Merge pull request #2066 from alfadb/fix/anthropic-stream-eof-failover

fix(gateway): Anthropic 流式 EOF 失败移交 + SSE error 帧标准化

backend/internal/service/gateway_service.go

@@ -11,6 +11,7 @@ import (
 	"io"
 	"log/slog"
 	mathrand "math/rand"
+	"net"
 	"net/http"
 	"net/url"
 	"os"
@@ -20,6 +21,7 @@ import (
 	"strconv"
 	"strings"
 	"sync/atomic"
+	"syscall"
 	"time"
 
 	"github.com/Wei-Shaw/sub2api/internal/config"
@@ -6520,6 +6522,49 @@ func (s *GatewayService) shouldFailoverOn400(respBody []byte) bool {
 	return false
 }
 
+// sanitizeStreamError 返回不含网络地址的客户端可见错误描述。
+// 默认 (*net.OpError).Error() 会拼接 Source/Addr 字段，泄露内部 IP/端口与上游
+// 服务器地址（例如 "read tcp 10.0.0.1:54321->52.1.2.3:443: read: connection
+// reset by peer"）。该函数只保留可识别的错误类别，原始 err 仍在调用点写入日志。
+func sanitizeStreamError(err error) string {
+	if err == nil {
+		return ""
+	}
+	switch {
+	case errors.Is(err, io.ErrUnexpectedEOF):
+		return "unexpected EOF"
+	case errors.Is(err, io.EOF):
+		return "EOF"
+	case errors.Is(err, context.Canceled):
+		return "canceled"
+	case errors.Is(err, context.DeadlineExceeded):
+		return "deadline exceeded"
+	case errors.Is(err, syscall.ECONNRESET):
+		return "connection reset by peer"
+	case errors.Is(err, syscall.ECONNABORTED):
+		return "connection aborted"
+	case errors.Is(err, syscall.ETIMEDOUT):
+		return "connection timed out"
+	case errors.Is(err, syscall.EPIPE):
+		return "broken pipe"
+	case errors.Is(err, syscall.ECONNREFUSED):
+		return "connection refused"
+	}
+	var netErr *net.OpError
+	if errors.As(err, &netErr) {
+		if netErr.Timeout() {
+			if netErr.Op != "" {
+				return netErr.Op + " timeout"
+			}
+			return "i/o timeout"
+		}
+		if netErr.Op != "" {
+			return netErr.Op + " network error"
+		}
+	}
+	return "upstream connection error"
+}
+
 // ExtractUpstreamErrorMessage 从上游响应体中提取错误消息
 // 支持 Claude 风格的错误格式：{"type":"error","error":{"type":"...","message":"..."}}
 func ExtractUpstreamErrorMessage(body []byte) string {
@@ -6957,14 +7002,31 @@ func (s *GatewayService) handleStreamingResponse(ctx context.Context, resp *http
 	}
 	lastDataAt := time.Now()
 
-	// 仅发送一次错误事件，避免多次写入导致协议混乱（写失败时尽力通知客户端）
+	// 仅发送一次错误事件，避免多次写入导致协议混乱（写失败时尽力通知客户端）。
+	// 事件格式遵循 Anthropic SSE 标准：{"type":"error","error":{"type":<reason>,"message":<message>}}
+	// 这样 Anthropic SDK / Claude Code 等客户端能按标准 error 类型解析，UI 能显示具体错误文案，
+	// 服务端 ExtractUpstreamErrorMessage 也能从透传的 body 中提取 message。
 	errorEventSent := false
-	sendErrorEvent := func(reason string) {
+	sendErrorEvent := func(reason, message string) {
 		if errorEventSent {
 			return
 		}
 		errorEventSent = true
-		_, _ = fmt.Fprintf(w, "event: error\ndata: {\"error\":\"%s\"}\n\n", reason)
+		if message == "" {
+			message = reason
+		}
+		body, err := json.Marshal(map[string]any{
+			"type": "error",
+			"error": map[string]string{
+				"type":    reason,
+				"message": message,
+			},
+		})
+		if err != nil {
+			// json.Marshal 不可能在已知 string-only 输入上失败，保守 fallback
+			body = []byte(fmt.Sprintf(`{"type":"error","error":{"type":%q,"message":%q}}`, reason, message))
+		}
+		_, _ = fmt.Fprintf(w, "event: error\ndata: %s\n\n", body)
 		flusher.Flush()
 	}
 
@@ -7124,10 +7186,32 @@ func (s *GatewayService) handleStreamingResponse(ctx context.Context, resp *http
 				// 客户端未断开，正常的错误处理
 				if errors.Is(ev.err, bufio.ErrTooLong) {
 					logger.LegacyPrintf("service.gateway", "SSE line too long: account=%d max_size=%d error=%v", account.ID, maxLineSize, ev.err)
-					sendErrorEvent("response_too_large")
+					sendErrorEvent("response_too_large", fmt.Sprintf("upstream SSE line exceeded %d bytes", maxLineSize))
 					return &streamingResult{usage: usage, firstTokenMs: firstTokenMs}, ev.err
 				}
-				sendErrorEvent("stream_read_error")
+				// 上游中途读错误（unexpected EOF / connection reset 等，常见于 HTTP/2 GOAWAY）：
+				// 若尚未向客户端写过任何字节，包成 UpstreamFailoverError 让 handler 层走 failover/重试。
+				// 已经开始写流时 SSE 协议无 resume，只能透传错误事件给客户端。
+				// 注意:面向客户端的 disconnectMsg 必须用 sanitizeStreamError 剥离地址,
+				// 默认 *net.OpError 的 Error() 会泄露内部 IP/端口和上游地址。完整 ev.err
+				// 仅在下方 LegacyPrintf 内部日志中保留供运维诊断。
+				disconnectMsg := "upstream stream disconnected: " + sanitizeStreamError(ev.err)
+				if !c.Writer.Written() {
+					logger.LegacyPrintf("service.gateway", "Upstream stream read error before any client output (account=%d), failing over: %v", account.ID, ev.err)
+					body, _ := json.Marshal(map[string]any{
+						"type": "error",
+						"error": map[string]string{
+							"type":    "upstream_disconnected",
+							"message": disconnectMsg,
+						},
+					})
+					return nil, &UpstreamFailoverError{
+						StatusCode:             http.StatusBadGateway,
+						ResponseBody:           body,
+						RetryableOnSameAccount: true,
+					}
+				}
+				sendErrorEvent("stream_read_error", disconnectMsg)
 				return &streamingResult{usage: usage, firstTokenMs: firstTokenMs}, fmt.Errorf("stream read error: %w", ev.err)
 			}
 			line := ev.line
@@ -7186,7 +7270,7 @@ func (s *GatewayService) handleStreamingResponse(ctx context.Context, resp *http
 			if s.rateLimitService != nil {
 				s.rateLimitService.HandleStreamTimeout(ctx, account, originalModel)
 			}
-			sendErrorEvent("stream_timeout")
+			sendErrorEvent("stream_timeout", fmt.Sprintf("upstream stream idle for %s", streamInterval))
 			return &streamingResult{usage: usage, firstTokenMs: firstTokenMs}, fmt.Errorf("stream data interval timeout")
 
 		case <-keepaliveCh: