feat: add best exit selection

docs/superpowers/specs/2026-05-01-best-exit-selection-design.md

@@ -0,0 +1,186 @@
+# Best Exit Selection Design
+
+## Goal
+
+Add a multi-exit tunnel strategy named `best` that always sends new connections through the currently best-quality exit. The feature should prevent traffic from continuing to use an exit whose latency or packet loss has degraded while the exit is still technically online.
+
+Existing connections must not be interrupted. Switching affects only new connections created after the runtime chain update is applied.
+
+## Current Context
+
+- Tunnel forwarding stores entry, chain, and exit nodes in `chain_tunnel`.
+- Multi-exit runtime chains are currently rendered as one GOST hop with multiple nodes.
+- GOST selectors support `fifo`, `round`, `rand`, and `hash`, plus fail filtering through `maxFails` and `failTimeout`.
+- The current fail filter only reacts to dial, handshake, or transport failures. It does not react to high latency when the exit is still reachable.
+- `tunnel_quality_prober` already runs panel-side TCP probes and stores tunnel quality history, but it currently probes representative nodes and does not drive runtime routing decisions.
+
+## User Decisions
+
+- Add a `best` option for multi-exit tunnels.
+- `best` means always choose the current best exit for new connections.
+- Score exits by end-to-end quality.
+- Keep the existing public probe target: `www.bing.com:443`.
+- Do not disrupt established connections.
+
+## Approach
+
+Implement `best` as a panel-driven control-plane strategy.
+
+The database stores the user's intended strategy as `best`. When the panel renders runtime GOST config for a `best` exit group, it sends a GOST selector strategy of `fifo`. The panel dynamically sorts the candidate exits so the current best exit is first. GOST then chooses the first node for new connections.
+
+This avoids adding active probing logic inside every GOST agent and reuses the existing panel-to-agent command path.
+
+## Components
+
+### Frontend
+
+The tunnel form adds `最优` to the multi-exit load strategy selector.
+
+- Label: `最优`
+- Value: `best`
+- Scope: tunnel forwarding exit groups, alongside `主备/fifo`, `轮询/round`, and `随机/rand`
+- Create and edit forms must submit and restore `best` unchanged.
+
+### Backend Data Model
+
+No schema change is required.
+
+The existing `chain_tunnel.strategy` column stores `best`. Repository and handler paths should preserve the value in API responses and updates.
+
+### Runtime Chain Rendering
+
+When building runtime chain config:
+
+- If the configured strategy is not `best`, keep existing behavior.
+- If the configured strategy is `best`, emit GOST selector strategy `fifo`.
+- Sort the target nodes using the panel's latest best-exit decision before rendering the node list.
+- If no quality decision exists yet, keep the saved node order.
+
+This preserves the user's `best` intent in storage while using a GOST selector that can execute the panel's sorted decision.
+
+### Quality Prober
+
+Extend `tunnel_quality_prober` to evaluate all candidates in `best` exit groups.
+
+For each chain owner node and candidate exit, measure:
+
+- Chain owner node to candidate exit using TCP ping.
+- Candidate exit to `www.bing.com:443` using TCP ping.
+
+For direct entry-to-exit tunnels, each entry node owns its own chain decision. For tunnels with intermediate chain hops, each node in the last hop group before the exits owns its own chain decision. This allows different entry or chain nodes to choose different best exits when their path quality differs.
+
+### Scoring
+
+Each exit candidate gets an end-to-end score for a specific chain owner node.
+
+- Total latency is the sum of owner-to-exit latency and exit-to-Bing latency.
+- Total loss combines both legs by success probability: `1 - (1 - lossA) * (1 - lossB)`.
+- Failed or unreachable candidates are sorted behind successful candidates.
+- The score should heavily penalize packet loss so that low-latency but lossy exits are not selected over stable exits.
+
+A practical scoring formula can be:
+
+```text
+score = totalLatencyMs + (totalLossPercent * lossPenaltyMsPerPercent)
+```
+
+Use `lossPenaltyMsPerPercent = 100` initially. For example, 5% loss adds 500ms to the score.
+
+### Switching Rules
+
+The panel should not update chains on every probe round.
+
+Switch only when all conditions are true:
+
+- The candidate best exit is different from the currently applied first exit.
+- The candidate is successful.
+- The candidate remains best for consecutive probe rounds.
+- The candidate beats the current exit by a minimum advantage threshold.
+- The chain owner node has passed a minimum switch cooldown.
+
+Initial constants:
+
+- Consecutive confirmations: 3 rounds.
+- Switch cooldown: 30 seconds per chain owner node.
+- Minimum advantage: the candidate score must improve by at least `max(20ms, currentScore * 0.15)`.
+
+If all exits fail, keep the current runtime order and do not issue a destructive update.
+
+### Runtime Update
+
+When a `best` chain owner node changes best exit:
+
+1. Rebuild that node's `chains_<tunnelID>` payload with the best exit first and remaining candidates sorted by quality for that node.
+2. Send `UpdateChains` to that chain owner node.
+3. Do not restart or update tunnel services.
+4. Record success or failure in logs and in the in-memory decision state.
+
+This affects only future connections. Existing TCP connections keep using the `net.Conn` created before the update and continue through their original exit.
+
+### Agent Safety Improvement
+
+The current agent `UpdateChains` path unregisters the old chain before registering the new chain. This does not kill existing connections, but it creates a small window where a new connection can fail because the chain name is temporarily absent.
+
+Improve the update path so it parses the new chain first and only replaces the registered chain after parsing succeeds. The replacement window should be as small as possible. If parsing fails, the old chain must remain active.
+
+## Error Handling
+
+- If probing one candidate fails, continue scoring other candidates.
+- If a chain owner node is offline or times out, skip decisions for that owner during the round instead of marking every candidate failed.
+- If a candidate has no successful required probe data, mark it failed for that round.
+- If `UpdateChains` fails, keep the current applied order and retry on a later round.
+- If the tunnel has one exit or an incomplete config, `best` behaves like the saved order and does not trigger dynamic switching.
+- If `monitor_tunnel_quality_enabled=false`, dynamic `best` switching pauses. The last applied runtime order remains in effect.
+
+## Observability
+
+The prober should maintain in-memory decision state per `best` tunnel and chain owner node.
+
+Useful fields:
+
+- Tunnel ID and chain owner node ID.
+- Current applied best exit node ID.
+- Candidate best exit node ID.
+- Candidate scores.
+- Last switch timestamp.
+- Last switch result.
+- Reason for not switching, such as cooldown, insufficient advantage, candidate unstable, or all exits failed.
+
+Initial UI scope is limited to supporting create, update, and display of the `best` strategy. A later enhancement can expose current best exit and candidate scores in the tunnel monitor view.
+
+## Testing
+
+Backend tests:
+
+- Score calculation orders candidates by latency and packet loss.
+- Packet loss penalty prevents lossy exits from winning only because latency is low.
+- All-failed candidates do not trigger a switch.
+- Consecutive confirmation and cooldown prevent flapping.
+- `strategy=best` persists in `chain_tunnel.strategy` and is returned by tunnel list/get APIs.
+- Runtime rendering maps `best` to GOST `fifo` and places the chosen best exit first.
+
+Agent tests:
+
+- `UpdateChains` parse failure keeps the old chain registered.
+- Successful `UpdateChains` updates the chain used by new connections.
+
+Frontend verification:
+
+- Tunnel form includes `最优` in the exit strategy selector.
+- Existing tunnels with `strategy=best` render correctly.
+- Create and update requests submit `best` unchanged.
+
+Verification commands:
+
+```bash
+(cd go-backend && go test ./...)
+(cd go-gost && go test ./...)
+(cd vite-frontend && pnpm run build)
+```
+
+## Non-Goals
+
+- Do not move existing live connections to a new exit.
+- Do not add per-tunnel custom probe targets in this phase.
+- Do not implement active best-exit probing inside GOST agents.
+- Do not add a detailed best-exit UI dashboard in this phase.

go-backend/internal/http/handler/federation.go

@@ -85,6 +85,81 @@ type federationRuntimeReleaseRoleRequest struct {
 	ResourceKey   string `json:"resourceKey"`
 }
 
+func federationRuntimeChainName(bindingID string) string {
+	bindingID = strings.TrimSpace(bindingID)
+	if bindingID == "" {
+		return ""
+	}
+	return "fed_chain_" + bindingID
+}
+
+func buildFederationMiddleChainConfig(chainName string, runtimeID int64, protocol, strategy string, targets []federationRuntimeTarget, interfaceName string) (map[string]interface{}, error) {
+	chainName = strings.TrimSpace(chainName)
+	if chainName == "" {
+		return nil, fmt.Errorf("chain name is required")
+	}
+	if len(targets) == 0 {
+		return nil, fmt.Errorf("targets are required for middle role")
+	}
+	protocol = defaultString(protocol, "tls")
+	nodeItems := make([]map[string]interface{}, 0, len(targets))
+	for i, target := range targets {
+		host := strings.TrimSpace(target.Host)
+		if host == "" || target.Port <= 0 {
+			return nil, fmt.Errorf("Invalid target")
+		}
+		targetProtocol := defaultString(target.Protocol, protocol)
+		connector := map[string]interface{}{
+			"type": "relay",
+		}
+		if isTCPTunnelProtocol(targetProtocol) {
+			connector["metadata"] = map[string]interface{}{
+				"nodelay":               true,
+				"mux.keepaliveInterval": "15s",
+				"mux.keepaliveTimeout":  "45s",
+			}
+		}
+		if isKCPTunnelProtocol(targetProtocol) {
+			connector["metadata"] = map[string]interface{}{
+				"connectTimeout": "30s",
+			}
+		}
+		nodeItems = append(nodeItems, map[string]interface{}{
+			"name":      fmt.Sprintf("node_%d", i+1),
+			"addr":      processServerAddress(fmt.Sprintf("%s:%d", host, target.Port)),
+			"connector": connector,
+			"dialer":    buildTunnelDialerConfig(targetProtocol),
+		})
+	}
+
+	chainData := map[string]interface{}{
+		"name": chainName,
+		"hops": []map[string]interface{}{
+			{
+				"name": fmt.Sprintf("hop_%d", runtimeID),
+				"selector": map[string]interface{}{
+					"strategy":    runtimeTunnelStrategy(strategy),
+					"maxFails":    1,
+					"failTimeout": int64(600000000000),
+				},
+				"nodes": nodeItems,
+			},
+		},
+	}
+	if strings.TrimSpace(interfaceName) != "" {
+		hops := chainData["hops"].([]map[string]interface{})
+		hops[0]["interface"] = interfaceName
+	}
+	return chainData, nil
+}
+
+func updateChainPayload(chainName string, chainData map[string]interface{}) map[string]interface{} {
+	return map[string]interface{}{
+		"chain": chainName,
+		"data":  chainData,
+	}
+}
+
 type federationRuntimeDiagnoseRequest struct {
 	IP       string `json:"ip"`
 	Port     int    `json:"port"`
@@ -145,8 +220,8 @@ type remoteUsageNodeItem struct {
 
 func buildFederationServiceConfig(serviceName, addr, protocol, role, chainName string, targetCount int, interfaceName string) map[string]interface{} {
 	service := map[string]interface{}{
-		"name":     serviceName,
-		"addr":     addr,
+		"name": serviceName,
+		"addr": addr,
 		"handler": map[string]interface{}{
 			"type": "relay",
 		},
@@ -1056,7 +1131,43 @@ func (h *Handler) federationRuntimeApplyRole(w http.ResponseWriter, r *http.Requ
 		return
 	}
 
