diff --git a/server/internal/proxy/convert/convert_test.go b/server/internal/proxy/convert/convert_test.go
index 4914fbc..fdfe812 100644
--- a/server/internal/proxy/convert/convert_test.go
+++ b/server/internal/proxy/convert/convert_test.go
@@ -1022,3 +1022,117 @@ func TestMessagesToChatReqToolOrder(t *testing.T) {
 		t.Fatalf("msg[0] missing tool_calls: %s", out)
 	}
 }
+
+// ---------------------------------------------------------------------------
+// 流式工具调用：三个此前缺失的方向（chat→responses / responses→chat / messages→responses）
+
+func TestStreamChatToResponsesToolCall(t *testing.T) {
+	tf := newChatToResponses().line
+	out := feedLines(t, tf, []string{
+		`data: {"id":"chatcmpl-1","object":"chat.completion.chunk","model":"gpt-4o","choices":[{"index":0,"delta":{"role":"assistant"},"finish_reason":null}]}` + "\n\n",
+		`data: {"id":"chatcmpl-1","object":"chat.completion.chunk","model":"gpt-4o","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"id":"call_9","type":"function","function":{"name":"get_weather","arguments":""}}]},"finish_reason":null}]}` + "\n\n",
+		`data: {"id":"chatcmpl-1","object":"chat.completion.chunk","model":"gpt-4o","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"{\"city\":\"北京\"}"}}]},"finish_reason":null}]}` + "\n\n",
+		`data: {"id":"chatcmpl-1","object":"chat.completion.chunk","model":"gpt-4o","choices":[{"index":0,"delta":{},"finish_reason":"tool_calls"}]}` + "\n\n",
+		`data: {"id":"chatcmpl-1","object":"chat.completion.chunk","model":"gpt-4o","choices":[],"usage":{"prompt_tokens":10,"completion_tokens":5}}` + "\n\n",
+		"data: [DONE]\n\n",
+	})
+	if !strings.Contains(out, "event: response.output_item.added") || !strings.Contains(out, `"name":"get_weather"`) {
+		t.Fatalf("missing function_call item added: %s", out)
+	}
+	if !strings.Contains(out, "event: response.function_call_arguments.delta") {
+		t.Fatalf("missing arguments delta: %s", out)
+	}
+	if !strings.Contains(out, "event: response.output_item.done") {
+		t.Fatalf("missing output_item.done: %s", out)
+	}
+	// completed 的 output 数组要携带完整参数，客户端不依赖事件也能拿到最终形态
+	if !strings.Contains(out, `"call_id":"call_9"`) {
+		t.Fatalf("missing call_id: %s", out)
+	}
+	if !strings.Contains(out, `\"city\":\"北京\"`) {
+		t.Fatalf("missing accumulated arguments in completed output: %s", out)
+	}
+	if !strings.Contains(out, `"usage":{"completion_tokens":5,"prompt_tokens":10}`) {
+		t.Fatalf("missing usage in completed: %s", out)
+	}
+}
+
+func TestStreamResponsesToChatToolCall(t *testing.T) {
+	tf := newResponsesToChat().line
+	out := feedLines(t, tf, []string{
+		"event: response.created\n",
+		`data: {"type":"response.created","response":{"id":"resp_1","model":"gpt-4o"}}` + "\n\n",
+		"event: response.output_item.added\n",
+		`data: {"type":"response.output_item.added","output_index":1,"item":{"type":"function_call","id":"fc_1","call_id":"call_9","name":"get_weather","arguments":""}}` + "\n\n",
+		"event: response.function_call_arguments.delta\n",
+		`data: {"type":"response.function_call_arguments.delta","item_id":"fc_1","delta":"{\"city\":\"北京\"}"}` + "\n\n",
+		"event: response.completed\n",
+		`data: {"type":"response.completed","response":{"id":"resp_1","status":"completed","usage":{"input_tokens":7,"output_tokens":8}}}` + "\n\n",
+	})
+	if !strings.Contains(out, `"tool_calls"`) || !strings.Contains(out, `"name":"get_weather"`) {
+		t.Fatalf("missing tool_calls chunk: %s", out)
+	}
+	if !strings.Contains(out, `"arguments":"{\"city\":\"北京\"}"`) {
