// 先冲刷已聚合部件,保持块间相对顺序。
+func anthInputItems(messages []aiwire.AnthMessage) ([]any, error) {
+	var items []any
+	for _, m := range messages {
+		var parts []map[string]any
+		flush := func() {
+			if len(parts) > 0 {
+				items = append(items, map[string]any{"role": m.Role, "content": parts})
+				parts = nil
+			}
+		}
+		for _, b := range m.Content.AllBlocks() {
+			item, part, err := anthBlockItem(m.Role, b)
+			if err != nil {
+				return nil, err
+			}
+			if part != nil {
+				parts = append(parts, part)
+			}
+			if item != nil {
+				flush()
+				items = append(items, item)
+			}
+		}
+		flush()
+	}
+	return items, nil
+}
+
+// anthBlockItem 把单个内容块转为独立项或消息部件(thinking 忽略,未知块拒绝)。
+func anthBlockItem(role string, b aiwire.AnthBlock) (item, part map[string]any, err error) {
+	switch b.Type {
+	case "text":
+		return nil, anthTextPart(role, b.Text), nil
+	case "image":
+		part, err = anthImageInput(b.Source)
+		return nil, part, err
+	case "tool_use":
+		return map[string]any{"type": "function_call", "call_id": b.ID,
+			"name": b.Name, "arguments": string(b.Input)}, nil, nil
+	case "tool_result":
+		return map[string]any{"type": "function_call_output",
+			"call_id": b.ToolUseID, "output": b.ResultText()}, nil, nil
+	case "thinking", "redacted_thinking":
+		return nil, nil, nil
+	default:
+		return nil, nil, ErrAiUnsupportedBlock
+	}
+}
+
+// anthTextPart 按角色选择 Responses 文本部件类型(assistant 历史为 output_text)。
+func anthTextPart(role, text string) map[string]any {
+	if role == "assistant" {
+		return map[string]any{"type": "output_text", "text": text}
+	}
+	return map[string]any{"type": "input_text", "text": text}
+}
+
+// anthImageInput 把 Anthropic image source 转为 Responses input_image 部件。
+func anthImageInput(source json.RawMessage) (map[string]any, error) {
+	var src struct {
+		Type      string `json:"type"`
+		MediaType string `json:"media_type"`
+		Data      string `json:"data"`
+		URL       string `json:"url"`
+	}
+	if err := json.Unmarshal(source, &src); err != nil {
+		return nil, fmt.Errorf("image source 解析失败: %w", err)
+	}
+	switch src.Type {
+	case "base64":
+		if src.MediaType == "" || src.Data == "" {
+			return nil, fmt.Errorf("image source 缺少 media_type 或 data")
+		}
+		return map[string]any{"type": "input_image",
+			"image_url": "data:" + src.MediaType + ";base64," + src.Data}, nil
+	case "url":
+		if src.URL == "" {
+			return nil, fmt.Errorf("image source 缺少 url")
+		}
+		return map[string]any{"type": "input_image", "image_url": src.URL}, nil
+	}
+	return nil, fmt.Errorf("不支持的 image source 类型 %q", src.Type)
+}
+
+func anthRespTools(tools []aiwire.AnthTool) []map[string]any {
+	if len(tools) == 0 {
+		return nil
+	}
+	out := make([]map[string]any, 0, len(tools))
+	for _, t := range tools {
+		out = append(out, map[string]any{"type": "function", "name": t.Name,
+			"description": t.Description, "parameters": t.InputSchema})
+	}
+	return out
+}
+
+// anthRespToolChoice 映射 {type:auto|any|tool|none,name} → Responses 形态。
+func anthRespToolChoice(raw json.RawMessage) any {
+	if len(raw) == 0 {
+		return nil
+	}
+	var tc struct {
+		Type string `json:"type"`
+		Name string `json:"name"`
+	}
+	if json.Unmarshal(raw, &tc) != nil {
+		return nil
+	}
+	switch tc.Type {
+	case "auto":
+		return "auto"
+	case "any":
+		return "required"
+	case "none":
+		return "none"
+	case "tool":
+		return map[string]string{"type": "function", "name": tc.Name}
+	}
+	return nil
+}
+
+// respPayload 是直通响应中本转换关心的子集(未知字段忽略)。
+type respPayload struct {
+	Model             string             `json:"model"`
+	Status            string             `json:"status"`
+	IncompleteDetails *respIncomplete    `json:"incomplete_details"`
