修复 AI 网关流式断流:错误透传、自动降级,max_tokens 可缺省

CHANGELOG.md

@@ -2,6 +2,20 @@
 
 格式遵循 [Keep a Changelog](https://keepachangelog.com/zh-CN/1.1.0/)，版本号遵循语义化版本。
 
+## [0.3.1] - 2026-07-13
+
+### Added
+
+- AI 网关流式断流自动降级：Messages 与 Chat Completions 流式请求在客户端尚未收到任何输出时遭遇上游断流，自动降级为非流式重做，结果按标准事件 / chunk 序列一次推送，调用日志记 `retries=1` 与降级标记。实测 OCI 兼容面对 `instructions` 与 `tools` 合计超约 64.5KB 的流式请求会静默断连（无错误事件，非流式正常，消息正文不计入），Claude Code 等大体量系统提示客户端极易触发；Responses 直通因初始事件已转发无法透明降级，日志记「上游流提前终止」；README 增补「已知上游限制：大 system 区流式断流」小节
+
+### Changed
+
+- Messages 的 `max_tokens` 改为可缺省：缺省或 ≤0 时按默认值 8192 放行（此前返回 400；部分客户端将该字段视为选填）
+
+### Fixed
+
+- AI 网关流式假成功：上游 error / `response.failed` 事件此前被吞掉，流提前 EOF 也被伪装成正常结束，调用日志呈现 200 · 0/0 且无错误信息；现 Messages 将上游失败转为 Anthropic `error` 事件，三个流式端点日志均记录上游错误消息或「上游流提前终止，未返回终态事件」
+
 ## [0.3.0] - 2026-07-13
 
 ### Added

README.md

@@ -216,6 +216,16 @@ AI 网关使用面板创建的独立密钥鉴权，支持 `Authorization: Bearer
 - Chat Completions 只承担协议转换与兼容修复；新能力优先在 Responses 和 Messages 提供
 - 单次请求最多尝试三个渠道；可重试错误会切换渠道，流式响应建立后不会换渠道重试
 
+### 已知上游限制：大 system 区流式断流
+
+实测（2026-07-13）OCI 兼容面对 `instructions` 与 `tools` 合计超约 64.5KB 的**流式**请求会在发出少量事件后静默断开连接（无任何错误事件；同请求非流式正常），与模型、字符集、消息正文大小均无关——消息正文（`input`）不计入该限制。Chat Completions 与 Messages 的 system/developer 提示会转换为 `instructions`，因此 Claude Code 等自带大体量系统提示与工具定义的客户端极易触发。
+
+网关侧应对：
+
+- Messages 与 Chat Completions 的流式请求在客户端尚未收到任何输出时遭遇上游断流，会自动降级为非流式重做，并按标准事件/chunk 序列一次推送；调用日志记 `retries=1` 与降级标记
+- Responses 直通因初始事件已转发、协议上无法透明降级，调用日志记「上游流提前终止」，客户端需自行回退非流式
+- 应急规避：将超长 system 内容移入首条 user 消息正文可绕过该限制（正文不计入），但语义有别，根治有待上游修复
+
 这里提供的是兼容接口而非 OpenAI / Anthropic 协议的完整实现。OCI OpenAI 兼容面的部分行为来自实测，未见 Oracle 文档承诺，可能随上游调整。路由与鉴权定义以 [Swagger YAML](docs/swagger.yaml) 或运行时 Swagger UI 为准；无法由 OpenAPI 完整表达的兼容边界列于上方。
 