+		t.Fatalf("missing arguments delta chunk: %s", out)
+	}
+	if !strings.Contains(out, `"finish_reason":"tool_calls"`) {
+		t.Fatalf("finish_reason should be tool_calls: %s", out)
+	}
+	if !strings.Contains(out, `"usage":{"input_tokens":7,"output_tokens":8}`) {
+		t.Fatalf("missing usage chunk: %s", out)
+	}
+	if !strings.Contains(out, "data: [DONE]") {
+		t.Fatalf("missing [DONE]: %s", out)
+	}
+}
+
+func TestStreamMessagesToResponsesToolCall(t *testing.T) {
+	tf := newMessagesToResponses().line
+	out := feedLines(t, tf, []string{
+		"event: message_start\n",
+		`data: {"type":"message_start","message":{"id":"msg_1","model":"claude-sonnet-5"}}` + "\n\n",
+		"event: content_block_start\n",
+		`data: {"type":"content_block_start","index":0,"content_block":{"type":"tool_use","id":"toolu_1","name":"get_weather"}}` + "\n\n",
+		"event: content_block_delta\n",
+		`data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"{\"city\":\"北京\"}"}}` + "\n\n",
+		"event: content_block_stop\n",
+		`data: {"type":"content_block_stop","index":0}` + "\n\n",
+		"event: message_delta\n",
+		`data: {"type":"message_delta","delta":{"stop_reason":"tool_use"},"usage":{"input_tokens":10,"output_tokens":5}}` + "\n\n",
+		"event: message_stop\n",
+		`data: {"type":"message_stop"}` + "\n\n",
+	})
+	if !strings.Contains(out, "event: response.output_item.added") || !strings.Contains(out, `"call_id":"toolu_1"`) {
+		t.Fatalf("missing function_call item added: %s", out)
+	}
+	if !strings.Contains(out, "event: response.function_call_arguments.delta") {
+		t.Fatalf("missing arguments delta: %s", out)
+	}
+	if !strings.Contains(out, "event: response.output_item.done") {
+		t.Fatalf("missing output_item.done: %s", out)
+	}
+	if !strings.Contains(out, `"call_id":"toolu_1"`) || !strings.Contains(out, `"arguments":"{\"city\":\"北京\"}"`) {
+		t.Fatalf("completed output array missing function_call with full arguments: %s", out)
+	}
+	if !strings.Contains(out, `"usage":{"input_tokens":10,"output_tokens":5}`) {
+		t.Fatalf("missing usage in completed: %s", out)
+	}
+}
+
+func TestStreamResponsesToChatTextStillWorks(t *testing.T) {
+	tf := newResponsesToChat().line
+	out := feedLines(t, tf, []string{
+		"event: response.created\n",
+		`data: {"type":"response.created","response":{"id":"resp_1","model":"gpt-4o"}}` + "\n\n",
+		"event: response.output_text.delta\n",
+		`data: {"type":"response.output_text.delta","delta":"hi"}` + "\n\n",
+		"event: response.completed\n",
+		`data: {"type":"response.completed","response":{"id":"resp_1","status":"completed","usage":{"input_tokens":1,"output_tokens":1}}}` + "\n\n",
+	})
+	if !strings.Contains(out, `"content":"hi"`) {
+		t.Fatalf("missing text chunk: %s", out)
+	}
+	if !strings.Contains(out, `"finish_reason":"stop"`) {
+		t.Fatalf("missing finish chunk: %s", out)
+	}
+}
diff --git a/server/internal/proxy/convert/stream.go b/server/internal/proxy/convert/stream.go
index fe2732a..07af4f7 100644
--- a/server/internal/proxy/convert/stream.go
+++ b/server/internal/proxy/convert/stream.go
@@ -2,6 +2,7 @@ package convert
 
 import (
 	"encoding/json"
+	"strconv"
 	"strings"
 )
 
@@ -443,9 +444,17 @@ type messagesToResponses struct {
 	model string
 	usage any
 	done  bool
+
+	// 工具调用：messages tool_use block index → function_call 输出条目。
+	// items 按首次出现顺序累积，response.completed 携带完整 output 数组。
+	items    []map[string]any
+	toolItem map[int]map[string]any
+	textSB   strings.Builder
 }
 