+	node, err := h.getNodeRecord(share.NodeID)
+	if err != nil {
+		response.WriteJSON(w, response.ErrDefault(err.Error()))
+		return
+	}
+
+	protocol := defaultString(req.Protocol, runtime.Protocol)
+	strategy := defaultString(req.Strategy, "round")
+	chainName := defaultString(runtime.ChainName, federationRuntimeChainName(runtime.BindingID))
+	if chainName == "" {
+		chainName = federationRuntimeChainName(fmt.Sprintf("%d", runtime.ID))
+	}
+	serviceName := fmt.Sprintf("fed_svc_%d", runtime.ID)
 	if runtime.Applied == 1 && strings.TrimSpace(runtime.BindingID) != "" {
+		if req.Role == "middle" && len(req.Targets) > 0 {
+			chainData, buildErr := buildFederationMiddleChainConfig(chainName, runtime.ID, protocol, strategy, req.Targets, node.InterfaceName)
+			if buildErr != nil {
+				response.WriteJSON(w, response.ErrDefault(buildErr.Error()))
+				return
+			}
+			if _, err := h.sendNodeCommand(share.NodeID, "UpdateChains", updateChainPayload(chainName, chainData), false, false); err != nil {
+				response.WriteJSON(w, response.ErrDefault(err.Error()))
+				return
+			}
+			targetBytes, _ := json.Marshal(req.Targets)
+			runtime.Role = req.Role
+			runtime.ChainName = chainName
+			runtime.Protocol = protocol
+			runtime.Strategy = strategy
+			runtime.Target = string(targetBytes)
+			runtime.Status = 1
+			runtime.UpdatedTime = time.Now().UnixMilli()
+			if err := h.repo.UpdatePeerShareRuntime(runtime); err != nil {
+				response.WriteJSON(w, response.Err(-2, err.Error()))
+				return
+			}
+		}
 		response.WriteJSON(w, response.OK(map[string]interface{}{
 			"bindingId":     runtime.BindingID,
 			"allocatedPort": runtime.Port,
@@ -1076,71 +1187,12 @@ func (h *Handler) federationRuntimeApplyRole(w http.ResponseWriter, r *http.Requ
 		}
 	}
 
-	node, err := h.getNodeRecord(share.NodeID)
-	if err != nil {
-		response.WriteJSON(w, response.ErrDefault(err.Error()))
-		return
-	}
-
-	protocol := defaultString(req.Protocol, runtime.Protocol)
-	strategy := defaultString(req.Strategy, "round")
-	chainName := fmt.Sprintf("fed_chain_%d", runtime.ID)
-	serviceName := fmt.Sprintf("fed_svc_%d", runtime.ID)
-
 	if req.Role == "middle" {
-		if len(req.Targets) == 0 {
-			response.WriteJSON(w, response.ErrDefault("targets are required for middle role"))
+		chainData, buildErr := buildFederationMiddleChainConfig(chainName, runtime.ID, protocol, strategy, req.Targets, node.InterfaceName)
+		if buildErr != nil {
+			response.WriteJSON(w, response.ErrDefault(buildErr.Error()))
 			return
 		}
-		nodeItems := make([]map[string]interface{}, 0, len(req.Targets))
-		for i, target := range req.Targets {
-			host := strings.TrimSpace(target.Host)
-			if host == "" || target.Port <= 0 {
-				response.WriteJSON(w, response.ErrDefault("Invalid target"))
-				return
-			}
-			targetProtocol := defaultString(target.Protocol, protocol)
-			connector := map[string]interface{}{
-				"type": "relay",
-			}
-			if isTCPTunnelProtocol(targetProtocol) {
-				connector["metadata"] = map[string]interface{}{
-					"nodelay":               true,
-					"mux.keepaliveInterval": "15s",
-					"mux.keepaliveTimeout":  "45s",
-				}
-			}
-			if isKCPTunnelProtocol(targetProtocol) {
-				connector["metadata"] = map[string]interface{}{
-					"connectTimeout": "30s",
-				}
-			}
-			nodeItems = append(nodeItems, map[string]interface{}{
-				"name":      fmt.Sprintf("node_%d", i+1),
-				"addr":      processServerAddress(fmt.Sprintf("%s:%d", host, target.Port)),
-				"connector": connector,
-				"dialer":    buildTunnelDialerConfig(targetProtocol),
-			})
-		}
-
-		chainData := map[string]interface{}{
-			"name": chainName,
-			"hops": []map[string]interface{}{
-				{
-					"name": fmt.Sprintf("hop_%d", runtime.ID),
-					"selector": map[string]interface{}{
-						"strategy":    strategy,
-						"maxFails":    1,
-						"failTimeout": int64(600000000000),
-					},
-					"nodes": nodeItems,
-				},
-			},
-		}
-		if strings.TrimSpace(node.InterfaceName) != "" {
-			hops := chainData["hops"].([]map[string]interface{})
-			hops[0]["interface"] = node.InterfaceName
-		}
 		if _, err := h.sendNodeCommand(share.NodeID, "AddChains", chainData, true, false); err != nil {
 			response.WriteJSON(w, response.ErrDefault(err.Error()))
 			return

go-backend/internal/http/handler/federation_runtime_test.go

@@ -281,6 +281,63 @@ func TestBuildFederationServiceConfig_NonTLSProtocol_NoNodelay(t *testing.T) {
 	}
 }
 
+func TestFederationRuntimeChainNameDerivesFromBindingID(t *testing.T) {
+	if got := federationRuntimeChainName("12"); got != "fed_chain_12" {
+		t.Fatalf("expected fed_chain_12, got %q", got)
+	}
+	if got := federationRuntimeChainName(" 12 "); got != "fed_chain_12" {
+		t.Fatalf("expected trimmed fed_chain_12, got %q", got)
+	}
+	if got := federationRuntimeChainName(""); got != "" {
+		t.Fatalf("expected blank binding ID to stay blank, got %q", got)
+	}
+}
+
+func TestBuildFederationMiddleChainConfigUsesExistingChainNameAndBestStrategy(t *testing.T) {
+	chainData, err := buildFederationMiddleChainConfig("fed_chain_12", 12, "tls", tunnelStrategyBest, []federationRuntimeTarget{
+		{Host: "10.0.0.31", Port: 30031, Protocol: "tls"},
+		{Host: "10.0.0.30", Port: 30030, Protocol: "tls"},
+	}, "")
+	if err != nil {
+		t.Fatalf("build chain: %v", err)
+	}
+	if chainData["name"] != "fed_chain_12" {
+		t.Fatalf("expected existing chain name, got %v", chainData["name"])
+	}
+	hops := chainData["hops"].([]map[string]interface{})
+	selector := hops[0]["selector"].(map[string]interface{})
+	if selector["strategy"] != bestExitRuntimeStrategy {
+		t.Fatalf("expected best strategy to map to fifo, got %v", selector["strategy"])
+	}
+	nodes := hops[0]["nodes"].([]map[string]interface{})
+	if nodes[0]["addr"] != "10.0.0.31:30031" || nodes[1]["addr"] != "10.0.0.30:30030" {
+		t.Fatalf("expected target order to be preserved, got %+v", nodes)
+	}
+}
+
+func TestUpdateChainPayloadWrapsChainDataForAgentUpdate(t *testing.T) {
+	chainData := map[string]interface{}{
+		"name": "fed_chain_12",
+		"hops": []map[string]interface{}{},
+	}
+
+	payload := updateChainPayload("fed_chain_12", chainData)
+	if len(payload) != 2 {
+		t.Fatalf("expected exact wrapper with 2 keys, got %+v", payload)
+	}
+	if payload["chain"] != "fed_chain_12" {
+		t.Fatalf("expected chain name in wrapper, got %v", payload["chain"])
+	}
+	wrappedData, ok := payload["data"].(map[string]interface{})
+	if !ok {
+		t.Fatalf("expected wrapped chain data map, got %T", payload["data"])
+	}
+	chainData["name"] = "fed_chain_12_updated"
+	if wrappedData["name"] != "fed_chain_12_updated" {
+		t.Fatalf("expected wrapper to preserve chainData identity, got %+v", wrappedData)
+	}
+}
+
 func TestFederationRuntimeReservePortRejectsWhenShareFlowExceeded(t *testing.T) {
 	r, err := repo.Open(filepath.Join(t.TempDir(), "panel.db"))
 	if err != nil {

go-backend/internal/http/handler/handler.go

@@ -52,6 +52,7 @@ type Handler struct {
 	nodeOnlineRedeploying    map[int64]struct{}
 
 	qualityProber *tunnelQualityProber
+	bestExit      *bestExitManager
 }
 
 const monitorTunnelQualityEnabledConfigKey = "monitor_tunnel_quality_enabled"
@@ -111,6 +112,7 @@ func New(repo *repo.Repository, jwtSecret string) *Handler {
 		nodeOnlineRedeployAt:     make(map[int64]time.Time),
 		nodeOnlineRedeployQueued: make(map[int64]struct{}),
 		nodeOnlineRedeploying:    make(map[int64]struct{}),
+		bestExit:                 newBestExitManager(),
 	}
 	h.healthCheck = health.NewChecker(repo, h.wsServer)
 	h.qualityProber = newTunnelQualityProber(h)