 ### 字段兼容矩阵
@@ -289,7 +299,7 @@ Chat Completions 会先转换为 Responses 请求，再把 OCI Responses 响应
 | `max_completion_tokens` | 🔄 | 转为 `max_output_tokens`，优先于 `max_tokens` |
 | `max_tokens` | 🔄 | 未提供 `max_completion_tokens` 时转为 `max_output_tokens` |
 | `temperature`、`top_p`、`parallel_tool_calls` | ◐ | 原值写入 Responses 请求，但不校验范围或模型能力 |
-| `stream` | 🔄 | OCI Responses SSE 桥接为 `chat.completion.chunk`，末尾补 `data: [DONE]` |
+| `stream` | 🔄 | OCI Responses SSE 桥接为 `chat.completion.chunk`，末尾补 `data: [DONE]`；上游断流且尚无输出时自动降级非流式重做，结果按 chunk 序列一次推送 |
 | `stream_options.include_usage` | 🔄 | 控制网关在终块后追加 `choices: []` 的 usage 块 |
 | `stream_options.include_obfuscation` | ⚠️ | 未建模，静默忽略 |
 | `tools[].type=function` | ◐ | `name`、`description`、`parameters` 支持；`function.strict` 被忽略 |
@@ -346,13 +356,13 @@ Chat Completions 会先转换为 Responses 请求，再把 OCI Responses 响应
 | 标准字段 | 状态 | 网关行为 |
 | --- | :---: | --- |
 | `model` | ◐ | 用于模型与渠道选择，但空字符串不会在 handler 中按参数错误拒绝，通常最终返回模型不存在 |
-| `max_tokens` | 🔄 | 必填且必须大于 0，转换为 `max_output_tokens` |
+| `max_tokens` | 🔄 | 可缺省（缺省或 ≤0 时按默认值 8192），转换为 `max_output_tokens` |
 | `messages` | ◐ | 必须非空；角色、交替顺序和空内容不做完整校验 |
 | `system` | ◐ | 支持字符串或 text 块数组；多个文本块直接拼接，`cache_control` 等附加字段被忽略 |
 | `temperature`、`top_p` | ◐ | 写入 Responses 请求，不做取值范围或模型能力校验 |
 | `top_k` | ⚠️ | 能解析但不会传给上游 |
 | `stop_sequences` | ⚠️ | 能解析但不会传给上游；响应 `stop_sequence` 恒为 `null` |
-| `stream` | 🔄 | Responses SSE 桥接为 Anthropic 事件序列 |
+| `stream` | 🔄 | Responses SSE 桥接为 Anthropic 事件序列；上游断流且尚无输出时自动降级非流式重做，结果按事件序列一次推送 |
 | `tools` | ◐ | 每个工具都转换成 Responses `function`；自定义客户端工具可用，Anthropic 服务端工具类型不保留原语义 |
 | `tool_choice` | ◐ | 支持 `auto`、`any`、`none`、具名 `tool`；`disable_parallel_tool_use` 等附加字段被忽略 |
 | `metadata` | ⚠️ | 能解析但不会传给上游 |
@@ -380,7 +390,7 @@ Chat Completions 会先转换为 Responses 请求，再把 OCI Responses 响应
 - `stop_reason` 只生成 `end_turn`、`tool_use`、`max_tokens`；`stop_sequence` 恒为 `null`
 - reasoning 不会生成 Anthropic `thinking` / `redacted_thinking` 块，也没有 signature
 - usage 只保留 `input_tokens`、`output_tokens` 和 `cache_read_input_tokens`，不提供 `cache_creation_input_tokens`
-- 流式输出标准事件骨架，但不生成 `thinking_delta`、`signature_delta` 或上游失败对应的 Anthropic `error` 事件
+- 流式输出标准事件骨架，但不生成 `thinking_delta` 和 `signature_delta`；上游错误事件与无终态断流会转成 Anthropic `error` 事件
 
 实现依据：[`anthresponses.go`](internal/service/anthresponses.go) · [`anthropic.go`](internal/aiwire/anthropic.go) · [`aigateway.go`](internal/api/aigateway.go)
 

docs/docs.go

@@ -79,7 +79,7 @@ const docTemplate = `{
                 "summary": "Anthropic Messages 兼容端点",
                 "parameters": [
                     {
-                        "description": "Anthropic messages 请求体(支持 stream;经 OCI OpenAI 兼容面直通)",
+                        "description": "Anthropic messages 请求体(支持 stream;经 OCI OpenAI 兼容面直通;max_tokens 可缺省,默认 8192)",
                         "name": "body",
                         "in": "body",
                         "required": true,

docs/swagger.json

@@ -72,7 +72,7 @@
                 "summary": "Anthropic Messages 兼容端点",
                 "parameters": [
                     {
-                        "description": "Anthropic messages 请求体(支持 stream;经 OCI OpenAI 兼容面直通)",
+                        "description": "Anthropic messages 请求体(支持 stream;经 OCI OpenAI 兼容面直通;max_tokens 可缺省,默认 8192)",
                         "name": "body",
                         "in": "body",
                         "required": true,

docs/swagger.yaml

@@ -617,7 +617,8 @@ paths:
   /ai/v1/messages:
     post:
       parameters:
-      - description: Anthropic messages 请求体(支持 stream;经 OCI OpenAI 兼容面直通)
+      - description: Anthropic messages 请求体(支持 stream;经 OCI OpenAI 兼容面直通;max_tokens
+          可缺省,默认 8192)
         in: body
         name: body
         required: true