-func newMessagesToResponses() *messagesToResponses { return &messagesToResponses{} }
+func newMessagesToResponses() *messagesToResponses {
+	return &messagesToResponses{toolItem: map[int]map[string]any{}}
+}
 
 func (t *messagesToResponses) line(line []byte) []byte {
 	isData, data, done := t.parseLine(line)
@@ -479,21 +488,89 @@ func (t *messagesToResponses) line(line []byte) []byte {
 				"id": "resp_" + strings.TrimPrefix(rid, "msg_"), "object": "response", "model": t.model, "status": "in_progress",
 			},
 		}))
+	case "content_block_start":
+		cb, _ := m["content_block"].(map[string]any)
+		if cb == nil || cb["type"] != "tool_use" {
+			return nil
+		}
+		blockIdx, _ := m["index"].(float64)
+		toolID, _ := cb["id"].(string)
+		name, _ := cb["name"].(string)
+		item := map[string]any{
+			"type": "function_call", "id": "fc_" + strconv.Itoa(len(t.items)), "call_id": toolID,
+			"name": name, "arguments": "", "status": "in_progress",
+		}
+		t.toolItem[int(blockIdx)] = item
+		oi := len(t.items)
+		t.items = append(t.items, item)
+		out = append(out, eventLine("response.output_item.added", map[string]any{
+			"type": "response.output_item.added", "output_index": oi, "item": item,
+		}))
 	case "content_block_delta":
 		delta, _ := m["delta"].(map[string]any)
-		text, _ := delta["text"].(string)
-		if text != "" {
-			out = append(out, eventLine("response.output_text.delta", map[string]any{
-				"type": "response.output_text.delta", "delta": text, "item_id": "msg_1", "output_index": 0, "content_index": 0,
+		deltaType, _ := delta["type"].(string)
+		if deltaType == "input_json_delta" {
+			blockIdx, _ := m["index"].(float64)
+			item, ok := t.toolItem[int(blockIdx)]
+			if !ok {
+				return nil
+			}
+			partial, _ := delta["partial_json"].(string)
+			if partial == "" {
+				return nil
+			}
+			item["arguments"] = item["arguments"].(string) + partial
+			out = append(out, eventLine("response.function_call_arguments.delta", map[string]any{
+				"type": "response.function_call_arguments.delta", "item_id": item["id"],
+				"output_index": itemIndexByID(t.items, item["id"].(string)), "delta": partial,
 			}))
+			return joinLines(out)
 		}
+		text, _ := delta["text"].(string)
+		if text == "" {
+			return nil
+		}
+		t.textSB.WriteString(text)
+		if len(t.items) == 0 || t.items[0]["type"] != "message" {
+			t.items = append([]map[string]any{{
+				"type": "message", "id": "msg_1", "role": "assistant", "status": "in_progress", "content": []any{},
+			}}, t.items...)
+		}
+		out = append(out, eventLine("response.output_text.delta", map[string]any{
+			"type": "response.output_text.delta", "delta": text, "item_id": "msg_1", "output_index": 0, "content_index": 0,
+		}))
+	case "content_block_stop":
+		blockIdx, _ := m["index"].(float64)
+		item, ok := t.toolItem[int(blockIdx)]
+		if !ok {
+			return nil
+		}
+		out = append(out, eventLine("response.output_item.done", map[string]any{
+			"type": "response.output_item.done", "output_index": itemIndexByID(t.items, item["id"].(string)),
+			"item": functionCallDone(item),
+		}))
 	case "message_stop":
 		if !t.done {
 			t.done = true
+			// 文本条目收尾（纯工具调用流可能没有文本条目）
+			if len(t.items) > 0 && t.items[0]["type"] == "message" {
+				out = append(out, eventLine("response.output_item.done", map[string]any{
+					"type": "response.output_item.done", "output_index": 0,
+					"item": messageItemDone("msg_1", t.textSB.String()),
+				}))
+			}
+			output := make([]any, 0, len(t.items))
+			for _, item := range t.items {