go-backend/internal/http/handler/mutations.go

@@ -8,6 +8,7 @@ import (
 	"encoding/json"
 	"errors"
 	"fmt"
+	"log"
 	"math/big"
 	"net"
 	"net/http"
@@ -3314,6 +3315,8 @@ func (h *Handler) applyFederationRuntime(state *tunnelCreateState, localDomain s
 			nextTargets := state.OutNodes
 			if hopIdx+1 < len(state.ChainHops) {
 				nextTargets = state.ChainHops[hopIdx+1]
+			} else {
+				nextTargets = h.orderBestExitTargets(state.TunnelID, chainNode.NodeID, nextTargets)
 			}
 			applyTargets := make([]client.RuntimeTarget, 0, len(nextTargets))
 			for _, target := range nextTargets {
@@ -3343,7 +3346,7 @@ func (h *Handler) applyFederationRuntime(state *tunnelCreateState, localDomain s
 				ResourceKey:   resourceKey,
 				Role:          "middle",
 				Protocol:      defaultString(chainNode.Protocol, "tls"),
-				Strategy:      defaultString(chainNode.Strategy, "round"),
+				Strategy:      runtimeStrategyForTargets(chainNode, nextTargets),
 				Targets:       applyTargets,
 			}
 			applyRes, err := fc.ApplyRole(remoteURL, remoteToken, localDomain, applyReq)
@@ -3457,6 +3460,8 @@ func (h *Handler) applyTunnelRuntimeWithMode(state *tunnelCreateState, upsert bo
 		targets := state.OutNodes
 		if len(state.ChainHops) > 0 {
 			targets = state.ChainHops[0]
+		} else {
+			targets = h.orderBestExitTargets(state.TunnelID, inNode.NodeID, targets)
 		}
 		chainData, err := buildTunnelChainConfig(state.TunnelID, inNode.NodeID, targets, state.Nodes, state.IPPreference)
 		if err != nil {
@@ -3472,11 +3477,13 @@ func (h *Handler) applyTunnelRuntimeWithMode(state *tunnelCreateState, upsert bo
 	}
 
 	for i, hop := range state.ChainHops {
-		nextTargets := state.OutNodes
-		if i+1 < len(state.ChainHops) {
-			nextTargets = state.ChainHops[i+1]
-		}
 		for _, chainNode := range hop {
+			nextTargets := state.OutNodes
+			if i+1 < len(state.ChainHops) {
+				nextTargets = state.ChainHops[i+1]
+			} else {
+				nextTargets = h.orderBestExitTargets(state.TunnelID, chainNode.NodeID, nextTargets)
+			}
 			node := state.Nodes[chainNode.NodeID]
 			if node != nil && (node.IsRemote == 1 || node.Status != 1) {
 				continue
@@ -3530,6 +3537,130 @@ func (h *Handler) applyTunnelChainOnNode(nodeID int64, chainData map[string]inte
 	return err
 }
 
+func (h *Handler) applyBestExitChainOrder(tunnelID, ownerNodeID int64, outNodes []chainNodeRecord, scores []bestExitCandidateScore, ipPreference string) error {
+	if h == nil {
+		log.Printf("best_exit: invalid chain update context tunnel=%d owner=%d", tunnelID, ownerNodeID)
+		return errors.New("invalid best exit chain update context")
+	}
+	if tunnelID <= 0 || ownerNodeID <= 0 || len(outNodes) == 0 {
+		log.Printf("best_exit: invalid chain update input tunnel=%d owner=%d exits=%d", tunnelID, ownerNodeID, len(outNodes))
+		return fmt.Errorf("invalid best exit chain update input tunnel=%d owner=%d exits=%d", tunnelID, ownerNodeID, len(outNodes))
+	}
+	targets := chainRecordsToRuntimeTargets(outNodes)
+	orderedIDs := make([]int64, 0, len(scores))
+	for _, score := range scores {
+		if score.ExitNodeID > 0 {
+			orderedIDs = append(orderedIDs, score.ExitNodeID)
+		}
+	}
+	targets = orderRuntimeTargetsByNodeID(targets, orderedIDs)
+	nodes := make(map[int64]*nodeRecord, len(targets)+1)
+	if owner, err := h.getNodeRecord(ownerNodeID); err == nil && owner != nil {
+		nodes[ownerNodeID] = owner
+	}
+	for _, target := range targets {
+		if node, err := h.getNodeRecord(target.NodeID); err == nil && node != nil {
+			nodes[target.NodeID] = node
+		}
+	}
+	owner := nodes[ownerNodeID]
+	if owner != nil && owner.IsRemote == 1 {
+		if err := h.applyRemoteBestExitChainOrder(tunnelID, ownerNodeID, owner, targets, nodes, ipPreference); err != nil {
+			log.Printf("best_exit: update remote federation chain failed tunnel=%d owner=%d err=%v", tunnelID, ownerNodeID, err)
+			return err
+		}
+		log.Printf("best_exit: updated remote federation chain tunnel=%d owner=%d best_exit=%d", tunnelID, ownerNodeID, targets[0].NodeID)
+		return nil
+	}
+	chainData, err := buildTunnelChainConfig(tunnelID, ownerNodeID, targets, nodes, ipPreference)
+	if err != nil {
+		log.Printf("best_exit: build chain failed tunnel=%d owner=%d err=%v", tunnelID, ownerNodeID, err)
+		return err
+	}
+	if err := h.applyTunnelChainOnNode(ownerNodeID, chainData, true); err != nil {
+		log.Printf("best_exit: update chain failed tunnel=%d owner=%d err=%v", tunnelID, ownerNodeID, err)
+		return err
+	}
+	log.Printf("best_exit: updated chain tunnel=%d owner=%d best_exit=%d", tunnelID, ownerNodeID, targets[0].NodeID)
+	return nil
+}
+
+func (h *Handler) applyRemoteBestExitChainOrder(tunnelID, ownerNodeID int64, owner *nodeRecord, targets []tunnelRuntimeNode, nodes map[int64]*nodeRecord, ipPreference string) error {
+	if h == nil || h.repo == nil || owner == nil {
+		return errors.New("invalid remote best exit update context")
+	}
+	bindings, err := h.repo.ListActiveFederationTunnelBindingsByTunnel(tunnelID)
+	if err != nil {
+		return err
+	}
+	var binding *repo.FederationTunnelBinding
+	for i := range bindings {
+		if bindings[i].NodeID == ownerNodeID && bindings[i].ChainType == 2 && bindings[i].Status == 1 {
+			binding = &bindings[i]
+			break
+		}
+	}
+	if binding == nil {
+		return fmt.Errorf("active federation middle binding not found for tunnel=%d owner=%d", tunnelID, ownerNodeID)
+	}
+
+	remoteURL := strings.TrimSpace(owner.RemoteURL)
+	if remoteURL == "" {
+		remoteURL = strings.TrimSpace(binding.RemoteURL)
+	}
+	remoteToken := strings.TrimSpace(owner.RemoteToken)
+	if remoteURL == "" || remoteToken == "" {
+		return errors.New("远程节点缺少共享配置")
+	}
+
+	applyTargets := make([]client.RuntimeTarget, 0, len(targets))
+	for _, target := range targets {
+		targetNode := nodes[target.NodeID]
+		if targetNode == nil {
+			return errors.New("节点不存在")
+		}
+		host, hostErr := selectTunnelDialHost(owner, targetNode, ipPreference, target.ConnectIP)
+		if hostErr != nil {
+			return hostErr
+		}
+		if target.Port <= 0 {
+			return errors.New("节点端口不能为空")
+		}
+		applyTargets = append(applyTargets, client.RuntimeTarget{
+			Host:     host,
+			Port:     target.Port,
+			Protocol: defaultString(target.Protocol, "tls"),
+		})
+	}
+
+	ownerRuntimeNode := tunnelRuntimeNode{NodeID: ownerNodeID, Protocol: "tls", Strategy: "round", ChainType: 2}
+	if chainRows, listErr := h.repo.ListChainNodesForTunnel(tunnelID); listErr == nil {
+		for _, row := range chainRows {
+			if row.NodeID == ownerNodeID && row.ChainType == 2 {
+				ownerRuntimeNode = tunnelRuntimeNode{
+					NodeID:    row.NodeID,
+					Protocol:  row.Protocol,
+					Strategy:  row.Strategy,
+					Inx:       int(row.Inx),
+					ChainType: row.ChainType,
+					Port:      row.Port,
+					ConnectIP: row.ConnectIP,
+				}
+				break
+			}
+		}
+	}
+
+	_, err = client.NewFederationClient().ApplyRole(remoteURL, remoteToken, h.federationLocalDomain(), client.RuntimeApplyRoleRequest{
+		ResourceKey: strings.TrimSpace(binding.ResourceKey),
+		Role:        "middle",
+		Protocol:    defaultString(ownerRuntimeNode.Protocol, "tls"),
+		Strategy:    runtimeStrategyForTargets(ownerRuntimeNode, targets),
+		Targets:     applyTargets,
+	})
+	return err
+}
+
 func (h *Handler) upsertTunnelChainOnNode(nodeID int64, chainData map[string]interface{}) error {
 	if h == nil {
 		return errors.New("invalid tunnel chain context")
@@ -3700,7 +3831,7 @@ func buildTunnelChainConfig(tunnelID int64, fromNodeID int64, targets []tunnelRu
 		})
 	}
 
-	strategy := defaultString(strings.TrimSpace(targets[0].Strategy), "round")
+	strategy := runtimeTunnelStrategy(defaultString(strings.TrimSpace(targets[0].Strategy), "round"))
 	hop := map[string]interface{}{
 		"name": fmt.Sprintf("hop_%d", tunnelID),
 		"selector": map[string]interface{}{
@@ -3720,6 +3851,24 @@ func buildTunnelChainConfig(tunnelID int64, fromNodeID int64, targets []tunnelRu
 	}, nil
 }
 