internal/api/aigateway.go

@@ -210,11 +210,15 @@ func sseHeaders(c *gin.Context) {
 	c.Writer.Flush()
 }
 
+// anthDefaultMaxTokens 是 max_tokens 缺省时的默认输出上限:Anthropic 协议
+// 该字段必填,但部分客户端当选填不传,按默认值放行而非 400 拒绝。
+const anthDefaultMaxTokens = 8192
+
 // messages 是 Anthropic /ai/v1/messages 端点。
 //
 // @Summary  Anthropic Messages 兼容端点
 // @Tags     AI 网关
-// @Param    body body object true "Anthropic messages 请求体(支持 stream;经 OCI OpenAI 兼容面直通)"
+// @Param    body body object true "Anthropic messages 请求体(支持 stream;经 OCI OpenAI 兼容面直通;max_tokens 可缺省,默认 8192)"
 // @Success  200 {object} map[string]any "Anthropic 兼容响应(流式为 SSE)"
 // @Router   /ai/v1/messages [post]
 func (h *aiGatewayHandler) messages(c *gin.Context) {
@@ -224,8 +228,7 @@ func (h *aiGatewayHandler) messages(c *gin.Context) {
 		return
 	}
 	if req.MaxTokens <= 0 {
-		aiError(c, http.StatusBadRequest, "invalid_request_error", "max_tokens 必填且需大于 0")
+		req.MaxTokens = anthDefaultMaxTokens
-		return
 	}
 	if len(req.Messages) == 0 {
 		aiError(c, http.StatusBadRequest, "invalid_request_error", "messages 不能为空")
@@ -281,10 +284,30 @@ func (h *aiGatewayHandler) streamAnthropic(c *gin.Context, body []byte, req aiwi
 	defer upstream.Close()
 	sseHeaders(c)
 	bridge := service.NewAnthRespBridge(aiRandID("msg_"), req.Model)
-	if err := forwardSSEData(upstream, func(data []byte) { writeAnthEvents(c, bridge.Feed(data)) }); err != nil {
+	emitted := 0
+	if err := forwardSSEData(upstream, func(data []byte) {
+		evs := bridge.Feed(data)
+		emitted += len(evs)
+		writeAnthEvents(c, evs)
+	}); err != nil {
 		entry.ErrMsg = err.Error()
 	}
+	// 上游断流且客户端尚未收到任何事件(OCI 兼容面对大请求 + 推理模型的
+	// 流式通道会在 reasoning 阶段掐断):降级非流式重做,结果按事件序列推送
+	if emitted == 0 && !bridge.SawTerminal() && c.Request.Context().Err() == nil {
+		if h.anthFallback(c, &entry, req) {
+			entry.Status = http.StatusOK
+			entry.LatencyMs = time.Since(start).Milliseconds()
+			callID := h.gw.LogCall(entry)
+			h.maybeLogContent(c, callID, "anthropic", req.Model, true, req, nil)
+			return
+		}
+	}
 	writeAnthEvents(c, bridge.Finish())
+	// 上游错误事件 / 提前终止:SSE 头已发出维持 200,错误落日志可查
+	if msg := bridge.Err(); msg != "" && entry.ErrMsg == "" {
+		entry.ErrMsg = msg
+	}
 	entry.Status = http.StatusOK
 	entry.LatencyMs = time.Since(start).Milliseconds()
 	usage := bridge.Usage()
@@ -295,6 +318,30 @@ func (h *aiGatewayHandler) streamAnthropic(c *gin.Context, body []byte, req aiwi
 	h.maybeLogContent(c, callID, "anthropic", req.Model, true, req, nil)
 }
 