+				if item["type"] == "message" {
+					output = append(output, messageItemDone("msg_1", t.textSB.String()))
+					continue
+				}
+				output = append(output, functionCallDone(item))
+			}
 			out = append(out, eventLine("response.completed", map[string]any{
 				"type": "response.completed",
 				"response": map[string]any{
-					"id": "resp_stream", "object": "response", "model": t.model, "status": "completed", "usage": t.usage,
+					"id": "resp_stream", "object": "response", "model": t.model, "status": "completed", "usage": t.usage, "output": output,
 				},
 			}))
 		}
@@ -507,9 +584,16 @@ func (t *messagesToResponses) line(line []byte) []byte {
 type responsesToChat struct {
 	sseState
 	id, model string
+
+	// 工具调用：responses function_call 条目 → chat tool_calls index（并行调用各占一个 index）
+	toolIdx  map[string]int
+	nextTool int
+	anyTool  bool
 }
 
-func newResponsesToChat() *responsesToChat { return &responsesToChat{} }
+func newResponsesToChat() *responsesToChat {
+	return &responsesToChat{toolIdx: map[string]int{}}
+}
 
 func (t *responsesToChat) line(line []byte) []byte {
 	isData, data, done := t.parseLine(line)
@@ -538,19 +622,61 @@ func (t *responsesToChat) line(line []byte) []byte {
 		}))
 	case "response.output_text.delta":
 		delta, _ := m["delta"].(string)
-		if delta != "" {
-			out = append(out, dataLine(map[string]any{
-				"id": "chatcmpl-" + strings.TrimPrefix(t.id, "resp_"), "object": "chat.completion.chunk", "model": t.model,
-				"choices": []any{map[string]any{"index": 0, "delta": map[string]any{"content": delta}, "finish_reason": nil}},
-			}))
+		if delta == "" {
+			return nil
 		}
-	case "response.completed":
 		out = append(out, dataLine(map[string]any{
 			"id": "chatcmpl-" + strings.TrimPrefix(t.id, "resp_"), "object": "chat.completion.chunk", "model": t.model,
-			"choices": []any{map[string]any{"index": 0, "delta": map[string]any{}, "finish_reason": "stop"}},
+			"choices": []any{map[string]any{"index": 0, "delta": map[string]any{"content": delta}, "finish_reason": nil}},
 		}))
-		if u, ok := m["response"].(map[string]any); ok {
-			if usage, ok := u["usage"]; ok {
+	case "response.output_item.added":
+		item, _ := m["item"].(map[string]any)
+		if item == nil || item["type"] != "function_call" {
+			return nil
+		}
+		itemID, _ := item["id"].(string)
+		callID, _ := item["call_id"].(string)
+		if callID == "" {
+			callID = itemID
+		}
+		name, _ := item["name"].(string)
+		idx := t.nextTool
+		t.nextTool++
+		t.toolIdx[itemID] = idx
+		t.anyTool = true
+		out = append(out, dataLine(map[string]any{
+			"id": "chatcmpl-" + strings.TrimPrefix(t.id, "resp_"), "object": "chat.completion.chunk", "model": t.model,
+			"choices": []any{map[string]any{"index": 0, "delta": map[string]any{
+				"tool_calls": []any{map[string]any{"index": idx, "id": callID, "type": "function", "function": map[string]any{"name": name, "arguments": ""}}},
+			}, "finish_reason": nil}},
+		}))
+	case "response.function_call_arguments.delta":
+		itemID, _ := m["item_id"].(string)
+		idx, ok := t.toolIdx[itemID]
+		if !ok {
+			return nil
+		}
+		delta, _ := m["delta"].(string)
+		if delta == "" {
+			return nil
+		}
+		out = append(out, dataLine(map[string]any{
+			"id": "chatcmpl-" + strings.TrimPrefix(t.id, "resp_"), "object": "chat.completion.chunk", "model": t.model,
+			"choices": []any{map[string]any{"index": 0, "delta": map[string]any{
+				"tool_calls": []any{map[string]any{"index": idx, "function": map[string]any{"arguments": delta}}},
+			}, "finish_reason": nil}},
+		}))
+	case "response.completed":
+		stop := "stop"
+		if t.anyTool {
+			stop = "tool_calls"
+		}
+		out = append(out, dataLine(map[string]any{
+			"id": "chatcmpl-" + strings.TrimPrefix(t.id, "resp_"), "object": "chat.completion.chunk", "model": t.model,
+			"choices": []any{map[string]any{"index": 0, "delta": map[string]any{}, "finish_reason": stop}},
+		}))
+		if resp, ok := m["response"].(map[string]any); ok {
+			if usage, ok := resp["usage"]; ok {
 				out = append(out, dataLine(map[string]any{
 					"id": "chatcmpl-" + strings.TrimPrefix(t.id, "resp_"), "object": "chat.completion.chunk", "model": t.model,
 					"choices": []any{}, "usage": usage,
@@ -571,9 +697,17 @@ type chatToResponses struct {
 	usage      any
 	finishSeen bool
 	done       bool
+
+	// 工具调用：chat tool_calls index → function_call 输出条目（arguments 就地累积）。
+	// items 按首次出现顺序累积，response.completed 携带完整 output 数组。
+	items    []map[string]any
+	toolItem map[int]map[string]any
+	textSB   strings.Builder
 }
 
-func newChatToResponses() *chatToResponses { return &chatToResponses{} }
+func newChatToResponses() *chatToResponses {
+	return &chatToResponses{toolItem: map[int]map[string]any{}}
+}
 
 func (t *chatToResponses) line(line []byte) []byte {
 	isData, data, done := t.parseLine(line)
@@ -584,12 +718,7 @@ func (t *chatToResponses) line(line []byte) []byte {
 		// 流结束兜底：finish 后 usage 未随块到达时在此补发 completed
 		if !t.done {
 			t.done = true
-			return eventLine("response.completed", map[string]any{
-				"type": "response.completed",
-				"response": map[string]any{
-					"id": "resp_stream", "object": "response", "model": t.model, "status": "completed", "usage": t.usage,
-				},
-			})
+			return t.completed()
 		}
 		return nil
 	}
@@ -616,24 +745,122 @@ func (t *chatToResponses) line(line []byte) []byte {
 	var out [][]byte
 	if role, _ := delta["role"].(string); role == "assistant" {
 		out = append(out, eventLine("response.created", map[string]any{
-			"type": "response.created",
+			"type":     "response.created",
 			"response": map[string]any{"id": "resp_stream", "object": "response", "model": t.model, "status": "in_progress"},
 		}))
 	}
 	if content, _ := delta["content"].(string); content != "" {
+		t.textSB.WriteString(content)
+		if len(t.items) == 0 || t.items[0]["type"] != "message" {
+			// 文本条目插入队首，保持 output_index 与 items 位置一致
+			t.items = append([]map[string]any{{
+				"type": "message", "id": "msg_1", "role": "assistant", "status": "in_progress", "content": []any{},
+			}}, t.items...)
+		}
 		out = append(out, eventLine("response.output_text.delta", map[string]any{
 			"type": "response.output_text.delta", "delta": content, "item_id": "msg_1", "output_index": 0, "content_index": 0,
 		}))
 	}
+	// 工具调用：新 index 开 function_call 条目（output_item.added），参数分段经
+	// function_call_arguments.delta 透出（与 OpenAI 原生 Responses 流的事件形状一致）
+	if tcs, ok := delta["tool_calls"].([]any); ok {
+		for _, tc := range tcs {
+			call, ok := tc.(map[string]any)
+			if !ok {
+				continue
+			}
+			ci, _ := call["index"].(float64)
+			ti := int(ci)
+			fn, _ := call["function"].(map[string]any)
+			name, _ := fn["name"].(string)
+			args, _ := fn["arguments"].(string)
+			item, seen := t.toolItem[ti]
+			if !seen {
+				callID, _ := call["id"].(string)
+				item = map[string]any{
+					"type": "function_call", "id": "fc_" + strconv.Itoa(len(t.items)), "call_id": callID,