+func (h *Handler) orderBestExitTargets(tunnelID, ownerNodeID int64, targets []tunnelRuntimeNode) []tunnelRuntimeNode {
+	if len(targets) <= 1 || !isBestTunnelStrategy(targets[0].Strategy) {
+		return append([]tunnelRuntimeNode(nil), targets...)
+	}
+	if h == nil || h.bestExit == nil {
+		return append([]tunnelRuntimeNode(nil), targets...)
+	}
+	return h.bestExit.orderTargets(bestExitOwnerKey{TunnelID: tunnelID, OwnerNodeID: ownerNodeID}, targets)
+}
+
+func runtimeStrategyForTargets(owner tunnelRuntimeNode, targets []tunnelRuntimeNode) string {
+	strategy := defaultString(owner.Strategy, "round")
+	if len(targets) > 0 {
+		strategy = defaultString(targets[0].Strategy, strategy)
+	}
+	return runtimeTunnelStrategy(strategy)
+}
+
 func buildTunnelChainServiceConfig(tunnelID int64, chainNode tunnelRuntimeNode, node *nodeRecord, nextHopCandidateCount int) []map[string]interface{} {
 	if node == nil {
 		return nil

go-backend/internal/http/handler/tunnel_best_exit.go

@@ -0,0 +1,432 @@
+package handler
+
+import (
+	"errors"
+	"sort"
+	"strings"
+	"sync"
+	"time"
+)
+
+const (
+	tunnelStrategyBest              = "best"
+	bestExitRuntimeStrategy         = "fifo"
+	bestExitPublicTargetHost        = "www.bing.com"
+	bestExitPublicTargetPort        = 443
+	bestExitLossPenaltyMsPerPercent = 100.0
+	bestExitConfirmationRounds      = 3
+	bestExitSwitchCooldown          = 30 * time.Second
+	bestExitApplyRetryCooldown      = bestExitSwitchCooldown
+	bestExitMinLatencyAdvantageMs   = 20.0
+	bestExitMinScoreAdvantageRatio  = 0.15
+)
+
+type bestExitOwnerKey struct {
+	TunnelID    int64
+	OwnerNodeID int64
+}
+
+type bestExitCandidateScore struct {
+	OwnerNodeID int64
+	ExitNodeID  int64
+	ExitName    string
+
+	OwnerToExitLatency float64
+	ExitToBingLatency  float64
+	OwnerToExitLoss    float64
+	ExitToBingLoss     float64
+	TotalLatency       float64
+	TotalLoss          float64
+	Score              float64
+	Success            bool
+	ErrorMessage       string
+}
+
+type bestExitSwitchDecision struct {
+	Switch     bool
+	ExitNodeID int64
+	Reason     string
+	Scores     []bestExitCandidateScore
+}
+
+type bestExitProbeFunc func(nodeID int64, ip string, port int, options diagnosisExecOptions) (latency float64, loss float64, err error)
+
+type bestExitProbeResult struct {
+	latency float64
+	loss    float64
+	err     error
+}
+
+type bestExitDecision struct {
+	AppliedExitNodeID          int64
+	PendingExitNodeID          int64
+	PendingCount               int
+	LastSwitchAt               time.Time
+	LastApplyFailureAt         time.Time
+	LastApplyFailureExitNodeID int64
+	LastReason                 string
+	Scores                     []bestExitCandidateScore
+}
+
+type bestExitManager struct {
+	mu        sync.Mutex
+	decisions map[bestExitOwnerKey]*bestExitDecision
+}
+
+func newBestExitManager() *bestExitManager {
+	return &bestExitManager{decisions: make(map[bestExitOwnerKey]*bestExitDecision)}
+}
+
+func isBestTunnelStrategy(strategy string) bool {
+	return strings.EqualFold(strings.TrimSpace(strategy), tunnelStrategyBest)
+}
+
+func runtimeTunnelStrategy(strategy string) string {
+	if isBestTunnelStrategy(strategy) {
+		return bestExitRuntimeStrategy
+	}
+	return strategy
+}
+
+func scoreBestExitCandidate(ownerNodeID int64, exit chainNodeRecord, ownerLatency, ownerLoss, publicLatency, publicLoss float64) bestExitCandidateScore {
+	totalLatency := ownerLatency + publicLatency
+	totalLoss := combineLossPercent(ownerLoss, publicLoss)
+	return bestExitCandidateScore{
+		OwnerNodeID:        ownerNodeID,
+		ExitNodeID:         exit.NodeID,
+		ExitName:           exit.NodeName,
+		OwnerToExitLatency: ownerLatency,
+		ExitToBingLatency:  publicLatency,
+		OwnerToExitLoss:    ownerLoss,
+		ExitToBingLoss:     publicLoss,
+		TotalLatency:       totalLatency,
+		TotalLoss:          totalLoss,
+		Score:              totalLatency + totalLoss*bestExitLossPenaltyMsPerPercent,
+		Success:            true,
+	}
+}
+
+func failedBestExitCandidate(ownerNodeID int64, exit chainNodeRecord, message string) bestExitCandidateScore {
+	return bestExitCandidateScore{
+		OwnerNodeID:  ownerNodeID,
+		ExitNodeID:   exit.NodeID,
+		ExitName:     exit.NodeName,
+		Success:      false,
+		ErrorMessage: message,
+	}
+}
+
+func combineLossPercent(a, b float64) float64 {
+	a = clampPercent(a)
+	b = clampPercent(b)
+	return (1 - (1-a/100.0)*(1-b/100.0)) * 100.0
+}
+
+func clampPercent(v float64) float64 {
+	if v < 0 {
+		return 0
+	}
+	if v > 100 {
+		return 100
+	}
+	return v
+}
+
+func sortBestExitScores(scores []bestExitCandidateScore) {
+	sort.SliceStable(scores, func(i, j int) bool {
+		return bestExitScoreLess(scores[i], scores[j])
+	})
+}
+
+func evaluateBestExitOwner(owner chainNodeRecord, exits []chainNodeRecord, nodes map[int64]*nodeRecord, ipPreference string, options diagnosisExecOptions, ping bestExitProbeFunc) []bestExitCandidateScore {
+	scores := make([]bestExitCandidateScore, 0, len(exits))
+	if owner.NodeID <= 0 || len(exits) == 0 || ping == nil {
+		return scores
+	}
+	ownerNode := nodes[owner.NodeID]
+	for _, exit := range exits {
+		exitNode := nodes[exit.NodeID]
+		if exitNode == nil {
+			scores = append(scores, failedBestExitCandidate(owner.NodeID, exit, "exit node unavailable"))
+			continue
+		}
+		targetIP, targetPort, resolveErr := resolveBestExitProbeTarget(ownerNode, exitNode, exit.Port, ipPreference, exit.ConnectIP)
+		if resolveErr != nil {
+			scores = append(scores, failedBestExitCandidate(owner.NodeID, exit, resolveErr.Error()))
+			continue
+		}
+		ownerLatency, ownerLoss, ownerErr := ping(owner.NodeID, targetIP, targetPort, options)
+		if ownerErr != nil {
+			scores = append(scores, failedBestExitCandidate(owner.NodeID, exit, ownerErr.Error()))
+			continue
+		}
+		publicLatency, publicLoss, publicErr := ping(exit.NodeID, bestExitPublicTargetHost, bestExitPublicTargetPort, options)
+		if publicErr != nil {
+			scores = append(scores, failedBestExitCandidate(owner.NodeID, exit, publicErr.Error()))
+			continue
+		}
+		scores = append(scores, scoreBestExitCandidate(owner.NodeID, exit, ownerLatency, ownerLoss, publicLatency, publicLoss))
+	}
+	sortBestExitScores(scores)
+	return scores
+}
+
+func resolveBestExitProbeTarget(fromNode, targetNode *nodeRecord, preferredPort int, ipPreference string, connectIP string) (string, int, error) {
+	if targetNode == nil {
+		return "", 0, errors.New("目标节点不存在")
+	}
+	host, err := selectTunnelDialHost(fromNode, targetNode, ipPreference, connectIP)
+	if err != nil {
+		return "", 0, err
+	}
+	if strings.TrimSpace(host) == "" {
+		return "", 0, errors.New("目标节点地址为空")
+	}
+	port := preferredPort
+	if port <= 0 {
+		port = firstPortFromRange(targetNode.PortRange)
+	}
+	if port <= 0 {
+		port = 443
+	}
+	return host, port, nil
+}
+
+func newBestExitRoundPinger(base bestExitProbeFunc) bestExitProbeFunc {
+	cache := make(map[int64]bestExitProbeResult)
+	return func(nodeID int64, ip string, port int, options diagnosisExecOptions) (float64, float64, error) {
+		if ip == bestExitPublicTargetHost && port == bestExitPublicTargetPort {
+			if cached, ok := cache[nodeID]; ok {
+				return cached.latency, cached.loss, cached.err
+			}
+			lat, loss, err := base(nodeID, ip, port, options)
+			cache[nodeID] = bestExitProbeResult{latency: lat, loss: loss, err: err}
+			return lat, loss, err
+		}
+		return base(nodeID, ip, port, options)
+	}
+}
+
+func bestExitChainOwners(inNodes []chainNodeRecord, chainHops [][]chainNodeRecord) []chainNodeRecord {
+	if len(chainHops) == 0 {
+		return inNodes
+	}
+	return chainHops[len(chainHops)-1]
+}
+
+func chainRecordsToRuntimeTargets(rows []chainNodeRecord) []tunnelRuntimeNode {
+	out := make([]tunnelRuntimeNode, 0, len(rows))
+	for _, row := range rows {
+		out = append(out, tunnelRuntimeNode{
+			NodeID:    row.NodeID,
+			Protocol:  row.Protocol,
+			Strategy:  row.Strategy,
+			Inx:       int(row.Inx),
+			ChainType: row.ChainType,
+			Port:      row.Port,
+			ConnectIP: row.ConnectIP,
+		})
+	}
+	return out
+}
+
+func orderRuntimeTargetsByNodeID(targets []tunnelRuntimeNode, orderedIDs []int64) []tunnelRuntimeNode {
+	out := append([]tunnelRuntimeNode(nil), targets...)
+	if len(out) <= 1 || len(orderedIDs) == 0 {
+		return out
+	}
+	positions := make(map[int64]int, len(orderedIDs))
+	for i, id := range orderedIDs {
+		if _, ok := positions[id]; !ok {
+			positions[id] = i
+		}
+	}
+	sort.SliceStable(out, func(i, j int) bool {
+		pi, iok := positions[out[i].NodeID]
+		pj, jok := positions[out[j].NodeID]
+		if iok != jok {
+			return iok
+		}
+		if iok && jok && pi != pj {
+			return pi < pj
+		}
+		return false
+	})
+	return out
+}
+
+func cloneBestExitScores(scores []bestExitCandidateScore) []bestExitCandidateScore {
+	return append([]bestExitCandidateScore(nil), scores...)
+}
+
+func bestExitDecisionResult(switchNow bool, exitNodeID int64, reason string, scores []bestExitCandidateScore) bestExitSwitchDecision {
+	return bestExitSwitchDecision{Switch: switchNow, ExitNodeID: exitNodeID, Reason: reason, Scores: cloneBestExitScores(scores)}
+}
+
+func bestExitScoreLess(a, b bestExitCandidateScore) bool {
+	if a.Success != b.Success {
+		return a.Success
+	}
+	if !a.Success && !b.Success {
+		return a.ExitNodeID < b.ExitNodeID
+	}
+	if a.Score != b.Score {
+		return a.Score < b.Score
+	}
+	return a.ExitNodeID < b.ExitNodeID
+}
+
+func bestExitHasMinimumAdvantage(candidate, current bestExitCandidateScore) bool {
+	if !candidate.Success {
+		return false
+	}
+	if !current.Success {
+		return true
+	}
+	improvement := current.Score - candidate.Score
+	threshold := current.Score * bestExitMinScoreAdvantageRatio
+	if threshold < bestExitMinLatencyAdvantageMs {
+		threshold = bestExitMinLatencyAdvantageMs
+	}
+	return improvement >= threshold
+}
+
+func (m *bestExitManager) setApplied(key bestExitOwnerKey, exitNodeID int64, at time.Time) {
+	m.mu.Lock()
+	defer m.mu.Unlock()
+	d := m.decisionLocked(key)
+	d.AppliedExitNodeID = exitNodeID
+	d.PendingExitNodeID = 0
+	d.PendingCount = 0
+	d.LastApplyFailureAt = time.Time{}
+	d.LastApplyFailureExitNodeID = 0
+	d.LastSwitchAt = at
+}
+
+func (m *bestExitManager) recordApplyFailure(key bestExitOwnerKey, exitNodeID int64, at time.Time) {
+	m.mu.Lock()
+	defer m.mu.Unlock()
+	d := m.decisionLocked(key)
+	d.LastApplyFailureAt = at
+	d.LastApplyFailureExitNodeID = exitNodeID
+	d.LastReason = "apply retry cooldown"
+}
+
+func (m *bestExitManager) ensureApplied(key bestExitOwnerKey, exitNodeID int64, at time.Time) {
+	if m == nil || exitNodeID <= 0 {
+		return
+	}
+	m.mu.Lock()
+	defer m.mu.Unlock()
+	d := m.decisionLocked(key)
+	if d.AppliedExitNodeID == 0 {
+		d.AppliedExitNodeID = exitNodeID
+		d.LastSwitchAt = at
+	}
+}
+
+func (m *bestExitManager) observeScores(key bestExitOwnerKey, scores []bestExitCandidateScore, now time.Time) bestExitSwitchDecision {
+	m.mu.Lock()
+	defer m.mu.Unlock()
+
+	ordered := append([]bestExitCandidateScore(nil), scores...)
+	sortBestExitScores(ordered)
+	d := m.decisionLocked(key)
+	d.Scores = cloneBestExitScores(ordered)
+
+	if len(ordered) == 0 || !ordered[0].Success {
+		d.LastReason = "all exits failed"
+		return bestExitDecisionResult(false, 0, d.LastReason, ordered)
+	}
+
+	candidate := ordered[0]
+	if d.AppliedExitNodeID == 0 {
+		d.AppliedExitNodeID = candidate.ExitNodeID
+		d.LastSwitchAt = now
+		d.LastReason = "initial best exit"
+		return bestExitDecisionResult(false, 0, d.LastReason, ordered)
+	}
+	if candidate.ExitNodeID == d.AppliedExitNodeID {
+		d.PendingExitNodeID = 0
+		d.PendingCount = 0
+		d.LastApplyFailureAt = time.Time{}
+		d.LastApplyFailureExitNodeID = 0
+		d.LastReason = "current exit remains best"
+		return bestExitDecisionResult(false, 0, d.LastReason, ordered)
+	}
+	if candidate.ExitNodeID == d.LastApplyFailureExitNodeID && !d.LastApplyFailureAt.IsZero() && now.Sub(d.LastApplyFailureAt) < bestExitApplyRetryCooldown {
+		d.LastReason = "apply retry cooldown"
+		return bestExitDecisionResult(false, 0, d.LastReason, ordered)
+	}
+	if now.Sub(d.LastSwitchAt) < bestExitSwitchCooldown {
+		d.LastReason = "cooldown"
+		return bestExitDecisionResult(false, 0, d.LastReason, ordered)
+	}
+
+	current := findBestExitScore(ordered, d.AppliedExitNodeID)
+	if !bestExitHasMinimumAdvantage(candidate, current) {
+		d.PendingExitNodeID = 0
+		d.PendingCount = 0
+		d.LastReason = "insufficient advantage"
+		return bestExitDecisionResult(false, 0, d.LastReason, ordered)
+	}
+
+	if d.PendingExitNodeID != candidate.ExitNodeID {
+		d.PendingExitNodeID = candidate.ExitNodeID
+		d.PendingCount = 1
+		d.LastReason = "candidate pending confirmation"
+		return bestExitDecisionResult(false, 0, d.LastReason, ordered)
+	}
+	d.PendingCount++
+	if d.PendingCount < bestExitConfirmationRounds {
+		d.LastReason = "candidate pending confirmation"
+		return bestExitDecisionResult(false, 0, d.LastReason, ordered)
+	}
+
+	d.LastReason = "switch confirmed"
+	return bestExitDecisionResult(true, candidate.ExitNodeID, d.LastReason, ordered)
+}
+
+func findBestExitScore(scores []bestExitCandidateScore, exitNodeID int64) bestExitCandidateScore {
+	for _, score := range scores {
+		if score.ExitNodeID == exitNodeID {
+			return score
+		}
+	}
+	return failedBestExitCandidate(0, chainNodeRecord{NodeID: exitNodeID}, "current exit has no successful score")
+}
+
+func (m *bestExitManager) decisionLocked(key bestExitOwnerKey) *bestExitDecision {
+	if d := m.decisions[key]; d != nil {
+		return d
+	}
+	d := &bestExitDecision{}
+	m.decisions[key] = d
+	return d
+}
+
+func (m *bestExitManager) orderTargets(key bestExitOwnerKey, targets []tunnelRuntimeNode) []tunnelRuntimeNode {
+	out := append([]tunnelRuntimeNode(nil), targets...)
+	if m == nil || len(out) <= 1 {
+		return out
+	}
+	m.mu.Lock()
+	applied := int64(0)
+	if d := m.decisions[key]; d != nil {
+		applied = d.AppliedExitNodeID
+	}
+	m.mu.Unlock()
+	if applied <= 0 {
+		return out
+	}
+	sort.SliceStable(out, func(i, j int) bool {
+		if out[i].NodeID == applied {
+			return true
+		}
+		if out[j].NodeID == applied {
+			return false
+		}
+		return false
+	})
+	return out
+}