+	Output            []respOutputItem   `json:"output"`
+	Usage             *aiwire.RespUsage  `json:"usage"`
+	Error             *map[string]string `json:"error"`
+}
+
+type respIncomplete struct {
+	Reason string `json:"reason"`
+}
+
+type respOutputItem struct {
+	Type      string `json:"type"`
+	CallID    string `json:"call_id"`
+	Name      string `json:"name"`
+	Arguments string `json:"arguments"`
+	Content   []struct {
+		Type string `json:"type"`
+		Text string `json:"text"`
+	} `json:"content"`
+}
+
+// ResponsesToAnthropic 把直通非流式响应转为 Anthropic Messages 响应。
+func ResponsesToAnthropic(payload []byte, msgID string) (*aiwire.MessagesResponse, error) {
+	var resp respPayload
+	if err := json.Unmarshal(payload, &resp); err != nil {
+		return nil, fmt.Errorf("解析上游响应: %w", err)
+	}
+	out := &aiwire.MessagesResponse{ID: msgID, Type: "message", Role: "assistant",
+		Model: resp.Model, Content: []aiwire.AnthBlock{}, StopReason: "end_turn"}
+	for _, item := range resp.Output {
+		switch item.Type {
+		case "message":
+			for _, part := range item.Content {
+				if part.Type == "output_text" && part.Text != "" {
+					out.Content = append(out.Content, aiwire.AnthBlock{Type: "text", Text: part.Text})
+				}
+			}
+		case "function_call":
+			out.Content = append(out.Content, aiwire.AnthBlock{Type: "tool_use", ID: item.CallID,
+				Name: item.Name, Input: argsToJSON(item.Arguments)})
+			out.StopReason = "tool_use"
+		}
+	}
+	if resp.Status == "incomplete" && resp.IncompleteDetails != nil &&
+		resp.IncompleteDetails.Reason == "max_output_tokens" {
+		out.StopReason = "max_tokens"
+	}
+	out.Usage = anthUsageFromResp(resp.Usage)
+	return out, nil
+}
+
+func anthUsageFromResp(u *aiwire.RespUsage) aiwire.AnthUsage {
+	if u == nil {
+		return aiwire.AnthUsage{}
+	}
+	return aiwire.AnthUsage{InputTokens: u.InputTokens, OutputTokens: u.OutputTokens,
+		CacheReadInputTokens: u.InputTokensDetails.CachedTokens}
+}
+
+// argsToJSON 保证 tool_use.input 是合法 JSON 对象(模型可能产出非法片段)。
+func argsToJSON(args string) json.RawMessage {
+	trimmed := strings.TrimSpace(args)
+	if trimmed == "" {
+		return json.RawMessage(`{}`)
+	}
+	if json.Valid([]byte(trimmed)) {
+		return json.RawMessage(trimmed)
+	}
+	b, _ := json.Marshal(map[string]string{"_raw": args})
+	return b
+}
+
+// ---- Anthropic 流式桥:Responses SSE 事件 → Anthropic 事件序列 ----
+
+// AnthEvent 是一条待写出的 Anthropic SSE 事件。
+type AnthEvent struct {
+	Event string
+	Data  any
+}
+
+// AnthRespBridge 把直通 SSE 事件流桥接为 Anthropic 事件序列:
+// message_start → content_block_start/delta/stop(text 与 tool_use 分块)→ message_delta → message_stop。
+// reasoning 系列事件丢弃。
+type AnthRespBridge struct {
+	id, model   string
+	started     bool
+	blockOpen   bool
+	blockIsTool bool
+	blockIndex  int
+	stopReason  string
+	usage       aiwire.AnthUsage
+}
+
+// NewAnthRespBridge 构造桥;id 为响应消息 ID。
+func NewAnthRespBridge(id, model string) *AnthRespBridge {
+	return &AnthRespBridge{id: id, model: model, blockIndex: -1, stopReason: "end_turn"}
+}
+
+// respStreamEvent 是直通 SSE data JSON 中桥关心的子集。
+type respStreamEvent struct {
+	Type     string          `json:"type"`
+	Delta    string          `json:"delta"`
+	Item     *respOutputItem `json:"item"`
+	Response *respPayload    `json:"response"`
+}
+
+// Feed 消费一行 SSE data JSON,返回应立即写出的事件。
+func (st *AnthRespBridge) Feed(data []byte) []AnthEvent {
+	var ev respStreamEvent
+	if json.Unmarshal(data, &ev) != nil {
+		return nil
+	}
+	var events []AnthEvent
+	if !st.started {
+		st.started = true
+		events = append(events, st.startEvent())
+	}
+	switch ev.Type {
+	case "response.output_item.added":
+		if ev.Item != nil && ev.Item.Type == "function_call" {
+			events = append(events, st.openBlock(true, ev.Item.CallID, ev.Item.Name)...)