+					"name": name, "arguments": "", "status": "in_progress",
+				}
+				t.toolItem[ti] = item
+				oi := len(t.items)
+				t.items = append(t.items, item)
+				out = append(out, eventLine("response.output_item.added", map[string]any{
+					"type": "response.output_item.added", "output_index": oi, "item": item,
+				}))
+			}
+			if args != "" {
+				item["arguments"] = item["arguments"].(string) + args
+				out = append(out, eventLine("response.function_call_arguments.delta", map[string]any{
+					"type": "response.function_call_arguments.delta", "item_id": item["id"],
+					"output_index": itemIndexByID(t.items, item["id"].(string)), "delta": args,
+				}))
+			}
+		}
+	}
 	// 上游 usage 块(choices 为空)通常晚于 finish_reason：此时再发 completed，携带 usage
 	if _, hasUsage := m["usage"]; hasUsage && t.finishSeen && !t.done {
 		t.done = true
-		out = append(out, eventLine("response.completed", map[string]any{
-			"type": "response.completed",
-			"response": map[string]any{
-				"id": "resp_stream", "object": "response", "model": t.model, "status": "completed", "usage": t.usage,
-			},
-		}))
+		out = append(out, t.completed())
 	}
 	return joinLines(out)
 }
+
+// itemIndexByID 返回条目在 items 中的位置（output_index）。未知 id 返回 0。
+func itemIndexByID(items []map[string]any, id string) int {
+	for i, it := range items {
+		if s, _ := it["id"].(string); s == id {
+			return i
+		}
+	}
+	return 0
+}
+
+// messageItemDone 文本条目的终态（流式期间 content 逐段累积，done 时一次性补齐）。
+func messageItemDone(id, text string) map[string]any {
+	return map[string]any{
+		"type": "message", "id": id, "role": "assistant", "status": "completed",
+		"content": []any{map[string]any{"type": "output_text", "text": text}},
+	}
+}
+
+// completed 收尾：先给每个条目发 output_item.done，再发带完整 output 数组的 response.completed。
+func (t *chatToResponses) completed() []byte {
+	var out [][]byte
+	for i, item := range t.items {
+		var final map[string]any
+		if item["type"] == "message" {
+			final = messageItemDone(item["id"].(string), t.textSB.String())
+		} else {
+			final = functionCallDone(item)
+		}
+		out = append(out, eventLine("response.output_item.done", map[string]any{
+			"type": "response.output_item.done", "output_index": i, "item": final,
+		}))
+	}
+	out = append(out, eventLine("response.completed", map[string]any{
+		"type":     "response.completed",
+		"response": map[string]any{"id": "resp_stream", "object": "response", "model": t.model, "status": "completed", "usage": t.usage, "output": t.outputArray()},
+	}))
+	return joinLines(out)
+}
+
+// outputArray 组装 response.completed 携带的 output 数组（与 output_item.done 的终态一致）。
+func (t *chatToResponses) outputArray() []any {
+	o := make([]any, 0, len(t.items))
+	for _, item := range t.items {
+		if item["type"] == "message" {
+			o = append(o, messageItemDone(item["id"].(string), t.textSB.String()))
+			continue
+		}
+		o = append(o, functionCallDone(item))
+	}
+	return o
+}
+
+// functionCallDone 工具调用条目的终态。
+func functionCallDone(item map[string]any) map[string]any {
+	return map[string]any{
+		"type": "function_call", "id": item["id"], "call_id": item["call_id"],
+		"name": item["name"], "arguments": item["arguments"],
+	}
+}
diff --git a/server/internal/proxy/gateway.go b/server/internal/proxy/gateway.go
index 9dff553..60005fd 100644
--- a/server/internal/proxy/gateway.go
+++ b/server/internal/proxy/gateway.go
@@ -128,6 +128,23 @@ func (g *Gateway) checkModelAllowed(c *gin.Context, u *store.User, model string)
 	return true
 }
 