go-backend/internal/http/handler/tunnel_best_exit_test.go

@@ -0,0 +1,420 @@
+package handler
+
+import (
+	"errors"
+	"testing"
+	"time"
+)
+
+var errBestExitProbeForTest = errors.New("probe failed")
+
+func TestBestExitScoreCombinesLatencyAndLoss(t *testing.T) {
+	exit := chainNodeRecord{NodeID: 30, NodeName: "exit-a"}
+	score := scoreBestExitCandidate(10, exit, 25, 2, 80, 3)
+
+	if !score.Success {
+		t.Fatalf("expected successful score")
+	}
+	if score.OwnerNodeID != 10 || score.ExitNodeID != 30 {
+		t.Fatalf("unexpected owner/exit ids: %+v", score)
+	}
+	if score.TotalLatency != 105 {
+		t.Fatalf("expected total latency 105, got %v", score.TotalLatency)
+	}
+	if score.TotalLoss < 4.9 || score.TotalLoss > 5.0 {
+		t.Fatalf("expected combined loss about 4.94, got %v", score.TotalLoss)
+	}
+	if score.Score < 599 || score.Score > 600 {
+		t.Fatalf("expected score about 599, got %v", score.Score)
+	}
+}
+
+func TestBestExitScorePenalizesLoss(t *testing.T) {
+	stable := scoreBestExitCandidate(10, chainNodeRecord{NodeID: 30}, 80, 0, 80, 0)
+	lowLatencyLossy := scoreBestExitCandidate(10, chainNodeRecord{NodeID: 31}, 10, 5, 10, 5)
+
+	if !bestExitScoreLess(stable, lowLatencyLossy) {
+		t.Fatalf("expected stable exit to beat low-latency lossy exit: stable=%+v lossy=%+v", stable, lowLatencyLossy)
+	}
+}
+
+func TestBestExitFailedCandidateSortsLast(t *testing.T) {
+	failed := failedBestExitCandidate(10, chainNodeRecord{NodeID: 30}, "dial timeout")
+	good := scoreBestExitCandidate(10, chainNodeRecord{NodeID: 31}, 100, 0, 100, 0)
+
+	scores := []bestExitCandidateScore{failed, good}
+	sortBestExitScores(scores)
+
+	if scores[0].ExitNodeID != 31 || scores[1].ExitNodeID != 30 {
+		t.Fatalf("expected good score first and failed score last, got %+v", scores)
+	}
+}
+
+func TestBestExitInitialObservationAppliesWithoutSwitch(t *testing.T) {
+	m := newBestExitManager()
+	key := bestExitOwnerKey{TunnelID: 7, OwnerNodeID: 10}
+	now := time.Unix(100, 0)
+	candidate := scoreBestExitCandidate(10, chainNodeRecord{NodeID: 31}, 40, 0, 60, 0)
+
+	decision := m.observeScores(key, []bestExitCandidateScore{candidate}, now)
+	if decision.Switch {
+		t.Fatalf("initial observation should not return switch: %+v", decision)
+	}
+	if m.decisions[key].AppliedExitNodeID != 31 {
+		t.Fatalf("expected applied exit 31, got %+v", m.decisions[key])
+	}
+}
+
+func TestBestExitDecisionRequiresMinimumAdvantage(t *testing.T) {
+	m := newBestExitManager()
+	key := bestExitOwnerKey{TunnelID: 7, OwnerNodeID: 10}
+	now := time.Unix(100, 0)
+	current := scoreBestExitCandidate(10, chainNodeRecord{NodeID: 30}, 100, 0, 100, 0)
+	candidate := scoreBestExitCandidate(10, chainNodeRecord{NodeID: 31}, 90, 0, 90, 0)
+
+	m.setApplied(key, 30, now.Add(-time.Minute))
+	for i := 0; i < bestExitConfirmationRounds+1; i++ {
+		decision := m.observeScores(key, []bestExitCandidateScore{candidate, current}, now.Add(time.Duration(i)*time.Second))
+		if decision.Switch {
+			t.Fatalf("candidate below minimum advantage should not switch after repeated observations: %+v", decision)
+		}
+	}
+	if m.decisions[key].AppliedExitNodeID != 30 {
+		t.Fatalf("expected applied exit to remain 30, got %+v", m.decisions[key])
+	}
+}
+
+func TestBestExitDecisionSwitchesWithMinimumAdvantage(t *testing.T) {
+	m := newBestExitManager()
+	key := bestExitOwnerKey{TunnelID: 7, OwnerNodeID: 10}
+	now := time.Unix(100, 0)
+	current := scoreBestExitCandidate(10, chainNodeRecord{NodeID: 30}, 100, 0, 100, 0)
+	candidate := scoreBestExitCandidate(10, chainNodeRecord{NodeID: 31}, 40, 0, 60, 0)
+
+	m.setApplied(key, 30, now.Add(-time.Minute))
+	for i := 0; i < bestExitConfirmationRounds-1; i++ {
+		decision := m.observeScores(key, []bestExitCandidateScore{candidate, current}, now.Add(time.Duration(i)*time.Second))
+		if decision.Switch {
+			t.Fatalf("candidate should wait for confirmations before switching: %+v", decision)
+		}
+	}
+	decision := m.observeScores(key, []bestExitCandidateScore{candidate, current}, now.Add((bestExitConfirmationRounds-1)*time.Second))
+	if !decision.Switch || decision.ExitNodeID != 31 {
+		t.Fatalf("candidate with enough advantage should switch after confirmations: %+v", decision)
+	}
+}
+
+func TestBestExitConfirmedSwitchDoesNotMarkAppliedUntilSetApplied(t *testing.T) {
+	m := newBestExitManager()
+	key := bestExitOwnerKey{TunnelID: 7, OwnerNodeID: 10}
+	now := time.Unix(100, 0)
+	current := scoreBestExitCandidate(10, chainNodeRecord{NodeID: 30}, 100, 0, 100, 0)
+	candidate := scoreBestExitCandidate(10, chainNodeRecord{NodeID: 31}, 40, 0, 60, 0)
+
+	m.setApplied(key, 30, now.Add(-time.Minute))
+	for i := 0; i < bestExitConfirmationRounds-1; i++ {
+		decision := m.observeScores(key, []bestExitCandidateScore{candidate, current}, now.Add(time.Duration(i)*time.Second))
+		if decision.Switch {
+			t.Fatalf("candidate should wait for confirmations before switching: %+v", decision)
+		}
+	}
+	decision := m.observeScores(key, []bestExitCandidateScore{candidate, current}, now.Add((bestExitConfirmationRounds-1)*time.Second))
+	if !decision.Switch || decision.ExitNodeID != 31 {
+		t.Fatalf("candidate with enough advantage should switch after confirmations: %+v", decision)
+	}
+	if m.decisions[key].AppliedExitNodeID != 30 {
+		t.Fatalf("confirmed switch should not mark applied before runtime update: %+v", m.decisions[key])
+	}
+
+	m.setApplied(key, decision.ExitNodeID, now.Add(time.Second))
+	if m.decisions[key].AppliedExitNodeID != 31 {
+		t.Fatalf("setApplied should commit confirmed switch: %+v", m.decisions[key])
+	}
+}
+
+func TestBestExitApplyFailureStartsRetryCooldownWithoutChangingAppliedExit(t *testing.T) {
+	m := newBestExitManager()
+	key := bestExitOwnerKey{TunnelID: 7, OwnerNodeID: 10}
+	now := time.Unix(100, 0)
+	current := scoreBestExitCandidate(10, chainNodeRecord{NodeID: 30}, 100, 0, 100, 0)
+	candidate := scoreBestExitCandidate(10, chainNodeRecord{NodeID: 31}, 40, 0, 60, 0)
+
+	m.setApplied(key, 30, now.Add(-time.Minute))
+	for i := 0; i < bestExitConfirmationRounds-1; i++ {
+		decision := m.observeScores(key, []bestExitCandidateScore{candidate, current}, now.Add(time.Duration(i)*time.Second))
+		if decision.Switch {
+			t.Fatalf("candidate should wait for confirmations before switching: %+v", decision)
+		}
+	}
+	confirmed := m.observeScores(key, []bestExitCandidateScore{candidate, current}, now.Add((bestExitConfirmationRounds-1)*time.Second))
+	if !confirmed.Switch || confirmed.ExitNodeID != 31 {
+		t.Fatalf("expected confirmed switch before apply failure: %+v", confirmed)
+	}
+
+	m.recordApplyFailure(key, confirmed.ExitNodeID, now.Add(bestExitConfirmationRounds*time.Second))
+	if m.decisions[key].AppliedExitNodeID != 30 {
+		t.Fatalf("apply failure should leave applied exit unchanged: %+v", m.decisions[key])
+	}
+
+	decision := m.observeScores(key, []bestExitCandidateScore{candidate, current}, now.Add((bestExitConfirmationRounds+1)*time.Second))
+	if decision.Switch {
+		t.Fatalf("apply retry cooldown should suppress immediate retry: %+v", decision)
+	}
+	if decision.Reason != "apply retry cooldown" {
+		t.Fatalf("expected apply retry cooldown reason, got %q", decision.Reason)
+	}
+
+	retry := m.observeScores(key, []bestExitCandidateScore{candidate, current}, now.Add(bestExitConfirmationRounds*time.Second+bestExitApplyRetryCooldown))
+	if !retry.Switch || retry.ExitNodeID != 31 {
+		t.Fatalf("expected retry after apply cooldown: %+v", retry)
+	}
+}
+
+func TestBestExitEnsureAppliedDoesNotOverrideExistingAppliedExit(t *testing.T) {
+	m := newBestExitManager()
+	key := bestExitOwnerKey{TunnelID: 7, OwnerNodeID: 10}
+	now := time.Unix(100, 0)
+
+	m.ensureApplied(key, 30, now)
+	if m.decisions[key].AppliedExitNodeID != 30 {
+		t.Fatalf("expected initial applied exit 30, got %+v", m.decisions[key])
+	}
+	if !m.decisions[key].LastSwitchAt.Equal(now) {
+		t.Fatalf("expected initial applied timestamp, got %+v", m.decisions[key])
+	}
+
+	m.ensureApplied(key, 31, now.Add(time.Minute))
+	if m.decisions[key].AppliedExitNodeID != 30 {
+		t.Fatalf("ensureApplied should not override existing applied exit: %+v", m.decisions[key])
+	}
+}
+
+func TestBestExitRoundPingerCachesPublicProbeOnly(t *testing.T) {
+	publicCalls := 0
+	ownerCalls := 0
+	pinger := newBestExitRoundPinger(func(nodeID int64, ip string, port int, _ diagnosisExecOptions) (float64, float64, error) {
+		if ip == bestExitPublicTargetHost && port == bestExitPublicTargetPort {
+			publicCalls++
+			return float64(nodeID), 0, nil
+		}
+		ownerCalls++
+		return float64(ownerCalls), 0, nil
+	})
+
+	if lat, _, err := pinger(30, bestExitPublicTargetHost, bestExitPublicTargetPort, diagnosisExecOptions{}); err != nil || lat != 30 {
+		t.Fatalf("unexpected first public ping result lat=%v err=%v", lat, err)
+	}
+	if lat, _, err := pinger(30, bestExitPublicTargetHost, bestExitPublicTargetPort, diagnosisExecOptions{}); err != nil || lat != 30 {
+		t.Fatalf("unexpected cached public ping result lat=%v err=%v", lat, err)