+// anthFallback 用非流式重做同一请求并把完整结果按事件序列推送;
+// 成功返回 true 并把渠道 / 用量 / 降级标注写入日志条目。
+func (h *aiGatewayHandler) anthFallback(c *gin.Context, entry *model.AiCallLog, req aiwire.MessagesRequest) bool {
+	req.Stream = false
+	body, err := service.AnthropicToResponsesBody(req)
+	if err != nil {
+		return false
+	}
+	payload, meta, err := h.gw.RespPassthrough(c.Request.Context(), body, req.Model, keyGroup(c))
+	if err != nil {
+		return false
+	}
+	out, err := service.ResponsesToAnthropic(payload, aiRandID("msg_"))
+	if err != nil {
+		return false
+	}
+	writeAnthEvents(c, service.AnthMessageEvents(out))
+	entry.ChannelID, entry.ChannelName = meta.ChannelID, meta.ChannelName
+	entry.Retries++
+	entry.ErrMsg = "流式上游断流,已降级非流式完成"
+	fillUsage(entry, service.RespPassthroughUsage(payload))
+	return true
+}
+
 // forwardSSEData 逐行读上游 SSE,把 data 行交给 emit 即时转换写出;
 // Anthropic 与 Chat Completions 两条流式桥共用。
 func forwardSSEData(upstream io.Reader, emit func([]byte)) error {
@@ -403,12 +450,31 @@ func (h *aiGatewayHandler) streamChat(c *gin.Context, body []byte, req aiwire.Ch
 	sseHeaders(c)
 	includeUsage := req.StreamOptions != nil && req.StreamOptions.IncludeUsage
 	bridge := service.NewChatRespBridge(aiRandID("chatcmpl-"), req.Model, time.Now().Unix(), includeUsage)
-	if err := forwardSSEData(upstream, func(data []byte) { writeChatChunks(c, bridge.Feed(data)) }); err != nil {
+	emitted := 0
+	if err := forwardSSEData(upstream, func(data []byte) {
+		chunks := bridge.Feed(data)
+		emitted += len(chunks)
+		writeChatChunks(c, chunks)
+	}); err != nil {
 		entry.ErrMsg = err.Error()
 	}
+	// 上游断流且客户端尚未收到任何块:降级非流式重做(成因同 messages 端点)
+	if emitted == 0 && bridge.Usage() == nil && c.Request.Context().Err() == nil {
+		if h.chatFallback(c, &entry, req, includeUsage) {
+			entry.Status = http.StatusOK
+			entry.LatencyMs = time.Since(start).Milliseconds()
+			callID := h.gw.LogCall(entry)
+			h.maybeLogContent(c, callID, "openai", req.Model, true, req, nil)
+			return
+		}
+	}
 	writeChatChunks(c, bridge.Finish())
 	c.Writer.WriteString("data: [DONE]\n\n")
 	c.Writer.Flush()
+	// 未见 completed 即结束:usage 缺失说明上游流异常提前终止,落日志可查
+	if bridge.Usage() == nil && entry.ErrMsg == "" {
+		entry.ErrMsg = "上游流提前终止,未返回终态事件"
+	}
 	entry.Status = http.StatusOK
 	entry.LatencyMs = time.Since(start).Milliseconds()
 	fillUsage(&entry, bridge.Usage())
@@ -416,6 +482,32 @@ func (h *aiGatewayHandler) streamChat(c *gin.Context, body []byte, req aiwire.Ch
 	h.maybeLogContent(c, callID, "openai", req.Model, true, req, nil)
 }
 
+// chatFallback 用非流式重做同一请求并把完整结果按 chunk 序列推送;
+// 成功返回 true 并把渠道 / 用量 / 降级标注写入日志条目。
+func (h *aiGatewayHandler) chatFallback(c *gin.Context, entry *model.AiCallLog, req aiwire.ChatRequest, includeUsage bool) bool {
+	req.Stream = false
+	body, err := service.ChatToResponsesBody(req)
+	if err != nil {
+		return false
+	}
+	payload, meta, err := h.gw.RespPassthrough(c.Request.Context(), body, req.Model, keyGroup(c))
+	if err != nil {
+		return false
+	}
+	out, err := service.ResponsesToChat(payload, aiRandID("chatcmpl-"), time.Now().Unix())
+	if err != nil {
+		return false
+	}
+	writeChatChunks(c, service.ChatResponseChunks(out, includeUsage))
+	c.Writer.WriteString("data: [DONE]\n\n")
+	c.Writer.Flush()
+	entry.ChannelID, entry.ChannelName = meta.ChannelID, meta.ChannelName
+	entry.Retries++
+	entry.ErrMsg = "流式上游断流,已降级非流式完成"
+	fillUsage(entry, service.RespPassthroughUsage(payload))
+	return true
+}
+
 // writeChatChunks 逐块写出 chunk 的 SSE data 行并 flush。
 func writeChatChunks(c *gin.Context, chunks []aiwire.ChatChunk) {
 	for _, ch := range chunks {
@@ -526,9 +618,14 @@ func (h *aiGatewayHandler) responsesPassthroughStream(c *gin.Context, body []byt
 	}
 	defer upstream.Close()
 	sseHeaders(c)
-	usage, err := forwardSSE(c, upstream)
+	usage, upErr, err := forwardSSE(c, upstream)
 	if err != nil {
 		entry.ErrMsg = err.Error()
+	} else if upErr != "" {
+		// 上游错误事件已原样转发给客户端,这里落日志可查
+		entry.ErrMsg = upErr
+	} else if usage == nil {
+		entry.ErrMsg = "上游流提前终止,未返回终态事件"
 	}
 	entry.Status = http.StatusOK
 	entry.LatencyMs = time.Since(start).Milliseconds()
@@ -538,10 +635,11 @@ func (h *aiGatewayHandler) responsesPassthroughStream(c *gin.Context, body []byt
 }
 