+// checkModelExists 拦截不存在或已禁用的模型。/v1/models 只列启用的模型，请求侧口径必须一致：
+// 禁用模型不能只是从列表消失——否则照样能调用、照样计费，管理员的下架操作形同虚设。
+func (g *Gateway) checkModelExists(c *gin.Context, model string) bool {
+	if model == "" {
+		apiError(c, http.StatusBadRequest, "invalid_request_error", "model is required")
+		return false
+	}
+	var n int64
+	g.db.Model(&store.Model{}).Where("name = ? AND enabled = ?", model, true).Count(&n)
+	if n == 0 {
+		apiError(c, http.StatusNotFound, "model_not_found",
+			fmt.Sprintf("The model %q does not exist or is not enabled", model))
+		return false
+	}
+	return true
+}
+
 func contains(list []string, s string) bool {
 	for _, v := range list {
 		if v == s {
diff --git a/server/internal/proxy/handlers.go b/server/internal/proxy/handlers.go
index b7e5f2f..6372c76 100644
--- a/server/internal/proxy/handlers.go
+++ b/server/internal/proxy/handlers.go
@@ -38,6 +38,9 @@ func (g *Gateway) chatCompletions(c *gin.Context) {
 	c.Set("protocol", convert.ProtoChat)
 	c.Set("model_name", br.Model)
 	g.recordRawRequest(c, u, body)
+	if !g.checkModelExists(c, br.Model) {
+		return
+	}
 	if !g.checkModelAllowed(c, u, br.Model) {
 		apiError(c, http.StatusForbidden, "model_not_allowed", "模型未对你开放，请联系管理员")
 		return
@@ -71,6 +74,9 @@ func (g *Gateway) responses(c *gin.Context) {
 	c.Set("protocol", convert.ProtoResponses)
 	c.Set("model_name", br.Model)
 	g.recordRawRequest(c, u, body)
+	if !g.checkModelExists(c, br.Model) {
+		return
+	}
 	if !g.checkModelAllowed(c, u, br.Model) {
 		apiError(c, http.StatusForbidden, "model_not_allowed", "模型未对你开放，请联系管理员")
 		return
@@ -104,6 +110,9 @@ func (g *Gateway) messages(c *gin.Context) {
 	c.Set("protocol", convert.ProtoMessages)
 	c.Set("model_name", br.Model)
 	g.recordRawRequest(c, u, body)
+	if !g.checkModelExists(c, br.Model) {
+		return
+	}
 	if !g.checkModelAllowed(c, u, br.Model) {
 		apiError(c, http.StatusForbidden, "model_not_allowed", "模型未对你开放，请联系管理员")
 		return
diff --git a/server/internal/proxy/passthrough.go b/server/internal/proxy/passthrough.go
index fc59134..1e7e86e 100644
--- a/server/internal/proxy/passthrough.go
+++ b/server/internal/proxy/passthrough.go
@@ -10,6 +10,7 @@ import (
 	"errors"
 	"fmt"
 	"io"
+	"log"
 	"net/http"
 	"strconv"
 	"strings"
@@ -305,7 +306,13 @@ func (g *Gateway) proxyOne(c *gin.Context, ch *store.Channel, plan *upstreamPlan
 		req.Header.Set("User-Agent", ua)
 	}
 	if plan.path == "/messages" {
+		// Anthropic 官方 API 只认 x-api-key（Bearer 是 OAuth token 的用法），
+		// 兼容网关两种都收；带上 anthropic-version 与客户端声明的 beta 头。
+		req.Header.Set("x-api-key", upKey)
 		req.Header.Set("anthropic-version", "2023-06-01")
+		if beta := c.GetHeader("anthropic-beta"); beta != "" {
+			req.Header.Set("anthropic-beta", beta)
+		}
 	}
 	for _, h := range []string{"OpenAI-Organization", "OpenAI-Project", "OpenAI-Beta"} {
 		if v := c.GetHeader(h); v != "" {
@@ -722,7 +729,9 @@ func (g *Gateway) finishUsage(c *gin.Context, ch *store.Channel, start time.Time
 			float64(cacheRead)/1e6*model.CacheReadPrice +
 			float64(cacheCreate)/1e6*model.CacheWritePriceOr()
 	} else {
-		cost = float64(in+cacheRead+cacheCreate)/1e6*0.15 + float64(out)/1e6*0.60 // 无定价模型时按示例价
+		// 模型在记账时已被删除（请求进行中被下架）：无法定价，按 0 记账留痕。
+		// 虚构单价扣用户的钱既不可审计也无法对账。
+		log.Printf("proxy: trace=%s model %q not found at billing time, cost recorded as 0", trace, mn)
 	}
 	// 落库的 input_tokens 存输入总量（含缓存读/写），与上游 message_start 口径一致，便于对账展示。
 	in += cacheRead + cacheCreate