+	}
+	if _, _, err := pinger(31, bestExitPublicTargetHost, bestExitPublicTargetPort, diagnosisExecOptions{}); err != nil {
+		t.Fatalf("unexpected second exit public ping err=%v", err)
+	}
+	if publicCalls != 2 {
+		t.Fatalf("expected public probes cached per exit node, got %d calls", publicCalls)
+	}
+
+	if _, _, err := pinger(10, "10.0.0.30", 30030, diagnosisExecOptions{}); err != nil {
+		t.Fatalf("unexpected owner ping err=%v", err)
+	}
+	if _, _, err := pinger(10, "10.0.0.30", 30030, diagnosisExecOptions{}); err != nil {
+		t.Fatalf("unexpected repeated owner ping err=%v", err)
+	}
+	if ownerCalls != 2 {
+		t.Fatalf("expected owner-to-exit probes not cached, got %d calls", ownerCalls)
+	}
+}
+
+func TestBestExitDecisionScoresAreDefensiveCopies(t *testing.T) {
+	m := newBestExitManager()
+	key := bestExitOwnerKey{TunnelID: 7, OwnerNodeID: 10}
+	now := time.Unix(100, 0)
+	candidate := scoreBestExitCandidate(10, chainNodeRecord{NodeID: 31}, 40, 0, 60, 0)
+	current := scoreBestExitCandidate(10, chainNodeRecord{NodeID: 30}, 100, 0, 100, 0)
+
+	decision := m.observeScores(key, []bestExitCandidateScore{candidate, current}, now)
+	decision.Scores[0].ExitNodeID = 99
+
+	if m.decisions[key].Scores[0].ExitNodeID != 31 {
+		t.Fatalf("decision scores mutation leaked into manager state: %+v", m.decisions[key].Scores)
+	}
+}
+
+func TestBestExitDecisionRequiresConfirmationsAndCooldown(t *testing.T) {
+	m := newBestExitManager()
+	key := bestExitOwnerKey{TunnelID: 7, OwnerNodeID: 10}
+	now := time.Unix(100, 0)
+	current := scoreBestExitCandidate(10, chainNodeRecord{NodeID: 30}, 100, 0, 100, 0)
+	candidate := scoreBestExitCandidate(10, chainNodeRecord{NodeID: 31}, 40, 0, 60, 0)
+
+	m.setApplied(key, 30, now.Add(-time.Minute))
+
+	if decision := m.observeScores(key, []bestExitCandidateScore{candidate, current}, now); decision.Switch {
+		t.Fatalf("first observation should not switch: %+v", decision)
+	}
+	if decision := m.observeScores(key, []bestExitCandidateScore{candidate, current}, now.Add(time.Second)); decision.Switch {
+		t.Fatalf("second observation should not switch: %+v", decision)
+	}
+	decision := m.observeScores(key, []bestExitCandidateScore{candidate, current}, now.Add(2*time.Second))
+	if !decision.Switch || decision.ExitNodeID != 31 {
+		t.Fatalf("third confirmed observation should switch to 31: %+v", decision)
+	}
+
+	betterAgain := scoreBestExitCandidate(10, chainNodeRecord{NodeID: 30}, 20, 0, 20, 0)
+	if decision := m.observeScores(key, []bestExitCandidateScore{betterAgain, candidate}, now.Add(3*time.Second)); decision.Switch {
+		t.Fatalf("cooldown should block immediate switch back: %+v", decision)
+	}
+}
+
+func TestBestExitOrderingUsesAppliedDecision(t *testing.T) {
+	m := newBestExitManager()
+	key := bestExitOwnerKey{TunnelID: 7, OwnerNodeID: 10}
+	m.setApplied(key, 31, time.Unix(100, 0))
+	targets := []tunnelRuntimeNode{
+		{NodeID: 30, Strategy: tunnelStrategyBest},
+		{NodeID: 31, Strategy: tunnelStrategyBest},
+		{NodeID: 32, Strategy: tunnelStrategyBest},
+	}
+
+	ordered := m.orderTargets(key, targets)
+	if ordered[0].NodeID != 31 || ordered[1].NodeID != 30 || ordered[2].NodeID != 32 {
+		t.Fatalf("unexpected order: %+v", ordered)
+	}
+	if targets[0].NodeID != 30 {
+		t.Fatalf("orderTargets mutated input: %+v", targets)
+	}
+}
+
+func TestBuildTunnelChainConfigMapsBestStrategyToFIFO(t *testing.T) {
+	nodes := map[int64]*nodeRecord{
+		10: {ID: 10, ServerIP: "10.0.0.10", ServerIPv4: "10.0.0.10", TCPListenAddr: "[::]"},
+		30: {ID: 30, ServerIP: "10.0.0.30", ServerIPv4: "10.0.0.30", TCPListenAddr: "[::]"},
+		31: {ID: 31, ServerIP: "10.0.0.31", ServerIPv4: "10.0.0.31", TCPListenAddr: "[::]"},
+	}
+	targets := []tunnelRuntimeNode{
+		{NodeID: 30, Port: 30030, Protocol: "tls", Strategy: tunnelStrategyBest, ChainType: 3},
+		{NodeID: 31, Port: 30031, Protocol: "tls", Strategy: tunnelStrategyBest, ChainType: 3},
+	}
+
+	chainData, err := buildTunnelChainConfig(77, 10, targets, nodes, "")
+	if err != nil {
+		t.Fatalf("build chain: %v", err)
+	}
+	hops := chainData["hops"].([]map[string]interface{})
+	selector := hops[0]["selector"].(map[string]interface{})
+	if selector["strategy"] != bestExitRuntimeStrategy {
+		t.Fatalf("expected best to render as fifo, got %v", selector["strategy"])
+	}
+}
+
+func TestHandlerOrdersBestExitTargetsForOwner(t *testing.T) {
+	h := &Handler{bestExit: newBestExitManager()}
+	key := bestExitOwnerKey{TunnelID: 77, OwnerNodeID: 10}
+	h.bestExit.setApplied(key, 31, time.Unix(100, 0))
+	targets := []tunnelRuntimeNode{
+		{NodeID: 30, Port: 30030, Strategy: tunnelStrategyBest},
+		{NodeID: 31, Port: 30031, Strategy: tunnelStrategyBest},
+	}
+
+	ordered := h.orderBestExitTargets(77, 10, targets)
+	if ordered[0].NodeID != 31 || ordered[1].NodeID != 30 {
+		t.Fatalf("unexpected ordered targets: %+v", ordered)
+	}
+}
+
+func TestRuntimeStrategyForTargetsMapsBestTargetStrategyToFIFO(t *testing.T) {
+	owner := tunnelRuntimeNode{Strategy: "round"}
+	targets := []tunnelRuntimeNode{{Strategy: tunnelStrategyBest}}
+
+	if got := runtimeStrategyForTargets(owner, targets); got != bestExitRuntimeStrategy {
+		t.Fatalf("expected best target strategy to map to fifo, got %q", got)
+	}
+}
+
+func TestRuntimeStrategyForTargetsPreservesNonBestTargetStrategy(t *testing.T) {
+	owner := tunnelRuntimeNode{Strategy: tunnelStrategyBest}
+	targets := []tunnelRuntimeNode{{Strategy: "round"}}
+
+	if got := runtimeStrategyForTargets(owner, targets); got != "round" {
+		t.Fatalf("expected target strategy round to remain unchanged, got %q", got)
+	}
+}
+
+func TestRuntimeStrategyForTargetsMapsBestOwnerStrategyWhenTargetsEmpty(t *testing.T) {
+	owner := tunnelRuntimeNode{Strategy: tunnelStrategyBest}
+
+	if got := runtimeStrategyForTargets(owner, nil); got != bestExitRuntimeStrategy {
+		t.Fatalf("expected best owner fallback strategy to map to fifo, got %q", got)
+	}
+}
+
+func TestEvaluateBestExitOwnerScoresAllCandidates(t *testing.T) {
+	owner := chainNodeRecord{NodeID: 10, NodeName: "entry"}
+	exits := []chainNodeRecord{
+		{NodeID: 30, NodeName: "exit-a", Port: 30030},
+		{NodeID: 31, NodeName: "exit-b", Port: 30031},
+	}
+	nodes := map[int64]*nodeRecord{
+		10: {ID: 10, ServerIP: "10.0.0.10", ServerIPv4: "10.0.0.10", TCPListenAddr: "[::]"},
+		30: {ID: 30, ServerIP: "10.0.0.30", ServerIPv4: "10.0.0.30", TCPListenAddr: "[::]"},
+		31: {ID: 31, ServerIP: "10.0.0.31", ServerIPv4: "10.0.0.31", TCPListenAddr: "[::]"},
+	}
+	pinger := func(nodeID int64, ip string, port int, _ diagnosisExecOptions) (float64, float64, error) {
+		switch {
+		case nodeID == 10 && port == 30030:
+			return 60, 0, nil
+		case nodeID == 10 && port == 30031:
+			return 20, 0, nil
+		case nodeID == 30 && ip == bestExitPublicTargetHost:
+			return 60, 0, nil
+		case nodeID == 31 && ip == bestExitPublicTargetHost:
+			return 20, 0, nil
+		default:
+			t.Fatalf("unexpected ping node=%d ip=%s port=%d", nodeID, ip, port)
+			return 0, 100, nil
+		}
+	}
+
+	scores := evaluateBestExitOwner(owner, exits, nodes, "", diagnosisExecOptions{}, pinger)
+	if len(scores) != 2 {
+		t.Fatalf("expected two scores, got %+v", scores)
+	}
+	if scores[0].ExitNodeID != 31 {
+		t.Fatalf("expected exit-b first, got %+v", scores)
+	}
+}
+
+func TestEvaluateBestExitOwnerMarksCandidateFailedWhenOwnerToExitFails(t *testing.T) {
+	owner := chainNodeRecord{NodeID: 10, NodeName: "entry"}
+	exits := []chainNodeRecord{{NodeID: 30, NodeName: "exit-a", Port: 30030}}
+	nodes := map[int64]*nodeRecord{
+		10: {ID: 10, ServerIP: "10.0.0.10", ServerIPv4: "10.0.0.10", TCPListenAddr: "[::]"},
+		30: {ID: 30, ServerIP: "10.0.0.30", ServerIPv4: "10.0.0.30", TCPListenAddr: "[::]"},
+	}
+	pinger := func(nodeID int64, ip string, port int, _ diagnosisExecOptions) (float64, float64, error) {
+		return 0, 100, errBestExitProbeForTest
+	}
+
+	scores := evaluateBestExitOwner(owner, exits, nodes, "", diagnosisExecOptions{}, pinger)
+	if len(scores) != 1 || scores[0].Success {
+		t.Fatalf("expected failed candidate, got %+v", scores)
+	}
+}
+
+func TestEvaluateBestExitOwnerMarksCandidateFailedWhenTargetResolutionFails(t *testing.T) {
+	owner := chainNodeRecord{NodeID: 10, NodeName: "entry"}
+	exits := []chainNodeRecord{{NodeID: 30, NodeName: "exit-v6", Port: 30030}}
+	nodes := map[int64]*nodeRecord{
+		10: {ID: 10, Name: "entry", ServerIP: "10.0.0.10", ServerIPv4: "10.0.0.10", TCPListenAddr: "[::]"},
+		30: {ID: 30, Name: "exit-v6", ServerIP: "2001:db8::30", ServerIPv6: "2001:db8::30", TCPListenAddr: "[::]"},
+	}
+	pinger := func(nodeID int64, ip string, port int, _ diagnosisExecOptions) (float64, float64, error) {
+		t.Fatalf("ping should not be called when target resolution fails: node=%d ip=%s port=%d", nodeID, ip, port)
+		return 0, 100, nil
+	}
+
+	scores := evaluateBestExitOwner(owner, exits, nodes, "v4", diagnosisExecOptions{}, pinger)
+	if len(scores) != 1 || scores[0].Success {
+		t.Fatalf("expected failed candidate, got %+v", scores)
+	}
+}

go-backend/internal/http/handler/tunnel_quality_prober.go

@@ -242,6 +242,7 @@ func (p *tunnelQualityProber) probeTunnel(tunnelID int64) {
 		pingTimeoutMS:  tunnelQualityPingTimeoutMs,
 		timeoutMessage: "探测超时",
 	}
+	p.probeBestExitOwners(tunnelID, inNodes, midNodesGrouped, outNodes, ipPreference, options)
 
 	switch tunnel.Type {
 	case 1:
@@ -375,6 +376,51 @@ func (p *tunnelQualityProber) probeTunnel(tunnelID int64) {
 	p.storeResult(snap)
 }
 
+func (p *tunnelQualityProber) probeBestExitOwners(tunnelID int64, inNodes []chainNodeRecord, chainHops [][]chainNodeRecord, outNodes []chainNodeRecord, ipPreference string, options diagnosisExecOptions) {
+	if p == nil || p.handler == nil || p.handler.bestExit == nil || len(outNodes) <= 1 {
+		return
+	}
+	if !isBestTunnelStrategy(outNodes[0].Strategy) {
+		return
+	}
+	owners := bestExitChainOwners(inNodes, chainHops)
+	if len(owners) == 0 {
+		return
+	}
+	nodeMap := make(map[int64]*nodeRecord, len(owners)+len(outNodes))
+	for _, owner := range owners {
+		if node, err := p.handler.getNodeRecord(owner.NodeID); err == nil && node != nil {
+			nodeMap[owner.NodeID] = node
+		}
+	}
+	for _, exit := range outNodes {
+		if node, err := p.handler.getNodeRecord(exit.NodeID); err == nil && node != nil {
+			nodeMap[exit.NodeID] = node
+		}
+	}
+	// This best-exit decision cache is per decision round; the display-oriented
+	// tunnel quality snapshot may still collect its own first-exit public probe.
+	roundPinger := newBestExitRoundPinger(p.tcpPingNode)
+	for _, owner := range owners {
+		if nodeMap[owner.NodeID] == nil {
+			continue
+		}
+		key := bestExitOwnerKey{TunnelID: tunnelID, OwnerNodeID: owner.NodeID}
+		p.handler.bestExit.ensureApplied(key, outNodes[0].NodeID, time.Now())
+		scores := evaluateBestExitOwner(owner, outNodes, nodeMap, ipPreference, options, roundPinger)
+		decision := p.handler.bestExit.observeScores(key, scores, time.Now())
+		if decision.Switch {
+			now := time.Now()
+			if err := p.handler.applyBestExitChainOrder(tunnelID, owner.NodeID, outNodes, decision.Scores, ipPreference); err != nil {
+				log.Printf("best_exit: switch apply failed tunnel=%d owner=%d exit=%d err=%v", tunnelID, owner.NodeID, decision.ExitNodeID, err)
+				p.handler.bestExit.recordApplyFailure(key, decision.ExitNodeID, now)
+				continue
+			}
+			p.handler.bestExit.setApplied(key, decision.ExitNodeID, time.Now())
+		}
+	}
+}
+
 func (p *tunnelQualityProber) tcpPingNode(nodeID int64, ip string, port int, options diagnosisExecOptions) (latency float64, loss float64, err error) {
 	h := p.handler
 	if h == nil {

go-gost/x/registry/chain.go

@@ -12,10 +12,25 @@ type chainRegistry struct {
 	registry[chain.Chainer]
 }
 
+func ReplaceChain(name string, v chain.Chainer) error {
+	if name == "" {
+		return nil
+	}
+	if r, ok := chainReg.(*chainRegistry); ok {
+		r.replace(name, v)
+		return nil
+	}
+	return chainReg.Register(name, v)
+}
+
 func (r *chainRegistry) Register(name string, v chain.Chainer) error {
 	return r.registry.Register(name, v)
 }
 
+func (r *chainRegistry) replace(name string, v chain.Chainer) {
+	r.m.Store(name, v)
+}
+
 func (r *chainRegistry) Get(name string) chain.Chainer {
 	if name != "" {
 		return &chainWrapper{name: name, r: r}

go-gost/x/registry/chain_test.go

@@ -0,0 +1,51 @@
+package registry
+
+import (
+	"context"
+	"net"
+	"testing"
+
+	"github.com/go-gost/core/chain"
+)
+
+type testChainer struct {
+	route chain.Route
+}
+
+func (c testChainer) Route(context.Context, string, string, ...chain.RouteOption) chain.Route {
+	return c.route
+}
+
+type testRoute struct {
+	nodes []*chain.Node
+}
+
+func (r testRoute) Dial(context.Context, string, string, ...chain.DialOption) (net.Conn, error) {
+	return nil, nil
+}
+
+func (r testRoute) Bind(context.Context, string, string, ...chain.BindOption) (net.Listener, error) {
+	return nil, nil
+}
+
+func (r testRoute) Nodes() []*chain.Node {
+	return r.nodes
+}
+
+func TestReplaceChainOverwritesExistingRegistration(t *testing.T) {
+	name := "replace_chain_tdd"
+	ChainRegistry().Unregister(name)
+	defer ChainRegistry().Unregister(name)
+
+	if err := ChainRegistry().Register(name, testChainer{route: testRoute{nodes: []*chain.Node{{Name: "old"}}}}); err != nil {
+		t.Fatalf("register old chain: %v", err)
+	}
+	if err := ReplaceChain(name, testChainer{route: testRoute{nodes: []*chain.Node{{Name: "new"}}}}); err != nil {
+		t.Fatalf("replace chain: %v", err)
+	}
+
+	route := ChainRegistry().Get(name).Route(context.Background(), "tcp", "example.com:443")
+	if route == nil || len(route.Nodes()) != 1 || route.Nodes()[0].Name != "new" {
+		t.Fatalf("expected replacement chain route, got %#v", route)
+	}
+}

go-gost/x/socket/chain.go

@@ -38,21 +38,21 @@ func createChain(req createChainRequest) error {
 }
 
 func updateChain(req updateChainRequest) error {
-
 	name := strings.TrimSpace(req.Chain)
-
-	if registry.ChainRegistry().IsRegistered(name) {
-		registry.ChainRegistry().Unregister(name)
+	if name == "" {
+		name = strings.TrimSpace(req.Data.Name)
+	}
+	if name == "" {
+		return errors.New("chain name is required")
 	}
 
 	req.Data.Name = name
-
 	v, err := parser.ParseChain(&req.Data, logger.Default())
 	if err != nil {
 		return errors.New("create chain " + name + " failed: " + err.Error())
 	}
 
-	if err := registry.ChainRegistry().Register(name, v); err != nil {
+	if err := registry.ReplaceChain(name, v); err != nil {
 		return errors.New("chain " + name + " already exists")
 	}
 

go-gost/x/socket/chain_test.go

@@ -0,0 +1,69 @@
+package socket
+
+import (
+	"testing"
+
+	corelogger "github.com/go-gost/core/logger"
+	"github.com/go-gost/x/config"
+	_ "github.com/go-gost/x/connector/relay"
+	_ "github.com/go-gost/x/dialer/tcp"
+	xlogger "github.com/go-gost/x/logger"
+	"github.com/go-gost/x/registry"
+)
+
+func TestUpdateChainParseFailureKeepsExistingChainRegistered(t *testing.T) {
+	corelogger.SetDefault(xlogger.Nop())
+
+	name := "chain_update_parse_failure_tdd"
+	originalConfig := config.Global()
+	defer config.Set(originalConfig)
+	registry.ChainRegistry().Unregister(name)
+	defer registry.ChainRegistry().Unregister(name)
+	config.Set(&config.Config{})
+
+	valid := config.ChainConfig{
+		Name: name,
+		Hops: []*config.HopConfig{{
+			Name: "hop-valid",
+			Nodes: []*config.NodeConfig{{
+				Name:      "node-valid",
+				Addr:      "127.0.0.1:443",
+				Connector: &config.ConnectorConfig{Type: "relay"},
+				Dialer:    &config.DialerConfig{Type: "tcp"},
+			}},
+		}},
+	}
+	if err := createChain(createChainRequest{Data: valid}); err != nil {
+		t.Fatalf("create valid chain: %v", err)
+	}
+	before := registry.ChainRegistry().Get(name)
+	if before == nil || !registry.ChainRegistry().IsRegistered(name) {
+		t.Fatalf("expected chain registered before update")
+	}
+
+	invalid := config.ChainConfig{
+		Hops: []*config.HopConfig{{
+			Name: "hop-invalid",
+			Nodes: []*config.NodeConfig{{
+				Name:      "node-invalid",
+				Addr:      "127.0.0.1:443",
+				Connector: &config.ConnectorConfig{Type: "connector-does-not-exist"},
+				Dialer:    &config.DialerConfig{Type: "tcp"},
+			}},
+		}},
+	}
+	err := updateChain(updateChainRequest{Chain: name, Data: invalid})
+	if err == nil {
+		t.Fatalf("expected invalid chain update to fail")
+	}
+	if !registry.ChainRegistry().IsRegistered(name) {
+		t.Fatalf("expected old chain to remain registered after failed update")
+	}
+	cfg := config.Global()
+	if len(cfg.Chains) != 1 || cfg.Chains[0] == nil || cfg.Chains[0].Name != name {
+		t.Fatalf("expected original chain config to remain, got %#v", cfg.Chains)
+	}
+	if got := cfg.Chains[0].Hops[0].Name; got != "hop-valid" {
+		t.Fatalf("expected original chain config to remain, got hop %q", got)
+	}
+}

vite-frontend/src/pages/tunnel.tsx

@@ -91,7 +91,7 @@ import {
 interface ChainTunnel {
   nodeId: number;
   protocol?: string; // 'tls' | 'wss' | 'tcp' | 'mtls' | 'mwss' | 'mtcp' | 'kcp' - 转发链协议
-  strategy?: string; // 'fifo' | 'round' | 'rand' - 仅转发链需要
+  strategy?: string; // 'fifo' | 'round' | 'rand' | 'best' - 仅转发链/多出口需要
   chainType?: number; // 1: 入口, 2: 转发链, 3: 出口
   inx?: number; // 转发链序号
   connectIp?: string; // 连接IP（多IP节点指定连接地址）
@@ -2901,6 +2901,7 @@ export default function TunnelPage() {
                                 <SelectItem key="fifo">主备</SelectItem>
                                 <SelectItem key="round">轮询</SelectItem>
                                 <SelectItem key="rand">随机</SelectItem>
+                                <SelectItem key="best">最优</SelectItem>
                               </Select>
                             </div>