 // forwardSSE 把上游 SSE 逐行转发给客户端,空行(事件边界)即 flush;
-// 顺带从 data 行提取 response.completed 的 usage。
+// 顺带从 data 行提取 response.completed 的 usage 与错误事件消息。
-func forwardSSE(c *gin.Context, upstream io.Reader) (*aiwire.Usage, error) {
+func forwardSSE(c *gin.Context, upstream io.Reader) (*aiwire.Usage, string, error) {
 	reader := bufio.NewReader(upstream)
 	var usage *aiwire.Usage
+	var upErr string
 	for {
 		line, err := reader.ReadBytes('\n')
 		if len(line) > 0 {
@@ -553,14 +651,17 @@ func forwardSSE(c *gin.Context, upstream io.Reader) (*aiwire.Usage, error) {
 				if u := service.RespStreamCompletedUsage(data); u != nil {
 					usage = u
 				}
+				if m := service.RespStreamErrorMsg(data); m != "" && upErr == "" {
+					upErr = m
+				}
 			}
 		}
 		if err != nil {
 			c.Writer.Flush()
 			if errors.Is(err, io.EOF) {
-				return usage, nil
+				return usage, upErr, nil
 			}
-			return usage, err
+			return usage, upErr, err
 		}
 	}
 }

internal/api/aigateway_test.go

@@ -85,6 +85,8 @@ func TestAiGatewayKeyModelRestrict(t *testing.T) {
 			`{"model":"meta.llama-3.3-70b-instruct","messages":[{"role":"user","content":"hi"}]}`, 404, deny},
 		{"chat completions 白名单内穿透", "limited-key-1234", "/ai/v1/chat/completions",
 			`{"model":"ghost-model","messages":[{"role":"user","content":"hi"}]}`, 404, pass},
+		{"messages 缺 max_tokens 按默认值放行", "open-key-12345", "/ai/v1/messages",
+			`{"model":"ghost-model","messages":[{"role":"user","content":"hi"}]}`, 404, pass},
 		{"chat completions 缺 messages 拒绝", "open-key-12345", "/ai/v1/chat/completions",
 			`{"model":"ghost-model"}`, 400, []string{"invalid_request_error"}},
 		{"chat completions 非 function 工具拒绝", "open-key-12345", "/ai/v1/chat/completions",

internal/service/airesponses.go

@@ -104,3 +104,25 @@ func RespStreamCompletedUsage(data []byte) *aiwire.Usage {
 	}
 	return RespPassthroughUsage(ev.Response)
 }
+
+// RespStreamErrorMsg 从一行 SSE data JSON 中提取 error / response.failed 事件的
+// 错误消息;非错误事件返回空串。流式直通据此把上游错误写入调用日志。
+func RespStreamErrorMsg(data []byte) string {
+	var ev respStreamEvent
+	if json.Unmarshal(data, &ev) != nil {
+		return ""
+	}
+	switch ev.Type {
+	case "error":
+		if ev.Message != "" {
+			return ev.Message
+		}
+		return "上游返回错误事件 error"
+	case "response.failed":
+		if m := respErrorMsg(ev.Response); m != "" {
+			return m
+		}
+		return "上游返回错误事件 response.failed"
+	}
+	return ""
+}

internal/service/anthresponses.go

@@ -258,7 +258,7 @@ type AnthEvent struct {
 
 // 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。
+// reasoning 系列事件丢弃;上游 error / response.failed 转 Anthropic error 事件透传。
 type AnthRespBridge struct {
 	id, model   string
 	started     bool
@@ -267,6 +267,11 @@ type AnthRespBridge struct {
 	blockIndex  int
 	stopReason  string
 	usage       aiwire.AnthUsage
+	// sawTerminal 标记收到过终态事件(completed/incomplete/failed);
+	// 上游流提前 EOF 时据此发 error 事件而非伪装正常结束
+	sawTerminal bool
+	// errMsg 记录上游错误消息,非空即本流已失败(供调用日志)
+	errMsg string
 }
 