@@ -793,6 +802,11 @@ func (g *Gateway) finishUsage(c *gin.Context, ch *store.Channel, start time.Time
 		RawResponse:         rawResp,
 		CreatedAt:           time.Now().UTC(),
 	})
+
+	// 请求级一行日志：usage_logs 落库是异步的且可能丢，排障时日志是唯一即时依据
+	log.Printf("proxy: trace=%s user=%d key=%d model=%q channel=%d proto=%s in=%d out=%d cache_r=%d cache_w=%d cost=%.6f status=%s http=%d latency=%dms err=%s",
+		traceStr, uidVal, kidVal, mn, chID, p, in, out, cacheRead, cacheCreate, cost, status,
+		c.Writer.Status(), latency, errCode)
 }
 
 // recordError 失败请求的记账（不产生扣费，status=error）。
diff --git a/server/internal/proxy/proxy_test.go b/server/internal/proxy/proxy_test.go
index 68e9c1c..475831d 100644
--- a/server/internal/proxy/proxy_test.go
+++ b/server/internal/proxy/proxy_test.go
@@ -605,3 +605,76 @@ func TestAllChannelsFailedAttributesLastChannel(t *testing.T) {
 		t.Fatalf("失败不应产生扣费：%+v", l)
 	}
 }
+
+// Anthropic 原生上游必须收到 x-api-key（官方 API 不接受 API key 走 Bearer）。
+func TestAnthropicUpstreamUsesXAPIKey(t *testing.T) {
+	var gotXAPIKey, gotAuth, gotVersion string
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		gotXAPIKey = r.Header.Get("x-api-key")
+		gotAuth = r.Header.Get("Authorization")
+		gotVersion = r.Header.Get("anthropic-version")
+		_, _ = io.WriteString(w, `{"id":"msg_1","type":"message","role":"assistant","model":"m1","content":[{"type":"text","text":"hi"}],"stop_reason":"end_turn","usage":{"input_tokens":1,"output_tokens":1}}`)
+	}))
+	defer srv.Close()
+
+	h := newHarness(t, srv.URL, 3000, true, nil)
+	// 把渠道改造成 Anthropic 原生（messages 协议）
+	if err := h.db.Model(&store.Channel{}).Where("id = ?", 1).Updates(map[string]any{
+		"provider": store.ChannelProviderAnthropic,
+		"formats":  []string{store.FormatMessages},
+	}).Error; err != nil {
+		t.Fatal(err)
+	}
+
+	// Anthropic 客户端走 x-api-key 头 + /v1/messages
+	w := httptest.NewRecorder()
+	c, _ := gin.CreateTestContext(w)
+	req := httptest.NewRequest(http.MethodPost, "/v1/messages", strings.NewReader(`{"model":"m1","messages":[{"role":"user","content":"hi"}]}`))
+	req.Header.Set("x-api-key", h.key)
+	req.Header.Set("anthropic-version", "2023-06-01")
+	req.Header.Set("anthropic-beta", "prompt-caching-2024-07-31")
+	c.Request = req
+	h.gw.Auth(c)
+	if !c.IsAborted() {
+		h.gw.Handle(c)
+	}
+	if w.Code != http.StatusOK {
+		t.Fatalf("messages 请求应成功，实际 %d: %s", w.Code, w.Body.String())
+	}
+	if gotXAPIKey != "upstream-key" {
+		t.Fatalf("上游应收到 x-api-key=upstream-key，实际 %q", gotXAPIKey)
+	}
+	if gotAuth != "Bearer upstream-key" {
+		t.Fatalf("兼容网关仍需 Bearer 头，实际 %q", gotAuth)
+	}
+	if gotVersion == "" {
+		t.Fatal("上游应收到 anthropic-version")
+	}
+}
+
+// 禁用/不存在的模型必须 404：不能只是从 /v1/models 列表消失、请求照样计费。
+func TestDisabledOrUnknownModelIsRejected(t *testing.T) {
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		t.Fatal("禁用模型不应触达上游")
+	}))
+	defer srv.Close()
+
+	h := newHarness(t, srv.URL, 3000, true, nil)
+	if err := h.db.Model(&store.Model{}).Where("id = ?", 1).Update("enabled", false).Error; err != nil {
+		t.Fatal(err)
+	}
+
+	w := h.run(t, nil, `{"model":"m1","messages":[{"role":"user","content":"hi"}]}`)
+	if w.Code != http.StatusNotFound {
+		t.Fatalf("禁用模型应 404，实际 %d: %s", w.Code, w.Body.String())
+	}
+	if !strings.Contains(w.Body.String(), "model_not_found") {
+		t.Fatalf("错误码应为 model_not_found: %s", w.Body.String())
+	}
+
+	// 空模型名直接 400
+	w = h.run(t, nil, `{"messages":[{"role":"user","content":"hi"}]}`)
+	if w.Code != http.StatusBadRequest {
+		t.Fatalf("空模型应 400，实际 %d: %s", w.Code, w.Body.String())
+	}
+}