fix: change SSE keepalive from event to comment format to unblock client watchdog

Change keepalive ping from SSE event (`event: ping\ndata: ...`) to SSE
comment (`: keepalive\n\n`) across all streaming handlers. SSE comments
keep HTTP connections alive but are not yielded by the Anthropic SDK,
so client-side idle watchdogs (e.g. Claude Code's 90s timeout) continue
counting down correctly when the upstream connection dies silently.

Author: danielzhou82

backend/internal/handler/gateway_helper.go

@@ -131,8 +131,10 @@ const (
 type SSEPingFormat string
 
 const (
-	// SSEPingFormatClaude is the Claude/Anthropic SSE ping format
-	SSEPingFormatClaude SSEPingFormat = "data: {\"type\": \"ping\"}\n\n"
+	// SSEPingFormatClaude uses SSE comment format: keeps HTTP connection alive without
+	// being yielded by Anthropic SDK, so client idle watchdogs work correctly.
+	// See streaming-keepalive-design.md
+	SSEPingFormatClaude SSEPingFormat = ": keepalive\n\n"
 	// SSEPingFormatNone indicates no ping should be sent (e.g., OpenAI has no ping spec)
 	SSEPingFormatNone SSEPingFormat = ""
 	// SSEPingFormatComment is an SSE comment ping for OpenAI/Codex CLI clients

backend/internal/service/antigravity_gateway_service.go

@@ -4053,9 +4053,9 @@ func (s *AntigravityGatewayService) handleClaudeStreamingResponse(c *gin.Context
 			if time.Since(lastDataAt) < keepaliveInterval {
 				continue
 			}
-			// SSE ping 事件：Anthropic 原生格式，客户端会正确处理，
-			// 同时保持连接活跃防止 Cloudflare Tunnel 等代理断开
-			if !cw.Fprintf("event: ping\ndata: {\"type\": \"ping\"}\n\n") {
+			// SSE 注释格式 keepalive：在 HTTP 层保持连接活跃，不干扰客户端 idle watchdog。
+			// 详见 streaming-keepalive-design.md
+			if !cw.Fprintf(": keepalive\n\n") {
 				logger.LegacyPrintf("service.antigravity_gateway", "Client disconnected during keepalive ping (antigravity claude), continuing to drain upstream for billing")
 				continue
 			}
@@ -4457,9 +4457,9 @@ func (s *AntigravityGatewayService) streamUpstreamResponse(c *gin.Context, resp
 			if time.Since(lastDataAt) < keepaliveInterval {
 				continue
 			}
-			// SSE ping 事件：Anthropic 原生格式，客户端会正确处理，
-			// 同时保持连接活跃防止 Cloudflare Tunnel 等代理断开
-			if !cw.Fprintf("event: ping\ndata: {\"type\": \"ping\"}\n\n") {
+			// SSE 注释格式 keepalive：在 HTTP 层保持连接活跃，不干扰客户端 idle watchdog。
+			// 详见 streaming-keepalive-design.md
+			if !cw.Fprintf(": keepalive\n\n") {
 				logger.LegacyPrintf("service.antigravity_gateway", "Client disconnected during keepalive ping (antigravity upstream), continuing to drain upstream for billing")
 				continue
 			}

backend/internal/service/gateway_service.go

@@ -6939,9 +6939,10 @@ func (s *GatewayService) handleStreamingResponse(ctx context.Context, resp *http
 			if time.Since(lastDataAt) < keepaliveInterval {
 				continue
 			}
-			// SSE ping 事件：Anthropic 原生格式，客户端会正确处理，
-			// 同时保持连接活跃防止 Cloudflare Tunnel 等代理断开
-			if _, werr := fmt.Fprint(w, "event: ping\ndata: {\"type\": \"ping\"}\n\n"); werr != nil {
+			// SSE 注释格式 keepalive：在 HTTP 层保持连接活跃（防止 Cloudflare Tunnel 等代理因空闲断开），
+			// 但不会被 Anthropic SDK yield 给消费者，不干扰客户端（如 Claude Code）的 idle watchdog。
+			// 详见 streaming-keepalive-design.md
+			if _, werr := fmt.Fprint(w, ": keepalive\n\n"); werr != nil {
 				clientDisconnected = true
 				logger.LegacyPrintf("service.gateway", "Client disconnected during keepalive ping, continuing to drain upstream for billing")
 				continue

backend/internal/service/openai_gateway_messages.go

@@ -552,8 +552,9 @@ func (s *OpenAIGatewayService) handleAnthropicStreamingResponse(
 			if time.Since(lastDataAt) < keepaliveInterval {
 				continue
 			}
-			// Send Anthropic-format ping event
-			if _, err := fmt.Fprint(c.Writer, "event: ping\ndata: {\"type\":\"ping\"}\n\n"); err != nil {
+			// SSE 注释格式 keepalive：在 HTTP 层保持连接活跃，不干扰客户端 idle watchdog。
+			// 详见 streaming-keepalive-design.md
+			if _, err := fmt.Fprint(c.Writer, ": keepalive\n\n"); err != nil {
 				// Client disconnected
 				logger.L().Info("openai messages stream: client disconnected during keepalive",
 					zap.String("request_id", requestID),