 // NewAnthRespBridge 构造桥;id 为响应消息 ID。
@@ -278,6 +283,7 @@ func NewAnthRespBridge(id, model string) *AnthRespBridge {
 type respStreamEvent struct {
 	Type     string          `json:"type"`
 	Delta    string          `json:"delta"`
+	Message  string          `json:"message"`
 	Item     *respOutputItem `json:"item"`
 	Response *respPayload    `json:"response"`
 }
@@ -288,27 +294,130 @@ func (st *AnthRespBridge) Feed(data []byte) []AnthEvent {
 	if json.Unmarshal(data, &ev) != nil {
 		return nil
 	}
-	var events []AnthEvent
+	if ev.Type == "error" || ev.Type == "response.failed" {
-	if !st.started {
+		return st.failWith(ev)
-		st.started = true
-		events = append(events, st.startEvent())
 	}
+	// message_start 延迟到首个可见输出事件:created / reasoning 阶段不向客户端
+	// 写任何字节,上游此段断流时 handler 才有降级非流式重做的无感窗口
+	var events []AnthEvent
 	switch ev.Type {
 	case "response.output_item.added":
 		if ev.Item != nil && ev.Item.Type == "function_call" {
+			events = append(events, st.ensureStarted()...)
 			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.ensureStarted()...)
 		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":
+	case "response.completed", "response.incomplete":
+		st.sawTerminal = true
 		st.finishFrom(ev.Response)
 	}
 	return events
 }
 
+// ensureStarted 在首个可见输出事件前补发 message_start(仅一次)。
+func (st *AnthRespBridge) ensureStarted() []AnthEvent {
+	if st.started {
+		return nil
+	}
+	st.started = true
+	return []AnthEvent{st.startEvent()}
+}
+
+// failWith 记录上游错误并产出 Anthropic error 事件(每流至多一次)。
+func (st *AnthRespBridge) failWith(ev respStreamEvent) []AnthEvent {
+	if st.errMsg != "" {
+		return nil
+	}
+	st.sawTerminal = true
+	msg := ev.Message
+	if ev.Type == "response.failed" {
+		st.finishFrom(ev.Response)
+		if m := respErrorMsg(ev.Response); m != "" {
+			msg = m
+		}
+	}
+	if msg == "" {
+		msg = "上游返回错误事件 " + ev.Type
+	}
+	st.errMsg = msg
+	return []AnthEvent{st.errorEvent()}
+}
+
+// respErrorMsg 提取 response.failed 载荷中的错误消息。
+func respErrorMsg(resp *respPayload) string {
+	if resp == nil || resp.Error == nil {
+		return ""
+	}
+	return (*resp.Error)["message"]
+}
+
+// errorEvent 按 Anthropic 流式协议产出 error 事件。
+func (st *AnthRespBridge) errorEvent() AnthEvent {
+	return AnthEvent{Event: "error", Data: map[string]any{
+		"type":  "error",
+		"error": map[string]string{"type": "api_error", "message": st.errMsg},
+	}}
+}
+
+// Err 返回上游错误消息;空串表示流正常(供调用日志)。
+func (st *AnthRespBridge) Err() string { return st.errMsg }
+
+// SawTerminal 报告是否收到过终态事件;false 即上游流提前终止。
+func (st *AnthRespBridge) SawTerminal() bool { return st.sawTerminal }
+
+// AnthMessageEvents 把完整 Messages 响应展开为标准事件序列,
+// 供流式上游断流后的非流式降级结果推送(客户端协议不变)。
+func AnthMessageEvents(m *aiwire.MessagesResponse) []AnthEvent {
+	events := []AnthEvent{{Event: "message_start", Data: map[string]any{
+		"type": "message_start",
+		"message": map[string]any{
+			"id": m.ID, "type": "message", "role": m.Role, "model": m.Model,
+			"content": []any{}, "stop_reason": nil,
+			"usage": map[string]int{"input_tokens": 0, "output_tokens": 0},
+		},
+	}}}
+	for i, b := range m.Content {
+		events = append(events, anthBlockEvents(i, b)...)
+	}
+	usage := map[string]int{"input_tokens": m.Usage.InputTokens, "output_tokens": m.Usage.OutputTokens}
+	if m.Usage.CacheReadInputTokens > 0 {
+		usage["cache_read_input_tokens"] = m.Usage.CacheReadInputTokens
+	}
+	events = append(events,
+		AnthEvent{Event: "message_delta", Data: map[string]any{
+			"type":  "message_delta",
+			"delta": map[string]any{"stop_reason": m.StopReason, "stop_sequence": nil},
+			"usage": usage,
+		}},
+		AnthEvent{Event: "message_stop", Data: map[string]any{"type": "message_stop"}})
+	return events
+}
+
+// anthBlockEvents 把一个内容块展开为 start / delta / stop 三事件。
+func anthBlockEvents(index int, b aiwire.AnthBlock) []AnthEvent {
+	var start, delta map[string]any
+	if b.Type == "tool_use" {
+		start = map[string]any{"type": "tool_use", "id": b.ID, "name": b.Name, "input": map[string]any{}}
+		delta = map[string]any{"type": "input_json_delta", "partial_json": string(b.Input)}
+	} else {
+		start = map[string]any{"type": "text", "text": ""}
+		delta = map[string]any{"type": "text_delta", "text": b.Text}
+	}
+	return []AnthEvent{
+		{Event: "content_block_start", Data: map[string]any{
+			"type": "content_block_start", "index": index, "content_block": start}},
+		{Event: "content_block_delta", Data: map[string]any{
+			"type": "content_block_delta", "index": index, "delta": delta}},
+		{Event: "content_block_stop", Data: map[string]any{
+			"type": "content_block_stop", "index": index}},
+	}
+}
+
 func (st *AnthRespBridge) textDelta(delta string) []AnthEvent {
 	var events []AnthEvent
 	if !st.blockOpen || st.blockIsTool {
@@ -378,7 +487,16 @@ func (st *AnthRespBridge) closeBlockEvent() AnthEvent {
 }
 