+			st.stopReason = "tool_use"
+		}
+	case "response.output_text.delta":
+		events = append(events, st.textDelta(ev.Delta)...)
+	case "response.function_call_arguments.delta":
+		events = append(events, st.argsDelta(ev.Delta)...)
+	case "response.completed", "response.incomplete", "response.failed":
+		st.finishFrom(ev.Response)
+	}
+	return events
+}
+
+func (st *AnthRespBridge) textDelta(delta string) []AnthEvent {
+	var events []AnthEvent
+	if !st.blockOpen || st.blockIsTool {
+		events = append(events, st.openBlock(false, "", "")...)
+	}
+	events = append(events, AnthEvent{Event: "content_block_delta", Data: map[string]any{
+		"type": "content_block_delta", "index": st.blockIndex,
+		"delta": map[string]string{"type": "text_delta", "text": delta},
+	}})
+	return events
+}
+
+func (st *AnthRespBridge) argsDelta(delta string) []AnthEvent {
+	if !st.blockOpen || !st.blockIsTool {
+		return nil
+	}
+	return []AnthEvent{{Event: "content_block_delta", Data: map[string]any{
+		"type": "content_block_delta", "index": st.blockIndex,
+		"delta": map[string]string{"type": "input_json_delta", "partial_json": delta},
+	}}}
+}
+
+// finishFrom 记录终态:usage 与 stop_reason(max_output_tokens 截断 → max_tokens)。
+func (st *AnthRespBridge) finishFrom(resp *respPayload) {
+	if resp == nil {
+		return
+	}
+	st.usage = anthUsageFromResp(resp.Usage)
+	if resp.Status == "incomplete" && resp.IncompleteDetails != nil &&
+		resp.IncompleteDetails.Reason == "max_output_tokens" {
+		st.stopReason = "max_tokens"
+	}
+}
+
+func (st *AnthRespBridge) startEvent() AnthEvent {
+	return AnthEvent{Event: "message_start", Data: map[string]any{
+		"type": "message_start",
+		"message": map[string]any{
+			"id": st.id, "type": "message", "role": "assistant", "model": st.model,
+			"content": []any{}, "stop_reason": nil,
+			"usage": map[string]int{"input_tokens": 0, "output_tokens": 0},
+		},
+	}}
+}
+
+// openBlock 关闭当前块并打开新块(text 或 tool_use)。
+func (st *AnthRespBridge) openBlock(isTool bool, toolID, toolName string) []AnthEvent {
+	var events []AnthEvent
+	if st.blockOpen {
+		events = append(events, st.closeBlockEvent())
+	}
+	st.blockOpen, st.blockIsTool = true, isTool
+	st.blockIndex++
+	block := map[string]any{"type": "text", "text": ""}
+	if isTool {
+		block = map[string]any{"type": "tool_use", "id": toolID, "name": toolName, "input": map[string]any{}}
+	}
+	events = append(events, AnthEvent{Event: "content_block_start", Data: map[string]any{
+		"type": "content_block_start", "index": st.blockIndex, "content_block": block,
+	}})
+	return events
+}
+
+func (st *AnthRespBridge) closeBlockEvent() AnthEvent {
+	return AnthEvent{Event: "content_block_stop", Data: map[string]any{
+		"type": "content_block_stop", "index": st.blockIndex,
+	}}
+}
+
+// Finish 在上游流结束后收尾:关块 → message_delta(stop_reason+usage)→ message_stop。
+func (st *AnthRespBridge) Finish() []AnthEvent {
+	var events []AnthEvent
+	if !st.started {
+		st.started = true
+		events = append(events, st.startEvent())
+	}
+	if st.blockOpen {
+		events = append(events, st.closeBlockEvent())
+		st.blockOpen = false
+	}
+	usage := map[string]int{"output_tokens": st.usage.OutputTokens}
+	if st.usage.InputTokens > 0 {
+		usage["input_tokens"] = st.usage.InputTokens
+	}
+	if st.usage.CacheReadInputTokens > 0 {
+		usage["cache_read_input_tokens"] = st.usage.CacheReadInputTokens
+	}
+	events = append(events,
+		AnthEvent{Event: "message_delta", Data: map[string]any{
+			"type":  "message_delta",
+			"delta": map[string]any{"stop_reason": st.stopReason, "stop_sequence": nil},
+			"usage": usage,
+		}},
+		AnthEvent{Event: "message_stop", Data: map[string]any{"type": "message_stop"}},
+	)
+	return events
+}
+
+// Usage 返回聚合到的用量(供调用日志)。
+func (st *AnthRespBridge) Usage() aiwire.AnthUsage { return st.usage }
@@ -0,0 +1,173 @@
+package service
+
+import (
+	"encoding/json"
+	"strings"
+	"testing"
+
+	"oci-portal/internal/aiwire"
+)
+
+func mustAnthReq(t *testing.T, raw string) aiwire.MessagesRequest {
+	t.Helper()
+	var req aiwire.MessagesRequest
+	if err := json.Unmarshal([]byte(raw), &req); err != nil {
+		t.Fatalf("解析请求: %v", err)
+	}
+	return req
+}
+
+// TestAnthropicToResponsesBody 断言 system/消息/工具/effort 的直通装配与 store 强制。
+func TestAnthropicToResponsesBody(t *testing.T) {
+	req := mustAnthReq(t, `{
+		"model": "xai.grok-4.3", "max_tokens": 128, "system": "你是助手", "stream": true,
+		"temperature": 0.5,
+		"messages": [
+			{"role": "user", "content": "东京天气?"},
+			{"role": "assistant", "content": [
+				{"type": "text", "text": "查询中"},
+				{"type": "tool_use", "id": "t1", "name": "get_weather", "input": {"city": "东京"}}
+			]},
+			{"role": "user", "content": [
+				{"type": "tool_result", "tool_use_id": "t1", "content": "晴 25 度"},
+				{"type": "text", "text": "继续"}
+			]}
+		],
+		"tools": [{"name": "get_weather", "description": "查天气", "input_schema": {"type": "object"}}],
+		"tool_choice": {"type": "any"},
+		"output_config": {"effort": "HIGH"}
+	}`)
+	payload, err := AnthropicToResponsesBody(req)
+	if err != nil {
+		t.Fatalf("AnthropicToResponsesBody: %v", err)
+	}
+	var body map[string]any
+	if err := json.Unmarshal(payload, &body); err != nil {
+		t.Fatalf("unmarshal: %v", err)
+	}
+	if body["model"] != "xai.grok-4.3" || body["max_output_tokens"] != float64(128) ||
+		body["instructions"] != "你是助手" || body["store"] != false || body["stream"] != true {
+		t.Fatalf("顶层字段装配错误: %v", body)
+	}
+	if body["tool_choice"] != "required" {
+		t.Fatalf("tool_choice = %v, want required", body["tool_choice"])
+	}
+	reasoning, _ := body["reasoning"].(map[string]any)
+	if reasoning["effort"] != "high" {
+		t.Fatalf("effort = %v, want high(小写透传)", reasoning)
+	}
+	input, _ := body["input"].([]any)
+	// user 消息、assistant 文本消息、function_call、function_call_output、末条 user 文本
+	if len(input) != 5 {
+		t.Fatalf("input 项数 = %d, want 5: %s", len(input), payload)
+	}
+	kinds := make([]string, 0, len(input))
+	for _, it := range input {
+		m := it.(map[string]any)
+		if ty, ok := m["type"].(string); ok {
+			kinds = append(kinds, ty)
+		} else {
+			kinds = append(kinds, "message/"+m["role"].(string))
+		}
+	}
+	want := "message/user,message/assistant,function_call,function_call_output,message/user"
+	if got := strings.Join(kinds, ","); got != want {
+		t.Fatalf("input 顺序 = %s, want %s", got, want)
+	}
+	if !strings.Contains(string(payload), `"output_text"`) {
+		t.Fatalf("assistant 历史文本应为 output_text 部件: %s", payload)
+	}
+}
+
+// TestAnthropicToResponsesBodyRejects 断言不支持的内容块拒绝与图片装配。
+func TestAnthropicToResponsesBodyRejects(t *testing.T) {
+	bad := mustAnthReq(t, `{"model":"m","max_tokens":8,"messages":[
+		{"role":"user","content":[{"type":"document","source":{}}]}]}`)
+	if _, err := AnthropicToResponsesBody(bad); err == nil {
+		t.Fatal("document 块应拒绝")
+	}
+	img := mustAnthReq(t, `{"model":"m","max_tokens":8,"messages":[
+		{"role":"user","content":[{"type":"image","source":{"type":"base64","media_type":"image/png","data":"QUJD"}}]}]}`)
+	payload, err := AnthropicToResponsesBody(img)
+	if err != nil || !strings.Contains(string(payload), "data:image/png;base64,QUJD") {
+		t.Fatalf("图片应转 data URI: %v %s", err, payload)
+	}
+}
+
+// TestResponsesToAnthropic 断言输出块、stop_reason 与 usage 的回转。
+func TestResponsesToAnthropic(t *testing.T) {
+	payload := []byte(`{"model":"xai.grok-4.3","status":"completed","output":[
+		{"type":"reasoning","summary":[]},
+		{"type":"message","content":[{"type":"output_text","text":"你好"}]},
+		{"type":"function_call","call_id":"c1","name":"get_weather","arguments":"{\"city\":\"东京\"}"}],
+		"usage":{"input_tokens":10,"output_tokens":5,"total_tokens":15,"input_tokens_details":{"cached_tokens":4}}}`)
+	out, err := ResponsesToAnthropic(payload, "msg_1")
+	if err != nil {
+		t.Fatalf("ResponsesToAnthropic: %v", err)
+	}
+	if len(out.Content) != 2 || out.Content[0].Type != "text" || out.Content[0].Text != "你好" {
+		t.Fatalf("content 装配错误: %+v", out.Content)
+	}
+	if out.Content[1].Type != "tool_use" || out.Content[1].ID != "c1" || string(out.Content[1].Input) != `{"city":"东京"}` {
+		t.Fatalf("tool_use 装配错误: %+v", out.Content[1])
+	}
+	if out.StopReason != "tool_use" {
+		t.Fatalf("stop_reason = %s, want tool_use", out.StopReason)
+	}
+	if out.Usage.InputTokens != 10 || out.Usage.OutputTokens != 5 || out.Usage.CacheReadInputTokens != 4 {
+		t.Fatalf("usage = %+v", out.Usage)
+	}
+
+	trunc := []byte(`{"model":"m","status":"incomplete","incomplete_details":{"reason":"max_output_tokens"},
+		"output":[{"type":"message","content":[{"type":"output_text","text":"半"}]}]}`)
+	out2, err := ResponsesToAnthropic(trunc, "msg_2")
+	if err != nil || out2.StopReason != "max_tokens" {
+		t.Fatalf("截断 stop_reason = %v, %v", out2, err)
+	}
+}
+
+func bridgeEventTypes(events []AnthEvent) string {
+	kinds := make([]string, 0, len(events))
+	for _, ev := range events {
+		kinds = append(kinds, ev.Event)
+	}
+	return strings.Join(kinds, ",")
+}
+
+// TestAnthRespBridge 断言流桥:文本增量、工具调用切块、reasoning 丢弃与收尾事件。
+func TestAnthRespBridge(t *testing.T) {
+	st := NewAnthRespBridge("msg_1", "m1")
+	var events []AnthEvent
+	feed := func(lines ...string) {
+		for _, l := range lines {
+			events = append(events, st.Feed([]byte(l))...)
+		}
+	}
+	feed(`{"type":"response.created","response":{"model":"m1"}}`,
+		`{"type":"response.reasoning_text.delta","delta":"思考中"}`,
+		`{"type":"response.output_text.delta","delta":"你"}`,
+		`{"type":"response.output_text.delta","delta":"好"}`,
+		`{"type":"response.output_item.added","item":{"type":"function_call","call_id":"c1","name":"f"}}`,
+		`{"type":"response.function_call_arguments.delta","delta":"{\"a\":1}"}`,
+		`{"type":"response.completed","response":{"status":"completed","usage":{"input_tokens":8,"output_tokens":4}}}`)
+	events = append(events, st.Finish()...)
+
+	got := bridgeEventTypes(events)
+	want := "message_start,content_block_start,content_block_delta,content_block_delta," +
+		"content_block_stop,content_block_start,content_block_delta,content_block_stop," +
+		"message_delta,message_stop"
+	if got != want {
+		t.Fatalf("事件序列:\n got %s\nwant %s", got, want)
+	}
+	if st.Usage().InputTokens != 8 || st.Usage().OutputTokens != 4 {
+		t.Fatalf("usage = %+v", st.Usage())
+	}
+}
+
+// TestAnthRespBridgeEmpty 断言空流也产出完整事件骨架。
+func TestAnthRespBridgeEmpty(t *testing.T) {
+	st := NewAnthRespBridge("msg_1", "m1")
+	if got := bridgeEventTypes(st.Finish()); got != "message_start,message_delta,message_stop" {
+		t.Errorf("空流事件序列 = %s", got)
+	}
+}