 // Finish 在上游流结束后收尾:关块 → message_delta(stop_reason+usage)→ message_stop。
+// 已发过 error 事件的流不再补终态;未见终态事件即 EOF 视为上游提前终止,
+// 发 error 事件而非伪装正常结束(否则客户端拿到"成功的空消息")。
 func (st *AnthRespBridge) Finish() []AnthEvent {
+	if st.errMsg != "" {
+		return nil
+	}
+	if !st.sawTerminal {
+		st.errMsg = "上游流提前终止,未返回终态事件"
+		return []AnthEvent{st.errorEvent()}
+	}
 	var events []AnthEvent
 	if !st.started {
 		st.started = true

internal/service/anthresponses_test.go

@@ -164,10 +164,105 @@ func TestAnthRespBridge(t *testing.T) {
 	}
 }
 
-// TestAnthRespBridgeEmpty 断言空流也产出完整事件骨架。
+// TestAnthRespBridgeFailure 断言异常流的错误透传:上游 error / response.failed
-func TestAnthRespBridgeEmpty(t *testing.T) {
+// 事件转 Anthropic error 事件,提前 EOF(未见终态)不再伪装正常结束。
+func TestAnthRespBridgeFailure(t *testing.T) {
+	cases := []struct {
+		name      string
+		lines     []string
+		wantKinds string
+		wantErr   string
+	}{
+		{
+			name:      "上游 error 事件透传",
+			lines:     []string{`{"type":"error","message":"model overloaded"}`},
+			wantKinds: "error",
+			wantErr:   "model overloaded",
+		},
+		{
+			name: "response.failed 提取错误消息",
+			lines: []string{
+				`{"type":"response.output_text.delta","delta":"你"}`,
+				`{"type":"response.failed","response":{"status":"failed","error":{"message":"content filtered"}}}`,
+			},
+			wantKinds: "message_start,content_block_start,content_block_delta,error",
+			wantErr:   "content filtered",
+		},
+		{
+			name:      "空流提前终止",
+			lines:     nil,
+			wantKinds: "error",
+			wantErr:   "上游流提前终止,未返回终态事件",
+		},
+		{
+			name: "输出中途 EOF 无终态",
+			lines: []string{
+				`{"type":"response.output_text.delta","delta":"你"}`,
+			},
+			wantKinds: "message_start,content_block_start,content_block_delta,error",
+			wantErr:   "上游流提前终止,未返回终态事件",
+		},
+	}
+	for _, tc := range cases {
+		t.Run(tc.name, func(t *testing.T) {
 			st := NewAnthRespBridge("msg_1", "m1")
-	if got := bridgeEventTypes(st.Finish()); got != "message_start,message_delta,message_stop" {
+			var events []AnthEvent
-		t.Errorf("空流事件序列 = %s", got)
+			for _, l := range tc.lines {
+				events = append(events, st.Feed([]byte(l))...)
+			}
+			events = append(events, st.Finish()...)
+			if got := bridgeEventTypes(events); got != tc.wantKinds {
+				t.Fatalf("事件序列 = %s, want %s", got, tc.wantKinds)
+			}
+			if st.Err() != tc.wantErr {
+				t.Fatalf("Err() = %q, want %q", st.Err(), tc.wantErr)
+			}
+		})
+	}
+}
+
+// TestRespStreamErrorMsg 断言直通流错误事件消息提取。
+func TestRespStreamErrorMsg(t *testing.T) {
+	cases := []struct {
+		name string
+		data string
+		want string
+	}{
+		{name: "error 事件", data: `{"type":"error","message":"boom"}`, want: "boom"},
+		{name: "error 无消息用占位", data: `{"type":"error"}`, want: "上游返回错误事件 error"},
+		{name: "failed 事件", data: `{"type":"response.failed","response":{"error":{"message":"bad"}}}`, want: "bad"},
+		{name: "正常事件返回空", data: `{"type":"response.completed","response":{}}`, want: ""},
+		{name: "非 JSON 返回空", data: `<html>`, want: ""},
+	}
+	for _, tc := range cases {
+		t.Run(tc.name, func(t *testing.T) {
+			if got := RespStreamErrorMsg([]byte(tc.data)); got != tc.want {
+				t.Fatalf("RespStreamErrorMsg = %q, want %q", got, tc.want)
+			}
+		})
+	}
+}
+
+// TestAnthMessageEvents 断言非流式降级结果展开的事件序列与 usage。
+func TestAnthMessageEvents(t *testing.T) {
+	msg := &aiwire.MessagesResponse{
+		ID: "msg_1", Type: "message", Role: "assistant", Model: "m1",
+		Content: []aiwire.AnthBlock{
+			{Type: "text", Text: "好"},
+			{Type: "tool_use", ID: "c1", Name: "f", Input: json.RawMessage(`{"a":1}`)},
+		},
+		StopReason: "tool_use",
+		Usage:      aiwire.AnthUsage{InputTokens: 9, OutputTokens: 3, CacheReadInputTokens: 5},
+	}
+	events := AnthMessageEvents(msg)
+	want := "message_start,content_block_start,content_block_delta,content_block_stop," +
+		"content_block_start,content_block_delta,content_block_stop,message_delta,message_stop"
+	if got := bridgeEventTypes(events); got != want {
+		t.Fatalf("事件序列:\n got %s\nwant %s", got, want)
+	}
+	delta := events[len(events)-2].Data.(map[string]any)
+	usage := delta["usage"].(map[string]int)
+	if usage["input_tokens"] != 9 || usage["output_tokens"] != 3 || usage["cache_read_input_tokens"] != 5 {
+		t.Fatalf("usage = %+v", usage)
 	}
 }

internal/service/chatresponses.go

@@ -358,3 +358,32 @@ func (b *ChatRespBridge) Finish() []aiwire.ChatChunk {
 
 // Usage 返回聚合到的用量(供调用日志),上游未报告时为 nil。
 func (b *ChatRespBridge) Usage() *aiwire.Usage { return b.usage }
+
+// ChatResponseChunks 把完整 Chat 响应展开为 chunk 序列(内容与工具调用 →
+// 终块 → 可选 usage 块),供流式上游断流后的非流式降级结果推送。
+func ChatResponseChunks(resp *aiwire.ChatResponse, includeUsage bool) []aiwire.ChatChunk {
+	if len(resp.Choices) == 0 {
+		return nil
+	}
+	choice := resp.Choices[0]
+	mk := func(delta aiwire.Delta, finish *string) aiwire.ChatChunk {
+		return aiwire.ChatChunk{ID: resp.ID, Object: "chat.completion.chunk", Created: resp.Created,
+			Model: resp.Model, Choices: []aiwire.ChunkChoice{{Index: 0, Delta: delta, FinishReason: finish}}}
+	}
+	delta := aiwire.Delta{Role: "assistant", Content: choice.Message.Content.Text}
+	for i, tc := range choice.Message.ToolCalls {
+		delta.ToolCalls = append(delta.ToolCalls, aiwire.ToolCallDelta{Index: i, ID: tc.ID,
+			Type: "function", Function: aiwire.FunctionCallDelta{Name: tc.Function.Name, Arguments: tc.Function.Arguments}})
+	}
+	finish := choice.FinishReason
+	chunks := []aiwire.ChatChunk{mk(delta, nil), mk(aiwire.Delta{}, &finish)}
+	if includeUsage {
+		usage := resp.Usage
+		if usage == nil {
+			usage = &aiwire.Usage{}
+		}
+		chunks = append(chunks, aiwire.ChatChunk{ID: resp.ID, Object: "chat.completion.chunk",
+			Created: resp.Created, Model: resp.Model, Choices: []aiwire.ChunkChoice{}, Usage: usage})
+	}
+	return chunks
+}